Use impl Future instead of Box<dyn Future>

[sdroege]

@A6GibKm

Any opinions on this one? I don't know how useful it is in practice. It might cause longer compile times and code bloat, or maybe the opposite.

We can't use async fn in traits because that gives compiler warnings (it's discouraged in public traits).

[sdroege]

An alternative here would be to do impl Future but still Box::pin the futures (which would be an implementation detail then). That hides this implementation detail and doesn't change anything with regards to generated code.

[sdroege]

So let's list the 4 options:

1. Keep things Pin<Box<dyn Future>>:
   • (-) Heap allocation, dynamic dispatch
   • (+) Nameable type, less code bloat
2. impl Future backed by Pin<Box<dyn Future>>:
   • (-) Heap allocation, dynamic dispatch, unnameable type
   • (+) Less code bloat, doesn't leak minor implementation details (boxed pinned future can be replaced by anything else later)
3. impl Future backed by the plain GioFuture
   • (-) Unnameable type, code bloat
   • (+) Doesn't leak minor implementation details, can be inlined (but questionable how useful: GioFuture is basically a oneshot channel)
4. async fn backed by the plain GioFuture
   • (-) Unnameable type, code bloat, gives compiler warnings in public traits
   • (+) Doesn't leak minor implementation details, can be inlined (but questionable how useful: GioFuture is basically a oneshot channel), Looks more natural

I think out of these, only 1. and 3. are really useful. The main question there is the actual amount of code bloat and how often the return type is used by name (there are some uses inside gio for converting futures into streams which might also be solvable differently).

I'd say to defer this to a future breaking release.