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