Matching of Self doesn't take lifetime parameters into account

[pcwalton]

The following code compiles:

struct Foo<'a,'b> {
    x: &'a int,
    y: &'b int,
}

trait Tr {
    fn foo(x: Self) {}
}

impl<'a,'b> Tr for Foo<'a,'b> {
    fn foo(x: Foo<'b,'a>) {} // <-- bad
}

fn main(){}

Nominating for 1.0, P-backcompat-lang. I may well just fix this along with #14022, since the solution is likely going to be similar to what @nikomatsakis pointed out, but I want to get it on the radar.

[pcwalton]

Hmm, OK. So the code does attempt to handle this, but I don't know if it is handled in the right way. Specifically, it creates fresh region variables for each region and checks to see whether they unify. The trouble is that Foo<'a,'b> does unify with Foo<'b,'a>—it unifies if and only if 'a and 'b are the same region. But this is precisely what @nikomatsakis claimed to be a problem in #14022. So I'm not sure what the right thing is to do here.

[pcwalton]

I'm going to assign to @nikomatsakis because I'm blocked on his opinion on what to do here.

[pnkfelix]

Assigning P-backcompat-lang, 1.0.

[pcwalton]

Right thing to do is to use ReSkolemized.

[nikomatsakis]

Argh. The right fix is to substitute some skolemized lifetimes in for the early bound regions before doing the unification rather than fresh variables.

[nikomatsakis]

Actually I think the problem is that we're just not calling resolve_regions_and_report_errors on the inference context