Yeah, but then there will be all that additional __sync_synchronize(), that are not needed if we use __atomic. And on the other hand, if we have that __sync_synchronize(), then we do not need __atomic...
On Wed, Dec 5, 2012 at 4:57 PM, Jakub Jelinek <ja...@redhat.com> wrote: > On Wed, Dec 05, 2012 at 04:53:44PM +0400, Dmitry Vyukov wrote: >> On Wed, Dec 5, 2012 at 4:42 PM, Jakub Jelinek <ja...@redhat.com> wrote: >> > On Wed, Dec 05, 2012 at 03:13:20PM +0400, Dmitry Vyukov wrote: >> >> I hope I addressed all your comments in this revision. >> >> I've fixed nand atomic operation, made atomic operations atomic again >> >> and added visibility attribute to interface functions. >> > >> > Mostly, still, __sync_lock_test_and_set isn't full barrier unlike most >> > other >> > __sync_* builtins, so either you'd need to use __atomic_exchange_n if >> > available (you can e.g. guard it with #ifdef __ATOMIC_SEQ_CST), or >> > if not available also use __sync_synchronize (). >> >> >> __atomic_xxx are available since gcc 4.7 and we use gcc 4.4 on some >> bots. I do not want to clutter the code with macros too much. > > Well, one option would be just a single #ifdef somewhere, > #ifndef __ATOMIC_SEQ_CST > #define __ATOMIC_RELEASE 0 > ... > #define __atomic_...(...) __sync_...(...) > ... > #endif > i.e. implement __atomic_* as macros using __sync_* if not available, > then just use __atomic_* in the code everywhere. > > Jakub
