Add support for use Trait::func

[obsgolem]

Rendered

This feature fully supplants rust-lang/rust#73001, allowing you to do things like:

use Default::default;

struct S {
    a: HashMap<i32, i32>,
}

impl S {
    fn new() -> S {
        S {
            a: default()
        }
    }
}

and more.

This is my first RFC, please forgive any missteps I make in the process.

Partially completes #1995.

[Evian-Zhang]

From The Rust Programming Language section 7.4:

Creating Idiomatic use Paths

Although both Listing 7-11 and 7-13 accomplish the same task, Listing 7-11 is the idiomatic way to bring a function into scope with use. Bringing the function’s parent module into scope with use means we have to specify the parent module when calling the function. Specifying the parent module when calling the function makes it clear that the function isn’t locally defined while still minimizing repetition of the full path.

We have been encouraged for a long time to always import the parent module when calling the function to distinguish locally defined functions and imported functions. However, the syntax sugar suggested in this RFC contradicts to this convention.

I think it is more appropriate to discuss this convention in the RFC to make it more clear how we should use this feature. :)

[crumblingstatue]

I'd like to reiterate here what I said on #1995

Many libraries define free functions for constructors to mathy types, like vectors.

Some examples:
https://docs.rs/glam/0.24.2/src/glam/f32/vec2.rs.html#12
https://docs.rs/egui/latest/egui/fn.pos2.html

It would be nice if instead of having to create wrapper functions, the user (or the lib author with pub use) could just do use Vec2::new as vec2;

This is some additional motivation that could go into the motivation section. Default::default is hardly the only use case for this.

[crumblingstatue]

That being said, I noticed that this RFC doesn't talk about importing inherent methods, which I think should be addressed.
If this RFC only proposes importing trait methods, but not inherent methods, then the use cases I mentioned above are nullified.
Why should we only support importing trait methods, but not inherent methods?

[obsgolem]

I discussed that in the future work section. use Type::method is out of scope for this RFC. The difficulty is impl blocks with arbitrary where clauses.

[kpreid]

The RFC text uses the term “method” where it should be using the term “associated function”.

• An associated function is a function defined in an impl or declared in a trait. (reference)
• A method is an associated function that has a self parameter (and therefore may be used in .func_name() method-call syntax). (reference)

I recommend that this be corrected to avoid creating confusion. In particular, the central use case of this feature is importing associated functions that are less concise to call because they are not methods, such as Default::default.

[fmease]

I'm surprised no one has mentioned generic traits yet. We would probably want to reject such kinds of imports but that should be stated explicitly in this RFC.

Consider:

use Trait::function;

trait Trait<'a, T, const N: usize> {
    fn function<'b, U, const M: usize>();
}

Unless we somehow concatenated the two list of generic parameters and where clauses which probably would have questionable semantics and other unforeseen consequences, there would be no way to explicitly specify the trait generics given only the associated item.

[fmease]

Well, maybe there's no need to reject those kinds of imports but I feel like they're a bit footgun-y but maybe that's just me. Since function would basically be <_ as Trait<'_, _, _>>::function, I'm just wondering in how many use cases / scenarios inference would actually succeed in practice and in how many there would be one param that can't inferred requiring users to switch from an imported associated function back to a more qualified path 🤔 Could be a non-issue though, I dunno.

[obsgolem]

So, with the desugaring I specify, it is unambiguous what should happen in this situation, which is exactly as you state. The compiler will try to infer all trait parameters, and if it can't it will throw an "unable to infer error". Then you just switch to the fully qualified syntax. This is exactly the recommendation we make for ambiguous inference right now.

[fmease]

If this PR only plans on adding support for importing associated functions, it should imo explicitly state explicitly that it excludes associated constants and types and add a short explainer for why that's the case. Note that implementing this feature for associated constants would probably be easier to do than for associated functions, so there's not much reason to exclude them, imho.

(Importing associated types on the other hand should probably be forbidden. Unless we reify the implicit Self type parameter (i.e., turn it into an explicit parameter), such an associated type Ty would be equivalent to <_ as Trait>::Ty which is impossible to solve since associated types are not injective in general.)

[obsgolem]

I've only ever seen this feature mentioned in with associated funcs. Writing in associated consts shouldn't be too hard, but I only want to do it if it is wanted.

[joshtriplett]

@obsgolem If you think associated constants would be straightforward to add to the RFC, please do go ahead and add it. If you think it'd be a challenge to add to the RFC, please mention it in future work instead.

[joshtriplett]

This seems like a notable limitation. We've talked about adding a compatibility mechanism that allows implementing supertrait methods via the subtrait, such that it would then be possible to extract a method from a trait into a supertrait without breaking compatibility with any user of the trait. This limitation would prevent that.

So, at the very least, if we add such a mechanism we'll have to fix this limitation at that time as well. Could you add a note to future work to that effect? "If we add a compatibility mechanism to implement a supertrait method when implementing its subtrait, without having to separately implement the supertrait (such that a new supertrait can be extracted from a trait without breaking compatibility), we would also need to lift the limitation on using a supertrait method via a subtrait."

[joshtriplett]

Suggested change

	occurs, a new item `m` is made available in the value namespace of the current module. Any attempts to call this item are treated calling the associated function explicitly qualified. As always, the `as` qualifier is optional, in which case the name of the new item is identical with the name of the associated function in the trait. In other words, the example:
	occurs, a new item `m` is made available in the value namespace of the current module. Any attempts to call this item are treated as calling the associated function explicitly qualified. As always, the `as` qualifier is optional, in which case the name of the new item is identical with the name of the associated function in the trait. In other words, the example:

[joshtriplett]

I'm going to go ahead and start the process of seeing if we have consensus on this:

@rfcbot merge

Please do either add support for associated constants inline in the RFC (if it's straightforward to do so) or mention them in future work (if not):

@rfcbot concern associated-constants

[rfcbot]

Team member @joshtriplett has proposed to merge this. The next step is review by the rest of the tagged team members:

• @joshtriplett
• @nikomatsakis
• @pnkfelix
• @scottmcm
• @tmandry

Concerns:

• associated-constants (Add support for use Trait::func #3591 (comment))

Once a majority of reviewers approve (and at most 2 approvals are outstanding), this will enter its final comment period. If you spot a major issue that hasn't been raised at any point in this process, please speak up!

cc @rust-lang/lang-advisors: FCP proposed for lang, please feel free to register concerns.
See this document for info about what commands tagged team members can give me.

[Veykril]

This seems badly phrased to me, use foo::self is generally disallowed.

error[E0429]: self imports are only allowed within a { } list

(its nitpicky I know, but it should still state use some_module::Trait::{self}; instead and rephrase it in terms of whats already allowed with trailing self imports)

[joshtriplett]

Perhaps jump straight to a use some_module::Trait::{self, method}; example, since that's the goal of this syntax?

[Veykril]

Actually, on second though, that is already allowed

trait T {}

mod module {
    use crate::T::{self};
}

compiles today as expected. So yes, it would make more sense to just jump straight to the use some_module::Trait::{self, method}; example and drop the rest.

[Veykril]

On an implementation side, I am assured that this feature is straightforward to implement.

Note that this may be the case for rustc, but not necessarily for alternative implementations. Wearing my rust-analyzer hat I'd expect this to be doable as well, but a bit annoying. Not saying this is a blocker (it's not impossible to implement) but mainly to keep that in mind until we attempt to implement it in r-a.