[llvm] [KnownFPClass] Refine positive zero and NaN result classes for sqrt (PR #214987)

[Matt Arsenault via llvm-commits]

================
@@ -556,15 +556,22 @@ KnownFPClass KnownFPClass::sqrt(const KnownFPClass &KnownSrc,
   if (KnownSrc.isKnownNever(fcSNan))
     Known.knownNot(fcSNan);
 
-  // Any negative value besides -0 returns a nan.
-  if (KnownSrc.isKnownNeverNaN() && KnownSrc.cannotBeOrderedLessThanZero())
+  // Any value < -0.0 returns NaN. However, a negative subnormal may be treated
+  // as zero depending on the input denormal mode.
+  if (KnownSrc.isKnownNever(fcNan | fcNegInf | fcNegNormal) &&
+      (KnownSrc.isKnownNever(fcNegSubnormal) || Mode.inputsAreZero()))
----------------
arsenm wrote:

This is a stronger use of the denormal mode than anywhere else here. Generally it's treated as a hazard; the value may be treated as 0 not that it must be. I think the langref rules assert that input flushing must definitely flush (compared to the lax output), but it's a fuzzy edge case. I'd prefer to take that conceptual leap separately 



https://github.com/llvm/llvm-project/pull/214987