1 /*
2  * kmp_taskdeps.cpp
3  */
4 
5 
6 //===----------------------------------------------------------------------===//
7 //
8 //                     The LLVM Compiler Infrastructure
9 //
10 // This file is dual licensed under the MIT and the University of Illinois Open
11 // Source Licenses. See LICENSE.txt for details.
12 //
13 //===----------------------------------------------------------------------===//
14 
15 
16 //#define KMP_SUPPORT_GRAPH_OUTPUT 1
17 
18 #include "kmp.h"
19 #include "kmp_io.h"
20 #include "kmp_wait_release.h"
21 
22 #if OMP_40_ENABLED
23 
24 // TODO: Improve memory allocation? keep a list of pre-allocated structures?
25 // allocate in blocks? re-use list finished list entries?
26 // TODO: don't use atomic ref counters for stack-allocated nodes.
27 // TODO: find an alternate to atomic refs for heap-allocated nodes?
28 // TODO: Finish graph output support
29 // TODO: kmp_lock_t seems a tad to big (and heavy weight) for this. Check other
30 // runtime locks
31 // TODO: Any ITT support needed?
32 
33 #ifdef KMP_SUPPORT_GRAPH_OUTPUT
34 static kmp_int32 kmp_node_id_seed = 0;
35 #endif
36 
37 static void __kmp_init_node(kmp_depnode_t *node) {
38   node->dn.task = NULL; // set to null initially, it will point to the right
39   // task once dependences have been processed
40   node->dn.successors = NULL;
41   __kmp_init_lock(&node->dn.lock);
42   node->dn.nrefs = 1; // init creates the first reference to the node
43 #ifdef KMP_SUPPORT_GRAPH_OUTPUT
44   node->dn.id = KMP_TEST_THEN_INC32(&kmp_node_id_seed);
45 #endif
46 }
47 
48 static inline kmp_depnode_t *__kmp_node_ref(kmp_depnode_t *node) {
49   KMP_TEST_THEN_INC32(&node->dn.nrefs);
50   return node;
51 }
52 
53 static inline void __kmp_node_deref(kmp_info_t *thread, kmp_depnode_t *node) {
54   if (!node)
55     return;
56 
57   kmp_int32 n = KMP_TEST_THEN_DEC32(&node->dn.nrefs) - 1;
58   if (n == 0) {
59     KMP_ASSERT(node->dn.nrefs == 0);
60 #if USE_FAST_MEMORY
61     __kmp_fast_free(thread, node);
62 #else
63     __kmp_thread_free(thread, node);
64 #endif
65   }
66 }
67 
68 #define KMP_ACQUIRE_DEPNODE(gtid, n) __kmp_acquire_lock(&(n)->dn.lock, (gtid))
69 #define KMP_RELEASE_DEPNODE(gtid, n) __kmp_release_lock(&(n)->dn.lock, (gtid))
70 
71 static void __kmp_depnode_list_free(kmp_info_t *thread, kmp_depnode_list *list);
72 
73 enum { KMP_DEPHASH_OTHER_SIZE = 97, KMP_DEPHASH_MASTER_SIZE = 997 };
74 
75 static inline kmp_int32 __kmp_dephash_hash(kmp_intptr_t addr, size_t hsize) {
76   // TODO alternate to try: set = (((Addr64)(addrUsefulBits * 9.618)) %
77   // m_num_sets );
78   return ((addr >> 6) ^ (addr >> 2)) % hsize;
79 }
80 
81 static kmp_dephash_t *__kmp_dephash_create(kmp_info_t *thread,
82                                            kmp_taskdata_t *current_task) {
83   kmp_dephash_t *h;
84 
85   size_t h_size;
86 
87   if (current_task->td_flags.tasktype == TASK_IMPLICIT)
88     h_size = KMP_DEPHASH_MASTER_SIZE;
89   else
90     h_size = KMP_DEPHASH_OTHER_SIZE;
91 
92   kmp_int32 size =
93       h_size * sizeof(kmp_dephash_entry_t *) + sizeof(kmp_dephash_t);
94 
95 #if USE_FAST_MEMORY
96   h = (kmp_dephash_t *)__kmp_fast_allocate(thread, size);
97 #else
98   h = (kmp_dephash_t *)__kmp_thread_malloc(thread, size);
99 #endif
100   h->size = h_size;
101 
102 #ifdef KMP_DEBUG
103   h->nelements = 0;
104   h->nconflicts = 0;
105 #endif
106   h->buckets = (kmp_dephash_entry **)(h + 1);
107 
108   for (size_t i = 0; i < h_size; i++)
109     h->buckets[i] = 0;
110 
111   return h;
112 }
113 
114 void __kmp_dephash_free_entries(kmp_info_t *thread, kmp_dephash_t *h) {
115   for (size_t i = 0; i < h->size; i++) {
116     if (h->buckets[i]) {
117       kmp_dephash_entry_t *next;
118       for (kmp_dephash_entry_t *entry = h->buckets[i]; entry; entry = next) {
119         next = entry->next_in_bucket;
120         __kmp_depnode_list_free(thread, entry->last_ins);
121         __kmp_node_deref(thread, entry->last_out);
122 #if USE_FAST_MEMORY
123         __kmp_fast_free(thread, entry);
124 #else
125         __kmp_thread_free(thread, entry);
126 #endif
127       }
128       h->buckets[i] = 0;
129     }
130   }
131 }
132 
133 void __kmp_dephash_free(kmp_info_t *thread, kmp_dephash_t *h) {
134   __kmp_dephash_free_entries(thread, h);
135 #if USE_FAST_MEMORY
136   __kmp_fast_free(thread, h);
137 #else
138   __kmp_thread_free(thread, h);
139 #endif
140 }
141 
142 static kmp_dephash_entry *
143 __kmp_dephash_find(kmp_info_t *thread, kmp_dephash_t *h, kmp_intptr_t addr) {
144   kmp_int32 bucket = __kmp_dephash_hash(addr, h->size);
145 
146   kmp_dephash_entry_t *entry;
147   for (entry = h->buckets[bucket]; entry; entry = entry->next_in_bucket)
148     if (entry->addr == addr)
149       break;
150 
151   if (entry == NULL) {
152 // create entry. This is only done by one thread so no locking required
153 #if USE_FAST_MEMORY
154     entry = (kmp_dephash_entry_t *)__kmp_fast_allocate(
155         thread, sizeof(kmp_dephash_entry_t));
156 #else
157     entry = (kmp_dephash_entry_t *)__kmp_thread_malloc(
158         thread, sizeof(kmp_dephash_entry_t));
159 #endif
160     entry->addr = addr;
161     entry->last_out = NULL;
162     entry->last_ins = NULL;
163     entry->next_in_bucket = h->buckets[bucket];
164     h->buckets[bucket] = entry;
165 #ifdef KMP_DEBUG
166     h->nelements++;
167     if (entry->next_in_bucket)
168       h->nconflicts++;
169 #endif
170   }
171   return entry;
172 }
173 
174 static kmp_depnode_list_t *__kmp_add_node(kmp_info_t *thread,
175                                           kmp_depnode_list_t *list,
176                                           kmp_depnode_t *node) {
177   kmp_depnode_list_t *new_head;
178 
179 #if USE_FAST_MEMORY
180   new_head = (kmp_depnode_list_t *)__kmp_fast_allocate(
181       thread, sizeof(kmp_depnode_list_t));
182 #else
183   new_head = (kmp_depnode_list_t *)__kmp_thread_malloc(
184       thread, sizeof(kmp_depnode_list_t));
185 #endif
186 
187   new_head->node = __kmp_node_ref(node);
188   new_head->next = list;
189 
190   return new_head;
191 }
192 
193 static void __kmp_depnode_list_free(kmp_info_t *thread,
194                                     kmp_depnode_list *list) {
195   kmp_depnode_list *next;
196 
197   for (; list; list = next) {
198     next = list->next;
199 
200     __kmp_node_deref(thread, list->node);
201 #if USE_FAST_MEMORY
202     __kmp_fast_free(thread, list);
203 #else
204     __kmp_thread_free(thread, list);
205 #endif
206   }
207 }
208 
209 static inline void __kmp_track_dependence(kmp_depnode_t *source,
210                                           kmp_depnode_t *sink,
211                                           kmp_task_t *sink_task) {
212 #ifdef KMP_SUPPORT_GRAPH_OUTPUT
213   kmp_taskdata_t *task_source = KMP_TASK_TO_TASKDATA(source->dn.task);
214   // do not use sink->dn.task as that is only filled after the dependencies
215   // are already processed!
216   kmp_taskdata_t *task_sink = KMP_TASK_TO_TASKDATA(sink_task);
217 
218   __kmp_printf("%d(%s) -> %d(%s)\n", source->dn.id,
219                task_source->td_ident->psource, sink->dn.id,
220                task_sink->td_ident->psource);
221 #endif
222 #if OMPT_SUPPORT && OMPT_TRACE
223   // OMPT tracks dependences between task (a=source, b=sink) in which
224   // task a blocks the execution of b through the ompt_new_dependence_callback
225   if (ompt_enabled &&
226       ompt_callbacks.ompt_callback(ompt_event_task_dependence_pair)) {
227     kmp_taskdata_t *task_source = KMP_TASK_TO_TASKDATA(source->dn.task);
228     kmp_taskdata_t *task_sink = KMP_TASK_TO_TASKDATA(sink_task);
229 
230     ompt_callbacks.ompt_callback(ompt_event_task_dependence_pair)(
231         task_source->ompt_task_info.task_id, task_sink->ompt_task_info.task_id);
232   }
233 #endif /* OMPT_SUPPORT && OMPT_TRACE */
234 }
235 
236 template <bool filter>
237 static inline kmp_int32
238 __kmp_process_deps(kmp_int32 gtid, kmp_depnode_t *node, kmp_dephash_t *hash,
239                    bool dep_barrier, kmp_int32 ndeps,
240                    kmp_depend_info_t *dep_list, kmp_task_t *task) {
241   KA_TRACE(30, ("__kmp_process_deps<%d>: T#%d processing %d dependencies : "
242                 "dep_barrier = %d\n",
243                 filter, gtid, ndeps, dep_barrier));
244 
245   kmp_info_t *thread = __kmp_threads[gtid];
246   kmp_int32 npredecessors = 0;
247   for (kmp_int32 i = 0; i < ndeps; i++) {
248     const kmp_depend_info_t *dep = &dep_list[i];
249 
250     KMP_DEBUG_ASSERT(dep->flags.in);
251 
252     if (filter && dep->base_addr == 0)
253       continue; // skip filtered entries
254 
255     kmp_dephash_entry_t *info =
256         __kmp_dephash_find(thread, hash, dep->base_addr);
257     kmp_depnode_t *last_out = info->last_out;
258 
259     if (dep->flags.out && info->last_ins) {
260       for (kmp_depnode_list_t *p = info->last_ins; p; p = p->next) {
261         kmp_depnode_t *indep = p->node;
262         if (indep->dn.task) {
263           KMP_ACQUIRE_DEPNODE(gtid, indep);
264           if (indep->dn.task) {
265             __kmp_track_dependence(indep, node, task);
266             indep->dn.successors =
267                 __kmp_add_node(thread, indep->dn.successors, node);
268             KA_TRACE(40, ("__kmp_process_deps<%d>: T#%d adding dependence from "
269                           "%p to %p\n",
270                           filter, gtid, KMP_TASK_TO_TASKDATA(indep->dn.task),
271                           KMP_TASK_TO_TASKDATA(task)));
272             npredecessors++;
273           }
274           KMP_RELEASE_DEPNODE(gtid, indep);
275         }
276       }
277 
278       __kmp_depnode_list_free(thread, info->last_ins);
279       info->last_ins = NULL;
280 
281     } else if (last_out && last_out->dn.task) {
282       KMP_ACQUIRE_DEPNODE(gtid, last_out);
283       if (last_out->dn.task) {
284         __kmp_track_dependence(last_out, node, task);
285         last_out->dn.successors =
286             __kmp_add_node(thread, last_out->dn.successors, node);
287         KA_TRACE(
288             40,
289             ("__kmp_process_deps<%d>: T#%d adding dependence from %p to %p\n",
290              filter, gtid, KMP_TASK_TO_TASKDATA(last_out->dn.task),
291              KMP_TASK_TO_TASKDATA(task)));
292 
293         npredecessors++;
294       }
295       KMP_RELEASE_DEPNODE(gtid, last_out);
296     }
297 
298     if (dep_barrier) {
299       // if this is a sync point in the serial sequence, then the previous
300       // outputs are guaranteed to be completed after
301       // the execution of this task so the previous output nodes can be cleared.
302       __kmp_node_deref(thread, last_out);
303       info->last_out = NULL;
304     } else {
305       if (dep->flags.out) {
306         __kmp_node_deref(thread, last_out);
307         info->last_out = __kmp_node_ref(node);
308       } else
309         info->last_ins = __kmp_add_node(thread, info->last_ins, node);
310     }
311   }
312 
313   KA_TRACE(30, ("__kmp_process_deps<%d>: T#%d found %d predecessors\n", filter,
314                 gtid, npredecessors));
315 
316   return npredecessors;
317 }
318 
319 #define NO_DEP_BARRIER (false)
320 #define DEP_BARRIER (true)
321 
322 // returns true if the task has any outstanding dependence
323 static bool __kmp_check_deps(kmp_int32 gtid, kmp_depnode_t *node,
324                              kmp_task_t *task, kmp_dephash_t *hash,
325                              bool dep_barrier, kmp_int32 ndeps,
326                              kmp_depend_info_t *dep_list,
327                              kmp_int32 ndeps_noalias,
328                              kmp_depend_info_t *noalias_dep_list) {
329   int i;
330 
331 #if KMP_DEBUG
332   kmp_taskdata_t *taskdata = KMP_TASK_TO_TASKDATA(task);
333 #endif
334   KA_TRACE(20, ("__kmp_check_deps: T#%d checking dependencies for task %p : %d "
335                 "possibly aliased dependencies, %d non-aliased depedencies : "
336                 "dep_barrier=%d .\n",
337                 gtid, taskdata, ndeps, ndeps_noalias, dep_barrier));
338 
339   // Filter deps in dep_list
340   // TODO: Different algorithm for large dep_list ( > 10 ? )
341   for (i = 0; i < ndeps; i++) {
342     if (dep_list[i].base_addr != 0)
343       for (int j = i + 1; j < ndeps; j++)
344         if (dep_list[i].base_addr == dep_list[j].base_addr) {
345           dep_list[i].flags.in |= dep_list[j].flags.in;
346           dep_list[i].flags.out |= dep_list[j].flags.out;
347           dep_list[j].base_addr = 0; // Mark j element as void
348         }
349   }
350 
351   // doesn't need to be atomic as no other thread is going to be accessing this
352   // node just yet.
353   // npredecessors is set -1 to ensure that none of the releasing tasks queues
354   // this task before we have finished processing all the dependencies
355   node->dn.npredecessors = -1;
356 
357   // used to pack all npredecessors additions into a single atomic operation at
358   // the end
359   int npredecessors;
360 
361   npredecessors = __kmp_process_deps<true>(gtid, node, hash, dep_barrier, ndeps,
362                                            dep_list, task);
363   npredecessors += __kmp_process_deps<false>(
364       gtid, node, hash, dep_barrier, ndeps_noalias, noalias_dep_list, task);
365 
366   node->dn.task = task;
367   KMP_MB();
368 
369   // Account for our initial fake value
370   npredecessors++;
371 
372   // Update predecessors and obtain current value to check if there are still
373   // any outstandig dependences (some tasks may have finished while we processed
374   // the dependences)
375   npredecessors = KMP_TEST_THEN_ADD32(&node->dn.npredecessors, npredecessors) +
376                   npredecessors;
377 
378   KA_TRACE(20, ("__kmp_check_deps: T#%d found %d predecessors for task %p \n",
379                 gtid, npredecessors, taskdata));
380 
381   // beyond this point the task could be queued (and executed) by a releasing
382   // task...
383   return npredecessors > 0 ? true : false;
384 }
385 
386 void __kmp_release_deps(kmp_int32 gtid, kmp_taskdata_t *task) {
387   kmp_info_t *thread = __kmp_threads[gtid];
388   kmp_depnode_t *node = task->td_depnode;
389 
390   if (task->td_dephash) {
391     KA_TRACE(
392         40, ("__kmp_release_deps: T#%d freeing dependencies hash of task %p.\n",
393              gtid, task));
394     __kmp_dephash_free(thread, task->td_dephash);
395     task->td_dephash = NULL;
396   }
397 
398   if (!node)
399     return;
400 
401   KA_TRACE(20, ("__kmp_release_deps: T#%d notifying successors of task %p.\n",
402                 gtid, task));
403 
404   KMP_ACQUIRE_DEPNODE(gtid, node);
405   node->dn.task =
406       NULL; // mark this task as finished, so no new dependencies are generated
407   KMP_RELEASE_DEPNODE(gtid, node);
408 
409   kmp_depnode_list_t *next;
410   for (kmp_depnode_list_t *p = node->dn.successors; p; p = next) {
411     kmp_depnode_t *successor = p->node;
412     kmp_int32 npredecessors =
413         KMP_TEST_THEN_DEC32(&successor->dn.npredecessors) - 1;
414 
415     // successor task can be NULL for wait_depends or because deps are still
416     // being processed
417     if (npredecessors == 0) {
418       KMP_MB();
419       if (successor->dn.task) {
420         KA_TRACE(20, ("__kmp_release_deps: T#%d successor %p of %p scheduled "
421                       "for execution.\n",
422                       gtid, successor->dn.task, task));
423         __kmp_omp_task(gtid, successor->dn.task, false);
424       }
425     }
426 
427     next = p->next;
428     __kmp_node_deref(thread, p->node);
429 #if USE_FAST_MEMORY
430     __kmp_fast_free(thread, p);
431 #else
432     __kmp_thread_free(thread, p);
433 #endif
434   }
435 
436   __kmp_node_deref(thread, node);
437 
438   KA_TRACE(
439       20,
440       ("__kmp_release_deps: T#%d all successors of %p notified of completion\n",
441        gtid, task));
442 }
443 
444 /*!
445 @ingroup TASKING
446 @param loc_ref location of the original task directive
447 @param gtid Global Thread ID of encountering thread
448 @param new_task task thunk allocated by __kmp_omp_task_alloc() for the ''new
449 task''
450 @param ndeps Number of depend items with possible aliasing
451 @param dep_list List of depend items with possible aliasing
452 @param ndeps_noalias Number of depend items with no aliasing
453 @param noalias_dep_list List of depend items with no aliasing
454 
455 @return Returns either TASK_CURRENT_NOT_QUEUED if the current task was not
456 suspendend and queued, or TASK_CURRENT_QUEUED if it was suspended and queued
457 
458 Schedule a non-thread-switchable task with dependences for execution
459 */
460 kmp_int32 __kmpc_omp_task_with_deps(ident_t *loc_ref, kmp_int32 gtid,
461                                     kmp_task_t *new_task, kmp_int32 ndeps,
462                                     kmp_depend_info_t *dep_list,
463                                     kmp_int32 ndeps_noalias,
464                                     kmp_depend_info_t *noalias_dep_list) {
465 
466   kmp_taskdata_t *new_taskdata = KMP_TASK_TO_TASKDATA(new_task);
467   KA_TRACE(10, ("__kmpc_omp_task_with_deps(enter): T#%d loc=%p task=%p\n", gtid,
468                 loc_ref, new_taskdata));
469 
470   kmp_info_t *thread = __kmp_threads[gtid];
471   kmp_taskdata_t *current_task = thread->th.th_current_task;
472 
473 #if OMPT_SUPPORT && OMPT_TRACE
474   /* OMPT grab all dependences if requested by the tool */
475   if (ompt_enabled && ndeps + ndeps_noalias > 0 &&
476       ompt_callbacks.ompt_callback(ompt_event_task_dependences)) {
477     kmp_int32 i;
478 
479     new_taskdata->ompt_task_info.ndeps = ndeps + ndeps_noalias;
480     new_taskdata->ompt_task_info.deps =
481         (ompt_task_dependence_t *)KMP_OMPT_DEPS_ALLOC(
482             thread, (ndeps + ndeps_noalias) * sizeof(ompt_task_dependence_t));
483 
484     KMP_ASSERT(new_taskdata->ompt_task_info.deps != NULL);
485 
486     for (i = 0; i < ndeps; i++) {
487       new_taskdata->ompt_task_info.deps[i].variable_addr =
488           (void *)dep_list[i].base_addr;
489       if (dep_list[i].flags.in && dep_list[i].flags.out)
490         new_taskdata->ompt_task_info.deps[i].dependence_flags =
491             ompt_task_dependence_type_inout;
492       else if (dep_list[i].flags.out)
493         new_taskdata->ompt_task_info.deps[i].dependence_flags =
494             ompt_task_dependence_type_out;
495       else if (dep_list[i].flags.in)
496         new_taskdata->ompt_task_info.deps[i].dependence_flags =
497             ompt_task_dependence_type_in;
498     }
499     for (i = 0; i < ndeps_noalias; i++) {
500       new_taskdata->ompt_task_info.deps[ndeps + i].variable_addr =
501           (void *)noalias_dep_list[i].base_addr;
502       if (noalias_dep_list[i].flags.in && noalias_dep_list[i].flags.out)
503         new_taskdata->ompt_task_info.deps[ndeps + i].dependence_flags =
504             ompt_task_dependence_type_inout;
505       else if (noalias_dep_list[i].flags.out)
506         new_taskdata->ompt_task_info.deps[ndeps + i].dependence_flags =
507             ompt_task_dependence_type_out;
508       else if (noalias_dep_list[i].flags.in)
509         new_taskdata->ompt_task_info.deps[ndeps + i].dependence_flags =
510             ompt_task_dependence_type_in;
511     }
512   }
513 #endif /* OMPT_SUPPORT && OMPT_TRACE */
514 
515   bool serial = current_task->td_flags.team_serial ||
516                 current_task->td_flags.tasking_ser ||
517                 current_task->td_flags.final;
518 #if OMP_45_ENABLED
519   kmp_task_team_t *task_team = thread->th.th_task_team;
520   serial = serial && !(task_team && task_team->tt.tt_found_proxy_tasks);
521 #endif
522 
523   if (!serial && (ndeps > 0 || ndeps_noalias > 0)) {
524     /* if no dependencies have been tracked yet, create the dependence hash */
525     if (current_task->td_dephash == NULL)
526       current_task->td_dephash = __kmp_dephash_create(thread, current_task);
527 
528 #if USE_FAST_MEMORY
529     kmp_depnode_t *node =
530         (kmp_depnode_t *)__kmp_fast_allocate(thread, sizeof(kmp_depnode_t));
531 #else
532     kmp_depnode_t *node =
533         (kmp_depnode_t *)__kmp_thread_malloc(thread, sizeof(kmp_depnode_t));
534 #endif
535 
536     __kmp_init_node(node);
537     new_taskdata->td_depnode = node;
538 
539     if (__kmp_check_deps(gtid, node, new_task, current_task->td_dephash,
540                          NO_DEP_BARRIER, ndeps, dep_list, ndeps_noalias,
541                          noalias_dep_list)) {
542       KA_TRACE(10, ("__kmpc_omp_task_with_deps(exit): T#%d task had blocking "
543                     "dependencies: "
544                     "loc=%p task=%p, return: TASK_CURRENT_NOT_QUEUED\n",
545                     gtid, loc_ref, new_taskdata));
546       return TASK_CURRENT_NOT_QUEUED;
547     }
548   } else {
549     KA_TRACE(10, ("__kmpc_omp_task_with_deps(exit): T#%d ignored dependencies "
550                   "for task (serialized)"
551                   "loc=%p task=%p\n",
552                   gtid, loc_ref, new_taskdata));
553   }
554 
555   KA_TRACE(10, ("__kmpc_omp_task_with_deps(exit): T#%d task had no blocking "
556                 "dependencies : "
557                 "loc=%p task=%p, transferring to __kmpc_omp_task\n",
558                 gtid, loc_ref, new_taskdata));
559 
560   return __kmpc_omp_task(loc_ref, gtid, new_task);
561 }
562 
563 /*!
564 @ingroup TASKING
565 @param loc_ref location of the original task directive
566 @param gtid Global Thread ID of encountering thread
567 @param ndeps Number of depend items with possible aliasing
568 @param dep_list List of depend items with possible aliasing
569 @param ndeps_noalias Number of depend items with no aliasing
570 @param noalias_dep_list List of depend items with no aliasing
571 
572 Blocks the current task until all specifies dependencies have been fulfilled.
573 */
574 void __kmpc_omp_wait_deps(ident_t *loc_ref, kmp_int32 gtid, kmp_int32 ndeps,
575                           kmp_depend_info_t *dep_list, kmp_int32 ndeps_noalias,
576                           kmp_depend_info_t *noalias_dep_list) {
577   KA_TRACE(10, ("__kmpc_omp_wait_deps(enter): T#%d loc=%p\n", gtid, loc_ref));
578 
579   if (ndeps == 0 && ndeps_noalias == 0) {
580     KA_TRACE(10, ("__kmpc_omp_wait_deps(exit): T#%d has no dependencies to "
581                   "wait upon : loc=%p\n",
582                   gtid, loc_ref));
583     return;
584   }
585 
586   kmp_info_t *thread = __kmp_threads[gtid];
587   kmp_taskdata_t *current_task = thread->th.th_current_task;
588 
589   // We can return immediately as:
590   // - dependences are not computed in serial teams (except with proxy tasks)
591   // - if the dephash is not yet created it means we have nothing to wait for
592   bool ignore = current_task->td_flags.team_serial ||
593                 current_task->td_flags.tasking_ser ||
594                 current_task->td_flags.final;
595 #if OMP_45_ENABLED
596   ignore = ignore && thread->th.th_task_team != NULL &&
597            thread->th.th_task_team->tt.tt_found_proxy_tasks == FALSE;
598 #endif
599   ignore = ignore || current_task->td_dephash == NULL;
600 
601   if (ignore) {
602     KA_TRACE(10, ("__kmpc_omp_wait_deps(exit): T#%d has no blocking "
603                   "dependencies : loc=%p\n",
604                   gtid, loc_ref));
605     return;
606   }
607 
608   kmp_depnode_t node;
609   __kmp_init_node(&node);
610 
611   if (!__kmp_check_deps(gtid, &node, NULL, current_task->td_dephash,
612                         DEP_BARRIER, ndeps, dep_list, ndeps_noalias,
613                         noalias_dep_list)) {
614     KA_TRACE(10, ("__kmpc_omp_wait_deps(exit): T#%d has no blocking "
615                   "dependencies : loc=%p\n",
616                   gtid, loc_ref));
617     return;
618   }
619 
620   int thread_finished = FALSE;
621   kmp_flag_32 flag((volatile kmp_uint32 *)&(node.dn.npredecessors), 0U);
622   while (node.dn.npredecessors > 0) {
623     flag.execute_tasks(thread, gtid, FALSE, &thread_finished,
624 #if USE_ITT_BUILD
625                        NULL,
626 #endif
627                        __kmp_task_stealing_constraint);
628   }
629 
630   KA_TRACE(10, ("__kmpc_omp_wait_deps(exit): T#%d finished waiting : loc=%p\n",
631                 gtid, loc_ref));
632 }
633 
634 #endif /* OMP_40_ENABLED */
635