[llvm] Take poison-generating attributes into account (PR #89138)

[Nikita Popov via llvm-commits]

================
@@ -496,14 +496,23 @@ class Instruction : public User,
   /// Drops metadata that may generate poison.
   void dropPoisonGeneratingMetadata();
 
-  /// Return true if this instruction has poison-generating flags or metadata.
-  bool hasPoisonGeneratingFlagsOrMetadata() const {
-    return hasPoisonGeneratingFlags() || hasPoisonGeneratingMetadata();
+  /// Return true if this instruction has poison-generating attribute.
+  bool hasPoisonGeneratingAttribute() const LLVM_READONLY;
----------------
nikic wrote:

```suggestion
  bool hasPoisonGeneratingReturnAttributes() const LLVM_READONLY;
```
Maybe? Just to make clear this won't report these attributes in argument position.

https://github.com/llvm/llvm-project/pull/89138