Introduce LocalTaskObj

[MajorBreakfast]

• Introduce LocalTaskObj and integrate it into LocalPool
• Export all the stuff form core::task under futures::task (some was previously exported under futures::executor) This is probably just a start. There're probably still some inconsistencies in the exports.
• This PR makes the tests in futures-executor/tests/local_pool.rs work

[cramertj]

@MajorBreakfast Pinged you on discord about this. I think, rather than a separate trait, I'd prefer to remove the Send bound on the UnsafeTask trait, and only introduce it in TaskObj::new. I also think I'd prefer for LocalTaskObj to live in core so that the conversion from TaskObj to LocalTaskObj doesn't depend on unstable representation details of std.

[MajorBreakfast]

I agree. It would then be very similar to how we do it with the UnsafeWake trait. The only thing different is that UnsafeWake needs bounds for Sync and Send because of the cloning as Waker. For this here however we, as you say, shouldn't add bounds to the trait, so that PinBox with a Future<Output = ()> + 'static can implement it

[MajorBreakfast]

Do we want to add SpawnLocalObjError to libcore as well? It's not absolutely necessary because the fields are public. (But we need it somewhere because SpawnObjError can't store a LocalTaskObj)

[cramertj]

@MajorBreakfast Hm, good question! I think either way is fine-- these are all going to be deprecated once unsized-rvalues + arbitrary-self-type-object-safety are fixed, so I don't think it matters particularly much either way.

[MajorBreakfast]

We will still need need the distinction for between LocalSpawnError and SpawnError (just without Obj in the name) because of the Send difference.

If you let me pick, then I'd put them into core because this way we can use transmute for conversion between them. LocalPools's spawn_obj calls spawn_local_obj internally, so this conversion is on a hot code path.

[cramertj]

@MajorBreakfast In the long term, I figured we wouldn't have LocalSpawnObjError or LocalSpawnError, only the current SpawnErrorKind. But either way is fine w/ me.

[MajorBreakfast]

Here is the PR that puts LocalTaskObj into libcore rust-lang/rust#51814

SpawnError gives us the task back when spawning isn't possible. This makes it possible to spawn it somewhere else. I don't think that it will go away. (And LocalSpawnError does the same for non-Send tasks)

[cramertj]

@MajorBreakfast It will have to go away once spawning goes through unsized rvalues, since there's no way to hand them back.

[MajorBreakfast]

That's a bummer.

So, this won't work? I thought the last element of a struct can be unsized?

pub struct SpawnError {
    pub kind: SpawnErrorKind,
    pub future: dyn Future<Output = ()> + Send,
}

Edit: I just looked at the RFC again. Functions will have to return a sized type. Then it's clear that it can't work

[MajorBreakfast]

The new nightly which now includes LocalTaskObj in libcore requires some fixes to the futures crate. Therefore this is now blocked on #1049

[MajorBreakfast]

I've updated the PR to use core::task::LocalTaskObj

[cramertj]

Nice! This is great!