[llvm] [AArch64] Use [SU]ADALP for partial (nx)v4i32 -> (nx)v2i64 add reduce. (PR #213915)

[Paul Walker via llvm-commits]

================
@@ -34755,11 +34743,23 @@ AArch64TargetLowering::LowerPARTIAL_REDUCE_MLA(SDValue Op,
 
   bool IsUnsigned = Op.getOpcode() == ISD::PARTIAL_REDUCE_UMLA;
 
+  // Attempt to fold v4i32 -> v2i64 via Neon *ADALP.
+  if (Subtarget->isNeonAvailable() && OrigResultVT == MVT::v2i64) {
+    unsigned Opc = IsUnsigned ? AArch64ISD::UADDLP : AArch64ISD::SADDLP;
+    if (ConvertToScalable)
+      DotNode = convertFromScalableVector(DAG, MVT::v4i32, DotNode);
+    SDValue Res = DAG.getNode(Opc, DL, OrigResultVT, DotNode);
+    return DAG.getNode(ISD::ADD, DL, OrigResultVT, OrigAcc, Res);
+  }
----------------
paulwalker-arm wrote:

Can you move this handling to the top? We already have a block for v16i8->v2i32, which I think you can repeat for v16i8->v2i64 when NEON is available and thus bypass SVE altogether.

https://github.com/llvm/llvm-project/pull/213915