xref: /freebsd/contrib/llvm-project/openmp/runtime/src/kmp_tasking.cpp (revision 1323ec571215a77ddd21294f0871979d5ad6b992)
1 /*
2  * kmp_tasking.cpp -- OpenMP 3.0 tasking support.
3  */
4 
5 //===----------------------------------------------------------------------===//
6 //
7 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
8 // See https://llvm.org/LICENSE.txt for license information.
9 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
10 //
11 //===----------------------------------------------------------------------===//
12 
13 #include "kmp.h"
14 #include "kmp_i18n.h"
15 #include "kmp_itt.h"
16 #include "kmp_stats.h"
17 #include "kmp_wait_release.h"
18 #include "kmp_taskdeps.h"
19 
20 #if OMPT_SUPPORT
21 #include "ompt-specific.h"
22 #endif
23 
24 /* forward declaration */
25 static void __kmp_enable_tasking(kmp_task_team_t *task_team,
26                                  kmp_info_t *this_thr);
27 static void __kmp_alloc_task_deque(kmp_info_t *thread,
28                                    kmp_thread_data_t *thread_data);
29 static int __kmp_realloc_task_threads_data(kmp_info_t *thread,
30                                            kmp_task_team_t *task_team);
31 static void __kmp_bottom_half_finish_proxy(kmp_int32 gtid, kmp_task_t *ptask);
32 
33 #ifdef BUILD_TIED_TASK_STACK
34 
35 //  __kmp_trace_task_stack: print the tied tasks from the task stack in order
36 //  from top do bottom
37 //
38 //  gtid: global thread identifier for thread containing stack
39 //  thread_data: thread data for task team thread containing stack
40 //  threshold: value above which the trace statement triggers
41 //  location: string identifying call site of this function (for trace)
42 static void __kmp_trace_task_stack(kmp_int32 gtid,
43                                    kmp_thread_data_t *thread_data,
44                                    int threshold, char *location) {
45   kmp_task_stack_t *task_stack = &thread_data->td.td_susp_tied_tasks;
46   kmp_taskdata_t **stack_top = task_stack->ts_top;
47   kmp_int32 entries = task_stack->ts_entries;
48   kmp_taskdata_t *tied_task;
49 
50   KA_TRACE(
51       threshold,
52       ("__kmp_trace_task_stack(start): location = %s, gtid = %d, entries = %d, "
53        "first_block = %p, stack_top = %p \n",
54        location, gtid, entries, task_stack->ts_first_block, stack_top));
55 
56   KMP_DEBUG_ASSERT(stack_top != NULL);
57   KMP_DEBUG_ASSERT(entries > 0);
58 
59   while (entries != 0) {
60     KMP_DEBUG_ASSERT(stack_top != &task_stack->ts_first_block.sb_block[0]);
61     // fix up ts_top if we need to pop from previous block
62     if (entries & TASK_STACK_INDEX_MASK == 0) {
63       kmp_stack_block_t *stack_block = (kmp_stack_block_t *)(stack_top);
64 
65       stack_block = stack_block->sb_prev;
66       stack_top = &stack_block->sb_block[TASK_STACK_BLOCK_SIZE];
67     }
68 
69     // finish bookkeeping
70     stack_top--;
71     entries--;
72 
73     tied_task = *stack_top;
74 
75     KMP_DEBUG_ASSERT(tied_task != NULL);
76     KMP_DEBUG_ASSERT(tied_task->td_flags.tasktype == TASK_TIED);
77 
78     KA_TRACE(threshold,
79              ("__kmp_trace_task_stack(%s):             gtid=%d, entry=%d, "
80               "stack_top=%p, tied_task=%p\n",
81               location, gtid, entries, stack_top, tied_task));
82   }
83   KMP_DEBUG_ASSERT(stack_top == &task_stack->ts_first_block.sb_block[0]);
84 
85   KA_TRACE(threshold,
86            ("__kmp_trace_task_stack(exit): location = %s, gtid = %d\n",
87             location, gtid));
88 }
89 
90 //  __kmp_init_task_stack: initialize the task stack for the first time
91 //  after a thread_data structure is created.
92 //  It should not be necessary to do this again (assuming the stack works).
93 //
94 //  gtid: global thread identifier of calling thread
95 //  thread_data: thread data for task team thread containing stack
96 static void __kmp_init_task_stack(kmp_int32 gtid,
97                                   kmp_thread_data_t *thread_data) {
98   kmp_task_stack_t *task_stack = &thread_data->td.td_susp_tied_tasks;
99   kmp_stack_block_t *first_block;
100 
101   // set up the first block of the stack
102   first_block = &task_stack->ts_first_block;
103   task_stack->ts_top = (kmp_taskdata_t **)first_block;
104   memset((void *)first_block, '\0',
105          TASK_STACK_BLOCK_SIZE * sizeof(kmp_taskdata_t *));
106 
107   // initialize the stack to be empty
108   task_stack->ts_entries = TASK_STACK_EMPTY;
109   first_block->sb_next = NULL;
110   first_block->sb_prev = NULL;
111 }
112 
113 //  __kmp_free_task_stack: free the task stack when thread_data is destroyed.
114 //
115 //  gtid: global thread identifier for calling thread
116 //  thread_data: thread info for thread containing stack
117 static void __kmp_free_task_stack(kmp_int32 gtid,
118                                   kmp_thread_data_t *thread_data) {
119   kmp_task_stack_t *task_stack = &thread_data->td.td_susp_tied_tasks;
120   kmp_stack_block_t *stack_block = &task_stack->ts_first_block;
121 
122   KMP_DEBUG_ASSERT(task_stack->ts_entries == TASK_STACK_EMPTY);
123   // free from the second block of the stack
124   while (stack_block != NULL) {
125     kmp_stack_block_t *next_block = (stack_block) ? stack_block->sb_next : NULL;
126 
127     stack_block->sb_next = NULL;
128     stack_block->sb_prev = NULL;
129     if (stack_block != &task_stack->ts_first_block) {
130       __kmp_thread_free(thread,
131                         stack_block); // free the block, if not the first
132     }
133     stack_block = next_block;
134   }
135   // initialize the stack to be empty
136   task_stack->ts_entries = 0;
137   task_stack->ts_top = NULL;
138 }
139 
140 //  __kmp_push_task_stack: Push the tied task onto the task stack.
141 //     Grow the stack if necessary by allocating another block.
142 //
143 //  gtid: global thread identifier for calling thread
144 //  thread: thread info for thread containing stack
145 //  tied_task: the task to push on the stack
146 static void __kmp_push_task_stack(kmp_int32 gtid, kmp_info_t *thread,
147                                   kmp_taskdata_t *tied_task) {
148   // GEH - need to consider what to do if tt_threads_data not allocated yet
149   kmp_thread_data_t *thread_data =
150       &thread->th.th_task_team->tt.tt_threads_data[__kmp_tid_from_gtid(gtid)];
151   kmp_task_stack_t *task_stack = &thread_data->td.td_susp_tied_tasks;
152 
153   if (tied_task->td_flags.team_serial || tied_task->td_flags.tasking_ser) {
154     return; // Don't push anything on stack if team or team tasks are serialized
155   }
156 
157   KMP_DEBUG_ASSERT(tied_task->td_flags.tasktype == TASK_TIED);
158   KMP_DEBUG_ASSERT(task_stack->ts_top != NULL);
159 
160   KA_TRACE(20,
161            ("__kmp_push_task_stack(enter): GTID: %d; THREAD: %p; TASK: %p\n",
162             gtid, thread, tied_task));
163   // Store entry
164   *(task_stack->ts_top) = tied_task;
165 
166   // Do bookkeeping for next push
167   task_stack->ts_top++;
168   task_stack->ts_entries++;
169 
170   if (task_stack->ts_entries & TASK_STACK_INDEX_MASK == 0) {
171     // Find beginning of this task block
172     kmp_stack_block_t *stack_block =
173         (kmp_stack_block_t *)(task_stack->ts_top - TASK_STACK_BLOCK_SIZE);
174 
175     // Check if we already have a block
176     if (stack_block->sb_next !=
177         NULL) { // reset ts_top to beginning of next block
178       task_stack->ts_top = &stack_block->sb_next->sb_block[0];
179     } else { // Alloc new block and link it up
180       kmp_stack_block_t *new_block = (kmp_stack_block_t *)__kmp_thread_calloc(
181           thread, sizeof(kmp_stack_block_t));
182 
183       task_stack->ts_top = &new_block->sb_block[0];
184       stack_block->sb_next = new_block;
185       new_block->sb_prev = stack_block;
186       new_block->sb_next = NULL;
187 
188       KA_TRACE(
189           30,
190           ("__kmp_push_task_stack(): GTID: %d; TASK: %p; Alloc new block: %p\n",
191            gtid, tied_task, new_block));
192     }
193   }
194   KA_TRACE(20, ("__kmp_push_task_stack(exit): GTID: %d; TASK: %p\n", gtid,
195                 tied_task));
196 }
197 
198 //  __kmp_pop_task_stack: Pop the tied task from the task stack.  Don't return
199 //  the task, just check to make sure it matches the ending task passed in.
200 //
201 //  gtid: global thread identifier for the calling thread
202 //  thread: thread info structure containing stack
203 //  tied_task: the task popped off the stack
204 //  ending_task: the task that is ending (should match popped task)
205 static void __kmp_pop_task_stack(kmp_int32 gtid, kmp_info_t *thread,
206                                  kmp_taskdata_t *ending_task) {
207   // GEH - need to consider what to do if tt_threads_data not allocated yet
208   kmp_thread_data_t *thread_data =
209       &thread->th.th_task_team->tt_threads_data[__kmp_tid_from_gtid(gtid)];
210   kmp_task_stack_t *task_stack = &thread_data->td.td_susp_tied_tasks;
211   kmp_taskdata_t *tied_task;
212 
213   if (ending_task->td_flags.team_serial || ending_task->td_flags.tasking_ser) {
214     // Don't pop anything from stack if team or team tasks are serialized
215     return;
216   }
217 
218   KMP_DEBUG_ASSERT(task_stack->ts_top != NULL);
219   KMP_DEBUG_ASSERT(task_stack->ts_entries > 0);
220 
221   KA_TRACE(20, ("__kmp_pop_task_stack(enter): GTID: %d; THREAD: %p\n", gtid,
222                 thread));
223 
224   // fix up ts_top if we need to pop from previous block
225   if (task_stack->ts_entries & TASK_STACK_INDEX_MASK == 0) {
226     kmp_stack_block_t *stack_block = (kmp_stack_block_t *)(task_stack->ts_top);
227 
228     stack_block = stack_block->sb_prev;
229     task_stack->ts_top = &stack_block->sb_block[TASK_STACK_BLOCK_SIZE];
230   }
231 
232   // finish bookkeeping
233   task_stack->ts_top--;
234   task_stack->ts_entries--;
235 
236   tied_task = *(task_stack->ts_top);
237 
238   KMP_DEBUG_ASSERT(tied_task != NULL);
239   KMP_DEBUG_ASSERT(tied_task->td_flags.tasktype == TASK_TIED);
240   KMP_DEBUG_ASSERT(tied_task == ending_task); // If we built the stack correctly
241 
242   KA_TRACE(20, ("__kmp_pop_task_stack(exit): GTID: %d; TASK: %p\n", gtid,
243                 tied_task));
244   return;
245 }
246 #endif /* BUILD_TIED_TASK_STACK */
247 
248 // returns 1 if new task is allowed to execute, 0 otherwise
249 // checks Task Scheduling constraint (if requested) and
250 // mutexinoutset dependencies if any
251 static bool __kmp_task_is_allowed(int gtid, const kmp_int32 is_constrained,
252                                   const kmp_taskdata_t *tasknew,
253                                   const kmp_taskdata_t *taskcurr) {
254   if (is_constrained && (tasknew->td_flags.tiedness == TASK_TIED)) {
255     // Check if the candidate obeys the Task Scheduling Constraints (TSC)
256     // only descendant of all deferred tied tasks can be scheduled, checking
257     // the last one is enough, as it in turn is the descendant of all others
258     kmp_taskdata_t *current = taskcurr->td_last_tied;
259     KMP_DEBUG_ASSERT(current != NULL);
260     // check if the task is not suspended on barrier
261     if (current->td_flags.tasktype == TASK_EXPLICIT ||
262         current->td_taskwait_thread > 0) { // <= 0 on barrier
263       kmp_int32 level = current->td_level;
264       kmp_taskdata_t *parent = tasknew->td_parent;
265       while (parent != current && parent->td_level > level) {
266         // check generation up to the level of the current task
267         parent = parent->td_parent;
268         KMP_DEBUG_ASSERT(parent != NULL);
269       }
270       if (parent != current)
271         return false;
272     }
273   }
274   // Check mutexinoutset dependencies, acquire locks
275   kmp_depnode_t *node = tasknew->td_depnode;
276   if (UNLIKELY(node && (node->dn.mtx_num_locks > 0))) {
277     for (int i = 0; i < node->dn.mtx_num_locks; ++i) {
278       KMP_DEBUG_ASSERT(node->dn.mtx_locks[i] != NULL);
279       if (__kmp_test_lock(node->dn.mtx_locks[i], gtid))
280         continue;
281       // could not get the lock, release previous locks
282       for (int j = i - 1; j >= 0; --j)
283         __kmp_release_lock(node->dn.mtx_locks[j], gtid);
284       return false;
285     }
286     // negative num_locks means all locks acquired successfully
287     node->dn.mtx_num_locks = -node->dn.mtx_num_locks;
288   }
289   return true;
290 }
291 
292 // __kmp_realloc_task_deque:
293 // Re-allocates a task deque for a particular thread, copies the content from
294 // the old deque and adjusts the necessary data structures relating to the
295 // deque. This operation must be done with the deque_lock being held
296 static void __kmp_realloc_task_deque(kmp_info_t *thread,
297                                      kmp_thread_data_t *thread_data) {
298   kmp_int32 size = TASK_DEQUE_SIZE(thread_data->td);
299   KMP_DEBUG_ASSERT(TCR_4(thread_data->td.td_deque_ntasks) == size);
300   kmp_int32 new_size = 2 * size;
301 
302   KE_TRACE(10, ("__kmp_realloc_task_deque: T#%d reallocating deque[from %d to "
303                 "%d] for thread_data %p\n",
304                 __kmp_gtid_from_thread(thread), size, new_size, thread_data));
305 
306   kmp_taskdata_t **new_deque =
307       (kmp_taskdata_t **)__kmp_allocate(new_size * sizeof(kmp_taskdata_t *));
308 
309   int i, j;
310   for (i = thread_data->td.td_deque_head, j = 0; j < size;
311        i = (i + 1) & TASK_DEQUE_MASK(thread_data->td), j++)
312     new_deque[j] = thread_data->td.td_deque[i];
313 
314   __kmp_free(thread_data->td.td_deque);
315 
316   thread_data->td.td_deque_head = 0;
317   thread_data->td.td_deque_tail = size;
318   thread_data->td.td_deque = new_deque;
319   thread_data->td.td_deque_size = new_size;
320 }
321 
322 //  __kmp_push_task: Add a task to the thread's deque
323 static kmp_int32 __kmp_push_task(kmp_int32 gtid, kmp_task_t *task) {
324   kmp_info_t *thread = __kmp_threads[gtid];
325   kmp_taskdata_t *taskdata = KMP_TASK_TO_TASKDATA(task);
326 
327   // We don't need to map to shadow gtid if it is already hidden helper thread
328   if (taskdata->td_flags.hidden_helper && !KMP_HIDDEN_HELPER_THREAD(gtid)) {
329     gtid = KMP_GTID_TO_SHADOW_GTID(gtid);
330     thread = __kmp_threads[gtid];
331   }
332 
333   kmp_task_team_t *task_team = thread->th.th_task_team;
334   kmp_int32 tid = __kmp_tid_from_gtid(gtid);
335   kmp_thread_data_t *thread_data;
336 
337   KA_TRACE(20,
338            ("__kmp_push_task: T#%d trying to push task %p.\n", gtid, taskdata));
339 
340   if (UNLIKELY(taskdata->td_flags.tiedness == TASK_UNTIED)) {
341     // untied task needs to increment counter so that the task structure is not
342     // freed prematurely
343     kmp_int32 counter = 1 + KMP_ATOMIC_INC(&taskdata->td_untied_count);
344     KMP_DEBUG_USE_VAR(counter);
345     KA_TRACE(
346         20,
347         ("__kmp_push_task: T#%d untied_count (%d) incremented for task %p\n",
348          gtid, counter, taskdata));
349   }
350 
351   // The first check avoids building task_team thread data if serialized
352   if (UNLIKELY(taskdata->td_flags.task_serial)) {
353     KA_TRACE(20, ("__kmp_push_task: T#%d team serialized; returning "
354                   "TASK_NOT_PUSHED for task %p\n",
355                   gtid, taskdata));
356     return TASK_NOT_PUSHED;
357   }
358 
359   // Now that serialized tasks have returned, we can assume that we are not in
360   // immediate exec mode
361   KMP_DEBUG_ASSERT(__kmp_tasking_mode != tskm_immediate_exec);
362   if (UNLIKELY(!KMP_TASKING_ENABLED(task_team))) {
363     __kmp_enable_tasking(task_team, thread);
364   }
365   KMP_DEBUG_ASSERT(TCR_4(task_team->tt.tt_found_tasks) == TRUE);
366   KMP_DEBUG_ASSERT(TCR_PTR(task_team->tt.tt_threads_data) != NULL);
367 
368   // Find tasking deque specific to encountering thread
369   thread_data = &task_team->tt.tt_threads_data[tid];
370 
371   // No lock needed since only owner can allocate. If the task is hidden_helper,
372   // we don't need it either because we have initialized the dequeue for hidden
373   // helper thread data.
374   if (UNLIKELY(thread_data->td.td_deque == NULL)) {
375     __kmp_alloc_task_deque(thread, thread_data);
376   }
377 
378   int locked = 0;
379   // Check if deque is full
380   if (TCR_4(thread_data->td.td_deque_ntasks) >=
381       TASK_DEQUE_SIZE(thread_data->td)) {
382     if (__kmp_enable_task_throttling &&
383         __kmp_task_is_allowed(gtid, __kmp_task_stealing_constraint, taskdata,
384                               thread->th.th_current_task)) {
385       KA_TRACE(20, ("__kmp_push_task: T#%d deque is full; returning "
386                     "TASK_NOT_PUSHED for task %p\n",
387                     gtid, taskdata));
388       return TASK_NOT_PUSHED;
389     } else {
390       __kmp_acquire_bootstrap_lock(&thread_data->td.td_deque_lock);
391       locked = 1;
392       if (TCR_4(thread_data->td.td_deque_ntasks) >=
393           TASK_DEQUE_SIZE(thread_data->td)) {
394         // expand deque to push the task which is not allowed to execute
395         __kmp_realloc_task_deque(thread, thread_data);
396       }
397     }
398   }
399   // Lock the deque for the task push operation
400   if (!locked) {
401     __kmp_acquire_bootstrap_lock(&thread_data->td.td_deque_lock);
402     // Need to recheck as we can get a proxy task from thread outside of OpenMP
403     if (TCR_4(thread_data->td.td_deque_ntasks) >=
404         TASK_DEQUE_SIZE(thread_data->td)) {
405       if (__kmp_enable_task_throttling &&
406           __kmp_task_is_allowed(gtid, __kmp_task_stealing_constraint, taskdata,
407                                 thread->th.th_current_task)) {
408         __kmp_release_bootstrap_lock(&thread_data->td.td_deque_lock);
409         KA_TRACE(20, ("__kmp_push_task: T#%d deque is full on 2nd check; "
410                       "returning TASK_NOT_PUSHED for task %p\n",
411                       gtid, taskdata));
412         return TASK_NOT_PUSHED;
413       } else {
414         // expand deque to push the task which is not allowed to execute
415         __kmp_realloc_task_deque(thread, thread_data);
416       }
417     }
418   }
419   // Must have room since no thread can add tasks but calling thread
420   KMP_DEBUG_ASSERT(TCR_4(thread_data->td.td_deque_ntasks) <
421                    TASK_DEQUE_SIZE(thread_data->td));
422 
423   thread_data->td.td_deque[thread_data->td.td_deque_tail] =
424       taskdata; // Push taskdata
425   // Wrap index.
426   thread_data->td.td_deque_tail =
427       (thread_data->td.td_deque_tail + 1) & TASK_DEQUE_MASK(thread_data->td);
428   TCW_4(thread_data->td.td_deque_ntasks,
429         TCR_4(thread_data->td.td_deque_ntasks) + 1); // Adjust task count
430   KMP_FSYNC_RELEASING(thread->th.th_current_task); // releasing self
431   KMP_FSYNC_RELEASING(taskdata); // releasing child
432   KA_TRACE(20, ("__kmp_push_task: T#%d returning TASK_SUCCESSFULLY_PUSHED: "
433                 "task=%p ntasks=%d head=%u tail=%u\n",
434                 gtid, taskdata, thread_data->td.td_deque_ntasks,
435                 thread_data->td.td_deque_head, thread_data->td.td_deque_tail));
436 
437   auto hidden_helper = taskdata->td_flags.hidden_helper;
438 
439   __kmp_release_bootstrap_lock(&thread_data->td.td_deque_lock);
440 
441   // Signal one worker thread to execute the task
442   if (UNLIKELY(hidden_helper)) {
443     // Wake hidden helper threads up if they're sleeping
444     __kmp_hidden_helper_worker_thread_signal();
445   }
446 
447   return TASK_SUCCESSFULLY_PUSHED;
448 }
449 
450 // __kmp_pop_current_task_from_thread: set up current task from called thread
451 // when team ends
452 //
453 // this_thr: thread structure to set current_task in.
454 void __kmp_pop_current_task_from_thread(kmp_info_t *this_thr) {
455   KF_TRACE(10, ("__kmp_pop_current_task_from_thread(enter): T#%d "
456                 "this_thread=%p, curtask=%p, "
457                 "curtask_parent=%p\n",
458                 0, this_thr, this_thr->th.th_current_task,
459                 this_thr->th.th_current_task->td_parent));
460 
461   this_thr->th.th_current_task = this_thr->th.th_current_task->td_parent;
462 
463   KF_TRACE(10, ("__kmp_pop_current_task_from_thread(exit): T#%d "
464                 "this_thread=%p, curtask=%p, "
465                 "curtask_parent=%p\n",
466                 0, this_thr, this_thr->th.th_current_task,
467                 this_thr->th.th_current_task->td_parent));
468 }
469 
470 // __kmp_push_current_task_to_thread: set up current task in called thread for a
471 // new team
472 //
473 // this_thr: thread structure to set up
474 // team: team for implicit task data
475 // tid: thread within team to set up
476 void __kmp_push_current_task_to_thread(kmp_info_t *this_thr, kmp_team_t *team,
477                                        int tid) {
478   // current task of the thread is a parent of the new just created implicit
479   // tasks of new team
480   KF_TRACE(10, ("__kmp_push_current_task_to_thread(enter): T#%d this_thread=%p "
481                 "curtask=%p "
482                 "parent_task=%p\n",
483                 tid, this_thr, this_thr->th.th_current_task,
484                 team->t.t_implicit_task_taskdata[tid].td_parent));
485 
486   KMP_DEBUG_ASSERT(this_thr != NULL);
487 
488   if (tid == 0) {
489     if (this_thr->th.th_current_task != &team->t.t_implicit_task_taskdata[0]) {
490       team->t.t_implicit_task_taskdata[0].td_parent =
491           this_thr->th.th_current_task;
492       this_thr->th.th_current_task = &team->t.t_implicit_task_taskdata[0];
493     }
494   } else {
495     team->t.t_implicit_task_taskdata[tid].td_parent =
496         team->t.t_implicit_task_taskdata[0].td_parent;
497     this_thr->th.th_current_task = &team->t.t_implicit_task_taskdata[tid];
498   }
499 
500   KF_TRACE(10, ("__kmp_push_current_task_to_thread(exit): T#%d this_thread=%p "
501                 "curtask=%p "
502                 "parent_task=%p\n",
503                 tid, this_thr, this_thr->th.th_current_task,
504                 team->t.t_implicit_task_taskdata[tid].td_parent));
505 }
506 
507 // __kmp_task_start: bookkeeping for a task starting execution
508 //
509 // GTID: global thread id of calling thread
510 // task: task starting execution
511 // current_task: task suspending
512 static void __kmp_task_start(kmp_int32 gtid, kmp_task_t *task,
513                              kmp_taskdata_t *current_task) {
514   kmp_taskdata_t *taskdata = KMP_TASK_TO_TASKDATA(task);
515   kmp_info_t *thread = __kmp_threads[gtid];
516 
517   KA_TRACE(10,
518            ("__kmp_task_start(enter): T#%d starting task %p: current_task=%p\n",
519             gtid, taskdata, current_task));
520 
521   KMP_DEBUG_ASSERT(taskdata->td_flags.tasktype == TASK_EXPLICIT);
522 
523   // mark currently executing task as suspended
524   // TODO: GEH - make sure root team implicit task is initialized properly.
525   // KMP_DEBUG_ASSERT( current_task -> td_flags.executing == 1 );
526   current_task->td_flags.executing = 0;
527 
528 // Add task to stack if tied
529 #ifdef BUILD_TIED_TASK_STACK
530   if (taskdata->td_flags.tiedness == TASK_TIED) {
531     __kmp_push_task_stack(gtid, thread, taskdata);
532   }
533 #endif /* BUILD_TIED_TASK_STACK */
534 
535   // mark starting task as executing and as current task
536   thread->th.th_current_task = taskdata;
537 
538   KMP_DEBUG_ASSERT(taskdata->td_flags.started == 0 ||
539                    taskdata->td_flags.tiedness == TASK_UNTIED);
540   KMP_DEBUG_ASSERT(taskdata->td_flags.executing == 0 ||
541                    taskdata->td_flags.tiedness == TASK_UNTIED);
542   taskdata->td_flags.started = 1;
543   taskdata->td_flags.executing = 1;
544   KMP_DEBUG_ASSERT(taskdata->td_flags.complete == 0);
545   KMP_DEBUG_ASSERT(taskdata->td_flags.freed == 0);
546 
547   // GEH TODO: shouldn't we pass some sort of location identifier here?
548   // APT: yes, we will pass location here.
549   // need to store current thread state (in a thread or taskdata structure)
550   // before setting work_state, otherwise wrong state is set after end of task
551 
552   KA_TRACE(10, ("__kmp_task_start(exit): T#%d task=%p\n", gtid, taskdata));
553 
554   return;
555 }
556 
557 #if OMPT_SUPPORT
558 //------------------------------------------------------------------------------
559 // __ompt_task_init:
560 //   Initialize OMPT fields maintained by a task. This will only be called after
561 //   ompt_start_tool, so we already know whether ompt is enabled or not.
562 
563 static inline void __ompt_task_init(kmp_taskdata_t *task, int tid) {
564   // The calls to __ompt_task_init already have the ompt_enabled condition.
565   task->ompt_task_info.task_data.value = 0;
566   task->ompt_task_info.frame.exit_frame = ompt_data_none;
567   task->ompt_task_info.frame.enter_frame = ompt_data_none;
568   task->ompt_task_info.frame.exit_frame_flags =
569       ompt_frame_runtime | ompt_frame_framepointer;
570   task->ompt_task_info.frame.enter_frame_flags =
571       ompt_frame_runtime | ompt_frame_framepointer;
572 }
573 
574 // __ompt_task_start:
575 //   Build and trigger task-begin event
576 static inline void __ompt_task_start(kmp_task_t *task,
577                                      kmp_taskdata_t *current_task,
578                                      kmp_int32 gtid) {
579   kmp_taskdata_t *taskdata = KMP_TASK_TO_TASKDATA(task);
580   ompt_task_status_t status = ompt_task_switch;
581   if (__kmp_threads[gtid]->th.ompt_thread_info.ompt_task_yielded) {
582     status = ompt_task_yield;
583     __kmp_threads[gtid]->th.ompt_thread_info.ompt_task_yielded = 0;
584   }
585   /* let OMPT know that we're about to run this task */
586   if (ompt_enabled.ompt_callback_task_schedule) {
587     ompt_callbacks.ompt_callback(ompt_callback_task_schedule)(
588         &(current_task->ompt_task_info.task_data), status,
589         &(taskdata->ompt_task_info.task_data));
590   }
591   taskdata->ompt_task_info.scheduling_parent = current_task;
592 }
593 
594 // __ompt_task_finish:
595 //   Build and trigger final task-schedule event
596 static inline void __ompt_task_finish(kmp_task_t *task,
597                                       kmp_taskdata_t *resumed_task,
598                                       ompt_task_status_t status) {
599   if (ompt_enabled.ompt_callback_task_schedule) {
600     kmp_taskdata_t *taskdata = KMP_TASK_TO_TASKDATA(task);
601     if (__kmp_omp_cancellation && taskdata->td_taskgroup &&
602         taskdata->td_taskgroup->cancel_request == cancel_taskgroup) {
603       status = ompt_task_cancel;
604     }
605 
606     /* let OMPT know that we're returning to the callee task */
607     ompt_callbacks.ompt_callback(ompt_callback_task_schedule)(
608         &(taskdata->ompt_task_info.task_data), status,
609         (resumed_task ? &(resumed_task->ompt_task_info.task_data) : NULL));
610   }
611 }
612 #endif
613 
614 template <bool ompt>
615 static void __kmpc_omp_task_begin_if0_template(ident_t *loc_ref, kmp_int32 gtid,
616                                                kmp_task_t *task,
617                                                void *frame_address,
618                                                void *return_address) {
619   kmp_taskdata_t *taskdata = KMP_TASK_TO_TASKDATA(task);
620   kmp_taskdata_t *current_task = __kmp_threads[gtid]->th.th_current_task;
621 
622   KA_TRACE(10, ("__kmpc_omp_task_begin_if0(enter): T#%d loc=%p task=%p "
623                 "current_task=%p\n",
624                 gtid, loc_ref, taskdata, current_task));
625 
626   if (UNLIKELY(taskdata->td_flags.tiedness == TASK_UNTIED)) {
627     // untied task needs to increment counter so that the task structure is not
628     // freed prematurely
629     kmp_int32 counter = 1 + KMP_ATOMIC_INC(&taskdata->td_untied_count);
630     KMP_DEBUG_USE_VAR(counter);
631     KA_TRACE(20, ("__kmpc_omp_task_begin_if0: T#%d untied_count (%d) "
632                   "incremented for task %p\n",
633                   gtid, counter, taskdata));
634   }
635 
636   taskdata->td_flags.task_serial =
637       1; // Execute this task immediately, not deferred.
638   __kmp_task_start(gtid, task, current_task);
639 
640 #if OMPT_SUPPORT
641   if (ompt) {
642     if (current_task->ompt_task_info.frame.enter_frame.ptr == NULL) {
643       current_task->ompt_task_info.frame.enter_frame.ptr =
644           taskdata->ompt_task_info.frame.exit_frame.ptr = frame_address;
645       current_task->ompt_task_info.frame.enter_frame_flags =
646           taskdata->ompt_task_info.frame.exit_frame_flags =
647               ompt_frame_application | ompt_frame_framepointer;
648     }
649     if (ompt_enabled.ompt_callback_task_create) {
650       ompt_task_info_t *parent_info = &(current_task->ompt_task_info);
651       ompt_callbacks.ompt_callback(ompt_callback_task_create)(
652           &(parent_info->task_data), &(parent_info->frame),
653           &(taskdata->ompt_task_info.task_data),
654           ompt_task_explicit | TASK_TYPE_DETAILS_FORMAT(taskdata), 0,
655           return_address);
656     }
657     __ompt_task_start(task, current_task, gtid);
658   }
659 #endif // OMPT_SUPPORT
660 
661   KA_TRACE(10, ("__kmpc_omp_task_begin_if0(exit): T#%d loc=%p task=%p,\n", gtid,
662                 loc_ref, taskdata));
663 }
664 
665 #if OMPT_SUPPORT
666 OMPT_NOINLINE
667 static void __kmpc_omp_task_begin_if0_ompt(ident_t *loc_ref, kmp_int32 gtid,
668                                            kmp_task_t *task,
669                                            void *frame_address,
670                                            void *return_address) {
671   __kmpc_omp_task_begin_if0_template<true>(loc_ref, gtid, task, frame_address,
672                                            return_address);
673 }
674 #endif // OMPT_SUPPORT
675 
676 // __kmpc_omp_task_begin_if0: report that a given serialized task has started
677 // execution
678 //
679 // loc_ref: source location information; points to beginning of task block.
680 // gtid: global thread number.
681 // task: task thunk for the started task.
682 void __kmpc_omp_task_begin_if0(ident_t *loc_ref, kmp_int32 gtid,
683                                kmp_task_t *task) {
684 #if OMPT_SUPPORT
685   if (UNLIKELY(ompt_enabled.enabled)) {
686     OMPT_STORE_RETURN_ADDRESS(gtid);
687     __kmpc_omp_task_begin_if0_ompt(loc_ref, gtid, task,
688                                    OMPT_GET_FRAME_ADDRESS(1),
689                                    OMPT_LOAD_RETURN_ADDRESS(gtid));
690     return;
691   }
692 #endif
693   __kmpc_omp_task_begin_if0_template<false>(loc_ref, gtid, task, NULL, NULL);
694 }
695 
696 #ifdef TASK_UNUSED
697 // __kmpc_omp_task_begin: report that a given task has started execution
698 // NEVER GENERATED BY COMPILER, DEPRECATED!!!
699 void __kmpc_omp_task_begin(ident_t *loc_ref, kmp_int32 gtid, kmp_task_t *task) {
700   kmp_taskdata_t *current_task = __kmp_threads[gtid]->th.th_current_task;
701 
702   KA_TRACE(
703       10,
704       ("__kmpc_omp_task_begin(enter): T#%d loc=%p task=%p current_task=%p\n",
705        gtid, loc_ref, KMP_TASK_TO_TASKDATA(task), current_task));
706 
707   __kmp_task_start(gtid, task, current_task);
708 
709   KA_TRACE(10, ("__kmpc_omp_task_begin(exit): T#%d loc=%p task=%p,\n", gtid,
710                 loc_ref, KMP_TASK_TO_TASKDATA(task)));
711   return;
712 }
713 #endif // TASK_UNUSED
714 
715 // __kmp_free_task: free the current task space and the space for shareds
716 //
717 // gtid: Global thread ID of calling thread
718 // taskdata: task to free
719 // thread: thread data structure of caller
720 static void __kmp_free_task(kmp_int32 gtid, kmp_taskdata_t *taskdata,
721                             kmp_info_t *thread) {
722   KA_TRACE(30, ("__kmp_free_task: T#%d freeing data from task %p\n", gtid,
723                 taskdata));
724 
725   // Check to make sure all flags and counters have the correct values
726   KMP_DEBUG_ASSERT(taskdata->td_flags.tasktype == TASK_EXPLICIT);
727   KMP_DEBUG_ASSERT(taskdata->td_flags.executing == 0);
728   KMP_DEBUG_ASSERT(taskdata->td_flags.complete == 1);
729   KMP_DEBUG_ASSERT(taskdata->td_flags.freed == 0);
730   KMP_DEBUG_ASSERT(taskdata->td_allocated_child_tasks == 0 ||
731                    taskdata->td_flags.task_serial == 1);
732   KMP_DEBUG_ASSERT(taskdata->td_incomplete_child_tasks == 0);
733 
734   taskdata->td_flags.freed = 1;
735 // deallocate the taskdata and shared variable blocks associated with this task
736 #if USE_FAST_MEMORY
737   __kmp_fast_free(thread, taskdata);
738 #else /* ! USE_FAST_MEMORY */
739   __kmp_thread_free(thread, taskdata);
740 #endif
741   KA_TRACE(20, ("__kmp_free_task: T#%d freed task %p\n", gtid, taskdata));
742 }
743 
744 // __kmp_free_task_and_ancestors: free the current task and ancestors without
745 // children
746 //
747 // gtid: Global thread ID of calling thread
748 // taskdata: task to free
749 // thread: thread data structure of caller
750 static void __kmp_free_task_and_ancestors(kmp_int32 gtid,
751                                           kmp_taskdata_t *taskdata,
752                                           kmp_info_t *thread) {
753   // Proxy tasks must always be allowed to free their parents
754   // because they can be run in background even in serial mode.
755   kmp_int32 team_serial =
756       (taskdata->td_flags.team_serial || taskdata->td_flags.tasking_ser) &&
757       !taskdata->td_flags.proxy;
758   KMP_DEBUG_ASSERT(taskdata->td_flags.tasktype == TASK_EXPLICIT);
759 
760   kmp_int32 children = KMP_ATOMIC_DEC(&taskdata->td_allocated_child_tasks) - 1;
761   KMP_DEBUG_ASSERT(children >= 0);
762 
763   // Now, go up the ancestor tree to see if any ancestors can now be freed.
764   while (children == 0) {
765     kmp_taskdata_t *parent_taskdata = taskdata->td_parent;
766 
767     KA_TRACE(20, ("__kmp_free_task_and_ancestors(enter): T#%d task %p complete "
768                   "and freeing itself\n",
769                   gtid, taskdata));
770 
771     // --- Deallocate my ancestor task ---
772     __kmp_free_task(gtid, taskdata, thread);
773 
774     taskdata = parent_taskdata;
775 
776     if (team_serial)
777       return;
778     // Stop checking ancestors at implicit task instead of walking up ancestor
779     // tree to avoid premature deallocation of ancestors.
780     if (taskdata->td_flags.tasktype == TASK_IMPLICIT) {
781       if (taskdata->td_dephash) { // do we need to cleanup dephash?
782         int children = KMP_ATOMIC_LD_ACQ(&taskdata->td_incomplete_child_tasks);
783         kmp_tasking_flags_t flags_old = taskdata->td_flags;
784         if (children == 0 && flags_old.complete == 1) {
785           kmp_tasking_flags_t flags_new = flags_old;
786           flags_new.complete = 0;
787           if (KMP_COMPARE_AND_STORE_ACQ32(
788                   RCAST(kmp_int32 *, &taskdata->td_flags),
789                   *RCAST(kmp_int32 *, &flags_old),
790                   *RCAST(kmp_int32 *, &flags_new))) {
791             KA_TRACE(100, ("__kmp_free_task_and_ancestors: T#%d cleans "
792                            "dephash of implicit task %p\n",
793                            gtid, taskdata));
794             // cleanup dephash of finished implicit task
795             __kmp_dephash_free_entries(thread, taskdata->td_dephash);
796           }
797         }
798       }
799       return;
800     }
801     // Predecrement simulated by "- 1" calculation
802     children = KMP_ATOMIC_DEC(&taskdata->td_allocated_child_tasks) - 1;
803     KMP_DEBUG_ASSERT(children >= 0);
804   }
805 
806   KA_TRACE(
807       20, ("__kmp_free_task_and_ancestors(exit): T#%d task %p has %d children; "
808            "not freeing it yet\n",
809            gtid, taskdata, children));
810 }
811 
812 // __kmp_task_finish: bookkeeping to do when a task finishes execution
813 //
814 // gtid: global thread ID for calling thread
815 // task: task to be finished
816 // resumed_task: task to be resumed.  (may be NULL if task is serialized)
817 //
818 // template<ompt>: effectively ompt_enabled.enabled!=0
819 // the version with ompt=false is inlined, allowing to optimize away all ompt
820 // code in this case
821 template <bool ompt>
822 static void __kmp_task_finish(kmp_int32 gtid, kmp_task_t *task,
823                               kmp_taskdata_t *resumed_task) {
824   kmp_taskdata_t *taskdata = KMP_TASK_TO_TASKDATA(task);
825   kmp_info_t *thread = __kmp_threads[gtid];
826   kmp_task_team_t *task_team =
827       thread->th.th_task_team; // might be NULL for serial teams...
828   kmp_int32 children = 0;
829 
830   KA_TRACE(10, ("__kmp_task_finish(enter): T#%d finishing task %p and resuming "
831                 "task %p\n",
832                 gtid, taskdata, resumed_task));
833 
834   KMP_DEBUG_ASSERT(taskdata->td_flags.tasktype == TASK_EXPLICIT);
835 
836 // Pop task from stack if tied
837 #ifdef BUILD_TIED_TASK_STACK
838   if (taskdata->td_flags.tiedness == TASK_TIED) {
839     __kmp_pop_task_stack(gtid, thread, taskdata);
840   }
841 #endif /* BUILD_TIED_TASK_STACK */
842 
843   if (UNLIKELY(taskdata->td_flags.tiedness == TASK_UNTIED)) {
844     // untied task needs to check the counter so that the task structure is not
845     // freed prematurely
846     kmp_int32 counter = KMP_ATOMIC_DEC(&taskdata->td_untied_count) - 1;
847     KA_TRACE(
848         20,
849         ("__kmp_task_finish: T#%d untied_count (%d) decremented for task %p\n",
850          gtid, counter, taskdata));
851     if (counter > 0) {
852       // untied task is not done, to be continued possibly by other thread, do
853       // not free it now
854       if (resumed_task == NULL) {
855         KMP_DEBUG_ASSERT(taskdata->td_flags.task_serial);
856         resumed_task = taskdata->td_parent; // In a serialized task, the resumed
857         // task is the parent
858       }
859       thread->th.th_current_task = resumed_task; // restore current_task
860       resumed_task->td_flags.executing = 1; // resume previous task
861       KA_TRACE(10, ("__kmp_task_finish(exit): T#%d partially done task %p, "
862                     "resuming task %p\n",
863                     gtid, taskdata, resumed_task));
864       return;
865     }
866   }
867 
868   // bookkeeping for resuming task:
869   // GEH - note tasking_ser => task_serial
870   KMP_DEBUG_ASSERT(
871       (taskdata->td_flags.tasking_ser || taskdata->td_flags.task_serial) ==
872       taskdata->td_flags.task_serial);
873   if (taskdata->td_flags.task_serial) {
874     if (resumed_task == NULL) {
875       resumed_task = taskdata->td_parent; // In a serialized task, the resumed
876       // task is the parent
877     }
878   } else {
879     KMP_DEBUG_ASSERT(resumed_task !=
880                      NULL); // verify that resumed task is passed as argument
881   }
882 
883   /* If the tasks' destructor thunk flag has been set, we need to invoke the
884      destructor thunk that has been generated by the compiler. The code is
885      placed here, since at this point other tasks might have been released
886      hence overlapping the destructor invocations with some other work in the
887      released tasks.  The OpenMP spec is not specific on when the destructors
888      are invoked, so we should be free to choose. */
889   if (UNLIKELY(taskdata->td_flags.destructors_thunk)) {
890     kmp_routine_entry_t destr_thunk = task->data1.destructors;
891     KMP_ASSERT(destr_thunk);
892     destr_thunk(gtid, task);
893   }
894 
895   KMP_DEBUG_ASSERT(taskdata->td_flags.complete == 0);
896   KMP_DEBUG_ASSERT(taskdata->td_flags.started == 1);
897   KMP_DEBUG_ASSERT(taskdata->td_flags.freed == 0);
898 
899   bool detach = false;
900   if (UNLIKELY(taskdata->td_flags.detachable == TASK_DETACHABLE)) {
901     if (taskdata->td_allow_completion_event.type ==
902         KMP_EVENT_ALLOW_COMPLETION) {
903       // event hasn't been fulfilled yet. Try to detach task.
904       __kmp_acquire_tas_lock(&taskdata->td_allow_completion_event.lock, gtid);
905       if (taskdata->td_allow_completion_event.type ==
906           KMP_EVENT_ALLOW_COMPLETION) {
907         // task finished execution
908         KMP_DEBUG_ASSERT(taskdata->td_flags.executing == 1);
909         taskdata->td_flags.executing = 0; // suspend the finishing task
910 
911 #if OMPT_SUPPORT
912         // For a detached task, which is not completed, we switch back
913         // the omp_fulfill_event signals completion
914         // locking is necessary to avoid a race with ompt_task_late_fulfill
915         if (ompt)
916           __ompt_task_finish(task, resumed_task, ompt_task_detach);
917 #endif
918 
919         // no access to taskdata after this point!
920         // __kmp_fulfill_event might free taskdata at any time from now
921 
922         taskdata->td_flags.proxy = TASK_PROXY; // proxify!
923         detach = true;
924       }
925       __kmp_release_tas_lock(&taskdata->td_allow_completion_event.lock, gtid);
926     }
927   }
928 
929   if (!detach) {
930     taskdata->td_flags.complete = 1; // mark the task as completed
931 
932 #if OMPT_SUPPORT
933     // This is not a detached task, we are done here
934     if (ompt)
935       __ompt_task_finish(task, resumed_task, ompt_task_complete);
936 #endif
937 
938     // Only need to keep track of count if team parallel and tasking not
939     // serialized, or task is detachable and event has already been fulfilled
940     if (!(taskdata->td_flags.team_serial || taskdata->td_flags.tasking_ser) ||
941         taskdata->td_flags.detachable == TASK_DETACHABLE ||
942         taskdata->td_flags.hidden_helper) {
943       __kmp_release_deps(gtid, taskdata);
944       // Predecrement simulated by "- 1" calculation
945       children =
946           KMP_ATOMIC_DEC(&taskdata->td_parent->td_incomplete_child_tasks) - 1;
947       KMP_DEBUG_ASSERT(children >= 0);
948       if (taskdata->td_taskgroup)
949         KMP_ATOMIC_DEC(&taskdata->td_taskgroup->count);
950     } else if (task_team && (task_team->tt.tt_found_proxy_tasks ||
951                              task_team->tt.tt_hidden_helper_task_encountered)) {
952       // if we found proxy or hidden helper tasks there could exist a dependency
953       // chain with the proxy task as origin
954       __kmp_release_deps(gtid, taskdata);
955     }
956     // td_flags.executing must be marked as 0 after __kmp_release_deps has been
957     // called. Othertwise, if a task is executed immediately from the
958     // release_deps code, the flag will be reset to 1 again by this same
959     // function
960     KMP_DEBUG_ASSERT(taskdata->td_flags.executing == 1);
961     taskdata->td_flags.executing = 0; // suspend the finishing task
962   }
963 
964   KA_TRACE(
965       20, ("__kmp_task_finish: T#%d finished task %p, %d incomplete children\n",
966            gtid, taskdata, children));
967 
968   // Free this task and then ancestor tasks if they have no children.
969   // Restore th_current_task first as suggested by John:
970   // johnmc: if an asynchronous inquiry peers into the runtime system
971   // it doesn't see the freed task as the current task.
972   thread->th.th_current_task = resumed_task;
973   if (!detach)
974     __kmp_free_task_and_ancestors(gtid, taskdata, thread);
975 
976   // TODO: GEH - make sure root team implicit task is initialized properly.
977   // KMP_DEBUG_ASSERT( resumed_task->td_flags.executing == 0 );
978   resumed_task->td_flags.executing = 1; // resume previous task
979 
980   KA_TRACE(
981       10, ("__kmp_task_finish(exit): T#%d finished task %p, resuming task %p\n",
982            gtid, taskdata, resumed_task));
983 
984   return;
985 }
986 
987 template <bool ompt>
988 static void __kmpc_omp_task_complete_if0_template(ident_t *loc_ref,
989                                                   kmp_int32 gtid,
990                                                   kmp_task_t *task) {
991   KA_TRACE(10, ("__kmpc_omp_task_complete_if0(enter): T#%d loc=%p task=%p\n",
992                 gtid, loc_ref, KMP_TASK_TO_TASKDATA(task)));
993   KMP_DEBUG_ASSERT(gtid >= 0);
994   // this routine will provide task to resume
995   __kmp_task_finish<ompt>(gtid, task, NULL);
996 
997   KA_TRACE(10, ("__kmpc_omp_task_complete_if0(exit): T#%d loc=%p task=%p\n",
998                 gtid, loc_ref, KMP_TASK_TO_TASKDATA(task)));
999 
1000 #if OMPT_SUPPORT
1001   if (ompt) {
1002     ompt_frame_t *ompt_frame;
1003     __ompt_get_task_info_internal(0, NULL, NULL, &ompt_frame, NULL, NULL);
1004     ompt_frame->enter_frame = ompt_data_none;
1005     ompt_frame->enter_frame_flags =
1006         ompt_frame_runtime | ompt_frame_framepointer;
1007   }
1008 #endif
1009 
1010   return;
1011 }
1012 
1013 #if OMPT_SUPPORT
1014 OMPT_NOINLINE
1015 void __kmpc_omp_task_complete_if0_ompt(ident_t *loc_ref, kmp_int32 gtid,
1016                                        kmp_task_t *task) {
1017   __kmpc_omp_task_complete_if0_template<true>(loc_ref, gtid, task);
1018 }
1019 #endif // OMPT_SUPPORT
1020 
1021 // __kmpc_omp_task_complete_if0: report that a task has completed execution
1022 //
1023 // loc_ref: source location information; points to end of task block.
1024 // gtid: global thread number.
1025 // task: task thunk for the completed task.
1026 void __kmpc_omp_task_complete_if0(ident_t *loc_ref, kmp_int32 gtid,
1027                                   kmp_task_t *task) {
1028 #if OMPT_SUPPORT
1029   if (UNLIKELY(ompt_enabled.enabled)) {
1030     __kmpc_omp_task_complete_if0_ompt(loc_ref, gtid, task);
1031     return;
1032   }
1033 #endif
1034   __kmpc_omp_task_complete_if0_template<false>(loc_ref, gtid, task);
1035 }
1036 
1037 #ifdef TASK_UNUSED
1038 // __kmpc_omp_task_complete: report that a task has completed execution
1039 // NEVER GENERATED BY COMPILER, DEPRECATED!!!
1040 void __kmpc_omp_task_complete(ident_t *loc_ref, kmp_int32 gtid,
1041                               kmp_task_t *task) {
1042   KA_TRACE(10, ("__kmpc_omp_task_complete(enter): T#%d loc=%p task=%p\n", gtid,
1043                 loc_ref, KMP_TASK_TO_TASKDATA(task)));
1044 
1045   __kmp_task_finish<false>(gtid, task,
1046                            NULL); // Not sure how to find task to resume
1047 
1048   KA_TRACE(10, ("__kmpc_omp_task_complete(exit): T#%d loc=%p task=%p\n", gtid,
1049                 loc_ref, KMP_TASK_TO_TASKDATA(task)));
1050   return;
1051 }
1052 #endif // TASK_UNUSED
1053 
1054 // __kmp_init_implicit_task: Initialize the appropriate fields in the implicit
1055 // task for a given thread
1056 //
1057 // loc_ref:  reference to source location of parallel region
1058 // this_thr:  thread data structure corresponding to implicit task
1059 // team: team for this_thr
1060 // tid: thread id of given thread within team
1061 // set_curr_task: TRUE if need to push current task to thread
1062 // NOTE: Routine does not set up the implicit task ICVS.  This is assumed to
1063 // have already been done elsewhere.
1064 // TODO: Get better loc_ref.  Value passed in may be NULL
1065 void __kmp_init_implicit_task(ident_t *loc_ref, kmp_info_t *this_thr,
1066                               kmp_team_t *team, int tid, int set_curr_task) {
1067   kmp_taskdata_t *task = &team->t.t_implicit_task_taskdata[tid];
1068 
1069   KF_TRACE(
1070       10,
1071       ("__kmp_init_implicit_task(enter): T#:%d team=%p task=%p, reinit=%s\n",
1072        tid, team, task, set_curr_task ? "TRUE" : "FALSE"));
1073 
1074   task->td_task_id = KMP_GEN_TASK_ID();
1075   task->td_team = team;
1076   //    task->td_parent   = NULL;  // fix for CQ230101 (broken parent task info
1077   //    in debugger)
1078   task->td_ident = loc_ref;
1079   task->td_taskwait_ident = NULL;
1080   task->td_taskwait_counter = 0;
1081   task->td_taskwait_thread = 0;
1082 
1083   task->td_flags.tiedness = TASK_TIED;
1084   task->td_flags.tasktype = TASK_IMPLICIT;
1085   task->td_flags.proxy = TASK_FULL;
1086 
1087   // All implicit tasks are executed immediately, not deferred
1088   task->td_flags.task_serial = 1;
1089   task->td_flags.tasking_ser = (__kmp_tasking_mode == tskm_immediate_exec);
1090   task->td_flags.team_serial = (team->t.t_serialized) ? 1 : 0;
1091 
1092   task->td_flags.started = 1;
1093   task->td_flags.executing = 1;
1094   task->td_flags.complete = 0;
1095   task->td_flags.freed = 0;
1096 
1097   task->td_depnode = NULL;
1098   task->td_last_tied = task;
1099   task->td_allow_completion_event.type = KMP_EVENT_UNINITIALIZED;
1100 
1101   if (set_curr_task) { // only do this init first time thread is created
1102     KMP_ATOMIC_ST_REL(&task->td_incomplete_child_tasks, 0);
1103     // Not used: don't need to deallocate implicit task
1104     KMP_ATOMIC_ST_REL(&task->td_allocated_child_tasks, 0);
1105     task->td_taskgroup = NULL; // An implicit task does not have taskgroup
1106     task->td_dephash = NULL;
1107     __kmp_push_current_task_to_thread(this_thr, team, tid);
1108   } else {
1109     KMP_DEBUG_ASSERT(task->td_incomplete_child_tasks == 0);
1110     KMP_DEBUG_ASSERT(task->td_allocated_child_tasks == 0);
1111   }
1112 
1113 #if OMPT_SUPPORT
1114   if (UNLIKELY(ompt_enabled.enabled))
1115     __ompt_task_init(task, tid);
1116 #endif
1117 
1118   KF_TRACE(10, ("__kmp_init_implicit_task(exit): T#:%d team=%p task=%p\n", tid,
1119                 team, task));
1120 }
1121 
1122 // __kmp_finish_implicit_task: Release resources associated to implicit tasks
1123 // at the end of parallel regions. Some resources are kept for reuse in the next
1124 // parallel region.
1125 //
1126 // thread:  thread data structure corresponding to implicit task
1127 void __kmp_finish_implicit_task(kmp_info_t *thread) {
1128   kmp_taskdata_t *task = thread->th.th_current_task;
1129   if (task->td_dephash) {
1130     int children;
1131     task->td_flags.complete = 1;
1132     children = KMP_ATOMIC_LD_ACQ(&task->td_incomplete_child_tasks);
1133     kmp_tasking_flags_t flags_old = task->td_flags;
1134     if (children == 0 && flags_old.complete == 1) {
1135       kmp_tasking_flags_t flags_new = flags_old;
1136       flags_new.complete = 0;
1137       if (KMP_COMPARE_AND_STORE_ACQ32(RCAST(kmp_int32 *, &task->td_flags),
1138                                       *RCAST(kmp_int32 *, &flags_old),
1139                                       *RCAST(kmp_int32 *, &flags_new))) {
1140         KA_TRACE(100, ("__kmp_finish_implicit_task: T#%d cleans "
1141                        "dephash of implicit task %p\n",
1142                        thread->th.th_info.ds.ds_gtid, task));
1143         __kmp_dephash_free_entries(thread, task->td_dephash);
1144       }
1145     }
1146   }
1147 }
1148 
1149 // __kmp_free_implicit_task: Release resources associated to implicit tasks
1150 // when these are destroyed regions
1151 //
1152 // thread:  thread data structure corresponding to implicit task
1153 void __kmp_free_implicit_task(kmp_info_t *thread) {
1154   kmp_taskdata_t *task = thread->th.th_current_task;
1155   if (task && task->td_dephash) {
1156     __kmp_dephash_free(thread, task->td_dephash);
1157     task->td_dephash = NULL;
1158   }
1159 }
1160 
1161 // Round up a size to a power of two specified by val: Used to insert padding
1162 // between structures co-allocated using a single malloc() call
1163 static size_t __kmp_round_up_to_val(size_t size, size_t val) {
1164   if (size & (val - 1)) {
1165     size &= ~(val - 1);
1166     if (size <= KMP_SIZE_T_MAX - val) {
1167       size += val; // Round up if there is no overflow.
1168     }
1169   }
1170   return size;
1171 } // __kmp_round_up_to_va
1172 
1173 // __kmp_task_alloc: Allocate the taskdata and task data structures for a task
1174 //
1175 // loc_ref: source location information
1176 // gtid: global thread number.
1177 // flags: include tiedness & task type (explicit vs. implicit) of the ''new''
1178 // task encountered. Converted from kmp_int32 to kmp_tasking_flags_t in routine.
1179 // sizeof_kmp_task_t:  Size in bytes of kmp_task_t data structure including
1180 // private vars accessed in task.
1181 // sizeof_shareds:  Size in bytes of array of pointers to shared vars accessed
1182 // in task.
1183 // task_entry: Pointer to task code entry point generated by compiler.
1184 // returns: a pointer to the allocated kmp_task_t structure (task).
1185 kmp_task_t *__kmp_task_alloc(ident_t *loc_ref, kmp_int32 gtid,
1186                              kmp_tasking_flags_t *flags,
1187                              size_t sizeof_kmp_task_t, size_t sizeof_shareds,
1188                              kmp_routine_entry_t task_entry) {
1189   kmp_task_t *task;
1190   kmp_taskdata_t *taskdata;
1191   kmp_info_t *thread = __kmp_threads[gtid];
1192   kmp_info_t *encountering_thread = thread;
1193   kmp_team_t *team = thread->th.th_team;
1194   kmp_taskdata_t *parent_task = thread->th.th_current_task;
1195   size_t shareds_offset;
1196 
1197   if (UNLIKELY(!TCR_4(__kmp_init_middle)))
1198     __kmp_middle_initialize();
1199 
1200   if (flags->hidden_helper) {
1201     if (__kmp_enable_hidden_helper) {
1202       if (!TCR_4(__kmp_init_hidden_helper))
1203         __kmp_hidden_helper_initialize();
1204 
1205       // For a hidden helper task encountered by a regular thread, we will push
1206       // the task to the (gtid%__kmp_hidden_helper_threads_num)-th hidden helper
1207       // thread.
1208       if (!KMP_HIDDEN_HELPER_THREAD(gtid)) {
1209         thread = __kmp_threads[KMP_GTID_TO_SHADOW_GTID(gtid)];
1210         // We don't change the parent-child relation for hidden helper task as
1211         // we need that to do per-task-region synchronization.
1212       }
1213     } else {
1214       // If the hidden helper task is not enabled, reset the flag to FALSE.
1215       flags->hidden_helper = FALSE;
1216     }
1217   }
1218 
1219   KA_TRACE(10, ("__kmp_task_alloc(enter): T#%d loc=%p, flags=(0x%x) "
1220                 "sizeof_task=%ld sizeof_shared=%ld entry=%p\n",
1221                 gtid, loc_ref, *((kmp_int32 *)flags), sizeof_kmp_task_t,
1222                 sizeof_shareds, task_entry));
1223 
1224   KMP_DEBUG_ASSERT(parent_task);
1225   if (parent_task->td_flags.final) {
1226     if (flags->merged_if0) {
1227     }
1228     flags->final = 1;
1229   }
1230 
1231   if (flags->tiedness == TASK_UNTIED && !team->t.t_serialized) {
1232     // Untied task encountered causes the TSC algorithm to check entire deque of
1233     // the victim thread. If no untied task encountered, then checking the head
1234     // of the deque should be enough.
1235     KMP_CHECK_UPDATE(
1236         encountering_thread->th.th_task_team->tt.tt_untied_task_encountered, 1);
1237   }
1238 
1239   // Detachable tasks are not proxy tasks yet but could be in the future. Doing
1240   // the tasking setup
1241   // when that happens is too late.
1242   if (UNLIKELY(flags->proxy == TASK_PROXY ||
1243                flags->detachable == TASK_DETACHABLE || flags->hidden_helper)) {
1244     if (flags->proxy == TASK_PROXY) {
1245       flags->tiedness = TASK_UNTIED;
1246       flags->merged_if0 = 1;
1247     }
1248     /* are we running in a sequential parallel or tskm_immediate_exec... we need
1249        tasking support enabled */
1250     if ((encountering_thread->th.th_task_team) == NULL) {
1251       /* This should only happen if the team is serialized
1252           setup a task team and propagate it to the thread */
1253       KMP_DEBUG_ASSERT(team->t.t_serialized);
1254       KA_TRACE(30,
1255                ("T#%d creating task team in __kmp_task_alloc for proxy task\n",
1256                 gtid));
1257       __kmp_task_team_setup(
1258           encountering_thread, team,
1259           1); // 1 indicates setup the current team regardless of nthreads
1260       encountering_thread->th.th_task_team =
1261           team->t.t_task_team[encountering_thread->th.th_task_state];
1262     }
1263     kmp_task_team_t *task_team = encountering_thread->th.th_task_team;
1264 
1265     /* tasking must be enabled now as the task might not be pushed */
1266     if (!KMP_TASKING_ENABLED(task_team)) {
1267       KA_TRACE(
1268           30,
1269           ("T#%d enabling tasking in __kmp_task_alloc for proxy task\n", gtid));
1270       __kmp_enable_tasking(task_team, encountering_thread);
1271       kmp_int32 tid = encountering_thread->th.th_info.ds.ds_tid;
1272       kmp_thread_data_t *thread_data = &task_team->tt.tt_threads_data[tid];
1273       // No lock needed since only owner can allocate
1274       if (thread_data->td.td_deque == NULL) {
1275         __kmp_alloc_task_deque(encountering_thread, thread_data);
1276       }
1277     }
1278 
1279     if ((flags->proxy == TASK_PROXY || flags->detachable == TASK_DETACHABLE) &&
1280         task_team->tt.tt_found_proxy_tasks == FALSE)
1281       TCW_4(task_team->tt.tt_found_proxy_tasks, TRUE);
1282     if (flags->hidden_helper &&
1283         task_team->tt.tt_hidden_helper_task_encountered == FALSE)
1284       TCW_4(task_team->tt.tt_hidden_helper_task_encountered, TRUE);
1285   }
1286 
1287   // Calculate shared structure offset including padding after kmp_task_t struct
1288   // to align pointers in shared struct
1289   shareds_offset = sizeof(kmp_taskdata_t) + sizeof_kmp_task_t;
1290   shareds_offset = __kmp_round_up_to_val(shareds_offset, sizeof(void *));
1291 
1292   // Allocate a kmp_taskdata_t block and a kmp_task_t block.
1293   KA_TRACE(30, ("__kmp_task_alloc: T#%d First malloc size: %ld\n", gtid,
1294                 shareds_offset));
1295   KA_TRACE(30, ("__kmp_task_alloc: T#%d Second malloc size: %ld\n", gtid,
1296                 sizeof_shareds));
1297 
1298   // Avoid double allocation here by combining shareds with taskdata
1299 #if USE_FAST_MEMORY
1300   taskdata = (kmp_taskdata_t *)__kmp_fast_allocate(
1301       encountering_thread, shareds_offset + sizeof_shareds);
1302 #else /* ! USE_FAST_MEMORY */
1303   taskdata = (kmp_taskdata_t *)__kmp_thread_malloc(
1304       encountering_thread, shareds_offset + sizeof_shareds);
1305 #endif /* USE_FAST_MEMORY */
1306 
1307   task = KMP_TASKDATA_TO_TASK(taskdata);
1308 
1309 // Make sure task & taskdata are aligned appropriately
1310 #if KMP_ARCH_X86 || KMP_ARCH_PPC64 || !KMP_HAVE_QUAD
1311   KMP_DEBUG_ASSERT((((kmp_uintptr_t)taskdata) & (sizeof(double) - 1)) == 0);
1312   KMP_DEBUG_ASSERT((((kmp_uintptr_t)task) & (sizeof(double) - 1)) == 0);
1313 #else
1314   KMP_DEBUG_ASSERT((((kmp_uintptr_t)taskdata) & (sizeof(_Quad) - 1)) == 0);
1315   KMP_DEBUG_ASSERT((((kmp_uintptr_t)task) & (sizeof(_Quad) - 1)) == 0);
1316 #endif
1317   if (sizeof_shareds > 0) {
1318     // Avoid double allocation here by combining shareds with taskdata
1319     task->shareds = &((char *)taskdata)[shareds_offset];
1320     // Make sure shareds struct is aligned to pointer size
1321     KMP_DEBUG_ASSERT((((kmp_uintptr_t)task->shareds) & (sizeof(void *) - 1)) ==
1322                      0);
1323   } else {
1324     task->shareds = NULL;
1325   }
1326   task->routine = task_entry;
1327   task->part_id = 0; // AC: Always start with 0 part id
1328 
1329   taskdata->td_task_id = KMP_GEN_TASK_ID();
1330   taskdata->td_team = thread->th.th_team;
1331   taskdata->td_alloc_thread = encountering_thread;
1332   taskdata->td_parent = parent_task;
1333   taskdata->td_level = parent_task->td_level + 1; // increment nesting level
1334   KMP_ATOMIC_ST_RLX(&taskdata->td_untied_count, 0);
1335   taskdata->td_ident = loc_ref;
1336   taskdata->td_taskwait_ident = NULL;
1337   taskdata->td_taskwait_counter = 0;
1338   taskdata->td_taskwait_thread = 0;
1339   KMP_DEBUG_ASSERT(taskdata->td_parent != NULL);
1340   // avoid copying icvs for proxy tasks
1341   if (flags->proxy == TASK_FULL)
1342     copy_icvs(&taskdata->td_icvs, &taskdata->td_parent->td_icvs);
1343 
1344   taskdata->td_flags = *flags;
1345   taskdata->encountering_gtid = gtid;
1346   taskdata->td_task_team = thread->th.th_task_team;
1347   taskdata->td_size_alloc = shareds_offset + sizeof_shareds;
1348   taskdata->td_flags.tasktype = TASK_EXPLICIT;
1349 
1350   // GEH - TODO: fix this to copy parent task's value of tasking_ser flag
1351   taskdata->td_flags.tasking_ser = (__kmp_tasking_mode == tskm_immediate_exec);
1352 
1353   // GEH - TODO: fix this to copy parent task's value of team_serial flag
1354   taskdata->td_flags.team_serial = (team->t.t_serialized) ? 1 : 0;
1355 
1356   // GEH - Note we serialize the task if the team is serialized to make sure
1357   // implicit parallel region tasks are not left until program termination to
1358   // execute. Also, it helps locality to execute immediately.
1359 
1360   taskdata->td_flags.task_serial =
1361       (parent_task->td_flags.final || taskdata->td_flags.team_serial ||
1362        taskdata->td_flags.tasking_ser || flags->merged_if0);
1363 
1364   taskdata->td_flags.started = 0;
1365   taskdata->td_flags.executing = 0;
1366   taskdata->td_flags.complete = 0;
1367   taskdata->td_flags.freed = 0;
1368 
1369   KMP_ATOMIC_ST_RLX(&taskdata->td_incomplete_child_tasks, 0);
1370   // start at one because counts current task and children
1371   KMP_ATOMIC_ST_RLX(&taskdata->td_allocated_child_tasks, 1);
1372   taskdata->td_taskgroup =
1373       parent_task->td_taskgroup; // task inherits taskgroup from the parent task
1374   taskdata->td_dephash = NULL;
1375   taskdata->td_depnode = NULL;
1376   if (flags->tiedness == TASK_UNTIED)
1377     taskdata->td_last_tied = NULL; // will be set when the task is scheduled
1378   else
1379     taskdata->td_last_tied = taskdata;
1380   taskdata->td_allow_completion_event.type = KMP_EVENT_UNINITIALIZED;
1381 #if OMPT_SUPPORT
1382   if (UNLIKELY(ompt_enabled.enabled))
1383     __ompt_task_init(taskdata, gtid);
1384 #endif
1385   // Only need to keep track of child task counts if team parallel and tasking
1386   // not serialized or if it is a proxy or detachable or hidden helper task
1387   if (flags->proxy == TASK_PROXY || flags->detachable == TASK_DETACHABLE ||
1388       flags->hidden_helper ||
1389       !(taskdata->td_flags.team_serial || taskdata->td_flags.tasking_ser)) {
1390     KMP_ATOMIC_INC(&parent_task->td_incomplete_child_tasks);
1391     if (parent_task->td_taskgroup)
1392       KMP_ATOMIC_INC(&parent_task->td_taskgroup->count);
1393     // Only need to keep track of allocated child tasks for explicit tasks since
1394     // implicit not deallocated
1395     if (taskdata->td_parent->td_flags.tasktype == TASK_EXPLICIT) {
1396       KMP_ATOMIC_INC(&taskdata->td_parent->td_allocated_child_tasks);
1397     }
1398     if (flags->hidden_helper) {
1399       taskdata->td_flags.task_serial = FALSE;
1400       // Increment the number of hidden helper tasks to be executed
1401       KMP_ATOMIC_INC(&__kmp_unexecuted_hidden_helper_tasks);
1402     }
1403   }
1404 
1405   KA_TRACE(20, ("__kmp_task_alloc(exit): T#%d created task %p parent=%p\n",
1406                 gtid, taskdata, taskdata->td_parent));
1407 
1408   return task;
1409 }
1410 
1411 kmp_task_t *__kmpc_omp_task_alloc(ident_t *loc_ref, kmp_int32 gtid,
1412                                   kmp_int32 flags, size_t sizeof_kmp_task_t,
1413                                   size_t sizeof_shareds,
1414                                   kmp_routine_entry_t task_entry) {
1415   kmp_task_t *retval;
1416   kmp_tasking_flags_t *input_flags = (kmp_tasking_flags_t *)&flags;
1417   __kmp_assert_valid_gtid(gtid);
1418   input_flags->native = FALSE;
1419   // __kmp_task_alloc() sets up all other runtime flags
1420   KA_TRACE(10, ("__kmpc_omp_task_alloc(enter): T#%d loc=%p, flags=(%s %s %s) "
1421                 "sizeof_task=%ld sizeof_shared=%ld entry=%p\n",
1422                 gtid, loc_ref, input_flags->tiedness ? "tied  " : "untied",
1423                 input_flags->proxy ? "proxy" : "",
1424                 input_flags->detachable ? "detachable" : "", sizeof_kmp_task_t,
1425                 sizeof_shareds, task_entry));
1426 
1427   retval = __kmp_task_alloc(loc_ref, gtid, input_flags, sizeof_kmp_task_t,
1428                             sizeof_shareds, task_entry);
1429 
1430   KA_TRACE(20, ("__kmpc_omp_task_alloc(exit): T#%d retval %p\n", gtid, retval));
1431 
1432   return retval;
1433 }
1434 
1435 kmp_task_t *__kmpc_omp_target_task_alloc(ident_t *loc_ref, kmp_int32 gtid,
1436                                          kmp_int32 flags,
1437                                          size_t sizeof_kmp_task_t,
1438                                          size_t sizeof_shareds,
1439                                          kmp_routine_entry_t task_entry,
1440                                          kmp_int64 device_id) {
1441   if (__kmp_enable_hidden_helper) {
1442     auto &input_flags = reinterpret_cast<kmp_tasking_flags_t &>(flags);
1443     input_flags.hidden_helper = TRUE;
1444     input_flags.tiedness = TASK_UNTIED;
1445   }
1446 
1447   return __kmpc_omp_task_alloc(loc_ref, gtid, flags, sizeof_kmp_task_t,
1448                                sizeof_shareds, task_entry);
1449 }
1450 
1451 /*!
1452 @ingroup TASKING
1453 @param loc_ref location of the original task directive
1454 @param gtid Global Thread ID of encountering thread
1455 @param new_task task thunk allocated by __kmpc_omp_task_alloc() for the ''new
1456 task''
1457 @param naffins Number of affinity items
1458 @param affin_list List of affinity items
1459 @return Returns non-zero if registering affinity information was not successful.
1460  Returns 0 if registration was successful
1461 This entry registers the affinity information attached to a task with the task
1462 thunk structure kmp_taskdata_t.
1463 */
1464 kmp_int32
1465 __kmpc_omp_reg_task_with_affinity(ident_t *loc_ref, kmp_int32 gtid,
1466                                   kmp_task_t *new_task, kmp_int32 naffins,
1467                                   kmp_task_affinity_info_t *affin_list) {
1468   return 0;
1469 }
1470 
1471 //  __kmp_invoke_task: invoke the specified task
1472 //
1473 // gtid: global thread ID of caller
1474 // task: the task to invoke
1475 // current_task: the task to resume after task invocation
1476 static void __kmp_invoke_task(kmp_int32 gtid, kmp_task_t *task,
1477                               kmp_taskdata_t *current_task) {
1478   kmp_taskdata_t *taskdata = KMP_TASK_TO_TASKDATA(task);
1479   kmp_info_t *thread;
1480   int discard = 0 /* false */;
1481   KA_TRACE(
1482       30, ("__kmp_invoke_task(enter): T#%d invoking task %p, current_task=%p\n",
1483            gtid, taskdata, current_task));
1484   KMP_DEBUG_ASSERT(task);
1485   if (UNLIKELY(taskdata->td_flags.proxy == TASK_PROXY &&
1486                taskdata->td_flags.complete == 1)) {
1487     // This is a proxy task that was already completed but it needs to run
1488     // its bottom-half finish
1489     KA_TRACE(
1490         30,
1491         ("__kmp_invoke_task: T#%d running bottom finish for proxy task %p\n",
1492          gtid, taskdata));
1493 
1494     __kmp_bottom_half_finish_proxy(gtid, task);
1495 
1496     KA_TRACE(30, ("__kmp_invoke_task(exit): T#%d completed bottom finish for "
1497                   "proxy task %p, resuming task %p\n",
1498                   gtid, taskdata, current_task));
1499 
1500     return;
1501   }
1502 
1503 #if OMPT_SUPPORT
1504   // For untied tasks, the first task executed only calls __kmpc_omp_task and
1505   // does not execute code.
1506   ompt_thread_info_t oldInfo;
1507   if (UNLIKELY(ompt_enabled.enabled)) {
1508     // Store the threads states and restore them after the task
1509     thread = __kmp_threads[gtid];
1510     oldInfo = thread->th.ompt_thread_info;
1511     thread->th.ompt_thread_info.wait_id = 0;
1512     thread->th.ompt_thread_info.state = (thread->th.th_team_serialized)
1513                                             ? ompt_state_work_serial
1514                                             : ompt_state_work_parallel;
1515     taskdata->ompt_task_info.frame.exit_frame.ptr = OMPT_GET_FRAME_ADDRESS(0);
1516   }
1517 #endif
1518 
1519   // Decreament the counter of hidden helper tasks to be executed
1520   if (taskdata->td_flags.hidden_helper) {
1521     // Hidden helper tasks can only be executed by hidden helper threads
1522     KMP_ASSERT(KMP_HIDDEN_HELPER_THREAD(gtid));
1523     KMP_ATOMIC_DEC(&__kmp_unexecuted_hidden_helper_tasks);
1524   }
1525 
1526   // Proxy tasks are not handled by the runtime
1527   if (taskdata->td_flags.proxy != TASK_PROXY) {
1528     __kmp_task_start(gtid, task, current_task); // OMPT only if not discarded
1529   }
1530 
1531   // TODO: cancel tasks if the parallel region has also been cancelled
1532   // TODO: check if this sequence can be hoisted above __kmp_task_start
1533   // if cancellation has been enabled for this run ...
1534   if (UNLIKELY(__kmp_omp_cancellation)) {
1535     thread = __kmp_threads[gtid];
1536     kmp_team_t *this_team = thread->th.th_team;
1537     kmp_taskgroup_t *taskgroup = taskdata->td_taskgroup;
1538     if ((taskgroup && taskgroup->cancel_request) ||
1539         (this_team->t.t_cancel_request == cancel_parallel)) {
1540 #if OMPT_SUPPORT && OMPT_OPTIONAL
1541       ompt_data_t *task_data;
1542       if (UNLIKELY(ompt_enabled.ompt_callback_cancel)) {
1543         __ompt_get_task_info_internal(0, NULL, &task_data, NULL, NULL, NULL);
1544         ompt_callbacks.ompt_callback(ompt_callback_cancel)(
1545             task_data,
1546             ((taskgroup && taskgroup->cancel_request) ? ompt_cancel_taskgroup
1547                                                       : ompt_cancel_parallel) |
1548                 ompt_cancel_discarded_task,
1549             NULL);
1550       }
1551 #endif
1552       KMP_COUNT_BLOCK(TASK_cancelled);
1553       // this task belongs to a task group and we need to cancel it
1554       discard = 1 /* true */;
1555     }
1556   }
1557 
1558   // Invoke the task routine and pass in relevant data.
1559   // Thunks generated by gcc take a different argument list.
1560   if (!discard) {
1561     if (taskdata->td_flags.tiedness == TASK_UNTIED) {
1562       taskdata->td_last_tied = current_task->td_last_tied;
1563       KMP_DEBUG_ASSERT(taskdata->td_last_tied);
1564     }
1565 #if KMP_STATS_ENABLED
1566     KMP_COUNT_BLOCK(TASK_executed);
1567     switch (KMP_GET_THREAD_STATE()) {
1568     case FORK_JOIN_BARRIER:
1569       KMP_PUSH_PARTITIONED_TIMER(OMP_task_join_bar);
1570       break;
1571     case PLAIN_BARRIER:
1572       KMP_PUSH_PARTITIONED_TIMER(OMP_task_plain_bar);
1573       break;
1574     case TASKYIELD:
1575       KMP_PUSH_PARTITIONED_TIMER(OMP_task_taskyield);
1576       break;
1577     case TASKWAIT:
1578       KMP_PUSH_PARTITIONED_TIMER(OMP_task_taskwait);
1579       break;
1580     case TASKGROUP:
1581       KMP_PUSH_PARTITIONED_TIMER(OMP_task_taskgroup);
1582       break;
1583     default:
1584       KMP_PUSH_PARTITIONED_TIMER(OMP_task_immediate);
1585       break;
1586     }
1587 #endif // KMP_STATS_ENABLED
1588 
1589 // OMPT task begin
1590 #if OMPT_SUPPORT
1591     if (UNLIKELY(ompt_enabled.enabled))
1592       __ompt_task_start(task, current_task, gtid);
1593 #endif
1594 
1595 #if OMPD_SUPPORT
1596     if (ompd_state & OMPD_ENABLE_BP)
1597       ompd_bp_task_begin();
1598 #endif
1599 
1600 #if USE_ITT_BUILD && USE_ITT_NOTIFY
1601     kmp_uint64 cur_time;
1602     kmp_int32 kmp_itt_count_task =
1603         __kmp_forkjoin_frames_mode == 3 && !taskdata->td_flags.task_serial &&
1604         current_task->td_flags.tasktype == TASK_IMPLICIT;
1605     if (kmp_itt_count_task) {
1606       thread = __kmp_threads[gtid];
1607       // Time outer level explicit task on barrier for adjusting imbalance time
1608       if (thread->th.th_bar_arrive_time)
1609         cur_time = __itt_get_timestamp();
1610       else
1611         kmp_itt_count_task = 0; // thread is not on a barrier - skip timing
1612     }
1613     KMP_FSYNC_ACQUIRED(taskdata); // acquired self (new task)
1614 #endif
1615 
1616 #ifdef KMP_GOMP_COMPAT
1617     if (taskdata->td_flags.native) {
1618       ((void (*)(void *))(*(task->routine)))(task->shareds);
1619     } else
1620 #endif /* KMP_GOMP_COMPAT */
1621     {
1622       (*(task->routine))(gtid, task);
1623     }
1624     KMP_POP_PARTITIONED_TIMER();
1625 
1626 #if USE_ITT_BUILD && USE_ITT_NOTIFY
1627     if (kmp_itt_count_task) {
1628       // Barrier imbalance - adjust arrive time with the task duration
1629       thread->th.th_bar_arrive_time += (__itt_get_timestamp() - cur_time);
1630     }
1631     KMP_FSYNC_CANCEL(taskdata); // destroy self (just executed)
1632     KMP_FSYNC_RELEASING(taskdata->td_parent); // releasing parent
1633 #endif
1634   }
1635 
1636 #if OMPD_SUPPORT
1637   if (ompd_state & OMPD_ENABLE_BP)
1638     ompd_bp_task_end();
1639 #endif
1640 
1641   // Proxy tasks are not handled by the runtime
1642   if (taskdata->td_flags.proxy != TASK_PROXY) {
1643 #if OMPT_SUPPORT
1644     if (UNLIKELY(ompt_enabled.enabled)) {
1645       thread->th.ompt_thread_info = oldInfo;
1646       if (taskdata->td_flags.tiedness == TASK_TIED) {
1647         taskdata->ompt_task_info.frame.exit_frame = ompt_data_none;
1648       }
1649       __kmp_task_finish<true>(gtid, task, current_task);
1650     } else
1651 #endif
1652       __kmp_task_finish<false>(gtid, task, current_task);
1653   }
1654 
1655   KA_TRACE(
1656       30,
1657       ("__kmp_invoke_task(exit): T#%d completed task %p, resuming task %p\n",
1658        gtid, taskdata, current_task));
1659   return;
1660 }
1661 
1662 // __kmpc_omp_task_parts: Schedule a thread-switchable task for execution
1663 //
1664 // loc_ref: location of original task pragma (ignored)
1665 // gtid: Global Thread ID of encountering thread
1666 // new_task: task thunk allocated by __kmp_omp_task_alloc() for the ''new task''
1667 // Returns:
1668 //    TASK_CURRENT_NOT_QUEUED (0) if did not suspend and queue current task to
1669 //    be resumed later.
1670 //    TASK_CURRENT_QUEUED (1) if suspended and queued the current task to be
1671 //    resumed later.
1672 kmp_int32 __kmpc_omp_task_parts(ident_t *loc_ref, kmp_int32 gtid,
1673                                 kmp_task_t *new_task) {
1674   kmp_taskdata_t *new_taskdata = KMP_TASK_TO_TASKDATA(new_task);
1675 
1676   KA_TRACE(10, ("__kmpc_omp_task_parts(enter): T#%d loc=%p task=%p\n", gtid,
1677                 loc_ref, new_taskdata));
1678 
1679 #if OMPT_SUPPORT
1680   kmp_taskdata_t *parent;
1681   if (UNLIKELY(ompt_enabled.enabled)) {
1682     parent = new_taskdata->td_parent;
1683     if (ompt_enabled.ompt_callback_task_create) {
1684       ompt_callbacks.ompt_callback(ompt_callback_task_create)(
1685           &(parent->ompt_task_info.task_data), &(parent->ompt_task_info.frame),
1686           &(new_taskdata->ompt_task_info.task_data), ompt_task_explicit, 0,
1687           OMPT_GET_RETURN_ADDRESS(0));
1688     }
1689   }
1690 #endif
1691 
1692   /* Should we execute the new task or queue it? For now, let's just always try
1693      to queue it.  If the queue fills up, then we'll execute it.  */
1694 
1695   if (__kmp_push_task(gtid, new_task) == TASK_NOT_PUSHED) // if cannot defer
1696   { // Execute this task immediately
1697     kmp_taskdata_t *current_task = __kmp_threads[gtid]->th.th_current_task;
1698     new_taskdata->td_flags.task_serial = 1;
1699     __kmp_invoke_task(gtid, new_task, current_task);
1700   }
1701 
1702   KA_TRACE(
1703       10,
1704       ("__kmpc_omp_task_parts(exit): T#%d returning TASK_CURRENT_NOT_QUEUED: "
1705        "loc=%p task=%p, return: TASK_CURRENT_NOT_QUEUED\n",
1706        gtid, loc_ref, new_taskdata));
1707 
1708 #if OMPT_SUPPORT
1709   if (UNLIKELY(ompt_enabled.enabled)) {
1710     parent->ompt_task_info.frame.enter_frame = ompt_data_none;
1711   }
1712 #endif
1713   return TASK_CURRENT_NOT_QUEUED;
1714 }
1715 
1716 // __kmp_omp_task: Schedule a non-thread-switchable task for execution
1717 //
1718 // gtid: Global Thread ID of encountering thread
1719 // new_task:non-thread-switchable task thunk allocated by __kmp_omp_task_alloc()
1720 // serialize_immediate: if TRUE then if the task is executed immediately its
1721 // execution will be serialized
1722 // Returns:
1723 //    TASK_CURRENT_NOT_QUEUED (0) if did not suspend and queue current task to
1724 //    be resumed later.
1725 //    TASK_CURRENT_QUEUED (1) if suspended and queued the current task to be
1726 //    resumed later.
1727 kmp_int32 __kmp_omp_task(kmp_int32 gtid, kmp_task_t *new_task,
1728                          bool serialize_immediate) {
1729   kmp_taskdata_t *new_taskdata = KMP_TASK_TO_TASKDATA(new_task);
1730 
1731   /* Should we execute the new task or queue it? For now, let's just always try
1732      to queue it.  If the queue fills up, then we'll execute it.  */
1733   if (new_taskdata->td_flags.proxy == TASK_PROXY ||
1734       __kmp_push_task(gtid, new_task) == TASK_NOT_PUSHED) // if cannot defer
1735   { // Execute this task immediately
1736     kmp_taskdata_t *current_task = __kmp_threads[gtid]->th.th_current_task;
1737     if (serialize_immediate)
1738       new_taskdata->td_flags.task_serial = 1;
1739     __kmp_invoke_task(gtid, new_task, current_task);
1740   }
1741 
1742   return TASK_CURRENT_NOT_QUEUED;
1743 }
1744 
1745 // __kmpc_omp_task: Wrapper around __kmp_omp_task to schedule a
1746 // non-thread-switchable task from the parent thread only!
1747 //
1748 // loc_ref: location of original task pragma (ignored)
1749 // gtid: Global Thread ID of encountering thread
1750 // new_task: non-thread-switchable task thunk allocated by
1751 // __kmp_omp_task_alloc()
1752 // Returns:
1753 //    TASK_CURRENT_NOT_QUEUED (0) if did not suspend and queue current task to
1754 //    be resumed later.
1755 //    TASK_CURRENT_QUEUED (1) if suspended and queued the current task to be
1756 //    resumed later.
1757 kmp_int32 __kmpc_omp_task(ident_t *loc_ref, kmp_int32 gtid,
1758                           kmp_task_t *new_task) {
1759   kmp_int32 res;
1760   KMP_SET_THREAD_STATE_BLOCK(EXPLICIT_TASK);
1761 
1762 #if KMP_DEBUG || OMPT_SUPPORT
1763   kmp_taskdata_t *new_taskdata = KMP_TASK_TO_TASKDATA(new_task);
1764 #endif
1765   KA_TRACE(10, ("__kmpc_omp_task(enter): T#%d loc=%p task=%p\n", gtid, loc_ref,
1766                 new_taskdata));
1767   __kmp_assert_valid_gtid(gtid);
1768 
1769 #if OMPT_SUPPORT
1770   kmp_taskdata_t *parent = NULL;
1771   if (UNLIKELY(ompt_enabled.enabled)) {
1772     if (!new_taskdata->td_flags.started) {
1773       OMPT_STORE_RETURN_ADDRESS(gtid);
1774       parent = new_taskdata->td_parent;
1775       if (!parent->ompt_task_info.frame.enter_frame.ptr) {
1776         parent->ompt_task_info.frame.enter_frame.ptr =
1777             OMPT_GET_FRAME_ADDRESS(0);
1778       }
1779       if (ompt_enabled.ompt_callback_task_create) {
1780         ompt_callbacks.ompt_callback(ompt_callback_task_create)(
1781             &(parent->ompt_task_info.task_data),
1782             &(parent->ompt_task_info.frame),
1783             &(new_taskdata->ompt_task_info.task_data),
1784             ompt_task_explicit | TASK_TYPE_DETAILS_FORMAT(new_taskdata), 0,
1785             OMPT_LOAD_RETURN_ADDRESS(gtid));
1786       }
1787     } else {
1788       // We are scheduling the continuation of an UNTIED task.
1789       // Scheduling back to the parent task.
1790       __ompt_task_finish(new_task,
1791                          new_taskdata->ompt_task_info.scheduling_parent,
1792                          ompt_task_switch);
1793       new_taskdata->ompt_task_info.frame.exit_frame = ompt_data_none;
1794     }
1795   }
1796 #endif
1797 
1798   res = __kmp_omp_task(gtid, new_task, true);
1799 
1800   KA_TRACE(10, ("__kmpc_omp_task(exit): T#%d returning "
1801                 "TASK_CURRENT_NOT_QUEUED: loc=%p task=%p\n",
1802                 gtid, loc_ref, new_taskdata));
1803 #if OMPT_SUPPORT
1804   if (UNLIKELY(ompt_enabled.enabled && parent != NULL)) {
1805     parent->ompt_task_info.frame.enter_frame = ompt_data_none;
1806   }
1807 #endif
1808   return res;
1809 }
1810 
1811 // __kmp_omp_taskloop_task: Wrapper around __kmp_omp_task to schedule
1812 // a taskloop task with the correct OMPT return address
1813 //
1814 // loc_ref: location of original task pragma (ignored)
1815 // gtid: Global Thread ID of encountering thread
1816 // new_task: non-thread-switchable task thunk allocated by
1817 // __kmp_omp_task_alloc()
1818 // codeptr_ra: return address for OMPT callback
1819 // Returns:
1820 //    TASK_CURRENT_NOT_QUEUED (0) if did not suspend and queue current task to
1821 //    be resumed later.
1822 //    TASK_CURRENT_QUEUED (1) if suspended and queued the current task to be
1823 //    resumed later.
1824 kmp_int32 __kmp_omp_taskloop_task(ident_t *loc_ref, kmp_int32 gtid,
1825                                   kmp_task_t *new_task, void *codeptr_ra) {
1826   kmp_int32 res;
1827   KMP_SET_THREAD_STATE_BLOCK(EXPLICIT_TASK);
1828 
1829 #if KMP_DEBUG || OMPT_SUPPORT
1830   kmp_taskdata_t *new_taskdata = KMP_TASK_TO_TASKDATA(new_task);
1831 #endif
1832   KA_TRACE(10, ("__kmpc_omp_task(enter): T#%d loc=%p task=%p\n", gtid, loc_ref,
1833                 new_taskdata));
1834 
1835 #if OMPT_SUPPORT
1836   kmp_taskdata_t *parent = NULL;
1837   if (UNLIKELY(ompt_enabled.enabled && !new_taskdata->td_flags.started)) {
1838     parent = new_taskdata->td_parent;
1839     if (!parent->ompt_task_info.frame.enter_frame.ptr)
1840       parent->ompt_task_info.frame.enter_frame.ptr = OMPT_GET_FRAME_ADDRESS(0);
1841     if (ompt_enabled.ompt_callback_task_create) {
1842       ompt_callbacks.ompt_callback(ompt_callback_task_create)(
1843           &(parent->ompt_task_info.task_data), &(parent->ompt_task_info.frame),
1844           &(new_taskdata->ompt_task_info.task_data),
1845           ompt_task_explicit | TASK_TYPE_DETAILS_FORMAT(new_taskdata), 0,
1846           codeptr_ra);
1847     }
1848   }
1849 #endif
1850 
1851   res = __kmp_omp_task(gtid, new_task, true);
1852 
1853   KA_TRACE(10, ("__kmpc_omp_task(exit): T#%d returning "
1854                 "TASK_CURRENT_NOT_QUEUED: loc=%p task=%p\n",
1855                 gtid, loc_ref, new_taskdata));
1856 #if OMPT_SUPPORT
1857   if (UNLIKELY(ompt_enabled.enabled && parent != NULL)) {
1858     parent->ompt_task_info.frame.enter_frame = ompt_data_none;
1859   }
1860 #endif
1861   return res;
1862 }
1863 
1864 template <bool ompt>
1865 static kmp_int32 __kmpc_omp_taskwait_template(ident_t *loc_ref, kmp_int32 gtid,
1866                                               void *frame_address,
1867                                               void *return_address) {
1868   kmp_taskdata_t *taskdata = nullptr;
1869   kmp_info_t *thread;
1870   int thread_finished = FALSE;
1871   KMP_SET_THREAD_STATE_BLOCK(TASKWAIT);
1872 
1873   KA_TRACE(10, ("__kmpc_omp_taskwait(enter): T#%d loc=%p\n", gtid, loc_ref));
1874   KMP_DEBUG_ASSERT(gtid >= 0);
1875 
1876   if (__kmp_tasking_mode != tskm_immediate_exec) {
1877     thread = __kmp_threads[gtid];
1878     taskdata = thread->th.th_current_task;
1879 
1880 #if OMPT_SUPPORT && OMPT_OPTIONAL
1881     ompt_data_t *my_task_data;
1882     ompt_data_t *my_parallel_data;
1883 
1884     if (ompt) {
1885       my_task_data = &(taskdata->ompt_task_info.task_data);
1886       my_parallel_data = OMPT_CUR_TEAM_DATA(thread);
1887 
1888       taskdata->ompt_task_info.frame.enter_frame.ptr = frame_address;
1889 
1890       if (ompt_enabled.ompt_callback_sync_region) {
1891         ompt_callbacks.ompt_callback(ompt_callback_sync_region)(
1892             ompt_sync_region_taskwait, ompt_scope_begin, my_parallel_data,
1893             my_task_data, return_address);
1894       }
1895 
1896       if (ompt_enabled.ompt_callback_sync_region_wait) {
1897         ompt_callbacks.ompt_callback(ompt_callback_sync_region_wait)(
1898             ompt_sync_region_taskwait, ompt_scope_begin, my_parallel_data,
1899             my_task_data, return_address);
1900       }
1901     }
1902 #endif // OMPT_SUPPORT && OMPT_OPTIONAL
1903 
1904 // Debugger: The taskwait is active. Store location and thread encountered the
1905 // taskwait.
1906 #if USE_ITT_BUILD
1907 // Note: These values are used by ITT events as well.
1908 #endif /* USE_ITT_BUILD */
1909     taskdata->td_taskwait_counter += 1;
1910     taskdata->td_taskwait_ident = loc_ref;
1911     taskdata->td_taskwait_thread = gtid + 1;
1912 
1913 #if USE_ITT_BUILD
1914     void *itt_sync_obj = NULL;
1915 #if USE_ITT_NOTIFY
1916     KMP_ITT_TASKWAIT_STARTING(itt_sync_obj);
1917 #endif /* USE_ITT_NOTIFY */
1918 #endif /* USE_ITT_BUILD */
1919 
1920     bool must_wait =
1921         !taskdata->td_flags.team_serial && !taskdata->td_flags.final;
1922 
1923     must_wait = must_wait || (thread->th.th_task_team != NULL &&
1924                               thread->th.th_task_team->tt.tt_found_proxy_tasks);
1925     // If hidden helper thread is encountered, we must enable wait here.
1926     must_wait =
1927         must_wait ||
1928         (__kmp_enable_hidden_helper && thread->th.th_task_team != NULL &&
1929          thread->th.th_task_team->tt.tt_hidden_helper_task_encountered);
1930 
1931     if (must_wait) {
1932       kmp_flag_32<false, false> flag(
1933           RCAST(std::atomic<kmp_uint32> *,
1934                 &(taskdata->td_incomplete_child_tasks)),
1935           0U);
1936       while (KMP_ATOMIC_LD_ACQ(&taskdata->td_incomplete_child_tasks) != 0) {
1937         flag.execute_tasks(thread, gtid, FALSE,
1938                            &thread_finished USE_ITT_BUILD_ARG(itt_sync_obj),
1939                            __kmp_task_stealing_constraint);
1940       }
1941     }
1942 #if USE_ITT_BUILD
1943     KMP_ITT_TASKWAIT_FINISHED(itt_sync_obj);
1944     KMP_FSYNC_ACQUIRED(taskdata); // acquire self - sync with children
1945 #endif /* USE_ITT_BUILD */
1946 
1947     // Debugger:  The taskwait is completed. Location remains, but thread is
1948     // negated.
1949     taskdata->td_taskwait_thread = -taskdata->td_taskwait_thread;
1950 
1951 #if OMPT_SUPPORT && OMPT_OPTIONAL
1952     if (ompt) {
1953       if (ompt_enabled.ompt_callback_sync_region_wait) {
1954         ompt_callbacks.ompt_callback(ompt_callback_sync_region_wait)(
1955             ompt_sync_region_taskwait, ompt_scope_end, my_parallel_data,
1956             my_task_data, return_address);
1957       }
1958       if (ompt_enabled.ompt_callback_sync_region) {
1959         ompt_callbacks.ompt_callback(ompt_callback_sync_region)(
1960             ompt_sync_region_taskwait, ompt_scope_end, my_parallel_data,
1961             my_task_data, return_address);
1962       }
1963       taskdata->ompt_task_info.frame.enter_frame = ompt_data_none;
1964     }
1965 #endif // OMPT_SUPPORT && OMPT_OPTIONAL
1966 
1967   }
1968 
1969   KA_TRACE(10, ("__kmpc_omp_taskwait(exit): T#%d task %p finished waiting, "
1970                 "returning TASK_CURRENT_NOT_QUEUED\n",
1971                 gtid, taskdata));
1972 
1973   return TASK_CURRENT_NOT_QUEUED;
1974 }
1975 
1976 #if OMPT_SUPPORT && OMPT_OPTIONAL
1977 OMPT_NOINLINE
1978 static kmp_int32 __kmpc_omp_taskwait_ompt(ident_t *loc_ref, kmp_int32 gtid,
1979                                           void *frame_address,
1980                                           void *return_address) {
1981   return __kmpc_omp_taskwait_template<true>(loc_ref, gtid, frame_address,
1982                                             return_address);
1983 }
1984 #endif // OMPT_SUPPORT && OMPT_OPTIONAL
1985 
1986 // __kmpc_omp_taskwait: Wait until all tasks generated by the current task are
1987 // complete
1988 kmp_int32 __kmpc_omp_taskwait(ident_t *loc_ref, kmp_int32 gtid) {
1989 #if OMPT_SUPPORT && OMPT_OPTIONAL
1990   if (UNLIKELY(ompt_enabled.enabled)) {
1991     OMPT_STORE_RETURN_ADDRESS(gtid);
1992     return __kmpc_omp_taskwait_ompt(loc_ref, gtid, OMPT_GET_FRAME_ADDRESS(0),
1993                                     OMPT_LOAD_RETURN_ADDRESS(gtid));
1994   }
1995 #endif
1996   return __kmpc_omp_taskwait_template<false>(loc_ref, gtid, NULL, NULL);
1997 }
1998 
1999 // __kmpc_omp_taskyield: switch to a different task
2000 kmp_int32 __kmpc_omp_taskyield(ident_t *loc_ref, kmp_int32 gtid, int end_part) {
2001   kmp_taskdata_t *taskdata = NULL;
2002   kmp_info_t *thread;
2003   int thread_finished = FALSE;
2004 
2005   KMP_COUNT_BLOCK(OMP_TASKYIELD);
2006   KMP_SET_THREAD_STATE_BLOCK(TASKYIELD);
2007 
2008   KA_TRACE(10, ("__kmpc_omp_taskyield(enter): T#%d loc=%p end_part = %d\n",
2009                 gtid, loc_ref, end_part));
2010   __kmp_assert_valid_gtid(gtid);
2011 
2012   if (__kmp_tasking_mode != tskm_immediate_exec && __kmp_init_parallel) {
2013     thread = __kmp_threads[gtid];
2014     taskdata = thread->th.th_current_task;
2015 // Should we model this as a task wait or not?
2016 // Debugger: The taskwait is active. Store location and thread encountered the
2017 // taskwait.
2018 #if USE_ITT_BUILD
2019 // Note: These values are used by ITT events as well.
2020 #endif /* USE_ITT_BUILD */
2021     taskdata->td_taskwait_counter += 1;
2022     taskdata->td_taskwait_ident = loc_ref;
2023     taskdata->td_taskwait_thread = gtid + 1;
2024 
2025 #if USE_ITT_BUILD
2026     void *itt_sync_obj = NULL;
2027 #if USE_ITT_NOTIFY
2028     KMP_ITT_TASKWAIT_STARTING(itt_sync_obj);
2029 #endif /* USE_ITT_NOTIFY */
2030 #endif /* USE_ITT_BUILD */
2031     if (!taskdata->td_flags.team_serial) {
2032       kmp_task_team_t *task_team = thread->th.th_task_team;
2033       if (task_team != NULL) {
2034         if (KMP_TASKING_ENABLED(task_team)) {
2035 #if OMPT_SUPPORT
2036           if (UNLIKELY(ompt_enabled.enabled))
2037             thread->th.ompt_thread_info.ompt_task_yielded = 1;
2038 #endif
2039           __kmp_execute_tasks_32(
2040               thread, gtid, (kmp_flag_32<> *)NULL, FALSE,
2041               &thread_finished USE_ITT_BUILD_ARG(itt_sync_obj),
2042               __kmp_task_stealing_constraint);
2043 #if OMPT_SUPPORT
2044           if (UNLIKELY(ompt_enabled.enabled))
2045             thread->th.ompt_thread_info.ompt_task_yielded = 0;
2046 #endif
2047         }
2048       }
2049     }
2050 #if USE_ITT_BUILD
2051     KMP_ITT_TASKWAIT_FINISHED(itt_sync_obj);
2052 #endif /* USE_ITT_BUILD */
2053 
2054     // Debugger:  The taskwait is completed. Location remains, but thread is
2055     // negated.
2056     taskdata->td_taskwait_thread = -taskdata->td_taskwait_thread;
2057   }
2058 
2059   KA_TRACE(10, ("__kmpc_omp_taskyield(exit): T#%d task %p resuming, "
2060                 "returning TASK_CURRENT_NOT_QUEUED\n",
2061                 gtid, taskdata));
2062 
2063   return TASK_CURRENT_NOT_QUEUED;
2064 }
2065 
2066 // Task Reduction implementation
2067 //
2068 // Note: initial implementation didn't take into account the possibility
2069 // to specify omp_orig for initializer of the UDR (user defined reduction).
2070 // Corrected implementation takes into account the omp_orig object.
2071 // Compiler is free to use old implementation if omp_orig is not specified.
2072 
2073 /*!
2074 @ingroup BASIC_TYPES
2075 @{
2076 */
2077 
2078 /*!
2079 Flags for special info per task reduction item.
2080 */
2081 typedef struct kmp_taskred_flags {
2082   /*! 1 - use lazy alloc/init (e.g. big objects, #tasks < #threads) */
2083   unsigned lazy_priv : 1;
2084   unsigned reserved31 : 31;
2085 } kmp_taskred_flags_t;
2086 
2087 /*!
2088 Internal struct for reduction data item related info set up by compiler.
2089 */
2090 typedef struct kmp_task_red_input {
2091   void *reduce_shar; /**< shared between tasks item to reduce into */
2092   size_t reduce_size; /**< size of data item in bytes */
2093   // three compiler-generated routines (init, fini are optional):
2094   void *reduce_init; /**< data initialization routine (single parameter) */
2095   void *reduce_fini; /**< data finalization routine */
2096   void *reduce_comb; /**< data combiner routine */
2097   kmp_taskred_flags_t flags; /**< flags for additional info from compiler */
2098 } kmp_task_red_input_t;
2099 
2100 /*!
2101 Internal struct for reduction data item related info saved by the library.
2102 */
2103 typedef struct kmp_taskred_data {
2104   void *reduce_shar; /**< shared between tasks item to reduce into */
2105   size_t reduce_size; /**< size of data item */
2106   kmp_taskred_flags_t flags; /**< flags for additional info from compiler */
2107   void *reduce_priv; /**< array of thread specific items */
2108   void *reduce_pend; /**< end of private data for faster comparison op */
2109   // three compiler-generated routines (init, fini are optional):
2110   void *reduce_comb; /**< data combiner routine */
2111   void *reduce_init; /**< data initialization routine (two parameters) */
2112   void *reduce_fini; /**< data finalization routine */
2113   void *reduce_orig; /**< original item (can be used in UDR initializer) */
2114 } kmp_taskred_data_t;
2115 
2116 /*!
2117 Internal struct for reduction data item related info set up by compiler.
2118 
2119 New interface: added reduce_orig field to provide omp_orig for UDR initializer.
2120 */
2121 typedef struct kmp_taskred_input {
2122   void *reduce_shar; /**< shared between tasks item to reduce into */
2123   void *reduce_orig; /**< original reduction item used for initialization */
2124   size_t reduce_size; /**< size of data item */
2125   // three compiler-generated routines (init, fini are optional):
2126   void *reduce_init; /**< data initialization routine (two parameters) */
2127   void *reduce_fini; /**< data finalization routine */
2128   void *reduce_comb; /**< data combiner routine */
2129   kmp_taskred_flags_t flags; /**< flags for additional info from compiler */
2130 } kmp_taskred_input_t;
2131 /*!
2132 @}
2133 */
2134 
2135 template <typename T> void __kmp_assign_orig(kmp_taskred_data_t &item, T &src);
2136 template <>
2137 void __kmp_assign_orig<kmp_task_red_input_t>(kmp_taskred_data_t &item,
2138                                              kmp_task_red_input_t &src) {
2139   item.reduce_orig = NULL;
2140 }
2141 template <>
2142 void __kmp_assign_orig<kmp_taskred_input_t>(kmp_taskred_data_t &item,
2143                                             kmp_taskred_input_t &src) {
2144   if (src.reduce_orig != NULL) {
2145     item.reduce_orig = src.reduce_orig;
2146   } else {
2147     item.reduce_orig = src.reduce_shar;
2148   } // non-NULL reduce_orig means new interface used
2149 }
2150 
2151 template <typename T> void __kmp_call_init(kmp_taskred_data_t &item, size_t j);
2152 template <>
2153 void __kmp_call_init<kmp_task_red_input_t>(kmp_taskred_data_t &item,
2154                                            size_t offset) {
2155   ((void (*)(void *))item.reduce_init)((char *)(item.reduce_priv) + offset);
2156 }
2157 template <>
2158 void __kmp_call_init<kmp_taskred_input_t>(kmp_taskred_data_t &item,
2159                                           size_t offset) {
2160   ((void (*)(void *, void *))item.reduce_init)(
2161       (char *)(item.reduce_priv) + offset, item.reduce_orig);
2162 }
2163 
2164 template <typename T>
2165 void *__kmp_task_reduction_init(int gtid, int num, T *data) {
2166   __kmp_assert_valid_gtid(gtid);
2167   kmp_info_t *thread = __kmp_threads[gtid];
2168   kmp_taskgroup_t *tg = thread->th.th_current_task->td_taskgroup;
2169   kmp_uint32 nth = thread->th.th_team_nproc;
2170   kmp_taskred_data_t *arr;
2171 
2172   // check input data just in case
2173   KMP_ASSERT(tg != NULL);
2174   KMP_ASSERT(data != NULL);
2175   KMP_ASSERT(num > 0);
2176   if (nth == 1) {
2177     KA_TRACE(10, ("__kmpc_task_reduction_init: T#%d, tg %p, exiting nth=1\n",
2178                   gtid, tg));
2179     return (void *)tg;
2180   }
2181   KA_TRACE(10, ("__kmpc_task_reduction_init: T#%d, taskgroup %p, #items %d\n",
2182                 gtid, tg, num));
2183   arr = (kmp_taskred_data_t *)__kmp_thread_malloc(
2184       thread, num * sizeof(kmp_taskred_data_t));
2185   for (int i = 0; i < num; ++i) {
2186     size_t size = data[i].reduce_size - 1;
2187     // round the size up to cache line per thread-specific item
2188     size += CACHE_LINE - size % CACHE_LINE;
2189     KMP_ASSERT(data[i].reduce_comb != NULL); // combiner is mandatory
2190     arr[i].reduce_shar = data[i].reduce_shar;
2191     arr[i].reduce_size = size;
2192     arr[i].flags = data[i].flags;
2193     arr[i].reduce_comb = data[i].reduce_comb;
2194     arr[i].reduce_init = data[i].reduce_init;
2195     arr[i].reduce_fini = data[i].reduce_fini;
2196     __kmp_assign_orig<T>(arr[i], data[i]);
2197     if (!arr[i].flags.lazy_priv) {
2198       // allocate cache-line aligned block and fill it with zeros
2199       arr[i].reduce_priv = __kmp_allocate(nth * size);
2200       arr[i].reduce_pend = (char *)(arr[i].reduce_priv) + nth * size;
2201       if (arr[i].reduce_init != NULL) {
2202         // initialize all thread-specific items
2203         for (size_t j = 0; j < nth; ++j) {
2204           __kmp_call_init<T>(arr[i], j * size);
2205         }
2206       }
2207     } else {
2208       // only allocate space for pointers now,
2209       // objects will be lazily allocated/initialized if/when requested
2210       // note that __kmp_allocate zeroes the allocated memory
2211       arr[i].reduce_priv = __kmp_allocate(nth * sizeof(void *));
2212     }
2213   }
2214   tg->reduce_data = (void *)arr;
2215   tg->reduce_num_data = num;
2216   return (void *)tg;
2217 }
2218 
2219 /*!
2220 @ingroup TASKING
2221 @param gtid      Global thread ID
2222 @param num       Number of data items to reduce
2223 @param data      Array of data for reduction
2224 @return The taskgroup identifier
2225 
2226 Initialize task reduction for the taskgroup.
2227 
2228 Note: this entry supposes the optional compiler-generated initializer routine
2229 has single parameter - pointer to object to be initialized. That means
2230 the reduction either does not use omp_orig object, or the omp_orig is accessible
2231 without help of the runtime library.
2232 */
2233 void *__kmpc_task_reduction_init(int gtid, int num, void *data) {
2234   return __kmp_task_reduction_init(gtid, num, (kmp_task_red_input_t *)data);
2235 }
2236 
2237 /*!
2238 @ingroup TASKING
2239 @param gtid      Global thread ID
2240 @param num       Number of data items to reduce
2241 @param data      Array of data for reduction
2242 @return The taskgroup identifier
2243 
2244 Initialize task reduction for the taskgroup.
2245 
2246 Note: this entry supposes the optional compiler-generated initializer routine
2247 has two parameters, pointer to object to be initialized and pointer to omp_orig
2248 */
2249 void *__kmpc_taskred_init(int gtid, int num, void *data) {
2250   return __kmp_task_reduction_init(gtid, num, (kmp_taskred_input_t *)data);
2251 }
2252 
2253 // Copy task reduction data (except for shared pointers).
2254 template <typename T>
2255 void __kmp_task_reduction_init_copy(kmp_info_t *thr, int num, T *data,
2256                                     kmp_taskgroup_t *tg, void *reduce_data) {
2257   kmp_taskred_data_t *arr;
2258   KA_TRACE(20, ("__kmp_task_reduction_init_copy: Th %p, init taskgroup %p,"
2259                 " from data %p\n",
2260                 thr, tg, reduce_data));
2261   arr = (kmp_taskred_data_t *)__kmp_thread_malloc(
2262       thr, num * sizeof(kmp_taskred_data_t));
2263   // threads will share private copies, thunk routines, sizes, flags, etc.:
2264   KMP_MEMCPY(arr, reduce_data, num * sizeof(kmp_taskred_data_t));
2265   for (int i = 0; i < num; ++i) {
2266     arr[i].reduce_shar = data[i].reduce_shar; // init unique shared pointers
2267   }
2268   tg->reduce_data = (void *)arr;
2269   tg->reduce_num_data = num;
2270 }
2271 
2272 /*!
2273 @ingroup TASKING
2274 @param gtid    Global thread ID
2275 @param tskgrp  The taskgroup ID (optional)
2276 @param data    Shared location of the item
2277 @return The pointer to per-thread data
2278 
2279 Get thread-specific location of data item
2280 */
2281 void *__kmpc_task_reduction_get_th_data(int gtid, void *tskgrp, void *data) {
2282   __kmp_assert_valid_gtid(gtid);
2283   kmp_info_t *thread = __kmp_threads[gtid];
2284   kmp_int32 nth = thread->th.th_team_nproc;
2285   if (nth == 1)
2286     return data; // nothing to do
2287 
2288   kmp_taskgroup_t *tg = (kmp_taskgroup_t *)tskgrp;
2289   if (tg == NULL)
2290     tg = thread->th.th_current_task->td_taskgroup;
2291   KMP_ASSERT(tg != NULL);
2292   kmp_taskred_data_t *arr = (kmp_taskred_data_t *)(tg->reduce_data);
2293   kmp_int32 num = tg->reduce_num_data;
2294   kmp_int32 tid = thread->th.th_info.ds.ds_tid;
2295 
2296   KMP_ASSERT(data != NULL);
2297   while (tg != NULL) {
2298     for (int i = 0; i < num; ++i) {
2299       if (!arr[i].flags.lazy_priv) {
2300         if (data == arr[i].reduce_shar ||
2301             (data >= arr[i].reduce_priv && data < arr[i].reduce_pend))
2302           return (char *)(arr[i].reduce_priv) + tid * arr[i].reduce_size;
2303       } else {
2304         // check shared location first
2305         void **p_priv = (void **)(arr[i].reduce_priv);
2306         if (data == arr[i].reduce_shar)
2307           goto found;
2308         // check if we get some thread specific location as parameter
2309         for (int j = 0; j < nth; ++j)
2310           if (data == p_priv[j])
2311             goto found;
2312         continue; // not found, continue search
2313       found:
2314         if (p_priv[tid] == NULL) {
2315           // allocate thread specific object lazily
2316           p_priv[tid] = __kmp_allocate(arr[i].reduce_size);
2317           if (arr[i].reduce_init != NULL) {
2318             if (arr[i].reduce_orig != NULL) { // new interface
2319               ((void (*)(void *, void *))arr[i].reduce_init)(
2320                   p_priv[tid], arr[i].reduce_orig);
2321             } else { // old interface (single parameter)
2322               ((void (*)(void *))arr[i].reduce_init)(p_priv[tid]);
2323             }
2324           }
2325         }
2326         return p_priv[tid];
2327       }
2328     }
2329     tg = tg->parent;
2330     arr = (kmp_taskred_data_t *)(tg->reduce_data);
2331     num = tg->reduce_num_data;
2332   }
2333   KMP_ASSERT2(0, "Unknown task reduction item");
2334   return NULL; // ERROR, this line never executed
2335 }
2336 
2337 // Finalize task reduction.
2338 // Called from __kmpc_end_taskgroup()
2339 static void __kmp_task_reduction_fini(kmp_info_t *th, kmp_taskgroup_t *tg) {
2340   kmp_int32 nth = th->th.th_team_nproc;
2341   KMP_DEBUG_ASSERT(nth > 1); // should not be called if nth == 1
2342   kmp_taskred_data_t *arr = (kmp_taskred_data_t *)tg->reduce_data;
2343   kmp_int32 num = tg->reduce_num_data;
2344   for (int i = 0; i < num; ++i) {
2345     void *sh_data = arr[i].reduce_shar;
2346     void (*f_fini)(void *) = (void (*)(void *))(arr[i].reduce_fini);
2347     void (*f_comb)(void *, void *) =
2348         (void (*)(void *, void *))(arr[i].reduce_comb);
2349     if (!arr[i].flags.lazy_priv) {
2350       void *pr_data = arr[i].reduce_priv;
2351       size_t size = arr[i].reduce_size;
2352       for (int j = 0; j < nth; ++j) {
2353         void *priv_data = (char *)pr_data + j * size;
2354         f_comb(sh_data, priv_data); // combine results
2355         if (f_fini)
2356           f_fini(priv_data); // finalize if needed
2357       }
2358     } else {
2359       void **pr_data = (void **)(arr[i].reduce_priv);
2360       for (int j = 0; j < nth; ++j) {
2361         if (pr_data[j] != NULL) {
2362           f_comb(sh_data, pr_data[j]); // combine results
2363           if (f_fini)
2364             f_fini(pr_data[j]); // finalize if needed
2365           __kmp_free(pr_data[j]);
2366         }
2367       }
2368     }
2369     __kmp_free(arr[i].reduce_priv);
2370   }
2371   __kmp_thread_free(th, arr);
2372   tg->reduce_data = NULL;
2373   tg->reduce_num_data = 0;
2374 }
2375 
2376 // Cleanup task reduction data for parallel or worksharing,
2377 // do not touch task private data other threads still working with.
2378 // Called from __kmpc_end_taskgroup()
2379 static void __kmp_task_reduction_clean(kmp_info_t *th, kmp_taskgroup_t *tg) {
2380   __kmp_thread_free(th, tg->reduce_data);
2381   tg->reduce_data = NULL;
2382   tg->reduce_num_data = 0;
2383 }
2384 
2385 template <typename T>
2386 void *__kmp_task_reduction_modifier_init(ident_t *loc, int gtid, int is_ws,
2387                                          int num, T *data) {
2388   __kmp_assert_valid_gtid(gtid);
2389   kmp_info_t *thr = __kmp_threads[gtid];
2390   kmp_int32 nth = thr->th.th_team_nproc;
2391   __kmpc_taskgroup(loc, gtid); // form new taskgroup first
2392   if (nth == 1) {
2393     KA_TRACE(10,
2394              ("__kmpc_reduction_modifier_init: T#%d, tg %p, exiting nth=1\n",
2395               gtid, thr->th.th_current_task->td_taskgroup));
2396     return (void *)thr->th.th_current_task->td_taskgroup;
2397   }
2398   kmp_team_t *team = thr->th.th_team;
2399   void *reduce_data;
2400   kmp_taskgroup_t *tg;
2401   reduce_data = KMP_ATOMIC_LD_RLX(&team->t.t_tg_reduce_data[is_ws]);
2402   if (reduce_data == NULL &&
2403       __kmp_atomic_compare_store(&team->t.t_tg_reduce_data[is_ws], reduce_data,
2404                                  (void *)1)) {
2405     // single thread enters this block to initialize common reduction data
2406     KMP_DEBUG_ASSERT(reduce_data == NULL);
2407     // first initialize own data, then make a copy other threads can use
2408     tg = (kmp_taskgroup_t *)__kmp_task_reduction_init<T>(gtid, num, data);
2409     reduce_data = __kmp_thread_malloc(thr, num * sizeof(kmp_taskred_data_t));
2410     KMP_MEMCPY(reduce_data, tg->reduce_data, num * sizeof(kmp_taskred_data_t));
2411     // fini counters should be 0 at this point
2412     KMP_DEBUG_ASSERT(KMP_ATOMIC_LD_RLX(&team->t.t_tg_fini_counter[0]) == 0);
2413     KMP_DEBUG_ASSERT(KMP_ATOMIC_LD_RLX(&team->t.t_tg_fini_counter[1]) == 0);
2414     KMP_ATOMIC_ST_REL(&team->t.t_tg_reduce_data[is_ws], reduce_data);
2415   } else {
2416     while (
2417         (reduce_data = KMP_ATOMIC_LD_ACQ(&team->t.t_tg_reduce_data[is_ws])) ==
2418         (void *)1) { // wait for task reduction initialization
2419       KMP_CPU_PAUSE();
2420     }
2421     KMP_DEBUG_ASSERT(reduce_data > (void *)1); // should be valid pointer here
2422     tg = thr->th.th_current_task->td_taskgroup;
2423     __kmp_task_reduction_init_copy<T>(thr, num, data, tg, reduce_data);
2424   }
2425   return tg;
2426 }
2427 
2428 /*!
2429 @ingroup TASKING
2430 @param loc       Source location info
2431 @param gtid      Global thread ID
2432 @param is_ws     Is 1 if the reduction is for worksharing, 0 otherwise
2433 @param num       Number of data items to reduce
2434 @param data      Array of data for reduction
2435 @return The taskgroup identifier
2436 
2437 Initialize task reduction for a parallel or worksharing.
2438 
2439 Note: this entry supposes the optional compiler-generated initializer routine
2440 has single parameter - pointer to object to be initialized. That means
2441 the reduction either does not use omp_orig object, or the omp_orig is accessible
2442 without help of the runtime library.
2443 */
2444 void *__kmpc_task_reduction_modifier_init(ident_t *loc, int gtid, int is_ws,
2445                                           int num, void *data) {
2446   return __kmp_task_reduction_modifier_init(loc, gtid, is_ws, num,
2447                                             (kmp_task_red_input_t *)data);
2448 }
2449 
2450 /*!
2451 @ingroup TASKING
2452 @param loc       Source location info
2453 @param gtid      Global thread ID
2454 @param is_ws     Is 1 if the reduction is for worksharing, 0 otherwise
2455 @param num       Number of data items to reduce
2456 @param data      Array of data for reduction
2457 @return The taskgroup identifier
2458 
2459 Initialize task reduction for a parallel or worksharing.
2460 
2461 Note: this entry supposes the optional compiler-generated initializer routine
2462 has two parameters, pointer to object to be initialized and pointer to omp_orig
2463 */
2464 void *__kmpc_taskred_modifier_init(ident_t *loc, int gtid, int is_ws, int num,
2465                                    void *data) {
2466   return __kmp_task_reduction_modifier_init(loc, gtid, is_ws, num,
2467                                             (kmp_taskred_input_t *)data);
2468 }
2469 
2470 /*!
2471 @ingroup TASKING
2472 @param loc       Source location info
2473 @param gtid      Global thread ID
2474 @param is_ws     Is 1 if the reduction is for worksharing, 0 otherwise
2475 
2476 Finalize task reduction for a parallel or worksharing.
2477 */
2478 void __kmpc_task_reduction_modifier_fini(ident_t *loc, int gtid, int is_ws) {
2479   __kmpc_end_taskgroup(loc, gtid);
2480 }
2481 
2482 // __kmpc_taskgroup: Start a new taskgroup
2483 void __kmpc_taskgroup(ident_t *loc, int gtid) {
2484   __kmp_assert_valid_gtid(gtid);
2485   kmp_info_t *thread = __kmp_threads[gtid];
2486   kmp_taskdata_t *taskdata = thread->th.th_current_task;
2487   kmp_taskgroup_t *tg_new =
2488       (kmp_taskgroup_t *)__kmp_thread_malloc(thread, sizeof(kmp_taskgroup_t));
2489   KA_TRACE(10, ("__kmpc_taskgroup: T#%d loc=%p group=%p\n", gtid, loc, tg_new));
2490   KMP_ATOMIC_ST_RLX(&tg_new->count, 0);
2491   KMP_ATOMIC_ST_RLX(&tg_new->cancel_request, cancel_noreq);
2492   tg_new->parent = taskdata->td_taskgroup;
2493   tg_new->reduce_data = NULL;
2494   tg_new->reduce_num_data = 0;
2495   tg_new->gomp_data = NULL;
2496   taskdata->td_taskgroup = tg_new;
2497 
2498 #if OMPT_SUPPORT && OMPT_OPTIONAL
2499   if (UNLIKELY(ompt_enabled.ompt_callback_sync_region)) {
2500     void *codeptr = OMPT_LOAD_RETURN_ADDRESS(gtid);
2501     if (!codeptr)
2502       codeptr = OMPT_GET_RETURN_ADDRESS(0);
2503     kmp_team_t *team = thread->th.th_team;
2504     ompt_data_t my_task_data = taskdata->ompt_task_info.task_data;
2505     // FIXME: I think this is wrong for lwt!
2506     ompt_data_t my_parallel_data = team->t.ompt_team_info.parallel_data;
2507 
2508     ompt_callbacks.ompt_callback(ompt_callback_sync_region)(
2509         ompt_sync_region_taskgroup, ompt_scope_begin, &(my_parallel_data),
2510         &(my_task_data), codeptr);
2511   }
2512 #endif
2513 }
2514 
2515 // __kmpc_end_taskgroup: Wait until all tasks generated by the current task
2516 //                       and its descendants are complete
2517 void __kmpc_end_taskgroup(ident_t *loc, int gtid) {
2518   __kmp_assert_valid_gtid(gtid);
2519   kmp_info_t *thread = __kmp_threads[gtid];
2520   kmp_taskdata_t *taskdata = thread->th.th_current_task;
2521   kmp_taskgroup_t *taskgroup = taskdata->td_taskgroup;
2522   int thread_finished = FALSE;
2523 
2524 #if OMPT_SUPPORT && OMPT_OPTIONAL
2525   kmp_team_t *team;
2526   ompt_data_t my_task_data;
2527   ompt_data_t my_parallel_data;
2528   void *codeptr = nullptr;
2529   if (UNLIKELY(ompt_enabled.enabled)) {
2530     team = thread->th.th_team;
2531     my_task_data = taskdata->ompt_task_info.task_data;
2532     // FIXME: I think this is wrong for lwt!
2533     my_parallel_data = team->t.ompt_team_info.parallel_data;
2534     codeptr = OMPT_LOAD_RETURN_ADDRESS(gtid);
2535     if (!codeptr)
2536       codeptr = OMPT_GET_RETURN_ADDRESS(0);
2537   }
2538 #endif
2539 
2540   KA_TRACE(10, ("__kmpc_end_taskgroup(enter): T#%d loc=%p\n", gtid, loc));
2541   KMP_DEBUG_ASSERT(taskgroup != NULL);
2542   KMP_SET_THREAD_STATE_BLOCK(TASKGROUP);
2543 
2544   if (__kmp_tasking_mode != tskm_immediate_exec) {
2545     // mark task as waiting not on a barrier
2546     taskdata->td_taskwait_counter += 1;
2547     taskdata->td_taskwait_ident = loc;
2548     taskdata->td_taskwait_thread = gtid + 1;
2549 #if USE_ITT_BUILD
2550     // For ITT the taskgroup wait is similar to taskwait until we need to
2551     // distinguish them
2552     void *itt_sync_obj = NULL;
2553 #if USE_ITT_NOTIFY
2554     KMP_ITT_TASKWAIT_STARTING(itt_sync_obj);
2555 #endif /* USE_ITT_NOTIFY */
2556 #endif /* USE_ITT_BUILD */
2557 
2558 #if OMPT_SUPPORT && OMPT_OPTIONAL
2559     if (UNLIKELY(ompt_enabled.ompt_callback_sync_region_wait)) {
2560       ompt_callbacks.ompt_callback(ompt_callback_sync_region_wait)(
2561           ompt_sync_region_taskgroup, ompt_scope_begin, &(my_parallel_data),
2562           &(my_task_data), codeptr);
2563     }
2564 #endif
2565 
2566     if (!taskdata->td_flags.team_serial ||
2567         (thread->th.th_task_team != NULL &&
2568          (thread->th.th_task_team->tt.tt_found_proxy_tasks ||
2569           thread->th.th_task_team->tt.tt_hidden_helper_task_encountered))) {
2570       kmp_flag_32<false, false> flag(
2571           RCAST(std::atomic<kmp_uint32> *, &(taskgroup->count)), 0U);
2572       while (KMP_ATOMIC_LD_ACQ(&taskgroup->count) != 0) {
2573         flag.execute_tasks(thread, gtid, FALSE,
2574                            &thread_finished USE_ITT_BUILD_ARG(itt_sync_obj),
2575                            __kmp_task_stealing_constraint);
2576       }
2577     }
2578     taskdata->td_taskwait_thread = -taskdata->td_taskwait_thread; // end waiting
2579 
2580 #if OMPT_SUPPORT && OMPT_OPTIONAL
2581     if (UNLIKELY(ompt_enabled.ompt_callback_sync_region_wait)) {
2582       ompt_callbacks.ompt_callback(ompt_callback_sync_region_wait)(
2583           ompt_sync_region_taskgroup, ompt_scope_end, &(my_parallel_data),
2584           &(my_task_data), codeptr);
2585     }
2586 #endif
2587 
2588 #if USE_ITT_BUILD
2589     KMP_ITT_TASKWAIT_FINISHED(itt_sync_obj);
2590     KMP_FSYNC_ACQUIRED(taskdata); // acquire self - sync with descendants
2591 #endif /* USE_ITT_BUILD */
2592   }
2593   KMP_DEBUG_ASSERT(taskgroup->count == 0);
2594 
2595   if (taskgroup->reduce_data != NULL &&
2596       !taskgroup->gomp_data) { // need to reduce?
2597     int cnt;
2598     void *reduce_data;
2599     kmp_team_t *t = thread->th.th_team;
2600     kmp_taskred_data_t *arr = (kmp_taskred_data_t *)taskgroup->reduce_data;
2601     // check if <priv> data of the first reduction variable shared for the team
2602     void *priv0 = arr[0].reduce_priv;
2603     if ((reduce_data = KMP_ATOMIC_LD_ACQ(&t->t.t_tg_reduce_data[0])) != NULL &&
2604         ((kmp_taskred_data_t *)reduce_data)[0].reduce_priv == priv0) {
2605       // finishing task reduction on parallel
2606       cnt = KMP_ATOMIC_INC(&t->t.t_tg_fini_counter[0]);
2607       if (cnt == thread->th.th_team_nproc - 1) {
2608         // we are the last thread passing __kmpc_reduction_modifier_fini()
2609         // finalize task reduction:
2610         __kmp_task_reduction_fini(thread, taskgroup);
2611         // cleanup fields in the team structure:
2612         // TODO: is relaxed store enough here (whole barrier should follow)?
2613         __kmp_thread_free(thread, reduce_data);
2614         KMP_ATOMIC_ST_REL(&t->t.t_tg_reduce_data[0], NULL);
2615         KMP_ATOMIC_ST_REL(&t->t.t_tg_fini_counter[0], 0);
2616       } else {
2617         // we are not the last thread passing __kmpc_reduction_modifier_fini(),
2618         // so do not finalize reduction, just clean own copy of the data
2619         __kmp_task_reduction_clean(thread, taskgroup);
2620       }
2621     } else if ((reduce_data = KMP_ATOMIC_LD_ACQ(&t->t.t_tg_reduce_data[1])) !=
2622                    NULL &&
2623                ((kmp_taskred_data_t *)reduce_data)[0].reduce_priv == priv0) {
2624       // finishing task reduction on worksharing
2625       cnt = KMP_ATOMIC_INC(&t->t.t_tg_fini_counter[1]);
2626       if (cnt == thread->th.th_team_nproc - 1) {
2627         // we are the last thread passing __kmpc_reduction_modifier_fini()
2628         __kmp_task_reduction_fini(thread, taskgroup);
2629         // cleanup fields in team structure:
2630         // TODO: is relaxed store enough here (whole barrier should follow)?
2631         __kmp_thread_free(thread, reduce_data);
2632         KMP_ATOMIC_ST_REL(&t->t.t_tg_reduce_data[1], NULL);
2633         KMP_ATOMIC_ST_REL(&t->t.t_tg_fini_counter[1], 0);
2634       } else {
2635         // we are not the last thread passing __kmpc_reduction_modifier_fini(),
2636         // so do not finalize reduction, just clean own copy of the data
2637         __kmp_task_reduction_clean(thread, taskgroup);
2638       }
2639     } else {
2640       // finishing task reduction on taskgroup
2641       __kmp_task_reduction_fini(thread, taskgroup);
2642     }
2643   }
2644   // Restore parent taskgroup for the current task
2645   taskdata->td_taskgroup = taskgroup->parent;
2646   __kmp_thread_free(thread, taskgroup);
2647 
2648   KA_TRACE(10, ("__kmpc_end_taskgroup(exit): T#%d task %p finished waiting\n",
2649                 gtid, taskdata));
2650 
2651 #if OMPT_SUPPORT && OMPT_OPTIONAL
2652   if (UNLIKELY(ompt_enabled.ompt_callback_sync_region)) {
2653     ompt_callbacks.ompt_callback(ompt_callback_sync_region)(
2654         ompt_sync_region_taskgroup, ompt_scope_end, &(my_parallel_data),
2655         &(my_task_data), codeptr);
2656   }
2657 #endif
2658 }
2659 
2660 // __kmp_remove_my_task: remove a task from my own deque
2661 static kmp_task_t *__kmp_remove_my_task(kmp_info_t *thread, kmp_int32 gtid,
2662                                         kmp_task_team_t *task_team,
2663                                         kmp_int32 is_constrained) {
2664   kmp_task_t *task;
2665   kmp_taskdata_t *taskdata;
2666   kmp_thread_data_t *thread_data;
2667   kmp_uint32 tail;
2668 
2669   KMP_DEBUG_ASSERT(__kmp_tasking_mode != tskm_immediate_exec);
2670   KMP_DEBUG_ASSERT(task_team->tt.tt_threads_data !=
2671                    NULL); // Caller should check this condition
2672 
2673   thread_data = &task_team->tt.tt_threads_data[__kmp_tid_from_gtid(gtid)];
2674 
2675   KA_TRACE(10, ("__kmp_remove_my_task(enter): T#%d ntasks=%d head=%u tail=%u\n",
2676                 gtid, thread_data->td.td_deque_ntasks,
2677                 thread_data->td.td_deque_head, thread_data->td.td_deque_tail));
2678 
2679   if (TCR_4(thread_data->td.td_deque_ntasks) == 0) {
2680     KA_TRACE(10,
2681              ("__kmp_remove_my_task(exit #1): T#%d No tasks to remove: "
2682               "ntasks=%d head=%u tail=%u\n",
2683               gtid, thread_data->td.td_deque_ntasks,
2684               thread_data->td.td_deque_head, thread_data->td.td_deque_tail));
2685     return NULL;
2686   }
2687 
2688   __kmp_acquire_bootstrap_lock(&thread_data->td.td_deque_lock);
2689 
2690   if (TCR_4(thread_data->td.td_deque_ntasks) == 0) {
2691     __kmp_release_bootstrap_lock(&thread_data->td.td_deque_lock);
2692     KA_TRACE(10,
2693              ("__kmp_remove_my_task(exit #2): T#%d No tasks to remove: "
2694               "ntasks=%d head=%u tail=%u\n",
2695               gtid, thread_data->td.td_deque_ntasks,
2696               thread_data->td.td_deque_head, thread_data->td.td_deque_tail));
2697     return NULL;
2698   }
2699 
2700   tail = (thread_data->td.td_deque_tail - 1) &
2701          TASK_DEQUE_MASK(thread_data->td); // Wrap index.
2702   taskdata = thread_data->td.td_deque[tail];
2703 
2704   if (!__kmp_task_is_allowed(gtid, is_constrained, taskdata,
2705                              thread->th.th_current_task)) {
2706     // The TSC does not allow to steal victim task
2707     __kmp_release_bootstrap_lock(&thread_data->td.td_deque_lock);
2708     KA_TRACE(10,
2709              ("__kmp_remove_my_task(exit #3): T#%d TSC blocks tail task: "
2710               "ntasks=%d head=%u tail=%u\n",
2711               gtid, thread_data->td.td_deque_ntasks,
2712               thread_data->td.td_deque_head, thread_data->td.td_deque_tail));
2713     return NULL;
2714   }
2715 
2716   thread_data->td.td_deque_tail = tail;
2717   TCW_4(thread_data->td.td_deque_ntasks, thread_data->td.td_deque_ntasks - 1);
2718 
2719   __kmp_release_bootstrap_lock(&thread_data->td.td_deque_lock);
2720 
2721   KA_TRACE(10, ("__kmp_remove_my_task(exit #4): T#%d task %p removed: "
2722                 "ntasks=%d head=%u tail=%u\n",
2723                 gtid, taskdata, thread_data->td.td_deque_ntasks,
2724                 thread_data->td.td_deque_head, thread_data->td.td_deque_tail));
2725 
2726   task = KMP_TASKDATA_TO_TASK(taskdata);
2727   return task;
2728 }
2729 
2730 // __kmp_steal_task: remove a task from another thread's deque
2731 // Assume that calling thread has already checked existence of
2732 // task_team thread_data before calling this routine.
2733 static kmp_task_t *__kmp_steal_task(kmp_info_t *victim_thr, kmp_int32 gtid,
2734                                     kmp_task_team_t *task_team,
2735                                     std::atomic<kmp_int32> *unfinished_threads,
2736                                     int *thread_finished,
2737                                     kmp_int32 is_constrained) {
2738   kmp_task_t *task;
2739   kmp_taskdata_t *taskdata;
2740   kmp_taskdata_t *current;
2741   kmp_thread_data_t *victim_td, *threads_data;
2742   kmp_int32 target;
2743   kmp_int32 victim_tid;
2744 
2745   KMP_DEBUG_ASSERT(__kmp_tasking_mode != tskm_immediate_exec);
2746 
2747   threads_data = task_team->tt.tt_threads_data;
2748   KMP_DEBUG_ASSERT(threads_data != NULL); // Caller should check this condition
2749 
2750   victim_tid = victim_thr->th.th_info.ds.ds_tid;
2751   victim_td = &threads_data[victim_tid];
2752 
2753   KA_TRACE(10, ("__kmp_steal_task(enter): T#%d try to steal from T#%d: "
2754                 "task_team=%p ntasks=%d head=%u tail=%u\n",
2755                 gtid, __kmp_gtid_from_thread(victim_thr), task_team,
2756                 victim_td->td.td_deque_ntasks, victim_td->td.td_deque_head,
2757                 victim_td->td.td_deque_tail));
2758 
2759   if (TCR_4(victim_td->td.td_deque_ntasks) == 0) {
2760     KA_TRACE(10, ("__kmp_steal_task(exit #1): T#%d could not steal from T#%d: "
2761                   "task_team=%p ntasks=%d head=%u tail=%u\n",
2762                   gtid, __kmp_gtid_from_thread(victim_thr), task_team,
2763                   victim_td->td.td_deque_ntasks, victim_td->td.td_deque_head,
2764                   victim_td->td.td_deque_tail));
2765     return NULL;
2766   }
2767 
2768   __kmp_acquire_bootstrap_lock(&victim_td->td.td_deque_lock);
2769 
2770   int ntasks = TCR_4(victim_td->td.td_deque_ntasks);
2771   // Check again after we acquire the lock
2772   if (ntasks == 0) {
2773     __kmp_release_bootstrap_lock(&victim_td->td.td_deque_lock);
2774     KA_TRACE(10, ("__kmp_steal_task(exit #2): T#%d could not steal from T#%d: "
2775                   "task_team=%p ntasks=%d head=%u tail=%u\n",
2776                   gtid, __kmp_gtid_from_thread(victim_thr), task_team, ntasks,
2777                   victim_td->td.td_deque_head, victim_td->td.td_deque_tail));
2778     return NULL;
2779   }
2780 
2781   KMP_DEBUG_ASSERT(victim_td->td.td_deque != NULL);
2782   current = __kmp_threads[gtid]->th.th_current_task;
2783   taskdata = victim_td->td.td_deque[victim_td->td.td_deque_head];
2784   if (__kmp_task_is_allowed(gtid, is_constrained, taskdata, current)) {
2785     // Bump head pointer and Wrap.
2786     victim_td->td.td_deque_head =
2787         (victim_td->td.td_deque_head + 1) & TASK_DEQUE_MASK(victim_td->td);
2788   } else {
2789     if (!task_team->tt.tt_untied_task_encountered) {
2790       // The TSC does not allow to steal victim task
2791       __kmp_release_bootstrap_lock(&victim_td->td.td_deque_lock);
2792       KA_TRACE(10, ("__kmp_steal_task(exit #3): T#%d could not steal from "
2793                     "T#%d: task_team=%p ntasks=%d head=%u tail=%u\n",
2794                     gtid, __kmp_gtid_from_thread(victim_thr), task_team, ntasks,
2795                     victim_td->td.td_deque_head, victim_td->td.td_deque_tail));
2796       return NULL;
2797     }
2798     int i;
2799     // walk through victim's deque trying to steal any task
2800     target = victim_td->td.td_deque_head;
2801     taskdata = NULL;
2802     for (i = 1; i < ntasks; ++i) {
2803       target = (target + 1) & TASK_DEQUE_MASK(victim_td->td);
2804       taskdata = victim_td->td.td_deque[target];
2805       if (__kmp_task_is_allowed(gtid, is_constrained, taskdata, current)) {
2806         break; // found victim task
2807       } else {
2808         taskdata = NULL;
2809       }
2810     }
2811     if (taskdata == NULL) {
2812       // No appropriate candidate to steal found
2813       __kmp_release_bootstrap_lock(&victim_td->td.td_deque_lock);
2814       KA_TRACE(10, ("__kmp_steal_task(exit #4): T#%d could not steal from "
2815                     "T#%d: task_team=%p ntasks=%d head=%u tail=%u\n",
2816                     gtid, __kmp_gtid_from_thread(victim_thr), task_team, ntasks,
2817                     victim_td->td.td_deque_head, victim_td->td.td_deque_tail));
2818       return NULL;
2819     }
2820     int prev = target;
2821     for (i = i + 1; i < ntasks; ++i) {
2822       // shift remaining tasks in the deque left by 1
2823       target = (target + 1) & TASK_DEQUE_MASK(victim_td->td);
2824       victim_td->td.td_deque[prev] = victim_td->td.td_deque[target];
2825       prev = target;
2826     }
2827     KMP_DEBUG_ASSERT(
2828         victim_td->td.td_deque_tail ==
2829         (kmp_uint32)((target + 1) & TASK_DEQUE_MASK(victim_td->td)));
2830     victim_td->td.td_deque_tail = target; // tail -= 1 (wrapped))
2831   }
2832   if (*thread_finished) {
2833     // We need to un-mark this victim as a finished victim.  This must be done
2834     // before releasing the lock, or else other threads (starting with the
2835     // primary thread victim) might be prematurely released from the barrier!!!
2836     kmp_int32 count;
2837 
2838     count = KMP_ATOMIC_INC(unfinished_threads);
2839 
2840     KA_TRACE(
2841         20,
2842         ("__kmp_steal_task: T#%d inc unfinished_threads to %d: task_team=%p\n",
2843          gtid, count + 1, task_team));
2844 
2845     *thread_finished = FALSE;
2846   }
2847   TCW_4(victim_td->td.td_deque_ntasks, ntasks - 1);
2848 
2849   __kmp_release_bootstrap_lock(&victim_td->td.td_deque_lock);
2850 
2851   KMP_COUNT_BLOCK(TASK_stolen);
2852   KA_TRACE(10,
2853            ("__kmp_steal_task(exit #5): T#%d stole task %p from T#%d: "
2854             "task_team=%p ntasks=%d head=%u tail=%u\n",
2855             gtid, taskdata, __kmp_gtid_from_thread(victim_thr), task_team,
2856             ntasks, victim_td->td.td_deque_head, victim_td->td.td_deque_tail));
2857 
2858   task = KMP_TASKDATA_TO_TASK(taskdata);
2859   return task;
2860 }
2861 
2862 // __kmp_execute_tasks_template: Choose and execute tasks until either the
2863 // condition is statisfied (return true) or there are none left (return false).
2864 //
2865 // final_spin is TRUE if this is the spin at the release barrier.
2866 // thread_finished indicates whether the thread is finished executing all
2867 // the tasks it has on its deque, and is at the release barrier.
2868 // spinner is the location on which to spin.
2869 // spinner == NULL means only execute a single task and return.
2870 // checker is the value to check to terminate the spin.
2871 template <class C>
2872 static inline int __kmp_execute_tasks_template(
2873     kmp_info_t *thread, kmp_int32 gtid, C *flag, int final_spin,
2874     int *thread_finished USE_ITT_BUILD_ARG(void *itt_sync_obj),
2875     kmp_int32 is_constrained) {
2876   kmp_task_team_t *task_team = thread->th.th_task_team;
2877   kmp_thread_data_t *threads_data;
2878   kmp_task_t *task;
2879   kmp_info_t *other_thread;
2880   kmp_taskdata_t *current_task = thread->th.th_current_task;
2881   std::atomic<kmp_int32> *unfinished_threads;
2882   kmp_int32 nthreads, victim_tid = -2, use_own_tasks = 1, new_victim = 0,
2883                       tid = thread->th.th_info.ds.ds_tid;
2884 
2885   KMP_DEBUG_ASSERT(__kmp_tasking_mode != tskm_immediate_exec);
2886   KMP_DEBUG_ASSERT(thread == __kmp_threads[gtid]);
2887 
2888   if (task_team == NULL || current_task == NULL)
2889     return FALSE;
2890 
2891   KA_TRACE(15, ("__kmp_execute_tasks_template(enter): T#%d final_spin=%d "
2892                 "*thread_finished=%d\n",
2893                 gtid, final_spin, *thread_finished));
2894 
2895   thread->th.th_reap_state = KMP_NOT_SAFE_TO_REAP;
2896   threads_data = (kmp_thread_data_t *)TCR_PTR(task_team->tt.tt_threads_data);
2897 
2898   KMP_DEBUG_ASSERT(threads_data != NULL);
2899 
2900   nthreads = task_team->tt.tt_nproc;
2901   unfinished_threads = &(task_team->tt.tt_unfinished_threads);
2902   KMP_DEBUG_ASSERT(nthreads > 1 || task_team->tt.tt_found_proxy_tasks ||
2903                    task_team->tt.tt_hidden_helper_task_encountered);
2904   KMP_DEBUG_ASSERT(*unfinished_threads >= 0);
2905 
2906   while (1) { // Outer loop keeps trying to find tasks in case of single thread
2907     // getting tasks from target constructs
2908     while (1) { // Inner loop to find a task and execute it
2909       task = NULL;
2910       if (use_own_tasks) { // check on own queue first
2911         task = __kmp_remove_my_task(thread, gtid, task_team, is_constrained);
2912       }
2913       if ((task == NULL) && (nthreads > 1)) { // Steal a task
2914         int asleep = 1;
2915         use_own_tasks = 0;
2916         // Try to steal from the last place I stole from successfully.
2917         if (victim_tid == -2) { // haven't stolen anything yet
2918           victim_tid = threads_data[tid].td.td_deque_last_stolen;
2919           if (victim_tid !=
2920               -1) // if we have a last stolen from victim, get the thread
2921             other_thread = threads_data[victim_tid].td.td_thr;
2922         }
2923         if (victim_tid != -1) { // found last victim
2924           asleep = 0;
2925         } else if (!new_victim) { // no recent steals and we haven't already
2926           // used a new victim; select a random thread
2927           do { // Find a different thread to steal work from.
2928             // Pick a random thread. Initial plan was to cycle through all the
2929             // threads, and only return if we tried to steal from every thread,
2930             // and failed.  Arch says that's not such a great idea.
2931             victim_tid = __kmp_get_random(thread) % (nthreads - 1);
2932             if (victim_tid >= tid) {
2933               ++victim_tid; // Adjusts random distribution to exclude self
2934             }
2935             // Found a potential victim
2936             other_thread = threads_data[victim_tid].td.td_thr;
2937             // There is a slight chance that __kmp_enable_tasking() did not wake
2938             // up all threads waiting at the barrier.  If victim is sleeping,
2939             // then wake it up. Since we were going to pay the cache miss
2940             // penalty for referencing another thread's kmp_info_t struct
2941             // anyway,
2942             // the check shouldn't cost too much performance at this point. In
2943             // extra barrier mode, tasks do not sleep at the separate tasking
2944             // barrier, so this isn't a problem.
2945             asleep = 0;
2946             if ((__kmp_tasking_mode == tskm_task_teams) &&
2947                 (__kmp_dflt_blocktime != KMP_MAX_BLOCKTIME) &&
2948                 (TCR_PTR(CCAST(void *, other_thread->th.th_sleep_loc)) !=
2949                  NULL)) {
2950               asleep = 1;
2951               __kmp_null_resume_wrapper(__kmp_gtid_from_thread(other_thread),
2952                                         other_thread->th.th_sleep_loc);
2953               // A sleeping thread should not have any tasks on it's queue.
2954               // There is a slight possibility that it resumes, steals a task
2955               // from another thread, which spawns more tasks, all in the time
2956               // that it takes this thread to check => don't write an assertion
2957               // that the victim's queue is empty.  Try stealing from a
2958               // different thread.
2959             }
2960           } while (asleep);
2961         }
2962 
2963         if (!asleep) {
2964           // We have a victim to try to steal from
2965           task = __kmp_steal_task(other_thread, gtid, task_team,
2966                                   unfinished_threads, thread_finished,
2967                                   is_constrained);
2968         }
2969         if (task != NULL) { // set last stolen to victim
2970           if (threads_data[tid].td.td_deque_last_stolen != victim_tid) {
2971             threads_data[tid].td.td_deque_last_stolen = victim_tid;
2972             // The pre-refactored code did not try more than 1 successful new
2973             // vicitm, unless the last one generated more local tasks;
2974             // new_victim keeps track of this
2975             new_victim = 1;
2976           }
2977         } else { // No tasks found; unset last_stolen
2978           KMP_CHECK_UPDATE(threads_data[tid].td.td_deque_last_stolen, -1);
2979           victim_tid = -2; // no successful victim found
2980         }
2981       }
2982 
2983       if (task == NULL)
2984         break; // break out of tasking loop
2985 
2986 // Found a task; execute it
2987 #if USE_ITT_BUILD && USE_ITT_NOTIFY
2988       if (__itt_sync_create_ptr || KMP_ITT_DEBUG) {
2989         if (itt_sync_obj == NULL) { // we are at fork barrier where we could not
2990           // get the object reliably
2991           itt_sync_obj = __kmp_itt_barrier_object(gtid, bs_forkjoin_barrier);
2992         }
2993         __kmp_itt_task_starting(itt_sync_obj);
2994       }
2995 #endif /* USE_ITT_BUILD && USE_ITT_NOTIFY */
2996       __kmp_invoke_task(gtid, task, current_task);
2997 #if USE_ITT_BUILD
2998       if (itt_sync_obj != NULL)
2999         __kmp_itt_task_finished(itt_sync_obj);
3000 #endif /* USE_ITT_BUILD */
3001       // If this thread is only partway through the barrier and the condition is
3002       // met, then return now, so that the barrier gather/release pattern can
3003       // proceed. If this thread is in the last spin loop in the barrier,
3004       // waiting to be released, we know that the termination condition will not
3005       // be satisfied, so don't waste any cycles checking it.
3006       if (flag == NULL || (!final_spin && flag->done_check())) {
3007         KA_TRACE(
3008             15,
3009             ("__kmp_execute_tasks_template: T#%d spin condition satisfied\n",
3010              gtid));
3011         return TRUE;
3012       }
3013       if (thread->th.th_task_team == NULL) {
3014         break;
3015       }
3016       KMP_YIELD(__kmp_library == library_throughput); // Yield before next task
3017       // If execution of a stolen task results in more tasks being placed on our
3018       // run queue, reset use_own_tasks
3019       if (!use_own_tasks && TCR_4(threads_data[tid].td.td_deque_ntasks) != 0) {
3020         KA_TRACE(20, ("__kmp_execute_tasks_template: T#%d stolen task spawned "
3021                       "other tasks, restart\n",
3022                       gtid));
3023         use_own_tasks = 1;
3024         new_victim = 0;
3025       }
3026     }
3027 
3028     // The task source has been exhausted. If in final spin loop of barrier,
3029     // check if termination condition is satisfied. The work queue may be empty
3030     // but there might be proxy tasks still executing.
3031     if (final_spin &&
3032         KMP_ATOMIC_LD_ACQ(&current_task->td_incomplete_child_tasks) == 0) {
3033       // First, decrement the #unfinished threads, if that has not already been
3034       // done.  This decrement might be to the spin location, and result in the
3035       // termination condition being satisfied.
3036       if (!*thread_finished) {
3037         kmp_int32 count;
3038 
3039         count = KMP_ATOMIC_DEC(unfinished_threads) - 1;
3040         KA_TRACE(20, ("__kmp_execute_tasks_template: T#%d dec "
3041                       "unfinished_threads to %d task_team=%p\n",
3042                       gtid, count, task_team));
3043         *thread_finished = TRUE;
3044       }
3045 
3046       // It is now unsafe to reference thread->th.th_team !!!
3047       // Decrementing task_team->tt.tt_unfinished_threads can allow the primary
3048       // thread to pass through the barrier, where it might reset each thread's
3049       // th.th_team field for the next parallel region. If we can steal more
3050       // work, we know that this has not happened yet.
3051       if (flag != NULL && flag->done_check()) {
3052         KA_TRACE(
3053             15,
3054             ("__kmp_execute_tasks_template: T#%d spin condition satisfied\n",
3055              gtid));
3056         return TRUE;
3057       }
3058     }
3059 
3060     // If this thread's task team is NULL, primary thread has recognized that
3061     // there are no more tasks; bail out
3062     if (thread->th.th_task_team == NULL) {
3063       KA_TRACE(15,
3064                ("__kmp_execute_tasks_template: T#%d no more tasks\n", gtid));
3065       return FALSE;
3066     }
3067 
3068     // We could be getting tasks from target constructs; if this is the only
3069     // thread, keep trying to execute tasks from own queue
3070     if (nthreads == 1 &&
3071         KMP_ATOMIC_LD_ACQ(&current_task->td_incomplete_child_tasks))
3072       use_own_tasks = 1;
3073     else {
3074       KA_TRACE(15,
3075                ("__kmp_execute_tasks_template: T#%d can't find work\n", gtid));
3076       return FALSE;
3077     }
3078   }
3079 }
3080 
3081 template <bool C, bool S>
3082 int __kmp_execute_tasks_32(
3083     kmp_info_t *thread, kmp_int32 gtid, kmp_flag_32<C, S> *flag, int final_spin,
3084     int *thread_finished USE_ITT_BUILD_ARG(void *itt_sync_obj),
3085     kmp_int32 is_constrained) {
3086   return __kmp_execute_tasks_template(
3087       thread, gtid, flag, final_spin,
3088       thread_finished USE_ITT_BUILD_ARG(itt_sync_obj), is_constrained);
3089 }
3090 
3091 template <bool C, bool S>
3092 int __kmp_execute_tasks_64(
3093     kmp_info_t *thread, kmp_int32 gtid, kmp_flag_64<C, S> *flag, int final_spin,
3094     int *thread_finished USE_ITT_BUILD_ARG(void *itt_sync_obj),
3095     kmp_int32 is_constrained) {
3096   return __kmp_execute_tasks_template(
3097       thread, gtid, flag, final_spin,
3098       thread_finished USE_ITT_BUILD_ARG(itt_sync_obj), is_constrained);
3099 }
3100 
3101 int __kmp_execute_tasks_oncore(
3102     kmp_info_t *thread, kmp_int32 gtid, kmp_flag_oncore *flag, int final_spin,
3103     int *thread_finished USE_ITT_BUILD_ARG(void *itt_sync_obj),
3104     kmp_int32 is_constrained) {
3105   return __kmp_execute_tasks_template(
3106       thread, gtid, flag, final_spin,
3107       thread_finished USE_ITT_BUILD_ARG(itt_sync_obj), is_constrained);
3108 }
3109 
3110 template int
3111 __kmp_execute_tasks_32<false, false>(kmp_info_t *, kmp_int32,
3112                                      kmp_flag_32<false, false> *, int,
3113                                      int *USE_ITT_BUILD_ARG(void *), kmp_int32);
3114 
3115 template int __kmp_execute_tasks_64<false, true>(kmp_info_t *, kmp_int32,
3116                                                  kmp_flag_64<false, true> *,
3117                                                  int,
3118                                                  int *USE_ITT_BUILD_ARG(void *),
3119                                                  kmp_int32);
3120 
3121 template int __kmp_execute_tasks_64<true, false>(kmp_info_t *, kmp_int32,
3122                                                  kmp_flag_64<true, false> *,
3123                                                  int,
3124                                                  int *USE_ITT_BUILD_ARG(void *),
3125                                                  kmp_int32);
3126 
3127 // __kmp_enable_tasking: Allocate task team and resume threads sleeping at the
3128 // next barrier so they can assist in executing enqueued tasks.
3129 // First thread in allocates the task team atomically.
3130 static void __kmp_enable_tasking(kmp_task_team_t *task_team,
3131                                  kmp_info_t *this_thr) {
3132   kmp_thread_data_t *threads_data;
3133   int nthreads, i, is_init_thread;
3134 
3135   KA_TRACE(10, ("__kmp_enable_tasking(enter): T#%d\n",
3136                 __kmp_gtid_from_thread(this_thr)));
3137 
3138   KMP_DEBUG_ASSERT(task_team != NULL);
3139   KMP_DEBUG_ASSERT(this_thr->th.th_team != NULL);
3140 
3141   nthreads = task_team->tt.tt_nproc;
3142   KMP_DEBUG_ASSERT(nthreads > 0);
3143   KMP_DEBUG_ASSERT(nthreads == this_thr->th.th_team->t.t_nproc);
3144 
3145   // Allocate or increase the size of threads_data if necessary
3146   is_init_thread = __kmp_realloc_task_threads_data(this_thr, task_team);
3147 
3148   if (!is_init_thread) {
3149     // Some other thread already set up the array.
3150     KA_TRACE(
3151         20,
3152         ("__kmp_enable_tasking(exit): T#%d: threads array already set up.\n",
3153          __kmp_gtid_from_thread(this_thr)));
3154     return;
3155   }
3156   threads_data = (kmp_thread_data_t *)TCR_PTR(task_team->tt.tt_threads_data);
3157   KMP_DEBUG_ASSERT(threads_data != NULL);
3158 
3159   if (__kmp_tasking_mode == tskm_task_teams &&
3160       (__kmp_dflt_blocktime != KMP_MAX_BLOCKTIME)) {
3161     // Release any threads sleeping at the barrier, so that they can steal
3162     // tasks and execute them.  In extra barrier mode, tasks do not sleep
3163     // at the separate tasking barrier, so this isn't a problem.
3164     for (i = 0; i < nthreads; i++) {
3165       volatile void *sleep_loc;
3166       kmp_info_t *thread = threads_data[i].td.td_thr;
3167 
3168       if (i == this_thr->th.th_info.ds.ds_tid) {
3169         continue;
3170       }
3171       // Since we haven't locked the thread's suspend mutex lock at this
3172       // point, there is a small window where a thread might be putting
3173       // itself to sleep, but hasn't set the th_sleep_loc field yet.
3174       // To work around this, __kmp_execute_tasks_template() periodically checks
3175       // see if other threads are sleeping (using the same random mechanism that
3176       // is used for task stealing) and awakens them if they are.
3177       if ((sleep_loc = TCR_PTR(CCAST(void *, thread->th.th_sleep_loc))) !=
3178           NULL) {
3179         KF_TRACE(50, ("__kmp_enable_tasking: T#%d waking up thread T#%d\n",
3180                       __kmp_gtid_from_thread(this_thr),
3181                       __kmp_gtid_from_thread(thread)));
3182         __kmp_null_resume_wrapper(__kmp_gtid_from_thread(thread), sleep_loc);
3183       } else {
3184         KF_TRACE(50, ("__kmp_enable_tasking: T#%d don't wake up thread T#%d\n",
3185                       __kmp_gtid_from_thread(this_thr),
3186                       __kmp_gtid_from_thread(thread)));
3187       }
3188     }
3189   }
3190 
3191   KA_TRACE(10, ("__kmp_enable_tasking(exit): T#%d\n",
3192                 __kmp_gtid_from_thread(this_thr)));
3193 }
3194 
3195 /* // TODO: Check the comment consistency
3196  * Utility routines for "task teams".  A task team (kmp_task_t) is kind of
3197  * like a shadow of the kmp_team_t data struct, with a different lifetime.
3198  * After a child * thread checks into a barrier and calls __kmp_release() from
3199  * the particular variant of __kmp_<barrier_kind>_barrier_gather(), it can no
3200  * longer assume that the kmp_team_t structure is intact (at any moment, the
3201  * primary thread may exit the barrier code and free the team data structure,
3202  * and return the threads to the thread pool).
3203  *
3204  * This does not work with the tasking code, as the thread is still
3205  * expected to participate in the execution of any tasks that may have been
3206  * spawned my a member of the team, and the thread still needs access to all
3207  * to each thread in the team, so that it can steal work from it.
3208  *
3209  * Enter the existence of the kmp_task_team_t struct.  It employs a reference
3210  * counting mechanism, and is allocated by the primary thread before calling
3211  * __kmp_<barrier_kind>_release, and then is release by the last thread to
3212  * exit __kmp_<barrier_kind>_release at the next barrier.  I.e. the lifetimes
3213  * of the kmp_task_team_t structs for consecutive barriers can overlap
3214  * (and will, unless the primary thread is the last thread to exit the barrier
3215  * release phase, which is not typical). The existence of such a struct is
3216  * useful outside the context of tasking.
3217  *
3218  * We currently use the existence of the threads array as an indicator that
3219  * tasks were spawned since the last barrier.  If the structure is to be
3220  * useful outside the context of tasking, then this will have to change, but
3221  * not setting the field minimizes the performance impact of tasking on
3222  * barriers, when no explicit tasks were spawned (pushed, actually).
3223  */
3224 
3225 static kmp_task_team_t *__kmp_free_task_teams =
3226     NULL; // Free list for task_team data structures
3227 // Lock for task team data structures
3228 kmp_bootstrap_lock_t __kmp_task_team_lock =
3229     KMP_BOOTSTRAP_LOCK_INITIALIZER(__kmp_task_team_lock);
3230 
3231 // __kmp_alloc_task_deque:
3232 // Allocates a task deque for a particular thread, and initialize the necessary
3233 // data structures relating to the deque.  This only happens once per thread
3234 // per task team since task teams are recycled. No lock is needed during
3235 // allocation since each thread allocates its own deque.
3236 static void __kmp_alloc_task_deque(kmp_info_t *thread,
3237                                    kmp_thread_data_t *thread_data) {
3238   __kmp_init_bootstrap_lock(&thread_data->td.td_deque_lock);
3239   KMP_DEBUG_ASSERT(thread_data->td.td_deque == NULL);
3240 
3241   // Initialize last stolen task field to "none"
3242   thread_data->td.td_deque_last_stolen = -1;
3243 
3244   KMP_DEBUG_ASSERT(TCR_4(thread_data->td.td_deque_ntasks) == 0);
3245   KMP_DEBUG_ASSERT(thread_data->td.td_deque_head == 0);
3246   KMP_DEBUG_ASSERT(thread_data->td.td_deque_tail == 0);
3247 
3248   KE_TRACE(
3249       10,
3250       ("__kmp_alloc_task_deque: T#%d allocating deque[%d] for thread_data %p\n",
3251        __kmp_gtid_from_thread(thread), INITIAL_TASK_DEQUE_SIZE, thread_data));
3252   // Allocate space for task deque, and zero the deque
3253   // Cannot use __kmp_thread_calloc() because threads not around for
3254   // kmp_reap_task_team( ).
3255   thread_data->td.td_deque = (kmp_taskdata_t **)__kmp_allocate(
3256       INITIAL_TASK_DEQUE_SIZE * sizeof(kmp_taskdata_t *));
3257   thread_data->td.td_deque_size = INITIAL_TASK_DEQUE_SIZE;
3258 }
3259 
3260 // __kmp_free_task_deque:
3261 // Deallocates a task deque for a particular thread. Happens at library
3262 // deallocation so don't need to reset all thread data fields.
3263 static void __kmp_free_task_deque(kmp_thread_data_t *thread_data) {
3264   if (thread_data->td.td_deque != NULL) {
3265     __kmp_acquire_bootstrap_lock(&thread_data->td.td_deque_lock);
3266     TCW_4(thread_data->td.td_deque_ntasks, 0);
3267     __kmp_free(thread_data->td.td_deque);
3268     thread_data->td.td_deque = NULL;
3269     __kmp_release_bootstrap_lock(&thread_data->td.td_deque_lock);
3270   }
3271 
3272 #ifdef BUILD_TIED_TASK_STACK
3273   // GEH: Figure out what to do here for td_susp_tied_tasks
3274   if (thread_data->td.td_susp_tied_tasks.ts_entries != TASK_STACK_EMPTY) {
3275     __kmp_free_task_stack(__kmp_thread_from_gtid(gtid), thread_data);
3276   }
3277 #endif // BUILD_TIED_TASK_STACK
3278 }
3279 
3280 // __kmp_realloc_task_threads_data:
3281 // Allocates a threads_data array for a task team, either by allocating an
3282 // initial array or enlarging an existing array.  Only the first thread to get
3283 // the lock allocs or enlarges the array and re-initializes the array elements.
3284 // That thread returns "TRUE", the rest return "FALSE".
3285 // Assumes that the new array size is given by task_team -> tt.tt_nproc.
3286 // The current size is given by task_team -> tt.tt_max_threads.
3287 static int __kmp_realloc_task_threads_data(kmp_info_t *thread,
3288                                            kmp_task_team_t *task_team) {
3289   kmp_thread_data_t **threads_data_p;
3290   kmp_int32 nthreads, maxthreads;
3291   int is_init_thread = FALSE;
3292 
3293   if (TCR_4(task_team->tt.tt_found_tasks)) {
3294     // Already reallocated and initialized.
3295     return FALSE;
3296   }
3297 
3298   threads_data_p = &task_team->tt.tt_threads_data;
3299   nthreads = task_team->tt.tt_nproc;
3300   maxthreads = task_team->tt.tt_max_threads;
3301 
3302   // All threads must lock when they encounter the first task of the implicit
3303   // task region to make sure threads_data fields are (re)initialized before
3304   // used.
3305   __kmp_acquire_bootstrap_lock(&task_team->tt.tt_threads_lock);
3306 
3307   if (!TCR_4(task_team->tt.tt_found_tasks)) {
3308     // first thread to enable tasking
3309     kmp_team_t *team = thread->th.th_team;
3310     int i;
3311 
3312     is_init_thread = TRUE;
3313     if (maxthreads < nthreads) {
3314 
3315       if (*threads_data_p != NULL) {
3316         kmp_thread_data_t *old_data = *threads_data_p;
3317         kmp_thread_data_t *new_data = NULL;
3318 
3319         KE_TRACE(
3320             10,
3321             ("__kmp_realloc_task_threads_data: T#%d reallocating "
3322              "threads data for task_team %p, new_size = %d, old_size = %d\n",
3323              __kmp_gtid_from_thread(thread), task_team, nthreads, maxthreads));
3324         // Reallocate threads_data to have more elements than current array
3325         // Cannot use __kmp_thread_realloc() because threads not around for
3326         // kmp_reap_task_team( ).  Note all new array entries are initialized
3327         // to zero by __kmp_allocate().
3328         new_data = (kmp_thread_data_t *)__kmp_allocate(
3329             nthreads * sizeof(kmp_thread_data_t));
3330         // copy old data to new data
3331         KMP_MEMCPY_S((void *)new_data, nthreads * sizeof(kmp_thread_data_t),
3332                      (void *)old_data, maxthreads * sizeof(kmp_thread_data_t));
3333 
3334 #ifdef BUILD_TIED_TASK_STACK
3335         // GEH: Figure out if this is the right thing to do
3336         for (i = maxthreads; i < nthreads; i++) {
3337           kmp_thread_data_t *thread_data = &(*threads_data_p)[i];
3338           __kmp_init_task_stack(__kmp_gtid_from_thread(thread), thread_data);
3339         }
3340 #endif // BUILD_TIED_TASK_STACK
3341        // Install the new data and free the old data
3342         (*threads_data_p) = new_data;
3343         __kmp_free(old_data);
3344       } else {
3345         KE_TRACE(10, ("__kmp_realloc_task_threads_data: T#%d allocating "
3346                       "threads data for task_team %p, size = %d\n",
3347                       __kmp_gtid_from_thread(thread), task_team, nthreads));
3348         // Make the initial allocate for threads_data array, and zero entries
3349         // Cannot use __kmp_thread_calloc() because threads not around for
3350         // kmp_reap_task_team( ).
3351         *threads_data_p = (kmp_thread_data_t *)__kmp_allocate(
3352             nthreads * sizeof(kmp_thread_data_t));
3353 #ifdef BUILD_TIED_TASK_STACK
3354         // GEH: Figure out if this is the right thing to do
3355         for (i = 0; i < nthreads; i++) {
3356           kmp_thread_data_t *thread_data = &(*threads_data_p)[i];
3357           __kmp_init_task_stack(__kmp_gtid_from_thread(thread), thread_data);
3358         }
3359 #endif // BUILD_TIED_TASK_STACK
3360       }
3361       task_team->tt.tt_max_threads = nthreads;
3362     } else {
3363       // If array has (more than) enough elements, go ahead and use it
3364       KMP_DEBUG_ASSERT(*threads_data_p != NULL);
3365     }
3366 
3367     // initialize threads_data pointers back to thread_info structures
3368     for (i = 0; i < nthreads; i++) {
3369       kmp_thread_data_t *thread_data = &(*threads_data_p)[i];
3370       thread_data->td.td_thr = team->t.t_threads[i];
3371 
3372       if (thread_data->td.td_deque_last_stolen >= nthreads) {
3373         // The last stolen field survives across teams / barrier, and the number
3374         // of threads may have changed.  It's possible (likely?) that a new
3375         // parallel region will exhibit the same behavior as previous region.
3376         thread_data->td.td_deque_last_stolen = -1;
3377       }
3378     }
3379 
3380     KMP_MB();
3381     TCW_SYNC_4(task_team->tt.tt_found_tasks, TRUE);
3382   }
3383 
3384   __kmp_release_bootstrap_lock(&task_team->tt.tt_threads_lock);
3385   return is_init_thread;
3386 }
3387 
3388 // __kmp_free_task_threads_data:
3389 // Deallocates a threads_data array for a task team, including any attached
3390 // tasking deques.  Only occurs at library shutdown.
3391 static void __kmp_free_task_threads_data(kmp_task_team_t *task_team) {
3392   __kmp_acquire_bootstrap_lock(&task_team->tt.tt_threads_lock);
3393   if (task_team->tt.tt_threads_data != NULL) {
3394     int i;
3395     for (i = 0; i < task_team->tt.tt_max_threads; i++) {
3396       __kmp_free_task_deque(&task_team->tt.tt_threads_data[i]);
3397     }
3398     __kmp_free(task_team->tt.tt_threads_data);
3399     task_team->tt.tt_threads_data = NULL;
3400   }
3401   __kmp_release_bootstrap_lock(&task_team->tt.tt_threads_lock);
3402 }
3403 
3404 // __kmp_allocate_task_team:
3405 // Allocates a task team associated with a specific team, taking it from
3406 // the global task team free list if possible.  Also initializes data
3407 // structures.
3408 static kmp_task_team_t *__kmp_allocate_task_team(kmp_info_t *thread,
3409                                                  kmp_team_t *team) {
3410   kmp_task_team_t *task_team = NULL;
3411   int nthreads;
3412 
3413   KA_TRACE(20, ("__kmp_allocate_task_team: T#%d entering; team = %p\n",
3414                 (thread ? __kmp_gtid_from_thread(thread) : -1), team));
3415 
3416   if (TCR_PTR(__kmp_free_task_teams) != NULL) {
3417     // Take a task team from the task team pool
3418     __kmp_acquire_bootstrap_lock(&__kmp_task_team_lock);
3419     if (__kmp_free_task_teams != NULL) {
3420       task_team = __kmp_free_task_teams;
3421       TCW_PTR(__kmp_free_task_teams, task_team->tt.tt_next);
3422       task_team->tt.tt_next = NULL;
3423     }
3424     __kmp_release_bootstrap_lock(&__kmp_task_team_lock);
3425   }
3426 
3427   if (task_team == NULL) {
3428     KE_TRACE(10, ("__kmp_allocate_task_team: T#%d allocating "
3429                   "task team for team %p\n",
3430                   __kmp_gtid_from_thread(thread), team));
3431     // Allocate a new task team if one is not available. Cannot use
3432     // __kmp_thread_malloc because threads not around for kmp_reap_task_team.
3433     task_team = (kmp_task_team_t *)__kmp_allocate(sizeof(kmp_task_team_t));
3434     __kmp_init_bootstrap_lock(&task_team->tt.tt_threads_lock);
3435 #if USE_ITT_BUILD && USE_ITT_NOTIFY && KMP_DEBUG
3436     // suppress race conditions detection on synchronization flags in debug mode
3437     // this helps to analyze library internals eliminating false positives
3438     __itt_suppress_mark_range(
3439         __itt_suppress_range, __itt_suppress_threading_errors,
3440         &task_team->tt.tt_found_tasks, sizeof(task_team->tt.tt_found_tasks));
3441     __itt_suppress_mark_range(__itt_suppress_range,
3442                               __itt_suppress_threading_errors,
3443                               CCAST(kmp_uint32 *, &task_team->tt.tt_active),
3444                               sizeof(task_team->tt.tt_active));
3445 #endif /* USE_ITT_BUILD && USE_ITT_NOTIFY && KMP_DEBUG */
3446     // Note: __kmp_allocate zeroes returned memory, othewise we would need:
3447     // task_team->tt.tt_threads_data = NULL;
3448     // task_team->tt.tt_max_threads = 0;
3449     // task_team->tt.tt_next = NULL;
3450   }
3451 
3452   TCW_4(task_team->tt.tt_found_tasks, FALSE);
3453   TCW_4(task_team->tt.tt_found_proxy_tasks, FALSE);
3454   task_team->tt.tt_nproc = nthreads = team->t.t_nproc;
3455 
3456   KMP_ATOMIC_ST_REL(&task_team->tt.tt_unfinished_threads, nthreads);
3457   TCW_4(task_team->tt.tt_hidden_helper_task_encountered, FALSE);
3458   TCW_4(task_team->tt.tt_active, TRUE);
3459 
3460   KA_TRACE(20, ("__kmp_allocate_task_team: T#%d exiting; task_team = %p "
3461                 "unfinished_threads init'd to %d\n",
3462                 (thread ? __kmp_gtid_from_thread(thread) : -1), task_team,
3463                 KMP_ATOMIC_LD_RLX(&task_team->tt.tt_unfinished_threads)));
3464   return task_team;
3465 }
3466 
3467 // __kmp_free_task_team:
3468 // Frees the task team associated with a specific thread, and adds it
3469 // to the global task team free list.
3470 void __kmp_free_task_team(kmp_info_t *thread, kmp_task_team_t *task_team) {
3471   KA_TRACE(20, ("__kmp_free_task_team: T#%d task_team = %p\n",
3472                 thread ? __kmp_gtid_from_thread(thread) : -1, task_team));
3473 
3474   // Put task team back on free list
3475   __kmp_acquire_bootstrap_lock(&__kmp_task_team_lock);
3476 
3477   KMP_DEBUG_ASSERT(task_team->tt.tt_next == NULL);
3478   task_team->tt.tt_next = __kmp_free_task_teams;
3479   TCW_PTR(__kmp_free_task_teams, task_team);
3480 
3481   __kmp_release_bootstrap_lock(&__kmp_task_team_lock);
3482 }
3483 
3484 // __kmp_reap_task_teams:
3485 // Free all the task teams on the task team free list.
3486 // Should only be done during library shutdown.
3487 // Cannot do anything that needs a thread structure or gtid since they are
3488 // already gone.
3489 void __kmp_reap_task_teams(void) {
3490   kmp_task_team_t *task_team;
3491 
3492   if (TCR_PTR(__kmp_free_task_teams) != NULL) {
3493     // Free all task_teams on the free list
3494     __kmp_acquire_bootstrap_lock(&__kmp_task_team_lock);
3495     while ((task_team = __kmp_free_task_teams) != NULL) {
3496       __kmp_free_task_teams = task_team->tt.tt_next;
3497       task_team->tt.tt_next = NULL;
3498 
3499       // Free threads_data if necessary
3500       if (task_team->tt.tt_threads_data != NULL) {
3501         __kmp_free_task_threads_data(task_team);
3502       }
3503       __kmp_free(task_team);
3504     }
3505     __kmp_release_bootstrap_lock(&__kmp_task_team_lock);
3506   }
3507 }
3508 
3509 // __kmp_wait_to_unref_task_teams:
3510 // Some threads could still be in the fork barrier release code, possibly
3511 // trying to steal tasks.  Wait for each thread to unreference its task team.
3512 void __kmp_wait_to_unref_task_teams(void) {
3513   kmp_info_t *thread;
3514   kmp_uint32 spins;
3515   int done;
3516 
3517   KMP_INIT_YIELD(spins);
3518 
3519   for (;;) {
3520     done = TRUE;
3521 
3522     // TODO: GEH - this may be is wrong because some sync would be necessary
3523     // in case threads are added to the pool during the traversal. Need to
3524     // verify that lock for thread pool is held when calling this routine.
3525     for (thread = CCAST(kmp_info_t *, __kmp_thread_pool); thread != NULL;
3526          thread = thread->th.th_next_pool) {
3527 #if KMP_OS_WINDOWS
3528       DWORD exit_val;
3529 #endif
3530       if (TCR_PTR(thread->th.th_task_team) == NULL) {
3531         KA_TRACE(10, ("__kmp_wait_to_unref_task_team: T#%d task_team == NULL\n",
3532                       __kmp_gtid_from_thread(thread)));
3533         continue;
3534       }
3535 #if KMP_OS_WINDOWS
3536       // TODO: GEH - add this check for Linux* OS / OS X* as well?
3537       if (!__kmp_is_thread_alive(thread, &exit_val)) {
3538         thread->th.th_task_team = NULL;
3539         continue;
3540       }
3541 #endif
3542 
3543       done = FALSE; // Because th_task_team pointer is not NULL for this thread
3544 
3545       KA_TRACE(10, ("__kmp_wait_to_unref_task_team: Waiting for T#%d to "
3546                     "unreference task_team\n",
3547                     __kmp_gtid_from_thread(thread)));
3548 
3549       if (__kmp_dflt_blocktime != KMP_MAX_BLOCKTIME) {
3550         volatile void *sleep_loc;
3551         // If the thread is sleeping, awaken it.
3552         if ((sleep_loc = TCR_PTR(CCAST(void *, thread->th.th_sleep_loc))) !=
3553             NULL) {
3554           KA_TRACE(
3555               10,
3556               ("__kmp_wait_to_unref_task_team: T#%d waking up thread T#%d\n",
3557                __kmp_gtid_from_thread(thread), __kmp_gtid_from_thread(thread)));
3558           __kmp_null_resume_wrapper(__kmp_gtid_from_thread(thread), sleep_loc);
3559         }
3560       }
3561     }
3562     if (done) {
3563       break;
3564     }
3565 
3566     // If oversubscribed or have waited a bit, yield.
3567     KMP_YIELD_OVERSUB_ELSE_SPIN(spins);
3568   }
3569 }
3570 
3571 // __kmp_task_team_setup:  Create a task_team for the current team, but use
3572 // an already created, unused one if it already exists.
3573 void __kmp_task_team_setup(kmp_info_t *this_thr, kmp_team_t *team, int always) {
3574   KMP_DEBUG_ASSERT(__kmp_tasking_mode != tskm_immediate_exec);
3575 
3576   // If this task_team hasn't been created yet, allocate it. It will be used in
3577   // the region after the next.
3578   // If it exists, it is the current task team and shouldn't be touched yet as
3579   // it may still be in use.
3580   if (team->t.t_task_team[this_thr->th.th_task_state] == NULL &&
3581       (always || team->t.t_nproc > 1)) {
3582     team->t.t_task_team[this_thr->th.th_task_state] =
3583         __kmp_allocate_task_team(this_thr, team);
3584     KA_TRACE(20, ("__kmp_task_team_setup: Primary T#%d created new task_team %p"
3585                   " for team %d at parity=%d\n",
3586                   __kmp_gtid_from_thread(this_thr),
3587                   team->t.t_task_team[this_thr->th.th_task_state], team->t.t_id,
3588                   this_thr->th.th_task_state));
3589   }
3590 
3591   // After threads exit the release, they will call sync, and then point to this
3592   // other task_team; make sure it is allocated and properly initialized. As
3593   // threads spin in the barrier release phase, they will continue to use the
3594   // previous task_team struct(above), until they receive the signal to stop
3595   // checking for tasks (they can't safely reference the kmp_team_t struct,
3596   // which could be reallocated by the primary thread). No task teams are formed
3597   // for serialized teams.
3598   if (team->t.t_nproc > 1) {
3599     int other_team = 1 - this_thr->th.th_task_state;
3600     KMP_DEBUG_ASSERT(other_team >= 0 && other_team < 2);
3601     if (team->t.t_task_team[other_team] == NULL) { // setup other team as well
3602       team->t.t_task_team[other_team] =
3603           __kmp_allocate_task_team(this_thr, team);
3604       KA_TRACE(20, ("__kmp_task_team_setup: Primary T#%d created second new "
3605                     "task_team %p for team %d at parity=%d\n",
3606                     __kmp_gtid_from_thread(this_thr),
3607                     team->t.t_task_team[other_team], team->t.t_id, other_team));
3608     } else { // Leave the old task team struct in place for the upcoming region;
3609       // adjust as needed
3610       kmp_task_team_t *task_team = team->t.t_task_team[other_team];
3611       if (!task_team->tt.tt_active ||
3612           team->t.t_nproc != task_team->tt.tt_nproc) {
3613         TCW_4(task_team->tt.tt_nproc, team->t.t_nproc);
3614         TCW_4(task_team->tt.tt_found_tasks, FALSE);
3615         TCW_4(task_team->tt.tt_found_proxy_tasks, FALSE);
3616         KMP_ATOMIC_ST_REL(&task_team->tt.tt_unfinished_threads,
3617                           team->t.t_nproc);
3618         TCW_4(task_team->tt.tt_active, TRUE);
3619       }
3620       // if team size has changed, the first thread to enable tasking will
3621       // realloc threads_data if necessary
3622       KA_TRACE(20, ("__kmp_task_team_setup: Primary T#%d reset next task_team "
3623                     "%p for team %d at parity=%d\n",
3624                     __kmp_gtid_from_thread(this_thr),
3625                     team->t.t_task_team[other_team], team->t.t_id, other_team));
3626     }
3627   }
3628 
3629   // For regular thread, task enabling should be called when the task is going
3630   // to be pushed to a dequeue. However, for the hidden helper thread, we need
3631   // it ahead of time so that some operations can be performed without race
3632   // condition.
3633   if (this_thr == __kmp_hidden_helper_main_thread) {
3634     for (int i = 0; i < 2; ++i) {
3635       kmp_task_team_t *task_team = team->t.t_task_team[i];
3636       if (KMP_TASKING_ENABLED(task_team)) {
3637         continue;
3638       }
3639       __kmp_enable_tasking(task_team, this_thr);
3640       for (int j = 0; j < task_team->tt.tt_nproc; ++j) {
3641         kmp_thread_data_t *thread_data = &task_team->tt.tt_threads_data[j];
3642         if (thread_data->td.td_deque == NULL) {
3643           __kmp_alloc_task_deque(__kmp_hidden_helper_threads[j], thread_data);
3644         }
3645       }
3646     }
3647   }
3648 }
3649 
3650 // __kmp_task_team_sync: Propagation of task team data from team to threads
3651 // which happens just after the release phase of a team barrier.  This may be
3652 // called by any thread, but only for teams with # threads > 1.
3653 void __kmp_task_team_sync(kmp_info_t *this_thr, kmp_team_t *team) {
3654   KMP_DEBUG_ASSERT(__kmp_tasking_mode != tskm_immediate_exec);
3655 
3656   // Toggle the th_task_state field, to switch which task_team this thread
3657   // refers to
3658   this_thr->th.th_task_state = (kmp_uint8)(1 - this_thr->th.th_task_state);
3659 
3660   // It is now safe to propagate the task team pointer from the team struct to
3661   // the current thread.
3662   TCW_PTR(this_thr->th.th_task_team,
3663           team->t.t_task_team[this_thr->th.th_task_state]);
3664   KA_TRACE(20,
3665            ("__kmp_task_team_sync: Thread T#%d task team switched to task_team "
3666             "%p from Team #%d (parity=%d)\n",
3667             __kmp_gtid_from_thread(this_thr), this_thr->th.th_task_team,
3668             team->t.t_id, this_thr->th.th_task_state));
3669 }
3670 
3671 // __kmp_task_team_wait: Primary thread waits for outstanding tasks after the
3672 // barrier gather phase. Only called by primary thread if #threads in team > 1
3673 // or if proxy tasks were created.
3674 //
3675 // wait is a flag that defaults to 1 (see kmp.h), but waiting can be turned off
3676 // by passing in 0 optionally as the last argument. When wait is zero, primary
3677 // thread does not wait for unfinished_threads to reach 0.
3678 void __kmp_task_team_wait(
3679     kmp_info_t *this_thr,
3680     kmp_team_t *team USE_ITT_BUILD_ARG(void *itt_sync_obj), int wait) {
3681   kmp_task_team_t *task_team = team->t.t_task_team[this_thr->th.th_task_state];
3682 
3683   KMP_DEBUG_ASSERT(__kmp_tasking_mode != tskm_immediate_exec);
3684   KMP_DEBUG_ASSERT(task_team == this_thr->th.th_task_team);
3685 
3686   if ((task_team != NULL) && KMP_TASKING_ENABLED(task_team)) {
3687     if (wait) {
3688       KA_TRACE(20, ("__kmp_task_team_wait: Primary T#%d waiting for all tasks "
3689                     "(for unfinished_threads to reach 0) on task_team = %p\n",
3690                     __kmp_gtid_from_thread(this_thr), task_team));
3691       // Worker threads may have dropped through to release phase, but could
3692       // still be executing tasks. Wait here for tasks to complete. To avoid
3693       // memory contention, only primary thread checks termination condition.
3694       kmp_flag_32<false, false> flag(
3695           RCAST(std::atomic<kmp_uint32> *,
3696                 &task_team->tt.tt_unfinished_threads),
3697           0U);
3698       flag.wait(this_thr, TRUE USE_ITT_BUILD_ARG(itt_sync_obj));
3699     }
3700     // Deactivate the old task team, so that the worker threads will stop
3701     // referencing it while spinning.
3702     KA_TRACE(
3703         20,
3704         ("__kmp_task_team_wait: Primary T#%d deactivating task_team %p: "
3705          "setting active to false, setting local and team's pointer to NULL\n",
3706          __kmp_gtid_from_thread(this_thr), task_team));
3707     KMP_DEBUG_ASSERT(task_team->tt.tt_nproc > 1 ||
3708                      task_team->tt.tt_found_proxy_tasks == TRUE);
3709     TCW_SYNC_4(task_team->tt.tt_found_proxy_tasks, FALSE);
3710     KMP_CHECK_UPDATE(task_team->tt.tt_untied_task_encountered, 0);
3711     TCW_SYNC_4(task_team->tt.tt_active, FALSE);
3712     KMP_MB();
3713 
3714     TCW_PTR(this_thr->th.th_task_team, NULL);
3715   }
3716 }
3717 
3718 // __kmp_tasking_barrier:
3719 // This routine is called only when __kmp_tasking_mode == tskm_extra_barrier.
3720 // Internal function to execute all tasks prior to a regular barrier or a join
3721 // barrier. It is a full barrier itself, which unfortunately turns regular
3722 // barriers into double barriers and join barriers into 1 1/2 barriers.
3723 void __kmp_tasking_barrier(kmp_team_t *team, kmp_info_t *thread, int gtid) {
3724   std::atomic<kmp_uint32> *spin = RCAST(
3725       std::atomic<kmp_uint32> *,
3726       &team->t.t_task_team[thread->th.th_task_state]->tt.tt_unfinished_threads);
3727   int flag = FALSE;
3728   KMP_DEBUG_ASSERT(__kmp_tasking_mode == tskm_extra_barrier);
3729 
3730 #if USE_ITT_BUILD
3731   KMP_FSYNC_SPIN_INIT(spin, NULL);
3732 #endif /* USE_ITT_BUILD */
3733   kmp_flag_32<false, false> spin_flag(spin, 0U);
3734   while (!spin_flag.execute_tasks(thread, gtid, TRUE,
3735                                   &flag USE_ITT_BUILD_ARG(NULL), 0)) {
3736 #if USE_ITT_BUILD
3737     // TODO: What about itt_sync_obj??
3738     KMP_FSYNC_SPIN_PREPARE(RCAST(void *, spin));
3739 #endif /* USE_ITT_BUILD */
3740 
3741     if (TCR_4(__kmp_global.g.g_done)) {
3742       if (__kmp_global.g.g_abort)
3743         __kmp_abort_thread();
3744       break;
3745     }
3746     KMP_YIELD(TRUE);
3747   }
3748 #if USE_ITT_BUILD
3749   KMP_FSYNC_SPIN_ACQUIRED(RCAST(void *, spin));
3750 #endif /* USE_ITT_BUILD */
3751 }
3752 
3753 // __kmp_give_task puts a task into a given thread queue if:
3754 //  - the queue for that thread was created
3755 //  - there's space in that queue
3756 // Because of this, __kmp_push_task needs to check if there's space after
3757 // getting the lock
3758 static bool __kmp_give_task(kmp_info_t *thread, kmp_int32 tid, kmp_task_t *task,
3759                             kmp_int32 pass) {
3760   kmp_taskdata_t *taskdata = KMP_TASK_TO_TASKDATA(task);
3761   kmp_task_team_t *task_team = taskdata->td_task_team;
3762 
3763   KA_TRACE(20, ("__kmp_give_task: trying to give task %p to thread %d.\n",
3764                 taskdata, tid));
3765 
3766   // If task_team is NULL something went really bad...
3767   KMP_DEBUG_ASSERT(task_team != NULL);
3768 
3769   bool result = false;
3770   kmp_thread_data_t *thread_data = &task_team->tt.tt_threads_data[tid];
3771 
3772   if (thread_data->td.td_deque == NULL) {
3773     // There's no queue in this thread, go find another one
3774     // We're guaranteed that at least one thread has a queue
3775     KA_TRACE(30,
3776              ("__kmp_give_task: thread %d has no queue while giving task %p.\n",
3777               tid, taskdata));
3778     return result;
3779   }
3780 
3781   if (TCR_4(thread_data->td.td_deque_ntasks) >=
3782       TASK_DEQUE_SIZE(thread_data->td)) {
3783     KA_TRACE(
3784         30,
3785         ("__kmp_give_task: queue is full while giving task %p to thread %d.\n",
3786          taskdata, tid));
3787 
3788     // if this deque is bigger than the pass ratio give a chance to another
3789     // thread
3790     if (TASK_DEQUE_SIZE(thread_data->td) / INITIAL_TASK_DEQUE_SIZE >= pass)
3791       return result;
3792 
3793     __kmp_acquire_bootstrap_lock(&thread_data->td.td_deque_lock);
3794     if (TCR_4(thread_data->td.td_deque_ntasks) >=
3795         TASK_DEQUE_SIZE(thread_data->td)) {
3796       // expand deque to push the task which is not allowed to execute
3797       __kmp_realloc_task_deque(thread, thread_data);
3798     }
3799 
3800   } else {
3801 
3802     __kmp_acquire_bootstrap_lock(&thread_data->td.td_deque_lock);
3803 
3804     if (TCR_4(thread_data->td.td_deque_ntasks) >=
3805         TASK_DEQUE_SIZE(thread_data->td)) {
3806       KA_TRACE(30, ("__kmp_give_task: queue is full while giving task %p to "
3807                     "thread %d.\n",
3808                     taskdata, tid));
3809 
3810       // if this deque is bigger than the pass ratio give a chance to another
3811       // thread
3812       if (TASK_DEQUE_SIZE(thread_data->td) / INITIAL_TASK_DEQUE_SIZE >= pass)
3813         goto release_and_exit;
3814 
3815       __kmp_realloc_task_deque(thread, thread_data);
3816     }
3817   }
3818 
3819   // lock is held here, and there is space in the deque
3820 
3821   thread_data->td.td_deque[thread_data->td.td_deque_tail] = taskdata;
3822   // Wrap index.
3823   thread_data->td.td_deque_tail =
3824       (thread_data->td.td_deque_tail + 1) & TASK_DEQUE_MASK(thread_data->td);
3825   TCW_4(thread_data->td.td_deque_ntasks,
3826         TCR_4(thread_data->td.td_deque_ntasks) + 1);
3827 
3828   result = true;
3829   KA_TRACE(30, ("__kmp_give_task: successfully gave task %p to thread %d.\n",
3830                 taskdata, tid));
3831 
3832 release_and_exit:
3833   __kmp_release_bootstrap_lock(&thread_data->td.td_deque_lock);
3834 
3835   return result;
3836 }
3837 
3838 #define PROXY_TASK_FLAG 0x40000000
3839 /* The finish of the proxy tasks is divided in two pieces:
3840     - the top half is the one that can be done from a thread outside the team
3841     - the bottom half must be run from a thread within the team
3842 
3843    In order to run the bottom half the task gets queued back into one of the
3844    threads of the team. Once the td_incomplete_child_task counter of the parent
3845    is decremented the threads can leave the barriers. So, the bottom half needs
3846    to be queued before the counter is decremented. The top half is therefore
3847    divided in two parts:
3848     - things that can be run before queuing the bottom half
3849     - things that must be run after queuing the bottom half
3850 
3851    This creates a second race as the bottom half can free the task before the
3852    second top half is executed. To avoid this we use the
3853    td_incomplete_child_task of the proxy task to synchronize the top and bottom
3854    half. */
3855 static void __kmp_first_top_half_finish_proxy(kmp_taskdata_t *taskdata) {
3856   KMP_DEBUG_ASSERT(taskdata->td_flags.tasktype == TASK_EXPLICIT);
3857   KMP_DEBUG_ASSERT(taskdata->td_flags.proxy == TASK_PROXY);
3858   KMP_DEBUG_ASSERT(taskdata->td_flags.complete == 0);
3859   KMP_DEBUG_ASSERT(taskdata->td_flags.freed == 0);
3860 
3861   taskdata->td_flags.complete = 1; // mark the task as completed
3862 
3863   if (taskdata->td_taskgroup)
3864     KMP_ATOMIC_DEC(&taskdata->td_taskgroup->count);
3865 
3866   // Create an imaginary children for this task so the bottom half cannot
3867   // release the task before we have completed the second top half
3868   KMP_ATOMIC_OR(&taskdata->td_incomplete_child_tasks, PROXY_TASK_FLAG);
3869 }
3870 
3871 static void __kmp_second_top_half_finish_proxy(kmp_taskdata_t *taskdata) {
3872   kmp_int32 children = 0;
3873 
3874   // Predecrement simulated by "- 1" calculation
3875   children =
3876       KMP_ATOMIC_DEC(&taskdata->td_parent->td_incomplete_child_tasks) - 1;
3877   KMP_DEBUG_ASSERT(children >= 0);
3878 
3879   // Remove the imaginary children
3880   KMP_ATOMIC_AND(&taskdata->td_incomplete_child_tasks, ~PROXY_TASK_FLAG);
3881 }
3882 
3883 static void __kmp_bottom_half_finish_proxy(kmp_int32 gtid, kmp_task_t *ptask) {
3884   kmp_taskdata_t *taskdata = KMP_TASK_TO_TASKDATA(ptask);
3885   kmp_info_t *thread = __kmp_threads[gtid];
3886 
3887   KMP_DEBUG_ASSERT(taskdata->td_flags.proxy == TASK_PROXY);
3888   KMP_DEBUG_ASSERT(taskdata->td_flags.complete ==
3889                    1); // top half must run before bottom half
3890 
3891   // We need to wait to make sure the top half is finished
3892   // Spinning here should be ok as this should happen quickly
3893   while ((KMP_ATOMIC_LD_ACQ(&taskdata->td_incomplete_child_tasks) &
3894           PROXY_TASK_FLAG) > 0)
3895     ;
3896 
3897   __kmp_release_deps(gtid, taskdata);
3898   __kmp_free_task_and_ancestors(gtid, taskdata, thread);
3899 }
3900 
3901 /*!
3902 @ingroup TASKING
3903 @param gtid Global Thread ID of encountering thread
3904 @param ptask Task which execution is completed
3905 
3906 Execute the completion of a proxy task from a thread of that is part of the
3907 team. Run first and bottom halves directly.
3908 */
3909 void __kmpc_proxy_task_completed(kmp_int32 gtid, kmp_task_t *ptask) {
3910   KMP_DEBUG_ASSERT(ptask != NULL);
3911   kmp_taskdata_t *taskdata = KMP_TASK_TO_TASKDATA(ptask);
3912   KA_TRACE(
3913       10, ("__kmp_proxy_task_completed(enter): T#%d proxy task %p completing\n",
3914            gtid, taskdata));
3915   __kmp_assert_valid_gtid(gtid);
3916   KMP_DEBUG_ASSERT(taskdata->td_flags.proxy == TASK_PROXY);
3917 
3918   __kmp_first_top_half_finish_proxy(taskdata);
3919   __kmp_second_top_half_finish_proxy(taskdata);
3920   __kmp_bottom_half_finish_proxy(gtid, ptask);
3921 
3922   KA_TRACE(10,
3923            ("__kmp_proxy_task_completed(exit): T#%d proxy task %p completing\n",
3924             gtid, taskdata));
3925 }
3926 
3927 void __kmpc_give_task(kmp_task_t *ptask, kmp_int32 start = 0) {
3928   KMP_DEBUG_ASSERT(ptask != NULL);
3929   kmp_taskdata_t *taskdata = KMP_TASK_TO_TASKDATA(ptask);
3930 
3931   // Enqueue task to complete bottom half completion from a thread within the
3932   // corresponding team
3933   kmp_team_t *team = taskdata->td_team;
3934   kmp_int32 nthreads = team->t.t_nproc;
3935   kmp_info_t *thread;
3936 
3937   // This should be similar to start_k = __kmp_get_random( thread ) % nthreads
3938   // but we cannot use __kmp_get_random here
3939   kmp_int32 start_k = start;
3940   kmp_int32 pass = 1;
3941   kmp_int32 k = start_k;
3942 
3943   do {
3944     // For now we're just linearly trying to find a thread
3945     thread = team->t.t_threads[k];
3946     k = (k + 1) % nthreads;
3947 
3948     // we did a full pass through all the threads
3949     if (k == start_k)
3950       pass = pass << 1;
3951 
3952   } while (!__kmp_give_task(thread, k, ptask, pass));
3953 }
3954 
3955 /*!
3956 @ingroup TASKING
3957 @param ptask Task which execution is completed
3958 
3959 Execute the completion of a proxy task from a thread that could not belong to
3960 the team.
3961 */
3962 void __kmpc_proxy_task_completed_ooo(kmp_task_t *ptask) {
3963   KMP_DEBUG_ASSERT(ptask != NULL);
3964   kmp_taskdata_t *taskdata = KMP_TASK_TO_TASKDATA(ptask);
3965 
3966   KA_TRACE(
3967       10,
3968       ("__kmp_proxy_task_completed_ooo(enter): proxy task completing ooo %p\n",
3969        taskdata));
3970 
3971   KMP_DEBUG_ASSERT(taskdata->td_flags.proxy == TASK_PROXY);
3972 
3973   __kmp_first_top_half_finish_proxy(taskdata);
3974 
3975   __kmpc_give_task(ptask);
3976 
3977   __kmp_second_top_half_finish_proxy(taskdata);
3978 
3979   KA_TRACE(
3980       10,
3981       ("__kmp_proxy_task_completed_ooo(exit): proxy task completing ooo %p\n",
3982        taskdata));
3983 }
3984 
3985 kmp_event_t *__kmpc_task_allow_completion_event(ident_t *loc_ref, int gtid,
3986                                                 kmp_task_t *task) {
3987   kmp_taskdata_t *td = KMP_TASK_TO_TASKDATA(task);
3988   if (td->td_allow_completion_event.type == KMP_EVENT_UNINITIALIZED) {
3989     td->td_allow_completion_event.type = KMP_EVENT_ALLOW_COMPLETION;
3990     td->td_allow_completion_event.ed.task = task;
3991     __kmp_init_tas_lock(&td->td_allow_completion_event.lock);
3992   }
3993   return &td->td_allow_completion_event;
3994 }
3995 
3996 void __kmp_fulfill_event(kmp_event_t *event) {
3997   if (event->type == KMP_EVENT_ALLOW_COMPLETION) {
3998     kmp_task_t *ptask = event->ed.task;
3999     kmp_taskdata_t *taskdata = KMP_TASK_TO_TASKDATA(ptask);
4000     bool detached = false;
4001     int gtid = __kmp_get_gtid();
4002 
4003     // The associated task might have completed or could be completing at this
4004     // point.
4005     // We need to take the lock to avoid races
4006     __kmp_acquire_tas_lock(&event->lock, gtid);
4007     if (taskdata->td_flags.proxy == TASK_PROXY) {
4008       detached = true;
4009     } else {
4010 #if OMPT_SUPPORT
4011       // The OMPT event must occur under mutual exclusion,
4012       // otherwise the tool might access ptask after free
4013       if (UNLIKELY(ompt_enabled.enabled))
4014         __ompt_task_finish(ptask, NULL, ompt_task_early_fulfill);
4015 #endif
4016     }
4017     event->type = KMP_EVENT_UNINITIALIZED;
4018     __kmp_release_tas_lock(&event->lock, gtid);
4019 
4020     if (detached) {
4021 #if OMPT_SUPPORT
4022       // We free ptask afterwards and know the task is finished,
4023       // so locking is not necessary
4024       if (UNLIKELY(ompt_enabled.enabled))
4025         __ompt_task_finish(ptask, NULL, ompt_task_late_fulfill);
4026 #endif
4027       // If the task detached complete the proxy task
4028       if (gtid >= 0) {
4029         kmp_team_t *team = taskdata->td_team;
4030         kmp_info_t *thread = __kmp_get_thread();
4031         if (thread->th.th_team == team) {
4032           __kmpc_proxy_task_completed(gtid, ptask);
4033           return;
4034         }
4035       }
4036 
4037       // fallback
4038       __kmpc_proxy_task_completed_ooo(ptask);
4039     }
4040   }
4041 }
4042 
4043 // __kmp_task_dup_alloc: Allocate the taskdata and make a copy of source task
4044 // for taskloop
4045 //
4046 // thread:   allocating thread
4047 // task_src: pointer to source task to be duplicated
4048 // returns:  a pointer to the allocated kmp_task_t structure (task).
4049 kmp_task_t *__kmp_task_dup_alloc(kmp_info_t *thread, kmp_task_t *task_src) {
4050   kmp_task_t *task;
4051   kmp_taskdata_t *taskdata;
4052   kmp_taskdata_t *taskdata_src = KMP_TASK_TO_TASKDATA(task_src);
4053   kmp_taskdata_t *parent_task = taskdata_src->td_parent; // same parent task
4054   size_t shareds_offset;
4055   size_t task_size;
4056 
4057   KA_TRACE(10, ("__kmp_task_dup_alloc(enter): Th %p, source task %p\n", thread,
4058                 task_src));
4059   KMP_DEBUG_ASSERT(taskdata_src->td_flags.proxy ==
4060                    TASK_FULL); // it should not be proxy task
4061   KMP_DEBUG_ASSERT(taskdata_src->td_flags.tasktype == TASK_EXPLICIT);
4062   task_size = taskdata_src->td_size_alloc;
4063 
4064   // Allocate a kmp_taskdata_t block and a kmp_task_t block.
4065   KA_TRACE(30, ("__kmp_task_dup_alloc: Th %p, malloc size %ld\n", thread,
4066                 task_size));
4067 #if USE_FAST_MEMORY
4068   taskdata = (kmp_taskdata_t *)__kmp_fast_allocate(thread, task_size);
4069 #else
4070   taskdata = (kmp_taskdata_t *)__kmp_thread_malloc(thread, task_size);
4071 #endif /* USE_FAST_MEMORY */
4072   KMP_MEMCPY(taskdata, taskdata_src, task_size);
4073 
4074   task = KMP_TASKDATA_TO_TASK(taskdata);
4075 
4076   // Initialize new task (only specific fields not affected by memcpy)
4077   taskdata->td_task_id = KMP_GEN_TASK_ID();
4078   if (task->shareds != NULL) { // need setup shareds pointer
4079     shareds_offset = (char *)task_src->shareds - (char *)taskdata_src;
4080     task->shareds = &((char *)taskdata)[shareds_offset];
4081     KMP_DEBUG_ASSERT((((kmp_uintptr_t)task->shareds) & (sizeof(void *) - 1)) ==
4082                      0);
4083   }
4084   taskdata->td_alloc_thread = thread;
4085   taskdata->td_parent = parent_task;
4086   // task inherits the taskgroup from the parent task
4087   taskdata->td_taskgroup = parent_task->td_taskgroup;
4088   // tied task needs to initialize the td_last_tied at creation,
4089   // untied one does this when it is scheduled for execution
4090   if (taskdata->td_flags.tiedness == TASK_TIED)
4091     taskdata->td_last_tied = taskdata;
4092 
4093   // Only need to keep track of child task counts if team parallel and tasking
4094   // not serialized
4095   if (!(taskdata->td_flags.team_serial || taskdata->td_flags.tasking_ser)) {
4096     KMP_ATOMIC_INC(&parent_task->td_incomplete_child_tasks);
4097     if (parent_task->td_taskgroup)
4098       KMP_ATOMIC_INC(&parent_task->td_taskgroup->count);
4099     // Only need to keep track of allocated child tasks for explicit tasks since
4100     // implicit not deallocated
4101     if (taskdata->td_parent->td_flags.tasktype == TASK_EXPLICIT)
4102       KMP_ATOMIC_INC(&taskdata->td_parent->td_allocated_child_tasks);
4103   }
4104 
4105   KA_TRACE(20,
4106            ("__kmp_task_dup_alloc(exit): Th %p, created task %p, parent=%p\n",
4107             thread, taskdata, taskdata->td_parent));
4108 #if OMPT_SUPPORT
4109   if (UNLIKELY(ompt_enabled.enabled))
4110     __ompt_task_init(taskdata, thread->th.th_info.ds.ds_gtid);
4111 #endif
4112   return task;
4113 }
4114 
4115 // Routine optionally generated by the compiler for setting the lastprivate flag
4116 // and calling needed constructors for private/firstprivate objects
4117 // (used to form taskloop tasks from pattern task)
4118 // Parameters: dest task, src task, lastprivate flag.
4119 typedef void (*p_task_dup_t)(kmp_task_t *, kmp_task_t *, kmp_int32);
4120 
4121 KMP_BUILD_ASSERT(sizeof(long) == 4 || sizeof(long) == 8);
4122 
4123 // class to encapsulate manipulating loop bounds in a taskloop task.
4124 // this abstracts away the Intel vs GOMP taskloop interface for setting/getting
4125 // the loop bound variables.
4126 class kmp_taskloop_bounds_t {
4127   kmp_task_t *task;
4128   const kmp_taskdata_t *taskdata;
4129   size_t lower_offset;
4130   size_t upper_offset;
4131 
4132 public:
4133   kmp_taskloop_bounds_t(kmp_task_t *_task, kmp_uint64 *lb, kmp_uint64 *ub)
4134       : task(_task), taskdata(KMP_TASK_TO_TASKDATA(task)),
4135         lower_offset((char *)lb - (char *)task),
4136         upper_offset((char *)ub - (char *)task) {
4137     KMP_DEBUG_ASSERT((char *)lb > (char *)_task);
4138     KMP_DEBUG_ASSERT((char *)ub > (char *)_task);
4139   }
4140   kmp_taskloop_bounds_t(kmp_task_t *_task, const kmp_taskloop_bounds_t &bounds)
4141       : task(_task), taskdata(KMP_TASK_TO_TASKDATA(_task)),
4142         lower_offset(bounds.lower_offset), upper_offset(bounds.upper_offset) {}
4143   size_t get_lower_offset() const { return lower_offset; }
4144   size_t get_upper_offset() const { return upper_offset; }
4145   kmp_uint64 get_lb() const {
4146     kmp_int64 retval;
4147 #if defined(KMP_GOMP_COMPAT)
4148     // Intel task just returns the lower bound normally
4149     if (!taskdata->td_flags.native) {
4150       retval = *(kmp_int64 *)((char *)task + lower_offset);
4151     } else {
4152       // GOMP task has to take into account the sizeof(long)
4153       if (taskdata->td_size_loop_bounds == 4) {
4154         kmp_int32 *lb = RCAST(kmp_int32 *, task->shareds);
4155         retval = (kmp_int64)*lb;
4156       } else {
4157         kmp_int64 *lb = RCAST(kmp_int64 *, task->shareds);
4158         retval = (kmp_int64)*lb;
4159       }
4160     }
4161 #else
4162     (void)taskdata;
4163     retval = *(kmp_int64 *)((char *)task + lower_offset);
4164 #endif // defined(KMP_GOMP_COMPAT)
4165     return retval;
4166   }
4167   kmp_uint64 get_ub() const {
4168     kmp_int64 retval;
4169 #if defined(KMP_GOMP_COMPAT)
4170     // Intel task just returns the upper bound normally
4171     if (!taskdata->td_flags.native) {
4172       retval = *(kmp_int64 *)((char *)task + upper_offset);
4173     } else {
4174       // GOMP task has to take into account the sizeof(long)
4175       if (taskdata->td_size_loop_bounds == 4) {
4176         kmp_int32 *ub = RCAST(kmp_int32 *, task->shareds) + 1;
4177         retval = (kmp_int64)*ub;
4178       } else {
4179         kmp_int64 *ub = RCAST(kmp_int64 *, task->shareds) + 1;
4180         retval = (kmp_int64)*ub;
4181       }
4182     }
4183 #else
4184     retval = *(kmp_int64 *)((char *)task + upper_offset);
4185 #endif // defined(KMP_GOMP_COMPAT)
4186     return retval;
4187   }
4188   void set_lb(kmp_uint64 lb) {
4189 #if defined(KMP_GOMP_COMPAT)
4190     // Intel task just sets the lower bound normally
4191     if (!taskdata->td_flags.native) {
4192       *(kmp_uint64 *)((char *)task + lower_offset) = lb;
4193     } else {
4194       // GOMP task has to take into account the sizeof(long)
4195       if (taskdata->td_size_loop_bounds == 4) {
4196         kmp_uint32 *lower = RCAST(kmp_uint32 *, task->shareds);
4197         *lower = (kmp_uint32)lb;
4198       } else {
4199         kmp_uint64 *lower = RCAST(kmp_uint64 *, task->shareds);
4200         *lower = (kmp_uint64)lb;
4201       }
4202     }
4203 #else
4204     *(kmp_uint64 *)((char *)task + lower_offset) = lb;
4205 #endif // defined(KMP_GOMP_COMPAT)
4206   }
4207   void set_ub(kmp_uint64 ub) {
4208 #if defined(KMP_GOMP_COMPAT)
4209     // Intel task just sets the upper bound normally
4210     if (!taskdata->td_flags.native) {
4211       *(kmp_uint64 *)((char *)task + upper_offset) = ub;
4212     } else {
4213       // GOMP task has to take into account the sizeof(long)
4214       if (taskdata->td_size_loop_bounds == 4) {
4215         kmp_uint32 *upper = RCAST(kmp_uint32 *, task->shareds) + 1;
4216         *upper = (kmp_uint32)ub;
4217       } else {
4218         kmp_uint64 *upper = RCAST(kmp_uint64 *, task->shareds) + 1;
4219         *upper = (kmp_uint64)ub;
4220       }
4221     }
4222 #else
4223     *(kmp_uint64 *)((char *)task + upper_offset) = ub;
4224 #endif // defined(KMP_GOMP_COMPAT)
4225   }
4226 };
4227 
4228 // __kmp_taskloop_linear: Start tasks of the taskloop linearly
4229 //
4230 // loc        Source location information
4231 // gtid       Global thread ID
4232 // task       Pattern task, exposes the loop iteration range
4233 // lb         Pointer to loop lower bound in task structure
4234 // ub         Pointer to loop upper bound in task structure
4235 // st         Loop stride
4236 // ub_glob    Global upper bound (used for lastprivate check)
4237 // num_tasks  Number of tasks to execute
4238 // grainsize  Number of loop iterations per task
4239 // extras     Number of chunks with grainsize+1 iterations
4240 // last_chunk Reduction of grainsize for last task
4241 // tc         Iterations count
4242 // task_dup   Tasks duplication routine
4243 // codeptr_ra Return address for OMPT events
4244 void __kmp_taskloop_linear(ident_t *loc, int gtid, kmp_task_t *task,
4245                            kmp_uint64 *lb, kmp_uint64 *ub, kmp_int64 st,
4246                            kmp_uint64 ub_glob, kmp_uint64 num_tasks,
4247                            kmp_uint64 grainsize, kmp_uint64 extras,
4248                            kmp_int64 last_chunk, kmp_uint64 tc,
4249 #if OMPT_SUPPORT
4250                            void *codeptr_ra,
4251 #endif
4252                            void *task_dup) {
4253   KMP_COUNT_BLOCK(OMP_TASKLOOP);
4254   KMP_TIME_PARTITIONED_BLOCK(OMP_taskloop_scheduling);
4255   p_task_dup_t ptask_dup = (p_task_dup_t)task_dup;
4256   // compiler provides global bounds here
4257   kmp_taskloop_bounds_t task_bounds(task, lb, ub);
4258   kmp_uint64 lower = task_bounds.get_lb();
4259   kmp_uint64 upper = task_bounds.get_ub();
4260   kmp_uint64 i;
4261   kmp_info_t *thread = __kmp_threads[gtid];
4262   kmp_taskdata_t *current_task = thread->th.th_current_task;
4263   kmp_task_t *next_task;
4264   kmp_int32 lastpriv = 0;
4265 
4266   KMP_DEBUG_ASSERT(tc == num_tasks * grainsize +
4267                              (last_chunk < 0 ? last_chunk : extras));
4268   KMP_DEBUG_ASSERT(num_tasks > extras);
4269   KMP_DEBUG_ASSERT(num_tasks > 0);
4270   KA_TRACE(20, ("__kmp_taskloop_linear: T#%d: %lld tasks, grainsize %lld, "
4271                 "extras %lld, last_chunk %lld, i=%lld,%lld(%d)%lld, dup %p\n",
4272                 gtid, num_tasks, grainsize, extras, last_chunk, lower, upper,
4273                 ub_glob, st, task_dup));
4274 
4275   // Launch num_tasks tasks, assign grainsize iterations each task
4276   for (i = 0; i < num_tasks; ++i) {
4277     kmp_uint64 chunk_minus_1;
4278     if (extras == 0) {
4279       chunk_minus_1 = grainsize - 1;
4280     } else {
4281       chunk_minus_1 = grainsize;
4282       --extras; // first extras iterations get bigger chunk (grainsize+1)
4283     }
4284     upper = lower + st * chunk_minus_1;
4285     if (upper > *ub) {
4286       upper = *ub;
4287     }
4288     if (i == num_tasks - 1) {
4289       // schedule the last task, set lastprivate flag if needed
4290       if (st == 1) { // most common case
4291         KMP_DEBUG_ASSERT(upper == *ub);
4292         if (upper == ub_glob)
4293           lastpriv = 1;
4294       } else if (st > 0) { // positive loop stride
4295         KMP_DEBUG_ASSERT((kmp_uint64)st > *ub - upper);
4296         if ((kmp_uint64)st > ub_glob - upper)
4297           lastpriv = 1;
4298       } else { // negative loop stride
4299         KMP_DEBUG_ASSERT(upper + st < *ub);
4300         if (upper - ub_glob < (kmp_uint64)(-st))
4301           lastpriv = 1;
4302       }
4303     }
4304     next_task = __kmp_task_dup_alloc(thread, task); // allocate new task
4305     kmp_taskdata_t *next_taskdata = KMP_TASK_TO_TASKDATA(next_task);
4306     kmp_taskloop_bounds_t next_task_bounds =
4307         kmp_taskloop_bounds_t(next_task, task_bounds);
4308 
4309     // adjust task-specific bounds
4310     next_task_bounds.set_lb(lower);
4311     if (next_taskdata->td_flags.native) {
4312       next_task_bounds.set_ub(upper + (st > 0 ? 1 : -1));
4313     } else {
4314       next_task_bounds.set_ub(upper);
4315     }
4316     if (ptask_dup != NULL) // set lastprivate flag, construct firstprivates,
4317                            // etc.
4318       ptask_dup(next_task, task, lastpriv);
4319     KA_TRACE(40,
4320              ("__kmp_taskloop_linear: T#%d; task #%llu: task %p: lower %lld, "
4321               "upper %lld stride %lld, (offsets %p %p)\n",
4322               gtid, i, next_task, lower, upper, st,
4323               next_task_bounds.get_lower_offset(),
4324               next_task_bounds.get_upper_offset()));
4325 #if OMPT_SUPPORT
4326     __kmp_omp_taskloop_task(NULL, gtid, next_task,
4327                             codeptr_ra); // schedule new task
4328 #else
4329     __kmp_omp_task(gtid, next_task, true); // schedule new task
4330 #endif
4331     lower = upper + st; // adjust lower bound for the next iteration
4332   }
4333   // free the pattern task and exit
4334   __kmp_task_start(gtid, task, current_task); // make internal bookkeeping
4335   // do not execute the pattern task, just do internal bookkeeping
4336   __kmp_task_finish<false>(gtid, task, current_task);
4337 }
4338 
4339 // Structure to keep taskloop parameters for auxiliary task
4340 // kept in the shareds of the task structure.
4341 typedef struct __taskloop_params {
4342   kmp_task_t *task;
4343   kmp_uint64 *lb;
4344   kmp_uint64 *ub;
4345   void *task_dup;
4346   kmp_int64 st;
4347   kmp_uint64 ub_glob;
4348   kmp_uint64 num_tasks;
4349   kmp_uint64 grainsize;
4350   kmp_uint64 extras;
4351   kmp_int64 last_chunk;
4352   kmp_uint64 tc;
4353   kmp_uint64 num_t_min;
4354 #if OMPT_SUPPORT
4355   void *codeptr_ra;
4356 #endif
4357 } __taskloop_params_t;
4358 
4359 void __kmp_taskloop_recur(ident_t *, int, kmp_task_t *, kmp_uint64 *,
4360                           kmp_uint64 *, kmp_int64, kmp_uint64, kmp_uint64,
4361                           kmp_uint64, kmp_uint64, kmp_int64, kmp_uint64,
4362                           kmp_uint64,
4363 #if OMPT_SUPPORT
4364                           void *,
4365 #endif
4366                           void *);
4367 
4368 // Execute part of the taskloop submitted as a task.
4369 int __kmp_taskloop_task(int gtid, void *ptask) {
4370   __taskloop_params_t *p =
4371       (__taskloop_params_t *)((kmp_task_t *)ptask)->shareds;
4372   kmp_task_t *task = p->task;
4373   kmp_uint64 *lb = p->lb;
4374   kmp_uint64 *ub = p->ub;
4375   void *task_dup = p->task_dup;
4376   //  p_task_dup_t ptask_dup = (p_task_dup_t)task_dup;
4377   kmp_int64 st = p->st;
4378   kmp_uint64 ub_glob = p->ub_glob;
4379   kmp_uint64 num_tasks = p->num_tasks;
4380   kmp_uint64 grainsize = p->grainsize;
4381   kmp_uint64 extras = p->extras;
4382   kmp_int64 last_chunk = p->last_chunk;
4383   kmp_uint64 tc = p->tc;
4384   kmp_uint64 num_t_min = p->num_t_min;
4385 #if OMPT_SUPPORT
4386   void *codeptr_ra = p->codeptr_ra;
4387 #endif
4388 #if KMP_DEBUG
4389   kmp_taskdata_t *taskdata = KMP_TASK_TO_TASKDATA(task);
4390   KMP_DEBUG_ASSERT(task != NULL);
4391   KA_TRACE(20,
4392            ("__kmp_taskloop_task: T#%d, task %p: %lld tasks, grainsize"
4393             " %lld, extras %lld, last_chunk %lld, i=%lld,%lld(%d), dup %p\n",
4394             gtid, taskdata, num_tasks, grainsize, extras, last_chunk, *lb, *ub,
4395             st, task_dup));
4396 #endif
4397   KMP_DEBUG_ASSERT(num_tasks * 2 + 1 > num_t_min);
4398   if (num_tasks > num_t_min)
4399     __kmp_taskloop_recur(NULL, gtid, task, lb, ub, st, ub_glob, num_tasks,
4400                          grainsize, extras, last_chunk, tc, num_t_min,
4401 #if OMPT_SUPPORT
4402                          codeptr_ra,
4403 #endif
4404                          task_dup);
4405   else
4406     __kmp_taskloop_linear(NULL, gtid, task, lb, ub, st, ub_glob, num_tasks,
4407                           grainsize, extras, last_chunk, tc,
4408 #if OMPT_SUPPORT
4409                           codeptr_ra,
4410 #endif
4411                           task_dup);
4412 
4413   KA_TRACE(40, ("__kmp_taskloop_task(exit): T#%d\n", gtid));
4414   return 0;
4415 }
4416 
4417 // Schedule part of the taskloop as a task,
4418 // execute the rest of the taskloop.
4419 //
4420 // loc        Source location information
4421 // gtid       Global thread ID
4422 // task       Pattern task, exposes the loop iteration range
4423 // lb         Pointer to loop lower bound in task structure
4424 // ub         Pointer to loop upper bound in task structure
4425 // st         Loop stride
4426 // ub_glob    Global upper bound (used for lastprivate check)
4427 // num_tasks  Number of tasks to execute
4428 // grainsize  Number of loop iterations per task
4429 // extras     Number of chunks with grainsize+1 iterations
4430 // last_chunk Reduction of grainsize for last task
4431 // tc         Iterations count
4432 // num_t_min  Threshold to launch tasks recursively
4433 // task_dup   Tasks duplication routine
4434 // codeptr_ra Return address for OMPT events
4435 void __kmp_taskloop_recur(ident_t *loc, int gtid, kmp_task_t *task,
4436                           kmp_uint64 *lb, kmp_uint64 *ub, kmp_int64 st,
4437                           kmp_uint64 ub_glob, kmp_uint64 num_tasks,
4438                           kmp_uint64 grainsize, kmp_uint64 extras,
4439                           kmp_int64 last_chunk, kmp_uint64 tc,
4440                           kmp_uint64 num_t_min,
4441 #if OMPT_SUPPORT
4442                           void *codeptr_ra,
4443 #endif
4444                           void *task_dup) {
4445   kmp_taskdata_t *taskdata = KMP_TASK_TO_TASKDATA(task);
4446   KMP_DEBUG_ASSERT(task != NULL);
4447   KMP_DEBUG_ASSERT(num_tasks > num_t_min);
4448   KA_TRACE(20,
4449            ("__kmp_taskloop_recur: T#%d, task %p: %lld tasks, grainsize"
4450             " %lld, extras %lld, last_chunk %lld, i=%lld,%lld(%d), dup %p\n",
4451             gtid, taskdata, num_tasks, grainsize, extras, last_chunk, *lb, *ub,
4452             st, task_dup));
4453   p_task_dup_t ptask_dup = (p_task_dup_t)task_dup;
4454   kmp_uint64 lower = *lb;
4455   kmp_info_t *thread = __kmp_threads[gtid];
4456   //  kmp_taskdata_t *current_task = thread->th.th_current_task;
4457   kmp_task_t *next_task;
4458   size_t lower_offset =
4459       (char *)lb - (char *)task; // remember offset of lb in the task structure
4460   size_t upper_offset =
4461       (char *)ub - (char *)task; // remember offset of ub in the task structure
4462 
4463   KMP_DEBUG_ASSERT(tc == num_tasks * grainsize +
4464                              (last_chunk < 0 ? last_chunk : extras));
4465   KMP_DEBUG_ASSERT(num_tasks > extras);
4466   KMP_DEBUG_ASSERT(num_tasks > 0);
4467 
4468   // split the loop in two halves
4469   kmp_uint64 lb1, ub0, tc0, tc1, ext0, ext1;
4470   kmp_int64 last_chunk0 = 0, last_chunk1 = 0;
4471   kmp_uint64 gr_size0 = grainsize;
4472   kmp_uint64 n_tsk0 = num_tasks >> 1; // num_tasks/2 to execute
4473   kmp_uint64 n_tsk1 = num_tasks - n_tsk0; // to schedule as a task
4474   if (last_chunk < 0) {
4475     ext0 = ext1 = 0;
4476     last_chunk1 = last_chunk;
4477     tc0 = grainsize * n_tsk0;
4478     tc1 = tc - tc0;
4479   } else if (n_tsk0 <= extras) {
4480     gr_size0++; // integrate extras into grainsize
4481     ext0 = 0; // no extra iters in 1st half
4482     ext1 = extras - n_tsk0; // remaining extras
4483     tc0 = gr_size0 * n_tsk0;
4484     tc1 = tc - tc0;
4485   } else { // n_tsk0 > extras
4486     ext1 = 0; // no extra iters in 2nd half
4487     ext0 = extras;
4488     tc1 = grainsize * n_tsk1;
4489     tc0 = tc - tc1;
4490   }
4491   ub0 = lower + st * (tc0 - 1);
4492   lb1 = ub0 + st;
4493 
4494   // create pattern task for 2nd half of the loop
4495   next_task = __kmp_task_dup_alloc(thread, task); // duplicate the task
4496   // adjust lower bound (upper bound is not changed) for the 2nd half
4497   *(kmp_uint64 *)((char *)next_task + lower_offset) = lb1;
4498   if (ptask_dup != NULL) // construct firstprivates, etc.
4499     ptask_dup(next_task, task, 0);
4500   *ub = ub0; // adjust upper bound for the 1st half
4501 
4502   // create auxiliary task for 2nd half of the loop
4503   // make sure new task has same parent task as the pattern task
4504   kmp_taskdata_t *current_task = thread->th.th_current_task;
4505   thread->th.th_current_task = taskdata->td_parent;
4506   kmp_task_t *new_task =
4507       __kmpc_omp_task_alloc(loc, gtid, 1, 3 * sizeof(void *),
4508                             sizeof(__taskloop_params_t), &__kmp_taskloop_task);
4509   // restore current task
4510   thread->th.th_current_task = current_task;
4511   __taskloop_params_t *p = (__taskloop_params_t *)new_task->shareds;
4512   p->task = next_task;
4513   p->lb = (kmp_uint64 *)((char *)next_task + lower_offset);
4514   p->ub = (kmp_uint64 *)((char *)next_task + upper_offset);
4515   p->task_dup = task_dup;
4516   p->st = st;
4517   p->ub_glob = ub_glob;
4518   p->num_tasks = n_tsk1;
4519   p->grainsize = grainsize;
4520   p->extras = ext1;
4521   p->last_chunk = last_chunk1;
4522   p->tc = tc1;
4523   p->num_t_min = num_t_min;
4524 #if OMPT_SUPPORT
4525   p->codeptr_ra = codeptr_ra;
4526 #endif
4527 
4528 #if OMPT_SUPPORT
4529   // schedule new task with correct return address for OMPT events
4530   __kmp_omp_taskloop_task(NULL, gtid, new_task, codeptr_ra);
4531 #else
4532   __kmp_omp_task(gtid, new_task, true); // schedule new task
4533 #endif
4534 
4535   // execute the 1st half of current subrange
4536   if (n_tsk0 > num_t_min)
4537     __kmp_taskloop_recur(loc, gtid, task, lb, ub, st, ub_glob, n_tsk0, gr_size0,
4538                          ext0, last_chunk0, tc0, num_t_min,
4539 #if OMPT_SUPPORT
4540                          codeptr_ra,
4541 #endif
4542                          task_dup);
4543   else
4544     __kmp_taskloop_linear(loc, gtid, task, lb, ub, st, ub_glob, n_tsk0,
4545                           gr_size0, ext0, last_chunk0, tc0,
4546 #if OMPT_SUPPORT
4547                           codeptr_ra,
4548 #endif
4549                           task_dup);
4550 
4551   KA_TRACE(40, ("__kmp_taskloop_recur(exit): T#%d\n", gtid));
4552 }
4553 
4554 static void __kmp_taskloop(ident_t *loc, int gtid, kmp_task_t *task, int if_val,
4555                            kmp_uint64 *lb, kmp_uint64 *ub, kmp_int64 st,
4556                            int nogroup, int sched, kmp_uint64 grainsize,
4557                            int modifier, void *task_dup) {
4558   kmp_taskdata_t *taskdata = KMP_TASK_TO_TASKDATA(task);
4559   KMP_DEBUG_ASSERT(task != NULL);
4560   if (nogroup == 0) {
4561 #if OMPT_SUPPORT && OMPT_OPTIONAL
4562     OMPT_STORE_RETURN_ADDRESS(gtid);
4563 #endif
4564     __kmpc_taskgroup(loc, gtid);
4565   }
4566 
4567   // =========================================================================
4568   // calculate loop parameters
4569   kmp_taskloop_bounds_t task_bounds(task, lb, ub);
4570   kmp_uint64 tc;
4571   // compiler provides global bounds here
4572   kmp_uint64 lower = task_bounds.get_lb();
4573   kmp_uint64 upper = task_bounds.get_ub();
4574   kmp_uint64 ub_glob = upper; // global upper used to calc lastprivate flag
4575   kmp_uint64 num_tasks = 0, extras = 0;
4576   kmp_int64 last_chunk =
4577       0; // reduce grainsize of last task by last_chunk in strict mode
4578   kmp_uint64 num_tasks_min = __kmp_taskloop_min_tasks;
4579   kmp_info_t *thread = __kmp_threads[gtid];
4580   kmp_taskdata_t *current_task = thread->th.th_current_task;
4581 
4582   KA_TRACE(20, ("__kmp_taskloop: T#%d, task %p, lb %lld, ub %lld, st %lld, "
4583                 "grain %llu(%d, %d), dup %p\n",
4584                 gtid, taskdata, lower, upper, st, grainsize, sched, modifier,
4585                 task_dup));
4586 
4587   // compute trip count
4588   if (st == 1) { // most common case
4589     tc = upper - lower + 1;
4590   } else if (st < 0) {
4591     tc = (lower - upper) / (-st) + 1;
4592   } else { // st > 0
4593     tc = (upper - lower) / st + 1;
4594   }
4595   if (tc == 0) {
4596     KA_TRACE(20, ("__kmp_taskloop(exit): T#%d zero-trip loop\n", gtid));
4597     // free the pattern task and exit
4598     __kmp_task_start(gtid, task, current_task);
4599     // do not execute anything for zero-trip loop
4600     __kmp_task_finish<false>(gtid, task, current_task);
4601     return;
4602   }
4603 
4604 #if OMPT_SUPPORT && OMPT_OPTIONAL
4605   ompt_team_info_t *team_info = __ompt_get_teaminfo(0, NULL);
4606   ompt_task_info_t *task_info = __ompt_get_task_info_object(0);
4607   if (ompt_enabled.ompt_callback_work) {
4608     ompt_callbacks.ompt_callback(ompt_callback_work)(
4609         ompt_work_taskloop, ompt_scope_begin, &(team_info->parallel_data),
4610         &(task_info->task_data), tc, OMPT_GET_RETURN_ADDRESS(0));
4611   }
4612 #endif
4613 
4614   if (num_tasks_min == 0)
4615     // TODO: can we choose better default heuristic?
4616     num_tasks_min =
4617         KMP_MIN(thread->th.th_team_nproc * 10, INITIAL_TASK_DEQUE_SIZE);
4618 
4619   // compute num_tasks/grainsize based on the input provided
4620   switch (sched) {
4621   case 0: // no schedule clause specified, we can choose the default
4622     // let's try to schedule (team_size*10) tasks
4623     grainsize = thread->th.th_team_nproc * 10;
4624     KMP_FALLTHROUGH();
4625   case 2: // num_tasks provided
4626     if (grainsize > tc) {
4627       num_tasks = tc; // too big num_tasks requested, adjust values
4628       grainsize = 1;
4629       extras = 0;
4630     } else {
4631       num_tasks = grainsize;
4632       grainsize = tc / num_tasks;
4633       extras = tc % num_tasks;
4634     }
4635     break;
4636   case 1: // grainsize provided
4637     if (grainsize > tc) {
4638       num_tasks = 1;
4639       grainsize = tc; // too big grainsize requested, adjust values
4640       extras = 0;
4641     } else {
4642       if (modifier) {
4643         num_tasks = (tc + grainsize - 1) / grainsize;
4644         last_chunk = tc - (num_tasks * grainsize);
4645         extras = 0;
4646       } else {
4647         num_tasks = tc / grainsize;
4648         // adjust grainsize for balanced distribution of iterations
4649         grainsize = tc / num_tasks;
4650         extras = tc % num_tasks;
4651       }
4652     }
4653     break;
4654   default:
4655     KMP_ASSERT2(0, "unknown scheduling of taskloop");
4656   }
4657 
4658   KMP_DEBUG_ASSERT(tc == num_tasks * grainsize +
4659                              (last_chunk < 0 ? last_chunk : extras));
4660   KMP_DEBUG_ASSERT(num_tasks > extras);
4661   KMP_DEBUG_ASSERT(num_tasks > 0);
4662   // =========================================================================
4663 
4664   // check if clause value first
4665   // Also require GOMP_taskloop to reduce to linear (taskdata->td_flags.native)
4666   if (if_val == 0) { // if(0) specified, mark task as serial
4667     taskdata->td_flags.task_serial = 1;
4668     taskdata->td_flags.tiedness = TASK_TIED; // AC: serial task cannot be untied
4669     // always start serial tasks linearly
4670     __kmp_taskloop_linear(loc, gtid, task, lb, ub, st, ub_glob, num_tasks,
4671                           grainsize, extras, last_chunk, tc,
4672 #if OMPT_SUPPORT
4673                           OMPT_GET_RETURN_ADDRESS(0),
4674 #endif
4675                           task_dup);
4676     // !taskdata->td_flags.native => currently force linear spawning of tasks
4677     // for GOMP_taskloop
4678   } else if (num_tasks > num_tasks_min && !taskdata->td_flags.native) {
4679     KA_TRACE(20, ("__kmp_taskloop: T#%d, go recursive: tc %llu, #tasks %llu"
4680                   "(%lld), grain %llu, extras %llu, last_chunk %lld\n",
4681                   gtid, tc, num_tasks, num_tasks_min, grainsize, extras,
4682                   last_chunk));
4683     __kmp_taskloop_recur(loc, gtid, task, lb, ub, st, ub_glob, num_tasks,
4684                          grainsize, extras, last_chunk, tc, num_tasks_min,
4685 #if OMPT_SUPPORT
4686                          OMPT_GET_RETURN_ADDRESS(0),
4687 #endif
4688                          task_dup);
4689   } else {
4690     KA_TRACE(20, ("__kmp_taskloop: T#%d, go linear: tc %llu, #tasks %llu"
4691                   "(%lld), grain %llu, extras %llu, last_chunk %lld\n",
4692                   gtid, tc, num_tasks, num_tasks_min, grainsize, extras,
4693                   last_chunk));
4694     __kmp_taskloop_linear(loc, gtid, task, lb, ub, st, ub_glob, num_tasks,
4695                           grainsize, extras, last_chunk, tc,
4696 #if OMPT_SUPPORT
4697                           OMPT_GET_RETURN_ADDRESS(0),
4698 #endif
4699                           task_dup);
4700   }
4701 
4702 #if OMPT_SUPPORT && OMPT_OPTIONAL
4703   if (ompt_enabled.ompt_callback_work) {
4704     ompt_callbacks.ompt_callback(ompt_callback_work)(
4705         ompt_work_taskloop, ompt_scope_end, &(team_info->parallel_data),
4706         &(task_info->task_data), tc, OMPT_GET_RETURN_ADDRESS(0));
4707   }
4708 #endif
4709 
4710   if (nogroup == 0) {
4711 #if OMPT_SUPPORT && OMPT_OPTIONAL
4712     OMPT_STORE_RETURN_ADDRESS(gtid);
4713 #endif
4714     __kmpc_end_taskgroup(loc, gtid);
4715   }
4716   KA_TRACE(20, ("__kmp_taskloop(exit): T#%d\n", gtid));
4717 }
4718 
4719 /*!
4720 @ingroup TASKING
4721 @param loc       Source location information
4722 @param gtid      Global thread ID
4723 @param task      Task structure
4724 @param if_val    Value of the if clause
4725 @param lb        Pointer to loop lower bound in task structure
4726 @param ub        Pointer to loop upper bound in task structure
4727 @param st        Loop stride
4728 @param nogroup   Flag, 1 if nogroup clause specified, 0 otherwise
4729 @param sched     Schedule specified 0/1/2 for none/grainsize/num_tasks
4730 @param grainsize Schedule value if specified
4731 @param task_dup  Tasks duplication routine
4732 
4733 Execute the taskloop construct.
4734 */
4735 void __kmpc_taskloop(ident_t *loc, int gtid, kmp_task_t *task, int if_val,
4736                      kmp_uint64 *lb, kmp_uint64 *ub, kmp_int64 st, int nogroup,
4737                      int sched, kmp_uint64 grainsize, void *task_dup) {
4738   __kmp_assert_valid_gtid(gtid);
4739   KA_TRACE(20, ("__kmpc_taskloop(enter): T#%d\n", gtid));
4740   __kmp_taskloop(loc, gtid, task, if_val, lb, ub, st, nogroup, sched, grainsize,
4741                  0, task_dup);
4742   KA_TRACE(20, ("__kmpc_taskloop(exit): T#%d\n", gtid));
4743 }
4744 
4745 /*!
4746 @ingroup TASKING
4747 @param loc       Source location information
4748 @param gtid      Global thread ID
4749 @param task      Task structure
4750 @param if_val    Value of the if clause
4751 @param lb        Pointer to loop lower bound in task structure
4752 @param ub        Pointer to loop upper bound in task structure
4753 @param st        Loop stride
4754 @param nogroup   Flag, 1 if nogroup clause specified, 0 otherwise
4755 @param sched     Schedule specified 0/1/2 for none/grainsize/num_tasks
4756 @param grainsize Schedule value if specified
4757 @param modifer   Modifier 'strict' for sched, 1 if present, 0 otherwise
4758 @param task_dup  Tasks duplication routine
4759 
4760 Execute the taskloop construct.
4761 */
4762 void __kmpc_taskloop_5(ident_t *loc, int gtid, kmp_task_t *task, int if_val,
4763                        kmp_uint64 *lb, kmp_uint64 *ub, kmp_int64 st,
4764                        int nogroup, int sched, kmp_uint64 grainsize,
4765                        int modifier, void *task_dup) {
4766   __kmp_assert_valid_gtid(gtid);
4767   KA_TRACE(20, ("__kmpc_taskloop_5(enter): T#%d\n", gtid));
4768   __kmp_taskloop(loc, gtid, task, if_val, lb, ub, st, nogroup, sched, grainsize,
4769                  modifier, task_dup);
4770   KA_TRACE(20, ("__kmpc_taskloop_5(exit): T#%d\n", gtid));
4771 }
4772