[X86][SSE] combineX86ShufflesConstants - early out for zeroable vectors (PR45443)

Shuffle combining can insert zero byte sized elements into the shuffle mask, which combineX86ShufflesConstants will attempt to fold without taking into account whether the byte-sized type is legal (e.g. AVX512F only targets).

If we have a full-zeroable vector then we should just return a zero version of the root type, otherwise if the type isn't valid we should bail.

Fixes PR45443

(cherry picked from commit e3b60597769f79a8abc19fb8ef1f321d9adc1358)

2 files changed, 28 insertions, 1 deletions

diff --git a/llvm/lib/Target/X86/X86ISelLowering.cpp b/llvm/lib/Target/X86/X86ISelLowering.cpp
index cbdd7135de4..60eefbc677d 100644
--- a/llvm/lib/Target/X86/X86ISelLowering.cpp
+++ b/llvm/lib/Target/X86/X86ISelLowering.cpp
@@ -33998,6 +33998,7 @@ static SDValue combineX86ShufflesConstants(ArrayRef<SDValue> Ops,
     return SDValue();
 
   // Shuffle the constant bits according to the mask.
+  SDLoc DL(Root);
   APInt UndefElts(NumMaskElts, 0);
   APInt ZeroElts(NumMaskElts, 0);
   APInt ConstantElts(NumMaskElts, 0);
@@ -34035,6 +34036,10 @@ static SDValue combineX86ShufflesConstants(ArrayRef<SDValue> Ops,
   }
   assert((UndefElts | ZeroElts | ConstantElts).isAllOnesValue());
 
+  // Attempt to create a zero vector.
+  if ((UndefElts | ZeroElts).isAllOnesValue())
+    return getZeroVector(Root.getSimpleValueType(), Subtarget, DAG, DL);
+
   // Create the constant data.
   MVT MaskSVT;
   if (VT.isFloatingPoint() && (MaskSizeInBits == 32 || MaskSizeInBits == 64))
@@ -34043,8 +34048,9 @@ static SDValue combineX86ShufflesConstants(ArrayRef<SDValue> Ops,
     MaskSVT = MVT::getIntegerVT(MaskSizeInBits);
 
   MVT MaskVT = MVT::getVectorVT(MaskSVT, NumMaskElts);
+  if (!DAG.getTargetLoweringInfo().isTypeLegal(MaskVT))
+    return SDValue();
 
-  SDLoc DL(Root);
   SDValue CstOp = getConstVector(ConstantBitData, UndefElts, MaskVT, DAG, DL);
   return DAG.getBitcast(VT, CstOp);
 }
diff --git a/llvm/test/CodeGen/X86/pr45443.ll b/llvm/test/CodeGen/X86/pr45443.ll
new file mode 100644
index 00000000000..1e40ab94e9c
--- /dev/null
+++ b/llvm/test/CodeGen/X86/pr45443.ll
@@ -0,0 +1,21 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
+; RUN: llc < %s -mtriple=i686-- -mattr=+avx512f | FileCheck %s --check-prefixes=CHECK,X86
+; RUN: llc < %s -mtriple=x86_64-- -mattr=+avx512f | FileCheck %s --check-prefixes=CHECK,X64
+
+define <16 x float> @PR45443() {
+; CHECK-LABEL: PR45443:
+; CHECK:       # %bb.0: # %bb
+; CHECK-NEXT:    vfmadd231ps {{.*#+}} zmm0 = (zmm0 * mem) + zmm0
+; CHECK-NEXT:    ret{{[l|q]}}
+bb:
+  %tmp = tail call <16 x i32> @llvm.x86.avx512.psll.d.512(<16 x i32> <i32 1090519040, i32 1090519040, i32 1090519040, i32 1090519040, i32 1090519040, i32 1090519040, i32 1090519040, i32 1090519040, i32 1090519040, i32 1090519040, i32 1090519040, i32 1090519040, i32 1090519040, i32 1090519040, i32 1090519040, i32 1090519040>, <4 x i32> <i32 1, i32 0, i32 undef, i32 undef>)
+  %tmp4 = tail call fast <16 x float> @llvm.fma.v16f32(<16 x float> undef, <16 x float> <float 0x3FE6300000000000, float 0x3FE6300000000000, float 0x3FE6300000000000, float 0x3FE6300000000000, float 0x3FE6300000000000, float 0x3FE6300000000000, float 0x3FE6300000000000, float 0x3FE6300000000000, float 0x3FE6300000000000, float 0x3FE6300000000000, float 0x3FE6300000000000, float 0x3FE6300000000000, float 0x3FE6300000000000, float 0x3FE6300000000000, float 0x3FE6300000000000, float 0x3FE6300000000000>, <16 x float> undef)
+  %tmp5 = icmp ult <16 x i32> %tmp, <i32 16777216, i32 16777216, i32 16777216, i32 16777216, i32 16777216, i32 16777216, i32 16777216, i32 16777216, i32 16777216, i32 16777216, i32 16777216, i32 16777216, i32 16777216, i32 16777216, i32 16777216, i32 16777216>
+  %tmp6 = and <16 x i32> %tmp, <i32 16777215, i32 16777215, i32 16777215, i32 16777215, i32 16777215, i32 16777215, i32 16777215, i32 16777215, i32 16777215, i32 16777215, i32 16777215, i32 16777215, i32 16777215, i32 16777215, i32 16777215, i32 16777215>
+  %tmp7 = icmp ne <16 x i32> %tmp6, zeroinitializer
+  %tmp8 = and <16 x i1> %tmp7, %tmp5
+  %tmp9 = select fast <16 x i1> %tmp8, <16 x float> <float 0xFFF0000000000000, float 0xFFF0000000000000, float 0xFFF0000000000000, float 0xFFF0000000000000, float 0xFFF0000000000000, float 0xFFF0000000000000, float 0xFFF0000000000000, float 0xFFF0000000000000, float 0xFFF0000000000000, float 0xFFF0000000000000, float 0xFFF0000000000000, float 0xFFF0000000000000, float 0xFFF0000000000000, float 0xFFF0000000000000, float 0xFFF0000000000000, float 0xFFF0000000000000>, <16 x float> %tmp4
+  ret <16 x float> %tmp9
+}
+declare <16 x float> @llvm.fma.v16f32(<16 x float>, <16 x float>, <16 x float>)
+declare <16 x i32> @llvm.x86.avx512.psll.d.512(<16 x i32>, <4 x i32>)