[AArch64][GlobalISel] Legalize G_INTRINSIC_TRUNC

Same patch as G_FCEIL etc.

Add the missing switch case in widenScalar, add G_INTRINSIC_TRUNC to the correct
rule in AArch64LegalizerInfo.cpp, and add a test.

llvm-svn: 359021

diff --git a/llvm/lib/CodeGen/GlobalISel/LegalizerHelper.cpp b/llvm/lib/CodeGen/GlobalISel/LegalizerHelper.cpp
index 4b592b3..9521e67 100644 (file)
--- a/llvm/lib/CodeGen/GlobalISel/LegalizerHelper.cpp
+++ b/llvm/lib/CodeGen/GlobalISel/LegalizerHelper.cpp
@@ -1327,6 +1327,7 @@ LegalizerHelper::widenScalar(MachineInstr &MI, unsigned TypeIdx, LLT WideTy) {
   case TargetOpcode::G_FEXP:
   case TargetOpcode::G_FEXP2:
   case TargetOpcode::G_FPOW:
+  case TargetOpcode::G_INTRINSIC_TRUNC:
     assert(TypeIdx == 0);
     Observer.changingInstr(MI);
 

diff --git a/llvm/lib/Target/AArch64/AArch64LegalizerInfo.cpp b/llvm/lib/Target/AArch64/AArch64LegalizerInfo.cpp
index d58eb51..1b3020e 100644 (file)
--- a/llvm/lib/Target/AArch64/AArch64LegalizerInfo.cpp
+++ b/llvm/lib/Target/AArch64/AArch64LegalizerInfo.cpp
@@ -131,7 +131,7 @@ AArch64LegalizerInfo::AArch64LegalizerInfo(const AArch64Subtarget &ST) {
   getActionDefinitionsBuilder(G_FREM).libcallFor({s32, s64});
 
   getActionDefinitionsBuilder(
-      {G_FCEIL, G_FABS, G_FSQRT, G_FFLOOR, G_FRINT, G_FMA})
+      {G_FCEIL, G_FABS, G_FSQRT, G_FFLOOR, G_FRINT, G_FMA, G_INTRINSIC_TRUNC})
       // If we don't have full FP16 support, then scalarize the elements of
       // vectors containing fp16 types.
       .fewerElementsIf(

diff --git a/llvm/test/CodeGen/AArch64/GlobalISel/legalize-intrinsic-trunc.mir b/llvm/test/CodeGen/AArch64/GlobalISel/legalize-intrinsic-trunc.mir
new file mode 100644 (file)
index 0000000..c552a55
--- /dev/null
+++ b/llvm/test/CodeGen/AArch64/GlobalISel/legalize-intrinsic-trunc.mir
@@ -0,0 +1,203 @@
+# NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py
+# RUN: llc %s -verify-machineinstrs -mtriple=aarch64-unknown-unknown -run-pass=legalizer -simplify-mir -aarch64-neon-syntax=apple -mattr=-fullfp16 -o - | FileCheck %s --check-prefix=NO-FP16
+# RUN: llc %s -verify-machineinstrs -mtriple=aarch64-unknown-unknown -run-pass=legalizer -simplify-mir -aarch64-neon-syntax=apple -mattr=+fullfp16 -o - | FileCheck %s --check-prefix=FP16
+
+...
+---
+name:            test_f16.intrinsic_trunc
+alignment:       2
+tracksRegLiveness: true
+body:             |
+  bb.0:
+    liveins: $h0
+    ; NO-FP16-LABEL: name: test_f16.intrinsic_trunc
+    ; NO-FP16: liveins: $h0
+    ; NO-FP16: [[COPY:%[0-9]+]]:_(s16) = COPY $h0
+    ; NO-FP16: [[FPEXT:%[0-9]+]]:_(s32) = G_FPEXT [[COPY]](s16)
+    ; NO-FP16: [[INTRINSIC_TRUNC:%[0-9]+]]:_(s32) = G_INTRINSIC_TRUNC [[FPEXT]]
+    ; NO-FP16: [[FPTRUNC:%[0-9]+]]:_(s16) = G_FPTRUNC [[INTRINSIC_TRUNC]](s32)
+    ; NO-FP16: $h0 = COPY [[FPTRUNC]](s16)
+    ; NO-FP16: RET_ReallyLR implicit $h0
+    ; FP16-LABEL: name: test_f16.intrinsic_trunc
+    ; FP16: liveins: $h0
+    ; FP16: [[COPY:%[0-9]+]]:_(s16) = COPY $h0
+    ; FP16: [[INTRINSIC_TRUNC:%[0-9]+]]:_(s16) = G_INTRINSIC_TRUNC [[COPY]]
+    ; FP16: $h0 = COPY [[INTRINSIC_TRUNC]](s16)
+    ; FP16: RET_ReallyLR implicit $h0
+    %0:_(s16) = COPY $h0
+    %1:_(s16) = G_INTRINSIC_TRUNC %0
+    $h0 = COPY %1(s16)
+    RET_ReallyLR implicit $h0
+
+...
+---
+name:            test_v4f16.intrinsic_trunc
+alignment:       2
+tracksRegLiveness: true
+machineFunctionInfo: {}
+body:             |
+  bb.0:
+    liveins: $d0
+
+    ; NO-FP16-LABEL: name: test_v4f16.intrinsic_trunc
+    ; NO-FP16: liveins: $d0
+    ; NO-FP16: [[COPY:%[0-9]+]]:_(<4 x s16>) = COPY $d0
+    ; NO-FP16: [[UV:%[0-9]+]]:_(s16), [[UV1:%[0-9]+]]:_(s16), [[UV2:%[0-9]+]]:_(s16), [[UV3:%[0-9]+]]:_(s16) = G_UNMERGE_VALUES [[COPY]](<4 x s16>)
+    ; NO-FP16: [[FPEXT:%[0-9]+]]:_(s32) = G_FPEXT [[UV]](s16)
+    ; NO-FP16: [[INTRINSIC_TRUNC:%[0-9]+]]:_(s32) = G_INTRINSIC_TRUNC [[FPEXT]]
+    ; NO-FP16: [[FPTRUNC:%[0-9]+]]:_(s16) = G_FPTRUNC [[INTRINSIC_TRUNC]](s32)
+    ; NO-FP16: [[FPEXT1:%[0-9]+]]:_(s32) = G_FPEXT [[UV1]](s16)
+    ; NO-FP16: [[INTRINSIC_TRUNC1:%[0-9]+]]:_(s32) = G_INTRINSIC_TRUNC [[FPEXT1]]
+    ; NO-FP16: [[FPTRUNC1:%[0-9]+]]:_(s16) = G_FPTRUNC [[INTRINSIC_TRUNC1]](s32)
+    ; NO-FP16: [[FPEXT2:%[0-9]+]]:_(s32) = G_FPEXT [[UV2]](s16)
+    ; NO-FP16: [[INTRINSIC_TRUNC2:%[0-9]+]]:_(s32) = G_INTRINSIC_TRUNC [[FPEXT2]]
+    ; NO-FP16: [[FPTRUNC2:%[0-9]+]]:_(s16) = G_FPTRUNC [[INTRINSIC_TRUNC2]](s32)
+    ; NO-FP16: [[FPEXT3:%[0-9]+]]:_(s32) = G_FPEXT [[UV3]](s16)
+    ; NO-FP16: [[INTRINSIC_TRUNC3:%[0-9]+]]:_(s32) = G_INTRINSIC_TRUNC [[FPEXT3]]
+    ; NO-FP16: [[FPTRUNC3:%[0-9]+]]:_(s16) = G_FPTRUNC [[INTRINSIC_TRUNC3]](s32)
+    ; NO-FP16: [[BUILD_VECTOR:%[0-9]+]]:_(<4 x s16>) = G_BUILD_VECTOR [[FPTRUNC]](s16), [[FPTRUNC1]](s16), [[FPTRUNC2]](s16), [[FPTRUNC3]](s16)
+    ; NO-FP16: $d0 = COPY [[BUILD_VECTOR]](<4 x s16>)
+    ; NO-FP16: RET_ReallyLR implicit $d0
+    ; FP16-LABEL: name: test_v4f16.intrinsic_trunc
+    ; FP16: liveins: $d0
+    ; FP16: [[COPY:%[0-9]+]]:_(<4 x s16>) = COPY $d0
+    ; FP16: [[INTRINSIC_TRUNC:%[0-9]+]]:_(<4 x s16>) = G_INTRINSIC_TRUNC [[COPY]]
+    ; FP16: $d0 = COPY [[INTRINSIC_TRUNC]](<4 x s16>)
+    ; FP16: RET_ReallyLR implicit $d0
+    %0:_(<4 x s16>) = COPY $d0
+    %1:_(<4 x s16>) = G_INTRINSIC_TRUNC %0
+    $d0 = COPY %1(<4 x s16>)
+    RET_ReallyLR implicit $d0
+
+...
+---
+name:            test_v8f16.intrinsic_trunc
+alignment:       2
+tracksRegLiveness: true
+machineFunctionInfo: {}
+body:             |
+  bb.0:
+    liveins: $q0
+
+    ; NO-FP16-LABEL: name: test_v8f16.intrinsic_trunc
+    ; NO-FP16: liveins: $q0
+    ; NO-FP16: [[COPY:%[0-9]+]]:_(<8 x s16>) = COPY $q0
+    ; NO-FP16: [[UV:%[0-9]+]]:_(s16), [[UV1:%[0-9]+]]:_(s16), [[UV2:%[0-9]+]]:_(s16), [[UV3:%[0-9]+]]:_(s16), [[UV4:%[0-9]+]]:_(s16), [[UV5:%[0-9]+]]:_(s16), [[UV6:%[0-9]+]]:_(s16), [[UV7:%[0-9]+]]:_(s16) = G_UNMERGE_VALUES [[COPY]](<8 x s16>)
+    ; NO-FP16: [[FPEXT:%[0-9]+]]:_(s32) = G_FPEXT [[UV]](s16)
+    ; NO-FP16: [[INTRINSIC_TRUNC:%[0-9]+]]:_(s32) = G_INTRINSIC_TRUNC [[FPEXT]]
+    ; NO-FP16: [[FPTRUNC:%[0-9]+]]:_(s16) = G_FPTRUNC [[INTRINSIC_TRUNC]](s32)
+    ; NO-FP16: [[FPEXT1:%[0-9]+]]:_(s32) = G_FPEXT [[UV1]](s16)
+    ; NO-FP16: [[INTRINSIC_TRUNC1:%[0-9]+]]:_(s32) = G_INTRINSIC_TRUNC [[FPEXT1]]
+    ; NO-FP16: [[FPTRUNC1:%[0-9]+]]:_(s16) = G_FPTRUNC [[INTRINSIC_TRUNC1]](s32)
+    ; NO-FP16: [[FPEXT2:%[0-9]+]]:_(s32) = G_FPEXT [[UV2]](s16)
+    ; NO-FP16: [[INTRINSIC_TRUNC2:%[0-9]+]]:_(s32) = G_INTRINSIC_TRUNC [[FPEXT2]]
+    ; NO-FP16: [[FPTRUNC2:%[0-9]+]]:_(s16) = G_FPTRUNC [[INTRINSIC_TRUNC2]](s32)
+    ; NO-FP16: [[FPEXT3:%[0-9]+]]:_(s32) = G_FPEXT [[UV3]](s16)
+    ; NO-FP16: [[INTRINSIC_TRUNC3:%[0-9]+]]:_(s32) = G_INTRINSIC_TRUNC [[FPEXT3]]
+    ; NO-FP16: [[FPTRUNC3:%[0-9]+]]:_(s16) = G_FPTRUNC [[INTRINSIC_TRUNC3]](s32)
+    ; NO-FP16: [[FPEXT4:%[0-9]+]]:_(s32) = G_FPEXT [[UV4]](s16)
+    ; NO-FP16: [[INTRINSIC_TRUNC4:%[0-9]+]]:_(s32) = G_INTRINSIC_TRUNC [[FPEXT4]]
+    ; NO-FP16: [[FPTRUNC4:%[0-9]+]]:_(s16) = G_FPTRUNC [[INTRINSIC_TRUNC4]](s32)
+    ; NO-FP16: [[FPEXT5:%[0-9]+]]:_(s32) = G_FPEXT [[UV5]](s16)
+    ; NO-FP16: [[INTRINSIC_TRUNC5:%[0-9]+]]:_(s32) = G_INTRINSIC_TRUNC [[FPEXT5]]
+    ; NO-FP16: [[FPTRUNC5:%[0-9]+]]:_(s16) = G_FPTRUNC [[INTRINSIC_TRUNC5]](s32)
+    ; NO-FP16: [[FPEXT6:%[0-9]+]]:_(s32) = G_FPEXT [[UV6]](s16)
+    ; NO-FP16: [[INTRINSIC_TRUNC6:%[0-9]+]]:_(s32) = G_INTRINSIC_TRUNC [[FPEXT6]]
+    ; NO-FP16: [[FPTRUNC6:%[0-9]+]]:_(s16) = G_FPTRUNC [[INTRINSIC_TRUNC6]](s32)
+    ; NO-FP16: [[FPEXT7:%[0-9]+]]:_(s32) = G_FPEXT [[UV7]](s16)
+    ; NO-FP16: [[INTRINSIC_TRUNC7:%[0-9]+]]:_(s32) = G_INTRINSIC_TRUNC [[FPEXT7]]
+    ; NO-FP16: [[FPTRUNC7:%[0-9]+]]:_(s16) = G_FPTRUNC [[INTRINSIC_TRUNC7]](s32)
+    ; NO-FP16: [[BUILD_VECTOR:%[0-9]+]]:_(<8 x s16>) = G_BUILD_VECTOR [[FPTRUNC]](s16), [[FPTRUNC1]](s16), [[FPTRUNC2]](s16), [[FPTRUNC3]](s16), [[FPTRUNC4]](s16), [[FPTRUNC5]](s16), [[FPTRUNC6]](s16), [[FPTRUNC7]](s16)
+    ; NO-FP16: $q0 = COPY [[BUILD_VECTOR]](<8 x s16>)
+    ; NO-FP16: RET_ReallyLR implicit $q0
+    ; FP16-LABEL: name: test_v8f16.intrinsic_trunc
+    ; FP16: liveins: $q0
+    ; FP16: [[COPY:%[0-9]+]]:_(<8 x s16>) = COPY $q0
+    ; FP16: [[INTRINSIC_TRUNC:%[0-9]+]]:_(<8 x s16>) = G_INTRINSIC_TRUNC [[COPY]]
+    ; FP16: $q0 = COPY [[INTRINSIC_TRUNC]](<8 x s16>)
+    ; FP16: RET_ReallyLR implicit $q0
+    %0:_(<8 x s16>) = COPY $q0
+    %1:_(<8 x s16>) = G_INTRINSIC_TRUNC %0
+    $q0 = COPY %1(<8 x s16>)
+    RET_ReallyLR implicit $q0
+
+...
+---
+name:            test_v2f32.intrinsic_trunc
+alignment:       2
+tracksRegLiveness: true
+machineFunctionInfo: {}
+body:             |
+  bb.0:
+    liveins: $d0
+
+    ; NO-FP16-LABEL: name: test_v2f32.intrinsic_trunc
+    ; NO-FP16: liveins: $d0
+    ; NO-FP16: [[COPY:%[0-9]+]]:_(<2 x s32>) = COPY $d0
+    ; NO-FP16: [[INTRINSIC_TRUNC:%[0-9]+]]:_(<2 x s32>) = G_INTRINSIC_TRUNC [[COPY]]
+    ; NO-FP16: $d0 = COPY [[INTRINSIC_TRUNC]](<2 x s32>)
+    ; NO-FP16: RET_ReallyLR implicit $d0
+    ; FP16-LABEL: name: test_v2f32.intrinsic_trunc
+    ; FP16: liveins: $d0
+    ; FP16: [[COPY:%[0-9]+]]:_(<2 x s32>) = COPY $d0
+    ; FP16: [[INTRINSIC_TRUNC:%[0-9]+]]:_(<2 x s32>) = G_INTRINSIC_TRUNC [[COPY]]
+    ; FP16: $d0 = COPY [[INTRINSIC_TRUNC]](<2 x s32>)
+    ; FP16: RET_ReallyLR implicit $d0
+    %0:_(<2 x s32>) = COPY $d0
+    %1:_(<2 x s32>) = G_INTRINSIC_TRUNC %0
+    $d0 = COPY %1(<2 x s32>)
+    RET_ReallyLR implicit $d0
+
+...
+---
+name:            test_v4f32.intrinsic_trunc
+alignment:       2
+tracksRegLiveness: true
+machineFunctionInfo: {}
+body:             |
+  bb.0:
+    liveins: $q0
+
+    ; NO-FP16-LABEL: name: test_v4f32.intrinsic_trunc
+    ; NO-FP16: liveins: $q0
+    ; NO-FP16: [[COPY:%[0-9]+]]:_(<4 x s32>) = COPY $q0
+    ; NO-FP16: [[INTRINSIC_TRUNC:%[0-9]+]]:_(<4 x s32>) = G_INTRINSIC_TRUNC [[COPY]]
+    ; NO-FP16: $q0 = COPY [[INTRINSIC_TRUNC]](<4 x s32>)
+    ; NO-FP16: RET_ReallyLR implicit $q0
+    ; FP16-LABEL: name: test_v4f32.intrinsic_trunc
+    ; FP16: liveins: $q0
+    ; FP16: [[COPY:%[0-9]+]]:_(<4 x s32>) = COPY $q0
+    ; FP16: [[INTRINSIC_TRUNC:%[0-9]+]]:_(<4 x s32>) = G_INTRINSIC_TRUNC [[COPY]]
+    ; FP16: $q0 = COPY [[INTRINSIC_TRUNC]](<4 x s32>)
+    ; FP16: RET_ReallyLR implicit $q0
+    %0:_(<4 x s32>) = COPY $q0
+    %1:_(<4 x s32>) = G_INTRINSIC_TRUNC %0
+    $q0 = COPY %1(<4 x s32>)
+    RET_ReallyLR implicit $q0
+
+...
+---
+name:            test_v2f64.intrinsic_trunc
+alignment:       2
+tracksRegLiveness: true
+machineFunctionInfo: {}
+body:             |
+  bb.0:
+    liveins: $q0
+
+    ; NO-FP16-LABEL: name: test_v2f64.intrinsic_trunc
+    ; NO-FP16: liveins: $q0
+    ; NO-FP16: [[COPY:%[0-9]+]]:_(<2 x s64>) = COPY $q0
+    ; NO-FP16: [[INTRINSIC_TRUNC:%[0-9]+]]:_(<2 x s64>) = G_INTRINSIC_TRUNC [[COPY]]
+    ; NO-FP16: $q0 = COPY [[INTRINSIC_TRUNC]](<2 x s64>)
+    ; NO-FP16: RET_ReallyLR implicit $q0
+    ; FP16-LABEL: name: test_v2f64.intrinsic_trunc
+    ; FP16: liveins: $q0
+    ; FP16: [[COPY:%[0-9]+]]:_(<2 x s64>) = COPY $q0
+    ; FP16: [[INTRINSIC_TRUNC:%[0-9]+]]:_(<2 x s64>) = G_INTRINSIC_TRUNC [[COPY]]
+    ; FP16: $q0 = COPY [[INTRINSIC_TRUNC]](<2 x s64>)
+    ; FP16: RET_ReallyLR implicit $q0
+    %0:_(<2 x s64>) = COPY $q0
+    %1:_(<2 x s64>) = G_INTRINSIC_TRUNC %0
+    $q0 = COPY %1(<2 x s64>)
+    RET_ReallyLR implicit $q0

diff --git a/llvm/test/CodeGen/AArch64/GlobalISel/legalizer-info-validation.mir b/llvm/test/CodeGen/AArch64/GlobalISel/legalizer-info-validation.mir
index e4a6f3c..f149a6c 100644 (file)
--- a/llvm/test/CodeGen/AArch64/GlobalISel/legalizer-info-validation.mir
+++ b/llvm/test/CodeGen/AArch64/GlobalISel/legalizer-info-validation.mir
@@ -88,7 +88,7 @@
 # DEBUG:      .. the first uncovered type index: 2, OK
 #
 # DEBUG-NEXT: G_INTRINSIC_TRUNC (opcode {{[0-9]+}}): 1 type index
-# DEBUG:      .. type index coverage check SKIPPED: no rules defined
+# DEBUG:      .. type index coverage check SKIPPED: user-defined predicate detected
 #
 # DEBUG-NEXT: G_INTRINSIC_ROUND (opcode {{[0-9]+}}): 1 type index
 # DEBUG:      .. type index coverage check SKIPPED: no rules defined