Beta Regression: cannot reference Self from where clause for inherent impl

[bytwise]

The following code compiles with rustc 1.0.0-beta (9854143cb 2015-04-02) (built 2015-04-02) but does not compile with rustc 1.0.0-beta.2 (e9080ec39 2015-04-16) (built 2015-04-16)

trait Foo {
    fn foo(&self);
}

struct Bar<T>(T);

impl<T> Bar<T> where Self: Foo {}

fn main() {}

[pnkfelix]

Specifically, the error message one gets (at least with my beta build) is:

% ./x86_64-apple-darwin/stage1/bin/rustc /tmp/r.rs
/tmp/r.rs:7:22: 7:26 error: use of undeclared type name `Self`
/tmp/r.rs:7 impl<T> Bar<T> where Self: Foo {}
                                 ^~~~
error: aborting due to previous error

[pnkfelix]

However, this may be deliberate. At least, I don't remember what the rules are about when Self is in scope, if at all, within impls -- I tend to only use it within trait definitions, myself.

In other words, one can certainly workaround this issue by writing the above impl as:

impl<T> Bar<T> where Bar<T>: Foo {}

[bombless]

No way, this is definitely wrong.
Self is a keyword, the compiler should never print error message like that.

[pnkfelix]

@bombless there's a difference nonetheless between a poor error message and an (unintentional) regression in the compiler itself. I'm trying to identify exactly what this is.

[pnkfelix]

(I have a strong suspicion that this change was injected by PR #23998, which was a fix for #23909. I am doing some builds to try to verify that suspicion now.)

[pnkfelix]

(no, from locally building at 926f38e it seems like PR #23998 is not the source of the "regression" here.)

Update: above was based on faulty input test case; i need to go back and check it again.

Update 2: Okay, it was definitely injected by 926f38e and thus due to PR #23998.

[Stebalien]

It is according to rust-lang/rfcs#522.

[pnkfelix]

@Stebalien it is (what) according to rust-lang/rfcs#522 ? Supposed to be allowed? Behaving as specified?

Update: In particular, there is plenty of wiggle room in rust-lang/rfcs#522 as written. It just says "When used inside an impl ..." without formally defining what "inside an impl" actually means, and thus a reasonable person might think it means "inside the curly braces associated with the impl"

Having said that, I am inclined to infer that you meant that RFC 522 is supposed to imply that this usage of Self is legal, and I personally am inclined to agree with that (inferred) opinion.

[arielb1]

@bombless

This is just the standard error message you get when you try to use Self outside of a trait, e.g.

fn foo(_: Self) {}

[Stebalien]

@Stebalien it is (what) according to rust-lang/rfcs#522 ? Supposed to be allowed? Behaving as specified?

IMHO, it should be allowed.

[pnkfelix]

cc @nrc : was this an intentional side-effect of #23998, or is this an inadvertent bug that we should fix?

(If its a bug, I'm assuming we can prioritizing fixing it based on how much breakage we observe on crates.io ?)

[bombless]

@arielb1
You're right, filed an issue #24968

[nrc]

@pnkfelix it was not deliberate. I don't really have an opinion on whether it is right or wrong as stands, I guess if you can write a bound on the concrete version of Self, you should be able to use Self itself.