Subj : Re: Optimization to Jeffrey Richter's COptex (Windows)?
To   : comp.programming.threads
From : Alexander Terekhov
Date : Thu Jan 27 2005 02:35 pm


SenderX wrote: ...

You've omitted timedlock(). Why so? ;-)

Well, timedlock() aside for a moment, with your waiters counting you 
increase the overhead of contention. More waiters in a batch results 
in more counting/read-modify-write. Get rid of waiters counting. With 
waiters bit maintained by you instead of counting, you'll have 
needless signaling from time to time (at the end of contention 
interval) and it may cause "spurious" wake at the beginning of next 
contention interval but that overhead is "constant" (it doesn't 
increase with increased contention).

Sorta optimizations meant to "compensate" some silly cas impls aside 
for a moment,

  atomic<int>      m_lock_status; // 0: free, 1/-1: locked/contention
  auto_reset_event m_retry_event; // prohibitively slow bin.sema/gate

  // ctor/dtor [w/o lazy event init] 

  bool trylock() throw() {
    return !m_lock_status.cas(0, 1, msync::acq); 
  }

  void lock() throw() {
    if (int lock_status = m_lock_status.cas(0, 1, msync::acq)) {
      do {
        while (lock_status < 0 || 
               lock_status = m_lock_status.cas(1, -1, msync::acq)) { 
          m_retry_event.wait(); // acquire msync
          lock_status = m_lock_status.load(msync::none);
          if (!lock_status) break;
        }
      } while (lock_status = m_lock_status.cas(0, -1, msync::acq));
    }
  }

  void unlock() throw() {
    if (m_lock_status.cas(1, 0, msync::rel) < 0) { 
      m_lock_status.store(0, msync::rel);
      m_retry_event.set(); // release msync
    }
  }

regards,
alexander.

.