Implicit tasks of worker threads might be reused before implicit-task-end event

Author: jprotze

runtime/src/kmp_barrier.cpp

@@ -1737,8 +1737,9 @@ __kmp_fork_barrier(int gtid, int tid)
     {
         if (this_thr->th.ompt_thread_info.state == ompt_state_wait_barrier_implicit) {
             int ds_tid = this_thr->th.th_info.ds.ds_tid;
-            ompt_task_data_t* tId = &(this_thr->th.th_current_task->ompt_task_info.task_data);
+            ompt_task_data_t* tId = (team)? &(this_thr->th.th_current_task->ompt_task_info.task_data) : &(this_thr->th.ompt_thread_info.task_data);
             ompt_parallel_data_t* pId = (team)? &(team->t.ompt_team_info.parallel_data) : &(this_thr->th.ompt_thread_info.parallel_data);
+//            ompt_parallel_data_t* pId = (team)? &(team->t.ompt_team_info.parallel_data) : NULL;
             this_thr->th.ompt_thread_info.state = ompt_state_overhead;
 #if OMPT_OPTIONAL
             void * codeptr = NULL;

runtime/src/kmp_runtime.cpp

@@ -5615,6 +5615,7 @@ __kmp_launch_thread( kmp_info_t *this_thr )
 
                 this_thr->th.ompt_thread_info.state = ompt_state_overhead;
                 this_thr->th.ompt_thread_info.parallel_data = (*pteam)->t.ompt_team_info.parallel_data;
+                this_thr->th.ompt_thread_info.task_data = *OMPT_CUR_TASK_DATA(this_thr);
             }
 #endif
             /* join barrier after parallel region */

runtime/src/kmp_wait_release.h

@@ -234,7 +234,8 @@ THIS function is called from
             }
         } else {        
                 pId = &(this_thr->th.ompt_thread_info.parallel_data);
-                tId = &(this_thr->th.th_current_task->ompt_task_info.task_data);
+//                pId = NULL;
+                tId = &(this_thr->th.ompt_thread_info.task_data);
         }
 #if OMPT_OPTIONAL
         if (ompt_entry_state == ompt_state_idle) {

runtime/src/ompt-internal.h

@@ -61,11 +61,12 @@ typedef struct ompt_lw_taskteam_s {
 
 
 typedef struct {
-    ompt_thread_data_t  thread_data;
+    ompt_thread_data_t    thread_data;
     ompt_parallel_data_t  parallel_data; /* stored here from implicit barrier-begin until implicit-task-end */
-    ompt_state_t        state;
-    ompt_wait_id_t      wait_id;
-    void                *idle_frame;
+    ompt_task_data_t      task_data; /* stored here from implicit barrier-begin until implicit-task-end */
+    ompt_state_t          state;
+    ompt_wait_id_t        wait_id;
+    void                  *idle_frame;
 } ompt_thread_info_t;
 
 

runtime/test/ompt/cancel/cancel_taskgroup.c

@@ -1,4 +1,4 @@
-// RUN:  %libomp-compile && env OMP_CANCELLATION=true %libomp-run | %sort-threads | FileCheck %s
+// RUN:  %libomp-compile && env OMP_CANCELLATION=true %libomp-run | %sort-threads | tee log.txt | FileCheck %s
 // REQUIRES: ompt
 
 #include "callback.h"
@@ -20,6 +20,7 @@ int main()
           printf("start execute task 1\n");
           OMPT_SIGNAL(condition);
           OMPT_WAIT(condition,2);
+          usleep(300); // master needs time to discard tasks 
           #pragma omp cancellation point taskgroup
           printf("end execute task 1\n");
         }