On Thu, Sep 06, 2012 at 01:33:11PM -0700, Benjamin De Kosnik wrote:
> Here's the patch as applied to trunk in rev. 191042. I'll apply it to
> 4.7 this weekend as long as nobody yelps.
> 2012-09-06 Thiago Macieira <thiago.macie...@intel.com>
>
> PR libstdc++/54172
> * libsupc++/guard.cc (__cxa_guard_acquire): Exit the loop earlier if
> we detect that another thread has had success. Don't compare_exchange
> from a finished state back to a waiting state. Comment.
>
> diff --git a/libstdc++-v3/libsupc++/guard.cc b/libstdc++-v3/libsupc++/guard.cc
> index adc9608..60165cd 100644
> --- a/libstdc++-v3/libsupc++/guard.cc
> +++ b/libstdc++-v3/libsupc++/guard.cc
> @@ -244,13 +244,13 @@ namespace __cxxabiv1
> if (__gthread_active_p ())
> {
> int *gi = (int *) (void *) g;
> - int expected(0);
> const int guard_bit = _GLIBCXX_GUARD_BIT;
> const int pending_bit = _GLIBCXX_GUARD_PENDING_BIT;
> const int waiting_bit = _GLIBCXX_GUARD_WAITING_BIT;
>
> while (1)
> {
> + int expected(0);
> if (__atomic_compare_exchange_n(gi, &expected, pending_bit, false,
> __ATOMIC_ACQ_REL,
> __ATOMIC_RELAXED))
Shouldn't this __ATOMIC_RELAXED be also __ATOMIC_ACQUIRE? If expected ends
up being guard_bit, then the code will return 0; right away.
> @@ -264,13 +264,26 @@ namespace __cxxabiv1
> // Already initialized.
> return 0;
> }
> +
> if (expected == pending_bit)
> {
> + // Use acquire here.
> int newv = expected | waiting_bit;
> if (!__atomic_compare_exchange_n(gi, &expected, newv, false,
> __ATOMIC_ACQ_REL,
> - __ATOMIC_RELAXED))
> - continue;
> + __ATOMIC_ACQUIRE))
> + {
> + if (expected == guard_bit)
> + {
> + // Make a thread that failed to set the
> + // waiting bit exit the function earlier,
> + // if it detects that another thread has
> + // successfully finished initialising.
> + return 0;
> + }
> + if (expected == 0)
> + continue;
> + }
>
> expected = newv;
> }
Jakub
