This change introduces a potential for UI starvation. By resetting wake_up_posted_ to false before calling ProcessTasks(), you allow a task running inside ProcessTasks() to trigger a new WakeUp() call that will immediately post another WM_NULL message. If tasks repeatedly schedule immediate follow-up work (a common pattern with Future.delayed(Duration.zero)), this can create a tight loop of processing WM_NULL messages, preventing the message loop from handling other messages like user input and making the application unresponsive. This appears to re-introduce the problem that the original timer-based workaround was trying to solve.

To fix this, wake_up_posted_ should be reset after ProcessTasks(). However, this would introduce a "lost wakeup" problem if another thread calls WakeUp() while ProcessTasks() is executing. This race condition would also need to be addressed for the solution to be correct.

That was the assumption originally. However if we make sure not to flood the message queue with unnecessary messages while ProcessTasks() is already pending the event messages are pumped correctly.

Should we update wake up posted after processing tasks though?

In other words if ProcessTasks asks for another wakeup, should we ignore that?

It might be worth a quick comment to explain this.

No. When wake up is requested during ProcessTasks() we need to schedule another wake-up message. Otherwise we would lose the wake-up request.

Let's say that dart code executed in ProcessTasks() calls Future.delayed(Duration.zero, () { ... }). If we update wake_up_posted_ after processing tasks, this will add the task to queue, but will not wake up the message loop. So the tasks will be only processed after something else wakes up the message loop (i.e. vsync waiter).

The original assumption that PostMessage always takes precedence before event messages does not seem to be correct. What was actually happening is that for every wake up request we have been posting wake up message, but then processed all scheduled tasks when handling just the first message alone. This might in turns schedules another task, but the task is performed too early (because there are other wake-up messages still in queue).

Thanks for the excellent explanations!