[X86] Fold insert_vector_elt(undef, elt, 0) --> scalar_to_vector(elt)

I had hoped to make this a generic fold in DAGCombine, but there's quite a few regressions in Thumb2 MVE that need addressing first.

Fixes regressions from D106675.

2 files changed, 12 insertions, 6 deletions

diff --git a/llvm/lib/Target/X86/X86ISelLowering.cpp b/llvm/lib/Target/X86/X86ISelLowering.cpp
index 88004c8..cb6a87a 100644
--- a/llvm/lib/Target/X86/X86ISelLowering.cpp
+++ b/llvm/lib/Target/X86/X86ISelLowering.cpp
@@ -47269,12 +47269,18 @@ static SDValue combineVectorInsert(SDNode *N, SelectionDAG &DAG,
                                    TargetLowering::DAGCombinerInfo &DCI,
                                    const X86Subtarget &Subtarget) {
   EVT VT = N->getValueType(0);
-  assert(((N->getOpcode() == X86ISD::PINSRB && VT == MVT::v16i8) ||
-          (N->getOpcode() == X86ISD::PINSRW && VT == MVT::v8i16) ||
-          N->getOpcode() == ISD::INSERT_VECTOR_ELT) &&
+  unsigned Opcode = N->getOpcode();
+  assert(((Opcode == X86ISD::PINSRB && VT == MVT::v16i8) ||
+          (Opcode == X86ISD::PINSRW && VT == MVT::v8i16) ||
+          Opcode == ISD::INSERT_VECTOR_ELT) &&
          "Unexpected vector insertion");
 
-  if (N->getOpcode() == X86ISD::PINSRB || N->getOpcode() == X86ISD::PINSRW) {
+  // Fold insert_vector_elt(undef, elt, 0) --> scalar_to_vector(elt).
+  if (Opcode == ISD::INSERT_VECTOR_ELT && N->getOperand(0).isUndef() &&
+      isNullConstant(N->getOperand(2)))
+    return DAG.getNode(ISD::SCALAR_TO_VECTOR, SDLoc(N), VT, N->getOperand(1));
+
+  if (Opcode == X86ISD::PINSRB || Opcode == X86ISD::PINSRW) {
     unsigned NumBitsPerElt = VT.getScalarSizeInBits();
     const TargetLowering &TLI = DAG.getTargetLoweringInfo();
     if (TLI.SimplifyDemandedBits(SDValue(N, 0),
diff --git a/llvm/test/CodeGen/X86/avx-insertelt.ll b/llvm/test/CodeGen/X86/avx-insertelt.ll
index 23b1a03..d3ac3f1 100644
--- a/llvm/test/CodeGen/X86/avx-insertelt.ll
+++ b/llvm/test/CodeGen/X86/avx-insertelt.ll
@@ -422,7 +422,7 @@ define <8 x i32> @insert_i32_two_elts_of_high_subvector(<8 x i32> %x, i32 %s) {
 define <4 x i64> @insert_i64_two_elts_of_high_subvector(<4 x i64> %x, i64 %s) {
 ; AVX-LABEL: insert_i64_two_elts_of_high_subvector:
 ; AVX:       # %bb.0:
-; AVX-NEXT:    vpinsrq $0, %rdi, %xmm0, %xmm1
+; AVX-NEXT:    vmovq %rdi, %xmm1
 ; AVX-NEXT:    vpinsrq $1, %rdi, %xmm1, %xmm1
 ; AVX-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm0
 ; AVX-NEXT:    retq
@@ -523,7 +523,7 @@ define <8 x i32> @insert_i32_two_elts_of_low_subvector(<8 x i32> %x, i32 %s) {
 define <4 x i64> @insert_i64_two_elts_of_low_subvector(<4 x i64> %x, i64 %s) {
 ; AVX-LABEL: insert_i64_two_elts_of_low_subvector:
 ; AVX:       # %bb.0:
-; AVX-NEXT:    vpinsrq $0, %rdi, %xmm0, %xmm1
+; AVX-NEXT:    vmovq %rdi, %xmm1
 ; AVX-NEXT:    vpinsrq $1, %rdi, %xmm1, %xmm1
 ; AVX-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
 ; AVX-NEXT:    retq