== fails to match types on associated type with two PartialEq<A / B> bounds

[SimonSapin]

rustc 1.12.0-nightly (47b3a98 2016-07-10)

pub struct A;
pub struct B;

pub trait Foo {
    type T: PartialEq<A> + PartialEq<B>;
}

pub fn generic<F: Foo>(t: F::T, a: A, b: B) -> bool {
    t == a && t == b
}

error: mismatched types [--explain E0308]
  --> a.rs:12:10
   |>
12 |>     t == a && t == b
   |>          ^ expected struct `B`, found struct `A`
note: expected type `B`
note:    found type `A`

error: aborting due to previous error

Changing the order of bounds in the associated type declaration changes which == causes a spurious "mismatched types" error.

Using UFCS works around the issue:

    PartialEq::<A>::eq(&t, &a) && t == b

---

The same issue exists with bounds like type T: PartialEq + PartialEq<B> to compare T with itself as well as another type.

[rkruppe]

cf. #24066

[eddyb]

Looks like this is caused by using find on the set of bounds from the trait, but that is just find the first one (in the order they are internally stored), this needs to filter and produce a vector of bounds or something similar.

cc @rust-lang/compiler

[rozbb]

Update: using filter and then calling candidates.extend on all matching bounds does not fix this issue.

[eddyb]

@doomrobo I believe your problem is this logic.
Right now it works fine because the if victim.candidate == other.candidate above would trigger for ObjectCandidate vs ObjectCandidate or ParamCandidate vs ParamCandidate.

However, it's incorrect with your changes as ParamCandidate(bound_a) is dropped in favor of ParamCandidate(bound_b) even if bound_a and bound_b are not equal.
The correct form would be more like this:

                // Arbitrarily give param candidates priority
                // over projection and object candidates.
                ObjectCandidate => true,
                ProjectionCandidate(_) => {
                    if let ProjectionCandidate(_) = other.candidate {
                        false
                    } else {
                        true
                    }
                },

[nikomatsakis]

I guess that you didn't just call filter, but are also packaging up the specific bound that was matched? Anyway, not having seen the full diff, the general direction of these changes make sense to me.

[nikomatsakis]

Related: #34912

[pnkfelix]

triage: P-medium