[llvm] [AArch64] Fix ReconstructShuffle for known vscale>1 (PR #205099)

[Gaëtan Bossu via llvm-commits]

================
@@ -14177,7 +14183,7 @@ SDValue AArch64TargetLowering::ReconstructShuffle(SDValue Op,
       continue;
     else if (V.getOpcode() != ISD::EXTRACT_VECTOR_ELT ||
              !isa<ConstantSDNode>(V.getOperand(1)) ||
-             V.getOperand(0).getValueType().isScalableVector()) {
+             !useNEONForVectorVT(V->getOperand(0).getValueType())) {
----------------
gbossu wrote:

There might be parts of `AArch64TargetLowering::ReconstructShuffle` that are safe for 256+ bits vectors, but tbh that function is very complex and I don't fully understand it. Unless there's a use case that comes up, I'd rather be safe and bail out for non-NEON sizes.

https://github.com/llvm/llvm-project/pull/205099