Regression in associated type checking for default impls

[jsmith628]

So I had a trait system in my (nightly) code for a while now that utilized specialization, but it no longer seems to be working after a recent update:

#![feature(specialization)]

trait Trait1 { type Selection: PartialEq; }

trait Trait2: PartialEq<Self> {}
impl<T:Trait2> Trait1 for T { default type Selection = T; }

warning: the feature `specialization` is incomplete and may not be safe to use and/or cause compiler crashes
 --> src/main.rs:1:12
  |
1 | #![feature(specialization)]
  |            ^^^^^^^^^^^^^^
  |
  = note: `#[warn(incomplete_features)]` on by default
  = note: see issue #31844 <https://github.com/rust-lang/rust/issues/31844> for more information

error[E0277]: can't compare `T` with `<T as Trait1>::Selection`
 --> src/main.rs:6:31
  |
3 | trait Trait1 { type Selection: PartialEq; }
  |                -------------------------- required by `Trait1::Selection`
...
6 | impl<T:Trait2> Trait1 for T { default type Selection = T; }
  |                               ^^^^^^^^^^^^^^^^^^^^^^^^^^^ no implementation for `T == <T as Trait1>::Selection`
  |
  = help: the trait `std::cmp::PartialEq<<T as Trait1>::Selection>` is not implemented for `T`
help: consider further restricting this bound
  |
6 | impl<T:Trait2 + std::cmp::PartialEq<<T as Trait1>::Selection>> Trait1 for T { default type Selection = T; }
  |               ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^

error: aborting due to previous error; 1 warning emitted

For more information about this error, try `rustc --explain E0277`.

Now, I haven't been following the development of the specialization feature too closely, so it is possible that this is somehow actually intended and I am missing something. However, given that this previously compiled and ran just fine a couple weeks ago and given that the specialization incomplete warning was only added recently, I am inclined to think that it is probably a regression of some sort.

Meta things

compiler version:

$rustc --version --verbose
rustc 1.46.0-nightly (7750c3d46 2020-06-26)
binary: rustc
commit-hash: 7750c3d46bc19784adb1ee6e37a5ec7e4cd7e772
commit-date: 2020-06-26
host: x86_64-unknown-linux-gnu
release: 1.46.0-nightly
LLVM version: 10.0

[tmandry]

Another MVCE (playground):

#![feature(specialization)]

#[derive(PartialEq)]
enum Never {}

trait Foo {
    type Assoc: PartialEq;
}

impl<T> Foo for T {
    default type Assoc = Never;
}

Regressed in 7058471, looks like #72788

[tmandry]

cc @matthewjasper, as far as I can tell this wasn't an intended outcome of the change?

[matthewjasper]

It wasn't really, it was just slightly more complex to have this still work after that PR.

[tmandry]

It wasn't really, it was just slightly more complex to have this still work after that PR.

@matthewjasper Would you mind describing at a high level what needs to be done?

[tmandry]

It looks like the problem stems from the bound predicate not being fully normalized when checking the defaulted associated type. From my reproducer above:

compare_projection_bounds: normalized predicate =
    Binder(TraitPredicate(<Never as std::cmp::PartialEq<<T as Foo>::Assoc>>))

which should instead be <Never as std::cmp::PartialEq<Never>>.

I'm not certain yet how this ought to be implemented. Perhaps a pass after normalization that looks for occurrences of the defaulted type (<T as Foo>::Assoc in this case) and replaces it with the defaulted type. Seems simple enough, in theory :) But maybe there are cases I'm not thinking of.