Remove unsound SystemParam impl for DeferredWorld

[ItsDoot]

Objective

The following test currently passes (i.e. no panic) on main:

#[test]
fn deferred_world_last() {
    #[derive(Resource)]
    struct FooRes;

    fn bad_system(foo: ResMut<FooRes>, mut world: DeferredWorld) {
        let foo = foo.into_inner();
        let foo2 = world.get_resource_mut::<FooRes>().unwrap().into_inner();
    }

    let mut world = World::new();
    world.insert_resource(FooRes);
    let mut bad_system = IntoSystem::into_system(bad_system);
    bad_system.initialize(&mut world);
    bad_system.run((), &mut world);
}

Solution

Remove the unsound SystemParam impl for DeferredWorld.

---

Migration Guide

The SystemParam implementation for DeferredWorld was removed due to unsoundness. Consider using Commands instead.

[JeanMertz]

Does this mean DeferredWorld can no longer be used in regular systems?

[chescock]

I think the only thing we need to do to make it sound is to check for existing access like

if system_meta.component_access_set.has_any_component_read()
    || system_meta.component_access_set.has_any_resource_read()
{
    panic!("Useful message")
}

Oh, and

system_meta.archetype_component_access.write_all();

Could we do that instead?

[hymm]

I think the only thing we need to do to make it sound is to check for existing access like

DeferredWorld pushes into the World's command queue, not one specific to the system and that requires exclusive access to the world.

[chescock]

Well, it was a little more complicated than I expected, but I created a competing PR that tries to leave the impl in place but make it sound: #17616

[chescock]

DeferredWorld pushes into the World's command queue, not one specific to the system and that requires exclusive access to the world.

Oh! I missed that. It might still be sound, though. If we declare that accessing the World's command queue requires write-all-components access, then no more than one system can claim it.

[alice-i-cecile]

On second thought, I'd prefer fixing this in 0.16 cycle.

[hymm]

Oh! I missed that. It might still be sound, though. If we declare that accessing the World's command queue requires write-all-components access, then no more than one system can claim it.

It might be sound, but when do the command queue get applied? I don't think there's a determined place when they ever will get applied. Just whenever the world's queue happen's to get applied. Fixing this properly would be to somehow pass a system stored command queue to the constructor of DeferredWorld.

[hymm]

My pref here is to remove this for now and it can just be readded once an appropriate pr shows up.

[alice-i-cecile]

Closing in favor of #17616.