Ensure a correct order between memory accesses.

Ensure that the scalar write access corresponds to the result of a load
instruction appears after the generic read access corresponds to the load
instruction.

llvm-svn: 186419

diff --git a/polly/include/polly/TempScopInfo.h b/polly/include/polly/TempScopInfo.h
index da84a71..56c72bf 100755 (executable)
--- a/polly/include/polly/TempScopInfo.h
+++ b/polly/include/polly/TempScopInfo.h
@@ -292,7 +292,10 @@ class TempScopInfo : public FunctionPass {
   ///
   /// @param Inst The instruction to be analyzed
   /// @param R    The SCoP region
-  void buildScalarDependences(Instruction *Inst, Region *R);
+  ///
+  /// @return     True if the Instruction is used in other BB and a scalar write
+  ///             Access is required.
+  bool buildScalarDependences(Instruction *Inst, Region *R);
 
   void buildAccessFunctions(Region &RefRegion, BasicBlock &BB);
 

diff --git a/polly/lib/Analysis/TempScopInfo.cpp b/polly/lib/Analysis/TempScopInfo.cpp
index 4a9733f..5aa8d96 100644 (file)
--- a/polly/lib/Analysis/TempScopInfo.cpp
+++ b/polly/lib/Analysis/TempScopInfo.cpp
@@ -104,11 +104,11 @@ void TempScop::printDetail(raw_ostream &OS, ScalarEvolution *SE, LoopInfo *LI,
   }
 }
 
-void TempScopInfo::buildScalarDependences(Instruction *Inst, Region *R) {
+bool TempScopInfo::buildScalarDependences(Instruction *Inst, Region *R) {
   // No need to translate these scalar dependences into polyhedral form, because
   // synthesizable scalars can be generated by the code generator.
   if (canSynthesize(Inst, LI, SE, R))
-    return;
+    return false;
 
   bool AnyCrossStmtUse = false;
   BasicBlock *ParentBB = Inst->getParent();
@@ -147,12 +147,7 @@ void TempScopInfo::buildScalarDependences(Instruction *Inst, Region *R) {
     AccFuncMap[UseParent].push_back(std::make_pair(ScalarAccess, U));
   }
 
-  // If the Instruction is used outside the statement, we need to build the
-  // write access.
-  if (AnyCrossStmtUse) {
-    IRAccess ScalarAccess(IRAccess::SCALARWRITE, Inst, ZeroOffset, 1, true);
-    AccFuncMap[ParentBB].push_back(std::make_pair(ScalarAccess, Inst));
-  }
+  return AnyCrossStmtUse;
 }
 
 IRAccess TempScopInfo::buildIRAccess(Instruction *Inst, Loop *L, Region *R) {
@@ -190,8 +185,12 @@ void TempScopInfo::buildAccessFunctions(Region &R, BasicBlock &BB) {
     if (isa<LoadInst>(Inst) || isa<StoreInst>(Inst))
       Functions.push_back(std::make_pair(buildIRAccess(Inst, L, &R), Inst));
 
-    if (!isa<StoreInst>(Inst))
-      buildScalarDependences(Inst, &R);
+    if (!isa<StoreInst>(Inst) && buildScalarDependences(Inst, &R)) {
+      // If the Instruction is used outside the statement, we need to build the
+      // write access.
+      IRAccess ScalarAccess(IRAccess::SCALARWRITE, Inst, ZeroOffset, 1, true);
+      Functions.push_back(std::make_pair(ScalarAccess, Inst));
+    }
   }
 
   if (Functions.empty())

diff --git a/polly/test/TempScop/inter_bb_scalar_dep.ll b/polly/test/TempScop/inter_bb_scalar_dep.ll
new file mode 100644 (file)
index 0000000..e83bf5b
--- /dev/null
+++ b/polly/test/TempScop/inter_bb_scalar_dep.ll
@@ -0,0 +1,54 @@
+; RUN: opt %loadPolly -basicaa -polly-analyze-ir -disable-polly-intra-scop-scalar-to-array -analyze < %s | FileCheck %s
+; RUN: opt %loadPolly -basicaa -polly-analyze-ir  -polly-codegen-scev -disable-polly-intra-scop-scalar-to-array -analyze < %s | FileCheck %s
+
+; void f(long A[], int N, int *init_ptr) {
+;   long i, j;
+;
+;   for (i = 0; i < N; ++i) {
+;     init = *init_ptr;
+;     for (i = 0; i < N; ++i) {
+;       A[i] = init + 2;
+;     }
+;   }
+; }
+
+target datalayout = "e-p:64:64:64-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:64:64-f32:32:32-f64:64:64-v64:64:64-v128:128:128-a0:0:64-s0:64:64-f80:128:128"
+target triple = "x86_64-unknown-linux-gnu"
+
+; Function Attrs: nounwind
+define void @f(i64* noalias %A, i64 %N, i64* noalias %init_ptr) #0 {
+entry:
+  br label %for.i
+
+for.i:                                            ; preds = %for.i.end, %entry
+  %indvar.i = phi i64 [ 0, %entry ], [ %indvar.i.next, %for.i.end ]
+  %indvar.i.next = add nsw i64 %indvar.i, 1
+  br label %entry.next
+
+entry.next:                                       ; preds = %for.i
+  %init = load i64* %init_ptr
+; CHECK: BB: entry.next
+; CHECK: Read init_ptr[0]
+; CHECK: Write init[0]
+  br label %for.j
+
+for.j:                                            ; preds = %for.j, %entry.next
+  %indvar.j = phi i64 [ 0, %entry.next ], [ %indvar.j.next, %for.j ]
+  %init_plus_two = add i64 %init, 2
+; CHECK: Read init[0]
+; CHECK: Write A[{0,+,8}<%for.j>]
+  %scevgep = getelementptr i64* %A, i64 %indvar.j
+  store i64 %init_plus_two, i64* %scevgep
+  %indvar.j.next = add nsw i64 %indvar.j, 1
+  %exitcond.j = icmp eq i64 %indvar.j.next, %N
+  br i1 %exitcond.j, label %for.i.end, label %for.j
+
+for.i.end:                                        ; preds = %for.j
+  %exitcond.i = icmp eq i64 %indvar.i.next, %N
+  br i1 %exitcond.i, label %return, label %for.i
+
+return:                                           ; preds = %for.i.end
+  ret void
+}
+
+attributes #0 = { nounwind }

diff --git a/polly/test/TempScop/intra_and_inter_bb_scalar_dep.ll b/polly/test/TempScop/intra_and_inter_bb_scalar_dep.ll
new file mode 100644 (file)
index 0000000..ce4c49f
--- /dev/null
+++ b/polly/test/TempScop/intra_and_inter_bb_scalar_dep.ll
@@ -0,0 +1,58 @@
+; RUN: opt %loadPolly -basicaa -polly-analyze-ir -disable-polly-intra-scop-scalar-to-array -analyze < %s | FileCheck %s
+; RUN: opt %loadPolly -basicaa -polly-analyze-ir -polly-codegen-scev -disable-polly-intra-scop-scalar-to-array -analyze < %s | FileCheck %s
+
+; void f(long A[], int N, int *init_ptr) {
+;   long i, j;
+;
+;   for (i = 0; i < N; ++i) {
+;     init = *init_ptr;
+;     for (i = 0; i < N; ++i) {
+;       init2 = *init_ptr;
+;       A[i] = init + init2;
+;     }
+;   }
+; }
+
+target datalayout = "e-p:64:64:64-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:64:64-f32:32:32-f64:64:64-v64:64:64-v128:128:128-a0:0:64-s0:64:64-f80:128:128"
+target triple = "x86_64-unknown-linux-gnu"
+
+; Function Attrs: nounwind
+define void @f(i64* noalias %A, i64 %N, i64* noalias %init_ptr) #0 {
+entry:
+  br label %for.i
+
+for.i:                                            ; preds = %for.i.end, %entry
+  %indvar.i = phi i64 [ 0, %entry ], [ %indvar.i.next, %for.i.end ]
+  %indvar.i.next = add nsw i64 %indvar.i, 1
+  br label %entry.next
+
+entry.next:                                       ; preds = %for.i
+  %init = load i64* %init_ptr
+; CHECK: BB: entry.next
+; CHECK: Read init_ptr[0]
+; CHECK: Write init[0]
+  br label %for.j
+
+for.j:                                            ; preds = %for.j, %entry.next
+  %indvar.j = phi i64 [ 0, %entry.next ], [ %indvar.j.next, %for.j ]
+  %init_2 = load i64* %init_ptr
+  %init_sum = add i64 %init, %init_2
+; CHECK: BB: for.j
+; CHECK: Read init[0]
+; CHECK: Read init_ptr[0]
+; CHECK: Write A[{0,+,8}<%for.j>]
+  %scevgep = getelementptr i64* %A, i64 %indvar.j
+  store i64 %init_sum, i64* %scevgep
+  %indvar.j.next = add nsw i64 %indvar.j, 1
+  %exitcond.j = icmp eq i64 %indvar.j.next, %N
+  br i1 %exitcond.j, label %for.i.end, label %for.j
+
+for.i.end:                                        ; preds = %for.j
+  %exitcond.i = icmp eq i64 %indvar.i.next, %N
+  br i1 %exitcond.i, label %return, label %for.i
+
+return:                                           ; preds = %for.i.end
+  ret void
+}
+
+attributes #0 = { nounwind }

diff --git a/polly/test/TempScop/intra_bb_scalar_dep.ll b/polly/test/TempScop/intra_bb_scalar_dep.ll
new file mode 100644 (file)
index 0000000..d6e4211
--- /dev/null
+++ b/polly/test/TempScop/intra_bb_scalar_dep.ll
@@ -0,0 +1,52 @@
+; RUN: opt %loadPolly -basicaa -polly-analyze-ir -disable-polly-intra-scop-scalar-to-array -analyze < %s | FileCheck %s
+; RUN: opt %loadPolly -basicaa -polly-analyze-ir  -polly-codegen-scev -disable-polly-intra-scop-scalar-to-array -analyze < %s | FileCheck %s
+
+; void f(long A[], int N, int *init_ptr) {
+;   long i, j;
+;
+;   for (i = 0; i < N; ++i) {
+;     for (i = 0; i < N; ++i) {
+;       init = *init_ptr;
+;       A[i] = init + 2;
+;     }
+;   }
+; }
+
+target datalayout = "e-p:64:64:64-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:64:64-f32:32:32-f64:64:64-v64:64:64-v128:128:128-a0:0:64-s0:64:64-f80:128:128"
+target triple = "x86_64-unknown-linux-gnu"
+
+; Function Attrs: nounwind
+define void @f(i64* noalias %A, i64 %N, i64* noalias %init_ptr) #0 {
+entry:
+  br label %for.i
+
+for.i:                                            ; preds = %for.i.end, %entry
+  %indvar.i = phi i64 [ 0, %entry ], [ %indvar.i.next, %for.i.end ]
+  %indvar.i.next = add nsw i64 %indvar.i, 1
+  br label %entry.next
+
+entry.next:                                       ; preds = %for.i
+  br label %for.j
+
+for.j:                                            ; preds = %for.j, %entry.next
+  %indvar.j = phi i64 [ 0, %entry.next ], [ %indvar.j.next, %for.j ]
+  %init = load i64* %init_ptr
+  %init_plus_two = add i64 %init, 2
+  %scevgep = getelementptr i64* %A, i64 %indvar.j
+  store i64 %init_plus_two, i64* %scevgep
+; CHECK: BB: for.j
+; CHECK: Read init_ptr[0]
+; CHECK: Write A[{0,+,8}<%for.j>]
+  %indvar.j.next = add nsw i64 %indvar.j, 1
+  %exitcond.j = icmp eq i64 %indvar.j.next, %N
+  br i1 %exitcond.j, label %for.i.end, label %for.j
+
+for.i.end:                                        ; preds = %for.j
+  %exitcond.i = icmp eq i64 %indvar.i.next, %N
+  br i1 %exitcond.i, label %return, label %for.i
+
+return:                                           ; preds = %for.i.end
+  ret void
+}
+
+attributes #0 = { nounwind }

diff --git a/polly/test/TempScop/scalar_to_array.ll b/polly/test/TempScop/scalar_to_array.ll
new file mode 100644 (file)
index 0000000..c2225c4
--- /dev/null
+++ b/polly/test/TempScop/scalar_to_array.ll
@@ -0,0 +1,186 @@
+; RUN: opt %loadPolly -basicaa -polly-analyze-ir -disable-polly-intra-scop-scalar-to-array -analyze < %s | FileCheck %s
+; RUN: opt %loadPolly -basicaa -polly-analyze-ir -disable-polly-intra-scop-scalar-to-array -polly-codegen-scev -analyze < %s | FileCheck %s
+
+; ModuleID = 'scalar_to_array.ll'
+target datalayout = "e-p:64:64:64-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:64:64-f32:32:32-f64:64:64-v64:64:64-v128:128:128-a0:0:64-s0:64:64-f80:128:128-n8:16:32:64"
+target triple = "x86_64-unknown-linux-gnu"
+
+@A = common global [1024 x float] zeroinitializer, align 8
+
+; CHECK: empty
+; Function Attrs: nounwind
+define i32 @empty() #0 {
+entry:
+  fence seq_cst
+  br label %for.cond
+
+for.cond:                                         ; preds = %for.inc, %entry
+  %indvar = phi i64 [ %indvar.next, %for.inc ], [ 0, %entry ]
+  %exitcond = icmp ne i64 %indvar, 1024
+  br i1 %exitcond, label %for.body, label %return
+
+for.body:                                         ; preds = %for.cond
+  br label %for.inc
+
+for.inc:                                          ; preds = %for.body
+  %indvar.next = add i64 %indvar, 1
+  br label %for.cond
+
+return:                                           ; preds = %for.cond
+  fence seq_cst
+  ret i32 0
+}
+
+; CHECK: array_access
+; Function Attrs: nounwind
+define i32 @array_access() #0 {
+entry:
+  fence seq_cst
+  br label %for.cond
+
+for.cond:                                         ; preds = %for.inc, %entry
+  %indvar = phi i64 [ %indvar.next, %for.inc ], [ 0, %entry ]
+  %exitcond = icmp ne i64 %indvar, 1024
+  br i1 %exitcond, label %for.body, label %return
+
+for.body:                                         ; preds = %for.cond
+  %arrayidx = getelementptr [1024 x float]* @A, i64 0, i64 %indvar
+  %float = uitofp i64 %indvar to float
+  store float %float, float* %arrayidx
+  br label %for.inc
+; CHECK: BB: for.body
+; CHECK-NOT: Read
+; CHECK: Write A[{0,+,4}<%for.cond>]
+
+for.inc:                                          ; preds = %for.body
+  %indvar.next = add i64 %indvar, 1
+  br label %for.cond
+
+return:                                           ; preds = %for.cond
+  fence seq_cst
+  ret i32 0
+}
+
+; Function Attrs: nounwind
+; CHECK: intra_scop_dep
+define i32 @intra_scop_dep() #0 {
+entry:
+  fence seq_cst
+  br label %for.cond
+
+for.cond:                                         ; preds = %for.inc, %entry
+  %indvar = phi i64 [ %indvar.next, %for.inc ], [ 0, %entry ]
+  %exitcond = icmp ne i64 %indvar, 1024
+  br i1 %exitcond, label %for.body.a, label %return
+
+for.body.a:                                       ; preds = %for.cond
+  %arrayidx = getelementptr [1024 x float]* @A, i64 0, i64 %indvar
+  %scalar = load float* %arrayidx
+  br label %for.body.b
+; CHECK: BB: for.body.a
+; CHECK: Read A[{0,+,4}<%for.cond>]
+; CHECK: Write scalar[0]
+
+for.body.b:                                       ; preds = %for.body.a
+  %arrayidx2 = getelementptr [1024 x float]* @A, i64 0, i64 %indvar
+  %float = uitofp i64 %indvar to float
+  %sum = fadd float %scalar, %float
+  store float %sum, float* %arrayidx2
+  br label %for.inc
+; CHECK: BB: for.body.b
+; CHECK: Read scalar[0]
+; CHECK: Write A[{0,+,4}<%for.cond>]
+
+for.inc:                                          ; preds = %for.body.b
+  %indvar.next = add i64 %indvar, 1
+  br label %for.cond
+
+return:                                           ; preds = %for.cond
+  fence seq_cst
+  ret i32 0
+}
+
+; It is not possible to have a scop which accesses a scalar element that is
+; a global variable. All global variables are pointers containing possibly
+; a single element. Hence they do not need to be handled anyways.
+; Please note that this is still required when scalar to array rewritting is
+; disabled.
+
+; CHECK: use_after_scop
+; Function Attrs: nounwind
+define i32 @use_after_scop() #0 {
+entry:
+  %scalar.s2a = alloca float
+  fence seq_cst
+  br label %for.head
+
+for.head:                                         ; preds = %for.inc, %entry
+  %indvar = phi i64 [ %indvar.next, %for.inc ], [ 0, %entry ]
+  br label %for.body
+
+for.body:                                         ; preds = %for.head
+  %arrayidx = getelementptr [1024 x float]* @A, i64 0, i64 %indvar
+  %scalar = load float* %arrayidx
+  store float %scalar, float* %scalar.s2a
+; Escaped uses are still required to be rewritten to stack variable.
+; CHECK: BB: for.body
+; CHECK: Read A[{0,+,4}<%for.head>]
+; CHECK: Write scalar.s2a[0]
+  br label %for.inc
+
+for.inc:                                          ; preds = %for.body
+  %indvar.next = add i64 %indvar, 1
+  %exitcond = icmp ne i64 %indvar, 1024
+  br i1 %exitcond, label %for.head, label %for.after
+
+for.after:                                        ; preds = %for.inc
+  %scalar.loadoutside = load float* %scalar.s2a
+  fence seq_cst
+  %return_value = fptosi float %scalar.loadoutside to i32
+  br label %return
+
+return:                                           ; preds = %for.after
+  ret i32 %return_value
+}
+
+; We currently do not transform scalar references, that have only read accesses
+; in the scop. There are two reasons for this:
+;
+;  o We don't introduce additional memory references which may yield to compile
+;    time overhead.
+;  o For integer values, such a translation may block the use of scalar
+;    evolution on those values.
+;
+; CHECK: before_scop
+; Function Attrs: nounwind
+define i32 @before_scop() #0 {
+entry:
+  br label %preheader
+
+preheader:                                        ; preds = %entry
+  %scalar = fadd float 4.000000e+00, 5.000000e+00
+  fence seq_cst
+  br label %for.cond
+
+for.cond:                                         ; preds = %for.inc, %preheader
+  %indvar = phi i64 [ %indvar.next, %for.inc ], [ 0, %preheader ]
+  %exitcond = icmp ne i64 %indvar, 1024
+  br i1 %exitcond, label %for.body, label %return
+
+for.body:                                         ; preds = %for.cond
+  %arrayidx = getelementptr [1024 x float]* @A, i64 0, i64 %indvar
+  store float %scalar, float* %arrayidx
+  br label %for.inc
+; CHECK: BB: for.body
+; CHECK: Write A[{0,+,4}<%for.cond>]
+
+for.inc:                                          ; preds = %for.body
+  %indvar.next = add i64 %indvar, 1
+  br label %for.cond
+
+return:                                           ; preds = %for.cond
+  fence seq_cst
+  ret i32 0
+}
+
+attributes #0 = { nounwind }

diff --git a/polly/test/TempScop/tempscop-printing.ll b/polly/test/TempScop/tempscop-printing.ll
index 3e990bd..04b5c67 100644 (file)
--- a/polly/test/TempScop/tempscop-printing.ll
+++ b/polly/test/TempScop/tempscop-printing.ll
@@ -1,4 +1,5 @@
 ; RUN: opt %loadPolly -basicaa -polly-analyze-ir -analyze < %s | FileCheck %s
+; RUN: opt %loadPolly -basicaa -polly-analyze-ir -analyze -disable-polly-intra-scop-scalar-to-array < %s | FileCheck %s -check-prefix=SCALARACCESS
 
 ; void f(long A[], int N, int *init_ptr) {
 ;   long i, j;
@@ -25,9 +26,12 @@ for.i:
 
 entry.next:
 ; CHECK: BB: entry.next
+; SCALARACCESS: BB: entry.next
   %init = load i64* %init_ptr
 ; CHECK:  Read init_ptr[0]
 ; CHECK:  Write init.s2a[0]
+; SCALARACCESS: Read init_ptr[0]
+; SCALARACCESS:  Write init[0]
   br label %for.j
 
 for.j:
@@ -35,6 +39,10 @@ for.j:
 ; CHECK: BB: for.j
 ; CHECK: Read init.s2a[0]
 ; CHECK: Write A[{0,+,8}<%for.j>]
+
+; SCALARACCESS: BB: for.j
+; SCALARACCESS: Read init
+; SCALARACCESS: Write A[{0,+,8}<%for.j>]
   %init_plus_two = add i64 %init, 2
   %scevgep = getelementptr i64* %A, i64 %indvar.j
   store i64 %init_plus_two, i64* %scevgep
@@ -49,3 +57,38 @@ for.i.end:
 return:
   ret void
 }
+
+define void @g(i64* noalias %A, i64 %N, i64* noalias %init_ptr) nounwind {
+entry:
+  br label %for.i
+
+for.i:
+  %indvar.i = phi i64 [ 0, %entry ], [ %indvar.i.next, %for.i.end ]
+  %indvar.i.next = add nsw i64 %indvar.i, 1
+  br label %entry.next
+
+entry.next:
+; SCALARACCESS: BB: entry.next
+  %init = load i64* %init_ptr
+; SCALARACCESS: Read init_ptr[0]
+; SCALARACCESS:  Write init[0]
+  br label %for.j
+
+for.j:
+; SCALARACCESS: BB: for.j
+  %indvar.j = phi i64 [ 0, %entry.next ], [ %indvar.j.next, %for.j ]
+  %scevgep = getelementptr i64* %A, i64 %indvar.j
+  store i64 %init, i64* %scevgep
+; SCALARACCESS: Read init
+; SCALARACCESS: Write A[{0,+,8}<%for.j>]
+  %indvar.j.next = add nsw i64 %indvar.j, 1
+  %exitcond.j = icmp eq i64 %indvar.j.next, %N
+  br i1 %exitcond.j, label %for.i.end, label %for.j
+
+for.i.end:
+  %exitcond.i = icmp eq i64 %indvar.i.next, %N
+  br i1 %exitcond.i, label %return, label %for.i
+
+return:
+  ret void
+}