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 #if ENABLE_LIBOMPTARGET
25 // Declaration of synchronization function from libomptarget.
26 extern "C" void __tgt_target_nowait_query(void **) KMP_WEAK_ATTRIBUTE_INTERNAL;
27 #endif
28 
29 /* forward declaration */
30 static void __kmp_enable_tasking(kmp_task_team_t *task_team,
31                                  kmp_info_t *this_thr);
32 static void __kmp_alloc_task_deque(kmp_info_t *thread,
33                                    kmp_thread_data_t *thread_data);
34 static int __kmp_realloc_task_threads_data(kmp_info_t *thread,
35                                            kmp_task_team_t *task_team);
36 static void __kmp_bottom_half_finish_proxy(kmp_int32 gtid, kmp_task_t *ptask);
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 (UNLIKELY(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_DEBUG_ASSERT(TCR_4(thread_data->td.td_deque_ntasks) == size);
305   kmp_int32 new_size = 2 * size;
306 
307   KE_TRACE(10, ("__kmp_realloc_task_deque: T#%d reallocating deque[from %d to "
308                 "%d] for thread_data %p\n",
309                 __kmp_gtid_from_thread(thread), size, new_size, thread_data));
310 
311   kmp_taskdata_t **new_deque =
312       (kmp_taskdata_t **)__kmp_allocate(new_size * sizeof(kmp_taskdata_t *));
313 
314   int i, j;
315   for (i = thread_data->td.td_deque_head, j = 0; j < size;
316        i = (i + 1) & TASK_DEQUE_MASK(thread_data->td), j++)
317     new_deque[j] = thread_data->td.td_deque[i];
318 
319   __kmp_free(thread_data->td.td_deque);
320 
321   thread_data->td.td_deque_head = 0;
322   thread_data->td.td_deque_tail = size;
323   thread_data->td.td_deque = new_deque;
324   thread_data->td.td_deque_size = new_size;
325 }
326 
327 static kmp_task_pri_t *__kmp_alloc_task_pri_list() {
328   kmp_task_pri_t *l = (kmp_task_pri_t *)__kmp_allocate(sizeof(kmp_task_pri_t));
329   kmp_thread_data_t *thread_data = &l->td;
330   __kmp_init_bootstrap_lock(&thread_data->td.td_deque_lock);
331   thread_data->td.td_deque_last_stolen = -1;
332   KE_TRACE(20, ("__kmp_alloc_task_pri_list: T#%d allocating deque[%d] "
333                 "for thread_data %p\n",
334                 __kmp_get_gtid(), INITIAL_TASK_DEQUE_SIZE, thread_data));
335   thread_data->td.td_deque = (kmp_taskdata_t **)__kmp_allocate(
336       INITIAL_TASK_DEQUE_SIZE * sizeof(kmp_taskdata_t *));
337   thread_data->td.td_deque_size = INITIAL_TASK_DEQUE_SIZE;
338   return l;
339 }
340 
341 // The function finds the deque of priority tasks with given priority, or
342 // allocates a new deque and put it into sorted (high -> low) list of deques.
343 // Deques of non-default priority tasks are shared between all threads in team,
344 // as opposed to per-thread deques of tasks with default priority.
345 // The function is called under the lock task_team->tt.tt_task_pri_lock.
346 static kmp_thread_data_t *
347 __kmp_get_priority_deque_data(kmp_task_team_t *task_team, kmp_int32 pri) {
348   kmp_thread_data_t *thread_data;
349   kmp_task_pri_t *lst = task_team->tt.tt_task_pri_list;
350   if (lst->priority == pri) {
351     // Found queue of tasks with given priority.
352     thread_data = &lst->td;
353   } else if (lst->priority < pri) {
354     // All current priority queues contain tasks with lower priority.
355     // Allocate new one for given priority tasks.
356     kmp_task_pri_t *list = __kmp_alloc_task_pri_list();
357     thread_data = &list->td;
358     list->priority = pri;
359     list->next = lst;
360     task_team->tt.tt_task_pri_list = list;
361   } else { // task_team->tt.tt_task_pri_list->priority > pri
362     kmp_task_pri_t *next_queue = lst->next;
363     while (next_queue && next_queue->priority > pri) {
364       lst = next_queue;
365       next_queue = lst->next;
366     }
367     // lst->priority > pri && (next == NULL || pri >= next->priority)
368     if (next_queue == NULL) {
369       // No queue with pri priority, need to allocate new one.
370       kmp_task_pri_t *list = __kmp_alloc_task_pri_list();
371       thread_data = &list->td;
372       list->priority = pri;
373       list->next = NULL;
374       lst->next = list;
375     } else if (next_queue->priority == pri) {
376       // Found queue of tasks with given priority.
377       thread_data = &next_queue->td;
378     } else { // lst->priority > pri > next->priority
379       // insert newly allocated between existed queues
380       kmp_task_pri_t *list = __kmp_alloc_task_pri_list();
381       thread_data = &list->td;
382       list->priority = pri;
383       list->next = next_queue;
384       lst->next = list;
385     }
386   }
387   return thread_data;
388 }
389 
390 //  __kmp_push_priority_task: Add a task to the team's priority task deque
391 static kmp_int32 __kmp_push_priority_task(kmp_int32 gtid, kmp_info_t *thread,
392                                           kmp_taskdata_t *taskdata,
393                                           kmp_task_team_t *task_team,
394                                           kmp_int32 pri) {
395   kmp_thread_data_t *thread_data = NULL;
396   KA_TRACE(20,
397            ("__kmp_push_priority_task: T#%d trying to push task %p, pri %d.\n",
398             gtid, taskdata, pri));
399 
400   // Find task queue specific to priority value
401   kmp_task_pri_t *lst = task_team->tt.tt_task_pri_list;
402   if (UNLIKELY(lst == NULL)) {
403     __kmp_acquire_bootstrap_lock(&task_team->tt.tt_task_pri_lock);
404     if (task_team->tt.tt_task_pri_list == NULL) {
405       // List of queues is still empty, allocate one.
406       kmp_task_pri_t *list = __kmp_alloc_task_pri_list();
407       thread_data = &list->td;
408       list->priority = pri;
409       list->next = NULL;
410       task_team->tt.tt_task_pri_list = list;
411     } else {
412       // Other thread initialized a queue. Check if it fits and get thread_data.
413       thread_data = __kmp_get_priority_deque_data(task_team, pri);
414     }
415     __kmp_release_bootstrap_lock(&task_team->tt.tt_task_pri_lock);
416   } else {
417     if (lst->priority == pri) {
418       // Found queue of tasks with given priority.
419       thread_data = &lst->td;
420     } else {
421       __kmp_acquire_bootstrap_lock(&task_team->tt.tt_task_pri_lock);
422       thread_data = __kmp_get_priority_deque_data(task_team, pri);
423       __kmp_release_bootstrap_lock(&task_team->tt.tt_task_pri_lock);
424     }
425   }
426   KMP_DEBUG_ASSERT(thread_data);
427 
428   __kmp_acquire_bootstrap_lock(&thread_data->td.td_deque_lock);
429   // Check if deque is full
430   if (TCR_4(thread_data->td.td_deque_ntasks) >=
431       TASK_DEQUE_SIZE(thread_data->td)) {
432     if (__kmp_enable_task_throttling &&
433         __kmp_task_is_allowed(gtid, __kmp_task_stealing_constraint, taskdata,
434                               thread->th.th_current_task)) {
435       __kmp_release_bootstrap_lock(&thread_data->td.td_deque_lock);
436       KA_TRACE(20, ("__kmp_push_priority_task: T#%d deque is full; returning "
437                     "TASK_NOT_PUSHED for task %p\n",
438                     gtid, taskdata));
439       return TASK_NOT_PUSHED;
440     } else {
441       // expand deque to push the task which is not allowed to execute
442       __kmp_realloc_task_deque(thread, thread_data);
443     }
444   }
445   KMP_DEBUG_ASSERT(TCR_4(thread_data->td.td_deque_ntasks) <
446                    TASK_DEQUE_SIZE(thread_data->td));
447   // Push taskdata.
448   thread_data->td.td_deque[thread_data->td.td_deque_tail] = taskdata;
449   // Wrap index.
450   thread_data->td.td_deque_tail =
451       (thread_data->td.td_deque_tail + 1) & TASK_DEQUE_MASK(thread_data->td);
452   TCW_4(thread_data->td.td_deque_ntasks,
453         TCR_4(thread_data->td.td_deque_ntasks) + 1); // Adjust task count
454   KMP_FSYNC_RELEASING(thread->th.th_current_task); // releasing self
455   KMP_FSYNC_RELEASING(taskdata); // releasing child
456   KA_TRACE(20, ("__kmp_push_priority_task: T#%d returning "
457                 "TASK_SUCCESSFULLY_PUSHED: task=%p ntasks=%d head=%u tail=%u\n",
458                 gtid, taskdata, thread_data->td.td_deque_ntasks,
459                 thread_data->td.td_deque_head, thread_data->td.td_deque_tail));
460   __kmp_release_bootstrap_lock(&thread_data->td.td_deque_lock);
461   task_team->tt.tt_num_task_pri++; // atomic inc
462   return TASK_SUCCESSFULLY_PUSHED;
463 }
464 
465 //  __kmp_push_task: Add a task to the thread's deque
466 static kmp_int32 __kmp_push_task(kmp_int32 gtid, kmp_task_t *task) {
467   kmp_info_t *thread = __kmp_threads[gtid];
468   kmp_taskdata_t *taskdata = KMP_TASK_TO_TASKDATA(task);
469 
470   // If we encounter a hidden helper task, and the current thread is not a
471   // hidden helper thread, we have to give the task to any hidden helper thread
472   // starting from its shadow one.
473   if (UNLIKELY(taskdata->td_flags.hidden_helper &&
474                !KMP_HIDDEN_HELPER_THREAD(gtid))) {
475     kmp_int32 shadow_gtid = KMP_GTID_TO_SHADOW_GTID(gtid);
476     __kmpc_give_task(task, __kmp_tid_from_gtid(shadow_gtid));
477     // Signal the hidden helper threads.
478     __kmp_hidden_helper_worker_thread_signal();
479     return TASK_SUCCESSFULLY_PUSHED;
480   }
481 
482   kmp_task_team_t *task_team = thread->th.th_task_team;
483   kmp_int32 tid = __kmp_tid_from_gtid(gtid);
484   kmp_thread_data_t *thread_data;
485 
486   KA_TRACE(20,
487            ("__kmp_push_task: T#%d trying to push task %p.\n", gtid, taskdata));
488 
489   if (UNLIKELY(taskdata->td_flags.tiedness == TASK_UNTIED)) {
490     // untied task needs to increment counter so that the task structure is not
491     // freed prematurely
492     kmp_int32 counter = 1 + KMP_ATOMIC_INC(&taskdata->td_untied_count);
493     KMP_DEBUG_USE_VAR(counter);
494     KA_TRACE(
495         20,
496         ("__kmp_push_task: T#%d untied_count (%d) incremented for task %p\n",
497          gtid, counter, taskdata));
498   }
499 
500   // The first check avoids building task_team thread data if serialized
501   if (UNLIKELY(taskdata->td_flags.task_serial)) {
502     KA_TRACE(20, ("__kmp_push_task: T#%d team serialized; returning "
503                   "TASK_NOT_PUSHED for task %p\n",
504                   gtid, taskdata));
505     return TASK_NOT_PUSHED;
506   }
507 
508   // Now that serialized tasks have returned, we can assume that we are not in
509   // immediate exec mode
510   KMP_DEBUG_ASSERT(__kmp_tasking_mode != tskm_immediate_exec);
511   if (UNLIKELY(!KMP_TASKING_ENABLED(task_team))) {
512     __kmp_enable_tasking(task_team, thread);
513   }
514   KMP_DEBUG_ASSERT(TCR_4(task_team->tt.tt_found_tasks) == TRUE);
515   KMP_DEBUG_ASSERT(TCR_PTR(task_team->tt.tt_threads_data) != NULL);
516 
517   if (taskdata->td_flags.priority_specified && task->data2.priority > 0 &&
518       __kmp_max_task_priority > 0) {
519     int pri = KMP_MIN(task->data2.priority, __kmp_max_task_priority);
520     return __kmp_push_priority_task(gtid, thread, taskdata, task_team, pri);
521   }
522 
523   // Find tasking deque specific to encountering thread
524   thread_data = &task_team->tt.tt_threads_data[tid];
525 
526   // No lock needed since only owner can allocate. If the task is hidden_helper,
527   // we don't need it either because we have initialized the dequeue for hidden
528   // helper thread data.
529   if (UNLIKELY(thread_data->td.td_deque == NULL)) {
530     __kmp_alloc_task_deque(thread, thread_data);
531   }
532 
533   int locked = 0;
534   // Check if deque is full
535   if (TCR_4(thread_data->td.td_deque_ntasks) >=
536       TASK_DEQUE_SIZE(thread_data->td)) {
537     if (__kmp_enable_task_throttling &&
538         __kmp_task_is_allowed(gtid, __kmp_task_stealing_constraint, taskdata,
539                               thread->th.th_current_task)) {
540       KA_TRACE(20, ("__kmp_push_task: T#%d deque is full; returning "
541                     "TASK_NOT_PUSHED for task %p\n",
542                     gtid, taskdata));
543       return TASK_NOT_PUSHED;
544     } else {
545       __kmp_acquire_bootstrap_lock(&thread_data->td.td_deque_lock);
546       locked = 1;
547       if (TCR_4(thread_data->td.td_deque_ntasks) >=
548           TASK_DEQUE_SIZE(thread_data->td)) {
549         // expand deque to push the task which is not allowed to execute
550         __kmp_realloc_task_deque(thread, thread_data);
551       }
552     }
553   }
554   // Lock the deque for the task push operation
555   if (!locked) {
556     __kmp_acquire_bootstrap_lock(&thread_data->td.td_deque_lock);
557     // Need to recheck as we can get a proxy task from thread outside of OpenMP
558     if (TCR_4(thread_data->td.td_deque_ntasks) >=
559         TASK_DEQUE_SIZE(thread_data->td)) {
560       if (__kmp_enable_task_throttling &&
561           __kmp_task_is_allowed(gtid, __kmp_task_stealing_constraint, taskdata,
562                                 thread->th.th_current_task)) {
563         __kmp_release_bootstrap_lock(&thread_data->td.td_deque_lock);
564         KA_TRACE(20, ("__kmp_push_task: T#%d deque is full on 2nd check; "
565                       "returning TASK_NOT_PUSHED for task %p\n",
566                       gtid, taskdata));
567         return TASK_NOT_PUSHED;
568       } else {
569         // expand deque to push the task which is not allowed to execute
570         __kmp_realloc_task_deque(thread, thread_data);
571       }
572     }
573   }
574   // Must have room since no thread can add tasks but calling thread
575   KMP_DEBUG_ASSERT(TCR_4(thread_data->td.td_deque_ntasks) <
576                    TASK_DEQUE_SIZE(thread_data->td));
577 
578   thread_data->td.td_deque[thread_data->td.td_deque_tail] =
579       taskdata; // Push taskdata
580   // Wrap index.
581   thread_data->td.td_deque_tail =
582       (thread_data->td.td_deque_tail + 1) & TASK_DEQUE_MASK(thread_data->td);
583   TCW_4(thread_data->td.td_deque_ntasks,
584         TCR_4(thread_data->td.td_deque_ntasks) + 1); // Adjust task count
585   KMP_FSYNC_RELEASING(thread->th.th_current_task); // releasing self
586   KMP_FSYNC_RELEASING(taskdata); // releasing child
587   KA_TRACE(20, ("__kmp_push_task: T#%d returning TASK_SUCCESSFULLY_PUSHED: "
588                 "task=%p ntasks=%d head=%u tail=%u\n",
589                 gtid, taskdata, thread_data->td.td_deque_ntasks,
590                 thread_data->td.td_deque_head, thread_data->td.td_deque_tail));
591 
592   __kmp_release_bootstrap_lock(&thread_data->td.td_deque_lock);
593 
594   return TASK_SUCCESSFULLY_PUSHED;
595 }
596 
597 // __kmp_pop_current_task_from_thread: set up current task from called thread
598 // when team ends
599 //
600 // this_thr: thread structure to set current_task in.
601 void __kmp_pop_current_task_from_thread(kmp_info_t *this_thr) {
602   KF_TRACE(10, ("__kmp_pop_current_task_from_thread(enter): T#%d "
603                 "this_thread=%p, curtask=%p, "
604                 "curtask_parent=%p\n",
605                 0, this_thr, this_thr->th.th_current_task,
606                 this_thr->th.th_current_task->td_parent));
607 
608   this_thr->th.th_current_task = this_thr->th.th_current_task->td_parent;
609 
610   KF_TRACE(10, ("__kmp_pop_current_task_from_thread(exit): T#%d "
611                 "this_thread=%p, curtask=%p, "
612                 "curtask_parent=%p\n",
613                 0, this_thr, this_thr->th.th_current_task,
614                 this_thr->th.th_current_task->td_parent));
615 }
616 
617 // __kmp_push_current_task_to_thread: set up current task in called thread for a
618 // new team
619 //
620 // this_thr: thread structure to set up
621 // team: team for implicit task data
622 // tid: thread within team to set up
623 void __kmp_push_current_task_to_thread(kmp_info_t *this_thr, kmp_team_t *team,
624                                        int tid) {
625   // current task of the thread is a parent of the new just created implicit
626   // tasks of new team
627   KF_TRACE(10, ("__kmp_push_current_task_to_thread(enter): T#%d this_thread=%p "
628                 "curtask=%p "
629                 "parent_task=%p\n",
630                 tid, this_thr, this_thr->th.th_current_task,
631                 team->t.t_implicit_task_taskdata[tid].td_parent));
632 
633   KMP_DEBUG_ASSERT(this_thr != NULL);
634 
635   if (tid == 0) {
636     if (this_thr->th.th_current_task != &team->t.t_implicit_task_taskdata[0]) {
637       team->t.t_implicit_task_taskdata[0].td_parent =
638           this_thr->th.th_current_task;
639       this_thr->th.th_current_task = &team->t.t_implicit_task_taskdata[0];
640     }
641   } else {
642     team->t.t_implicit_task_taskdata[tid].td_parent =
643         team->t.t_implicit_task_taskdata[0].td_parent;
644     this_thr->th.th_current_task = &team->t.t_implicit_task_taskdata[tid];
645   }
646 
647   KF_TRACE(10, ("__kmp_push_current_task_to_thread(exit): T#%d this_thread=%p "
648                 "curtask=%p "
649                 "parent_task=%p\n",
650                 tid, this_thr, this_thr->th.th_current_task,
651                 team->t.t_implicit_task_taskdata[tid].td_parent));
652 }
653 
654 // __kmp_task_start: bookkeeping for a task starting execution
655 //
656 // GTID: global thread id of calling thread
657 // task: task starting execution
658 // current_task: task suspending
659 static void __kmp_task_start(kmp_int32 gtid, kmp_task_t *task,
660                              kmp_taskdata_t *current_task) {
661   kmp_taskdata_t *taskdata = KMP_TASK_TO_TASKDATA(task);
662   kmp_info_t *thread = __kmp_threads[gtid];
663 
664   KA_TRACE(10,
665            ("__kmp_task_start(enter): T#%d starting task %p: current_task=%p\n",
666             gtid, taskdata, current_task));
667 
668   KMP_DEBUG_ASSERT(taskdata->td_flags.tasktype == TASK_EXPLICIT);
669 
670   // mark currently executing task as suspended
671   // TODO: GEH - make sure root team implicit task is initialized properly.
672   // KMP_DEBUG_ASSERT( current_task -> td_flags.executing == 1 );
673   current_task->td_flags.executing = 0;
674 
675 // Add task to stack if tied
676 #ifdef BUILD_TIED_TASK_STACK
677   if (taskdata->td_flags.tiedness == TASK_TIED) {
678     __kmp_push_task_stack(gtid, thread, taskdata);
679   }
680 #endif /* BUILD_TIED_TASK_STACK */
681 
682   // mark starting task as executing and as current task
683   thread->th.th_current_task = taskdata;
684 
685   KMP_DEBUG_ASSERT(taskdata->td_flags.started == 0 ||
686                    taskdata->td_flags.tiedness == TASK_UNTIED);
687   KMP_DEBUG_ASSERT(taskdata->td_flags.executing == 0 ||
688                    taskdata->td_flags.tiedness == TASK_UNTIED);
689   taskdata->td_flags.started = 1;
690   taskdata->td_flags.executing = 1;
691   KMP_DEBUG_ASSERT(taskdata->td_flags.complete == 0);
692   KMP_DEBUG_ASSERT(taskdata->td_flags.freed == 0);
693 
694   // GEH TODO: shouldn't we pass some sort of location identifier here?
695   // APT: yes, we will pass location here.
696   // need to store current thread state (in a thread or taskdata structure)
697   // before setting work_state, otherwise wrong state is set after end of task
698 
699   KA_TRACE(10, ("__kmp_task_start(exit): T#%d task=%p\n", gtid, taskdata));
700 
701   return;
702 }
703 
704 #if OMPT_SUPPORT
705 //------------------------------------------------------------------------------
706 // __ompt_task_init:
707 //   Initialize OMPT fields maintained by a task. This will only be called after
708 //   ompt_start_tool, so we already know whether ompt is enabled or not.
709 
710 static inline void __ompt_task_init(kmp_taskdata_t *task, int tid) {
711   // The calls to __ompt_task_init already have the ompt_enabled condition.
712   task->ompt_task_info.task_data.value = 0;
713   task->ompt_task_info.frame.exit_frame = ompt_data_none;
714   task->ompt_task_info.frame.enter_frame = ompt_data_none;
715   task->ompt_task_info.frame.exit_frame_flags =
716       ompt_frame_runtime | ompt_frame_framepointer;
717   task->ompt_task_info.frame.enter_frame_flags =
718       ompt_frame_runtime | ompt_frame_framepointer;
719   task->ompt_task_info.dispatch_chunk.start = 0;
720   task->ompt_task_info.dispatch_chunk.iterations = 0;
721 }
722 
723 // __ompt_task_start:
724 //   Build and trigger task-begin event
725 static inline void __ompt_task_start(kmp_task_t *task,
726                                      kmp_taskdata_t *current_task,
727                                      kmp_int32 gtid) {
728   kmp_taskdata_t *taskdata = KMP_TASK_TO_TASKDATA(task);
729   ompt_task_status_t status = ompt_task_switch;
730   if (__kmp_threads[gtid]->th.ompt_thread_info.ompt_task_yielded) {
731     status = ompt_task_yield;
732     __kmp_threads[gtid]->th.ompt_thread_info.ompt_task_yielded = 0;
733   }
734   /* let OMPT know that we're about to run this task */
735   if (ompt_enabled.ompt_callback_task_schedule) {
736     ompt_callbacks.ompt_callback(ompt_callback_task_schedule)(
737         &(current_task->ompt_task_info.task_data), status,
738         &(taskdata->ompt_task_info.task_data));
739   }
740   taskdata->ompt_task_info.scheduling_parent = current_task;
741 }
742 
743 // __ompt_task_finish:
744 //   Build and trigger final task-schedule event
745 static inline void __ompt_task_finish(kmp_task_t *task,
746                                       kmp_taskdata_t *resumed_task,
747                                       ompt_task_status_t status) {
748   if (ompt_enabled.ompt_callback_task_schedule) {
749     kmp_taskdata_t *taskdata = KMP_TASK_TO_TASKDATA(task);
750     if (__kmp_omp_cancellation && taskdata->td_taskgroup &&
751         taskdata->td_taskgroup->cancel_request == cancel_taskgroup) {
752       status = ompt_task_cancel;
753     }
754 
755     /* let OMPT know that we're returning to the callee task */
756     ompt_callbacks.ompt_callback(ompt_callback_task_schedule)(
757         &(taskdata->ompt_task_info.task_data), status,
758         (resumed_task ? &(resumed_task->ompt_task_info.task_data) : NULL));
759   }
760 }
761 #endif
762 
763 template <bool ompt>
764 static void __kmpc_omp_task_begin_if0_template(ident_t *loc_ref, kmp_int32 gtid,
765                                                kmp_task_t *task,
766                                                void *frame_address,
767                                                void *return_address) {
768   kmp_taskdata_t *taskdata = KMP_TASK_TO_TASKDATA(task);
769   kmp_taskdata_t *current_task = __kmp_threads[gtid]->th.th_current_task;
770 
771   KA_TRACE(10, ("__kmpc_omp_task_begin_if0(enter): T#%d loc=%p task=%p "
772                 "current_task=%p\n",
773                 gtid, loc_ref, taskdata, current_task));
774 
775   if (UNLIKELY(taskdata->td_flags.tiedness == TASK_UNTIED)) {
776     // untied task needs to increment counter so that the task structure is not
777     // freed prematurely
778     kmp_int32 counter = 1 + KMP_ATOMIC_INC(&taskdata->td_untied_count);
779     KMP_DEBUG_USE_VAR(counter);
780     KA_TRACE(20, ("__kmpc_omp_task_begin_if0: T#%d untied_count (%d) "
781                   "incremented for task %p\n",
782                   gtid, counter, taskdata));
783   }
784 
785   taskdata->td_flags.task_serial =
786       1; // Execute this task immediately, not deferred.
787   __kmp_task_start(gtid, task, current_task);
788 
789 #if OMPT_SUPPORT
790   if (ompt) {
791     if (current_task->ompt_task_info.frame.enter_frame.ptr == NULL) {
792       current_task->ompt_task_info.frame.enter_frame.ptr =
793           taskdata->ompt_task_info.frame.exit_frame.ptr = frame_address;
794       current_task->ompt_task_info.frame.enter_frame_flags =
795           taskdata->ompt_task_info.frame.exit_frame_flags =
796               ompt_frame_application | ompt_frame_framepointer;
797     }
798     if (ompt_enabled.ompt_callback_task_create) {
799       ompt_task_info_t *parent_info = &(current_task->ompt_task_info);
800       ompt_callbacks.ompt_callback(ompt_callback_task_create)(
801           &(parent_info->task_data), &(parent_info->frame),
802           &(taskdata->ompt_task_info.task_data),
803           ompt_task_explicit | TASK_TYPE_DETAILS_FORMAT(taskdata), 0,
804           return_address);
805     }
806     __ompt_task_start(task, current_task, gtid);
807   }
808 #endif // OMPT_SUPPORT
809 
810   KA_TRACE(10, ("__kmpc_omp_task_begin_if0(exit): T#%d loc=%p task=%p,\n", gtid,
811                 loc_ref, taskdata));
812 }
813 
814 #if OMPT_SUPPORT
815 OMPT_NOINLINE
816 static void __kmpc_omp_task_begin_if0_ompt(ident_t *loc_ref, kmp_int32 gtid,
817                                            kmp_task_t *task,
818                                            void *frame_address,
819                                            void *return_address) {
820   __kmpc_omp_task_begin_if0_template<true>(loc_ref, gtid, task, frame_address,
821                                            return_address);
822 }
823 #endif // OMPT_SUPPORT
824 
825 // __kmpc_omp_task_begin_if0: report that a given serialized task has started
826 // execution
827 //
828 // loc_ref: source location information; points to beginning of task block.
829 // gtid: global thread number.
830 // task: task thunk for the started task.
831 void __kmpc_omp_task_begin_if0(ident_t *loc_ref, kmp_int32 gtid,
832                                kmp_task_t *task) {
833 #if OMPT_SUPPORT
834   if (UNLIKELY(ompt_enabled.enabled)) {
835     OMPT_STORE_RETURN_ADDRESS(gtid);
836     __kmpc_omp_task_begin_if0_ompt(loc_ref, gtid, task,
837                                    OMPT_GET_FRAME_ADDRESS(1),
838                                    OMPT_LOAD_RETURN_ADDRESS(gtid));
839     return;
840   }
841 #endif
842   __kmpc_omp_task_begin_if0_template<false>(loc_ref, gtid, task, NULL, NULL);
843 }
844 
845 #ifdef TASK_UNUSED
846 // __kmpc_omp_task_begin: report that a given task has started execution
847 // NEVER GENERATED BY COMPILER, DEPRECATED!!!
848 void __kmpc_omp_task_begin(ident_t *loc_ref, kmp_int32 gtid, kmp_task_t *task) {
849   kmp_taskdata_t *current_task = __kmp_threads[gtid]->th.th_current_task;
850 
851   KA_TRACE(
852       10,
853       ("__kmpc_omp_task_begin(enter): T#%d loc=%p task=%p current_task=%p\n",
854        gtid, loc_ref, KMP_TASK_TO_TASKDATA(task), current_task));
855 
856   __kmp_task_start(gtid, task, current_task);
857 
858   KA_TRACE(10, ("__kmpc_omp_task_begin(exit): T#%d loc=%p task=%p,\n", gtid,
859                 loc_ref, KMP_TASK_TO_TASKDATA(task)));
860   return;
861 }
862 #endif // TASK_UNUSED
863 
864 // __kmp_free_task: free the current task space and the space for shareds
865 //
866 // gtid: Global thread ID of calling thread
867 // taskdata: task to free
868 // thread: thread data structure of caller
869 static void __kmp_free_task(kmp_int32 gtid, kmp_taskdata_t *taskdata,
870                             kmp_info_t *thread) {
871   KA_TRACE(30, ("__kmp_free_task: T#%d freeing data from task %p\n", gtid,
872                 taskdata));
873 
874   // Check to make sure all flags and counters have the correct values
875   KMP_DEBUG_ASSERT(taskdata->td_flags.tasktype == TASK_EXPLICIT);
876   KMP_DEBUG_ASSERT(taskdata->td_flags.executing == 0);
877   KMP_DEBUG_ASSERT(taskdata->td_flags.complete == 1);
878   KMP_DEBUG_ASSERT(taskdata->td_flags.freed == 0);
879   KMP_DEBUG_ASSERT(taskdata->td_allocated_child_tasks == 0 ||
880                    taskdata->td_flags.task_serial == 1);
881   KMP_DEBUG_ASSERT(taskdata->td_incomplete_child_tasks == 0);
882   kmp_task_t *task = KMP_TASKDATA_TO_TASK(taskdata);
883   // Clear data to not be re-used later by mistake.
884   task->data1.destructors = NULL;
885   task->data2.priority = 0;
886 
887   taskdata->td_flags.freed = 1;
888 // deallocate the taskdata and shared variable blocks associated with this task
889 #if USE_FAST_MEMORY
890   __kmp_fast_free(thread, taskdata);
891 #else /* ! USE_FAST_MEMORY */
892   __kmp_thread_free(thread, taskdata);
893 #endif
894   KA_TRACE(20, ("__kmp_free_task: T#%d freed task %p\n", gtid, taskdata));
895 }
896 
897 // __kmp_free_task_and_ancestors: free the current task and ancestors without
898 // children
899 //
900 // gtid: Global thread ID of calling thread
901 // taskdata: task to free
902 // thread: thread data structure of caller
903 static void __kmp_free_task_and_ancestors(kmp_int32 gtid,
904                                           kmp_taskdata_t *taskdata,
905                                           kmp_info_t *thread) {
906   // Proxy tasks must always be allowed to free their parents
907   // because they can be run in background even in serial mode.
908   kmp_int32 team_serial =
909       (taskdata->td_flags.team_serial || taskdata->td_flags.tasking_ser) &&
910       !taskdata->td_flags.proxy;
911   KMP_DEBUG_ASSERT(taskdata->td_flags.tasktype == TASK_EXPLICIT);
912 
913   kmp_int32 children = KMP_ATOMIC_DEC(&taskdata->td_allocated_child_tasks) - 1;
914   KMP_DEBUG_ASSERT(children >= 0);
915 
916   // Now, go up the ancestor tree to see if any ancestors can now be freed.
917   while (children == 0) {
918     kmp_taskdata_t *parent_taskdata = taskdata->td_parent;
919 
920     KA_TRACE(20, ("__kmp_free_task_and_ancestors(enter): T#%d task %p complete "
921                   "and freeing itself\n",
922                   gtid, taskdata));
923 
924     // --- Deallocate my ancestor task ---
925     __kmp_free_task(gtid, taskdata, thread);
926 
927     taskdata = parent_taskdata;
928 
929     if (team_serial)
930       return;
931     // Stop checking ancestors at implicit task instead of walking up ancestor
932     // tree to avoid premature deallocation of ancestors.
933     if (taskdata->td_flags.tasktype == TASK_IMPLICIT) {
934       if (taskdata->td_dephash) { // do we need to cleanup dephash?
935         int children = KMP_ATOMIC_LD_ACQ(&taskdata->td_incomplete_child_tasks);
936         kmp_tasking_flags_t flags_old = taskdata->td_flags;
937         if (children == 0 && flags_old.complete == 1) {
938           kmp_tasking_flags_t flags_new = flags_old;
939           flags_new.complete = 0;
940           if (KMP_COMPARE_AND_STORE_ACQ32(
941                   RCAST(kmp_int32 *, &taskdata->td_flags),
942                   *RCAST(kmp_int32 *, &flags_old),
943                   *RCAST(kmp_int32 *, &flags_new))) {
944             KA_TRACE(100, ("__kmp_free_task_and_ancestors: T#%d cleans "
945                            "dephash of implicit task %p\n",
946                            gtid, taskdata));
947             // cleanup dephash of finished implicit task
948             __kmp_dephash_free_entries(thread, taskdata->td_dephash);
949           }
950         }
951       }
952       return;
953     }
954     // Predecrement simulated by "- 1" calculation
955     children = KMP_ATOMIC_DEC(&taskdata->td_allocated_child_tasks) - 1;
956     KMP_DEBUG_ASSERT(children >= 0);
957   }
958 
959   KA_TRACE(
960       20, ("__kmp_free_task_and_ancestors(exit): T#%d task %p has %d children; "
961            "not freeing it yet\n",
962            gtid, taskdata, children));
963 }
964 
965 // Only need to keep track of child task counts if any of the following:
966 // 1. team parallel and tasking not serialized;
967 // 2. it is a proxy or detachable or hidden helper task
968 // 3. the children counter of its parent task is greater than 0.
969 // The reason for the 3rd one is for serialized team that found detached task,
970 // hidden helper task, T. In this case, the execution of T is still deferred,
971 // and it is also possible that a regular task depends on T. In this case, if we
972 // don't track the children, task synchronization will be broken.
973 static bool __kmp_track_children_task(kmp_taskdata_t *taskdata) {
974   kmp_tasking_flags_t flags = taskdata->td_flags;
975   bool ret = !(flags.team_serial || flags.tasking_ser);
976   ret = ret || flags.proxy == TASK_PROXY ||
977         flags.detachable == TASK_DETACHABLE || flags.hidden_helper;
978   ret = ret ||
979         KMP_ATOMIC_LD_ACQ(&taskdata->td_parent->td_incomplete_child_tasks) > 0;
980   return ret;
981 }
982 
983 // __kmp_task_finish: bookkeeping to do when a task finishes execution
984 //
985 // gtid: global thread ID for calling thread
986 // task: task to be finished
987 // resumed_task: task to be resumed.  (may be NULL if task is serialized)
988 //
989 // template<ompt>: effectively ompt_enabled.enabled!=0
990 // the version with ompt=false is inlined, allowing to optimize away all ompt
991 // code in this case
992 template <bool ompt>
993 static void __kmp_task_finish(kmp_int32 gtid, kmp_task_t *task,
994                               kmp_taskdata_t *resumed_task) {
995   kmp_taskdata_t *taskdata = KMP_TASK_TO_TASKDATA(task);
996   kmp_info_t *thread = __kmp_threads[gtid];
997   kmp_task_team_t *task_team =
998       thread->th.th_task_team; // might be NULL for serial teams...
999 #if KMP_DEBUG
1000   kmp_int32 children = 0;
1001 #endif
1002   KA_TRACE(10, ("__kmp_task_finish(enter): T#%d finishing task %p and resuming "
1003                 "task %p\n",
1004                 gtid, taskdata, resumed_task));
1005 
1006   KMP_DEBUG_ASSERT(taskdata->td_flags.tasktype == TASK_EXPLICIT);
1007 
1008 // Pop task from stack if tied
1009 #ifdef BUILD_TIED_TASK_STACK
1010   if (taskdata->td_flags.tiedness == TASK_TIED) {
1011     __kmp_pop_task_stack(gtid, thread, taskdata);
1012   }
1013 #endif /* BUILD_TIED_TASK_STACK */
1014 
1015   if (UNLIKELY(taskdata->td_flags.tiedness == TASK_UNTIED)) {
1016     // untied task needs to check the counter so that the task structure is not
1017     // freed prematurely
1018     kmp_int32 counter = KMP_ATOMIC_DEC(&taskdata->td_untied_count) - 1;
1019     KA_TRACE(
1020         20,
1021         ("__kmp_task_finish: T#%d untied_count (%d) decremented for task %p\n",
1022          gtid, counter, taskdata));
1023     if (counter > 0) {
1024       // untied task is not done, to be continued possibly by other thread, do
1025       // not free it now
1026       if (resumed_task == NULL) {
1027         KMP_DEBUG_ASSERT(taskdata->td_flags.task_serial);
1028         resumed_task = taskdata->td_parent; // In a serialized task, the resumed
1029         // task is the parent
1030       }
1031       thread->th.th_current_task = resumed_task; // restore current_task
1032       resumed_task->td_flags.executing = 1; // resume previous task
1033       KA_TRACE(10, ("__kmp_task_finish(exit): T#%d partially done task %p, "
1034                     "resuming task %p\n",
1035                     gtid, taskdata, resumed_task));
1036       return;
1037     }
1038   }
1039 
1040   // bookkeeping for resuming task:
1041   // GEH - note tasking_ser => task_serial
1042   KMP_DEBUG_ASSERT(
1043       (taskdata->td_flags.tasking_ser || taskdata->td_flags.task_serial) ==
1044       taskdata->td_flags.task_serial);
1045   if (taskdata->td_flags.task_serial) {
1046     if (resumed_task == NULL) {
1047       resumed_task = taskdata->td_parent; // In a serialized task, the resumed
1048       // task is the parent
1049     }
1050   } else {
1051     KMP_DEBUG_ASSERT(resumed_task !=
1052                      NULL); // verify that resumed task is passed as argument
1053   }
1054 
1055   /* If the tasks' destructor thunk flag has been set, we need to invoke the
1056      destructor thunk that has been generated by the compiler. The code is
1057      placed here, since at this point other tasks might have been released
1058      hence overlapping the destructor invocations with some other work in the
1059      released tasks.  The OpenMP spec is not specific on when the destructors
1060      are invoked, so we should be free to choose. */
1061   if (UNLIKELY(taskdata->td_flags.destructors_thunk)) {
1062     kmp_routine_entry_t destr_thunk = task->data1.destructors;
1063     KMP_ASSERT(destr_thunk);
1064     destr_thunk(gtid, task);
1065   }
1066 
1067   KMP_DEBUG_ASSERT(taskdata->td_flags.complete == 0);
1068   KMP_DEBUG_ASSERT(taskdata->td_flags.started == 1);
1069   KMP_DEBUG_ASSERT(taskdata->td_flags.freed == 0);
1070 
1071   bool completed = true;
1072   if (UNLIKELY(taskdata->td_flags.detachable == TASK_DETACHABLE)) {
1073     if (taskdata->td_allow_completion_event.type ==
1074         KMP_EVENT_ALLOW_COMPLETION) {
1075       // event hasn't been fulfilled yet. Try to detach task.
1076       __kmp_acquire_tas_lock(&taskdata->td_allow_completion_event.lock, gtid);
1077       if (taskdata->td_allow_completion_event.type ==
1078           KMP_EVENT_ALLOW_COMPLETION) {
1079         // task finished execution
1080         KMP_DEBUG_ASSERT(taskdata->td_flags.executing == 1);
1081         taskdata->td_flags.executing = 0; // suspend the finishing task
1082 
1083 #if OMPT_SUPPORT
1084         // For a detached task, which is not completed, we switch back
1085         // the omp_fulfill_event signals completion
1086         // locking is necessary to avoid a race with ompt_task_late_fulfill
1087         if (ompt)
1088           __ompt_task_finish(task, resumed_task, ompt_task_detach);
1089 #endif
1090 
1091         // no access to taskdata after this point!
1092         // __kmp_fulfill_event might free taskdata at any time from now
1093 
1094         taskdata->td_flags.proxy = TASK_PROXY; // proxify!
1095         completed = false;
1096       }
1097       __kmp_release_tas_lock(&taskdata->td_allow_completion_event.lock, gtid);
1098     }
1099   }
1100 
1101   // Tasks with valid target async handles must be re-enqueued.
1102   if (taskdata->td_target_data.async_handle != NULL) {
1103     // Note: no need to translate gtid to its shadow. If the current thread is a
1104     // hidden helper one, then the gtid is already correct. Otherwise, hidden
1105     // helper threads are disabled, and gtid refers to a OpenMP thread.
1106     __kmpc_give_task(task, __kmp_tid_from_gtid(gtid));
1107     if (KMP_HIDDEN_HELPER_THREAD(gtid))
1108       __kmp_hidden_helper_worker_thread_signal();
1109     completed = false;
1110   }
1111 
1112   if (completed) {
1113     taskdata->td_flags.complete = 1; // mark the task as completed
1114 
1115 #if OMPT_SUPPORT
1116     // This is not a detached task, we are done here
1117     if (ompt)
1118       __ompt_task_finish(task, resumed_task, ompt_task_complete);
1119 #endif
1120     // TODO: What would be the balance between the conditions in the function
1121     // and an atomic operation?
1122     if (__kmp_track_children_task(taskdata)) {
1123       __kmp_release_deps(gtid, taskdata);
1124       // Predecrement simulated by "- 1" calculation
1125 #if KMP_DEBUG
1126       children = -1 +
1127 #endif
1128           KMP_ATOMIC_DEC(&taskdata->td_parent->td_incomplete_child_tasks);
1129       KMP_DEBUG_ASSERT(children >= 0);
1130       if (taskdata->td_taskgroup)
1131         KMP_ATOMIC_DEC(&taskdata->td_taskgroup->count);
1132     } else if (task_team && (task_team->tt.tt_found_proxy_tasks ||
1133                              task_team->tt.tt_hidden_helper_task_encountered)) {
1134       // if we found proxy or hidden helper tasks there could exist a dependency
1135       // chain with the proxy task as origin
1136       __kmp_release_deps(gtid, taskdata);
1137     }
1138     // td_flags.executing must be marked as 0 after __kmp_release_deps has been
1139     // called. Othertwise, if a task is executed immediately from the
1140     // release_deps code, the flag will be reset to 1 again by this same
1141     // function
1142     KMP_DEBUG_ASSERT(taskdata->td_flags.executing == 1);
1143     taskdata->td_flags.executing = 0; // suspend the finishing task
1144 
1145     // Decrement the counter of hidden helper tasks to be executed.
1146     if (taskdata->td_flags.hidden_helper) {
1147       // Hidden helper tasks can only be executed by hidden helper threads.
1148       KMP_ASSERT(KMP_HIDDEN_HELPER_THREAD(gtid));
1149       KMP_ATOMIC_DEC(&__kmp_unexecuted_hidden_helper_tasks);
1150     }
1151   }
1152 
1153   KA_TRACE(
1154       20, ("__kmp_task_finish: T#%d finished task %p, %d incomplete children\n",
1155            gtid, taskdata, children));
1156 
1157   // Free this task and then ancestor tasks if they have no children.
1158   // Restore th_current_task first as suggested by John:
1159   // johnmc: if an asynchronous inquiry peers into the runtime system
1160   // it doesn't see the freed task as the current task.
1161   thread->th.th_current_task = resumed_task;
1162   if (completed)
1163     __kmp_free_task_and_ancestors(gtid, taskdata, thread);
1164 
1165   // TODO: GEH - make sure root team implicit task is initialized properly.
1166   // KMP_DEBUG_ASSERT( resumed_task->td_flags.executing == 0 );
1167   resumed_task->td_flags.executing = 1; // resume previous task
1168 
1169   KA_TRACE(
1170       10, ("__kmp_task_finish(exit): T#%d finished task %p, resuming task %p\n",
1171            gtid, taskdata, resumed_task));
1172 
1173   return;
1174 }
1175 
1176 template <bool ompt>
1177 static void __kmpc_omp_task_complete_if0_template(ident_t *loc_ref,
1178                                                   kmp_int32 gtid,
1179                                                   kmp_task_t *task) {
1180   KA_TRACE(10, ("__kmpc_omp_task_complete_if0(enter): T#%d loc=%p task=%p\n",
1181                 gtid, loc_ref, KMP_TASK_TO_TASKDATA(task)));
1182   KMP_DEBUG_ASSERT(gtid >= 0);
1183   // this routine will provide task to resume
1184   __kmp_task_finish<ompt>(gtid, task, NULL);
1185 
1186   KA_TRACE(10, ("__kmpc_omp_task_complete_if0(exit): T#%d loc=%p task=%p\n",
1187                 gtid, loc_ref, KMP_TASK_TO_TASKDATA(task)));
1188 
1189 #if OMPT_SUPPORT
1190   if (ompt) {
1191     ompt_frame_t *ompt_frame;
1192     __ompt_get_task_info_internal(0, NULL, NULL, &ompt_frame, NULL, NULL);
1193     ompt_frame->enter_frame = ompt_data_none;
1194     ompt_frame->enter_frame_flags =
1195         ompt_frame_runtime | ompt_frame_framepointer;
1196   }
1197 #endif
1198 
1199   return;
1200 }
1201 
1202 #if OMPT_SUPPORT
1203 OMPT_NOINLINE
1204 void __kmpc_omp_task_complete_if0_ompt(ident_t *loc_ref, kmp_int32 gtid,
1205                                        kmp_task_t *task) {
1206   __kmpc_omp_task_complete_if0_template<true>(loc_ref, gtid, task);
1207 }
1208 #endif // OMPT_SUPPORT
1209 
1210 // __kmpc_omp_task_complete_if0: report that a task has completed execution
1211 //
1212 // loc_ref: source location information; points to end of task block.
1213 // gtid: global thread number.
1214 // task: task thunk for the completed task.
1215 void __kmpc_omp_task_complete_if0(ident_t *loc_ref, kmp_int32 gtid,
1216                                   kmp_task_t *task) {
1217 #if OMPT_SUPPORT
1218   if (UNLIKELY(ompt_enabled.enabled)) {
1219     __kmpc_omp_task_complete_if0_ompt(loc_ref, gtid, task);
1220     return;
1221   }
1222 #endif
1223   __kmpc_omp_task_complete_if0_template<false>(loc_ref, gtid, task);
1224 }
1225 
1226 #ifdef TASK_UNUSED
1227 // __kmpc_omp_task_complete: report that a task has completed execution
1228 // NEVER GENERATED BY COMPILER, DEPRECATED!!!
1229 void __kmpc_omp_task_complete(ident_t *loc_ref, kmp_int32 gtid,
1230                               kmp_task_t *task) {
1231   KA_TRACE(10, ("__kmpc_omp_task_complete(enter): T#%d loc=%p task=%p\n", gtid,
1232                 loc_ref, KMP_TASK_TO_TASKDATA(task)));
1233 
1234   __kmp_task_finish<false>(gtid, task,
1235                            NULL); // Not sure how to find task to resume
1236 
1237   KA_TRACE(10, ("__kmpc_omp_task_complete(exit): T#%d loc=%p task=%p\n", gtid,
1238                 loc_ref, KMP_TASK_TO_TASKDATA(task)));
1239   return;
1240 }
1241 #endif // TASK_UNUSED
1242 
1243 // __kmp_init_implicit_task: Initialize the appropriate fields in the implicit
1244 // task for a given thread
1245 //
1246 // loc_ref:  reference to source location of parallel region
1247 // this_thr:  thread data structure corresponding to implicit task
1248 // team: team for this_thr
1249 // tid: thread id of given thread within team
1250 // set_curr_task: TRUE if need to push current task to thread
1251 // NOTE: Routine does not set up the implicit task ICVS.  This is assumed to
1252 // have already been done elsewhere.
1253 // TODO: Get better loc_ref.  Value passed in may be NULL
1254 void __kmp_init_implicit_task(ident_t *loc_ref, kmp_info_t *this_thr,
1255                               kmp_team_t *team, int tid, int set_curr_task) {
1256   kmp_taskdata_t *task = &team->t.t_implicit_task_taskdata[tid];
1257 
1258   KF_TRACE(
1259       10,
1260       ("__kmp_init_implicit_task(enter): T#:%d team=%p task=%p, reinit=%s\n",
1261        tid, team, task, set_curr_task ? "TRUE" : "FALSE"));
1262 
1263   task->td_task_id = KMP_GEN_TASK_ID();
1264   task->td_team = team;
1265   //    task->td_parent   = NULL;  // fix for CQ230101 (broken parent task info
1266   //    in debugger)
1267   task->td_ident = loc_ref;
1268   task->td_taskwait_ident = NULL;
1269   task->td_taskwait_counter = 0;
1270   task->td_taskwait_thread = 0;
1271 
1272   task->td_flags.tiedness = TASK_TIED;
1273   task->td_flags.tasktype = TASK_IMPLICIT;
1274   task->td_flags.proxy = TASK_FULL;
1275 
1276   // All implicit tasks are executed immediately, not deferred
1277   task->td_flags.task_serial = 1;
1278   task->td_flags.tasking_ser = (__kmp_tasking_mode == tskm_immediate_exec);
1279   task->td_flags.team_serial = (team->t.t_serialized) ? 1 : 0;
1280 
1281   task->td_flags.started = 1;
1282   task->td_flags.executing = 1;
1283   task->td_flags.complete = 0;
1284   task->td_flags.freed = 0;
1285 
1286   task->td_depnode = NULL;
1287   task->td_last_tied = task;
1288   task->td_allow_completion_event.type = KMP_EVENT_UNINITIALIZED;
1289 
1290   if (set_curr_task) { // only do this init first time thread is created
1291     KMP_ATOMIC_ST_REL(&task->td_incomplete_child_tasks, 0);
1292     // Not used: don't need to deallocate implicit task
1293     KMP_ATOMIC_ST_REL(&task->td_allocated_child_tasks, 0);
1294     task->td_taskgroup = NULL; // An implicit task does not have taskgroup
1295     task->td_dephash = NULL;
1296     __kmp_push_current_task_to_thread(this_thr, team, tid);
1297   } else {
1298     KMP_DEBUG_ASSERT(task->td_incomplete_child_tasks == 0);
1299     KMP_DEBUG_ASSERT(task->td_allocated_child_tasks == 0);
1300   }
1301 
1302 #if OMPT_SUPPORT
1303   if (UNLIKELY(ompt_enabled.enabled))
1304     __ompt_task_init(task, tid);
1305 #endif
1306 
1307   KF_TRACE(10, ("__kmp_init_implicit_task(exit): T#:%d team=%p task=%p\n", tid,
1308                 team, task));
1309 }
1310 
1311 // __kmp_finish_implicit_task: Release resources associated to implicit tasks
1312 // at the end of parallel regions. Some resources are kept for reuse in the next
1313 // parallel region.
1314 //
1315 // thread:  thread data structure corresponding to implicit task
1316 void __kmp_finish_implicit_task(kmp_info_t *thread) {
1317   kmp_taskdata_t *task = thread->th.th_current_task;
1318   if (task->td_dephash) {
1319     int children;
1320     task->td_flags.complete = 1;
1321     children = KMP_ATOMIC_LD_ACQ(&task->td_incomplete_child_tasks);
1322     kmp_tasking_flags_t flags_old = task->td_flags;
1323     if (children == 0 && flags_old.complete == 1) {
1324       kmp_tasking_flags_t flags_new = flags_old;
1325       flags_new.complete = 0;
1326       if (KMP_COMPARE_AND_STORE_ACQ32(RCAST(kmp_int32 *, &task->td_flags),
1327                                       *RCAST(kmp_int32 *, &flags_old),
1328                                       *RCAST(kmp_int32 *, &flags_new))) {
1329         KA_TRACE(100, ("__kmp_finish_implicit_task: T#%d cleans "
1330                        "dephash of implicit task %p\n",
1331                        thread->th.th_info.ds.ds_gtid, task));
1332         __kmp_dephash_free_entries(thread, task->td_dephash);
1333       }
1334     }
1335   }
1336 }
1337 
1338 // __kmp_free_implicit_task: Release resources associated to implicit tasks
1339 // when these are destroyed regions
1340 //
1341 // thread:  thread data structure corresponding to implicit task
1342 void __kmp_free_implicit_task(kmp_info_t *thread) {
1343   kmp_taskdata_t *task = thread->th.th_current_task;
1344   if (task && task->td_dephash) {
1345     __kmp_dephash_free(thread, task->td_dephash);
1346     task->td_dephash = NULL;
1347   }
1348 }
1349 
1350 // Round up a size to a power of two specified by val: Used to insert padding
1351 // between structures co-allocated using a single malloc() call
1352 static size_t __kmp_round_up_to_val(size_t size, size_t val) {
1353   if (size & (val - 1)) {
1354     size &= ~(val - 1);
1355     if (size <= KMP_SIZE_T_MAX - val) {
1356       size += val; // Round up if there is no overflow.
1357     }
1358   }
1359   return size;
1360 } // __kmp_round_up_to_va
1361 
1362 // __kmp_task_alloc: Allocate the taskdata and task data structures for a task
1363 //
1364 // loc_ref: source location information
1365 // gtid: global thread number.
1366 // flags: include tiedness & task type (explicit vs. implicit) of the ''new''
1367 // task encountered. Converted from kmp_int32 to kmp_tasking_flags_t in routine.
1368 // sizeof_kmp_task_t:  Size in bytes of kmp_task_t data structure including
1369 // private vars accessed in task.
1370 // sizeof_shareds:  Size in bytes of array of pointers to shared vars accessed
1371 // in task.
1372 // task_entry: Pointer to task code entry point generated by compiler.
1373 // returns: a pointer to the allocated kmp_task_t structure (task).
1374 kmp_task_t *__kmp_task_alloc(ident_t *loc_ref, kmp_int32 gtid,
1375                              kmp_tasking_flags_t *flags,
1376                              size_t sizeof_kmp_task_t, size_t sizeof_shareds,
1377                              kmp_routine_entry_t task_entry) {
1378   kmp_task_t *task;
1379   kmp_taskdata_t *taskdata;
1380   kmp_info_t *thread = __kmp_threads[gtid];
1381   kmp_team_t *team = thread->th.th_team;
1382   kmp_taskdata_t *parent_task = thread->th.th_current_task;
1383   size_t shareds_offset;
1384 
1385   if (UNLIKELY(!TCR_4(__kmp_init_middle)))
1386     __kmp_middle_initialize();
1387 
1388   if (flags->hidden_helper) {
1389     if (__kmp_enable_hidden_helper) {
1390       if (!TCR_4(__kmp_init_hidden_helper))
1391         __kmp_hidden_helper_initialize();
1392     } else {
1393       // If the hidden helper task is not enabled, reset the flag to FALSE.
1394       flags->hidden_helper = FALSE;
1395     }
1396   }
1397 
1398   KA_TRACE(10, ("__kmp_task_alloc(enter): T#%d loc=%p, flags=(0x%x) "
1399                 "sizeof_task=%ld sizeof_shared=%ld entry=%p\n",
1400                 gtid, loc_ref, *((kmp_int32 *)flags), sizeof_kmp_task_t,
1401                 sizeof_shareds, task_entry));
1402 
1403   KMP_DEBUG_ASSERT(parent_task);
1404   if (parent_task->td_flags.final) {
1405     if (flags->merged_if0) {
1406     }
1407     flags->final = 1;
1408   }
1409 
1410   if (flags->tiedness == TASK_UNTIED && !team->t.t_serialized) {
1411     // Untied task encountered causes the TSC algorithm to check entire deque of
1412     // the victim thread. If no untied task encountered, then checking the head
1413     // of the deque should be enough.
1414     KMP_CHECK_UPDATE(thread->th.th_task_team->tt.tt_untied_task_encountered, 1);
1415   }
1416 
1417   // Detachable tasks are not proxy tasks yet but could be in the future. Doing
1418   // the tasking setup
1419   // when that happens is too late.
1420   if (UNLIKELY(flags->proxy == TASK_PROXY ||
1421                flags->detachable == TASK_DETACHABLE || flags->hidden_helper)) {
1422     if (flags->proxy == TASK_PROXY) {
1423       flags->tiedness = TASK_UNTIED;
1424       flags->merged_if0 = 1;
1425     }
1426     /* are we running in a sequential parallel or tskm_immediate_exec... we need
1427        tasking support enabled */
1428     if ((thread->th.th_task_team) == NULL) {
1429       /* This should only happen if the team is serialized
1430           setup a task team and propagate it to the thread */
1431       KMP_DEBUG_ASSERT(team->t.t_serialized);
1432       KA_TRACE(30,
1433                ("T#%d creating task team in __kmp_task_alloc for proxy task\n",
1434                 gtid));
1435       // 1 indicates setup the current team regardless of nthreads
1436       __kmp_task_team_setup(thread, team, 1);
1437       thread->th.th_task_team = team->t.t_task_team[thread->th.th_task_state];
1438     }
1439     kmp_task_team_t *task_team = thread->th.th_task_team;
1440 
1441     /* tasking must be enabled now as the task might not be pushed */
1442     if (!KMP_TASKING_ENABLED(task_team)) {
1443       KA_TRACE(
1444           30,
1445           ("T#%d enabling tasking in __kmp_task_alloc for proxy task\n", gtid));
1446       __kmp_enable_tasking(task_team, thread);
1447       kmp_int32 tid = thread->th.th_info.ds.ds_tid;
1448       kmp_thread_data_t *thread_data = &task_team->tt.tt_threads_data[tid];
1449       // No lock needed since only owner can allocate
1450       if (thread_data->td.td_deque == NULL) {
1451         __kmp_alloc_task_deque(thread, thread_data);
1452       }
1453     }
1454 
1455     if ((flags->proxy == TASK_PROXY || flags->detachable == TASK_DETACHABLE) &&
1456         task_team->tt.tt_found_proxy_tasks == FALSE)
1457       TCW_4(task_team->tt.tt_found_proxy_tasks, TRUE);
1458     if (flags->hidden_helper &&
1459         task_team->tt.tt_hidden_helper_task_encountered == FALSE)
1460       TCW_4(task_team->tt.tt_hidden_helper_task_encountered, TRUE);
1461   }
1462 
1463   // Calculate shared structure offset including padding after kmp_task_t struct
1464   // to align pointers in shared struct
1465   shareds_offset = sizeof(kmp_taskdata_t) + sizeof_kmp_task_t;
1466   shareds_offset = __kmp_round_up_to_val(shareds_offset, sizeof(void *));
1467 
1468   // Allocate a kmp_taskdata_t block and a kmp_task_t block.
1469   KA_TRACE(30, ("__kmp_task_alloc: T#%d First malloc size: %ld\n", gtid,
1470                 shareds_offset));
1471   KA_TRACE(30, ("__kmp_task_alloc: T#%d Second malloc size: %ld\n", gtid,
1472                 sizeof_shareds));
1473 
1474   // Avoid double allocation here by combining shareds with taskdata
1475 #if USE_FAST_MEMORY
1476   taskdata = (kmp_taskdata_t *)__kmp_fast_allocate(thread, shareds_offset +
1477                                                                sizeof_shareds);
1478 #else /* ! USE_FAST_MEMORY */
1479   taskdata = (kmp_taskdata_t *)__kmp_thread_malloc(thread, shareds_offset +
1480                                                                sizeof_shareds);
1481 #endif /* USE_FAST_MEMORY */
1482 
1483   task = KMP_TASKDATA_TO_TASK(taskdata);
1484 
1485 // Make sure task & taskdata are aligned appropriately
1486 #if KMP_ARCH_X86 || KMP_ARCH_PPC64 || !KMP_HAVE_QUAD
1487   KMP_DEBUG_ASSERT((((kmp_uintptr_t)taskdata) & (sizeof(double) - 1)) == 0);
1488   KMP_DEBUG_ASSERT((((kmp_uintptr_t)task) & (sizeof(double) - 1)) == 0);
1489 #else
1490   KMP_DEBUG_ASSERT((((kmp_uintptr_t)taskdata) & (sizeof(_Quad) - 1)) == 0);
1491   KMP_DEBUG_ASSERT((((kmp_uintptr_t)task) & (sizeof(_Quad) - 1)) == 0);
1492 #endif
1493   if (sizeof_shareds > 0) {
1494     // Avoid double allocation here by combining shareds with taskdata
1495     task->shareds = &((char *)taskdata)[shareds_offset];
1496     // Make sure shareds struct is aligned to pointer size
1497     KMP_DEBUG_ASSERT((((kmp_uintptr_t)task->shareds) & (sizeof(void *) - 1)) ==
1498                      0);
1499   } else {
1500     task->shareds = NULL;
1501   }
1502   task->routine = task_entry;
1503   task->part_id = 0; // AC: Always start with 0 part id
1504 
1505   taskdata->td_task_id = KMP_GEN_TASK_ID();
1506   taskdata->td_team = thread->th.th_team;
1507   taskdata->td_alloc_thread = thread;
1508   taskdata->td_parent = parent_task;
1509   taskdata->td_level = parent_task->td_level + 1; // increment nesting level
1510   KMP_ATOMIC_ST_RLX(&taskdata->td_untied_count, 0);
1511   taskdata->td_ident = loc_ref;
1512   taskdata->td_taskwait_ident = NULL;
1513   taskdata->td_taskwait_counter = 0;
1514   taskdata->td_taskwait_thread = 0;
1515   KMP_DEBUG_ASSERT(taskdata->td_parent != NULL);
1516   // avoid copying icvs for proxy tasks
1517   if (flags->proxy == TASK_FULL)
1518     copy_icvs(&taskdata->td_icvs, &taskdata->td_parent->td_icvs);
1519 
1520   taskdata->td_flags = *flags;
1521   taskdata->td_task_team = thread->th.th_task_team;
1522   taskdata->td_size_alloc = shareds_offset + sizeof_shareds;
1523   taskdata->td_flags.tasktype = TASK_EXPLICIT;
1524   // If it is hidden helper task, we need to set the team and task team
1525   // correspondingly.
1526   if (flags->hidden_helper) {
1527     kmp_info_t *shadow_thread = __kmp_threads[KMP_GTID_TO_SHADOW_GTID(gtid)];
1528     taskdata->td_team = shadow_thread->th.th_team;
1529     taskdata->td_task_team = shadow_thread->th.th_task_team;
1530   }
1531 
1532   // GEH - TODO: fix this to copy parent task's value of tasking_ser flag
1533   taskdata->td_flags.tasking_ser = (__kmp_tasking_mode == tskm_immediate_exec);
1534 
1535   // GEH - TODO: fix this to copy parent task's value of team_serial flag
1536   taskdata->td_flags.team_serial = (team->t.t_serialized) ? 1 : 0;
1537 
1538   // GEH - Note we serialize the task if the team is serialized to make sure
1539   // implicit parallel region tasks are not left until program termination to
1540   // execute. Also, it helps locality to execute immediately.
1541 
1542   taskdata->td_flags.task_serial =
1543       (parent_task->td_flags.final || taskdata->td_flags.team_serial ||
1544        taskdata->td_flags.tasking_ser || flags->merged_if0);
1545 
1546   taskdata->td_flags.started = 0;
1547   taskdata->td_flags.executing = 0;
1548   taskdata->td_flags.complete = 0;
1549   taskdata->td_flags.freed = 0;
1550 
1551   KMP_ATOMIC_ST_RLX(&taskdata->td_incomplete_child_tasks, 0);
1552   // start at one because counts current task and children
1553   KMP_ATOMIC_ST_RLX(&taskdata->td_allocated_child_tasks, 1);
1554   taskdata->td_taskgroup =
1555       parent_task->td_taskgroup; // task inherits taskgroup from the parent task
1556   taskdata->td_dephash = NULL;
1557   taskdata->td_depnode = NULL;
1558   taskdata->td_target_data.async_handle = NULL;
1559   if (flags->tiedness == TASK_UNTIED)
1560     taskdata->td_last_tied = NULL; // will be set when the task is scheduled
1561   else
1562     taskdata->td_last_tied = taskdata;
1563   taskdata->td_allow_completion_event.type = KMP_EVENT_UNINITIALIZED;
1564 #if OMPT_SUPPORT
1565   if (UNLIKELY(ompt_enabled.enabled))
1566     __ompt_task_init(taskdata, gtid);
1567 #endif
1568   // TODO: What would be the balance between the conditions in the function and
1569   // an atomic operation?
1570   if (__kmp_track_children_task(taskdata)) {
1571     KMP_ATOMIC_INC(&parent_task->td_incomplete_child_tasks);
1572     if (parent_task->td_taskgroup)
1573       KMP_ATOMIC_INC(&parent_task->td_taskgroup->count);
1574     // Only need to keep track of allocated child tasks for explicit tasks since
1575     // implicit not deallocated
1576     if (taskdata->td_parent->td_flags.tasktype == TASK_EXPLICIT) {
1577       KMP_ATOMIC_INC(&taskdata->td_parent->td_allocated_child_tasks);
1578     }
1579     if (flags->hidden_helper) {
1580       taskdata->td_flags.task_serial = FALSE;
1581       // Increment the number of hidden helper tasks to be executed
1582       KMP_ATOMIC_INC(&__kmp_unexecuted_hidden_helper_tasks);
1583     }
1584   }
1585 
1586   KA_TRACE(20, ("__kmp_task_alloc(exit): T#%d created task %p parent=%p\n",
1587                 gtid, taskdata, taskdata->td_parent));
1588 
1589   return task;
1590 }
1591 
1592 kmp_task_t *__kmpc_omp_task_alloc(ident_t *loc_ref, kmp_int32 gtid,
1593                                   kmp_int32 flags, size_t sizeof_kmp_task_t,
1594                                   size_t sizeof_shareds,
1595                                   kmp_routine_entry_t task_entry) {
1596   kmp_task_t *retval;
1597   kmp_tasking_flags_t *input_flags = (kmp_tasking_flags_t *)&flags;
1598   __kmp_assert_valid_gtid(gtid);
1599   input_flags->native = FALSE;
1600   // __kmp_task_alloc() sets up all other runtime flags
1601   KA_TRACE(10, ("__kmpc_omp_task_alloc(enter): T#%d loc=%p, flags=(%s %s %s) "
1602                 "sizeof_task=%ld sizeof_shared=%ld entry=%p\n",
1603                 gtid, loc_ref, input_flags->tiedness ? "tied  " : "untied",
1604                 input_flags->proxy ? "proxy" : "",
1605                 input_flags->detachable ? "detachable" : "", sizeof_kmp_task_t,
1606                 sizeof_shareds, task_entry));
1607 
1608   retval = __kmp_task_alloc(loc_ref, gtid, input_flags, sizeof_kmp_task_t,
1609                             sizeof_shareds, task_entry);
1610 
1611   KA_TRACE(20, ("__kmpc_omp_task_alloc(exit): T#%d retval %p\n", gtid, retval));
1612 
1613   return retval;
1614 }
1615 
1616 kmp_task_t *__kmpc_omp_target_task_alloc(ident_t *loc_ref, kmp_int32 gtid,
1617                                          kmp_int32 flags,
1618                                          size_t sizeof_kmp_task_t,
1619                                          size_t sizeof_shareds,
1620                                          kmp_routine_entry_t task_entry,
1621                                          kmp_int64 device_id) {
1622   auto &input_flags = reinterpret_cast<kmp_tasking_flags_t &>(flags);
1623   // target task is untied defined in the specification
1624   input_flags.tiedness = TASK_UNTIED;
1625 
1626   if (__kmp_enable_hidden_helper)
1627     input_flags.hidden_helper = TRUE;
1628 
1629   return __kmpc_omp_task_alloc(loc_ref, gtid, flags, sizeof_kmp_task_t,
1630                                sizeof_shareds, task_entry);
1631 }
1632 
1633 /*!
1634 @ingroup TASKING
1635 @param loc_ref location of the original task directive
1636 @param gtid Global Thread ID of encountering thread
1637 @param new_task task thunk allocated by __kmpc_omp_task_alloc() for the ''new
1638 task''
1639 @param naffins Number of affinity items
1640 @param affin_list List of affinity items
1641 @return Returns non-zero if registering affinity information was not successful.
1642  Returns 0 if registration was successful
1643 This entry registers the affinity information attached to a task with the task
1644 thunk structure kmp_taskdata_t.
1645 */
1646 kmp_int32
1647 __kmpc_omp_reg_task_with_affinity(ident_t *loc_ref, kmp_int32 gtid,
1648                                   kmp_task_t *new_task, kmp_int32 naffins,
1649                                   kmp_task_affinity_info_t *affin_list) {
1650   return 0;
1651 }
1652 
1653 //  __kmp_invoke_task: invoke the specified task
1654 //
1655 // gtid: global thread ID of caller
1656 // task: the task to invoke
1657 // current_task: the task to resume after task invocation
1658 static void __kmp_invoke_task(kmp_int32 gtid, kmp_task_t *task,
1659                               kmp_taskdata_t *current_task) {
1660   kmp_taskdata_t *taskdata = KMP_TASK_TO_TASKDATA(task);
1661   kmp_info_t *thread;
1662   int discard = 0 /* false */;
1663   KA_TRACE(
1664       30, ("__kmp_invoke_task(enter): T#%d invoking task %p, current_task=%p\n",
1665            gtid, taskdata, current_task));
1666   KMP_DEBUG_ASSERT(task);
1667   if (UNLIKELY(taskdata->td_flags.proxy == TASK_PROXY &&
1668                taskdata->td_flags.complete == 1)) {
1669     // This is a proxy task that was already completed but it needs to run
1670     // its bottom-half finish
1671     KA_TRACE(
1672         30,
1673         ("__kmp_invoke_task: T#%d running bottom finish for proxy task %p\n",
1674          gtid, taskdata));
1675 
1676     __kmp_bottom_half_finish_proxy(gtid, task);
1677 
1678     KA_TRACE(30, ("__kmp_invoke_task(exit): T#%d completed bottom finish for "
1679                   "proxy task %p, resuming task %p\n",
1680                   gtid, taskdata, current_task));
1681 
1682     return;
1683   }
1684 
1685 #if OMPT_SUPPORT
1686   // For untied tasks, the first task executed only calls __kmpc_omp_task and
1687   // does not execute code.
1688   ompt_thread_info_t oldInfo;
1689   if (UNLIKELY(ompt_enabled.enabled)) {
1690     // Store the threads states and restore them after the task
1691     thread = __kmp_threads[gtid];
1692     oldInfo = thread->th.ompt_thread_info;
1693     thread->th.ompt_thread_info.wait_id = 0;
1694     thread->th.ompt_thread_info.state = (thread->th.th_team_serialized)
1695                                             ? ompt_state_work_serial
1696                                             : ompt_state_work_parallel;
1697     taskdata->ompt_task_info.frame.exit_frame.ptr = OMPT_GET_FRAME_ADDRESS(0);
1698   }
1699 #endif
1700 
1701   // Proxy tasks are not handled by the runtime
1702   if (taskdata->td_flags.proxy != TASK_PROXY) {
1703     __kmp_task_start(gtid, task, current_task); // OMPT only if not discarded
1704   }
1705 
1706   // TODO: cancel tasks if the parallel region has also been cancelled
1707   // TODO: check if this sequence can be hoisted above __kmp_task_start
1708   // if cancellation has been enabled for this run ...
1709   if (UNLIKELY(__kmp_omp_cancellation)) {
1710     thread = __kmp_threads[gtid];
1711     kmp_team_t *this_team = thread->th.th_team;
1712     kmp_taskgroup_t *taskgroup = taskdata->td_taskgroup;
1713     if ((taskgroup && taskgroup->cancel_request) ||
1714         (this_team->t.t_cancel_request == cancel_parallel)) {
1715 #if OMPT_SUPPORT && OMPT_OPTIONAL
1716       ompt_data_t *task_data;
1717       if (UNLIKELY(ompt_enabled.ompt_callback_cancel)) {
1718         __ompt_get_task_info_internal(0, NULL, &task_data, NULL, NULL, NULL);
1719         ompt_callbacks.ompt_callback(ompt_callback_cancel)(
1720             task_data,
1721             ((taskgroup && taskgroup->cancel_request) ? ompt_cancel_taskgroup
1722                                                       : ompt_cancel_parallel) |
1723                 ompt_cancel_discarded_task,
1724             NULL);
1725       }
1726 #endif
1727       KMP_COUNT_BLOCK(TASK_cancelled);
1728       // this task belongs to a task group and we need to cancel it
1729       discard = 1 /* true */;
1730     }
1731   }
1732 
1733   // Invoke the task routine and pass in relevant data.
1734   // Thunks generated by gcc take a different argument list.
1735   if (!discard) {
1736     if (taskdata->td_flags.tiedness == TASK_UNTIED) {
1737       taskdata->td_last_tied = current_task->td_last_tied;
1738       KMP_DEBUG_ASSERT(taskdata->td_last_tied);
1739     }
1740 #if KMP_STATS_ENABLED
1741     KMP_COUNT_BLOCK(TASK_executed);
1742     switch (KMP_GET_THREAD_STATE()) {
1743     case FORK_JOIN_BARRIER:
1744       KMP_PUSH_PARTITIONED_TIMER(OMP_task_join_bar);
1745       break;
1746     case PLAIN_BARRIER:
1747       KMP_PUSH_PARTITIONED_TIMER(OMP_task_plain_bar);
1748       break;
1749     case TASKYIELD:
1750       KMP_PUSH_PARTITIONED_TIMER(OMP_task_taskyield);
1751       break;
1752     case TASKWAIT:
1753       KMP_PUSH_PARTITIONED_TIMER(OMP_task_taskwait);
1754       break;
1755     case TASKGROUP:
1756       KMP_PUSH_PARTITIONED_TIMER(OMP_task_taskgroup);
1757       break;
1758     default:
1759       KMP_PUSH_PARTITIONED_TIMER(OMP_task_immediate);
1760       break;
1761     }
1762 #endif // KMP_STATS_ENABLED
1763 
1764 // OMPT task begin
1765 #if OMPT_SUPPORT
1766     if (UNLIKELY(ompt_enabled.enabled))
1767       __ompt_task_start(task, current_task, gtid);
1768 #endif
1769 #if OMPT_SUPPORT && OMPT_OPTIONAL
1770     if (UNLIKELY(ompt_enabled.ompt_callback_dispatch &&
1771                  taskdata->ompt_task_info.dispatch_chunk.iterations > 0)) {
1772       ompt_data_t instance = ompt_data_none;
1773       instance.ptr = &(taskdata->ompt_task_info.dispatch_chunk);
1774       ompt_team_info_t *team_info = __ompt_get_teaminfo(0, NULL);
1775       ompt_callbacks.ompt_callback(ompt_callback_dispatch)(
1776           &(team_info->parallel_data), &(taskdata->ompt_task_info.task_data),
1777           ompt_dispatch_taskloop_chunk, instance);
1778       taskdata->ompt_task_info.dispatch_chunk = {0, 0};
1779     }
1780 #endif // OMPT_SUPPORT && OMPT_OPTIONAL
1781 
1782 #if OMPD_SUPPORT
1783     if (ompd_state & OMPD_ENABLE_BP)
1784       ompd_bp_task_begin();
1785 #endif
1786 
1787 #if USE_ITT_BUILD && USE_ITT_NOTIFY
1788     kmp_uint64 cur_time;
1789     kmp_int32 kmp_itt_count_task =
1790         __kmp_forkjoin_frames_mode == 3 && !taskdata->td_flags.task_serial &&
1791         current_task->td_flags.tasktype == TASK_IMPLICIT;
1792     if (kmp_itt_count_task) {
1793       thread = __kmp_threads[gtid];
1794       // Time outer level explicit task on barrier for adjusting imbalance time
1795       if (thread->th.th_bar_arrive_time)
1796         cur_time = __itt_get_timestamp();
1797       else
1798         kmp_itt_count_task = 0; // thread is not on a barrier - skip timing
1799     }
1800     KMP_FSYNC_ACQUIRED(taskdata); // acquired self (new task)
1801 #endif
1802 
1803 #if ENABLE_LIBOMPTARGET
1804     if (taskdata->td_target_data.async_handle != NULL) {
1805       // If we have a valid target async handle, that means that we have already
1806       // executed the task routine once. We must query for the handle completion
1807       // instead of re-executing the routine.
1808       __tgt_target_nowait_query(&taskdata->td_target_data.async_handle);
1809     } else
1810 #endif
1811     if (task->routine != NULL) {
1812 #ifdef KMP_GOMP_COMPAT
1813       if (taskdata->td_flags.native) {
1814         ((void (*)(void *))(*(task->routine)))(task->shareds);
1815       } else
1816 #endif /* KMP_GOMP_COMPAT */
1817       {
1818         (*(task->routine))(gtid, task);
1819       }
1820     }
1821     KMP_POP_PARTITIONED_TIMER();
1822 
1823 #if USE_ITT_BUILD && USE_ITT_NOTIFY
1824     if (kmp_itt_count_task) {
1825       // Barrier imbalance - adjust arrive time with the task duration
1826       thread->th.th_bar_arrive_time += (__itt_get_timestamp() - cur_time);
1827     }
1828     KMP_FSYNC_CANCEL(taskdata); // destroy self (just executed)
1829     KMP_FSYNC_RELEASING(taskdata->td_parent); // releasing parent
1830 #endif
1831   }
1832 
1833 #if OMPD_SUPPORT
1834   if (ompd_state & OMPD_ENABLE_BP)
1835     ompd_bp_task_end();
1836 #endif
1837 
1838   // Proxy tasks are not handled by the runtime
1839   if (taskdata->td_flags.proxy != TASK_PROXY) {
1840 #if OMPT_SUPPORT
1841     if (UNLIKELY(ompt_enabled.enabled)) {
1842       thread->th.ompt_thread_info = oldInfo;
1843       if (taskdata->td_flags.tiedness == TASK_TIED) {
1844         taskdata->ompt_task_info.frame.exit_frame = ompt_data_none;
1845       }
1846       __kmp_task_finish<true>(gtid, task, current_task);
1847     } else
1848 #endif
1849       __kmp_task_finish<false>(gtid, task, current_task);
1850   }
1851 
1852   KA_TRACE(
1853       30,
1854       ("__kmp_invoke_task(exit): T#%d completed task %p, resuming task %p\n",
1855        gtid, taskdata, current_task));
1856   return;
1857 }
1858 
1859 // __kmpc_omp_task_parts: Schedule a thread-switchable task for execution
1860 //
1861 // loc_ref: location of original task pragma (ignored)
1862 // gtid: Global Thread ID of encountering thread
1863 // new_task: task thunk allocated by __kmp_omp_task_alloc() for the ''new task''
1864 // Returns:
1865 //    TASK_CURRENT_NOT_QUEUED (0) if did not suspend and queue current task to
1866 //    be resumed later.
1867 //    TASK_CURRENT_QUEUED (1) if suspended and queued the current task to be
1868 //    resumed later.
1869 kmp_int32 __kmpc_omp_task_parts(ident_t *loc_ref, kmp_int32 gtid,
1870                                 kmp_task_t *new_task) {
1871   kmp_taskdata_t *new_taskdata = KMP_TASK_TO_TASKDATA(new_task);
1872 
1873   KA_TRACE(10, ("__kmpc_omp_task_parts(enter): T#%d loc=%p task=%p\n", gtid,
1874                 loc_ref, new_taskdata));
1875 
1876 #if OMPT_SUPPORT
1877   kmp_taskdata_t *parent;
1878   if (UNLIKELY(ompt_enabled.enabled)) {
1879     parent = new_taskdata->td_parent;
1880     if (ompt_enabled.ompt_callback_task_create) {
1881       ompt_callbacks.ompt_callback(ompt_callback_task_create)(
1882           &(parent->ompt_task_info.task_data), &(parent->ompt_task_info.frame),
1883           &(new_taskdata->ompt_task_info.task_data), ompt_task_explicit, 0,
1884           OMPT_GET_RETURN_ADDRESS(0));
1885     }
1886   }
1887 #endif
1888 
1889   /* Should we execute the new task or queue it? For now, let's just always try
1890      to queue it.  If the queue fills up, then we'll execute it.  */
1891 
1892   if (__kmp_push_task(gtid, new_task) == TASK_NOT_PUSHED) // if cannot defer
1893   { // Execute this task immediately
1894     kmp_taskdata_t *current_task = __kmp_threads[gtid]->th.th_current_task;
1895     new_taskdata->td_flags.task_serial = 1;
1896     __kmp_invoke_task(gtid, new_task, current_task);
1897   }
1898 
1899   KA_TRACE(
1900       10,
1901       ("__kmpc_omp_task_parts(exit): T#%d returning TASK_CURRENT_NOT_QUEUED: "
1902        "loc=%p task=%p, return: TASK_CURRENT_NOT_QUEUED\n",
1903        gtid, loc_ref, new_taskdata));
1904 
1905 #if OMPT_SUPPORT
1906   if (UNLIKELY(ompt_enabled.enabled)) {
1907     parent->ompt_task_info.frame.enter_frame = ompt_data_none;
1908   }
1909 #endif
1910   return TASK_CURRENT_NOT_QUEUED;
1911 }
1912 
1913 // __kmp_omp_task: Schedule a non-thread-switchable task for execution
1914 //
1915 // gtid: Global Thread ID of encountering thread
1916 // new_task:non-thread-switchable task thunk allocated by __kmp_omp_task_alloc()
1917 // serialize_immediate: if TRUE then if the task is executed immediately its
1918 // execution will be serialized
1919 // Returns:
1920 //    TASK_CURRENT_NOT_QUEUED (0) if did not suspend and queue current task to
1921 //    be resumed later.
1922 //    TASK_CURRENT_QUEUED (1) if suspended and queued the current task to be
1923 //    resumed later.
1924 kmp_int32 __kmp_omp_task(kmp_int32 gtid, kmp_task_t *new_task,
1925                          bool serialize_immediate) {
1926   kmp_taskdata_t *new_taskdata = KMP_TASK_TO_TASKDATA(new_task);
1927 
1928   /* Should we execute the new task or queue it? For now, let's just always try
1929      to queue it.  If the queue fills up, then we'll execute it.  */
1930   if (new_taskdata->td_flags.proxy == TASK_PROXY ||
1931       __kmp_push_task(gtid, new_task) == TASK_NOT_PUSHED) // if cannot defer
1932   { // Execute this task immediately
1933     kmp_taskdata_t *current_task = __kmp_threads[gtid]->th.th_current_task;
1934     if (serialize_immediate)
1935       new_taskdata->td_flags.task_serial = 1;
1936     __kmp_invoke_task(gtid, new_task, current_task);
1937   } else if (__kmp_dflt_blocktime != KMP_MAX_BLOCKTIME &&
1938              __kmp_wpolicy_passive) {
1939     kmp_info_t *this_thr = __kmp_threads[gtid];
1940     kmp_team_t *team = this_thr->th.th_team;
1941     kmp_int32 nthreads = this_thr->th.th_team_nproc;
1942     for (int i = 0; i < nthreads; ++i) {
1943       kmp_info_t *thread = team->t.t_threads[i];
1944       if (thread == this_thr)
1945         continue;
1946       if (thread->th.th_sleep_loc != NULL) {
1947         __kmp_null_resume_wrapper(thread);
1948         break; // awake one thread at a time
1949       }
1950     }
1951   }
1952   return TASK_CURRENT_NOT_QUEUED;
1953 }
1954 
1955 // __kmpc_omp_task: Wrapper around __kmp_omp_task to schedule a
1956 // non-thread-switchable task from the parent thread only!
1957 //
1958 // loc_ref: location of original task pragma (ignored)
1959 // gtid: Global Thread ID of encountering thread
1960 // new_task: non-thread-switchable task thunk allocated by
1961 // __kmp_omp_task_alloc()
1962 // Returns:
1963 //    TASK_CURRENT_NOT_QUEUED (0) if did not suspend and queue current task to
1964 //    be resumed later.
1965 //    TASK_CURRENT_QUEUED (1) if suspended and queued the current task to be
1966 //    resumed later.
1967 kmp_int32 __kmpc_omp_task(ident_t *loc_ref, kmp_int32 gtid,
1968                           kmp_task_t *new_task) {
1969   kmp_int32 res;
1970   KMP_SET_THREAD_STATE_BLOCK(EXPLICIT_TASK);
1971 
1972 #if KMP_DEBUG || OMPT_SUPPORT
1973   kmp_taskdata_t *new_taskdata = KMP_TASK_TO_TASKDATA(new_task);
1974 #endif
1975   KA_TRACE(10, ("__kmpc_omp_task(enter): T#%d loc=%p task=%p\n", gtid, loc_ref,
1976                 new_taskdata));
1977   __kmp_assert_valid_gtid(gtid);
1978 
1979 #if OMPT_SUPPORT
1980   kmp_taskdata_t *parent = NULL;
1981   if (UNLIKELY(ompt_enabled.enabled)) {
1982     if (!new_taskdata->td_flags.started) {
1983       OMPT_STORE_RETURN_ADDRESS(gtid);
1984       parent = new_taskdata->td_parent;
1985       if (!parent->ompt_task_info.frame.enter_frame.ptr) {
1986         parent->ompt_task_info.frame.enter_frame.ptr =
1987             OMPT_GET_FRAME_ADDRESS(0);
1988       }
1989       if (ompt_enabled.ompt_callback_task_create) {
1990         ompt_callbacks.ompt_callback(ompt_callback_task_create)(
1991             &(parent->ompt_task_info.task_data),
1992             &(parent->ompt_task_info.frame),
1993             &(new_taskdata->ompt_task_info.task_data),
1994             ompt_task_explicit | TASK_TYPE_DETAILS_FORMAT(new_taskdata), 0,
1995             OMPT_LOAD_RETURN_ADDRESS(gtid));
1996       }
1997     } else {
1998       // We are scheduling the continuation of an UNTIED task.
1999       // Scheduling back to the parent task.
2000       __ompt_task_finish(new_task,
2001                          new_taskdata->ompt_task_info.scheduling_parent,
2002                          ompt_task_switch);
2003       new_taskdata->ompt_task_info.frame.exit_frame = ompt_data_none;
2004     }
2005   }
2006 #endif
2007 
2008   res = __kmp_omp_task(gtid, new_task, true);
2009 
2010   KA_TRACE(10, ("__kmpc_omp_task(exit): T#%d returning "
2011                 "TASK_CURRENT_NOT_QUEUED: loc=%p task=%p\n",
2012                 gtid, loc_ref, new_taskdata));
2013 #if OMPT_SUPPORT
2014   if (UNLIKELY(ompt_enabled.enabled && parent != NULL)) {
2015     parent->ompt_task_info.frame.enter_frame = ompt_data_none;
2016   }
2017 #endif
2018   return res;
2019 }
2020 
2021 // __kmp_omp_taskloop_task: Wrapper around __kmp_omp_task to schedule
2022 // a taskloop task with the correct OMPT return address
2023 //
2024 // loc_ref: location of original task pragma (ignored)
2025 // gtid: Global Thread ID of encountering thread
2026 // new_task: non-thread-switchable task thunk allocated by
2027 // __kmp_omp_task_alloc()
2028 // codeptr_ra: return address for OMPT callback
2029 // Returns:
2030 //    TASK_CURRENT_NOT_QUEUED (0) if did not suspend and queue current task to
2031 //    be resumed later.
2032 //    TASK_CURRENT_QUEUED (1) if suspended and queued the current task to be
2033 //    resumed later.
2034 kmp_int32 __kmp_omp_taskloop_task(ident_t *loc_ref, kmp_int32 gtid,
2035                                   kmp_task_t *new_task, void *codeptr_ra) {
2036   kmp_int32 res;
2037   KMP_SET_THREAD_STATE_BLOCK(EXPLICIT_TASK);
2038 
2039 #if KMP_DEBUG || OMPT_SUPPORT
2040   kmp_taskdata_t *new_taskdata = KMP_TASK_TO_TASKDATA(new_task);
2041 #endif
2042   KA_TRACE(10, ("__kmpc_omp_task(enter): T#%d loc=%p task=%p\n", gtid, loc_ref,
2043                 new_taskdata));
2044 
2045 #if OMPT_SUPPORT
2046   kmp_taskdata_t *parent = NULL;
2047   if (UNLIKELY(ompt_enabled.enabled && !new_taskdata->td_flags.started)) {
2048     parent = new_taskdata->td_parent;
2049     if (!parent->ompt_task_info.frame.enter_frame.ptr)
2050       parent->ompt_task_info.frame.enter_frame.ptr = OMPT_GET_FRAME_ADDRESS(0);
2051     if (ompt_enabled.ompt_callback_task_create) {
2052       ompt_callbacks.ompt_callback(ompt_callback_task_create)(
2053           &(parent->ompt_task_info.task_data), &(parent->ompt_task_info.frame),
2054           &(new_taskdata->ompt_task_info.task_data),
2055           ompt_task_explicit | TASK_TYPE_DETAILS_FORMAT(new_taskdata), 0,
2056           codeptr_ra);
2057     }
2058   }
2059 #endif
2060 
2061   res = __kmp_omp_task(gtid, new_task, true);
2062 
2063   KA_TRACE(10, ("__kmpc_omp_task(exit): T#%d returning "
2064                 "TASK_CURRENT_NOT_QUEUED: loc=%p task=%p\n",
2065                 gtid, loc_ref, new_taskdata));
2066 #if OMPT_SUPPORT
2067   if (UNLIKELY(ompt_enabled.enabled && parent != NULL)) {
2068     parent->ompt_task_info.frame.enter_frame = ompt_data_none;
2069   }
2070 #endif
2071   return res;
2072 }
2073 
2074 template <bool ompt>
2075 static kmp_int32 __kmpc_omp_taskwait_template(ident_t *loc_ref, kmp_int32 gtid,
2076                                               void *frame_address,
2077                                               void *return_address) {
2078   kmp_taskdata_t *taskdata = nullptr;
2079   kmp_info_t *thread;
2080   int thread_finished = FALSE;
2081   KMP_SET_THREAD_STATE_BLOCK(TASKWAIT);
2082 
2083   KA_TRACE(10, ("__kmpc_omp_taskwait(enter): T#%d loc=%p\n", gtid, loc_ref));
2084   KMP_DEBUG_ASSERT(gtid >= 0);
2085 
2086   if (__kmp_tasking_mode != tskm_immediate_exec) {
2087     thread = __kmp_threads[gtid];
2088     taskdata = thread->th.th_current_task;
2089 
2090 #if OMPT_SUPPORT && OMPT_OPTIONAL
2091     ompt_data_t *my_task_data;
2092     ompt_data_t *my_parallel_data;
2093 
2094     if (ompt) {
2095       my_task_data = &(taskdata->ompt_task_info.task_data);
2096       my_parallel_data = OMPT_CUR_TEAM_DATA(thread);
2097 
2098       taskdata->ompt_task_info.frame.enter_frame.ptr = frame_address;
2099 
2100       if (ompt_enabled.ompt_callback_sync_region) {
2101         ompt_callbacks.ompt_callback(ompt_callback_sync_region)(
2102             ompt_sync_region_taskwait, ompt_scope_begin, my_parallel_data,
2103             my_task_data, return_address);
2104       }
2105 
2106       if (ompt_enabled.ompt_callback_sync_region_wait) {
2107         ompt_callbacks.ompt_callback(ompt_callback_sync_region_wait)(
2108             ompt_sync_region_taskwait, ompt_scope_begin, my_parallel_data,
2109             my_task_data, return_address);
2110       }
2111     }
2112 #endif // OMPT_SUPPORT && OMPT_OPTIONAL
2113 
2114 // Debugger: The taskwait is active. Store location and thread encountered the
2115 // taskwait.
2116 #if USE_ITT_BUILD
2117 // Note: These values are used by ITT events as well.
2118 #endif /* USE_ITT_BUILD */
2119     taskdata->td_taskwait_counter += 1;
2120     taskdata->td_taskwait_ident = loc_ref;
2121     taskdata->td_taskwait_thread = gtid + 1;
2122 
2123 #if USE_ITT_BUILD
2124     void *itt_sync_obj = NULL;
2125 #if USE_ITT_NOTIFY
2126     KMP_ITT_TASKWAIT_STARTING(itt_sync_obj);
2127 #endif /* USE_ITT_NOTIFY */
2128 #endif /* USE_ITT_BUILD */
2129 
2130     bool must_wait =
2131         !taskdata->td_flags.team_serial && !taskdata->td_flags.final;
2132 
2133     must_wait = must_wait || (thread->th.th_task_team != NULL &&
2134                               thread->th.th_task_team->tt.tt_found_proxy_tasks);
2135     // If hidden helper thread is encountered, we must enable wait here.
2136     must_wait =
2137         must_wait ||
2138         (__kmp_enable_hidden_helper && thread->th.th_task_team != NULL &&
2139          thread->th.th_task_team->tt.tt_hidden_helper_task_encountered);
2140 
2141     if (must_wait) {
2142       kmp_flag_32<false, false> flag(
2143           RCAST(std::atomic<kmp_uint32> *,
2144                 &(taskdata->td_incomplete_child_tasks)),
2145           0U);
2146       while (KMP_ATOMIC_LD_ACQ(&taskdata->td_incomplete_child_tasks) != 0) {
2147         flag.execute_tasks(thread, gtid, FALSE,
2148                            &thread_finished USE_ITT_BUILD_ARG(itt_sync_obj),
2149                            __kmp_task_stealing_constraint);
2150       }
2151     }
2152 #if USE_ITT_BUILD
2153     KMP_ITT_TASKWAIT_FINISHED(itt_sync_obj);
2154     KMP_FSYNC_ACQUIRED(taskdata); // acquire self - sync with children
2155 #endif /* USE_ITT_BUILD */
2156 
2157     // Debugger:  The taskwait is completed. Location remains, but thread is
2158     // negated.
2159     taskdata->td_taskwait_thread = -taskdata->td_taskwait_thread;
2160 
2161 #if OMPT_SUPPORT && OMPT_OPTIONAL
2162     if (ompt) {
2163       if (ompt_enabled.ompt_callback_sync_region_wait) {
2164         ompt_callbacks.ompt_callback(ompt_callback_sync_region_wait)(
2165             ompt_sync_region_taskwait, ompt_scope_end, my_parallel_data,
2166             my_task_data, return_address);
2167       }
2168       if (ompt_enabled.ompt_callback_sync_region) {
2169         ompt_callbacks.ompt_callback(ompt_callback_sync_region)(
2170             ompt_sync_region_taskwait, ompt_scope_end, my_parallel_data,
2171             my_task_data, return_address);
2172       }
2173       taskdata->ompt_task_info.frame.enter_frame = ompt_data_none;
2174     }
2175 #endif // OMPT_SUPPORT && OMPT_OPTIONAL
2176 
2177   }
2178 
2179   KA_TRACE(10, ("__kmpc_omp_taskwait(exit): T#%d task %p finished waiting, "
2180                 "returning TASK_CURRENT_NOT_QUEUED\n",
2181                 gtid, taskdata));
2182 
2183   return TASK_CURRENT_NOT_QUEUED;
2184 }
2185 
2186 #if OMPT_SUPPORT && OMPT_OPTIONAL
2187 OMPT_NOINLINE
2188 static kmp_int32 __kmpc_omp_taskwait_ompt(ident_t *loc_ref, kmp_int32 gtid,
2189                                           void *frame_address,
2190                                           void *return_address) {
2191   return __kmpc_omp_taskwait_template<true>(loc_ref, gtid, frame_address,
2192                                             return_address);
2193 }
2194 #endif // OMPT_SUPPORT && OMPT_OPTIONAL
2195 
2196 // __kmpc_omp_taskwait: Wait until all tasks generated by the current task are
2197 // complete
2198 kmp_int32 __kmpc_omp_taskwait(ident_t *loc_ref, kmp_int32 gtid) {
2199 #if OMPT_SUPPORT && OMPT_OPTIONAL
2200   if (UNLIKELY(ompt_enabled.enabled)) {
2201     OMPT_STORE_RETURN_ADDRESS(gtid);
2202     return __kmpc_omp_taskwait_ompt(loc_ref, gtid, OMPT_GET_FRAME_ADDRESS(0),
2203                                     OMPT_LOAD_RETURN_ADDRESS(gtid));
2204   }
2205 #endif
2206   return __kmpc_omp_taskwait_template<false>(loc_ref, gtid, NULL, NULL);
2207 }
2208 
2209 // __kmpc_omp_taskyield: switch to a different task
2210 kmp_int32 __kmpc_omp_taskyield(ident_t *loc_ref, kmp_int32 gtid, int end_part) {
2211   kmp_taskdata_t *taskdata = NULL;
2212   kmp_info_t *thread;
2213   int thread_finished = FALSE;
2214 
2215   KMP_COUNT_BLOCK(OMP_TASKYIELD);
2216   KMP_SET_THREAD_STATE_BLOCK(TASKYIELD);
2217 
2218   KA_TRACE(10, ("__kmpc_omp_taskyield(enter): T#%d loc=%p end_part = %d\n",
2219                 gtid, loc_ref, end_part));
2220   __kmp_assert_valid_gtid(gtid);
2221 
2222   if (__kmp_tasking_mode != tskm_immediate_exec && __kmp_init_parallel) {
2223     thread = __kmp_threads[gtid];
2224     taskdata = thread->th.th_current_task;
2225 // Should we model this as a task wait or not?
2226 // Debugger: The taskwait is active. Store location and thread encountered the
2227 // taskwait.
2228 #if USE_ITT_BUILD
2229 // Note: These values are used by ITT events as well.
2230 #endif /* USE_ITT_BUILD */
2231     taskdata->td_taskwait_counter += 1;
2232     taskdata->td_taskwait_ident = loc_ref;
2233     taskdata->td_taskwait_thread = gtid + 1;
2234 
2235 #if USE_ITT_BUILD
2236     void *itt_sync_obj = NULL;
2237 #if USE_ITT_NOTIFY
2238     KMP_ITT_TASKWAIT_STARTING(itt_sync_obj);
2239 #endif /* USE_ITT_NOTIFY */
2240 #endif /* USE_ITT_BUILD */
2241     if (!taskdata->td_flags.team_serial) {
2242       kmp_task_team_t *task_team = thread->th.th_task_team;
2243       if (task_team != NULL) {
2244         if (KMP_TASKING_ENABLED(task_team)) {
2245 #if OMPT_SUPPORT
2246           if (UNLIKELY(ompt_enabled.enabled))
2247             thread->th.ompt_thread_info.ompt_task_yielded = 1;
2248 #endif
2249           __kmp_execute_tasks_32(
2250               thread, gtid, (kmp_flag_32<> *)NULL, FALSE,
2251               &thread_finished USE_ITT_BUILD_ARG(itt_sync_obj),
2252               __kmp_task_stealing_constraint);
2253 #if OMPT_SUPPORT
2254           if (UNLIKELY(ompt_enabled.enabled))
2255             thread->th.ompt_thread_info.ompt_task_yielded = 0;
2256 #endif
2257         }
2258       }
2259     }
2260 #if USE_ITT_BUILD
2261     KMP_ITT_TASKWAIT_FINISHED(itt_sync_obj);
2262 #endif /* USE_ITT_BUILD */
2263 
2264     // Debugger:  The taskwait is completed. Location remains, but thread is
2265     // negated.
2266     taskdata->td_taskwait_thread = -taskdata->td_taskwait_thread;
2267   }
2268 
2269   KA_TRACE(10, ("__kmpc_omp_taskyield(exit): T#%d task %p resuming, "
2270                 "returning TASK_CURRENT_NOT_QUEUED\n",
2271                 gtid, taskdata));
2272 
2273   return TASK_CURRENT_NOT_QUEUED;
2274 }
2275 
2276 // Task Reduction implementation
2277 //
2278 // Note: initial implementation didn't take into account the possibility
2279 // to specify omp_orig for initializer of the UDR (user defined reduction).
2280 // Corrected implementation takes into account the omp_orig object.
2281 // Compiler is free to use old implementation if omp_orig is not specified.
2282 
2283 /*!
2284 @ingroup BASIC_TYPES
2285 @{
2286 */
2287 
2288 /*!
2289 Flags for special info per task reduction item.
2290 */
2291 typedef struct kmp_taskred_flags {
2292   /*! 1 - use lazy alloc/init (e.g. big objects, num tasks < num threads) */
2293   unsigned lazy_priv : 1;
2294   unsigned reserved31 : 31;
2295 } kmp_taskred_flags_t;
2296 
2297 /*!
2298 Internal struct for reduction data item related info set up by compiler.
2299 */
2300 typedef struct kmp_task_red_input {
2301   void *reduce_shar; /**< shared between tasks item to reduce into */
2302   size_t reduce_size; /**< size of data item in bytes */
2303   // three compiler-generated routines (init, fini are optional):
2304   void *reduce_init; /**< data initialization routine (single parameter) */
2305   void *reduce_fini; /**< data finalization routine */
2306   void *reduce_comb; /**< data combiner routine */
2307   kmp_taskred_flags_t flags; /**< flags for additional info from compiler */
2308 } kmp_task_red_input_t;
2309 
2310 /*!
2311 Internal struct for reduction data item related info saved by the library.
2312 */
2313 typedef struct kmp_taskred_data {
2314   void *reduce_shar; /**< shared between tasks item to reduce into */
2315   size_t reduce_size; /**< size of data item */
2316   kmp_taskred_flags_t flags; /**< flags for additional info from compiler */
2317   void *reduce_priv; /**< array of thread specific items */
2318   void *reduce_pend; /**< end of private data for faster comparison op */
2319   // three compiler-generated routines (init, fini are optional):
2320   void *reduce_comb; /**< data combiner routine */
2321   void *reduce_init; /**< data initialization routine (two parameters) */
2322   void *reduce_fini; /**< data finalization routine */
2323   void *reduce_orig; /**< original item (can be used in UDR initializer) */
2324 } kmp_taskred_data_t;
2325 
2326 /*!
2327 Internal struct for reduction data item related info set up by compiler.
2328 
2329 New interface: added reduce_orig field to provide omp_orig for UDR initializer.
2330 */
2331 typedef struct kmp_taskred_input {
2332   void *reduce_shar; /**< shared between tasks item to reduce into */
2333   void *reduce_orig; /**< original reduction item used for initialization */
2334   size_t reduce_size; /**< size of data item */
2335   // three compiler-generated routines (init, fini are optional):
2336   void *reduce_init; /**< data initialization routine (two parameters) */
2337   void *reduce_fini; /**< data finalization routine */
2338   void *reduce_comb; /**< data combiner routine */
2339   kmp_taskred_flags_t flags; /**< flags for additional info from compiler */
2340 } kmp_taskred_input_t;
2341 /*!
2342 @}
2343 */
2344 
2345 template <typename T> void __kmp_assign_orig(kmp_taskred_data_t &item, T &src);
2346 template <>
2347 void __kmp_assign_orig<kmp_task_red_input_t>(kmp_taskred_data_t &item,
2348                                              kmp_task_red_input_t &src) {
2349   item.reduce_orig = NULL;
2350 }
2351 template <>
2352 void __kmp_assign_orig<kmp_taskred_input_t>(kmp_taskred_data_t &item,
2353                                             kmp_taskred_input_t &src) {
2354   if (src.reduce_orig != NULL) {
2355     item.reduce_orig = src.reduce_orig;
2356   } else {
2357     item.reduce_orig = src.reduce_shar;
2358   } // non-NULL reduce_orig means new interface used
2359 }
2360 
2361 template <typename T> void __kmp_call_init(kmp_taskred_data_t &item, size_t j);
2362 template <>
2363 void __kmp_call_init<kmp_task_red_input_t>(kmp_taskred_data_t &item,
2364                                            size_t offset) {
2365   ((void (*)(void *))item.reduce_init)((char *)(item.reduce_priv) + offset);
2366 }
2367 template <>
2368 void __kmp_call_init<kmp_taskred_input_t>(kmp_taskred_data_t &item,
2369                                           size_t offset) {
2370   ((void (*)(void *, void *))item.reduce_init)(
2371       (char *)(item.reduce_priv) + offset, item.reduce_orig);
2372 }
2373 
2374 template <typename T>
2375 void *__kmp_task_reduction_init(int gtid, int num, T *data) {
2376   __kmp_assert_valid_gtid(gtid);
2377   kmp_info_t *thread = __kmp_threads[gtid];
2378   kmp_taskgroup_t *tg = thread->th.th_current_task->td_taskgroup;
2379   kmp_uint32 nth = thread->th.th_team_nproc;
2380   kmp_taskred_data_t *arr;
2381 
2382   // check input data just in case
2383   KMP_ASSERT(tg != NULL);
2384   KMP_ASSERT(data != NULL);
2385   KMP_ASSERT(num > 0);
2386   if (nth == 1) {
2387     KA_TRACE(10, ("__kmpc_task_reduction_init: T#%d, tg %p, exiting nth=1\n",
2388                   gtid, tg));
2389     return (void *)tg;
2390   }
2391   KA_TRACE(10, ("__kmpc_task_reduction_init: T#%d, taskgroup %p, #items %d\n",
2392                 gtid, tg, num));
2393   arr = (kmp_taskred_data_t *)__kmp_thread_malloc(
2394       thread, num * sizeof(kmp_taskred_data_t));
2395   for (int i = 0; i < num; ++i) {
2396     size_t size = data[i].reduce_size - 1;
2397     // round the size up to cache line per thread-specific item
2398     size += CACHE_LINE - size % CACHE_LINE;
2399     KMP_ASSERT(data[i].reduce_comb != NULL); // combiner is mandatory
2400     arr[i].reduce_shar = data[i].reduce_shar;
2401     arr[i].reduce_size = size;
2402     arr[i].flags = data[i].flags;
2403     arr[i].reduce_comb = data[i].reduce_comb;
2404     arr[i].reduce_init = data[i].reduce_init;
2405     arr[i].reduce_fini = data[i].reduce_fini;
2406     __kmp_assign_orig<T>(arr[i], data[i]);
2407     if (!arr[i].flags.lazy_priv) {
2408       // allocate cache-line aligned block and fill it with zeros
2409       arr[i].reduce_priv = __kmp_allocate(nth * size);
2410       arr[i].reduce_pend = (char *)(arr[i].reduce_priv) + nth * size;
2411       if (arr[i].reduce_init != NULL) {
2412         // initialize all thread-specific items
2413         for (size_t j = 0; j < nth; ++j) {
2414           __kmp_call_init<T>(arr[i], j * size);
2415         }
2416       }
2417     } else {
2418       // only allocate space for pointers now,
2419       // objects will be lazily allocated/initialized if/when requested
2420       // note that __kmp_allocate zeroes the allocated memory
2421       arr[i].reduce_priv = __kmp_allocate(nth * sizeof(void *));
2422     }
2423   }
2424   tg->reduce_data = (void *)arr;
2425   tg->reduce_num_data = num;
2426   return (void *)tg;
2427 }
2428 
2429 /*!
2430 @ingroup TASKING
2431 @param gtid      Global thread ID
2432 @param num       Number of data items to reduce
2433 @param data      Array of data for reduction
2434 @return The taskgroup identifier
2435 
2436 Initialize task reduction for the taskgroup.
2437 
2438 Note: this entry supposes the optional compiler-generated initializer routine
2439 has single parameter - pointer to object to be initialized. That means
2440 the reduction either does not use omp_orig object, or the omp_orig is accessible
2441 without help of the runtime library.
2442 */
2443 void *__kmpc_task_reduction_init(int gtid, int num, void *data) {
2444   return __kmp_task_reduction_init(gtid, num, (kmp_task_red_input_t *)data);
2445 }
2446 
2447 /*!
2448 @ingroup TASKING
2449 @param gtid      Global thread ID
2450 @param num       Number of data items to reduce
2451 @param data      Array of data for reduction
2452 @return The taskgroup identifier
2453 
2454 Initialize task reduction for the taskgroup.
2455 
2456 Note: this entry supposes the optional compiler-generated initializer routine
2457 has two parameters, pointer to object to be initialized and pointer to omp_orig
2458 */
2459 void *__kmpc_taskred_init(int gtid, int num, void *data) {
2460   return __kmp_task_reduction_init(gtid, num, (kmp_taskred_input_t *)data);
2461 }
2462 
2463 // Copy task reduction data (except for shared pointers).
2464 template <typename T>
2465 void __kmp_task_reduction_init_copy(kmp_info_t *thr, int num, T *data,
2466                                     kmp_taskgroup_t *tg, void *reduce_data) {
2467   kmp_taskred_data_t *arr;
2468   KA_TRACE(20, ("__kmp_task_reduction_init_copy: Th %p, init taskgroup %p,"
2469                 " from data %p\n",
2470                 thr, tg, reduce_data));
2471   arr = (kmp_taskred_data_t *)__kmp_thread_malloc(
2472       thr, num * sizeof(kmp_taskred_data_t));
2473   // threads will share private copies, thunk routines, sizes, flags, etc.:
2474   KMP_MEMCPY(arr, reduce_data, num * sizeof(kmp_taskred_data_t));
2475   for (int i = 0; i < num; ++i) {
2476     arr[i].reduce_shar = data[i].reduce_shar; // init unique shared pointers
2477   }
2478   tg->reduce_data = (void *)arr;
2479   tg->reduce_num_data = num;
2480 }
2481 
2482 /*!
2483 @ingroup TASKING
2484 @param gtid    Global thread ID
2485 @param tskgrp  The taskgroup ID (optional)
2486 @param data    Shared location of the item
2487 @return The pointer to per-thread data
2488 
2489 Get thread-specific location of data item
2490 */
2491 void *__kmpc_task_reduction_get_th_data(int gtid, void *tskgrp, void *data) {
2492   __kmp_assert_valid_gtid(gtid);
2493   kmp_info_t *thread = __kmp_threads[gtid];
2494   kmp_int32 nth = thread->th.th_team_nproc;
2495   if (nth == 1)
2496     return data; // nothing to do
2497 
2498   kmp_taskgroup_t *tg = (kmp_taskgroup_t *)tskgrp;
2499   if (tg == NULL)
2500     tg = thread->th.th_current_task->td_taskgroup;
2501   KMP_ASSERT(tg != NULL);
2502   kmp_taskred_data_t *arr = (kmp_taskred_data_t *)(tg->reduce_data);
2503   kmp_int32 num = tg->reduce_num_data;
2504   kmp_int32 tid = thread->th.th_info.ds.ds_tid;
2505 
2506   KMP_ASSERT(data != NULL);
2507   while (tg != NULL) {
2508     for (int i = 0; i < num; ++i) {
2509       if (!arr[i].flags.lazy_priv) {
2510         if (data == arr[i].reduce_shar ||
2511             (data >= arr[i].reduce_priv && data < arr[i].reduce_pend))
2512           return (char *)(arr[i].reduce_priv) + tid * arr[i].reduce_size;
2513       } else {
2514         // check shared location first
2515         void **p_priv = (void **)(arr[i].reduce_priv);
2516         if (data == arr[i].reduce_shar)
2517           goto found;
2518         // check if we get some thread specific location as parameter
2519         for (int j = 0; j < nth; ++j)
2520           if (data == p_priv[j])
2521             goto found;
2522         continue; // not found, continue search
2523       found:
2524         if (p_priv[tid] == NULL) {
2525           // allocate thread specific object lazily
2526           p_priv[tid] = __kmp_allocate(arr[i].reduce_size);
2527           if (arr[i].reduce_init != NULL) {
2528             if (arr[i].reduce_orig != NULL) { // new interface
2529               ((void (*)(void *, void *))arr[i].reduce_init)(
2530                   p_priv[tid], arr[i].reduce_orig);
2531             } else { // old interface (single parameter)
2532               ((void (*)(void *))arr[i].reduce_init)(p_priv[tid]);
2533             }
2534           }
2535         }
2536         return p_priv[tid];
2537       }
2538     }
2539     tg = tg->parent;
2540     arr = (kmp_taskred_data_t *)(tg->reduce_data);
2541     num = tg->reduce_num_data;
2542   }
2543   KMP_ASSERT2(0, "Unknown task reduction item");
2544   return NULL; // ERROR, this line never executed
2545 }
2546 
2547 // Finalize task reduction.
2548 // Called from __kmpc_end_taskgroup()
2549 static void __kmp_task_reduction_fini(kmp_info_t *th, kmp_taskgroup_t *tg) {
2550   kmp_int32 nth = th->th.th_team_nproc;
2551   KMP_DEBUG_ASSERT(nth > 1); // should not be called if nth == 1
2552   kmp_taskred_data_t *arr = (kmp_taskred_data_t *)tg->reduce_data;
2553   kmp_int32 num = tg->reduce_num_data;
2554   for (int i = 0; i < num; ++i) {
2555     void *sh_data = arr[i].reduce_shar;
2556     void (*f_fini)(void *) = (void (*)(void *))(arr[i].reduce_fini);
2557     void (*f_comb)(void *, void *) =
2558         (void (*)(void *, void *))(arr[i].reduce_comb);
2559     if (!arr[i].flags.lazy_priv) {
2560       void *pr_data = arr[i].reduce_priv;
2561       size_t size = arr[i].reduce_size;
2562       for (int j = 0; j < nth; ++j) {
2563         void *priv_data = (char *)pr_data + j * size;
2564         f_comb(sh_data, priv_data); // combine results
2565         if (f_fini)
2566           f_fini(priv_data); // finalize if needed
2567       }
2568     } else {
2569       void **pr_data = (void **)(arr[i].reduce_priv);
2570       for (int j = 0; j < nth; ++j) {
2571         if (pr_data[j] != NULL) {
2572           f_comb(sh_data, pr_data[j]); // combine results
2573           if (f_fini)
2574             f_fini(pr_data[j]); // finalize if needed
2575           __kmp_free(pr_data[j]);
2576         }
2577       }
2578     }
2579     __kmp_free(arr[i].reduce_priv);
2580   }
2581   __kmp_thread_free(th, arr);
2582   tg->reduce_data = NULL;
2583   tg->reduce_num_data = 0;
2584 }
2585 
2586 // Cleanup task reduction data for parallel or worksharing,
2587 // do not touch task private data other threads still working with.
2588 // Called from __kmpc_end_taskgroup()
2589 static void __kmp_task_reduction_clean(kmp_info_t *th, kmp_taskgroup_t *tg) {
2590   __kmp_thread_free(th, tg->reduce_data);
2591   tg->reduce_data = NULL;
2592   tg->reduce_num_data = 0;
2593 }
2594 
2595 template <typename T>
2596 void *__kmp_task_reduction_modifier_init(ident_t *loc, int gtid, int is_ws,
2597                                          int num, T *data) {
2598   __kmp_assert_valid_gtid(gtid);
2599   kmp_info_t *thr = __kmp_threads[gtid];
2600   kmp_int32 nth = thr->th.th_team_nproc;
2601   __kmpc_taskgroup(loc, gtid); // form new taskgroup first
2602   if (nth == 1) {
2603     KA_TRACE(10,
2604              ("__kmpc_reduction_modifier_init: T#%d, tg %p, exiting nth=1\n",
2605               gtid, thr->th.th_current_task->td_taskgroup));
2606     return (void *)thr->th.th_current_task->td_taskgroup;
2607   }
2608   kmp_team_t *team = thr->th.th_team;
2609   void *reduce_data;
2610   kmp_taskgroup_t *tg;
2611   reduce_data = KMP_ATOMIC_LD_RLX(&team->t.t_tg_reduce_data[is_ws]);
2612   if (reduce_data == NULL &&
2613       __kmp_atomic_compare_store(&team->t.t_tg_reduce_data[is_ws], reduce_data,
2614                                  (void *)1)) {
2615     // single thread enters this block to initialize common reduction data
2616     KMP_DEBUG_ASSERT(reduce_data == NULL);
2617     // first initialize own data, then make a copy other threads can use
2618     tg = (kmp_taskgroup_t *)__kmp_task_reduction_init<T>(gtid, num, data);
2619     reduce_data = __kmp_thread_malloc(thr, num * sizeof(kmp_taskred_data_t));
2620     KMP_MEMCPY(reduce_data, tg->reduce_data, num * sizeof(kmp_taskred_data_t));
2621     // fini counters should be 0 at this point
2622     KMP_DEBUG_ASSERT(KMP_ATOMIC_LD_RLX(&team->t.t_tg_fini_counter[0]) == 0);
2623     KMP_DEBUG_ASSERT(KMP_ATOMIC_LD_RLX(&team->t.t_tg_fini_counter[1]) == 0);
2624     KMP_ATOMIC_ST_REL(&team->t.t_tg_reduce_data[is_ws], reduce_data);
2625   } else {
2626     while (
2627         (reduce_data = KMP_ATOMIC_LD_ACQ(&team->t.t_tg_reduce_data[is_ws])) ==
2628         (void *)1) { // wait for task reduction initialization
2629       KMP_CPU_PAUSE();
2630     }
2631     KMP_DEBUG_ASSERT(reduce_data > (void *)1); // should be valid pointer here
2632     tg = thr->th.th_current_task->td_taskgroup;
2633     __kmp_task_reduction_init_copy<T>(thr, num, data, tg, reduce_data);
2634   }
2635   return tg;
2636 }
2637 
2638 /*!
2639 @ingroup TASKING
2640 @param loc       Source location info
2641 @param gtid      Global thread ID
2642 @param is_ws     Is 1 if the reduction is for worksharing, 0 otherwise
2643 @param num       Number of data items to reduce
2644 @param data      Array of data for reduction
2645 @return The taskgroup identifier
2646 
2647 Initialize task reduction for a parallel or worksharing.
2648 
2649 Note: this entry supposes the optional compiler-generated initializer routine
2650 has single parameter - pointer to object to be initialized. That means
2651 the reduction either does not use omp_orig object, or the omp_orig is accessible
2652 without help of the runtime library.
2653 */
2654 void *__kmpc_task_reduction_modifier_init(ident_t *loc, int gtid, int is_ws,
2655                                           int num, void *data) {
2656   return __kmp_task_reduction_modifier_init(loc, gtid, is_ws, num,
2657                                             (kmp_task_red_input_t *)data);
2658 }
2659 
2660 /*!
2661 @ingroup TASKING
2662 @param loc       Source location info
2663 @param gtid      Global thread ID
2664 @param is_ws     Is 1 if the reduction is for worksharing, 0 otherwise
2665 @param num       Number of data items to reduce
2666 @param data      Array of data for reduction
2667 @return The taskgroup identifier
2668 
2669 Initialize task reduction for a parallel or worksharing.
2670 
2671 Note: this entry supposes the optional compiler-generated initializer routine
2672 has two parameters, pointer to object to be initialized and pointer to omp_orig
2673 */
2674 void *__kmpc_taskred_modifier_init(ident_t *loc, int gtid, int is_ws, int num,
2675                                    void *data) {
2676   return __kmp_task_reduction_modifier_init(loc, gtid, is_ws, num,
2677                                             (kmp_taskred_input_t *)data);
2678 }
2679 
2680 /*!
2681 @ingroup TASKING
2682 @param loc       Source location info
2683 @param gtid      Global thread ID
2684 @param is_ws     Is 1 if the reduction is for worksharing, 0 otherwise
2685 
2686 Finalize task reduction for a parallel or worksharing.
2687 */
2688 void __kmpc_task_reduction_modifier_fini(ident_t *loc, int gtid, int is_ws) {
2689   __kmpc_end_taskgroup(loc, gtid);
2690 }
2691 
2692 // __kmpc_taskgroup: Start a new taskgroup
2693 void __kmpc_taskgroup(ident_t *loc, int gtid) {
2694   __kmp_assert_valid_gtid(gtid);
2695   kmp_info_t *thread = __kmp_threads[gtid];
2696   kmp_taskdata_t *taskdata = thread->th.th_current_task;
2697   kmp_taskgroup_t *tg_new =
2698       (kmp_taskgroup_t *)__kmp_thread_malloc(thread, sizeof(kmp_taskgroup_t));
2699   KA_TRACE(10, ("__kmpc_taskgroup: T#%d loc=%p group=%p\n", gtid, loc, tg_new));
2700   KMP_ATOMIC_ST_RLX(&tg_new->count, 0);
2701   KMP_ATOMIC_ST_RLX(&tg_new->cancel_request, cancel_noreq);
2702   tg_new->parent = taskdata->td_taskgroup;
2703   tg_new->reduce_data = NULL;
2704   tg_new->reduce_num_data = 0;
2705   tg_new->gomp_data = NULL;
2706   taskdata->td_taskgroup = tg_new;
2707 
2708 #if OMPT_SUPPORT && OMPT_OPTIONAL
2709   if (UNLIKELY(ompt_enabled.ompt_callback_sync_region)) {
2710     void *codeptr = OMPT_LOAD_RETURN_ADDRESS(gtid);
2711     if (!codeptr)
2712       codeptr = OMPT_GET_RETURN_ADDRESS(0);
2713     kmp_team_t *team = thread->th.th_team;
2714     ompt_data_t my_task_data = taskdata->ompt_task_info.task_data;
2715     // FIXME: I think this is wrong for lwt!
2716     ompt_data_t my_parallel_data = team->t.ompt_team_info.parallel_data;
2717 
2718     ompt_callbacks.ompt_callback(ompt_callback_sync_region)(
2719         ompt_sync_region_taskgroup, ompt_scope_begin, &(my_parallel_data),
2720         &(my_task_data), codeptr);
2721   }
2722 #endif
2723 }
2724 
2725 // __kmpc_end_taskgroup: Wait until all tasks generated by the current task
2726 //                       and its descendants are complete
2727 void __kmpc_end_taskgroup(ident_t *loc, int gtid) {
2728   __kmp_assert_valid_gtid(gtid);
2729   kmp_info_t *thread = __kmp_threads[gtid];
2730   kmp_taskdata_t *taskdata = thread->th.th_current_task;
2731   kmp_taskgroup_t *taskgroup = taskdata->td_taskgroup;
2732   int thread_finished = FALSE;
2733 
2734 #if OMPT_SUPPORT && OMPT_OPTIONAL
2735   kmp_team_t *team;
2736   ompt_data_t my_task_data;
2737   ompt_data_t my_parallel_data;
2738   void *codeptr = nullptr;
2739   if (UNLIKELY(ompt_enabled.enabled)) {
2740     team = thread->th.th_team;
2741     my_task_data = taskdata->ompt_task_info.task_data;
2742     // FIXME: I think this is wrong for lwt!
2743     my_parallel_data = team->t.ompt_team_info.parallel_data;
2744     codeptr = OMPT_LOAD_RETURN_ADDRESS(gtid);
2745     if (!codeptr)
2746       codeptr = OMPT_GET_RETURN_ADDRESS(0);
2747   }
2748 #endif
2749 
2750   KA_TRACE(10, ("__kmpc_end_taskgroup(enter): T#%d loc=%p\n", gtid, loc));
2751   KMP_DEBUG_ASSERT(taskgroup != NULL);
2752   KMP_SET_THREAD_STATE_BLOCK(TASKGROUP);
2753 
2754   if (__kmp_tasking_mode != tskm_immediate_exec) {
2755     // mark task as waiting not on a barrier
2756     taskdata->td_taskwait_counter += 1;
2757     taskdata->td_taskwait_ident = loc;
2758     taskdata->td_taskwait_thread = gtid + 1;
2759 #if USE_ITT_BUILD
2760     // For ITT the taskgroup wait is similar to taskwait until we need to
2761     // distinguish them
2762     void *itt_sync_obj = NULL;
2763 #if USE_ITT_NOTIFY
2764     KMP_ITT_TASKWAIT_STARTING(itt_sync_obj);
2765 #endif /* USE_ITT_NOTIFY */
2766 #endif /* USE_ITT_BUILD */
2767 
2768 #if OMPT_SUPPORT && OMPT_OPTIONAL
2769     if (UNLIKELY(ompt_enabled.ompt_callback_sync_region_wait)) {
2770       ompt_callbacks.ompt_callback(ompt_callback_sync_region_wait)(
2771           ompt_sync_region_taskgroup, ompt_scope_begin, &(my_parallel_data),
2772           &(my_task_data), codeptr);
2773     }
2774 #endif
2775 
2776     if (!taskdata->td_flags.team_serial ||
2777         (thread->th.th_task_team != NULL &&
2778          (thread->th.th_task_team->tt.tt_found_proxy_tasks ||
2779           thread->th.th_task_team->tt.tt_hidden_helper_task_encountered))) {
2780       kmp_flag_32<false, false> flag(
2781           RCAST(std::atomic<kmp_uint32> *, &(taskgroup->count)), 0U);
2782       while (KMP_ATOMIC_LD_ACQ(&taskgroup->count) != 0) {
2783         flag.execute_tasks(thread, gtid, FALSE,
2784                            &thread_finished USE_ITT_BUILD_ARG(itt_sync_obj),
2785                            __kmp_task_stealing_constraint);
2786       }
2787     }
2788     taskdata->td_taskwait_thread = -taskdata->td_taskwait_thread; // end waiting
2789 
2790 #if OMPT_SUPPORT && OMPT_OPTIONAL
2791     if (UNLIKELY(ompt_enabled.ompt_callback_sync_region_wait)) {
2792       ompt_callbacks.ompt_callback(ompt_callback_sync_region_wait)(
2793           ompt_sync_region_taskgroup, ompt_scope_end, &(my_parallel_data),
2794           &(my_task_data), codeptr);
2795     }
2796 #endif
2797 
2798 #if USE_ITT_BUILD
2799     KMP_ITT_TASKWAIT_FINISHED(itt_sync_obj);
2800     KMP_FSYNC_ACQUIRED(taskdata); // acquire self - sync with descendants
2801 #endif /* USE_ITT_BUILD */
2802   }
2803   KMP_DEBUG_ASSERT(taskgroup->count == 0);
2804 
2805   if (taskgroup->reduce_data != NULL &&
2806       !taskgroup->gomp_data) { // need to reduce?
2807     int cnt;
2808     void *reduce_data;
2809     kmp_team_t *t = thread->th.th_team;
2810     kmp_taskred_data_t *arr = (kmp_taskred_data_t *)taskgroup->reduce_data;
2811     // check if <priv> data of the first reduction variable shared for the team
2812     void *priv0 = arr[0].reduce_priv;
2813     if ((reduce_data = KMP_ATOMIC_LD_ACQ(&t->t.t_tg_reduce_data[0])) != NULL &&
2814         ((kmp_taskred_data_t *)reduce_data)[0].reduce_priv == priv0) {
2815       // finishing task reduction on parallel
2816       cnt = KMP_ATOMIC_INC(&t->t.t_tg_fini_counter[0]);
2817       if (cnt == thread->th.th_team_nproc - 1) {
2818         // we are the last thread passing __kmpc_reduction_modifier_fini()
2819         // finalize task reduction:
2820         __kmp_task_reduction_fini(thread, taskgroup);
2821         // cleanup fields in the team structure:
2822         // TODO: is relaxed store enough here (whole barrier should follow)?
2823         __kmp_thread_free(thread, reduce_data);
2824         KMP_ATOMIC_ST_REL(&t->t.t_tg_reduce_data[0], NULL);
2825         KMP_ATOMIC_ST_REL(&t->t.t_tg_fini_counter[0], 0);
2826       } else {
2827         // we are not the last thread passing __kmpc_reduction_modifier_fini(),
2828         // so do not finalize reduction, just clean own copy of the data
2829         __kmp_task_reduction_clean(thread, taskgroup);
2830       }
2831     } else if ((reduce_data = KMP_ATOMIC_LD_ACQ(&t->t.t_tg_reduce_data[1])) !=
2832                    NULL &&
2833                ((kmp_taskred_data_t *)reduce_data)[0].reduce_priv == priv0) {
2834       // finishing task reduction on worksharing
2835       cnt = KMP_ATOMIC_INC(&t->t.t_tg_fini_counter[1]);
2836       if (cnt == thread->th.th_team_nproc - 1) {
2837         // we are the last thread passing __kmpc_reduction_modifier_fini()
2838         __kmp_task_reduction_fini(thread, taskgroup);
2839         // cleanup fields in team structure:
2840         // TODO: is relaxed store enough here (whole barrier should follow)?
2841         __kmp_thread_free(thread, reduce_data);
2842         KMP_ATOMIC_ST_REL(&t->t.t_tg_reduce_data[1], NULL);
2843         KMP_ATOMIC_ST_REL(&t->t.t_tg_fini_counter[1], 0);
2844       } else {
2845         // we are not the last thread passing __kmpc_reduction_modifier_fini(),
2846         // so do not finalize reduction, just clean own copy of the data
2847         __kmp_task_reduction_clean(thread, taskgroup);
2848       }
2849     } else {
2850       // finishing task reduction on taskgroup
2851       __kmp_task_reduction_fini(thread, taskgroup);
2852     }
2853   }
2854   // Restore parent taskgroup for the current task
2855   taskdata->td_taskgroup = taskgroup->parent;
2856   __kmp_thread_free(thread, taskgroup);
2857 
2858   KA_TRACE(10, ("__kmpc_end_taskgroup(exit): T#%d task %p finished waiting\n",
2859                 gtid, taskdata));
2860 
2861 #if OMPT_SUPPORT && OMPT_OPTIONAL
2862   if (UNLIKELY(ompt_enabled.ompt_callback_sync_region)) {
2863     ompt_callbacks.ompt_callback(ompt_callback_sync_region)(
2864         ompt_sync_region_taskgroup, ompt_scope_end, &(my_parallel_data),
2865         &(my_task_data), codeptr);
2866   }
2867 #endif
2868 }
2869 
2870 static kmp_task_t *__kmp_get_priority_task(kmp_int32 gtid,
2871                                            kmp_task_team_t *task_team,
2872                                            kmp_int32 is_constrained) {
2873   kmp_task_t *task = NULL;
2874   kmp_taskdata_t *taskdata;
2875   kmp_taskdata_t *current;
2876   kmp_thread_data_t *thread_data;
2877   int ntasks = task_team->tt.tt_num_task_pri;
2878   if (ntasks == 0) {
2879     KA_TRACE(
2880         20, ("__kmp_get_priority_task(exit #1): T#%d No tasks to get\n", gtid));
2881     return NULL;
2882   }
2883   do {
2884     // decrement num_tasks to "reserve" one task to get for execution
2885     if (__kmp_atomic_compare_store(&task_team->tt.tt_num_task_pri, ntasks,
2886                                    ntasks - 1))
2887       break;
2888   } while (ntasks > 0);
2889   if (ntasks == 0) {
2890     KA_TRACE(20, ("__kmp_get_priority_task(exit #2): T#%d No tasks to get\n",
2891                   __kmp_get_gtid()));
2892     return NULL;
2893   }
2894   // We got a "ticket" to get a "reserved" priority task
2895   int deque_ntasks;
2896   kmp_task_pri_t *list = task_team->tt.tt_task_pri_list;
2897   do {
2898     KMP_ASSERT(list != NULL);
2899     thread_data = &list->td;
2900     __kmp_acquire_bootstrap_lock(&thread_data->td.td_deque_lock);
2901     deque_ntasks = thread_data->td.td_deque_ntasks;
2902     if (deque_ntasks == 0) {
2903       __kmp_release_bootstrap_lock(&thread_data->td.td_deque_lock);
2904       KA_TRACE(20, ("__kmp_get_priority_task: T#%d No tasks to get from %p\n",
2905                     __kmp_get_gtid(), thread_data));
2906       list = list->next;
2907     }
2908   } while (deque_ntasks == 0);
2909   KMP_DEBUG_ASSERT(deque_ntasks);
2910   int target = thread_data->td.td_deque_head;
2911   current = __kmp_threads[gtid]->th.th_current_task;
2912   taskdata = thread_data->td.td_deque[target];
2913   if (__kmp_task_is_allowed(gtid, is_constrained, taskdata, current)) {
2914     // Bump head pointer and Wrap.
2915     thread_data->td.td_deque_head =
2916         (target + 1) & TASK_DEQUE_MASK(thread_data->td);
2917   } else {
2918     if (!task_team->tt.tt_untied_task_encountered) {
2919       // The TSC does not allow to steal victim task
2920       __kmp_release_bootstrap_lock(&thread_data->td.td_deque_lock);
2921       KA_TRACE(20, ("__kmp_get_priority_task(exit #3): T#%d could not get task "
2922                     "from %p: task_team=%p ntasks=%d head=%u tail=%u\n",
2923                     gtid, thread_data, task_team, deque_ntasks, target,
2924                     thread_data->td.td_deque_tail));
2925       task_team->tt.tt_num_task_pri++; // atomic inc, restore value
2926       return NULL;
2927     }
2928     int i;
2929     // walk through the deque trying to steal any task
2930     taskdata = NULL;
2931     for (i = 1; i < deque_ntasks; ++i) {
2932       target = (target + 1) & TASK_DEQUE_MASK(thread_data->td);
2933       taskdata = thread_data->td.td_deque[target];
2934       if (__kmp_task_is_allowed(gtid, is_constrained, taskdata, current)) {
2935         break; // found task to execute
2936       } else {
2937         taskdata = NULL;
2938       }
2939     }
2940     if (taskdata == NULL) {
2941       // No appropriate candidate found to execute
2942       __kmp_release_bootstrap_lock(&thread_data->td.td_deque_lock);
2943       KA_TRACE(
2944           10, ("__kmp_get_priority_task(exit #4): T#%d could not get task from "
2945                "%p: task_team=%p ntasks=%d head=%u tail=%u\n",
2946                gtid, thread_data, task_team, deque_ntasks,
2947                thread_data->td.td_deque_head, thread_data->td.td_deque_tail));
2948       task_team->tt.tt_num_task_pri++; // atomic inc, restore value
2949       return NULL;
2950     }
2951     int prev = target;
2952     for (i = i + 1; i < deque_ntasks; ++i) {
2953       // shift remaining tasks in the deque left by 1
2954       target = (target + 1) & TASK_DEQUE_MASK(thread_data->td);
2955       thread_data->td.td_deque[prev] = thread_data->td.td_deque[target];
2956       prev = target;
2957     }
2958     KMP_DEBUG_ASSERT(
2959         thread_data->td.td_deque_tail ==
2960         (kmp_uint32)((target + 1) & TASK_DEQUE_MASK(thread_data->td)));
2961     thread_data->td.td_deque_tail = target; // tail -= 1 (wrapped))
2962   }
2963   thread_data->td.td_deque_ntasks = deque_ntasks - 1;
2964   __kmp_release_bootstrap_lock(&thread_data->td.td_deque_lock);
2965   task = KMP_TASKDATA_TO_TASK(taskdata);
2966   return task;
2967 }
2968 
2969 // __kmp_remove_my_task: remove a task from my own deque
2970 static kmp_task_t *__kmp_remove_my_task(kmp_info_t *thread, kmp_int32 gtid,
2971                                         kmp_task_team_t *task_team,
2972                                         kmp_int32 is_constrained) {
2973   kmp_task_t *task;
2974   kmp_taskdata_t *taskdata;
2975   kmp_thread_data_t *thread_data;
2976   kmp_uint32 tail;
2977 
2978   KMP_DEBUG_ASSERT(__kmp_tasking_mode != tskm_immediate_exec);
2979   KMP_DEBUG_ASSERT(task_team->tt.tt_threads_data !=
2980                    NULL); // Caller should check this condition
2981 
2982   thread_data = &task_team->tt.tt_threads_data[__kmp_tid_from_gtid(gtid)];
2983 
2984   KA_TRACE(10, ("__kmp_remove_my_task(enter): T#%d ntasks=%d head=%u tail=%u\n",
2985                 gtid, thread_data->td.td_deque_ntasks,
2986                 thread_data->td.td_deque_head, thread_data->td.td_deque_tail));
2987 
2988   if (TCR_4(thread_data->td.td_deque_ntasks) == 0) {
2989     KA_TRACE(10,
2990              ("__kmp_remove_my_task(exit #1): T#%d No tasks to remove: "
2991               "ntasks=%d head=%u tail=%u\n",
2992               gtid, thread_data->td.td_deque_ntasks,
2993               thread_data->td.td_deque_head, thread_data->td.td_deque_tail));
2994     return NULL;
2995   }
2996 
2997   __kmp_acquire_bootstrap_lock(&thread_data->td.td_deque_lock);
2998 
2999   if (TCR_4(thread_data->td.td_deque_ntasks) == 0) {
3000     __kmp_release_bootstrap_lock(&thread_data->td.td_deque_lock);
3001     KA_TRACE(10,
3002              ("__kmp_remove_my_task(exit #2): T#%d No tasks to remove: "
3003               "ntasks=%d head=%u tail=%u\n",
3004               gtid, thread_data->td.td_deque_ntasks,
3005               thread_data->td.td_deque_head, thread_data->td.td_deque_tail));
3006     return NULL;
3007   }
3008 
3009   tail = (thread_data->td.td_deque_tail - 1) &
3010          TASK_DEQUE_MASK(thread_data->td); // Wrap index.
3011   taskdata = thread_data->td.td_deque[tail];
3012 
3013   if (!__kmp_task_is_allowed(gtid, is_constrained, taskdata,
3014                              thread->th.th_current_task)) {
3015     // The TSC does not allow to steal victim task
3016     __kmp_release_bootstrap_lock(&thread_data->td.td_deque_lock);
3017     KA_TRACE(10,
3018              ("__kmp_remove_my_task(exit #3): T#%d TSC blocks tail task: "
3019               "ntasks=%d head=%u tail=%u\n",
3020               gtid, thread_data->td.td_deque_ntasks,
3021               thread_data->td.td_deque_head, thread_data->td.td_deque_tail));
3022     return NULL;
3023   }
3024 
3025   thread_data->td.td_deque_tail = tail;
3026   TCW_4(thread_data->td.td_deque_ntasks, thread_data->td.td_deque_ntasks - 1);
3027 
3028   __kmp_release_bootstrap_lock(&thread_data->td.td_deque_lock);
3029 
3030   KA_TRACE(10, ("__kmp_remove_my_task(exit #4): T#%d task %p removed: "
3031                 "ntasks=%d head=%u tail=%u\n",
3032                 gtid, taskdata, thread_data->td.td_deque_ntasks,
3033                 thread_data->td.td_deque_head, thread_data->td.td_deque_tail));
3034 
3035   task = KMP_TASKDATA_TO_TASK(taskdata);
3036   return task;
3037 }
3038 
3039 // __kmp_steal_task: remove a task from another thread's deque
3040 // Assume that calling thread has already checked existence of
3041 // task_team thread_data before calling this routine.
3042 static kmp_task_t *__kmp_steal_task(kmp_info_t *victim_thr, kmp_int32 gtid,
3043                                     kmp_task_team_t *task_team,
3044                                     std::atomic<kmp_int32> *unfinished_threads,
3045                                     int *thread_finished,
3046                                     kmp_int32 is_constrained) {
3047   kmp_task_t *task;
3048   kmp_taskdata_t *taskdata;
3049   kmp_taskdata_t *current;
3050   kmp_thread_data_t *victim_td, *threads_data;
3051   kmp_int32 target;
3052   kmp_int32 victim_tid;
3053 
3054   KMP_DEBUG_ASSERT(__kmp_tasking_mode != tskm_immediate_exec);
3055 
3056   threads_data = task_team->tt.tt_threads_data;
3057   KMP_DEBUG_ASSERT(threads_data != NULL); // Caller should check this condition
3058 
3059   victim_tid = victim_thr->th.th_info.ds.ds_tid;
3060   victim_td = &threads_data[victim_tid];
3061 
3062   KA_TRACE(10, ("__kmp_steal_task(enter): T#%d try to steal from T#%d: "
3063                 "task_team=%p ntasks=%d head=%u tail=%u\n",
3064                 gtid, __kmp_gtid_from_thread(victim_thr), task_team,
3065                 victim_td->td.td_deque_ntasks, victim_td->td.td_deque_head,
3066                 victim_td->td.td_deque_tail));
3067 
3068   if (TCR_4(victim_td->td.td_deque_ntasks) == 0) {
3069     KA_TRACE(10, ("__kmp_steal_task(exit #1): T#%d could not steal from T#%d: "
3070                   "task_team=%p ntasks=%d head=%u tail=%u\n",
3071                   gtid, __kmp_gtid_from_thread(victim_thr), task_team,
3072                   victim_td->td.td_deque_ntasks, victim_td->td.td_deque_head,
3073                   victim_td->td.td_deque_tail));
3074     return NULL;
3075   }
3076 
3077   __kmp_acquire_bootstrap_lock(&victim_td->td.td_deque_lock);
3078 
3079   int ntasks = TCR_4(victim_td->td.td_deque_ntasks);
3080   // Check again after we acquire the lock
3081   if (ntasks == 0) {
3082     __kmp_release_bootstrap_lock(&victim_td->td.td_deque_lock);
3083     KA_TRACE(10, ("__kmp_steal_task(exit #2): T#%d could not steal from T#%d: "
3084                   "task_team=%p ntasks=%d head=%u tail=%u\n",
3085                   gtid, __kmp_gtid_from_thread(victim_thr), task_team, ntasks,
3086                   victim_td->td.td_deque_head, victim_td->td.td_deque_tail));
3087     return NULL;
3088   }
3089 
3090   KMP_DEBUG_ASSERT(victim_td->td.td_deque != NULL);
3091   current = __kmp_threads[gtid]->th.th_current_task;
3092   taskdata = victim_td->td.td_deque[victim_td->td.td_deque_head];
3093   if (__kmp_task_is_allowed(gtid, is_constrained, taskdata, current)) {
3094     // Bump head pointer and Wrap.
3095     victim_td->td.td_deque_head =
3096         (victim_td->td.td_deque_head + 1) & TASK_DEQUE_MASK(victim_td->td);
3097   } else {
3098     if (!task_team->tt.tt_untied_task_encountered) {
3099       // The TSC does not allow to steal victim task
3100       __kmp_release_bootstrap_lock(&victim_td->td.td_deque_lock);
3101       KA_TRACE(10, ("__kmp_steal_task(exit #3): T#%d could not steal from "
3102                     "T#%d: task_team=%p ntasks=%d head=%u tail=%u\n",
3103                     gtid, __kmp_gtid_from_thread(victim_thr), task_team, ntasks,
3104                     victim_td->td.td_deque_head, victim_td->td.td_deque_tail));
3105       return NULL;
3106     }
3107     int i;
3108     // walk through victim's deque trying to steal any task
3109     target = victim_td->td.td_deque_head;
3110     taskdata = NULL;
3111     for (i = 1; i < ntasks; ++i) {
3112       target = (target + 1) & TASK_DEQUE_MASK(victim_td->td);
3113       taskdata = victim_td->td.td_deque[target];
3114       if (__kmp_task_is_allowed(gtid, is_constrained, taskdata, current)) {
3115         break; // found victim task
3116       } else {
3117         taskdata = NULL;
3118       }
3119     }
3120     if (taskdata == NULL) {
3121       // No appropriate candidate to steal found
3122       __kmp_release_bootstrap_lock(&victim_td->td.td_deque_lock);
3123       KA_TRACE(10, ("__kmp_steal_task(exit #4): T#%d could not steal from "
3124                     "T#%d: task_team=%p ntasks=%d head=%u tail=%u\n",
3125                     gtid, __kmp_gtid_from_thread(victim_thr), task_team, ntasks,
3126                     victim_td->td.td_deque_head, victim_td->td.td_deque_tail));
3127       return NULL;
3128     }
3129     int prev = target;
3130     for (i = i + 1; i < ntasks; ++i) {
3131       // shift remaining tasks in the deque left by 1
3132       target = (target + 1) & TASK_DEQUE_MASK(victim_td->td);
3133       victim_td->td.td_deque[prev] = victim_td->td.td_deque[target];
3134       prev = target;
3135     }
3136     KMP_DEBUG_ASSERT(
3137         victim_td->td.td_deque_tail ==
3138         (kmp_uint32)((target + 1) & TASK_DEQUE_MASK(victim_td->td)));
3139     victim_td->td.td_deque_tail = target; // tail -= 1 (wrapped))
3140   }
3141   if (*thread_finished) {
3142     // We need to un-mark this victim as a finished victim.  This must be done
3143     // before releasing the lock, or else other threads (starting with the
3144     // primary thread victim) might be prematurely released from the barrier!!!
3145 #if KMP_DEBUG
3146     kmp_int32 count =
3147 #endif
3148         KMP_ATOMIC_INC(unfinished_threads);
3149     KA_TRACE(
3150         20,
3151         ("__kmp_steal_task: T#%d inc unfinished_threads to %d: task_team=%p\n",
3152          gtid, count + 1, task_team));
3153     *thread_finished = FALSE;
3154   }
3155   TCW_4(victim_td->td.td_deque_ntasks, ntasks - 1);
3156 
3157   __kmp_release_bootstrap_lock(&victim_td->td.td_deque_lock);
3158 
3159   KMP_COUNT_BLOCK(TASK_stolen);
3160   KA_TRACE(10,
3161            ("__kmp_steal_task(exit #5): T#%d stole task %p from T#%d: "
3162             "task_team=%p ntasks=%d head=%u tail=%u\n",
3163             gtid, taskdata, __kmp_gtid_from_thread(victim_thr), task_team,
3164             ntasks, victim_td->td.td_deque_head, victim_td->td.td_deque_tail));
3165 
3166   task = KMP_TASKDATA_TO_TASK(taskdata);
3167   return task;
3168 }
3169 
3170 // __kmp_execute_tasks_template: Choose and execute tasks until either the
3171 // condition is statisfied (return true) or there are none left (return false).
3172 //
3173 // final_spin is TRUE if this is the spin at the release barrier.
3174 // thread_finished indicates whether the thread is finished executing all
3175 // the tasks it has on its deque, and is at the release barrier.
3176 // spinner is the location on which to spin.
3177 // spinner == NULL means only execute a single task and return.
3178 // checker is the value to check to terminate the spin.
3179 template <class C>
3180 static inline int __kmp_execute_tasks_template(
3181     kmp_info_t *thread, kmp_int32 gtid, C *flag, int final_spin,
3182     int *thread_finished USE_ITT_BUILD_ARG(void *itt_sync_obj),
3183     kmp_int32 is_constrained) {
3184   kmp_task_team_t *task_team = thread->th.th_task_team;
3185   kmp_thread_data_t *threads_data;
3186   kmp_task_t *task;
3187   kmp_info_t *other_thread;
3188   kmp_taskdata_t *current_task = thread->th.th_current_task;
3189   std::atomic<kmp_int32> *unfinished_threads;
3190   kmp_int32 nthreads, victim_tid = -2, use_own_tasks = 1, new_victim = 0,
3191                       tid = thread->th.th_info.ds.ds_tid;
3192 
3193   KMP_DEBUG_ASSERT(__kmp_tasking_mode != tskm_immediate_exec);
3194   KMP_DEBUG_ASSERT(thread == __kmp_threads[gtid]);
3195 
3196   if (task_team == NULL || current_task == NULL)
3197     return FALSE;
3198 
3199   KA_TRACE(15, ("__kmp_execute_tasks_template(enter): T#%d final_spin=%d "
3200                 "*thread_finished=%d\n",
3201                 gtid, final_spin, *thread_finished));
3202 
3203   thread->th.th_reap_state = KMP_NOT_SAFE_TO_REAP;
3204   threads_data = (kmp_thread_data_t *)TCR_PTR(task_team->tt.tt_threads_data);
3205 
3206   KMP_DEBUG_ASSERT(threads_data != NULL);
3207 
3208   nthreads = task_team->tt.tt_nproc;
3209   unfinished_threads = &(task_team->tt.tt_unfinished_threads);
3210   KMP_DEBUG_ASSERT(nthreads > 1 || task_team->tt.tt_found_proxy_tasks ||
3211                    task_team->tt.tt_hidden_helper_task_encountered);
3212   KMP_DEBUG_ASSERT(*unfinished_threads >= 0);
3213 
3214   while (1) { // Outer loop keeps trying to find tasks in case of single thread
3215     // getting tasks from target constructs
3216     while (1) { // Inner loop to find a task and execute it
3217       task = NULL;
3218       if (task_team->tt.tt_num_task_pri) { // get priority task first
3219         task = __kmp_get_priority_task(gtid, task_team, is_constrained);
3220       }
3221       if (task == NULL && use_own_tasks) { // check own queue next
3222         task = __kmp_remove_my_task(thread, gtid, task_team, is_constrained);
3223       }
3224       if ((task == NULL) && (nthreads > 1)) { // Steal a task finally
3225         int asleep = 1;
3226         use_own_tasks = 0;
3227         // Try to steal from the last place I stole from successfully.
3228         if (victim_tid == -2) { // haven't stolen anything yet
3229           victim_tid = threads_data[tid].td.td_deque_last_stolen;
3230           if (victim_tid !=
3231               -1) // if we have a last stolen from victim, get the thread
3232             other_thread = threads_data[victim_tid].td.td_thr;
3233         }
3234         if (victim_tid != -1) { // found last victim
3235           asleep = 0;
3236         } else if (!new_victim) { // no recent steals and we haven't already
3237           // used a new victim; select a random thread
3238           do { // Find a different thread to steal work from.
3239             // Pick a random thread. Initial plan was to cycle through all the
3240             // threads, and only return if we tried to steal from every thread,
3241             // and failed.  Arch says that's not such a great idea.
3242             victim_tid = __kmp_get_random(thread) % (nthreads - 1);
3243             if (victim_tid >= tid) {
3244               ++victim_tid; // Adjusts random distribution to exclude self
3245             }
3246             // Found a potential victim
3247             other_thread = threads_data[victim_tid].td.td_thr;
3248             // There is a slight chance that __kmp_enable_tasking() did not wake
3249             // up all threads waiting at the barrier.  If victim is sleeping,
3250             // then wake it up. Since we were going to pay the cache miss
3251             // penalty for referencing another thread's kmp_info_t struct
3252             // anyway,
3253             // the check shouldn't cost too much performance at this point. In
3254             // extra barrier mode, tasks do not sleep at the separate tasking
3255             // barrier, so this isn't a problem.
3256             asleep = 0;
3257             if ((__kmp_tasking_mode == tskm_task_teams) &&
3258                 (__kmp_dflt_blocktime != KMP_MAX_BLOCKTIME) &&
3259                 (TCR_PTR(CCAST(void *, other_thread->th.th_sleep_loc)) !=
3260                  NULL)) {
3261               asleep = 1;
3262               __kmp_null_resume_wrapper(other_thread);
3263               // A sleeping thread should not have any tasks on it's queue.
3264               // There is a slight possibility that it resumes, steals a task
3265               // from another thread, which spawns more tasks, all in the time
3266               // that it takes this thread to check => don't write an assertion
3267               // that the victim's queue is empty.  Try stealing from a
3268               // different thread.
3269             }
3270           } while (asleep);
3271         }
3272 
3273         if (!asleep) {
3274           // We have a victim to try to steal from
3275           task = __kmp_steal_task(other_thread, gtid, task_team,
3276                                   unfinished_threads, thread_finished,
3277                                   is_constrained);
3278         }
3279         if (task != NULL) { // set last stolen to victim
3280           if (threads_data[tid].td.td_deque_last_stolen != victim_tid) {
3281             threads_data[tid].td.td_deque_last_stolen = victim_tid;
3282             // The pre-refactored code did not try more than 1 successful new
3283             // vicitm, unless the last one generated more local tasks;
3284             // new_victim keeps track of this
3285             new_victim = 1;
3286           }
3287         } else { // No tasks found; unset last_stolen
3288           KMP_CHECK_UPDATE(threads_data[tid].td.td_deque_last_stolen, -1);
3289           victim_tid = -2; // no successful victim found
3290         }
3291       }
3292 
3293       if (task == NULL)
3294         break; // break out of tasking loop
3295 
3296 // Found a task; execute it
3297 #if USE_ITT_BUILD && USE_ITT_NOTIFY
3298       if (__itt_sync_create_ptr || KMP_ITT_DEBUG) {
3299         if (itt_sync_obj == NULL) { // we are at fork barrier where we could not
3300           // get the object reliably
3301           itt_sync_obj = __kmp_itt_barrier_object(gtid, bs_forkjoin_barrier);
3302         }
3303         __kmp_itt_task_starting(itt_sync_obj);
3304       }
3305 #endif /* USE_ITT_BUILD && USE_ITT_NOTIFY */
3306       __kmp_invoke_task(gtid, task, current_task);
3307 #if USE_ITT_BUILD
3308       if (itt_sync_obj != NULL)
3309         __kmp_itt_task_finished(itt_sync_obj);
3310 #endif /* USE_ITT_BUILD */
3311       // If this thread is only partway through the barrier and the condition is
3312       // met, then return now, so that the barrier gather/release pattern can
3313       // proceed. If this thread is in the last spin loop in the barrier,
3314       // waiting to be released, we know that the termination condition will not
3315       // be satisfied, so don't waste any cycles checking it.
3316       if (flag == NULL || (!final_spin && flag->done_check())) {
3317         KA_TRACE(
3318             15,
3319             ("__kmp_execute_tasks_template: T#%d spin condition satisfied\n",
3320              gtid));
3321         return TRUE;
3322       }
3323       if (thread->th.th_task_team == NULL) {
3324         break;
3325       }
3326       KMP_YIELD(__kmp_library == library_throughput); // Yield before next task
3327       // If execution of a stolen task results in more tasks being placed on our
3328       // run queue, reset use_own_tasks
3329       if (!use_own_tasks && TCR_4(threads_data[tid].td.td_deque_ntasks) != 0) {
3330         KA_TRACE(20, ("__kmp_execute_tasks_template: T#%d stolen task spawned "
3331                       "other tasks, restart\n",
3332                       gtid));
3333         use_own_tasks = 1;
3334         new_victim = 0;
3335       }
3336     }
3337 
3338     // The task source has been exhausted. If in final spin loop of barrier,
3339     // check if termination condition is satisfied. The work queue may be empty
3340     // but there might be proxy tasks still executing.
3341     if (final_spin &&
3342         KMP_ATOMIC_LD_ACQ(&current_task->td_incomplete_child_tasks) == 0) {
3343       // First, decrement the #unfinished threads, if that has not already been
3344       // done.  This decrement might be to the spin location, and result in the
3345       // termination condition being satisfied.
3346       if (!*thread_finished) {
3347 #if KMP_DEBUG
3348         kmp_int32 count = -1 +
3349 #endif
3350             KMP_ATOMIC_DEC(unfinished_threads);
3351         KA_TRACE(20, ("__kmp_execute_tasks_template: T#%d dec "
3352                       "unfinished_threads to %d task_team=%p\n",
3353                       gtid, count, task_team));
3354         *thread_finished = TRUE;
3355       }
3356 
3357       // It is now unsafe to reference thread->th.th_team !!!
3358       // Decrementing task_team->tt.tt_unfinished_threads can allow the primary
3359       // thread to pass through the barrier, where it might reset each thread's
3360       // th.th_team field for the next parallel region. If we can steal more
3361       // work, we know that this has not happened yet.
3362       if (flag != NULL && flag->done_check()) {
3363         KA_TRACE(
3364             15,
3365             ("__kmp_execute_tasks_template: T#%d spin condition satisfied\n",
3366              gtid));
3367         return TRUE;
3368       }
3369     }
3370 
3371     // If this thread's task team is NULL, primary thread has recognized that
3372     // there are no more tasks; bail out
3373     if (thread->th.th_task_team == NULL) {
3374       KA_TRACE(15,
3375                ("__kmp_execute_tasks_template: T#%d no more tasks\n", gtid));
3376       return FALSE;
3377     }
3378 
3379     // Check the flag again to see if it has already done in case to be trapped
3380     // into infinite loop when a if0 task depends on a hidden helper task
3381     // outside any parallel region. Detached tasks are not impacted in this case
3382     // because the only thread executing this function has to execute the proxy
3383     // task so it is in another code path that has the same check.
3384     if (flag == NULL || (!final_spin && flag->done_check())) {
3385       KA_TRACE(15,
3386                ("__kmp_execute_tasks_template: T#%d spin condition satisfied\n",
3387                 gtid));
3388       return TRUE;
3389     }
3390 
3391     // We could be getting tasks from target constructs; if this is the only
3392     // thread, keep trying to execute tasks from own queue
3393     if (nthreads == 1 &&
3394         KMP_ATOMIC_LD_ACQ(&current_task->td_incomplete_child_tasks))
3395       use_own_tasks = 1;
3396     else {
3397       KA_TRACE(15,
3398                ("__kmp_execute_tasks_template: T#%d can't find work\n", gtid));
3399       return FALSE;
3400     }
3401   }
3402 }
3403 
3404 template <bool C, bool S>
3405 int __kmp_execute_tasks_32(
3406     kmp_info_t *thread, kmp_int32 gtid, kmp_flag_32<C, S> *flag, int final_spin,
3407     int *thread_finished USE_ITT_BUILD_ARG(void *itt_sync_obj),
3408     kmp_int32 is_constrained) {
3409   return __kmp_execute_tasks_template(
3410       thread, gtid, flag, final_spin,
3411       thread_finished USE_ITT_BUILD_ARG(itt_sync_obj), is_constrained);
3412 }
3413 
3414 template <bool C, bool S>
3415 int __kmp_execute_tasks_64(
3416     kmp_info_t *thread, kmp_int32 gtid, kmp_flag_64<C, S> *flag, int final_spin,
3417     int *thread_finished USE_ITT_BUILD_ARG(void *itt_sync_obj),
3418     kmp_int32 is_constrained) {
3419   return __kmp_execute_tasks_template(
3420       thread, gtid, flag, final_spin,
3421       thread_finished USE_ITT_BUILD_ARG(itt_sync_obj), is_constrained);
3422 }
3423 
3424 template <bool C, bool S>
3425 int __kmp_atomic_execute_tasks_64(
3426     kmp_info_t *thread, kmp_int32 gtid, kmp_atomic_flag_64<C, S> *flag,
3427     int final_spin, int *thread_finished USE_ITT_BUILD_ARG(void *itt_sync_obj),
3428     kmp_int32 is_constrained) {
3429   return __kmp_execute_tasks_template(
3430       thread, gtid, flag, final_spin,
3431       thread_finished USE_ITT_BUILD_ARG(itt_sync_obj), is_constrained);
3432 }
3433 
3434 int __kmp_execute_tasks_oncore(
3435     kmp_info_t *thread, kmp_int32 gtid, kmp_flag_oncore *flag, int final_spin,
3436     int *thread_finished USE_ITT_BUILD_ARG(void *itt_sync_obj),
3437     kmp_int32 is_constrained) {
3438   return __kmp_execute_tasks_template(
3439       thread, gtid, flag, final_spin,
3440       thread_finished USE_ITT_BUILD_ARG(itt_sync_obj), is_constrained);
3441 }
3442 
3443 template int
3444 __kmp_execute_tasks_32<false, false>(kmp_info_t *, kmp_int32,
3445                                      kmp_flag_32<false, false> *, int,
3446                                      int *USE_ITT_BUILD_ARG(void *), kmp_int32);
3447 
3448 template int __kmp_execute_tasks_64<false, true>(kmp_info_t *, kmp_int32,
3449                                                  kmp_flag_64<false, true> *,
3450                                                  int,
3451                                                  int *USE_ITT_BUILD_ARG(void *),
3452                                                  kmp_int32);
3453 
3454 template int __kmp_execute_tasks_64<true, false>(kmp_info_t *, kmp_int32,
3455                                                  kmp_flag_64<true, false> *,
3456                                                  int,
3457                                                  int *USE_ITT_BUILD_ARG(void *),
3458                                                  kmp_int32);
3459 
3460 template int __kmp_atomic_execute_tasks_64<false, true>(
3461     kmp_info_t *, kmp_int32, kmp_atomic_flag_64<false, true> *, int,
3462     int *USE_ITT_BUILD_ARG(void *), kmp_int32);
3463 
3464 template int __kmp_atomic_execute_tasks_64<true, false>(
3465     kmp_info_t *, kmp_int32, kmp_atomic_flag_64<true, false> *, int,
3466     int *USE_ITT_BUILD_ARG(void *), kmp_int32);
3467 
3468 // __kmp_enable_tasking: Allocate task team and resume threads sleeping at the
3469 // next barrier so they can assist in executing enqueued tasks.
3470 // First thread in allocates the task team atomically.
3471 static void __kmp_enable_tasking(kmp_task_team_t *task_team,
3472                                  kmp_info_t *this_thr) {
3473   kmp_thread_data_t *threads_data;
3474   int nthreads, i, is_init_thread;
3475 
3476   KA_TRACE(10, ("__kmp_enable_tasking(enter): T#%d\n",
3477                 __kmp_gtid_from_thread(this_thr)));
3478 
3479   KMP_DEBUG_ASSERT(task_team != NULL);
3480   KMP_DEBUG_ASSERT(this_thr->th.th_team != NULL);
3481 
3482   nthreads = task_team->tt.tt_nproc;
3483   KMP_DEBUG_ASSERT(nthreads > 0);
3484   KMP_DEBUG_ASSERT(nthreads == this_thr->th.th_team->t.t_nproc);
3485 
3486   // Allocate or increase the size of threads_data if necessary
3487   is_init_thread = __kmp_realloc_task_threads_data(this_thr, task_team);
3488 
3489   if (!is_init_thread) {
3490     // Some other thread already set up the array.
3491     KA_TRACE(
3492         20,
3493         ("__kmp_enable_tasking(exit): T#%d: threads array already set up.\n",
3494          __kmp_gtid_from_thread(this_thr)));
3495     return;
3496   }
3497   threads_data = (kmp_thread_data_t *)TCR_PTR(task_team->tt.tt_threads_data);
3498   KMP_DEBUG_ASSERT(threads_data != NULL);
3499 
3500   if (__kmp_tasking_mode == tskm_task_teams &&
3501       (__kmp_dflt_blocktime != KMP_MAX_BLOCKTIME)) {
3502     // Release any threads sleeping at the barrier, so that they can steal
3503     // tasks and execute them.  In extra barrier mode, tasks do not sleep
3504     // at the separate tasking barrier, so this isn't a problem.
3505     for (i = 0; i < nthreads; i++) {
3506       void *sleep_loc;
3507       kmp_info_t *thread = threads_data[i].td.td_thr;
3508 
3509       if (i == this_thr->th.th_info.ds.ds_tid) {
3510         continue;
3511       }
3512       // Since we haven't locked the thread's suspend mutex lock at this
3513       // point, there is a small window where a thread might be putting
3514       // itself to sleep, but hasn't set the th_sleep_loc field yet.
3515       // To work around this, __kmp_execute_tasks_template() periodically checks
3516       // see if other threads are sleeping (using the same random mechanism that
3517       // is used for task stealing) and awakens them if they are.
3518       if ((sleep_loc = TCR_PTR(CCAST(void *, thread->th.th_sleep_loc))) !=
3519           NULL) {
3520         KF_TRACE(50, ("__kmp_enable_tasking: T#%d waking up thread T#%d\n",
3521                       __kmp_gtid_from_thread(this_thr),
3522                       __kmp_gtid_from_thread(thread)));
3523         __kmp_null_resume_wrapper(thread);
3524       } else {
3525         KF_TRACE(50, ("__kmp_enable_tasking: T#%d don't wake up thread T#%d\n",
3526                       __kmp_gtid_from_thread(this_thr),
3527                       __kmp_gtid_from_thread(thread)));
3528       }
3529     }
3530   }
3531 
3532   KA_TRACE(10, ("__kmp_enable_tasking(exit): T#%d\n",
3533                 __kmp_gtid_from_thread(this_thr)));
3534 }
3535 
3536 /* // TODO: Check the comment consistency
3537  * Utility routines for "task teams".  A task team (kmp_task_t) is kind of
3538  * like a shadow of the kmp_team_t data struct, with a different lifetime.
3539  * After a child * thread checks into a barrier and calls __kmp_release() from
3540  * the particular variant of __kmp_<barrier_kind>_barrier_gather(), it can no
3541  * longer assume that the kmp_team_t structure is intact (at any moment, the
3542  * primary thread may exit the barrier code and free the team data structure,
3543  * and return the threads to the thread pool).
3544  *
3545  * This does not work with the tasking code, as the thread is still
3546  * expected to participate in the execution of any tasks that may have been
3547  * spawned my a member of the team, and the thread still needs access to all
3548  * to each thread in the team, so that it can steal work from it.
3549  *
3550  * Enter the existence of the kmp_task_team_t struct.  It employs a reference
3551  * counting mechanism, and is allocated by the primary thread before calling
3552  * __kmp_<barrier_kind>_release, and then is release by the last thread to
3553  * exit __kmp_<barrier_kind>_release at the next barrier.  I.e. the lifetimes
3554  * of the kmp_task_team_t structs for consecutive barriers can overlap
3555  * (and will, unless the primary thread is the last thread to exit the barrier
3556  * release phase, which is not typical). The existence of such a struct is
3557  * useful outside the context of tasking.
3558  *
3559  * We currently use the existence of the threads array as an indicator that
3560  * tasks were spawned since the last barrier.  If the structure is to be
3561  * useful outside the context of tasking, then this will have to change, but
3562  * not setting the field minimizes the performance impact of tasking on
3563  * barriers, when no explicit tasks were spawned (pushed, actually).
3564  */
3565 
3566 static kmp_task_team_t *__kmp_free_task_teams =
3567     NULL; // Free list for task_team data structures
3568 // Lock for task team data structures
3569 kmp_bootstrap_lock_t __kmp_task_team_lock =
3570     KMP_BOOTSTRAP_LOCK_INITIALIZER(__kmp_task_team_lock);
3571 
3572 // __kmp_alloc_task_deque:
3573 // Allocates a task deque for a particular thread, and initialize the necessary
3574 // data structures relating to the deque.  This only happens once per thread
3575 // per task team since task teams are recycled. No lock is needed during
3576 // allocation since each thread allocates its own deque.
3577 static void __kmp_alloc_task_deque(kmp_info_t *thread,
3578                                    kmp_thread_data_t *thread_data) {
3579   __kmp_init_bootstrap_lock(&thread_data->td.td_deque_lock);
3580   KMP_DEBUG_ASSERT(thread_data->td.td_deque == NULL);
3581 
3582   // Initialize last stolen task field to "none"
3583   thread_data->td.td_deque_last_stolen = -1;
3584 
3585   KMP_DEBUG_ASSERT(TCR_4(thread_data->td.td_deque_ntasks) == 0);
3586   KMP_DEBUG_ASSERT(thread_data->td.td_deque_head == 0);
3587   KMP_DEBUG_ASSERT(thread_data->td.td_deque_tail == 0);
3588 
3589   KE_TRACE(
3590       10,
3591       ("__kmp_alloc_task_deque: T#%d allocating deque[%d] for thread_data %p\n",
3592        __kmp_gtid_from_thread(thread), INITIAL_TASK_DEQUE_SIZE, thread_data));
3593   // Allocate space for task deque, and zero the deque
3594   // Cannot use __kmp_thread_calloc() because threads not around for
3595   // kmp_reap_task_team( ).
3596   thread_data->td.td_deque = (kmp_taskdata_t **)__kmp_allocate(
3597       INITIAL_TASK_DEQUE_SIZE * sizeof(kmp_taskdata_t *));
3598   thread_data->td.td_deque_size = INITIAL_TASK_DEQUE_SIZE;
3599 }
3600 
3601 // __kmp_free_task_deque:
3602 // Deallocates a task deque for a particular thread. Happens at library
3603 // deallocation so don't need to reset all thread data fields.
3604 static void __kmp_free_task_deque(kmp_thread_data_t *thread_data) {
3605   if (thread_data->td.td_deque != NULL) {
3606     __kmp_acquire_bootstrap_lock(&thread_data->td.td_deque_lock);
3607     TCW_4(thread_data->td.td_deque_ntasks, 0);
3608     __kmp_free(thread_data->td.td_deque);
3609     thread_data->td.td_deque = NULL;
3610     __kmp_release_bootstrap_lock(&thread_data->td.td_deque_lock);
3611   }
3612 
3613 #ifdef BUILD_TIED_TASK_STACK
3614   // GEH: Figure out what to do here for td_susp_tied_tasks
3615   if (thread_data->td.td_susp_tied_tasks.ts_entries != TASK_STACK_EMPTY) {
3616     __kmp_free_task_stack(__kmp_thread_from_gtid(gtid), thread_data);
3617   }
3618 #endif // BUILD_TIED_TASK_STACK
3619 }
3620 
3621 // __kmp_realloc_task_threads_data:
3622 // Allocates a threads_data array for a task team, either by allocating an
3623 // initial array or enlarging an existing array.  Only the first thread to get
3624 // the lock allocs or enlarges the array and re-initializes the array elements.
3625 // That thread returns "TRUE", the rest return "FALSE".
3626 // Assumes that the new array size is given by task_team -> tt.tt_nproc.
3627 // The current size is given by task_team -> tt.tt_max_threads.
3628 static int __kmp_realloc_task_threads_data(kmp_info_t *thread,
3629                                            kmp_task_team_t *task_team) {
3630   kmp_thread_data_t **threads_data_p;
3631   kmp_int32 nthreads, maxthreads;
3632   int is_init_thread = FALSE;
3633 
3634   if (TCR_4(task_team->tt.tt_found_tasks)) {
3635     // Already reallocated and initialized.
3636     return FALSE;
3637   }
3638 
3639   threads_data_p = &task_team->tt.tt_threads_data;
3640   nthreads = task_team->tt.tt_nproc;
3641   maxthreads = task_team->tt.tt_max_threads;
3642 
3643   // All threads must lock when they encounter the first task of the implicit
3644   // task region to make sure threads_data fields are (re)initialized before
3645   // used.
3646   __kmp_acquire_bootstrap_lock(&task_team->tt.tt_threads_lock);
3647 
3648   if (!TCR_4(task_team->tt.tt_found_tasks)) {
3649     // first thread to enable tasking
3650     kmp_team_t *team = thread->th.th_team;
3651     int i;
3652 
3653     is_init_thread = TRUE;
3654     if (maxthreads < nthreads) {
3655 
3656       if (*threads_data_p != NULL) {
3657         kmp_thread_data_t *old_data = *threads_data_p;
3658         kmp_thread_data_t *new_data = NULL;
3659 
3660         KE_TRACE(
3661             10,
3662             ("__kmp_realloc_task_threads_data: T#%d reallocating "
3663              "threads data for task_team %p, new_size = %d, old_size = %d\n",
3664              __kmp_gtid_from_thread(thread), task_team, nthreads, maxthreads));
3665         // Reallocate threads_data to have more elements than current array
3666         // Cannot use __kmp_thread_realloc() because threads not around for
3667         // kmp_reap_task_team( ).  Note all new array entries are initialized
3668         // to zero by __kmp_allocate().
3669         new_data = (kmp_thread_data_t *)__kmp_allocate(
3670             nthreads * sizeof(kmp_thread_data_t));
3671         // copy old data to new data
3672         KMP_MEMCPY_S((void *)new_data, nthreads * sizeof(kmp_thread_data_t),
3673                      (void *)old_data, maxthreads * sizeof(kmp_thread_data_t));
3674 
3675 #ifdef BUILD_TIED_TASK_STACK
3676         // GEH: Figure out if this is the right thing to do
3677         for (i = maxthreads; i < nthreads; i++) {
3678           kmp_thread_data_t *thread_data = &(*threads_data_p)[i];
3679           __kmp_init_task_stack(__kmp_gtid_from_thread(thread), thread_data);
3680         }
3681 #endif // BUILD_TIED_TASK_STACK
3682        // Install the new data and free the old data
3683         (*threads_data_p) = new_data;
3684         __kmp_free(old_data);
3685       } else {
3686         KE_TRACE(10, ("__kmp_realloc_task_threads_data: T#%d allocating "
3687                       "threads data for task_team %p, size = %d\n",
3688                       __kmp_gtid_from_thread(thread), task_team, nthreads));
3689         // Make the initial allocate for threads_data array, and zero entries
3690         // Cannot use __kmp_thread_calloc() because threads not around for
3691         // kmp_reap_task_team( ).
3692         *threads_data_p = (kmp_thread_data_t *)__kmp_allocate(
3693             nthreads * sizeof(kmp_thread_data_t));
3694 #ifdef BUILD_TIED_TASK_STACK
3695         // GEH: Figure out if this is the right thing to do
3696         for (i = 0; i < nthreads; i++) {
3697           kmp_thread_data_t *thread_data = &(*threads_data_p)[i];
3698           __kmp_init_task_stack(__kmp_gtid_from_thread(thread), thread_data);
3699         }
3700 #endif // BUILD_TIED_TASK_STACK
3701       }
3702       task_team->tt.tt_max_threads = nthreads;
3703     } else {
3704       // If array has (more than) enough elements, go ahead and use it
3705       KMP_DEBUG_ASSERT(*threads_data_p != NULL);
3706     }
3707 
3708     // initialize threads_data pointers back to thread_info structures
3709     for (i = 0; i < nthreads; i++) {
3710       kmp_thread_data_t *thread_data = &(*threads_data_p)[i];
3711       thread_data->td.td_thr = team->t.t_threads[i];
3712 
3713       if (thread_data->td.td_deque_last_stolen >= nthreads) {
3714         // The last stolen field survives across teams / barrier, and the number
3715         // of threads may have changed.  It's possible (likely?) that a new
3716         // parallel region will exhibit the same behavior as previous region.
3717         thread_data->td.td_deque_last_stolen = -1;
3718       }
3719     }
3720 
3721     KMP_MB();
3722     TCW_SYNC_4(task_team->tt.tt_found_tasks, TRUE);
3723   }
3724 
3725   __kmp_release_bootstrap_lock(&task_team->tt.tt_threads_lock);
3726   return is_init_thread;
3727 }
3728 
3729 // __kmp_free_task_threads_data:
3730 // Deallocates a threads_data array for a task team, including any attached
3731 // tasking deques.  Only occurs at library shutdown.
3732 static void __kmp_free_task_threads_data(kmp_task_team_t *task_team) {
3733   __kmp_acquire_bootstrap_lock(&task_team->tt.tt_threads_lock);
3734   if (task_team->tt.tt_threads_data != NULL) {
3735     int i;
3736     for (i = 0; i < task_team->tt.tt_max_threads; i++) {
3737       __kmp_free_task_deque(&task_team->tt.tt_threads_data[i]);
3738     }
3739     __kmp_free(task_team->tt.tt_threads_data);
3740     task_team->tt.tt_threads_data = NULL;
3741   }
3742   __kmp_release_bootstrap_lock(&task_team->tt.tt_threads_lock);
3743 }
3744 
3745 // __kmp_free_task_pri_list:
3746 // Deallocates tasking deques used for priority tasks.
3747 // Only occurs at library shutdown.
3748 static void __kmp_free_task_pri_list(kmp_task_team_t *task_team) {
3749   __kmp_acquire_bootstrap_lock(&task_team->tt.tt_task_pri_lock);
3750   if (task_team->tt.tt_task_pri_list != NULL) {
3751     kmp_task_pri_t *list = task_team->tt.tt_task_pri_list;
3752     while (list != NULL) {
3753       kmp_task_pri_t *next = list->next;
3754       __kmp_free_task_deque(&list->td);
3755       __kmp_free(list);
3756       list = next;
3757     }
3758     task_team->tt.tt_task_pri_list = NULL;
3759   }
3760   __kmp_release_bootstrap_lock(&task_team->tt.tt_task_pri_lock);
3761 }
3762 
3763 // __kmp_allocate_task_team:
3764 // Allocates a task team associated with a specific team, taking it from
3765 // the global task team free list if possible.  Also initializes data
3766 // structures.
3767 static kmp_task_team_t *__kmp_allocate_task_team(kmp_info_t *thread,
3768                                                  kmp_team_t *team) {
3769   kmp_task_team_t *task_team = NULL;
3770   int nthreads;
3771 
3772   KA_TRACE(20, ("__kmp_allocate_task_team: T#%d entering; team = %p\n",
3773                 (thread ? __kmp_gtid_from_thread(thread) : -1), team));
3774 
3775   if (TCR_PTR(__kmp_free_task_teams) != NULL) {
3776     // Take a task team from the task team pool
3777     __kmp_acquire_bootstrap_lock(&__kmp_task_team_lock);
3778     if (__kmp_free_task_teams != NULL) {
3779       task_team = __kmp_free_task_teams;
3780       TCW_PTR(__kmp_free_task_teams, task_team->tt.tt_next);
3781       task_team->tt.tt_next = NULL;
3782     }
3783     __kmp_release_bootstrap_lock(&__kmp_task_team_lock);
3784   }
3785 
3786   if (task_team == NULL) {
3787     KE_TRACE(10, ("__kmp_allocate_task_team: T#%d allocating "
3788                   "task team for team %p\n",
3789                   __kmp_gtid_from_thread(thread), team));
3790     // Allocate a new task team if one is not available. Cannot use
3791     // __kmp_thread_malloc because threads not around for kmp_reap_task_team.
3792     task_team = (kmp_task_team_t *)__kmp_allocate(sizeof(kmp_task_team_t));
3793     __kmp_init_bootstrap_lock(&task_team->tt.tt_threads_lock);
3794     __kmp_init_bootstrap_lock(&task_team->tt.tt_task_pri_lock);
3795 #if USE_ITT_BUILD && USE_ITT_NOTIFY && KMP_DEBUG
3796     // suppress race conditions detection on synchronization flags in debug mode
3797     // this helps to analyze library internals eliminating false positives
3798     __itt_suppress_mark_range(
3799         __itt_suppress_range, __itt_suppress_threading_errors,
3800         &task_team->tt.tt_found_tasks, sizeof(task_team->tt.tt_found_tasks));
3801     __itt_suppress_mark_range(__itt_suppress_range,
3802                               __itt_suppress_threading_errors,
3803                               CCAST(kmp_uint32 *, &task_team->tt.tt_active),
3804                               sizeof(task_team->tt.tt_active));
3805 #endif /* USE_ITT_BUILD && USE_ITT_NOTIFY && KMP_DEBUG */
3806     // Note: __kmp_allocate zeroes returned memory, othewise we would need:
3807     // task_team->tt.tt_threads_data = NULL;
3808     // task_team->tt.tt_max_threads = 0;
3809     // task_team->tt.tt_next = NULL;
3810   }
3811 
3812   TCW_4(task_team->tt.tt_found_tasks, FALSE);
3813   TCW_4(task_team->tt.tt_found_proxy_tasks, FALSE);
3814   TCW_4(task_team->tt.tt_hidden_helper_task_encountered, FALSE);
3815   task_team->tt.tt_nproc = nthreads = team->t.t_nproc;
3816 
3817   KMP_ATOMIC_ST_REL(&task_team->tt.tt_unfinished_threads, nthreads);
3818   TCW_4(task_team->tt.tt_hidden_helper_task_encountered, FALSE);
3819   TCW_4(task_team->tt.tt_active, TRUE);
3820 
3821   KA_TRACE(20, ("__kmp_allocate_task_team: T#%d exiting; task_team = %p "
3822                 "unfinished_threads init'd to %d\n",
3823                 (thread ? __kmp_gtid_from_thread(thread) : -1), task_team,
3824                 KMP_ATOMIC_LD_RLX(&task_team->tt.tt_unfinished_threads)));
3825   return task_team;
3826 }
3827 
3828 // __kmp_free_task_team:
3829 // Frees the task team associated with a specific thread, and adds it
3830 // to the global task team free list.
3831 void __kmp_free_task_team(kmp_info_t *thread, kmp_task_team_t *task_team) {
3832   KA_TRACE(20, ("__kmp_free_task_team: T#%d task_team = %p\n",
3833                 thread ? __kmp_gtid_from_thread(thread) : -1, task_team));
3834 
3835   // Put task team back on free list
3836   __kmp_acquire_bootstrap_lock(&__kmp_task_team_lock);
3837 
3838   KMP_DEBUG_ASSERT(task_team->tt.tt_next == NULL);
3839   task_team->tt.tt_next = __kmp_free_task_teams;
3840   TCW_PTR(__kmp_free_task_teams, task_team);
3841 
3842   __kmp_release_bootstrap_lock(&__kmp_task_team_lock);
3843 }
3844 
3845 // __kmp_reap_task_teams:
3846 // Free all the task teams on the task team free list.
3847 // Should only be done during library shutdown.
3848 // Cannot do anything that needs a thread structure or gtid since they are
3849 // already gone.
3850 void __kmp_reap_task_teams(void) {
3851   kmp_task_team_t *task_team;
3852 
3853   if (TCR_PTR(__kmp_free_task_teams) != NULL) {
3854     // Free all task_teams on the free list
3855     __kmp_acquire_bootstrap_lock(&__kmp_task_team_lock);
3856     while ((task_team = __kmp_free_task_teams) != NULL) {
3857       __kmp_free_task_teams = task_team->tt.tt_next;
3858       task_team->tt.tt_next = NULL;
3859 
3860       // Free threads_data if necessary
3861       if (task_team->tt.tt_threads_data != NULL) {
3862         __kmp_free_task_threads_data(task_team);
3863       }
3864       if (task_team->tt.tt_task_pri_list != NULL) {
3865         __kmp_free_task_pri_list(task_team);
3866       }
3867       __kmp_free(task_team);
3868     }
3869     __kmp_release_bootstrap_lock(&__kmp_task_team_lock);
3870   }
3871 }
3872 
3873 // __kmp_wait_to_unref_task_teams:
3874 // Some threads could still be in the fork barrier release code, possibly
3875 // trying to steal tasks.  Wait for each thread to unreference its task team.
3876 void __kmp_wait_to_unref_task_teams(void) {
3877   kmp_info_t *thread;
3878   kmp_uint32 spins;
3879   kmp_uint64 time;
3880   int done;
3881 
3882   KMP_INIT_YIELD(spins);
3883   KMP_INIT_BACKOFF(time);
3884 
3885   for (;;) {
3886     done = TRUE;
3887 
3888     // TODO: GEH - this may be is wrong because some sync would be necessary
3889     // in case threads are added to the pool during the traversal. Need to
3890     // verify that lock for thread pool is held when calling this routine.
3891     for (thread = CCAST(kmp_info_t *, __kmp_thread_pool); thread != NULL;
3892          thread = thread->th.th_next_pool) {
3893 #if KMP_OS_WINDOWS
3894       DWORD exit_val;
3895 #endif
3896       if (TCR_PTR(thread->th.th_task_team) == NULL) {
3897         KA_TRACE(10, ("__kmp_wait_to_unref_task_team: T#%d task_team == NULL\n",
3898                       __kmp_gtid_from_thread(thread)));
3899         continue;
3900       }
3901 #if KMP_OS_WINDOWS
3902       // TODO: GEH - add this check for Linux* OS / OS X* as well?
3903       if (!__kmp_is_thread_alive(thread, &exit_val)) {
3904         thread->th.th_task_team = NULL;
3905         continue;
3906       }
3907 #endif
3908 
3909       done = FALSE; // Because th_task_team pointer is not NULL for this thread
3910 
3911       KA_TRACE(10, ("__kmp_wait_to_unref_task_team: Waiting for T#%d to "
3912                     "unreference task_team\n",
3913                     __kmp_gtid_from_thread(thread)));
3914 
3915       if (__kmp_dflt_blocktime != KMP_MAX_BLOCKTIME) {
3916         void *sleep_loc;
3917         // If the thread is sleeping, awaken it.
3918         if ((sleep_loc = TCR_PTR(CCAST(void *, thread->th.th_sleep_loc))) !=
3919             NULL) {
3920           KA_TRACE(
3921               10,
3922               ("__kmp_wait_to_unref_task_team: T#%d waking up thread T#%d\n",
3923                __kmp_gtid_from_thread(thread), __kmp_gtid_from_thread(thread)));
3924           __kmp_null_resume_wrapper(thread);
3925         }
3926       }
3927     }
3928     if (done) {
3929       break;
3930     }
3931 
3932     // If oversubscribed or have waited a bit, yield.
3933     KMP_YIELD_OVERSUB_ELSE_SPIN(spins, time);
3934   }
3935 }
3936 
3937 // __kmp_task_team_setup:  Create a task_team for the current team, but use
3938 // an already created, unused one if it already exists.
3939 void __kmp_task_team_setup(kmp_info_t *this_thr, kmp_team_t *team, int always) {
3940   KMP_DEBUG_ASSERT(__kmp_tasking_mode != tskm_immediate_exec);
3941 
3942   // If this task_team hasn't been created yet, allocate it. It will be used in
3943   // the region after the next.
3944   // If it exists, it is the current task team and shouldn't be touched yet as
3945   // it may still be in use.
3946   if (team->t.t_task_team[this_thr->th.th_task_state] == NULL &&
3947       (always || team->t.t_nproc > 1)) {
3948     team->t.t_task_team[this_thr->th.th_task_state] =
3949         __kmp_allocate_task_team(this_thr, team);
3950     KA_TRACE(20, ("__kmp_task_team_setup: Primary T#%d created new task_team %p"
3951                   " for team %d at parity=%d\n",
3952                   __kmp_gtid_from_thread(this_thr),
3953                   team->t.t_task_team[this_thr->th.th_task_state], team->t.t_id,
3954                   this_thr->th.th_task_state));
3955   }
3956 
3957   // After threads exit the release, they will call sync, and then point to this
3958   // other task_team; make sure it is allocated and properly initialized. As
3959   // threads spin in the barrier release phase, they will continue to use the
3960   // previous task_team struct(above), until they receive the signal to stop
3961   // checking for tasks (they can't safely reference the kmp_team_t struct,
3962   // which could be reallocated by the primary thread). No task teams are formed
3963   // for serialized teams.
3964   if (team->t.t_nproc > 1) {
3965     int other_team = 1 - this_thr->th.th_task_state;
3966     KMP_DEBUG_ASSERT(other_team >= 0 && other_team < 2);
3967     if (team->t.t_task_team[other_team] == NULL) { // setup other team as well
3968       team->t.t_task_team[other_team] =
3969           __kmp_allocate_task_team(this_thr, team);
3970       KA_TRACE(20, ("__kmp_task_team_setup: Primary T#%d created second new "
3971                     "task_team %p for team %d at parity=%d\n",
3972                     __kmp_gtid_from_thread(this_thr),
3973                     team->t.t_task_team[other_team], team->t.t_id, other_team));
3974     } else { // Leave the old task team struct in place for the upcoming region;
3975       // adjust as needed
3976       kmp_task_team_t *task_team = team->t.t_task_team[other_team];
3977       if (!task_team->tt.tt_active ||
3978           team->t.t_nproc != task_team->tt.tt_nproc) {
3979         TCW_4(task_team->tt.tt_nproc, team->t.t_nproc);
3980         TCW_4(task_team->tt.tt_found_tasks, FALSE);
3981         TCW_4(task_team->tt.tt_found_proxy_tasks, FALSE);
3982         TCW_4(task_team->tt.tt_hidden_helper_task_encountered, FALSE);
3983         KMP_ATOMIC_ST_REL(&task_team->tt.tt_unfinished_threads,
3984                           team->t.t_nproc);
3985         TCW_4(task_team->tt.tt_active, TRUE);
3986       }
3987       // if team size has changed, the first thread to enable tasking will
3988       // realloc threads_data if necessary
3989       KA_TRACE(20, ("__kmp_task_team_setup: Primary T#%d reset next task_team "
3990                     "%p for team %d at parity=%d\n",
3991                     __kmp_gtid_from_thread(this_thr),
3992                     team->t.t_task_team[other_team], team->t.t_id, other_team));
3993     }
3994   }
3995 
3996   // For regular thread, task enabling should be called when the task is going
3997   // to be pushed to a dequeue. However, for the hidden helper thread, we need
3998   // it ahead of time so that some operations can be performed without race
3999   // condition.
4000   if (this_thr == __kmp_hidden_helper_main_thread) {
4001     for (int i = 0; i < 2; ++i) {
4002       kmp_task_team_t *task_team = team->t.t_task_team[i];
4003       if (KMP_TASKING_ENABLED(task_team)) {
4004         continue;
4005       }
4006       __kmp_enable_tasking(task_team, this_thr);
4007       for (int j = 0; j < task_team->tt.tt_nproc; ++j) {
4008         kmp_thread_data_t *thread_data = &task_team->tt.tt_threads_data[j];
4009         if (thread_data->td.td_deque == NULL) {
4010           __kmp_alloc_task_deque(__kmp_hidden_helper_threads[j], thread_data);
4011         }
4012       }
4013     }
4014   }
4015 }
4016 
4017 // __kmp_task_team_sync: Propagation of task team data from team to threads
4018 // which happens just after the release phase of a team barrier.  This may be
4019 // called by any thread, but only for teams with # threads > 1.
4020 void __kmp_task_team_sync(kmp_info_t *this_thr, kmp_team_t *team) {
4021   KMP_DEBUG_ASSERT(__kmp_tasking_mode != tskm_immediate_exec);
4022 
4023   // Toggle the th_task_state field, to switch which task_team this thread
4024   // refers to
4025   this_thr->th.th_task_state = (kmp_uint8)(1 - this_thr->th.th_task_state);
4026 
4027   // It is now safe to propagate the task team pointer from the team struct to
4028   // the current thread.
4029   TCW_PTR(this_thr->th.th_task_team,
4030           team->t.t_task_team[this_thr->th.th_task_state]);
4031   KA_TRACE(20,
4032            ("__kmp_task_team_sync: Thread T#%d task team switched to task_team "
4033             "%p from Team #%d (parity=%d)\n",
4034             __kmp_gtid_from_thread(this_thr), this_thr->th.th_task_team,
4035             team->t.t_id, this_thr->th.th_task_state));
4036 }
4037 
4038 // __kmp_task_team_wait: Primary thread waits for outstanding tasks after the
4039 // barrier gather phase. Only called by primary thread if #threads in team > 1
4040 // or if proxy tasks were created.
4041 //
4042 // wait is a flag that defaults to 1 (see kmp.h), but waiting can be turned off
4043 // by passing in 0 optionally as the last argument. When wait is zero, primary
4044 // thread does not wait for unfinished_threads to reach 0.
4045 void __kmp_task_team_wait(
4046     kmp_info_t *this_thr,
4047     kmp_team_t *team USE_ITT_BUILD_ARG(void *itt_sync_obj), int wait) {
4048   kmp_task_team_t *task_team = team->t.t_task_team[this_thr->th.th_task_state];
4049 
4050   KMP_DEBUG_ASSERT(__kmp_tasking_mode != tskm_immediate_exec);
4051   KMP_DEBUG_ASSERT(task_team == this_thr->th.th_task_team);
4052 
4053   if ((task_team != NULL) && KMP_TASKING_ENABLED(task_team)) {
4054     if (wait) {
4055       KA_TRACE(20, ("__kmp_task_team_wait: Primary T#%d waiting for all tasks "
4056                     "(for unfinished_threads to reach 0) on task_team = %p\n",
4057                     __kmp_gtid_from_thread(this_thr), task_team));
4058       // Worker threads may have dropped through to release phase, but could
4059       // still be executing tasks. Wait here for tasks to complete. To avoid
4060       // memory contention, only primary thread checks termination condition.
4061       kmp_flag_32<false, false> flag(
4062           RCAST(std::atomic<kmp_uint32> *,
4063                 &task_team->tt.tt_unfinished_threads),
4064           0U);
4065       flag.wait(this_thr, TRUE USE_ITT_BUILD_ARG(itt_sync_obj));
4066     }
4067     // Deactivate the old task team, so that the worker threads will stop
4068     // referencing it while spinning.
4069     KA_TRACE(
4070         20,
4071         ("__kmp_task_team_wait: Primary T#%d deactivating task_team %p: "
4072          "setting active to false, setting local and team's pointer to NULL\n",
4073          __kmp_gtid_from_thread(this_thr), task_team));
4074     KMP_DEBUG_ASSERT(task_team->tt.tt_nproc > 1 ||
4075                      task_team->tt.tt_found_proxy_tasks == TRUE ||
4076                      task_team->tt.tt_hidden_helper_task_encountered == TRUE);
4077     TCW_SYNC_4(task_team->tt.tt_found_proxy_tasks, FALSE);
4078     TCW_SYNC_4(task_team->tt.tt_hidden_helper_task_encountered, FALSE);
4079     KMP_CHECK_UPDATE(task_team->tt.tt_untied_task_encountered, 0);
4080     TCW_SYNC_4(task_team->tt.tt_active, FALSE);
4081     KMP_MB();
4082 
4083     TCW_PTR(this_thr->th.th_task_team, NULL);
4084   }
4085 }
4086 
4087 // __kmp_tasking_barrier:
4088 // This routine is called only when __kmp_tasking_mode == tskm_extra_barrier.
4089 // Internal function to execute all tasks prior to a regular barrier or a join
4090 // barrier. It is a full barrier itself, which unfortunately turns regular
4091 // barriers into double barriers and join barriers into 1 1/2 barriers.
4092 void __kmp_tasking_barrier(kmp_team_t *team, kmp_info_t *thread, int gtid) {
4093   std::atomic<kmp_uint32> *spin = RCAST(
4094       std::atomic<kmp_uint32> *,
4095       &team->t.t_task_team[thread->th.th_task_state]->tt.tt_unfinished_threads);
4096   int flag = FALSE;
4097   KMP_DEBUG_ASSERT(__kmp_tasking_mode == tskm_extra_barrier);
4098 
4099 #if USE_ITT_BUILD
4100   KMP_FSYNC_SPIN_INIT(spin, NULL);
4101 #endif /* USE_ITT_BUILD */
4102   kmp_flag_32<false, false> spin_flag(spin, 0U);
4103   while (!spin_flag.execute_tasks(thread, gtid, TRUE,
4104                                   &flag USE_ITT_BUILD_ARG(NULL), 0)) {
4105 #if USE_ITT_BUILD
4106     // TODO: What about itt_sync_obj??
4107     KMP_FSYNC_SPIN_PREPARE(RCAST(void *, spin));
4108 #endif /* USE_ITT_BUILD */
4109 
4110     if (TCR_4(__kmp_global.g.g_done)) {
4111       if (__kmp_global.g.g_abort)
4112         __kmp_abort_thread();
4113       break;
4114     }
4115     KMP_YIELD(TRUE);
4116   }
4117 #if USE_ITT_BUILD
4118   KMP_FSYNC_SPIN_ACQUIRED(RCAST(void *, spin));
4119 #endif /* USE_ITT_BUILD */
4120 }
4121 
4122 // __kmp_give_task puts a task into a given thread queue if:
4123 //  - the queue for that thread was created
4124 //  - there's space in that queue
4125 // Because of this, __kmp_push_task needs to check if there's space after
4126 // getting the lock
4127 static bool __kmp_give_task(kmp_info_t *thread, kmp_int32 tid, kmp_task_t *task,
4128                             kmp_int32 pass) {
4129   kmp_taskdata_t *taskdata = KMP_TASK_TO_TASKDATA(task);
4130   kmp_task_team_t *task_team = taskdata->td_task_team;
4131 
4132   KA_TRACE(20, ("__kmp_give_task: trying to give task %p to thread %d.\n",
4133                 taskdata, tid));
4134 
4135   // If task_team is NULL something went really bad...
4136   KMP_DEBUG_ASSERT(task_team != NULL);
4137 
4138   bool result = false;
4139   kmp_thread_data_t *thread_data = &task_team->tt.tt_threads_data[tid];
4140 
4141   if (thread_data->td.td_deque == NULL) {
4142     // There's no queue in this thread, go find another one
4143     // We're guaranteed that at least one thread has a queue
4144     KA_TRACE(30,
4145              ("__kmp_give_task: thread %d has no queue while giving task %p.\n",
4146               tid, taskdata));
4147     return result;
4148   }
4149 
4150   if (TCR_4(thread_data->td.td_deque_ntasks) >=
4151       TASK_DEQUE_SIZE(thread_data->td)) {
4152     KA_TRACE(
4153         30,
4154         ("__kmp_give_task: queue is full while giving task %p to thread %d.\n",
4155          taskdata, tid));
4156 
4157     // if this deque is bigger than the pass ratio give a chance to another
4158     // thread
4159     if (TASK_DEQUE_SIZE(thread_data->td) / INITIAL_TASK_DEQUE_SIZE >= pass)
4160       return result;
4161 
4162     __kmp_acquire_bootstrap_lock(&thread_data->td.td_deque_lock);
4163     if (TCR_4(thread_data->td.td_deque_ntasks) >=
4164         TASK_DEQUE_SIZE(thread_data->td)) {
4165       // expand deque to push the task which is not allowed to execute
4166       __kmp_realloc_task_deque(thread, thread_data);
4167     }
4168 
4169   } else {
4170 
4171     __kmp_acquire_bootstrap_lock(&thread_data->td.td_deque_lock);
4172 
4173     if (TCR_4(thread_data->td.td_deque_ntasks) >=
4174         TASK_DEQUE_SIZE(thread_data->td)) {
4175       KA_TRACE(30, ("__kmp_give_task: queue is full while giving task %p to "
4176                     "thread %d.\n",
4177                     taskdata, tid));
4178 
4179       // if this deque is bigger than the pass ratio give a chance to another
4180       // thread
4181       if (TASK_DEQUE_SIZE(thread_data->td) / INITIAL_TASK_DEQUE_SIZE >= pass)
4182         goto release_and_exit;
4183 
4184       __kmp_realloc_task_deque(thread, thread_data);
4185     }
4186   }
4187 
4188   // lock is held here, and there is space in the deque
4189 
4190   thread_data->td.td_deque[thread_data->td.td_deque_tail] = taskdata;
4191   // Wrap index.
4192   thread_data->td.td_deque_tail =
4193       (thread_data->td.td_deque_tail + 1) & TASK_DEQUE_MASK(thread_data->td);
4194   TCW_4(thread_data->td.td_deque_ntasks,
4195         TCR_4(thread_data->td.td_deque_ntasks) + 1);
4196 
4197   result = true;
4198   KA_TRACE(30, ("__kmp_give_task: successfully gave task %p to thread %d.\n",
4199                 taskdata, tid));
4200 
4201 release_and_exit:
4202   __kmp_release_bootstrap_lock(&thread_data->td.td_deque_lock);
4203 
4204   return result;
4205 }
4206 
4207 #define PROXY_TASK_FLAG 0x40000000
4208 /* The finish of the proxy tasks is divided in two pieces:
4209     - the top half is the one that can be done from a thread outside the team
4210     - the bottom half must be run from a thread within the team
4211 
4212    In order to run the bottom half the task gets queued back into one of the
4213    threads of the team. Once the td_incomplete_child_task counter of the parent
4214    is decremented the threads can leave the barriers. So, the bottom half needs
4215    to be queued before the counter is decremented. The top half is therefore
4216    divided in two parts:
4217     - things that can be run before queuing the bottom half
4218     - things that must be run after queuing the bottom half
4219 
4220    This creates a second race as the bottom half can free the task before the
4221    second top half is executed. To avoid this we use the
4222    td_incomplete_child_task of the proxy task to synchronize the top and bottom
4223    half. */
4224 static void __kmp_first_top_half_finish_proxy(kmp_taskdata_t *taskdata) {
4225   KMP_DEBUG_ASSERT(taskdata->td_flags.tasktype == TASK_EXPLICIT);
4226   KMP_DEBUG_ASSERT(taskdata->td_flags.proxy == TASK_PROXY);
4227   KMP_DEBUG_ASSERT(taskdata->td_flags.complete == 0);
4228   KMP_DEBUG_ASSERT(taskdata->td_flags.freed == 0);
4229 
4230   taskdata->td_flags.complete = 1; // mark the task as completed
4231 
4232   if (taskdata->td_taskgroup)
4233     KMP_ATOMIC_DEC(&taskdata->td_taskgroup->count);
4234 
4235   // Create an imaginary children for this task so the bottom half cannot
4236   // release the task before we have completed the second top half
4237   KMP_ATOMIC_OR(&taskdata->td_incomplete_child_tasks, PROXY_TASK_FLAG);
4238 }
4239 
4240 static void __kmp_second_top_half_finish_proxy(kmp_taskdata_t *taskdata) {
4241 #if KMP_DEBUG
4242   kmp_int32 children = 0;
4243   // Predecrement simulated by "- 1" calculation
4244   children = -1 +
4245 #endif
4246       KMP_ATOMIC_DEC(&taskdata->td_parent->td_incomplete_child_tasks);
4247   KMP_DEBUG_ASSERT(children >= 0);
4248 
4249   // Remove the imaginary children
4250   KMP_ATOMIC_AND(&taskdata->td_incomplete_child_tasks, ~PROXY_TASK_FLAG);
4251 }
4252 
4253 static void __kmp_bottom_half_finish_proxy(kmp_int32 gtid, kmp_task_t *ptask) {
4254   kmp_taskdata_t *taskdata = KMP_TASK_TO_TASKDATA(ptask);
4255   kmp_info_t *thread = __kmp_threads[gtid];
4256 
4257   KMP_DEBUG_ASSERT(taskdata->td_flags.proxy == TASK_PROXY);
4258   KMP_DEBUG_ASSERT(taskdata->td_flags.complete ==
4259                    1); // top half must run before bottom half
4260 
4261   // We need to wait to make sure the top half is finished
4262   // Spinning here should be ok as this should happen quickly
4263   while ((KMP_ATOMIC_LD_ACQ(&taskdata->td_incomplete_child_tasks) &
4264           PROXY_TASK_FLAG) > 0)
4265     ;
4266 
4267   __kmp_release_deps(gtid, taskdata);
4268   __kmp_free_task_and_ancestors(gtid, taskdata, thread);
4269 }
4270 
4271 /*!
4272 @ingroup TASKING
4273 @param gtid Global Thread ID of encountering thread
4274 @param ptask Task which execution is completed
4275 
4276 Execute the completion of a proxy task from a thread of that is part of the
4277 team. Run first and bottom halves directly.
4278 */
4279 void __kmpc_proxy_task_completed(kmp_int32 gtid, kmp_task_t *ptask) {
4280   KMP_DEBUG_ASSERT(ptask != NULL);
4281   kmp_taskdata_t *taskdata = KMP_TASK_TO_TASKDATA(ptask);
4282   KA_TRACE(
4283       10, ("__kmp_proxy_task_completed(enter): T#%d proxy task %p completing\n",
4284            gtid, taskdata));
4285   __kmp_assert_valid_gtid(gtid);
4286   KMP_DEBUG_ASSERT(taskdata->td_flags.proxy == TASK_PROXY);
4287 
4288   __kmp_first_top_half_finish_proxy(taskdata);
4289   __kmp_second_top_half_finish_proxy(taskdata);
4290   __kmp_bottom_half_finish_proxy(gtid, ptask);
4291 
4292   KA_TRACE(10,
4293            ("__kmp_proxy_task_completed(exit): T#%d proxy task %p completing\n",
4294             gtid, taskdata));
4295 }
4296 
4297 void __kmpc_give_task(kmp_task_t *ptask, kmp_int32 start = 0) {
4298   KMP_DEBUG_ASSERT(ptask != NULL);
4299   kmp_taskdata_t *taskdata = KMP_TASK_TO_TASKDATA(ptask);
4300 
4301   // Enqueue task to complete bottom half completion from a thread within the
4302   // corresponding team
4303   kmp_team_t *team = taskdata->td_team;
4304   kmp_int32 nthreads = team->t.t_nproc;
4305   kmp_info_t *thread;
4306 
4307   // This should be similar to start_k = __kmp_get_random( thread ) % nthreads
4308   // but we cannot use __kmp_get_random here
4309   kmp_int32 start_k = start % nthreads;
4310   kmp_int32 pass = 1;
4311   kmp_int32 k = start_k;
4312 
4313   do {
4314     // For now we're just linearly trying to find a thread
4315     thread = team->t.t_threads[k];
4316     k = (k + 1) % nthreads;
4317 
4318     // we did a full pass through all the threads
4319     if (k == start_k)
4320       pass = pass << 1;
4321 
4322   } while (!__kmp_give_task(thread, k, ptask, pass));
4323 
4324   if (__kmp_dflt_blocktime != KMP_MAX_BLOCKTIME && __kmp_wpolicy_passive) {
4325     // awake at least one thread to execute given task
4326     for (int i = 0; i < nthreads; ++i) {
4327       thread = team->t.t_threads[i];
4328       if (thread->th.th_sleep_loc != NULL) {
4329         __kmp_null_resume_wrapper(thread);
4330         break;
4331       }
4332     }
4333   }
4334 }
4335 
4336 /*!
4337 @ingroup TASKING
4338 @param ptask Task which execution is completed
4339 
4340 Execute the completion of a proxy task from a thread that could not belong to
4341 the team.
4342 */
4343 void __kmpc_proxy_task_completed_ooo(kmp_task_t *ptask) {
4344   KMP_DEBUG_ASSERT(ptask != NULL);
4345   kmp_taskdata_t *taskdata = KMP_TASK_TO_TASKDATA(ptask);
4346 
4347   KA_TRACE(
4348       10,
4349       ("__kmp_proxy_task_completed_ooo(enter): proxy task completing ooo %p\n",
4350        taskdata));
4351 
4352   KMP_DEBUG_ASSERT(taskdata->td_flags.proxy == TASK_PROXY);
4353 
4354   __kmp_first_top_half_finish_proxy(taskdata);
4355 
4356   __kmpc_give_task(ptask);
4357 
4358   __kmp_second_top_half_finish_proxy(taskdata);
4359 
4360   KA_TRACE(
4361       10,
4362       ("__kmp_proxy_task_completed_ooo(exit): proxy task completing ooo %p\n",
4363        taskdata));
4364 }
4365 
4366 kmp_event_t *__kmpc_task_allow_completion_event(ident_t *loc_ref, int gtid,
4367                                                 kmp_task_t *task) {
4368   kmp_taskdata_t *td = KMP_TASK_TO_TASKDATA(task);
4369   if (td->td_allow_completion_event.type == KMP_EVENT_UNINITIALIZED) {
4370     td->td_allow_completion_event.type = KMP_EVENT_ALLOW_COMPLETION;
4371     td->td_allow_completion_event.ed.task = task;
4372     __kmp_init_tas_lock(&td->td_allow_completion_event.lock);
4373   }
4374   return &td->td_allow_completion_event;
4375 }
4376 
4377 void __kmp_fulfill_event(kmp_event_t *event) {
4378   if (event->type == KMP_EVENT_ALLOW_COMPLETION) {
4379     kmp_task_t *ptask = event->ed.task;
4380     kmp_taskdata_t *taskdata = KMP_TASK_TO_TASKDATA(ptask);
4381     bool detached = false;
4382     int gtid = __kmp_get_gtid();
4383 
4384     // The associated task might have completed or could be completing at this
4385     // point.
4386     // We need to take the lock to avoid races
4387     __kmp_acquire_tas_lock(&event->lock, gtid);
4388     if (taskdata->td_flags.proxy == TASK_PROXY) {
4389       detached = true;
4390     } else {
4391 #if OMPT_SUPPORT
4392       // The OMPT event must occur under mutual exclusion,
4393       // otherwise the tool might access ptask after free
4394       if (UNLIKELY(ompt_enabled.enabled))
4395         __ompt_task_finish(ptask, NULL, ompt_task_early_fulfill);
4396 #endif
4397     }
4398     event->type = KMP_EVENT_UNINITIALIZED;
4399     __kmp_release_tas_lock(&event->lock, gtid);
4400 
4401     if (detached) {
4402 #if OMPT_SUPPORT
4403       // We free ptask afterwards and know the task is finished,
4404       // so locking is not necessary
4405       if (UNLIKELY(ompt_enabled.enabled))
4406         __ompt_task_finish(ptask, NULL, ompt_task_late_fulfill);
4407 #endif
4408       // If the task detached complete the proxy task
4409       if (gtid >= 0) {
4410         kmp_team_t *team = taskdata->td_team;
4411         kmp_info_t *thread = __kmp_get_thread();
4412         if (thread->th.th_team == team) {
4413           __kmpc_proxy_task_completed(gtid, ptask);
4414           return;
4415         }
4416       }
4417 
4418       // fallback
4419       __kmpc_proxy_task_completed_ooo(ptask);
4420     }
4421   }
4422 }
4423 
4424 // __kmp_task_dup_alloc: Allocate the taskdata and make a copy of source task
4425 // for taskloop
4426 //
4427 // thread:   allocating thread
4428 // task_src: pointer to source task to be duplicated
4429 // returns:  a pointer to the allocated kmp_task_t structure (task).
4430 kmp_task_t *__kmp_task_dup_alloc(kmp_info_t *thread, kmp_task_t *task_src) {
4431   kmp_task_t *task;
4432   kmp_taskdata_t *taskdata;
4433   kmp_taskdata_t *taskdata_src = KMP_TASK_TO_TASKDATA(task_src);
4434   kmp_taskdata_t *parent_task = taskdata_src->td_parent; // same parent task
4435   size_t shareds_offset;
4436   size_t task_size;
4437 
4438   KA_TRACE(10, ("__kmp_task_dup_alloc(enter): Th %p, source task %p\n", thread,
4439                 task_src));
4440   KMP_DEBUG_ASSERT(taskdata_src->td_flags.proxy ==
4441                    TASK_FULL); // it should not be proxy task
4442   KMP_DEBUG_ASSERT(taskdata_src->td_flags.tasktype == TASK_EXPLICIT);
4443   task_size = taskdata_src->td_size_alloc;
4444 
4445   // Allocate a kmp_taskdata_t block and a kmp_task_t block.
4446   KA_TRACE(30, ("__kmp_task_dup_alloc: Th %p, malloc size %ld\n", thread,
4447                 task_size));
4448 #if USE_FAST_MEMORY
4449   taskdata = (kmp_taskdata_t *)__kmp_fast_allocate(thread, task_size);
4450 #else
4451   taskdata = (kmp_taskdata_t *)__kmp_thread_malloc(thread, task_size);
4452 #endif /* USE_FAST_MEMORY */
4453   KMP_MEMCPY(taskdata, taskdata_src, task_size);
4454 
4455   task = KMP_TASKDATA_TO_TASK(taskdata);
4456 
4457   // Initialize new task (only specific fields not affected by memcpy)
4458   taskdata->td_task_id = KMP_GEN_TASK_ID();
4459   if (task->shareds != NULL) { // need setup shareds pointer
4460     shareds_offset = (char *)task_src->shareds - (char *)taskdata_src;
4461     task->shareds = &((char *)taskdata)[shareds_offset];
4462     KMP_DEBUG_ASSERT((((kmp_uintptr_t)task->shareds) & (sizeof(void *) - 1)) ==
4463                      0);
4464   }
4465   taskdata->td_alloc_thread = thread;
4466   taskdata->td_parent = parent_task;
4467   // task inherits the taskgroup from the parent task
4468   taskdata->td_taskgroup = parent_task->td_taskgroup;
4469   // tied task needs to initialize the td_last_tied at creation,
4470   // untied one does this when it is scheduled for execution
4471   if (taskdata->td_flags.tiedness == TASK_TIED)
4472     taskdata->td_last_tied = taskdata;
4473 
4474   // Only need to keep track of child task counts if team parallel and tasking
4475   // not serialized
4476   if (!(taskdata->td_flags.team_serial || taskdata->td_flags.tasking_ser)) {
4477     KMP_ATOMIC_INC(&parent_task->td_incomplete_child_tasks);
4478     if (parent_task->td_taskgroup)
4479       KMP_ATOMIC_INC(&parent_task->td_taskgroup->count);
4480     // Only need to keep track of allocated child tasks for explicit tasks since
4481     // implicit not deallocated
4482     if (taskdata->td_parent->td_flags.tasktype == TASK_EXPLICIT)
4483       KMP_ATOMIC_INC(&taskdata->td_parent->td_allocated_child_tasks);
4484   }
4485 
4486   KA_TRACE(20,
4487            ("__kmp_task_dup_alloc(exit): Th %p, created task %p, parent=%p\n",
4488             thread, taskdata, taskdata->td_parent));
4489 #if OMPT_SUPPORT
4490   if (UNLIKELY(ompt_enabled.enabled))
4491     __ompt_task_init(taskdata, thread->th.th_info.ds.ds_gtid);
4492 #endif
4493   return task;
4494 }
4495 
4496 // Routine optionally generated by the compiler for setting the lastprivate flag
4497 // and calling needed constructors for private/firstprivate objects
4498 // (used to form taskloop tasks from pattern task)
4499 // Parameters: dest task, src task, lastprivate flag.
4500 typedef void (*p_task_dup_t)(kmp_task_t *, kmp_task_t *, kmp_int32);
4501 
4502 KMP_BUILD_ASSERT(sizeof(long) == 4 || sizeof(long) == 8);
4503 
4504 // class to encapsulate manipulating loop bounds in a taskloop task.
4505 // this abstracts away the Intel vs GOMP taskloop interface for setting/getting
4506 // the loop bound variables.
4507 class kmp_taskloop_bounds_t {
4508   kmp_task_t *task;
4509   const kmp_taskdata_t *taskdata;
4510   size_t lower_offset;
4511   size_t upper_offset;
4512 
4513 public:
4514   kmp_taskloop_bounds_t(kmp_task_t *_task, kmp_uint64 *lb, kmp_uint64 *ub)
4515       : task(_task), taskdata(KMP_TASK_TO_TASKDATA(task)),
4516         lower_offset((char *)lb - (char *)task),
4517         upper_offset((char *)ub - (char *)task) {
4518     KMP_DEBUG_ASSERT((char *)lb > (char *)_task);
4519     KMP_DEBUG_ASSERT((char *)ub > (char *)_task);
4520   }
4521   kmp_taskloop_bounds_t(kmp_task_t *_task, const kmp_taskloop_bounds_t &bounds)
4522       : task(_task), taskdata(KMP_TASK_TO_TASKDATA(_task)),
4523         lower_offset(bounds.lower_offset), upper_offset(bounds.upper_offset) {}
4524   size_t get_lower_offset() const { return lower_offset; }
4525   size_t get_upper_offset() const { return upper_offset; }
4526   kmp_uint64 get_lb() const {
4527     kmp_int64 retval;
4528 #if defined(KMP_GOMP_COMPAT)
4529     // Intel task just returns the lower bound normally
4530     if (!taskdata->td_flags.native) {
4531       retval = *(kmp_int64 *)((char *)task + lower_offset);
4532     } else {
4533       // GOMP task has to take into account the sizeof(long)
4534       if (taskdata->td_size_loop_bounds == 4) {
4535         kmp_int32 *lb = RCAST(kmp_int32 *, task->shareds);
4536         retval = (kmp_int64)*lb;
4537       } else {
4538         kmp_int64 *lb = RCAST(kmp_int64 *, task->shareds);
4539         retval = (kmp_int64)*lb;
4540       }
4541     }
4542 #else
4543     (void)taskdata;
4544     retval = *(kmp_int64 *)((char *)task + lower_offset);
4545 #endif // defined(KMP_GOMP_COMPAT)
4546     return retval;
4547   }
4548   kmp_uint64 get_ub() const {
4549     kmp_int64 retval;
4550 #if defined(KMP_GOMP_COMPAT)
4551     // Intel task just returns the upper bound normally
4552     if (!taskdata->td_flags.native) {
4553       retval = *(kmp_int64 *)((char *)task + upper_offset);
4554     } else {
4555       // GOMP task has to take into account the sizeof(long)
4556       if (taskdata->td_size_loop_bounds == 4) {
4557         kmp_int32 *ub = RCAST(kmp_int32 *, task->shareds) + 1;
4558         retval = (kmp_int64)*ub;
4559       } else {
4560         kmp_int64 *ub = RCAST(kmp_int64 *, task->shareds) + 1;
4561         retval = (kmp_int64)*ub;
4562       }
4563     }
4564 #else
4565     retval = *(kmp_int64 *)((char *)task + upper_offset);
4566 #endif // defined(KMP_GOMP_COMPAT)
4567     return retval;
4568   }
4569   void set_lb(kmp_uint64 lb) {
4570 #if defined(KMP_GOMP_COMPAT)
4571     // Intel task just sets the lower bound normally
4572     if (!taskdata->td_flags.native) {
4573       *(kmp_uint64 *)((char *)task + lower_offset) = lb;
4574     } else {
4575       // GOMP task has to take into account the sizeof(long)
4576       if (taskdata->td_size_loop_bounds == 4) {
4577         kmp_uint32 *lower = RCAST(kmp_uint32 *, task->shareds);
4578         *lower = (kmp_uint32)lb;
4579       } else {
4580         kmp_uint64 *lower = RCAST(kmp_uint64 *, task->shareds);
4581         *lower = (kmp_uint64)lb;
4582       }
4583     }
4584 #else
4585     *(kmp_uint64 *)((char *)task + lower_offset) = lb;
4586 #endif // defined(KMP_GOMP_COMPAT)
4587   }
4588   void set_ub(kmp_uint64 ub) {
4589 #if defined(KMP_GOMP_COMPAT)
4590     // Intel task just sets the upper bound normally
4591     if (!taskdata->td_flags.native) {
4592       *(kmp_uint64 *)((char *)task + upper_offset) = ub;
4593     } else {
4594       // GOMP task has to take into account the sizeof(long)
4595       if (taskdata->td_size_loop_bounds == 4) {
4596         kmp_uint32 *upper = RCAST(kmp_uint32 *, task->shareds) + 1;
4597         *upper = (kmp_uint32)ub;
4598       } else {
4599         kmp_uint64 *upper = RCAST(kmp_uint64 *, task->shareds) + 1;
4600         *upper = (kmp_uint64)ub;
4601       }
4602     }
4603 #else
4604     *(kmp_uint64 *)((char *)task + upper_offset) = ub;
4605 #endif // defined(KMP_GOMP_COMPAT)
4606   }
4607 };
4608 
4609 // __kmp_taskloop_linear: Start tasks of the taskloop linearly
4610 //
4611 // loc        Source location information
4612 // gtid       Global thread ID
4613 // task       Pattern task, exposes the loop iteration range
4614 // lb         Pointer to loop lower bound in task structure
4615 // ub         Pointer to loop upper bound in task structure
4616 // st         Loop stride
4617 // ub_glob    Global upper bound (used for lastprivate check)
4618 // num_tasks  Number of tasks to execute
4619 // grainsize  Number of loop iterations per task
4620 // extras     Number of chunks with grainsize+1 iterations
4621 // last_chunk Reduction of grainsize for last task
4622 // tc         Iterations count
4623 // task_dup   Tasks duplication routine
4624 // codeptr_ra Return address for OMPT events
4625 void __kmp_taskloop_linear(ident_t *loc, int gtid, kmp_task_t *task,
4626                            kmp_uint64 *lb, kmp_uint64 *ub, kmp_int64 st,
4627                            kmp_uint64 ub_glob, kmp_uint64 num_tasks,
4628                            kmp_uint64 grainsize, kmp_uint64 extras,
4629                            kmp_int64 last_chunk, kmp_uint64 tc,
4630 #if OMPT_SUPPORT
4631                            void *codeptr_ra,
4632 #endif
4633                            void *task_dup) {
4634   KMP_COUNT_BLOCK(OMP_TASKLOOP);
4635   KMP_TIME_PARTITIONED_BLOCK(OMP_taskloop_scheduling);
4636   p_task_dup_t ptask_dup = (p_task_dup_t)task_dup;
4637   // compiler provides global bounds here
4638   kmp_taskloop_bounds_t task_bounds(task, lb, ub);
4639   kmp_uint64 lower = task_bounds.get_lb();
4640   kmp_uint64 upper = task_bounds.get_ub();
4641   kmp_uint64 i;
4642   kmp_info_t *thread = __kmp_threads[gtid];
4643   kmp_taskdata_t *current_task = thread->th.th_current_task;
4644   kmp_task_t *next_task;
4645   kmp_int32 lastpriv = 0;
4646 
4647   KMP_DEBUG_ASSERT(tc == num_tasks * grainsize +
4648                              (last_chunk < 0 ? last_chunk : extras));
4649   KMP_DEBUG_ASSERT(num_tasks > extras);
4650   KMP_DEBUG_ASSERT(num_tasks > 0);
4651   KA_TRACE(20, ("__kmp_taskloop_linear: T#%d: %lld tasks, grainsize %lld, "
4652                 "extras %lld, last_chunk %lld, i=%lld,%lld(%d)%lld, dup %p\n",
4653                 gtid, num_tasks, grainsize, extras, last_chunk, lower, upper,
4654                 ub_glob, st, task_dup));
4655 
4656   // Launch num_tasks tasks, assign grainsize iterations each task
4657   for (i = 0; i < num_tasks; ++i) {
4658     kmp_uint64 chunk_minus_1;
4659     if (extras == 0) {
4660       chunk_minus_1 = grainsize - 1;
4661     } else {
4662       chunk_minus_1 = grainsize;
4663       --extras; // first extras iterations get bigger chunk (grainsize+1)
4664     }
4665     upper = lower + st * chunk_minus_1;
4666     if (upper > *ub) {
4667       upper = *ub;
4668     }
4669     if (i == num_tasks - 1) {
4670       // schedule the last task, set lastprivate flag if needed
4671       if (st == 1) { // most common case
4672         KMP_DEBUG_ASSERT(upper == *ub);
4673         if (upper == ub_glob)
4674           lastpriv = 1;
4675       } else if (st > 0) { // positive loop stride
4676         KMP_DEBUG_ASSERT((kmp_uint64)st > *ub - upper);
4677         if ((kmp_uint64)st > ub_glob - upper)
4678           lastpriv = 1;
4679       } else { // negative loop stride
4680         KMP_DEBUG_ASSERT(upper + st < *ub);
4681         if (upper - ub_glob < (kmp_uint64)(-st))
4682           lastpriv = 1;
4683       }
4684     }
4685     next_task = __kmp_task_dup_alloc(thread, task); // allocate new task
4686     kmp_taskdata_t *next_taskdata = KMP_TASK_TO_TASKDATA(next_task);
4687     kmp_taskloop_bounds_t next_task_bounds =
4688         kmp_taskloop_bounds_t(next_task, task_bounds);
4689 
4690     // adjust task-specific bounds
4691     next_task_bounds.set_lb(lower);
4692     if (next_taskdata->td_flags.native) {
4693       next_task_bounds.set_ub(upper + (st > 0 ? 1 : -1));
4694     } else {
4695       next_task_bounds.set_ub(upper);
4696     }
4697     if (ptask_dup != NULL) // set lastprivate flag, construct firstprivates,
4698                            // etc.
4699       ptask_dup(next_task, task, lastpriv);
4700     KA_TRACE(40,
4701              ("__kmp_taskloop_linear: T#%d; task #%llu: task %p: lower %lld, "
4702               "upper %lld stride %lld, (offsets %p %p)\n",
4703               gtid, i, next_task, lower, upper, st,
4704               next_task_bounds.get_lower_offset(),
4705               next_task_bounds.get_upper_offset()));
4706 #if OMPT_SUPPORT
4707     __kmp_omp_taskloop_task(NULL, gtid, next_task,
4708                             codeptr_ra); // schedule new task
4709 #if OMPT_OPTIONAL
4710     if (ompt_enabled.ompt_callback_dispatch) {
4711       OMPT_GET_DISPATCH_CHUNK(next_taskdata->ompt_task_info.dispatch_chunk,
4712                               lower, upper, st);
4713     }
4714 #endif // OMPT_OPTIONAL
4715 #else
4716     __kmp_omp_task(gtid, next_task, true); // schedule new task
4717 #endif
4718     lower = upper + st; // adjust lower bound for the next iteration
4719   }
4720   // free the pattern task and exit
4721   __kmp_task_start(gtid, task, current_task); // make internal bookkeeping
4722   // do not execute the pattern task, just do internal bookkeeping
4723   __kmp_task_finish<false>(gtid, task, current_task);
4724 }
4725 
4726 // Structure to keep taskloop parameters for auxiliary task
4727 // kept in the shareds of the task structure.
4728 typedef struct __taskloop_params {
4729   kmp_task_t *task;
4730   kmp_uint64 *lb;
4731   kmp_uint64 *ub;
4732   void *task_dup;
4733   kmp_int64 st;
4734   kmp_uint64 ub_glob;
4735   kmp_uint64 num_tasks;
4736   kmp_uint64 grainsize;
4737   kmp_uint64 extras;
4738   kmp_int64 last_chunk;
4739   kmp_uint64 tc;
4740   kmp_uint64 num_t_min;
4741 #if OMPT_SUPPORT
4742   void *codeptr_ra;
4743 #endif
4744 } __taskloop_params_t;
4745 
4746 void __kmp_taskloop_recur(ident_t *, int, kmp_task_t *, kmp_uint64 *,
4747                           kmp_uint64 *, kmp_int64, kmp_uint64, kmp_uint64,
4748                           kmp_uint64, kmp_uint64, kmp_int64, kmp_uint64,
4749                           kmp_uint64,
4750 #if OMPT_SUPPORT
4751                           void *,
4752 #endif
4753                           void *);
4754 
4755 // Execute part of the taskloop submitted as a task.
4756 int __kmp_taskloop_task(int gtid, void *ptask) {
4757   __taskloop_params_t *p =
4758       (__taskloop_params_t *)((kmp_task_t *)ptask)->shareds;
4759   kmp_task_t *task = p->task;
4760   kmp_uint64 *lb = p->lb;
4761   kmp_uint64 *ub = p->ub;
4762   void *task_dup = p->task_dup;
4763   //  p_task_dup_t ptask_dup = (p_task_dup_t)task_dup;
4764   kmp_int64 st = p->st;
4765   kmp_uint64 ub_glob = p->ub_glob;
4766   kmp_uint64 num_tasks = p->num_tasks;
4767   kmp_uint64 grainsize = p->grainsize;
4768   kmp_uint64 extras = p->extras;
4769   kmp_int64 last_chunk = p->last_chunk;
4770   kmp_uint64 tc = p->tc;
4771   kmp_uint64 num_t_min = p->num_t_min;
4772 #if OMPT_SUPPORT
4773   void *codeptr_ra = p->codeptr_ra;
4774 #endif
4775 #if KMP_DEBUG
4776   kmp_taskdata_t *taskdata = KMP_TASK_TO_TASKDATA(task);
4777   KMP_DEBUG_ASSERT(task != NULL);
4778   KA_TRACE(20,
4779            ("__kmp_taskloop_task: T#%d, task %p: %lld tasks, grainsize"
4780             " %lld, extras %lld, last_chunk %lld, i=%lld,%lld(%d), dup %p\n",
4781             gtid, taskdata, num_tasks, grainsize, extras, last_chunk, *lb, *ub,
4782             st, task_dup));
4783 #endif
4784   KMP_DEBUG_ASSERT(num_tasks * 2 + 1 > num_t_min);
4785   if (num_tasks > num_t_min)
4786     __kmp_taskloop_recur(NULL, gtid, task, lb, ub, st, ub_glob, num_tasks,
4787                          grainsize, extras, last_chunk, tc, num_t_min,
4788 #if OMPT_SUPPORT
4789                          codeptr_ra,
4790 #endif
4791                          task_dup);
4792   else
4793     __kmp_taskloop_linear(NULL, gtid, task, lb, ub, st, ub_glob, num_tasks,
4794                           grainsize, extras, last_chunk, tc,
4795 #if OMPT_SUPPORT
4796                           codeptr_ra,
4797 #endif
4798                           task_dup);
4799 
4800   KA_TRACE(40, ("__kmp_taskloop_task(exit): T#%d\n", gtid));
4801   return 0;
4802 }
4803 
4804 // Schedule part of the taskloop as a task,
4805 // execute the rest of the taskloop.
4806 //
4807 // loc        Source location information
4808 // gtid       Global thread ID
4809 // task       Pattern task, exposes the loop iteration range
4810 // lb         Pointer to loop lower bound in task structure
4811 // ub         Pointer to loop upper bound in task structure
4812 // st         Loop stride
4813 // ub_glob    Global upper bound (used for lastprivate check)
4814 // num_tasks  Number of tasks to execute
4815 // grainsize  Number of loop iterations per task
4816 // extras     Number of chunks with grainsize+1 iterations
4817 // last_chunk Reduction of grainsize for last task
4818 // tc         Iterations count
4819 // num_t_min  Threshold to launch tasks recursively
4820 // task_dup   Tasks duplication routine
4821 // codeptr_ra Return address for OMPT events
4822 void __kmp_taskloop_recur(ident_t *loc, int gtid, kmp_task_t *task,
4823                           kmp_uint64 *lb, kmp_uint64 *ub, kmp_int64 st,
4824                           kmp_uint64 ub_glob, kmp_uint64 num_tasks,
4825                           kmp_uint64 grainsize, kmp_uint64 extras,
4826                           kmp_int64 last_chunk, kmp_uint64 tc,
4827                           kmp_uint64 num_t_min,
4828 #if OMPT_SUPPORT
4829                           void *codeptr_ra,
4830 #endif
4831                           void *task_dup) {
4832   kmp_taskdata_t *taskdata = KMP_TASK_TO_TASKDATA(task);
4833   KMP_DEBUG_ASSERT(task != NULL);
4834   KMP_DEBUG_ASSERT(num_tasks > num_t_min);
4835   KA_TRACE(20,
4836            ("__kmp_taskloop_recur: T#%d, task %p: %lld tasks, grainsize"
4837             " %lld, extras %lld, last_chunk %lld, i=%lld,%lld(%d), dup %p\n",
4838             gtid, taskdata, num_tasks, grainsize, extras, last_chunk, *lb, *ub,
4839             st, task_dup));
4840   p_task_dup_t ptask_dup = (p_task_dup_t)task_dup;
4841   kmp_uint64 lower = *lb;
4842   kmp_info_t *thread = __kmp_threads[gtid];
4843   //  kmp_taskdata_t *current_task = thread->th.th_current_task;
4844   kmp_task_t *next_task;
4845   size_t lower_offset =
4846       (char *)lb - (char *)task; // remember offset of lb in the task structure
4847   size_t upper_offset =
4848       (char *)ub - (char *)task; // remember offset of ub in the task structure
4849 
4850   KMP_DEBUG_ASSERT(tc == num_tasks * grainsize +
4851                              (last_chunk < 0 ? last_chunk : extras));
4852   KMP_DEBUG_ASSERT(num_tasks > extras);
4853   KMP_DEBUG_ASSERT(num_tasks > 0);
4854 
4855   // split the loop in two halves
4856   kmp_uint64 lb1, ub0, tc0, tc1, ext0, ext1;
4857   kmp_int64 last_chunk0 = 0, last_chunk1 = 0;
4858   kmp_uint64 gr_size0 = grainsize;
4859   kmp_uint64 n_tsk0 = num_tasks >> 1; // num_tasks/2 to execute
4860   kmp_uint64 n_tsk1 = num_tasks - n_tsk0; // to schedule as a task
4861   if (last_chunk < 0) {
4862     ext0 = ext1 = 0;
4863     last_chunk1 = last_chunk;
4864     tc0 = grainsize * n_tsk0;
4865     tc1 = tc - tc0;
4866   } else if (n_tsk0 <= extras) {
4867     gr_size0++; // integrate extras into grainsize
4868     ext0 = 0; // no extra iters in 1st half
4869     ext1 = extras - n_tsk0; // remaining extras
4870     tc0 = gr_size0 * n_tsk0;
4871     tc1 = tc - tc0;
4872   } else { // n_tsk0 > extras
4873     ext1 = 0; // no extra iters in 2nd half
4874     ext0 = extras;
4875     tc1 = grainsize * n_tsk1;
4876     tc0 = tc - tc1;
4877   }
4878   ub0 = lower + st * (tc0 - 1);
4879   lb1 = ub0 + st;
4880 
4881   // create pattern task for 2nd half of the loop
4882   next_task = __kmp_task_dup_alloc(thread, task); // duplicate the task
4883   // adjust lower bound (upper bound is not changed) for the 2nd half
4884   *(kmp_uint64 *)((char *)next_task + lower_offset) = lb1;
4885   if (ptask_dup != NULL) // construct firstprivates, etc.
4886     ptask_dup(next_task, task, 0);
4887   *ub = ub0; // adjust upper bound for the 1st half
4888 
4889   // create auxiliary task for 2nd half of the loop
4890   // make sure new task has same parent task as the pattern task
4891   kmp_taskdata_t *current_task = thread->th.th_current_task;
4892   thread->th.th_current_task = taskdata->td_parent;
4893   kmp_task_t *new_task =
4894       __kmpc_omp_task_alloc(loc, gtid, 1, 3 * sizeof(void *),
4895                             sizeof(__taskloop_params_t), &__kmp_taskloop_task);
4896   // restore current task
4897   thread->th.th_current_task = current_task;
4898   __taskloop_params_t *p = (__taskloop_params_t *)new_task->shareds;
4899   p->task = next_task;
4900   p->lb = (kmp_uint64 *)((char *)next_task + lower_offset);
4901   p->ub = (kmp_uint64 *)((char *)next_task + upper_offset);
4902   p->task_dup = task_dup;
4903   p->st = st;
4904   p->ub_glob = ub_glob;
4905   p->num_tasks = n_tsk1;
4906   p->grainsize = grainsize;
4907   p->extras = ext1;
4908   p->last_chunk = last_chunk1;
4909   p->tc = tc1;
4910   p->num_t_min = num_t_min;
4911 #if OMPT_SUPPORT
4912   p->codeptr_ra = codeptr_ra;
4913 #endif
4914 
4915 #if OMPT_SUPPORT
4916   // schedule new task with correct return address for OMPT events
4917   __kmp_omp_taskloop_task(NULL, gtid, new_task, codeptr_ra);
4918 #else
4919   __kmp_omp_task(gtid, new_task, true); // schedule new task
4920 #endif
4921 
4922   // execute the 1st half of current subrange
4923   if (n_tsk0 > num_t_min)
4924     __kmp_taskloop_recur(loc, gtid, task, lb, ub, st, ub_glob, n_tsk0, gr_size0,
4925                          ext0, last_chunk0, tc0, num_t_min,
4926 #if OMPT_SUPPORT
4927                          codeptr_ra,
4928 #endif
4929                          task_dup);
4930   else
4931     __kmp_taskloop_linear(loc, gtid, task, lb, ub, st, ub_glob, n_tsk0,
4932                           gr_size0, ext0, last_chunk0, tc0,
4933 #if OMPT_SUPPORT
4934                           codeptr_ra,
4935 #endif
4936                           task_dup);
4937 
4938   KA_TRACE(40, ("__kmp_taskloop_recur(exit): T#%d\n", gtid));
4939 }
4940 
4941 static void __kmp_taskloop(ident_t *loc, int gtid, kmp_task_t *task, int if_val,
4942                            kmp_uint64 *lb, kmp_uint64 *ub, kmp_int64 st,
4943                            int nogroup, int sched, kmp_uint64 grainsize,
4944                            int modifier, void *task_dup) {
4945   kmp_taskdata_t *taskdata = KMP_TASK_TO_TASKDATA(task);
4946   KMP_DEBUG_ASSERT(task != NULL);
4947   if (nogroup == 0) {
4948 #if OMPT_SUPPORT && OMPT_OPTIONAL
4949     OMPT_STORE_RETURN_ADDRESS(gtid);
4950 #endif
4951     __kmpc_taskgroup(loc, gtid);
4952   }
4953 
4954   // =========================================================================
4955   // calculate loop parameters
4956   kmp_taskloop_bounds_t task_bounds(task, lb, ub);
4957   kmp_uint64 tc;
4958   // compiler provides global bounds here
4959   kmp_uint64 lower = task_bounds.get_lb();
4960   kmp_uint64 upper = task_bounds.get_ub();
4961   kmp_uint64 ub_glob = upper; // global upper used to calc lastprivate flag
4962   kmp_uint64 num_tasks = 0, extras = 0;
4963   kmp_int64 last_chunk =
4964       0; // reduce grainsize of last task by last_chunk in strict mode
4965   kmp_uint64 num_tasks_min = __kmp_taskloop_min_tasks;
4966   kmp_info_t *thread = __kmp_threads[gtid];
4967   kmp_taskdata_t *current_task = thread->th.th_current_task;
4968 
4969   KA_TRACE(20, ("__kmp_taskloop: T#%d, task %p, lb %lld, ub %lld, st %lld, "
4970                 "grain %llu(%d, %d), dup %p\n",
4971                 gtid, taskdata, lower, upper, st, grainsize, sched, modifier,
4972                 task_dup));
4973 
4974   // compute trip count
4975   if (st == 1) { // most common case
4976     tc = upper - lower + 1;
4977   } else if (st < 0) {
4978     tc = (lower - upper) / (-st) + 1;
4979   } else { // st > 0
4980     tc = (upper - lower) / st + 1;
4981   }
4982   if (tc == 0) {
4983     KA_TRACE(20, ("__kmp_taskloop(exit): T#%d zero-trip loop\n", gtid));
4984     // free the pattern task and exit
4985     __kmp_task_start(gtid, task, current_task);
4986     // do not execute anything for zero-trip loop
4987     __kmp_task_finish<false>(gtid, task, current_task);
4988     return;
4989   }
4990 
4991 #if OMPT_SUPPORT && OMPT_OPTIONAL
4992   ompt_team_info_t *team_info = __ompt_get_teaminfo(0, NULL);
4993   ompt_task_info_t *task_info = __ompt_get_task_info_object(0);
4994   if (ompt_enabled.ompt_callback_work) {
4995     ompt_callbacks.ompt_callback(ompt_callback_work)(
4996         ompt_work_taskloop, ompt_scope_begin, &(team_info->parallel_data),
4997         &(task_info->task_data), tc, OMPT_GET_RETURN_ADDRESS(0));
4998   }
4999 #endif
5000 
5001   if (num_tasks_min == 0)
5002     // TODO: can we choose better default heuristic?
5003     num_tasks_min =
5004         KMP_MIN(thread->th.th_team_nproc * 10, INITIAL_TASK_DEQUE_SIZE);
5005 
5006   // compute num_tasks/grainsize based on the input provided
5007   switch (sched) {
5008   case 0: // no schedule clause specified, we can choose the default
5009     // let's try to schedule (team_size*10) tasks
5010     grainsize = thread->th.th_team_nproc * 10;
5011     KMP_FALLTHROUGH();
5012   case 2: // num_tasks provided
5013     if (grainsize > tc) {
5014       num_tasks = tc; // too big num_tasks requested, adjust values
5015       grainsize = 1;
5016       extras = 0;
5017     } else {
5018       num_tasks = grainsize;
5019       grainsize = tc / num_tasks;
5020       extras = tc % num_tasks;
5021     }
5022     break;
5023   case 1: // grainsize provided
5024     if (grainsize > tc) {
5025       num_tasks = 1;
5026       grainsize = tc; // too big grainsize requested, adjust values
5027       extras = 0;
5028     } else {
5029       if (modifier) {
5030         num_tasks = (tc + grainsize - 1) / grainsize;
5031         last_chunk = tc - (num_tasks * grainsize);
5032         extras = 0;
5033       } else {
5034         num_tasks = tc / grainsize;
5035         // adjust grainsize for balanced distribution of iterations
5036         grainsize = tc / num_tasks;
5037         extras = tc % num_tasks;
5038       }
5039     }
5040     break;
5041   default:
5042     KMP_ASSERT2(0, "unknown scheduling of taskloop");
5043   }
5044 
5045   KMP_DEBUG_ASSERT(tc == num_tasks * grainsize +
5046                              (last_chunk < 0 ? last_chunk : extras));
5047   KMP_DEBUG_ASSERT(num_tasks > extras);
5048   KMP_DEBUG_ASSERT(num_tasks > 0);
5049   // =========================================================================
5050 
5051   // check if clause value first
5052   // Also require GOMP_taskloop to reduce to linear (taskdata->td_flags.native)
5053   if (if_val == 0) { // if(0) specified, mark task as serial
5054     taskdata->td_flags.task_serial = 1;
5055     taskdata->td_flags.tiedness = TASK_TIED; // AC: serial task cannot be untied
5056     // always start serial tasks linearly
5057     __kmp_taskloop_linear(loc, gtid, task, lb, ub, st, ub_glob, num_tasks,
5058                           grainsize, extras, last_chunk, tc,
5059 #if OMPT_SUPPORT
5060                           OMPT_GET_RETURN_ADDRESS(0),
5061 #endif
5062                           task_dup);
5063     // !taskdata->td_flags.native => currently force linear spawning of tasks
5064     // for GOMP_taskloop
5065   } else if (num_tasks > num_tasks_min && !taskdata->td_flags.native) {
5066     KA_TRACE(20, ("__kmp_taskloop: T#%d, go recursive: tc %llu, #tasks %llu"
5067                   "(%lld), grain %llu, extras %llu, last_chunk %lld\n",
5068                   gtid, tc, num_tasks, num_tasks_min, grainsize, extras,
5069                   last_chunk));
5070     __kmp_taskloop_recur(loc, gtid, task, lb, ub, st, ub_glob, num_tasks,
5071                          grainsize, extras, last_chunk, tc, num_tasks_min,
5072 #if OMPT_SUPPORT
5073                          OMPT_GET_RETURN_ADDRESS(0),
5074 #endif
5075                          task_dup);
5076   } else {
5077     KA_TRACE(20, ("__kmp_taskloop: T#%d, go linear: tc %llu, #tasks %llu"
5078                   "(%lld), grain %llu, extras %llu, last_chunk %lld\n",
5079                   gtid, tc, num_tasks, num_tasks_min, grainsize, extras,
5080                   last_chunk));
5081     __kmp_taskloop_linear(loc, gtid, task, lb, ub, st, ub_glob, num_tasks,
5082                           grainsize, extras, last_chunk, tc,
5083 #if OMPT_SUPPORT
5084                           OMPT_GET_RETURN_ADDRESS(0),
5085 #endif
5086                           task_dup);
5087   }
5088 
5089 #if OMPT_SUPPORT && OMPT_OPTIONAL
5090   if (ompt_enabled.ompt_callback_work) {
5091     ompt_callbacks.ompt_callback(ompt_callback_work)(
5092         ompt_work_taskloop, ompt_scope_end, &(team_info->parallel_data),
5093         &(task_info->task_data), tc, OMPT_GET_RETURN_ADDRESS(0));
5094   }
5095 #endif
5096 
5097   if (nogroup == 0) {
5098 #if OMPT_SUPPORT && OMPT_OPTIONAL
5099     OMPT_STORE_RETURN_ADDRESS(gtid);
5100 #endif
5101     __kmpc_end_taskgroup(loc, gtid);
5102   }
5103   KA_TRACE(20, ("__kmp_taskloop(exit): T#%d\n", gtid));
5104 }
5105 
5106 /*!
5107 @ingroup TASKING
5108 @param loc       Source location information
5109 @param gtid      Global thread ID
5110 @param task      Task structure
5111 @param if_val    Value of the if clause
5112 @param lb        Pointer to loop lower bound in task structure
5113 @param ub        Pointer to loop upper bound in task structure
5114 @param st        Loop stride
5115 @param nogroup   Flag, 1 if nogroup clause specified, 0 otherwise
5116 @param sched     Schedule specified 0/1/2 for none/grainsize/num_tasks
5117 @param grainsize Schedule value if specified
5118 @param task_dup  Tasks duplication routine
5119 
5120 Execute the taskloop construct.
5121 */
5122 void __kmpc_taskloop(ident_t *loc, int gtid, kmp_task_t *task, int if_val,
5123                      kmp_uint64 *lb, kmp_uint64 *ub, kmp_int64 st, int nogroup,
5124                      int sched, kmp_uint64 grainsize, void *task_dup) {
5125   __kmp_assert_valid_gtid(gtid);
5126   KA_TRACE(20, ("__kmpc_taskloop(enter): T#%d\n", gtid));
5127   __kmp_taskloop(loc, gtid, task, if_val, lb, ub, st, nogroup, sched, grainsize,
5128                  0, task_dup);
5129   KA_TRACE(20, ("__kmpc_taskloop(exit): T#%d\n", gtid));
5130 }
5131 
5132 /*!
5133 @ingroup TASKING
5134 @param loc       Source location information
5135 @param gtid      Global thread ID
5136 @param task      Task structure
5137 @param if_val    Value of the if clause
5138 @param lb        Pointer to loop lower bound in task structure
5139 @param ub        Pointer to loop upper bound in task structure
5140 @param st        Loop stride
5141 @param nogroup   Flag, 1 if nogroup clause specified, 0 otherwise
5142 @param sched     Schedule specified 0/1/2 for none/grainsize/num_tasks
5143 @param grainsize Schedule value if specified
5144 @param modifier  Modifier 'strict' for sched, 1 if present, 0 otherwise
5145 @param task_dup  Tasks duplication routine
5146 
5147 Execute the taskloop construct.
5148 */
5149 void __kmpc_taskloop_5(ident_t *loc, int gtid, kmp_task_t *task, int if_val,
5150                        kmp_uint64 *lb, kmp_uint64 *ub, kmp_int64 st,
5151                        int nogroup, int sched, kmp_uint64 grainsize,
5152                        int modifier, void *task_dup) {
5153   __kmp_assert_valid_gtid(gtid);
5154   KA_TRACE(20, ("__kmpc_taskloop_5(enter): T#%d\n", gtid));
5155   __kmp_taskloop(loc, gtid, task, if_val, lb, ub, st, nogroup, sched, grainsize,
5156                  modifier, task_dup);
5157   KA_TRACE(20, ("__kmpc_taskloop_5(exit): T#%d\n", gtid));
5158 }
5159 
5160 /*!
5161 @ingroup TASKING
5162 @param gtid Global Thread ID of current thread
5163 @return Returns a pointer to the thread's current task async handle. If no task
5164 is present or gtid is invalid, returns NULL.
5165 
5166 Acqurires a pointer to the target async handle from the current task.
5167 */
5168 void **__kmpc_omp_get_target_async_handle_ptr(kmp_int32 gtid) {
5169   if (gtid == KMP_GTID_DNE)
5170     return NULL;
5171 
5172   kmp_info_t *thread = __kmp_thread_from_gtid(gtid);
5173   kmp_taskdata_t *taskdata = thread->th.th_current_task;
5174 
5175   if (!taskdata)
5176     return NULL;
5177 
5178   return &taskdata->td_target_data.async_handle;
5179 }
5180 
5181 /*!
5182 @ingroup TASKING
5183 @param gtid Global Thread ID of current thread
5184 @return Returns TRUE if the current task being executed of the given thread has
5185 a task team allocated to it. Otherwise, returns FALSE.
5186 
5187 Checks if the current thread has a task team.
5188 */
5189 bool __kmpc_omp_has_task_team(kmp_int32 gtid) {
5190   if (gtid == KMP_GTID_DNE)
5191     return FALSE;
5192 
5193   kmp_info_t *thread = __kmp_thread_from_gtid(gtid);
5194   kmp_taskdata_t *taskdata = thread->th.th_current_task;
5195 
5196   if (!taskdata)
5197     return FALSE;
5198 
5199   return taskdata->td_task_team != NULL;
5200 }
5201