Regression around `Trait::method -> Self`

[wycats]

I have this trait definition:

pub trait LoadableMessage {
    fn load<R: Reader>(reader: &mut R) -> IoResult<Self> {
        let mut stream = InputStream::new();
        LoadableMessage::load_from_stream(&mut stream)
    }
}

And this usage:

pub fn load<'a, M: LoadableMessage, R: Reader>(reader: &mut R) -> IoResult<M> {
    LoadableMessage::load(reader)
}

And I get this error upon compilation:

-gnu/release/deps --extern nix=/home/wycats/Code/skylight-rust/target/x86_64-unknown-linux-gnu/release/deps/libnix-13dc7f065ef11e3c.rlib`
/home/wycats/Code/skylight-rust/libs/buffoon/src/lib.rs:19:5: 19:34 error: mismatched types: expected `core::result::Result<M,std::io::IoError>`, found `core::result::Result<Self,std::io::IoError>` (expected type parameter, found Self)
/home/wycats/Code/skylight-rust/libs/buffoon/src/lib.rs:19     LoadableMessage::load(reader)
                                                               ^~~~~~~~~~~~~~~~~~~~~~~~~~~~~

This was compiling recently (I believe last week), but does not compile with today's nightly.

[nrc]

cc me

[nrc]

This is a weird bug: here is a minimal test case:

pub trait Foo {
    fn load_from() -> Box<Self>;
    fn load() -> Box<Self> {
        Foo::load_from()
    }
}

pub fn load<M: Foo>() -> Box<M> {
    Foo::load()
}

but, replace Foo::load_from() by fail!(); and this compiles just fine. And yet the error message is on the call to Foo::load() (and commenting that out, this compiles fine too).

[nrc]

And calling load_from directly works too - it seems to need the default method in there

[nrc]

The regressing patch seems to be ce342f5. Which is ... impossible? It only touches files in trans and has nothing to do with type inference. But, I have triple checked this - it really seems to be this commit which regresses the reduced test case.

[nrc]

And of course patching without that commit does not fix things, so I was right about it being impossible. Now I am just confused about my bisection.

[nrc]

OK, now I actually have a sensible candidate: 2c0f876

[nrc]

This is probably a dup of #18187

[nrc]

So, this (https://github.com/nick29581/rust/compare/rust-lang:master...nick29581:infer-bug?expand=1) fixes the problem, but I suspect it is not the right fix. I imagine the candidates.ambiguous = true; (probably, perhaps also the shallow_resolve, if that has side effects) should be moved out of this method some how. But I'm not clear on exactly how this should work. @nikomatsakis help?

[nikomatsakis]

Interesting, I'll look.

[nikomatsakis]

OK, so, what's happening is this:

1. When checking the default method definition, we see a call to $0::load_from(), where we don't know what $0 is. We look around and find only the where clause Self : Foo in scope, so we decide to use that.
2. Because $0 : Foo does not contain any type parameters, we place it in the global result cache.
3. Later, we find the result in the cache and hence do not search.

This is why the bug goes away when you remove the definition. It would also go away if there were at most one impl.

I'm debating what the best fix is. Clearly, when the result is due to a where clause, we should place the result in the local cache. That would fix the problem. I could also imagine that we might not want to employ a where clause when there are inference variables -- but then again, why not? If there are no impls, it's the only plausible way to resolve this trait obligation.