[llvm] [AArch64] Eliminate Common Subexpression of CSEL by Reassociation (PR #121350)

[Marius Kamp via llvm-commits]

================
@@ -24838,6 +24838,122 @@ static SDValue foldCSELOfCSEL(SDNode *Op, SelectionDAG &DAG) {
   return DAG.getNode(AArch64ISD::CSEL, DL, VT, L, R, CCValue, Cond);
 }
 
+// Reassociate the true/false expressions of a CSEL instruction to obtain a
+// common subexpression with the comparison instruction. For example, change
+// (CSEL (ADD (ADD x y) -c) f LO (SUBS x c)) to
+// (CSEL (ADD (SUBS x c) y) f LO (SUBS x c)) such that (SUBS x c) is a common
+// subexpression.
+static SDValue reassociateCSELOperandsForCSE(SDNode *N, SelectionDAG &DAG) {
+  SDValue SubsNode = N->getOperand(3);
+  if (SubsNode.getOpcode() != AArch64ISD::SUBS || !SubsNode.hasOneUse())
+    return SDValue();
+  auto *CmpOpConst = dyn_cast<ConstantSDNode>(SubsNode.getOperand(1));
+  if (!CmpOpConst)
+    return SDValue();
+
+  auto CC = static_cast<AArch64CC::CondCode>(N->getConstantOperandVal(2));
+  bool IsEquality = CC == AArch64CC::EQ || CC == AArch64CC::NE;
+  if (IsEquality && !CmpOpConst->isZero())
----------------
mskamp wrote:

Yes, we can also handle eq and ne. I've also implemented this.

I agree that smaller patches are sometimes better. I'll take a look into non-constant operands, which would probably require some more restructuring, and start working on a follow-up patch. For this PR, it is probably easier to stick to constants to keep the PR concise.

https://github.com/llvm/llvm-project/pull/121350