[llvm-branch-commits] [llvm] InstCombine: Improve SimplifyDemandedFPClass min/max handling (PR #175382)

[Matt Arsenault via llvm-branch-commits]

================
@@ -2329,29 +2329,23 @@ Value *InstCombinerImpl::SimplifyDemandedUseFPClass(Instruction *I,
         KnownRHS.knownNot(fcNan);
       }
 
-      KnownFPClass::MinMaxKind OpKind;
+      bool OrderedZeroSign = !FMF.noSignedZeros();
----------------
arsenm wrote:

For this particular case it enables additional folds. But yes, we have an additional pre-existing bug that needs to be fixed in computeKnownFPClass for these. Not making it more conservative could introduce new poison results 

https://github.com/llvm/llvm-project/pull/175382