[llvm] [polly] [InstSimplify] Fix incorrect poison propagation when folding phi (PR #96631)

[Ralf Jung via llvm-commits]

================
@@ -5278,7 +5278,14 @@ static Value *simplifyPHINode(PHINode *PN, ArrayRef<Value *> IncomingValues,
     // If we have a PHI node like phi(X, undef, X), where X is defined by some
     // instruction, we cannot return X as the result of the PHI node unless it
     // dominates the PHI block.
-    return valueDominatesPHI(CommonValue, PN, Q.DT) ? CommonValue : nullptr;
+    if (!valueDominatesPHI(CommonValue, PN, Q.DT))
+      return nullptr;
+
+    // Make sure we do not replace an undef value with poison.
+    if (HasUndefInput &&
+        !isGuaranteedNotToBePoison(CommonValue, Q.AC, Q.CxtI, Q.DT))
----------------
RalfJung wrote:

This is even a double-negation ("not guaranteed to not be poison"), so it's definitely not just you^^

https://github.com/llvm/llvm-project/pull/96631