[OpenMP] libomp: fix mutexinoutset dependence for proxy tasks

[AndreyChurbanov]

Summary

Once kmp_task_finish is not executed for proxy tasks,
move mutexinoutset dependency code to kmp_release_deps
which is executed for all task kinds.

[tianshilei1992]

LGTM.

openmp/runtime/src/kmp_taskdeps.h

93

Why is it "unlikely"? Because mutexinoutset is a rare case?

[AndreyChurbanov]

openmp/runtime/src/kmp_taskdeps.h

93

Why is it "unlikely"? Because mutexinoutset is a rare case?

Right. We still haven't seen real code that uses mutexinoutset dependence. Even if such code exist there are much more codes those do not use mutexinoutset. So the performance of this code block looks lesser priority comparing to quicker skip the code when it is not needed.

[tianshilei1992]

openmp/runtime/src/kmp_taskdeps.h

93

Gotcha. Thanks for the information.

openmp/runtime/src/kmp_taskdeps.h

#else
   __kmp_thread_free(thread, h);
 #endif
 }
 
 static inline void __kmp_release_deps(kmp_int32 gtid, kmp_taskdata_t *task) {
   kmp_info_t *thread = __kmp_threads[gtid];
   kmp_depnode_t *node = task->td_depnode;
 
+  // Check mutexinoutset dependencies, release locks
+  if (UNLIKELY(node && (node->dn.mtx_num_locks < 0))) {
+    // negative num_locks means all locks were acquired
+    node->dn.mtx_num_locks = -node->dn.mtx_num_locks;
+    for (int i = node->dn.mtx_num_locks - 1; i >= 0; --i) {
+      KMP_DEBUG_ASSERT(node->dn.mtx_locks[i] != NULL);
+      __kmp_release_lock(node->dn.mtx_locks[i], gtid);
+    }
+  }
+
   if (task->td_dephash) {
     KA_TRACE(
         40, ("__kmp_release_deps: T#%d freeing dependencies hash of task %p.\n",
              gtid, task));
     __kmp_dephash_free(thread, task->td_dephash);
     task->td_dephash = NULL;
   }
 

openmp/runtime/src/kmp_tasking.cpp

if (counter > 0) {
       resumed_task->td_flags.executing = 1; // resume previous task
       KA_TRACE(10, ("__kmp_task_finish(exit): T#%d partially done task %p, "
                     "resuming task %p\n",
                     gtid, taskdata, resumed_task));
       return;
     }
   }
 
-  // Check mutexinoutset dependencies, release locks
-  kmp_depnode_t *node = taskdata->td_depnode;
-  if (node && (node->dn.mtx_num_locks < 0)) {
-    // negative num_locks means all locks were acquired
-    node->dn.mtx_num_locks = -node->dn.mtx_num_locks;
-    for (int i = node->dn.mtx_num_locks - 1; i >= 0; --i) {
-      KMP_DEBUG_ASSERT(node->dn.mtx_locks[i] != NULL);
-      __kmp_release_lock(node->dn.mtx_locks[i], gtid);
-    }
-  }
-
   // bookkeeping for resuming task:
   // GEH - note tasking_ser => task_serial
   KMP_DEBUG_ASSERT(
       (taskdata->td_flags.tasking_ser || taskdata->td_flags.task_serial) ==
       taskdata->td_flags.task_serial);
   if (taskdata->td_flags.task_serial) {
     if (resumed_task == NULL) {
       resumed_task = taskdata->td_parent; // In a serialized task, the resumed