#[track_caller] should inherit through specialization ancestors, too?

[anp]

The current implementation of #[track_caller] allows trait methods to have the annotation, including at their definition site. When applied in a trait definition, the attribute has effect on all impls. @eddyb raised the question in code review of whether you should also be able to add the attribute without adding an implementation as part of a specialization chain.

To me this seems like the correct semantics given what we'd like the non-specialized behavior to be. It's probably a little complicated to implement.

[eddyb]

whether you should also be able to add the attribute without adding an implementation as part of a specialization chain

To be clear, what I was thinking of is something like this:

trait Foo {
    fn foo();
}

impl<T> Foo for T {
    #[track_caller]
    default fn foo() { panic!("default") }
}

impl Foo for () {
    fn foo() { panic!("final") }
}

IIUC (and I suppose you could add a testcase like this), the current implementation results in the "final" method not being #[track_caller], only the T != () "default" will have it.

And my suggestion was that long-term we might want "final" to get #[track_caller] from specializing "default".

The implementation might not be difficult, I just am not sure what the "ancestors" specialization graph traversal lets you do, yet.
It might be even simpler than the current implementation, which looks only at the trait for an inherited #[track_caller] (as specialization infra may do things like matching by name, all by itself).

[anp]

Hm, I see yeah. This "just works" for default trait impls/overriding trait impls right now, so I imagine the trick would be to run should_inherit_track_caller in a loop on specialization ancestors?