> Not at first glance. The reason I ask is the SableVM Java VM. The only > requirement it makes of an architecture is that it can implement the > "compare_and_swap" routine. Unfortunately this is implemented as a single > function, so it would seem necessary to lock the whole VM while performing > the op, which is probably going to kill performance.
what do you mean "as a single function"? i don't think the performance impact will be as great as you seem to be suggesting. you may have noticed that the kernel uses a lock hash for this purpose. this may help lessen lock contention if you are doing a lot of compare-and-swap ops on different addresses. fwiw we've ported a number of applications to hppa where we've had to use this multi-op compare_and_swap. while i've not done any serious benchmarking, things seem to work ok in general.... randolph
