[llvm] [InstCombine] Fix for folding select-like `shufflevector` into floating point binary operators. (PR #85452)

[Matt Arsenault via llvm-commits]

================
@@ -2159,6 +2160,18 @@ static Instruction *foldSelectShuffleWith1Binop(ShuffleVectorInst &Shuf) {
   if (!IdC)
     return nullptr;
 
+  Value *X = Op0IsBinop ? Op1 : Op0;
+
+  // Prevent folding in the case the non-binop operand might have NaN values.
+  // If X can have NaN elements then we have that the floating point math
+  // operation in the transformed code may not preserve the exact NaN
+  // bit-pattern -- e.g. `fadd sNaN, 0.0 -> qNaN`.
+  // This makes the transformation incorrect since the original program would
+  // have preserved the exact NaN bit-pattern.
+  // Avoid the folding if X can have NaN elements.
+  if (Shuf.getType()->isFPOrFPVectorTy() && !isKnownNeverNaN(X, 0, SQ))
----------------
arsenm wrote:

Shut can't be a scalar FP type, so could just query the element type 

https://github.com/llvm/llvm-project/pull/85452