> I'm not convinced it's really useful.

It doesn't seem a lot useful until you *really* need it. I've got 2 highly parallel programs that took advantage of this pattern.


> the wait is O(C) (appending to C wait queues) and wakeup
> is O(CT) (C removal from a T-length deque).

That's another thing that I wanted to address, but I didn't want to change the data structure in this patch.

The complexity for `notify` and `notify_all` is the same as before.