Rfcs: Inherent trait implementations

[frewsxcv]

I've been working a lot with traits lately with this rust-geo pull request.

In that pull request, it splits up concrete geospatial types and geospatial operations. A simplified example:

struct PointType {
    x: f32,
    y: f32,
}

trait PointTrait {
    fn x(&self) -> f32;
    fn y(&self) -> f32;
    fn distance<P: PointTrait>(&self, other: P) -> f32 {
        // distance algorithm
    }
}

impl PointTrait for PointType {
    fn x(&self) -> f32 { self.x }
    fn y(&self) -> f32 { self.y }
}

Now, if someone ever wanted to use this, they'd have to import both the type and the trait:

use geo::{PointType, PointTrait};
let p1 = PointType { x: 1, y: 1 };
let p2 = PointType { x: 2, y: 2 };
println!("distance: {}", p1.distance(p2));

Most of the time, the user is always going to want the associated methods associated with PointType, so it's a little unfortunate that the user has to add another import whenever they want this functionality.

It'd be cool if I could make a trait implementation as exposed so that the triat methods are always available on that specific type. Something like:

exposed impl PointTrait for PointType {
    fn x(&self) -> f32 { self.x }
    fn y(&self) -> f32 { self.y }
}

(There's probably a better keyword to use here than exposed, or maybe some other way to indicate this.)

Then the user can just do:

use geo::PointType;
let p1 = PointType { x: 1, y: 1 };
let p2 = PointType { x: 2, y: 2 };
println!("distance: {}", p1.distance(p2));

I haven't thought too long about this, so I might be overlooking something here. I don't feel strongly at all about this, just expressing a thought.

[Ms2ger]

Please explain what happens when calling point.a() in the following scenario:

crate A {
    struct PointType {
        x: f32,
        y: f32,
    }
}

crate B {
    trait OnePointTrait {
        fn a(&self) -> f32;
        fn b(&self) -> f32;
    }

    exposed impl OnePointTrait for PointType {
        fn a(&self) -> f32 { self.x }
        fn b(&self) -> f32 { self.y }
    }
}

crate C {
    trait OtherPointTrait {
        fn a(&self) -> f32;
        fn b(&self) -> f32;
    }

    exposed impl OtherPointTrait for PointType {
        fn a(&self) -> f32 { self.y }
        fn b(&self) -> f32 { self.x }
    }
}

[frewsxcv]

Could add a restriction such that exposed only works on concrete types defined in the same crate.

On a side note, this could also help out the 'num-traits' crate which IIRC reexports all the trait impls in the 'num' crate.

[oli-obk]

You can always put all the traits into one module and do use module::* if all you care about is less typing.

Also, advanced enough IDEs (or rustfix) will be able to automatically insert the necessary imports (I'm working on this at this very moment)

[eddyb]

The correct term here, IMO, is "inherent", and this would have to obey the rules of inherent impls.
That is, unless we get some privacy-based flexibility, the type must be defined in the crate of the impl.

[SimonSapin]

use module::* still need to be repeated in every module that uses the type in question.

Agree that "inherent" is the correct term. One way to think of this feature is syntactic sugar to generate an impl with #[inline] inherent methods that call the trait methods of the same name (and takes the same arguments, have the same return type, …). Then naturally this can only be used where new inherent methods can be defined.

[Wopple]

IMO, the extra effort of having to write an additional import is worth the flexibility of being able to define your own implementations of functions. I can imagine a case where you want to use a struct but you don't like the implementation of a trait provided by the crate. If that implementation was inherent, how would you opt out of using it in favor of an implementation with the same signature of your choosing?

Edit: never mind (see below)

[frewsxcv]

I can imagine a case where you want to use a struct but you don't like the implementation of a trait provided by the crate. If that implementation was inherent, how would you opt out of using it in favor of an implementation with the same signature of your choosing?

I'm not sure how this idea makes your concern any worse.

e.g. It's already possible for PointType to implement Display, and if you don't like the implementation of Display provided by the crate, you can't do much about it.

[Stebalien]

Inherit impls actually came up when discussing "thin" pointers to traits. One idea (that I liked) was to allow structs to declare which traits they must implement:

struct MyStruct: Trait1 + Trait2 + Trait3; // or struct MyStruct impl Trait1 ...

This would have two side effects (not including thin pointer stuff).

1. The struct must implement these traits (checked by the compiler).
2. The user would be able to call MyStruct.trait1_method() without importing Trait1.

The nice thing about this alternative is that it keeps everything in one place.

[SimonSapin]

@Stebalien This is an OK alternative, but it seems a bit redundant with the impl blocks which already name traits and types together. And don’t see how "keeping everything in one place" is valuable when impl MyStruct { fn inherent() } can already be "far" from struct MyStruct

[burdges]

I'd think the syntax for your exposed impl .. should be adding a pub reference to the trait to the inherent impl for the type, like

trait PointTrait { .. }

impl PointTrait for PointType { .. }

impl PointType {
    pub use PointTrait;
}

or maybe pub trait PointTrait.

Now use module::PointType brings PointTrait into scope as PointType::PointTrait, so it does not conflict with another item named PointTrait, but method notation still works on PointType.

[Wopple]

@frewsxcv In the following example, it prints 0 or 1 depending on which trait you import:

struct MyStruct {
    val: u32
}

trait Trait1 {
    fn get_val(&self) -> u32;
}

trait Trait2 {
    fn get_val(&self) -> u32;
}

impl Trait1 for MyStruct {
    fn get_val(&self) -> u32 { self.val }
}

impl Trait2 for MyStruct {
    fn get_val(&self) -> u32 { self.val + 1 }
}

mod test {
    use ::MyStruct;
//    use ::Trait1;
    use ::Trait2;

    fn run() {
        let s = MyStruct { val: 0 };
        println!("val: {:?}", s.get_val());
    }
}

it can't compile if you import both because the get_val function becomes ambiguous.

[frewsxcv]

@Wopple You can differentiate via:

Trait1::get_val(&s)

EDIT: I originally wrote this, but it's wrong:

(s as Trait1).get_val()

[Wopple]

@frewsxcv I see, thanks!

[mark-i-m]

Perhaps something like pub impl would be useful? A nice side effect of this is that you can have some methods be inherent and not others, even within the Trait...

[mark-i-m]

Also, pub is already a reserved word, whereas inherent and exposed are not, TMK...

[ideasman42]

One of the down-sides to the current situation (needing to explicitly use the traits), that it makes using traits less attractive - even when they make sense logically. In some code I just don't use traits because of the extra hassles of ensuring they're in the name-space and just impl SomeStruct {...}, the down-side to this is the functions on multiple structs can get out of sync without any warning, see stackexchange question on this topic.

[mark-i-m]

In some code I just don't use traits because of the extra hassles of ensuring they're in the name-space

Hmm... I can see why it would be annoying having to find the correct trait, but I have never found this to be an excessive "hassle" before... at least, not to the degree that I would rather not use traits. Could you give an example of where this was a huge burden?

[newpavlov]

I would love to see this feature implemented, as for example in the RustCrypto I have to re-export traits in every crate to make it more convenient for users, or otherwise they'll have to list additional crate (containing traits) as explicit dependency. Also it makes harder to write documentation, as I have to explain how traits are organized to users who do not want to write generic code and just want to call several methods and have stuff done.

Overall it's a very annoying papercut for crate authors and users.

[burdges]

It's never been clear to me why RustCrypto uses so many crates instead of features. Yes, crates provide an easier to understand separation, but if the crates all reside in the same repository anyways then most advantages seem moot.

There is no shortage of applications for this language feature though of course, so I'm not objecting to anything. :)

We rand into issues when discussing use in items like impls a few months ago, so the impl MyStruct { pub use Trait; .. } syntax might run into similar difficulties, or maybe it would fix them. I donno if pub impl Trait for MyStruct { .. makes sense, but maybe. And the struct declaring traits it must implement sounds workable.

[Diggsey]

I would like to see this feature too: I'm currently looking at generating traits and types from javascript IDL files, and I have to choose between generating inherent methods vs trait methods for interfaces, and neither corresponds cleanly to what the javascript API actually looks like.

[stepancheg]

I want to point to (now closed) "custom prelude" RFC https://github.com/rust-lang/rfcs/pull/890 Should that RFC implemented, the described problem could be solved with a custom prelude: the client could import PointTrait into their crate prelude.

[idanarye]

If I could add my suggestion, I would like to see a syntax like this:

impl PointTrait for PointType {
    fn x(&self) -> f32 { self.x }
    pub fn y(&self) -> f32 { self.y }
}

You'd have to use PointTrait in order to access point.x, but point.y will be available in other modules even if they don't use PointTrait. Of course, you can only use pub if the impl block is defined in the same crate as PointType.

The advantages of marking the specific function rather then the impl statement are:

• Just like with non-trait impl blocks, pub fn will be visible outside the module and just fn without pub will only be visible inside the module. (Though in this case you can use the trait to get access to these methods)
• Finer picking of which functions are inherent and which are trait-only.
• Conflicts are easier to resolve - just pick which of the methods are the inherent one.
• Speaking of conflicts - the error messages will be nicer. Instead of something like "_conflicting inherent implementation of trait A and trait B on type T: method foo is defined in both traits_" the error message will be more like "_inherent methodfoo` is defined multiple times_".

[idanarye]

I've implemented my syntax with a proc_macro: https://github.com/idanarye/rust-inherent-pub