[clang] [Clang][Sema] Allow counted_by on void* in GNU mode (PR #164737)

[Yeoul Na via cfe-commits]

================
@@ -132,9 +132,20 @@ bool Sema::CheckCountedByAttrOnField(FieldDecl *FD, Expr *E, bool CountInBytes,
     // `BoundsSafetyCheckUseOfCountAttrPtr`
     //
     // * When the pointee type is always an incomplete type (e.g.
-    // `void`) the attribute is disallowed by this method because we know the
-    // type can never be completed so there's no reason to allow it.
-    InvalidTypeKind = CountedByInvalidPointeeTypeKind::INCOMPLETE;
+    // `void` in strict C mode) the attribute is disallowed by this method
+    // because we know the type can never be completed so there's no reason
+    // to allow it.
+    //
+    // Exception: In GNU mode, void has an implicit size of 1 byte for pointer
+    // arithmetic. Therefore, counted_by on void* is allowed as a GNU extension
+    // and behaves equivalently to sized_by (treating the count as bytes).
+    bool IsVoidPtrInGNUMode = PointeeTy->isVoidType() && getLangOpts().GNUMode;
----------------
rapidsna wrote:

I agree with @kees. From my perspective, allowing `counted_by` on `void *` without `-fbounds-safety` means we would have to allow it with `-fbounds-safety` as well (in GNU mode), since we want the same annotated headers to be consumable by the compiler with or without `-fbounds-safety` (except in some inevitable cases like inline function bodies in headers).

I think this is actually a reasonable compromise because treating `void *` as having a stride of `1` is a de facto standard in C.

However, in general, we cannot change the behavior of -fbounds-safety solely because the Linux community pushes back strongly. I think the issue is that they don't have the full context for -fbounds-safety, so decisions about individual attributes without that broader context can seem overly opinionated (while this change isn't not unreasonable). I will be working to improve this situation.

https://github.com/llvm/llvm-project/pull/164737