[llvm] [InstCombine] Ensure Safe Handling of Flags in foldFNegIntoConstant (PR #94148)

[Matt Arsenault via llvm-commits]

================
@@ -2653,8 +2653,15 @@ static Instruction *foldFNegIntoConstant(Instruction &I, const DataLayout &DL) {
   // Fold negation into constant operand.
   // -(X * C) --> X * (-C)
   if (match(FNegOp, m_FMul(m_Value(X), m_Constant(C))))
-    if (Constant *NegC = ConstantFoldUnaryOpOperand(Instruction::FNeg, C, DL))
+    if (Constant *NegC = ConstantFoldUnaryOpOperand(Instruction::FNeg, C, DL)) {
+      if (match(C, m_AnyZeroFP()) && I.getFastMathFlags().noInfs()) {
----------------
arsenm wrote:

Directionally this feels wrong. A constant, but not known literal 0/-0 may end up resolving to a 0 (e.g. degenerate constant expression casts or undef). I think this should first be written to handle the flags conservatively regardless of the value and we can try to refine for the known-non-0 cases if necessary 

https://github.com/llvm/llvm-project/pull/94148