Jakub Jelinek wrote:
Tested with libgomp testsuite and looked at performance numbers of Sho's
omp_fib.c and libgomp.c/sort-1.c, unfortunately the differences looked to be
in the noise. But, e.g. on omp_fib.c strace -f shows that the number of
futex FUTEX_WAKE syscalls went down a lot (from ~ 750
On 07/15/2011 02:37 AM, Jakub Jelinek wrote:
> Any comments? Can anyone see meassurable differences on some benchmark?
>
> 2011-07-15 Jakub Jelinek
>
> * config/linux/wait.h (do_spin): New inline, largely copied
> from do_wait, just don't do futex_wait here, instead return true if
On Fri, Jul 15, 2011 at 12:02:06PM +0200, Paolo Bonzini wrote:
> On 07/15/2011 11:37 AM, Jakub Jelinek wrote:
> >While __sync_lock_test_and_set isn't a full barrier on all targets,
> >I hope it doesn't matter, because the inline caller has already done one
> >__sync_val_compare_and_swap which is a
On 07/15/2011 11:37 AM, Jakub Jelinek wrote:
While __sync_lock_test_and_set isn't a full barrier on all targets,
I hope it doesn't matter, because the inline caller has already done one
__sync_val_compare_and_swap which is a full barrier.
Why not take the occasion to add the __sync_swap extensi
Hi!
If lock contention is high, but all locks are held for relatively short time
and no threads actually goes into futex_wait, we still completely
unnecessarily do lots of futex_wake syscalls.
On Linux with futex, our mutexes have 3 states:
0 - unlocked
1 - locked, uncontended
2 - locked, contend
