[llvm] [AArch64][SME] Enable dynamic shuffle for fixed length types. (PR #72490)

[Sander de Smalen via llvm-commits]

================
@@ -26123,26 +26123,50 @@ static SDValue GenerateFixedLengthSVETBL(SDValue Op, SDValue Op1, SDValue Op2,
 
   // Ignore two operands if no SVE2 or all index numbers couldn't
   // be represented.
-  if (!IsSingleOp && (!Subtarget.hasSVE2() || MinSVESize != MaxSVESize))
+  if (!IsSingleOp && !Subtarget.hasSVE2())
     return SDValue();
 
   EVT VTOp1 = Op.getOperand(0).getValueType();
   unsigned BitsPerElt = VTOp1.getVectorElementType().getSizeInBits();
   unsigned IndexLen = MinSVESize / BitsPerElt;
   unsigned ElementsPerVectorReg = VTOp1.getVectorNumElements();
   uint64_t MaxOffset = APInt(BitsPerElt, -1, false).getZExtValue();
+  EVT MaskEltType = VTOp1.getVectorElementType().changeTypeToInteger();
+  EVT MaskType = EVT::getVectorVT(*DAG.getContext(), MaskEltType, IndexLen);
+  bool MinMaxEqual = (MinSVESize == MaxSVESize);
   assert(ElementsPerVectorReg <= IndexLen && ShuffleMask.size() <= IndexLen &&
          "Incorrectly legalised shuffle operation");
 
   SmallVector<SDValue, 8> TBLMask;
+  // If MinSVESize is not equal to MaxSVESize then we need to know which
+  // TBL mask element needs adjustment.
+  SmallVector<SDValue, 8> MulByVLMask;
----------------
sdesmalen-arm wrote:

nit: Should this be named `AddRuntimeVLMask` (because it _adds_ the runtime vector length to the indices)

https://github.com/llvm/llvm-project/pull/72490