[llvm-branch-commits] [llvm] InstCombine: Implement SimplifyDemandedFPClass for sqrt (PR #173883)

[Matt Arsenault via llvm-branch-commits]

================
@@ -2244,6 +2244,37 @@ Value *InstCombinerImpl::SimplifyDemandedUseFPClass(Value *V,
       FPClassTest ValidResults = DemandedMask & Known.KnownFPClasses;
       return getFPClassConstant(VTy, ValidResults, /*IsCanonicalizing=*/true);
     }
+    case Intrinsic::sqrt: {
+      FPClassTest DemandedSrcMask =
+          DemandedMask & (fcNegZero | fcPositive | fcNan);
+
+      if (DemandedMask & fcNan)
+        DemandedSrcMask |= (fcNegative & ~fcNegZero);
+
+      KnownFPClass KnownSrc;
+      if (SimplifyDemandedFPClass(I, 0, DemandedSrcMask, KnownSrc, Depth + 1))
+        return I;
+
+      Type *EltTy = VTy->getScalarType();
+      DenormalMode Mode = F.getDenormalMode(EltTy->getFltSemantics());
+
+      // sqrt(-x) = nan, but be careful of negative subnormals flushed to 0.
+      if (KnownSrc.isKnownNever(fcPositive) &&
+          KnownSrc.isKnownNeverLogicalZero(Mode))
+        return ConstantFP::getQNaN(VTy);
----------------
arsenm wrote:

No, that would only apply if the result had an implied sign. This is just performing the negative input - > nan fold 

https://github.com/llvm/llvm-project/pull/173883