covariant method return types on trait impls?

[pnkfelix]

Do trait method signatures have to match exactly, or is there some subtyping relationship that is supposed to hold between an impl's methods and the methods of the trait definition itself?

If there is a subtyping relationship, then I would expect the following to compile:

#![crate_type="lib"]

// illustrating subtype relationship: typeof(g) <: typeof(f) ...
fn foo<'a,'b>(ignored: Option<&'a &'b ()>, f: || -> &'a int) { unimplemented!() }
fn bar<'a,'b>(ignored: Option<&'a &'b ()>, g: || -> &'b int) { foo(ignored, g) }

// ... but does that subtype relationship apply to the return types here?
trait T       { fn m<'a,'b>(self, ignored: Option<&'a &'b ()>) -> &'a int; }
impl T for () { fn m<'a,'b>(self, ignored: Option<&'a &'b ()>) -> &'b int { unimplemented!() } }

(it currently does not compile; it complains of the mismatch between the lifetime parameters.)

If there is not a subtyping relationship, then I would not expect the following to compile:

#![crate_type="lib"]

trait T<'a>                       { fn m(&'a self) -> &'a int; }
impl<'ignored> T<'ignored> for () { fn m<'a>(&'a self) -> &'a int { unimplemented!() } }

(it currently does compile.)

Note that the latter case seems sound to me, but the fact that this works in the latter case but does not for the former case just makes me worry about what the rules actually are that we are enforcing with respect to how much an impl needs to conform with its corresponding trait.

[pnkfelix]

cc @nikomatsakis

[pnkfelix]

((The semi-good news is that whatever the oddity is that leads us to accept the second example in the description, whatever that oddity is, I have not yet been able to construct an example via that oddity that breaks type soundness. Not yet.))

[nikomatsakis]

@pnkfelix there is a subtyping relationship, but I don't think it properly takes the relationships between free lifetimes into account.

[steveklabnik]

Triage: another very old ticket. @nikomatsakis @pnkfelix has this question been resolved over the past few years?

[nikomatsakis]

I think @pnkfelix's confusion may in fact be #18937, which I am supposed to be fixing. In any case, I'm going to close this issue, since I think the big question (must impl signatures match exactly) is basically de facto settled to "no, some amount of subtyping is permitted".