[llvm] [AArch64][SVE] Add partial reduction SDNodes (PR #117185)

[Sander de Smalen via llvm-commits]

================
@@ -22011,138 +22010,188 @@ static SDValue tryCombineWhileLo(SDNode *N,
   return SDValue(N, 0);
 }
 
-SDValue tryLowerPartialReductionToDot(SDNode *N,
-                                      const AArch64Subtarget *Subtarget,
-                                      SelectionDAG &DAG) {
-
-  assert(N->getOpcode() == ISD::INTRINSIC_WO_CHAIN &&
-         getIntrinsicID(N) ==
-             Intrinsic::experimental_vector_partial_reduce_add &&
-         "Expected a partial reduction node");
-
-  bool Scalable = N->getValueType(0).isScalableVector();
-  if (Scalable && !Subtarget->isSVEorStreamingSVEAvailable())
+SDValue tryCombinePartialReduceMLAMulOp(SDValue &Op0, SDValue &Op1,
+                                        SDValue &Op2, SelectionDAG &DAG,
+                                        SDLoc &DL) {
+  // Makes PARTIAL_REDUCE_MLA(Acc, MUL(EXT(MulOpLHS), EXT(MulOpRHS)), Splat (1))
+  // into PARTIAL_REDUCE_MLA(Acc, EXT(MulOpLHS), EXT(MulOpRHS))
+  if (Op1->getOpcode() != ISD::MUL)
     return SDValue();
-  if (!Scalable && (!Subtarget->isNeonAvailable() || !Subtarget->hasDotProd()))
+
+  SDValue ExtMulOpLHS = Op1->getOperand(0);
+  SDValue ExtMulOpRHS = Op1->getOperand(1);
+  unsigned ExtMulOpLHSOpcode = ExtMulOpLHS->getOpcode();
+  unsigned ExtMulOpRHSOpcode = ExtMulOpRHS->getOpcode();
+  if (!ISD::isExtOpcode(ExtMulOpLHSOpcode) ||
+      !ISD::isExtOpcode(ExtMulOpRHSOpcode))
     return SDValue();
 
-  SDLoc DL(N);
+  SDValue MulOpLHS = ExtMulOpLHS->getOperand(0);
+  SDValue MulOpRHS = ExtMulOpRHS->getOperand(0);
+  EVT MulOpLHSVT = MulOpLHS.getValueType();
+  if (MulOpLHSVT != MulOpRHS.getValueType())
+    return SDValue();
----------------
sdesmalen-arm wrote:

With the way you've currently written this function, my understanding is that this code can be removed, because this code does not try to fold any of the extends into the PARTIAL_REDUCE operation. i.e. al it is doing is:
```
PARTIAL_REDUCE_UMLA(acc, MUL(lhs, rhs), SPLAT(1))
->
PARTIAL_REDUCE_UMLA(acc, lhs, rhs)
```

https://github.com/llvm/llvm-project/pull/117185