DynMetadata equality is not documented

[alercah]

Location

https://doc.rust-lang.org/std/ptr/struct.DynMetadata.html

Summary

There is no documentation as far as I should see that defines the behaviour of the PartialEq impl for DynMetadata. I think the guarantees that are made or not made should be documented. For instance, are vtables for a given concrete type guaranteed to be unique? People will assume they are, and that == will tell you if the types match. It would not surprise me to learn that this is wrong.

[SimonSapin]

Vtables are likely not unique at the moment. Yes, documenting that explicitly sounds good. Related: #69757

[nbdd0121]

Is it necessary to have PartialEq implementaton on DynMetadata?

[SimonSapin]

All raw pointers (including to trait objects) implement PartialEq, and they conceptually do it by comparing (ptr as *const (), metadata(ptr)). For that formulation to work, we’d need DynMetadata: PartialEq. This is currently reflected in Pointee::Metadata having a PartialEq bound.

However the current implementation of raw pointer comparison is just bitwise comparison and does not go through metadata(), so it would technically be possible to remove this bound and impl. I’m not sure it’s desirable, though.

[RalfJung]

One can compare vtable pointers via wide raw pointer comparison, so for better or worse I think we should also allow it on DynMetadata. That might be a good place to document the footguns related to comparison of vtables.