Unclear introduction of 'static with APIT

[CAD97]

Given the code:

struct Event;
struct Span;

impl Event {
    fn span(&self) -> &Span {
        &Span
    }
}

impl Span {
    fn fields(&self) -> impl '_ + Iterator<Item = (&'static str, &'_ str)> {
        None.into_iter()
    }
}

fn show_event<'a>(event: &'a Event) {
    show_fields(event.span().fields());
}

fn show_fields<'a, 'data>(fields: impl 'a + Iterator<Item = (&'data str, &'data str)>) {
    for (_name, _value) in fields {
        todo!();
    }
}

The current output is:

error[E0759]: `event` has lifetime `'a` but it needs to satisfy a `'static` lifetime requirement
  --> src/lib.rs:17:23
   |
16 | fn show_event<'a>(event: &'a Event) {
   |                          --------- this data with lifetime `'a`...
17 |     show_fields(event.span().fields());
   |                 ----- ^^^^
   |                 |
   |                 ...is captured here...
   |
note: ...and is required to live as long as `'static` here
  --> src/lib.rs:17:5
   |
17 |     show_fields(event.span().fields());
   |     ^^^^^^^^^^^

It's not immediately clear from this error where the 'static bound is introduced, especially when not as reduced. I only noticed that the 'static was probably being introduced by the APIT during reduction to file this issue.

It would be helpful to have a note/hint on the argument of the function that introduces the 'static bound. (And, as a side note, I'm not even sure if the 'static bound can be eliminated.)

This was tested to be the output on 1.56.0-nightly (2021-09-01 50171c3).

[CAD97]

Thanks to users on urlo, we've both figured out exactly where the 'static obligation comes from as well as how to avoid it.

This is a tricky situation to figure out. In summary:

• Span::fields is returning impl 'a + Iterator<Item = (&'static str, &'a str)>, and
• show_fields wants impl Iterator<Item = (&'d str, &'d str)>.

Because return position impl Trait is invariant (because it could potentially be fulfilled by a type which is invariant over the lifetime(s)), we fail to unify the two types, because we can't have 'd = 'static and 'd = 'a at the same time.

rustc then chooses to explain the mismatch as "has lifetime 'a but needs to satisfy 'static, which is... not all that illuminating to where the 'static came from, thus the diagnostic issue.

I do not know how best to expose this logical chain to the user, but I think it's necessary to give more information to pin in onto variance. The error message "blames" show_fields for introducing the 'static obligation, where it's not really at fault, since it doesn't require the argument is 'static; the 'static comes from trying to unify an invariant short lifetime with an invariant long lifetime.

[BGR360]

@rustbot label +A-variance