On Fri, Aug 29, 2014 at 10:40:57AM -0700, Richard Henderson wrote:
> On 08/06/2014 03:05 AM, Varvara Rainchik wrote:
> > * libgomp.h (gomp_thread): For non TLS case create thread data.
> > * team.c (create_non_tls_thread_data): New function.
> >
> >
> > ---
> > diff --git a/libgomp/libgomp.h b/libgomp/libgomp.h
> > index a1482cc..cf3ec8f 100644
> > --- a/libgomp/libgomp.h
> > +++ b/libgomp/libgomp.h
> > @@ -479,9 +479,15 @@ static inline struct gomp_thread *gomp_thread (void)
> > }
> > #else
> > extern pthread_key_t gomp_tls_key;
> > +extern struct gomp_thread *create_non_tls_thread_data (void);
> > static inline struct gomp_thread *gomp_thread (void)
> > {
> > - return pthread_getspecific (gomp_tls_key);
> > + struct gomp_thread *thr = pthread_getspecific (gomp_tls_key);
> > + if (thr == NULL)
> > + {
> > + thr = create_non_tls_thread_data ();
> > + }
> > + return thr;
> > }
>
> This should never happen.
I guess it can happen if you mix up explicit pthread_create and libgomp APIs.
initialize_team will only initialize it in the initial thread, while if you
use #pragma omp ... or omp_* calls from a thread created with
pthread_create, in the !HAVE_TLS case pthread_getspecific will return NULL.
Now, the patch doesn't handle that case completely though (and is badly
formatted), the problem is that if we allocate in the !HAVE_TLS case
in non-initial thread the TLS data, we want to free them again, so that
would mean pthread_key_create with non-NULL destructor, and then we need to
differentiate in between the 3 cases - key equal to &initial_thread_tls_data
(would need to move out of the block context), no freeing needed, thread
created by gomp_thread_start, no freeing needed, otherwise free.
> The thread-specific data is set in gomp_thread_start and initialize_team.
>
> Where are you getting a call to gomp_thread that hasn't been through one of
> those functions?
Jakub
