Regression in new inference scheme

[bkoropoff]

I was having a great deal of difficulty getting my library (https://github.com/bkoropoff/rust-jlens) to build on newer versions of rustc, even after updating it to deal with generalized lifetime bounds. Reverting to a rustc prior to the new inference scheme resolved the issue. I came up with the following reduced test case:

struct Chain<'a:'b,'b> {
    x: &'a int,
    y: Option<&'b Chain<'a,'b>>
}

impl<'a,'b> Chain<'a,'b> {
    fn leaf(x: &'a int) -> Chain<'a,'b> { Chain { x: x, y: None } }
}

fn callme<'a,'b> (y: Chain<'a,'b>, x: &'a int, f: <'c>|Chain<'a,'c>|) {
    f(Chain { x: x, y: Some(&y) })
}

fn main() {
    let x = 5i;
    let y = Chain::leaf(&x);
    callme(y, &x, |_| {});
}

This seems to build fine prior to the new inference scheme landing and gives the following error after:

test.rs:12:30: 12:31 error: `y` does not live long enough
test.rs:12     f(Chain { x: x, y: Some(&y) })
                                        ^
test.rs:11:71: 13:2 note: reference must be valid for the lifetime 'b as defined on the block at 11:70...
test.rs:11 fn callme<'a,'b> (y: Chain<'a,'b>, x: &'a int, f: <'c>|Chain<'a,'c>|) {
test.rs:12     f(Chain { x: x, y: Some(&y) })
test.rs:13 }
test.rs:11:71: 13:2 note: ...but borrowed value is only valid for the block at 11:70
test.rs:11 fn callme<'a,'b> (y: Chain<'a,'b>, x: &'a int, f: <'c>|Chain<'a,'c>|) {
test.rs:12     f(Chain { x: x, y: Some(&y) })
test.rs:13 }

cc @nikomatsakis

[nikomatsakis]

Without having investigated deeply, I think the problem is #3598. But I will investigate now to be sure.

[nikomatsakis]

Preliminary investigation suggests that yes, the issue is #3598. In this case, the old type checker was erroneously accepting the program. The reason is that chain winds up being invariant w/r/t 'b because Option is invariant with respect to its type argument. I've often worked around this locally by not using Option and making a specialized type. Here is an example from the Rust code base:

https://github.com/rust-lang/rust/blob/master/src/librustc/middle/resolve_lifetime.rs#L58

That said, I am hoping to open an RFC about support covariance soon. I want to do some local experiments, but I think the codebase is now at a point where it should be easy to introduce.

[nikomatsakis]

In my branch which fixes #3598, this test compiles.

[bkoropoff]

@nikomatsakis Is your branch for issue #3598 published anywhere yet? It would be interesting to try building my library with it. In any case, I'll try your suggestion of using a specialized enum. Thanks for taking a look at this so quickly!

[bkoropoff]

Using a specialized enum worked. Since this is not actually a regression but the result of an apparent soundness bug being fixed, closing issue.