[X86] combineX86ShuffleChain - don't fold to truncate(concat(V1,V2)) if it was already a PACK op

Fixes #55050

(cherry picked from commit e8305c0b8f492272446d517e67d451f390d35dbe)

2 files changed, 49 insertions, 1 deletions

diff --git a/llvm/lib/Target/X86/X86ISelLowering.cpp b/llvm/lib/Target/X86/X86ISelLowering.cpp
index 682932b8f3e6..8bb7e81e19bb 100644
--- a/llvm/lib/Target/X86/X86ISelLowering.cpp
+++ b/llvm/lib/Target/X86/X86ISelLowering.cpp
@@ -37558,7 +37558,11 @@ static SDValue combineX86ShuffleChain(ArrayRef<SDValue> Inputs, SDValue Root,
          (RootVT.is128BitVector() && Subtarget.hasVLX())) &&
         (MaskEltSizeInBits > 8 || Subtarget.hasBWI()) &&
         isSequentialOrUndefInRange(Mask, 0, NumMaskElts, 0, 2)) {
-      if (Depth == 0 && Root.getOpcode() == ISD::TRUNCATE)
+      // Bail if this was already a truncation or PACK node.
+      // We sometimes fail to match PACK if we demand known undef elements.
+      if (Depth == 0 && (Root.getOpcode() == ISD::TRUNCATE ||
+                         Root.getOpcode() == X86ISD::PACKSS ||
+                         Root.getOpcode() == X86ISD::PACKUS))
         return SDValue(); // Nothing to do!
       ShuffleSrcVT = MVT::getIntegerVT(MaskEltSizeInBits * 2);
       ShuffleSrcVT = MVT::getVectorVT(ShuffleSrcVT, NumMaskElts / 2);
diff --git a/llvm/test/CodeGen/X86/vector-shuffle-combining-avx512bwvl.ll b/llvm/test/CodeGen/X86/vector-shuffle-combining-avx512bwvl.ll
index be65effbc724..3be4e6777d6f 100644
--- a/llvm/test/CodeGen/X86/vector-shuffle-combining-avx512bwvl.ll
+++ b/llvm/test/CodeGen/X86/vector-shuffle-combining-avx512bwvl.ll
@@ -174,3 +174,47 @@ define <8 x i32> @PR46393(<8 x i16> %a0, i8 %a1) {
   %sel = select <8 x i1> %mask, <8 x i32> %shl, <8 x i32> zeroinitializer
   ret <8 x i32> %sel
 }
+
+define i64 @PR55050() {
+; X86-LABEL: PR55050:
+; X86:       # %bb.0: # %entry
+; X86-NEXT:    xorl %eax, %eax
+; X86-NEXT:    testb %al, %al
+; X86-NEXT:    jne .LBB10_2
+; X86-NEXT:  # %bb.1: # %if
+; X86-NEXT:    xorl %eax, %eax
+; X86-NEXT:  .LBB10_2: # %exit
+; X86-NEXT:    movl %eax, %edx
+; X86-NEXT:    retl
+;
+; X64-LABEL: PR55050:
+; X64:       # %bb.0: # %entry
+; X64-NEXT:    xorl %eax, %eax
+; X64-NEXT:    testb %al, %al
+; X64-NEXT:    xorl %eax, %eax
+; X64-NEXT:    retq
+entry:
+  %i275 = call <2 x i64> @llvm.x86.sse2.psad.bw(<16 x i8> undef, <16 x i8> zeroinitializer)
+  %i277 = call <2 x i64> @llvm.x86.sse2.psad.bw(<16 x i8> undef, <16 x i8> zeroinitializer)
+  br i1 undef, label %exit, label %if
+
+if:
+  %i298 = bitcast <2 x i64> %i275 to <4 x i32>
+  %i299 = bitcast <2 x i64> %i277 to <4 x i32>
+  %i300 = shufflevector <4 x i32> %i298, <4 x i32> %i299, <4 x i32> <i32 0, i32 2, i32 4, i32 6>
+  %i339 = call <8 x i16> @llvm.x86.sse41.packusdw(<4 x i32> %i300, <4 x i32> undef)
+  %i354 = shufflevector <8 x i16> %i339, <8 x i16> undef, <8 x i32> <i32 0, i32 undef, i32 2, i32 undef, i32 4, i32 undef, i32 6, i32 undef>
+  %i356 = call <16 x i8> @llvm.x86.sse2.packuswb.128(<8 x i16> %i354, <8 x i16> undef)
+  %i357 = shufflevector <16 x i8> %i356, <16 x i8> zeroinitializer, <16 x i32> <i32 6, i32 5, i32 4, i32 16, i32 2, i32 1, i32 0, i32 16, i32 10, i32 9, i32 8, i32 16, i32 16, i32 16, i32 16, i32 16>
+  %i361 = extractelement <16 x i8> %i357, i64 8
+  %i360 = and i8 %i361, 63
+  %i379 = zext i8 %i360 to i64
+  br label %exit
+
+exit:
+  %res = phi i64 [ %i379, %if ], [ 0, %entry ]
+  ret i64 %res
+}
+declare <2 x i64> @llvm.x86.sse2.psad.bw(<16 x i8>, <16 x i8>)
+declare <16 x i8> @llvm.x86.sse2.packuswb.128(<8 x i16>, <8 x i16>)
+declare <8 x i16> @llvm.x86.sse41.packusdw(<4 x i32>, <4 x i32>)