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 #include "tsan_annotations.h"
25 
26 /* forward declaration */
27 static void __kmp_enable_tasking(kmp_task_team_t *task_team,
28                                  kmp_info_t *this_thr);
29 static void __kmp_alloc_task_deque(kmp_info_t *thread,
30                                    kmp_thread_data_t *thread_data);
31 static int __kmp_realloc_task_threads_data(kmp_info_t *thread,
32                                            kmp_task_team_t *task_team);
33 
34 #if OMP_45_ENABLED
35 static void __kmp_bottom_half_finish_proxy(kmp_int32 gtid, kmp_task_t *ptask);
36 #endif
37 
38 #ifdef BUILD_TIED_TASK_STACK
39 
40 //  __kmp_trace_task_stack: print the tied tasks from the task stack in order
41 //  from top do bottom
42 //
43 //  gtid: global thread identifier for thread containing stack
44 //  thread_data: thread data for task team thread containing stack
45 //  threshold: value above which the trace statement triggers
46 //  location: string identifying call site of this function (for trace)
47 static void __kmp_trace_task_stack(kmp_int32 gtid,
48                                    kmp_thread_data_t *thread_data,
49                                    int threshold, char *location) {
50   kmp_task_stack_t *task_stack = &thread_data->td.td_susp_tied_tasks;
51   kmp_taskdata_t **stack_top = task_stack->ts_top;
52   kmp_int32 entries = task_stack->ts_entries;
53   kmp_taskdata_t *tied_task;
54 
55   KA_TRACE(
56       threshold,
57       ("__kmp_trace_task_stack(start): location = %s, gtid = %d, entries = %d, "
58        "first_block = %p, stack_top = %p \n",
59        location, gtid, entries, task_stack->ts_first_block, stack_top));
60 
61   KMP_DEBUG_ASSERT(stack_top != NULL);
62   KMP_DEBUG_ASSERT(entries > 0);
63 
64   while (entries != 0) {
65     KMP_DEBUG_ASSERT(stack_top != &task_stack->ts_first_block.sb_block[0]);
66     // fix up ts_top if we need to pop from previous block
67     if (entries & TASK_STACK_INDEX_MASK == 0) {
68       kmp_stack_block_t *stack_block = (kmp_stack_block_t *)(stack_top);
69 
70       stack_block = stack_block->sb_prev;
71       stack_top = &stack_block->sb_block[TASK_STACK_BLOCK_SIZE];
72     }
73 
74     // finish bookkeeping
75     stack_top--;
76     entries--;
77 
78     tied_task = *stack_top;
79 
80     KMP_DEBUG_ASSERT(tied_task != NULL);
81     KMP_DEBUG_ASSERT(tied_task->td_flags.tasktype == TASK_TIED);
82 
83     KA_TRACE(threshold,
84              ("__kmp_trace_task_stack(%s):             gtid=%d, entry=%d, "
85               "stack_top=%p, tied_task=%p\n",
86               location, gtid, entries, stack_top, tied_task));
87   }
88   KMP_DEBUG_ASSERT(stack_top == &task_stack->ts_first_block.sb_block[0]);
89 
90   KA_TRACE(threshold,
91            ("__kmp_trace_task_stack(exit): location = %s, gtid = %d\n",
92             location, gtid));
93 }
94 
95 //  __kmp_init_task_stack: initialize the task stack for the first time
96 //  after a thread_data structure is created.
97 //  It should not be necessary to do this again (assuming the stack works).
98 //
99 //  gtid: global thread identifier of calling thread
100 //  thread_data: thread data for task team thread containing stack
101 static void __kmp_init_task_stack(kmp_int32 gtid,
102                                   kmp_thread_data_t *thread_data) {
103   kmp_task_stack_t *task_stack = &thread_data->td.td_susp_tied_tasks;
104   kmp_stack_block_t *first_block;
105 
106   // set up the first block of the stack
107   first_block = &task_stack->ts_first_block;
108   task_stack->ts_top = (kmp_taskdata_t **)first_block;
109   memset((void *)first_block, '\0',
110          TASK_STACK_BLOCK_SIZE * sizeof(kmp_taskdata_t *));
111 
112   // initialize the stack to be empty
113   task_stack->ts_entries = TASK_STACK_EMPTY;
114   first_block->sb_next = NULL;
115   first_block->sb_prev = NULL;
116 }
117 
118 //  __kmp_free_task_stack: free the task stack when thread_data is destroyed.
119 //
120 //  gtid: global thread identifier for calling thread
121 //  thread_data: thread info for thread containing stack
122 static void __kmp_free_task_stack(kmp_int32 gtid,
123                                   kmp_thread_data_t *thread_data) {
124   kmp_task_stack_t *task_stack = &thread_data->td.td_susp_tied_tasks;
125   kmp_stack_block_t *stack_block = &task_stack->ts_first_block;
126 
127   KMP_DEBUG_ASSERT(task_stack->ts_entries == TASK_STACK_EMPTY);
128   // free from the second block of the stack
129   while (stack_block != NULL) {
130     kmp_stack_block_t *next_block = (stack_block) ? stack_block->sb_next : NULL;
131 
132     stack_block->sb_next = NULL;
133     stack_block->sb_prev = NULL;
134     if (stack_block != &task_stack->ts_first_block) {
135       __kmp_thread_free(thread,
136                         stack_block); // free the block, if not the first
137     }
138     stack_block = next_block;
139   }
140   // initialize the stack to be empty
141   task_stack->ts_entries = 0;
142   task_stack->ts_top = NULL;
143 }
144 
145 //  __kmp_push_task_stack: Push the tied task onto the task stack.
146 //     Grow the stack if necessary by allocating another block.
147 //
148 //  gtid: global thread identifier for calling thread
149 //  thread: thread info for thread containing stack
150 //  tied_task: the task to push on the stack
151 static void __kmp_push_task_stack(kmp_int32 gtid, kmp_info_t *thread,
152                                   kmp_taskdata_t *tied_task) {
153   // GEH - need to consider what to do if tt_threads_data not allocated yet
154   kmp_thread_data_t *thread_data =
155       &thread->th.th_task_team->tt.tt_threads_data[__kmp_tid_from_gtid(gtid)];
156   kmp_task_stack_t *task_stack = &thread_data->td.td_susp_tied_tasks;
157 
158   if (tied_task->td_flags.team_serial || tied_task->td_flags.tasking_ser) {
159     return; // Don't push anything on stack if team or team tasks are serialized
160   }
161 
162   KMP_DEBUG_ASSERT(tied_task->td_flags.tasktype == TASK_TIED);
163   KMP_DEBUG_ASSERT(task_stack->ts_top != NULL);
164 
165   KA_TRACE(20,
166            ("__kmp_push_task_stack(enter): GTID: %d; THREAD: %p; TASK: %p\n",
167             gtid, thread, tied_task));
168   // Store entry
169   *(task_stack->ts_top) = tied_task;
170 
171   // Do bookkeeping for next push
172   task_stack->ts_top++;
173   task_stack->ts_entries++;
174 
175   if (task_stack->ts_entries & TASK_STACK_INDEX_MASK == 0) {
176     // Find beginning of this task block
177     kmp_stack_block_t *stack_block =
178         (kmp_stack_block_t *)(task_stack->ts_top - TASK_STACK_BLOCK_SIZE);
179 
180     // Check if we already have a block
181     if (stack_block->sb_next !=
182         NULL) { // reset ts_top to beginning of next block
183       task_stack->ts_top = &stack_block->sb_next->sb_block[0];
184     } else { // Alloc new block and link it up
185       kmp_stack_block_t *new_block = (kmp_stack_block_t *)__kmp_thread_calloc(
186           thread, sizeof(kmp_stack_block_t));
187 
188       task_stack->ts_top = &new_block->sb_block[0];
189       stack_block->sb_next = new_block;
190       new_block->sb_prev = stack_block;
191       new_block->sb_next = NULL;
192 
193       KA_TRACE(
194           30,
195           ("__kmp_push_task_stack(): GTID: %d; TASK: %p; Alloc new block: %p\n",
196            gtid, tied_task, new_block));
197     }
198   }
199   KA_TRACE(20, ("__kmp_push_task_stack(exit): GTID: %d; TASK: %p\n", gtid,
200                 tied_task));
201 }
202 
203 //  __kmp_pop_task_stack: Pop the tied task from the task stack.  Don't return
204 //  the task, just check to make sure it matches the ending task passed in.
205 //
206 //  gtid: global thread identifier for the calling thread
207 //  thread: thread info structure containing stack
208 //  tied_task: the task popped off the stack
209 //  ending_task: the task that is ending (should match popped task)
210 static void __kmp_pop_task_stack(kmp_int32 gtid, kmp_info_t *thread,
211                                  kmp_taskdata_t *ending_task) {
212   // GEH - need to consider what to do if tt_threads_data not allocated yet
213   kmp_thread_data_t *thread_data =
214       &thread->th.th_task_team->tt_threads_data[__kmp_tid_from_gtid(gtid)];
215   kmp_task_stack_t *task_stack = &thread_data->td.td_susp_tied_tasks;
216   kmp_taskdata_t *tied_task;
217 
218   if (ending_task->td_flags.team_serial || ending_task->td_flags.tasking_ser) {
219     // Don't pop anything from stack if team or team tasks are serialized
220     return;
221   }
222 
223   KMP_DEBUG_ASSERT(task_stack->ts_top != NULL);
224   KMP_DEBUG_ASSERT(task_stack->ts_entries > 0);
225 
226   KA_TRACE(20, ("__kmp_pop_task_stack(enter): GTID: %d; THREAD: %p\n", gtid,
227                 thread));
228 
229   // fix up ts_top if we need to pop from previous block
230   if (task_stack->ts_entries & TASK_STACK_INDEX_MASK == 0) {
231     kmp_stack_block_t *stack_block = (kmp_stack_block_t *)(task_stack->ts_top);
232 
233     stack_block = stack_block->sb_prev;
234     task_stack->ts_top = &stack_block->sb_block[TASK_STACK_BLOCK_SIZE];
235   }
236 
237   // finish bookkeeping
238   task_stack->ts_top--;
239   task_stack->ts_entries--;
240 
241   tied_task = *(task_stack->ts_top);
242 
243   KMP_DEBUG_ASSERT(tied_task != NULL);
244   KMP_DEBUG_ASSERT(tied_task->td_flags.tasktype == TASK_TIED);
245   KMP_DEBUG_ASSERT(tied_task == ending_task); // If we built the stack correctly
246 
247   KA_TRACE(20, ("__kmp_pop_task_stack(exit): GTID: %d; TASK: %p\n", gtid,
248                 tied_task));
249   return;
250 }
251 #endif /* BUILD_TIED_TASK_STACK */
252 
253 // returns 1 if new task is allowed to execute, 0 otherwise
254 // checks Task Scheduling constraint (if requested) and
255 // mutexinoutset dependencies if any
256 static bool __kmp_task_is_allowed(int gtid, const kmp_int32 is_constrained,
257                                   const kmp_taskdata_t *tasknew,
258                                   const kmp_taskdata_t *taskcurr) {
259   if (is_constrained && (tasknew->td_flags.tiedness == TASK_TIED)) {
260     // Check if the candidate obeys the Task Scheduling Constraints (TSC)
261     // only descendant of all deferred tied tasks can be scheduled, checking
262     // the last one is enough, as it in turn is the descendant of all others
263     kmp_taskdata_t *current = taskcurr->td_last_tied;
264     KMP_DEBUG_ASSERT(current != NULL);
265     // check if the task is not suspended on barrier
266     if (current->td_flags.tasktype == TASK_EXPLICIT ||
267         current->td_taskwait_thread > 0) { // <= 0 on barrier
268       kmp_int32 level = current->td_level;
269       kmp_taskdata_t *parent = tasknew->td_parent;
270       while (parent != current && parent->td_level > level) {
271         // check generation up to the level of the current task
272         parent = parent->td_parent;
273         KMP_DEBUG_ASSERT(parent != NULL);
274       }
275       if (parent != current)
276         return false;
277     }
278   }
279   // Check mutexinoutset dependencies, acquire locks
280   kmp_depnode_t *node = tasknew->td_depnode;
281   if (node && (node->dn.mtx_num_locks > 0)) {
282     for (int i = 0; i < node->dn.mtx_num_locks; ++i) {
283       KMP_DEBUG_ASSERT(node->dn.mtx_locks[i] != NULL);
284       if (__kmp_test_lock(node->dn.mtx_locks[i], gtid))
285         continue;
286       // could not get the lock, release previous locks
287       for (int j = i - 1; j >= 0; --j)
288         __kmp_release_lock(node->dn.mtx_locks[j], gtid);
289       return false;
290     }
291     // negative num_locks means all locks acquired successfully
292     node->dn.mtx_num_locks = -node->dn.mtx_num_locks;
293   }
294   return true;
295 }
296 
297 // __kmp_realloc_task_deque:
298 // Re-allocates a task deque for a particular thread, copies the content from
299 // the old deque and adjusts the necessary data structures relating to the
300 // deque. This operation must be done with the deque_lock being held
301 static void __kmp_realloc_task_deque(kmp_info_t *thread,
302                                      kmp_thread_data_t *thread_data) {
303   kmp_int32 size = TASK_DEQUE_SIZE(thread_data->td);
304   kmp_int32 new_size = 2 * size;
305 
306   KE_TRACE(10, ("__kmp_realloc_task_deque: T#%d reallocating deque[from %d to "
307                 "%d] for thread_data %p\n",
308                 __kmp_gtid_from_thread(thread), size, new_size, thread_data));
309 
310   kmp_taskdata_t **new_deque =
311       (kmp_taskdata_t **)__kmp_allocate(new_size * sizeof(kmp_taskdata_t *));
312 
313   int i, j;
314   for (i = thread_data->td.td_deque_head, j = 0; j < size;
315        i = (i + 1) & TASK_DEQUE_MASK(thread_data->td), j++)
316     new_deque[j] = thread_data->td.td_deque[i];
317 
318   __kmp_free(thread_data->td.td_deque);
319 
320   thread_data->td.td_deque_head = 0;
321   thread_data->td.td_deque_tail = size;
322   thread_data->td.td_deque = new_deque;
323   thread_data->td.td_deque_size = new_size;
324 }
325 
326 //  __kmp_push_task: Add a task to the thread's deque
327 static kmp_int32 __kmp_push_task(kmp_int32 gtid, kmp_task_t *task) {
328   kmp_info_t *thread = __kmp_threads[gtid];
329   kmp_taskdata_t *taskdata = KMP_TASK_TO_TASKDATA(task);
330   kmp_task_team_t *task_team = thread->th.th_task_team;
331   kmp_int32 tid = __kmp_tid_from_gtid(gtid);
332   kmp_thread_data_t *thread_data;
333 
334   KA_TRACE(20,
335            ("__kmp_push_task: T#%d trying to push task %p.\n", gtid, taskdata));
336 
337   if (taskdata->td_flags.tiedness == TASK_UNTIED) {
338     // untied task needs to increment counter so that the task structure is not
339     // freed prematurely
340     kmp_int32 counter = 1 + KMP_ATOMIC_INC(&taskdata->td_untied_count);
341     KMP_DEBUG_USE_VAR(counter);
342     KA_TRACE(
343         20,
344         ("__kmp_push_task: T#%d untied_count (%d) incremented for task %p\n",
345          gtid, counter, taskdata));
346   }
347 
348   // The first check avoids building task_team thread data if serialized
349   if (taskdata->td_flags.task_serial) {
350     KA_TRACE(20, ("__kmp_push_task: T#%d team serialized; returning "
351                   "TASK_NOT_PUSHED for task %p\n",
352                   gtid, taskdata));
353     return TASK_NOT_PUSHED;
354   }
355 
356   // Now that serialized tasks have returned, we can assume that we are not in
357   // immediate exec mode
358   KMP_DEBUG_ASSERT(__kmp_tasking_mode != tskm_immediate_exec);
359   if (!KMP_TASKING_ENABLED(task_team)) {
360     __kmp_enable_tasking(task_team, thread);
361   }
362   KMP_DEBUG_ASSERT(TCR_4(task_team->tt.tt_found_tasks) == TRUE);
363   KMP_DEBUG_ASSERT(TCR_PTR(task_team->tt.tt_threads_data) != NULL);
364 
365   // Find tasking deque specific to encountering thread
366   thread_data = &task_team->tt.tt_threads_data[tid];
367 
368   // No lock needed since only owner can allocate
369   if (thread_data->td.td_deque == NULL) {
370     __kmp_alloc_task_deque(thread, thread_data);
371   }
372 
373   int locked = 0;
374   // Check if deque is full
375   if (TCR_4(thread_data->td.td_deque_ntasks) >=
376       TASK_DEQUE_SIZE(thread_data->td)) {
377     if (__kmp_task_is_allowed(gtid, __kmp_task_stealing_constraint, taskdata,
378                               thread->th.th_current_task)) {
379       KA_TRACE(20, ("__kmp_push_task: T#%d deque is full; returning "
380                     "TASK_NOT_PUSHED for task %p\n",
381                     gtid, taskdata));
382       return TASK_NOT_PUSHED;
383     } else {
384       __kmp_acquire_bootstrap_lock(&thread_data->td.td_deque_lock);
385       locked = 1;
386       // expand deque to push the task which is not allowed to execute
387       __kmp_realloc_task_deque(thread, thread_data);
388     }
389   }
390   // Lock the deque for the task push operation
391   if (!locked) {
392     __kmp_acquire_bootstrap_lock(&thread_data->td.td_deque_lock);
393 #if OMP_45_ENABLED
394     // Need to recheck as we can get a proxy task from thread outside of OpenMP
395     if (TCR_4(thread_data->td.td_deque_ntasks) >=
396         TASK_DEQUE_SIZE(thread_data->td)) {
397       if (__kmp_task_is_allowed(gtid, __kmp_task_stealing_constraint, taskdata,
398                                 thread->th.th_current_task)) {
399         __kmp_release_bootstrap_lock(&thread_data->td.td_deque_lock);
400         KA_TRACE(20, ("__kmp_push_task: T#%d deque is full on 2nd check; "
401                       "returning TASK_NOT_PUSHED for task %p\n",
402                       gtid, taskdata));
403         return TASK_NOT_PUSHED;
404       } else {
405         // expand deque to push the task which is not allowed to execute
406         __kmp_realloc_task_deque(thread, thread_data);
407       }
408     }
409 #endif
410   }
411   // Must have room since no thread can add tasks but calling thread
412   KMP_DEBUG_ASSERT(TCR_4(thread_data->td.td_deque_ntasks) <
413                    TASK_DEQUE_SIZE(thread_data->td));
414 
415   thread_data->td.td_deque[thread_data->td.td_deque_tail] =
416       taskdata; // Push taskdata
417   // Wrap index.
418   thread_data->td.td_deque_tail =
419       (thread_data->td.td_deque_tail + 1) & TASK_DEQUE_MASK(thread_data->td);
420   TCW_4(thread_data->td.td_deque_ntasks,
421         TCR_4(thread_data->td.td_deque_ntasks) + 1); // Adjust task count
422 
423   KA_TRACE(20, ("__kmp_push_task: T#%d returning TASK_SUCCESSFULLY_PUSHED: "
424                 "task=%p ntasks=%d head=%u tail=%u\n",
425                 gtid, taskdata, thread_data->td.td_deque_ntasks,
426                 thread_data->td.td_deque_head, thread_data->td.td_deque_tail));
427 
428   __kmp_release_bootstrap_lock(&thread_data->td.td_deque_lock);
429 
430   return TASK_SUCCESSFULLY_PUSHED;
431 }
432 
433 // __kmp_pop_current_task_from_thread: set up current task from called thread
434 // when team ends
435 //
436 // this_thr: thread structure to set current_task in.
437 void __kmp_pop_current_task_from_thread(kmp_info_t *this_thr) {
438   KF_TRACE(10, ("__kmp_pop_current_task_from_thread(enter): T#%d "
439                 "this_thread=%p, curtask=%p, "
440                 "curtask_parent=%p\n",
441                 0, this_thr, this_thr->th.th_current_task,
442                 this_thr->th.th_current_task->td_parent));
443 
444   this_thr->th.th_current_task = this_thr->th.th_current_task->td_parent;
445 
446   KF_TRACE(10, ("__kmp_pop_current_task_from_thread(exit): T#%d "
447                 "this_thread=%p, curtask=%p, "
448                 "curtask_parent=%p\n",
449                 0, this_thr, this_thr->th.th_current_task,
450                 this_thr->th.th_current_task->td_parent));
451 }
452 
453 // __kmp_push_current_task_to_thread: set up current task in called thread for a
454 // new team
455 //
456 // this_thr: thread structure to set up
457 // team: team for implicit task data
458 // tid: thread within team to set up
459 void __kmp_push_current_task_to_thread(kmp_info_t *this_thr, kmp_team_t *team,
460                                        int tid) {
461   // current task of the thread is a parent of the new just created implicit
462   // tasks of new team
463   KF_TRACE(10, ("__kmp_push_current_task_to_thread(enter): T#%d this_thread=%p "
464                 "curtask=%p "
465                 "parent_task=%p\n",
466                 tid, this_thr, this_thr->th.th_current_task,
467                 team->t.t_implicit_task_taskdata[tid].td_parent));
468 
469   KMP_DEBUG_ASSERT(this_thr != NULL);
470 
471   if (tid == 0) {
472     if (this_thr->th.th_current_task != &team->t.t_implicit_task_taskdata[0]) {
473       team->t.t_implicit_task_taskdata[0].td_parent =
474           this_thr->th.th_current_task;
475       this_thr->th.th_current_task = &team->t.t_implicit_task_taskdata[0];
476     }
477   } else {
478     team->t.t_implicit_task_taskdata[tid].td_parent =
479         team->t.t_implicit_task_taskdata[0].td_parent;
480     this_thr->th.th_current_task = &team->t.t_implicit_task_taskdata[tid];
481   }
482 
483   KF_TRACE(10, ("__kmp_push_current_task_to_thread(exit): T#%d this_thread=%p "
484                 "curtask=%p "
485                 "parent_task=%p\n",
486                 tid, this_thr, this_thr->th.th_current_task,
487                 team->t.t_implicit_task_taskdata[tid].td_parent));
488 }
489 
490 // __kmp_task_start: bookkeeping for a task starting execution
491 //
492 // GTID: global thread id of calling thread
493 // task: task starting execution
494 // current_task: task suspending
495 static void __kmp_task_start(kmp_int32 gtid, kmp_task_t *task,
496                              kmp_taskdata_t *current_task) {
497   kmp_taskdata_t *taskdata = KMP_TASK_TO_TASKDATA(task);
498   kmp_info_t *thread = __kmp_threads[gtid];
499 
500   KA_TRACE(10,
501            ("__kmp_task_start(enter): T#%d starting task %p: current_task=%p\n",
502             gtid, taskdata, current_task));
503 
504   KMP_DEBUG_ASSERT(taskdata->td_flags.tasktype == TASK_EXPLICIT);
505 
506   // mark currently executing task as suspended
507   // TODO: GEH - make sure root team implicit task is initialized properly.
508   // KMP_DEBUG_ASSERT( current_task -> td_flags.executing == 1 );
509   current_task->td_flags.executing = 0;
510 
511 // Add task to stack if tied
512 #ifdef BUILD_TIED_TASK_STACK
513   if (taskdata->td_flags.tiedness == TASK_TIED) {
514     __kmp_push_task_stack(gtid, thread, taskdata);
515   }
516 #endif /* BUILD_TIED_TASK_STACK */
517 
518   // mark starting task as executing and as current task
519   thread->th.th_current_task = taskdata;
520 
521   KMP_DEBUG_ASSERT(taskdata->td_flags.started == 0 ||
522                    taskdata->td_flags.tiedness == TASK_UNTIED);
523   KMP_DEBUG_ASSERT(taskdata->td_flags.executing == 0 ||
524                    taskdata->td_flags.tiedness == TASK_UNTIED);
525   taskdata->td_flags.started = 1;
526   taskdata->td_flags.executing = 1;
527   KMP_DEBUG_ASSERT(taskdata->td_flags.complete == 0);
528   KMP_DEBUG_ASSERT(taskdata->td_flags.freed == 0);
529 
530   // GEH TODO: shouldn't we pass some sort of location identifier here?
531   // APT: yes, we will pass location here.
532   // need to store current thread state (in a thread or taskdata structure)
533   // before setting work_state, otherwise wrong state is set after end of task
534 
535   KA_TRACE(10, ("__kmp_task_start(exit): T#%d task=%p\n", gtid, taskdata));
536 
537   return;
538 }
539 
540 #if OMPT_SUPPORT
541 //------------------------------------------------------------------------------
542 // __ompt_task_init:
543 //   Initialize OMPT fields maintained by a task. This will only be called after
544 //   ompt_start_tool, so we already know whether ompt is enabled or not.
545 
546 static inline void __ompt_task_init(kmp_taskdata_t *task, int tid) {
547   // The calls to __ompt_task_init already have the ompt_enabled condition.
548   task->ompt_task_info.task_data.value = 0;
549   task->ompt_task_info.frame.exit_frame = ompt_data_none;
550   task->ompt_task_info.frame.enter_frame = ompt_data_none;
551   task->ompt_task_info.frame.exit_frame_flags = ompt_frame_runtime | ompt_frame_framepointer;
552   task->ompt_task_info.frame.enter_frame_flags = ompt_frame_runtime | ompt_frame_framepointer;
553 #if OMP_40_ENABLED
554   task->ompt_task_info.ndeps = 0;
555   task->ompt_task_info.deps = NULL;
556 #endif /* OMP_40_ENABLED */
557 }
558 
559 // __ompt_task_start:
560 //   Build and trigger task-begin event
561 static inline void __ompt_task_start(kmp_task_t *task,
562                                      kmp_taskdata_t *current_task,
563                                      kmp_int32 gtid) {
564   kmp_taskdata_t *taskdata = KMP_TASK_TO_TASKDATA(task);
565   ompt_task_status_t status = ompt_task_switch;
566   if (__kmp_threads[gtid]->th.ompt_thread_info.ompt_task_yielded) {
567     status = ompt_task_yield;
568     __kmp_threads[gtid]->th.ompt_thread_info.ompt_task_yielded = 0;
569   }
570   /* let OMPT know that we're about to run this task */
571   if (ompt_enabled.ompt_callback_task_schedule) {
572     ompt_callbacks.ompt_callback(ompt_callback_task_schedule)(
573         &(current_task->ompt_task_info.task_data), status,
574         &(taskdata->ompt_task_info.task_data));
575   }
576   taskdata->ompt_task_info.scheduling_parent = current_task;
577 }
578 
579 // __ompt_task_finish:
580 //   Build and trigger final task-schedule event
581 static inline void
582 __ompt_task_finish(kmp_task_t *task, kmp_taskdata_t *resumed_task,
583                    ompt_task_status_t status = ompt_task_complete) {
584   kmp_taskdata_t *taskdata = KMP_TASK_TO_TASKDATA(task);
585   if (__kmp_omp_cancellation && taskdata->td_taskgroup &&
586       taskdata->td_taskgroup->cancel_request == cancel_taskgroup) {
587     status = ompt_task_cancel;
588   }
589 
590   /* let OMPT know that we're returning to the callee task */
591   if (ompt_enabled.ompt_callback_task_schedule) {
592     ompt_callbacks.ompt_callback(ompt_callback_task_schedule)(
593         &(taskdata->ompt_task_info.task_data), status,
594         &((resumed_task ? resumed_task
595                         : (taskdata->ompt_task_info.scheduling_parent
596                                ? taskdata->ompt_task_info.scheduling_parent
597                                : taskdata->td_parent))
598               ->ompt_task_info.task_data));
599   }
600 }
601 #endif
602 
603 template <bool ompt>
604 static void __kmpc_omp_task_begin_if0_template(ident_t *loc_ref, kmp_int32 gtid,
605                                                kmp_task_t *task,
606                                                void *frame_address,
607                                                void *return_address) {
608   kmp_taskdata_t *taskdata = KMP_TASK_TO_TASKDATA(task);
609   kmp_taskdata_t *current_task = __kmp_threads[gtid]->th.th_current_task;
610 
611   KA_TRACE(10, ("__kmpc_omp_task_begin_if0(enter): T#%d loc=%p task=%p "
612                 "current_task=%p\n",
613                 gtid, loc_ref, taskdata, current_task));
614 
615   if (taskdata->td_flags.tiedness == TASK_UNTIED) {
616     // untied task needs to increment counter so that the task structure is not
617     // freed prematurely
618     kmp_int32 counter = 1 + KMP_ATOMIC_INC(&taskdata->td_untied_count);
619     KMP_DEBUG_USE_VAR(counter);
620     KA_TRACE(20, ("__kmpc_omp_task_begin_if0: T#%d untied_count (%d) "
621                   "incremented for task %p\n",
622                   gtid, counter, taskdata));
623   }
624 
625   taskdata->td_flags.task_serial =
626       1; // Execute this task immediately, not deferred.
627   __kmp_task_start(gtid, task, current_task);
628 
629 #if OMPT_SUPPORT
630   if (ompt) {
631     if (current_task->ompt_task_info.frame.enter_frame.ptr == NULL) {
632       current_task->ompt_task_info.frame.enter_frame.ptr =
633           taskdata->ompt_task_info.frame.exit_frame.ptr = frame_address;
634       current_task->ompt_task_info.frame.enter_frame_flags =
635           taskdata->ompt_task_info.frame.exit_frame_flags = ompt_frame_application | ompt_frame_framepointer;
636     }
637     if (ompt_enabled.ompt_callback_task_create) {
638       ompt_task_info_t *parent_info = &(current_task->ompt_task_info);
639       ompt_callbacks.ompt_callback(ompt_callback_task_create)(
640           &(parent_info->task_data), &(parent_info->frame),
641           &(taskdata->ompt_task_info.task_data),
642           ompt_task_explicit | TASK_TYPE_DETAILS_FORMAT(taskdata), 0,
643           return_address);
644     }
645     __ompt_task_start(task, current_task, gtid);
646   }
647 #endif // OMPT_SUPPORT
648 
649   KA_TRACE(10, ("__kmpc_omp_task_begin_if0(exit): T#%d loc=%p task=%p,\n", gtid,
650                 loc_ref, taskdata));
651 }
652 
653 #if OMPT_SUPPORT
654 OMPT_NOINLINE
655 static void __kmpc_omp_task_begin_if0_ompt(ident_t *loc_ref, kmp_int32 gtid,
656                                            kmp_task_t *task,
657                                            void *frame_address,
658                                            void *return_address) {
659   __kmpc_omp_task_begin_if0_template<true>(loc_ref, gtid, task, frame_address,
660                                            return_address);
661 }
662 #endif // OMPT_SUPPORT
663 
664 // __kmpc_omp_task_begin_if0: report that a given serialized task has started
665 // execution
666 //
667 // loc_ref: source location information; points to beginning of task block.
668 // gtid: global thread number.
669 // task: task thunk for the started task.
670 void __kmpc_omp_task_begin_if0(ident_t *loc_ref, kmp_int32 gtid,
671                                kmp_task_t *task) {
672 #if OMPT_SUPPORT
673   if (UNLIKELY(ompt_enabled.enabled)) {
674     OMPT_STORE_RETURN_ADDRESS(gtid);
675     __kmpc_omp_task_begin_if0_ompt(loc_ref, gtid, task,
676                                    OMPT_GET_FRAME_ADDRESS(1),
677                                    OMPT_LOAD_RETURN_ADDRESS(gtid));
678     return;
679   }
680 #endif
681   __kmpc_omp_task_begin_if0_template<false>(loc_ref, gtid, task, NULL, NULL);
682 }
683 
684 #ifdef TASK_UNUSED
685 // __kmpc_omp_task_begin: report that a given task has started execution
686 // NEVER GENERATED BY COMPILER, DEPRECATED!!!
687 void __kmpc_omp_task_begin(ident_t *loc_ref, kmp_int32 gtid, kmp_task_t *task) {
688   kmp_taskdata_t *current_task = __kmp_threads[gtid]->th.th_current_task;
689 
690   KA_TRACE(
691       10,
692       ("__kmpc_omp_task_begin(enter): T#%d loc=%p task=%p current_task=%p\n",
693        gtid, loc_ref, KMP_TASK_TO_TASKDATA(task), current_task));
694 
695   __kmp_task_start(gtid, task, current_task);
696 
697   KA_TRACE(10, ("__kmpc_omp_task_begin(exit): T#%d loc=%p task=%p,\n", gtid,
698                 loc_ref, KMP_TASK_TO_TASKDATA(task)));
699   return;
700 }
701 #endif // TASK_UNUSED
702 
703 // __kmp_free_task: free the current task space and the space for shareds
704 //
705 // gtid: Global thread ID of calling thread
706 // taskdata: task to free
707 // thread: thread data structure of caller
708 static void __kmp_free_task(kmp_int32 gtid, kmp_taskdata_t *taskdata,
709                             kmp_info_t *thread) {
710   KA_TRACE(30, ("__kmp_free_task: T#%d freeing data from task %p\n", gtid,
711                 taskdata));
712 
713   // Check to make sure all flags and counters have the correct values
714   KMP_DEBUG_ASSERT(taskdata->td_flags.tasktype == TASK_EXPLICIT);
715   KMP_DEBUG_ASSERT(taskdata->td_flags.executing == 0);
716   KMP_DEBUG_ASSERT(taskdata->td_flags.complete == 1);
717   KMP_DEBUG_ASSERT(taskdata->td_flags.freed == 0);
718   KMP_DEBUG_ASSERT(taskdata->td_allocated_child_tasks == 0 ||
719                    taskdata->td_flags.task_serial == 1);
720   KMP_DEBUG_ASSERT(taskdata->td_incomplete_child_tasks == 0);
721 
722   taskdata->td_flags.freed = 1;
723   ANNOTATE_HAPPENS_BEFORE(taskdata);
724 // deallocate the taskdata and shared variable blocks associated with this task
725 #if USE_FAST_MEMORY
726   __kmp_fast_free(thread, taskdata);
727 #else /* ! USE_FAST_MEMORY */
728   __kmp_thread_free(thread, taskdata);
729 #endif
730 
731   KA_TRACE(20, ("__kmp_free_task: T#%d freed task %p\n", gtid, taskdata));
732 }
733 
734 // __kmp_free_task_and_ancestors: free the current task and ancestors without
735 // children
736 //
737 // gtid: Global thread ID of calling thread
738 // taskdata: task to free
739 // thread: thread data structure of caller
740 static void __kmp_free_task_and_ancestors(kmp_int32 gtid,
741                                           kmp_taskdata_t *taskdata,
742                                           kmp_info_t *thread) {
743 #if OMP_45_ENABLED
744   // Proxy tasks must always be allowed to free their parents
745   // because they can be run in background even in serial mode.
746   kmp_int32 team_serial =
747       (taskdata->td_flags.team_serial || taskdata->td_flags.tasking_ser) &&
748       !taskdata->td_flags.proxy;
749 #else
750   kmp_int32 team_serial =
751       taskdata->td_flags.team_serial || taskdata->td_flags.tasking_ser;
752 #endif
753   KMP_DEBUG_ASSERT(taskdata->td_flags.tasktype == TASK_EXPLICIT);
754 
755   kmp_int32 children = KMP_ATOMIC_DEC(&taskdata->td_allocated_child_tasks) - 1;
756   KMP_DEBUG_ASSERT(children >= 0);
757 
758   // Now, go up the ancestor tree to see if any ancestors can now be freed.
759   while (children == 0) {
760     kmp_taskdata_t *parent_taskdata = taskdata->td_parent;
761 
762     KA_TRACE(20, ("__kmp_free_task_and_ancestors(enter): T#%d task %p complete "
763                   "and freeing itself\n",
764                   gtid, taskdata));
765 
766     // --- Deallocate my ancestor task ---
767     __kmp_free_task(gtid, taskdata, thread);
768 
769     taskdata = parent_taskdata;
770 
771     if (team_serial)
772       return;
773     // Stop checking ancestors at implicit task instead of walking up ancestor
774     // tree to avoid premature deallocation of ancestors.
775     if (taskdata->td_flags.tasktype == TASK_IMPLICIT) {
776       if (taskdata->td_dephash) { // do we need to cleanup dephash?
777         int children = KMP_ATOMIC_LD_ACQ(&taskdata->td_incomplete_child_tasks);
778         kmp_tasking_flags_t flags_old = taskdata->td_flags;
779         if (children == 0 && flags_old.complete == 1) {
780           kmp_tasking_flags_t flags_new = flags_old;
781           flags_new.complete = 0;
782           if (KMP_COMPARE_AND_STORE_ACQ32(
783                   RCAST(kmp_int32 *, &taskdata->td_flags),
784                   *RCAST(kmp_int32 *, &flags_old),
785                   *RCAST(kmp_int32 *, &flags_new))) {
786             KA_TRACE(100, ("__kmp_free_task_and_ancestors: T#%d cleans "
787                            "dephash of implicit task %p\n",
788                            gtid, taskdata));
789             // cleanup dephash of finished implicit task
790             __kmp_dephash_free_entries(thread, taskdata->td_dephash);
791           }
792         }
793       }
794       return;
795     }
796     // Predecrement simulated by "- 1" calculation
797     children = KMP_ATOMIC_DEC(&taskdata->td_allocated_child_tasks) - 1;
798     KMP_DEBUG_ASSERT(children >= 0);
799   }
800 
801   KA_TRACE(
802       20, ("__kmp_free_task_and_ancestors(exit): T#%d task %p has %d children; "
803            "not freeing it yet\n",
804            gtid, taskdata, children));
805 }
806 
807 // __kmp_task_finish: bookkeeping to do when a task finishes execution
808 //
809 // gtid: global thread ID for calling thread
810 // task: task to be finished
811 // resumed_task: task to be resumed.  (may be NULL if task is serialized)
812 template <bool ompt>
813 static void __kmp_task_finish(kmp_int32 gtid, kmp_task_t *task,
814                               kmp_taskdata_t *resumed_task) {
815   kmp_taskdata_t *taskdata = KMP_TASK_TO_TASKDATA(task);
816   kmp_info_t *thread = __kmp_threads[gtid];
817 #if OMP_45_ENABLED
818   kmp_task_team_t *task_team =
819       thread->th.th_task_team; // might be NULL for serial teams...
820 #endif // OMP_45_ENABLED
821   kmp_int32 children = 0;
822 
823   KA_TRACE(10, ("__kmp_task_finish(enter): T#%d finishing task %p and resuming "
824                 "task %p\n",
825                 gtid, taskdata, resumed_task));
826 
827   KMP_DEBUG_ASSERT(taskdata->td_flags.tasktype == TASK_EXPLICIT);
828 
829 // Pop task from stack if tied
830 #ifdef BUILD_TIED_TASK_STACK
831   if (taskdata->td_flags.tiedness == TASK_TIED) {
832     __kmp_pop_task_stack(gtid, thread, taskdata);
833   }
834 #endif /* BUILD_TIED_TASK_STACK */
835 
836   if (taskdata->td_flags.tiedness == TASK_UNTIED) {
837     // untied task needs to check the counter so that the task structure is not
838     // freed prematurely
839     kmp_int32 counter = KMP_ATOMIC_DEC(&taskdata->td_untied_count) - 1;
840     KA_TRACE(
841         20,
842         ("__kmp_task_finish: T#%d untied_count (%d) decremented for task %p\n",
843          gtid, counter, taskdata));
844     if (counter > 0) {
845       // untied task is not done, to be continued possibly by other thread, do
846       // not free it now
847       if (resumed_task == NULL) {
848         KMP_DEBUG_ASSERT(taskdata->td_flags.task_serial);
849         resumed_task = taskdata->td_parent; // In a serialized task, the resumed
850         // task is the parent
851       }
852       thread->th.th_current_task = resumed_task; // restore current_task
853       resumed_task->td_flags.executing = 1; // resume previous task
854       KA_TRACE(10, ("__kmp_task_finish(exit): T#%d partially done task %p, "
855                     "resuming task %p\n",
856                     gtid, taskdata, resumed_task));
857       return;
858     }
859   }
860 #if OMPT_SUPPORT
861   if (ompt)
862     __ompt_task_finish(task, resumed_task);
863 #endif
864 
865   // Check mutexinoutset dependencies, release locks
866   kmp_depnode_t *node = taskdata->td_depnode;
867   if (node && (node->dn.mtx_num_locks < 0)) {
868     // negative num_locks means all locks were acquired
869     node->dn.mtx_num_locks = -node->dn.mtx_num_locks;
870     for (int i = node->dn.mtx_num_locks - 1; i >= 0; --i) {
871       KMP_DEBUG_ASSERT(node->dn.mtx_locks[i] != NULL);
872       __kmp_release_lock(node->dn.mtx_locks[i], gtid);
873     }
874   }
875 
876   KMP_DEBUG_ASSERT(taskdata->td_flags.complete == 0);
877   taskdata->td_flags.complete = 1; // mark the task as completed
878   KMP_DEBUG_ASSERT(taskdata->td_flags.started == 1);
879   KMP_DEBUG_ASSERT(taskdata->td_flags.freed == 0);
880 
881   // Only need to keep track of count if team parallel and tasking not
882   // serialized
883   if (!(taskdata->td_flags.team_serial || taskdata->td_flags.tasking_ser)) {
884     // Predecrement simulated by "- 1" calculation
885     children =
886         KMP_ATOMIC_DEC(&taskdata->td_parent->td_incomplete_child_tasks) - 1;
887     KMP_DEBUG_ASSERT(children >= 0);
888 #if OMP_40_ENABLED
889     if (taskdata->td_taskgroup)
890       KMP_ATOMIC_DEC(&taskdata->td_taskgroup->count);
891     __kmp_release_deps(gtid, taskdata);
892 #if OMP_45_ENABLED
893   } else if (task_team && task_team->tt.tt_found_proxy_tasks) {
894     // if we found proxy tasks there could exist a dependency chain
895     // with the proxy task as origin
896     __kmp_release_deps(gtid, taskdata);
897 #endif // OMP_45_ENABLED
898 #endif // OMP_40_ENABLED
899   }
900 
901   // td_flags.executing must be marked as 0 after __kmp_release_deps has been
902   // called. Othertwise, if a task is executed immediately from the release_deps
903   // code, the flag will be reset to 1 again by this same function
904   KMP_DEBUG_ASSERT(taskdata->td_flags.executing == 1);
905   taskdata->td_flags.executing = 0; // suspend the finishing task
906 
907   KA_TRACE(
908       20, ("__kmp_task_finish: T#%d finished task %p, %d incomplete children\n",
909            gtid, taskdata, children));
910 
911 #if OMP_40_ENABLED
912   /* If the tasks' destructor thunk flag has been set, we need to invoke the
913      destructor thunk that has been generated by the compiler. The code is
914      placed here, since at this point other tasks might have been released
915      hence overlapping the destructor invokations with some other work in the
916      released tasks.  The OpenMP spec is not specific on when the destructors
917      are invoked, so we should be free to choose. */
918   if (taskdata->td_flags.destructors_thunk) {
919     kmp_routine_entry_t destr_thunk = task->data1.destructors;
920     KMP_ASSERT(destr_thunk);
921     destr_thunk(gtid, task);
922   }
923 #endif // OMP_40_ENABLED
924 
925   // bookkeeping for resuming task:
926   // GEH - note tasking_ser => task_serial
927   KMP_DEBUG_ASSERT(
928       (taskdata->td_flags.tasking_ser || taskdata->td_flags.task_serial) ==
929       taskdata->td_flags.task_serial);
930   if (taskdata->td_flags.task_serial) {
931     if (resumed_task == NULL) {
932       resumed_task = taskdata->td_parent; // In a serialized task, the resumed
933       // task is the parent
934     }
935   } else {
936     KMP_DEBUG_ASSERT(resumed_task !=
937                      NULL); // verify that resumed task is passed as arguemnt
938   }
939 
940   // Free this task and then ancestor tasks if they have no children.
941   // Restore th_current_task first as suggested by John:
942   // johnmc: if an asynchronous inquiry peers into the runtime system
943   // it doesn't see the freed task as the current task.
944   thread->th.th_current_task = resumed_task;
945   __kmp_free_task_and_ancestors(gtid, taskdata, thread);
946 
947   // TODO: GEH - make sure root team implicit task is initialized properly.
948   // KMP_DEBUG_ASSERT( resumed_task->td_flags.executing == 0 );
949   resumed_task->td_flags.executing = 1; // resume previous task
950 
951   KA_TRACE(
952       10, ("__kmp_task_finish(exit): T#%d finished task %p, resuming task %p\n",
953            gtid, taskdata, resumed_task));
954 
955   return;
956 }
957 
958 template <bool ompt>
959 static void __kmpc_omp_task_complete_if0_template(ident_t *loc_ref,
960                                                   kmp_int32 gtid,
961                                                   kmp_task_t *task) {
962   KA_TRACE(10, ("__kmpc_omp_task_complete_if0(enter): T#%d loc=%p task=%p\n",
963                 gtid, loc_ref, KMP_TASK_TO_TASKDATA(task)));
964   // this routine will provide task to resume
965   __kmp_task_finish<ompt>(gtid, task, NULL);
966 
967   KA_TRACE(10, ("__kmpc_omp_task_complete_if0(exit): T#%d loc=%p task=%p\n",
968                 gtid, loc_ref, KMP_TASK_TO_TASKDATA(task)));
969 
970 #if OMPT_SUPPORT
971   if (ompt) {
972     ompt_frame_t *ompt_frame;
973     __ompt_get_task_info_internal(0, NULL, NULL, &ompt_frame, NULL, NULL);
974     ompt_frame->enter_frame = ompt_data_none;
975     ompt_frame->enter_frame_flags = ompt_frame_runtime | ompt_frame_framepointer;
976   }
977 #endif
978 
979   return;
980 }
981 
982 #if OMPT_SUPPORT
983 OMPT_NOINLINE
984 void __kmpc_omp_task_complete_if0_ompt(ident_t *loc_ref, kmp_int32 gtid,
985                                        kmp_task_t *task) {
986   __kmpc_omp_task_complete_if0_template<true>(loc_ref, gtid, task);
987 }
988 #endif // OMPT_SUPPORT
989 
990 // __kmpc_omp_task_complete_if0: report that a task has completed execution
991 //
992 // loc_ref: source location information; points to end of task block.
993 // gtid: global thread number.
994 // task: task thunk for the completed task.
995 void __kmpc_omp_task_complete_if0(ident_t *loc_ref, kmp_int32 gtid,
996                                   kmp_task_t *task) {
997 #if OMPT_SUPPORT
998   if (UNLIKELY(ompt_enabled.enabled)) {
999     __kmpc_omp_task_complete_if0_ompt(loc_ref, gtid, task);
1000     return;
1001   }
1002 #endif
1003   __kmpc_omp_task_complete_if0_template<false>(loc_ref, gtid, task);
1004 }
1005 
1006 #ifdef TASK_UNUSED
1007 // __kmpc_omp_task_complete: report that a task has completed execution
1008 // NEVER GENERATED BY COMPILER, DEPRECATED!!!
1009 void __kmpc_omp_task_complete(ident_t *loc_ref, kmp_int32 gtid,
1010                               kmp_task_t *task) {
1011   KA_TRACE(10, ("__kmpc_omp_task_complete(enter): T#%d loc=%p task=%p\n", gtid,
1012                 loc_ref, KMP_TASK_TO_TASKDATA(task)));
1013 
1014   __kmp_task_finish<false>(gtid, task,
1015                            NULL); // Not sure how to find task to resume
1016 
1017   KA_TRACE(10, ("__kmpc_omp_task_complete(exit): T#%d loc=%p task=%p\n", gtid,
1018                 loc_ref, KMP_TASK_TO_TASKDATA(task)));
1019   return;
1020 }
1021 #endif // TASK_UNUSED
1022 
1023 // __kmp_init_implicit_task: Initialize the appropriate fields in the implicit
1024 // task for a given thread
1025 //
1026 // loc_ref:  reference to source location of parallel region
1027 // this_thr:  thread data structure corresponding to implicit task
1028 // team: team for this_thr
1029 // tid: thread id of given thread within team
1030 // set_curr_task: TRUE if need to push current task to thread
1031 // NOTE: Routine does not set up the implicit task ICVS.  This is assumed to
1032 // have already been done elsewhere.
1033 // TODO: Get better loc_ref.  Value passed in may be NULL
1034 void __kmp_init_implicit_task(ident_t *loc_ref, kmp_info_t *this_thr,
1035                               kmp_team_t *team, int tid, int set_curr_task) {
1036   kmp_taskdata_t *task = &team->t.t_implicit_task_taskdata[tid];
1037 
1038   KF_TRACE(
1039       10,
1040       ("__kmp_init_implicit_task(enter): T#:%d team=%p task=%p, reinit=%s\n",
1041        tid, team, task, set_curr_task ? "TRUE" : "FALSE"));
1042 
1043   task->td_task_id = KMP_GEN_TASK_ID();
1044   task->td_team = team;
1045   //    task->td_parent   = NULL;  // fix for CQ230101 (broken parent task info
1046   //    in debugger)
1047   task->td_ident = loc_ref;
1048   task->td_taskwait_ident = NULL;
1049   task->td_taskwait_counter = 0;
1050   task->td_taskwait_thread = 0;
1051 
1052   task->td_flags.tiedness = TASK_TIED;
1053   task->td_flags.tasktype = TASK_IMPLICIT;
1054 #if OMP_45_ENABLED
1055   task->td_flags.proxy = TASK_FULL;
1056 #endif
1057 
1058   // All implicit tasks are executed immediately, not deferred
1059   task->td_flags.task_serial = 1;
1060   task->td_flags.tasking_ser = (__kmp_tasking_mode == tskm_immediate_exec);
1061   task->td_flags.team_serial = (team->t.t_serialized) ? 1 : 0;
1062 
1063   task->td_flags.started = 1;
1064   task->td_flags.executing = 1;
1065   task->td_flags.complete = 0;
1066   task->td_flags.freed = 0;
1067 
1068 #if OMP_40_ENABLED
1069   task->td_depnode = NULL;
1070 #endif
1071   task->td_last_tied = task;
1072 
1073   if (set_curr_task) { // only do this init first time thread is created
1074     KMP_ATOMIC_ST_REL(&task->td_incomplete_child_tasks, 0);
1075     // Not used: don't need to deallocate implicit task
1076     KMP_ATOMIC_ST_REL(&task->td_allocated_child_tasks, 0);
1077 #if OMP_40_ENABLED
1078     task->td_taskgroup = NULL; // An implicit task does not have taskgroup
1079     task->td_dephash = NULL;
1080 #endif
1081     __kmp_push_current_task_to_thread(this_thr, team, tid);
1082   } else {
1083     KMP_DEBUG_ASSERT(task->td_incomplete_child_tasks == 0);
1084     KMP_DEBUG_ASSERT(task->td_allocated_child_tasks == 0);
1085   }
1086 
1087 #if OMPT_SUPPORT
1088   if (UNLIKELY(ompt_enabled.enabled))
1089     __ompt_task_init(task, tid);
1090 #endif
1091 
1092   KF_TRACE(10, ("__kmp_init_implicit_task(exit): T#:%d team=%p task=%p\n", tid,
1093                 team, task));
1094 }
1095 
1096 // __kmp_finish_implicit_task: Release resources associated to implicit tasks
1097 // at the end of parallel regions. Some resources are kept for reuse in the next
1098 // parallel region.
1099 //
1100 // thread:  thread data structure corresponding to implicit task
1101 void __kmp_finish_implicit_task(kmp_info_t *thread) {
1102   kmp_taskdata_t *task = thread->th.th_current_task;
1103   if (task->td_dephash) {
1104     int children;
1105     task->td_flags.complete = 1;
1106     children = KMP_ATOMIC_LD_ACQ(&task->td_incomplete_child_tasks);
1107     kmp_tasking_flags_t flags_old = task->td_flags;
1108     if (children == 0 && flags_old.complete == 1) {
1109       kmp_tasking_flags_t flags_new = flags_old;
1110       flags_new.complete = 0;
1111       if (KMP_COMPARE_AND_STORE_ACQ32(RCAST(kmp_int32 *, &task->td_flags),
1112                                       *RCAST(kmp_int32 *, &flags_old),
1113                                       *RCAST(kmp_int32 *, &flags_new))) {
1114         KA_TRACE(100, ("__kmp_finish_implicit_task: T#%d cleans "
1115                        "dephash of implicit task %p\n",
1116                        thread->th.th_info.ds.ds_gtid, task));
1117         __kmp_dephash_free_entries(thread, task->td_dephash);
1118       }
1119     }
1120   }
1121 }
1122 
1123 // __kmp_free_implicit_task: Release resources associated to implicit tasks
1124 // when these are destroyed regions
1125 //
1126 // thread:  thread data structure corresponding to implicit task
1127 void __kmp_free_implicit_task(kmp_info_t *thread) {
1128   kmp_taskdata_t *task = thread->th.th_current_task;
1129   if (task && task->td_dephash) {
1130     __kmp_dephash_free(thread, task->td_dephash);
1131     task->td_dephash = NULL;
1132   }
1133 }
1134 
1135 // Round up a size to a power of two specified by val: Used to insert padding
1136 // between structures co-allocated using a single malloc() call
1137 static size_t __kmp_round_up_to_val(size_t size, size_t val) {
1138   if (size & (val - 1)) {
1139     size &= ~(val - 1);
1140     if (size <= KMP_SIZE_T_MAX - val) {
1141       size += val; // Round up if there is no overflow.
1142     }
1143   }
1144   return size;
1145 } // __kmp_round_up_to_va
1146 
1147 // __kmp_task_alloc: Allocate the taskdata and task data structures for a task
1148 //
1149 // loc_ref: source location information
1150 // gtid: global thread number.
1151 // flags: include tiedness & task type (explicit vs. implicit) of the ''new''
1152 // task encountered. Converted from kmp_int32 to kmp_tasking_flags_t in routine.
1153 // sizeof_kmp_task_t:  Size in bytes of kmp_task_t data structure including
1154 // private vars accessed in task.
1155 // sizeof_shareds:  Size in bytes of array of pointers to shared vars accessed
1156 // in task.
1157 // task_entry: Pointer to task code entry point generated by compiler.
1158 // returns: a pointer to the allocated kmp_task_t structure (task).
1159 kmp_task_t *__kmp_task_alloc(ident_t *loc_ref, kmp_int32 gtid,
1160                              kmp_tasking_flags_t *flags,
1161                              size_t sizeof_kmp_task_t, size_t sizeof_shareds,
1162                              kmp_routine_entry_t task_entry) {
1163   kmp_task_t *task;
1164   kmp_taskdata_t *taskdata;
1165   kmp_info_t *thread = __kmp_threads[gtid];
1166   kmp_team_t *team = thread->th.th_team;
1167   kmp_taskdata_t *parent_task = thread->th.th_current_task;
1168   size_t shareds_offset;
1169 
1170   if (!TCR_4(__kmp_init_middle))
1171     __kmp_middle_initialize();
1172 
1173   KA_TRACE(10, ("__kmp_task_alloc(enter): T#%d loc=%p, flags=(0x%x) "
1174                 "sizeof_task=%ld sizeof_shared=%ld entry=%p\n",
1175                 gtid, loc_ref, *((kmp_int32 *)flags), sizeof_kmp_task_t,
1176                 sizeof_shareds, task_entry));
1177 
1178   if (parent_task->td_flags.final) {
1179     if (flags->merged_if0) {
1180     }
1181     flags->final = 1;
1182   }
1183   if (flags->tiedness == TASK_UNTIED && !team->t.t_serialized) {
1184     // Untied task encountered causes the TSC algorithm to check entire deque of
1185     // the victim thread. If no untied task encountered, then checking the head
1186     // of the deque should be enough.
1187     KMP_CHECK_UPDATE(thread->th.th_task_team->tt.tt_untied_task_encountered, 1);
1188   }
1189 
1190 #if OMP_45_ENABLED
1191   if (flags->proxy == TASK_PROXY) {
1192     flags->tiedness = TASK_UNTIED;
1193     flags->merged_if0 = 1;
1194 
1195     /* are we running in a sequential parallel or tskm_immediate_exec... we need
1196        tasking support enabled */
1197     if ((thread->th.th_task_team) == NULL) {
1198       /* This should only happen if the team is serialized
1199           setup a task team and propagate it to the thread */
1200       KMP_DEBUG_ASSERT(team->t.t_serialized);
1201       KA_TRACE(30,
1202                ("T#%d creating task team in __kmp_task_alloc for proxy task\n",
1203                 gtid));
1204       __kmp_task_team_setup(
1205           thread, team,
1206           1); // 1 indicates setup the current team regardless of nthreads
1207       thread->th.th_task_team = team->t.t_task_team[thread->th.th_task_state];
1208     }
1209     kmp_task_team_t *task_team = thread->th.th_task_team;
1210 
1211     /* tasking must be enabled now as the task might not be pushed */
1212     if (!KMP_TASKING_ENABLED(task_team)) {
1213       KA_TRACE(
1214           30,
1215           ("T#%d enabling tasking in __kmp_task_alloc for proxy task\n", gtid));
1216       __kmp_enable_tasking(task_team, thread);
1217       kmp_int32 tid = thread->th.th_info.ds.ds_tid;
1218       kmp_thread_data_t *thread_data = &task_team->tt.tt_threads_data[tid];
1219       // No lock needed since only owner can allocate
1220       if (thread_data->td.td_deque == NULL) {
1221         __kmp_alloc_task_deque(thread, thread_data);
1222       }
1223     }
1224 
1225     if (task_team->tt.tt_found_proxy_tasks == FALSE)
1226       TCW_4(task_team->tt.tt_found_proxy_tasks, TRUE);
1227   }
1228 #endif
1229 
1230   // Calculate shared structure offset including padding after kmp_task_t struct
1231   // to align pointers in shared struct
1232   shareds_offset = sizeof(kmp_taskdata_t) + sizeof_kmp_task_t;
1233   shareds_offset = __kmp_round_up_to_val(shareds_offset, sizeof(void *));
1234 
1235   // Allocate a kmp_taskdata_t block and a kmp_task_t block.
1236   KA_TRACE(30, ("__kmp_task_alloc: T#%d First malloc size: %ld\n", gtid,
1237                 shareds_offset));
1238   KA_TRACE(30, ("__kmp_task_alloc: T#%d Second malloc size: %ld\n", gtid,
1239                 sizeof_shareds));
1240 
1241 // Avoid double allocation here by combining shareds with taskdata
1242 #if USE_FAST_MEMORY
1243   taskdata = (kmp_taskdata_t *)__kmp_fast_allocate(thread, shareds_offset +
1244                                                                sizeof_shareds);
1245 #else /* ! USE_FAST_MEMORY */
1246   taskdata = (kmp_taskdata_t *)__kmp_thread_malloc(thread, shareds_offset +
1247                                                                sizeof_shareds);
1248 #endif /* USE_FAST_MEMORY */
1249   ANNOTATE_HAPPENS_AFTER(taskdata);
1250 
1251   task = KMP_TASKDATA_TO_TASK(taskdata);
1252 
1253 // Make sure task & taskdata are aligned appropriately
1254 #if KMP_ARCH_X86 || KMP_ARCH_PPC64 || !KMP_HAVE_QUAD
1255   KMP_DEBUG_ASSERT((((kmp_uintptr_t)taskdata) & (sizeof(double) - 1)) == 0);
1256   KMP_DEBUG_ASSERT((((kmp_uintptr_t)task) & (sizeof(double) - 1)) == 0);
1257 #else
1258   KMP_DEBUG_ASSERT((((kmp_uintptr_t)taskdata) & (sizeof(_Quad) - 1)) == 0);
1259   KMP_DEBUG_ASSERT((((kmp_uintptr_t)task) & (sizeof(_Quad) - 1)) == 0);
1260 #endif
1261   if (sizeof_shareds > 0) {
1262     // Avoid double allocation here by combining shareds with taskdata
1263     task->shareds = &((char *)taskdata)[shareds_offset];
1264     // Make sure shareds struct is aligned to pointer size
1265     KMP_DEBUG_ASSERT((((kmp_uintptr_t)task->shareds) & (sizeof(void *) - 1)) ==
1266                      0);
1267   } else {
1268     task->shareds = NULL;
1269   }
1270   task->routine = task_entry;
1271   task->part_id = 0; // AC: Always start with 0 part id
1272 
1273   taskdata->td_task_id = KMP_GEN_TASK_ID();
1274   taskdata->td_team = team;
1275   taskdata->td_alloc_thread = thread;
1276   taskdata->td_parent = parent_task;
1277   taskdata->td_level = parent_task->td_level + 1; // increment nesting level
1278   KMP_ATOMIC_ST_RLX(&taskdata->td_untied_count, 0);
1279   taskdata->td_ident = loc_ref;
1280   taskdata->td_taskwait_ident = NULL;
1281   taskdata->td_taskwait_counter = 0;
1282   taskdata->td_taskwait_thread = 0;
1283   KMP_DEBUG_ASSERT(taskdata->td_parent != NULL);
1284 #if OMP_45_ENABLED
1285   // avoid copying icvs for proxy tasks
1286   if (flags->proxy == TASK_FULL)
1287 #endif
1288     copy_icvs(&taskdata->td_icvs, &taskdata->td_parent->td_icvs);
1289 
1290   taskdata->td_flags.tiedness = flags->tiedness;
1291   taskdata->td_flags.final = flags->final;
1292   taskdata->td_flags.merged_if0 = flags->merged_if0;
1293 #if OMP_40_ENABLED
1294   taskdata->td_flags.destructors_thunk = flags->destructors_thunk;
1295 #endif // OMP_40_ENABLED
1296 #if OMP_45_ENABLED
1297   taskdata->td_flags.proxy = flags->proxy;
1298   taskdata->td_task_team = thread->th.th_task_team;
1299   taskdata->td_size_alloc = shareds_offset + sizeof_shareds;
1300 #endif
1301   taskdata->td_flags.tasktype = TASK_EXPLICIT;
1302 
1303   // GEH - TODO: fix this to copy parent task's value of tasking_ser flag
1304   taskdata->td_flags.tasking_ser = (__kmp_tasking_mode == tskm_immediate_exec);
1305 
1306   // GEH - TODO: fix this to copy parent task's value of team_serial flag
1307   taskdata->td_flags.team_serial = (team->t.t_serialized) ? 1 : 0;
1308 
1309   // GEH - Note we serialize the task if the team is serialized to make sure
1310   // implicit parallel region tasks are not left until program termination to
1311   // execute. Also, it helps locality to execute immediately.
1312 
1313   taskdata->td_flags.task_serial =
1314       (parent_task->td_flags.final || taskdata->td_flags.team_serial ||
1315        taskdata->td_flags.tasking_ser);
1316 
1317   taskdata->td_flags.started = 0;
1318   taskdata->td_flags.executing = 0;
1319   taskdata->td_flags.complete = 0;
1320   taskdata->td_flags.freed = 0;
1321 
1322   taskdata->td_flags.native = flags->native;
1323 
1324   KMP_ATOMIC_ST_RLX(&taskdata->td_incomplete_child_tasks, 0);
1325   // start at one because counts current task and children
1326   KMP_ATOMIC_ST_RLX(&taskdata->td_allocated_child_tasks, 1);
1327 #if OMP_40_ENABLED
1328   taskdata->td_taskgroup =
1329       parent_task->td_taskgroup; // task inherits taskgroup from the parent task
1330   taskdata->td_dephash = NULL;
1331   taskdata->td_depnode = NULL;
1332 #endif
1333   if (flags->tiedness == TASK_UNTIED)
1334     taskdata->td_last_tied = NULL; // will be set when the task is scheduled
1335   else
1336     taskdata->td_last_tied = taskdata;
1337 
1338 #if OMPT_SUPPORT
1339   if (UNLIKELY(ompt_enabled.enabled))
1340     __ompt_task_init(taskdata, gtid);
1341 #endif
1342 // Only need to keep track of child task counts if team parallel and tasking not
1343 // serialized or if it is a proxy task
1344 #if OMP_45_ENABLED
1345   if (flags->proxy == TASK_PROXY ||
1346       !(taskdata->td_flags.team_serial || taskdata->td_flags.tasking_ser))
1347 #else
1348   if (!(taskdata->td_flags.team_serial || taskdata->td_flags.tasking_ser))
1349 #endif
1350   {
1351     KMP_ATOMIC_INC(&parent_task->td_incomplete_child_tasks);
1352 #if OMP_40_ENABLED
1353     if (parent_task->td_taskgroup)
1354       KMP_ATOMIC_INC(&parent_task->td_taskgroup->count);
1355 #endif
1356     // Only need to keep track of allocated child tasks for explicit tasks since
1357     // implicit not deallocated
1358     if (taskdata->td_parent->td_flags.tasktype == TASK_EXPLICIT) {
1359       KMP_ATOMIC_INC(&taskdata->td_parent->td_allocated_child_tasks);
1360     }
1361   }
1362 
1363   KA_TRACE(20, ("__kmp_task_alloc(exit): T#%d created task %p parent=%p\n",
1364                 gtid, taskdata, taskdata->td_parent));
1365   ANNOTATE_HAPPENS_BEFORE(task);
1366 
1367   return task;
1368 }
1369 
1370 kmp_task_t *__kmpc_omp_task_alloc(ident_t *loc_ref, kmp_int32 gtid,
1371                                   kmp_int32 flags, size_t sizeof_kmp_task_t,
1372                                   size_t sizeof_shareds,
1373                                   kmp_routine_entry_t task_entry) {
1374   kmp_task_t *retval;
1375   kmp_tasking_flags_t *input_flags = (kmp_tasking_flags_t *)&flags;
1376 
1377   input_flags->native = FALSE;
1378 // __kmp_task_alloc() sets up all other runtime flags
1379 
1380 #if OMP_45_ENABLED
1381   KA_TRACE(10, ("__kmpc_omp_task_alloc(enter): T#%d loc=%p, flags=(%s %s) "
1382                 "sizeof_task=%ld sizeof_shared=%ld entry=%p\n",
1383                 gtid, loc_ref, input_flags->tiedness ? "tied  " : "untied",
1384                 input_flags->proxy ? "proxy" : "", sizeof_kmp_task_t,
1385                 sizeof_shareds, task_entry));
1386 #else
1387   KA_TRACE(10, ("__kmpc_omp_task_alloc(enter): T#%d loc=%p, flags=(%s) "
1388                 "sizeof_task=%ld sizeof_shared=%ld entry=%p\n",
1389                 gtid, loc_ref, input_flags->tiedness ? "tied  " : "untied",
1390                 sizeof_kmp_task_t, sizeof_shareds, task_entry));
1391 #endif
1392 
1393   retval = __kmp_task_alloc(loc_ref, gtid, input_flags, sizeof_kmp_task_t,
1394                             sizeof_shareds, task_entry);
1395 
1396   KA_TRACE(20, ("__kmpc_omp_task_alloc(exit): T#%d retval %p\n", gtid, retval));
1397 
1398   return retval;
1399 }
1400 
1401 #if OMP_50_ENABLED
1402 /*!
1403 @ingroup TASKING
1404 @param loc_ref location of the original task directive
1405 @param gtid Global Thread ID of encountering thread
1406 @param new_task task thunk allocated by __kmpc_omp_task_alloc() for the ''new
1407 task''
1408 @param naffins Number of affinity items
1409 @param affin_list List of affinity items
1410 @return Returns non-zero if registering affinity information was not successful.
1411  Returns 0 if registration was successful
1412 This entry registers the affinity information attached to a task with the task
1413 thunk structure kmp_taskdata_t.
1414 */
1415 kmp_int32
1416 __kmpc_omp_reg_task_with_affinity(ident_t *loc_ref, kmp_int32 gtid,
1417                                   kmp_task_t *new_task, kmp_int32 naffins,
1418                                   kmp_task_affinity_info_t *affin_list) {
1419   return 0;
1420 }
1421 #endif
1422 
1423 //  __kmp_invoke_task: invoke the specified task
1424 //
1425 // gtid: global thread ID of caller
1426 // task: the task to invoke
1427 // current_task: the task to resume after task invokation
1428 static void __kmp_invoke_task(kmp_int32 gtid, kmp_task_t *task,
1429                               kmp_taskdata_t *current_task) {
1430   kmp_taskdata_t *taskdata = KMP_TASK_TO_TASKDATA(task);
1431   kmp_info_t *thread;
1432 #if OMP_40_ENABLED
1433   int discard = 0 /* false */;
1434 #endif
1435   KA_TRACE(
1436       30, ("__kmp_invoke_task(enter): T#%d invoking task %p, current_task=%p\n",
1437            gtid, taskdata, current_task));
1438   KMP_DEBUG_ASSERT(task);
1439 #if OMP_45_ENABLED
1440   if (taskdata->td_flags.proxy == TASK_PROXY &&
1441       taskdata->td_flags.complete == 1) {
1442     // This is a proxy task that was already completed but it needs to run
1443     // its bottom-half finish
1444     KA_TRACE(
1445         30,
1446         ("__kmp_invoke_task: T#%d running bottom finish for proxy task %p\n",
1447          gtid, taskdata));
1448 
1449     __kmp_bottom_half_finish_proxy(gtid, task);
1450 
1451     KA_TRACE(30, ("__kmp_invoke_task(exit): T#%d completed bottom finish for "
1452                   "proxy task %p, resuming task %p\n",
1453                   gtid, taskdata, current_task));
1454 
1455     return;
1456   }
1457 #endif
1458 
1459 #if OMPT_SUPPORT
1460   // For untied tasks, the first task executed only calls __kmpc_omp_task and
1461   // does not execute code.
1462   ompt_thread_info_t oldInfo;
1463   if (UNLIKELY(ompt_enabled.enabled)) {
1464     // Store the threads states and restore them after the task
1465     thread = __kmp_threads[gtid];
1466     oldInfo = thread->th.ompt_thread_info;
1467     thread->th.ompt_thread_info.wait_id = 0;
1468     thread->th.ompt_thread_info.state = (thread->th.th_team_serialized)
1469                                             ? ompt_state_work_serial
1470                                             : ompt_state_work_parallel;
1471     taskdata->ompt_task_info.frame.exit_frame.ptr = OMPT_GET_FRAME_ADDRESS(0);
1472   }
1473 #endif
1474 
1475 #if OMP_45_ENABLED
1476   // Proxy tasks are not handled by the runtime
1477   if (taskdata->td_flags.proxy != TASK_PROXY) {
1478 #endif
1479     ANNOTATE_HAPPENS_AFTER(task);
1480     __kmp_task_start(gtid, task, current_task); // OMPT only if not discarded
1481 #if OMP_45_ENABLED
1482   }
1483 #endif
1484 
1485 #if OMP_40_ENABLED
1486   // TODO: cancel tasks if the parallel region has also been cancelled
1487   // TODO: check if this sequence can be hoisted above __kmp_task_start
1488   // if cancellation has been enabled for this run ...
1489   if (__kmp_omp_cancellation) {
1490     thread = __kmp_threads[gtid];
1491     kmp_team_t *this_team = thread->th.th_team;
1492     kmp_taskgroup_t *taskgroup = taskdata->td_taskgroup;
1493     if ((taskgroup && taskgroup->cancel_request) ||
1494         (this_team->t.t_cancel_request == cancel_parallel)) {
1495 #if OMPT_SUPPORT && OMPT_OPTIONAL
1496       ompt_data_t *task_data;
1497       if (UNLIKELY(ompt_enabled.ompt_callback_cancel)) {
1498         __ompt_get_task_info_internal(0, NULL, &task_data, NULL, NULL, NULL);
1499         ompt_callbacks.ompt_callback(ompt_callback_cancel)(
1500             task_data,
1501             ((taskgroup && taskgroup->cancel_request) ? ompt_cancel_taskgroup
1502                                                       : ompt_cancel_parallel) |
1503                 ompt_cancel_discarded_task,
1504             NULL);
1505       }
1506 #endif
1507       KMP_COUNT_BLOCK(TASK_cancelled);
1508       // this task belongs to a task group and we need to cancel it
1509       discard = 1 /* true */;
1510     }
1511   }
1512 
1513   // Invoke the task routine and pass in relevant data.
1514   // Thunks generated by gcc take a different argument list.
1515   if (!discard) {
1516     if (taskdata->td_flags.tiedness == TASK_UNTIED) {
1517       taskdata->td_last_tied = current_task->td_last_tied;
1518       KMP_DEBUG_ASSERT(taskdata->td_last_tied);
1519     }
1520 #if KMP_STATS_ENABLED
1521     KMP_COUNT_BLOCK(TASK_executed);
1522     switch (KMP_GET_THREAD_STATE()) {
1523     case FORK_JOIN_BARRIER:
1524       KMP_PUSH_PARTITIONED_TIMER(OMP_task_join_bar);
1525       break;
1526     case PLAIN_BARRIER:
1527       KMP_PUSH_PARTITIONED_TIMER(OMP_task_plain_bar);
1528       break;
1529     case TASKYIELD:
1530       KMP_PUSH_PARTITIONED_TIMER(OMP_task_taskyield);
1531       break;
1532     case TASKWAIT:
1533       KMP_PUSH_PARTITIONED_TIMER(OMP_task_taskwait);
1534       break;
1535     case TASKGROUP:
1536       KMP_PUSH_PARTITIONED_TIMER(OMP_task_taskgroup);
1537       break;
1538     default:
1539       KMP_PUSH_PARTITIONED_TIMER(OMP_task_immediate);
1540       break;
1541     }
1542 #endif // KMP_STATS_ENABLED
1543 #endif // OMP_40_ENABLED
1544 
1545 // OMPT task begin
1546 #if OMPT_SUPPORT
1547     if (UNLIKELY(ompt_enabled.enabled))
1548       __ompt_task_start(task, current_task, gtid);
1549 #endif
1550 
1551 #if USE_ITT_BUILD && USE_ITT_NOTIFY
1552     kmp_uint64 cur_time;
1553     kmp_int32 kmp_itt_count_task =
1554         __kmp_forkjoin_frames_mode == 3 && !taskdata->td_flags.task_serial &&
1555         current_task->td_flags.tasktype == TASK_IMPLICIT;
1556     if (kmp_itt_count_task) {
1557       thread = __kmp_threads[gtid];
1558       // Time outer level explicit task on barrier for adjusting imbalance time
1559       if (thread->th.th_bar_arrive_time)
1560         cur_time = __itt_get_timestamp();
1561       else
1562         kmp_itt_count_task = 0; // thread is not on a barrier - skip timing
1563     }
1564 #endif
1565 
1566 #ifdef KMP_GOMP_COMPAT
1567     if (taskdata->td_flags.native) {
1568       ((void (*)(void *))(*(task->routine)))(task->shareds);
1569     } else
1570 #endif /* KMP_GOMP_COMPAT */
1571     {
1572       (*(task->routine))(gtid, task);
1573     }
1574     KMP_POP_PARTITIONED_TIMER();
1575 
1576 #if USE_ITT_BUILD && USE_ITT_NOTIFY
1577     if (kmp_itt_count_task) {
1578       // Barrier imbalance - adjust arrive time with the task duration
1579       thread->th.th_bar_arrive_time += (__itt_get_timestamp() - cur_time);
1580     }
1581 #endif
1582 
1583 #if OMP_40_ENABLED
1584   }
1585 #endif // OMP_40_ENABLED
1586 
1587 
1588 #if OMP_45_ENABLED
1589   // Proxy tasks are not handled by the runtime
1590   if (taskdata->td_flags.proxy != TASK_PROXY) {
1591 #endif
1592     ANNOTATE_HAPPENS_BEFORE(taskdata->td_parent);
1593 #if OMPT_SUPPORT
1594     if (UNLIKELY(ompt_enabled.enabled)) {
1595       thread->th.ompt_thread_info = oldInfo;
1596       if (taskdata->td_flags.tiedness == TASK_TIED) {
1597         taskdata->ompt_task_info.frame.exit_frame = ompt_data_none;
1598       }
1599       __kmp_task_finish<true>(gtid, task, current_task);
1600     } else
1601 #endif
1602       __kmp_task_finish<false>(gtid, task, current_task);
1603 #if OMP_45_ENABLED
1604   }
1605 #endif
1606 
1607   KA_TRACE(
1608       30,
1609       ("__kmp_invoke_task(exit): T#%d completed task %p, resuming task %p\n",
1610        gtid, taskdata, current_task));
1611   return;
1612 }
1613 
1614 // __kmpc_omp_task_parts: Schedule a thread-switchable task for execution
1615 //
1616 // loc_ref: location of original task pragma (ignored)
1617 // gtid: Global Thread ID of encountering thread
1618 // new_task: task thunk allocated by __kmp_omp_task_alloc() for the ''new task''
1619 // Returns:
1620 //    TASK_CURRENT_NOT_QUEUED (0) if did not suspend and queue current task to
1621 //    be resumed later.
1622 //    TASK_CURRENT_QUEUED (1) if suspended and queued the current task to be
1623 //    resumed later.
1624 kmp_int32 __kmpc_omp_task_parts(ident_t *loc_ref, kmp_int32 gtid,
1625                                 kmp_task_t *new_task) {
1626   kmp_taskdata_t *new_taskdata = KMP_TASK_TO_TASKDATA(new_task);
1627 
1628   KA_TRACE(10, ("__kmpc_omp_task_parts(enter): T#%d loc=%p task=%p\n", gtid,
1629                 loc_ref, new_taskdata));
1630 
1631 #if OMPT_SUPPORT
1632   kmp_taskdata_t *parent;
1633   if (UNLIKELY(ompt_enabled.enabled)) {
1634     parent = new_taskdata->td_parent;
1635     if (ompt_enabled.ompt_callback_task_create) {
1636       ompt_data_t task_data = ompt_data_none;
1637       ompt_callbacks.ompt_callback(ompt_callback_task_create)(
1638           parent ? &(parent->ompt_task_info.task_data) : &task_data,
1639           parent ? &(parent->ompt_task_info.frame) : NULL,
1640           &(new_taskdata->ompt_task_info.task_data), ompt_task_explicit, 0,
1641           OMPT_GET_RETURN_ADDRESS(0));
1642     }
1643   }
1644 #endif
1645 
1646   /* Should we execute the new task or queue it? For now, let's just always try
1647      to queue it.  If the queue fills up, then we'll execute it.  */
1648 
1649   if (__kmp_push_task(gtid, new_task) == TASK_NOT_PUSHED) // if cannot defer
1650   { // Execute this task immediately
1651     kmp_taskdata_t *current_task = __kmp_threads[gtid]->th.th_current_task;
1652     new_taskdata->td_flags.task_serial = 1;
1653     __kmp_invoke_task(gtid, new_task, current_task);
1654   }
1655 
1656   KA_TRACE(
1657       10,
1658       ("__kmpc_omp_task_parts(exit): T#%d returning TASK_CURRENT_NOT_QUEUED: "
1659        "loc=%p task=%p, return: TASK_CURRENT_NOT_QUEUED\n",
1660        gtid, loc_ref, new_taskdata));
1661 
1662   ANNOTATE_HAPPENS_BEFORE(new_task);
1663 #if OMPT_SUPPORT
1664   if (UNLIKELY(ompt_enabled.enabled)) {
1665     parent->ompt_task_info.frame.enter_frame = ompt_data_none;
1666   }
1667 #endif
1668   return TASK_CURRENT_NOT_QUEUED;
1669 }
1670 
1671 // __kmp_omp_task: Schedule a non-thread-switchable task for execution
1672 //
1673 // gtid: Global Thread ID of encountering thread
1674 // new_task:non-thread-switchable task thunk allocated by __kmp_omp_task_alloc()
1675 // serialize_immediate: if TRUE then if the task is executed immediately its
1676 // execution will be serialized
1677 // Returns:
1678 //    TASK_CURRENT_NOT_QUEUED (0) if did not suspend and queue current task to
1679 //    be resumed later.
1680 //    TASK_CURRENT_QUEUED (1) if suspended and queued the current task to be
1681 //    resumed later.
1682 kmp_int32 __kmp_omp_task(kmp_int32 gtid, kmp_task_t *new_task,
1683                          bool serialize_immediate) {
1684   kmp_taskdata_t *new_taskdata = KMP_TASK_TO_TASKDATA(new_task);
1685 
1686 /* Should we execute the new task or queue it? For now, let's just always try to
1687    queue it.  If the queue fills up, then we'll execute it.  */
1688 #if OMP_45_ENABLED
1689   if (new_taskdata->td_flags.proxy == TASK_PROXY ||
1690       __kmp_push_task(gtid, new_task) == TASK_NOT_PUSHED) // if cannot defer
1691 #else
1692   if (__kmp_push_task(gtid, new_task) == TASK_NOT_PUSHED) // if cannot defer
1693 #endif
1694   { // Execute this task immediately
1695     kmp_taskdata_t *current_task = __kmp_threads[gtid]->th.th_current_task;
1696     if (serialize_immediate)
1697       new_taskdata->td_flags.task_serial = 1;
1698     __kmp_invoke_task(gtid, new_task, current_task);
1699   }
1700 
1701   ANNOTATE_HAPPENS_BEFORE(new_task);
1702   return TASK_CURRENT_NOT_QUEUED;
1703 }
1704 
1705 // __kmpc_omp_task: Wrapper around __kmp_omp_task to schedule a
1706 // non-thread-switchable task from the parent thread only!
1707 //
1708 // loc_ref: location of original task pragma (ignored)
1709 // gtid: Global Thread ID of encountering thread
1710 // new_task: non-thread-switchable task thunk allocated by
1711 // __kmp_omp_task_alloc()
1712 // Returns:
1713 //    TASK_CURRENT_NOT_QUEUED (0) if did not suspend and queue current task to
1714 //    be resumed later.
1715 //    TASK_CURRENT_QUEUED (1) if suspended and queued the current task to be
1716 //    resumed later.
1717 kmp_int32 __kmpc_omp_task(ident_t *loc_ref, kmp_int32 gtid,
1718                           kmp_task_t *new_task) {
1719   kmp_int32 res;
1720   KMP_SET_THREAD_STATE_BLOCK(EXPLICIT_TASK);
1721 
1722 #if KMP_DEBUG || OMPT_SUPPORT
1723   kmp_taskdata_t *new_taskdata = KMP_TASK_TO_TASKDATA(new_task);
1724 #endif
1725   KA_TRACE(10, ("__kmpc_omp_task(enter): T#%d loc=%p task=%p\n", gtid, loc_ref,
1726                 new_taskdata));
1727 
1728 #if OMPT_SUPPORT
1729   kmp_taskdata_t *parent = NULL;
1730   if (UNLIKELY(ompt_enabled.enabled)) {
1731     if (!new_taskdata->td_flags.started) {
1732       OMPT_STORE_RETURN_ADDRESS(gtid);
1733       parent = new_taskdata->td_parent;
1734       if (!parent->ompt_task_info.frame.enter_frame.ptr) {
1735         parent->ompt_task_info.frame.enter_frame.ptr = OMPT_GET_FRAME_ADDRESS(0);
1736       }
1737       if (ompt_enabled.ompt_callback_task_create) {
1738         ompt_data_t task_data = ompt_data_none;
1739         ompt_callbacks.ompt_callback(ompt_callback_task_create)(
1740             parent ? &(parent->ompt_task_info.task_data) : &task_data,
1741             parent ? &(parent->ompt_task_info.frame) : NULL,
1742             &(new_taskdata->ompt_task_info.task_data),
1743             ompt_task_explicit | TASK_TYPE_DETAILS_FORMAT(new_taskdata), 0,
1744             OMPT_LOAD_RETURN_ADDRESS(gtid));
1745       }
1746     } else {
1747       // We are scheduling the continuation of an UNTIED task.
1748       // Scheduling back to the parent task.
1749       __ompt_task_finish(new_task,
1750                          new_taskdata->ompt_task_info.scheduling_parent,
1751                          ompt_task_switch);
1752       new_taskdata->ompt_task_info.frame.exit_frame = ompt_data_none;
1753     }
1754   }
1755 #endif
1756 
1757   res = __kmp_omp_task(gtid, new_task, true);
1758 
1759   KA_TRACE(10, ("__kmpc_omp_task(exit): T#%d returning "
1760                 "TASK_CURRENT_NOT_QUEUED: loc=%p task=%p\n",
1761                 gtid, loc_ref, new_taskdata));
1762 #if OMPT_SUPPORT
1763   if (UNLIKELY(ompt_enabled.enabled && parent != NULL)) {
1764     parent->ompt_task_info.frame.enter_frame = ompt_data_none;
1765   }
1766 #endif
1767   return res;
1768 }
1769 
1770 // __kmp_omp_taskloop_task: Wrapper around __kmp_omp_task to schedule
1771 // a taskloop task with the correct OMPT return address
1772 //
1773 // loc_ref: location of original task pragma (ignored)
1774 // gtid: Global Thread ID of encountering thread
1775 // new_task: non-thread-switchable task thunk allocated by
1776 // __kmp_omp_task_alloc()
1777 // codeptr_ra: return address for OMPT callback
1778 // Returns:
1779 //    TASK_CURRENT_NOT_QUEUED (0) if did not suspend and queue current task to
1780 //    be resumed later.
1781 //    TASK_CURRENT_QUEUED (1) if suspended and queued the current task to be
1782 //    resumed later.
1783 kmp_int32 __kmp_omp_taskloop_task(ident_t *loc_ref, kmp_int32 gtid,
1784                                   kmp_task_t *new_task, void *codeptr_ra) {
1785   kmp_int32 res;
1786   KMP_SET_THREAD_STATE_BLOCK(EXPLICIT_TASK);
1787 
1788 #if KMP_DEBUG || OMPT_SUPPORT
1789   kmp_taskdata_t *new_taskdata = KMP_TASK_TO_TASKDATA(new_task);
1790 #endif
1791   KA_TRACE(10, ("__kmpc_omp_task(enter): T#%d loc=%p task=%p\n", gtid, loc_ref,
1792                 new_taskdata));
1793 
1794 #if OMPT_SUPPORT
1795   kmp_taskdata_t *parent = NULL;
1796   if (UNLIKELY(ompt_enabled.enabled && !new_taskdata->td_flags.started)) {
1797     parent = new_taskdata->td_parent;
1798     if (!parent->ompt_task_info.frame.enter_frame.ptr)
1799       parent->ompt_task_info.frame.enter_frame.ptr = OMPT_GET_FRAME_ADDRESS(0);
1800     if (ompt_enabled.ompt_callback_task_create) {
1801       ompt_data_t task_data = ompt_data_none;
1802       ompt_callbacks.ompt_callback(ompt_callback_task_create)(
1803           parent ? &(parent->ompt_task_info.task_data) : &task_data,
1804           parent ? &(parent->ompt_task_info.frame) : NULL,
1805           &(new_taskdata->ompt_task_info.task_data),
1806           ompt_task_explicit | TASK_TYPE_DETAILS_FORMAT(new_taskdata), 0,
1807           codeptr_ra);
1808     }
1809   }
1810 #endif
1811 
1812   res = __kmp_omp_task(gtid, new_task, true);
1813 
1814   KA_TRACE(10, ("__kmpc_omp_task(exit): T#%d returning "
1815                 "TASK_CURRENT_NOT_QUEUED: loc=%p task=%p\n",
1816                 gtid, loc_ref, new_taskdata));
1817 #if OMPT_SUPPORT
1818   if (UNLIKELY(ompt_enabled.enabled && parent != NULL)) {
1819     parent->ompt_task_info.frame.enter_frame = ompt_data_none;
1820   }
1821 #endif
1822   return res;
1823 }
1824 
1825 template <bool ompt>
1826 static kmp_int32 __kmpc_omp_taskwait_template(ident_t *loc_ref, kmp_int32 gtid,
1827                                               void *frame_address,
1828                                               void *return_address) {
1829   kmp_taskdata_t *taskdata;
1830   kmp_info_t *thread;
1831   int thread_finished = FALSE;
1832   KMP_SET_THREAD_STATE_BLOCK(TASKWAIT);
1833 
1834   KA_TRACE(10, ("__kmpc_omp_taskwait(enter): T#%d loc=%p\n", gtid, loc_ref));
1835 
1836   if (__kmp_tasking_mode != tskm_immediate_exec) {
1837     thread = __kmp_threads[gtid];
1838     taskdata = thread->th.th_current_task;
1839 
1840 #if OMPT_SUPPORT && OMPT_OPTIONAL
1841     ompt_data_t *my_task_data;
1842     ompt_data_t *my_parallel_data;
1843 
1844     if (ompt) {
1845       my_task_data = &(taskdata->ompt_task_info.task_data);
1846       my_parallel_data = OMPT_CUR_TEAM_DATA(thread);
1847 
1848       taskdata->ompt_task_info.frame.enter_frame.ptr = frame_address;
1849 
1850       if (ompt_enabled.ompt_callback_sync_region) {
1851         ompt_callbacks.ompt_callback(ompt_callback_sync_region)(
1852             ompt_sync_region_taskwait, ompt_scope_begin, my_parallel_data,
1853             my_task_data, return_address);
1854       }
1855 
1856       if (ompt_enabled.ompt_callback_sync_region_wait) {
1857         ompt_callbacks.ompt_callback(ompt_callback_sync_region_wait)(
1858             ompt_sync_region_taskwait, ompt_scope_begin, my_parallel_data,
1859             my_task_data, return_address);
1860       }
1861     }
1862 #endif // OMPT_SUPPORT && OMPT_OPTIONAL
1863 
1864 // Debugger: The taskwait is active. Store location and thread encountered the
1865 // taskwait.
1866 #if USE_ITT_BUILD
1867 // Note: These values are used by ITT events as well.
1868 #endif /* USE_ITT_BUILD */
1869     taskdata->td_taskwait_counter += 1;
1870     taskdata->td_taskwait_ident = loc_ref;
1871     taskdata->td_taskwait_thread = gtid + 1;
1872 
1873 #if USE_ITT_BUILD
1874     void *itt_sync_obj = __kmp_itt_taskwait_object(gtid);
1875     if (itt_sync_obj != NULL)
1876       __kmp_itt_taskwait_starting(gtid, itt_sync_obj);
1877 #endif /* USE_ITT_BUILD */
1878 
1879     bool must_wait =
1880         !taskdata->td_flags.team_serial && !taskdata->td_flags.final;
1881 
1882 #if OMP_45_ENABLED
1883     must_wait = must_wait || (thread->th.th_task_team != NULL &&
1884                               thread->th.th_task_team->tt.tt_found_proxy_tasks);
1885 #endif
1886     if (must_wait) {
1887       kmp_flag_32 flag(RCAST(std::atomic<kmp_uint32> *,
1888                              &(taskdata->td_incomplete_child_tasks)),
1889                        0U);
1890       while (KMP_ATOMIC_LD_ACQ(&taskdata->td_incomplete_child_tasks) != 0) {
1891         flag.execute_tasks(thread, gtid, FALSE,
1892                            &thread_finished USE_ITT_BUILD_ARG(itt_sync_obj),
1893                            __kmp_task_stealing_constraint);
1894       }
1895     }
1896 #if USE_ITT_BUILD
1897     if (itt_sync_obj != NULL)
1898       __kmp_itt_taskwait_finished(gtid, itt_sync_obj);
1899 #endif /* USE_ITT_BUILD */
1900 
1901     // Debugger:  The taskwait is completed. Location remains, but thread is
1902     // negated.
1903     taskdata->td_taskwait_thread = -taskdata->td_taskwait_thread;
1904 
1905 #if OMPT_SUPPORT && OMPT_OPTIONAL
1906     if (ompt) {
1907       if (ompt_enabled.ompt_callback_sync_region_wait) {
1908         ompt_callbacks.ompt_callback(ompt_callback_sync_region_wait)(
1909             ompt_sync_region_taskwait, ompt_scope_end, my_parallel_data,
1910             my_task_data, return_address);
1911       }
1912       if (ompt_enabled.ompt_callback_sync_region) {
1913         ompt_callbacks.ompt_callback(ompt_callback_sync_region)(
1914             ompt_sync_region_taskwait, ompt_scope_end, my_parallel_data,
1915             my_task_data, return_address);
1916       }
1917       taskdata->ompt_task_info.frame.enter_frame = ompt_data_none;
1918     }
1919 #endif // OMPT_SUPPORT && OMPT_OPTIONAL
1920 
1921     ANNOTATE_HAPPENS_AFTER(taskdata);
1922   }
1923 
1924   KA_TRACE(10, ("__kmpc_omp_taskwait(exit): T#%d task %p finished waiting, "
1925                 "returning TASK_CURRENT_NOT_QUEUED\n",
1926                 gtid, taskdata));
1927 
1928   return TASK_CURRENT_NOT_QUEUED;
1929 }
1930 
1931 #if OMPT_SUPPORT && OMPT_OPTIONAL
1932 OMPT_NOINLINE
1933 static kmp_int32 __kmpc_omp_taskwait_ompt(ident_t *loc_ref, kmp_int32 gtid,
1934                                           void *frame_address,
1935                                           void *return_address) {
1936   return __kmpc_omp_taskwait_template<true>(loc_ref, gtid, frame_address,
1937                                             return_address);
1938 }
1939 #endif // OMPT_SUPPORT && OMPT_OPTIONAL
1940 
1941 // __kmpc_omp_taskwait: Wait until all tasks generated by the current task are
1942 // complete
1943 kmp_int32 __kmpc_omp_taskwait(ident_t *loc_ref, kmp_int32 gtid) {
1944 #if OMPT_SUPPORT && OMPT_OPTIONAL
1945   if (UNLIKELY(ompt_enabled.enabled)) {
1946     OMPT_STORE_RETURN_ADDRESS(gtid);
1947     return __kmpc_omp_taskwait_ompt(loc_ref, gtid, OMPT_GET_FRAME_ADDRESS(0),
1948                                     OMPT_LOAD_RETURN_ADDRESS(gtid));
1949   }
1950 #endif
1951   return __kmpc_omp_taskwait_template<false>(loc_ref, gtid, NULL, NULL);
1952 }
1953 
1954 // __kmpc_omp_taskyield: switch to a different task
1955 kmp_int32 __kmpc_omp_taskyield(ident_t *loc_ref, kmp_int32 gtid, int end_part) {
1956   kmp_taskdata_t *taskdata;
1957   kmp_info_t *thread;
1958   int thread_finished = FALSE;
1959 
1960   KMP_COUNT_BLOCK(OMP_TASKYIELD);
1961   KMP_SET_THREAD_STATE_BLOCK(TASKYIELD);
1962 
1963   KA_TRACE(10, ("__kmpc_omp_taskyield(enter): T#%d loc=%p end_part = %d\n",
1964                 gtid, loc_ref, end_part));
1965 
1966   if (__kmp_tasking_mode != tskm_immediate_exec && __kmp_init_parallel) {
1967     thread = __kmp_threads[gtid];
1968     taskdata = thread->th.th_current_task;
1969 // Should we model this as a task wait or not?
1970 // Debugger: The taskwait is active. Store location and thread encountered the
1971 // taskwait.
1972 #if USE_ITT_BUILD
1973 // Note: These values are used by ITT events as well.
1974 #endif /* USE_ITT_BUILD */
1975     taskdata->td_taskwait_counter += 1;
1976     taskdata->td_taskwait_ident = loc_ref;
1977     taskdata->td_taskwait_thread = gtid + 1;
1978 
1979 #if USE_ITT_BUILD
1980     void *itt_sync_obj = __kmp_itt_taskwait_object(gtid);
1981     if (itt_sync_obj != NULL)
1982       __kmp_itt_taskwait_starting(gtid, itt_sync_obj);
1983 #endif /* USE_ITT_BUILD */
1984     if (!taskdata->td_flags.team_serial) {
1985       kmp_task_team_t *task_team = thread->th.th_task_team;
1986       if (task_team != NULL) {
1987         if (KMP_TASKING_ENABLED(task_team)) {
1988 #if OMPT_SUPPORT
1989           if (UNLIKELY(ompt_enabled.enabled))
1990             thread->th.ompt_thread_info.ompt_task_yielded = 1;
1991 #endif
1992           __kmp_execute_tasks_32(
1993               thread, gtid, NULL, FALSE,
1994               &thread_finished USE_ITT_BUILD_ARG(itt_sync_obj),
1995               __kmp_task_stealing_constraint);
1996 #if OMPT_SUPPORT
1997           if (UNLIKELY(ompt_enabled.enabled))
1998             thread->th.ompt_thread_info.ompt_task_yielded = 0;
1999 #endif
2000         }
2001       }
2002     }
2003 #if USE_ITT_BUILD
2004     if (itt_sync_obj != NULL)
2005       __kmp_itt_taskwait_finished(gtid, itt_sync_obj);
2006 #endif /* USE_ITT_BUILD */
2007 
2008     // Debugger:  The taskwait is completed. Location remains, but thread is
2009     // negated.
2010     taskdata->td_taskwait_thread = -taskdata->td_taskwait_thread;
2011   }
2012 
2013   KA_TRACE(10, ("__kmpc_omp_taskyield(exit): T#%d task %p resuming, "
2014                 "returning TASK_CURRENT_NOT_QUEUED\n",
2015                 gtid, taskdata));
2016 
2017   return TASK_CURRENT_NOT_QUEUED;
2018 }
2019 
2020 #if OMP_50_ENABLED
2021 // Task Reduction implementation
2022 
2023 typedef struct kmp_task_red_flags {
2024   unsigned lazy_priv : 1; // hint: (1) use lazy allocation (big objects)
2025   unsigned reserved31 : 31;
2026 } kmp_task_red_flags_t;
2027 
2028 // internal structure for reduction data item related info
2029 typedef struct kmp_task_red_data {
2030   void *reduce_shar; // shared reduction item
2031   size_t reduce_size; // size of data item
2032   void *reduce_priv; // thread specific data
2033   void *reduce_pend; // end of private data for comparison op
2034   void *reduce_init; // data initialization routine
2035   void *reduce_fini; // data finalization routine
2036   void *reduce_comb; // data combiner routine
2037   kmp_task_red_flags_t flags; // flags for additional info from compiler
2038 } kmp_task_red_data_t;
2039 
2040 // structure sent us by compiler - one per reduction item
2041 typedef struct kmp_task_red_input {
2042   void *reduce_shar; // shared reduction item
2043   size_t reduce_size; // size of data item
2044   void *reduce_init; // data initialization routine
2045   void *reduce_fini; // data finalization routine
2046   void *reduce_comb; // data combiner routine
2047   kmp_task_red_flags_t flags; // flags for additional info from compiler
2048 } kmp_task_red_input_t;
2049 
2050 /*!
2051 @ingroup TASKING
2052 @param gtid      Global thread ID
2053 @param num       Number of data items to reduce
2054 @param data      Array of data for reduction
2055 @return The taskgroup identifier
2056 
2057 Initialize task reduction for the taskgroup.
2058 */
2059 void *__kmpc_task_reduction_init(int gtid, int num, void *data) {
2060   kmp_info_t *thread = __kmp_threads[gtid];
2061   kmp_taskgroup_t *tg = thread->th.th_current_task->td_taskgroup;
2062   kmp_int32 nth = thread->th.th_team_nproc;
2063   kmp_task_red_input_t *input = (kmp_task_red_input_t *)data;
2064   kmp_task_red_data_t *arr;
2065 
2066   // check input data just in case
2067   KMP_ASSERT(tg != NULL);
2068   KMP_ASSERT(data != NULL);
2069   KMP_ASSERT(num > 0);
2070   if (nth == 1) {
2071     KA_TRACE(10, ("__kmpc_task_reduction_init: T#%d, tg %p, exiting nth=1\n",
2072                   gtid, tg));
2073     return (void *)tg;
2074   }
2075   KA_TRACE(10, ("__kmpc_task_reduction_init: T#%d, taskgroup %p, #items %d\n",
2076                 gtid, tg, num));
2077   arr = (kmp_task_red_data_t *)__kmp_thread_malloc(
2078       thread, num * sizeof(kmp_task_red_data_t));
2079   for (int i = 0; i < num; ++i) {
2080     void (*f_init)(void *) = (void (*)(void *))(input[i].reduce_init);
2081     size_t size = input[i].reduce_size - 1;
2082     // round the size up to cache line per thread-specific item
2083     size += CACHE_LINE - size % CACHE_LINE;
2084     KMP_ASSERT(input[i].reduce_comb != NULL); // combiner is mandatory
2085     arr[i].reduce_shar = input[i].reduce_shar;
2086     arr[i].reduce_size = size;
2087     arr[i].reduce_init = input[i].reduce_init;
2088     arr[i].reduce_fini = input[i].reduce_fini;
2089     arr[i].reduce_comb = input[i].reduce_comb;
2090     arr[i].flags = input[i].flags;
2091     if (!input[i].flags.lazy_priv) {
2092       // allocate cache-line aligned block and fill it with zeros
2093       arr[i].reduce_priv = __kmp_allocate(nth * size);
2094       arr[i].reduce_pend = (char *)(arr[i].reduce_priv) + nth * size;
2095       if (f_init != NULL) {
2096         // initialize thread-specific items
2097         for (int j = 0; j < nth; ++j) {
2098           f_init((char *)(arr[i].reduce_priv) + j * size);
2099         }
2100       }
2101     } else {
2102       // only allocate space for pointers now,
2103       // objects will be lazily allocated/initialized once requested
2104       arr[i].reduce_priv = __kmp_allocate(nth * sizeof(void *));
2105     }
2106   }
2107   tg->reduce_data = (void *)arr;
2108   tg->reduce_num_data = num;
2109   return (void *)tg;
2110 }
2111 
2112 /*!
2113 @ingroup TASKING
2114 @param gtid    Global thread ID
2115 @param tskgrp  The taskgroup ID (optional)
2116 @param data    Shared location of the item
2117 @return The pointer to per-thread data
2118 
2119 Get thread-specific location of data item
2120 */
2121 void *__kmpc_task_reduction_get_th_data(int gtid, void *tskgrp, void *data) {
2122   kmp_info_t *thread = __kmp_threads[gtid];
2123   kmp_int32 nth = thread->th.th_team_nproc;
2124   if (nth == 1)
2125     return data; // nothing to do
2126 
2127   kmp_taskgroup_t *tg = (kmp_taskgroup_t *)tskgrp;
2128   if (tg == NULL)
2129     tg = thread->th.th_current_task->td_taskgroup;
2130   KMP_ASSERT(tg != NULL);
2131   kmp_task_red_data_t *arr = (kmp_task_red_data_t *)(tg->reduce_data);
2132   kmp_int32 num = tg->reduce_num_data;
2133   kmp_int32 tid = thread->th.th_info.ds.ds_tid;
2134 
2135   KMP_ASSERT(data != NULL);
2136   while (tg != NULL) {
2137     for (int i = 0; i < num; ++i) {
2138       if (!arr[i].flags.lazy_priv) {
2139         if (data == arr[i].reduce_shar ||
2140             (data >= arr[i].reduce_priv && data < arr[i].reduce_pend))
2141           return (char *)(arr[i].reduce_priv) + tid * arr[i].reduce_size;
2142       } else {
2143         // check shared location first
2144         void **p_priv = (void **)(arr[i].reduce_priv);
2145         if (data == arr[i].reduce_shar)
2146           goto found;
2147         // check if we get some thread specific location as parameter
2148         for (int j = 0; j < nth; ++j)
2149           if (data == p_priv[j])
2150             goto found;
2151         continue; // not found, continue search
2152       found:
2153         if (p_priv[tid] == NULL) {
2154           // allocate thread specific object lazily
2155           void (*f_init)(void *) = (void (*)(void *))(arr[i].reduce_init);
2156           p_priv[tid] = __kmp_allocate(arr[i].reduce_size);
2157           if (f_init != NULL) {
2158             f_init(p_priv[tid]);
2159           }
2160         }
2161         return p_priv[tid];
2162       }
2163     }
2164     tg = tg->parent;
2165     arr = (kmp_task_red_data_t *)(tg->reduce_data);
2166     num = tg->reduce_num_data;
2167   }
2168   KMP_ASSERT2(0, "Unknown task reduction item");
2169   return NULL; // ERROR, this line never executed
2170 }
2171 
2172 // Finalize task reduction.
2173 // Called from __kmpc_end_taskgroup()
2174 static void __kmp_task_reduction_fini(kmp_info_t *th, kmp_taskgroup_t *tg) {
2175   kmp_int32 nth = th->th.th_team_nproc;
2176   KMP_DEBUG_ASSERT(nth > 1); // should not be called if nth == 1
2177   kmp_task_red_data_t *arr = (kmp_task_red_data_t *)tg->reduce_data;
2178   kmp_int32 num = tg->reduce_num_data;
2179   for (int i = 0; i < num; ++i) {
2180     void *sh_data = arr[i].reduce_shar;
2181     void (*f_fini)(void *) = (void (*)(void *))(arr[i].reduce_fini);
2182     void (*f_comb)(void *, void *) =
2183         (void (*)(void *, void *))(arr[i].reduce_comb);
2184     if (!arr[i].flags.lazy_priv) {
2185       void *pr_data = arr[i].reduce_priv;
2186       size_t size = arr[i].reduce_size;
2187       for (int j = 0; j < nth; ++j) {
2188         void *priv_data = (char *)pr_data + j * size;
2189         f_comb(sh_data, priv_data); // combine results
2190         if (f_fini)
2191           f_fini(priv_data); // finalize if needed
2192       }
2193     } else {
2194       void **pr_data = (void **)(arr[i].reduce_priv);
2195       for (int j = 0; j < nth; ++j) {
2196         if (pr_data[j] != NULL) {
2197           f_comb(sh_data, pr_data[j]); // combine results
2198           if (f_fini)
2199             f_fini(pr_data[j]); // finalize if needed
2200           __kmp_free(pr_data[j]);
2201         }
2202       }
2203     }
2204     __kmp_free(arr[i].reduce_priv);
2205   }
2206   __kmp_thread_free(th, arr);
2207   tg->reduce_data = NULL;
2208   tg->reduce_num_data = 0;
2209 }
2210 #endif
2211 
2212 #if OMP_40_ENABLED
2213 // __kmpc_taskgroup: Start a new taskgroup
2214 void __kmpc_taskgroup(ident_t *loc, int gtid) {
2215   kmp_info_t *thread = __kmp_threads[gtid];
2216   kmp_taskdata_t *taskdata = thread->th.th_current_task;
2217   kmp_taskgroup_t *tg_new =
2218       (kmp_taskgroup_t *)__kmp_thread_malloc(thread, sizeof(kmp_taskgroup_t));
2219   KA_TRACE(10, ("__kmpc_taskgroup: T#%d loc=%p group=%p\n", gtid, loc, tg_new));
2220   KMP_ATOMIC_ST_RLX(&tg_new->count, 0);
2221   KMP_ATOMIC_ST_RLX(&tg_new->cancel_request, cancel_noreq);
2222   tg_new->parent = taskdata->td_taskgroup;
2223 #if OMP_50_ENABLED
2224   tg_new->reduce_data = NULL;
2225   tg_new->reduce_num_data = 0;
2226 #endif
2227   taskdata->td_taskgroup = tg_new;
2228 
2229 #if OMPT_SUPPORT && OMPT_OPTIONAL
2230   if (UNLIKELY(ompt_enabled.ompt_callback_sync_region)) {
2231     void *codeptr = OMPT_LOAD_RETURN_ADDRESS(gtid);
2232     if (!codeptr)
2233       codeptr = OMPT_GET_RETURN_ADDRESS(0);
2234     kmp_team_t *team = thread->th.th_team;
2235     ompt_data_t my_task_data = taskdata->ompt_task_info.task_data;
2236     // FIXME: I think this is wrong for lwt!
2237     ompt_data_t my_parallel_data = team->t.ompt_team_info.parallel_data;
2238 
2239     ompt_callbacks.ompt_callback(ompt_callback_sync_region)(
2240         ompt_sync_region_taskgroup, ompt_scope_begin, &(my_parallel_data),
2241         &(my_task_data), codeptr);
2242   }
2243 #endif
2244 }
2245 
2246 // __kmpc_end_taskgroup: Wait until all tasks generated by the current task
2247 //                       and its descendants are complete
2248 void __kmpc_end_taskgroup(ident_t *loc, int gtid) {
2249   kmp_info_t *thread = __kmp_threads[gtid];
2250   kmp_taskdata_t *taskdata = thread->th.th_current_task;
2251   kmp_taskgroup_t *taskgroup = taskdata->td_taskgroup;
2252   int thread_finished = FALSE;
2253 
2254 #if OMPT_SUPPORT && OMPT_OPTIONAL
2255   kmp_team_t *team;
2256   ompt_data_t my_task_data;
2257   ompt_data_t my_parallel_data;
2258   void *codeptr;
2259   if (UNLIKELY(ompt_enabled.enabled)) {
2260     team = thread->th.th_team;
2261     my_task_data = taskdata->ompt_task_info.task_data;
2262     // FIXME: I think this is wrong for lwt!
2263     my_parallel_data = team->t.ompt_team_info.parallel_data;
2264     codeptr = OMPT_LOAD_RETURN_ADDRESS(gtid);
2265     if (!codeptr)
2266       codeptr = OMPT_GET_RETURN_ADDRESS(0);
2267   }
2268 #endif
2269 
2270   KA_TRACE(10, ("__kmpc_end_taskgroup(enter): T#%d loc=%p\n", gtid, loc));
2271   KMP_DEBUG_ASSERT(taskgroup != NULL);
2272   KMP_SET_THREAD_STATE_BLOCK(TASKGROUP);
2273 
2274   if (__kmp_tasking_mode != tskm_immediate_exec) {
2275     // mark task as waiting not on a barrier
2276     taskdata->td_taskwait_counter += 1;
2277     taskdata->td_taskwait_ident = loc;
2278     taskdata->td_taskwait_thread = gtid + 1;
2279 #if USE_ITT_BUILD
2280     // For ITT the taskgroup wait is similar to taskwait until we need to
2281     // distinguish them
2282     void *itt_sync_obj = __kmp_itt_taskwait_object(gtid);
2283     if (itt_sync_obj != NULL)
2284       __kmp_itt_taskwait_starting(gtid, itt_sync_obj);
2285 #endif /* USE_ITT_BUILD */
2286 
2287 #if OMPT_SUPPORT && OMPT_OPTIONAL
2288     if (UNLIKELY(ompt_enabled.ompt_callback_sync_region_wait)) {
2289       ompt_callbacks.ompt_callback(ompt_callback_sync_region_wait)(
2290           ompt_sync_region_taskgroup, ompt_scope_begin, &(my_parallel_data),
2291           &(my_task_data), codeptr);
2292     }
2293 #endif
2294 
2295 #if OMP_45_ENABLED
2296     if (!taskdata->td_flags.team_serial ||
2297         (thread->th.th_task_team != NULL &&
2298          thread->th.th_task_team->tt.tt_found_proxy_tasks))
2299 #else
2300     if (!taskdata->td_flags.team_serial)
2301 #endif
2302     {
2303       kmp_flag_32 flag(RCAST(std::atomic<kmp_uint32> *, &(taskgroup->count)),
2304                        0U);
2305       while (KMP_ATOMIC_LD_ACQ(&taskgroup->count) != 0) {
2306         flag.execute_tasks(thread, gtid, FALSE,
2307                            &thread_finished USE_ITT_BUILD_ARG(itt_sync_obj),
2308                            __kmp_task_stealing_constraint);
2309       }
2310     }
2311     taskdata->td_taskwait_thread = -taskdata->td_taskwait_thread; // end waiting
2312 
2313 #if OMPT_SUPPORT && OMPT_OPTIONAL
2314     if (UNLIKELY(ompt_enabled.ompt_callback_sync_region_wait)) {
2315       ompt_callbacks.ompt_callback(ompt_callback_sync_region_wait)(
2316           ompt_sync_region_taskgroup, ompt_scope_end, &(my_parallel_data),
2317           &(my_task_data), codeptr);
2318     }
2319 #endif
2320 
2321 #if USE_ITT_BUILD
2322     if (itt_sync_obj != NULL)
2323       __kmp_itt_taskwait_finished(gtid, itt_sync_obj);
2324 #endif /* USE_ITT_BUILD */
2325   }
2326   KMP_DEBUG_ASSERT(taskgroup->count == 0);
2327 
2328 #if OMP_50_ENABLED
2329   if (taskgroup->reduce_data != NULL) // need to reduce?
2330     __kmp_task_reduction_fini(thread, taskgroup);
2331 #endif
2332   // Restore parent taskgroup for the current task
2333   taskdata->td_taskgroup = taskgroup->parent;
2334   __kmp_thread_free(thread, taskgroup);
2335 
2336   KA_TRACE(10, ("__kmpc_end_taskgroup(exit): T#%d task %p finished waiting\n",
2337                 gtid, taskdata));
2338   ANNOTATE_HAPPENS_AFTER(taskdata);
2339 
2340 #if OMPT_SUPPORT && OMPT_OPTIONAL
2341   if (UNLIKELY(ompt_enabled.ompt_callback_sync_region)) {
2342     ompt_callbacks.ompt_callback(ompt_callback_sync_region)(
2343         ompt_sync_region_taskgroup, ompt_scope_end, &(my_parallel_data),
2344         &(my_task_data), codeptr);
2345   }
2346 #endif
2347 }
2348 #endif
2349 
2350 // __kmp_remove_my_task: remove a task from my own deque
2351 static kmp_task_t *__kmp_remove_my_task(kmp_info_t *thread, kmp_int32 gtid,
2352                                         kmp_task_team_t *task_team,
2353                                         kmp_int32 is_constrained) {
2354   kmp_task_t *task;
2355   kmp_taskdata_t *taskdata;
2356   kmp_thread_data_t *thread_data;
2357   kmp_uint32 tail;
2358 
2359   KMP_DEBUG_ASSERT(__kmp_tasking_mode != tskm_immediate_exec);
2360   KMP_DEBUG_ASSERT(task_team->tt.tt_threads_data !=
2361                    NULL); // Caller should check this condition
2362 
2363   thread_data = &task_team->tt.tt_threads_data[__kmp_tid_from_gtid(gtid)];
2364 
2365   KA_TRACE(10, ("__kmp_remove_my_task(enter): T#%d ntasks=%d head=%u tail=%u\n",
2366                 gtid, thread_data->td.td_deque_ntasks,
2367                 thread_data->td.td_deque_head, thread_data->td.td_deque_tail));
2368 
2369   if (TCR_4(thread_data->td.td_deque_ntasks) == 0) {
2370     KA_TRACE(10,
2371              ("__kmp_remove_my_task(exit #1): T#%d No tasks to remove: "
2372               "ntasks=%d head=%u tail=%u\n",
2373               gtid, thread_data->td.td_deque_ntasks,
2374               thread_data->td.td_deque_head, thread_data->td.td_deque_tail));
2375     return NULL;
2376   }
2377 
2378   __kmp_acquire_bootstrap_lock(&thread_data->td.td_deque_lock);
2379 
2380   if (TCR_4(thread_data->td.td_deque_ntasks) == 0) {
2381     __kmp_release_bootstrap_lock(&thread_data->td.td_deque_lock);
2382     KA_TRACE(10,
2383              ("__kmp_remove_my_task(exit #2): T#%d No tasks to remove: "
2384               "ntasks=%d head=%u tail=%u\n",
2385               gtid, thread_data->td.td_deque_ntasks,
2386               thread_data->td.td_deque_head, thread_data->td.td_deque_tail));
2387     return NULL;
2388   }
2389 
2390   tail = (thread_data->td.td_deque_tail - 1) &
2391          TASK_DEQUE_MASK(thread_data->td); // Wrap index.
2392   taskdata = thread_data->td.td_deque[tail];
2393 
2394   if (!__kmp_task_is_allowed(gtid, is_constrained, taskdata,
2395                              thread->th.th_current_task)) {
2396     // The TSC does not allow to steal victim task
2397     __kmp_release_bootstrap_lock(&thread_data->td.td_deque_lock);
2398     KA_TRACE(10,
2399              ("__kmp_remove_my_task(exit #3): T#%d TSC blocks tail task: "
2400               "ntasks=%d head=%u tail=%u\n",
2401               gtid, thread_data->td.td_deque_ntasks,
2402               thread_data->td.td_deque_head, thread_data->td.td_deque_tail));
2403     return NULL;
2404   }
2405 
2406   thread_data->td.td_deque_tail = tail;
2407   TCW_4(thread_data->td.td_deque_ntasks, thread_data->td.td_deque_ntasks - 1);
2408 
2409   __kmp_release_bootstrap_lock(&thread_data->td.td_deque_lock);
2410 
2411   KA_TRACE(10, ("__kmp_remove_my_task(exit #4): T#%d task %p removed: "
2412                 "ntasks=%d head=%u tail=%u\n",
2413                 gtid, taskdata, thread_data->td.td_deque_ntasks,
2414                 thread_data->td.td_deque_head, thread_data->td.td_deque_tail));
2415 
2416   task = KMP_TASKDATA_TO_TASK(taskdata);
2417   return task;
2418 }
2419 
2420 // __kmp_steal_task: remove a task from another thread's deque
2421 // Assume that calling thread has already checked existence of
2422 // task_team thread_data before calling this routine.
2423 static kmp_task_t *__kmp_steal_task(kmp_info_t *victim_thr, kmp_int32 gtid,
2424                                     kmp_task_team_t *task_team,
2425                                     std::atomic<kmp_int32> *unfinished_threads,
2426                                     int *thread_finished,
2427                                     kmp_int32 is_constrained) {
2428   kmp_task_t *task;
2429   kmp_taskdata_t *taskdata;
2430   kmp_taskdata_t *current;
2431   kmp_thread_data_t *victim_td, *threads_data;
2432   kmp_int32 target;
2433   kmp_int32 victim_tid;
2434 
2435   KMP_DEBUG_ASSERT(__kmp_tasking_mode != tskm_immediate_exec);
2436 
2437   threads_data = task_team->tt.tt_threads_data;
2438   KMP_DEBUG_ASSERT(threads_data != NULL); // Caller should check this condition
2439 
2440   victim_tid = victim_thr->th.th_info.ds.ds_tid;
2441   victim_td = &threads_data[victim_tid];
2442 
2443   KA_TRACE(10, ("__kmp_steal_task(enter): T#%d try to steal from T#%d: "
2444                 "task_team=%p ntasks=%d head=%u tail=%u\n",
2445                 gtid, __kmp_gtid_from_thread(victim_thr), task_team,
2446                 victim_td->td.td_deque_ntasks, victim_td->td.td_deque_head,
2447                 victim_td->td.td_deque_tail));
2448 
2449   if (TCR_4(victim_td->td.td_deque_ntasks) == 0) {
2450     KA_TRACE(10, ("__kmp_steal_task(exit #1): T#%d could not steal from T#%d: "
2451                   "task_team=%p ntasks=%d head=%u tail=%u\n",
2452                   gtid, __kmp_gtid_from_thread(victim_thr), task_team,
2453                   victim_td->td.td_deque_ntasks, victim_td->td.td_deque_head,
2454                   victim_td->td.td_deque_tail));
2455     return NULL;
2456   }
2457 
2458   __kmp_acquire_bootstrap_lock(&victim_td->td.td_deque_lock);
2459 
2460   int ntasks = TCR_4(victim_td->td.td_deque_ntasks);
2461   // Check again after we acquire the lock
2462   if (ntasks == 0) {
2463     __kmp_release_bootstrap_lock(&victim_td->td.td_deque_lock);
2464     KA_TRACE(10, ("__kmp_steal_task(exit #2): T#%d could not steal from T#%d: "
2465                   "task_team=%p ntasks=%d head=%u tail=%u\n",
2466                   gtid, __kmp_gtid_from_thread(victim_thr), task_team, ntasks,
2467                   victim_td->td.td_deque_head, victim_td->td.td_deque_tail));
2468     return NULL;
2469   }
2470 
2471   KMP_DEBUG_ASSERT(victim_td->td.td_deque != NULL);
2472   current = __kmp_threads[gtid]->th.th_current_task;
2473   taskdata = victim_td->td.td_deque[victim_td->td.td_deque_head];
2474   if (__kmp_task_is_allowed(gtid, is_constrained, taskdata, current)) {
2475     // Bump head pointer and Wrap.
2476     victim_td->td.td_deque_head =
2477         (victim_td->td.td_deque_head + 1) & TASK_DEQUE_MASK(victim_td->td);
2478   } else {
2479     if (!task_team->tt.tt_untied_task_encountered) {
2480       // The TSC does not allow to steal victim task
2481       __kmp_release_bootstrap_lock(&victim_td->td.td_deque_lock);
2482       KA_TRACE(10, ("__kmp_steal_task(exit #3): T#%d could not steal from "
2483                     "T#%d: task_team=%p ntasks=%d head=%u tail=%u\n",
2484                     gtid, __kmp_gtid_from_thread(victim_thr), task_team, ntasks,
2485                     victim_td->td.td_deque_head, victim_td->td.td_deque_tail));
2486       return NULL;
2487     }
2488     int i;
2489     // walk through victim's deque trying to steal any task
2490     target = victim_td->td.td_deque_head;
2491     taskdata = NULL;
2492     for (i = 1; i < ntasks; ++i) {
2493       target = (target + 1) & TASK_DEQUE_MASK(victim_td->td);
2494       taskdata = victim_td->td.td_deque[target];
2495       if (__kmp_task_is_allowed(gtid, is_constrained, taskdata, current)) {
2496         break; // found victim task
2497       } else {
2498         taskdata = NULL;
2499       }
2500     }
2501     if (taskdata == NULL) {
2502       // No appropriate candidate to steal found
2503       __kmp_release_bootstrap_lock(&victim_td->td.td_deque_lock);
2504       KA_TRACE(10, ("__kmp_steal_task(exit #4): T#%d could not steal from "
2505                     "T#%d: task_team=%p ntasks=%d head=%u tail=%u\n",
2506                     gtid, __kmp_gtid_from_thread(victim_thr), task_team, ntasks,
2507                     victim_td->td.td_deque_head, victim_td->td.td_deque_tail));
2508       return NULL;
2509     }
2510     int prev = target;
2511     for (i = i + 1; i < ntasks; ++i) {
2512       // shift remaining tasks in the deque left by 1
2513       target = (target + 1) & TASK_DEQUE_MASK(victim_td->td);
2514       victim_td->td.td_deque[prev] = victim_td->td.td_deque[target];
2515       prev = target;
2516     }
2517     KMP_DEBUG_ASSERT(
2518         victim_td->td.td_deque_tail ==
2519         (kmp_uint32)((target + 1) & TASK_DEQUE_MASK(victim_td->td)));
2520     victim_td->td.td_deque_tail = target; // tail -= 1 (wrapped))
2521   }
2522   if (*thread_finished) {
2523     // We need to un-mark this victim as a finished victim.  This must be done
2524     // before releasing the lock, or else other threads (starting with the
2525     // master victim) might be prematurely released from the barrier!!!
2526     kmp_int32 count;
2527 
2528     count = KMP_ATOMIC_INC(unfinished_threads);
2529 
2530     KA_TRACE(
2531         20,
2532         ("__kmp_steal_task: T#%d inc unfinished_threads to %d: task_team=%p\n",
2533          gtid, count + 1, task_team));
2534 
2535     *thread_finished = FALSE;
2536   }
2537   TCW_4(victim_td->td.td_deque_ntasks, ntasks - 1);
2538 
2539   __kmp_release_bootstrap_lock(&victim_td->td.td_deque_lock);
2540 
2541   KMP_COUNT_BLOCK(TASK_stolen);
2542   KA_TRACE(10,
2543            ("__kmp_steal_task(exit #5): T#%d stole task %p from T#%d: "
2544             "task_team=%p ntasks=%d head=%u tail=%u\n",
2545             gtid, taskdata, __kmp_gtid_from_thread(victim_thr), task_team,
2546             ntasks, victim_td->td.td_deque_head, victim_td->td.td_deque_tail));
2547 
2548   task = KMP_TASKDATA_TO_TASK(taskdata);
2549   return task;
2550 }
2551 
2552 // __kmp_execute_tasks_template: Choose and execute tasks until either the
2553 // condition is statisfied (return true) or there are none left (return false).
2554 //
2555 // final_spin is TRUE if this is the spin at the release barrier.
2556 // thread_finished indicates whether the thread is finished executing all
2557 // the tasks it has on its deque, and is at the release barrier.
2558 // spinner is the location on which to spin.
2559 // spinner == NULL means only execute a single task and return.
2560 // checker is the value to check to terminate the spin.
2561 template <class C>
2562 static inline int __kmp_execute_tasks_template(
2563     kmp_info_t *thread, kmp_int32 gtid, C *flag, int final_spin,
2564     int *thread_finished USE_ITT_BUILD_ARG(void *itt_sync_obj),
2565     kmp_int32 is_constrained) {
2566   kmp_task_team_t *task_team = thread->th.th_task_team;
2567   kmp_thread_data_t *threads_data;
2568   kmp_task_t *task;
2569   kmp_info_t *other_thread;
2570   kmp_taskdata_t *current_task = thread->th.th_current_task;
2571   std::atomic<kmp_int32> *unfinished_threads;
2572   kmp_int32 nthreads, victim_tid = -2, use_own_tasks = 1, new_victim = 0,
2573                       tid = thread->th.th_info.ds.ds_tid;
2574 
2575   KMP_DEBUG_ASSERT(__kmp_tasking_mode != tskm_immediate_exec);
2576   KMP_DEBUG_ASSERT(thread == __kmp_threads[gtid]);
2577 
2578   if (task_team == NULL || current_task == NULL)
2579     return FALSE;
2580 
2581   KA_TRACE(15, ("__kmp_execute_tasks_template(enter): T#%d final_spin=%d "
2582                 "*thread_finished=%d\n",
2583                 gtid, final_spin, *thread_finished));
2584 
2585   thread->th.th_reap_state = KMP_NOT_SAFE_TO_REAP;
2586   threads_data = (kmp_thread_data_t *)TCR_PTR(task_team->tt.tt_threads_data);
2587   KMP_DEBUG_ASSERT(threads_data != NULL);
2588 
2589   nthreads = task_team->tt.tt_nproc;
2590   unfinished_threads = &(task_team->tt.tt_unfinished_threads);
2591 #if OMP_45_ENABLED
2592   KMP_DEBUG_ASSERT(nthreads > 1 || task_team->tt.tt_found_proxy_tasks);
2593 #else
2594   KMP_DEBUG_ASSERT(nthreads > 1);
2595 #endif
2596   KMP_DEBUG_ASSERT(*unfinished_threads >= 0);
2597 
2598   while (1) { // Outer loop keeps trying to find tasks in case of single thread
2599     // getting tasks from target constructs
2600     while (1) { // Inner loop to find a task and execute it
2601       task = NULL;
2602       if (use_own_tasks) { // check on own queue first
2603         task = __kmp_remove_my_task(thread, gtid, task_team, is_constrained);
2604       }
2605       if ((task == NULL) && (nthreads > 1)) { // Steal a task
2606         int asleep = 1;
2607         use_own_tasks = 0;
2608         // Try to steal from the last place I stole from successfully.
2609         if (victim_tid == -2) { // haven't stolen anything yet
2610           victim_tid = threads_data[tid].td.td_deque_last_stolen;
2611           if (victim_tid !=
2612               -1) // if we have a last stolen from victim, get the thread
2613             other_thread = threads_data[victim_tid].td.td_thr;
2614         }
2615         if (victim_tid != -1) { // found last victim
2616           asleep = 0;
2617         } else if (!new_victim) { // no recent steals and we haven't already
2618           // used a new victim; select a random thread
2619           do { // Find a different thread to steal work from.
2620             // Pick a random thread. Initial plan was to cycle through all the
2621             // threads, and only return if we tried to steal from every thread,
2622             // and failed.  Arch says that's not such a great idea.
2623             victim_tid = __kmp_get_random(thread) % (nthreads - 1);
2624             if (victim_tid >= tid) {
2625               ++victim_tid; // Adjusts random distribution to exclude self
2626             }
2627             // Found a potential victim
2628             other_thread = threads_data[victim_tid].td.td_thr;
2629             // There is a slight chance that __kmp_enable_tasking() did not wake
2630             // up all threads waiting at the barrier.  If victim is sleeping,
2631             // then wake it up. Since we were going to pay the cache miss
2632             // penalty for referencing another thread's kmp_info_t struct
2633             // anyway,
2634             // the check shouldn't cost too much performance at this point. In
2635             // extra barrier mode, tasks do not sleep at the separate tasking
2636             // barrier, so this isn't a problem.
2637             asleep = 0;
2638             if ((__kmp_tasking_mode == tskm_task_teams) &&
2639                 (__kmp_dflt_blocktime != KMP_MAX_BLOCKTIME) &&
2640                 (TCR_PTR(CCAST(void *, other_thread->th.th_sleep_loc)) !=
2641                  NULL)) {
2642               asleep = 1;
2643               __kmp_null_resume_wrapper(__kmp_gtid_from_thread(other_thread),
2644                                         other_thread->th.th_sleep_loc);
2645               // A sleeping thread should not have any tasks on it's queue.
2646               // There is a slight possibility that it resumes, steals a task
2647               // from another thread, which spawns more tasks, all in the time
2648               // that it takes this thread to check => don't write an assertion
2649               // that the victim's queue is empty.  Try stealing from a
2650               // different thread.
2651             }
2652           } while (asleep);
2653         }
2654 
2655         if (!asleep) {
2656           // We have a victim to try to steal from
2657           task = __kmp_steal_task(other_thread, gtid, task_team,
2658                                   unfinished_threads, thread_finished,
2659                                   is_constrained);
2660         }
2661         if (task != NULL) { // set last stolen to victim
2662           if (threads_data[tid].td.td_deque_last_stolen != victim_tid) {
2663             threads_data[tid].td.td_deque_last_stolen = victim_tid;
2664             // The pre-refactored code did not try more than 1 successful new
2665             // vicitm, unless the last one generated more local tasks;
2666             // new_victim keeps track of this
2667             new_victim = 1;
2668           }
2669         } else { // No tasks found; unset last_stolen
2670           KMP_CHECK_UPDATE(threads_data[tid].td.td_deque_last_stolen, -1);
2671           victim_tid = -2; // no successful victim found
2672         }
2673       }
2674 
2675       if (task == NULL) // break out of tasking loop
2676         break;
2677 
2678 // Found a task; execute it
2679 #if USE_ITT_BUILD && USE_ITT_NOTIFY
2680       if (__itt_sync_create_ptr || KMP_ITT_DEBUG) {
2681         if (itt_sync_obj == NULL) { // we are at fork barrier where we could not
2682           // get the object reliably
2683           itt_sync_obj = __kmp_itt_barrier_object(gtid, bs_forkjoin_barrier);
2684         }
2685         __kmp_itt_task_starting(itt_sync_obj);
2686       }
2687 #endif /* USE_ITT_BUILD && USE_ITT_NOTIFY */
2688       __kmp_invoke_task(gtid, task, current_task);
2689 #if USE_ITT_BUILD
2690       if (itt_sync_obj != NULL)
2691         __kmp_itt_task_finished(itt_sync_obj);
2692 #endif /* USE_ITT_BUILD */
2693       // If this thread is only partway through the barrier and the condition is
2694       // met, then return now, so that the barrier gather/release pattern can
2695       // proceed. If this thread is in the last spin loop in the barrier,
2696       // waiting to be released, we know that the termination condition will not
2697       // be satisified, so don't waste any cycles checking it.
2698       if (flag == NULL || (!final_spin && flag->done_check())) {
2699         KA_TRACE(
2700             15,
2701             ("__kmp_execute_tasks_template: T#%d spin condition satisfied\n",
2702              gtid));
2703         return TRUE;
2704       }
2705       if (thread->th.th_task_team == NULL) {
2706         break;
2707       }
2708       KMP_YIELD(__kmp_library == library_throughput); // Yield before next task
2709       // If execution of a stolen task results in more tasks being placed on our
2710       // run queue, reset use_own_tasks
2711       if (!use_own_tasks && TCR_4(threads_data[tid].td.td_deque_ntasks) != 0) {
2712         KA_TRACE(20, ("__kmp_execute_tasks_template: T#%d stolen task spawned "
2713                       "other tasks, restart\n",
2714                       gtid));
2715         use_own_tasks = 1;
2716         new_victim = 0;
2717       }
2718     }
2719 
2720 // The task source has been exhausted. If in final spin loop of barrier, check
2721 // if termination condition is satisfied.
2722 #if OMP_45_ENABLED
2723     // The work queue may be empty but there might be proxy tasks still
2724     // executing
2725     if (final_spin &&
2726         KMP_ATOMIC_LD_ACQ(&current_task->td_incomplete_child_tasks) == 0)
2727 #else
2728     if (final_spin)
2729 #endif
2730     {
2731       // First, decrement the #unfinished threads, if that has not already been
2732       // done.  This decrement might be to the spin location, and result in the
2733       // termination condition being satisfied.
2734       if (!*thread_finished) {
2735         kmp_int32 count;
2736 
2737         count = KMP_ATOMIC_DEC(unfinished_threads) - 1;
2738         KA_TRACE(20, ("__kmp_execute_tasks_template: T#%d dec "
2739                       "unfinished_threads to %d task_team=%p\n",
2740                       gtid, count, task_team));
2741         *thread_finished = TRUE;
2742       }
2743 
2744       // It is now unsafe to reference thread->th.th_team !!!
2745       // Decrementing task_team->tt.tt_unfinished_threads can allow the master
2746       // thread to pass through the barrier, where it might reset each thread's
2747       // th.th_team field for the next parallel region. If we can steal more
2748       // work, we know that this has not happened yet.
2749       if (flag != NULL && flag->done_check()) {
2750         KA_TRACE(
2751             15,
2752             ("__kmp_execute_tasks_template: T#%d spin condition satisfied\n",
2753              gtid));
2754         return TRUE;
2755       }
2756     }
2757 
2758     // If this thread's task team is NULL, master has recognized that there are
2759     // no more tasks; bail out
2760     if (thread->th.th_task_team == NULL) {
2761       KA_TRACE(15,
2762                ("__kmp_execute_tasks_template: T#%d no more tasks\n", gtid));
2763       return FALSE;
2764     }
2765 
2766 #if OMP_45_ENABLED
2767     // We could be getting tasks from target constructs; if this is the only
2768     // thread, keep trying to execute tasks from own queue
2769     if (nthreads == 1)
2770       use_own_tasks = 1;
2771     else
2772 #endif
2773     {
2774       KA_TRACE(15,
2775                ("__kmp_execute_tasks_template: T#%d can't find work\n", gtid));
2776       return FALSE;
2777     }
2778   }
2779 }
2780 
2781 int __kmp_execute_tasks_32(
2782     kmp_info_t *thread, kmp_int32 gtid, kmp_flag_32 *flag, int final_spin,
2783     int *thread_finished USE_ITT_BUILD_ARG(void *itt_sync_obj),
2784     kmp_int32 is_constrained) {
2785   return __kmp_execute_tasks_template(
2786       thread, gtid, flag, final_spin,
2787       thread_finished USE_ITT_BUILD_ARG(itt_sync_obj), is_constrained);
2788 }
2789 
2790 int __kmp_execute_tasks_64(
2791     kmp_info_t *thread, kmp_int32 gtid, kmp_flag_64 *flag, int final_spin,
2792     int *thread_finished USE_ITT_BUILD_ARG(void *itt_sync_obj),
2793     kmp_int32 is_constrained) {
2794   return __kmp_execute_tasks_template(
2795       thread, gtid, flag, final_spin,
2796       thread_finished USE_ITT_BUILD_ARG(itt_sync_obj), is_constrained);
2797 }
2798 
2799 int __kmp_execute_tasks_oncore(
2800     kmp_info_t *thread, kmp_int32 gtid, kmp_flag_oncore *flag, int final_spin,
2801     int *thread_finished USE_ITT_BUILD_ARG(void *itt_sync_obj),
2802     kmp_int32 is_constrained) {
2803   return __kmp_execute_tasks_template(
2804       thread, gtid, flag, final_spin,
2805       thread_finished USE_ITT_BUILD_ARG(itt_sync_obj), is_constrained);
2806 }
2807 
2808 // __kmp_enable_tasking: Allocate task team and resume threads sleeping at the
2809 // next barrier so they can assist in executing enqueued tasks.
2810 // First thread in allocates the task team atomically.
2811 static void __kmp_enable_tasking(kmp_task_team_t *task_team,
2812                                  kmp_info_t *this_thr) {
2813   kmp_thread_data_t *threads_data;
2814   int nthreads, i, is_init_thread;
2815 
2816   KA_TRACE(10, ("__kmp_enable_tasking(enter): T#%d\n",
2817                 __kmp_gtid_from_thread(this_thr)));
2818 
2819   KMP_DEBUG_ASSERT(task_team != NULL);
2820   KMP_DEBUG_ASSERT(this_thr->th.th_team != NULL);
2821 
2822   nthreads = task_team->tt.tt_nproc;
2823   KMP_DEBUG_ASSERT(nthreads > 0);
2824   KMP_DEBUG_ASSERT(nthreads == this_thr->th.th_team->t.t_nproc);
2825 
2826   // Allocate or increase the size of threads_data if necessary
2827   is_init_thread = __kmp_realloc_task_threads_data(this_thr, task_team);
2828 
2829   if (!is_init_thread) {
2830     // Some other thread already set up the array.
2831     KA_TRACE(
2832         20,
2833         ("__kmp_enable_tasking(exit): T#%d: threads array already set up.\n",
2834          __kmp_gtid_from_thread(this_thr)));
2835     return;
2836   }
2837   threads_data = (kmp_thread_data_t *)TCR_PTR(task_team->tt.tt_threads_data);
2838   KMP_DEBUG_ASSERT(threads_data != NULL);
2839 
2840   if (__kmp_tasking_mode == tskm_task_teams &&
2841       (__kmp_dflt_blocktime != KMP_MAX_BLOCKTIME)) {
2842     // Release any threads sleeping at the barrier, so that they can steal
2843     // tasks and execute them.  In extra barrier mode, tasks do not sleep
2844     // at the separate tasking barrier, so this isn't a problem.
2845     for (i = 0; i < nthreads; i++) {
2846       volatile void *sleep_loc;
2847       kmp_info_t *thread = threads_data[i].td.td_thr;
2848 
2849       if (i == this_thr->th.th_info.ds.ds_tid) {
2850         continue;
2851       }
2852       // Since we haven't locked the thread's suspend mutex lock at this
2853       // point, there is a small window where a thread might be putting
2854       // itself to sleep, but hasn't set the th_sleep_loc field yet.
2855       // To work around this, __kmp_execute_tasks_template() periodically checks
2856       // see if other threads are sleeping (using the same random mechanism that
2857       // is used for task stealing) and awakens them if they are.
2858       if ((sleep_loc = TCR_PTR(CCAST(void *, thread->th.th_sleep_loc))) !=
2859           NULL) {
2860         KF_TRACE(50, ("__kmp_enable_tasking: T#%d waking up thread T#%d\n",
2861                       __kmp_gtid_from_thread(this_thr),
2862                       __kmp_gtid_from_thread(thread)));
2863         __kmp_null_resume_wrapper(__kmp_gtid_from_thread(thread), sleep_loc);
2864       } else {
2865         KF_TRACE(50, ("__kmp_enable_tasking: T#%d don't wake up thread T#%d\n",
2866                       __kmp_gtid_from_thread(this_thr),
2867                       __kmp_gtid_from_thread(thread)));
2868       }
2869     }
2870   }
2871 
2872   KA_TRACE(10, ("__kmp_enable_tasking(exit): T#%d\n",
2873                 __kmp_gtid_from_thread(this_thr)));
2874 }
2875 
2876 /* // TODO: Check the comment consistency
2877  * Utility routines for "task teams".  A task team (kmp_task_t) is kind of
2878  * like a shadow of the kmp_team_t data struct, with a different lifetime.
2879  * After a child * thread checks into a barrier and calls __kmp_release() from
2880  * the particular variant of __kmp_<barrier_kind>_barrier_gather(), it can no
2881  * longer assume that the kmp_team_t structure is intact (at any moment, the
2882  * master thread may exit the barrier code and free the team data structure,
2883  * and return the threads to the thread pool).
2884  *
2885  * This does not work with the the tasking code, as the thread is still
2886  * expected to participate in the execution of any tasks that may have been
2887  * spawned my a member of the team, and the thread still needs access to all
2888  * to each thread in the team, so that it can steal work from it.
2889  *
2890  * Enter the existence of the kmp_task_team_t struct.  It employs a reference
2891  * counting mechanims, and is allocated by the master thread before calling
2892  * __kmp_<barrier_kind>_release, and then is release by the last thread to
2893  * exit __kmp_<barrier_kind>_release at the next barrier.  I.e. the lifetimes
2894  * of the kmp_task_team_t structs for consecutive barriers can overlap
2895  * (and will, unless the master thread is the last thread to exit the barrier
2896  * release phase, which is not typical).
2897  *
2898  * The existence of such a struct is useful outside the context of tasking,
2899  * but for now, I'm trying to keep it specific to the OMP_30_ENABLED macro,
2900  * so that any performance differences show up when comparing the 2.5 vs. 3.0
2901  * libraries.
2902  *
2903  * We currently use the existence of the threads array as an indicator that
2904  * tasks were spawned since the last barrier.  If the structure is to be
2905  * useful outside the context of tasking, then this will have to change, but
2906  * not settting the field minimizes the performance impact of tasking on
2907  * barriers, when no explicit tasks were spawned (pushed, actually).
2908  */
2909 
2910 static kmp_task_team_t *__kmp_free_task_teams =
2911     NULL; // Free list for task_team data structures
2912 // Lock for task team data structures
2913 kmp_bootstrap_lock_t __kmp_task_team_lock =
2914     KMP_BOOTSTRAP_LOCK_INITIALIZER(__kmp_task_team_lock);
2915 
2916 // __kmp_alloc_task_deque:
2917 // Allocates a task deque for a particular thread, and initialize the necessary
2918 // data structures relating to the deque.  This only happens once per thread
2919 // per task team since task teams are recycled. No lock is needed during
2920 // allocation since each thread allocates its own deque.
2921 static void __kmp_alloc_task_deque(kmp_info_t *thread,
2922                                    kmp_thread_data_t *thread_data) {
2923   __kmp_init_bootstrap_lock(&thread_data->td.td_deque_lock);
2924   KMP_DEBUG_ASSERT(thread_data->td.td_deque == NULL);
2925 
2926   // Initialize last stolen task field to "none"
2927   thread_data->td.td_deque_last_stolen = -1;
2928 
2929   KMP_DEBUG_ASSERT(TCR_4(thread_data->td.td_deque_ntasks) == 0);
2930   KMP_DEBUG_ASSERT(thread_data->td.td_deque_head == 0);
2931   KMP_DEBUG_ASSERT(thread_data->td.td_deque_tail == 0);
2932 
2933   KE_TRACE(
2934       10,
2935       ("__kmp_alloc_task_deque: T#%d allocating deque[%d] for thread_data %p\n",
2936        __kmp_gtid_from_thread(thread), INITIAL_TASK_DEQUE_SIZE, thread_data));
2937   // Allocate space for task deque, and zero the deque
2938   // Cannot use __kmp_thread_calloc() because threads not around for
2939   // kmp_reap_task_team( ).
2940   thread_data->td.td_deque = (kmp_taskdata_t **)__kmp_allocate(
2941       INITIAL_TASK_DEQUE_SIZE * sizeof(kmp_taskdata_t *));
2942   thread_data->td.td_deque_size = INITIAL_TASK_DEQUE_SIZE;
2943 }
2944 
2945 // __kmp_free_task_deque:
2946 // Deallocates a task deque for a particular thread. Happens at library
2947 // deallocation so don't need to reset all thread data fields.
2948 static void __kmp_free_task_deque(kmp_thread_data_t *thread_data) {
2949   if (thread_data->td.td_deque != NULL) {
2950     __kmp_acquire_bootstrap_lock(&thread_data->td.td_deque_lock);
2951     TCW_4(thread_data->td.td_deque_ntasks, 0);
2952     __kmp_free(thread_data->td.td_deque);
2953     thread_data->td.td_deque = NULL;
2954     __kmp_release_bootstrap_lock(&thread_data->td.td_deque_lock);
2955   }
2956 
2957 #ifdef BUILD_TIED_TASK_STACK
2958   // GEH: Figure out what to do here for td_susp_tied_tasks
2959   if (thread_data->td.td_susp_tied_tasks.ts_entries != TASK_STACK_EMPTY) {
2960     __kmp_free_task_stack(__kmp_thread_from_gtid(gtid), thread_data);
2961   }
2962 #endif // BUILD_TIED_TASK_STACK
2963 }
2964 
2965 // __kmp_realloc_task_threads_data:
2966 // Allocates a threads_data array for a task team, either by allocating an
2967 // initial array or enlarging an existing array.  Only the first thread to get
2968 // the lock allocs or enlarges the array and re-initializes the array eleemnts.
2969 // That thread returns "TRUE", the rest return "FALSE".
2970 // Assumes that the new array size is given by task_team -> tt.tt_nproc.
2971 // The current size is given by task_team -> tt.tt_max_threads.
2972 static int __kmp_realloc_task_threads_data(kmp_info_t *thread,
2973                                            kmp_task_team_t *task_team) {
2974   kmp_thread_data_t **threads_data_p;
2975   kmp_int32 nthreads, maxthreads;
2976   int is_init_thread = FALSE;
2977 
2978   if (TCR_4(task_team->tt.tt_found_tasks)) {
2979     // Already reallocated and initialized.
2980     return FALSE;
2981   }
2982 
2983   threads_data_p = &task_team->tt.tt_threads_data;
2984   nthreads = task_team->tt.tt_nproc;
2985   maxthreads = task_team->tt.tt_max_threads;
2986 
2987   // All threads must lock when they encounter the first task of the implicit
2988   // task region to make sure threads_data fields are (re)initialized before
2989   // used.
2990   __kmp_acquire_bootstrap_lock(&task_team->tt.tt_threads_lock);
2991 
2992   if (!TCR_4(task_team->tt.tt_found_tasks)) {
2993     // first thread to enable tasking
2994     kmp_team_t *team = thread->th.th_team;
2995     int i;
2996 
2997     is_init_thread = TRUE;
2998     if (maxthreads < nthreads) {
2999 
3000       if (*threads_data_p != NULL) {
3001         kmp_thread_data_t *old_data = *threads_data_p;
3002         kmp_thread_data_t *new_data = NULL;
3003 
3004         KE_TRACE(
3005             10,
3006             ("__kmp_realloc_task_threads_data: T#%d reallocating "
3007              "threads data for task_team %p, new_size = %d, old_size = %d\n",
3008              __kmp_gtid_from_thread(thread), task_team, nthreads, maxthreads));
3009         // Reallocate threads_data to have more elements than current array
3010         // Cannot use __kmp_thread_realloc() because threads not around for
3011         // kmp_reap_task_team( ).  Note all new array entries are initialized
3012         // to zero by __kmp_allocate().
3013         new_data = (kmp_thread_data_t *)__kmp_allocate(
3014             nthreads * sizeof(kmp_thread_data_t));
3015         // copy old data to new data
3016         KMP_MEMCPY_S((void *)new_data, nthreads * sizeof(kmp_thread_data_t),
3017                      (void *)old_data, maxthreads * sizeof(kmp_thread_data_t));
3018 
3019 #ifdef BUILD_TIED_TASK_STACK
3020         // GEH: Figure out if this is the right thing to do
3021         for (i = maxthreads; i < nthreads; i++) {
3022           kmp_thread_data_t *thread_data = &(*threads_data_p)[i];
3023           __kmp_init_task_stack(__kmp_gtid_from_thread(thread), thread_data);
3024         }
3025 #endif // BUILD_TIED_TASK_STACK
3026         // Install the new data and free the old data
3027         (*threads_data_p) = new_data;
3028         __kmp_free(old_data);
3029       } else {
3030         KE_TRACE(10, ("__kmp_realloc_task_threads_data: T#%d allocating "
3031                       "threads data for task_team %p, size = %d\n",
3032                       __kmp_gtid_from_thread(thread), task_team, nthreads));
3033         // Make the initial allocate for threads_data array, and zero entries
3034         // Cannot use __kmp_thread_calloc() because threads not around for
3035         // kmp_reap_task_team( ).
3036         ANNOTATE_IGNORE_WRITES_BEGIN();
3037         *threads_data_p = (kmp_thread_data_t *)__kmp_allocate(
3038             nthreads * sizeof(kmp_thread_data_t));
3039         ANNOTATE_IGNORE_WRITES_END();
3040 #ifdef BUILD_TIED_TASK_STACK
3041         // GEH: Figure out if this is the right thing to do
3042         for (i = 0; i < nthreads; i++) {
3043           kmp_thread_data_t *thread_data = &(*threads_data_p)[i];
3044           __kmp_init_task_stack(__kmp_gtid_from_thread(thread), thread_data);
3045         }
3046 #endif // BUILD_TIED_TASK_STACK
3047       }
3048       task_team->tt.tt_max_threads = nthreads;
3049     } else {
3050       // If array has (more than) enough elements, go ahead and use it
3051       KMP_DEBUG_ASSERT(*threads_data_p != NULL);
3052     }
3053 
3054     // initialize threads_data pointers back to thread_info structures
3055     for (i = 0; i < nthreads; i++) {
3056       kmp_thread_data_t *thread_data = &(*threads_data_p)[i];
3057       thread_data->td.td_thr = team->t.t_threads[i];
3058 
3059       if (thread_data->td.td_deque_last_stolen >= nthreads) {
3060         // The last stolen field survives across teams / barrier, and the number
3061         // of threads may have changed.  It's possible (likely?) that a new
3062         // parallel region will exhibit the same behavior as previous region.
3063         thread_data->td.td_deque_last_stolen = -1;
3064       }
3065     }
3066 
3067     KMP_MB();
3068     TCW_SYNC_4(task_team->tt.tt_found_tasks, TRUE);
3069   }
3070 
3071   __kmp_release_bootstrap_lock(&task_team->tt.tt_threads_lock);
3072   return is_init_thread;
3073 }
3074 
3075 // __kmp_free_task_threads_data:
3076 // Deallocates a threads_data array for a task team, including any attached
3077 // tasking deques.  Only occurs at library shutdown.
3078 static void __kmp_free_task_threads_data(kmp_task_team_t *task_team) {
3079   __kmp_acquire_bootstrap_lock(&task_team->tt.tt_threads_lock);
3080   if (task_team->tt.tt_threads_data != NULL) {
3081     int i;
3082     for (i = 0; i < task_team->tt.tt_max_threads; i++) {
3083       __kmp_free_task_deque(&task_team->tt.tt_threads_data[i]);
3084     }
3085     __kmp_free(task_team->tt.tt_threads_data);
3086     task_team->tt.tt_threads_data = NULL;
3087   }
3088   __kmp_release_bootstrap_lock(&task_team->tt.tt_threads_lock);
3089 }
3090 
3091 // __kmp_allocate_task_team:
3092 // Allocates a task team associated with a specific team, taking it from
3093 // the global task team free list if possible.  Also initializes data
3094 // structures.
3095 static kmp_task_team_t *__kmp_allocate_task_team(kmp_info_t *thread,
3096                                                  kmp_team_t *team) {
3097   kmp_task_team_t *task_team = NULL;
3098   int nthreads;
3099 
3100   KA_TRACE(20, ("__kmp_allocate_task_team: T#%d entering; team = %p\n",
3101                 (thread ? __kmp_gtid_from_thread(thread) : -1), team));
3102 
3103   if (TCR_PTR(__kmp_free_task_teams) != NULL) {
3104     // Take a task team from the task team pool
3105     __kmp_acquire_bootstrap_lock(&__kmp_task_team_lock);
3106     if (__kmp_free_task_teams != NULL) {
3107       task_team = __kmp_free_task_teams;
3108       TCW_PTR(__kmp_free_task_teams, task_team->tt.tt_next);
3109       task_team->tt.tt_next = NULL;
3110     }
3111     __kmp_release_bootstrap_lock(&__kmp_task_team_lock);
3112   }
3113 
3114   if (task_team == NULL) {
3115     KE_TRACE(10, ("__kmp_allocate_task_team: T#%d allocating "
3116                   "task team for team %p\n",
3117                   __kmp_gtid_from_thread(thread), team));
3118     // Allocate a new task team if one is not available.
3119     // Cannot use __kmp_thread_malloc() because threads not around for
3120     // kmp_reap_task_team( ).
3121     task_team = (kmp_task_team_t *)__kmp_allocate(sizeof(kmp_task_team_t));
3122     __kmp_init_bootstrap_lock(&task_team->tt.tt_threads_lock);
3123     // AC: __kmp_allocate zeroes returned memory
3124     // task_team -> tt.tt_threads_data = NULL;
3125     // task_team -> tt.tt_max_threads = 0;
3126     // task_team -> tt.tt_next = NULL;
3127   }
3128 
3129   TCW_4(task_team->tt.tt_found_tasks, FALSE);
3130 #if OMP_45_ENABLED
3131   TCW_4(task_team->tt.tt_found_proxy_tasks, FALSE);
3132 #endif
3133   task_team->tt.tt_nproc = nthreads = team->t.t_nproc;
3134 
3135   KMP_ATOMIC_ST_REL(&task_team->tt.tt_unfinished_threads, nthreads);
3136   TCW_4(task_team->tt.tt_active, TRUE);
3137 
3138   KA_TRACE(20, ("__kmp_allocate_task_team: T#%d exiting; task_team = %p "
3139                 "unfinished_threads init'd to %d\n",
3140                 (thread ? __kmp_gtid_from_thread(thread) : -1), task_team,
3141                 KMP_ATOMIC_LD_RLX(&task_team->tt.tt_unfinished_threads)));
3142   return task_team;
3143 }
3144 
3145 // __kmp_free_task_team:
3146 // Frees the task team associated with a specific thread, and adds it
3147 // to the global task team free list.
3148 void __kmp_free_task_team(kmp_info_t *thread, kmp_task_team_t *task_team) {
3149   KA_TRACE(20, ("__kmp_free_task_team: T#%d task_team = %p\n",
3150                 thread ? __kmp_gtid_from_thread(thread) : -1, task_team));
3151 
3152   // Put task team back on free list
3153   __kmp_acquire_bootstrap_lock(&__kmp_task_team_lock);
3154 
3155   KMP_DEBUG_ASSERT(task_team->tt.tt_next == NULL);
3156   task_team->tt.tt_next = __kmp_free_task_teams;
3157   TCW_PTR(__kmp_free_task_teams, task_team);
3158 
3159   __kmp_release_bootstrap_lock(&__kmp_task_team_lock);
3160 }
3161 
3162 // __kmp_reap_task_teams:
3163 // Free all the task teams on the task team free list.
3164 // Should only be done during library shutdown.
3165 // Cannot do anything that needs a thread structure or gtid since they are
3166 // already gone.
3167 void __kmp_reap_task_teams(void) {
3168   kmp_task_team_t *task_team;
3169 
3170   if (TCR_PTR(__kmp_free_task_teams) != NULL) {
3171     // Free all task_teams on the free list
3172     __kmp_acquire_bootstrap_lock(&__kmp_task_team_lock);
3173     while ((task_team = __kmp_free_task_teams) != NULL) {
3174       __kmp_free_task_teams = task_team->tt.tt_next;
3175       task_team->tt.tt_next = NULL;
3176 
3177       // Free threads_data if necessary
3178       if (task_team->tt.tt_threads_data != NULL) {
3179         __kmp_free_task_threads_data(task_team);
3180       }
3181       __kmp_free(task_team);
3182     }
3183     __kmp_release_bootstrap_lock(&__kmp_task_team_lock);
3184   }
3185 }
3186 
3187 // __kmp_wait_to_unref_task_teams:
3188 // Some threads could still be in the fork barrier release code, possibly
3189 // trying to steal tasks.  Wait for each thread to unreference its task team.
3190 void __kmp_wait_to_unref_task_teams(void) {
3191   kmp_info_t *thread;
3192   kmp_uint32 spins;
3193   int done;
3194 
3195   KMP_INIT_YIELD(spins);
3196 
3197   for (;;) {
3198     done = TRUE;
3199 
3200     // TODO: GEH - this may be is wrong because some sync would be necessary
3201     // in case threads are added to the pool during the traversal. Need to
3202     // verify that lock for thread pool is held when calling this routine.
3203     for (thread = CCAST(kmp_info_t *, __kmp_thread_pool); thread != NULL;
3204          thread = thread->th.th_next_pool) {
3205 #if KMP_OS_WINDOWS
3206       DWORD exit_val;
3207 #endif
3208       if (TCR_PTR(thread->th.th_task_team) == NULL) {
3209         KA_TRACE(10, ("__kmp_wait_to_unref_task_team: T#%d task_team == NULL\n",
3210                       __kmp_gtid_from_thread(thread)));
3211         continue;
3212       }
3213 #if KMP_OS_WINDOWS
3214       // TODO: GEH - add this check for Linux* OS / OS X* as well?
3215       if (!__kmp_is_thread_alive(thread, &exit_val)) {
3216         thread->th.th_task_team = NULL;
3217         continue;
3218       }
3219 #endif
3220 
3221       done = FALSE; // Because th_task_team pointer is not NULL for this thread
3222 
3223       KA_TRACE(10, ("__kmp_wait_to_unref_task_team: Waiting for T#%d to "
3224                     "unreference task_team\n",
3225                     __kmp_gtid_from_thread(thread)));
3226 
3227       if (__kmp_dflt_blocktime != KMP_MAX_BLOCKTIME) {
3228         volatile void *sleep_loc;
3229         // If the thread is sleeping, awaken it.
3230         if ((sleep_loc = TCR_PTR(CCAST(void *, thread->th.th_sleep_loc))) !=
3231             NULL) {
3232           KA_TRACE(
3233               10,
3234               ("__kmp_wait_to_unref_task_team: T#%d waking up thread T#%d\n",
3235                __kmp_gtid_from_thread(thread), __kmp_gtid_from_thread(thread)));
3236           __kmp_null_resume_wrapper(__kmp_gtid_from_thread(thread), sleep_loc);
3237         }
3238       }
3239     }
3240     if (done) {
3241       break;
3242     }
3243 
3244     // If oversubscribed or have waited a bit, yield.
3245     KMP_YIELD_OVERSUB_ELSE_SPIN(spins);
3246   }
3247 }
3248 
3249 // __kmp_task_team_setup:  Create a task_team for the current team, but use
3250 // an already created, unused one if it already exists.
3251 void __kmp_task_team_setup(kmp_info_t *this_thr, kmp_team_t *team, int always) {
3252   KMP_DEBUG_ASSERT(__kmp_tasking_mode != tskm_immediate_exec);
3253 
3254   // If this task_team hasn't been created yet, allocate it. It will be used in
3255   // the region after the next.
3256   // If it exists, it is the current task team and shouldn't be touched yet as
3257   // it may still be in use.
3258   if (team->t.t_task_team[this_thr->th.th_task_state] == NULL &&
3259       (always || team->t.t_nproc > 1)) {
3260     team->t.t_task_team[this_thr->th.th_task_state] =
3261         __kmp_allocate_task_team(this_thr, team);
3262     KA_TRACE(20, ("__kmp_task_team_setup: Master T#%d created new task_team %p "
3263                   "for team %d at parity=%d\n",
3264                   __kmp_gtid_from_thread(this_thr),
3265                   team->t.t_task_team[this_thr->th.th_task_state],
3266                   ((team != NULL) ? team->t.t_id : -1),
3267                   this_thr->th.th_task_state));
3268   }
3269 
3270   // After threads exit the release, they will call sync, and then point to this
3271   // other task_team; make sure it is allocated and properly initialized. As
3272   // threads spin in the barrier release phase, they will continue to use the
3273   // previous task_team struct(above), until they receive the signal to stop
3274   // checking for tasks (they can't safely reference the kmp_team_t struct,
3275   // which could be reallocated by the master thread). No task teams are formed
3276   // for serialized teams.
3277   if (team->t.t_nproc > 1) {
3278     int other_team = 1 - this_thr->th.th_task_state;
3279     if (team->t.t_task_team[other_team] == NULL) { // setup other team as well
3280       team->t.t_task_team[other_team] =
3281           __kmp_allocate_task_team(this_thr, team);
3282       KA_TRACE(20, ("__kmp_task_team_setup: Master T#%d created second new "
3283                     "task_team %p for team %d at parity=%d\n",
3284                     __kmp_gtid_from_thread(this_thr),
3285                     team->t.t_task_team[other_team],
3286                     ((team != NULL) ? team->t.t_id : -1), other_team));
3287     } else { // Leave the old task team struct in place for the upcoming region;
3288       // adjust as needed
3289       kmp_task_team_t *task_team = team->t.t_task_team[other_team];
3290       if (!task_team->tt.tt_active ||
3291           team->t.t_nproc != task_team->tt.tt_nproc) {
3292         TCW_4(task_team->tt.tt_nproc, team->t.t_nproc);
3293         TCW_4(task_team->tt.tt_found_tasks, FALSE);
3294 #if OMP_45_ENABLED
3295         TCW_4(task_team->tt.tt_found_proxy_tasks, FALSE);
3296 #endif
3297         KMP_ATOMIC_ST_REL(&task_team->tt.tt_unfinished_threads,
3298                           team->t.t_nproc);
3299         TCW_4(task_team->tt.tt_active, TRUE);
3300       }
3301       // if team size has changed, the first thread to enable tasking will
3302       // realloc threads_data if necessary
3303       KA_TRACE(20, ("__kmp_task_team_setup: Master T#%d reset next task_team "
3304                     "%p for team %d at parity=%d\n",
3305                     __kmp_gtid_from_thread(this_thr),
3306                     team->t.t_task_team[other_team],
3307                     ((team != NULL) ? team->t.t_id : -1), other_team));
3308     }
3309   }
3310 }
3311 
3312 // __kmp_task_team_sync: Propagation of task team data from team to threads
3313 // which happens just after the release phase of a team barrier.  This may be
3314 // called by any thread, but only for teams with # threads > 1.
3315 void __kmp_task_team_sync(kmp_info_t *this_thr, kmp_team_t *team) {
3316   KMP_DEBUG_ASSERT(__kmp_tasking_mode != tskm_immediate_exec);
3317 
3318   // Toggle the th_task_state field, to switch which task_team this thread
3319   // refers to
3320   this_thr->th.th_task_state = 1 - this_thr->th.th_task_state;
3321   // It is now safe to propagate the task team pointer from the team struct to
3322   // the current thread.
3323   TCW_PTR(this_thr->th.th_task_team,
3324           team->t.t_task_team[this_thr->th.th_task_state]);
3325   KA_TRACE(20,
3326            ("__kmp_task_team_sync: Thread T#%d task team switched to task_team "
3327             "%p from Team #%d (parity=%d)\n",
3328             __kmp_gtid_from_thread(this_thr), this_thr->th.th_task_team,
3329             ((team != NULL) ? team->t.t_id : -1), this_thr->th.th_task_state));
3330 }
3331 
3332 // __kmp_task_team_wait: Master thread waits for outstanding tasks after the
3333 // barrier gather phase. Only called by master thread if #threads in team > 1 or
3334 // if proxy tasks were created.
3335 //
3336 // wait is a flag that defaults to 1 (see kmp.h), but waiting can be turned off
3337 // by passing in 0 optionally as the last argument. When wait is zero, master
3338 // thread does not wait for unfinished_threads to reach 0.
3339 void __kmp_task_team_wait(
3340     kmp_info_t *this_thr,
3341     kmp_team_t *team USE_ITT_BUILD_ARG(void *itt_sync_obj), int wait) {
3342   kmp_task_team_t *task_team = team->t.t_task_team[this_thr->th.th_task_state];
3343 
3344   KMP_DEBUG_ASSERT(__kmp_tasking_mode != tskm_immediate_exec);
3345   KMP_DEBUG_ASSERT(task_team == this_thr->th.th_task_team);
3346 
3347   if ((task_team != NULL) && KMP_TASKING_ENABLED(task_team)) {
3348     if (wait) {
3349       KA_TRACE(20, ("__kmp_task_team_wait: Master T#%d waiting for all tasks "
3350                     "(for unfinished_threads to reach 0) on task_team = %p\n",
3351                     __kmp_gtid_from_thread(this_thr), task_team));
3352       // Worker threads may have dropped through to release phase, but could
3353       // still be executing tasks. Wait here for tasks to complete. To avoid
3354       // memory contention, only master thread checks termination condition.
3355       kmp_flag_32 flag(RCAST(std::atomic<kmp_uint32> *,
3356                              &task_team->tt.tt_unfinished_threads),
3357                        0U);
3358       flag.wait(this_thr, TRUE USE_ITT_BUILD_ARG(itt_sync_obj));
3359     }
3360     // Deactivate the old task team, so that the worker threads will stop
3361     // referencing it while spinning.
3362     KA_TRACE(
3363         20,
3364         ("__kmp_task_team_wait: Master T#%d deactivating task_team %p: "
3365          "setting active to false, setting local and team's pointer to NULL\n",
3366          __kmp_gtid_from_thread(this_thr), task_team));
3367 #if OMP_45_ENABLED
3368     KMP_DEBUG_ASSERT(task_team->tt.tt_nproc > 1 ||
3369                      task_team->tt.tt_found_proxy_tasks == TRUE);
3370     TCW_SYNC_4(task_team->tt.tt_found_proxy_tasks, FALSE);
3371 #else
3372     KMP_DEBUG_ASSERT(task_team->tt.tt_nproc > 1);
3373 #endif
3374     KMP_CHECK_UPDATE(task_team->tt.tt_untied_task_encountered, 0);
3375     TCW_SYNC_4(task_team->tt.tt_active, FALSE);
3376     KMP_MB();
3377 
3378     TCW_PTR(this_thr->th.th_task_team, NULL);
3379   }
3380 }
3381 
3382 // __kmp_tasking_barrier:
3383 // This routine may only called when __kmp_tasking_mode == tskm_extra_barrier.
3384 // Internal function to execute all tasks prior to a regular barrier or a join
3385 // barrier. It is a full barrier itself, which unfortunately turns regular
3386 // barriers into double barriers and join barriers into 1 1/2 barriers.
3387 void __kmp_tasking_barrier(kmp_team_t *team, kmp_info_t *thread, int gtid) {
3388   std::atomic<kmp_uint32> *spin = RCAST(
3389       std::atomic<kmp_uint32> *,
3390       &team->t.t_task_team[thread->th.th_task_state]->tt.tt_unfinished_threads);
3391   int flag = FALSE;
3392   KMP_DEBUG_ASSERT(__kmp_tasking_mode == tskm_extra_barrier);
3393 
3394 #if USE_ITT_BUILD
3395   KMP_FSYNC_SPIN_INIT(spin, NULL);
3396 #endif /* USE_ITT_BUILD */
3397   kmp_flag_32 spin_flag(spin, 0U);
3398   while (!spin_flag.execute_tasks(thread, gtid, TRUE,
3399                                   &flag USE_ITT_BUILD_ARG(NULL), 0)) {
3400 #if USE_ITT_BUILD
3401     // TODO: What about itt_sync_obj??
3402     KMP_FSYNC_SPIN_PREPARE(RCAST(void *, spin));
3403 #endif /* USE_ITT_BUILD */
3404 
3405     if (TCR_4(__kmp_global.g.g_done)) {
3406       if (__kmp_global.g.g_abort)
3407         __kmp_abort_thread();
3408       break;
3409     }
3410     KMP_YIELD(TRUE);
3411   }
3412 #if USE_ITT_BUILD
3413   KMP_FSYNC_SPIN_ACQUIRED(RCAST(void *, spin));
3414 #endif /* USE_ITT_BUILD */
3415 }
3416 
3417 #if OMP_45_ENABLED
3418 
3419 // __kmp_give_task puts a task into a given thread queue if:
3420 //  - the queue for that thread was created
3421 //  - there's space in that queue
3422 // Because of this, __kmp_push_task needs to check if there's space after
3423 // getting the lock
3424 static bool __kmp_give_task(kmp_info_t *thread, kmp_int32 tid, kmp_task_t *task,
3425                             kmp_int32 pass) {
3426   kmp_taskdata_t *taskdata = KMP_TASK_TO_TASKDATA(task);
3427   kmp_task_team_t *task_team = taskdata->td_task_team;
3428 
3429   KA_TRACE(20, ("__kmp_give_task: trying to give task %p to thread %d.\n",
3430                 taskdata, tid));
3431 
3432   // If task_team is NULL something went really bad...
3433   KMP_DEBUG_ASSERT(task_team != NULL);
3434 
3435   bool result = false;
3436   kmp_thread_data_t *thread_data = &task_team->tt.tt_threads_data[tid];
3437 
3438   if (thread_data->td.td_deque == NULL) {
3439     // There's no queue in this thread, go find another one
3440     // We're guaranteed that at least one thread has a queue
3441     KA_TRACE(30,
3442              ("__kmp_give_task: thread %d has no queue while giving task %p.\n",
3443               tid, taskdata));
3444     return result;
3445   }
3446 
3447   if (TCR_4(thread_data->td.td_deque_ntasks) >=
3448       TASK_DEQUE_SIZE(thread_data->td)) {
3449     KA_TRACE(
3450         30,
3451         ("__kmp_give_task: queue is full while giving task %p to thread %d.\n",
3452          taskdata, tid));
3453 
3454     // if this deque is bigger than the pass ratio give a chance to another
3455     // thread
3456     if (TASK_DEQUE_SIZE(thread_data->td) / INITIAL_TASK_DEQUE_SIZE >= pass)
3457       return result;
3458 
3459     __kmp_acquire_bootstrap_lock(&thread_data->td.td_deque_lock);
3460     __kmp_realloc_task_deque(thread, thread_data);
3461 
3462   } else {
3463 
3464     __kmp_acquire_bootstrap_lock(&thread_data->td.td_deque_lock);
3465 
3466     if (TCR_4(thread_data->td.td_deque_ntasks) >=
3467         TASK_DEQUE_SIZE(thread_data->td)) {
3468       KA_TRACE(30, ("__kmp_give_task: queue is full while giving task %p to "
3469                     "thread %d.\n",
3470                     taskdata, tid));
3471 
3472       // if this deque is bigger than the pass ratio give a chance to another
3473       // thread
3474       if (TASK_DEQUE_SIZE(thread_data->td) / INITIAL_TASK_DEQUE_SIZE >= pass)
3475         goto release_and_exit;
3476 
3477       __kmp_realloc_task_deque(thread, thread_data);
3478     }
3479   }
3480 
3481   // lock is held here, and there is space in the deque
3482 
3483   thread_data->td.td_deque[thread_data->td.td_deque_tail] = taskdata;
3484   // Wrap index.
3485   thread_data->td.td_deque_tail =
3486       (thread_data->td.td_deque_tail + 1) & TASK_DEQUE_MASK(thread_data->td);
3487   TCW_4(thread_data->td.td_deque_ntasks,
3488         TCR_4(thread_data->td.td_deque_ntasks) + 1);
3489 
3490   result = true;
3491   KA_TRACE(30, ("__kmp_give_task: successfully gave task %p to thread %d.\n",
3492                 taskdata, tid));
3493 
3494 release_and_exit:
3495   __kmp_release_bootstrap_lock(&thread_data->td.td_deque_lock);
3496 
3497   return result;
3498 }
3499 
3500 /* The finish of the proxy tasks is divided in two pieces:
3501     - the top half is the one that can be done from a thread outside the team
3502     - the bottom half must be run from a thread within the team
3503 
3504    In order to run the bottom half the task gets queued back into one of the
3505    threads of the team. Once the td_incomplete_child_task counter of the parent
3506    is decremented the threads can leave the barriers. So, the bottom half needs
3507    to be queued before the counter is decremented. The top half is therefore
3508    divided in two parts:
3509     - things that can be run before queuing the bottom half
3510     - things that must be run after queuing the bottom half
3511 
3512    This creates a second race as the bottom half can free the task before the
3513    second top half is executed. To avoid this we use the
3514    td_incomplete_child_task of the proxy task to synchronize the top and bottom
3515    half. */
3516 static void __kmp_first_top_half_finish_proxy(kmp_taskdata_t *taskdata) {
3517   KMP_DEBUG_ASSERT(taskdata->td_flags.tasktype == TASK_EXPLICIT);
3518   KMP_DEBUG_ASSERT(taskdata->td_flags.proxy == TASK_PROXY);
3519   KMP_DEBUG_ASSERT(taskdata->td_flags.complete == 0);
3520   KMP_DEBUG_ASSERT(taskdata->td_flags.freed == 0);
3521 
3522   taskdata->td_flags.complete = 1; // mark the task as completed
3523 
3524   if (taskdata->td_taskgroup)
3525     KMP_ATOMIC_DEC(&taskdata->td_taskgroup->count);
3526 
3527   // Create an imaginary children for this task so the bottom half cannot
3528   // release the task before we have completed the second top half
3529   KMP_ATOMIC_INC(&taskdata->td_incomplete_child_tasks);
3530 }
3531 
3532 static void __kmp_second_top_half_finish_proxy(kmp_taskdata_t *taskdata) {
3533   kmp_int32 children = 0;
3534 
3535   // Predecrement simulated by "- 1" calculation
3536   children =
3537       KMP_ATOMIC_DEC(&taskdata->td_parent->td_incomplete_child_tasks) - 1;
3538   KMP_DEBUG_ASSERT(children >= 0);
3539 
3540   // Remove the imaginary children
3541   KMP_ATOMIC_DEC(&taskdata->td_incomplete_child_tasks);
3542 }
3543 
3544 static void __kmp_bottom_half_finish_proxy(kmp_int32 gtid, kmp_task_t *ptask) {
3545   kmp_taskdata_t *taskdata = KMP_TASK_TO_TASKDATA(ptask);
3546   kmp_info_t *thread = __kmp_threads[gtid];
3547 
3548   KMP_DEBUG_ASSERT(taskdata->td_flags.proxy == TASK_PROXY);
3549   KMP_DEBUG_ASSERT(taskdata->td_flags.complete ==
3550                    1); // top half must run before bottom half
3551 
3552   // We need to wait to make sure the top half is finished
3553   // Spinning here should be ok as this should happen quickly
3554   while (KMP_ATOMIC_LD_ACQ(&taskdata->td_incomplete_child_tasks) > 0)
3555     ;
3556 
3557   __kmp_release_deps(gtid, taskdata);
3558   __kmp_free_task_and_ancestors(gtid, taskdata, thread);
3559 }
3560 
3561 /*!
3562 @ingroup TASKING
3563 @param gtid Global Thread ID of encountering thread
3564 @param ptask Task which execution is completed
3565 
3566 Execute the completation of a proxy task from a thread of that is part of the
3567 team. Run first and bottom halves directly.
3568 */
3569 void __kmpc_proxy_task_completed(kmp_int32 gtid, kmp_task_t *ptask) {
3570   KMP_DEBUG_ASSERT(ptask != NULL);
3571   kmp_taskdata_t *taskdata = KMP_TASK_TO_TASKDATA(ptask);
3572   KA_TRACE(
3573       10, ("__kmp_proxy_task_completed(enter): T#%d proxy task %p completing\n",
3574            gtid, taskdata));
3575 
3576   KMP_DEBUG_ASSERT(taskdata->td_flags.proxy == TASK_PROXY);
3577 
3578   __kmp_first_top_half_finish_proxy(taskdata);
3579   __kmp_second_top_half_finish_proxy(taskdata);
3580   __kmp_bottom_half_finish_proxy(gtid, ptask);
3581 
3582   KA_TRACE(10,
3583            ("__kmp_proxy_task_completed(exit): T#%d proxy task %p completing\n",
3584             gtid, taskdata));
3585 }
3586 
3587 /*!
3588 @ingroup TASKING
3589 @param ptask Task which execution is completed
3590 
3591 Execute the completation of a proxy task from a thread that could not belong to
3592 the team.
3593 */
3594 void __kmpc_proxy_task_completed_ooo(kmp_task_t *ptask) {
3595   KMP_DEBUG_ASSERT(ptask != NULL);
3596   kmp_taskdata_t *taskdata = KMP_TASK_TO_TASKDATA(ptask);
3597 
3598   KA_TRACE(
3599       10,
3600       ("__kmp_proxy_task_completed_ooo(enter): proxy task completing ooo %p\n",
3601        taskdata));
3602 
3603   KMP_DEBUG_ASSERT(taskdata->td_flags.proxy == TASK_PROXY);
3604 
3605   __kmp_first_top_half_finish_proxy(taskdata);
3606 
3607   // Enqueue task to complete bottom half completion from a thread within the
3608   // corresponding team
3609   kmp_team_t *team = taskdata->td_team;
3610   kmp_int32 nthreads = team->t.t_nproc;
3611   kmp_info_t *thread;
3612 
3613   // This should be similar to start_k = __kmp_get_random( thread ) % nthreads
3614   // but we cannot use __kmp_get_random here
3615   kmp_int32 start_k = 0;
3616   kmp_int32 pass = 1;
3617   kmp_int32 k = start_k;
3618 
3619   do {
3620     // For now we're just linearly trying to find a thread
3621     thread = team->t.t_threads[k];
3622     k = (k + 1) % nthreads;
3623 
3624     // we did a full pass through all the threads
3625     if (k == start_k)
3626       pass = pass << 1;
3627 
3628   } while (!__kmp_give_task(thread, k, ptask, pass));
3629 
3630   __kmp_second_top_half_finish_proxy(taskdata);
3631 
3632   KA_TRACE(
3633       10,
3634       ("__kmp_proxy_task_completed_ooo(exit): proxy task completing ooo %p\n",
3635        taskdata));
3636 }
3637 
3638 // __kmp_task_dup_alloc: Allocate the taskdata and make a copy of source task
3639 // for taskloop
3640 //
3641 // thread:   allocating thread
3642 // task_src: pointer to source task to be duplicated
3643 // returns:  a pointer to the allocated kmp_task_t structure (task).
3644 kmp_task_t *__kmp_task_dup_alloc(kmp_info_t *thread, kmp_task_t *task_src) {
3645   kmp_task_t *task;
3646   kmp_taskdata_t *taskdata;
3647   kmp_taskdata_t *taskdata_src;
3648   kmp_taskdata_t *parent_task = thread->th.th_current_task;
3649   size_t shareds_offset;
3650   size_t task_size;
3651 
3652   KA_TRACE(10, ("__kmp_task_dup_alloc(enter): Th %p, source task %p\n", thread,
3653                 task_src));
3654   taskdata_src = KMP_TASK_TO_TASKDATA(task_src);
3655   KMP_DEBUG_ASSERT(taskdata_src->td_flags.proxy ==
3656                    TASK_FULL); // it should not be proxy task
3657   KMP_DEBUG_ASSERT(taskdata_src->td_flags.tasktype == TASK_EXPLICIT);
3658   task_size = taskdata_src->td_size_alloc;
3659 
3660   // Allocate a kmp_taskdata_t block and a kmp_task_t block.
3661   KA_TRACE(30, ("__kmp_task_dup_alloc: Th %p, malloc size %ld\n", thread,
3662                 task_size));
3663 #if USE_FAST_MEMORY
3664   taskdata = (kmp_taskdata_t *)__kmp_fast_allocate(thread, task_size);
3665 #else
3666   taskdata = (kmp_taskdata_t *)__kmp_thread_malloc(thread, task_size);
3667 #endif /* USE_FAST_MEMORY */
3668   KMP_MEMCPY(taskdata, taskdata_src, task_size);
3669 
3670   task = KMP_TASKDATA_TO_TASK(taskdata);
3671 
3672   // Initialize new task (only specific fields not affected by memcpy)
3673   taskdata->td_task_id = KMP_GEN_TASK_ID();
3674   if (task->shareds != NULL) { // need setup shareds pointer
3675     shareds_offset = (char *)task_src->shareds - (char *)taskdata_src;
3676     task->shareds = &((char *)taskdata)[shareds_offset];
3677     KMP_DEBUG_ASSERT((((kmp_uintptr_t)task->shareds) & (sizeof(void *) - 1)) ==
3678                      0);
3679   }
3680   taskdata->td_alloc_thread = thread;
3681   taskdata->td_parent = parent_task;
3682   taskdata->td_taskgroup =
3683       parent_task
3684           ->td_taskgroup; // task inherits the taskgroup from the parent task
3685 
3686   // Only need to keep track of child task counts if team parallel and tasking
3687   // not serialized
3688   if (!(taskdata->td_flags.team_serial || taskdata->td_flags.tasking_ser)) {
3689     KMP_ATOMIC_INC(&parent_task->td_incomplete_child_tasks);
3690     if (parent_task->td_taskgroup)
3691       KMP_ATOMIC_INC(&parent_task->td_taskgroup->count);
3692     // Only need to keep track of allocated child tasks for explicit tasks since
3693     // implicit not deallocated
3694     if (taskdata->td_parent->td_flags.tasktype == TASK_EXPLICIT)
3695       KMP_ATOMIC_INC(&taskdata->td_parent->td_allocated_child_tasks);
3696   }
3697 
3698   KA_TRACE(20,
3699            ("__kmp_task_dup_alloc(exit): Th %p, created task %p, parent=%p\n",
3700             thread, taskdata, taskdata->td_parent));
3701 #if OMPT_SUPPORT
3702   if (UNLIKELY(ompt_enabled.enabled))
3703     __ompt_task_init(taskdata, thread->th.th_info.ds.ds_gtid);
3704 #endif
3705   return task;
3706 }
3707 
3708 // Routine optionally generated by the compiler for setting the lastprivate flag
3709 // and calling needed constructors for private/firstprivate objects
3710 // (used to form taskloop tasks from pattern task)
3711 // Parameters: dest task, src task, lastprivate flag.
3712 typedef void (*p_task_dup_t)(kmp_task_t *, kmp_task_t *, kmp_int32);
3713 
3714 KMP_BUILD_ASSERT(sizeof(long) == 4 || sizeof(long) == 8);
3715 
3716 // class to encapsulate manipulating loop bounds in a taskloop task.
3717 // this abstracts away the Intel vs GOMP taskloop interface for setting/getting
3718 // the loop bound variables.
3719 class kmp_taskloop_bounds_t {
3720   kmp_task_t *task;
3721   const kmp_taskdata_t *taskdata;
3722   size_t lower_offset;
3723   size_t upper_offset;
3724 
3725 public:
3726   kmp_taskloop_bounds_t(kmp_task_t *_task, kmp_uint64 *lb, kmp_uint64 *ub)
3727       : task(_task), taskdata(KMP_TASK_TO_TASKDATA(task)),
3728         lower_offset((char *)lb - (char *)task),
3729         upper_offset((char *)ub - (char *)task) {
3730     KMP_DEBUG_ASSERT((char *)lb > (char *)_task);
3731     KMP_DEBUG_ASSERT((char *)ub > (char *)_task);
3732   }
3733   kmp_taskloop_bounds_t(kmp_task_t *_task, const kmp_taskloop_bounds_t &bounds)
3734       : task(_task), taskdata(KMP_TASK_TO_TASKDATA(_task)),
3735         lower_offset(bounds.lower_offset), upper_offset(bounds.upper_offset) {}
3736   size_t get_lower_offset() const { return lower_offset; }
3737   size_t get_upper_offset() const { return upper_offset; }
3738   kmp_uint64 get_lb() const {
3739     kmp_int64 retval;
3740 #if defined(KMP_GOMP_COMPAT)
3741     // Intel task just returns the lower bound normally
3742     if (!taskdata->td_flags.native) {
3743       retval = *(kmp_int64 *)((char *)task + lower_offset);
3744     } else {
3745       // GOMP task has to take into account the sizeof(long)
3746       if (taskdata->td_size_loop_bounds == 4) {
3747         kmp_int32 *lb = RCAST(kmp_int32 *, task->shareds);
3748         retval = (kmp_int64)*lb;
3749       } else {
3750         kmp_int64 *lb = RCAST(kmp_int64 *, task->shareds);
3751         retval = (kmp_int64)*lb;
3752       }
3753     }
3754 #else
3755     retval = *(kmp_int64 *)((char *)task + lower_offset);
3756 #endif // defined(KMP_GOMP_COMPAT)
3757     return retval;
3758   }
3759   kmp_uint64 get_ub() const {
3760     kmp_int64 retval;
3761 #if defined(KMP_GOMP_COMPAT)
3762     // Intel task just returns the upper bound normally
3763     if (!taskdata->td_flags.native) {
3764       retval = *(kmp_int64 *)((char *)task + upper_offset);
3765     } else {
3766       // GOMP task has to take into account the sizeof(long)
3767       if (taskdata->td_size_loop_bounds == 4) {
3768         kmp_int32 *ub = RCAST(kmp_int32 *, task->shareds) + 1;
3769         retval = (kmp_int64)*ub;
3770       } else {
3771         kmp_int64 *ub = RCAST(kmp_int64 *, task->shareds) + 1;
3772         retval = (kmp_int64)*ub;
3773       }
3774     }
3775 #else
3776     retval = *(kmp_int64 *)((char *)task + upper_offset);
3777 #endif // defined(KMP_GOMP_COMPAT)
3778     return retval;
3779   }
3780   void set_lb(kmp_uint64 lb) {
3781 #if defined(KMP_GOMP_COMPAT)
3782     // Intel task just sets the lower bound normally
3783     if (!taskdata->td_flags.native) {
3784       *(kmp_uint64 *)((char *)task + lower_offset) = lb;
3785     } else {
3786       // GOMP task has to take into account the sizeof(long)
3787       if (taskdata->td_size_loop_bounds == 4) {
3788         kmp_uint32 *lower = RCAST(kmp_uint32 *, task->shareds);
3789         *lower = (kmp_uint32)lb;
3790       } else {
3791         kmp_uint64 *lower = RCAST(kmp_uint64 *, task->shareds);
3792         *lower = (kmp_uint64)lb;
3793       }
3794     }
3795 #else
3796     *(kmp_uint64 *)((char *)task + lower_offset) = lb;
3797 #endif // defined(KMP_GOMP_COMPAT)
3798   }
3799   void set_ub(kmp_uint64 ub) {
3800 #if defined(KMP_GOMP_COMPAT)
3801     // Intel task just sets the upper bound normally
3802     if (!taskdata->td_flags.native) {
3803       *(kmp_uint64 *)((char *)task + upper_offset) = ub;
3804     } else {
3805       // GOMP task has to take into account the sizeof(long)
3806       if (taskdata->td_size_loop_bounds == 4) {
3807         kmp_uint32 *upper = RCAST(kmp_uint32 *, task->shareds) + 1;
3808         *upper = (kmp_uint32)ub;
3809       } else {
3810         kmp_uint64 *upper = RCAST(kmp_uint64 *, task->shareds) + 1;
3811         *upper = (kmp_uint64)ub;
3812       }
3813     }
3814 #else
3815     *(kmp_uint64 *)((char *)task + upper_offset) = ub;
3816 #endif // defined(KMP_GOMP_COMPAT)
3817   }
3818 };
3819 
3820 // __kmp_taskloop_linear: Start tasks of the taskloop linearly
3821 //
3822 // loc        Source location information
3823 // gtid       Global thread ID
3824 // task       Pattern task, exposes the loop iteration range
3825 // lb         Pointer to loop lower bound in task structure
3826 // ub         Pointer to loop upper bound in task structure
3827 // st         Loop stride
3828 // ub_glob    Global upper bound (used for lastprivate check)
3829 // num_tasks  Number of tasks to execute
3830 // grainsize  Number of loop iterations per task
3831 // extras     Number of chunks with grainsize+1 iterations
3832 // tc         Iterations count
3833 // task_dup   Tasks duplication routine
3834 // codeptr_ra Return address for OMPT events
3835 void __kmp_taskloop_linear(ident_t *loc, int gtid, kmp_task_t *task,
3836                            kmp_uint64 *lb, kmp_uint64 *ub, kmp_int64 st,
3837                            kmp_uint64 ub_glob, kmp_uint64 num_tasks,
3838                            kmp_uint64 grainsize, kmp_uint64 extras,
3839                            kmp_uint64 tc,
3840 #if OMPT_SUPPORT
3841                            void *codeptr_ra,
3842 #endif
3843                            void *task_dup) {
3844   KMP_COUNT_BLOCK(OMP_TASKLOOP);
3845   KMP_TIME_PARTITIONED_BLOCK(OMP_taskloop_scheduling);
3846   p_task_dup_t ptask_dup = (p_task_dup_t)task_dup;
3847   // compiler provides global bounds here
3848   kmp_taskloop_bounds_t task_bounds(task, lb, ub);
3849   kmp_uint64 lower = task_bounds.get_lb();
3850   kmp_uint64 upper = task_bounds.get_ub();
3851   kmp_uint64 i;
3852   kmp_info_t *thread = __kmp_threads[gtid];
3853   kmp_taskdata_t *current_task = thread->th.th_current_task;
3854   kmp_task_t *next_task;
3855   kmp_int32 lastpriv = 0;
3856 
3857   KMP_DEBUG_ASSERT(tc == num_tasks * grainsize + extras);
3858   KMP_DEBUG_ASSERT(num_tasks > extras);
3859   KMP_DEBUG_ASSERT(num_tasks > 0);
3860   KA_TRACE(20, ("__kmp_taskloop_linear: T#%d: %lld tasks, grainsize %lld, "
3861                 "extras %lld, i=%lld,%lld(%d)%lld, dup %p\n",
3862                 gtid, num_tasks, grainsize, extras, lower, upper, ub_glob, st,
3863                 task_dup));
3864 
3865   // Launch num_tasks tasks, assign grainsize iterations each task
3866   for (i = 0; i < num_tasks; ++i) {
3867     kmp_uint64 chunk_minus_1;
3868     if (extras == 0) {
3869       chunk_minus_1 = grainsize - 1;
3870     } else {
3871       chunk_minus_1 = grainsize;
3872       --extras; // first extras iterations get bigger chunk (grainsize+1)
3873     }
3874     upper = lower + st * chunk_minus_1;
3875     if (i == num_tasks - 1) {
3876       // schedule the last task, set lastprivate flag if needed
3877       if (st == 1) { // most common case
3878         KMP_DEBUG_ASSERT(upper == *ub);
3879         if (upper == ub_glob)
3880           lastpriv = 1;
3881       } else if (st > 0) { // positive loop stride
3882         KMP_DEBUG_ASSERT((kmp_uint64)st > *ub - upper);
3883         if ((kmp_uint64)st > ub_glob - upper)
3884           lastpriv = 1;
3885       } else { // negative loop stride
3886         KMP_DEBUG_ASSERT(upper + st < *ub);
3887         if (upper - ub_glob < (kmp_uint64)(-st))
3888           lastpriv = 1;
3889       }
3890     }
3891     next_task = __kmp_task_dup_alloc(thread, task); // allocate new task
3892     kmp_taskdata_t *next_taskdata = KMP_TASK_TO_TASKDATA(next_task);
3893     kmp_taskloop_bounds_t next_task_bounds =
3894         kmp_taskloop_bounds_t(next_task, task_bounds);
3895 
3896     // adjust task-specific bounds
3897     next_task_bounds.set_lb(lower);
3898     if (next_taskdata->td_flags.native) {
3899       next_task_bounds.set_ub(upper + (st > 0 ? 1 : -1));
3900     } else {
3901       next_task_bounds.set_ub(upper);
3902     }
3903     if (ptask_dup != NULL) // set lastprivate flag, construct fistprivates, etc.
3904       ptask_dup(next_task, task, lastpriv);
3905     KA_TRACE(40,
3906              ("__kmp_taskloop_linear: T#%d; task #%llu: task %p: lower %lld, "
3907               "upper %lld stride %lld, (offsets %p %p)\n",
3908               gtid, i, next_task, lower, upper, st,
3909               next_task_bounds.get_lower_offset(),
3910               next_task_bounds.get_upper_offset()));
3911 #if OMPT_SUPPORT
3912     __kmp_omp_taskloop_task(NULL, gtid, next_task,
3913                            codeptr_ra); // schedule new task
3914 #else
3915     __kmp_omp_task(gtid, next_task, true); // schedule new task
3916 #endif
3917     lower = upper + st; // adjust lower bound for the next iteration
3918   }
3919   // free the pattern task and exit
3920   __kmp_task_start(gtid, task, current_task); // make internal bookkeeping
3921   // do not execute the pattern task, just do internal bookkeeping
3922   __kmp_task_finish<false>(gtid, task, current_task);
3923 }
3924 
3925 // Structure to keep taskloop parameters for auxiliary task
3926 // kept in the shareds of the task structure.
3927 typedef struct __taskloop_params {
3928   kmp_task_t *task;
3929   kmp_uint64 *lb;
3930   kmp_uint64 *ub;
3931   void *task_dup;
3932   kmp_int64 st;
3933   kmp_uint64 ub_glob;
3934   kmp_uint64 num_tasks;
3935   kmp_uint64 grainsize;
3936   kmp_uint64 extras;
3937   kmp_uint64 tc;
3938   kmp_uint64 num_t_min;
3939 #if OMPT_SUPPORT
3940   void *codeptr_ra;
3941 #endif
3942 } __taskloop_params_t;
3943 
3944 void __kmp_taskloop_recur(ident_t *, int, kmp_task_t *, kmp_uint64 *,
3945                           kmp_uint64 *, kmp_int64, kmp_uint64, kmp_uint64,
3946                           kmp_uint64, kmp_uint64, kmp_uint64, kmp_uint64,
3947 #if OMPT_SUPPORT
3948                           void *,
3949 #endif
3950                           void *);
3951 
3952 // Execute part of the the taskloop submitted as a task.
3953 int __kmp_taskloop_task(int gtid, void *ptask) {
3954   __taskloop_params_t *p =
3955       (__taskloop_params_t *)((kmp_task_t *)ptask)->shareds;
3956   kmp_task_t *task = p->task;
3957   kmp_uint64 *lb = p->lb;
3958   kmp_uint64 *ub = p->ub;
3959   void *task_dup = p->task_dup;
3960   //  p_task_dup_t ptask_dup = (p_task_dup_t)task_dup;
3961   kmp_int64 st = p->st;
3962   kmp_uint64 ub_glob = p->ub_glob;
3963   kmp_uint64 num_tasks = p->num_tasks;
3964   kmp_uint64 grainsize = p->grainsize;
3965   kmp_uint64 extras = p->extras;
3966   kmp_uint64 tc = p->tc;
3967   kmp_uint64 num_t_min = p->num_t_min;
3968 #if OMPT_SUPPORT
3969   void *codeptr_ra = p->codeptr_ra;
3970 #endif
3971 #if KMP_DEBUG
3972   kmp_taskdata_t *taskdata = KMP_TASK_TO_TASKDATA(task);
3973   KMP_DEBUG_ASSERT(task != NULL);
3974   KA_TRACE(20, ("__kmp_taskloop_task: T#%d, task %p: %lld tasks, grainsize"
3975                 " %lld, extras %lld, i=%lld,%lld(%d), dup %p\n",
3976                 gtid, taskdata, num_tasks, grainsize, extras, *lb, *ub, st,
3977                 task_dup));
3978 #endif
3979   KMP_DEBUG_ASSERT(num_tasks * 2 + 1 > num_t_min);
3980   if (num_tasks > num_t_min)
3981     __kmp_taskloop_recur(NULL, gtid, task, lb, ub, st, ub_glob, num_tasks,
3982                          grainsize, extras, tc, num_t_min,
3983 #if OMPT_SUPPORT
3984                          codeptr_ra,
3985 #endif
3986                          task_dup);
3987   else
3988     __kmp_taskloop_linear(NULL, gtid, task, lb, ub, st, ub_glob, num_tasks,
3989                           grainsize, extras, tc,
3990 #if OMPT_SUPPORT
3991                           codeptr_ra,
3992 #endif
3993                           task_dup);
3994 
3995   KA_TRACE(40, ("__kmp_taskloop_task(exit): T#%d\n", gtid));
3996   return 0;
3997 }
3998 
3999 // Schedule part of the the taskloop as a task,
4000 // execute the rest of the the taskloop.
4001 //
4002 // loc        Source location information
4003 // gtid       Global thread ID
4004 // task       Pattern task, exposes the loop iteration range
4005 // lb         Pointer to loop lower bound in task structure
4006 // ub         Pointer to loop upper bound in task structure
4007 // st         Loop stride
4008 // ub_glob    Global upper bound (used for lastprivate check)
4009 // num_tasks  Number of tasks to execute
4010 // grainsize  Number of loop iterations per task
4011 // extras     Number of chunks with grainsize+1 iterations
4012 // tc         Iterations count
4013 // num_t_min  Threashold to launch tasks recursively
4014 // task_dup   Tasks duplication routine
4015 // codeptr_ra Return address for OMPT events
4016 void __kmp_taskloop_recur(ident_t *loc, int gtid, kmp_task_t *task,
4017                           kmp_uint64 *lb, kmp_uint64 *ub, kmp_int64 st,
4018                           kmp_uint64 ub_glob, kmp_uint64 num_tasks,
4019                           kmp_uint64 grainsize, kmp_uint64 extras,
4020                           kmp_uint64 tc, kmp_uint64 num_t_min,
4021 #if OMPT_SUPPORT
4022                           void *codeptr_ra,
4023 #endif
4024                           void *task_dup) {
4025 #if KMP_DEBUG
4026   kmp_taskdata_t *taskdata = KMP_TASK_TO_TASKDATA(task);
4027   KMP_DEBUG_ASSERT(task != NULL);
4028   KMP_DEBUG_ASSERT(num_tasks > num_t_min);
4029   KA_TRACE(20, ("__kmp_taskloop_recur: T#%d, task %p: %lld tasks, grainsize"
4030                 " %lld, extras %lld, i=%lld,%lld(%d), dup %p\n",
4031                 gtid, taskdata, num_tasks, grainsize, extras, *lb, *ub, st,
4032                 task_dup));
4033 #endif
4034   p_task_dup_t ptask_dup = (p_task_dup_t)task_dup;
4035   kmp_uint64 lower = *lb;
4036   kmp_info_t *thread = __kmp_threads[gtid];
4037   //  kmp_taskdata_t *current_task = thread->th.th_current_task;
4038   kmp_task_t *next_task;
4039   size_t lower_offset =
4040       (char *)lb - (char *)task; // remember offset of lb in the task structure
4041   size_t upper_offset =
4042       (char *)ub - (char *)task; // remember offset of ub in the task structure
4043 
4044   KMP_DEBUG_ASSERT(tc == num_tasks * grainsize + extras);
4045   KMP_DEBUG_ASSERT(num_tasks > extras);
4046   KMP_DEBUG_ASSERT(num_tasks > 0);
4047 
4048   // split the loop in two halves
4049   kmp_uint64 lb1, ub0, tc0, tc1, ext0, ext1;
4050   kmp_uint64 gr_size0 = grainsize;
4051   kmp_uint64 n_tsk0 = num_tasks >> 1; // num_tasks/2 to execute
4052   kmp_uint64 n_tsk1 = num_tasks - n_tsk0; // to schedule as a task
4053   if (n_tsk0 <= extras) {
4054     gr_size0++; // integrate extras into grainsize
4055     ext0 = 0; // no extra iters in 1st half
4056     ext1 = extras - n_tsk0; // remaining extras
4057     tc0 = gr_size0 * n_tsk0;
4058     tc1 = tc - tc0;
4059   } else { // n_tsk0 > extras
4060     ext1 = 0; // no extra iters in 2nd half
4061     ext0 = extras;
4062     tc1 = grainsize * n_tsk1;
4063     tc0 = tc - tc1;
4064   }
4065   ub0 = lower + st * (tc0 - 1);
4066   lb1 = ub0 + st;
4067 
4068   // create pattern task for 2nd half of the loop
4069   next_task = __kmp_task_dup_alloc(thread, task); // duplicate the task
4070   // adjust lower bound (upper bound is not changed) for the 2nd half
4071   *(kmp_uint64 *)((char *)next_task + lower_offset) = lb1;
4072   if (ptask_dup != NULL) // construct fistprivates, etc.
4073     ptask_dup(next_task, task, 0);
4074   *ub = ub0; // adjust upper bound for the 1st half
4075 
4076   // create auxiliary task for 2nd half of the loop
4077   kmp_task_t *new_task =
4078       __kmpc_omp_task_alloc(loc, gtid, 1, 3 * sizeof(void *),
4079                             sizeof(__taskloop_params_t), &__kmp_taskloop_task);
4080   __taskloop_params_t *p = (__taskloop_params_t *)new_task->shareds;
4081   p->task = next_task;
4082   p->lb = (kmp_uint64 *)((char *)next_task + lower_offset);
4083   p->ub = (kmp_uint64 *)((char *)next_task + upper_offset);
4084   p->task_dup = task_dup;
4085   p->st = st;
4086   p->ub_glob = ub_glob;
4087   p->num_tasks = n_tsk1;
4088   p->grainsize = grainsize;
4089   p->extras = ext1;
4090   p->tc = tc1;
4091   p->num_t_min = num_t_min;
4092 #if OMPT_SUPPORT
4093   p->codeptr_ra = codeptr_ra;
4094 #endif
4095 
4096 #if OMPT_SUPPORT
4097   // schedule new task with correct return address for OMPT events
4098   __kmp_omp_taskloop_task(NULL, gtid, new_task, codeptr_ra);
4099 #else
4100   __kmp_omp_task(gtid, new_task, true); // schedule new task
4101 #endif
4102 
4103   // execute the 1st half of current subrange
4104   if (n_tsk0 > num_t_min)
4105     __kmp_taskloop_recur(loc, gtid, task, lb, ub, st, ub_glob, n_tsk0, gr_size0,
4106                          ext0, tc0, num_t_min,
4107 #if OMPT_SUPPORT
4108                          codeptr_ra,
4109 #endif
4110                          task_dup);
4111   else
4112     __kmp_taskloop_linear(loc, gtid, task, lb, ub, st, ub_glob, n_tsk0,
4113                           gr_size0, ext0, tc0,
4114 #if OMPT_SUPPORT
4115                           codeptr_ra,
4116 #endif
4117                           task_dup);
4118 
4119   KA_TRACE(40, ("__kmpc_taskloop_recur(exit): T#%d\n", gtid));
4120 }
4121 
4122 /*!
4123 @ingroup TASKING
4124 @param loc       Source location information
4125 @param gtid      Global thread ID
4126 @param task      Task structure
4127 @param if_val    Value of the if clause
4128 @param lb        Pointer to loop lower bound in task structure
4129 @param ub        Pointer to loop upper bound in task structure
4130 @param st        Loop stride
4131 @param nogroup   Flag, 1 if no taskgroup needs to be added, 0 otherwise
4132 @param sched     Schedule specified 0/1/2 for none/grainsize/num_tasks
4133 @param grainsize Schedule value if specified
4134 @param task_dup  Tasks duplication routine
4135 
4136 Execute the taskloop construct.
4137 */
4138 void __kmpc_taskloop(ident_t *loc, int gtid, kmp_task_t *task, int if_val,
4139                      kmp_uint64 *lb, kmp_uint64 *ub, kmp_int64 st, int nogroup,
4140                      int sched, kmp_uint64 grainsize, void *task_dup) {
4141   kmp_taskdata_t *taskdata = KMP_TASK_TO_TASKDATA(task);
4142   KMP_DEBUG_ASSERT(task != NULL);
4143 
4144   if (nogroup == 0) {
4145 #if OMPT_SUPPORT && OMPT_OPTIONAL
4146     OMPT_STORE_RETURN_ADDRESS(gtid);
4147 #endif
4148     __kmpc_taskgroup(loc, gtid);
4149   }
4150 
4151   // =========================================================================
4152   // calculate loop parameters
4153   kmp_taskloop_bounds_t task_bounds(task, lb, ub);
4154   kmp_uint64 tc;
4155   // compiler provides global bounds here
4156   kmp_uint64 lower = task_bounds.get_lb();
4157   kmp_uint64 upper = task_bounds.get_ub();
4158   kmp_uint64 ub_glob = upper; // global upper used to calc lastprivate flag
4159   kmp_uint64 num_tasks = 0, extras = 0;
4160   kmp_uint64 num_tasks_min = __kmp_taskloop_min_tasks;
4161   kmp_info_t *thread = __kmp_threads[gtid];
4162   kmp_taskdata_t *current_task = thread->th.th_current_task;
4163 
4164   KA_TRACE(20, ("__kmpc_taskloop: T#%d, task %p, lb %lld, ub %lld, st %lld, "
4165                 "grain %llu(%d), dup %p\n",
4166                 gtid, taskdata, lower, upper, st, grainsize, sched, task_dup));
4167 
4168   // compute trip count
4169   if (st == 1) { // most common case
4170     tc = upper - lower + 1;
4171   } else if (st < 0) {
4172     tc = (lower - upper) / (-st) + 1;
4173   } else { // st > 0
4174     tc = (upper - lower) / st + 1;
4175   }
4176   if (tc == 0) {
4177     KA_TRACE(20, ("__kmpc_taskloop(exit): T#%d zero-trip loop\n", gtid));
4178     // free the pattern task and exit
4179     __kmp_task_start(gtid, task, current_task);
4180     // do not execute anything for zero-trip loop
4181     __kmp_task_finish<false>(gtid, task, current_task);
4182     return;
4183   }
4184 
4185 #if OMPT_SUPPORT && OMPT_OPTIONAL
4186   ompt_team_info_t *team_info = __ompt_get_teaminfo(0, NULL);
4187   ompt_task_info_t *task_info = __ompt_get_task_info_object(0);
4188   if (ompt_enabled.ompt_callback_work) {
4189     ompt_callbacks.ompt_callback(ompt_callback_work)(
4190         ompt_work_taskloop, ompt_scope_begin, &(team_info->parallel_data),
4191         &(task_info->task_data), tc, OMPT_GET_RETURN_ADDRESS(0));
4192   }
4193 #endif
4194 
4195   if (num_tasks_min == 0)
4196     // TODO: can we choose better default heuristic?
4197     num_tasks_min =
4198         KMP_MIN(thread->th.th_team_nproc * 10, INITIAL_TASK_DEQUE_SIZE);
4199 
4200   // compute num_tasks/grainsize based on the input provided
4201   switch (sched) {
4202   case 0: // no schedule clause specified, we can choose the default
4203     // let's try to schedule (team_size*10) tasks
4204     grainsize = thread->th.th_team_nproc * 10;
4205     KMP_FALLTHROUGH();
4206   case 2: // num_tasks provided
4207     if (grainsize > tc) {
4208       num_tasks = tc; // too big num_tasks requested, adjust values
4209       grainsize = 1;
4210       extras = 0;
4211     } else {
4212       num_tasks = grainsize;
4213       grainsize = tc / num_tasks;
4214       extras = tc % num_tasks;
4215     }
4216     break;
4217   case 1: // grainsize provided
4218     if (grainsize > tc) {
4219       num_tasks = 1; // too big grainsize requested, adjust values
4220       grainsize = tc;
4221       extras = 0;
4222     } else {
4223       num_tasks = tc / grainsize;
4224       // adjust grainsize for balanced distribution of iterations
4225       grainsize = tc / num_tasks;
4226       extras = tc % num_tasks;
4227     }
4228     break;
4229   default:
4230     KMP_ASSERT2(0, "unknown scheduling of taskloop");
4231   }
4232   KMP_DEBUG_ASSERT(tc == num_tasks * grainsize + extras);
4233   KMP_DEBUG_ASSERT(num_tasks > extras);
4234   KMP_DEBUG_ASSERT(num_tasks > 0);
4235   // =========================================================================
4236 
4237   // check if clause value first
4238   // Also require GOMP_taskloop to reduce to linear (taskdata->td_flags.native)
4239   if (if_val == 0) { // if(0) specified, mark task as serial
4240     taskdata->td_flags.task_serial = 1;
4241     taskdata->td_flags.tiedness = TASK_TIED; // AC: serial task cannot be untied
4242     // always start serial tasks linearly
4243     __kmp_taskloop_linear(loc, gtid, task, lb, ub, st, ub_glob, num_tasks,
4244                           grainsize, extras, tc,
4245 #if OMPT_SUPPORT
4246                           OMPT_GET_RETURN_ADDRESS(0),
4247 #endif
4248                           task_dup);
4249     // !taskdata->td_flags.native => currently force linear spawning of tasks
4250     // for GOMP_taskloop
4251   } else if (num_tasks > num_tasks_min && !taskdata->td_flags.native) {
4252     KA_TRACE(20, ("__kmpc_taskloop: T#%d, go recursive: tc %llu, #tasks %llu"
4253                   "(%lld), grain %llu, extras %llu\n",
4254                   gtid, tc, num_tasks, num_tasks_min, grainsize, extras));
4255     __kmp_taskloop_recur(loc, gtid, task, lb, ub, st, ub_glob, num_tasks,
4256                          grainsize, extras, tc, num_tasks_min,
4257 #if OMPT_SUPPORT
4258                          OMPT_GET_RETURN_ADDRESS(0),
4259 #endif
4260                          task_dup);
4261   } else {
4262     KA_TRACE(20, ("__kmpc_taskloop: T#%d, go linear: tc %llu, #tasks %llu"
4263                   "(%lld), grain %llu, extras %llu\n",
4264                   gtid, tc, num_tasks, num_tasks_min, grainsize, extras));
4265     __kmp_taskloop_linear(loc, gtid, task, lb, ub, st, ub_glob, num_tasks,
4266                           grainsize, extras, tc,
4267 #if OMPT_SUPPORT
4268                           OMPT_GET_RETURN_ADDRESS(0),
4269 #endif
4270                           task_dup);
4271   }
4272 
4273 #if OMPT_SUPPORT && OMPT_OPTIONAL
4274   if (ompt_enabled.ompt_callback_work) {
4275     ompt_callbacks.ompt_callback(ompt_callback_work)(
4276         ompt_work_taskloop, ompt_scope_end, &(team_info->parallel_data),
4277         &(task_info->task_data), tc, OMPT_GET_RETURN_ADDRESS(0));
4278   }
4279 #endif
4280 
4281   if (nogroup == 0) {
4282 #if OMPT_SUPPORT && OMPT_OPTIONAL
4283     OMPT_STORE_RETURN_ADDRESS(gtid);
4284 #endif
4285     __kmpc_end_taskgroup(loc, gtid);
4286   }
4287   KA_TRACE(20, ("__kmpc_taskloop(exit): T#%d\n", gtid));
4288 }
4289 
4290 #endif
4291