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