Bug fix: removed zeroing of the th.th_task_state for master thread at start of nested parallel.

[jlpeyton]

Summary

The th.th_task_state for the master thread at the start of a nested parallel should not be zeroed in kmp_allocate_team() because it is later put in the stack of states in kmp_fork_call() for further re-use after exiting the nested region. It is zeroed after being put in the stack.

[AndreyChurbanov]

LGTM

openmp/trunk/runtime/src/kmp_runtime.c

# if KMP_AFFINITY_SUPPORTED
                 __kmp_partition_places( team );
             }
 # else
             if ( team->t.t_proc_bind != new_proc_bind ) {
                 team->t.t_proc_bind = new_proc_bind;
             }
 # endif /* KMP_AFFINITY_SUPPORTED */
 #endif /* OMP_40_ENABLED */
-
-            if (level) {
-                for(f = 0; f < new_nproc; ++f) {
-                    team->t.t_threads[f]->th.th_task_state = 0;
-                }
-            }
         }
         else if( team->t.t_nproc > new_nproc ) {
             KA_TRACE( 20, ("__kmp_allocate_team: decreasing hot team thread count to %d\n", new_nproc ));
 
             team->t.t_size_changed = 1;
             if ( __kmp_tasking_mode != tskm_immediate_exec ) {
                 // Signal the worker threads (esp. extra ones) to stop looking for tasks while spin waiting.
                 // The task teams are reference counted and will be deallocated by the last worker thread.
#endif // KMP_NESTED_HOT_TEAMS
                         task_team->tt.tt_unfinished_threads = new_nproc;
                         task_team->tt.tt_ref_ct = new_nproc - 1;
                     }
                 }
             }
 
             /* update the remaining threads */
             if (level) {
-                for(f = 0; f < new_nproc; ++f) {
+                team->t.t_threads[0]->th.th_team_nproc = new_nproc;
+                for(f = 1; f < new_nproc; ++f) {
                     team->t.t_threads[f]->th.th_team_nproc = new_nproc;
                     team->t.t_threads[f]->th.th_task_state = 0;
                 }
             }
             else {
                 for(f = 0; f < new_nproc; ++f) {
                     team->t.t_threads[f]->th.th_team_nproc = new_nproc;
                 }
#endif // KMP_NESTED_HOT_TEAMS
                         team->t.t_task_team[tt_idx] = NULL;
                     }
                 }
             }
 
             /* reinitialize the threads */
             KMP_DEBUG_ASSERT(team->t.t_nproc == new_nproc);
             if (level) {
+                int old_state = team->t.t_threads[0]->th.th_task_state;
                 for (f=0;  f < team->t.t_nproc; ++f)
                     __kmp_initialize_info( team->t.t_threads[ f ], team, f, __kmp_gtid_from_tid( f, team ) );
+                // th_task_state for master thread will be put in stack of states in __kmp_fork_call()
+                // before zeroing, for workers it was just zeroed in __kmp_initialize_info()
+                team->t.t_threads[0]->th.th_task_state = old_state;
             }
             else {
                 int old_state = team->t.t_threads[0]->th.th_task_state;
                 for (f=0;  f<team->t.t_nproc; ++f) {
                     __kmp_initialize_info( team->t.t_threads[ f ], team, f, __kmp_gtid_from_tid( f, team ) );
                     team->t.t_threads[f]->th.th_task_state = old_state;
                     team->t.t_threads[f]->th.th_task_team = team->t.t_task_team[old_state];
                 }
#if OMP_40_ENABLED
                 thr->th.th_teams_microtask = master->th.th_teams_microtask;
                 thr->th.th_teams_level     = master->th.th_teams_level;
                 thr->th.th_teams_size      = master->th.th_teams_size;
             }
         }
 #endif /* OMP_40_ENABLED */
 #if KMP_NESTED_HOT_TEAMS
         if( level ) {
-            // Sync task (TODO: and barrier?) state for nested hot teams, not needed for outermost hot team.
+            // Sync barrier state for nested hot teams, not needed for outermost hot team.
             for( f = 1; f < new_nproc; ++f ) {
                 kmp_info_t *thr = team->t.t_threads[f];
-                thr->th.th_task_state = 0;
                 int b;
                 kmp_balign_t * balign = thr->th.th_bar;
                 for( b = 0; b < bs_last_barrier; ++ b ) {
                     balign[ b ].bb.b_arrived        = team->t.t_bar[ b ].b_arrived;
                     KMP_DEBUG_ASSERT(balign[b].bb.wait_flag != KMP_BARRIER_PARENT_FLAG);
 #if USE_DEBUGGER
                     balign[ b ].bb.b_worker_arrived = team->t.t_bar[ b ].b_team_arrived;
 #endif