1 /*
2  * kmp_taskdeps.cpp
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 //#define KMP_SUPPORT_GRAPH_OUTPUT 1
14 
15 #include "kmp.h"
16 #include "kmp_io.h"
17 #include "kmp_wait_release.h"
18 #include "kmp_taskdeps.h"
19 #if OMPT_SUPPORT
20 #include "ompt-specific.h"
21 #endif
22 
23 // TODO: Improve memory allocation? keep a list of pre-allocated structures?
24 // 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
29 // runtime locks
30 // TODO: Any ITT support needed?
31 
32 #ifdef KMP_SUPPORT_GRAPH_OUTPUT
33 static std::atomic<kmp_int32> kmp_node_id_seed = ATOMIC_VAR_INIT(0);
34 #endif
35 
36 static void __kmp_init_node(kmp_depnode_t *node) {
37   node->dn.successors = NULL;
38   node->dn.task = NULL; // will point to the right task
39   // once dependences have been processed
40   for (int i = 0; i < MAX_MTX_DEPS; ++i)
41     node->dn.mtx_locks[i] = NULL;
42   node->dn.mtx_num_locks = 0;
43   __kmp_init_lock(&node->dn.lock);
44   KMP_ATOMIC_ST_RLX(&node->dn.nrefs, 1); // init creates the first reference
45 #ifdef KMP_SUPPORT_GRAPH_OUTPUT
46   node->dn.id = KMP_ATOMIC_INC(&kmp_node_id_seed);
47 #endif
48 }
49 
50 static inline kmp_depnode_t *__kmp_node_ref(kmp_depnode_t *node) {
51   KMP_ATOMIC_INC(&node->dn.nrefs);
52   return node;
53 }
54 
55 enum { KMP_DEPHASH_OTHER_SIZE = 97, KMP_DEPHASH_MASTER_SIZE = 997 };
56 
57 size_t sizes[] = {997, 2003, 4001, 8191, 16001, 32003, 64007, 131071, 270029};
58 const size_t MAX_GEN = 8;
59 
60 static inline size_t __kmp_dephash_hash(kmp_intptr_t addr, size_t hsize) {
61   // TODO alternate to try: set = (((Addr64)(addrUsefulBits * 9.618)) %
62   // m_num_sets );
63   return ((addr >> 6) ^ (addr >> 2)) % hsize;
64 }
65 
66 static kmp_dephash_t *__kmp_dephash_extend(kmp_info_t *thread,
67                                            kmp_dephash_t *current_dephash) {
68   kmp_dephash_t *h;
69 
70   size_t gen = current_dephash->generation + 1;
71   if (gen >= MAX_GEN)
72     return current_dephash;
73   size_t new_size = sizes[gen];
74 
75   size_t size_to_allocate =
76       new_size * sizeof(kmp_dephash_entry_t *) + sizeof(kmp_dephash_t);
77 
78 #if USE_FAST_MEMORY
79   h = (kmp_dephash_t *)__kmp_fast_allocate(thread, size_to_allocate);
80 #else
81   h = (kmp_dephash_t *)__kmp_thread_malloc(thread, size_to_allocate);
82 #endif
83 
84   h->size = new_size;
85   h->nelements = current_dephash->nelements;
86   h->buckets = (kmp_dephash_entry **)(h + 1);
87   h->generation = gen;
88   h->nconflicts = 0;
89 
90   // make sure buckets are properly initialized
91   for (size_t i = 0; i < new_size; i++) {
92     h->buckets[i] = NULL;
93   }
94 
95   // insert existing elements in the new table
96   for (size_t i = 0; i < current_dephash->size; i++) {
97     kmp_dephash_entry_t *next, *entry;
98     for (entry = current_dephash->buckets[i]; entry; entry = next) {
99       next = entry->next_in_bucket;
100       // Compute the new hash using the new size, and insert the entry in
101       // the new bucket.
102       size_t new_bucket = __kmp_dephash_hash(entry->addr, h->size);
103       entry->next_in_bucket = h->buckets[new_bucket];
104       if (entry->next_in_bucket) {
105         h->nconflicts++;
106       }
107       h->buckets[new_bucket] = entry;
108     }
109   }
110 
111   // Free old hash table
112 #if USE_FAST_MEMORY
113   __kmp_fast_free(thread, current_dephash);
114 #else
115   __kmp_thread_free(thread, current_dephash);
116 #endif
117 
118   return h;
119 }
120 
121 static kmp_dephash_t *__kmp_dephash_create(kmp_info_t *thread,
122                                            kmp_taskdata_t *current_task) {
123   kmp_dephash_t *h;
124 
125   size_t h_size;
126 
127   if (current_task->td_flags.tasktype == TASK_IMPLICIT)
128     h_size = KMP_DEPHASH_MASTER_SIZE;
129   else
130     h_size = KMP_DEPHASH_OTHER_SIZE;
131 
132   size_t size = h_size * sizeof(kmp_dephash_entry_t *) + sizeof(kmp_dephash_t);
133 
134 #if USE_FAST_MEMORY
135   h = (kmp_dephash_t *)__kmp_fast_allocate(thread, size);
136 #else
137   h = (kmp_dephash_t *)__kmp_thread_malloc(thread, size);
138 #endif
139   h->size = h_size;
140 
141   h->generation = 0;
142   h->nelements = 0;
143   h->nconflicts = 0;
144   h->buckets = (kmp_dephash_entry **)(h + 1);
145 
146   for (size_t i = 0; i < h_size; i++)
147     h->buckets[i] = 0;
148 
149   return h;
150 }
151 
152 static kmp_dephash_entry *__kmp_dephash_find(kmp_info_t *thread,
153                                              kmp_dephash_t **hash,
154                                              kmp_intptr_t addr) {
155   kmp_dephash_t *h = *hash;
156   if (h->nelements != 0 && h->nconflicts / h->size >= 1) {
157     *hash = __kmp_dephash_extend(thread, h);
158     h = *hash;
159   }
160   size_t bucket = __kmp_dephash_hash(addr, h->size);
161 
162   kmp_dephash_entry_t *entry;
163   for (entry = h->buckets[bucket]; entry; entry = entry->next_in_bucket)
164     if (entry->addr == addr)
165       break;
166 
167   if (entry == NULL) {
168 // create entry. This is only done by one thread so no locking required
169 #if USE_FAST_MEMORY
170     entry = (kmp_dephash_entry_t *)__kmp_fast_allocate(
171         thread, sizeof(kmp_dephash_entry_t));
172 #else
173     entry = (kmp_dephash_entry_t *)__kmp_thread_malloc(
174         thread, sizeof(kmp_dephash_entry_t));
175 #endif
176     entry->addr = addr;
177     entry->last_out = NULL;
178     entry->last_set = NULL;
179     entry->prev_set = NULL;
180     entry->last_flag = 0;
181     entry->mtx_lock = NULL;
182     entry->next_in_bucket = h->buckets[bucket];
183     h->buckets[bucket] = entry;
184     h->nelements++;
185     if (entry->next_in_bucket)
186       h->nconflicts++;
187   }
188   return entry;
189 }
190 
191 static kmp_depnode_list_t *__kmp_add_node(kmp_info_t *thread,
192                                           kmp_depnode_list_t *list,
193                                           kmp_depnode_t *node) {
194   kmp_depnode_list_t *new_head;
195 
196 #if USE_FAST_MEMORY
197   new_head = (kmp_depnode_list_t *)__kmp_fast_allocate(
198       thread, sizeof(kmp_depnode_list_t));
199 #else
200   new_head = (kmp_depnode_list_t *)__kmp_thread_malloc(
201       thread, sizeof(kmp_depnode_list_t));
202 #endif
203 
204   new_head->node = __kmp_node_ref(node);
205   new_head->next = list;
206 
207   return new_head;
208 }
209 
210 static inline void __kmp_track_dependence(kmp_int32 gtid, kmp_depnode_t *source,
211                                           kmp_depnode_t *sink,
212                                           kmp_task_t *sink_task) {
213 #ifdef KMP_SUPPORT_GRAPH_OUTPUT
214   kmp_taskdata_t *task_source = KMP_TASK_TO_TASKDATA(source->dn.task);
215   // do not use sink->dn.task as that is only filled after the dependences
216   // are already processed!
217   kmp_taskdata_t *task_sink = KMP_TASK_TO_TASKDATA(sink_task);
218 
219   __kmp_printf("%d(%s) -> %d(%s)\n", source->dn.id,
220                task_source->td_ident->psource, sink->dn.id,
221                task_sink->td_ident->psource);
222 #endif
223 #if OMPT_SUPPORT && OMPT_OPTIONAL
224   /* OMPT tracks dependences between task (a=source, b=sink) in which
225      task a blocks the execution of b through the ompt_new_dependence_callback
226      */
227   if (ompt_enabled.ompt_callback_task_dependence) {
228     kmp_taskdata_t *task_source = KMP_TASK_TO_TASKDATA(source->dn.task);
229     ompt_data_t *sink_data;
230     if (sink_task)
231       sink_data = &(KMP_TASK_TO_TASKDATA(sink_task)->ompt_task_info.task_data);
232     else
233       sink_data = &__kmp_threads[gtid]->th.ompt_thread_info.task_data;
234 
235     ompt_callbacks.ompt_callback(ompt_callback_task_dependence)(
236         &(task_source->ompt_task_info.task_data), sink_data);
237   }
238 #endif /* OMPT_SUPPORT && OMPT_OPTIONAL */
239 }
240 
241 static inline kmp_int32
242 __kmp_depnode_link_successor(kmp_int32 gtid, kmp_info_t *thread,
243                              kmp_task_t *task, kmp_depnode_t *node,
244                              kmp_depnode_list_t *plist) {
245   if (!plist)
246     return 0;
247   kmp_int32 npredecessors = 0;
248   // link node as successor of list elements
249   for (kmp_depnode_list_t *p = plist; p; p = p->next) {
250     kmp_depnode_t *dep = p->node;
251     if (dep->dn.task) {
252       KMP_ACQUIRE_DEPNODE(gtid, dep);
253       if (dep->dn.task) {
254         __kmp_track_dependence(gtid, dep, node, task);
255         dep->dn.successors = __kmp_add_node(thread, dep->dn.successors, node);
256         KA_TRACE(40, ("__kmp_process_deps: T#%d adding dependence from %p to "
257                       "%p\n",
258                       gtid, KMP_TASK_TO_TASKDATA(dep->dn.task),
259                       KMP_TASK_TO_TASKDATA(task)));
260         npredecessors++;
261       }
262       KMP_RELEASE_DEPNODE(gtid, dep);
263     }
264   }
265   return npredecessors;
266 }
267 
268 static inline kmp_int32 __kmp_depnode_link_successor(kmp_int32 gtid,
269                                                      kmp_info_t *thread,
270                                                      kmp_task_t *task,
271                                                      kmp_depnode_t *source,
272                                                      kmp_depnode_t *sink) {
273   if (!sink)
274     return 0;
275   kmp_int32 npredecessors = 0;
276   if (sink->dn.task) {
277     // synchronously add source to sink' list of successors
278     KMP_ACQUIRE_DEPNODE(gtid, sink);
279     if (sink->dn.task) {
280       __kmp_track_dependence(gtid, sink, source, task);
281       sink->dn.successors = __kmp_add_node(thread, sink->dn.successors, source);
282       KA_TRACE(40, ("__kmp_process_deps: T#%d adding dependence from %p to "
283                     "%p\n",
284                     gtid, KMP_TASK_TO_TASKDATA(sink->dn.task),
285                     KMP_TASK_TO_TASKDATA(task)));
286       npredecessors++;
287     }
288     KMP_RELEASE_DEPNODE(gtid, sink);
289   }
290   return npredecessors;
291 }
292 
293 template <bool filter>
294 static inline kmp_int32
295 __kmp_process_deps(kmp_int32 gtid, kmp_depnode_t *node, kmp_dephash_t **hash,
296                    bool dep_barrier, kmp_int32 ndeps,
297                    kmp_depend_info_t *dep_list, kmp_task_t *task) {
298   KA_TRACE(30, ("__kmp_process_deps<%d>: T#%d processing %d dependences : "
299                 "dep_barrier = %d\n",
300                 filter, gtid, ndeps, dep_barrier));
301 
302   kmp_info_t *thread = __kmp_threads[gtid];
303   kmp_int32 npredecessors = 0;
304   for (kmp_int32 i = 0; i < ndeps; i++) {
305     const kmp_depend_info_t *dep = &dep_list[i];
306 
307     if (filter && dep->base_addr == 0)
308       continue; // skip filtered entries
309 
310     kmp_dephash_entry_t *info =
311         __kmp_dephash_find(thread, hash, dep->base_addr);
312     kmp_depnode_t *last_out = info->last_out;
313     kmp_depnode_list_t *last_set = info->last_set;
314     kmp_depnode_list_t *prev_set = info->prev_set;
315 
316     if (dep->flags.out) { // out or inout --> clean lists if any
317       if (last_set) {
318         npredecessors +=
319             __kmp_depnode_link_successor(gtid, thread, task, node, last_set);
320         __kmp_depnode_list_free(thread, last_set);
321         __kmp_depnode_list_free(thread, prev_set);
322         info->last_set = NULL;
323         info->prev_set = NULL;
324         info->last_flag = 0; // no sets in this dephash entry
325       } else {
326         npredecessors +=
327             __kmp_depnode_link_successor(gtid, thread, task, node, last_out);
328       }
329       __kmp_node_deref(thread, last_out);
330       if (!dep_barrier) {
331         info->last_out = __kmp_node_ref(node);
332       } else {
333         // if this is a sync point in the serial sequence, then the previous
334         // outputs are guaranteed to be completed after the execution of this
335         // task so the previous output nodes can be cleared.
336         info->last_out = NULL;
337       }
338     } else { // either IN or MTX or SET
339       if (info->last_flag == 0 || info->last_flag == dep->flag) {
340         // last_set either didn't exist or of same dep kind
341         // link node as successor of the last_out if any
342         npredecessors +=
343             __kmp_depnode_link_successor(gtid, thread, task, node, last_out);
344         // link node as successor of all nodes in the prev_set if any
345         npredecessors +=
346             __kmp_depnode_link_successor(gtid, thread, task, node, prev_set);
347       } else { // last_set is of different dep kind, make it prev_set
348         // link node as successor of all nodes in the last_set
349         npredecessors +=
350             __kmp_depnode_link_successor(gtid, thread, task, node, last_set);
351         // clean last_out if any
352         __kmp_node_deref(thread, last_out);
353         info->last_out = NULL;
354         // clean prev_set if any
355         __kmp_depnode_list_free(thread, prev_set);
356         // move last_set to prev_set, new last_set will be allocated
357         info->prev_set = last_set;
358         info->last_set = NULL;
359       }
360       info->last_flag = dep->flag; // store dep kind of the last_set
361       info->last_set = __kmp_add_node(thread, info->last_set, node);
362 
363       // check if we are processing MTX dependency
364       if (dep->flag == KMP_DEP_MTX) {
365         if (info->mtx_lock == NULL) {
366           info->mtx_lock = (kmp_lock_t *)__kmp_allocate(sizeof(kmp_lock_t));
367           __kmp_init_lock(info->mtx_lock);
368         }
369         KMP_DEBUG_ASSERT(node->dn.mtx_num_locks < MAX_MTX_DEPS);
370         kmp_int32 m;
371         // Save lock in node's array
372         for (m = 0; m < MAX_MTX_DEPS; ++m) {
373           // sort pointers in decreasing order to avoid potential livelock
374           if (node->dn.mtx_locks[m] < info->mtx_lock) {
375             KMP_DEBUG_ASSERT(!node->dn.mtx_locks[node->dn.mtx_num_locks]);
376             for (int n = node->dn.mtx_num_locks; n > m; --n) {
377               // shift right all lesser non-NULL pointers
378               KMP_DEBUG_ASSERT(node->dn.mtx_locks[n - 1] != NULL);
379               node->dn.mtx_locks[n] = node->dn.mtx_locks[n - 1];
380             }
381             node->dn.mtx_locks[m] = info->mtx_lock;
382             break;
383           }
384         }
385         KMP_DEBUG_ASSERT(m < MAX_MTX_DEPS); // must break from loop
386         node->dn.mtx_num_locks++;
387       }
388     }
389   }
390   KA_TRACE(30, ("__kmp_process_deps<%d>: T#%d found %d predecessors\n", filter,
391                 gtid, npredecessors));
392   return npredecessors;
393 }
394 
395 #define NO_DEP_BARRIER (false)
396 #define DEP_BARRIER (true)
397 
398 // returns true if the task has any outstanding dependence
399 static bool __kmp_check_deps(kmp_int32 gtid, kmp_depnode_t *node,
400                              kmp_task_t *task, kmp_dephash_t **hash,
401                              bool dep_barrier, kmp_int32 ndeps,
402                              kmp_depend_info_t *dep_list,
403                              kmp_int32 ndeps_noalias,
404                              kmp_depend_info_t *noalias_dep_list) {
405   int i, n_mtxs = 0;
406 #if KMP_DEBUG
407   kmp_taskdata_t *taskdata = KMP_TASK_TO_TASKDATA(task);
408 #endif
409   KA_TRACE(20, ("__kmp_check_deps: T#%d checking dependences for task %p : %d "
410                 "possibly aliased dependences, %d non-aliased dependences : "
411                 "dep_barrier=%d .\n",
412                 gtid, taskdata, ndeps, ndeps_noalias, dep_barrier));
413 
414   // Filter deps in dep_list
415   // TODO: Different algorithm for large dep_list ( > 10 ? )
416   for (i = 0; i < ndeps; i++) {
417     if (dep_list[i].base_addr != 0) {
418       KMP_DEBUG_ASSERT(
419           dep_list[i].flag == KMP_DEP_IN || dep_list[i].flag == KMP_DEP_OUT ||
420           dep_list[i].flag == KMP_DEP_INOUT ||
421           dep_list[i].flag == KMP_DEP_MTX || dep_list[i].flag == KMP_DEP_SET);
422       for (int j = i + 1; j < ndeps; j++) {
423         if (dep_list[i].base_addr == dep_list[j].base_addr) {
424           if (dep_list[i].flag != dep_list[j].flag) {
425             // two different dependences on same address work identical to OUT
426             dep_list[i].flag = KMP_DEP_OUT;
427           }
428           dep_list[j].base_addr = 0; // Mark j element as void
429         }
430       }
431       if (dep_list[i].flag == KMP_DEP_MTX) {
432         // limit number of mtx deps to MAX_MTX_DEPS per node
433         if (n_mtxs < MAX_MTX_DEPS && task != NULL) {
434           ++n_mtxs;
435         } else {
436           dep_list[i].flag = KMP_DEP_OUT; // downgrade mutexinoutset to inout
437         }
438       }
439     }
440   }
441 
442   // doesn't need to be atomic as no other thread is going to be accessing this
443   // node just yet.
444   // npredecessors is set -1 to ensure that none of the releasing tasks queues
445   // this task before we have finished processing all the dependences
446   node->dn.npredecessors = -1;
447 
448   // used to pack all npredecessors additions into a single atomic operation at
449   // the end
450   int npredecessors;
451 
452   npredecessors = __kmp_process_deps<true>(gtid, node, hash, dep_barrier, ndeps,
453                                            dep_list, task);
454   npredecessors += __kmp_process_deps<false>(
455       gtid, node, hash, dep_barrier, ndeps_noalias, noalias_dep_list, task);
456 
457   node->dn.task = task;
458   KMP_MB();
459 
460   // Account for our initial fake value
461   npredecessors++;
462 
463   // Update predecessors and obtain current value to check if there are still
464   // any outstanding dependences (some tasks may have finished while we
465   // processed the dependences)
466   npredecessors =
467       node->dn.npredecessors.fetch_add(npredecessors) + npredecessors;
468 
469   KA_TRACE(20, ("__kmp_check_deps: T#%d found %d predecessors for task %p \n",
470                 gtid, npredecessors, taskdata));
471 
472   // beyond this point the task could be queued (and executed) by a releasing
473   // task...
474   return npredecessors > 0 ? true : false;
475 }
476 
477 /*!
478 @ingroup TASKING
479 @param loc_ref location of the original task directive
480 @param gtid Global Thread ID of encountering thread
481 @param new_task task thunk allocated by __kmp_omp_task_alloc() for the ''new
482 task''
483 @param ndeps Number of depend items with possible aliasing
484 @param dep_list List of depend items with possible aliasing
485 @param ndeps_noalias Number of depend items with no aliasing
486 @param noalias_dep_list List of depend items with no aliasing
487 
488 @return Returns either TASK_CURRENT_NOT_QUEUED if the current task was not
489 suspended and queued, or TASK_CURRENT_QUEUED if it was suspended and queued
490 
491 Schedule a non-thread-switchable task with dependences for execution
492 */
493 kmp_int32 __kmpc_omp_task_with_deps(ident_t *loc_ref, kmp_int32 gtid,
494                                     kmp_task_t *new_task, kmp_int32 ndeps,
495                                     kmp_depend_info_t *dep_list,
496                                     kmp_int32 ndeps_noalias,
497                                     kmp_depend_info_t *noalias_dep_list) {
498 
499   kmp_taskdata_t *new_taskdata = KMP_TASK_TO_TASKDATA(new_task);
500   KA_TRACE(10, ("__kmpc_omp_task_with_deps(enter): T#%d loc=%p task=%p\n", gtid,
501                 loc_ref, new_taskdata));
502   __kmp_assert_valid_gtid(gtid);
503   kmp_info_t *thread = __kmp_threads[gtid];
504   kmp_taskdata_t *current_task = thread->th.th_current_task;
505 
506 #if OMPT_SUPPORT
507   if (ompt_enabled.enabled) {
508     if (!current_task->ompt_task_info.frame.enter_frame.ptr)
509       current_task->ompt_task_info.frame.enter_frame.ptr =
510           OMPT_GET_FRAME_ADDRESS(0);
511     if (ompt_enabled.ompt_callback_task_create) {
512       ompt_callbacks.ompt_callback(ompt_callback_task_create)(
513           &(current_task->ompt_task_info.task_data),
514           &(current_task->ompt_task_info.frame),
515           &(new_taskdata->ompt_task_info.task_data),
516           ompt_task_explicit | TASK_TYPE_DETAILS_FORMAT(new_taskdata), 1,
517           OMPT_LOAD_OR_GET_RETURN_ADDRESS(gtid));
518     }
519 
520     new_taskdata->ompt_task_info.frame.enter_frame.ptr =
521         OMPT_GET_FRAME_ADDRESS(0);
522   }
523 
524 #if OMPT_OPTIONAL
525   /* OMPT grab all dependences if requested by the tool */
526   if (ndeps + ndeps_noalias > 0 && ompt_enabled.ompt_callback_dependences) {
527     kmp_int32 i;
528 
529     int ompt_ndeps = ndeps + ndeps_noalias;
530     ompt_dependence_t *ompt_deps = (ompt_dependence_t *)KMP_OMPT_DEPS_ALLOC(
531         thread, (ndeps + ndeps_noalias) * sizeof(ompt_dependence_t));
532 
533     KMP_ASSERT(ompt_deps != NULL);
534 
535     for (i = 0; i < ndeps; i++) {
536       ompt_deps[i].variable.ptr = (void *)dep_list[i].base_addr;
537       if (dep_list[i].flags.in && dep_list[i].flags.out)
538         ompt_deps[i].dependence_type = ompt_dependence_type_inout;
539       else if (dep_list[i].flags.out)
540         ompt_deps[i].dependence_type = ompt_dependence_type_out;
541       else if (dep_list[i].flags.in)
542         ompt_deps[i].dependence_type = ompt_dependence_type_in;
543       else if (dep_list[i].flags.mtx)
544         ompt_deps[i].dependence_type = ompt_dependence_type_mutexinoutset;
545       else if (dep_list[i].flags.set)
546         ompt_deps[i].dependence_type = ompt_dependence_type_inoutset;
547     }
548     for (i = 0; i < ndeps_noalias; i++) {
549       ompt_deps[ndeps + i].variable.ptr = (void *)noalias_dep_list[i].base_addr;
550       if (noalias_dep_list[i].flags.in && noalias_dep_list[i].flags.out)
551         ompt_deps[ndeps + i].dependence_type = ompt_dependence_type_inout;
552       else if (noalias_dep_list[i].flags.out)
553         ompt_deps[ndeps + i].dependence_type = ompt_dependence_type_out;
554       else if (noalias_dep_list[i].flags.in)
555         ompt_deps[ndeps + i].dependence_type = ompt_dependence_type_in;
556       else if (noalias_dep_list[i].flags.mtx)
557         ompt_deps[ndeps + i].dependence_type =
558             ompt_dependence_type_mutexinoutset;
559       else if (noalias_dep_list[i].flags.set)
560         ompt_deps[ndeps + i].dependence_type = ompt_dependence_type_inoutset;
561     }
562     ompt_callbacks.ompt_callback(ompt_callback_dependences)(
563         &(new_taskdata->ompt_task_info.task_data), ompt_deps, ompt_ndeps);
564     /* We can now free the allocated memory for the dependences */
565     /* For OMPD we might want to delay the free until end of this function */
566     KMP_OMPT_DEPS_FREE(thread, ompt_deps);
567   }
568 #endif /* OMPT_OPTIONAL */
569 #endif /* OMPT_SUPPORT */
570 
571   bool serial = current_task->td_flags.team_serial ||
572                 current_task->td_flags.tasking_ser ||
573                 current_task->td_flags.final;
574   kmp_task_team_t *task_team = thread->th.th_task_team;
575   serial = serial &&
576            !(task_team && (task_team->tt.tt_found_proxy_tasks ||
577                            task_team->tt.tt_hidden_helper_task_encountered));
578 
579   if (!serial && (ndeps > 0 || ndeps_noalias > 0)) {
580     /* if no dependences have been tracked yet, create the dependence hash */
581     if (current_task->td_dephash == NULL)
582       current_task->td_dephash = __kmp_dephash_create(thread, current_task);
583 
584 #if USE_FAST_MEMORY
585     kmp_depnode_t *node =
586         (kmp_depnode_t *)__kmp_fast_allocate(thread, sizeof(kmp_depnode_t));
587 #else
588     kmp_depnode_t *node =
589         (kmp_depnode_t *)__kmp_thread_malloc(thread, sizeof(kmp_depnode_t));
590 #endif
591 
592     __kmp_init_node(node);
593     new_taskdata->td_depnode = node;
594 
595     if (__kmp_check_deps(gtid, node, new_task, &current_task->td_dephash,
596                          NO_DEP_BARRIER, ndeps, dep_list, ndeps_noalias,
597                          noalias_dep_list)) {
598       KA_TRACE(10, ("__kmpc_omp_task_with_deps(exit): T#%d task had blocking "
599                     "dependences: "
600                     "loc=%p task=%p, return: TASK_CURRENT_NOT_QUEUED\n",
601                     gtid, loc_ref, new_taskdata));
602 #if OMPT_SUPPORT
603       if (ompt_enabled.enabled) {
604         current_task->ompt_task_info.frame.enter_frame = ompt_data_none;
605       }
606 #endif
607       return TASK_CURRENT_NOT_QUEUED;
608     }
609   } else {
610     KA_TRACE(10, ("__kmpc_omp_task_with_deps(exit): T#%d ignored dependences "
611                   "for task (serialized) loc=%p task=%p\n",
612                   gtid, loc_ref, new_taskdata));
613   }
614 
615   KA_TRACE(10, ("__kmpc_omp_task_with_deps(exit): T#%d task had no blocking "
616                 "dependences : "
617                 "loc=%p task=%p, transferring to __kmp_omp_task\n",
618                 gtid, loc_ref, new_taskdata));
619 
620   kmp_int32 ret = __kmp_omp_task(gtid, new_task, true);
621 #if OMPT_SUPPORT
622   if (ompt_enabled.enabled) {
623     current_task->ompt_task_info.frame.enter_frame = ompt_data_none;
624   }
625 #endif
626   return ret;
627 }
628 
629 #if OMPT_SUPPORT
630 void __ompt_taskwait_dep_finish(kmp_taskdata_t *current_task,
631                                 ompt_data_t *taskwait_task_data) {
632   if (ompt_enabled.ompt_callback_task_schedule) {
633     ompt_callbacks.ompt_callback(ompt_callback_task_schedule)(
634         taskwait_task_data, ompt_taskwait_complete, NULL);
635   }
636   current_task->ompt_task_info.frame.enter_frame.ptr = NULL;
637   *taskwait_task_data = ompt_data_none;
638 }
639 #endif /* OMPT_SUPPORT */
640 
641 /*!
642 @ingroup TASKING
643 @param loc_ref location of the original task directive
644 @param gtid Global Thread ID of encountering thread
645 @param ndeps Number of depend items with possible aliasing
646 @param dep_list List of depend items with possible aliasing
647 @param ndeps_noalias Number of depend items with no aliasing
648 @param noalias_dep_list List of depend items with no aliasing
649 
650 Blocks the current task until all specifies dependences have been fulfilled.
651 */
652 void __kmpc_omp_wait_deps(ident_t *loc_ref, kmp_int32 gtid, kmp_int32 ndeps,
653                           kmp_depend_info_t *dep_list, kmp_int32 ndeps_noalias,
654                           kmp_depend_info_t *noalias_dep_list) {
655   KA_TRACE(10, ("__kmpc_omp_wait_deps(enter): T#%d loc=%p\n", gtid, loc_ref));
656 
657   if (ndeps == 0 && ndeps_noalias == 0) {
658     KA_TRACE(10, ("__kmpc_omp_wait_deps(exit): T#%d has no dependences to "
659                   "wait upon : loc=%p\n",
660                   gtid, loc_ref));
661     return;
662   }
663   __kmp_assert_valid_gtid(gtid);
664   kmp_info_t *thread = __kmp_threads[gtid];
665   kmp_taskdata_t *current_task = thread->th.th_current_task;
666 
667 #if OMPT_SUPPORT
668   // this function represents a taskwait construct with depend clause
669   // We signal 4 events:
670   //  - creation of the taskwait task
671   //  - dependences of the taskwait task
672   //  - schedule and finish of the taskwait task
673   ompt_data_t *taskwait_task_data = &thread->th.ompt_thread_info.task_data;
674   KMP_ASSERT(taskwait_task_data->ptr == NULL);
675   if (ompt_enabled.enabled) {
676     if (!current_task->ompt_task_info.frame.enter_frame.ptr)
677       current_task->ompt_task_info.frame.enter_frame.ptr =
678           OMPT_GET_FRAME_ADDRESS(0);
679     if (ompt_enabled.ompt_callback_task_create) {
680       ompt_callbacks.ompt_callback(ompt_callback_task_create)(
681           &(current_task->ompt_task_info.task_data),
682           &(current_task->ompt_task_info.frame), taskwait_task_data,
683           ompt_task_taskwait | ompt_task_undeferred | ompt_task_mergeable, 1,
684           OMPT_LOAD_OR_GET_RETURN_ADDRESS(gtid));
685     }
686   }
687 
688 #if OMPT_OPTIONAL
689   /* OMPT grab all dependences if requested by the tool */
690   if (ndeps + ndeps_noalias > 0 && ompt_enabled.ompt_callback_dependences) {
691     kmp_int32 i;
692 
693     int ompt_ndeps = ndeps + ndeps_noalias;
694     ompt_dependence_t *ompt_deps = (ompt_dependence_t *)KMP_OMPT_DEPS_ALLOC(
695         thread, (ndeps + ndeps_noalias) * sizeof(ompt_dependence_t));
696 
697     KMP_ASSERT(ompt_deps != NULL);
698 
699     for (i = 0; i < ndeps; i++) {
700       ompt_deps[i].variable.ptr = (void *)dep_list[i].base_addr;
701       if (dep_list[i].flags.in && dep_list[i].flags.out)
702         ompt_deps[i].dependence_type = ompt_dependence_type_inout;
703       else if (dep_list[i].flags.out)
704         ompt_deps[i].dependence_type = ompt_dependence_type_out;
705       else if (dep_list[i].flags.in)
706         ompt_deps[i].dependence_type = ompt_dependence_type_in;
707       else if (dep_list[i].flags.mtx)
708         ompt_deps[ndeps + i].dependence_type =
709             ompt_dependence_type_mutexinoutset;
710       else if (dep_list[i].flags.set)
711         ompt_deps[ndeps + i].dependence_type = ompt_dependence_type_inoutset;
712     }
713     for (i = 0; i < ndeps_noalias; i++) {
714       ompt_deps[ndeps + i].variable.ptr = (void *)noalias_dep_list[i].base_addr;
715       if (noalias_dep_list[i].flags.in && noalias_dep_list[i].flags.out)
716         ompt_deps[ndeps + i].dependence_type = ompt_dependence_type_inout;
717       else if (noalias_dep_list[i].flags.out)
718         ompt_deps[ndeps + i].dependence_type = ompt_dependence_type_out;
719       else if (noalias_dep_list[i].flags.in)
720         ompt_deps[ndeps + i].dependence_type = ompt_dependence_type_in;
721       else if (noalias_dep_list[i].flags.mtx)
722         ompt_deps[ndeps + i].dependence_type =
723             ompt_dependence_type_mutexinoutset;
724       else if (noalias_dep_list[i].flags.set)
725         ompt_deps[ndeps + i].dependence_type = ompt_dependence_type_inoutset;
726     }
727     ompt_callbacks.ompt_callback(ompt_callback_dependences)(
728         taskwait_task_data, ompt_deps, ompt_ndeps);
729     /* We can now free the allocated memory for the dependences */
730     /* For OMPD we might want to delay the free until end of this function */
731     KMP_OMPT_DEPS_FREE(thread, ompt_deps);
732     ompt_deps = NULL;
733   }
734 #endif /* OMPT_OPTIONAL */
735 #endif /* OMPT_SUPPORT */
736 
737   // We can return immediately as:
738   // - dependences are not computed in serial teams (except with proxy tasks)
739   // - if the dephash is not yet created it means we have nothing to wait for
740   bool ignore = current_task->td_flags.team_serial ||
741                 current_task->td_flags.tasking_ser ||
742                 current_task->td_flags.final;
743   ignore = ignore && thread->th.th_task_team != NULL &&
744            thread->th.th_task_team->tt.tt_found_proxy_tasks == FALSE;
745   ignore = ignore || current_task->td_dephash == NULL;
746 
747   if (ignore) {
748     KA_TRACE(10, ("__kmpc_omp_wait_deps(exit): T#%d has no blocking "
749                   "dependences : loc=%p\n",
750                   gtid, loc_ref));
751 #if OMPT_SUPPORT
752     __ompt_taskwait_dep_finish(current_task, taskwait_task_data);
753 #endif /* OMPT_SUPPORT */
754     return;
755   }
756 
757   kmp_depnode_t node = {0};
758   __kmp_init_node(&node);
759   // the stack owns the node
760   __kmp_node_ref(&node);
761 
762   if (!__kmp_check_deps(gtid, &node, NULL, &current_task->td_dephash,
763                         DEP_BARRIER, ndeps, dep_list, ndeps_noalias,
764                         noalias_dep_list)) {
765     KA_TRACE(10, ("__kmpc_omp_wait_deps(exit): T#%d has no blocking "
766                   "dependences : loc=%p\n",
767                   gtid, loc_ref));
768 #if OMPT_SUPPORT
769     __ompt_taskwait_dep_finish(current_task, taskwait_task_data);
770 #endif /* OMPT_SUPPORT */
771     return;
772   }
773 
774   int thread_finished = FALSE;
775   kmp_flag_32<false, false> flag(
776       (std::atomic<kmp_uint32> *)&node.dn.npredecessors, 0U);
777   while (node.dn.npredecessors > 0) {
778     flag.execute_tasks(thread, gtid, FALSE,
779                        &thread_finished USE_ITT_BUILD_ARG(NULL),
780                        __kmp_task_stealing_constraint);
781   }
782 
783 #if OMPT_SUPPORT
784   __ompt_taskwait_dep_finish(current_task, taskwait_task_data);
785 #endif /* OMPT_SUPPORT */
786   KA_TRACE(10, ("__kmpc_omp_wait_deps(exit): T#%d finished waiting : loc=%p\n",
787                 gtid, loc_ref));
788 }
789