[llvm] [AArch64] Optimize two large shifts and a combine into a single combine and shift (PR #99480)

[David Green via llvm-commits]

================
@@ -14198,9 +14198,43 @@ SDValue AArch64TargetLowering::LowerCONCAT_VECTORS(SDValue Op,
                                    !Subtarget->isNeonAvailable()))
     return LowerFixedLengthConcatVectorsToSVE(Op, DAG);
 
-  assert(Op.getValueType().isScalableVector() &&
-         isTypeLegal(Op.getValueType()) &&
-         "Expected legal scalable vector type!");
+  if (!Op.getValueType().isScalableVector()) {
----------------
davemgreen wrote:

Could this be done in performConcatVectorsCombine instead? It should hopefully be very similar code, but avoids the need to mark the nodes as custom.

https://github.com/llvm/llvm-project/pull/99480