[llvm] [AArch64][SVE2p1] Allow more uses of mask in performActiveLaneMaskCombine (PR #159360)

[Kerry McLaughlin via llvm-commits]

================
@@ -18774,7 +18774,7 @@ static SDValue performVecReduceAddCombineWithUADDLP(SDNode *N,
 static SDValue
 performActiveLaneMaskCombine(SDNode *N, TargetLowering::DAGCombinerInfo &DCI,
                              const AArch64Subtarget *ST) {
-  if (DCI.isBeforeLegalize())
+  if (DCI.isBeforeLegalize() && !!DCI.isBeforeLegalizeOps())
----------------
kmclaughlin-arm wrote:

I've now removed this altogether, since finding that a concat with 4+ inputs will be split after this combine without changing the stage at which it applies (although I've also made changes to create multiple concat_vectors in the latest commit too).

https://github.com/llvm/llvm-project/pull/159360