24 comments

[ 4.4 ms ] story [ 71.3 ms ] thread
(comment deleted)
This post reminds me of a problem I've hit with my Objective-Rust project[1] (Rust code that's interoperable with Objective-C), to which I haven't found a good solution.

The closest analog to Objective-C's @protocols in Rust are traits. I would love to be able to transliterate the following Objective-C:

  @protocol SomeProtocol
  // methods
  @end

  void demo() {
    // Just a demo of how to use SomeProtocol with a var:
    id<SomeProtocol> variable = nil;
  }
To the following Rust:

  #[objrs(protocol)]
  trait SomeProtocol {
    // methods
  }

  fn demo() {
    // Just a demo of how to use SomeProtocol with a var:
    let variable: *mut Id<SomeProtocol> = 0 as *mut _;
  }

  // The following is provided by objrs:
  struct Id<T: ?Sized>(core::marker::PhantomData<T>, Opaque);
  extern "C" {
    type Opaque;
  }
Unfortunately, RFC 255[2] breaks this. This fails because SomeProtocol is not object safe. Trait objects in Rust can be pretty frustrating in Rust, as this post illustrates.

I've had to workaround this in objrs by doing something like the following:

  #[objrs(protocol)]
  trait SomeProtocol {
    // methods
  }
  // The macro above generates stuff kinda like this:
  struct SomeProtocolId;
  impl SomeProtocol for SomeProtocolId {}
  impl<T: ?Sized + SomeProtocol> SomeProtocol for Id<T> {}

  fn demo() {
    // Now you have to know when to use SomeProtocol vs SomeProtocolId:
    let variable: *mut Id<SomeProtocolId> = 0 as *mut _;
  }
I'm going to see if there are some places where this blog post might simplify some things in my code... it's got some interesting ideas.

[1]: https://gitlab.com/objrs/objrs

[2]: https://github.com/rust-lang/rfcs/blob/master/text/0255-obje...

I don't know enough Rust to completely understand what you are talking about with object safety, but this is a very interesting project. How do you plan to handle Objective-C's runtime features? Even if you don't plan to implement advanced features like objc_allocateClassPair, one thing you will have to deal with rather quickly is informal protocols and optional methods, which are resolved at runtime.
Runtime modifications to the Objective-C runtime (e.g., objc_allocateClassPair) will have to go through libobjc, just like they do in Objective-C. I could provide friendlier interfaces to libobjc, but I don't plan to.

objrs already supports optional methods (for externally-defined protocols), though using them is the same as in Objective-C (you still have to call respondsToSelector: first). I've thought about adding an additional API that simplifies this a bit (i.e., a method that returns an Option, where it's None if the class doesn't respond to the method, or Some(fn) if it does (where fn is a closure you can use to invoke the function)).

Ideally you want `Id` to actually be parameterized on traits, not trait object types that happen to have the same syntax (now deprecated) as their corresponding traits. Of course, that's not currently possible. But you may be interested in this repo for planning an RFC to eventually add such functionality:

https://github.com/Centril/rfc-trait-parametric-polymorphism...

After being created 8 months ago, it's sort of inactive, but hopefully that'll change once the compiler's feature-implementation backlog starts catching up – in particular, once "chalk in rustc" becomes a reality.

Oh neat! I hadn't seen that one, but it's literally item #20 on my Rust wishlist. I'll have to file an issue to give some motivating examples. Thank you for pointing it out to me!
IMHO, Rust just should take size of the largest implementation of Animal for `impl Animal`, and viola — we can make them static. No heap is good for embedded.
Would this not require all impls to be in the same translation unit or the linker has to be explicitly aware of it?
In case of embedded firmware, all impls will be in same unit. In case of a program with linked libraries, only main program will know sizes of all possible impls.
Note that it's quite possible for a trait to have infinitely many implementations in any given program. For example, Debug is implemented by u64, Box<u64>, Box<Box<u64>>, ..., and so on. While these happen to have the same size, it's not hard to imagine a situation where this is not the case.
Your situation is already sufficiently impossible because it shows that for a given trait, the set of implementors is often (countably) infinite.
However, the number of implementors actually instantiated in any given program is necessarily finite. You could write a program that would make it infinite, e.g.

    trait Foo { fn foo(); }

    impl<T> Foo for T {
        fn foo() {
            <Box<T> as Foo>::foo();
        }
    }
...but the compiler can't compile it: since Rust implements generics solely through monomorphization, that would require generating an infinitely large binary :)
> Rust implements generics solely through monomorphization

That doesn't sound right. Trait objects use dynamic dispatch.

You can obviously use an `enum` for this, and it's not so clunky that I think Rust should add yet another feature for it.
> Rust, not being an object-oriented language, doesn't quite do inheritence like the others. In C++ or Java, subclasses extend superclasses. In Rust, structs implement traits.

I wouldn't say that Rust isn't object oriented. It just takes a different approach to some things. Inheritance isn't really a strictly required feature of the object oriented code design. Many for instance pointed out, that composition is preferable:

https://en.wikipedia.org/wiki/Composition_over_inheritance

Rust is clearly not object-oriented. It has very minimal support for late binding, and no support for reflection or dynamic libraries.

This isn't a critique, it's just a statement of fact. No Rust designer will say they set out to build a new object-oriented language.

Late binding is the only one of those features commonly associated with object-orientation, and Rust is just as good as C++ on that front.
You can do compile-time reflection, which is very popular.
Why boxed? just change to vec::<&dyn animal> and push reference of Dog to vec ,it should work. Isnot &dyn animal trait object?
It works, but the point of the article remains the same because &dyn Animal is also a fat pointer.
Very insightful I enjoyed this very much