Clean closure and task-local-state in jl_finish_task

[vchuravy]

Implements the second part of @vtjnash idea in #47405 (comment)

Particularly motivated by @quinnj observation that it is easy to add a return nothing to a task, but the arguments
to the task need to be handled differently. (x-ref #40626 (comment))

[gbaraldi]

Can someone get a reference to the fields from julia? Or just to the content?

[vchuravy]

Can someone get a reference to the fields from julia? Or just to the content?

If you mean &ct.tls. No they are just normal fields in a struct.

[NHDaly]

👍 LGTM

Can you also add a test that confirms that objects captured by a Task's closure are GC'd when the Task is finished?
Something like

function foo(x::Vector)
    @async length(x)
end
a = []
r = WeakRef(a)
t = foo(a)
a = nothing
wait(t)
GC.gc()
@test r.value === nothing

or somethign like that?

[vchuravy]

@nanosoldier runtests()

[nanosoldier]

Your package evaluation job has completed - possible new issues were detected. A full report can be found here.

[vchuravy]

Two packages observe the nothing

AdvancedPS v0.4.3: fail vs. ok
ChannelBuffers v0.2.0: fail vs. ok

[vtjnash]

AdvancedPS uses LibTask, which violates deep assumptions necessary for fast code, and therefore is undefined behavior and unsupported entirely.

ChannelBuffers is accessing the fields of a closure via reflection, which is an unstable API–though could be easily re-written to use a stable ABI version by making it a struct instead. I suppose though that this is probably valid code however.

[JeffBezanson]

I think the only non-breaking way to do this is to make tls->current_task like a weak reference, such that when that task is finished and current_task is the only reference, some fields inside the task can be cleared.