[OpenMP] Fix for https://bugs.llvm.org/show_bug.cgi?id=45904.

[AndreyChurbanov]

Summary

Fixed task bookkeeping for serialized detached tasks.

Not sure how to properly request the OpenMP 5.0 option for the test; hardcoded the -fopenmp-version=50 option for now. If somebody knows a better way, please advise.

[ABataev]

-fopenmp-version=50 is the right way to provide a version of OpenMP.

[jdoerfert]

LGTM, one comment.

openmp/runtime/src/kmp_tasking.cpp

918–919

Update the comment please.

[Hahnfeld]

Does -fopenmp-version work with the Intel Compiler? How will this work if the GCC eventually gets support for OpenMP 5.0?

[AndreyChurbanov]

Comments addressed.

[AndreyChurbanov]

In D79944#2037911, @Hahnfeld wrote:

Does -fopenmp-version work with the Intel Compiler? How will this work if the GCC eventually gets support for OpenMP 5.0?

Thanks for comment. I added Intel legacy compiler icc to the list of unsupported compilers as it supports neither detach clause nor -fopenmp-version option; and listed particular gcc versions, so that next version could be tested if they will implement the support of detach clause.

[protze.joachim]

Can FileCheck deal with tests, which potentially deadlock/stall?

This was the initial behavior of the test provided in the issue.

[jdoerfert]

In D79944#2041506, @protze.joachim wrote:

Can FileCheck deal with tests, which potentially deadlock/stall?

This was the initial behavior of the test provided in the issue.

It will deadlock/stall with the program, thus the problem will show up.

[protze.joachim]

I guess my question was: can lit and the build bots deal with such deadlock? Do they employ timeouts?
Or is there some commonly used mechanism of timeout, which we need to include in the runtime?

[jdoerfert]

In D79944#2044461, @protze.joachim wrote:

I guess my question was: can lit and the build bots deal with such deadlock? Do they employ timeouts?
Or is there some commonly used mechanism of timeout, which we need to include in the runtime?

The bots deal with this as we accidentally run into endless loops in the compiler more often than one would think ;)

I guess lit itself will just hang, maybe there is a timeout option builtin or you can manually wrap it in a timeout script if you need to.

openmp/runtime/src/kmp_tasking.cpp

if (taskdata->td_allow_completion_event.type ==
       }
       __kmp_release_tas_lock(&taskdata->td_allow_completion_event.lock, gtid);
     }
   }
 
   if (!detach) {
     taskdata->td_flags.complete = 1; // mark the task as completed
 
     // Only need to keep track of count if team parallel and tasking not
-    // serialized
-    if (!(taskdata->td_flags.team_serial || taskdata->td_flags.tasking_ser)) {
+    // serialized, or task is detachable and event has already been fulfilled 
+    if (!(taskdata->td_flags.team_serial || taskdata->td_flags.tasking_ser) ||
+        taskdata->td_flags.detachable == TASK_DETACHABLE) {
       // Predecrement simulated by "- 1" calculation
       children =
           KMP_ATOMIC_DEC(&taskdata->td_parent->td_incomplete_child_tasks) - 1;
       KMP_DEBUG_ASSERT(children >= 0);
       if (taskdata->td_taskgroup)
         KMP_ATOMIC_DEC(&taskdata->td_taskgroup->count);
       __kmp_release_deps(gtid, taskdata);
     } else if (task_team && task_team->tt.tt_found_proxy_tasks) {

openmp/runtime/test/tasking/omp_detach_taskwait.c

+// RUN: %libomp-compile -fopenmp-version=50 && env OMP_NUM_THREADS='3' %libomp-run
+// RUN: %libomp-compile -fopenmp-version=50 && env OMP_NUM_THREADS='1' %libomp-run
+
+// Checked gcc 9.2 still does not support detach clause on task construct.
+// UNSUPPORTED: gcc-4, gcc-5, gcc-6, gcc-7, gcc-8, gcc-9
+// clang supports detach clause since version 11.
+// UNSUPPORTED: clang-10, clang-9, clang-8, clang-7
+// icc compiler does not support detach clause.
+// UNSUPPORTED: icc
+
+#include <omp.h>
+
+int main()
+{
+  #pragma omp parallel
+  #pragma omp master
+  {
+    omp_event_handle_t event;
+    #pragma omp task detach(event)
+    {
+      omp_fulfill_event(event);
+    }
+    #pragma omp taskwait
+  }
+  return 0;
+}