[llvm] ValueTracking: Improve handling of fadd in computeKnownFPClass. (PR #174123)

[Matt Arsenault via llvm-commits]

================
@@ -5640,18 +5644,29 @@ void computeKnownFPClass(const Value *V, const APInt &DemandedElts,
         if (KnownLHS.cannotBeOrderedLessThanZero() &&
             KnownRHS.cannotBeOrderedLessThanZero())
           Known.knownNot(KnownFPClass::OrderedLessThanZeroMask);
+        if (KnownLHS.cannotBeOrderedGreaterThanZero() &&
+            KnownRHS.cannotBeOrderedGreaterThanZero())
+          Known.knownNot(KnownFPClass::OrderedGreaterThanZeroMask);
+
         if (!F)
           break;
 
         const fltSemantics &FltSem =
             Op->getType()->getScalarType()->getFltSemantics();
         DenormalMode Mode = F->getDenormalMode(FltSem);
 
+        // Doubling 0 will give the same 0.
+        if (SelfAdd && KnownRHS.isKnownNeverLogicalPosZero(Mode) &&
+            (Mode.Output == DenormalMode::IEEE ||
+             Mode.Output == DenormalMode::PreserveSign))
+          Known.knownNot(fcPosZero);
+
         // (fadd x, 0.0) is guaranteed to return +0.0, not -0.0.
         if ((KnownLHS.isKnownNeverLogicalNegZero(Mode) ||
              KnownRHS.isKnownNeverLogicalNegZero(Mode)) &&
             // Make sure output negative denormal can't flush to -0
-            outputDenormalIsIEEEOrPosZero(*F, Op->getType()))
+            (Mode.Output == DenormalMode::IEEE ||
----------------
arsenm wrote:

No? If the sign is preserved this will flush to -0

https://github.com/llvm/llvm-project/pull/174123