Mismatched assumptions for interaction between associated type and lifetime in trait

[huonw]

pub trait Foo<'a> {
    type Bar: std::fmt::Debug;
    fn baz(&'a self) -> Self::Bar;
}

pub struct X {
    name : String,
}

impl<'a> Foo<'a> for X {
    type Bar = &'a String;
    fn baz(&'a self) -> &'a String {
        &self.name
    }
}

pub fn foo<T: for<'a> Foo<'a>>(x: T) {
    let y = x.baz();
    drop(x);
    // for T = X, this is a &String pointing to dropped memory
    println!("{:?}", y); 
}

pub fn main() {
    foo(X { name: "foo".to_string() });
}

This segfaults.

The impl is allowing &'a String because 'a is in scope, but the use in foo assumes that Bar is independent of 'a.

(Noticed in http://stackoverflow.com/q/29745134/1256624 .)

[huonw]

triage: P-high

cc @nikomatsakis

[bluss]

petgraph uses this kind of trait to provide generic "neighbor vertex" iterators.

[arielb1]

I think a cleaner example is:

pub trait Foo<'a> {
    type Bar;
}

impl<'a, T> Foo<'a> for T { type Bar = &'a T; }

fn denormalise<'a, T>(t: &'a T) -> <T as Foo<'a>>::Bar {
    t
}

pub fn free_and_use<T: for<'a> Foo<'a>,
                    F: for<'a> FnOnce(<T as Foo<'a>>::Bar)>(x: T, f: F) {
    let y;
    'body: loop { // lifetime annotations added for clarity
        's: loop { y = denormalise(&x); break }
        drop(x);
        return f(y);
    }
}

pub fn main() {
    free_and_use("foo".to_string(), |s| println!("{}", s));
}

It is clear here that the declaration and implementation of Foo is not doing anything wrong, but free_and_use wrongly assumes that <T as Foo<'s>>::Bar : 'body, which does not make a lot of sense.

[nikomatsakis]

Yes, I see the problem. It lies in some of the logic around type parameter outlives rules -- we assume type parameters outlive the current fn, and we assume the same of projections, which is clearly wrong. I've been wanting to revisit these rules about projections and how long they live (#23442) though not for soundness reasons.

[nikomatsakis]

triage: P-high

[nikomatsakis]

Just FYI -- I've been working steadily on this. I'm going to try and writeup my thoughts "soon" (famous last words).

[nikomatsakis]

This original example is now giving me an error on stable: http://is.gd/bidora

The question is, did we add a test?

[nikomatsakis]

The test associated-types-outlives.rs claims to be a regression test for this issue. Closing!