[X86] getFauxShuffleMask - always match insert_subvector(insert_subvector(undef,sub,0),sub,c) 'subvector splat' patterns (#130115)

The plan is to remove the vXi64 cross lane shuffle constraint entirely, but this special 'splat' case was easy to handle while I fight the remaining regressions.

Author: RKSimon

llvm/lib/Target/X86/X86ISelLowering.cpp

@@ -6153,12 +6153,13 @@ static bool getFauxShuffleMask(SDValue N, const APInt &DemandedElts,
       return true;
     }
     // Handle CONCAT(SUB0, SUB1).
-    // Limit this to vXi64 vector cases to make the most of cross lane shuffles.
+    // Limit to vXi64/splat cases to make the most of cross lane shuffles.
     if (Depth > 0 && InsertIdx == NumSubElts && NumElts == (2 * NumSubElts) &&
-        NumBitsPerElt == 64 && Src.getOpcode() == ISD::INSERT_SUBVECTOR &&
+        Src.getOpcode() == ISD::INSERT_SUBVECTOR &&
         Src.getOperand(0).isUndef() &&
         Src.getOperand(1).getValueType() == SubVT &&
         Src.getConstantOperandVal(2) == 0 &&
+        (NumBitsPerElt == 64 || Src.getOperand(1) == Sub) &&
         SDNode::areOnlyUsersOf({N.getNode(), Src.getNode()}, Sub.getNode())) {
       for (int i = 0; i != (int)NumSubElts; ++i)
         Mask.push_back(i);

llvm/test/CodeGen/X86/vector-partial-undef.ll

@@ -150,8 +150,7 @@ define <8 x i32> @xor_undef_elts_alt(<4 x i32> %x) {
 ; AVX-LABEL: xor_undef_elts_alt:
 ; AVX:       # %bb.0:
 ; AVX-NEXT:    # kill: def $xmm0 killed $xmm0 def $ymm0
-; AVX-NEXT:    vinsertf128 $1, %xmm0, %ymm0, %ymm0
-; AVX-NEXT:    vmovaps {{.*#+}} ymm1 = [6,1,5,4,3,2,0,7]
+; AVX-NEXT:    vmovaps {{.*#+}} ymm1 = [2,1,1,0,3,2,0,3]
 ; AVX-NEXT:    vpermps %ymm0, %ymm1, %ymm0
 ; AVX-NEXT:    vxorps {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm0, %ymm0
 ; AVX-NEXT:    retq