Don't rely on Box being in the prelude

[Kestrer]

This code:

use async_trait::async_trait;

struct Box;

#[async_trait]
trait Foo {
    async fn foo(&self);
}

#[async_trait]
impl Foo for () {
    async fn foo(&self) {}
}

Fails because async-trait uses Box instead of ::std::boxed::Box ::alloc::boxed::Box in its macro expansion.

[taiki-e]

If use ::std::boxed::Box path, cannot use async-trait on no-std environment.

[Kestrer]

Oh, then it should use ::alloc::boxed::Box.

[taiki-e]

Yeah, but note that alloc crate is not available without an explicit extern crate alloc.

[taiki-e]

note that alloc crate is not available without an explicit extern crate alloc.

The way to make this works correctly is a bit tricky. It's available in function body by importing like extern crate alloc as _alloc, but doing the same in the position of the return value can cause _alloc to conflict with _alloc generated by other #[async_trait]. Probably it needs a trick, such as including the hash value of the input AST in the imported crate name.

[dtolnay]

I would like to leave this as currently implemented. Nobody should redefine Box and then expect to have boxed trait objects in the same module using a different Box.