[OpenMP][Tools] Fix handling of initial-task-end

[protze.joachim]

Summary

Latest OpenMP spec says parallel_data is NULL for initial/implicit-task-end.
We nevertheless need to cleanup the ParallelData here, as there is no other
callback for the end of the implicit parallel region. We can use the reference
stored in the TaskData.

[dreachem]

What happens for a teams construct? The implicit tasks in that case are all initial tasks. Should this Data->Team->Delete() apply there as well, or only for the implicit parallel region for the whole program?

[protze.joachim]

In D114005#3135157, @dreachem wrote:

What happens for a teams construct? The implicit tasks in that case are all initial tasks. Should this Data->Team->Delete() apply there as well, or only for the implicit parallel region for the whole program?

I think, support for teams is a separate issue, I created D114020 for this.
Looking at the implicit flushes section, I think we need to add an ordering constraint for initial-task-end events and the teams-end event. ("after all initial-task-end events" instead of "after the thread executes its initial-task-end event")

[dreachem]

Thanks. This LGTM.

[RitanyaB]

Hi, @protze.joachim.
We have a dependency on this patch, therefore we'd like to know if you'll be committing it or if it's been held back for unknown reasons.

openmp/tools/archer/ompt-tsan.cpp

static void ompt_tsan_implicit_task(ompt_scope_endpoint_t endpoint,
   case ompt_scope_end: {
     TaskData *Data = ToTaskData(task_data);
 #ifdef DEBUG
     assert(Data->freed == 0 && "Implicit task end should only be called once!");
     Data->freed = 1;
 #endif
     assert(Data->RefCount == 1 &&
            "All tasks should have finished at the implicit barrier!");
-    Data->Delete();
     if (type & ompt_task_initial) {
-      ToParallelData(parallel_data)->Delete();
+      Data->Team->Delete();
     }
+    Data->Delete();
     TsanFuncExit();
     break;
   }
   case ompt_scope_beginend:
     // Should not occur according to OpenMP 5.1
     // Tested in OMPT tests
     break;
   }