[OpenMP] Fix shutdown hang bug

[jlpeyton]

Summary

The deadlock happens when the primary thread gets initz lock and tries to join
the worker thread which waits for the same lock in TLS key destructor.
The patch removes the lock and the code of setting TLS value which needed
the lock. Also removed setting TLS from __kmp_unregister_root_current_thread.

[AndreyChurbanov]

LGTM

openmp/runtime/src/kmp_runtime.cpp

#if OMPT_SUPPORT
     // the runtime is shutting down so we won't report any events
     thread->th.ompt_thread_info.state = ompt_state_undefined;
 #endif
     __kmp_task_team_wait(thread, team USE_ITT_BUILD_ARG(NULL));
   }
 
   __kmp_reset_root(gtid, root);
 
-  /* free up this thread slot */
-  __kmp_gtid_set_specific(KMP_GTID_DNE);
-#ifdef KMP_TDATA_GTID
-  __kmp_gtid = KMP_GTID_DNE;
-#endif
-
   KMP_MB();
   KC_TRACE(10,
            ("__kmp_unregister_root_current_thread: T#%d unregistered\n", gtid));
 
   __kmp_release_bootstrap_lock(&__kmp_forkjoin_lock);
 }
 
 #if KMP_OS_WINDOWS

 #if KMP_COMPILER_ICC
 #pragma warning(pop)
 #endif
 
   KA_TRACE(30, ("__kmp_internal_end_dest: T#%d\n", gtid));
   /* NOTE: the gtid is stored as gitd+1 in the thread-local-storage
    * this is because 0 is reserved for the nothing-stored case */
 
-  /* josh: One reason for setting the gtid specific data even when it is being
-     destroyed by pthread is to allow gtid lookup through thread specific data
-     (__kmp_gtid_get_specific).  Some of the code, especially stat code,
-     that gets executed in the call to __kmp_internal_end_thread, actually
-     gets the gtid through the thread specific data.  Setting it here seems
-     rather inelegant and perhaps wrong, but allows __kmp_internal_end_thread
-     to run smoothly.
-     todo: get rid of this after we remove the dependence on
-     __kmp_gtid_get_specific  */
-  if (gtid >= 0 && KMP_UBER_GTID(gtid))
-    __kmp_gtid_set_specific(gtid);
-#ifdef KMP_TDATA_GTID
-  __kmp_gtid = gtid;
-#endif
   __kmp_internal_end_thread(gtid);
 }
 
 #if KMP_OS_UNIX && KMP_DYNAMIC_LIB
 
 __attribute__((destructor)) void __kmp_internal_end_dtor(void) {
   __kmp_internal_end_atexit();
 }