Regression from 1.20 to 1.21 stable -> stable with StreamingIterator code

[mikeyhew]

The following code compiles with Rust version 1.20, but not with 1.21 and above:

pub trait StreamingIterator<'a>: 'a {
    type Item: 'a;

    fn next(&'a mut self) -> Option<Self::Item>;
}

pub trait StreamingIteratorExt: for<'a> StreamingIterator<'a> {}

impl<I> StreamingIteratorExt for I where
    I: for<'a>StreamingIterator<'a>
{}

Error in Rust 1.21+:

error[E0310]: the parameter type `I` may not live long enough
  --> src/lib.rs:14:9
   |
14 | impl<I> StreamingIteratorExt for I where
   |      -  ^^^^^^^^^^^^^^^^^^^^
   |      |
   |      help: consider adding an explicit lifetime bound `I: 'static`...
   |
note: ...so that the type `I` will meet its required lifetime bounds
  --> src/lib.rs:14:9
   |
14 | impl<I> StreamingIteratorExt for I where
   |         ^^^^^^^^^^^^^^^^^^^^

Note the Self: 'a bound on StreamingIterator. If I remove it, the error goes away.

[Mark-Simulacrum]

@eddyb Could this be the same root cause as this regression even though it has a different version rang on it?

[eddyb]

I doubt it, empty lifetime vs 'static. I don't know what happens here, this is HRTB for<'a> T: 'a which I think is T: 'static. cc @nikomatsakis

[nikomatsakis]

triage: P-high

[nikomatsakis]

OK, so, I had not looked so closely at this example before. Now I see why the compiler is upset. There is this where-clause that I: for<'a> StreamingIterator<'a>. We currently enforce WF rules that state that, in this case, the where-clauses declared on the trait must hold. We translate this to imply that for<'a> I: 'a (since StreamingIterator<'a> has the where-clause Self: 'a). We then decide that the only way this could be true is if I: 'static. To some extent, the mystery is why this was ever accepted.

Now, as it happens, I think that this handling is not very good. The implied bounds and other work we've got going on would allow us to accept this program, I believe. But with the compiler's current handling of higher-ranked trait bounds, I think it's working as I expect.

@Mark-Simulacrum how far back can we bisect? Oh, we have a nightly bisection tool, right? I'd be interested (if the data can be readily obtained) to find out when this behavior changed.

[Mark-Simulacrum]

We can bisect 90 days back by commit, and with some work by nightly (i.e., different bisectors -- mine can do only 90 days as of now, whereas nightlies should still be possible indefinitely). I don't have time in the short term to run that bisection but @est31 might be able to...

[est31]

Sad to hear that it was decided to delete PR artifacts after 90 days... I think it should be at least 6 months, maybe even a year until deletion.

Well we were lucky that this particular regression is inside that window. The regressing commit:

426711d 'Auto merge of #43786 - scalexm:issue-43784, r=nikomatsakis"'