Trait bounds on struct type parameters aren't always enforced (depending on the order of impls)

[SSheldon]

The following compiles without issue, even though the impl of Trait2 is missing a trait bound that T implements Bound:

pub trait Bound { }
pub struct Foo<T: Bound>;

pub trait Trait1 { }
impl<T: Bound> Trait1 for Foo<T> { }

pub trait Trait2 { }
impl<T> Trait2 for Foo<T> { }

fn main() { }

However, if you switch the order of the impls...

pub trait Bound { }
pub struct Foo<T: Bound>;

pub trait Trait2 { }
impl<T> Trait2 for Foo<T> { }

pub trait Trait1 { }
impl<T: Bound> Trait1 for Foo<T> { }

fn main() { }

You suddenly start getting this error:

error: the trait `Bound` is not implemented for the type `T` [E0277]
impl<T> Trait2 for Foo<T> { }
^~~~~~~~~~~~~~~~~~~~~~~~~~~~~
error: aborting due to previous error

I did not expect the ordering of the impls to matter here. I'm using rustc 1.0.0-nightly (1d00c545e 2015-01-30 19:56:34 +0000).

[pnkfelix]

P-high, not 1.0. (But definitely a weird bug that we would like to fix.)

[nikomatsakis]

cc me

[SSheldon]

@nikomatsakis, @pnkfelix: should this be P-backcompat-lang since the compiler accepts code that it should not?

[pnkfelix]

@SSheldon I think we decided this is one of those edge cases where the amount of breakage (that fixing this is expected to incur) was deemed acceptable for fixing post 1.0.

At the time this was last triaged, there was a general rule that if something was tagged P-backcompat-lang, then it must also be set for the 1.0 milestone (at the latest).

However, since then during recent triage meetings, we have been more willing to tag things as P-backcompat-lang even if they were not attached to the 1.0 milestone.

So, with that in mind:

triage: P-backcompat-lang, not P-high

[arielb1]

This causes an RFC1214 warning in 1.4+.

[nikomatsakis]

triage: P-medium

Looks like this is already fixed by RFC 1214, as @arielb1 reports.

[nikomatsakis]

Tagging as E-needstest, since we should be able to add a test with WARN messages (and perhaps a #[rustc_error] annotation on main)

[nikomatsakis]

Tagging as mentor just because I can give advice on how to craft a suitable test (nmatsakis on IRC)

[bltavares]

Triaging:

Both examples causes error on nightly rustc 1.11.0-nightly (5c2a5d449 2016-06-11), but this seems to be the expected behaviour.

<anon>:2:16: 2:17 error: parameter `T` is never used [E0392]
<anon>:2 pub struct Foo<T: Bound>;
                        ^
<anon>:2:16: 2:17 help: see the detailed explanation for E0392
<anon>:2:16: 2:17 help: consider removing `T` or using a marker such as `core::marker::PhantomData`
<anon>:8:9: 8:15 error: the trait `Bound` is not implemented for the type `T` [E0277]
<anon>:8 impl<T> Trait2 for Foo<T> { }
                 ^~~~~~
<anon>:8:9: 8:15 help: see the detailed explanation for E0277
<anon>:8:9: 8:15 note: required by `Foo`
error: aborting due to 2 previous errors

@nikomatsakis I would like to help with the test if you don't mind giving initial directions.

[SSheldon]

@bltavares, that repro case was from pre-1.0, here's the updated repro case:

pub trait Bound { }
pub struct Foo<T: Bound>(T);

pub trait Trait1 { }
impl<T: Bound> Trait1 for Foo<T> { }

pub trait Trait2 { }
impl<T> Trait2 for Foo<T> { }

fn main() { }

That code will compile with rustc 1.0.0 (a59de37e9 2015-05-13), but fails to compile with rustc 1.9.0 (e4e8b6668 2016-05-18).

[brson]

@arielb1 This seems to have been fixed by RFC 1214. Can it be closed?

[pnkfelix]

@brson I think it needs a test. I probably should not be assigned to it since I don't plan to spend time making a test.

[martinhath]

So all there is to do is add a test which checks that Ssheldons code does not compile?

[nikomatsakis]

@martinhath yes :)

[SSheldon]

Fixed now with a test merged in #36941