[llvm] [X86] Fold VPERMV(MASK,CONCAT(LO,HI)) -> VPERMV3(WIDEN(LO),MASK',WIDEN(HI)) (PR #129708)

[Simon Pilgrim via llvm-commits]

================
@@ -42607,6 +42607,43 @@ static SDValue combineTargetShuffle(SDValue N, const SDLoc &DL,
 
     return SDValue();
   }
+  case X86ISD::VPERMV: {
+    // Combine VPERMV to VPERMV3 if the source operand can be freely split.
+    SmallVector<int, 32> Mask;
+    SmallVector<SDValue, 2> SrcOps, SubOps;
+    SDValue Src = peekThroughBitcasts(N.getOperand(1));
+    if ((Subtarget.hasVLX() ||
+         (VT.is512BitVector() && Subtarget.hasAVX512())) &&
----------------
RKSimon wrote:

We already have a 512-bit VPERMV node at this point - useAVX512Regs shouldn't be necessary, but we can add it as an additional precaution if you want.

https://github.com/llvm/llvm-project/pull/129708