RFC: Allow re-exporting associated items with `use`

[Gankro]

Extend use to work in trait definitions and impl blocks for re-exporting associated
items under different names. This provides a clean way to rename items harmlessly, and
provides a more ergonomic way to provide the same functionality under different names
when necessary to satisfy an API.

This enables

trait Iterator {
    fn len_hint(&self) -> (usize, Option<usize>) {
        (0, None)
    }

    #[deprecated("renamed to len_hint")]
    use Self::len_hint as size_hint;
}

Rendered

[Gankro]

Added an example.

[seanmonstar]

I greatly prefer the alternative, where one could do fn foo = Self::bar or whatever.

[alexcrichton]

I would personally find this somewhat surprising in that an attribute is now greatly affecting name lookup and method resolution, where attributes normally don't have much effect on these sorts of phases of the compiler. I would also probably prefer a more first-class fn foo(...) = bar or something like that, but I also don't personally feel so motivated to add this at this time.

[tshepang]

I only know of 2:

• size_hint -> len_hint
• ExactSizeIterator -> ExactLenIterator

What are the others?

[Gankro]

• There's a lot of mut_* *_mut things you could argue.
• min_align_of is being renamed to align_of, removing align_of entirely: rust-lang/rust#25646
• connect -> join: #1102
• slice::tail/init: #1058

Just off the top of my head

[tshepang]

Some of those (I don't know about the rest) are Unstable though

[Gankro]

Only tail/init are unstable, and people are proposing renaming other methods in that RFC (first/last)

[tshepang]

diverse. Set --> diverse set

[petrochenkov]

use bar as foo; works today to some extent.
http://is.gd/SQpfGX

It's similar to the proposed fn foo = bar; because it works on the name level and can be adjusted to work in more places.

[eddyb]

Keep in mind this is an issue specifically for methods (the RFC could definitely be more clear about that).
Adding use syntax for methods could work, but the attribute method was proposed because it requires no syntax changes.

[jmesmon]

I'd be also be in favor of generalizing use to work for this case: it is what we would already use (no pun intended) in other cases for renaming things. Is there something specific about how methods (vs functions and other names) are treated that makes expanding use more problematic?

[bluss]

@jmesmon The old name needs to be deprecated and the new name not, and rust must not allow implementing both method names. If that works, then it's fine.

[bluss]

The attribute acts like a rename and deprecation of the old value. Deprecations already use attributes, so that part is fine. The novel thing is just redirecting an old name, specified in the rename, to the new method.

[arthurprs]

I'm strongly in favor of this as it allows easier decisions ongoing.

But shouldn't this (optionally) go into the deprecated attribute itself?

[jmesmon]

@bluss I suppose I'd expect one to have a non-deprecated implementation (of, in this case, the method) and a use to add the old, deprecated naming that was marked as deprecated. I suppose this also means we'd need to allow names created via use to be deprecated (if that isn't already allowed).

[llogiq]

Currently, this could be done with

#[deprecated(since = "1.2.3")]
fn deprecated_method(...) { replacement_method(...) }

I don't think that situation is so bad that we need a new language feature to handle it, but then I'm not the one maintaining the language. 😉

[bluss]

@llogiq That is not enough to solve size_hint renaming in a satisfactory way.

[CloudiDust]

I am for extending use (or fn, to a lesser extent). AFAIK, no renaming in Rust automatically deprecates the old name currently. And I am of the opinion that it should also be the case here. (Though an accompanying deprecation is usually wanted.) Then if all we do is a simple renaming, we should extend the existing language construct use, instead of introducing a new attribute.

@arthurprs, IMHO, extending deprecated would put the new name in the attribute and it will be harder to notice. Also the syntax may become weird if multiple renamings happen. A renamed attribute attached to the new name is clearer.

[Gankro]

I've completely rewritten the RFC to just use use per feedback.

[petrochenkov]

Hm, why do type aliases even exist if they just duplicate the functionality of use?

use B as A;
type A = B;

Well, because type can do smarter things than use, like:

type S<T> = SU<T, DefaultU>;

Can the same logic be applied to function aliases? Do we ever need something like:

fn f<T> = g<T, DefaultU>;
// or
fn h<'a> = k<'a, 'a>;

?
If yes, then a specialized tool for function aliases would be a better solution than use. (But it can, of course, coexist with use, like type does.)

[arthurprs]

This is good @Gankro 👍

[nrc]

-1 Why not just leave the deprecated method as just a call to the new method? This seems like a lot of added complexity (especially when reading code) for not very much gain. If this needs a solution (and I'm not sure it does), then it seems it should be done with tooling, rather than with a language feature.

[bluss]

@nrc: We need this for renaming/deprecating of methods that you usually chain / delegate off. The example that sparked this RFC was Iterator::size_hint.

[eddyb]

@nrc that doesn't work. However, a simple attribute for renaming a method would be enough. It would just work (and you should keep it unstable while designing a better solution).

If @Gankro wants to go the use route... It might be possible to get it somewhere. Maybe it could be useful for other things.

But it's a lot more specification and implementation work than this RFC assumes:
You have to make associated items of traits importable and exportable without breaking backwards compatibility (keep in mind use std::clone::Clone::*; compiles right now and imports nothing).
You have to allow non-re-export uses of use inside trait definitions.

Finally, you have to define re-export semantics in trait definitions that happen to also allow the kind of renaming desired here, which requires traits not only to expose those items, but also catch their implementations (only when the re-export is same-trait).

What would re-exporting something that isn't an associated item do? Just behave as if the trait was a module, and ignore them on method lookup or in <T as Trait>::associated paths?

In any case, 👎 for this RFC as it is written right now, it's missing a lot of details and it would not be prudent to accept it in this state. That said, I'm warming up to this use of use, as long as all ramifications are investigated.

[nrc]

"doesn't work" seems to mean causes some loss of efficiency and possible confusion for implementers. The trade off for size_hint in particular seems to be - having a sub-optimal name or some degree of performance loss (and implementer confusion), satisfying that trade-off doesn't seem worth extra language-level complexity.

[pnkfelix]

Am I right in my interpretation that everyone who is saying "that doesn't work" in response to @llogiq and @nrc is basing their statement on a (hypothetical) requirement that when deprecating an API due to a trivial rename, one wants to enforce that old_name and new_name are in sync?

NOTE: I used the word "trivial" not to make the problem sound easy, but rather to stress the fact that connecting up such synchronization in that specific case (where it is just a rename, no parameter reordering or conversions) seems to be tightly wound up with the feature being added here.

It seems to me like there are other ways, perhaps lint-based, to enforce such synchronization that, IMO, would have a prayer of generalizing to more complex scenarios.

Update: Off the top of my head, here's one idea: If the deprecated trait method can be expressed in terms of a default method that calls the new one in some manner, then we add an attribute to the deprecated method in the trait that says "warn on override". Where would that go wrong ? Implementers would be told they need to now implement the new method (if they have not already), and they would be told not to provide their own impl of the defaulted one anymore.

[nikomatsakis]

Hear ye, hear ye. This RFC is now entering final comment period.

[bluss]

@nrc: Ah but correcting a slightly worse name doesn't warrant the trouble that an imperfect rename involves. I think we need “perfect” name deprecations to even consider deprecations just because of these kinds of small blemishes in the varnish.

What I think of renaming size_hint specifically is that we shouldn't waste time renaming things that are so out of the way and are neither a user learning issue nor a user misunderstanding issue.

[Gankro]

@pnkfelix Yes, this RFC is largely trying to address the problem of keeping old and new names in sync, particularly when previous defaults prevent you from just building a default cycle or redirection.

[liigo]

+1, this is a nice solution

[nikomatsakis]

Using use feels pretty inconsistent to me. If nothing else, this feels like it ought to be a pub use --- but in general I feel like use is used to bring things into scope. It seems very reasonable to me that one might want to allow use in traits/impls just for that purpose, as you can with blocks.

This comment is not meant as an endorsement of the underlying functionality. I have to think over how much work and complexity it is (as @eddyb suggests).

[huonw]

I like this functionality in theory. But it is relatively complicated and only serves... one use-case (size_hint) now, and the circumstances where it is needed seem relatively rare? Are there others that people have in mind?

[arthurprs]

Not really sure but we could generalize it to support renaming traits and structures as well.

[eddyb]

@arthurprs But that's already possible!
Or do you mean deprecating a pub use while keeping the original item undeprecated?
In which case, if it doesn't work already, it's an order of magnitude less work (and I wouldn't even open a RFC for that, call it a bug fix).

[Gankro]

I have definitely deprecated re-exports of modules and types in std before.

[arthurprs]

Ya, I'm felling stupid right now, sorry.

[huonw]

@Gankro it doesn't work.

// deprecated.rs
#![crate_type = "lib"]
#![staged_api]
#![feature(staged_api)]
#![stable(feature = "x", since = "1.0.0")]

#[stable(feature="x", since="1.0.0")]
pub mod foo {
    #[stable(feature="x", since="1.0.0")]
    pub fn f() {}
}

#[deprecated(since="1.1.0")]
#[stable(feature="x", since="1.0.0")]
pub use foo as bar;

extern crate deprecated;

fn main() {
    deprecated::foo::f();
    deprecated::bar::f();
}

All compiles without a peep.

[Gankro]

Hmm... maybe I hallucinated that...

[Gankro]

This proposal has enough issues that I'm just going to close it. I think the right idea is in here somewhere, but I'm not the one to find it.