[libc++] Implement `std::condition_variable_any::wait[_for/until]` overloads that take `stop_token`

[huixie90]

Summary

• This is section 32.6.4 of P0660R10
• https://eel.is/c++draft/thread.condvarany.intwait

[huixie90]

CI

[Mordante]

In general looks good. I have some minor comments. I think it would be good to keep track how often predicates are called and whether that matches the (minimal) expectation. For example when a stop_requested() == true the predicate is called exactly one time.

libcxx/include/condition_variable

106

Can you align the // since C++20 ?

319

328

libcxx/test/std/thread/thread.condition/thread.condition.condvarany/wait_for_token_pred.pass.cpp

96

This is our typical code style.

135

I wonder here and at other places whether 2 ms is long enough.
Are we sure this is changed after the wait_for is really in it's loop? Or can the flag be set before entering the loop.

Would it make sense to validate how often the predicate is executed? When it's less than 2 the loop was executed 0 times.

libcxx/test/std/thread/thread.condition/thread.condition.condvarany/wait_token_pred.pass.cpp

43

Please include <concepts>.

[ldionne]

Commandeering per discussion w/ Hui.

libcxx/include/condition_variable

95

Let's mark this as experimental.

[huixie90]

address feedback rebase

[huixie90]

CI

[huixie90]

libcxx/include/condition_variable

319

this is a member function call . not std::wait_until

328

this is a member function call, not std::wait_until

[huixie90]

rebase

[huixie90]

FTM

[huixie90]

CI

[huixie90]

CI

[huixie90]

CI

[huixie90]

CI

[huixie90]

guard stop_token headers with no thread macro

[huixie90]

ci

[ldionne]

libcxx/test/std/thread/thread.condition/thread.condition.condvarany/wait_for_token_pred.pass.cpp

135

If the thread creation is really slow, this test could be vacuous (it would test the same as the test that does request_stop() before we wait_for). Would it make sense to instead do something like:

auto thread = support::make_test_thread([&]() {
    // wait until we're ready
    some_atomic.wait(...);
    ss.request_stop();

    while (!done) {
        cv.notify_all();
        std::this_thread::sleep_for(2ms);
      }
});

std::same_as<bool> auto r = cv.wait_for(lock, ss.get_token(), 1h, [&]() {
  some_atomic.notify_all(); // here signal that we're ready for the other thread to request a stop
  return false;
});
assert(!r);
done = true;
thread.join();

Does this make sense at all? That way the pred will be the one to trigger the fact that we request_stop() on the other thread. Would that be more robust?

libcxx/test/std/thread/thread.condition/thread.condition.condvarany/wait_token_pred.pass.cpp

80

Whatever we do for the other test we should do here too.

libcxx/test/std/thread/thread.condition/thread.condition.condvarany/wait_until_token_pred.pass.cpp

130

And here.

[huixie90]

address feedback

[ldionne]

LGTM with minor nitpick. You can ignore the CI failures since they are unrelated and have been fixed since then. Thanks!

libcxx/docs/Status/Cxx20.rst

51