Hello
I detected this issue as I was updating the patches to send to the mailing
list. I have not created a bug report.
When the CAS operation fails and expected == guard_bit, __cxa_guard_acquire
will return immediately indicating that the initialisation has already
succeeded. However, it's missing the acquire barrier for the changes done on
the other thread, to match the release barrier from __cxa_guard_release.
That is:
thread A thread B
load.acq == 0 load.acq == 0
__cxa_guard_acquire __cxa_guard_acquire
CAS(0 -> 256) success
__cxa_guard_release
store.rel(1)
CAS(0 ->256) fails
At this point, we must synchronise with the store-release from thread A.
--
Thiago Macieira - thiago.macieira (AT) intel.com
Software Architect - Intel Open Source Technology Center
Intel Sweden AB - Registration Number: 556189-6027
Knarrarnäsgatan 15, 164 40 Kista, Stockholm, Sweden
2012-08-30 Thiago Macieira <thiago.macie...@intel.com>
* libsupc++/guard.cc (__cxa_guard_acquire): must use acquire semantics
in case of failure, to acquire changes done by the other thread
---
libstdc++-v3/libsupc++/guard.cc | 2 +-
1 file changed, 1 insertion(+), 1 deletion(-)
diff --git a/libstdc++-v3/libsupc++/guard.cc b/libstdc++-v3/libsupc++/guard.cc
index 36352e7..73d7221 100644
--- a/libstdc++-v3/libsupc++/guard.cc
+++ b/libstdc++-v3/libsupc++/guard.cc
@@ -253,7 +253,7 @@ namespace __cxxabiv1
int expected(0);
if (__atomic_compare_exchange_n(gi, &expected, pending_bit, false,
__ATOMIC_ACQ_REL,
- __ATOMIC_RELAXED))
+ __ATOMIC_ACQUIRE))
{
// This thread should do the initialization.
return 1;
--
1.7.11.4
