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