[llvm-branch-commits] [llvm] InstCombine: Handle fdiv in SimplifyDemandedFPClass (PR #175946)

[Matt Arsenault via llvm-branch-commits]

================
@@ -2331,6 +2331,136 @@ Value *InstCombinerImpl::SimplifyDemandedUseFPClass(Instruction *I,
     FPClassTest ValidResults = DemandedMask & Known.KnownFPClasses;
     return getFPClassConstant(VTy, ValidResults, /*IsCanonicalizing=*/true);
   }
+  case Instruction::FDiv: {
+    Value *X = I->getOperand(0);
+    Value *Y = I->getOperand(1);
+    if (X == Y && isGuaranteedNotToBeUndef(X, SQ.AC, CxtI, SQ.DT, Depth + 1)) {
+      // If the source is 0, inf or nan, the result is a nan
+
+      Value *IsZeroOrNan = Builder.CreateFCmpFMF(
+          FCmpInst::FCMP_UEQ, I->getOperand(0), ConstantFP::getZero(VTy), FMF);
+
+      Value *Fabs =
+          Builder.CreateUnaryIntrinsic(Intrinsic::fabs, I->getOperand(0), FMF);
+      Value *IsInfOrNan = Builder.CreateFCmpFMF(
+          FCmpInst::FCMP_UEQ, Fabs, ConstantFP::getInfinity(VTy), FMF);
+
+      Value *IsInfOrZeroOrNan = Builder.CreateOr(IsInfOrNan, IsZeroOrNan);
----------------
arsenm wrote:

That doesn't have the same behavior. is.fpclass doesn't canonicalize the input, emitting fcmp gets the correct behavior regardless of denormal mode. instcombine will only fold to is.fpclass depending on the mode 

https://github.com/llvm/llvm-project/pull/175946