user type annotations are captured post normalization

[nikomatsakis]

This example compiles but should not. Haven't investigated deeply.

#![feature(nll)]

trait Foo { 
    type Item;
}

impl<'a, u32> Foo for &'a u32 {
    type Item = &'a i32;
}

fn main() {
    let a = 22_i32;
    let x: <&'static u32 as Foo>::Item = &a;
}

cc #47184

[nikomatsakis]

I think in general we probably need to rework a bit how the NLL checker is handling user-given type annotations to cover normalizations. My expectation is that we should move over to the strategy of instantiating the user-types with fresh type variables so that we can run the normalize routine on them, and then adapt the relate_tys code to handle unbound type variables. We've done most of the legwork here already so that should be too hard to do.

[pnkfelix]

Discussed at NLL weekly meeting. Assigning to @nikomatsakis as primary person to resolve this. Assigning to self as a kind of backup plan since I know that @nikomatsakis has some conflicts this week that will impede their ability to actually hack on this problem in the short term.

[nikomatsakis]

The problem is that we are capturing these types after they've been normalized. I think the best fix would be to capture the types before they've been normalized, and then have the NLL checker do the normalization. I'm not 100% sure how much of a pain this is going to be though; it might be a bit of a pain in some cases.

[nikomatsakis]

It also would require #55093 to land first.

[nikomatsakis]

OK, digging a bit deeper. This is going to be an awful pain to fix =)