[llvm] [LLVM][AArch64] Refactor lowering of fixed length integer setcc operations. (PR #132434)

[Paul Walker via llvm-commits]

================
@@ -25268,6 +25216,14 @@ static SDValue performSETCCCombine(SDNode *N,
   if (SDValue V = performOrXorChainCombine(N, DAG))
     return V;
 
+  EVT CmpVT = LHS.getValueType();
+
+  APInt SplatLHSVal;
+  if (CmpVT.isInteger() && Cond == ISD::SETGT &&
+      ISD::isConstantSplatVector(LHS.getNode(), SplatLHSVal) &&
+      SplatLHSVal.isOne())
+    return DAG.getSetCC(DL, VT, DAG.getConstant(0, DL, CmpVT), RHS, ISD::SETGE);
----------------
paulwalker-arm wrote:

I added this combine because it proved awkward to detect splat(1) across all the NEON types within a PatFrag.  I think this is again because we're performing isel of constants during legalisation when it would be better to use splat_vector universally.  Please let me know if you've a better solution, otherwise I'm happy to circle back after making the DAG more amenable.

https://github.com/llvm/llvm-project/pull/132434