[X86] canonicalizeShuffleWithOp - add handling for X86ISD::VPERMV nodes (#127625)

Handle different src/mask operand ordering of X86ISD::VPERMV nodes

Author: RKSimon

llvm/lib/Target/X86/X86ISelLowering.cpp

@@ -41629,23 +41629,28 @@ static SDValue canonicalizeShuffleWithOp(SDValue N, SelectionDAG &DAG,
   case X86ISD::PSHUFD:
   case X86ISD::PSHUFHW:
   case X86ISD::PSHUFLW:
+  case X86ISD::VPERMV:
   case X86ISD::VPERMI:
   case X86ISD::VPERMILPI: {
-    if (N.getOperand(0).getValueType() == ShuffleVT &&
-        N->isOnlyUserOf(N.getOperand(0).getNode())) {
-      SDValue N0 = peekThroughOneUseBitcasts(N.getOperand(0));
+    unsigned SrcIdx = Opc == X86ISD::VPERMV ? 1 : 0;
+    if (N.getOperand(SrcIdx).getValueType() == ShuffleVT &&
+        N->isOnlyUserOf(N.getOperand(SrcIdx).getNode())) {
+      SDValue N0 = peekThroughOneUseBitcasts(N.getOperand(SrcIdx));
       unsigned SrcOpcode = N0.getOpcode();
       EVT OpVT = N0.getValueType();
       if (TLI.isBinOp(SrcOpcode) && IsSafeToMoveShuffle(N0, SrcOpcode)) {
         SDValue Op00 = peekThroughOneUseBitcasts(N0.getOperand(0));
         SDValue Op01 = peekThroughOneUseBitcasts(N0.getOperand(1));
-        bool FoldShuf = Opc != X86ISD::VPERMI;
+        bool FoldShuf = Opc != X86ISD::VPERMI && Opc != X86ISD::VPERMV;
         if (IsMergeableWithShuffle(Op00, FoldShuf) ||
             IsMergeableWithShuffle(Op01, FoldShuf)) {
           SDValue LHS, RHS;
           Op00 = DAG.getBitcast(ShuffleVT, Op00);
           Op01 = DAG.getBitcast(ShuffleVT, Op01);
-          if (N.getNumOperands() == 2) {
+          if (Opc == X86ISD::VPERMV) {
+            LHS = DAG.getNode(Opc, DL, ShuffleVT, N.getOperand(0), Op00);
+            RHS = DAG.getNode(Opc, DL, ShuffleVT, N.getOperand(0), Op01);
+          } else if (N.getNumOperands() == 2) {
             LHS = DAG.getNode(Opc, DL, ShuffleVT, Op00, N.getOperand(1));
             RHS = DAG.getNode(Opc, DL, ShuffleVT, Op01, N.getOperand(1));
           } else {
@@ -41661,11 +41666,13 @@ static SDValue canonicalizeShuffleWithOp(SDValue N, SelectionDAG &DAG,
       if (SrcOpcode == ISD::SINT_TO_FP && IsSafeToMoveShuffle(N0, SrcOpcode) &&
           OpVT.getScalarSizeInBits() ==
               N0.getOperand(0).getScalarValueSizeInBits()) {
-        SDValue Op00 = DAG.getBitcast(ShuffleVT, N0.getOperand(0));
-        SDValue Res =
-            N.getNumOperands() == 2
-                ? DAG.getNode(Opc, DL, ShuffleVT, Op00, N.getOperand(1))
-                : DAG.getNode(Opc, DL, ShuffleVT, Op00);
+        SDValue Res = DAG.getBitcast(ShuffleVT, N0.getOperand(0));
+        if (Opc == X86ISD::VPERMV)
+          Res = DAG.getNode(Opc, DL, ShuffleVT, N.getOperand(0), Res);
+        else if (N.getNumOperands() == 2)
+          Res = DAG.getNode(Opc, DL, ShuffleVT, Res, N.getOperand(1));
+        else
+          Res = DAG.getNode(Opc, DL, ShuffleVT, Res);
         Res = DAG.getBitcast(N0.getOperand(0).getValueType(), Res);
         return DAG.getBitcast(ShuffleVT, DAG.getNode(SrcOpcode, DL, OpVT, Res));
       }

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

@@ -151,9 +151,9 @@ define <8 x i32> @xor_undef_elts_alt(<4 x i32> %x) {
 ; AVX:       # %bb.0:
 ; AVX-NEXT:    # kill: def $xmm0 killed $xmm0 def $ymm0
 ; AVX-NEXT:    vinsertf128 $1, %xmm0, %ymm0, %ymm0
-; AVX-NEXT:    vxorps {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm0, %ymm0
 ; AVX-NEXT:    vmovaps {{.*#+}} ymm1 = [6,1,5,4,3,2,0,7]
 ; AVX-NEXT:    vpermps %ymm0, %ymm1, %ymm0
+; AVX-NEXT:    vxorps {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm0, %ymm0
 ; AVX-NEXT:    retq
   %extend = shufflevector <4 x i32> %x, <4 x i32> undef, <8 x i32> <i32 undef, i32 undef, i32 2, i32 3, i32 0, i32 1, i32 undef, i32 undef>
   %bogus_bo = xor <8 x i32> %extend, <i32 42, i32 43, i32 undef, i32 undef, i32 undef, i32 undef, i32 44, i32 12>

llvm/test/CodeGen/X86/vector-shuffle-combining.ll

@@ -2469,10 +2469,10 @@ define <8 x i32> @combine_unneeded_subvector1(<8 x i32> %a) {
 ;
 ; AVX2-FAST-ALL-LABEL: combine_unneeded_subvector1:
 ; AVX2-FAST-ALL:       # %bb.0:
-; AVX2-FAST-ALL-NEXT:    vpaddd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm0, %ymm0
 ; AVX2-FAST-ALL-NEXT:    vbroadcasti128 {{.*#+}} ymm1 = [7,6,5,4,7,6,5,4]
 ; AVX2-FAST-ALL-NEXT:    # ymm1 = mem[0,1,0,1]
 ; AVX2-FAST-ALL-NEXT:    vpermd %ymm0, %ymm1, %ymm0
+; AVX2-FAST-ALL-NEXT:    vpaddd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm0, %ymm0
 ; AVX2-FAST-ALL-NEXT:    retq
 ;
 ; AVX2-FAST-PERLANE-LABEL: combine_unneeded_subvector1: