[OpenMP] libomp: Fix dist barrier for nested parallel/team

[t-msn]

Summary

Some code path in nested parallelism miss updates for dist barrier
which could lead hang. Add appropriate function calls.

[t-msn]

I noticed this when running check-openmp with dist barrier by using https://reviews.llvm.org/D122645
btw, it seems parallel tests run does not work when using dist barrier, but I couldn't figure out the reason.

[tlwilmar]

Hi -- can you tell me which specific tests failed reliably? Was it one particular arch? I want to make sure I'm covering those tests and will give your change a try and review soon. Thanks!

[t-msn]

Hi
I saw check-openmp hang on both x86 and arm64. Adding timeout value to lit option gives me the following result:

$ env LIT_OPTS="--show-xfail --timeout=60 -j 1"  CHECK_OPENMP_ENV="KMP_FORKJOIN_BARRIER_PATTERN=dist,dist KMP_PLAIN_BARRIER_PATTERN=dist,dist KMP_REDUCTION_BARRIER_PATTERN=dist,dist" ninja check-openmp
...
Timed Out Tests (15):
  libomp :: affinity/bug-nested.c
  libomp :: affinity/format/fields_values.c
  libomp :: affinity/format/nested.c
  libomp :: affinity/format/nested2.c
  libomp :: affinity/format/nested_mixed.c
  libomp :: env/omp_thread_limit.c
  libomp :: ompt/misc/threads_nested.c
  libomp :: ompt/parallel/nested.c
  libomp :: ompt/parallel/nested_lwt.c
  libomp :: ompt/parallel/nested_thread_num.c
  libomp :: ompt/parallel/nested_threadnum.c
  libomp :: ompt/teams/parallel_team.c
  libomp :: parallel/omp_nested.c
  libomp :: teams/kmp_num_teams.c
  libomp :: worksharing/for/omp_monotonic_schedule_set_get.c

As these tests are related to nest/teams, I spotted suspected code path. With this patch, these timeout failures are gone.
(In addition to above timeouts, ompt/synchronization/reduction/tree_reduce.c is failed. I believe this is a test problem and partially addressed in D123359)

openmp/runtime/src/kmp_runtime.cpp

for (i = 1; i < team->t.t_nproc; i++) {
           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
         }
       }
     }
+    if (__kmp_barrier_release_pattern[bs_forkjoin_barrier] == bp_dist_bar) {
+      __kmp_add_threads_to_team(team, team->t.t_nproc);
+    }
 
 #if KMP_AFFINITY_SUPPORTED
     // Do not partition the places list for teams construct workers who
     // haven't actually been forked to do real work yet. This partitioning
     // will take place in the parallel region nested within the teams construct.
     if (!fork_teams_workers) {
       __kmp_partition_places(team);
     }

 #endif
         }
         if (__kmp_tasking_mode != tskm_immediate_exec) {
           // Synchronize thread's task state
           other_threads[i]->th.th_task_state = master_th->th.th_task_state;
         }
       }
     }
+    if (__kmp_barrier_release_pattern[bs_forkjoin_barrier] == bp_dist_bar) {
+      // Reset barrier status
+      __kmp_resize_dist_barrier(team, team->t.t_nproc, team->t.t_nproc);
+    }
 
 #if OMPT_SUPPORT
     if (ompt_enabled.enabled) {
       __kmp_join_ompt(gtid, master_th, parent_team, &ompt_parallel_data,
                       OMPT_INVOKER(fork_context) | ompt_parallel_team, codeptr);
     }
 #endif