Warning shown despite --cap-lints=allow: E0122 "trait bounds are not (yet) enforced in type definitions"

[SimonSapin]

Test case:

type Foo<T: Clone> = Option<T>;

fn main() {}

With rustc 1.20.0-nightly (696412d 2017-07-06), output when compiling with --cap-lints=allow:

warning[E0122]: trait bounds are not (yet) enforced in type definitions
 --> a.rs:1:1
  |
1 | type Foo<T: Clone> = Option<T>;
  | ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^

--cap-lints=allow correctly stops the dead code / unused type alias warning from being shown, but the trait bounds not enforced warning is shown anyway.

Per #34596 and #40640 (comment), this might be on purpose. I think this is the wrong decision. The point of cap-lints is to allow Cargo to inhibit warnings in external dependencies. When building Servo it is not useful to spam the console with many warnings, future-proofing or not, about code that Servo has not direct control over.

CC @nikomatsakis @aturon

[nagisa]

Fairly related discussion

[Mark-Simulacrum]

Nominating for discussion. Not sure if compiler team or lang team is best, though, so tagging both...

[aturon]

From the lang team's point of view: definitely a bug.

[nikomatsakis]

Couple of bugs here:

First off, we should not be ignoring the bounds in types. And sometimes they are needed. I'd like to fix this by not ignoring them, but I worry about the backcompat issues. Unknown risk there.

Second, if we must warn, we should be emitting a regular lint that can be silenced.

Plausibly, we should just remove the warning (eliminating the immediate problem), and prioritize trying to do a proper fix here.

[eddyb]

Somewhat related to not ignoring bounds (but in the other direction): #44075.

[nikomatsakis]

After some discussion in compiler team, decided to move this to a "future compatibility" lint, one that we will hopefully make more precise over time. Right now the check is not very good. At least this means you won't have to see it. We need some mentoring instructions though!

[nikomatsakis]

Some tips here: https://forge.rust-lang.org/rustc-bug-fix-procedure.html

[nikomatsakis]

triage: P-medium

[weiznich]

I've hit this recently.
It seems like it is possible to write code that requires exactly this bound. So it is quite bad that we cannot suppress this warning.

Example code:



trait Foo {
    type A;
}

struct StructA;
struct StructB;

impl Foo for StructA {
    type A = ();
}

// Warning about not enforced trait bounds
type B<A: Foo> = A::A;

// Error with the message:  associated type `A` not found for `A`
// -> Trait bound is required
//type C<A> = A::A

type D = B<StructA>;
// As the warning says this works
type E = B<StructB>;

[nikomatsakis]

Agreed, this is poor. There just hasn't been a lot of energy to try and improve the situation.

[Enselic]

Fixed in nightly-2017-09-22, probably by #44627. We should add a regression test for this particular case before we close this issue however, since I don't think the test in the PR detects a regression for the problem described in this issue. The problem in this issue is output on stderr, not that compilation fails (rustc exit code is 0 both with and without the fix in the PR).

@rustbot label E-needs-test