On 12/11/13 20:13, Samuel Thibault wrote: > Hello, > > OK, I believe that'll work indeed. What really makes it work is this: > > Justus Winter, le Mon 11 Nov 2013 21:12:34 +0100, a écrit : >> + /* Decrement nreqthreads. */ >> + unsigned int tc = __atomic_sub_fetch (&thread_counts, 1, >> + __ATOMIC_RELAXED); >> + if (NREQTHREADS (tc) == 0) > > You have atomicity between the decrementation and testing whether the > resulting value is 0. That's why we were keeping the spinlock along the > test. The same happens with the test on totalthreads below. > > Now, there is no need for the two counters to be updated coherently, > they are used for different reasons, and don't actually interfere. Also, > mixing them into one int reduces the maximum number of threads to 65535, > which is not so big.
Couldn't 8 byte integers be used here? That would give us the same number of maximum threads. Cheers, Emilio > So I'd say please move to using two unsigned ints. Yes, it means > two atomic operations on thread creation/destruction, but that's an > expensive operation already anyway (and we're not supposed to do that so > often). Readability of the code is more important than that :) > > Samuel > >
