[X86] canCreateUndefOrPoisonForTargetNode/isGuaranteedNotToBeUndefOrPoisonForTargetNode - add X86ISD::INSERTPS handling

[RKSimon]

X86ISD::INSERTPS shuffles can't create undef/poison itself, allowing us to fold freeze(insertps(x,y,i)) -> insertps(freeze(x),freeze(y),i)

[llvmbot]

@llvm/pr-subscribers-backend-x86

Author: Simon Pilgrim (RKSimon)

Changes

X86ISD::INSERTPS shuffles can't create undef/poison itself, allowing us to fold freeze(insertps(x,y,i)) -> insertps(freeze(x),freeze(y),i)

---

Full diff: https://github.com/llvm/llvm-project/pull/161234.diff

2 Files Affected:

• (modified) llvm/lib/Target/X86/X86ISelLowering.cpp (+2)
• (modified) llvm/test/CodeGen/X86/vector-shuffle-combining-sse41.ll (+1-4)

diff --git a/llvm/lib/Target/X86/X86ISelLowering.cpp b/llvm/lib/Target/X86/X86ISelLowering.cpp
index 292eab77e2002..cd04ff5bc7ef4 100644
--- a/llvm/lib/Target/X86/X86ISelLowering.cpp
+++ b/llvm/lib/Target/X86/X86ISelLowering.cpp
@@ -45169,6 +45169,7 @@ bool X86TargetLowering::isGuaranteedNotToBeUndefOrPoisonForTargetNode(
   case X86ISD::Wrapper:
   case X86ISD::WrapperRIP:
     return true;
+  case X86ISD::INSERTPS:
   case X86ISD::BLENDI:
   case X86ISD::PSHUFB:
   case X86ISD::PSHUFD:
@@ -45239,6 +45240,7 @@ bool X86TargetLowering::canCreateUndefOrPoisonForTargetNode(
   case X86ISD::BLENDV:
     return false;
   // SSE target shuffles.
+  case X86ISD::INSERTPS:
   case X86ISD::PSHUFB:
   case X86ISD::PSHUFD:
   case X86ISD::UNPCKL:
diff --git a/llvm/test/CodeGen/X86/vector-shuffle-combining-sse41.ll b/llvm/test/CodeGen/X86/vector-shuffle-combining-sse41.ll
index bec33492bbf1e..3590c4d027be7 100644
--- a/llvm/test/CodeGen/X86/vector-shuffle-combining-sse41.ll
+++ b/llvm/test/CodeGen/X86/vector-shuffle-combining-sse41.ll
@@ -62,15 +62,12 @@ define <4 x i32> @combine_blend_of_permutes_v4i32(<2 x i64> %a0, <2 x i64> %a1)
 define <4 x float> @freeze_insertps(<4 x float> %a0, <4 x float> %a1) {
 ; SSE-LABEL: freeze_insertps:
 ; SSE:       # %bb.0:
-; SSE-NEXT:    insertps {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[2,3]
-; SSE-NEXT:    insertps {{.*#+}} xmm1 = xmm0[1],xmm1[1,2,3]
 ; SSE-NEXT:    movaps %xmm1, %xmm0
 ; SSE-NEXT:    retq
 ;
 ; AVX-LABEL: freeze_insertps:
 ; AVX:       # %bb.0:
-; AVX-NEXT:    vinsertps {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[2,3]
-; AVX-NEXT:    vinsertps {{.*#+}} xmm0 = xmm0[1],xmm1[1,2,3]
+; AVX-NEXT:    vmovaps %xmm1, %xmm0
 ; AVX-NEXT:    retq
   %s0 = call <4 x float> @llvm.x86.sse41.insertps(<4 x float> %a0, <4 x float> %a1, i8 16)
   %f0 = freeze <4 x float> %s0