ICE with lifetime-generic type alias impl trait

[jonas-schievink]

#![feature(type_alias_impl_trait)]

pub type Successors<'a> = impl Iterator<Item = &'a ()>;

pub fn f<'a>() -> Successors<'a> {
    None.into_iter()
}

trait Tr {
    type Item;
}

impl<'a> Tr for &'a () {
    type Item = Successors<'a>;
}

fn ohno<'a>() -> <&'a () as Tr>::Item {
    None.into_iter()
}

gives:

error: internal compiler error: free_scope: DefId(0:3 ~ playground[ad24]::Successors[0]) not recognized by the region scope tree for None / Some(DefId(0:13 ~ playground[ad24]::ohno[0]))

My guess is that this has the same underlying cause as #56445, but this one doesn't involve associated consts or arrays at all, so it might be different still.

[Aaron1011]

This appears to be due to a bad interaction between #66178 and associated type projections involving lifetimes.

When we normalize the function signature during type-checking, we project <&'a () as Tr>::Item to Successors<'_#0r>. We then convert '_#0r back to the 'generic' region for Successors, which is 'a.

However, this 'a is a distinct early bound region, different from the 'a in fn ohno<'a>() (changing the regions to not all be called 'a in the example makes this clearer). This causes issues during lexical_region_resolve, since we've essentially pulled a region out of thin air.

This issue doesn't seem to come up for type parameters (playground with lifetimes replaced with types). I'm not sure if this is because our handling of type parameters is correct for this kind of projection, or if we need a more complicated example to break things.

[JohnTitor]

The ICE no longer happens with the latest nightly (note that the playground's nightly isn't the latest). That delay_span_bug was removed by #72362. Marked as E-needs-test.