http://gcc.gnu.org/bugzilla/show_bug.cgi?id=48076
--- Comment #2 from Richard Henderson <rth at gcc dot gnu.org> 2012-11-27
18:01:17 UTC ---
Are you sure this isn't a false-positive?
The way I read this code, it is certainly possible for the optimizer
(or the processor) to prefetch emutls_key before the load of offset:
__gthread_key_t prefetch = emutls_key;
pointer offset = obj->loc.offset;
if (__builtin_expect (offset == 0, 0))
{
...
}
struct __emutls_array *arr = __gthread_getspecific (prefetch);
But the compiler should see the memory barriers within the if path
and insert
if (__builtin_expect (offset == 0, 0))
{
...
__gthread_mutex_unlock (&emutls_mutex);
prefetch = emutls_key;
}
and the processor had better cancel any speculative prefetch when
it sees the explicit barriers.
I'm also assuming here that Android is using the same gthr-posix.h
that is used by desktop glibc linux, and so there isn't a mistake
in the gthread macros causing a lack of barrier...
