[llvm] [ValueTracking] Fix computeKnownFPClass handling of nsz (PR #186315)

[James Y Knight via llvm-commits]

================
@@ -5118,9 +5118,14 @@ void computeKnownFPClass(const Value *V, const APInt &DemandedElts,
   if (Depth == MaxAnalysisRecursionDepth)
     return;
 
+  // True only for ops where NSZ may relax the result zero sign.
+  bool AllowNSZResultRelaxation = false;
   const unsigned Opc = Op->getOpcode();
   switch (Opc) {
   case Instruction::FNeg: {
+    // fneg(fabs(...)) always produces -0 for zero inputs.
+    AllowNSZResultRelaxation =
+        !match(Op->getOperand(0), m_Intrinsic<Intrinsic::fabs>(m_Value()));
----------------
jyknight wrote:

What justifies such special case at all? AFAIK, "fneg nsz" is permitted to return either of plus or minus zero when it gets plus or minus zero inputs -- independent of what operation that value came from.

https://github.com/llvm/llvm-project/pull/186315