1 /*
2  * kmp_csupport.cpp -- kfront linkage support for OpenMP.
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_IMP
14 #include "omp.h" /* extern "C" declarations of user-visible routines */
15 #include "kmp.h"
16 #include "kmp_error.h"
17 #include "kmp_i18n.h"
18 #include "kmp_itt.h"
19 #include "kmp_lock.h"
20 #include "kmp_stats.h"
21 #include "ompt-specific.h"
22 
23 #define MAX_MESSAGE 512
24 
25 // flags will be used in future, e.g. to implement openmp_strict library
26 // restrictions
27 
28 /*!
29  * @ingroup STARTUP_SHUTDOWN
30  * @param loc   in   source location information
31  * @param flags in   for future use (currently ignored)
32  *
33  * Initialize the runtime library. This call is optional; if it is not made then
34  * it will be implicitly called by attempts to use other library functions.
35  */
36 void __kmpc_begin(ident_t *loc, kmp_int32 flags) {
37   // By default __kmpc_begin() is no-op.
38   char *env;
39   if ((env = getenv("KMP_INITIAL_THREAD_BIND")) != NULL &&
40       __kmp_str_match_true(env)) {
41     __kmp_middle_initialize();
42     KC_TRACE(10, ("__kmpc_begin: middle initialization called\n"));
43   } else if (__kmp_ignore_mppbeg() == FALSE) {
44     // By default __kmp_ignore_mppbeg() returns TRUE.
45     __kmp_internal_begin();
46     KC_TRACE(10, ("__kmpc_begin: called\n"));
47   }
48 }
49 
50 /*!
51  * @ingroup STARTUP_SHUTDOWN
52  * @param loc source location information
53  *
54  * Shutdown the runtime library. This is also optional, and even if called will
55  * not do anything unless the `KMP_IGNORE_MPPEND` environment variable is set to
56  * zero.
57  */
58 void __kmpc_end(ident_t *loc) {
59   // By default, __kmp_ignore_mppend() returns TRUE which makes __kmpc_end()
60   // call no-op. However, this can be overridden with KMP_IGNORE_MPPEND
61   // environment variable. If KMP_IGNORE_MPPEND is 0, __kmp_ignore_mppend()
62   // returns FALSE and __kmpc_end() will unregister this root (it can cause
63   // library shut down).
64   if (__kmp_ignore_mppend() == FALSE) {
65     KC_TRACE(10, ("__kmpc_end: called\n"));
66     KA_TRACE(30, ("__kmpc_end\n"));
67 
68     __kmp_internal_end_thread(-1);
69   }
70 #if KMP_OS_WINDOWS && OMPT_SUPPORT
71   // Normal exit process on Windows does not allow worker threads of the final
72   // parallel region to finish reporting their events, so shutting down the
73   // library here fixes the issue at least for the cases where __kmpc_end() is
74   // placed properly.
75   if (ompt_enabled.enabled)
76     __kmp_internal_end_library(__kmp_gtid_get_specific());
77 #endif
78 }
79 
80 /*!
81 @ingroup THREAD_STATES
82 @param loc Source location information.
83 @return The global thread index of the active thread.
84 
85 This function can be called in any context.
86 
87 If the runtime has ony been entered at the outermost level from a
88 single (necessarily non-OpenMP<sup>*</sup>) thread, then the thread number is
89 that which would be returned by omp_get_thread_num() in the outermost
90 active parallel construct. (Or zero if there is no active parallel
91 construct, since the master thread is necessarily thread zero).
92 
93 If multiple non-OpenMP threads all enter an OpenMP construct then this
94 will be a unique thread identifier among all the threads created by
95 the OpenMP runtime (but the value cannot be defined in terms of
96 OpenMP thread ids returned by omp_get_thread_num()).
97 */
98 kmp_int32 __kmpc_global_thread_num(ident_t *loc) {
99   kmp_int32 gtid = __kmp_entry_gtid();
100 
101   KC_TRACE(10, ("__kmpc_global_thread_num: T#%d\n", gtid));
102 
103   return gtid;
104 }
105 
106 /*!
107 @ingroup THREAD_STATES
108 @param loc Source location information.
109 @return The number of threads under control of the OpenMP<sup>*</sup> runtime
110 
111 This function can be called in any context.
112 It returns the total number of threads under the control of the OpenMP runtime.
113 That is not a number that can be determined by any OpenMP standard calls, since
114 the library may be called from more than one non-OpenMP thread, and this
115 reflects the total over all such calls. Similarly the runtime maintains
116 underlying threads even when they are not active (since the cost of creating
117 and destroying OS threads is high), this call counts all such threads even if
118 they are not waiting for work.
119 */
120 kmp_int32 __kmpc_global_num_threads(ident_t *loc) {
121   KC_TRACE(10,
122            ("__kmpc_global_num_threads: num_threads = %d\n", __kmp_all_nth));
123 
124   return TCR_4(__kmp_all_nth);
125 }
126 
127 /*!
128 @ingroup THREAD_STATES
129 @param loc Source location information.
130 @return The thread number of the calling thread in the innermost active parallel
131 construct.
132 */
133 kmp_int32 __kmpc_bound_thread_num(ident_t *loc) {
134   KC_TRACE(10, ("__kmpc_bound_thread_num: called\n"));
135   return __kmp_tid_from_gtid(__kmp_entry_gtid());
136 }
137 
138 /*!
139 @ingroup THREAD_STATES
140 @param loc Source location information.
141 @return The number of threads in the innermost active parallel construct.
142 */
143 kmp_int32 __kmpc_bound_num_threads(ident_t *loc) {
144   KC_TRACE(10, ("__kmpc_bound_num_threads: called\n"));
145 
146   return __kmp_entry_thread()->th.th_team->t.t_nproc;
147 }
148 
149 /*!
150  * @ingroup DEPRECATED
151  * @param loc location description
152  *
153  * This function need not be called. It always returns TRUE.
154  */
155 kmp_int32 __kmpc_ok_to_fork(ident_t *loc) {
156 #ifndef KMP_DEBUG
157 
158   return TRUE;
159 
160 #else
161 
162   const char *semi2;
163   const char *semi3;
164   int line_no;
165 
166   if (__kmp_par_range == 0) {
167     return TRUE;
168   }
169   semi2 = loc->psource;
170   if (semi2 == NULL) {
171     return TRUE;
172   }
173   semi2 = strchr(semi2, ';');
174   if (semi2 == NULL) {
175     return TRUE;
176   }
177   semi2 = strchr(semi2 + 1, ';');
178   if (semi2 == NULL) {
179     return TRUE;
180   }
181   if (__kmp_par_range_filename[0]) {
182     const char *name = semi2 - 1;
183     while ((name > loc->psource) && (*name != '/') && (*name != ';')) {
184       name--;
185     }
186     if ((*name == '/') || (*name == ';')) {
187       name++;
188     }
189     if (strncmp(__kmp_par_range_filename, name, semi2 - name)) {
190       return __kmp_par_range < 0;
191     }
192   }
193   semi3 = strchr(semi2 + 1, ';');
194   if (__kmp_par_range_routine[0]) {
195     if ((semi3 != NULL) && (semi3 > semi2) &&
196         (strncmp(__kmp_par_range_routine, semi2 + 1, semi3 - semi2 - 1))) {
197       return __kmp_par_range < 0;
198     }
199   }
200   if (KMP_SSCANF(semi3 + 1, "%d", &line_no) == 1) {
201     if ((line_no >= __kmp_par_range_lb) && (line_no <= __kmp_par_range_ub)) {
202       return __kmp_par_range > 0;
203     }
204     return __kmp_par_range < 0;
205   }
206   return TRUE;
207 
208 #endif /* KMP_DEBUG */
209 }
210 
211 /*!
212 @ingroup THREAD_STATES
213 @param loc Source location information.
214 @return 1 if this thread is executing inside an active parallel region, zero if
215 not.
216 */
217 kmp_int32 __kmpc_in_parallel(ident_t *loc) {
218   return __kmp_entry_thread()->th.th_root->r.r_active;
219 }
220 
221 /*!
222 @ingroup PARALLEL
223 @param loc source location information
224 @param global_tid global thread number
225 @param num_threads number of threads requested for this parallel construct
226 
227 Set the number of threads to be used by the next fork spawned by this thread.
228 This call is only required if the parallel construct has a `num_threads` clause.
229 */
230 void __kmpc_push_num_threads(ident_t *loc, kmp_int32 global_tid,
231                              kmp_int32 num_threads) {
232   KA_TRACE(20, ("__kmpc_push_num_threads: enter T#%d num_threads=%d\n",
233                 global_tid, num_threads));
234   __kmp_assert_valid_gtid(global_tid);
235   __kmp_push_num_threads(loc, global_tid, num_threads);
236 }
237 
238 void __kmpc_pop_num_threads(ident_t *loc, kmp_int32 global_tid) {
239   KA_TRACE(20, ("__kmpc_pop_num_threads: enter\n"));
240   /* the num_threads are automatically popped */
241 }
242 
243 void __kmpc_push_proc_bind(ident_t *loc, kmp_int32 global_tid,
244                            kmp_int32 proc_bind) {
245   KA_TRACE(20, ("__kmpc_push_proc_bind: enter T#%d proc_bind=%d\n", global_tid,
246                 proc_bind));
247   __kmp_assert_valid_gtid(global_tid);
248   __kmp_push_proc_bind(loc, global_tid, (kmp_proc_bind_t)proc_bind);
249 }
250 
251 /*!
252 @ingroup PARALLEL
253 @param loc  source location information
254 @param argc  total number of arguments in the ellipsis
255 @param microtask  pointer to callback routine consisting of outlined parallel
256 construct
257 @param ...  pointers to shared variables that aren't global
258 
259 Do the actual fork and call the microtask in the relevant number of threads.
260 */
261 void __kmpc_fork_call(ident_t *loc, kmp_int32 argc, kmpc_micro microtask, ...) {
262   int gtid = __kmp_entry_gtid();
263 
264 #if (KMP_STATS_ENABLED)
265   // If we were in a serial region, then stop the serial timer, record
266   // the event, and start parallel region timer
267   stats_state_e previous_state = KMP_GET_THREAD_STATE();
268   if (previous_state == stats_state_e::SERIAL_REGION) {
269     KMP_EXCHANGE_PARTITIONED_TIMER(OMP_parallel_overhead);
270   } else {
271     KMP_PUSH_PARTITIONED_TIMER(OMP_parallel_overhead);
272   }
273   int inParallel = __kmpc_in_parallel(loc);
274   if (inParallel) {
275     KMP_COUNT_BLOCK(OMP_NESTED_PARALLEL);
276   } else {
277     KMP_COUNT_BLOCK(OMP_PARALLEL);
278   }
279 #endif
280 
281   // maybe to save thr_state is enough here
282   {
283     va_list ap;
284     va_start(ap, microtask);
285 
286 #if OMPT_SUPPORT
287     ompt_frame_t *ompt_frame;
288     if (ompt_enabled.enabled) {
289       kmp_info_t *master_th = __kmp_threads[gtid];
290       kmp_team_t *parent_team = master_th->th.th_team;
291       ompt_lw_taskteam_t *lwt = parent_team->t.ompt_serialized_team_info;
292       if (lwt)
293         ompt_frame = &(lwt->ompt_task_info.frame);
294       else {
295         int tid = __kmp_tid_from_gtid(gtid);
296         ompt_frame = &(
297             parent_team->t.t_implicit_task_taskdata[tid].ompt_task_info.frame);
298       }
299       ompt_frame->enter_frame.ptr = OMPT_GET_FRAME_ADDRESS(0);
300       OMPT_STORE_RETURN_ADDRESS(gtid);
301     }
302 #endif
303 
304 #if INCLUDE_SSC_MARKS
305     SSC_MARK_FORKING();
306 #endif
307     __kmp_fork_call(loc, gtid, fork_context_intel, argc,
308                     VOLATILE_CAST(microtask_t) microtask, // "wrapped" task
309                     VOLATILE_CAST(launch_t) __kmp_invoke_task_func,
310                     kmp_va_addr_of(ap));
311 #if INCLUDE_SSC_MARKS
312     SSC_MARK_JOINING();
313 #endif
314     __kmp_join_call(loc, gtid
315 #if OMPT_SUPPORT
316                     ,
317                     fork_context_intel
318 #endif
319                     );
320 
321     va_end(ap);
322   }
323 
324 #if KMP_STATS_ENABLED
325   if (previous_state == stats_state_e::SERIAL_REGION) {
326     KMP_EXCHANGE_PARTITIONED_TIMER(OMP_serial);
327     KMP_SET_THREAD_STATE(previous_state);
328   } else {
329     KMP_POP_PARTITIONED_TIMER();
330   }
331 #endif // KMP_STATS_ENABLED
332 }
333 
334 /*!
335 @ingroup PARALLEL
336 @param loc source location information
337 @param global_tid global thread number
338 @param num_teams number of teams requested for the teams construct
339 @param num_threads number of threads per team requested for the teams construct
340 
341 Set the number of teams to be used by the teams construct.
342 This call is only required if the teams construct has a `num_teams` clause
343 or a `thread_limit` clause (or both).
344 */
345 void __kmpc_push_num_teams(ident_t *loc, kmp_int32 global_tid,
346                            kmp_int32 num_teams, kmp_int32 num_threads) {
347   KA_TRACE(20,
348            ("__kmpc_push_num_teams: enter T#%d num_teams=%d num_threads=%d\n",
349             global_tid, num_teams, num_threads));
350   __kmp_assert_valid_gtid(global_tid);
351   __kmp_push_num_teams(loc, global_tid, num_teams, num_threads);
352 }
353 
354 /*!
355 @ingroup PARALLEL
356 @param loc  source location information
357 @param argc  total number of arguments in the ellipsis
358 @param microtask  pointer to callback routine consisting of outlined teams
359 construct
360 @param ...  pointers to shared variables that aren't global
361 
362 Do the actual fork and call the microtask in the relevant number of threads.
363 */
364 void __kmpc_fork_teams(ident_t *loc, kmp_int32 argc, kmpc_micro microtask,
365                        ...) {
366   int gtid = __kmp_entry_gtid();
367   kmp_info_t *this_thr = __kmp_threads[gtid];
368   va_list ap;
369   va_start(ap, microtask);
370 
371 #if KMP_STATS_ENABLED
372   KMP_COUNT_BLOCK(OMP_TEAMS);
373   stats_state_e previous_state = KMP_GET_THREAD_STATE();
374   if (previous_state == stats_state_e::SERIAL_REGION) {
375     KMP_EXCHANGE_PARTITIONED_TIMER(OMP_teams_overhead);
376   } else {
377     KMP_PUSH_PARTITIONED_TIMER(OMP_teams_overhead);
378   }
379 #endif
380 
381   // remember teams entry point and nesting level
382   this_thr->th.th_teams_microtask = microtask;
383   this_thr->th.th_teams_level =
384       this_thr->th.th_team->t.t_level; // AC: can be >0 on host
385 
386 #if OMPT_SUPPORT
387   kmp_team_t *parent_team = this_thr->th.th_team;
388   int tid = __kmp_tid_from_gtid(gtid);
389   if (ompt_enabled.enabled) {
390     parent_team->t.t_implicit_task_taskdata[tid]
391         .ompt_task_info.frame.enter_frame.ptr = OMPT_GET_FRAME_ADDRESS(0);
392   }
393   OMPT_STORE_RETURN_ADDRESS(gtid);
394 #endif
395 
396   // check if __kmpc_push_num_teams called, set default number of teams
397   // otherwise
398   if (this_thr->th.th_teams_size.nteams == 0) {
399     __kmp_push_num_teams(loc, gtid, 0, 0);
400   }
401   KMP_DEBUG_ASSERT(this_thr->th.th_set_nproc >= 1);
402   KMP_DEBUG_ASSERT(this_thr->th.th_teams_size.nteams >= 1);
403   KMP_DEBUG_ASSERT(this_thr->th.th_teams_size.nth >= 1);
404 
405   __kmp_fork_call(
406       loc, gtid, fork_context_intel, argc,
407       VOLATILE_CAST(microtask_t) __kmp_teams_master, // "wrapped" task
408       VOLATILE_CAST(launch_t) __kmp_invoke_teams_master, kmp_va_addr_of(ap));
409   __kmp_join_call(loc, gtid
410 #if OMPT_SUPPORT
411                   ,
412                   fork_context_intel
413 #endif
414                   );
415 
416   // Pop current CG root off list
417   KMP_DEBUG_ASSERT(this_thr->th.th_cg_roots);
418   kmp_cg_root_t *tmp = this_thr->th.th_cg_roots;
419   this_thr->th.th_cg_roots = tmp->up;
420   KA_TRACE(100, ("__kmpc_fork_teams: Thread %p popping node %p and moving up"
421                  " to node %p. cg_nthreads was %d\n",
422                  this_thr, tmp, this_thr->th.th_cg_roots, tmp->cg_nthreads));
423   KMP_DEBUG_ASSERT(tmp->cg_nthreads);
424   int i = tmp->cg_nthreads--;
425   if (i == 1) { // check is we are the last thread in CG (not always the case)
426     __kmp_free(tmp);
427   }
428   // Restore current task's thread_limit from CG root
429   KMP_DEBUG_ASSERT(this_thr->th.th_cg_roots);
430   this_thr->th.th_current_task->td_icvs.thread_limit =
431       this_thr->th.th_cg_roots->cg_thread_limit;
432 
433   this_thr->th.th_teams_microtask = NULL;
434   this_thr->th.th_teams_level = 0;
435   *(kmp_int64 *)(&this_thr->th.th_teams_size) = 0L;
436   va_end(ap);
437 #if KMP_STATS_ENABLED
438   if (previous_state == stats_state_e::SERIAL_REGION) {
439     KMP_EXCHANGE_PARTITIONED_TIMER(OMP_serial);
440     KMP_SET_THREAD_STATE(previous_state);
441   } else {
442     KMP_POP_PARTITIONED_TIMER();
443   }
444 #endif // KMP_STATS_ENABLED
445 }
446 
447 // I don't think this function should ever have been exported.
448 // The __kmpc_ prefix was misapplied.  I'm fairly certain that no generated
449 // openmp code ever called it, but it's been exported from the RTL for so
450 // long that I'm afraid to remove the definition.
451 int __kmpc_invoke_task_func(int gtid) { return __kmp_invoke_task_func(gtid); }
452 
453 /*!
454 @ingroup PARALLEL
455 @param loc  source location information
456 @param global_tid  global thread number
457 
458 Enter a serialized parallel construct. This interface is used to handle a
459 conditional parallel region, like this,
460 @code
461 #pragma omp parallel if (condition)
462 @endcode
463 when the condition is false.
464 */
465 void __kmpc_serialized_parallel(ident_t *loc, kmp_int32 global_tid) {
466   // The implementation is now in kmp_runtime.cpp so that it can share static
467   // functions with kmp_fork_call since the tasks to be done are similar in
468   // each case.
469   __kmp_assert_valid_gtid(global_tid);
470 #if OMPT_SUPPORT
471   OMPT_STORE_RETURN_ADDRESS(global_tid);
472 #endif
473   __kmp_serialized_parallel(loc, global_tid);
474 }
475 
476 /*!
477 @ingroup PARALLEL
478 @param loc  source location information
479 @param global_tid  global thread number
480 
481 Leave a serialized parallel construct.
482 */
483 void __kmpc_end_serialized_parallel(ident_t *loc, kmp_int32 global_tid) {
484   kmp_internal_control_t *top;
485   kmp_info_t *this_thr;
486   kmp_team_t *serial_team;
487 
488   KC_TRACE(10,
489            ("__kmpc_end_serialized_parallel: called by T#%d\n", global_tid));
490 
491   /* skip all this code for autopar serialized loops since it results in
492      unacceptable overhead */
493   if (loc != NULL && (loc->flags & KMP_IDENT_AUTOPAR))
494     return;
495 
496   // Not autopar code
497   __kmp_assert_valid_gtid(global_tid);
498   if (!TCR_4(__kmp_init_parallel))
499     __kmp_parallel_initialize();
500 
501   __kmp_resume_if_soft_paused();
502 
503   this_thr = __kmp_threads[global_tid];
504   serial_team = this_thr->th.th_serial_team;
505 
506   kmp_task_team_t *task_team = this_thr->th.th_task_team;
507   // we need to wait for the proxy tasks before finishing the thread
508   if (task_team != NULL && task_team->tt.tt_found_proxy_tasks)
509     __kmp_task_team_wait(this_thr, serial_team USE_ITT_BUILD_ARG(NULL));
510 
511   KMP_MB();
512   KMP_DEBUG_ASSERT(serial_team);
513   KMP_ASSERT(serial_team->t.t_serialized);
514   KMP_DEBUG_ASSERT(this_thr->th.th_team == serial_team);
515   KMP_DEBUG_ASSERT(serial_team != this_thr->th.th_root->r.r_root_team);
516   KMP_DEBUG_ASSERT(serial_team->t.t_threads);
517   KMP_DEBUG_ASSERT(serial_team->t.t_threads[0] == this_thr);
518 
519 #if OMPT_SUPPORT
520   if (ompt_enabled.enabled &&
521       this_thr->th.ompt_thread_info.state != ompt_state_overhead) {
522     OMPT_CUR_TASK_INFO(this_thr)->frame.exit_frame = ompt_data_none;
523     if (ompt_enabled.ompt_callback_implicit_task) {
524       ompt_callbacks.ompt_callback(ompt_callback_implicit_task)(
525           ompt_scope_end, NULL, OMPT_CUR_TASK_DATA(this_thr), 1,
526           OMPT_CUR_TASK_INFO(this_thr)->thread_num, ompt_task_implicit);
527     }
528 
529     // reset clear the task id only after unlinking the task
530     ompt_data_t *parent_task_data;
531     __ompt_get_task_info_internal(1, NULL, &parent_task_data, NULL, NULL, NULL);
532 
533     if (ompt_enabled.ompt_callback_parallel_end) {
534       ompt_callbacks.ompt_callback(ompt_callback_parallel_end)(
535           &(serial_team->t.ompt_team_info.parallel_data), parent_task_data,
536           ompt_parallel_invoker_program | ompt_parallel_team,
537           OMPT_LOAD_RETURN_ADDRESS(global_tid));
538     }
539     __ompt_lw_taskteam_unlink(this_thr);
540     this_thr->th.ompt_thread_info.state = ompt_state_overhead;
541   }
542 #endif
543 
544   /* If necessary, pop the internal control stack values and replace the team
545    * values */
546   top = serial_team->t.t_control_stack_top;
547   if (top && top->serial_nesting_level == serial_team->t.t_serialized) {
548     copy_icvs(&serial_team->t.t_threads[0]->th.th_current_task->td_icvs, top);
549     serial_team->t.t_control_stack_top = top->next;
550     __kmp_free(top);
551   }
552 
553   // if( serial_team -> t.t_serialized > 1 )
554   serial_team->t.t_level--;
555 
556   /* pop dispatch buffers stack */
557   KMP_DEBUG_ASSERT(serial_team->t.t_dispatch->th_disp_buffer);
558   {
559     dispatch_private_info_t *disp_buffer =
560         serial_team->t.t_dispatch->th_disp_buffer;
561     serial_team->t.t_dispatch->th_disp_buffer =
562         serial_team->t.t_dispatch->th_disp_buffer->next;
563     __kmp_free(disp_buffer);
564   }
565   this_thr->th.th_def_allocator = serial_team->t.t_def_allocator; // restore
566 
567   --serial_team->t.t_serialized;
568   if (serial_team->t.t_serialized == 0) {
569 
570 /* return to the parallel section */
571 
572 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
573     if (__kmp_inherit_fp_control && serial_team->t.t_fp_control_saved) {
574       __kmp_clear_x87_fpu_status_word();
575       __kmp_load_x87_fpu_control_word(&serial_team->t.t_x87_fpu_control_word);
576       __kmp_load_mxcsr(&serial_team->t.t_mxcsr);
577     }
578 #endif /* KMP_ARCH_X86 || KMP_ARCH_X86_64 */
579 
580     this_thr->th.th_team = serial_team->t.t_parent;
581     this_thr->th.th_info.ds.ds_tid = serial_team->t.t_master_tid;
582 
583     /* restore values cached in the thread */
584     this_thr->th.th_team_nproc = serial_team->t.t_parent->t.t_nproc; /*  JPH */
585     this_thr->th.th_team_master =
586         serial_team->t.t_parent->t.t_threads[0]; /* JPH */
587     this_thr->th.th_team_serialized = this_thr->th.th_team->t.t_serialized;
588 
589     /* TODO the below shouldn't need to be adjusted for serialized teams */
590     this_thr->th.th_dispatch =
591         &this_thr->th.th_team->t.t_dispatch[serial_team->t.t_master_tid];
592 
593     __kmp_pop_current_task_from_thread(this_thr);
594 
595     KMP_ASSERT(this_thr->th.th_current_task->td_flags.executing == 0);
596     this_thr->th.th_current_task->td_flags.executing = 1;
597 
598     if (__kmp_tasking_mode != tskm_immediate_exec) {
599       // Copy the task team from the new child / old parent team to the thread.
600       this_thr->th.th_task_team =
601           this_thr->th.th_team->t.t_task_team[this_thr->th.th_task_state];
602       KA_TRACE(20,
603                ("__kmpc_end_serialized_parallel: T#%d restoring task_team %p / "
604                 "team %p\n",
605                 global_tid, this_thr->th.th_task_team, this_thr->th.th_team));
606     }
607   } else {
608     if (__kmp_tasking_mode != tskm_immediate_exec) {
609       KA_TRACE(20, ("__kmpc_end_serialized_parallel: T#%d decreasing nesting "
610                     "depth of serial team %p to %d\n",
611                     global_tid, serial_team, serial_team->t.t_serialized));
612     }
613   }
614 
615   if (__kmp_env_consistency_check)
616     __kmp_pop_parallel(global_tid, NULL);
617 #if OMPT_SUPPORT
618   if (ompt_enabled.enabled)
619     this_thr->th.ompt_thread_info.state =
620         ((this_thr->th.th_team_serialized) ? ompt_state_work_serial
621                                            : ompt_state_work_parallel);
622 #endif
623 }
624 
625 /*!
626 @ingroup SYNCHRONIZATION
627 @param loc  source location information.
628 
629 Execute <tt>flush</tt>. This is implemented as a full memory fence. (Though
630 depending on the memory ordering convention obeyed by the compiler
631 even that may not be necessary).
632 */
633 void __kmpc_flush(ident_t *loc) {
634   KC_TRACE(10, ("__kmpc_flush: called\n"));
635 
636   /* need explicit __mf() here since use volatile instead in library */
637   KMP_MB(); /* Flush all pending memory write invalidates.  */
638 
639 #if (KMP_ARCH_X86 || KMP_ARCH_X86_64)
640 #if KMP_MIC
641 // fence-style instructions do not exist, but lock; xaddl $0,(%rsp) can be used.
642 // We shouldn't need it, though, since the ABI rules require that
643 // * If the compiler generates NGO stores it also generates the fence
644 // * If users hand-code NGO stores they should insert the fence
645 // therefore no incomplete unordered stores should be visible.
646 #else
647   // C74404
648   // This is to address non-temporal store instructions (sfence needed).
649   // The clflush instruction is addressed either (mfence needed).
650   // Probably the non-temporal load monvtdqa instruction should also be
651   // addressed.
652   // mfence is a SSE2 instruction. Do not execute it if CPU is not SSE2.
653   if (!__kmp_cpuinfo.initialized) {
654     __kmp_query_cpuid(&__kmp_cpuinfo);
655   }
656   if (!__kmp_cpuinfo.sse2) {
657     // CPU cannot execute SSE2 instructions.
658   } else {
659 #if KMP_COMPILER_ICC
660     _mm_mfence();
661 #elif KMP_COMPILER_MSVC
662     MemoryBarrier();
663 #else
664     __sync_synchronize();
665 #endif // KMP_COMPILER_ICC
666   }
667 #endif // KMP_MIC
668 #elif (KMP_ARCH_ARM || KMP_ARCH_AARCH64 || KMP_ARCH_MIPS || KMP_ARCH_MIPS64 || \
669        KMP_ARCH_RISCV64)
670 // Nothing to see here move along
671 #elif KMP_ARCH_PPC64
672 // Nothing needed here (we have a real MB above).
673 #else
674 #error Unknown or unsupported architecture
675 #endif
676 
677 #if OMPT_SUPPORT && OMPT_OPTIONAL
678   if (ompt_enabled.ompt_callback_flush) {
679     ompt_callbacks.ompt_callback(ompt_callback_flush)(
680         __ompt_get_thread_data_internal(), OMPT_GET_RETURN_ADDRESS(0));
681   }
682 #endif
683 }
684 
685 /* -------------------------------------------------------------------------- */
686 /*!
687 @ingroup SYNCHRONIZATION
688 @param loc source location information
689 @param global_tid thread id.
690 
691 Execute a barrier.
692 */
693 void __kmpc_barrier(ident_t *loc, kmp_int32 global_tid) {
694   KMP_COUNT_BLOCK(OMP_BARRIER);
695   KC_TRACE(10, ("__kmpc_barrier: called T#%d\n", global_tid));
696   __kmp_assert_valid_gtid(global_tid);
697 
698   if (!TCR_4(__kmp_init_parallel))
699     __kmp_parallel_initialize();
700 
701   __kmp_resume_if_soft_paused();
702 
703   if (__kmp_env_consistency_check) {
704     if (loc == 0) {
705       KMP_WARNING(ConstructIdentInvalid); // ??? What does it mean for the user?
706     }
707     __kmp_check_barrier(global_tid, ct_barrier, loc);
708   }
709 
710 #if OMPT_SUPPORT
711   ompt_frame_t *ompt_frame;
712   if (ompt_enabled.enabled) {
713     __ompt_get_task_info_internal(0, NULL, NULL, &ompt_frame, NULL, NULL);
714     if (ompt_frame->enter_frame.ptr == NULL)
715       ompt_frame->enter_frame.ptr = OMPT_GET_FRAME_ADDRESS(0);
716     OMPT_STORE_RETURN_ADDRESS(global_tid);
717   }
718 #endif
719   __kmp_threads[global_tid]->th.th_ident = loc;
720   // TODO: explicit barrier_wait_id:
721   //   this function is called when 'barrier' directive is present or
722   //   implicit barrier at the end of a worksharing construct.
723   // 1) better to add a per-thread barrier counter to a thread data structure
724   // 2) set to 0 when a new team is created
725   // 4) no sync is required
726 
727   __kmp_barrier(bs_plain_barrier, global_tid, FALSE, 0, NULL, NULL);
728 #if OMPT_SUPPORT && OMPT_OPTIONAL
729   if (ompt_enabled.enabled) {
730     ompt_frame->enter_frame = ompt_data_none;
731   }
732 #endif
733 }
734 
735 /* The BARRIER for a MASTER section is always explicit   */
736 /*!
737 @ingroup WORK_SHARING
738 @param loc  source location information.
739 @param global_tid  global thread number .
740 @return 1 if this thread should execute the <tt>master</tt> block, 0 otherwise.
741 */
742 kmp_int32 __kmpc_master(ident_t *loc, kmp_int32 global_tid) {
743   int status = 0;
744 
745   KC_TRACE(10, ("__kmpc_master: called T#%d\n", global_tid));
746   __kmp_assert_valid_gtid(global_tid);
747 
748   if (!TCR_4(__kmp_init_parallel))
749     __kmp_parallel_initialize();
750 
751   __kmp_resume_if_soft_paused();
752 
753   if (KMP_MASTER_GTID(global_tid)) {
754     KMP_COUNT_BLOCK(OMP_MASTER);
755     KMP_PUSH_PARTITIONED_TIMER(OMP_master);
756     status = 1;
757   }
758 
759 #if OMPT_SUPPORT && OMPT_OPTIONAL
760   if (status) {
761     if (ompt_enabled.ompt_callback_masked) {
762       kmp_info_t *this_thr = __kmp_threads[global_tid];
763       kmp_team_t *team = this_thr->th.th_team;
764 
765       int tid = __kmp_tid_from_gtid(global_tid);
766       ompt_callbacks.ompt_callback(ompt_callback_masked)(
767           ompt_scope_begin, &(team->t.ompt_team_info.parallel_data),
768           &(team->t.t_implicit_task_taskdata[tid].ompt_task_info.task_data),
769           OMPT_GET_RETURN_ADDRESS(0));
770     }
771   }
772 #endif
773 
774   if (__kmp_env_consistency_check) {
775 #if KMP_USE_DYNAMIC_LOCK
776     if (status)
777       __kmp_push_sync(global_tid, ct_master, loc, NULL, 0);
778     else
779       __kmp_check_sync(global_tid, ct_master, loc, NULL, 0);
780 #else
781     if (status)
782       __kmp_push_sync(global_tid, ct_master, loc, NULL);
783     else
784       __kmp_check_sync(global_tid, ct_master, loc, NULL);
785 #endif
786   }
787 
788   return status;
789 }
790 
791 /*!
792 @ingroup WORK_SHARING
793 @param loc  source location information.
794 @param global_tid  global thread number .
795 
796 Mark the end of a <tt>master</tt> region. This should only be called by the
797 thread that executes the <tt>master</tt> region.
798 */
799 void __kmpc_end_master(ident_t *loc, kmp_int32 global_tid) {
800   KC_TRACE(10, ("__kmpc_end_master: called T#%d\n", global_tid));
801   __kmp_assert_valid_gtid(global_tid);
802   KMP_DEBUG_ASSERT(KMP_MASTER_GTID(global_tid));
803   KMP_POP_PARTITIONED_TIMER();
804 
805 #if OMPT_SUPPORT && OMPT_OPTIONAL
806   kmp_info_t *this_thr = __kmp_threads[global_tid];
807   kmp_team_t *team = this_thr->th.th_team;
808   if (ompt_enabled.ompt_callback_masked) {
809     int tid = __kmp_tid_from_gtid(global_tid);
810     ompt_callbacks.ompt_callback(ompt_callback_masked)(
811         ompt_scope_end, &(team->t.ompt_team_info.parallel_data),
812         &(team->t.t_implicit_task_taskdata[tid].ompt_task_info.task_data),
813         OMPT_GET_RETURN_ADDRESS(0));
814   }
815 #endif
816 
817   if (__kmp_env_consistency_check) {
818     if (KMP_MASTER_GTID(global_tid))
819       __kmp_pop_sync(global_tid, ct_master, loc);
820   }
821 }
822 
823 /*!
824 @ingroup WORK_SHARING
825 @param loc  source location information.
826 @param gtid  global thread number.
827 
828 Start execution of an <tt>ordered</tt> construct.
829 */
830 void __kmpc_ordered(ident_t *loc, kmp_int32 gtid) {
831   int cid = 0;
832   kmp_info_t *th;
833   KMP_DEBUG_ASSERT(__kmp_init_serial);
834 
835   KC_TRACE(10, ("__kmpc_ordered: called T#%d\n", gtid));
836   __kmp_assert_valid_gtid(gtid);
837 
838   if (!TCR_4(__kmp_init_parallel))
839     __kmp_parallel_initialize();
840 
841   __kmp_resume_if_soft_paused();
842 
843 #if USE_ITT_BUILD
844   __kmp_itt_ordered_prep(gtid);
845 // TODO: ordered_wait_id
846 #endif /* USE_ITT_BUILD */
847 
848   th = __kmp_threads[gtid];
849 
850 #if OMPT_SUPPORT && OMPT_OPTIONAL
851   kmp_team_t *team;
852   ompt_wait_id_t lck;
853   void *codeptr_ra;
854   if (ompt_enabled.enabled) {
855     OMPT_STORE_RETURN_ADDRESS(gtid);
856     team = __kmp_team_from_gtid(gtid);
857     lck = (ompt_wait_id_t)(uintptr_t)&team->t.t_ordered.dt.t_value;
858     /* OMPT state update */
859     th->th.ompt_thread_info.wait_id = lck;
860     th->th.ompt_thread_info.state = ompt_state_wait_ordered;
861 
862     /* OMPT event callback */
863     codeptr_ra = OMPT_LOAD_RETURN_ADDRESS(gtid);
864     if (ompt_enabled.ompt_callback_mutex_acquire) {
865       ompt_callbacks.ompt_callback(ompt_callback_mutex_acquire)(
866           ompt_mutex_ordered, omp_lock_hint_none, kmp_mutex_impl_spin, lck,
867           codeptr_ra);
868     }
869   }
870 #endif
871 
872   if (th->th.th_dispatch->th_deo_fcn != 0)
873     (*th->th.th_dispatch->th_deo_fcn)(&gtid, &cid, loc);
874   else
875     __kmp_parallel_deo(&gtid, &cid, loc);
876 
877 #if OMPT_SUPPORT && OMPT_OPTIONAL
878   if (ompt_enabled.enabled) {
879     /* OMPT state update */
880     th->th.ompt_thread_info.state = ompt_state_work_parallel;
881     th->th.ompt_thread_info.wait_id = 0;
882 
883     /* OMPT event callback */
884     if (ompt_enabled.ompt_callback_mutex_acquired) {
885       ompt_callbacks.ompt_callback(ompt_callback_mutex_acquired)(
886           ompt_mutex_ordered, (ompt_wait_id_t)(uintptr_t)lck, codeptr_ra);
887     }
888   }
889 #endif
890 
891 #if USE_ITT_BUILD
892   __kmp_itt_ordered_start(gtid);
893 #endif /* USE_ITT_BUILD */
894 }
895 
896 /*!
897 @ingroup WORK_SHARING
898 @param loc  source location information.
899 @param gtid  global thread number.
900 
901 End execution of an <tt>ordered</tt> construct.
902 */
903 void __kmpc_end_ordered(ident_t *loc, kmp_int32 gtid) {
904   int cid = 0;
905   kmp_info_t *th;
906 
907   KC_TRACE(10, ("__kmpc_end_ordered: called T#%d\n", gtid));
908   __kmp_assert_valid_gtid(gtid);
909 
910 #if USE_ITT_BUILD
911   __kmp_itt_ordered_end(gtid);
912 // TODO: ordered_wait_id
913 #endif /* USE_ITT_BUILD */
914 
915   th = __kmp_threads[gtid];
916 
917   if (th->th.th_dispatch->th_dxo_fcn != 0)
918     (*th->th.th_dispatch->th_dxo_fcn)(&gtid, &cid, loc);
919   else
920     __kmp_parallel_dxo(&gtid, &cid, loc);
921 
922 #if OMPT_SUPPORT && OMPT_OPTIONAL
923   OMPT_STORE_RETURN_ADDRESS(gtid);
924   if (ompt_enabled.ompt_callback_mutex_released) {
925     ompt_callbacks.ompt_callback(ompt_callback_mutex_released)(
926         ompt_mutex_ordered,
927         (ompt_wait_id_t)(uintptr_t)&__kmp_team_from_gtid(gtid)
928             ->t.t_ordered.dt.t_value,
929         OMPT_LOAD_RETURN_ADDRESS(gtid));
930   }
931 #endif
932 }
933 
934 #if KMP_USE_DYNAMIC_LOCK
935 
936 static __forceinline void
937 __kmp_init_indirect_csptr(kmp_critical_name *crit, ident_t const *loc,
938                           kmp_int32 gtid, kmp_indirect_locktag_t tag) {
939   // Pointer to the allocated indirect lock is written to crit, while indexing
940   // is ignored.
941   void *idx;
942   kmp_indirect_lock_t **lck;
943   lck = (kmp_indirect_lock_t **)crit;
944   kmp_indirect_lock_t *ilk = __kmp_allocate_indirect_lock(&idx, gtid, tag);
945   KMP_I_LOCK_FUNC(ilk, init)(ilk->lock);
946   KMP_SET_I_LOCK_LOCATION(ilk, loc);
947   KMP_SET_I_LOCK_FLAGS(ilk, kmp_lf_critical_section);
948   KA_TRACE(20,
949            ("__kmp_init_indirect_csptr: initialized indirect lock #%d\n", tag));
950 #if USE_ITT_BUILD
951   __kmp_itt_critical_creating(ilk->lock, loc);
952 #endif
953   int status = KMP_COMPARE_AND_STORE_PTR(lck, nullptr, ilk);
954   if (status == 0) {
955 #if USE_ITT_BUILD
956     __kmp_itt_critical_destroyed(ilk->lock);
957 #endif
958     // We don't really need to destroy the unclaimed lock here since it will be
959     // cleaned up at program exit.
960     // KMP_D_LOCK_FUNC(&idx, destroy)((kmp_dyna_lock_t *)&idx);
961   }
962   KMP_DEBUG_ASSERT(*lck != NULL);
963 }
964 
965 // Fast-path acquire tas lock
966 #define KMP_ACQUIRE_TAS_LOCK(lock, gtid)                                       \
967   {                                                                            \
968     kmp_tas_lock_t *l = (kmp_tas_lock_t *)lock;                                \
969     kmp_int32 tas_free = KMP_LOCK_FREE(tas);                                   \
970     kmp_int32 tas_busy = KMP_LOCK_BUSY(gtid + 1, tas);                         \
971     if (KMP_ATOMIC_LD_RLX(&l->lk.poll) != tas_free ||                          \
972         !__kmp_atomic_compare_store_acq(&l->lk.poll, tas_free, tas_busy)) {    \
973       kmp_uint32 spins;                                                        \
974       KMP_FSYNC_PREPARE(l);                                                    \
975       KMP_INIT_YIELD(spins);                                                   \
976       kmp_backoff_t backoff = __kmp_spin_backoff_params;                       \
977       do {                                                                     \
978         if (TCR_4(__kmp_nth) >                                                 \
979             (__kmp_avail_proc ? __kmp_avail_proc : __kmp_xproc)) {             \
980           KMP_YIELD(TRUE);                                                     \
981         } else {                                                               \
982           KMP_YIELD_SPIN(spins);                                               \
983         }                                                                      \
984         __kmp_spin_backoff(&backoff);                                          \
985       } while (                                                                \
986           KMP_ATOMIC_LD_RLX(&l->lk.poll) != tas_free ||                        \
987           !__kmp_atomic_compare_store_acq(&l->lk.poll, tas_free, tas_busy));   \
988     }                                                                          \
989     KMP_FSYNC_ACQUIRED(l);                                                     \
990   }
991 
992 // Fast-path test tas lock
993 #define KMP_TEST_TAS_LOCK(lock, gtid, rc)                                      \
994   {                                                                            \
995     kmp_tas_lock_t *l = (kmp_tas_lock_t *)lock;                                \
996     kmp_int32 tas_free = KMP_LOCK_FREE(tas);                                   \
997     kmp_int32 tas_busy = KMP_LOCK_BUSY(gtid + 1, tas);                         \
998     rc = KMP_ATOMIC_LD_RLX(&l->lk.poll) == tas_free &&                         \
999          __kmp_atomic_compare_store_acq(&l->lk.poll, tas_free, tas_busy);      \
1000   }
1001 
1002 // Fast-path release tas lock
1003 #define KMP_RELEASE_TAS_LOCK(lock, gtid)                                       \
1004   { KMP_ATOMIC_ST_REL(&((kmp_tas_lock_t *)lock)->lk.poll, KMP_LOCK_FREE(tas)); }
1005 
1006 #if KMP_USE_FUTEX
1007 
1008 #include <sys/syscall.h>
1009 #include <unistd.h>
1010 #ifndef FUTEX_WAIT
1011 #define FUTEX_WAIT 0
1012 #endif
1013 #ifndef FUTEX_WAKE
1014 #define FUTEX_WAKE 1
1015 #endif
1016 
1017 // Fast-path acquire futex lock
1018 #define KMP_ACQUIRE_FUTEX_LOCK(lock, gtid)                                     \
1019   {                                                                            \
1020     kmp_futex_lock_t *ftx = (kmp_futex_lock_t *)lock;                          \
1021     kmp_int32 gtid_code = (gtid + 1) << 1;                                     \
1022     KMP_MB();                                                                  \
1023     KMP_FSYNC_PREPARE(ftx);                                                    \
1024     kmp_int32 poll_val;                                                        \
1025     while ((poll_val = KMP_COMPARE_AND_STORE_RET32(                            \
1026                 &(ftx->lk.poll), KMP_LOCK_FREE(futex),                         \
1027                 KMP_LOCK_BUSY(gtid_code, futex))) != KMP_LOCK_FREE(futex)) {   \
1028       kmp_int32 cond = KMP_LOCK_STRIP(poll_val) & 1;                           \
1029       if (!cond) {                                                             \
1030         if (!KMP_COMPARE_AND_STORE_RET32(&(ftx->lk.poll), poll_val,            \
1031                                          poll_val |                            \
1032                                              KMP_LOCK_BUSY(1, futex))) {       \
1033           continue;                                                            \
1034         }                                                                      \
1035         poll_val |= KMP_LOCK_BUSY(1, futex);                                   \
1036       }                                                                        \
1037       kmp_int32 rc;                                                            \
1038       if ((rc = syscall(__NR_futex, &(ftx->lk.poll), FUTEX_WAIT, poll_val,     \
1039                         NULL, NULL, 0)) != 0) {                                \
1040         continue;                                                              \
1041       }                                                                        \
1042       gtid_code |= 1;                                                          \
1043     }                                                                          \
1044     KMP_FSYNC_ACQUIRED(ftx);                                                   \
1045   }
1046 
1047 // Fast-path test futex lock
1048 #define KMP_TEST_FUTEX_LOCK(lock, gtid, rc)                                    \
1049   {                                                                            \
1050     kmp_futex_lock_t *ftx = (kmp_futex_lock_t *)lock;                          \
1051     if (KMP_COMPARE_AND_STORE_ACQ32(&(ftx->lk.poll), KMP_LOCK_FREE(futex),     \
1052                                     KMP_LOCK_BUSY(gtid + 1 << 1, futex))) {    \
1053       KMP_FSYNC_ACQUIRED(ftx);                                                 \
1054       rc = TRUE;                                                               \
1055     } else {                                                                   \
1056       rc = FALSE;                                                              \
1057     }                                                                          \
1058   }
1059 
1060 // Fast-path release futex lock
1061 #define KMP_RELEASE_FUTEX_LOCK(lock, gtid)                                     \
1062   {                                                                            \
1063     kmp_futex_lock_t *ftx = (kmp_futex_lock_t *)lock;                          \
1064     KMP_MB();                                                                  \
1065     KMP_FSYNC_RELEASING(ftx);                                                  \
1066     kmp_int32 poll_val =                                                       \
1067         KMP_XCHG_FIXED32(&(ftx->lk.poll), KMP_LOCK_FREE(futex));               \
1068     if (KMP_LOCK_STRIP(poll_val) & 1) {                                        \
1069       syscall(__NR_futex, &(ftx->lk.poll), FUTEX_WAKE,                         \
1070               KMP_LOCK_BUSY(1, futex), NULL, NULL, 0);                         \
1071     }                                                                          \
1072     KMP_MB();                                                                  \
1073     KMP_YIELD_OVERSUB();                                                       \
1074   }
1075 
1076 #endif // KMP_USE_FUTEX
1077 
1078 #else // KMP_USE_DYNAMIC_LOCK
1079 
1080 static kmp_user_lock_p __kmp_get_critical_section_ptr(kmp_critical_name *crit,
1081                                                       ident_t const *loc,
1082                                                       kmp_int32 gtid) {
1083   kmp_user_lock_p *lck_pp = (kmp_user_lock_p *)crit;
1084 
1085   // Because of the double-check, the following load doesn't need to be volatile
1086   kmp_user_lock_p lck = (kmp_user_lock_p)TCR_PTR(*lck_pp);
1087 
1088   if (lck == NULL) {
1089     void *idx;
1090 
1091     // Allocate & initialize the lock.
1092     // Remember alloc'ed locks in table in order to free them in __kmp_cleanup()
1093     lck = __kmp_user_lock_allocate(&idx, gtid, kmp_lf_critical_section);
1094     __kmp_init_user_lock_with_checks(lck);
1095     __kmp_set_user_lock_location(lck, loc);
1096 #if USE_ITT_BUILD
1097     __kmp_itt_critical_creating(lck);
1098 // __kmp_itt_critical_creating() should be called *before* the first usage
1099 // of underlying lock. It is the only place where we can guarantee it. There
1100 // are chances the lock will destroyed with no usage, but it is not a
1101 // problem, because this is not real event seen by user but rather setting
1102 // name for object (lock). See more details in kmp_itt.h.
1103 #endif /* USE_ITT_BUILD */
1104 
1105     // Use a cmpxchg instruction to slam the start of the critical section with
1106     // the lock pointer.  If another thread beat us to it, deallocate the lock,
1107     // and use the lock that the other thread allocated.
1108     int status = KMP_COMPARE_AND_STORE_PTR(lck_pp, 0, lck);
1109 
1110     if (status == 0) {
1111 // Deallocate the lock and reload the value.
1112 #if USE_ITT_BUILD
1113       __kmp_itt_critical_destroyed(lck);
1114 // Let ITT know the lock is destroyed and the same memory location may be reused
1115 // for another purpose.
1116 #endif /* USE_ITT_BUILD */
1117       __kmp_destroy_user_lock_with_checks(lck);
1118       __kmp_user_lock_free(&idx, gtid, lck);
1119       lck = (kmp_user_lock_p)TCR_PTR(*lck_pp);
1120       KMP_DEBUG_ASSERT(lck != NULL);
1121     }
1122   }
1123   return lck;
1124 }
1125 
1126 #endif // KMP_USE_DYNAMIC_LOCK
1127 
1128 /*!
1129 @ingroup WORK_SHARING
1130 @param loc  source location information.
1131 @param global_tid  global thread number.
1132 @param crit identity of the critical section. This could be a pointer to a lock
1133 associated with the critical section, or some other suitably unique value.
1134 
1135 Enter code protected by a `critical` construct.
1136 This function blocks until the executing thread can enter the critical section.
1137 */
1138 void __kmpc_critical(ident_t *loc, kmp_int32 global_tid,
1139                      kmp_critical_name *crit) {
1140 #if KMP_USE_DYNAMIC_LOCK
1141 #if OMPT_SUPPORT && OMPT_OPTIONAL
1142   OMPT_STORE_RETURN_ADDRESS(global_tid);
1143 #endif // OMPT_SUPPORT
1144   __kmpc_critical_with_hint(loc, global_tid, crit, omp_lock_hint_none);
1145 #else
1146   KMP_COUNT_BLOCK(OMP_CRITICAL);
1147 #if OMPT_SUPPORT && OMPT_OPTIONAL
1148   ompt_state_t prev_state = ompt_state_undefined;
1149   ompt_thread_info_t ti;
1150 #endif
1151   kmp_user_lock_p lck;
1152 
1153   KC_TRACE(10, ("__kmpc_critical: called T#%d\n", global_tid));
1154   __kmp_assert_valid_gtid(global_tid);
1155 
1156   // TODO: add THR_OVHD_STATE
1157 
1158   KMP_PUSH_PARTITIONED_TIMER(OMP_critical_wait);
1159   KMP_CHECK_USER_LOCK_INIT();
1160 
1161   if ((__kmp_user_lock_kind == lk_tas) &&
1162       (sizeof(lck->tas.lk.poll) <= OMP_CRITICAL_SIZE)) {
1163     lck = (kmp_user_lock_p)crit;
1164   }
1165 #if KMP_USE_FUTEX
1166   else if ((__kmp_user_lock_kind == lk_futex) &&
1167            (sizeof(lck->futex.lk.poll) <= OMP_CRITICAL_SIZE)) {
1168     lck = (kmp_user_lock_p)crit;
1169   }
1170 #endif
1171   else { // ticket, queuing or drdpa
1172     lck = __kmp_get_critical_section_ptr(crit, loc, global_tid);
1173   }
1174 
1175   if (__kmp_env_consistency_check)
1176     __kmp_push_sync(global_tid, ct_critical, loc, lck);
1177 
1178 // since the critical directive binds to all threads, not just the current
1179 // team we have to check this even if we are in a serialized team.
1180 // also, even if we are the uber thread, we still have to conduct the lock,
1181 // as we have to contend with sibling threads.
1182 
1183 #if USE_ITT_BUILD
1184   __kmp_itt_critical_acquiring(lck);
1185 #endif /* USE_ITT_BUILD */
1186 #if OMPT_SUPPORT && OMPT_OPTIONAL
1187   OMPT_STORE_RETURN_ADDRESS(gtid);
1188   void *codeptr_ra = NULL;
1189   if (ompt_enabled.enabled) {
1190     ti = __kmp_threads[global_tid]->th.ompt_thread_info;
1191     /* OMPT state update */
1192     prev_state = ti.state;
1193     ti.wait_id = (ompt_wait_id_t)(uintptr_t)lck;
1194     ti.state = ompt_state_wait_critical;
1195 
1196     /* OMPT event callback */
1197     codeptr_ra = OMPT_LOAD_RETURN_ADDRESS(gtid);
1198     if (ompt_enabled.ompt_callback_mutex_acquire) {
1199       ompt_callbacks.ompt_callback(ompt_callback_mutex_acquire)(
1200           ompt_mutex_critical, omp_lock_hint_none, __ompt_get_mutex_impl_type(),
1201           (ompt_wait_id_t)(uintptr_t)lck, codeptr_ra);
1202     }
1203   }
1204 #endif
1205   // Value of 'crit' should be good for using as a critical_id of the critical
1206   // section directive.
1207   __kmp_acquire_user_lock_with_checks(lck, global_tid);
1208 
1209 #if USE_ITT_BUILD
1210   __kmp_itt_critical_acquired(lck);
1211 #endif /* USE_ITT_BUILD */
1212 #if OMPT_SUPPORT && OMPT_OPTIONAL
1213   if (ompt_enabled.enabled) {
1214     /* OMPT state update */
1215     ti.state = prev_state;
1216     ti.wait_id = 0;
1217 
1218     /* OMPT event callback */
1219     if (ompt_enabled.ompt_callback_mutex_acquired) {
1220       ompt_callbacks.ompt_callback(ompt_callback_mutex_acquired)(
1221           ompt_mutex_critical, (ompt_wait_id_t)(uintptr_t)lck, codeptr_ra);
1222     }
1223   }
1224 #endif
1225   KMP_POP_PARTITIONED_TIMER();
1226 
1227   KMP_PUSH_PARTITIONED_TIMER(OMP_critical);
1228   KA_TRACE(15, ("__kmpc_critical: done T#%d\n", global_tid));
1229 #endif // KMP_USE_DYNAMIC_LOCK
1230 }
1231 
1232 #if KMP_USE_DYNAMIC_LOCK
1233 
1234 // Converts the given hint to an internal lock implementation
1235 static __forceinline kmp_dyna_lockseq_t __kmp_map_hint_to_lock(uintptr_t hint) {
1236 #if KMP_USE_TSX
1237 #define KMP_TSX_LOCK(seq) lockseq_##seq
1238 #else
1239 #define KMP_TSX_LOCK(seq) __kmp_user_lock_seq
1240 #endif
1241 
1242 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
1243 #define KMP_CPUINFO_RTM (__kmp_cpuinfo.rtm)
1244 #else
1245 #define KMP_CPUINFO_RTM 0
1246 #endif
1247 
1248   // Hints that do not require further logic
1249   if (hint & kmp_lock_hint_hle)
1250     return KMP_TSX_LOCK(hle);
1251   if (hint & kmp_lock_hint_rtm)
1252     return KMP_CPUINFO_RTM ? KMP_TSX_LOCK(rtm) : __kmp_user_lock_seq;
1253   if (hint & kmp_lock_hint_adaptive)
1254     return KMP_CPUINFO_RTM ? KMP_TSX_LOCK(adaptive) : __kmp_user_lock_seq;
1255 
1256   // Rule out conflicting hints first by returning the default lock
1257   if ((hint & omp_lock_hint_contended) && (hint & omp_lock_hint_uncontended))
1258     return __kmp_user_lock_seq;
1259   if ((hint & omp_lock_hint_speculative) &&
1260       (hint & omp_lock_hint_nonspeculative))
1261     return __kmp_user_lock_seq;
1262 
1263   // Do not even consider speculation when it appears to be contended
1264   if (hint & omp_lock_hint_contended)
1265     return lockseq_queuing;
1266 
1267   // Uncontended lock without speculation
1268   if ((hint & omp_lock_hint_uncontended) && !(hint & omp_lock_hint_speculative))
1269     return lockseq_tas;
1270 
1271   // HLE lock for speculation
1272   if (hint & omp_lock_hint_speculative)
1273     return KMP_TSX_LOCK(hle);
1274 
1275   return __kmp_user_lock_seq;
1276 }
1277 
1278 #if OMPT_SUPPORT && OMPT_OPTIONAL
1279 #if KMP_USE_DYNAMIC_LOCK
1280 static kmp_mutex_impl_t
1281 __ompt_get_mutex_impl_type(void *user_lock, kmp_indirect_lock_t *ilock = 0) {
1282   if (user_lock) {
1283     switch (KMP_EXTRACT_D_TAG(user_lock)) {
1284     case 0:
1285       break;
1286 #if KMP_USE_FUTEX
1287     case locktag_futex:
1288       return kmp_mutex_impl_queuing;
1289 #endif
1290     case locktag_tas:
1291       return kmp_mutex_impl_spin;
1292 #if KMP_USE_TSX
1293     case locktag_hle:
1294       return kmp_mutex_impl_speculative;
1295 #endif
1296     default:
1297       return kmp_mutex_impl_none;
1298     }
1299     ilock = KMP_LOOKUP_I_LOCK(user_lock);
1300   }
1301   KMP_ASSERT(ilock);
1302   switch (ilock->type) {
1303 #if KMP_USE_TSX
1304   case locktag_adaptive:
1305   case locktag_rtm:
1306     return kmp_mutex_impl_speculative;
1307 #endif
1308   case locktag_nested_tas:
1309     return kmp_mutex_impl_spin;
1310 #if KMP_USE_FUTEX
1311   case locktag_nested_futex:
1312 #endif
1313   case locktag_ticket:
1314   case locktag_queuing:
1315   case locktag_drdpa:
1316   case locktag_nested_ticket:
1317   case locktag_nested_queuing:
1318   case locktag_nested_drdpa:
1319     return kmp_mutex_impl_queuing;
1320   default:
1321     return kmp_mutex_impl_none;
1322   }
1323 }
1324 #else
1325 // For locks without dynamic binding
1326 static kmp_mutex_impl_t __ompt_get_mutex_impl_type() {
1327   switch (__kmp_user_lock_kind) {
1328   case lk_tas:
1329     return kmp_mutex_impl_spin;
1330 #if KMP_USE_FUTEX
1331   case lk_futex:
1332 #endif
1333   case lk_ticket:
1334   case lk_queuing:
1335   case lk_drdpa:
1336     return kmp_mutex_impl_queuing;
1337 #if KMP_USE_TSX
1338   case lk_hle:
1339   case lk_rtm:
1340   case lk_adaptive:
1341     return kmp_mutex_impl_speculative;
1342 #endif
1343   default:
1344     return kmp_mutex_impl_none;
1345   }
1346 }
1347 #endif // KMP_USE_DYNAMIC_LOCK
1348 #endif // OMPT_SUPPORT && OMPT_OPTIONAL
1349 
1350 /*!
1351 @ingroup WORK_SHARING
1352 @param loc  source location information.
1353 @param global_tid  global thread number.
1354 @param crit identity of the critical section. This could be a pointer to a lock
1355 associated with the critical section, or some other suitably unique value.
1356 @param hint the lock hint.
1357 
1358 Enter code protected by a `critical` construct with a hint. The hint value is
1359 used to suggest a lock implementation. This function blocks until the executing
1360 thread can enter the critical section unless the hint suggests use of
1361 speculative execution and the hardware supports it.
1362 */
1363 void __kmpc_critical_with_hint(ident_t *loc, kmp_int32 global_tid,
1364                                kmp_critical_name *crit, uint32_t hint) {
1365   KMP_COUNT_BLOCK(OMP_CRITICAL);
1366   kmp_user_lock_p lck;
1367 #if OMPT_SUPPORT && OMPT_OPTIONAL
1368   ompt_state_t prev_state = ompt_state_undefined;
1369   ompt_thread_info_t ti;
1370   // This is the case, if called from __kmpc_critical:
1371   void *codeptr = OMPT_LOAD_RETURN_ADDRESS(global_tid);
1372   if (!codeptr)
1373     codeptr = OMPT_GET_RETURN_ADDRESS(0);
1374 #endif
1375 
1376   KC_TRACE(10, ("__kmpc_critical: called T#%d\n", global_tid));
1377   __kmp_assert_valid_gtid(global_tid);
1378 
1379   kmp_dyna_lock_t *lk = (kmp_dyna_lock_t *)crit;
1380   // Check if it is initialized.
1381   KMP_PUSH_PARTITIONED_TIMER(OMP_critical_wait);
1382   if (*lk == 0) {
1383     kmp_dyna_lockseq_t lckseq = __kmp_map_hint_to_lock(hint);
1384     if (KMP_IS_D_LOCK(lckseq)) {
1385       KMP_COMPARE_AND_STORE_ACQ32((volatile kmp_int32 *)crit, 0,
1386                                   KMP_GET_D_TAG(lckseq));
1387     } else {
1388       __kmp_init_indirect_csptr(crit, loc, global_tid, KMP_GET_I_TAG(lckseq));
1389     }
1390   }
1391   // Branch for accessing the actual lock object and set operation. This
1392   // branching is inevitable since this lock initialization does not follow the
1393   // normal dispatch path (lock table is not used).
1394   if (KMP_EXTRACT_D_TAG(lk) != 0) {
1395     lck = (kmp_user_lock_p)lk;
1396     if (__kmp_env_consistency_check) {
1397       __kmp_push_sync(global_tid, ct_critical, loc, lck,
1398                       __kmp_map_hint_to_lock(hint));
1399     }
1400 #if USE_ITT_BUILD
1401     __kmp_itt_critical_acquiring(lck);
1402 #endif
1403 #if OMPT_SUPPORT && OMPT_OPTIONAL
1404     if (ompt_enabled.enabled) {
1405       ti = __kmp_threads[global_tid]->th.ompt_thread_info;
1406       /* OMPT state update */
1407       prev_state = ti.state;
1408       ti.wait_id = (ompt_wait_id_t)(uintptr_t)lck;
1409       ti.state = ompt_state_wait_critical;
1410 
1411       /* OMPT event callback */
1412       if (ompt_enabled.ompt_callback_mutex_acquire) {
1413         ompt_callbacks.ompt_callback(ompt_callback_mutex_acquire)(
1414             ompt_mutex_critical, (unsigned int)hint,
1415             __ompt_get_mutex_impl_type(crit), (ompt_wait_id_t)(uintptr_t)lck,
1416             codeptr);
1417       }
1418     }
1419 #endif
1420 #if KMP_USE_INLINED_TAS
1421     if (__kmp_user_lock_seq == lockseq_tas && !__kmp_env_consistency_check) {
1422       KMP_ACQUIRE_TAS_LOCK(lck, global_tid);
1423     } else
1424 #elif KMP_USE_INLINED_FUTEX
1425     if (__kmp_user_lock_seq == lockseq_futex && !__kmp_env_consistency_check) {
1426       KMP_ACQUIRE_FUTEX_LOCK(lck, global_tid);
1427     } else
1428 #endif
1429     {
1430       KMP_D_LOCK_FUNC(lk, set)(lk, global_tid);
1431     }
1432   } else {
1433     kmp_indirect_lock_t *ilk = *((kmp_indirect_lock_t **)lk);
1434     lck = ilk->lock;
1435     if (__kmp_env_consistency_check) {
1436       __kmp_push_sync(global_tid, ct_critical, loc, lck,
1437                       __kmp_map_hint_to_lock(hint));
1438     }
1439 #if USE_ITT_BUILD
1440     __kmp_itt_critical_acquiring(lck);
1441 #endif
1442 #if OMPT_SUPPORT && OMPT_OPTIONAL
1443     if (ompt_enabled.enabled) {
1444       ti = __kmp_threads[global_tid]->th.ompt_thread_info;
1445       /* OMPT state update */
1446       prev_state = ti.state;
1447       ti.wait_id = (ompt_wait_id_t)(uintptr_t)lck;
1448       ti.state = ompt_state_wait_critical;
1449 
1450       /* OMPT event callback */
1451       if (ompt_enabled.ompt_callback_mutex_acquire) {
1452         ompt_callbacks.ompt_callback(ompt_callback_mutex_acquire)(
1453             ompt_mutex_critical, (unsigned int)hint,
1454             __ompt_get_mutex_impl_type(0, ilk), (ompt_wait_id_t)(uintptr_t)lck,
1455             codeptr);
1456       }
1457     }
1458 #endif
1459     KMP_I_LOCK_FUNC(ilk, set)(lck, global_tid);
1460   }
1461   KMP_POP_PARTITIONED_TIMER();
1462 
1463 #if USE_ITT_BUILD
1464   __kmp_itt_critical_acquired(lck);
1465 #endif /* USE_ITT_BUILD */
1466 #if OMPT_SUPPORT && OMPT_OPTIONAL
1467   if (ompt_enabled.enabled) {
1468     /* OMPT state update */
1469     ti.state = prev_state;
1470     ti.wait_id = 0;
1471 
1472     /* OMPT event callback */
1473     if (ompt_enabled.ompt_callback_mutex_acquired) {
1474       ompt_callbacks.ompt_callback(ompt_callback_mutex_acquired)(
1475           ompt_mutex_critical, (ompt_wait_id_t)(uintptr_t)lck, codeptr);
1476     }
1477   }
1478 #endif
1479 
1480   KMP_PUSH_PARTITIONED_TIMER(OMP_critical);
1481   KA_TRACE(15, ("__kmpc_critical: done T#%d\n", global_tid));
1482 } // __kmpc_critical_with_hint
1483 
1484 #endif // KMP_USE_DYNAMIC_LOCK
1485 
1486 /*!
1487 @ingroup WORK_SHARING
1488 @param loc  source location information.
1489 @param global_tid  global thread number .
1490 @param crit identity of the critical section. This could be a pointer to a lock
1491 associated with the critical section, or some other suitably unique value.
1492 
1493 Leave a critical section, releasing any lock that was held during its execution.
1494 */
1495 void __kmpc_end_critical(ident_t *loc, kmp_int32 global_tid,
1496                          kmp_critical_name *crit) {
1497   kmp_user_lock_p lck;
1498 
1499   KC_TRACE(10, ("__kmpc_end_critical: called T#%d\n", global_tid));
1500 
1501 #if KMP_USE_DYNAMIC_LOCK
1502   if (KMP_IS_D_LOCK(__kmp_user_lock_seq)) {
1503     lck = (kmp_user_lock_p)crit;
1504     KMP_ASSERT(lck != NULL);
1505     if (__kmp_env_consistency_check) {
1506       __kmp_pop_sync(global_tid, ct_critical, loc);
1507     }
1508 #if USE_ITT_BUILD
1509     __kmp_itt_critical_releasing(lck);
1510 #endif
1511 #if KMP_USE_INLINED_TAS
1512     if (__kmp_user_lock_seq == lockseq_tas && !__kmp_env_consistency_check) {
1513       KMP_RELEASE_TAS_LOCK(lck, global_tid);
1514     } else
1515 #elif KMP_USE_INLINED_FUTEX
1516     if (__kmp_user_lock_seq == lockseq_futex && !__kmp_env_consistency_check) {
1517       KMP_RELEASE_FUTEX_LOCK(lck, global_tid);
1518     } else
1519 #endif
1520     {
1521       KMP_D_LOCK_FUNC(lck, unset)((kmp_dyna_lock_t *)lck, global_tid);
1522     }
1523   } else {
1524     kmp_indirect_lock_t *ilk =
1525         (kmp_indirect_lock_t *)TCR_PTR(*((kmp_indirect_lock_t **)crit));
1526     KMP_ASSERT(ilk != NULL);
1527     lck = ilk->lock;
1528     if (__kmp_env_consistency_check) {
1529       __kmp_pop_sync(global_tid, ct_critical, loc);
1530     }
1531 #if USE_ITT_BUILD
1532     __kmp_itt_critical_releasing(lck);
1533 #endif
1534     KMP_I_LOCK_FUNC(ilk, unset)(lck, global_tid);
1535   }
1536 
1537 #else // KMP_USE_DYNAMIC_LOCK
1538 
1539   if ((__kmp_user_lock_kind == lk_tas) &&
1540       (sizeof(lck->tas.lk.poll) <= OMP_CRITICAL_SIZE)) {
1541     lck = (kmp_user_lock_p)crit;
1542   }
1543 #if KMP_USE_FUTEX
1544   else if ((__kmp_user_lock_kind == lk_futex) &&
1545            (sizeof(lck->futex.lk.poll) <= OMP_CRITICAL_SIZE)) {
1546     lck = (kmp_user_lock_p)crit;
1547   }
1548 #endif
1549   else { // ticket, queuing or drdpa
1550     lck = (kmp_user_lock_p)TCR_PTR(*((kmp_user_lock_p *)crit));
1551   }
1552 
1553   KMP_ASSERT(lck != NULL);
1554 
1555   if (__kmp_env_consistency_check)
1556     __kmp_pop_sync(global_tid, ct_critical, loc);
1557 
1558 #if USE_ITT_BUILD
1559   __kmp_itt_critical_releasing(lck);
1560 #endif /* USE_ITT_BUILD */
1561   // Value of 'crit' should be good for using as a critical_id of the critical
1562   // section directive.
1563   __kmp_release_user_lock_with_checks(lck, global_tid);
1564 
1565 #endif // KMP_USE_DYNAMIC_LOCK
1566 
1567 #if OMPT_SUPPORT && OMPT_OPTIONAL
1568   /* OMPT release event triggers after lock is released; place here to trigger
1569    * for all #if branches */
1570   OMPT_STORE_RETURN_ADDRESS(global_tid);
1571   if (ompt_enabled.ompt_callback_mutex_released) {
1572     ompt_callbacks.ompt_callback(ompt_callback_mutex_released)(
1573         ompt_mutex_critical, (ompt_wait_id_t)(uintptr_t)lck,
1574         OMPT_LOAD_RETURN_ADDRESS(0));
1575   }
1576 #endif
1577 
1578   KMP_POP_PARTITIONED_TIMER();
1579   KA_TRACE(15, ("__kmpc_end_critical: done T#%d\n", global_tid));
1580 }
1581 
1582 /*!
1583 @ingroup SYNCHRONIZATION
1584 @param loc source location information
1585 @param global_tid thread id.
1586 @return one if the thread should execute the master block, zero otherwise
1587 
1588 Start execution of a combined barrier and master. The barrier is executed inside
1589 this function.
1590 */
1591 kmp_int32 __kmpc_barrier_master(ident_t *loc, kmp_int32 global_tid) {
1592   int status;
1593   KC_TRACE(10, ("__kmpc_barrier_master: called T#%d\n", global_tid));
1594   __kmp_assert_valid_gtid(global_tid);
1595 
1596   if (!TCR_4(__kmp_init_parallel))
1597     __kmp_parallel_initialize();
1598 
1599   __kmp_resume_if_soft_paused();
1600 
1601   if (__kmp_env_consistency_check)
1602     __kmp_check_barrier(global_tid, ct_barrier, loc);
1603 
1604 #if OMPT_SUPPORT
1605   ompt_frame_t *ompt_frame;
1606   if (ompt_enabled.enabled) {
1607     __ompt_get_task_info_internal(0, NULL, NULL, &ompt_frame, NULL, NULL);
1608     if (ompt_frame->enter_frame.ptr == NULL)
1609       ompt_frame->enter_frame.ptr = OMPT_GET_FRAME_ADDRESS(0);
1610     OMPT_STORE_RETURN_ADDRESS(global_tid);
1611   }
1612 #endif
1613 #if USE_ITT_NOTIFY
1614   __kmp_threads[global_tid]->th.th_ident = loc;
1615 #endif
1616   status = __kmp_barrier(bs_plain_barrier, global_tid, TRUE, 0, NULL, NULL);
1617 #if OMPT_SUPPORT && OMPT_OPTIONAL
1618   if (ompt_enabled.enabled) {
1619     ompt_frame->enter_frame = ompt_data_none;
1620   }
1621 #endif
1622 
1623   return (status != 0) ? 0 : 1;
1624 }
1625 
1626 /*!
1627 @ingroup SYNCHRONIZATION
1628 @param loc source location information
1629 @param global_tid thread id.
1630 
1631 Complete the execution of a combined barrier and master. This function should
1632 only be called at the completion of the <tt>master</tt> code. Other threads will
1633 still be waiting at the barrier and this call releases them.
1634 */
1635 void __kmpc_end_barrier_master(ident_t *loc, kmp_int32 global_tid) {
1636   KC_TRACE(10, ("__kmpc_end_barrier_master: called T#%d\n", global_tid));
1637   __kmp_assert_valid_gtid(global_tid);
1638   __kmp_end_split_barrier(bs_plain_barrier, global_tid);
1639 }
1640 
1641 /*!
1642 @ingroup SYNCHRONIZATION
1643 @param loc source location information
1644 @param global_tid thread id.
1645 @return one if the thread should execute the master block, zero otherwise
1646 
1647 Start execution of a combined barrier and master(nowait) construct.
1648 The barrier is executed inside this function.
1649 There is no equivalent "end" function, since the
1650 */
1651 kmp_int32 __kmpc_barrier_master_nowait(ident_t *loc, kmp_int32 global_tid) {
1652   kmp_int32 ret;
1653   KC_TRACE(10, ("__kmpc_barrier_master_nowait: called T#%d\n", global_tid));
1654   __kmp_assert_valid_gtid(global_tid);
1655 
1656   if (!TCR_4(__kmp_init_parallel))
1657     __kmp_parallel_initialize();
1658 
1659   __kmp_resume_if_soft_paused();
1660 
1661   if (__kmp_env_consistency_check) {
1662     if (loc == 0) {
1663       KMP_WARNING(ConstructIdentInvalid); // ??? What does it mean for the user?
1664     }
1665     __kmp_check_barrier(global_tid, ct_barrier, loc);
1666   }
1667 
1668 #if OMPT_SUPPORT
1669   ompt_frame_t *ompt_frame;
1670   if (ompt_enabled.enabled) {
1671     __ompt_get_task_info_internal(0, NULL, NULL, &ompt_frame, NULL, NULL);
1672     if (ompt_frame->enter_frame.ptr == NULL)
1673       ompt_frame->enter_frame.ptr = OMPT_GET_FRAME_ADDRESS(0);
1674     OMPT_STORE_RETURN_ADDRESS(global_tid);
1675   }
1676 #endif
1677 #if USE_ITT_NOTIFY
1678   __kmp_threads[global_tid]->th.th_ident = loc;
1679 #endif
1680   __kmp_barrier(bs_plain_barrier, global_tid, FALSE, 0, NULL, NULL);
1681 #if OMPT_SUPPORT && OMPT_OPTIONAL
1682   if (ompt_enabled.enabled) {
1683     ompt_frame->enter_frame = ompt_data_none;
1684   }
1685 #endif
1686 
1687   ret = __kmpc_master(loc, global_tid);
1688 
1689   if (__kmp_env_consistency_check) {
1690     /*  there's no __kmpc_end_master called; so the (stats) */
1691     /*  actions of __kmpc_end_master are done here          */
1692     if (ret) {
1693       /* only one thread should do the pop since only */
1694       /* one did the push (see __kmpc_master())       */
1695       __kmp_pop_sync(global_tid, ct_master, loc);
1696     }
1697   }
1698 
1699   return (ret);
1700 }
1701 
1702 /* The BARRIER for a SINGLE process section is always explicit   */
1703 /*!
1704 @ingroup WORK_SHARING
1705 @param loc  source location information
1706 @param global_tid  global thread number
1707 @return One if this thread should execute the single construct, zero otherwise.
1708 
1709 Test whether to execute a <tt>single</tt> construct.
1710 There are no implicit barriers in the two "single" calls, rather the compiler
1711 should introduce an explicit barrier if it is required.
1712 */
1713 
1714 kmp_int32 __kmpc_single(ident_t *loc, kmp_int32 global_tid) {
1715   __kmp_assert_valid_gtid(global_tid);
1716   kmp_int32 rc = __kmp_enter_single(global_tid, loc, TRUE);
1717 
1718   if (rc) {
1719     // We are going to execute the single statement, so we should count it.
1720     KMP_COUNT_BLOCK(OMP_SINGLE);
1721     KMP_PUSH_PARTITIONED_TIMER(OMP_single);
1722   }
1723 
1724 #if OMPT_SUPPORT && OMPT_OPTIONAL
1725   kmp_info_t *this_thr = __kmp_threads[global_tid];
1726   kmp_team_t *team = this_thr->th.th_team;
1727   int tid = __kmp_tid_from_gtid(global_tid);
1728 
1729   if (ompt_enabled.enabled) {
1730     if (rc) {
1731       if (ompt_enabled.ompt_callback_work) {
1732         ompt_callbacks.ompt_callback(ompt_callback_work)(
1733             ompt_work_single_executor, ompt_scope_begin,
1734             &(team->t.ompt_team_info.parallel_data),
1735             &(team->t.t_implicit_task_taskdata[tid].ompt_task_info.task_data),
1736             1, OMPT_GET_RETURN_ADDRESS(0));
1737       }
1738     } else {
1739       if (ompt_enabled.ompt_callback_work) {
1740         ompt_callbacks.ompt_callback(ompt_callback_work)(
1741             ompt_work_single_other, ompt_scope_begin,
1742             &(team->t.ompt_team_info.parallel_data),
1743             &(team->t.t_implicit_task_taskdata[tid].ompt_task_info.task_data),
1744             1, OMPT_GET_RETURN_ADDRESS(0));
1745         ompt_callbacks.ompt_callback(ompt_callback_work)(
1746             ompt_work_single_other, ompt_scope_end,
1747             &(team->t.ompt_team_info.parallel_data),
1748             &(team->t.t_implicit_task_taskdata[tid].ompt_task_info.task_data),
1749             1, OMPT_GET_RETURN_ADDRESS(0));
1750       }
1751     }
1752   }
1753 #endif
1754 
1755   return rc;
1756 }
1757 
1758 /*!
1759 @ingroup WORK_SHARING
1760 @param loc  source location information
1761 @param global_tid  global thread number
1762 
1763 Mark the end of a <tt>single</tt> construct.  This function should
1764 only be called by the thread that executed the block of code protected
1765 by the `single` construct.
1766 */
1767 void __kmpc_end_single(ident_t *loc, kmp_int32 global_tid) {
1768   __kmp_assert_valid_gtid(global_tid);
1769   __kmp_exit_single(global_tid);
1770   KMP_POP_PARTITIONED_TIMER();
1771 
1772 #if OMPT_SUPPORT && OMPT_OPTIONAL
1773   kmp_info_t *this_thr = __kmp_threads[global_tid];
1774   kmp_team_t *team = this_thr->th.th_team;
1775   int tid = __kmp_tid_from_gtid(global_tid);
1776 
1777   if (ompt_enabled.ompt_callback_work) {
1778     ompt_callbacks.ompt_callback(ompt_callback_work)(
1779         ompt_work_single_executor, ompt_scope_end,
1780         &(team->t.ompt_team_info.parallel_data),
1781         &(team->t.t_implicit_task_taskdata[tid].ompt_task_info.task_data), 1,
1782         OMPT_GET_RETURN_ADDRESS(0));
1783   }
1784 #endif
1785 }
1786 
1787 /*!
1788 @ingroup WORK_SHARING
1789 @param loc Source location
1790 @param global_tid Global thread id
1791 
1792 Mark the end of a statically scheduled loop.
1793 */
1794 void __kmpc_for_static_fini(ident_t *loc, kmp_int32 global_tid) {
1795   KMP_POP_PARTITIONED_TIMER();
1796   KE_TRACE(10, ("__kmpc_for_static_fini called T#%d\n", global_tid));
1797 
1798 #if OMPT_SUPPORT && OMPT_OPTIONAL
1799   if (ompt_enabled.ompt_callback_work) {
1800     ompt_work_t ompt_work_type = ompt_work_loop;
1801     ompt_team_info_t *team_info = __ompt_get_teaminfo(0, NULL);
1802     ompt_task_info_t *task_info = __ompt_get_task_info_object(0);
1803     // Determine workshare type
1804     if (loc != NULL) {
1805       if ((loc->flags & KMP_IDENT_WORK_LOOP) != 0) {
1806         ompt_work_type = ompt_work_loop;
1807       } else if ((loc->flags & KMP_IDENT_WORK_SECTIONS) != 0) {
1808         ompt_work_type = ompt_work_sections;
1809       } else if ((loc->flags & KMP_IDENT_WORK_DISTRIBUTE) != 0) {
1810         ompt_work_type = ompt_work_distribute;
1811       } else {
1812         // use default set above.
1813         // a warning about this case is provided in __kmpc_for_static_init
1814       }
1815       KMP_DEBUG_ASSERT(ompt_work_type);
1816     }
1817     ompt_callbacks.ompt_callback(ompt_callback_work)(
1818         ompt_work_type, ompt_scope_end, &(team_info->parallel_data),
1819         &(task_info->task_data), 0, OMPT_GET_RETURN_ADDRESS(0));
1820   }
1821 #endif
1822   if (__kmp_env_consistency_check)
1823     __kmp_pop_workshare(global_tid, ct_pdo, loc);
1824 }
1825 
1826 // User routines which take C-style arguments (call by value)
1827 // different from the Fortran equivalent routines
1828 
1829 void ompc_set_num_threads(int arg) {
1830   // !!!!! TODO: check the per-task binding
1831   __kmp_set_num_threads(arg, __kmp_entry_gtid());
1832 }
1833 
1834 void ompc_set_dynamic(int flag) {
1835   kmp_info_t *thread;
1836 
1837   /* For the thread-private implementation of the internal controls */
1838   thread = __kmp_entry_thread();
1839 
1840   __kmp_save_internal_controls(thread);
1841 
1842   set__dynamic(thread, flag ? TRUE : FALSE);
1843 }
1844 
1845 void ompc_set_nested(int flag) {
1846   kmp_info_t *thread;
1847 
1848   /* For the thread-private internal controls implementation */
1849   thread = __kmp_entry_thread();
1850 
1851   __kmp_save_internal_controls(thread);
1852 
1853   set__max_active_levels(thread, flag ? __kmp_dflt_max_active_levels : 1);
1854 }
1855 
1856 void ompc_set_max_active_levels(int max_active_levels) {
1857   /* TO DO */
1858   /* we want per-task implementation of this internal control */
1859 
1860   /* For the per-thread internal controls implementation */
1861   __kmp_set_max_active_levels(__kmp_entry_gtid(), max_active_levels);
1862 }
1863 
1864 void ompc_set_schedule(omp_sched_t kind, int modifier) {
1865   // !!!!! TODO: check the per-task binding
1866   __kmp_set_schedule(__kmp_entry_gtid(), (kmp_sched_t)kind, modifier);
1867 }
1868 
1869 int ompc_get_ancestor_thread_num(int level) {
1870   return __kmp_get_ancestor_thread_num(__kmp_entry_gtid(), level);
1871 }
1872 
1873 int ompc_get_team_size(int level) {
1874   return __kmp_get_team_size(__kmp_entry_gtid(), level);
1875 }
1876 
1877 /* OpenMP 5.0 Affinity Format API */
1878 
1879 void ompc_set_affinity_format(char const *format) {
1880   if (!__kmp_init_serial) {
1881     __kmp_serial_initialize();
1882   }
1883   __kmp_strncpy_truncate(__kmp_affinity_format, KMP_AFFINITY_FORMAT_SIZE,
1884                          format, KMP_STRLEN(format) + 1);
1885 }
1886 
1887 size_t ompc_get_affinity_format(char *buffer, size_t size) {
1888   size_t format_size;
1889   if (!__kmp_init_serial) {
1890     __kmp_serial_initialize();
1891   }
1892   format_size = KMP_STRLEN(__kmp_affinity_format);
1893   if (buffer && size) {
1894     __kmp_strncpy_truncate(buffer, size, __kmp_affinity_format,
1895                            format_size + 1);
1896   }
1897   return format_size;
1898 }
1899 
1900 void ompc_display_affinity(char const *format) {
1901   int gtid;
1902   if (!TCR_4(__kmp_init_middle)) {
1903     __kmp_middle_initialize();
1904   }
1905   gtid = __kmp_get_gtid();
1906   __kmp_aux_display_affinity(gtid, format);
1907 }
1908 
1909 size_t ompc_capture_affinity(char *buffer, size_t buf_size,
1910                              char const *format) {
1911   int gtid;
1912   size_t num_required;
1913   kmp_str_buf_t capture_buf;
1914   if (!TCR_4(__kmp_init_middle)) {
1915     __kmp_middle_initialize();
1916   }
1917   gtid = __kmp_get_gtid();
1918   __kmp_str_buf_init(&capture_buf);
1919   num_required = __kmp_aux_capture_affinity(gtid, format, &capture_buf);
1920   if (buffer && buf_size) {
1921     __kmp_strncpy_truncate(buffer, buf_size, capture_buf.str,
1922                            capture_buf.used + 1);
1923   }
1924   __kmp_str_buf_free(&capture_buf);
1925   return num_required;
1926 }
1927 
1928 void kmpc_set_stacksize(int arg) {
1929   // __kmp_aux_set_stacksize initializes the library if needed
1930   __kmp_aux_set_stacksize(arg);
1931 }
1932 
1933 void kmpc_set_stacksize_s(size_t arg) {
1934   // __kmp_aux_set_stacksize initializes the library if needed
1935   __kmp_aux_set_stacksize(arg);
1936 }
1937 
1938 void kmpc_set_blocktime(int arg) {
1939   int gtid, tid;
1940   kmp_info_t *thread;
1941 
1942   gtid = __kmp_entry_gtid();
1943   tid = __kmp_tid_from_gtid(gtid);
1944   thread = __kmp_thread_from_gtid(gtid);
1945 
1946   __kmp_aux_set_blocktime(arg, thread, tid);
1947 }
1948 
1949 void kmpc_set_library(int arg) {
1950   // __kmp_user_set_library initializes the library if needed
1951   __kmp_user_set_library((enum library_type)arg);
1952 }
1953 
1954 void kmpc_set_defaults(char const *str) {
1955   // __kmp_aux_set_defaults initializes the library if needed
1956   __kmp_aux_set_defaults(str, KMP_STRLEN(str));
1957 }
1958 
1959 void kmpc_set_disp_num_buffers(int arg) {
1960   // ignore after initialization because some teams have already
1961   // allocated dispatch buffers
1962   if (__kmp_init_serial == 0 && arg > 0)
1963     __kmp_dispatch_num_buffers = arg;
1964 }
1965 
1966 int kmpc_set_affinity_mask_proc(int proc, void **mask) {
1967 #if defined(KMP_STUB) || !KMP_AFFINITY_SUPPORTED
1968   return -1;
1969 #else
1970   if (!TCR_4(__kmp_init_middle)) {
1971     __kmp_middle_initialize();
1972   }
1973   return __kmp_aux_set_affinity_mask_proc(proc, mask);
1974 #endif
1975 }
1976 
1977 int kmpc_unset_affinity_mask_proc(int proc, void **mask) {
1978 #if defined(KMP_STUB) || !KMP_AFFINITY_SUPPORTED
1979   return -1;
1980 #else
1981   if (!TCR_4(__kmp_init_middle)) {
1982     __kmp_middle_initialize();
1983   }
1984   return __kmp_aux_unset_affinity_mask_proc(proc, mask);
1985 #endif
1986 }
1987 
1988 int kmpc_get_affinity_mask_proc(int proc, void **mask) {
1989 #if defined(KMP_STUB) || !KMP_AFFINITY_SUPPORTED
1990   return -1;
1991 #else
1992   if (!TCR_4(__kmp_init_middle)) {
1993     __kmp_middle_initialize();
1994   }
1995   return __kmp_aux_get_affinity_mask_proc(proc, mask);
1996 #endif
1997 }
1998 
1999 /* -------------------------------------------------------------------------- */
2000 /*!
2001 @ingroup THREADPRIVATE
2002 @param loc       source location information
2003 @param gtid      global thread number
2004 @param cpy_size  size of the cpy_data buffer
2005 @param cpy_data  pointer to data to be copied
2006 @param cpy_func  helper function to call for copying data
2007 @param didit     flag variable: 1=single thread; 0=not single thread
2008 
2009 __kmpc_copyprivate implements the interface for the private data broadcast
2010 needed for the copyprivate clause associated with a single region in an
2011 OpenMP<sup>*</sup> program (both C and Fortran).
2012 All threads participating in the parallel region call this routine.
2013 One of the threads (called the single thread) should have the <tt>didit</tt>
2014 variable set to 1 and all other threads should have that variable set to 0.
2015 All threads pass a pointer to a data buffer (cpy_data) that they have built.
2016 
2017 The OpenMP specification forbids the use of nowait on the single region when a
2018 copyprivate clause is present. However, @ref __kmpc_copyprivate implements a
2019 barrier internally to avoid race conditions, so the code generation for the
2020 single region should avoid generating a barrier after the call to @ref
2021 __kmpc_copyprivate.
2022 
2023 The <tt>gtid</tt> parameter is the global thread id for the current thread.
2024 The <tt>loc</tt> parameter is a pointer to source location information.
2025 
2026 Internal implementation: The single thread will first copy its descriptor
2027 address (cpy_data) to a team-private location, then the other threads will each
2028 call the function pointed to by the parameter cpy_func, which carries out the
2029 copy by copying the data using the cpy_data buffer.
2030 
2031 The cpy_func routine used for the copy and the contents of the data area defined
2032 by cpy_data and cpy_size may be built in any fashion that will allow the copy
2033 to be done. For instance, the cpy_data buffer can hold the actual data to be
2034 copied or it may hold a list of pointers to the data. The cpy_func routine must
2035 interpret the cpy_data buffer appropriately.
2036 
2037 The interface to cpy_func is as follows:
2038 @code
2039 void cpy_func( void *destination, void *source )
2040 @endcode
2041 where void *destination is the cpy_data pointer for the thread being copied to
2042 and void *source is the cpy_data pointer for the thread being copied from.
2043 */
2044 void __kmpc_copyprivate(ident_t *loc, kmp_int32 gtid, size_t cpy_size,
2045                         void *cpy_data, void (*cpy_func)(void *, void *),
2046                         kmp_int32 didit) {
2047   void **data_ptr;
2048   KC_TRACE(10, ("__kmpc_copyprivate: called T#%d\n", gtid));
2049   __kmp_assert_valid_gtid(gtid);
2050 
2051   KMP_MB();
2052 
2053   data_ptr = &__kmp_team_from_gtid(gtid)->t.t_copypriv_data;
2054 
2055   if (__kmp_env_consistency_check) {
2056     if (loc == 0) {
2057       KMP_WARNING(ConstructIdentInvalid);
2058     }
2059   }
2060 
2061   // ToDo: Optimize the following two barriers into some kind of split barrier
2062 
2063   if (didit)
2064     *data_ptr = cpy_data;
2065 
2066 #if OMPT_SUPPORT
2067   ompt_frame_t *ompt_frame;
2068   if (ompt_enabled.enabled) {
2069     __ompt_get_task_info_internal(0, NULL, NULL, &ompt_frame, NULL, NULL);
2070     if (ompt_frame->enter_frame.ptr == NULL)
2071       ompt_frame->enter_frame.ptr = OMPT_GET_FRAME_ADDRESS(0);
2072     OMPT_STORE_RETURN_ADDRESS(gtid);
2073   }
2074 #endif
2075 /* This barrier is not a barrier region boundary */
2076 #if USE_ITT_NOTIFY
2077   __kmp_threads[gtid]->th.th_ident = loc;
2078 #endif
2079   __kmp_barrier(bs_plain_barrier, gtid, FALSE, 0, NULL, NULL);
2080 
2081   if (!didit)
2082     (*cpy_func)(cpy_data, *data_ptr);
2083 
2084 // Consider next barrier a user-visible barrier for barrier region boundaries
2085 // Nesting checks are already handled by the single construct checks
2086 
2087 #if OMPT_SUPPORT
2088   if (ompt_enabled.enabled) {
2089     OMPT_STORE_RETURN_ADDRESS(gtid);
2090   }
2091 #endif
2092 #if USE_ITT_NOTIFY
2093   __kmp_threads[gtid]->th.th_ident = loc; // TODO: check if it is needed (e.g.
2094 // tasks can overwrite the location)
2095 #endif
2096   __kmp_barrier(bs_plain_barrier, gtid, FALSE, 0, NULL, NULL);
2097 #if OMPT_SUPPORT && OMPT_OPTIONAL
2098   if (ompt_enabled.enabled) {
2099     ompt_frame->enter_frame = ompt_data_none;
2100   }
2101 #endif
2102 }
2103 
2104 /* -------------------------------------------------------------------------- */
2105 
2106 #define INIT_LOCK __kmp_init_user_lock_with_checks
2107 #define INIT_NESTED_LOCK __kmp_init_nested_user_lock_with_checks
2108 #define ACQUIRE_LOCK __kmp_acquire_user_lock_with_checks
2109 #define ACQUIRE_LOCK_TIMED __kmp_acquire_user_lock_with_checks_timed
2110 #define ACQUIRE_NESTED_LOCK __kmp_acquire_nested_user_lock_with_checks
2111 #define ACQUIRE_NESTED_LOCK_TIMED                                              \
2112   __kmp_acquire_nested_user_lock_with_checks_timed
2113 #define RELEASE_LOCK __kmp_release_user_lock_with_checks
2114 #define RELEASE_NESTED_LOCK __kmp_release_nested_user_lock_with_checks
2115 #define TEST_LOCK __kmp_test_user_lock_with_checks
2116 #define TEST_NESTED_LOCK __kmp_test_nested_user_lock_with_checks
2117 #define DESTROY_LOCK __kmp_destroy_user_lock_with_checks
2118 #define DESTROY_NESTED_LOCK __kmp_destroy_nested_user_lock_with_checks
2119 
2120 // TODO: Make check abort messages use location info & pass it into
2121 // with_checks routines
2122 
2123 #if KMP_USE_DYNAMIC_LOCK
2124 
2125 // internal lock initializer
2126 static __forceinline void __kmp_init_lock_with_hint(ident_t *loc, void **lock,
2127                                                     kmp_dyna_lockseq_t seq) {
2128   if (KMP_IS_D_LOCK(seq)) {
2129     KMP_INIT_D_LOCK(lock, seq);
2130 #if USE_ITT_BUILD
2131     __kmp_itt_lock_creating((kmp_user_lock_p)lock, NULL);
2132 #endif
2133   } else {
2134     KMP_INIT_I_LOCK(lock, seq);
2135 #if USE_ITT_BUILD
2136     kmp_indirect_lock_t *ilk = KMP_LOOKUP_I_LOCK(lock);
2137     __kmp_itt_lock_creating(ilk->lock, loc);
2138 #endif
2139   }
2140 }
2141 
2142 // internal nest lock initializer
2143 static __forceinline void
2144 __kmp_init_nest_lock_with_hint(ident_t *loc, void **lock,
2145                                kmp_dyna_lockseq_t seq) {
2146 #if KMP_USE_TSX
2147   // Don't have nested lock implementation for speculative locks
2148   if (seq == lockseq_hle || seq == lockseq_rtm || seq == lockseq_adaptive)
2149     seq = __kmp_user_lock_seq;
2150 #endif
2151   switch (seq) {
2152   case lockseq_tas:
2153     seq = lockseq_nested_tas;
2154     break;
2155 #if KMP_USE_FUTEX
2156   case lockseq_futex:
2157     seq = lockseq_nested_futex;
2158     break;
2159 #endif
2160   case lockseq_ticket:
2161     seq = lockseq_nested_ticket;
2162     break;
2163   case lockseq_queuing:
2164     seq = lockseq_nested_queuing;
2165     break;
2166   case lockseq_drdpa:
2167     seq = lockseq_nested_drdpa;
2168     break;
2169   default:
2170     seq = lockseq_nested_queuing;
2171   }
2172   KMP_INIT_I_LOCK(lock, seq);
2173 #if USE_ITT_BUILD
2174   kmp_indirect_lock_t *ilk = KMP_LOOKUP_I_LOCK(lock);
2175   __kmp_itt_lock_creating(ilk->lock, loc);
2176 #endif
2177 }
2178 
2179 /* initialize the lock with a hint */
2180 void __kmpc_init_lock_with_hint(ident_t *loc, kmp_int32 gtid, void **user_lock,
2181                                 uintptr_t hint) {
2182   KMP_DEBUG_ASSERT(__kmp_init_serial);
2183   if (__kmp_env_consistency_check && user_lock == NULL) {
2184     KMP_FATAL(LockIsUninitialized, "omp_init_lock_with_hint");
2185   }
2186 
2187   __kmp_init_lock_with_hint(loc, user_lock, __kmp_map_hint_to_lock(hint));
2188 
2189 #if OMPT_SUPPORT && OMPT_OPTIONAL
2190   // This is the case, if called from omp_init_lock_with_hint:
2191   void *codeptr = OMPT_LOAD_RETURN_ADDRESS(gtid);
2192   if (!codeptr)
2193     codeptr = OMPT_GET_RETURN_ADDRESS(0);
2194   if (ompt_enabled.ompt_callback_lock_init) {
2195     ompt_callbacks.ompt_callback(ompt_callback_lock_init)(
2196         ompt_mutex_lock, (omp_lock_hint_t)hint,
2197         __ompt_get_mutex_impl_type(user_lock),
2198         (ompt_wait_id_t)(uintptr_t)user_lock, codeptr);
2199   }
2200 #endif
2201 }
2202 
2203 /* initialize the lock with a hint */
2204 void __kmpc_init_nest_lock_with_hint(ident_t *loc, kmp_int32 gtid,
2205                                      void **user_lock, uintptr_t hint) {
2206   KMP_DEBUG_ASSERT(__kmp_init_serial);
2207   if (__kmp_env_consistency_check && user_lock == NULL) {
2208     KMP_FATAL(LockIsUninitialized, "omp_init_nest_lock_with_hint");
2209   }
2210 
2211   __kmp_init_nest_lock_with_hint(loc, user_lock, __kmp_map_hint_to_lock(hint));
2212 
2213 #if OMPT_SUPPORT && OMPT_OPTIONAL
2214   // This is the case, if called from omp_init_lock_with_hint:
2215   void *codeptr = OMPT_LOAD_RETURN_ADDRESS(gtid);
2216   if (!codeptr)
2217     codeptr = OMPT_GET_RETURN_ADDRESS(0);
2218   if (ompt_enabled.ompt_callback_lock_init) {
2219     ompt_callbacks.ompt_callback(ompt_callback_lock_init)(
2220         ompt_mutex_nest_lock, (omp_lock_hint_t)hint,
2221         __ompt_get_mutex_impl_type(user_lock),
2222         (ompt_wait_id_t)(uintptr_t)user_lock, codeptr);
2223   }
2224 #endif
2225 }
2226 
2227 #endif // KMP_USE_DYNAMIC_LOCK
2228 
2229 /* initialize the lock */
2230 void __kmpc_init_lock(ident_t *loc, kmp_int32 gtid, void **user_lock) {
2231 #if KMP_USE_DYNAMIC_LOCK
2232 
2233   KMP_DEBUG_ASSERT(__kmp_init_serial);
2234   if (__kmp_env_consistency_check && user_lock == NULL) {
2235     KMP_FATAL(LockIsUninitialized, "omp_init_lock");
2236   }
2237   __kmp_init_lock_with_hint(loc, user_lock, __kmp_user_lock_seq);
2238 
2239 #if OMPT_SUPPORT && OMPT_OPTIONAL
2240   // This is the case, if called from omp_init_lock_with_hint:
2241   void *codeptr = OMPT_LOAD_RETURN_ADDRESS(gtid);
2242   if (!codeptr)
2243     codeptr = OMPT_GET_RETURN_ADDRESS(0);
2244   if (ompt_enabled.ompt_callback_lock_init) {
2245     ompt_callbacks.ompt_callback(ompt_callback_lock_init)(
2246         ompt_mutex_lock, omp_lock_hint_none,
2247         __ompt_get_mutex_impl_type(user_lock),
2248         (ompt_wait_id_t)(uintptr_t)user_lock, codeptr);
2249   }
2250 #endif
2251 
2252 #else // KMP_USE_DYNAMIC_LOCK
2253 
2254   static char const *const func = "omp_init_lock";
2255   kmp_user_lock_p lck;
2256   KMP_DEBUG_ASSERT(__kmp_init_serial);
2257 
2258   if (__kmp_env_consistency_check) {
2259     if (user_lock == NULL) {
2260       KMP_FATAL(LockIsUninitialized, func);
2261     }
2262   }
2263 
2264   KMP_CHECK_USER_LOCK_INIT();
2265 
2266   if ((__kmp_user_lock_kind == lk_tas) &&
2267       (sizeof(lck->tas.lk.poll) <= OMP_LOCK_T_SIZE)) {
2268     lck = (kmp_user_lock_p)user_lock;
2269   }
2270 #if KMP_USE_FUTEX
2271   else if ((__kmp_user_lock_kind == lk_futex) &&
2272            (sizeof(lck->futex.lk.poll) <= OMP_LOCK_T_SIZE)) {
2273     lck = (kmp_user_lock_p)user_lock;
2274   }
2275 #endif
2276   else {
2277     lck = __kmp_user_lock_allocate(user_lock, gtid, 0);
2278   }
2279   INIT_LOCK(lck);
2280   __kmp_set_user_lock_location(lck, loc);
2281 
2282 #if OMPT_SUPPORT && OMPT_OPTIONAL
2283   // This is the case, if called from omp_init_lock_with_hint:
2284   void *codeptr = OMPT_LOAD_RETURN_ADDRESS(gtid);
2285   if (!codeptr)
2286     codeptr = OMPT_GET_RETURN_ADDRESS(0);
2287   if (ompt_enabled.ompt_callback_lock_init) {
2288     ompt_callbacks.ompt_callback(ompt_callback_lock_init)(
2289         ompt_mutex_lock, omp_lock_hint_none, __ompt_get_mutex_impl_type(),
2290         (ompt_wait_id_t)(uintptr_t)user_lock, codeptr);
2291   }
2292 #endif
2293 
2294 #if USE_ITT_BUILD
2295   __kmp_itt_lock_creating(lck);
2296 #endif /* USE_ITT_BUILD */
2297 
2298 #endif // KMP_USE_DYNAMIC_LOCK
2299 } // __kmpc_init_lock
2300 
2301 /* initialize the lock */
2302 void __kmpc_init_nest_lock(ident_t *loc, kmp_int32 gtid, void **user_lock) {
2303 #if KMP_USE_DYNAMIC_LOCK
2304 
2305   KMP_DEBUG_ASSERT(__kmp_init_serial);
2306   if (__kmp_env_consistency_check && user_lock == NULL) {
2307     KMP_FATAL(LockIsUninitialized, "omp_init_nest_lock");
2308   }
2309   __kmp_init_nest_lock_with_hint(loc, user_lock, __kmp_user_lock_seq);
2310 
2311 #if OMPT_SUPPORT && OMPT_OPTIONAL
2312   // This is the case, if called from omp_init_lock_with_hint:
2313   void *codeptr = OMPT_LOAD_RETURN_ADDRESS(gtid);
2314   if (!codeptr)
2315     codeptr = OMPT_GET_RETURN_ADDRESS(0);
2316   if (ompt_enabled.ompt_callback_lock_init) {
2317     ompt_callbacks.ompt_callback(ompt_callback_lock_init)(
2318         ompt_mutex_nest_lock, omp_lock_hint_none,
2319         __ompt_get_mutex_impl_type(user_lock),
2320         (ompt_wait_id_t)(uintptr_t)user_lock, codeptr);
2321   }
2322 #endif
2323 
2324 #else // KMP_USE_DYNAMIC_LOCK
2325 
2326   static char const *const func = "omp_init_nest_lock";
2327   kmp_user_lock_p lck;
2328   KMP_DEBUG_ASSERT(__kmp_init_serial);
2329 
2330   if (__kmp_env_consistency_check) {
2331     if (user_lock == NULL) {
2332       KMP_FATAL(LockIsUninitialized, func);
2333     }
2334   }
2335 
2336   KMP_CHECK_USER_LOCK_INIT();
2337 
2338   if ((__kmp_user_lock_kind == lk_tas) &&
2339       (sizeof(lck->tas.lk.poll) + sizeof(lck->tas.lk.depth_locked) <=
2340        OMP_NEST_LOCK_T_SIZE)) {
2341     lck = (kmp_user_lock_p)user_lock;
2342   }
2343 #if KMP_USE_FUTEX
2344   else if ((__kmp_user_lock_kind == lk_futex) &&
2345            (sizeof(lck->futex.lk.poll) + sizeof(lck->futex.lk.depth_locked) <=
2346             OMP_NEST_LOCK_T_SIZE)) {
2347     lck = (kmp_user_lock_p)user_lock;
2348   }
2349 #endif
2350   else {
2351     lck = __kmp_user_lock_allocate(user_lock, gtid, 0);
2352   }
2353 
2354   INIT_NESTED_LOCK(lck);
2355   __kmp_set_user_lock_location(lck, loc);
2356 
2357 #if OMPT_SUPPORT && OMPT_OPTIONAL
2358   // This is the case, if called from omp_init_lock_with_hint:
2359   void *codeptr = OMPT_LOAD_RETURN_ADDRESS(gtid);
2360   if (!codeptr)
2361     codeptr = OMPT_GET_RETURN_ADDRESS(0);
2362   if (ompt_enabled.ompt_callback_lock_init) {
2363     ompt_callbacks.ompt_callback(ompt_callback_lock_init)(
2364         ompt_mutex_nest_lock, omp_lock_hint_none, __ompt_get_mutex_impl_type(),
2365         (ompt_wait_id_t)(uintptr_t)user_lock, codeptr);
2366   }
2367 #endif
2368 
2369 #if USE_ITT_BUILD
2370   __kmp_itt_lock_creating(lck);
2371 #endif /* USE_ITT_BUILD */
2372 
2373 #endif // KMP_USE_DYNAMIC_LOCK
2374 } // __kmpc_init_nest_lock
2375 
2376 void __kmpc_destroy_lock(ident_t *loc, kmp_int32 gtid, void **user_lock) {
2377 #if KMP_USE_DYNAMIC_LOCK
2378 
2379 #if USE_ITT_BUILD
2380   kmp_user_lock_p lck;
2381   if (KMP_EXTRACT_D_TAG(user_lock) == 0) {
2382     lck = ((kmp_indirect_lock_t *)KMP_LOOKUP_I_LOCK(user_lock))->lock;
2383   } else {
2384     lck = (kmp_user_lock_p)user_lock;
2385   }
2386   __kmp_itt_lock_destroyed(lck);
2387 #endif
2388 #if OMPT_SUPPORT && OMPT_OPTIONAL
2389   // This is the case, if called from omp_init_lock_with_hint:
2390   void *codeptr = OMPT_LOAD_RETURN_ADDRESS(gtid);
2391   if (!codeptr)
2392     codeptr = OMPT_GET_RETURN_ADDRESS(0);
2393   if (ompt_enabled.ompt_callback_lock_destroy) {
2394     kmp_user_lock_p lck;
2395     if (KMP_EXTRACT_D_TAG(user_lock) == 0) {
2396       lck = ((kmp_indirect_lock_t *)KMP_LOOKUP_I_LOCK(user_lock))->lock;
2397     } else {
2398       lck = (kmp_user_lock_p)user_lock;
2399     }
2400     ompt_callbacks.ompt_callback(ompt_callback_lock_destroy)(
2401         ompt_mutex_lock, (ompt_wait_id_t)(uintptr_t)user_lock, codeptr);
2402   }
2403 #endif
2404   KMP_D_LOCK_FUNC(user_lock, destroy)((kmp_dyna_lock_t *)user_lock);
2405 #else
2406   kmp_user_lock_p lck;
2407 
2408   if ((__kmp_user_lock_kind == lk_tas) &&
2409       (sizeof(lck->tas.lk.poll) <= OMP_LOCK_T_SIZE)) {
2410     lck = (kmp_user_lock_p)user_lock;
2411   }
2412 #if KMP_USE_FUTEX
2413   else if ((__kmp_user_lock_kind == lk_futex) &&
2414            (sizeof(lck->futex.lk.poll) <= OMP_LOCK_T_SIZE)) {
2415     lck = (kmp_user_lock_p)user_lock;
2416   }
2417 #endif
2418   else {
2419     lck = __kmp_lookup_user_lock(user_lock, "omp_destroy_lock");
2420   }
2421 
2422 #if OMPT_SUPPORT && OMPT_OPTIONAL
2423   // This is the case, if called from omp_init_lock_with_hint:
2424   void *codeptr = OMPT_LOAD_RETURN_ADDRESS(gtid);
2425   if (!codeptr)
2426     codeptr = OMPT_GET_RETURN_ADDRESS(0);
2427   if (ompt_enabled.ompt_callback_lock_destroy) {
2428     ompt_callbacks.ompt_callback(ompt_callback_lock_destroy)(
2429         ompt_mutex_lock, (ompt_wait_id_t)(uintptr_t)user_lock, codeptr);
2430   }
2431 #endif
2432 
2433 #if USE_ITT_BUILD
2434   __kmp_itt_lock_destroyed(lck);
2435 #endif /* USE_ITT_BUILD */
2436   DESTROY_LOCK(lck);
2437 
2438   if ((__kmp_user_lock_kind == lk_tas) &&
2439       (sizeof(lck->tas.lk.poll) <= OMP_LOCK_T_SIZE)) {
2440     ;
2441   }
2442 #if KMP_USE_FUTEX
2443   else if ((__kmp_user_lock_kind == lk_futex) &&
2444            (sizeof(lck->futex.lk.poll) <= OMP_LOCK_T_SIZE)) {
2445     ;
2446   }
2447 #endif
2448   else {
2449     __kmp_user_lock_free(user_lock, gtid, lck);
2450   }
2451 #endif // KMP_USE_DYNAMIC_LOCK
2452 } // __kmpc_destroy_lock
2453 
2454 /* destroy the lock */
2455 void __kmpc_destroy_nest_lock(ident_t *loc, kmp_int32 gtid, void **user_lock) {
2456 #if KMP_USE_DYNAMIC_LOCK
2457 
2458 #if USE_ITT_BUILD
2459   kmp_indirect_lock_t *ilk = KMP_LOOKUP_I_LOCK(user_lock);
2460   __kmp_itt_lock_destroyed(ilk->lock);
2461 #endif
2462 #if OMPT_SUPPORT && OMPT_OPTIONAL
2463   // This is the case, if called from omp_init_lock_with_hint:
2464   void *codeptr = OMPT_LOAD_RETURN_ADDRESS(gtid);
2465   if (!codeptr)
2466     codeptr = OMPT_GET_RETURN_ADDRESS(0);
2467   if (ompt_enabled.ompt_callback_lock_destroy) {
2468     ompt_callbacks.ompt_callback(ompt_callback_lock_destroy)(
2469         ompt_mutex_nest_lock, (ompt_wait_id_t)(uintptr_t)user_lock, codeptr);
2470   }
2471 #endif
2472   KMP_D_LOCK_FUNC(user_lock, destroy)((kmp_dyna_lock_t *)user_lock);
2473 
2474 #else // KMP_USE_DYNAMIC_LOCK
2475 
2476   kmp_user_lock_p lck;
2477 
2478   if ((__kmp_user_lock_kind == lk_tas) &&
2479       (sizeof(lck->tas.lk.poll) + sizeof(lck->tas.lk.depth_locked) <=
2480        OMP_NEST_LOCK_T_SIZE)) {
2481     lck = (kmp_user_lock_p)user_lock;
2482   }
2483 #if KMP_USE_FUTEX
2484   else if ((__kmp_user_lock_kind == lk_futex) &&
2485            (sizeof(lck->futex.lk.poll) + sizeof(lck->futex.lk.depth_locked) <=
2486             OMP_NEST_LOCK_T_SIZE)) {
2487     lck = (kmp_user_lock_p)user_lock;
2488   }
2489 #endif
2490   else {
2491     lck = __kmp_lookup_user_lock(user_lock, "omp_destroy_nest_lock");
2492   }
2493 
2494 #if OMPT_SUPPORT && OMPT_OPTIONAL
2495   // This is the case, if called from omp_init_lock_with_hint:
2496   void *codeptr = OMPT_LOAD_RETURN_ADDRESS(gtid);
2497   if (!codeptr)
2498     codeptr = OMPT_GET_RETURN_ADDRESS(0);
2499   if (ompt_enabled.ompt_callback_lock_destroy) {
2500     ompt_callbacks.ompt_callback(ompt_callback_lock_destroy)(
2501         ompt_mutex_nest_lock, (ompt_wait_id_t)(uintptr_t)user_lock, codeptr);
2502   }
2503 #endif
2504 
2505 #if USE_ITT_BUILD
2506   __kmp_itt_lock_destroyed(lck);
2507 #endif /* USE_ITT_BUILD */
2508 
2509   DESTROY_NESTED_LOCK(lck);
2510 
2511   if ((__kmp_user_lock_kind == lk_tas) &&
2512       (sizeof(lck->tas.lk.poll) + sizeof(lck->tas.lk.depth_locked) <=
2513        OMP_NEST_LOCK_T_SIZE)) {
2514     ;
2515   }
2516 #if KMP_USE_FUTEX
2517   else if ((__kmp_user_lock_kind == lk_futex) &&
2518            (sizeof(lck->futex.lk.poll) + sizeof(lck->futex.lk.depth_locked) <=
2519             OMP_NEST_LOCK_T_SIZE)) {
2520     ;
2521   }
2522 #endif
2523   else {
2524     __kmp_user_lock_free(user_lock, gtid, lck);
2525   }
2526 #endif // KMP_USE_DYNAMIC_LOCK
2527 } // __kmpc_destroy_nest_lock
2528 
2529 void __kmpc_set_lock(ident_t *loc, kmp_int32 gtid, void **user_lock) {
2530   KMP_COUNT_BLOCK(OMP_set_lock);
2531 #if KMP_USE_DYNAMIC_LOCK
2532   int tag = KMP_EXTRACT_D_TAG(user_lock);
2533 #if USE_ITT_BUILD
2534   __kmp_itt_lock_acquiring(
2535       (kmp_user_lock_p)
2536           user_lock); // itt function will get to the right lock object.
2537 #endif
2538 #if OMPT_SUPPORT && OMPT_OPTIONAL
2539   // This is the case, if called from omp_init_lock_with_hint:
2540   void *codeptr = OMPT_LOAD_RETURN_ADDRESS(gtid);
2541   if (!codeptr)
2542     codeptr = OMPT_GET_RETURN_ADDRESS(0);
2543   if (ompt_enabled.ompt_callback_mutex_acquire) {
2544     ompt_callbacks.ompt_callback(ompt_callback_mutex_acquire)(
2545         ompt_mutex_lock, omp_lock_hint_none,
2546         __ompt_get_mutex_impl_type(user_lock),
2547         (ompt_wait_id_t)(uintptr_t)user_lock, codeptr);
2548   }
2549 #endif
2550 #if KMP_USE_INLINED_TAS
2551   if (tag == locktag_tas && !__kmp_env_consistency_check) {
2552     KMP_ACQUIRE_TAS_LOCK(user_lock, gtid);
2553   } else
2554 #elif KMP_USE_INLINED_FUTEX
2555   if (tag == locktag_futex && !__kmp_env_consistency_check) {
2556     KMP_ACQUIRE_FUTEX_LOCK(user_lock, gtid);
2557   } else
2558 #endif
2559   {
2560     __kmp_direct_set[tag]((kmp_dyna_lock_t *)user_lock, gtid);
2561   }
2562 #if USE_ITT_BUILD
2563   __kmp_itt_lock_acquired((kmp_user_lock_p)user_lock);
2564 #endif
2565 #if OMPT_SUPPORT && OMPT_OPTIONAL
2566   if (ompt_enabled.ompt_callback_mutex_acquired) {
2567     ompt_callbacks.ompt_callback(ompt_callback_mutex_acquired)(
2568         ompt_mutex_lock, (ompt_wait_id_t)(uintptr_t)user_lock, codeptr);
2569   }
2570 #endif
2571 
2572 #else // KMP_USE_DYNAMIC_LOCK
2573 
2574   kmp_user_lock_p lck;
2575 
2576   if ((__kmp_user_lock_kind == lk_tas) &&
2577       (sizeof(lck->tas.lk.poll) <= OMP_LOCK_T_SIZE)) {
2578     lck = (kmp_user_lock_p)user_lock;
2579   }
2580 #if KMP_USE_FUTEX
2581   else if ((__kmp_user_lock_kind == lk_futex) &&
2582            (sizeof(lck->futex.lk.poll) <= OMP_LOCK_T_SIZE)) {
2583     lck = (kmp_user_lock_p)user_lock;
2584   }
2585 #endif
2586   else {
2587     lck = __kmp_lookup_user_lock(user_lock, "omp_set_lock");
2588   }
2589 
2590 #if USE_ITT_BUILD
2591   __kmp_itt_lock_acquiring(lck);
2592 #endif /* USE_ITT_BUILD */
2593 #if OMPT_SUPPORT && OMPT_OPTIONAL
2594   // This is the case, if called from omp_init_lock_with_hint:
2595   void *codeptr = OMPT_LOAD_RETURN_ADDRESS(gtid);
2596   if (!codeptr)
2597     codeptr = OMPT_GET_RETURN_ADDRESS(0);
2598   if (ompt_enabled.ompt_callback_mutex_acquire) {
2599     ompt_callbacks.ompt_callback(ompt_callback_mutex_acquire)(
2600         ompt_mutex_lock, omp_lock_hint_none, __ompt_get_mutex_impl_type(),
2601         (ompt_wait_id_t)(uintptr_t)lck, codeptr);
2602   }
2603 #endif
2604 
2605   ACQUIRE_LOCK(lck, gtid);
2606 
2607 #if USE_ITT_BUILD
2608   __kmp_itt_lock_acquired(lck);
2609 #endif /* USE_ITT_BUILD */
2610 
2611 #if OMPT_SUPPORT && OMPT_OPTIONAL
2612   if (ompt_enabled.ompt_callback_mutex_acquired) {
2613     ompt_callbacks.ompt_callback(ompt_callback_mutex_acquired)(
2614         ompt_mutex_lock, (ompt_wait_id_t)(uintptr_t)lck, codeptr);
2615   }
2616 #endif
2617 
2618 #endif // KMP_USE_DYNAMIC_LOCK
2619 }
2620 
2621 void __kmpc_set_nest_lock(ident_t *loc, kmp_int32 gtid, void **user_lock) {
2622 #if KMP_USE_DYNAMIC_LOCK
2623 
2624 #if USE_ITT_BUILD
2625   __kmp_itt_lock_acquiring((kmp_user_lock_p)user_lock);
2626 #endif
2627 #if OMPT_SUPPORT && OMPT_OPTIONAL
2628   // This is the case, if called from omp_init_lock_with_hint:
2629   void *codeptr = OMPT_LOAD_RETURN_ADDRESS(gtid);
2630   if (!codeptr)
2631     codeptr = OMPT_GET_RETURN_ADDRESS(0);
2632   if (ompt_enabled.enabled) {
2633     if (ompt_enabled.ompt_callback_mutex_acquire) {
2634       ompt_callbacks.ompt_callback(ompt_callback_mutex_acquire)(
2635           ompt_mutex_nest_lock, omp_lock_hint_none,
2636           __ompt_get_mutex_impl_type(user_lock),
2637           (ompt_wait_id_t)(uintptr_t)user_lock, codeptr);
2638     }
2639   }
2640 #endif
2641   int acquire_status =
2642       KMP_D_LOCK_FUNC(user_lock, set)((kmp_dyna_lock_t *)user_lock, gtid);
2643   (void) acquire_status;
2644 #if USE_ITT_BUILD
2645   __kmp_itt_lock_acquired((kmp_user_lock_p)user_lock);
2646 #endif
2647 
2648 #if OMPT_SUPPORT && OMPT_OPTIONAL
2649   if (ompt_enabled.enabled) {
2650     if (acquire_status == KMP_LOCK_ACQUIRED_FIRST) {
2651       if (ompt_enabled.ompt_callback_mutex_acquired) {
2652         // lock_first
2653         ompt_callbacks.ompt_callback(ompt_callback_mutex_acquired)(
2654             ompt_mutex_nest_lock, (ompt_wait_id_t)(uintptr_t)user_lock,
2655             codeptr);
2656       }
2657     } else {
2658       if (ompt_enabled.ompt_callback_nest_lock) {
2659         // lock_next
2660         ompt_callbacks.ompt_callback(ompt_callback_nest_lock)(
2661             ompt_scope_begin, (ompt_wait_id_t)(uintptr_t)user_lock, codeptr);
2662       }
2663     }
2664   }
2665 #endif
2666 
2667 #else // KMP_USE_DYNAMIC_LOCK
2668   int acquire_status;
2669   kmp_user_lock_p lck;
2670 
2671   if ((__kmp_user_lock_kind == lk_tas) &&
2672       (sizeof(lck->tas.lk.poll) + sizeof(lck->tas.lk.depth_locked) <=
2673        OMP_NEST_LOCK_T_SIZE)) {
2674     lck = (kmp_user_lock_p)user_lock;
2675   }
2676 #if KMP_USE_FUTEX
2677   else if ((__kmp_user_lock_kind == lk_futex) &&
2678            (sizeof(lck->futex.lk.poll) + sizeof(lck->futex.lk.depth_locked) <=
2679             OMP_NEST_LOCK_T_SIZE)) {
2680     lck = (kmp_user_lock_p)user_lock;
2681   }
2682 #endif
2683   else {
2684     lck = __kmp_lookup_user_lock(user_lock, "omp_set_nest_lock");
2685   }
2686 
2687 #if USE_ITT_BUILD
2688   __kmp_itt_lock_acquiring(lck);
2689 #endif /* USE_ITT_BUILD */
2690 #if OMPT_SUPPORT && OMPT_OPTIONAL
2691   // This is the case, if called from omp_init_lock_with_hint:
2692   void *codeptr = OMPT_LOAD_RETURN_ADDRESS(gtid);
2693   if (!codeptr)
2694     codeptr = OMPT_GET_RETURN_ADDRESS(0);
2695   if (ompt_enabled.enabled) {
2696     if (ompt_enabled.ompt_callback_mutex_acquire) {
2697       ompt_callbacks.ompt_callback(ompt_callback_mutex_acquire)(
2698           ompt_mutex_nest_lock, omp_lock_hint_none,
2699           __ompt_get_mutex_impl_type(), (ompt_wait_id_t)(uintptr_t)lck,
2700           codeptr);
2701     }
2702   }
2703 #endif
2704 
2705   ACQUIRE_NESTED_LOCK(lck, gtid, &acquire_status);
2706 
2707 #if USE_ITT_BUILD
2708   __kmp_itt_lock_acquired(lck);
2709 #endif /* USE_ITT_BUILD */
2710 
2711 #if OMPT_SUPPORT && OMPT_OPTIONAL
2712   if (ompt_enabled.enabled) {
2713     if (acquire_status == KMP_LOCK_ACQUIRED_FIRST) {
2714       if (ompt_enabled.ompt_callback_mutex_acquired) {
2715         // lock_first
2716         ompt_callbacks.ompt_callback(ompt_callback_mutex_acquired)(
2717             ompt_mutex_nest_lock, (ompt_wait_id_t)(uintptr_t)lck, codeptr);
2718       }
2719     } else {
2720       if (ompt_enabled.ompt_callback_nest_lock) {
2721         // lock_next
2722         ompt_callbacks.ompt_callback(ompt_callback_nest_lock)(
2723             ompt_scope_begin, (ompt_wait_id_t)(uintptr_t)lck, codeptr);
2724       }
2725     }
2726   }
2727 #endif
2728 
2729 #endif // KMP_USE_DYNAMIC_LOCK
2730 }
2731 
2732 void __kmpc_unset_lock(ident_t *loc, kmp_int32 gtid, void **user_lock) {
2733 #if KMP_USE_DYNAMIC_LOCK
2734 
2735   int tag = KMP_EXTRACT_D_TAG(user_lock);
2736 #if USE_ITT_BUILD
2737   __kmp_itt_lock_releasing((kmp_user_lock_p)user_lock);
2738 #endif
2739 #if KMP_USE_INLINED_TAS
2740   if (tag == locktag_tas && !__kmp_env_consistency_check) {
2741     KMP_RELEASE_TAS_LOCK(user_lock, gtid);
2742   } else
2743 #elif KMP_USE_INLINED_FUTEX
2744   if (tag == locktag_futex && !__kmp_env_consistency_check) {
2745     KMP_RELEASE_FUTEX_LOCK(user_lock, gtid);
2746   } else
2747 #endif
2748   {
2749     __kmp_direct_unset[tag]((kmp_dyna_lock_t *)user_lock, gtid);
2750   }
2751 
2752 #if OMPT_SUPPORT && OMPT_OPTIONAL
2753   // This is the case, if called from omp_init_lock_with_hint:
2754   void *codeptr = OMPT_LOAD_RETURN_ADDRESS(gtid);
2755   if (!codeptr)
2756     codeptr = OMPT_GET_RETURN_ADDRESS(0);
2757   if (ompt_enabled.ompt_callback_mutex_released) {
2758     ompt_callbacks.ompt_callback(ompt_callback_mutex_released)(
2759         ompt_mutex_lock, (ompt_wait_id_t)(uintptr_t)user_lock, codeptr);
2760   }
2761 #endif
2762 
2763 #else // KMP_USE_DYNAMIC_LOCK
2764 
2765   kmp_user_lock_p lck;
2766 
2767   /* Can't use serial interval since not block structured */
2768   /* release the lock */
2769 
2770   if ((__kmp_user_lock_kind == lk_tas) &&
2771       (sizeof(lck->tas.lk.poll) <= OMP_LOCK_T_SIZE)) {
2772 #if KMP_OS_LINUX &&                                                            \
2773     (KMP_ARCH_X86 || KMP_ARCH_X86_64 || KMP_ARCH_ARM || KMP_ARCH_AARCH64)
2774 // "fast" path implemented to fix customer performance issue
2775 #if USE_ITT_BUILD
2776     __kmp_itt_lock_releasing((kmp_user_lock_p)user_lock);
2777 #endif /* USE_ITT_BUILD */
2778     TCW_4(((kmp_user_lock_p)user_lock)->tas.lk.poll, 0);
2779     KMP_MB();
2780 
2781 #if OMPT_SUPPORT && OMPT_OPTIONAL
2782     // This is the case, if called from omp_init_lock_with_hint:
2783     void *codeptr = OMPT_LOAD_RETURN_ADDRESS(gtid);
2784     if (!codeptr)
2785       codeptr = OMPT_GET_RETURN_ADDRESS(0);
2786     if (ompt_enabled.ompt_callback_mutex_released) {
2787       ompt_callbacks.ompt_callback(ompt_callback_mutex_released)(
2788           ompt_mutex_lock, (ompt_wait_id_t)(uintptr_t)lck, codeptr);
2789     }
2790 #endif
2791 
2792     return;
2793 #else
2794     lck = (kmp_user_lock_p)user_lock;
2795 #endif
2796   }
2797 #if KMP_USE_FUTEX
2798   else if ((__kmp_user_lock_kind == lk_futex) &&
2799            (sizeof(lck->futex.lk.poll) <= OMP_LOCK_T_SIZE)) {
2800     lck = (kmp_user_lock_p)user_lock;
2801   }
2802 #endif
2803   else {
2804     lck = __kmp_lookup_user_lock(user_lock, "omp_unset_lock");
2805   }
2806 
2807 #if USE_ITT_BUILD
2808   __kmp_itt_lock_releasing(lck);
2809 #endif /* USE_ITT_BUILD */
2810 
2811   RELEASE_LOCK(lck, gtid);
2812 
2813 #if OMPT_SUPPORT && OMPT_OPTIONAL
2814   // This is the case, if called from omp_init_lock_with_hint:
2815   void *codeptr = OMPT_LOAD_RETURN_ADDRESS(gtid);
2816   if (!codeptr)
2817     codeptr = OMPT_GET_RETURN_ADDRESS(0);
2818   if (ompt_enabled.ompt_callback_mutex_released) {
2819     ompt_callbacks.ompt_callback(ompt_callback_mutex_released)(
2820         ompt_mutex_lock, (ompt_wait_id_t)(uintptr_t)lck, codeptr);
2821   }
2822 #endif
2823 
2824 #endif // KMP_USE_DYNAMIC_LOCK
2825 }
2826 
2827 /* release the lock */
2828 void __kmpc_unset_nest_lock(ident_t *loc, kmp_int32 gtid, void **user_lock) {
2829 #if KMP_USE_DYNAMIC_LOCK
2830 
2831 #if USE_ITT_BUILD
2832   __kmp_itt_lock_releasing((kmp_user_lock_p)user_lock);
2833 #endif
2834   int release_status =
2835       KMP_D_LOCK_FUNC(user_lock, unset)((kmp_dyna_lock_t *)user_lock, gtid);
2836   (void) release_status;
2837 
2838 #if OMPT_SUPPORT && OMPT_OPTIONAL
2839   // This is the case, if called from omp_init_lock_with_hint:
2840   void *codeptr = OMPT_LOAD_RETURN_ADDRESS(gtid);
2841   if (!codeptr)
2842     codeptr = OMPT_GET_RETURN_ADDRESS(0);
2843   if (ompt_enabled.enabled) {
2844     if (release_status == KMP_LOCK_RELEASED) {
2845       if (ompt_enabled.ompt_callback_mutex_released) {
2846         // release_lock_last
2847         ompt_callbacks.ompt_callback(ompt_callback_mutex_released)(
2848             ompt_mutex_nest_lock, (ompt_wait_id_t)(uintptr_t)user_lock,
2849             codeptr);
2850       }
2851     } else if (ompt_enabled.ompt_callback_nest_lock) {
2852       // release_lock_prev
2853       ompt_callbacks.ompt_callback(ompt_callback_nest_lock)(
2854           ompt_scope_end, (ompt_wait_id_t)(uintptr_t)user_lock, codeptr);
2855     }
2856   }
2857 #endif
2858 
2859 #else // KMP_USE_DYNAMIC_LOCK
2860 
2861   kmp_user_lock_p lck;
2862 
2863   /* Can't use serial interval since not block structured */
2864 
2865   if ((__kmp_user_lock_kind == lk_tas) &&
2866       (sizeof(lck->tas.lk.poll) + sizeof(lck->tas.lk.depth_locked) <=
2867        OMP_NEST_LOCK_T_SIZE)) {
2868 #if KMP_OS_LINUX &&                                                            \
2869     (KMP_ARCH_X86 || KMP_ARCH_X86_64 || KMP_ARCH_ARM || KMP_ARCH_AARCH64)
2870     // "fast" path implemented to fix customer performance issue
2871     kmp_tas_lock_t *tl = (kmp_tas_lock_t *)user_lock;
2872 #if USE_ITT_BUILD
2873     __kmp_itt_lock_releasing((kmp_user_lock_p)user_lock);
2874 #endif /* USE_ITT_BUILD */
2875 
2876 #if OMPT_SUPPORT && OMPT_OPTIONAL
2877     int release_status = KMP_LOCK_STILL_HELD;
2878 #endif
2879 
2880     if (--(tl->lk.depth_locked) == 0) {
2881       TCW_4(tl->lk.poll, 0);
2882 #if OMPT_SUPPORT && OMPT_OPTIONAL
2883       release_status = KMP_LOCK_RELEASED;
2884 #endif
2885     }
2886     KMP_MB();
2887 
2888 #if OMPT_SUPPORT && OMPT_OPTIONAL
2889     // This is the case, if called from omp_init_lock_with_hint:
2890     void *codeptr = OMPT_LOAD_RETURN_ADDRESS(gtid);
2891     if (!codeptr)
2892       codeptr = OMPT_GET_RETURN_ADDRESS(0);
2893     if (ompt_enabled.enabled) {
2894       if (release_status == KMP_LOCK_RELEASED) {
2895         if (ompt_enabled.ompt_callback_mutex_released) {
2896           // release_lock_last
2897           ompt_callbacks.ompt_callback(ompt_callback_mutex_released)(
2898               ompt_mutex_nest_lock, (ompt_wait_id_t)(uintptr_t)lck, codeptr);
2899         }
2900       } else if (ompt_enabled.ompt_callback_nest_lock) {
2901         // release_lock_previous
2902         ompt_callbacks.ompt_callback(ompt_callback_nest_lock)(
2903             ompt_mutex_scope_end, (ompt_wait_id_t)(uintptr_t)lck, codeptr);
2904       }
2905     }
2906 #endif
2907 
2908     return;
2909 #else
2910     lck = (kmp_user_lock_p)user_lock;
2911 #endif
2912   }
2913 #if KMP_USE_FUTEX
2914   else if ((__kmp_user_lock_kind == lk_futex) &&
2915            (sizeof(lck->futex.lk.poll) + sizeof(lck->futex.lk.depth_locked) <=
2916             OMP_NEST_LOCK_T_SIZE)) {
2917     lck = (kmp_user_lock_p)user_lock;
2918   }
2919 #endif
2920   else {
2921     lck = __kmp_lookup_user_lock(user_lock, "omp_unset_nest_lock");
2922   }
2923 
2924 #if USE_ITT_BUILD
2925   __kmp_itt_lock_releasing(lck);
2926 #endif /* USE_ITT_BUILD */
2927 
2928   int release_status;
2929   release_status = RELEASE_NESTED_LOCK(lck, gtid);
2930 #if OMPT_SUPPORT && OMPT_OPTIONAL
2931   // This is the case, if called from omp_init_lock_with_hint:
2932   void *codeptr = OMPT_LOAD_RETURN_ADDRESS(gtid);
2933   if (!codeptr)
2934     codeptr = OMPT_GET_RETURN_ADDRESS(0);
2935   if (ompt_enabled.enabled) {
2936     if (release_status == KMP_LOCK_RELEASED) {
2937       if (ompt_enabled.ompt_callback_mutex_released) {
2938         // release_lock_last
2939         ompt_callbacks.ompt_callback(ompt_callback_mutex_released)(
2940             ompt_mutex_nest_lock, (ompt_wait_id_t)(uintptr_t)lck, codeptr);
2941       }
2942     } else if (ompt_enabled.ompt_callback_nest_lock) {
2943       // release_lock_previous
2944       ompt_callbacks.ompt_callback(ompt_callback_nest_lock)(
2945           ompt_mutex_scope_end, (ompt_wait_id_t)(uintptr_t)lck, codeptr);
2946     }
2947   }
2948 #endif
2949 
2950 #endif // KMP_USE_DYNAMIC_LOCK
2951 }
2952 
2953 /* try to acquire the lock */
2954 int __kmpc_test_lock(ident_t *loc, kmp_int32 gtid, void **user_lock) {
2955   KMP_COUNT_BLOCK(OMP_test_lock);
2956 
2957 #if KMP_USE_DYNAMIC_LOCK
2958   int rc;
2959   int tag = KMP_EXTRACT_D_TAG(user_lock);
2960 #if USE_ITT_BUILD
2961   __kmp_itt_lock_acquiring((kmp_user_lock_p)user_lock);
2962 #endif
2963 #if OMPT_SUPPORT && OMPT_OPTIONAL
2964   // This is the case, if called from omp_init_lock_with_hint:
2965   void *codeptr = OMPT_LOAD_RETURN_ADDRESS(gtid);
2966   if (!codeptr)
2967     codeptr = OMPT_GET_RETURN_ADDRESS(0);
2968   if (ompt_enabled.ompt_callback_mutex_acquire) {
2969     ompt_callbacks.ompt_callback(ompt_callback_mutex_acquire)(
2970         ompt_mutex_lock, omp_lock_hint_none,
2971         __ompt_get_mutex_impl_type(user_lock),
2972         (ompt_wait_id_t)(uintptr_t)user_lock, codeptr);
2973   }
2974 #endif
2975 #if KMP_USE_INLINED_TAS
2976   if (tag == locktag_tas && !__kmp_env_consistency_check) {
2977     KMP_TEST_TAS_LOCK(user_lock, gtid, rc);
2978   } else
2979 #elif KMP_USE_INLINED_FUTEX
2980   if (tag == locktag_futex && !__kmp_env_consistency_check) {
2981     KMP_TEST_FUTEX_LOCK(user_lock, gtid, rc);
2982   } else
2983 #endif
2984   {
2985     rc = __kmp_direct_test[tag]((kmp_dyna_lock_t *)user_lock, gtid);
2986   }
2987   if (rc) {
2988 #if USE_ITT_BUILD
2989     __kmp_itt_lock_acquired((kmp_user_lock_p)user_lock);
2990 #endif
2991 #if OMPT_SUPPORT && OMPT_OPTIONAL
2992     if (ompt_enabled.ompt_callback_mutex_acquired) {
2993       ompt_callbacks.ompt_callback(ompt_callback_mutex_acquired)(
2994           ompt_mutex_lock, (ompt_wait_id_t)(uintptr_t)user_lock, codeptr);
2995     }
2996 #endif
2997     return FTN_TRUE;
2998   } else {
2999 #if USE_ITT_BUILD
3000     __kmp_itt_lock_cancelled((kmp_user_lock_p)user_lock);
3001 #endif
3002     return FTN_FALSE;
3003   }
3004 
3005 #else // KMP_USE_DYNAMIC_LOCK
3006 
3007   kmp_user_lock_p lck;
3008   int rc;
3009 
3010   if ((__kmp_user_lock_kind == lk_tas) &&
3011       (sizeof(lck->tas.lk.poll) <= OMP_LOCK_T_SIZE)) {
3012     lck = (kmp_user_lock_p)user_lock;
3013   }
3014 #if KMP_USE_FUTEX
3015   else if ((__kmp_user_lock_kind == lk_futex) &&
3016            (sizeof(lck->futex.lk.poll) <= OMP_LOCK_T_SIZE)) {
3017     lck = (kmp_user_lock_p)user_lock;
3018   }
3019 #endif
3020   else {
3021     lck = __kmp_lookup_user_lock(user_lock, "omp_test_lock");
3022   }
3023 
3024 #if USE_ITT_BUILD
3025   __kmp_itt_lock_acquiring(lck);
3026 #endif /* USE_ITT_BUILD */
3027 #if OMPT_SUPPORT && OMPT_OPTIONAL
3028   // This is the case, if called from omp_init_lock_with_hint:
3029   void *codeptr = OMPT_LOAD_RETURN_ADDRESS(gtid);
3030   if (!codeptr)
3031     codeptr = OMPT_GET_RETURN_ADDRESS(0);
3032   if (ompt_enabled.ompt_callback_mutex_acquire) {
3033     ompt_callbacks.ompt_callback(ompt_callback_mutex_acquire)(
3034         ompt_mutex_lock, omp_lock_hint_none, __ompt_get_mutex_impl_type(),
3035         (ompt_wait_id_t)(uintptr_t)lck, codeptr);
3036   }
3037 #endif
3038 
3039   rc = TEST_LOCK(lck, gtid);
3040 #if USE_ITT_BUILD
3041   if (rc) {
3042     __kmp_itt_lock_acquired(lck);
3043   } else {
3044     __kmp_itt_lock_cancelled(lck);
3045   }
3046 #endif /* USE_ITT_BUILD */
3047 #if OMPT_SUPPORT && OMPT_OPTIONAL
3048   if (rc && ompt_enabled.ompt_callback_mutex_acquired) {
3049     ompt_callbacks.ompt_callback(ompt_callback_mutex_acquired)(
3050         ompt_mutex_lock, (ompt_wait_id_t)(uintptr_t)lck, codeptr);
3051   }
3052 #endif
3053 
3054   return (rc ? FTN_TRUE : FTN_FALSE);
3055 
3056 /* Can't use serial interval since not block structured */
3057 
3058 #endif // KMP_USE_DYNAMIC_LOCK
3059 }
3060 
3061 /* try to acquire the lock */
3062 int __kmpc_test_nest_lock(ident_t *loc, kmp_int32 gtid, void **user_lock) {
3063 #if KMP_USE_DYNAMIC_LOCK
3064   int rc;
3065 #if USE_ITT_BUILD
3066   __kmp_itt_lock_acquiring((kmp_user_lock_p)user_lock);
3067 #endif
3068 #if OMPT_SUPPORT && OMPT_OPTIONAL
3069   // This is the case, if called from omp_init_lock_with_hint:
3070   void *codeptr = OMPT_LOAD_RETURN_ADDRESS(gtid);
3071   if (!codeptr)
3072     codeptr = OMPT_GET_RETURN_ADDRESS(0);
3073   if (ompt_enabled.ompt_callback_mutex_acquire) {
3074     ompt_callbacks.ompt_callback(ompt_callback_mutex_acquire)(
3075         ompt_mutex_nest_lock, omp_lock_hint_none,
3076         __ompt_get_mutex_impl_type(user_lock),
3077         (ompt_wait_id_t)(uintptr_t)user_lock, codeptr);
3078   }
3079 #endif
3080   rc = KMP_D_LOCK_FUNC(user_lock, test)((kmp_dyna_lock_t *)user_lock, gtid);
3081 #if USE_ITT_BUILD
3082   if (rc) {
3083     __kmp_itt_lock_acquired((kmp_user_lock_p)user_lock);
3084   } else {
3085     __kmp_itt_lock_cancelled((kmp_user_lock_p)user_lock);
3086   }
3087 #endif
3088 #if OMPT_SUPPORT && OMPT_OPTIONAL
3089   if (ompt_enabled.enabled && rc) {
3090     if (rc == 1) {
3091       if (ompt_enabled.ompt_callback_mutex_acquired) {
3092         // lock_first
3093         ompt_callbacks.ompt_callback(ompt_callback_mutex_acquired)(
3094             ompt_mutex_nest_lock, (ompt_wait_id_t)(uintptr_t)user_lock,
3095             codeptr);
3096       }
3097     } else {
3098       if (ompt_enabled.ompt_callback_nest_lock) {
3099         // lock_next
3100         ompt_callbacks.ompt_callback(ompt_callback_nest_lock)(
3101             ompt_scope_begin, (ompt_wait_id_t)(uintptr_t)user_lock, codeptr);
3102       }
3103     }
3104   }
3105 #endif
3106   return rc;
3107 
3108 #else // KMP_USE_DYNAMIC_LOCK
3109 
3110   kmp_user_lock_p lck;
3111   int rc;
3112 
3113   if ((__kmp_user_lock_kind == lk_tas) &&
3114       (sizeof(lck->tas.lk.poll) + sizeof(lck->tas.lk.depth_locked) <=
3115        OMP_NEST_LOCK_T_SIZE)) {
3116     lck = (kmp_user_lock_p)user_lock;
3117   }
3118 #if KMP_USE_FUTEX
3119   else if ((__kmp_user_lock_kind == lk_futex) &&
3120            (sizeof(lck->futex.lk.poll) + sizeof(lck->futex.lk.depth_locked) <=
3121             OMP_NEST_LOCK_T_SIZE)) {
3122     lck = (kmp_user_lock_p)user_lock;
3123   }
3124 #endif
3125   else {
3126     lck = __kmp_lookup_user_lock(user_lock, "omp_test_nest_lock");
3127   }
3128 
3129 #if USE_ITT_BUILD
3130   __kmp_itt_lock_acquiring(lck);
3131 #endif /* USE_ITT_BUILD */
3132 
3133 #if OMPT_SUPPORT && OMPT_OPTIONAL
3134   // This is the case, if called from omp_init_lock_with_hint:
3135   void *codeptr = OMPT_LOAD_RETURN_ADDRESS(gtid);
3136   if (!codeptr)
3137     codeptr = OMPT_GET_RETURN_ADDRESS(0);
3138   if (ompt_enabled.enabled) &&
3139         ompt_enabled.ompt_callback_mutex_acquire) {
3140       ompt_callbacks.ompt_callback(ompt_callback_mutex_acquire)(
3141           ompt_mutex_nest_lock, omp_lock_hint_none,
3142           __ompt_get_mutex_impl_type(), (ompt_wait_id_t)(uintptr_t)lck,
3143           codeptr);
3144     }
3145 #endif
3146 
3147   rc = TEST_NESTED_LOCK(lck, gtid);
3148 #if USE_ITT_BUILD
3149   if (rc) {
3150     __kmp_itt_lock_acquired(lck);
3151   } else {
3152     __kmp_itt_lock_cancelled(lck);
3153   }
3154 #endif /* USE_ITT_BUILD */
3155 #if OMPT_SUPPORT && OMPT_OPTIONAL
3156   if (ompt_enabled.enabled && rc) {
3157     if (rc == 1) {
3158       if (ompt_enabled.ompt_callback_mutex_acquired) {
3159         // lock_first
3160         ompt_callbacks.ompt_callback(ompt_callback_mutex_acquired)(
3161             ompt_mutex_nest_lock, (ompt_wait_id_t)(uintptr_t)lck, codeptr);
3162       }
3163     } else {
3164       if (ompt_enabled.ompt_callback_nest_lock) {
3165         // lock_next
3166         ompt_callbacks.ompt_callback(ompt_callback_nest_lock)(
3167             ompt_mutex_scope_begin, (ompt_wait_id_t)(uintptr_t)lck, codeptr);
3168       }
3169     }
3170   }
3171 #endif
3172   return rc;
3173 
3174 /* Can't use serial interval since not block structured */
3175 
3176 #endif // KMP_USE_DYNAMIC_LOCK
3177 }
3178 
3179 // Interface to fast scalable reduce methods routines
3180 
3181 // keep the selected method in a thread local structure for cross-function
3182 // usage: will be used in __kmpc_end_reduce* functions;
3183 // another solution: to re-determine the method one more time in
3184 // __kmpc_end_reduce* functions (new prototype required then)
3185 // AT: which solution is better?
3186 #define __KMP_SET_REDUCTION_METHOD(gtid, rmethod)                              \
3187   ((__kmp_threads[(gtid)]->th.th_local.packed_reduction_method) = (rmethod))
3188 
3189 #define __KMP_GET_REDUCTION_METHOD(gtid)                                       \
3190   (__kmp_threads[(gtid)]->th.th_local.packed_reduction_method)
3191 
3192 // description of the packed_reduction_method variable: look at the macros in
3193 // kmp.h
3194 
3195 // used in a critical section reduce block
3196 static __forceinline void
3197 __kmp_enter_critical_section_reduce_block(ident_t *loc, kmp_int32 global_tid,
3198                                           kmp_critical_name *crit) {
3199 
3200   // this lock was visible to a customer and to the threading profile tool as a
3201   // serial overhead span (although it's used for an internal purpose only)
3202   //            why was it visible in previous implementation?
3203   //            should we keep it visible in new reduce block?
3204   kmp_user_lock_p lck;
3205 
3206 #if KMP_USE_DYNAMIC_LOCK
3207 
3208   kmp_dyna_lock_t *lk = (kmp_dyna_lock_t *)crit;
3209   // Check if it is initialized.
3210   if (*lk == 0) {
3211     if (KMP_IS_D_LOCK(__kmp_user_lock_seq)) {
3212       KMP_COMPARE_AND_STORE_ACQ32((volatile kmp_int32 *)crit, 0,
3213                                   KMP_GET_D_TAG(__kmp_user_lock_seq));
3214     } else {
3215       __kmp_init_indirect_csptr(crit, loc, global_tid,
3216                                 KMP_GET_I_TAG(__kmp_user_lock_seq));
3217     }
3218   }
3219   // Branch for accessing the actual lock object and set operation. This
3220   // branching is inevitable since this lock initialization does not follow the
3221   // normal dispatch path (lock table is not used).
3222   if (KMP_EXTRACT_D_TAG(lk) != 0) {
3223     lck = (kmp_user_lock_p)lk;
3224     KMP_DEBUG_ASSERT(lck != NULL);
3225     if (__kmp_env_consistency_check) {
3226       __kmp_push_sync(global_tid, ct_critical, loc, lck, __kmp_user_lock_seq);
3227     }
3228     KMP_D_LOCK_FUNC(lk, set)(lk, global_tid);
3229   } else {
3230     kmp_indirect_lock_t *ilk = *((kmp_indirect_lock_t **)lk);
3231     lck = ilk->lock;
3232     KMP_DEBUG_ASSERT(lck != NULL);
3233     if (__kmp_env_consistency_check) {
3234       __kmp_push_sync(global_tid, ct_critical, loc, lck, __kmp_user_lock_seq);
3235     }
3236     KMP_I_LOCK_FUNC(ilk, set)(lck, global_tid);
3237   }
3238 
3239 #else // KMP_USE_DYNAMIC_LOCK
3240 
3241   // We know that the fast reduction code is only emitted by Intel compilers
3242   // with 32 byte critical sections. If there isn't enough space, then we
3243   // have to use a pointer.
3244   if (__kmp_base_user_lock_size <= INTEL_CRITICAL_SIZE) {
3245     lck = (kmp_user_lock_p)crit;
3246   } else {
3247     lck = __kmp_get_critical_section_ptr(crit, loc, global_tid);
3248   }
3249   KMP_DEBUG_ASSERT(lck != NULL);
3250 
3251   if (__kmp_env_consistency_check)
3252     __kmp_push_sync(global_tid, ct_critical, loc, lck);
3253 
3254   __kmp_acquire_user_lock_with_checks(lck, global_tid);
3255 
3256 #endif // KMP_USE_DYNAMIC_LOCK
3257 }
3258 
3259 // used in a critical section reduce block
3260 static __forceinline void
3261 __kmp_end_critical_section_reduce_block(ident_t *loc, kmp_int32 global_tid,
3262                                         kmp_critical_name *crit) {
3263 
3264   kmp_user_lock_p lck;
3265 
3266 #if KMP_USE_DYNAMIC_LOCK
3267 
3268   if (KMP_IS_D_LOCK(__kmp_user_lock_seq)) {
3269     lck = (kmp_user_lock_p)crit;
3270     if (__kmp_env_consistency_check)
3271       __kmp_pop_sync(global_tid, ct_critical, loc);
3272     KMP_D_LOCK_FUNC(lck, unset)((kmp_dyna_lock_t *)lck, global_tid);
3273   } else {
3274     kmp_indirect_lock_t *ilk =
3275         (kmp_indirect_lock_t *)TCR_PTR(*((kmp_indirect_lock_t **)crit));
3276     if (__kmp_env_consistency_check)
3277       __kmp_pop_sync(global_tid, ct_critical, loc);
3278     KMP_I_LOCK_FUNC(ilk, unset)(ilk->lock, global_tid);
3279   }
3280 
3281 #else // KMP_USE_DYNAMIC_LOCK
3282 
3283   // We know that the fast reduction code is only emitted by Intel compilers
3284   // with 32 byte critical sections. If there isn't enough space, then we have
3285   // to use a pointer.
3286   if (__kmp_base_user_lock_size > 32) {
3287     lck = *((kmp_user_lock_p *)crit);
3288     KMP_ASSERT(lck != NULL);
3289   } else {
3290     lck = (kmp_user_lock_p)crit;
3291   }
3292 
3293   if (__kmp_env_consistency_check)
3294     __kmp_pop_sync(global_tid, ct_critical, loc);
3295 
3296   __kmp_release_user_lock_with_checks(lck, global_tid);
3297 
3298 #endif // KMP_USE_DYNAMIC_LOCK
3299 } // __kmp_end_critical_section_reduce_block
3300 
3301 static __forceinline int
3302 __kmp_swap_teams_for_teams_reduction(kmp_info_t *th, kmp_team_t **team_p,
3303                                      int *task_state) {
3304   kmp_team_t *team;
3305 
3306   // Check if we are inside the teams construct?
3307   if (th->th.th_teams_microtask) {
3308     *team_p = team = th->th.th_team;
3309     if (team->t.t_level == th->th.th_teams_level) {
3310       // This is reduction at teams construct.
3311       KMP_DEBUG_ASSERT(!th->th.th_info.ds.ds_tid); // AC: check that tid == 0
3312       // Let's swap teams temporarily for the reduction.
3313       th->th.th_info.ds.ds_tid = team->t.t_master_tid;
3314       th->th.th_team = team->t.t_parent;
3315       th->th.th_team_nproc = th->th.th_team->t.t_nproc;
3316       th->th.th_task_team = th->th.th_team->t.t_task_team[0];
3317       *task_state = th->th.th_task_state;
3318       th->th.th_task_state = 0;
3319 
3320       return 1;
3321     }
3322   }
3323   return 0;
3324 }
3325 
3326 static __forceinline void
3327 __kmp_restore_swapped_teams(kmp_info_t *th, kmp_team_t *team, int task_state) {
3328   // Restore thread structure swapped in __kmp_swap_teams_for_teams_reduction.
3329   th->th.th_info.ds.ds_tid = 0;
3330   th->th.th_team = team;
3331   th->th.th_team_nproc = team->t.t_nproc;
3332   th->th.th_task_team = team->t.t_task_team[task_state];
3333   th->th.th_task_state = task_state;
3334 }
3335 
3336 /* 2.a.i. Reduce Block without a terminating barrier */
3337 /*!
3338 @ingroup SYNCHRONIZATION
3339 @param loc source location information
3340 @param global_tid global thread number
3341 @param num_vars number of items (variables) to be reduced
3342 @param reduce_size size of data in bytes to be reduced
3343 @param reduce_data pointer to data to be reduced
3344 @param reduce_func callback function providing reduction operation on two
3345 operands and returning result of reduction in lhs_data
3346 @param lck pointer to the unique lock data structure
3347 @result 1 for the master thread, 0 for all other team threads, 2 for all team
3348 threads if atomic reduction needed
3349 
3350 The nowait version is used for a reduce clause with the nowait argument.
3351 */
3352 kmp_int32
3353 __kmpc_reduce_nowait(ident_t *loc, kmp_int32 global_tid, kmp_int32 num_vars,
3354                      size_t reduce_size, void *reduce_data,
3355                      void (*reduce_func)(void *lhs_data, void *rhs_data),
3356                      kmp_critical_name *lck) {
3357 
3358   KMP_COUNT_BLOCK(REDUCE_nowait);
3359   int retval = 0;
3360   PACKED_REDUCTION_METHOD_T packed_reduction_method;
3361   kmp_info_t *th;
3362   kmp_team_t *team;
3363   int teams_swapped = 0, task_state;
3364   KA_TRACE(10, ("__kmpc_reduce_nowait() enter: called T#%d\n", global_tid));
3365   __kmp_assert_valid_gtid(global_tid);
3366 
3367   // why do we need this initialization here at all?
3368   // Reduction clause can not be used as a stand-alone directive.
3369 
3370   // do not call __kmp_serial_initialize(), it will be called by
3371   // __kmp_parallel_initialize() if needed
3372   // possible detection of false-positive race by the threadchecker ???
3373   if (!TCR_4(__kmp_init_parallel))
3374     __kmp_parallel_initialize();
3375 
3376   __kmp_resume_if_soft_paused();
3377 
3378 // check correctness of reduce block nesting
3379 #if KMP_USE_DYNAMIC_LOCK
3380   if (__kmp_env_consistency_check)
3381     __kmp_push_sync(global_tid, ct_reduce, loc, NULL, 0);
3382 #else
3383   if (__kmp_env_consistency_check)
3384     __kmp_push_sync(global_tid, ct_reduce, loc, NULL);
3385 #endif
3386 
3387   th = __kmp_thread_from_gtid(global_tid);
3388   teams_swapped = __kmp_swap_teams_for_teams_reduction(th, &team, &task_state);
3389 
3390   // packed_reduction_method value will be reused by __kmp_end_reduce* function,
3391   // the value should be kept in a variable
3392   // the variable should be either a construct-specific or thread-specific
3393   // property, not a team specific property
3394   //     (a thread can reach the next reduce block on the next construct, reduce
3395   //     method may differ on the next construct)
3396   // an ident_t "loc" parameter could be used as a construct-specific property
3397   // (what if loc == 0?)
3398   //     (if both construct-specific and team-specific variables were shared,
3399   //     then unness extra syncs should be needed)
3400   // a thread-specific variable is better regarding two issues above (next
3401   // construct and extra syncs)
3402   // a thread-specific "th_local.reduction_method" variable is used currently
3403   // each thread executes 'determine' and 'set' lines (no need to execute by one
3404   // thread, to avoid unness extra syncs)
3405 
3406   packed_reduction_method = __kmp_determine_reduction_method(
3407       loc, global_tid, num_vars, reduce_size, reduce_data, reduce_func, lck);
3408   __KMP_SET_REDUCTION_METHOD(global_tid, packed_reduction_method);
3409 
3410   OMPT_REDUCTION_DECL(th, global_tid);
3411   if (packed_reduction_method == critical_reduce_block) {
3412 
3413     OMPT_REDUCTION_BEGIN;
3414 
3415     __kmp_enter_critical_section_reduce_block(loc, global_tid, lck);
3416     retval = 1;
3417 
3418   } else if (packed_reduction_method == empty_reduce_block) {
3419 
3420     OMPT_REDUCTION_BEGIN;
3421 
3422     // usage: if team size == 1, no synchronization is required ( Intel
3423     // platforms only )
3424     retval = 1;
3425 
3426   } else if (packed_reduction_method == atomic_reduce_block) {
3427 
3428     retval = 2;
3429 
3430     // all threads should do this pop here (because __kmpc_end_reduce_nowait()
3431     // won't be called by the code gen)
3432     //     (it's not quite good, because the checking block has been closed by
3433     //     this 'pop',
3434     //      but atomic operation has not been executed yet, will be executed
3435     //      slightly later, literally on next instruction)
3436     if (__kmp_env_consistency_check)
3437       __kmp_pop_sync(global_tid, ct_reduce, loc);
3438 
3439   } else if (TEST_REDUCTION_METHOD(packed_reduction_method,
3440                                    tree_reduce_block)) {
3441 
3442 // AT: performance issue: a real barrier here
3443 // AT:     (if master goes slow, other threads are blocked here waiting for the
3444 // master to come and release them)
3445 // AT:     (it's not what a customer might expect specifying NOWAIT clause)
3446 // AT:     (specifying NOWAIT won't result in improvement of performance, it'll
3447 // be confusing to a customer)
3448 // AT: another implementation of *barrier_gather*nowait() (or some other design)
3449 // might go faster and be more in line with sense of NOWAIT
3450 // AT: TO DO: do epcc test and compare times
3451 
3452 // this barrier should be invisible to a customer and to the threading profile
3453 // tool (it's neither a terminating barrier nor customer's code, it's
3454 // used for an internal purpose)
3455 #if OMPT_SUPPORT
3456     // JP: can this barrier potentially leed to task scheduling?
3457     // JP: as long as there is a barrier in the implementation, OMPT should and
3458     // will provide the barrier events
3459     //         so we set-up the necessary frame/return addresses.
3460     ompt_frame_t *ompt_frame;
3461     if (ompt_enabled.enabled) {
3462       __ompt_get_task_info_internal(0, NULL, NULL, &ompt_frame, NULL, NULL);
3463       if (ompt_frame->enter_frame.ptr == NULL)
3464         ompt_frame->enter_frame.ptr = OMPT_GET_FRAME_ADDRESS(0);
3465       OMPT_STORE_RETURN_ADDRESS(global_tid);
3466     }
3467 #endif
3468 #if USE_ITT_NOTIFY
3469     __kmp_threads[global_tid]->th.th_ident = loc;
3470 #endif
3471     retval =
3472         __kmp_barrier(UNPACK_REDUCTION_BARRIER(packed_reduction_method),
3473                       global_tid, FALSE, reduce_size, reduce_data, reduce_func);
3474     retval = (retval != 0) ? (0) : (1);
3475 #if OMPT_SUPPORT && OMPT_OPTIONAL
3476     if (ompt_enabled.enabled) {
3477       ompt_frame->enter_frame = ompt_data_none;
3478     }
3479 #endif
3480 
3481     // all other workers except master should do this pop here
3482     //     ( none of other workers will get to __kmpc_end_reduce_nowait() )
3483     if (__kmp_env_consistency_check) {
3484       if (retval == 0) {
3485         __kmp_pop_sync(global_tid, ct_reduce, loc);
3486       }
3487     }
3488 
3489   } else {
3490 
3491     // should never reach this block
3492     KMP_ASSERT(0); // "unexpected method"
3493   }
3494   if (teams_swapped) {
3495     __kmp_restore_swapped_teams(th, team, task_state);
3496   }
3497   KA_TRACE(
3498       10,
3499       ("__kmpc_reduce_nowait() exit: called T#%d: method %08x, returns %08x\n",
3500        global_tid, packed_reduction_method, retval));
3501 
3502   return retval;
3503 }
3504 
3505 /*!
3506 @ingroup SYNCHRONIZATION
3507 @param loc source location information
3508 @param global_tid global thread id.
3509 @param lck pointer to the unique lock data structure
3510 
3511 Finish the execution of a reduce nowait.
3512 */
3513 void __kmpc_end_reduce_nowait(ident_t *loc, kmp_int32 global_tid,
3514                               kmp_critical_name *lck) {
3515 
3516   PACKED_REDUCTION_METHOD_T packed_reduction_method;
3517 
3518   KA_TRACE(10, ("__kmpc_end_reduce_nowait() enter: called T#%d\n", global_tid));
3519   __kmp_assert_valid_gtid(global_tid);
3520 
3521   packed_reduction_method = __KMP_GET_REDUCTION_METHOD(global_tid);
3522 
3523   OMPT_REDUCTION_DECL(__kmp_thread_from_gtid(global_tid), global_tid);
3524 
3525   if (packed_reduction_method == critical_reduce_block) {
3526 
3527     __kmp_end_critical_section_reduce_block(loc, global_tid, lck);
3528     OMPT_REDUCTION_END;
3529 
3530   } else if (packed_reduction_method == empty_reduce_block) {
3531 
3532     // usage: if team size == 1, no synchronization is required ( on Intel
3533     // platforms only )
3534 
3535     OMPT_REDUCTION_END;
3536 
3537   } else if (packed_reduction_method == atomic_reduce_block) {
3538 
3539     // neither master nor other workers should get here
3540     //     (code gen does not generate this call in case 2: atomic reduce block)
3541     // actually it's better to remove this elseif at all;
3542     // after removal this value will checked by the 'else' and will assert
3543 
3544   } else if (TEST_REDUCTION_METHOD(packed_reduction_method,
3545                                    tree_reduce_block)) {
3546 
3547     // only master gets here
3548     // OMPT: tree reduction is annotated in the barrier code
3549 
3550   } else {
3551 
3552     // should never reach this block
3553     KMP_ASSERT(0); // "unexpected method"
3554   }
3555 
3556   if (__kmp_env_consistency_check)
3557     __kmp_pop_sync(global_tid, ct_reduce, loc);
3558 
3559   KA_TRACE(10, ("__kmpc_end_reduce_nowait() exit: called T#%d: method %08x\n",
3560                 global_tid, packed_reduction_method));
3561 
3562   return;
3563 }
3564 
3565 /* 2.a.ii. Reduce Block with a terminating barrier */
3566 
3567 /*!
3568 @ingroup SYNCHRONIZATION
3569 @param loc source location information
3570 @param global_tid global thread number
3571 @param num_vars number of items (variables) to be reduced
3572 @param reduce_size size of data in bytes to be reduced
3573 @param reduce_data pointer to data to be reduced
3574 @param reduce_func callback function providing reduction operation on two
3575 operands and returning result of reduction in lhs_data
3576 @param lck pointer to the unique lock data structure
3577 @result 1 for the master thread, 0 for all other team threads, 2 for all team
3578 threads if atomic reduction needed
3579 
3580 A blocking reduce that includes an implicit barrier.
3581 */
3582 kmp_int32 __kmpc_reduce(ident_t *loc, kmp_int32 global_tid, kmp_int32 num_vars,
3583                         size_t reduce_size, void *reduce_data,
3584                         void (*reduce_func)(void *lhs_data, void *rhs_data),
3585                         kmp_critical_name *lck) {
3586   KMP_COUNT_BLOCK(REDUCE_wait);
3587   int retval = 0;
3588   PACKED_REDUCTION_METHOD_T packed_reduction_method;
3589   kmp_info_t *th;
3590   kmp_team_t *team;
3591   int teams_swapped = 0, task_state;
3592 
3593   KA_TRACE(10, ("__kmpc_reduce() enter: called T#%d\n", global_tid));
3594   __kmp_assert_valid_gtid(global_tid);
3595 
3596   // why do we need this initialization here at all?
3597   // Reduction clause can not be a stand-alone directive.
3598 
3599   // do not call __kmp_serial_initialize(), it will be called by
3600   // __kmp_parallel_initialize() if needed
3601   // possible detection of false-positive race by the threadchecker ???
3602   if (!TCR_4(__kmp_init_parallel))
3603     __kmp_parallel_initialize();
3604 
3605   __kmp_resume_if_soft_paused();
3606 
3607 // check correctness of reduce block nesting
3608 #if KMP_USE_DYNAMIC_LOCK
3609   if (__kmp_env_consistency_check)
3610     __kmp_push_sync(global_tid, ct_reduce, loc, NULL, 0);
3611 #else
3612   if (__kmp_env_consistency_check)
3613     __kmp_push_sync(global_tid, ct_reduce, loc, NULL);
3614 #endif
3615 
3616   th = __kmp_thread_from_gtid(global_tid);
3617   teams_swapped = __kmp_swap_teams_for_teams_reduction(th, &team, &task_state);
3618 
3619   packed_reduction_method = __kmp_determine_reduction_method(
3620       loc, global_tid, num_vars, reduce_size, reduce_data, reduce_func, lck);
3621   __KMP_SET_REDUCTION_METHOD(global_tid, packed_reduction_method);
3622 
3623   OMPT_REDUCTION_DECL(th, global_tid);
3624 
3625   if (packed_reduction_method == critical_reduce_block) {
3626 
3627     OMPT_REDUCTION_BEGIN;
3628     __kmp_enter_critical_section_reduce_block(loc, global_tid, lck);
3629     retval = 1;
3630 
3631   } else if (packed_reduction_method == empty_reduce_block) {
3632 
3633     OMPT_REDUCTION_BEGIN;
3634     // usage: if team size == 1, no synchronization is required ( Intel
3635     // platforms only )
3636     retval = 1;
3637 
3638   } else if (packed_reduction_method == atomic_reduce_block) {
3639 
3640     retval = 2;
3641 
3642   } else if (TEST_REDUCTION_METHOD(packed_reduction_method,
3643                                    tree_reduce_block)) {
3644 
3645 // case tree_reduce_block:
3646 // this barrier should be visible to a customer and to the threading profile
3647 // tool (it's a terminating barrier on constructs if NOWAIT not specified)
3648 #if OMPT_SUPPORT
3649     ompt_frame_t *ompt_frame;
3650     if (ompt_enabled.enabled) {
3651       __ompt_get_task_info_internal(0, NULL, NULL, &ompt_frame, NULL, NULL);
3652       if (ompt_frame->enter_frame.ptr == NULL)
3653         ompt_frame->enter_frame.ptr = OMPT_GET_FRAME_ADDRESS(0);
3654       OMPT_STORE_RETURN_ADDRESS(global_tid);
3655     }
3656 #endif
3657 #if USE_ITT_NOTIFY
3658     __kmp_threads[global_tid]->th.th_ident =
3659         loc; // needed for correct notification of frames
3660 #endif
3661     retval =
3662         __kmp_barrier(UNPACK_REDUCTION_BARRIER(packed_reduction_method),
3663                       global_tid, TRUE, reduce_size, reduce_data, reduce_func);
3664     retval = (retval != 0) ? (0) : (1);
3665 #if OMPT_SUPPORT && OMPT_OPTIONAL
3666     if (ompt_enabled.enabled) {
3667       ompt_frame->enter_frame = ompt_data_none;
3668     }
3669 #endif
3670 
3671     // all other workers except master should do this pop here
3672     // ( none of other workers except master will enter __kmpc_end_reduce() )
3673     if (__kmp_env_consistency_check) {
3674       if (retval == 0) { // 0: all other workers; 1: master
3675         __kmp_pop_sync(global_tid, ct_reduce, loc);
3676       }
3677     }
3678 
3679   } else {
3680 
3681     // should never reach this block
3682     KMP_ASSERT(0); // "unexpected method"
3683   }
3684   if (teams_swapped) {
3685     __kmp_restore_swapped_teams(th, team, task_state);
3686   }
3687 
3688   KA_TRACE(10,
3689            ("__kmpc_reduce() exit: called T#%d: method %08x, returns %08x\n",
3690             global_tid, packed_reduction_method, retval));
3691   return retval;
3692 }
3693 
3694 /*!
3695 @ingroup SYNCHRONIZATION
3696 @param loc source location information
3697 @param global_tid global thread id.
3698 @param lck pointer to the unique lock data structure
3699 
3700 Finish the execution of a blocking reduce.
3701 The <tt>lck</tt> pointer must be the same as that used in the corresponding
3702 start function.
3703 */
3704 void __kmpc_end_reduce(ident_t *loc, kmp_int32 global_tid,
3705                        kmp_critical_name *lck) {
3706 
3707   PACKED_REDUCTION_METHOD_T packed_reduction_method;
3708   kmp_info_t *th;
3709   kmp_team_t *team;
3710   int teams_swapped = 0, task_state;
3711 
3712   KA_TRACE(10, ("__kmpc_end_reduce() enter: called T#%d\n", global_tid));
3713   __kmp_assert_valid_gtid(global_tid);
3714 
3715   th = __kmp_thread_from_gtid(global_tid);
3716   teams_swapped = __kmp_swap_teams_for_teams_reduction(th, &team, &task_state);
3717 
3718   packed_reduction_method = __KMP_GET_REDUCTION_METHOD(global_tid);
3719 
3720   // this barrier should be visible to a customer and to the threading profile
3721   // tool (it's a terminating barrier on constructs if NOWAIT not specified)
3722   OMPT_REDUCTION_DECL(th, global_tid);
3723 
3724   if (packed_reduction_method == critical_reduce_block) {
3725     __kmp_end_critical_section_reduce_block(loc, global_tid, lck);
3726 
3727     OMPT_REDUCTION_END;
3728 
3729 // TODO: implicit barrier: should be exposed
3730 #if OMPT_SUPPORT
3731     ompt_frame_t *ompt_frame;
3732     if (ompt_enabled.enabled) {
3733       __ompt_get_task_info_internal(0, NULL, NULL, &ompt_frame, NULL, NULL);
3734       if (ompt_frame->enter_frame.ptr == NULL)
3735         ompt_frame->enter_frame.ptr = OMPT_GET_FRAME_ADDRESS(0);
3736       OMPT_STORE_RETURN_ADDRESS(global_tid);
3737     }
3738 #endif
3739 #if USE_ITT_NOTIFY
3740     __kmp_threads[global_tid]->th.th_ident = loc;
3741 #endif
3742     __kmp_barrier(bs_plain_barrier, global_tid, FALSE, 0, NULL, NULL);
3743 #if OMPT_SUPPORT && OMPT_OPTIONAL
3744     if (ompt_enabled.enabled) {
3745       ompt_frame->enter_frame = ompt_data_none;
3746     }
3747 #endif
3748 
3749   } else if (packed_reduction_method == empty_reduce_block) {
3750 
3751     OMPT_REDUCTION_END;
3752 
3753 // usage: if team size==1, no synchronization is required (Intel platforms only)
3754 
3755 // TODO: implicit barrier: should be exposed
3756 #if OMPT_SUPPORT
3757     ompt_frame_t *ompt_frame;
3758     if (ompt_enabled.enabled) {
3759       __ompt_get_task_info_internal(0, NULL, NULL, &ompt_frame, NULL, NULL);
3760       if (ompt_frame->enter_frame.ptr == NULL)
3761         ompt_frame->enter_frame.ptr = OMPT_GET_FRAME_ADDRESS(0);
3762       OMPT_STORE_RETURN_ADDRESS(global_tid);
3763     }
3764 #endif
3765 #if USE_ITT_NOTIFY
3766     __kmp_threads[global_tid]->th.th_ident = loc;
3767 #endif
3768     __kmp_barrier(bs_plain_barrier, global_tid, FALSE, 0, NULL, NULL);
3769 #if OMPT_SUPPORT && OMPT_OPTIONAL
3770     if (ompt_enabled.enabled) {
3771       ompt_frame->enter_frame = ompt_data_none;
3772     }
3773 #endif
3774 
3775   } else if (packed_reduction_method == atomic_reduce_block) {
3776 
3777 #if OMPT_SUPPORT
3778     ompt_frame_t *ompt_frame;
3779     if (ompt_enabled.enabled) {
3780       __ompt_get_task_info_internal(0, NULL, NULL, &ompt_frame, NULL, NULL);
3781       if (ompt_frame->enter_frame.ptr == NULL)
3782         ompt_frame->enter_frame.ptr = OMPT_GET_FRAME_ADDRESS(0);
3783       OMPT_STORE_RETURN_ADDRESS(global_tid);
3784     }
3785 #endif
3786 // TODO: implicit barrier: should be exposed
3787 #if USE_ITT_NOTIFY
3788     __kmp_threads[global_tid]->th.th_ident = loc;
3789 #endif
3790     __kmp_barrier(bs_plain_barrier, global_tid, FALSE, 0, NULL, NULL);
3791 #if OMPT_SUPPORT && OMPT_OPTIONAL
3792     if (ompt_enabled.enabled) {
3793       ompt_frame->enter_frame = ompt_data_none;
3794     }
3795 #endif
3796 
3797   } else if (TEST_REDUCTION_METHOD(packed_reduction_method,
3798                                    tree_reduce_block)) {
3799 
3800     // only master executes here (master releases all other workers)
3801     __kmp_end_split_barrier(UNPACK_REDUCTION_BARRIER(packed_reduction_method),
3802                             global_tid);
3803 
3804   } else {
3805 
3806     // should never reach this block
3807     KMP_ASSERT(0); // "unexpected method"
3808   }
3809   if (teams_swapped) {
3810     __kmp_restore_swapped_teams(th, team, task_state);
3811   }
3812 
3813   if (__kmp_env_consistency_check)
3814     __kmp_pop_sync(global_tid, ct_reduce, loc);
3815 
3816   KA_TRACE(10, ("__kmpc_end_reduce() exit: called T#%d: method %08x\n",
3817                 global_tid, packed_reduction_method));
3818 
3819   return;
3820 }
3821 
3822 #undef __KMP_GET_REDUCTION_METHOD
3823 #undef __KMP_SET_REDUCTION_METHOD
3824 
3825 /* end of interface to fast scalable reduce routines */
3826 
3827 kmp_uint64 __kmpc_get_taskid() {
3828 
3829   kmp_int32 gtid;
3830   kmp_info_t *thread;
3831 
3832   gtid = __kmp_get_gtid();
3833   if (gtid < 0) {
3834     return 0;
3835   }
3836   thread = __kmp_thread_from_gtid(gtid);
3837   return thread->th.th_current_task->td_task_id;
3838 
3839 } // __kmpc_get_taskid
3840 
3841 kmp_uint64 __kmpc_get_parent_taskid() {
3842 
3843   kmp_int32 gtid;
3844   kmp_info_t *thread;
3845   kmp_taskdata_t *parent_task;
3846 
3847   gtid = __kmp_get_gtid();
3848   if (gtid < 0) {
3849     return 0;
3850   }
3851   thread = __kmp_thread_from_gtid(gtid);
3852   parent_task = thread->th.th_current_task->td_parent;
3853   return (parent_task == NULL ? 0 : parent_task->td_task_id);
3854 
3855 } // __kmpc_get_parent_taskid
3856 
3857 /*!
3858 @ingroup WORK_SHARING
3859 @param loc  source location information.
3860 @param gtid  global thread number.
3861 @param num_dims  number of associated doacross loops.
3862 @param dims  info on loops bounds.
3863 
3864 Initialize doacross loop information.
3865 Expect compiler send us inclusive bounds,
3866 e.g. for(i=2;i<9;i+=2) lo=2, up=8, st=2.
3867 */
3868 void __kmpc_doacross_init(ident_t *loc, int gtid, int num_dims,
3869                           const struct kmp_dim *dims) {
3870   __kmp_assert_valid_gtid(gtid);
3871   int j, idx;
3872   kmp_int64 last, trace_count;
3873   kmp_info_t *th = __kmp_threads[gtid];
3874   kmp_team_t *team = th->th.th_team;
3875   kmp_uint32 *flags;
3876   kmp_disp_t *pr_buf = th->th.th_dispatch;
3877   dispatch_shared_info_t *sh_buf;
3878 
3879   KA_TRACE(
3880       20,
3881       ("__kmpc_doacross_init() enter: called T#%d, num dims %d, active %d\n",
3882        gtid, num_dims, !team->t.t_serialized));
3883   KMP_DEBUG_ASSERT(dims != NULL);
3884   KMP_DEBUG_ASSERT(num_dims > 0);
3885 
3886   if (team->t.t_serialized) {
3887     KA_TRACE(20, ("__kmpc_doacross_init() exit: serialized team\n"));
3888     return; // no dependencies if team is serialized
3889   }
3890   KMP_DEBUG_ASSERT(team->t.t_nproc > 1);
3891   idx = pr_buf->th_doacross_buf_idx++; // Increment index of shared buffer for
3892   // the next loop
3893   sh_buf = &team->t.t_disp_buffer[idx % __kmp_dispatch_num_buffers];
3894 
3895   // Save bounds info into allocated private buffer
3896   KMP_DEBUG_ASSERT(pr_buf->th_doacross_info == NULL);
3897   pr_buf->th_doacross_info = (kmp_int64 *)__kmp_thread_malloc(
3898       th, sizeof(kmp_int64) * (4 * num_dims + 1));
3899   KMP_DEBUG_ASSERT(pr_buf->th_doacross_info != NULL);
3900   pr_buf->th_doacross_info[0] =
3901       (kmp_int64)num_dims; // first element is number of dimensions
3902   // Save also address of num_done in order to access it later without knowing
3903   // the buffer index
3904   pr_buf->th_doacross_info[1] = (kmp_int64)&sh_buf->doacross_num_done;
3905   pr_buf->th_doacross_info[2] = dims[0].lo;
3906   pr_buf->th_doacross_info[3] = dims[0].up;
3907   pr_buf->th_doacross_info[4] = dims[0].st;
3908   last = 5;
3909   for (j = 1; j < num_dims; ++j) {
3910     kmp_int64
3911         range_length; // To keep ranges of all dimensions but the first dims[0]
3912     if (dims[j].st == 1) { // most common case
3913       // AC: should we care of ranges bigger than LLONG_MAX? (not for now)
3914       range_length = dims[j].up - dims[j].lo + 1;
3915     } else {
3916       if (dims[j].st > 0) {
3917         KMP_DEBUG_ASSERT(dims[j].up > dims[j].lo);
3918         range_length = (kmp_uint64)(dims[j].up - dims[j].lo) / dims[j].st + 1;
3919       } else { // negative increment
3920         KMP_DEBUG_ASSERT(dims[j].lo > dims[j].up);
3921         range_length =
3922             (kmp_uint64)(dims[j].lo - dims[j].up) / (-dims[j].st) + 1;
3923       }
3924     }
3925     pr_buf->th_doacross_info[last++] = range_length;
3926     pr_buf->th_doacross_info[last++] = dims[j].lo;
3927     pr_buf->th_doacross_info[last++] = dims[j].up;
3928     pr_buf->th_doacross_info[last++] = dims[j].st;
3929   }
3930 
3931   // Compute total trip count.
3932   // Start with range of dims[0] which we don't need to keep in the buffer.
3933   if (dims[0].st == 1) { // most common case
3934     trace_count = dims[0].up - dims[0].lo + 1;
3935   } else if (dims[0].st > 0) {
3936     KMP_DEBUG_ASSERT(dims[0].up > dims[0].lo);
3937     trace_count = (kmp_uint64)(dims[0].up - dims[0].lo) / dims[0].st + 1;
3938   } else { // negative increment
3939     KMP_DEBUG_ASSERT(dims[0].lo > dims[0].up);
3940     trace_count = (kmp_uint64)(dims[0].lo - dims[0].up) / (-dims[0].st) + 1;
3941   }
3942   for (j = 1; j < num_dims; ++j) {
3943     trace_count *= pr_buf->th_doacross_info[4 * j + 1]; // use kept ranges
3944   }
3945   KMP_DEBUG_ASSERT(trace_count > 0);
3946 
3947   // Check if shared buffer is not occupied by other loop (idx -
3948   // __kmp_dispatch_num_buffers)
3949   if (idx != sh_buf->doacross_buf_idx) {
3950     // Shared buffer is occupied, wait for it to be free
3951     __kmp_wait_4((volatile kmp_uint32 *)&sh_buf->doacross_buf_idx, idx,
3952                  __kmp_eq_4, NULL);
3953   }
3954 #if KMP_32_BIT_ARCH
3955   // Check if we are the first thread. After the CAS the first thread gets 0,
3956   // others get 1 if initialization is in progress, allocated pointer otherwise.
3957   // Treat pointer as volatile integer (value 0 or 1) until memory is allocated.
3958   flags = (kmp_uint32 *)KMP_COMPARE_AND_STORE_RET32(
3959       (volatile kmp_int32 *)&sh_buf->doacross_flags, NULL, 1);
3960 #else
3961   flags = (kmp_uint32 *)KMP_COMPARE_AND_STORE_RET64(
3962       (volatile kmp_int64 *)&sh_buf->doacross_flags, NULL, 1LL);
3963 #endif
3964   if (flags == NULL) {
3965     // we are the first thread, allocate the array of flags
3966     size_t size = trace_count / 8 + 8; // in bytes, use single bit per iteration
3967     flags = (kmp_uint32 *)__kmp_thread_calloc(th, size, 1);
3968     KMP_MB();
3969     sh_buf->doacross_flags = flags;
3970   } else if (flags == (kmp_uint32 *)1) {
3971 #if KMP_32_BIT_ARCH
3972     // initialization is still in progress, need to wait
3973     while (*(volatile kmp_int32 *)&sh_buf->doacross_flags == 1)
3974 #else
3975     while (*(volatile kmp_int64 *)&sh_buf->doacross_flags == 1LL)
3976 #endif
3977       KMP_YIELD(TRUE);
3978     KMP_MB();
3979   } else {
3980     KMP_MB();
3981   }
3982   KMP_DEBUG_ASSERT(sh_buf->doacross_flags > (kmp_uint32 *)1); // check ptr value
3983   pr_buf->th_doacross_flags =
3984       sh_buf->doacross_flags; // save private copy in order to not
3985   // touch shared buffer on each iteration
3986   KA_TRACE(20, ("__kmpc_doacross_init() exit: T#%d\n", gtid));
3987 }
3988 
3989 void __kmpc_doacross_wait(ident_t *loc, int gtid, const kmp_int64 *vec) {
3990   __kmp_assert_valid_gtid(gtid);
3991   kmp_int32 shft, num_dims, i;
3992   kmp_uint32 flag;
3993   kmp_int64 iter_number; // iteration number of "collapsed" loop nest
3994   kmp_info_t *th = __kmp_threads[gtid];
3995   kmp_team_t *team = th->th.th_team;
3996   kmp_disp_t *pr_buf;
3997   kmp_int64 lo, up, st;
3998 
3999   KA_TRACE(20, ("__kmpc_doacross_wait() enter: called T#%d\n", gtid));
4000   if (team->t.t_serialized) {
4001     KA_TRACE(20, ("__kmpc_doacross_wait() exit: serialized team\n"));
4002     return; // no dependencies if team is serialized
4003   }
4004 
4005   // calculate sequential iteration number and check out-of-bounds condition
4006   pr_buf = th->th.th_dispatch;
4007   KMP_DEBUG_ASSERT(pr_buf->th_doacross_info != NULL);
4008   num_dims = pr_buf->th_doacross_info[0];
4009   lo = pr_buf->th_doacross_info[2];
4010   up = pr_buf->th_doacross_info[3];
4011   st = pr_buf->th_doacross_info[4];
4012 #if OMPT_SUPPORT && OMPT_OPTIONAL
4013   ompt_dependence_t deps[num_dims];
4014 #endif
4015   if (st == 1) { // most common case
4016     if (vec[0] < lo || vec[0] > up) {
4017       KA_TRACE(20, ("__kmpc_doacross_wait() exit: T#%d iter %lld is out of "
4018                     "bounds [%lld,%lld]\n",
4019                     gtid, vec[0], lo, up));
4020       return;
4021     }
4022     iter_number = vec[0] - lo;
4023   } else if (st > 0) {
4024     if (vec[0] < lo || vec[0] > up) {
4025       KA_TRACE(20, ("__kmpc_doacross_wait() exit: T#%d iter %lld is out of "
4026                     "bounds [%lld,%lld]\n",
4027                     gtid, vec[0], lo, up));
4028       return;
4029     }
4030     iter_number = (kmp_uint64)(vec[0] - lo) / st;
4031   } else { // negative increment
4032     if (vec[0] > lo || vec[0] < up) {
4033       KA_TRACE(20, ("__kmpc_doacross_wait() exit: T#%d iter %lld is out of "
4034                     "bounds [%lld,%lld]\n",
4035                     gtid, vec[0], lo, up));
4036       return;
4037     }
4038     iter_number = (kmp_uint64)(lo - vec[0]) / (-st);
4039   }
4040 #if OMPT_SUPPORT && OMPT_OPTIONAL
4041   deps[0].variable.value = iter_number;
4042   deps[0].dependence_type = ompt_dependence_type_sink;
4043 #endif
4044   for (i = 1; i < num_dims; ++i) {
4045     kmp_int64 iter, ln;
4046     kmp_int32 j = i * 4;
4047     ln = pr_buf->th_doacross_info[j + 1];
4048     lo = pr_buf->th_doacross_info[j + 2];
4049     up = pr_buf->th_doacross_info[j + 3];
4050     st = pr_buf->th_doacross_info[j + 4];
4051     if (st == 1) {
4052       if (vec[i] < lo || vec[i] > up) {
4053         KA_TRACE(20, ("__kmpc_doacross_wait() exit: T#%d iter %lld is out of "
4054                       "bounds [%lld,%lld]\n",
4055                       gtid, vec[i], lo, up));
4056         return;
4057       }
4058       iter = vec[i] - lo;
4059     } else if (st > 0) {
4060       if (vec[i] < lo || vec[i] > up) {
4061         KA_TRACE(20, ("__kmpc_doacross_wait() exit: T#%d iter %lld is out of "
4062                       "bounds [%lld,%lld]\n",
4063                       gtid, vec[i], lo, up));
4064         return;
4065       }
4066       iter = (kmp_uint64)(vec[i] - lo) / st;
4067     } else { // st < 0
4068       if (vec[i] > lo || vec[i] < up) {
4069         KA_TRACE(20, ("__kmpc_doacross_wait() exit: T#%d iter %lld is out of "
4070                       "bounds [%lld,%lld]\n",
4071                       gtid, vec[i], lo, up));
4072         return;
4073       }
4074       iter = (kmp_uint64)(lo - vec[i]) / (-st);
4075     }
4076     iter_number = iter + ln * iter_number;
4077 #if OMPT_SUPPORT && OMPT_OPTIONAL
4078     deps[i].variable.value = iter;
4079     deps[i].dependence_type = ompt_dependence_type_sink;
4080 #endif
4081   }
4082   shft = iter_number % 32; // use 32-bit granularity
4083   iter_number >>= 5; // divided by 32
4084   flag = 1 << shft;
4085   while ((flag & pr_buf->th_doacross_flags[iter_number]) == 0) {
4086     KMP_YIELD(TRUE);
4087   }
4088   KMP_MB();
4089 #if OMPT_SUPPORT && OMPT_OPTIONAL
4090   if (ompt_enabled.ompt_callback_dependences) {
4091     ompt_callbacks.ompt_callback(ompt_callback_dependences)(
4092         &(OMPT_CUR_TASK_INFO(th)->task_data), deps, num_dims);
4093   }
4094 #endif
4095   KA_TRACE(20,
4096            ("__kmpc_doacross_wait() exit: T#%d wait for iter %lld completed\n",
4097             gtid, (iter_number << 5) + shft));
4098 }
4099 
4100 void __kmpc_doacross_post(ident_t *loc, int gtid, const kmp_int64 *vec) {
4101   __kmp_assert_valid_gtid(gtid);
4102   kmp_int32 shft, num_dims, i;
4103   kmp_uint32 flag;
4104   kmp_int64 iter_number; // iteration number of "collapsed" loop nest
4105   kmp_info_t *th = __kmp_threads[gtid];
4106   kmp_team_t *team = th->th.th_team;
4107   kmp_disp_t *pr_buf;
4108   kmp_int64 lo, st;
4109 
4110   KA_TRACE(20, ("__kmpc_doacross_post() enter: called T#%d\n", gtid));
4111   if (team->t.t_serialized) {
4112     KA_TRACE(20, ("__kmpc_doacross_post() exit: serialized team\n"));
4113     return; // no dependencies if team is serialized
4114   }
4115 
4116   // calculate sequential iteration number (same as in "wait" but no
4117   // out-of-bounds checks)
4118   pr_buf = th->th.th_dispatch;
4119   KMP_DEBUG_ASSERT(pr_buf->th_doacross_info != NULL);
4120   num_dims = pr_buf->th_doacross_info[0];
4121   lo = pr_buf->th_doacross_info[2];
4122   st = pr_buf->th_doacross_info[4];
4123 #if OMPT_SUPPORT && OMPT_OPTIONAL
4124   ompt_dependence_t deps[num_dims];
4125 #endif
4126   if (st == 1) { // most common case
4127     iter_number = vec[0] - lo;
4128   } else if (st > 0) {
4129     iter_number = (kmp_uint64)(vec[0] - lo) / st;
4130   } else { // negative increment
4131     iter_number = (kmp_uint64)(lo - vec[0]) / (-st);
4132   }
4133 #if OMPT_SUPPORT && OMPT_OPTIONAL
4134   deps[0].variable.value = iter_number;
4135   deps[0].dependence_type = ompt_dependence_type_source;
4136 #endif
4137   for (i = 1; i < num_dims; ++i) {
4138     kmp_int64 iter, ln;
4139     kmp_int32 j = i * 4;
4140     ln = pr_buf->th_doacross_info[j + 1];
4141     lo = pr_buf->th_doacross_info[j + 2];
4142     st = pr_buf->th_doacross_info[j + 4];
4143     if (st == 1) {
4144       iter = vec[i] - lo;
4145     } else if (st > 0) {
4146       iter = (kmp_uint64)(vec[i] - lo) / st;
4147     } else { // st < 0
4148       iter = (kmp_uint64)(lo - vec[i]) / (-st);
4149     }
4150     iter_number = iter + ln * iter_number;
4151 #if OMPT_SUPPORT && OMPT_OPTIONAL
4152     deps[i].variable.value = iter;
4153     deps[i].dependence_type = ompt_dependence_type_source;
4154 #endif
4155   }
4156 #if OMPT_SUPPORT && OMPT_OPTIONAL
4157   if (ompt_enabled.ompt_callback_dependences) {
4158     ompt_callbacks.ompt_callback(ompt_callback_dependences)(
4159         &(OMPT_CUR_TASK_INFO(th)->task_data), deps, num_dims);
4160   }
4161 #endif
4162   shft = iter_number % 32; // use 32-bit granularity
4163   iter_number >>= 5; // divided by 32
4164   flag = 1 << shft;
4165   KMP_MB();
4166   if ((flag & pr_buf->th_doacross_flags[iter_number]) == 0)
4167     KMP_TEST_THEN_OR32(&pr_buf->th_doacross_flags[iter_number], flag);
4168   KA_TRACE(20, ("__kmpc_doacross_post() exit: T#%d iter %lld posted\n", gtid,
4169                 (iter_number << 5) + shft));
4170 }
4171 
4172 void __kmpc_doacross_fini(ident_t *loc, int gtid) {
4173   __kmp_assert_valid_gtid(gtid);
4174   kmp_int32 num_done;
4175   kmp_info_t *th = __kmp_threads[gtid];
4176   kmp_team_t *team = th->th.th_team;
4177   kmp_disp_t *pr_buf = th->th.th_dispatch;
4178 
4179   KA_TRACE(20, ("__kmpc_doacross_fini() enter: called T#%d\n", gtid));
4180   if (team->t.t_serialized) {
4181     KA_TRACE(20, ("__kmpc_doacross_fini() exit: serialized team %p\n", team));
4182     return; // nothing to do
4183   }
4184   num_done = KMP_TEST_THEN_INC32((kmp_int32 *)pr_buf->th_doacross_info[1]) + 1;
4185   if (num_done == th->th.th_team_nproc) {
4186     // we are the last thread, need to free shared resources
4187     int idx = pr_buf->th_doacross_buf_idx - 1;
4188     dispatch_shared_info_t *sh_buf =
4189         &team->t.t_disp_buffer[idx % __kmp_dispatch_num_buffers];
4190     KMP_DEBUG_ASSERT(pr_buf->th_doacross_info[1] ==
4191                      (kmp_int64)&sh_buf->doacross_num_done);
4192     KMP_DEBUG_ASSERT(num_done == sh_buf->doacross_num_done);
4193     KMP_DEBUG_ASSERT(idx == sh_buf->doacross_buf_idx);
4194     __kmp_thread_free(th, CCAST(kmp_uint32 *, sh_buf->doacross_flags));
4195     sh_buf->doacross_flags = NULL;
4196     sh_buf->doacross_num_done = 0;
4197     sh_buf->doacross_buf_idx +=
4198         __kmp_dispatch_num_buffers; // free buffer for future re-use
4199   }
4200   // free private resources (need to keep buffer index forever)
4201   pr_buf->th_doacross_flags = NULL;
4202   __kmp_thread_free(th, (void *)pr_buf->th_doacross_info);
4203   pr_buf->th_doacross_info = NULL;
4204   KA_TRACE(20, ("__kmpc_doacross_fini() exit: T#%d\n", gtid));
4205 }
4206 
4207 /* omp_alloc/omp_calloc/omp_free only defined for C/C++, not for Fortran */
4208 void *omp_alloc(size_t size, omp_allocator_handle_t allocator) {
4209   return __kmpc_alloc(__kmp_entry_gtid(), size, allocator);
4210 }
4211 
4212 void *omp_calloc(size_t nmemb, size_t size, omp_allocator_handle_t allocator) {
4213   return __kmpc_calloc(__kmp_entry_gtid(), nmemb, size, allocator);
4214 }
4215 
4216 void *omp_realloc(void *ptr, size_t size, omp_allocator_handle_t allocator,
4217                   omp_allocator_handle_t free_allocator) {
4218   return __kmpc_realloc(__kmp_entry_gtid(), ptr, size, allocator,
4219                         free_allocator);
4220 }
4221 
4222 void omp_free(void *ptr, omp_allocator_handle_t allocator) {
4223   __kmpc_free(__kmp_entry_gtid(), ptr, allocator);
4224 }
4225 
4226 int __kmpc_get_target_offload(void) {
4227   if (!__kmp_init_serial) {
4228     __kmp_serial_initialize();
4229   }
4230   return __kmp_target_offload;
4231 }
4232 
4233 int __kmpc_pause_resource(kmp_pause_status_t level) {
4234   if (!__kmp_init_serial) {
4235     return 1; // Can't pause if runtime is not initialized
4236   }
4237   return __kmp_pause_resource(level);
4238 }
4239