[X86] canCreateUndefOrPoisonForTargetNode - PMADDWD/PMADDUBSW nodes don't create poison

Help with regression identified on #84924

Author: RKSimon

llvm/lib/Target/X86/X86ISelLowering.cpp

@@ -43315,6 +43315,9 @@ bool X86TargetLowering::canCreateUndefOrPoisonForTargetNode(
     bool PoisonOnly, bool ConsiderFlags, unsigned Depth) const {
 
   switch (Op.getOpcode()) {
+  // SSE vector multiplies are either inbounds or saturate.
+  case X86ISD::VPMADDUBSW:
+  case X86ISD::VPMADDWD:
   // SSE vector shifts handle out of bounds shift amounts.
   case X86ISD::VSHLI:
   case X86ISD::VSRLI:

llvm/test/CodeGen/X86/combine-pmadd.ll

@@ -81,10 +81,11 @@ define <8 x i32> @combine_pmaddwd_concat_freeze(<8 x i16> %a0, <8 x i16> %a1) {
 ;
 ; AVX2-LABEL: combine_pmaddwd_concat_freeze:
 ; AVX2:       # %bb.0:
+; AVX2-NEXT:    # kill: def $xmm0 killed $xmm0 def $ymm0
 ; AVX2-NEXT:    vpbroadcastw {{.*#+}} xmm2 = [1,1,1,1,1,1,1,1]
-; AVX2-NEXT:    vpmaddwd %xmm2, %xmm0, %xmm0
-; AVX2-NEXT:    vpmaddwd %xmm2, %xmm1, %xmm1
 ; AVX2-NEXT:    vinserti128 $1, %xmm1, %ymm0, %ymm0
+; AVX2-NEXT:    vinserti128 $1, %xmm2, %ymm2, %ymm1
+; AVX2-NEXT:    vpmaddwd %ymm1, %ymm0, %ymm0
 ; AVX2-NEXT:    retq
   %lo = call <4 x i32> @llvm.x86.sse2.pmadd.wd(<8 x i16> %a0, <8 x i16> <i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1>)
   %hi = call <4 x i32> @llvm.x86.sse2.pmadd.wd(<8 x i16> %a1, <8 x i16> <i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1>)
@@ -227,10 +228,11 @@ define <16 x i16> @combine_pmaddubsw_concat_freeze(<16 x i8> %a0, <16 x i8> %a1)
 ;
 ; AVX2-LABEL: combine_pmaddubsw_concat_freeze:
 ; AVX2:       # %bb.0:
+; AVX2-NEXT:    # kill: def $xmm0 killed $xmm0 def $ymm0
 ; AVX2-NEXT:    vpbroadcastb {{.*#+}} xmm2 = [1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1]
-; AVX2-NEXT:    vpmaddubsw %xmm2, %xmm0, %xmm0
-; AVX2-NEXT:    vpmaddubsw %xmm2, %xmm1, %xmm1
 ; AVX2-NEXT:    vinserti128 $1, %xmm1, %ymm0, %ymm0
+; AVX2-NEXT:    vinserti128 $1, %xmm2, %ymm2, %ymm1
+; AVX2-NEXT:    vpmaddubsw %ymm1, %ymm0, %ymm0
 ; AVX2-NEXT:    retq
   %lo = call <8 x i16> @llvm.x86.ssse3.pmadd.ub.sw.128(<16 x i8> %a0, <16 x i8> <i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1>)
   %hi = call <8 x i16> @llvm.x86.ssse3.pmadd.ub.sw.128(<16 x i8> %a1, <16 x i8> <i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1, i8 1>)