moonbitlang / async / 378

78%

push

github

perf(event_loop): short-circuit timer scan when no timers are pending

EventLoop::wait_for_event runs twice per scheduler tick:

  1. self.timers.iter().head() to compute the poll.wait timeout
  2. for timer in self.timers to fire any expired timer callbacks

When the program does no timer work (IO-only loops, pure compute via
pause, queue producer/consumer loops, etc.) the timers SortedSet is
empty but the two iterator-creation and ms_since_epoch() calls still
fire on every tick. A callgrind profile of a pure-pause workload
(single coroutine, pause() x 2.5M) attributes ~7% of all instructions
executed to the timer scan.

Gate both spots on self.timers.is_empty(). When timers are pending the
extra branch is cheap; when they are not, the entire iterator
allocation + ms_since_epoch + loop setup drops out.

pause_loop bench (native, 3-run median):
  baseline: 637 ms
  patched : 469 ms  (-26.4%)

6 of 6 new or added lines in 1 file covered. (100.0%)

3326 of 4283 relevant lines covered (77.66%)

78646.21 hits per line