Prevent double boxing in boxed

[stepancheg]

future.boxed() creates a new box even if future itself is already a Box<Future>.

This can be inconvenient in this use case:

type SomeFuture = ...;
fn make_future() -> SomeFuture { ... }

fn accept_future<F : Future>(f: F) {
    f.boxed() ...
}

accept_future(make_future());

If SomeFuture is an alias for Box<Future> then accept_future makes a Box<Box<Future>>. Which could be worked around (e. g. by creating another function like

fn accept_future_boxed(future: Box<Future>) { ... }

but that would be inconvenient, and still won't prevent accidental wrapping futures in double boxes.

boxed could implemented like this:

fn boxed<F : Future + Sized + 'static>(f: F) -> Box<Future> {
    if TypeId::of::<F>() == TypeId::of::<Box<Future>>() {
        assert!(mem::size_of::<F>() == mem::size_of::<Box<Future>>());
        unsafe {
            let mut r: Box<Future> = mem::uninitialized();
            ptr::copy_nonoverlapping(&f as *const F, &mut r as *mut Box<Future> as *mut u8 as *mut F, 1);
            mem::forget(f);
            r
        }
    } else {
        Box::new(f)
    }
}

PR #512 implements boxed in Future and Stream this way.

[3Hren]

This possibly can be implemented without RTTI using something like Boxed trait using specializations.

[stepancheg]

This possibly can be implemented without RTTI using something like Boxed trait using specializations.

First, TypeId::of is not RTTI, because function parameter is a type, not a value. It is compile time type indentification, and condition is optimized away by compiler.

Second, if you know how to implement it with Boxed trait, please show me how. I tried, and failed.

[3Hren]

First, TypeId::of is not RTTI, because function parameter is a type, not a value. It is compile time type indentification, and condition is optimized away by compiler.

I'm not sure (but likely so). Docs/ASM proof?

Second, if you know how to implement it with Boxed trait, please show me how. I tried, and failed.

If I understood the issue correctly, here is the proof of concept: #513.

[3Hren]

Your tests from #512 passes. Decided not to copy/paste them nevertheless.

[stepancheg]

I'm not sure (but likely so). Docs/ASM proof?

https://godbolt.org/g/JdxAes

If I understood the issue correctly, here is the proof of concept: #513.

Thanks. It doesn't work in stable though.

[Kixunil]

Specialisation would be much nicer.

[stepancheg]

@Kixunil it doesn't work in stable.

Anyway, it doesn't matter how it is implemented internally, from user point view nothing will change when in couple of months implementation will be changed to specialization.

[Kixunil]

@stjepang I know. I feel like everything is blocked on specialisation these days... I agree with implementing it with unsafe now and changing it later.

[alexcrichton]

Closing due to #512 (comment)