loss of implied bounds after normalization

[aliemjay]

This doesn't compile although it should:

trait Identity { type Ty; }
impl<'a, 'b> Identity for &'a &'b () { type Ty = &'a &'b (); }
fn test<'a, 'b>(_: <&'a &'b () as Identity>::Ty) {}
//~^ ERROR lifetime mismatch

It does pass if the impl is changed to

impl<T> Identity for T { type Ty = T; }

The issue here is that when we try to normalize <&'a &'b () as Identity>::Ty, the result is &'?1 &'?2 (), where ['?1, '?2] are new region variables that are unified with ['a, 'b], respectively. However, when we try to compute the implied bounds after normalization, we get '?2: '?1 instead of the desired bound 'b: 'a.

@rustbot label C-bug T-types A-implied-bounds

[aliemjay]

The PR was reverted.