rapidsna wrote: FWIW, in the context of `-fbounds-safety` (https://github.com/llvm/llvm-project/blob/main/clang/docs/BoundsSafety.rst), I think the best is to have the bounds-annotated types to be part of the canonical type system because they affect the semantics and the codegen (again with `-fbounds-safety`) and the annotations should also apply to rvalues (e.g., function return types).
That said, sugar types happen to work fine overall because where the compiler is about to lose the type sugar (e.g., assigning it to a variable with type that doesn't have the same attribute), the model naturally prevented it by either emitting an error for incompatible type conversion, or promoting the type into a wide pointer type (i.e., `__bidi_indexable` for lvalue-to-rvalue conversion on `__counted_by`). That being said, we still encountered some cases where the type sugars are lost inconsistently due to implementation issues (e.g., getting unqualified type also removes the type sugar depending on the position it appertains to). So again, ideally, I like it to be part of the type system. However, without `-fbounds-safety`, `__counted_by` is merely used as a helper for sanitizers and the `__builtin_dynamic_object_size`, so it would make more sense to be a sugar type. Then, the question is whether we want the attribute to be parsed different with or without `-fbounds-safety` and would that be a problem. https://github.com/llvm/llvm-project/pull/78000 _______________________________________________ cfe-commits mailing list cfe-commits@lists.llvm.org https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
