object lifetime bound defaults in trait bindings are incompletely implemented

[nikomatsakis]

In #63376, we fixed the behavior of default object lifetime bounds in "binding arguments", as described in this gist. For example the following type;

dyn Foo<Bar = dyn Baz>

will now default dyn Baz to dyn Baz + 'static.

However, the rules were incompletely implemented. If the trait has a lifetime parameter, we currently require an explicit lifetime bound. The expected behavior is to use the bounds declared on the associated type to set the default. So, given Foo like so:

trait Foo<'a> {
  type Bar: 'a + ?Sized;
}

then dyn Foo<'x, Bar = dyn Baz> would default to dyn Baz + 'a. Presently, however, it gives an error.

NB. There is one "design decision" to be made; does a where clause like where Self::Bar: 'x adjust the default? And how does that work when extended to GATs? My personal opinion is that we should limit ourselves to bounds declared on the associated type itself (and hence where Self::Bar: 'x would not count towards the default object lifetime bound), but this merits a bit of further discussion. This is largely because the complexity to specify and implement the behavior around where clauses and with GATs is kind feels out of proportion with the benefit.

[nikomatsakis]

(Note also: PR #63376 has not landed yet and this behavior is not finalized as of this writing. I just wanted an issue I could link to.)

[Centril]

When considering object lifetime bound defaults holistically, I think we should also consider the interactions with associated type defaults and trait objects. cc #29661.