[llvm] [X86] Fold generic ADD/SUB with constants to X86ISD::SUB/ADD (PR #164316)

[via llvm-commits]

================
@@ -57631,8 +57631,25 @@ static SDValue combineX86AddSub(SDNode *N, SelectionDAG &DAG,
       DCI.CombineTo(GenericAddSub, Op);
     }
   };
-  MatchGeneric(LHS, RHS, false);
-  MatchGeneric(RHS, LHS, X86ISD::SUB == N->getOpcode());
+  MatchGeneric(GenericOpc, LHS, RHS, false);
+  MatchGeneric(GenericOpc, RHS, LHS, X86ISD::SUB == N->getOpcode());
+
+  if (ConstantSDNode *Const = dyn_cast<ConstantSDNode>(RHS)) {
+    SDValue NegC = DAG.getConstant(-Const->getAPIntValue(), DL, VT);
+    if (X86ISD::SUB == N->getOpcode()) {
+      // With LHS - C, fold LHS + (-C)
----------------
brandonxin wrote:

What we see here is an `X86ISD::SUB`, which sets `eflags` differently than `X86ISD::ADD`, so an `X86ISD::SUB LHS, C` cannot be canonicalized to `X86ISD::ADD LHS, -C`. Therefore we explicitly check for the `X86ISD::SUB` pattern here.

In contrast, a generic `ISD::SUB LHS, C` is equivalent to `ISD::ADD LHS, -C` and is canonicalized by the DAG, so we don't need to look for the `ISD::SUB LHS, C` to fold.

https://github.com/llvm/llvm-project/pull/164316