Silent demotion of lifetime in analysis causes confusing message

[dhardy]

See StackOverflow question.

Summary: the compiler accepts a trait impl function returning a &'static str when the trait requires only &'a str, 'a being the life of the object, but then assumes the lifetime of the returned value is only &'a str despite direct call (not via trait).

Possible solutions I see:

1. Do not demote the lifetime of a function when called via the object (not the trait), but keep both versions of the function (i.e. lifetime &'static when called via the object and lifetime &'a when called via the trait). This is likely confusing, so (2) may be better.
2. Make it an error to implement a trait function with too long a lifetime.

[huonw]

cc @nikomatsakis; this is the "subtyping" rule, right?

[nikomatsakis]

Yes, this is the subtyping rule. I'm not sure whether changes are needed here, but if we were to change anything I would want to get stricter about the correspondence between the trait and impl.

[nikomatsakis]

There are other places -- notably where-clauses -- where we absolutely cannot require a 1-to-1 correspondence. But in method signatures we likely could. It's a bit trickier than it might sound.

[nikomatsakis]

Nominating. We have to decide what semantics we want here I guess.

[pnkfelix]

P-high, not required for 1.0. (But we would take a patch if we happen to get one before 1.0; otherwise the current semantics is stuck).

[steveklabnik]

But we would take a patch if we happen to get one before 1.0; otherwise the current semantics is stuck).

Well, a loooong time after 1.0, I guess this is stuck here, closing.