[llvm] [SelectionDAG] Fix fcmp fold for new min/max semantics (PR #223655)

[Matt Arsenault via llvm-commits]

================
@@ -7040,12 +7036,20 @@ static SDValue foldAndOrOfSETCC(SDNode *LogicOp, SelectionDAG &DAG) {
                                TLI.isOperationLegal(ISD::FMINNUM_IEEE, OpVT);
   bool isFMAXNUMFMINNUM = TLI.isOperationLegalOrCustom(ISD::FMAXNUM, OpVT) &&
                           TLI.isOperationLegalOrCustom(ISD::FMINNUM, OpVT);
+  // A custom FMINIMUMNUM/FMAXIMUMNUM may be more expensive than the compares,
+  // so only use one if the target considers forming it profitable.
+  bool isFMAXIMUMNUMFMINIMUMNUM =
+      (TLI.isOperationLegal(ISD::FMAXIMUMNUM, OpVT) &&
+       TLI.isOperationLegal(ISD::FMINIMUMNUM, OpVT)) ||
+      (TLI.isOperationLegalOrCustom(ISD::FMAXIMUMNUM, OpVT) &&
+       TLI.isOperationLegalOrCustom(ISD::FMINIMUMNUM, OpVT) &&
+       TLI.isProfitableToCombineMinNumMaxNum(OpVT));
----------------
arsenm wrote:

I am disappointed that isProfitableToCombineMinNumMaxNum exists, and I think it should be removed. Given the X86 comment in it's implementation, and that PSEUDO_FMIN/FMAX now exist, I think we would be better off removing it and adding the PSEUDO_* cases to these combines as preferred if available 

https://github.com/llvm/llvm-project/pull/223655