[llvm] 65031c1 - [SVE][InstCombine] Don't convert calls to fp binop intrinsics to instructions when strictfp is required.

[Paul Walker via llvm-commits]

Author: Paul Walker
Date: 2023-04-06T10:23:44Z
New Revision: 65031c189bc818ee449294b2f02abc55a5f88691

URL: https://github.com/llvm/llvm-project/commit/65031c189bc818ee449294b2f02abc55a5f88691
DIFF: https://github.com/llvm/llvm-project/commit/65031c189bc818ee449294b2f02abc55a5f88691.diff

LOG: [SVE][InstCombine] Don't convert calls to fp binop intrinsics to instructions when strictfp is required.

There's no support to lower scalable vector constrained operations for SVE
and thus without this change we'll crash during code generation.

Differential Revision: https://reviews.llvm.org/D147600

Added: 
    

Modified: 
    llvm/lib/Target/AArch64/AArch64TargetTransformInfo.cpp
    llvm/test/Transforms/InstCombine/AArch64/sve-intrinsic-strictfp.ll

Removed: 
    


################################################################################
diff  --git a/llvm/lib/Target/AArch64/AArch64TargetTransformInfo.cpp b/llvm/lib/Target/AArch64/AArch64TargetTransformInfo.cpp
index df8eff2812cf2..7c57d6ca1aec0 100644
--- a/llvm/lib/Target/AArch64/AArch64TargetTransformInfo.cpp
+++ b/llvm/lib/Target/AArch64/AArch64TargetTransformInfo.cpp
@@ -1178,6 +1178,10 @@ static Instruction::BinaryOps intrinsicIDToBinOpCode(unsigned Intrinsic) {
 
 static std::optional<Instruction *>
 instCombineSVEVectorBinOp(InstCombiner &IC, IntrinsicInst &II) {
+  // Bail due to missing support for ISD::STRICT_ scalable vector operations.
+  if (II.isStrictFP())
+    return std::nullopt;
+
   auto *OpPredicate = II.getOperand(0);
   auto BinOpCode = intrinsicIDToBinOpCode(II.getIntrinsicID());
   if (BinOpCode == Instruction::BinaryOpsEnd ||

diff  --git a/llvm/test/Transforms/InstCombine/AArch64/sve-intrinsic-strictfp.ll b/llvm/test/Transforms/InstCombine/AArch64/sve-intrinsic-strictfp.ll
index 0272ffb3e2d54..91b47410e8cb3 100644
--- a/llvm/test/Transforms/InstCombine/AArch64/sve-intrinsic-strictfp.ll
+++ b/llvm/test/Transforms/InstCombine/AArch64/sve-intrinsic-strictfp.ll
@@ -8,8 +8,9 @@ target triple = "aarch64-unknown-linux-gnu"
 define <vscale x 2 x double> @replace_fadd_intrinsic_double_strictfp(<vscale x 2 x double> %a, <vscale x 2 x double> %b) #0 {
 ; CHECK: Function Attrs: strictfp
 ; CHECK-LABEL: @replace_fadd_intrinsic_double_strictfp(
-; CHECK-NEXT:    [[TMP1:%.*]] = fadd <vscale x 2 x double> [[A:%.*]], [[B:%.*]]
-; CHECK-NEXT:    ret <vscale x 2 x double> [[TMP1]]
+; CHECK-NEXT:    [[TMP1:%.*]] = tail call <vscale x 2 x i1> @llvm.aarch64.sve.ptrue.nxv2i1(i32 31) #[[ATTR2:[0-9]+]]
+; CHECK-NEXT:    [[TMP2:%.*]] = tail call <vscale x 2 x double> @llvm.aarch64.sve.fadd.nxv2f64(<vscale x 2 x i1> [[TMP1]], <vscale x 2 x double> [[A:%.*]], <vscale x 2 x double> [[B:%.*]]) #[[ATTR2]]
+; CHECK-NEXT:    ret <vscale x 2 x double> [[TMP2]]
 ;
   %1 = tail call <vscale x 2 x i1> @llvm.aarch64.sve.ptrue.nxv2i1(i32 31) #1
   %2 = tail call <vscale x 2 x double> @llvm.aarch64.sve.fadd.nxv2f64(<vscale x 2 x i1> %1, <vscale x 2 x double> %a, <vscale x 2 x double> %b) #1
@@ -21,8 +22,9 @@ define <vscale x 2 x double> @replace_fadd_intrinsic_double_strictfp(<vscale x 2
 define <vscale x 2 x double> @call_replace_fadd_intrinsic_double_strictfp(<vscale x 2 x double> %a, <vscale x 2 x double> %b) #0 {
 ; CHECK: Function Attrs: strictfp
 ; CHECK-LABEL: @call_replace_fadd_intrinsic_double_strictfp(
-; CHECK-NEXT:    [[DOTSTRICT:%.*]] = call <vscale x 2 x double> @llvm.experimental.constrained.fadd.nxv2f64(<vscale x 2 x double> [[A:%.*]], <vscale x 2 x double> [[B:%.*]], metadata !"round.tonearest", metadata !"fpexcept.ignore") #[[ATTR3:[0-9]+]]
-; CHECK-NEXT:    ret <vscale x 2 x double> [[DOTSTRICT]]
+; CHECK-NEXT:    [[TMP1:%.*]] = call <vscale x 2 x i1> @llvm.aarch64.sve.ptrue.nxv2i1(i32 31) #[[ATTR2]]
+; CHECK-NEXT:    [[TMP2:%.*]] = call <vscale x 2 x double> @llvm.aarch64.sve.fadd.nxv2f64(<vscale x 2 x i1> [[TMP1]], <vscale x 2 x double> [[A:%.*]], <vscale x 2 x double> [[B:%.*]]) #[[ATTR2]]
+; CHECK-NEXT:    ret <vscale x 2 x double> [[TMP2]]
 ;
   %1 = call <vscale x 2 x double> @replace_fadd_intrinsic_double_strictfp(<vscale x 2 x double> %a, <vscale x 2 x double> %b) #1
   ret <vscale x 2 x double> %1
@@ -33,8 +35,9 @@ define <vscale x 2 x double> @call_replace_fadd_intrinsic_double_strictfp(<vscal
 define <vscale x 2 x double> @replace_fmul_intrinsic_double_strictfp(<vscale x 2 x double> %a, <vscale x 2 x double> %b) #0 {
 ; CHECK: Function Attrs: strictfp
 ; CHECK-LABEL: @replace_fmul_intrinsic_double_strictfp(
-; CHECK-NEXT:    [[TMP1:%.*]] = fmul <vscale x 2 x double> [[A:%.*]], [[B:%.*]]
-; CHECK-NEXT:    ret <vscale x 2 x double> [[TMP1]]
+; CHECK-NEXT:    [[TMP1:%.*]] = tail call <vscale x 2 x i1> @llvm.aarch64.sve.ptrue.nxv2i1(i32 31) #[[ATTR2]]
+; CHECK-NEXT:    [[TMP2:%.*]] = tail call <vscale x 2 x double> @llvm.aarch64.sve.fmul.nxv2f64(<vscale x 2 x i1> [[TMP1]], <vscale x 2 x double> [[A:%.*]], <vscale x 2 x double> [[B:%.*]]) #[[ATTR2]]
+; CHECK-NEXT:    ret <vscale x 2 x double> [[TMP2]]
 ;
   %1 = tail call <vscale x 2 x i1> @llvm.aarch64.sve.ptrue.nxv2i1(i32 31) #1
   %2 = tail call <vscale x 2 x double> @llvm.aarch64.sve.fmul.nxv2f64(<vscale x 2 x i1> %1, <vscale x 2 x double> %a, <vscale x 2 x double> %b) #1
@@ -46,8 +49,9 @@ define <vscale x 2 x double> @replace_fmul_intrinsic_double_strictfp(<vscale x 2
 define <vscale x 2 x double> @call_replace_fmul_intrinsic_double_strictfp(<vscale x 2 x double> %a, <vscale x 2 x double> %b) #0 {
 ; CHECK: Function Attrs: strictfp
 ; CHECK-LABEL: @call_replace_fmul_intrinsic_double_strictfp(
-; CHECK-NEXT:    [[DOTSTRICT:%.*]] = call <vscale x 2 x double> @llvm.experimental.constrained.fmul.nxv2f64(<vscale x 2 x double> [[A:%.*]], <vscale x 2 x double> [[B:%.*]], metadata !"round.tonearest", metadata !"fpexcept.ignore") #[[ATTR3]]
-; CHECK-NEXT:    ret <vscale x 2 x double> [[DOTSTRICT]]
+; CHECK-NEXT:    [[TMP1:%.*]] = call <vscale x 2 x i1> @llvm.aarch64.sve.ptrue.nxv2i1(i32 31) #[[ATTR2]]
+; CHECK-NEXT:    [[TMP2:%.*]] = call <vscale x 2 x double> @llvm.aarch64.sve.fmul.nxv2f64(<vscale x 2 x i1> [[TMP1]], <vscale x 2 x double> [[A:%.*]], <vscale x 2 x double> [[B:%.*]]) #[[ATTR2]]
+; CHECK-NEXT:    ret <vscale x 2 x double> [[TMP2]]
 ;
   %1 = call <vscale x 2 x double> @replace_fmul_intrinsic_double_strictfp(<vscale x 2 x double> %a, <vscale x 2 x double> %b) #1
   ret <vscale x 2 x double> %1
@@ -58,8 +62,9 @@ define <vscale x 2 x double> @call_replace_fmul_intrinsic_double_strictfp(<vscal
 define <vscale x 2 x double> @replace_fsub_intrinsic_double_strictfp(<vscale x 2 x double> %a, <vscale x 2 x double> %b) #0 {
 ; CHECK: Function Attrs: strictfp
 ; CHECK-LABEL: @replace_fsub_intrinsic_double_strictfp(
-; CHECK-NEXT:    [[TMP1:%.*]] = fsub <vscale x 2 x double> [[A:%.*]], [[B:%.*]]
-; CHECK-NEXT:    ret <vscale x 2 x double> [[TMP1]]
+; CHECK-NEXT:    [[TMP1:%.*]] = tail call <vscale x 2 x i1> @llvm.aarch64.sve.ptrue.nxv2i1(i32 31) #[[ATTR2]]
+; CHECK-NEXT:    [[TMP2:%.*]] = tail call <vscale x 2 x double> @llvm.aarch64.sve.fsub.nxv2f64(<vscale x 2 x i1> [[TMP1]], <vscale x 2 x double> [[A:%.*]], <vscale x 2 x double> [[B:%.*]]) #[[ATTR2]]
+; CHECK-NEXT:    ret <vscale x 2 x double> [[TMP2]]
 ;
   %1 = tail call <vscale x 2 x i1> @llvm.aarch64.sve.ptrue.nxv2i1(i32 31) #1
   %2 = tail call <vscale x 2 x double> @llvm.aarch64.sve.fsub.nxv2f64(<vscale x 2 x i1> %1, <vscale x 2 x double> %a, <vscale x 2 x double> %b) #1
@@ -71,8 +76,9 @@ define <vscale x 2 x double> @replace_fsub_intrinsic_double_strictfp(<vscale x 2
 define <vscale x 2 x double> @call_replace_fsub_intrinsic_double_strictfp(<vscale x 2 x double> %a, <vscale x 2 x double> %b) #0 {
 ; CHECK: Function Attrs: strictfp
 ; CHECK-LABEL: @call_replace_fsub_intrinsic_double_strictfp(
-; CHECK-NEXT:    [[DOTSTRICT:%.*]] = call <vscale x 2 x double> @llvm.experimental.constrained.fsub.nxv2f64(<vscale x 2 x double> [[A:%.*]], <vscale x 2 x double> [[B:%.*]], metadata !"round.tonearest", metadata !"fpexcept.ignore") #[[ATTR3]]
-; CHECK-NEXT:    ret <vscale x 2 x double> [[DOTSTRICT]]
+; CHECK-NEXT:    [[TMP1:%.*]] = call <vscale x 2 x i1> @llvm.aarch64.sve.ptrue.nxv2i1(i32 31) #[[ATTR2]]
+; CHECK-NEXT:    [[TMP2:%.*]] = call <vscale x 2 x double> @llvm.aarch64.sve.fsub.nxv2f64(<vscale x 2 x i1> [[TMP1]], <vscale x 2 x double> [[A:%.*]], <vscale x 2 x double> [[B:%.*]]) #[[ATTR2]]
+; CHECK-NEXT:    ret <vscale x 2 x double> [[TMP2]]
 ;
   %1 = call <vscale x 2 x double> @replace_fsub_intrinsic_double_strictfp(<vscale x 2 x double> %a, <vscale x 2 x double> %b) #1
   ret <vscale x 2 x double> %1
@@ -89,6 +95,5 @@ attributes #1 = { strictfp }
 ;.
 ; CHECK: attributes #[[ATTR0:[0-9]+]] = { strictfp "target-features"="+sve" }
 ; CHECK: attributes #[[ATTR1:[0-9]+]] = { nocallback nofree nosync nounwind willreturn memory(none) }
-; CHECK: attributes #[[ATTR2:[0-9]+]] = { nocallback nofree nosync nounwind willreturn memory(inaccessiblemem: readwrite) }
-; CHECK: attributes #[[ATTR3]] = { strictfp }
+; CHECK: attributes #[[ATTR2]] = { strictfp }
 ;.