fix deadlock when awaiting a sender completes inline with set_stopped

[ericniebler]

fixes #2036

[mika-fischer]

@ericniebler FYI, the current changes in this PR (i.e. b8cd4b6) do fix both #2036 and #2041 here!

[ericniebler]

Hooray! Thanks for letting me know.

[mika-fischer]

I want to add that the changes in this PR broke sender_to for stdexec::task when the environment has no start_scheduler, at least in MSVC. This works in main, but now gives a hard compile error:

struct void_rcvr {
    using receiver_concept = stdexec::receiver_tag;
    auto get_env() const noexcept { return stdexec::env<>{}; }
    void set_value() && noexcept {}
};
static_assert(!stdexec::sender_to<stdexec::task<void>, void_rcvr>);

Incidentally, sender_in was already broken in the main branch before these changes. This gives a hard compile error both in main and this PR:

EDIT: Ignore the last part, I misunderstood sender_in semantics.

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.

[ericniebler]

/ok to test f3ae543

[ericniebler]

/ok to test 23793ef

[ericniebler]

/ok to test b86e0a5

[ericniebler]

/ok to test 7dc615b

[ericniebler]

I want to add that the changes in this PR broke sender_to for stdexec::task when the environment has no start_scheduler

i think the fact that it compiled before this pr is a bug. you'll get an error if you try to actually connect that receiver with the task. just like the on algorithm, task needs a start scheduler.

UPDATE: see for example https://godbolt.org/z/E5j9beGq1

[mika-fischer]

static_assert(!stdexec::sender_to<stdexec::task<void>, void_rcvr>);

i think the fact that it compiled before this pr is a bug. you'll get an error if you try to actually connect that receiver with the task. just like the on algorithm, task needs a start scheduler.

I think you might have missed the ! in the static_assert above.

I don't say that sender_to<task<void>, void_rcvr> should be true, I'm saying it should be false, which it was in main, but now in this PR it just doesn't compile at all anymore, which seems wrong.

EDIT: See https://godbolt.org/z/M5M459jzo

[ericniebler]

Ah thanks for clarifying. I'll look into it.