xref: /freebsd/contrib/llvm-project/openmp/runtime/src/kmp_taskdeps.cpp (revision a50d73d5782a351ad83e8d1f84d11720a12e70d3)
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 #if USE_ITT_BUILD && USE_ITT_NOTIFY
49   __itt_sync_create(node, "OMP task dep node", NULL, 0);
50 #endif
51 }
52 
53 static inline kmp_depnode_t *__kmp_node_ref(kmp_depnode_t *node) {
54   KMP_ATOMIC_INC(&node->dn.nrefs);
55   return node;
56 }
57 
58 enum { KMP_DEPHASH_OTHER_SIZE = 97, KMP_DEPHASH_MASTER_SIZE = 997 };
59 
60 size_t sizes[] = {997, 2003, 4001, 8191, 16001, 32003, 64007, 131071, 270029};
61 const size_t MAX_GEN = 8;
62 
63 static inline size_t __kmp_dephash_hash(kmp_intptr_t addr, size_t hsize) {
64   // TODO alternate to try: set = (((Addr64)(addrUsefulBits * 9.618)) %
65   // m_num_sets );
66   return ((addr >> 6) ^ (addr >> 2)) % hsize;
67 }
68 
69 static kmp_dephash_t *__kmp_dephash_extend(kmp_info_t *thread,
70                                            kmp_dephash_t *current_dephash) {
71   kmp_dephash_t *h;
72 
73   size_t gen = current_dephash->generation + 1;
74   if (gen >= MAX_GEN)
75     return current_dephash;
76   size_t new_size = sizes[gen];
77 
78   size_t size_to_allocate =
79       new_size * sizeof(kmp_dephash_entry_t *) + sizeof(kmp_dephash_t);
80 
81 #if USE_FAST_MEMORY
82   h = (kmp_dephash_t *)__kmp_fast_allocate(thread, size_to_allocate);
83 #else
84   h = (kmp_dephash_t *)__kmp_thread_malloc(thread, size_to_allocate);
85 #endif
86 
87   h->size = new_size;
88   h->nelements = current_dephash->nelements;
89   h->buckets = (kmp_dephash_entry **)(h + 1);
90   h->generation = gen;
91   h->nconflicts = 0;
92   h->last_all = current_dephash->last_all;
93 
94   // make sure buckets are properly initialized
95   for (size_t i = 0; i < new_size; i++) {
96     h->buckets[i] = NULL;
97   }
98 
99   // insert existing elements in the new table
100   for (size_t i = 0; i < current_dephash->size; i++) {
101     kmp_dephash_entry_t *next, *entry;
102     for (entry = current_dephash->buckets[i]; entry; entry = next) {
103       next = entry->next_in_bucket;
104       // Compute the new hash using the new size, and insert the entry in
105       // the new bucket.
106       size_t new_bucket = __kmp_dephash_hash(entry->addr, h->size);
107       entry->next_in_bucket = h->buckets[new_bucket];
108       if (entry->next_in_bucket) {
109         h->nconflicts++;
110       }
111       h->buckets[new_bucket] = entry;
112     }
113   }
114 
115   // Free old hash table
116 #if USE_FAST_MEMORY
117   __kmp_fast_free(thread, current_dephash);
118 #else
119   __kmp_thread_free(thread, current_dephash);
120 #endif
121 
122   return h;
123 }
124 
125 static kmp_dephash_t *__kmp_dephash_create(kmp_info_t *thread,
126                                            kmp_taskdata_t *current_task) {
127   kmp_dephash_t *h;
128 
129   size_t h_size;
130 
131   if (current_task->td_flags.tasktype == TASK_IMPLICIT)
132     h_size = KMP_DEPHASH_MASTER_SIZE;
133   else
134     h_size = KMP_DEPHASH_OTHER_SIZE;
135 
136   size_t size = h_size * sizeof(kmp_dephash_entry_t *) + sizeof(kmp_dephash_t);
137 
138 #if USE_FAST_MEMORY
139   h = (kmp_dephash_t *)__kmp_fast_allocate(thread, size);
140 #else
141   h = (kmp_dephash_t *)__kmp_thread_malloc(thread, size);
142 #endif
143   h->size = h_size;
144 
145   h->generation = 0;
146   h->nelements = 0;
147   h->nconflicts = 0;
148   h->buckets = (kmp_dephash_entry **)(h + 1);
149   h->last_all = NULL;
150 
151   for (size_t i = 0; i < h_size; i++)
152     h->buckets[i] = 0;
153 
154   return h;
155 }
156 
157 static kmp_dephash_entry *__kmp_dephash_find(kmp_info_t *thread,
158                                              kmp_dephash_t **hash,
159                                              kmp_intptr_t addr) {
160   kmp_dephash_t *h = *hash;
161   if (h->nelements != 0 && h->nconflicts / h->size >= 1) {
162     *hash = __kmp_dephash_extend(thread, h);
163     h = *hash;
164   }
165   size_t bucket = __kmp_dephash_hash(addr, h->size);
166 
167   kmp_dephash_entry_t *entry;
168   for (entry = h->buckets[bucket]; entry; entry = entry->next_in_bucket)
169     if (entry->addr == addr)
170       break;
171 
172   if (entry == NULL) {
173 // create entry. This is only done by one thread so no locking required
174 #if USE_FAST_MEMORY
175     entry = (kmp_dephash_entry_t *)__kmp_fast_allocate(
176         thread, sizeof(kmp_dephash_entry_t));
177 #else
178     entry = (kmp_dephash_entry_t *)__kmp_thread_malloc(
179         thread, sizeof(kmp_dephash_entry_t));
180 #endif
181     entry->addr = addr;
182     if (!h->last_all) // no predecessor task with omp_all_memory dependence
183       entry->last_out = NULL;
184     else // else link the omp_all_memory depnode to the new entry
185       entry->last_out = __kmp_node_ref(h->last_all);
186     entry->last_set = NULL;
187     entry->prev_set = NULL;
188     entry->last_flag = 0;
189     entry->mtx_lock = NULL;
190     entry->next_in_bucket = h->buckets[bucket];
191     h->buckets[bucket] = entry;
192     h->nelements++;
193     if (entry->next_in_bucket)
194       h->nconflicts++;
195   }
196   return entry;
197 }
198 
199 static kmp_depnode_list_t *__kmp_add_node(kmp_info_t *thread,
200                                           kmp_depnode_list_t *list,
201                                           kmp_depnode_t *node) {
202   kmp_depnode_list_t *new_head;
203 
204 #if USE_FAST_MEMORY
205   new_head = (kmp_depnode_list_t *)__kmp_fast_allocate(
206       thread, sizeof(kmp_depnode_list_t));
207 #else
208   new_head = (kmp_depnode_list_t *)__kmp_thread_malloc(
209       thread, sizeof(kmp_depnode_list_t));
210 #endif
211 
212   new_head->node = __kmp_node_ref(node);
213   new_head->next = list;
214 
215   return new_head;
216 }
217 
218 static inline void __kmp_track_dependence(kmp_int32 gtid, kmp_depnode_t *source,
219                                           kmp_depnode_t *sink,
220                                           kmp_task_t *sink_task) {
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 dependences
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,
228                task_source->td_ident->psource, sink->dn.id,
229                task_sink->td_ident->psource);
230 #endif
231 #if OMPT_SUPPORT && OMPT_OPTIONAL
232   /* OMPT tracks dependences between task (a=source, b=sink) in which
233      task a blocks the execution of b through the ompt_new_dependence_callback
234      */
235   if (ompt_enabled.ompt_callback_task_dependence) {
236     kmp_taskdata_t *task_source = KMP_TASK_TO_TASKDATA(source->dn.task);
237     ompt_data_t *sink_data;
238     if (sink_task)
239       sink_data = &(KMP_TASK_TO_TASKDATA(sink_task)->ompt_task_info.task_data);
240     else
241       sink_data = &__kmp_threads[gtid]->th.ompt_thread_info.task_data;
242 
243     ompt_callbacks.ompt_callback(ompt_callback_task_dependence)(
244         &(task_source->ompt_task_info.task_data), sink_data);
245   }
246 #endif /* OMPT_SUPPORT && OMPT_OPTIONAL */
247 }
248 
249 static inline kmp_int32
250 __kmp_depnode_link_successor(kmp_int32 gtid, kmp_info_t *thread,
251                              kmp_task_t *task, kmp_depnode_t *node,
252                              kmp_depnode_list_t *plist) {
253   if (!plist)
254     return 0;
255   kmp_int32 npredecessors = 0;
256   // link node as successor of list elements
257   for (kmp_depnode_list_t *p = plist; p; p = p->next) {
258     kmp_depnode_t *dep = p->node;
259     if (dep->dn.task) {
260       KMP_ACQUIRE_DEPNODE(gtid, dep);
261       if (dep->dn.task) {
262         __kmp_track_dependence(gtid, dep, node, task);
263         dep->dn.successors = __kmp_add_node(thread, dep->dn.successors, node);
264         KA_TRACE(40, ("__kmp_process_deps: T#%d adding dependence from %p to "
265                       "%p\n",
266                       gtid, KMP_TASK_TO_TASKDATA(dep->dn.task),
267                       KMP_TASK_TO_TASKDATA(task)));
268         npredecessors++;
269       }
270       KMP_RELEASE_DEPNODE(gtid, dep);
271     }
272   }
273   return npredecessors;
274 }
275 
276 static inline kmp_int32 __kmp_depnode_link_successor(kmp_int32 gtid,
277                                                      kmp_info_t *thread,
278                                                      kmp_task_t *task,
279                                                      kmp_depnode_t *source,
280                                                      kmp_depnode_t *sink) {
281   if (!sink)
282     return 0;
283   kmp_int32 npredecessors = 0;
284   if (sink->dn.task) {
285     // synchronously add source to sink' list of successors
286     KMP_ACQUIRE_DEPNODE(gtid, sink);
287     if (sink->dn.task) {
288       __kmp_track_dependence(gtid, sink, source, task);
289       sink->dn.successors = __kmp_add_node(thread, sink->dn.successors, source);
290       KA_TRACE(40, ("__kmp_process_deps: T#%d adding dependence from %p to "
291                     "%p\n",
292                     gtid, KMP_TASK_TO_TASKDATA(sink->dn.task),
293                     KMP_TASK_TO_TASKDATA(task)));
294       npredecessors++;
295     }
296     KMP_RELEASE_DEPNODE(gtid, sink);
297   }
298   return npredecessors;
299 }
300 
301 static inline kmp_int32
302 __kmp_process_dep_all(kmp_int32 gtid, kmp_depnode_t *node, kmp_dephash_t *h,
303                       bool dep_barrier, kmp_task_t *task) {
304   KA_TRACE(30, ("__kmp_process_dep_all: T#%d processing dep_all, "
305                 "dep_barrier = %d\n",
306                 gtid, dep_barrier));
307   kmp_info_t *thread = __kmp_threads[gtid];
308   kmp_int32 npredecessors = 0;
309 
310   // process previous omp_all_memory node if any
311   npredecessors +=
312       __kmp_depnode_link_successor(gtid, thread, task, node, h->last_all);
313   __kmp_node_deref(thread, h->last_all);
314   if (!dep_barrier) {
315     h->last_all = __kmp_node_ref(node);
316   } else {
317     // if this is a sync point in the serial sequence, then the previous
318     // outputs are guaranteed to be completed after the execution of this
319     // task so the previous output nodes can be cleared.
320     h->last_all = NULL;
321   }
322 
323   // process all regular dependences
324   for (size_t i = 0; i < h->size; i++) {
325     kmp_dephash_entry_t *info = h->buckets[i];
326     if (!info) // skip empty slots in dephash
327       continue;
328     for (; info; info = info->next_in_bucket) {
329       // for each entry the omp_all_memory works as OUT dependence
330       kmp_depnode_t *last_out = info->last_out;
331       kmp_depnode_list_t *last_set = info->last_set;
332       kmp_depnode_list_t *prev_set = info->prev_set;
333       if (last_set) {
334         npredecessors +=
335             __kmp_depnode_link_successor(gtid, thread, task, node, last_set);
336         __kmp_depnode_list_free(thread, last_set);
337         __kmp_depnode_list_free(thread, prev_set);
338         info->last_set = NULL;
339         info->prev_set = NULL;
340         info->last_flag = 0; // no sets in this dephash entry
341       } else {
342         npredecessors +=
343             __kmp_depnode_link_successor(gtid, thread, task, node, last_out);
344       }
345       __kmp_node_deref(thread, last_out);
346       if (!dep_barrier) {
347         info->last_out = __kmp_node_ref(node);
348       } else {
349         info->last_out = NULL;
350       }
351     }
352   }
353   KA_TRACE(30, ("__kmp_process_dep_all: T#%d found %d predecessors\n", gtid,
354                 npredecessors));
355   return npredecessors;
356 }
357 
358 template <bool filter>
359 static inline kmp_int32
360 __kmp_process_deps(kmp_int32 gtid, kmp_depnode_t *node, kmp_dephash_t **hash,
361                    bool dep_barrier, kmp_int32 ndeps,
362                    kmp_depend_info_t *dep_list, kmp_task_t *task) {
363   KA_TRACE(30, ("__kmp_process_deps<%d>: T#%d processing %d dependences : "
364                 "dep_barrier = %d\n",
365                 filter, gtid, ndeps, dep_barrier));
366 
367   kmp_info_t *thread = __kmp_threads[gtid];
368   kmp_int32 npredecessors = 0;
369   for (kmp_int32 i = 0; i < ndeps; i++) {
370     const kmp_depend_info_t *dep = &dep_list[i];
371 
372     if (filter && dep->base_addr == 0)
373       continue; // skip filtered entries
374 
375     kmp_dephash_entry_t *info =
376         __kmp_dephash_find(thread, hash, dep->base_addr);
377     kmp_depnode_t *last_out = info->last_out;
378     kmp_depnode_list_t *last_set = info->last_set;
379     kmp_depnode_list_t *prev_set = info->prev_set;
380 
381     if (dep->flags.out) { // out or inout --> clean lists if any
382       if (last_set) {
383         npredecessors +=
384             __kmp_depnode_link_successor(gtid, thread, task, node, last_set);
385         __kmp_depnode_list_free(thread, last_set);
386         __kmp_depnode_list_free(thread, prev_set);
387         info->last_set = NULL;
388         info->prev_set = NULL;
389         info->last_flag = 0; // no sets in this dephash entry
390       } else {
391         npredecessors +=
392             __kmp_depnode_link_successor(gtid, thread, task, node, last_out);
393       }
394       __kmp_node_deref(thread, last_out);
395       if (!dep_barrier) {
396         info->last_out = __kmp_node_ref(node);
397       } else {
398         // if this is a sync point in the serial sequence, then the previous
399         // outputs are guaranteed to be completed after the execution of this
400         // task so the previous output nodes can be cleared.
401         info->last_out = NULL;
402       }
403     } else { // either IN or MTX or SET
404       if (info->last_flag == 0 || info->last_flag == dep->flag) {
405         // last_set either didn't exist or of same dep kind
406         // link node as successor of the last_out if any
407         npredecessors +=
408             __kmp_depnode_link_successor(gtid, thread, task, node, last_out);
409         // link node as successor of all nodes in the prev_set if any
410         npredecessors +=
411             __kmp_depnode_link_successor(gtid, thread, task, node, prev_set);
412         if (dep_barrier) {
413           // clean last_out and prev_set if any; don't touch last_set
414           __kmp_node_deref(thread, last_out);
415           info->last_out = NULL;
416           __kmp_depnode_list_free(thread, prev_set);
417           info->prev_set = NULL;
418         }
419       } else { // last_set is of different dep kind, make it prev_set
420         // link node as successor of all nodes in the last_set
421         npredecessors +=
422             __kmp_depnode_link_successor(gtid, thread, task, node, last_set);
423         // clean last_out if any
424         __kmp_node_deref(thread, last_out);
425         info->last_out = NULL;
426         // clean prev_set if any
427         __kmp_depnode_list_free(thread, prev_set);
428         if (!dep_barrier) {
429           // move last_set to prev_set, new last_set will be allocated
430           info->prev_set = last_set;
431         } else {
432           info->prev_set = NULL;
433           info->last_flag = 0;
434         }
435         info->last_set = NULL;
436       }
437       // for dep_barrier last_flag value should remain:
438       // 0 if last_set is empty, unchanged otherwise
439       if (!dep_barrier) {
440         info->last_flag = dep->flag; // store dep kind of the last_set
441         info->last_set = __kmp_add_node(thread, info->last_set, node);
442       }
443       // check if we are processing MTX dependency
444       if (dep->flag == KMP_DEP_MTX) {
445         if (info->mtx_lock == NULL) {
446           info->mtx_lock = (kmp_lock_t *)__kmp_allocate(sizeof(kmp_lock_t));
447           __kmp_init_lock(info->mtx_lock);
448         }
449         KMP_DEBUG_ASSERT(node->dn.mtx_num_locks < MAX_MTX_DEPS);
450         kmp_int32 m;
451         // Save lock in node's array
452         for (m = 0; m < MAX_MTX_DEPS; ++m) {
453           // sort pointers in decreasing order to avoid potential livelock
454           if (node->dn.mtx_locks[m] < info->mtx_lock) {
455             KMP_DEBUG_ASSERT(!node->dn.mtx_locks[node->dn.mtx_num_locks]);
456             for (int n = node->dn.mtx_num_locks; n > m; --n) {
457               // shift right all lesser non-NULL pointers
458               KMP_DEBUG_ASSERT(node->dn.mtx_locks[n - 1] != NULL);
459               node->dn.mtx_locks[n] = node->dn.mtx_locks[n - 1];
460             }
461             node->dn.mtx_locks[m] = info->mtx_lock;
462             break;
463           }
464         }
465         KMP_DEBUG_ASSERT(m < MAX_MTX_DEPS); // must break from loop
466         node->dn.mtx_num_locks++;
467       }
468     }
469   }
470   KA_TRACE(30, ("__kmp_process_deps<%d>: T#%d found %d predecessors\n", filter,
471                 gtid, npredecessors));
472   return npredecessors;
473 }
474 
475 #define NO_DEP_BARRIER (false)
476 #define DEP_BARRIER (true)
477 
478 // returns true if the task has any outstanding dependence
479 static bool __kmp_check_deps(kmp_int32 gtid, kmp_depnode_t *node,
480                              kmp_task_t *task, kmp_dephash_t **hash,
481                              bool dep_barrier, kmp_int32 ndeps,
482                              kmp_depend_info_t *dep_list,
483                              kmp_int32 ndeps_noalias,
484                              kmp_depend_info_t *noalias_dep_list) {
485   int i, n_mtxs = 0, dep_all = 0;
486 #if KMP_DEBUG
487   kmp_taskdata_t *taskdata = KMP_TASK_TO_TASKDATA(task);
488 #endif
489   KA_TRACE(20, ("__kmp_check_deps: T#%d checking dependences for task %p : %d "
490                 "possibly aliased dependences, %d non-aliased dependences : "
491                 "dep_barrier=%d .\n",
492                 gtid, taskdata, ndeps, ndeps_noalias, dep_barrier));
493 
494   // Filter deps in dep_list
495   // TODO: Different algorithm for large dep_list ( > 10 ? )
496   for (i = 0; i < ndeps; i++) {
497     if (dep_list[i].base_addr != 0 &&
498         dep_list[i].base_addr != (kmp_intptr_t)KMP_SIZE_T_MAX) {
499       KMP_DEBUG_ASSERT(
500           dep_list[i].flag == KMP_DEP_IN || dep_list[i].flag == KMP_DEP_OUT ||
501           dep_list[i].flag == KMP_DEP_INOUT ||
502           dep_list[i].flag == KMP_DEP_MTX || dep_list[i].flag == KMP_DEP_SET);
503       for (int j = i + 1; j < ndeps; j++) {
504         if (dep_list[i].base_addr == dep_list[j].base_addr) {
505           if (dep_list[i].flag != dep_list[j].flag) {
506             // two different dependences on same address work identical to OUT
507             dep_list[i].flag = KMP_DEP_OUT;
508           }
509           dep_list[j].base_addr = 0; // Mark j element as void
510         }
511       }
512       if (dep_list[i].flag == KMP_DEP_MTX) {
513         // limit number of mtx deps to MAX_MTX_DEPS per node
514         if (n_mtxs < MAX_MTX_DEPS && task != NULL) {
515           ++n_mtxs;
516         } else {
517           dep_list[i].flag = KMP_DEP_OUT; // downgrade mutexinoutset to inout
518         }
519       }
520     } else if (dep_list[i].flag == KMP_DEP_ALL ||
521                dep_list[i].base_addr == (kmp_intptr_t)KMP_SIZE_T_MAX) {
522       // omp_all_memory dependence can be marked by compiler by either
523       // (addr=0 && flag=0x80) (flag KMP_DEP_ALL), or (addr=-1).
524       // omp_all_memory overrides all other dependences if any
525       dep_all = 1;
526       break;
527     }
528   }
529 
530   // doesn't need to be atomic as no other thread is going to be accessing this
531   // node just yet.
532   // npredecessors is set -1 to ensure that none of the releasing tasks queues
533   // this task before we have finished processing all the dependences
534   node->dn.npredecessors = -1;
535 
536   // used to pack all npredecessors additions into a single atomic operation at
537   // the end
538   int npredecessors;
539 
540   if (!dep_all) { // regular dependences
541     npredecessors = __kmp_process_deps<true>(gtid, node, hash, dep_barrier,
542                                              ndeps, dep_list, task);
543     npredecessors += __kmp_process_deps<false>(
544         gtid, node, hash, dep_barrier, ndeps_noalias, noalias_dep_list, task);
545   } else { // omp_all_memory dependence
546     npredecessors = __kmp_process_dep_all(gtid, node, *hash, dep_barrier, task);
547   }
548 
549   node->dn.task = task;
550   KMP_MB();
551 
552   // Account for our initial fake value
553   npredecessors++;
554 
555   // Update predecessors and obtain current value to check if there are still
556   // any outstanding dependences (some tasks may have finished while we
557   // processed the dependences)
558   npredecessors =
559       node->dn.npredecessors.fetch_add(npredecessors) + npredecessors;
560 
561   KA_TRACE(20, ("__kmp_check_deps: T#%d found %d predecessors for task %p \n",
562                 gtid, npredecessors, taskdata));
563 
564   // beyond this point the task could be queued (and executed) by a releasing
565   // task...
566   return npredecessors > 0 ? true : false;
567 }
568 
569 /*!
570 @ingroup TASKING
571 @param loc_ref location of the original task directive
572 @param gtid Global Thread ID of encountering thread
573 @param new_task task thunk allocated by __kmp_omp_task_alloc() for the ''new
574 task''
575 @param ndeps Number of depend items with possible aliasing
576 @param dep_list List of depend items with possible aliasing
577 @param ndeps_noalias Number of depend items with no aliasing
578 @param noalias_dep_list List of depend items with no aliasing
579 
580 @return Returns either TASK_CURRENT_NOT_QUEUED if the current task was not
581 suspended and queued, or TASK_CURRENT_QUEUED if it was suspended and queued
582 
583 Schedule a non-thread-switchable task with dependences for execution
584 */
585 kmp_int32 __kmpc_omp_task_with_deps(ident_t *loc_ref, kmp_int32 gtid,
586                                     kmp_task_t *new_task, kmp_int32 ndeps,
587                                     kmp_depend_info_t *dep_list,
588                                     kmp_int32 ndeps_noalias,
589                                     kmp_depend_info_t *noalias_dep_list) {
590 
591   kmp_taskdata_t *new_taskdata = KMP_TASK_TO_TASKDATA(new_task);
592   KA_TRACE(10, ("__kmpc_omp_task_with_deps(enter): T#%d loc=%p task=%p\n", gtid,
593                 loc_ref, new_taskdata));
594   __kmp_assert_valid_gtid(gtid);
595   kmp_info_t *thread = __kmp_threads[gtid];
596   kmp_taskdata_t *current_task = thread->th.th_current_task;
597 
598 #if OMPT_SUPPORT
599   if (ompt_enabled.enabled) {
600     if (!current_task->ompt_task_info.frame.enter_frame.ptr)
601       current_task->ompt_task_info.frame.enter_frame.ptr =
602           OMPT_GET_FRAME_ADDRESS(0);
603     if (ompt_enabled.ompt_callback_task_create) {
604       ompt_callbacks.ompt_callback(ompt_callback_task_create)(
605           &(current_task->ompt_task_info.task_data),
606           &(current_task->ompt_task_info.frame),
607           &(new_taskdata->ompt_task_info.task_data),
608           ompt_task_explicit | TASK_TYPE_DETAILS_FORMAT(new_taskdata), 1,
609           OMPT_LOAD_OR_GET_RETURN_ADDRESS(gtid));
610     }
611 
612     new_taskdata->ompt_task_info.frame.enter_frame.ptr =
613         OMPT_GET_FRAME_ADDRESS(0);
614   }
615 
616 #if OMPT_OPTIONAL
617   /* OMPT grab all dependences if requested by the tool */
618   if (ndeps + ndeps_noalias > 0 && ompt_enabled.ompt_callback_dependences) {
619     kmp_int32 i;
620 
621     int ompt_ndeps = ndeps + ndeps_noalias;
622     ompt_dependence_t *ompt_deps = (ompt_dependence_t *)KMP_OMPT_DEPS_ALLOC(
623         thread, (ndeps + ndeps_noalias) * sizeof(ompt_dependence_t));
624 
625     KMP_ASSERT(ompt_deps != NULL);
626 
627     for (i = 0; i < ndeps; i++) {
628       ompt_deps[i].variable.ptr = (void *)dep_list[i].base_addr;
629       if (dep_list[i].flags.in && dep_list[i].flags.out)
630         ompt_deps[i].dependence_type = ompt_dependence_type_inout;
631       else if (dep_list[i].flags.out)
632         ompt_deps[i].dependence_type = ompt_dependence_type_out;
633       else if (dep_list[i].flags.in)
634         ompt_deps[i].dependence_type = ompt_dependence_type_in;
635       else if (dep_list[i].flags.mtx)
636         ompt_deps[i].dependence_type = ompt_dependence_type_mutexinoutset;
637       else if (dep_list[i].flags.set)
638         ompt_deps[i].dependence_type = ompt_dependence_type_inoutset;
639     }
640     for (i = 0; i < ndeps_noalias; i++) {
641       ompt_deps[ndeps + i].variable.ptr = (void *)noalias_dep_list[i].base_addr;
642       if (noalias_dep_list[i].flags.in && noalias_dep_list[i].flags.out)
643         ompt_deps[ndeps + i].dependence_type = ompt_dependence_type_inout;
644       else if (noalias_dep_list[i].flags.out)
645         ompt_deps[ndeps + i].dependence_type = ompt_dependence_type_out;
646       else if (noalias_dep_list[i].flags.in)
647         ompt_deps[ndeps + i].dependence_type = ompt_dependence_type_in;
648       else if (noalias_dep_list[i].flags.mtx)
649         ompt_deps[ndeps + i].dependence_type =
650             ompt_dependence_type_mutexinoutset;
651       else if (noalias_dep_list[i].flags.set)
652         ompt_deps[ndeps + i].dependence_type = ompt_dependence_type_inoutset;
653     }
654     ompt_callbacks.ompt_callback(ompt_callback_dependences)(
655         &(new_taskdata->ompt_task_info.task_data), ompt_deps, ompt_ndeps);
656     /* We can now free the allocated memory for the dependences */
657     /* For OMPD we might want to delay the free until end of this function */
658     KMP_OMPT_DEPS_FREE(thread, ompt_deps);
659   }
660 #endif /* OMPT_OPTIONAL */
661 #endif /* OMPT_SUPPORT */
662 
663   bool serial = current_task->td_flags.team_serial ||
664                 current_task->td_flags.tasking_ser ||
665                 current_task->td_flags.final;
666   kmp_task_team_t *task_team = thread->th.th_task_team;
667   serial = serial &&
668            !(task_team && (task_team->tt.tt_found_proxy_tasks ||
669                            task_team->tt.tt_hidden_helper_task_encountered));
670 
671   if (!serial && (ndeps > 0 || ndeps_noalias > 0)) {
672     /* if no dependences have been tracked yet, create the dependence hash */
673     if (current_task->td_dephash == NULL)
674       current_task->td_dephash = __kmp_dephash_create(thread, current_task);
675 
676 #if USE_FAST_MEMORY
677     kmp_depnode_t *node =
678         (kmp_depnode_t *)__kmp_fast_allocate(thread, sizeof(kmp_depnode_t));
679 #else
680     kmp_depnode_t *node =
681         (kmp_depnode_t *)__kmp_thread_malloc(thread, sizeof(kmp_depnode_t));
682 #endif
683 
684     __kmp_init_node(node);
685     new_taskdata->td_depnode = node;
686 
687     if (__kmp_check_deps(gtid, node, new_task, &current_task->td_dephash,
688                          NO_DEP_BARRIER, ndeps, dep_list, ndeps_noalias,
689                          noalias_dep_list)) {
690       KA_TRACE(10, ("__kmpc_omp_task_with_deps(exit): T#%d task had blocking "
691                     "dependences: "
692                     "loc=%p task=%p, return: TASK_CURRENT_NOT_QUEUED\n",
693                     gtid, loc_ref, new_taskdata));
694 #if OMPT_SUPPORT
695       if (ompt_enabled.enabled) {
696         current_task->ompt_task_info.frame.enter_frame = ompt_data_none;
697       }
698 #endif
699       return TASK_CURRENT_NOT_QUEUED;
700     }
701   } else {
702     KA_TRACE(10, ("__kmpc_omp_task_with_deps(exit): T#%d ignored dependences "
703                   "for task (serialized) loc=%p task=%p\n",
704                   gtid, loc_ref, new_taskdata));
705   }
706 
707   KA_TRACE(10, ("__kmpc_omp_task_with_deps(exit): T#%d task had no blocking "
708                 "dependences : "
709                 "loc=%p task=%p, transferring to __kmp_omp_task\n",
710                 gtid, loc_ref, new_taskdata));
711 
712   kmp_int32 ret = __kmp_omp_task(gtid, new_task, true);
713 #if OMPT_SUPPORT
714   if (ompt_enabled.enabled) {
715     current_task->ompt_task_info.frame.enter_frame = ompt_data_none;
716   }
717 #endif
718   return ret;
719 }
720 
721 #if OMPT_SUPPORT
722 void __ompt_taskwait_dep_finish(kmp_taskdata_t *current_task,
723                                 ompt_data_t *taskwait_task_data) {
724   if (ompt_enabled.ompt_callback_task_schedule) {
725     ompt_callbacks.ompt_callback(ompt_callback_task_schedule)(
726         taskwait_task_data, ompt_taskwait_complete, NULL);
727   }
728   current_task->ompt_task_info.frame.enter_frame.ptr = NULL;
729   *taskwait_task_data = ompt_data_none;
730 }
731 #endif /* OMPT_SUPPORT */
732 
733 /*!
734 @ingroup TASKING
735 @param loc_ref location of the original task directive
736 @param gtid Global Thread ID of encountering thread
737 @param ndeps Number of depend items with possible aliasing
738 @param dep_list List of depend items with possible aliasing
739 @param ndeps_noalias Number of depend items with no aliasing
740 @param noalias_dep_list List of depend items with no aliasing
741 
742 Blocks the current task until all specifies dependences have been fulfilled.
743 */
744 void __kmpc_omp_wait_deps(ident_t *loc_ref, kmp_int32 gtid, kmp_int32 ndeps,
745                           kmp_depend_info_t *dep_list, kmp_int32 ndeps_noalias,
746                           kmp_depend_info_t *noalias_dep_list) {
747   KA_TRACE(10, ("__kmpc_omp_wait_deps(enter): T#%d loc=%p\n", gtid, loc_ref));
748 
749   if (ndeps == 0 && ndeps_noalias == 0) {
750     KA_TRACE(10, ("__kmpc_omp_wait_deps(exit): T#%d has no dependences to "
751                   "wait upon : loc=%p\n",
752                   gtid, loc_ref));
753     return;
754   }
755   __kmp_assert_valid_gtid(gtid);
756   kmp_info_t *thread = __kmp_threads[gtid];
757   kmp_taskdata_t *current_task = thread->th.th_current_task;
758 
759 #if OMPT_SUPPORT
760   // this function represents a taskwait construct with depend clause
761   // We signal 4 events:
762   //  - creation of the taskwait task
763   //  - dependences of the taskwait task
764   //  - schedule and finish of the taskwait task
765   ompt_data_t *taskwait_task_data = &thread->th.ompt_thread_info.task_data;
766   KMP_ASSERT(taskwait_task_data->ptr == NULL);
767   if (ompt_enabled.enabled) {
768     if (!current_task->ompt_task_info.frame.enter_frame.ptr)
769       current_task->ompt_task_info.frame.enter_frame.ptr =
770           OMPT_GET_FRAME_ADDRESS(0);
771     if (ompt_enabled.ompt_callback_task_create) {
772       ompt_callbacks.ompt_callback(ompt_callback_task_create)(
773           &(current_task->ompt_task_info.task_data),
774           &(current_task->ompt_task_info.frame), taskwait_task_data,
775           ompt_task_taskwait | ompt_task_undeferred | ompt_task_mergeable, 1,
776           OMPT_LOAD_OR_GET_RETURN_ADDRESS(gtid));
777     }
778   }
779 
780 #if OMPT_OPTIONAL
781   /* OMPT grab all dependences if requested by the tool */
782   if (ndeps + ndeps_noalias > 0 && ompt_enabled.ompt_callback_dependences) {
783     kmp_int32 i;
784 
785     int ompt_ndeps = ndeps + ndeps_noalias;
786     ompt_dependence_t *ompt_deps = (ompt_dependence_t *)KMP_OMPT_DEPS_ALLOC(
787         thread, (ndeps + ndeps_noalias) * sizeof(ompt_dependence_t));
788 
789     KMP_ASSERT(ompt_deps != NULL);
790 
791     for (i = 0; i < ndeps; i++) {
792       ompt_deps[i].variable.ptr = (void *)dep_list[i].base_addr;
793       if (dep_list[i].flags.in && dep_list[i].flags.out)
794         ompt_deps[i].dependence_type = ompt_dependence_type_inout;
795       else if (dep_list[i].flags.out)
796         ompt_deps[i].dependence_type = ompt_dependence_type_out;
797       else if (dep_list[i].flags.in)
798         ompt_deps[i].dependence_type = ompt_dependence_type_in;
799       else if (dep_list[i].flags.mtx)
800         ompt_deps[ndeps + i].dependence_type =
801             ompt_dependence_type_mutexinoutset;
802       else if (dep_list[i].flags.set)
803         ompt_deps[ndeps + i].dependence_type = ompt_dependence_type_inoutset;
804     }
805     for (i = 0; i < ndeps_noalias; i++) {
806       ompt_deps[ndeps + i].variable.ptr = (void *)noalias_dep_list[i].base_addr;
807       if (noalias_dep_list[i].flags.in && noalias_dep_list[i].flags.out)
808         ompt_deps[ndeps + i].dependence_type = ompt_dependence_type_inout;
809       else if (noalias_dep_list[i].flags.out)
810         ompt_deps[ndeps + i].dependence_type = ompt_dependence_type_out;
811       else if (noalias_dep_list[i].flags.in)
812         ompt_deps[ndeps + i].dependence_type = ompt_dependence_type_in;
813       else if (noalias_dep_list[i].flags.mtx)
814         ompt_deps[ndeps + i].dependence_type =
815             ompt_dependence_type_mutexinoutset;
816       else if (noalias_dep_list[i].flags.set)
817         ompt_deps[ndeps + i].dependence_type = ompt_dependence_type_inoutset;
818     }
819     ompt_callbacks.ompt_callback(ompt_callback_dependences)(
820         taskwait_task_data, ompt_deps, ompt_ndeps);
821     /* We can now free the allocated memory for the dependences */
822     /* For OMPD we might want to delay the free until end of this function */
823     KMP_OMPT_DEPS_FREE(thread, ompt_deps);
824     ompt_deps = NULL;
825   }
826 #endif /* OMPT_OPTIONAL */
827 #endif /* OMPT_SUPPORT */
828 
829   // We can return immediately as:
830   // - dependences are not computed in serial teams (except with proxy tasks)
831   // - if the dephash is not yet created it means we have nothing to wait for
832   bool ignore = current_task->td_flags.team_serial ||
833                 current_task->td_flags.tasking_ser ||
834                 current_task->td_flags.final;
835   ignore =
836       ignore && thread->th.th_task_team != NULL &&
837       thread->th.th_task_team->tt.tt_found_proxy_tasks == FALSE &&
838       thread->th.th_task_team->tt.tt_hidden_helper_task_encountered == FALSE;
839   ignore = ignore || current_task->td_dephash == NULL;
840 
841   if (ignore) {
842     KA_TRACE(10, ("__kmpc_omp_wait_deps(exit): T#%d has no blocking "
843                   "dependences : loc=%p\n",
844                   gtid, loc_ref));
845 #if OMPT_SUPPORT
846     __ompt_taskwait_dep_finish(current_task, taskwait_task_data);
847 #endif /* OMPT_SUPPORT */
848     return;
849   }
850 
851   kmp_depnode_t node = {0};
852   __kmp_init_node(&node);
853 
854   if (!__kmp_check_deps(gtid, &node, NULL, &current_task->td_dephash,
855                         DEP_BARRIER, ndeps, dep_list, ndeps_noalias,
856                         noalias_dep_list)) {
857     KA_TRACE(10, ("__kmpc_omp_wait_deps(exit): T#%d has no blocking "
858                   "dependences : loc=%p\n",
859                   gtid, loc_ref));
860 #if OMPT_SUPPORT
861     __ompt_taskwait_dep_finish(current_task, taskwait_task_data);
862 #endif /* OMPT_SUPPORT */
863     return;
864   }
865 
866   int thread_finished = FALSE;
867   kmp_flag_32<false, false> flag(
868       (std::atomic<kmp_uint32> *)&node.dn.npredecessors, 0U);
869   while (node.dn.npredecessors > 0) {
870     flag.execute_tasks(thread, gtid, FALSE,
871                        &thread_finished USE_ITT_BUILD_ARG(NULL),
872                        __kmp_task_stealing_constraint);
873   }
874 
875 #if OMPT_SUPPORT
876   __ompt_taskwait_dep_finish(current_task, taskwait_task_data);
877 #endif /* OMPT_SUPPORT */
878   KA_TRACE(10, ("__kmpc_omp_wait_deps(exit): T#%d finished waiting : loc=%p\n",
879                 gtid, loc_ref));
880 }
881