1 /*
2  * kmp_runtime.cpp -- KPTS runtime support library
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 #include "kmp.h"
14 #include "kmp_affinity.h"
15 #include "kmp_atomic.h"
16 #include "kmp_environment.h"
17 #include "kmp_error.h"
18 #include "kmp_i18n.h"
19 #include "kmp_io.h"
20 #include "kmp_itt.h"
21 #include "kmp_settings.h"
22 #include "kmp_stats.h"
23 #include "kmp_str.h"
24 #include "kmp_wait_release.h"
25 #include "kmp_wrapper_getpid.h"
26 #include "kmp_dispatch.h"
27 #if KMP_USE_HIER_SCHED
28 #include "kmp_dispatch_hier.h"
29 #endif
30 
31 #if OMPT_SUPPORT
32 #include "ompt-specific.h"
33 #endif
34 
35 #if OMP_PROFILING_SUPPORT
36 #include "llvm/Support/TimeProfiler.h"
37 static char *ProfileTraceFile = nullptr;
38 #endif
39 
40 /* these are temporary issues to be dealt with */
41 #define KMP_USE_PRCTL 0
42 
43 #if KMP_OS_WINDOWS
44 #include <process.h>
45 #endif
46 
47 #include "tsan_annotations.h"
48 
49 #if KMP_OS_WINDOWS
50 // windows does not need include files as it doesn't use shared memory
51 #else
52 #include <sys/mman.h>
53 #include <sys/stat.h>
54 #include <fcntl.h>
55 #define SHM_SIZE 1024
56 #endif
57 
58 #if defined(KMP_GOMP_COMPAT)
59 char const __kmp_version_alt_comp[] =
60     KMP_VERSION_PREFIX "alternative compiler support: yes";
61 #endif /* defined(KMP_GOMP_COMPAT) */
62 
63 char const __kmp_version_omp_api[] =
64     KMP_VERSION_PREFIX "API version: 5.0 (201611)";
65 
66 #ifdef KMP_DEBUG
67 char const __kmp_version_lock[] =
68     KMP_VERSION_PREFIX "lock type: run time selectable";
69 #endif /* KMP_DEBUG */
70 
71 #define KMP_MIN(x, y) ((x) < (y) ? (x) : (y))
72 
73 /* ------------------------------------------------------------------------ */
74 
75 #if KMP_USE_MONITOR
76 kmp_info_t __kmp_monitor;
77 #endif
78 
79 /* Forward declarations */
80 
81 void __kmp_cleanup(void);
82 
83 static void __kmp_initialize_info(kmp_info_t *, kmp_team_t *, int tid,
84                                   int gtid);
85 static void __kmp_initialize_team(kmp_team_t *team, int new_nproc,
86                                   kmp_internal_control_t *new_icvs,
87                                   ident_t *loc);
88 #if KMP_AFFINITY_SUPPORTED
89 static void __kmp_partition_places(kmp_team_t *team,
90                                    int update_master_only = 0);
91 #endif
92 static void __kmp_do_serial_initialize(void);
93 void __kmp_fork_barrier(int gtid, int tid);
94 void __kmp_join_barrier(int gtid);
95 void __kmp_setup_icv_copy(kmp_team_t *team, int new_nproc,
96                           kmp_internal_control_t *new_icvs, ident_t *loc);
97 
98 #ifdef USE_LOAD_BALANCE
99 static int __kmp_load_balance_nproc(kmp_root_t *root, int set_nproc);
100 #endif
101 
102 static int __kmp_expand_threads(int nNeed);
103 #if KMP_OS_WINDOWS
104 static int __kmp_unregister_root_other_thread(int gtid);
105 #endif
106 static void __kmp_reap_thread(kmp_info_t *thread, int is_root);
107 kmp_info_t *__kmp_thread_pool_insert_pt = NULL;
108 
109 /* Calculate the identifier of the current thread */
110 /* fast (and somewhat portable) way to get unique identifier of executing
111    thread. Returns KMP_GTID_DNE if we haven't been assigned a gtid. */
112 int __kmp_get_global_thread_id() {
113   int i;
114   kmp_info_t **other_threads;
115   size_t stack_data;
116   char *stack_addr;
117   size_t stack_size;
118   char *stack_base;
119 
120   KA_TRACE(
121       1000,
122       ("*** __kmp_get_global_thread_id: entering, nproc=%d  all_nproc=%d\n",
123        __kmp_nth, __kmp_all_nth));
124 
125   /* JPH - to handle the case where __kmpc_end(0) is called immediately prior to
126      a parallel region, made it return KMP_GTID_DNE to force serial_initialize
127      by caller. Had to handle KMP_GTID_DNE at all call-sites, or else guarantee
128      __kmp_init_gtid for this to work. */
129 
130   if (!TCR_4(__kmp_init_gtid))
131     return KMP_GTID_DNE;
132 
133 #ifdef KMP_TDATA_GTID
134   if (TCR_4(__kmp_gtid_mode) >= 3) {
135     KA_TRACE(1000, ("*** __kmp_get_global_thread_id: using TDATA\n"));
136     return __kmp_gtid;
137   }
138 #endif
139   if (TCR_4(__kmp_gtid_mode) >= 2) {
140     KA_TRACE(1000, ("*** __kmp_get_global_thread_id: using keyed TLS\n"));
141     return __kmp_gtid_get_specific();
142   }
143   KA_TRACE(1000, ("*** __kmp_get_global_thread_id: using internal alg.\n"));
144 
145   stack_addr = (char *)&stack_data;
146   other_threads = __kmp_threads;
147 
148   /* ATT: The code below is a source of potential bugs due to unsynchronized
149      access to __kmp_threads array. For example:
150      1. Current thread loads other_threads[i] to thr and checks it, it is
151         non-NULL.
152      2. Current thread is suspended by OS.
153      3. Another thread unregisters and finishes (debug versions of free()
154         may fill memory with something like 0xEF).
155      4. Current thread is resumed.
156      5. Current thread reads junk from *thr.
157      TODO: Fix it.  --ln  */
158 
159   for (i = 0; i < __kmp_threads_capacity; i++) {
160 
161     kmp_info_t *thr = (kmp_info_t *)TCR_SYNC_PTR(other_threads[i]);
162     if (!thr)
163       continue;
164 
165     stack_size = (size_t)TCR_PTR(thr->th.th_info.ds.ds_stacksize);
166     stack_base = (char *)TCR_PTR(thr->th.th_info.ds.ds_stackbase);
167 
168     /* stack grows down -- search through all of the active threads */
169 
170     if (stack_addr <= stack_base) {
171       size_t stack_diff = stack_base - stack_addr;
172 
173       if (stack_diff <= stack_size) {
174         /* The only way we can be closer than the allocated */
175         /* stack size is if we are running on this thread. */
176         KMP_DEBUG_ASSERT(__kmp_gtid_get_specific() == i);
177         return i;
178       }
179     }
180   }
181 
182   /* get specific to try and determine our gtid */
183   KA_TRACE(1000,
184            ("*** __kmp_get_global_thread_id: internal alg. failed to find "
185             "thread, using TLS\n"));
186   i = __kmp_gtid_get_specific();
187 
188   /*fprintf( stderr, "=== %d\n", i );  */ /* GROO */
189 
190   /* if we havn't been assigned a gtid, then return code */
191   if (i < 0)
192     return i;
193 
194   /* dynamically updated stack window for uber threads to avoid get_specific
195      call */
196   if (!TCR_4(other_threads[i]->th.th_info.ds.ds_stackgrow)) {
197     KMP_FATAL(StackOverflow, i);
198   }
199 
200   stack_base = (char *)other_threads[i]->th.th_info.ds.ds_stackbase;
201   if (stack_addr > stack_base) {
202     TCW_PTR(other_threads[i]->th.th_info.ds.ds_stackbase, stack_addr);
203     TCW_PTR(other_threads[i]->th.th_info.ds.ds_stacksize,
204             other_threads[i]->th.th_info.ds.ds_stacksize + stack_addr -
205                 stack_base);
206   } else {
207     TCW_PTR(other_threads[i]->th.th_info.ds.ds_stacksize,
208             stack_base - stack_addr);
209   }
210 
211   /* Reprint stack bounds for ubermaster since they have been refined */
212   if (__kmp_storage_map) {
213     char *stack_end = (char *)other_threads[i]->th.th_info.ds.ds_stackbase;
214     char *stack_beg = stack_end - other_threads[i]->th.th_info.ds.ds_stacksize;
215     __kmp_print_storage_map_gtid(i, stack_beg, stack_end,
216                                  other_threads[i]->th.th_info.ds.ds_stacksize,
217                                  "th_%d stack (refinement)", i);
218   }
219   return i;
220 }
221 
222 int __kmp_get_global_thread_id_reg() {
223   int gtid;
224 
225   if (!__kmp_init_serial) {
226     gtid = KMP_GTID_DNE;
227   } else
228 #ifdef KMP_TDATA_GTID
229       if (TCR_4(__kmp_gtid_mode) >= 3) {
230     KA_TRACE(1000, ("*** __kmp_get_global_thread_id_reg: using TDATA\n"));
231     gtid = __kmp_gtid;
232   } else
233 #endif
234       if (TCR_4(__kmp_gtid_mode) >= 2) {
235     KA_TRACE(1000, ("*** __kmp_get_global_thread_id_reg: using keyed TLS\n"));
236     gtid = __kmp_gtid_get_specific();
237   } else {
238     KA_TRACE(1000,
239              ("*** __kmp_get_global_thread_id_reg: using internal alg.\n"));
240     gtid = __kmp_get_global_thread_id();
241   }
242 
243   /* we must be a new uber master sibling thread */
244   if (gtid == KMP_GTID_DNE) {
245     KA_TRACE(10,
246              ("__kmp_get_global_thread_id_reg: Encountered new root thread. "
247               "Registering a new gtid.\n"));
248     __kmp_acquire_bootstrap_lock(&__kmp_initz_lock);
249     if (!__kmp_init_serial) {
250       __kmp_do_serial_initialize();
251       gtid = __kmp_gtid_get_specific();
252     } else {
253       gtid = __kmp_register_root(FALSE);
254     }
255     __kmp_release_bootstrap_lock(&__kmp_initz_lock);
256     /*__kmp_printf( "+++ %d\n", gtid ); */ /* GROO */
257   }
258 
259   KMP_DEBUG_ASSERT(gtid >= 0);
260 
261   return gtid;
262 }
263 
264 /* caller must hold forkjoin_lock */
265 void __kmp_check_stack_overlap(kmp_info_t *th) {
266   int f;
267   char *stack_beg = NULL;
268   char *stack_end = NULL;
269   int gtid;
270 
271   KA_TRACE(10, ("__kmp_check_stack_overlap: called\n"));
272   if (__kmp_storage_map) {
273     stack_end = (char *)th->th.th_info.ds.ds_stackbase;
274     stack_beg = stack_end - th->th.th_info.ds.ds_stacksize;
275 
276     gtid = __kmp_gtid_from_thread(th);
277 
278     if (gtid == KMP_GTID_MONITOR) {
279       __kmp_print_storage_map_gtid(
280           gtid, stack_beg, stack_end, th->th.th_info.ds.ds_stacksize,
281           "th_%s stack (%s)", "mon",
282           (th->th.th_info.ds.ds_stackgrow) ? "initial" : "actual");
283     } else {
284       __kmp_print_storage_map_gtid(
285           gtid, stack_beg, stack_end, th->th.th_info.ds.ds_stacksize,
286           "th_%d stack (%s)", gtid,
287           (th->th.th_info.ds.ds_stackgrow) ? "initial" : "actual");
288     }
289   }
290 
291   /* No point in checking ubermaster threads since they use refinement and
292    * cannot overlap */
293   gtid = __kmp_gtid_from_thread(th);
294   if (__kmp_env_checks == TRUE && !KMP_UBER_GTID(gtid)) {
295     KA_TRACE(10,
296              ("__kmp_check_stack_overlap: performing extensive checking\n"));
297     if (stack_beg == NULL) {
298       stack_end = (char *)th->th.th_info.ds.ds_stackbase;
299       stack_beg = stack_end - th->th.th_info.ds.ds_stacksize;
300     }
301 
302     for (f = 0; f < __kmp_threads_capacity; f++) {
303       kmp_info_t *f_th = (kmp_info_t *)TCR_SYNC_PTR(__kmp_threads[f]);
304 
305       if (f_th && f_th != th) {
306         char *other_stack_end =
307             (char *)TCR_PTR(f_th->th.th_info.ds.ds_stackbase);
308         char *other_stack_beg =
309             other_stack_end - (size_t)TCR_PTR(f_th->th.th_info.ds.ds_stacksize);
310         if ((stack_beg > other_stack_beg && stack_beg < other_stack_end) ||
311             (stack_end > other_stack_beg && stack_end < other_stack_end)) {
312 
313           /* Print the other stack values before the abort */
314           if (__kmp_storage_map)
315             __kmp_print_storage_map_gtid(
316                 -1, other_stack_beg, other_stack_end,
317                 (size_t)TCR_PTR(f_th->th.th_info.ds.ds_stacksize),
318                 "th_%d stack (overlapped)", __kmp_gtid_from_thread(f_th));
319 
320           __kmp_fatal(KMP_MSG(StackOverlap), KMP_HNT(ChangeStackLimit),
321                       __kmp_msg_null);
322         }
323       }
324     }
325   }
326   KA_TRACE(10, ("__kmp_check_stack_overlap: returning\n"));
327 }
328 
329 /* ------------------------------------------------------------------------ */
330 
331 void __kmp_infinite_loop(void) {
332   static int done = FALSE;
333 
334   while (!done) {
335     KMP_YIELD(TRUE);
336   }
337 }
338 
339 #define MAX_MESSAGE 512
340 
341 void __kmp_print_storage_map_gtid(int gtid, void *p1, void *p2, size_t size,
342                                   char const *format, ...) {
343   char buffer[MAX_MESSAGE];
344   va_list ap;
345 
346   va_start(ap, format);
347   KMP_SNPRINTF(buffer, sizeof(buffer), "OMP storage map: %p %p%8lu %s\n", p1,
348                p2, (unsigned long)size, format);
349   __kmp_acquire_bootstrap_lock(&__kmp_stdio_lock);
350   __kmp_vprintf(kmp_err, buffer, ap);
351 #if KMP_PRINT_DATA_PLACEMENT
352   int node;
353   if (gtid >= 0) {
354     if (p1 <= p2 && (char *)p2 - (char *)p1 == size) {
355       if (__kmp_storage_map_verbose) {
356         node = __kmp_get_host_node(p1);
357         if (node < 0) /* doesn't work, so don't try this next time */
358           __kmp_storage_map_verbose = FALSE;
359         else {
360           char *last;
361           int lastNode;
362           int localProc = __kmp_get_cpu_from_gtid(gtid);
363 
364           const int page_size = KMP_GET_PAGE_SIZE();
365 
366           p1 = (void *)((size_t)p1 & ~((size_t)page_size - 1));
367           p2 = (void *)(((size_t)p2 - 1) & ~((size_t)page_size - 1));
368           if (localProc >= 0)
369             __kmp_printf_no_lock("  GTID %d localNode %d\n", gtid,
370                                  localProc >> 1);
371           else
372             __kmp_printf_no_lock("  GTID %d\n", gtid);
373 #if KMP_USE_PRCTL
374           /* The more elaborate format is disabled for now because of the prctl
375            * hanging bug. */
376           do {
377             last = p1;
378             lastNode = node;
379             /* This loop collates adjacent pages with the same host node. */
380             do {
381               (char *)p1 += page_size;
382             } while (p1 <= p2 && (node = __kmp_get_host_node(p1)) == lastNode);
383             __kmp_printf_no_lock("    %p-%p memNode %d\n", last, (char *)p1 - 1,
384                                  lastNode);
385           } while (p1 <= p2);
386 #else
387           __kmp_printf_no_lock("    %p-%p memNode %d\n", p1,
388                                (char *)p1 + (page_size - 1),
389                                __kmp_get_host_node(p1));
390           if (p1 < p2) {
391             __kmp_printf_no_lock("    %p-%p memNode %d\n", p2,
392                                  (char *)p2 + (page_size - 1),
393                                  __kmp_get_host_node(p2));
394           }
395 #endif
396         }
397       }
398     } else
399       __kmp_printf_no_lock("  %s\n", KMP_I18N_STR(StorageMapWarning));
400   }
401 #endif /* KMP_PRINT_DATA_PLACEMENT */
402   __kmp_release_bootstrap_lock(&__kmp_stdio_lock);
403 }
404 
405 void __kmp_warn(char const *format, ...) {
406   char buffer[MAX_MESSAGE];
407   va_list ap;
408 
409   if (__kmp_generate_warnings == kmp_warnings_off) {
410     return;
411   }
412 
413   va_start(ap, format);
414 
415   KMP_SNPRINTF(buffer, sizeof(buffer), "OMP warning: %s\n", format);
416   __kmp_acquire_bootstrap_lock(&__kmp_stdio_lock);
417   __kmp_vprintf(kmp_err, buffer, ap);
418   __kmp_release_bootstrap_lock(&__kmp_stdio_lock);
419 
420   va_end(ap);
421 }
422 
423 void __kmp_abort_process() {
424   // Later threads may stall here, but that's ok because abort() will kill them.
425   __kmp_acquire_bootstrap_lock(&__kmp_exit_lock);
426 
427   if (__kmp_debug_buf) {
428     __kmp_dump_debug_buffer();
429   }
430 
431   if (KMP_OS_WINDOWS) {
432     // Let other threads know of abnormal termination and prevent deadlock
433     // if abort happened during library initialization or shutdown
434     __kmp_global.g.g_abort = SIGABRT;
435 
436     /* On Windows* OS by default abort() causes pop-up error box, which stalls
437        nightly testing. Unfortunately, we cannot reliably suppress pop-up error
438        boxes. _set_abort_behavior() works well, but this function is not
439        available in VS7 (this is not problem for DLL, but it is a problem for
440        static OpenMP RTL). SetErrorMode (and so, timelimit utility) does not
441        help, at least in some versions of MS C RTL.
442 
443        It seems following sequence is the only way to simulate abort() and
444        avoid pop-up error box. */
445     raise(SIGABRT);
446     _exit(3); // Just in case, if signal ignored, exit anyway.
447   } else {
448     __kmp_unregister_library();
449     abort();
450   }
451 
452   __kmp_infinite_loop();
453   __kmp_release_bootstrap_lock(&__kmp_exit_lock);
454 
455 } // __kmp_abort_process
456 
457 void __kmp_abort_thread(void) {
458   // TODO: Eliminate g_abort global variable and this function.
459   // In case of abort just call abort(), it will kill all the threads.
460   __kmp_infinite_loop();
461 } // __kmp_abort_thread
462 
463 /* Print out the storage map for the major kmp_info_t thread data structures
464    that are allocated together. */
465 
466 static void __kmp_print_thread_storage_map(kmp_info_t *thr, int gtid) {
467   __kmp_print_storage_map_gtid(gtid, thr, thr + 1, sizeof(kmp_info_t), "th_%d",
468                                gtid);
469 
470   __kmp_print_storage_map_gtid(gtid, &thr->th.th_info, &thr->th.th_team,
471                                sizeof(kmp_desc_t), "th_%d.th_info", gtid);
472 
473   __kmp_print_storage_map_gtid(gtid, &thr->th.th_local, &thr->th.th_pri_head,
474                                sizeof(kmp_local_t), "th_%d.th_local", gtid);
475 
476   __kmp_print_storage_map_gtid(
477       gtid, &thr->th.th_bar[0], &thr->th.th_bar[bs_last_barrier],
478       sizeof(kmp_balign_t) * bs_last_barrier, "th_%d.th_bar", gtid);
479 
480   __kmp_print_storage_map_gtid(gtid, &thr->th.th_bar[bs_plain_barrier],
481                                &thr->th.th_bar[bs_plain_barrier + 1],
482                                sizeof(kmp_balign_t), "th_%d.th_bar[plain]",
483                                gtid);
484 
485   __kmp_print_storage_map_gtid(gtid, &thr->th.th_bar[bs_forkjoin_barrier],
486                                &thr->th.th_bar[bs_forkjoin_barrier + 1],
487                                sizeof(kmp_balign_t), "th_%d.th_bar[forkjoin]",
488                                gtid);
489 
490 #if KMP_FAST_REDUCTION_BARRIER
491   __kmp_print_storage_map_gtid(gtid, &thr->th.th_bar[bs_reduction_barrier],
492                                &thr->th.th_bar[bs_reduction_barrier + 1],
493                                sizeof(kmp_balign_t), "th_%d.th_bar[reduction]",
494                                gtid);
495 #endif // KMP_FAST_REDUCTION_BARRIER
496 }
497 
498 /* Print out the storage map for the major kmp_team_t team data structures
499    that are allocated together. */
500 
501 static void __kmp_print_team_storage_map(const char *header, kmp_team_t *team,
502                                          int team_id, int num_thr) {
503   int num_disp_buff = team->t.t_max_nproc > 1 ? __kmp_dispatch_num_buffers : 2;
504   __kmp_print_storage_map_gtid(-1, team, team + 1, sizeof(kmp_team_t), "%s_%d",
505                                header, team_id);
506 
507   __kmp_print_storage_map_gtid(-1, &team->t.t_bar[0],
508                                &team->t.t_bar[bs_last_barrier],
509                                sizeof(kmp_balign_team_t) * bs_last_barrier,
510                                "%s_%d.t_bar", header, team_id);
511 
512   __kmp_print_storage_map_gtid(-1, &team->t.t_bar[bs_plain_barrier],
513                                &team->t.t_bar[bs_plain_barrier + 1],
514                                sizeof(kmp_balign_team_t), "%s_%d.t_bar[plain]",
515                                header, team_id);
516 
517   __kmp_print_storage_map_gtid(-1, &team->t.t_bar[bs_forkjoin_barrier],
518                                &team->t.t_bar[bs_forkjoin_barrier + 1],
519                                sizeof(kmp_balign_team_t),
520                                "%s_%d.t_bar[forkjoin]", header, team_id);
521 
522 #if KMP_FAST_REDUCTION_BARRIER
523   __kmp_print_storage_map_gtid(-1, &team->t.t_bar[bs_reduction_barrier],
524                                &team->t.t_bar[bs_reduction_barrier + 1],
525                                sizeof(kmp_balign_team_t),
526                                "%s_%d.t_bar[reduction]", header, team_id);
527 #endif // KMP_FAST_REDUCTION_BARRIER
528 
529   __kmp_print_storage_map_gtid(
530       -1, &team->t.t_dispatch[0], &team->t.t_dispatch[num_thr],
531       sizeof(kmp_disp_t) * num_thr, "%s_%d.t_dispatch", header, team_id);
532 
533   __kmp_print_storage_map_gtid(
534       -1, &team->t.t_threads[0], &team->t.t_threads[num_thr],
535       sizeof(kmp_info_t *) * num_thr, "%s_%d.t_threads", header, team_id);
536 
537   __kmp_print_storage_map_gtid(-1, &team->t.t_disp_buffer[0],
538                                &team->t.t_disp_buffer[num_disp_buff],
539                                sizeof(dispatch_shared_info_t) * num_disp_buff,
540                                "%s_%d.t_disp_buffer", header, team_id);
541 }
542 
543 static void __kmp_init_allocator() {
544   __kmp_init_memkind();
545   __kmp_init_target_mem();
546 }
547 static void __kmp_fini_allocator() { __kmp_fini_memkind(); }
548 
549 /* ------------------------------------------------------------------------ */
550 
551 #if KMP_DYNAMIC_LIB
552 #if KMP_OS_WINDOWS
553 
554 BOOL WINAPI DllMain(HINSTANCE hInstDLL, DWORD fdwReason, LPVOID lpReserved) {
555   //__kmp_acquire_bootstrap_lock( &__kmp_initz_lock );
556 
557   switch (fdwReason) {
558 
559   case DLL_PROCESS_ATTACH:
560     KA_TRACE(10, ("DllMain: PROCESS_ATTACH\n"));
561 
562     return TRUE;
563 
564   case DLL_PROCESS_DETACH:
565     KA_TRACE(10, ("DllMain: PROCESS_DETACH T#%d\n", __kmp_gtid_get_specific()));
566 
567     // According to Windows* documentation for DllMain entry point:
568     // for DLL_PROCESS_DETACH, lpReserved is used for telling the difference:
569     //   lpReserved == NULL when FreeLibrary() is called,
570     //   lpReserved != NULL when the process is terminated.
571     // When FreeLibrary() is called, worker threads remain alive. So the
572     // runtime's state is consistent and executing proper shutdown is OK.
573     // When the process is terminated, worker threads have exited or been
574     // forcefully terminated by the OS and only the shutdown thread remains.
575     // This can leave the runtime in an inconsistent state.
576     // Hence, only attempt proper cleanup when FreeLibrary() is called.
577     // Otherwise, rely on OS to reclaim resources.
578     if (lpReserved == NULL)
579       __kmp_internal_end_library(__kmp_gtid_get_specific());
580 
581     return TRUE;
582 
583   case DLL_THREAD_ATTACH:
584     KA_TRACE(10, ("DllMain: THREAD_ATTACH\n"));
585 
586     /* if we want to register new siblings all the time here call
587      * __kmp_get_gtid(); */
588     return TRUE;
589 
590   case DLL_THREAD_DETACH:
591     KA_TRACE(10, ("DllMain: THREAD_DETACH T#%d\n", __kmp_gtid_get_specific()));
592 
593     __kmp_internal_end_thread(__kmp_gtid_get_specific());
594     return TRUE;
595   }
596 
597   return TRUE;
598 }
599 
600 #endif /* KMP_OS_WINDOWS */
601 #endif /* KMP_DYNAMIC_LIB */
602 
603 /* __kmp_parallel_deo -- Wait until it's our turn. */
604 void __kmp_parallel_deo(int *gtid_ref, int *cid_ref, ident_t *loc_ref) {
605   int gtid = *gtid_ref;
606 #ifdef BUILD_PARALLEL_ORDERED
607   kmp_team_t *team = __kmp_team_from_gtid(gtid);
608 #endif /* BUILD_PARALLEL_ORDERED */
609 
610   if (__kmp_env_consistency_check) {
611     if (__kmp_threads[gtid]->th.th_root->r.r_active)
612 #if KMP_USE_DYNAMIC_LOCK
613       __kmp_push_sync(gtid, ct_ordered_in_parallel, loc_ref, NULL, 0);
614 #else
615       __kmp_push_sync(gtid, ct_ordered_in_parallel, loc_ref, NULL);
616 #endif
617   }
618 #ifdef BUILD_PARALLEL_ORDERED
619   if (!team->t.t_serialized) {
620     KMP_MB();
621     KMP_WAIT(&team->t.t_ordered.dt.t_value, __kmp_tid_from_gtid(gtid), KMP_EQ,
622              NULL);
623     KMP_MB();
624   }
625 #endif /* BUILD_PARALLEL_ORDERED */
626 }
627 
628 /* __kmp_parallel_dxo -- Signal the next task. */
629 void __kmp_parallel_dxo(int *gtid_ref, int *cid_ref, ident_t *loc_ref) {
630   int gtid = *gtid_ref;
631 #ifdef BUILD_PARALLEL_ORDERED
632   int tid = __kmp_tid_from_gtid(gtid);
633   kmp_team_t *team = __kmp_team_from_gtid(gtid);
634 #endif /* BUILD_PARALLEL_ORDERED */
635 
636   if (__kmp_env_consistency_check) {
637     if (__kmp_threads[gtid]->th.th_root->r.r_active)
638       __kmp_pop_sync(gtid, ct_ordered_in_parallel, loc_ref);
639   }
640 #ifdef BUILD_PARALLEL_ORDERED
641   if (!team->t.t_serialized) {
642     KMP_MB(); /* Flush all pending memory write invalidates.  */
643 
644     /* use the tid of the next thread in this team */
645     /* TODO replace with general release procedure */
646     team->t.t_ordered.dt.t_value = ((tid + 1) % team->t.t_nproc);
647 
648     KMP_MB(); /* Flush all pending memory write invalidates.  */
649   }
650 #endif /* BUILD_PARALLEL_ORDERED */
651 }
652 
653 /* ------------------------------------------------------------------------ */
654 /* The BARRIER for a SINGLE process section is always explicit   */
655 
656 int __kmp_enter_single(int gtid, ident_t *id_ref, int push_ws) {
657   int status;
658   kmp_info_t *th;
659   kmp_team_t *team;
660 
661   if (!TCR_4(__kmp_init_parallel))
662     __kmp_parallel_initialize();
663   __kmp_resume_if_soft_paused();
664 
665   th = __kmp_threads[gtid];
666   team = th->th.th_team;
667   status = 0;
668 
669   th->th.th_ident = id_ref;
670 
671   if (team->t.t_serialized) {
672     status = 1;
673   } else {
674     kmp_int32 old_this = th->th.th_local.this_construct;
675 
676     ++th->th.th_local.this_construct;
677     /* try to set team count to thread count--success means thread got the
678        single block */
679     /* TODO: Should this be acquire or release? */
680     if (team->t.t_construct == old_this) {
681       status = __kmp_atomic_compare_store_acq(&team->t.t_construct, old_this,
682                                               th->th.th_local.this_construct);
683     }
684 #if USE_ITT_BUILD
685     if (__itt_metadata_add_ptr && __kmp_forkjoin_frames_mode == 3 &&
686         KMP_MASTER_GTID(gtid) && th->th.th_teams_microtask == NULL &&
687         team->t.t_active_level == 1) {
688       // Only report metadata by primary thread of active team at level 1
689       __kmp_itt_metadata_single(id_ref);
690     }
691 #endif /* USE_ITT_BUILD */
692   }
693 
694   if (__kmp_env_consistency_check) {
695     if (status && push_ws) {
696       __kmp_push_workshare(gtid, ct_psingle, id_ref);
697     } else {
698       __kmp_check_workshare(gtid, ct_psingle, id_ref);
699     }
700   }
701 #if USE_ITT_BUILD
702   if (status) {
703     __kmp_itt_single_start(gtid);
704   }
705 #endif /* USE_ITT_BUILD */
706   return status;
707 }
708 
709 void __kmp_exit_single(int gtid) {
710 #if USE_ITT_BUILD
711   __kmp_itt_single_end(gtid);
712 #endif /* USE_ITT_BUILD */
713   if (__kmp_env_consistency_check)
714     __kmp_pop_workshare(gtid, ct_psingle, NULL);
715 }
716 
717 /* determine if we can go parallel or must use a serialized parallel region and
718  * how many threads we can use
719  * set_nproc is the number of threads requested for the team
720  * returns 0 if we should serialize or only use one thread,
721  * otherwise the number of threads to use
722  * The forkjoin lock is held by the caller. */
723 static int __kmp_reserve_threads(kmp_root_t *root, kmp_team_t *parent_team,
724                                  int master_tid, int set_nthreads,
725                                  int enter_teams) {
726   int capacity;
727   int new_nthreads;
728   KMP_DEBUG_ASSERT(__kmp_init_serial);
729   KMP_DEBUG_ASSERT(root && parent_team);
730   kmp_info_t *this_thr = parent_team->t.t_threads[master_tid];
731 
732   // If dyn-var is set, dynamically adjust the number of desired threads,
733   // according to the method specified by dynamic_mode.
734   new_nthreads = set_nthreads;
735   if (!get__dynamic_2(parent_team, master_tid)) {
736     ;
737   }
738 #ifdef USE_LOAD_BALANCE
739   else if (__kmp_global.g.g_dynamic_mode == dynamic_load_balance) {
740     new_nthreads = __kmp_load_balance_nproc(root, set_nthreads);
741     if (new_nthreads == 1) {
742       KC_TRACE(10, ("__kmp_reserve_threads: T#%d load balance reduced "
743                     "reservation to 1 thread\n",
744                     master_tid));
745       return 1;
746     }
747     if (new_nthreads < set_nthreads) {
748       KC_TRACE(10, ("__kmp_reserve_threads: T#%d load balance reduced "
749                     "reservation to %d threads\n",
750                     master_tid, new_nthreads));
751     }
752   }
753 #endif /* USE_LOAD_BALANCE */
754   else if (__kmp_global.g.g_dynamic_mode == dynamic_thread_limit) {
755     new_nthreads = __kmp_avail_proc - __kmp_nth +
756                    (root->r.r_active ? 1 : root->r.r_hot_team->t.t_nproc);
757     if (new_nthreads <= 1) {
758       KC_TRACE(10, ("__kmp_reserve_threads: T#%d thread limit reduced "
759                     "reservation to 1 thread\n",
760                     master_tid));
761       return 1;
762     }
763     if (new_nthreads < set_nthreads) {
764       KC_TRACE(10, ("__kmp_reserve_threads: T#%d thread limit reduced "
765                     "reservation to %d threads\n",
766                     master_tid, new_nthreads));
767     } else {
768       new_nthreads = set_nthreads;
769     }
770   } else if (__kmp_global.g.g_dynamic_mode == dynamic_random) {
771     if (set_nthreads > 2) {
772       new_nthreads = __kmp_get_random(parent_team->t.t_threads[master_tid]);
773       new_nthreads = (new_nthreads % set_nthreads) + 1;
774       if (new_nthreads == 1) {
775         KC_TRACE(10, ("__kmp_reserve_threads: T#%d dynamic random reduced "
776                       "reservation to 1 thread\n",
777                       master_tid));
778         return 1;
779       }
780       if (new_nthreads < set_nthreads) {
781         KC_TRACE(10, ("__kmp_reserve_threads: T#%d dynamic random reduced "
782                       "reservation to %d threads\n",
783                       master_tid, new_nthreads));
784       }
785     }
786   } else {
787     KMP_ASSERT(0);
788   }
789 
790   // Respect KMP_ALL_THREADS/KMP_DEVICE_THREAD_LIMIT.
791   if (__kmp_nth + new_nthreads -
792           (root->r.r_active ? 1 : root->r.r_hot_team->t.t_nproc) >
793       __kmp_max_nth) {
794     int tl_nthreads = __kmp_max_nth - __kmp_nth +
795                       (root->r.r_active ? 1 : root->r.r_hot_team->t.t_nproc);
796     if (tl_nthreads <= 0) {
797       tl_nthreads = 1;
798     }
799 
800     // If dyn-var is false, emit a 1-time warning.
801     if (!get__dynamic_2(parent_team, master_tid) && (!__kmp_reserve_warn)) {
802       __kmp_reserve_warn = 1;
803       __kmp_msg(kmp_ms_warning,
804                 KMP_MSG(CantFormThrTeam, set_nthreads, tl_nthreads),
805                 KMP_HNT(Unset_ALL_THREADS), __kmp_msg_null);
806     }
807     if (tl_nthreads == 1) {
808       KC_TRACE(10, ("__kmp_reserve_threads: T#%d KMP_DEVICE_THREAD_LIMIT "
809                     "reduced reservation to 1 thread\n",
810                     master_tid));
811       return 1;
812     }
813     KC_TRACE(10, ("__kmp_reserve_threads: T#%d KMP_DEVICE_THREAD_LIMIT reduced "
814                   "reservation to %d threads\n",
815                   master_tid, tl_nthreads));
816     new_nthreads = tl_nthreads;
817   }
818 
819   // Respect OMP_THREAD_LIMIT
820   int cg_nthreads = this_thr->th.th_cg_roots->cg_nthreads;
821   int max_cg_threads = this_thr->th.th_cg_roots->cg_thread_limit;
822   if (cg_nthreads + new_nthreads -
823           (root->r.r_active ? 1 : root->r.r_hot_team->t.t_nproc) >
824       max_cg_threads) {
825     int tl_nthreads = max_cg_threads - cg_nthreads +
826                       (root->r.r_active ? 1 : root->r.r_hot_team->t.t_nproc);
827     if (tl_nthreads <= 0) {
828       tl_nthreads = 1;
829     }
830 
831     // If dyn-var is false, emit a 1-time warning.
832     if (!get__dynamic_2(parent_team, master_tid) && (!__kmp_reserve_warn)) {
833       __kmp_reserve_warn = 1;
834       __kmp_msg(kmp_ms_warning,
835                 KMP_MSG(CantFormThrTeam, set_nthreads, tl_nthreads),
836                 KMP_HNT(Unset_ALL_THREADS), __kmp_msg_null);
837     }
838     if (tl_nthreads == 1) {
839       KC_TRACE(10, ("__kmp_reserve_threads: T#%d OMP_THREAD_LIMIT "
840                     "reduced reservation to 1 thread\n",
841                     master_tid));
842       return 1;
843     }
844     KC_TRACE(10, ("__kmp_reserve_threads: T#%d OMP_THREAD_LIMIT reduced "
845                   "reservation to %d threads\n",
846                   master_tid, tl_nthreads));
847     new_nthreads = tl_nthreads;
848   }
849 
850   // Check if the threads array is large enough, or needs expanding.
851   // See comment in __kmp_register_root() about the adjustment if
852   // __kmp_threads[0] == NULL.
853   capacity = __kmp_threads_capacity;
854   if (TCR_PTR(__kmp_threads[0]) == NULL) {
855     --capacity;
856   }
857   // If it is not for initializing the hidden helper team, we need to take
858   // __kmp_hidden_helper_threads_num out of the capacity because it is included
859   // in __kmp_threads_capacity.
860   if (__kmp_enable_hidden_helper && !TCR_4(__kmp_init_hidden_helper_threads)) {
861     capacity -= __kmp_hidden_helper_threads_num;
862   }
863   if (__kmp_nth + new_nthreads -
864           (root->r.r_active ? 1 : root->r.r_hot_team->t.t_nproc) >
865       capacity) {
866     // Expand the threads array.
867     int slotsRequired = __kmp_nth + new_nthreads -
868                         (root->r.r_active ? 1 : root->r.r_hot_team->t.t_nproc) -
869                         capacity;
870     int slotsAdded = __kmp_expand_threads(slotsRequired);
871     if (slotsAdded < slotsRequired) {
872       // The threads array was not expanded enough.
873       new_nthreads -= (slotsRequired - slotsAdded);
874       KMP_ASSERT(new_nthreads >= 1);
875 
876       // If dyn-var is false, emit a 1-time warning.
877       if (!get__dynamic_2(parent_team, master_tid) && (!__kmp_reserve_warn)) {
878         __kmp_reserve_warn = 1;
879         if (__kmp_tp_cached) {
880           __kmp_msg(kmp_ms_warning,
881                     KMP_MSG(CantFormThrTeam, set_nthreads, new_nthreads),
882                     KMP_HNT(Set_ALL_THREADPRIVATE, __kmp_tp_capacity),
883                     KMP_HNT(PossibleSystemLimitOnThreads), __kmp_msg_null);
884         } else {
885           __kmp_msg(kmp_ms_warning,
886                     KMP_MSG(CantFormThrTeam, set_nthreads, new_nthreads),
887                     KMP_HNT(SystemLimitOnThreads), __kmp_msg_null);
888         }
889       }
890     }
891   }
892 
893 #ifdef KMP_DEBUG
894   if (new_nthreads == 1) {
895     KC_TRACE(10,
896              ("__kmp_reserve_threads: T#%d serializing team after reclaiming "
897               "dead roots and rechecking; requested %d threads\n",
898               __kmp_get_gtid(), set_nthreads));
899   } else {
900     KC_TRACE(10, ("__kmp_reserve_threads: T#%d allocating %d threads; requested"
901                   " %d threads\n",
902                   __kmp_get_gtid(), new_nthreads, set_nthreads));
903   }
904 #endif // KMP_DEBUG
905   return new_nthreads;
906 }
907 
908 /* Allocate threads from the thread pool and assign them to the new team. We are
909    assured that there are enough threads available, because we checked on that
910    earlier within critical section forkjoin */
911 static void __kmp_fork_team_threads(kmp_root_t *root, kmp_team_t *team,
912                                     kmp_info_t *master_th, int master_gtid) {
913   int i;
914   int use_hot_team;
915 
916   KA_TRACE(10, ("__kmp_fork_team_threads: new_nprocs = %d\n", team->t.t_nproc));
917   KMP_DEBUG_ASSERT(master_gtid == __kmp_get_gtid());
918   KMP_MB();
919 
920   /* first, let's setup the primary thread */
921   master_th->th.th_info.ds.ds_tid = 0;
922   master_th->th.th_team = team;
923   master_th->th.th_team_nproc = team->t.t_nproc;
924   master_th->th.th_team_master = master_th;
925   master_th->th.th_team_serialized = FALSE;
926   master_th->th.th_dispatch = &team->t.t_dispatch[0];
927 
928 /* make sure we are not the optimized hot team */
929 #if KMP_NESTED_HOT_TEAMS
930   use_hot_team = 0;
931   kmp_hot_team_ptr_t *hot_teams = master_th->th.th_hot_teams;
932   if (hot_teams) { // hot teams array is not allocated if
933     // KMP_HOT_TEAMS_MAX_LEVEL=0
934     int level = team->t.t_active_level - 1; // index in array of hot teams
935     if (master_th->th.th_teams_microtask) { // are we inside the teams?
936       if (master_th->th.th_teams_size.nteams > 1) {
937         ++level; // level was not increased in teams construct for
938         // team_of_masters
939       }
940       if (team->t.t_pkfn != (microtask_t)__kmp_teams_master &&
941           master_th->th.th_teams_level == team->t.t_level) {
942         ++level; // level was not increased in teams construct for
943         // team_of_workers before the parallel
944       } // team->t.t_level will be increased inside parallel
945     }
946     if (level < __kmp_hot_teams_max_level) {
947       if (hot_teams[level].hot_team) {
948         // hot team has already been allocated for given level
949         KMP_DEBUG_ASSERT(hot_teams[level].hot_team == team);
950         use_hot_team = 1; // the team is ready to use
951       } else {
952         use_hot_team = 0; // AC: threads are not allocated yet
953         hot_teams[level].hot_team = team; // remember new hot team
954         hot_teams[level].hot_team_nth = team->t.t_nproc;
955       }
956     } else {
957       use_hot_team = 0;
958     }
959   }
960 #else
961   use_hot_team = team == root->r.r_hot_team;
962 #endif
963   if (!use_hot_team) {
964 
965     /* install the primary thread */
966     team->t.t_threads[0] = master_th;
967     __kmp_initialize_info(master_th, team, 0, master_gtid);
968 
969     /* now, install the worker threads */
970     for (i = 1; i < team->t.t_nproc; i++) {
971 
972       /* fork or reallocate a new thread and install it in team */
973       kmp_info_t *thr = __kmp_allocate_thread(root, team, i);
974       team->t.t_threads[i] = thr;
975       KMP_DEBUG_ASSERT(thr);
976       KMP_DEBUG_ASSERT(thr->th.th_team == team);
977       /* align team and thread arrived states */
978       KA_TRACE(20, ("__kmp_fork_team_threads: T#%d(%d:%d) init arrived "
979                     "T#%d(%d:%d) join =%llu, plain=%llu\n",
980                     __kmp_gtid_from_tid(0, team), team->t.t_id, 0,
981                     __kmp_gtid_from_tid(i, team), team->t.t_id, i,
982                     team->t.t_bar[bs_forkjoin_barrier].b_arrived,
983                     team->t.t_bar[bs_plain_barrier].b_arrived));
984       thr->th.th_teams_microtask = master_th->th.th_teams_microtask;
985       thr->th.th_teams_level = master_th->th.th_teams_level;
986       thr->th.th_teams_size = master_th->th.th_teams_size;
987       { // Initialize threads' barrier data.
988         int b;
989         kmp_balign_t *balign = team->t.t_threads[i]->th.th_bar;
990         for (b = 0; b < bs_last_barrier; ++b) {
991           balign[b].bb.b_arrived = team->t.t_bar[b].b_arrived;
992           KMP_DEBUG_ASSERT(balign[b].bb.wait_flag != KMP_BARRIER_PARENT_FLAG);
993 #if USE_DEBUGGER
994           balign[b].bb.b_worker_arrived = team->t.t_bar[b].b_team_arrived;
995 #endif
996         }
997       }
998     }
999 
1000 #if KMP_AFFINITY_SUPPORTED
1001     __kmp_partition_places(team);
1002 #endif
1003   }
1004 
1005   if (__kmp_display_affinity && team->t.t_display_affinity != 1) {
1006     for (i = 0; i < team->t.t_nproc; i++) {
1007       kmp_info_t *thr = team->t.t_threads[i];
1008       if (thr->th.th_prev_num_threads != team->t.t_nproc ||
1009           thr->th.th_prev_level != team->t.t_level) {
1010         team->t.t_display_affinity = 1;
1011         break;
1012       }
1013     }
1014   }
1015 
1016   KMP_MB();
1017 }
1018 
1019 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
1020 // Propagate any changes to the floating point control registers out to the team
1021 // We try to avoid unnecessary writes to the relevant cache line in the team
1022 // structure, so we don't make changes unless they are needed.
1023 inline static void propagateFPControl(kmp_team_t *team) {
1024   if (__kmp_inherit_fp_control) {
1025     kmp_int16 x87_fpu_control_word;
1026     kmp_uint32 mxcsr;
1027 
1028     // Get primary thread's values of FPU control flags (both X87 and vector)
1029     __kmp_store_x87_fpu_control_word(&x87_fpu_control_word);
1030     __kmp_store_mxcsr(&mxcsr);
1031     mxcsr &= KMP_X86_MXCSR_MASK;
1032 
1033     // There is no point looking at t_fp_control_saved here.
1034     // If it is TRUE, we still have to update the values if they are different
1035     // from those we now have. If it is FALSE we didn't save anything yet, but
1036     // our objective is the same. We have to ensure that the values in the team
1037     // are the same as those we have.
1038     // So, this code achieves what we need whether or not t_fp_control_saved is
1039     // true. By checking whether the value needs updating we avoid unnecessary
1040     // writes that would put the cache-line into a written state, causing all
1041     // threads in the team to have to read it again.
1042     KMP_CHECK_UPDATE(team->t.t_x87_fpu_control_word, x87_fpu_control_word);
1043     KMP_CHECK_UPDATE(team->t.t_mxcsr, mxcsr);
1044     // Although we don't use this value, other code in the runtime wants to know
1045     // whether it should restore them. So we must ensure it is correct.
1046     KMP_CHECK_UPDATE(team->t.t_fp_control_saved, TRUE);
1047   } else {
1048     // Similarly here. Don't write to this cache-line in the team structure
1049     // unless we have to.
1050     KMP_CHECK_UPDATE(team->t.t_fp_control_saved, FALSE);
1051   }
1052 }
1053 
1054 // Do the opposite, setting the hardware registers to the updated values from
1055 // the team.
1056 inline static void updateHWFPControl(kmp_team_t *team) {
1057   if (__kmp_inherit_fp_control && team->t.t_fp_control_saved) {
1058     // Only reset the fp control regs if they have been changed in the team.
1059     // the parallel region that we are exiting.
1060     kmp_int16 x87_fpu_control_word;
1061     kmp_uint32 mxcsr;
1062     __kmp_store_x87_fpu_control_word(&x87_fpu_control_word);
1063     __kmp_store_mxcsr(&mxcsr);
1064     mxcsr &= KMP_X86_MXCSR_MASK;
1065 
1066     if (team->t.t_x87_fpu_control_word != x87_fpu_control_word) {
1067       __kmp_clear_x87_fpu_status_word();
1068       __kmp_load_x87_fpu_control_word(&team->t.t_x87_fpu_control_word);
1069     }
1070 
1071     if (team->t.t_mxcsr != mxcsr) {
1072       __kmp_load_mxcsr(&team->t.t_mxcsr);
1073     }
1074   }
1075 }
1076 #else
1077 #define propagateFPControl(x) ((void)0)
1078 #define updateHWFPControl(x) ((void)0)
1079 #endif /* KMP_ARCH_X86 || KMP_ARCH_X86_64 */
1080 
1081 static void __kmp_alloc_argv_entries(int argc, kmp_team_t *team,
1082                                      int realloc); // forward declaration
1083 
1084 /* Run a parallel region that has been serialized, so runs only in a team of the
1085    single primary thread. */
1086 void __kmp_serialized_parallel(ident_t *loc, kmp_int32 global_tid) {
1087   kmp_info_t *this_thr;
1088   kmp_team_t *serial_team;
1089 
1090   KC_TRACE(10, ("__kmpc_serialized_parallel: called by T#%d\n", global_tid));
1091 
1092   /* Skip all this code for autopar serialized loops since it results in
1093      unacceptable overhead */
1094   if (loc != NULL && (loc->flags & KMP_IDENT_AUTOPAR))
1095     return;
1096 
1097   if (!TCR_4(__kmp_init_parallel))
1098     __kmp_parallel_initialize();
1099   __kmp_resume_if_soft_paused();
1100 
1101   this_thr = __kmp_threads[global_tid];
1102   serial_team = this_thr->th.th_serial_team;
1103 
1104   /* utilize the serialized team held by this thread */
1105   KMP_DEBUG_ASSERT(serial_team);
1106   KMP_MB();
1107 
1108   if (__kmp_tasking_mode != tskm_immediate_exec) {
1109     KMP_DEBUG_ASSERT(
1110         this_thr->th.th_task_team ==
1111         this_thr->th.th_team->t.t_task_team[this_thr->th.th_task_state]);
1112     KMP_DEBUG_ASSERT(serial_team->t.t_task_team[this_thr->th.th_task_state] ==
1113                      NULL);
1114     KA_TRACE(20, ("__kmpc_serialized_parallel: T#%d pushing task_team %p / "
1115                   "team %p, new task_team = NULL\n",
1116                   global_tid, this_thr->th.th_task_team, this_thr->th.th_team));
1117     this_thr->th.th_task_team = NULL;
1118   }
1119 
1120   kmp_proc_bind_t proc_bind = this_thr->th.th_set_proc_bind;
1121   if (this_thr->th.th_current_task->td_icvs.proc_bind == proc_bind_false) {
1122     proc_bind = proc_bind_false;
1123   } else if (proc_bind == proc_bind_default) {
1124     // No proc_bind clause was specified, so use the current value
1125     // of proc-bind-var for this parallel region.
1126     proc_bind = this_thr->th.th_current_task->td_icvs.proc_bind;
1127   }
1128   // Reset for next parallel region
1129   this_thr->th.th_set_proc_bind = proc_bind_default;
1130 
1131 #if OMPT_SUPPORT
1132   ompt_data_t ompt_parallel_data = ompt_data_none;
1133   ompt_data_t *implicit_task_data;
1134   void *codeptr = OMPT_LOAD_RETURN_ADDRESS(global_tid);
1135   if (ompt_enabled.enabled &&
1136       this_thr->th.ompt_thread_info.state != ompt_state_overhead) {
1137 
1138     ompt_task_info_t *parent_task_info;
1139     parent_task_info = OMPT_CUR_TASK_INFO(this_thr);
1140 
1141     parent_task_info->frame.enter_frame.ptr = OMPT_GET_FRAME_ADDRESS(0);
1142     if (ompt_enabled.ompt_callback_parallel_begin) {
1143       int team_size = 1;
1144 
1145       ompt_callbacks.ompt_callback(ompt_callback_parallel_begin)(
1146           &(parent_task_info->task_data), &(parent_task_info->frame),
1147           &ompt_parallel_data, team_size,
1148           ompt_parallel_invoker_program | ompt_parallel_team, codeptr);
1149     }
1150   }
1151 #endif // OMPT_SUPPORT
1152 
1153   if (this_thr->th.th_team != serial_team) {
1154     // Nested level will be an index in the nested nthreads array
1155     int level = this_thr->th.th_team->t.t_level;
1156 
1157     if (serial_team->t.t_serialized) {
1158       /* this serial team was already used
1159          TODO increase performance by making this locks more specific */
1160       kmp_team_t *new_team;
1161 
1162       __kmp_acquire_bootstrap_lock(&__kmp_forkjoin_lock);
1163 
1164       new_team =
1165           __kmp_allocate_team(this_thr->th.th_root, 1, 1,
1166 #if OMPT_SUPPORT
1167                               ompt_parallel_data,
1168 #endif
1169                               proc_bind, &this_thr->th.th_current_task->td_icvs,
1170                               0 USE_NESTED_HOT_ARG(NULL));
1171       __kmp_release_bootstrap_lock(&__kmp_forkjoin_lock);
1172       KMP_ASSERT(new_team);
1173 
1174       /* setup new serialized team and install it */
1175       new_team->t.t_threads[0] = this_thr;
1176       new_team->t.t_parent = this_thr->th.th_team;
1177       serial_team = new_team;
1178       this_thr->th.th_serial_team = serial_team;
1179 
1180       KF_TRACE(
1181           10,
1182           ("__kmpc_serialized_parallel: T#%d allocated new serial team %p\n",
1183            global_tid, serial_team));
1184 
1185       /* TODO the above breaks the requirement that if we run out of resources,
1186          then we can still guarantee that serialized teams are ok, since we may
1187          need to allocate a new one */
1188     } else {
1189       KF_TRACE(
1190           10,
1191           ("__kmpc_serialized_parallel: T#%d reusing cached serial team %p\n",
1192            global_tid, serial_team));
1193     }
1194 
1195     /* we have to initialize this serial team */
1196     KMP_DEBUG_ASSERT(serial_team->t.t_threads);
1197     KMP_DEBUG_ASSERT(serial_team->t.t_threads[0] == this_thr);
1198     KMP_DEBUG_ASSERT(this_thr->th.th_team != serial_team);
1199     serial_team->t.t_ident = loc;
1200     serial_team->t.t_serialized = 1;
1201     serial_team->t.t_nproc = 1;
1202     serial_team->t.t_parent = this_thr->th.th_team;
1203     serial_team->t.t_sched.sched = this_thr->th.th_team->t.t_sched.sched;
1204     this_thr->th.th_team = serial_team;
1205     serial_team->t.t_master_tid = this_thr->th.th_info.ds.ds_tid;
1206 
1207     KF_TRACE(10, ("__kmpc_serialized_parallel: T#d curtask=%p\n", global_tid,
1208                   this_thr->th.th_current_task));
1209     KMP_ASSERT(this_thr->th.th_current_task->td_flags.executing == 1);
1210     this_thr->th.th_current_task->td_flags.executing = 0;
1211 
1212     __kmp_push_current_task_to_thread(this_thr, serial_team, 0);
1213 
1214     /* TODO: GEH: do ICVs work for nested serialized teams? Don't we need an
1215        implicit task for each serialized task represented by
1216        team->t.t_serialized? */
1217     copy_icvs(&this_thr->th.th_current_task->td_icvs,
1218               &this_thr->th.th_current_task->td_parent->td_icvs);
1219 
1220     // Thread value exists in the nested nthreads array for the next nested
1221     // level
1222     if (__kmp_nested_nth.used && (level + 1 < __kmp_nested_nth.used)) {
1223       this_thr->th.th_current_task->td_icvs.nproc =
1224           __kmp_nested_nth.nth[level + 1];
1225     }
1226 
1227     if (__kmp_nested_proc_bind.used &&
1228         (level + 1 < __kmp_nested_proc_bind.used)) {
1229       this_thr->th.th_current_task->td_icvs.proc_bind =
1230           __kmp_nested_proc_bind.bind_types[level + 1];
1231     }
1232 
1233 #if USE_DEBUGGER
1234     serial_team->t.t_pkfn = (microtask_t)(~0); // For the debugger.
1235 #endif
1236     this_thr->th.th_info.ds.ds_tid = 0;
1237 
1238     /* set thread cache values */
1239     this_thr->th.th_team_nproc = 1;
1240     this_thr->th.th_team_master = this_thr;
1241     this_thr->th.th_team_serialized = 1;
1242 
1243     serial_team->t.t_level = serial_team->t.t_parent->t.t_level + 1;
1244     serial_team->t.t_active_level = serial_team->t.t_parent->t.t_active_level;
1245     serial_team->t.t_def_allocator = this_thr->th.th_def_allocator; // save
1246 
1247     propagateFPControl(serial_team);
1248 
1249     /* check if we need to allocate dispatch buffers stack */
1250     KMP_DEBUG_ASSERT(serial_team->t.t_dispatch);
1251     if (!serial_team->t.t_dispatch->th_disp_buffer) {
1252       serial_team->t.t_dispatch->th_disp_buffer =
1253           (dispatch_private_info_t *)__kmp_allocate(
1254               sizeof(dispatch_private_info_t));
1255     }
1256     this_thr->th.th_dispatch = serial_team->t.t_dispatch;
1257 
1258     KMP_MB();
1259 
1260   } else {
1261     /* this serialized team is already being used,
1262      * that's fine, just add another nested level */
1263     KMP_DEBUG_ASSERT(this_thr->th.th_team == serial_team);
1264     KMP_DEBUG_ASSERT(serial_team->t.t_threads);
1265     KMP_DEBUG_ASSERT(serial_team->t.t_threads[0] == this_thr);
1266     ++serial_team->t.t_serialized;
1267     this_thr->th.th_team_serialized = serial_team->t.t_serialized;
1268 
1269     // Nested level will be an index in the nested nthreads array
1270     int level = this_thr->th.th_team->t.t_level;
1271     // Thread value exists in the nested nthreads array for the next nested
1272     // level
1273     if (__kmp_nested_nth.used && (level + 1 < __kmp_nested_nth.used)) {
1274       this_thr->th.th_current_task->td_icvs.nproc =
1275           __kmp_nested_nth.nth[level + 1];
1276     }
1277     serial_team->t.t_level++;
1278     KF_TRACE(10, ("__kmpc_serialized_parallel: T#%d increasing nesting level "
1279                   "of serial team %p to %d\n",
1280                   global_tid, serial_team, serial_team->t.t_level));
1281 
1282     /* allocate/push dispatch buffers stack */
1283     KMP_DEBUG_ASSERT(serial_team->t.t_dispatch);
1284     {
1285       dispatch_private_info_t *disp_buffer =
1286           (dispatch_private_info_t *)__kmp_allocate(
1287               sizeof(dispatch_private_info_t));
1288       disp_buffer->next = serial_team->t.t_dispatch->th_disp_buffer;
1289       serial_team->t.t_dispatch->th_disp_buffer = disp_buffer;
1290     }
1291     this_thr->th.th_dispatch = serial_team->t.t_dispatch;
1292 
1293     KMP_MB();
1294   }
1295   KMP_CHECK_UPDATE(serial_team->t.t_cancel_request, cancel_noreq);
1296 
1297   // Perform the display affinity functionality for
1298   // serialized parallel regions
1299   if (__kmp_display_affinity) {
1300     if (this_thr->th.th_prev_level != serial_team->t.t_level ||
1301         this_thr->th.th_prev_num_threads != 1) {
1302       // NULL means use the affinity-format-var ICV
1303       __kmp_aux_display_affinity(global_tid, NULL);
1304       this_thr->th.th_prev_level = serial_team->t.t_level;
1305       this_thr->th.th_prev_num_threads = 1;
1306     }
1307   }
1308 
1309   if (__kmp_env_consistency_check)
1310     __kmp_push_parallel(global_tid, NULL);
1311 #if OMPT_SUPPORT
1312   serial_team->t.ompt_team_info.master_return_address = codeptr;
1313   if (ompt_enabled.enabled &&
1314       this_thr->th.ompt_thread_info.state != ompt_state_overhead) {
1315     OMPT_CUR_TASK_INFO(this_thr)->frame.exit_frame.ptr =
1316         OMPT_GET_FRAME_ADDRESS(0);
1317 
1318     ompt_lw_taskteam_t lw_taskteam;
1319     __ompt_lw_taskteam_init(&lw_taskteam, this_thr, global_tid,
1320                             &ompt_parallel_data, codeptr);
1321 
1322     __ompt_lw_taskteam_link(&lw_taskteam, this_thr, 1);
1323     // don't use lw_taskteam after linking. content was swaped
1324 
1325     /* OMPT implicit task begin */
1326     implicit_task_data = OMPT_CUR_TASK_DATA(this_thr);
1327     if (ompt_enabled.ompt_callback_implicit_task) {
1328       ompt_callbacks.ompt_callback(ompt_callback_implicit_task)(
1329           ompt_scope_begin, OMPT_CUR_TEAM_DATA(this_thr),
1330           OMPT_CUR_TASK_DATA(this_thr), 1, __kmp_tid_from_gtid(global_tid),
1331           ompt_task_implicit); // TODO: Can this be ompt_task_initial?
1332       OMPT_CUR_TASK_INFO(this_thr)->thread_num =
1333           __kmp_tid_from_gtid(global_tid);
1334     }
1335 
1336     /* OMPT state */
1337     this_thr->th.ompt_thread_info.state = ompt_state_work_parallel;
1338     OMPT_CUR_TASK_INFO(this_thr)->frame.exit_frame.ptr =
1339         OMPT_GET_FRAME_ADDRESS(0);
1340   }
1341 #endif
1342 }
1343 
1344 /* most of the work for a fork */
1345 /* return true if we really went parallel, false if serialized */
1346 int __kmp_fork_call(ident_t *loc, int gtid,
1347                     enum fork_context_e call_context, // Intel, GNU, ...
1348                     kmp_int32 argc, microtask_t microtask, launch_t invoker,
1349                     kmp_va_list ap) {
1350   void **argv;
1351   int i;
1352   int master_tid;
1353   int master_this_cons;
1354   kmp_team_t *team;
1355   kmp_team_t *parent_team;
1356   kmp_info_t *master_th;
1357   kmp_root_t *root;
1358   int nthreads;
1359   int master_active;
1360   int master_set_numthreads;
1361   int level;
1362   int active_level;
1363   int teams_level;
1364 #if KMP_NESTED_HOT_TEAMS
1365   kmp_hot_team_ptr_t **p_hot_teams;
1366 #endif
1367   { // KMP_TIME_BLOCK
1368     KMP_TIME_DEVELOPER_PARTITIONED_BLOCK(KMP_fork_call);
1369     KMP_COUNT_VALUE(OMP_PARALLEL_args, argc);
1370 
1371     KA_TRACE(20, ("__kmp_fork_call: enter T#%d\n", gtid));
1372     if (__kmp_stkpadding > 0 && __kmp_root[gtid] != NULL) {
1373       /* Some systems prefer the stack for the root thread(s) to start with */
1374       /* some gap from the parent stack to prevent false sharing. */
1375       void *dummy = KMP_ALLOCA(__kmp_stkpadding);
1376       /* These 2 lines below are so this does not get optimized out */
1377       if (__kmp_stkpadding > KMP_MAX_STKPADDING)
1378         __kmp_stkpadding += (short)((kmp_int64)dummy);
1379     }
1380 
1381     /* initialize if needed */
1382     KMP_DEBUG_ASSERT(
1383         __kmp_init_serial); // AC: potentially unsafe, not in sync with shutdown
1384     if (!TCR_4(__kmp_init_parallel))
1385       __kmp_parallel_initialize();
1386     __kmp_resume_if_soft_paused();
1387 
1388     /* setup current data */
1389     master_th = __kmp_threads[gtid]; // AC: potentially unsafe, not in sync with
1390     // shutdown
1391     parent_team = master_th->th.th_team;
1392     master_tid = master_th->th.th_info.ds.ds_tid;
1393     master_this_cons = master_th->th.th_local.this_construct;
1394     root = master_th->th.th_root;
1395     master_active = root->r.r_active;
1396     master_set_numthreads = master_th->th.th_set_nproc;
1397 
1398 #if OMPT_SUPPORT
1399     ompt_data_t ompt_parallel_data = ompt_data_none;
1400     ompt_data_t *parent_task_data;
1401     ompt_frame_t *ompt_frame;
1402     ompt_data_t *implicit_task_data;
1403     void *return_address = NULL;
1404 
1405     if (ompt_enabled.enabled) {
1406       __ompt_get_task_info_internal(0, NULL, &parent_task_data, &ompt_frame,
1407                                     NULL, NULL);
1408       return_address = OMPT_LOAD_RETURN_ADDRESS(gtid);
1409     }
1410 #endif
1411 
1412     // Nested level will be an index in the nested nthreads array
1413     level = parent_team->t.t_level;
1414     // used to launch non-serial teams even if nested is not allowed
1415     active_level = parent_team->t.t_active_level;
1416     // needed to check nesting inside the teams
1417     teams_level = master_th->th.th_teams_level;
1418 #if KMP_NESTED_HOT_TEAMS
1419     p_hot_teams = &master_th->th.th_hot_teams;
1420     if (*p_hot_teams == NULL && __kmp_hot_teams_max_level > 0) {
1421       *p_hot_teams = (kmp_hot_team_ptr_t *)__kmp_allocate(
1422           sizeof(kmp_hot_team_ptr_t) * __kmp_hot_teams_max_level);
1423       (*p_hot_teams)[0].hot_team = root->r.r_hot_team;
1424       // it is either actual or not needed (when active_level > 0)
1425       (*p_hot_teams)[0].hot_team_nth = 1;
1426     }
1427 #endif
1428 
1429 #if OMPT_SUPPORT
1430     if (ompt_enabled.enabled) {
1431       if (ompt_enabled.ompt_callback_parallel_begin) {
1432         int team_size = master_set_numthreads
1433                             ? master_set_numthreads
1434                             : get__nproc_2(parent_team, master_tid);
1435         int flags = OMPT_INVOKER(call_context) |
1436                     ((microtask == (microtask_t)__kmp_teams_master)
1437                          ? ompt_parallel_league
1438                          : ompt_parallel_team);
1439         ompt_callbacks.ompt_callback(ompt_callback_parallel_begin)(
1440             parent_task_data, ompt_frame, &ompt_parallel_data, team_size, flags,
1441             return_address);
1442       }
1443       master_th->th.ompt_thread_info.state = ompt_state_overhead;
1444     }
1445 #endif
1446 
1447     master_th->th.th_ident = loc;
1448 
1449     if (master_th->th.th_teams_microtask && ap &&
1450         microtask != (microtask_t)__kmp_teams_master && level == teams_level) {
1451       // AC: This is start of parallel that is nested inside teams construct.
1452       // The team is actual (hot), all workers are ready at the fork barrier.
1453       // No lock needed to initialize the team a bit, then free workers.
1454       parent_team->t.t_ident = loc;
1455       __kmp_alloc_argv_entries(argc, parent_team, TRUE);
1456       parent_team->t.t_argc = argc;
1457       argv = (void **)parent_team->t.t_argv;
1458       for (i = argc - 1; i >= 0; --i)
1459         *argv++ = va_arg(kmp_va_deref(ap), void *);
1460       // Increment our nested depth levels, but not increase the serialization
1461       if (parent_team == master_th->th.th_serial_team) {
1462         // AC: we are in serialized parallel
1463         __kmpc_serialized_parallel(loc, gtid);
1464         KMP_DEBUG_ASSERT(parent_team->t.t_serialized > 1);
1465 
1466         if (call_context == fork_context_gnu) {
1467           // AC: need to decrement t_serialized for enquiry functions to work
1468           // correctly, will restore at join time
1469           parent_team->t.t_serialized--;
1470           return TRUE;
1471         }
1472 
1473 #if OMPT_SUPPORT
1474         void *dummy;
1475         void **exit_frame_p;
1476 
1477         ompt_lw_taskteam_t lw_taskteam;
1478 
1479         if (ompt_enabled.enabled) {
1480           __ompt_lw_taskteam_init(&lw_taskteam, master_th, gtid,
1481                                   &ompt_parallel_data, return_address);
1482           exit_frame_p = &(lw_taskteam.ompt_task_info.frame.exit_frame.ptr);
1483 
1484           __ompt_lw_taskteam_link(&lw_taskteam, master_th, 0);
1485           // don't use lw_taskteam after linking. content was swaped
1486 
1487           /* OMPT implicit task begin */
1488           implicit_task_data = OMPT_CUR_TASK_DATA(master_th);
1489           if (ompt_enabled.ompt_callback_implicit_task) {
1490             OMPT_CUR_TASK_INFO(master_th)->thread_num =
1491                 __kmp_tid_from_gtid(gtid);
1492             ompt_callbacks.ompt_callback(ompt_callback_implicit_task)(
1493                 ompt_scope_begin, OMPT_CUR_TEAM_DATA(master_th),
1494                 implicit_task_data, 1,
1495                 OMPT_CUR_TASK_INFO(master_th)->thread_num, ompt_task_implicit);
1496           }
1497 
1498           /* OMPT state */
1499           master_th->th.ompt_thread_info.state = ompt_state_work_parallel;
1500         } else {
1501           exit_frame_p = &dummy;
1502         }
1503 #endif
1504         // AC: need to decrement t_serialized for enquiry functions to work
1505         // correctly, will restore at join time
1506         parent_team->t.t_serialized--;
1507 
1508         {
1509           KMP_TIME_PARTITIONED_BLOCK(OMP_parallel);
1510           KMP_SET_THREAD_STATE_BLOCK(IMPLICIT_TASK);
1511           __kmp_invoke_microtask(microtask, gtid, 0, argc, parent_team->t.t_argv
1512 #if OMPT_SUPPORT
1513                                  ,
1514                                  exit_frame_p
1515 #endif
1516           );
1517         }
1518 
1519 #if OMPT_SUPPORT
1520         if (ompt_enabled.enabled) {
1521           *exit_frame_p = NULL;
1522           OMPT_CUR_TASK_INFO(master_th)->frame.exit_frame = ompt_data_none;
1523           if (ompt_enabled.ompt_callback_implicit_task) {
1524             ompt_callbacks.ompt_callback(ompt_callback_implicit_task)(
1525                 ompt_scope_end, NULL, implicit_task_data, 1,
1526                 OMPT_CUR_TASK_INFO(master_th)->thread_num, ompt_task_implicit);
1527           }
1528           ompt_parallel_data = *OMPT_CUR_TEAM_DATA(master_th);
1529           __ompt_lw_taskteam_unlink(master_th);
1530           if (ompt_enabled.ompt_callback_parallel_end) {
1531             ompt_callbacks.ompt_callback(ompt_callback_parallel_end)(
1532                 &ompt_parallel_data, OMPT_CUR_TASK_DATA(master_th),
1533                 OMPT_INVOKER(call_context) | ompt_parallel_team,
1534                 return_address);
1535           }
1536           master_th->th.ompt_thread_info.state = ompt_state_overhead;
1537         }
1538 #endif
1539         return TRUE;
1540       }
1541 
1542       parent_team->t.t_pkfn = microtask;
1543       parent_team->t.t_invoke = invoker;
1544       KMP_ATOMIC_INC(&root->r.r_in_parallel);
1545       parent_team->t.t_active_level++;
1546       parent_team->t.t_level++;
1547       parent_team->t.t_def_allocator = master_th->th.th_def_allocator; // save
1548 
1549 #if OMPT_SUPPORT
1550       if (ompt_enabled.enabled) {
1551         ompt_lw_taskteam_t lw_taskteam;
1552         __ompt_lw_taskteam_init(&lw_taskteam, master_th, gtid,
1553                                 &ompt_parallel_data, return_address);
1554         __ompt_lw_taskteam_link(&lw_taskteam, master_th, 1, true);
1555       }
1556 #endif
1557 
1558       /* Change number of threads in the team if requested */
1559       if (master_set_numthreads) { // The parallel has num_threads clause
1560         if (master_set_numthreads < master_th->th.th_teams_size.nth) {
1561           // AC: only can reduce number of threads dynamically, can't increase
1562           kmp_info_t **other_threads = parent_team->t.t_threads;
1563           parent_team->t.t_nproc = master_set_numthreads;
1564           for (i = 0; i < master_set_numthreads; ++i) {
1565             other_threads[i]->th.th_team_nproc = master_set_numthreads;
1566           }
1567           // Keep extra threads hot in the team for possible next parallels
1568         }
1569         master_th->th.th_set_nproc = 0;
1570       }
1571 
1572 #if USE_DEBUGGER
1573       if (__kmp_debugging) { // Let debugger override number of threads.
1574         int nth = __kmp_omp_num_threads(loc);
1575         if (nth > 0) { // 0 means debugger doesn't want to change num threads
1576           master_set_numthreads = nth;
1577         }
1578       }
1579 #endif
1580 
1581 #if USE_ITT_BUILD && USE_ITT_NOTIFY
1582       if (((__itt_frame_submit_v3_ptr && __itt_get_timestamp_ptr) ||
1583            KMP_ITT_DEBUG) &&
1584           __kmp_forkjoin_frames_mode == 3 &&
1585           parent_team->t.t_active_level == 1 // only report frames at level 1
1586           && master_th->th.th_teams_size.nteams == 1) {
1587         kmp_uint64 tmp_time = __itt_get_timestamp();
1588         master_th->th.th_frame_time = tmp_time;
1589         parent_team->t.t_region_time = tmp_time;
1590       }
1591       if (__itt_stack_caller_create_ptr) {
1592         KMP_DEBUG_ASSERT(parent_team->t.t_stack_id == NULL);
1593         // create new stack stitching id before entering fork barrier
1594         parent_team->t.t_stack_id = __kmp_itt_stack_caller_create();
1595       }
1596 #endif /* USE_ITT_BUILD && USE_ITT_NOTIFY */
1597 
1598       KF_TRACE(10, ("__kmp_fork_call: before internal fork: root=%p, team=%p, "
1599                     "master_th=%p, gtid=%d\n",
1600                     root, parent_team, master_th, gtid));
1601       __kmp_internal_fork(loc, gtid, parent_team);
1602       KF_TRACE(10, ("__kmp_fork_call: after internal fork: root=%p, team=%p, "
1603                     "master_th=%p, gtid=%d\n",
1604                     root, parent_team, master_th, gtid));
1605 
1606       if (call_context == fork_context_gnu)
1607         return TRUE;
1608 
1609       /* Invoke microtask for PRIMARY thread */
1610       KA_TRACE(20, ("__kmp_fork_call: T#%d(%d:0) invoke microtask = %p\n", gtid,
1611                     parent_team->t.t_id, parent_team->t.t_pkfn));
1612 
1613       if (!parent_team->t.t_invoke(gtid)) {
1614         KMP_ASSERT2(0, "cannot invoke microtask for PRIMARY thread");
1615       }
1616       KA_TRACE(20, ("__kmp_fork_call: T#%d(%d:0) done microtask = %p\n", gtid,
1617                     parent_team->t.t_id, parent_team->t.t_pkfn));
1618       KMP_MB(); /* Flush all pending memory write invalidates.  */
1619 
1620       KA_TRACE(20, ("__kmp_fork_call: parallel exit T#%d\n", gtid));
1621 
1622       return TRUE;
1623     } // Parallel closely nested in teams construct
1624 
1625 #if KMP_DEBUG
1626     if (__kmp_tasking_mode != tskm_immediate_exec) {
1627       KMP_DEBUG_ASSERT(master_th->th.th_task_team ==
1628                        parent_team->t.t_task_team[master_th->th.th_task_state]);
1629     }
1630 #endif
1631 
1632     int enter_teams = 0;
1633     if (parent_team->t.t_active_level >=
1634         master_th->th.th_current_task->td_icvs.max_active_levels) {
1635       nthreads = 1;
1636     } else {
1637       enter_teams = ((ap == NULL && active_level == 0) ||
1638                      (ap && teams_level > 0 && teams_level == level));
1639       nthreads =
1640           master_set_numthreads
1641               ? master_set_numthreads
1642               : get__nproc_2(
1643                     parent_team,
1644                     master_tid); // TODO: get nproc directly from current task
1645 
1646       // Check if we need to take forkjoin lock? (no need for serialized
1647       // parallel out of teams construct). This code moved here from
1648       // __kmp_reserve_threads() to speedup nested serialized parallels.
1649       if (nthreads > 1) {
1650         if ((get__max_active_levels(master_th) == 1 &&
1651              (root->r.r_in_parallel && !enter_teams)) ||
1652             (__kmp_library == library_serial)) {
1653           KC_TRACE(10, ("__kmp_fork_call: T#%d serializing team; requested %d"
1654                         " threads\n",
1655                         gtid, nthreads));
1656           nthreads = 1;
1657         }
1658       }
1659       if (nthreads > 1) {
1660         /* determine how many new threads we can use */
1661         __kmp_acquire_bootstrap_lock(&__kmp_forkjoin_lock);
1662         /* AC: If we execute teams from parallel region (on host), then teams
1663            should be created but each can only have 1 thread if nesting is
1664            disabled. If teams called from serial region, then teams and their
1665            threads should be created regardless of the nesting setting. */
1666         nthreads = __kmp_reserve_threads(root, parent_team, master_tid,
1667                                          nthreads, enter_teams);
1668         if (nthreads == 1) {
1669           // Free lock for single thread execution here; for multi-thread
1670           // execution it will be freed later after team of threads created
1671           // and initialized
1672           __kmp_release_bootstrap_lock(&__kmp_forkjoin_lock);
1673         }
1674       }
1675     }
1676     KMP_DEBUG_ASSERT(nthreads > 0);
1677 
1678     // If we temporarily changed the set number of threads then restore it now
1679     master_th->th.th_set_nproc = 0;
1680 
1681     /* create a serialized parallel region? */
1682     if (nthreads == 1) {
1683 /* josh todo: hypothetical question: what do we do for OS X*? */
1684 #if KMP_OS_LINUX &&                                                            \
1685     (KMP_ARCH_X86 || KMP_ARCH_X86_64 || KMP_ARCH_ARM || KMP_ARCH_AARCH64)
1686       void *args[argc];
1687 #else
1688       void **args = (void **)KMP_ALLOCA(argc * sizeof(void *));
1689 #endif /* KMP_OS_LINUX && ( KMP_ARCH_X86 || KMP_ARCH_X86_64 || KMP_ARCH_ARM || \
1690           KMP_ARCH_AARCH64) */
1691 
1692       KA_TRACE(20,
1693                ("__kmp_fork_call: T#%d serializing parallel region\n", gtid));
1694 
1695       __kmpc_serialized_parallel(loc, gtid);
1696 
1697       if (call_context == fork_context_intel) {
1698         /* TODO this sucks, use the compiler itself to pass args! :) */
1699         master_th->th.th_serial_team->t.t_ident = loc;
1700         if (!ap) {
1701           // revert change made in __kmpc_serialized_parallel()
1702           master_th->th.th_serial_team->t.t_level--;
1703           // Get args from parent team for teams construct
1704 
1705 #if OMPT_SUPPORT
1706           void *dummy;
1707           void **exit_frame_p;
1708           ompt_task_info_t *task_info;
1709 
1710           ompt_lw_taskteam_t lw_taskteam;
1711 
1712           if (ompt_enabled.enabled) {
1713             __ompt_lw_taskteam_init(&lw_taskteam, master_th, gtid,
1714                                     &ompt_parallel_data, return_address);
1715 
1716             __ompt_lw_taskteam_link(&lw_taskteam, master_th, 0);
1717             // don't use lw_taskteam after linking. content was swaped
1718 
1719             task_info = OMPT_CUR_TASK_INFO(master_th);
1720             exit_frame_p = &(task_info->frame.exit_frame.ptr);
1721             if (ompt_enabled.ompt_callback_implicit_task) {
1722               OMPT_CUR_TASK_INFO(master_th)->thread_num =
1723                   __kmp_tid_from_gtid(gtid);
1724               ompt_callbacks.ompt_callback(ompt_callback_implicit_task)(
1725                   ompt_scope_begin, OMPT_CUR_TEAM_DATA(master_th),
1726                   &(task_info->task_data), 1,
1727                   OMPT_CUR_TASK_INFO(master_th)->thread_num,
1728                   ompt_task_implicit);
1729             }
1730 
1731             /* OMPT state */
1732             master_th->th.ompt_thread_info.state = ompt_state_work_parallel;
1733           } else {
1734             exit_frame_p = &dummy;
1735           }
1736 #endif
1737 
1738           {
1739             KMP_TIME_PARTITIONED_BLOCK(OMP_parallel);
1740             KMP_SET_THREAD_STATE_BLOCK(IMPLICIT_TASK);
1741             __kmp_invoke_microtask(microtask, gtid, 0, argc,
1742                                    parent_team->t.t_argv
1743 #if OMPT_SUPPORT
1744                                    ,
1745                                    exit_frame_p
1746 #endif
1747             );
1748           }
1749 
1750 #if OMPT_SUPPORT
1751           if (ompt_enabled.enabled) {
1752             *exit_frame_p = NULL;
1753             if (ompt_enabled.ompt_callback_implicit_task) {
1754               ompt_callbacks.ompt_callback(ompt_callback_implicit_task)(
1755                   ompt_scope_end, NULL, &(task_info->task_data), 1,
1756                   OMPT_CUR_TASK_INFO(master_th)->thread_num,
1757                   ompt_task_implicit);
1758             }
1759             ompt_parallel_data = *OMPT_CUR_TEAM_DATA(master_th);
1760             __ompt_lw_taskteam_unlink(master_th);
1761             if (ompt_enabled.ompt_callback_parallel_end) {
1762               ompt_callbacks.ompt_callback(ompt_callback_parallel_end)(
1763                   &ompt_parallel_data, parent_task_data,
1764                   OMPT_INVOKER(call_context) | ompt_parallel_team,
1765                   return_address);
1766             }
1767             master_th->th.ompt_thread_info.state = ompt_state_overhead;
1768           }
1769 #endif
1770         } else if (microtask == (microtask_t)__kmp_teams_master) {
1771           KMP_DEBUG_ASSERT(master_th->th.th_team ==
1772                            master_th->th.th_serial_team);
1773           team = master_th->th.th_team;
1774           // team->t.t_pkfn = microtask;
1775           team->t.t_invoke = invoker;
1776           __kmp_alloc_argv_entries(argc, team, TRUE);
1777           team->t.t_argc = argc;
1778           argv = (void **)team->t.t_argv;
1779           if (ap) {
1780             for (i = argc - 1; i >= 0; --i)
1781               *argv++ = va_arg(kmp_va_deref(ap), void *);
1782           } else {
1783             for (i = 0; i < argc; ++i)
1784               // Get args from parent team for teams construct
1785               argv[i] = parent_team->t.t_argv[i];
1786           }
1787           // AC: revert change made in __kmpc_serialized_parallel()
1788           //     because initial code in teams should have level=0
1789           team->t.t_level--;
1790           // AC: call special invoker for outer "parallel" of teams construct
1791           invoker(gtid);
1792 #if OMPT_SUPPORT
1793           if (ompt_enabled.enabled) {
1794             ompt_task_info_t *task_info = OMPT_CUR_TASK_INFO(master_th);
1795             if (ompt_enabled.ompt_callback_implicit_task) {
1796               ompt_callbacks.ompt_callback(ompt_callback_implicit_task)(
1797                   ompt_scope_end, NULL, &(task_info->task_data), 0,
1798                   OMPT_CUR_TASK_INFO(master_th)->thread_num, ompt_task_initial);
1799             }
1800             if (ompt_enabled.ompt_callback_parallel_end) {
1801               ompt_callbacks.ompt_callback(ompt_callback_parallel_end)(
1802                   &ompt_parallel_data, parent_task_data,
1803                   OMPT_INVOKER(call_context) | ompt_parallel_league,
1804                   return_address);
1805             }
1806             master_th->th.ompt_thread_info.state = ompt_state_overhead;
1807           }
1808 #endif
1809         } else {
1810           argv = args;
1811           for (i = argc - 1; i >= 0; --i)
1812             *argv++ = va_arg(kmp_va_deref(ap), void *);
1813           KMP_MB();
1814 
1815 #if OMPT_SUPPORT
1816           void *dummy;
1817           void **exit_frame_p;
1818           ompt_task_info_t *task_info;
1819 
1820           ompt_lw_taskteam_t lw_taskteam;
1821 
1822           if (ompt_enabled.enabled) {
1823             __ompt_lw_taskteam_init(&lw_taskteam, master_th, gtid,
1824                                     &ompt_parallel_data, return_address);
1825             __ompt_lw_taskteam_link(&lw_taskteam, master_th, 0);
1826             // don't use lw_taskteam after linking. content was swaped
1827             task_info = OMPT_CUR_TASK_INFO(master_th);
1828             exit_frame_p = &(task_info->frame.exit_frame.ptr);
1829 
1830             /* OMPT implicit task begin */
1831             implicit_task_data = OMPT_CUR_TASK_DATA(master_th);
1832             if (ompt_enabled.ompt_callback_implicit_task) {
1833               ompt_callbacks.ompt_callback(ompt_callback_implicit_task)(
1834                   ompt_scope_begin, OMPT_CUR_TEAM_DATA(master_th),
1835                   implicit_task_data, 1, __kmp_tid_from_gtid(gtid),
1836                   ompt_task_implicit);
1837               OMPT_CUR_TASK_INFO(master_th)->thread_num =
1838                   __kmp_tid_from_gtid(gtid);
1839             }
1840 
1841             /* OMPT state */
1842             master_th->th.ompt_thread_info.state = ompt_state_work_parallel;
1843           } else {
1844             exit_frame_p = &dummy;
1845           }
1846 #endif
1847 
1848           {
1849             KMP_TIME_PARTITIONED_BLOCK(OMP_parallel);
1850             KMP_SET_THREAD_STATE_BLOCK(IMPLICIT_TASK);
1851             __kmp_invoke_microtask(microtask, gtid, 0, argc, args
1852 #if OMPT_SUPPORT
1853                                    ,
1854                                    exit_frame_p
1855 #endif
1856             );
1857           }
1858 
1859 #if OMPT_SUPPORT
1860           if (ompt_enabled.enabled) {
1861             *exit_frame_p = NULL;
1862             if (ompt_enabled.ompt_callback_implicit_task) {
1863               ompt_callbacks.ompt_callback(ompt_callback_implicit_task)(
1864                   ompt_scope_end, NULL, &(task_info->task_data), 1,
1865                   OMPT_CUR_TASK_INFO(master_th)->thread_num,
1866                   ompt_task_implicit);
1867             }
1868 
1869             ompt_parallel_data = *OMPT_CUR_TEAM_DATA(master_th);
1870             __ompt_lw_taskteam_unlink(master_th);
1871             if (ompt_enabled.ompt_callback_parallel_end) {
1872               ompt_callbacks.ompt_callback(ompt_callback_parallel_end)(
1873                   &ompt_parallel_data, parent_task_data,
1874                   OMPT_INVOKER(call_context) | ompt_parallel_team,
1875                   return_address);
1876             }
1877             master_th->th.ompt_thread_info.state = ompt_state_overhead;
1878           }
1879 #endif
1880         }
1881       } else if (call_context == fork_context_gnu) {
1882 #if OMPT_SUPPORT
1883         ompt_lw_taskteam_t lwt;
1884         __ompt_lw_taskteam_init(&lwt, master_th, gtid, &ompt_parallel_data,
1885                                 return_address);
1886 
1887         lwt.ompt_task_info.frame.exit_frame = ompt_data_none;
1888         __ompt_lw_taskteam_link(&lwt, master_th, 1);
1889 // don't use lw_taskteam after linking. content was swaped
1890 #endif
1891 
1892         // we were called from GNU native code
1893         KA_TRACE(20, ("__kmp_fork_call: T#%d serial exit\n", gtid));
1894         return FALSE;
1895       } else {
1896         KMP_ASSERT2(call_context < fork_context_last,
1897                     "__kmp_fork_call: unknown fork_context parameter");
1898       }
1899 
1900       KA_TRACE(20, ("__kmp_fork_call: T#%d serial exit\n", gtid));
1901       KMP_MB();
1902       return FALSE;
1903     } // if (nthreads == 1)
1904 
1905     // GEH: only modify the executing flag in the case when not serialized
1906     //      serialized case is handled in kmpc_serialized_parallel
1907     KF_TRACE(10, ("__kmp_fork_call: parent_team_aclevel=%d, master_th=%p, "
1908                   "curtask=%p, curtask_max_aclevel=%d\n",
1909                   parent_team->t.t_active_level, master_th,
1910                   master_th->th.th_current_task,
1911                   master_th->th.th_current_task->td_icvs.max_active_levels));
1912     // TODO: GEH - cannot do this assertion because root thread not set up as
1913     // executing
1914     // KMP_ASSERT( master_th->th.th_current_task->td_flags.executing == 1 );
1915     master_th->th.th_current_task->td_flags.executing = 0;
1916 
1917     if (!master_th->th.th_teams_microtask || level > teams_level) {
1918       /* Increment our nested depth level */
1919       KMP_ATOMIC_INC(&root->r.r_in_parallel);
1920     }
1921 
1922     // See if we need to make a copy of the ICVs.
1923     int nthreads_icv = master_th->th.th_current_task->td_icvs.nproc;
1924     if ((level + 1 < __kmp_nested_nth.used) &&
1925         (__kmp_nested_nth.nth[level + 1] != nthreads_icv)) {
1926       nthreads_icv = __kmp_nested_nth.nth[level + 1];
1927     } else {
1928       nthreads_icv = 0; // don't update
1929     }
1930 
1931     // Figure out the proc_bind_policy for the new team.
1932     kmp_proc_bind_t proc_bind = master_th->th.th_set_proc_bind;
1933     kmp_proc_bind_t proc_bind_icv =
1934         proc_bind_default; // proc_bind_default means don't update
1935     if (master_th->th.th_current_task->td_icvs.proc_bind == proc_bind_false) {
1936       proc_bind = proc_bind_false;
1937     } else {
1938       if (proc_bind == proc_bind_default) {
1939         // No proc_bind clause specified; use current proc-bind-var for this
1940         // parallel region
1941         proc_bind = master_th->th.th_current_task->td_icvs.proc_bind;
1942       }
1943       /* else: The proc_bind policy was specified explicitly on parallel clause.
1944          This overrides proc-bind-var for this parallel region, but does not
1945          change proc-bind-var. */
1946       // Figure the value of proc-bind-var for the child threads.
1947       if ((level + 1 < __kmp_nested_proc_bind.used) &&
1948           (__kmp_nested_proc_bind.bind_types[level + 1] !=
1949            master_th->th.th_current_task->td_icvs.proc_bind)) {
1950         proc_bind_icv = __kmp_nested_proc_bind.bind_types[level + 1];
1951       }
1952     }
1953 
1954     // Reset for next parallel region
1955     master_th->th.th_set_proc_bind = proc_bind_default;
1956 
1957     if ((nthreads_icv > 0) || (proc_bind_icv != proc_bind_default)) {
1958       kmp_internal_control_t new_icvs;
1959       copy_icvs(&new_icvs, &master_th->th.th_current_task->td_icvs);
1960       new_icvs.next = NULL;
1961       if (nthreads_icv > 0) {
1962         new_icvs.nproc = nthreads_icv;
1963       }
1964       if (proc_bind_icv != proc_bind_default) {
1965         new_icvs.proc_bind = proc_bind_icv;
1966       }
1967 
1968       /* allocate a new parallel team */
1969       KF_TRACE(10, ("__kmp_fork_call: before __kmp_allocate_team\n"));
1970       team = __kmp_allocate_team(root, nthreads, nthreads,
1971 #if OMPT_SUPPORT
1972                                  ompt_parallel_data,
1973 #endif
1974                                  proc_bind, &new_icvs,
1975                                  argc USE_NESTED_HOT_ARG(master_th));
1976     } else {
1977       /* allocate a new parallel team */
1978       KF_TRACE(10, ("__kmp_fork_call: before __kmp_allocate_team\n"));
1979       team = __kmp_allocate_team(root, nthreads, nthreads,
1980 #if OMPT_SUPPORT
1981                                  ompt_parallel_data,
1982 #endif
1983                                  proc_bind,
1984                                  &master_th->th.th_current_task->td_icvs,
1985                                  argc USE_NESTED_HOT_ARG(master_th));
1986     }
1987     KF_TRACE(
1988         10, ("__kmp_fork_call: after __kmp_allocate_team - team = %p\n", team));
1989 
1990     /* setup the new team */
1991     KMP_CHECK_UPDATE(team->t.t_master_tid, master_tid);
1992     KMP_CHECK_UPDATE(team->t.t_master_this_cons, master_this_cons);
1993     KMP_CHECK_UPDATE(team->t.t_ident, loc);
1994     KMP_CHECK_UPDATE(team->t.t_parent, parent_team);
1995     KMP_CHECK_UPDATE_SYNC(team->t.t_pkfn, microtask);
1996 #if OMPT_SUPPORT
1997     KMP_CHECK_UPDATE_SYNC(team->t.ompt_team_info.master_return_address,
1998                           return_address);
1999 #endif
2000     KMP_CHECK_UPDATE(team->t.t_invoke, invoker); // TODO move to root, maybe
2001     // TODO: parent_team->t.t_level == INT_MAX ???
2002     if (!master_th->th.th_teams_microtask || level > teams_level) {
2003       int new_level = parent_team->t.t_level + 1;
2004       KMP_CHECK_UPDATE(team->t.t_level, new_level);
2005       new_level = parent_team->t.t_active_level + 1;
2006       KMP_CHECK_UPDATE(team->t.t_active_level, new_level);
2007     } else {
2008       // AC: Do not increase parallel level at start of the teams construct
2009       int new_level = parent_team->t.t_level;
2010       KMP_CHECK_UPDATE(team->t.t_level, new_level);
2011       new_level = parent_team->t.t_active_level;
2012       KMP_CHECK_UPDATE(team->t.t_active_level, new_level);
2013     }
2014     kmp_r_sched_t new_sched = get__sched_2(parent_team, master_tid);
2015     // set primary thread's schedule as new run-time schedule
2016     KMP_CHECK_UPDATE(team->t.t_sched.sched, new_sched.sched);
2017 
2018     KMP_CHECK_UPDATE(team->t.t_cancel_request, cancel_noreq);
2019     KMP_CHECK_UPDATE(team->t.t_def_allocator, master_th->th.th_def_allocator);
2020 
2021     // Update the floating point rounding in the team if required.
2022     propagateFPControl(team);
2023 
2024     if (__kmp_tasking_mode != tskm_immediate_exec) {
2025       // Set primary thread's task team to team's task team. Unless this is hot
2026       // team, it should be NULL.
2027       KMP_DEBUG_ASSERT(master_th->th.th_task_team ==
2028                        parent_team->t.t_task_team[master_th->th.th_task_state]);
2029       KA_TRACE(20, ("__kmp_fork_call: Primary T#%d pushing task_team %p / team "
2030                     "%p, new task_team %p / team %p\n",
2031                     __kmp_gtid_from_thread(master_th),
2032                     master_th->th.th_task_team, parent_team,
2033                     team->t.t_task_team[master_th->th.th_task_state], team));
2034 
2035       if (active_level || master_th->th.th_task_team) {
2036         // Take a memo of primary thread's task_state
2037         KMP_DEBUG_ASSERT(master_th->th.th_task_state_memo_stack);
2038         if (master_th->th.th_task_state_top >=
2039             master_th->th.th_task_state_stack_sz) { // increase size
2040           kmp_uint32 new_size = 2 * master_th->th.th_task_state_stack_sz;
2041           kmp_uint8 *old_stack, *new_stack;
2042           kmp_uint32 i;
2043           new_stack = (kmp_uint8 *)__kmp_allocate(new_size);
2044           for (i = 0; i < master_th->th.th_task_state_stack_sz; ++i) {
2045             new_stack[i] = master_th->th.th_task_state_memo_stack[i];
2046           }
2047           for (i = master_th->th.th_task_state_stack_sz; i < new_size;
2048                ++i) { // zero-init rest of stack
2049             new_stack[i] = 0;
2050           }
2051           old_stack = master_th->th.th_task_state_memo_stack;
2052           master_th->th.th_task_state_memo_stack = new_stack;
2053           master_th->th.th_task_state_stack_sz = new_size;
2054           __kmp_free(old_stack);
2055         }
2056         // Store primary thread's task_state on stack
2057         master_th->th
2058             .th_task_state_memo_stack[master_th->th.th_task_state_top] =
2059             master_th->th.th_task_state;
2060         master_th->th.th_task_state_top++;
2061 #if KMP_NESTED_HOT_TEAMS
2062         if (master_th->th.th_hot_teams &&
2063             active_level < __kmp_hot_teams_max_level &&
2064             team == master_th->th.th_hot_teams[active_level].hot_team) {
2065           // Restore primary thread's nested state if nested hot team
2066           master_th->th.th_task_state =
2067               master_th->th
2068                   .th_task_state_memo_stack[master_th->th.th_task_state_top];
2069         } else {
2070 #endif
2071           master_th->th.th_task_state = 0;
2072 #if KMP_NESTED_HOT_TEAMS
2073         }
2074 #endif
2075       }
2076 #if !KMP_NESTED_HOT_TEAMS
2077       KMP_DEBUG_ASSERT((master_th->th.th_task_team == NULL) ||
2078                        (team == root->r.r_hot_team));
2079 #endif
2080     }
2081 
2082     KA_TRACE(
2083         20,
2084         ("__kmp_fork_call: T#%d(%d:%d)->(%d:0) created a team of %d threads\n",
2085          gtid, parent_team->t.t_id, team->t.t_master_tid, team->t.t_id,
2086          team->t.t_nproc));
2087     KMP_DEBUG_ASSERT(team != root->r.r_hot_team ||
2088                      (team->t.t_master_tid == 0 &&
2089                       (team->t.t_parent == root->r.r_root_team ||
2090                        team->t.t_parent->t.t_serialized)));
2091     KMP_MB();
2092 
2093     /* now, setup the arguments */
2094     argv = (void **)team->t.t_argv;
2095     if (ap) {
2096       for (i = argc - 1; i >= 0; --i) {
2097         void *new_argv = va_arg(kmp_va_deref(ap), void *);
2098         KMP_CHECK_UPDATE(*argv, new_argv);
2099         argv++;
2100       }
2101     } else {
2102       for (i = 0; i < argc; ++i) {
2103         // Get args from parent team for teams construct
2104         KMP_CHECK_UPDATE(argv[i], team->t.t_parent->t.t_argv[i]);
2105       }
2106     }
2107 
2108     /* now actually fork the threads */
2109     KMP_CHECK_UPDATE(team->t.t_master_active, master_active);
2110     if (!root->r.r_active) // Only do assignment if it prevents cache ping-pong
2111       root->r.r_active = TRUE;
2112 
2113     __kmp_fork_team_threads(root, team, master_th, gtid);
2114     __kmp_setup_icv_copy(team, nthreads,
2115                          &master_th->th.th_current_task->td_icvs, loc);
2116 
2117 #if OMPT_SUPPORT
2118     master_th->th.ompt_thread_info.state = ompt_state_work_parallel;
2119 #endif
2120 
2121     __kmp_release_bootstrap_lock(&__kmp_forkjoin_lock);
2122 
2123 #if USE_ITT_BUILD
2124     if (team->t.t_active_level == 1 // only report frames at level 1
2125         && !master_th->th.th_teams_microtask) { // not in teams construct
2126 #if USE_ITT_NOTIFY
2127       if ((__itt_frame_submit_v3_ptr || KMP_ITT_DEBUG) &&
2128           (__kmp_forkjoin_frames_mode == 3 ||
2129            __kmp_forkjoin_frames_mode == 1)) {
2130         kmp_uint64 tmp_time = 0;
2131         if (__itt_get_timestamp_ptr)
2132           tmp_time = __itt_get_timestamp();
2133         // Internal fork - report frame begin
2134         master_th->th.th_frame_time = tmp_time;
2135         if (__kmp_forkjoin_frames_mode == 3)
2136           team->t.t_region_time = tmp_time;
2137       } else
2138 // only one notification scheme (either "submit" or "forking/joined", not both)
2139 #endif /* USE_ITT_NOTIFY */
2140           if ((__itt_frame_begin_v3_ptr || KMP_ITT_DEBUG) &&
2141               __kmp_forkjoin_frames && !__kmp_forkjoin_frames_mode) {
2142         // Mark start of "parallel" region for Intel(R) VTune(TM) analyzer.
2143         __kmp_itt_region_forking(gtid, team->t.t_nproc, 0);
2144       }
2145     }
2146 #endif /* USE_ITT_BUILD */
2147 
2148     /* now go on and do the work */
2149     KMP_DEBUG_ASSERT(team == __kmp_threads[gtid]->th.th_team);
2150     KMP_MB();
2151     KF_TRACE(10,
2152              ("__kmp_internal_fork : root=%p, team=%p, master_th=%p, gtid=%d\n",
2153               root, team, master_th, gtid));
2154 
2155 #if USE_ITT_BUILD
2156     if (__itt_stack_caller_create_ptr) {
2157       // create new stack stitching id before entering fork barrier
2158       if (!enter_teams) {
2159         KMP_DEBUG_ASSERT(team->t.t_stack_id == NULL);
2160         team->t.t_stack_id = __kmp_itt_stack_caller_create();
2161       } else if (parent_team->t.t_serialized) {
2162         // keep stack stitching id in the serialized parent_team;
2163         // current team will be used for parallel inside the teams;
2164         // if parent_team is active, then it already keeps stack stitching id
2165         // for the league of teams
2166         KMP_DEBUG_ASSERT(parent_team->t.t_stack_id == NULL);
2167         parent_team->t.t_stack_id = __kmp_itt_stack_caller_create();
2168       }
2169     }
2170 #endif /* USE_ITT_BUILD */
2171 
2172     // AC: skip __kmp_internal_fork at teams construct, let only primary
2173     // threads execute
2174     if (ap) {
2175       __kmp_internal_fork(loc, gtid, team);
2176       KF_TRACE(10, ("__kmp_internal_fork : after : root=%p, team=%p, "
2177                     "master_th=%p, gtid=%d\n",
2178                     root, team, master_th, gtid));
2179     }
2180 
2181     if (call_context == fork_context_gnu) {
2182       KA_TRACE(20, ("__kmp_fork_call: parallel exit T#%d\n", gtid));
2183       return TRUE;
2184     }
2185 
2186     /* Invoke microtask for PRIMARY thread */
2187     KA_TRACE(20, ("__kmp_fork_call: T#%d(%d:0) invoke microtask = %p\n", gtid,
2188                   team->t.t_id, team->t.t_pkfn));
2189   } // END of timer KMP_fork_call block
2190 
2191 #if KMP_STATS_ENABLED
2192   // If beginning a teams construct, then change thread state
2193   stats_state_e previous_state = KMP_GET_THREAD_STATE();
2194   if (!ap) {
2195     KMP_SET_THREAD_STATE(stats_state_e::TEAMS_REGION);
2196   }
2197 #endif
2198 
2199   if (!team->t.t_invoke(gtid)) {
2200     KMP_ASSERT2(0, "cannot invoke microtask for PRIMARY thread");
2201   }
2202 
2203 #if KMP_STATS_ENABLED
2204   // If was beginning of a teams construct, then reset thread state
2205   if (!ap) {
2206     KMP_SET_THREAD_STATE(previous_state);
2207   }
2208 #endif
2209 
2210   KA_TRACE(20, ("__kmp_fork_call: T#%d(%d:0) done microtask = %p\n", gtid,
2211                 team->t.t_id, team->t.t_pkfn));
2212   KMP_MB(); /* Flush all pending memory write invalidates.  */
2213 
2214   KA_TRACE(20, ("__kmp_fork_call: parallel exit T#%d\n", gtid));
2215 
2216 #if OMPT_SUPPORT
2217   if (ompt_enabled.enabled) {
2218     master_th->th.ompt_thread_info.state = ompt_state_overhead;
2219   }
2220 #endif
2221 
2222   return TRUE;
2223 }
2224 
2225 #if OMPT_SUPPORT
2226 static inline void __kmp_join_restore_state(kmp_info_t *thread,
2227                                             kmp_team_t *team) {
2228   // restore state outside the region
2229   thread->th.ompt_thread_info.state =
2230       ((team->t.t_serialized) ? ompt_state_work_serial
2231                               : ompt_state_work_parallel);
2232 }
2233 
2234 static inline void __kmp_join_ompt(int gtid, kmp_info_t *thread,
2235                                    kmp_team_t *team, ompt_data_t *parallel_data,
2236                                    int flags, void *codeptr) {
2237   ompt_task_info_t *task_info = __ompt_get_task_info_object(0);
2238   if (ompt_enabled.ompt_callback_parallel_end) {
2239     ompt_callbacks.ompt_callback(ompt_callback_parallel_end)(
2240         parallel_data, &(task_info->task_data), flags, codeptr);
2241   }
2242 
2243   task_info->frame.enter_frame = ompt_data_none;
2244   __kmp_join_restore_state(thread, team);
2245 }
2246 #endif
2247 
2248 void __kmp_join_call(ident_t *loc, int gtid
2249 #if OMPT_SUPPORT
2250                      ,
2251                      enum fork_context_e fork_context
2252 #endif
2253                      ,
2254                      int exit_teams) {
2255   KMP_TIME_DEVELOPER_PARTITIONED_BLOCK(KMP_join_call);
2256   kmp_team_t *team;
2257   kmp_team_t *parent_team;
2258   kmp_info_t *master_th;
2259   kmp_root_t *root;
2260   int master_active;
2261 
2262   KA_TRACE(20, ("__kmp_join_call: enter T#%d\n", gtid));
2263 
2264   /* setup current data */
2265   master_th = __kmp_threads[gtid];
2266   root = master_th->th.th_root;
2267   team = master_th->th.th_team;
2268   parent_team = team->t.t_parent;
2269 
2270   master_th->th.th_ident = loc;
2271 
2272 #if OMPT_SUPPORT
2273   void *team_microtask = (void *)team->t.t_pkfn;
2274   // For GOMP interface with serialized parallel, need the
2275   // __kmpc_end_serialized_parallel to call hooks for OMPT end-implicit-task
2276   // and end-parallel events.
2277   if (ompt_enabled.enabled &&
2278       !(team->t.t_serialized && fork_context == fork_context_gnu)) {
2279     master_th->th.ompt_thread_info.state = ompt_state_overhead;
2280   }
2281 #endif
2282 
2283 #if KMP_DEBUG
2284   if (__kmp_tasking_mode != tskm_immediate_exec && !exit_teams) {
2285     KA_TRACE(20, ("__kmp_join_call: T#%d, old team = %p old task_team = %p, "
2286                   "th_task_team = %p\n",
2287                   __kmp_gtid_from_thread(master_th), team,
2288                   team->t.t_task_team[master_th->th.th_task_state],
2289                   master_th->th.th_task_team));
2290     KMP_DEBUG_ASSERT(master_th->th.th_task_team ==
2291                      team->t.t_task_team[master_th->th.th_task_state]);
2292   }
2293 #endif
2294 
2295   if (team->t.t_serialized) {
2296     if (master_th->th.th_teams_microtask) {
2297       // We are in teams construct
2298       int level = team->t.t_level;
2299       int tlevel = master_th->th.th_teams_level;
2300       if (level == tlevel) {
2301         // AC: we haven't incremented it earlier at start of teams construct,
2302         //     so do it here - at the end of teams construct
2303         team->t.t_level++;
2304       } else if (level == tlevel + 1) {
2305         // AC: we are exiting parallel inside teams, need to increment
2306         // serialization in order to restore it in the next call to
2307         // __kmpc_end_serialized_parallel
2308         team->t.t_serialized++;
2309       }
2310     }
2311     __kmpc_end_serialized_parallel(loc, gtid);
2312 
2313 #if OMPT_SUPPORT
2314     if (ompt_enabled.enabled) {
2315       __kmp_join_restore_state(master_th, parent_team);
2316     }
2317 #endif
2318 
2319     return;
2320   }
2321 
2322   master_active = team->t.t_master_active;
2323 
2324   if (!exit_teams) {
2325     // AC: No barrier for internal teams at exit from teams construct.
2326     //     But there is barrier for external team (league).
2327     __kmp_internal_join(loc, gtid, team);
2328 #if USE_ITT_BUILD
2329     if (__itt_stack_caller_create_ptr) {
2330       KMP_DEBUG_ASSERT(team->t.t_stack_id != NULL);
2331       // destroy the stack stitching id after join barrier
2332       __kmp_itt_stack_caller_destroy((__itt_caller)team->t.t_stack_id);
2333       team->t.t_stack_id = NULL;
2334     }
2335 #endif
2336   } else {
2337     master_th->th.th_task_state =
2338         0; // AC: no tasking in teams (out of any parallel)
2339 #if USE_ITT_BUILD
2340     if (__itt_stack_caller_create_ptr && parent_team->t.t_serialized) {
2341       KMP_DEBUG_ASSERT(parent_team->t.t_stack_id != NULL);
2342       // destroy the stack stitching id on exit from the teams construct
2343       // if parent_team is active, then the id will be destroyed later on
2344       // by master of the league of teams
2345       __kmp_itt_stack_caller_destroy((__itt_caller)parent_team->t.t_stack_id);
2346       parent_team->t.t_stack_id = NULL;
2347     }
2348 #endif
2349   }
2350 
2351   KMP_MB();
2352 
2353 #if OMPT_SUPPORT
2354   ompt_data_t *parallel_data = &(team->t.ompt_team_info.parallel_data);
2355   void *codeptr = team->t.ompt_team_info.master_return_address;
2356 #endif
2357 
2358 #if USE_ITT_BUILD
2359   // Mark end of "parallel" region for Intel(R) VTune(TM) analyzer.
2360   if (team->t.t_active_level == 1 &&
2361       (!master_th->th.th_teams_microtask || /* not in teams construct */
2362        master_th->th.th_teams_size.nteams == 1)) {
2363     master_th->th.th_ident = loc;
2364     // only one notification scheme (either "submit" or "forking/joined", not
2365     // both)
2366     if ((__itt_frame_submit_v3_ptr || KMP_ITT_DEBUG) &&
2367         __kmp_forkjoin_frames_mode == 3)
2368       __kmp_itt_frame_submit(gtid, team->t.t_region_time,
2369                              master_th->th.th_frame_time, 0, loc,
2370                              master_th->th.th_team_nproc, 1);
2371     else if ((__itt_frame_end_v3_ptr || KMP_ITT_DEBUG) &&
2372              !__kmp_forkjoin_frames_mode && __kmp_forkjoin_frames)
2373       __kmp_itt_region_joined(gtid);
2374   } // active_level == 1
2375 #endif /* USE_ITT_BUILD */
2376 
2377   if (master_th->th.th_teams_microtask && !exit_teams &&
2378       team->t.t_pkfn != (microtask_t)__kmp_teams_master &&
2379       team->t.t_level == master_th->th.th_teams_level + 1) {
2380 // AC: We need to leave the team structure intact at the end of parallel
2381 // inside the teams construct, so that at the next parallel same (hot) team
2382 // works, only adjust nesting levels
2383 #if OMPT_SUPPORT
2384     ompt_data_t ompt_parallel_data = ompt_data_none;
2385     if (ompt_enabled.enabled) {
2386       ompt_task_info_t *task_info = __ompt_get_task_info_object(0);
2387       if (ompt_enabled.ompt_callback_implicit_task) {
2388         int ompt_team_size = team->t.t_nproc;
2389         ompt_callbacks.ompt_callback(ompt_callback_implicit_task)(
2390             ompt_scope_end, NULL, &(task_info->task_data), ompt_team_size,
2391             OMPT_CUR_TASK_INFO(master_th)->thread_num, ompt_task_implicit);
2392       }
2393       task_info->frame.exit_frame = ompt_data_none;
2394       task_info->task_data = ompt_data_none;
2395       ompt_parallel_data = *OMPT_CUR_TEAM_DATA(master_th);
2396       __ompt_lw_taskteam_unlink(master_th);
2397     }
2398 #endif
2399     /* Decrement our nested depth level */
2400     team->t.t_level--;
2401     team->t.t_active_level--;
2402     KMP_ATOMIC_DEC(&root->r.r_in_parallel);
2403 
2404     // Restore number of threads in the team if needed. This code relies on
2405     // the proper adjustment of th_teams_size.nth after the fork in
2406     // __kmp_teams_master on each teams primary thread in the case that
2407     // __kmp_reserve_threads reduced it.
2408     if (master_th->th.th_team_nproc < master_th->th.th_teams_size.nth) {
2409       int old_num = master_th->th.th_team_nproc;
2410       int new_num = master_th->th.th_teams_size.nth;
2411       kmp_info_t **other_threads = team->t.t_threads;
2412       team->t.t_nproc = new_num;
2413       for (int i = 0; i < old_num; ++i) {
2414         other_threads[i]->th.th_team_nproc = new_num;
2415       }
2416       // Adjust states of non-used threads of the team
2417       for (int i = old_num; i < new_num; ++i) {
2418         // Re-initialize thread's barrier data.
2419         KMP_DEBUG_ASSERT(other_threads[i]);
2420         kmp_balign_t *balign = other_threads[i]->th.th_bar;
2421         for (int b = 0; b < bs_last_barrier; ++b) {
2422           balign[b].bb.b_arrived = team->t.t_bar[b].b_arrived;
2423           KMP_DEBUG_ASSERT(balign[b].bb.wait_flag != KMP_BARRIER_PARENT_FLAG);
2424 #if USE_DEBUGGER
2425           balign[b].bb.b_worker_arrived = team->t.t_bar[b].b_team_arrived;
2426 #endif
2427         }
2428         if (__kmp_tasking_mode != tskm_immediate_exec) {
2429           // Synchronize thread's task state
2430           other_threads[i]->th.th_task_state = master_th->th.th_task_state;
2431         }
2432       }
2433     }
2434 
2435 #if OMPT_SUPPORT
2436     if (ompt_enabled.enabled) {
2437       __kmp_join_ompt(gtid, master_th, parent_team, &ompt_parallel_data,
2438                       OMPT_INVOKER(fork_context) | ompt_parallel_team, codeptr);
2439     }
2440 #endif
2441 
2442     return;
2443   }
2444 
2445   /* do cleanup and restore the parent team */
2446   master_th->th.th_info.ds.ds_tid = team->t.t_master_tid;
2447   master_th->th.th_local.this_construct = team->t.t_master_this_cons;
2448 
2449   master_th->th.th_dispatch = &parent_team->t.t_dispatch[team->t.t_master_tid];
2450 
2451   /* jc: The following lock has instructions with REL and ACQ semantics,
2452      separating the parallel user code called in this parallel region
2453      from the serial user code called after this function returns. */
2454   __kmp_acquire_bootstrap_lock(&__kmp_forkjoin_lock);
2455 
2456   if (!master_th->th.th_teams_microtask ||
2457       team->t.t_level > master_th->th.th_teams_level) {
2458     /* Decrement our nested depth level */
2459     KMP_ATOMIC_DEC(&root->r.r_in_parallel);
2460   }
2461   KMP_DEBUG_ASSERT(root->r.r_in_parallel >= 0);
2462 
2463 #if OMPT_SUPPORT
2464   if (ompt_enabled.enabled) {
2465     ompt_task_info_t *task_info = __ompt_get_task_info_object(0);
2466     if (ompt_enabled.ompt_callback_implicit_task) {
2467       int flags = (team_microtask == (void *)__kmp_teams_master)
2468                       ? ompt_task_initial
2469                       : ompt_task_implicit;
2470       int ompt_team_size = (flags == ompt_task_initial) ? 0 : team->t.t_nproc;
2471       ompt_callbacks.ompt_callback(ompt_callback_implicit_task)(
2472           ompt_scope_end, NULL, &(task_info->task_data), ompt_team_size,
2473           OMPT_CUR_TASK_INFO(master_th)->thread_num, flags);
2474     }
2475     task_info->frame.exit_frame = ompt_data_none;
2476     task_info->task_data = ompt_data_none;
2477   }
2478 #endif
2479 
2480   KF_TRACE(10, ("__kmp_join_call1: T#%d, this_thread=%p team=%p\n", 0,
2481                 master_th, team));
2482   __kmp_pop_current_task_from_thread(master_th);
2483 
2484 #if KMP_AFFINITY_SUPPORTED
2485   // Restore master thread's partition.
2486   master_th->th.th_first_place = team->t.t_first_place;
2487   master_th->th.th_last_place = team->t.t_last_place;
2488 #endif // KMP_AFFINITY_SUPPORTED
2489   master_th->th.th_def_allocator = team->t.t_def_allocator;
2490 
2491   updateHWFPControl(team);
2492 
2493   if (root->r.r_active != master_active)
2494     root->r.r_active = master_active;
2495 
2496   __kmp_free_team(root, team USE_NESTED_HOT_ARG(
2497                             master_th)); // this will free worker threads
2498 
2499   /* this race was fun to find. make sure the following is in the critical
2500      region otherwise assertions may fail occasionally since the old team may be
2501      reallocated and the hierarchy appears inconsistent. it is actually safe to
2502      run and won't cause any bugs, but will cause those assertion failures. it's
2503      only one deref&assign so might as well put this in the critical region */
2504   master_th->th.th_team = parent_team;
2505   master_th->th.th_team_nproc = parent_team->t.t_nproc;
2506   master_th->th.th_team_master = parent_team->t.t_threads[0];
2507   master_th->th.th_team_serialized = parent_team->t.t_serialized;
2508 
2509   /* restore serialized team, if need be */
2510   if (parent_team->t.t_serialized &&
2511       parent_team != master_th->th.th_serial_team &&
2512       parent_team != root->r.r_root_team) {
2513     __kmp_free_team(root,
2514                     master_th->th.th_serial_team USE_NESTED_HOT_ARG(NULL));
2515     master_th->th.th_serial_team = parent_team;
2516   }
2517 
2518   if (__kmp_tasking_mode != tskm_immediate_exec) {
2519     if (master_th->th.th_task_state_top >
2520         0) { // Restore task state from memo stack
2521       KMP_DEBUG_ASSERT(master_th->th.th_task_state_memo_stack);
2522       // Remember primary thread's state if we re-use this nested hot team
2523       master_th->th.th_task_state_memo_stack[master_th->th.th_task_state_top] =
2524           master_th->th.th_task_state;
2525       --master_th->th.th_task_state_top; // pop
2526       // Now restore state at this level
2527       master_th->th.th_task_state =
2528           master_th->th
2529               .th_task_state_memo_stack[master_th->th.th_task_state_top];
2530     }
2531     // Copy the task team from the parent team to the primary thread
2532     master_th->th.th_task_team =
2533         parent_team->t.t_task_team[master_th->th.th_task_state];
2534     KA_TRACE(20,
2535              ("__kmp_join_call: Primary T#%d restoring task_team %p, team %p\n",
2536               __kmp_gtid_from_thread(master_th), master_th->th.th_task_team,
2537               parent_team));
2538   }
2539 
2540   // TODO: GEH - cannot do this assertion because root thread not set up as
2541   // executing
2542   // KMP_ASSERT( master_th->th.th_current_task->td_flags.executing == 0 );
2543   master_th->th.th_current_task->td_flags.executing = 1;
2544 
2545   __kmp_release_bootstrap_lock(&__kmp_forkjoin_lock);
2546 
2547 #if OMPT_SUPPORT
2548   int flags =
2549       OMPT_INVOKER(fork_context) |
2550       ((team_microtask == (void *)__kmp_teams_master) ? ompt_parallel_league
2551                                                       : ompt_parallel_team);
2552   if (ompt_enabled.enabled) {
2553     __kmp_join_ompt(gtid, master_th, parent_team, parallel_data, flags,
2554                     codeptr);
2555   }
2556 #endif
2557 
2558   KMP_MB();
2559   KA_TRACE(20, ("__kmp_join_call: exit T#%d\n", gtid));
2560 }
2561 
2562 /* Check whether we should push an internal control record onto the
2563    serial team stack.  If so, do it.  */
2564 void __kmp_save_internal_controls(kmp_info_t *thread) {
2565 
2566   if (thread->th.th_team != thread->th.th_serial_team) {
2567     return;
2568   }
2569   if (thread->th.th_team->t.t_serialized > 1) {
2570     int push = 0;
2571 
2572     if (thread->th.th_team->t.t_control_stack_top == NULL) {
2573       push = 1;
2574     } else {
2575       if (thread->th.th_team->t.t_control_stack_top->serial_nesting_level !=
2576           thread->th.th_team->t.t_serialized) {
2577         push = 1;
2578       }
2579     }
2580     if (push) { /* push a record on the serial team's stack */
2581       kmp_internal_control_t *control =
2582           (kmp_internal_control_t *)__kmp_allocate(
2583               sizeof(kmp_internal_control_t));
2584 
2585       copy_icvs(control, &thread->th.th_current_task->td_icvs);
2586 
2587       control->serial_nesting_level = thread->th.th_team->t.t_serialized;
2588 
2589       control->next = thread->th.th_team->t.t_control_stack_top;
2590       thread->th.th_team->t.t_control_stack_top = control;
2591     }
2592   }
2593 }
2594 
2595 /* Changes set_nproc */
2596 void __kmp_set_num_threads(int new_nth, int gtid) {
2597   kmp_info_t *thread;
2598   kmp_root_t *root;
2599 
2600   KF_TRACE(10, ("__kmp_set_num_threads: new __kmp_nth = %d\n", new_nth));
2601   KMP_DEBUG_ASSERT(__kmp_init_serial);
2602 
2603   if (new_nth < 1)
2604     new_nth = 1;
2605   else if (new_nth > __kmp_max_nth)
2606     new_nth = __kmp_max_nth;
2607 
2608   KMP_COUNT_VALUE(OMP_set_numthreads, new_nth);
2609   thread = __kmp_threads[gtid];
2610   if (thread->th.th_current_task->td_icvs.nproc == new_nth)
2611     return; // nothing to do
2612 
2613   __kmp_save_internal_controls(thread);
2614 
2615   set__nproc(thread, new_nth);
2616 
2617   // If this omp_set_num_threads() call will cause the hot team size to be
2618   // reduced (in the absence of a num_threads clause), then reduce it now,
2619   // rather than waiting for the next parallel region.
2620   root = thread->th.th_root;
2621   if (__kmp_init_parallel && (!root->r.r_active) &&
2622       (root->r.r_hot_team->t.t_nproc > new_nth)
2623 #if KMP_NESTED_HOT_TEAMS
2624       && __kmp_hot_teams_max_level && !__kmp_hot_teams_mode
2625 #endif
2626   ) {
2627     kmp_team_t *hot_team = root->r.r_hot_team;
2628     int f;
2629 
2630     __kmp_acquire_bootstrap_lock(&__kmp_forkjoin_lock);
2631 
2632     // Release the extra threads we don't need any more.
2633     for (f = new_nth; f < hot_team->t.t_nproc; f++) {
2634       KMP_DEBUG_ASSERT(hot_team->t.t_threads[f] != NULL);
2635       if (__kmp_tasking_mode != tskm_immediate_exec) {
2636         // When decreasing team size, threads no longer in the team should unref
2637         // task team.
2638         hot_team->t.t_threads[f]->th.th_task_team = NULL;
2639       }
2640       __kmp_free_thread(hot_team->t.t_threads[f]);
2641       hot_team->t.t_threads[f] = NULL;
2642     }
2643     hot_team->t.t_nproc = new_nth;
2644 #if KMP_NESTED_HOT_TEAMS
2645     if (thread->th.th_hot_teams) {
2646       KMP_DEBUG_ASSERT(hot_team == thread->th.th_hot_teams[0].hot_team);
2647       thread->th.th_hot_teams[0].hot_team_nth = new_nth;
2648     }
2649 #endif
2650 
2651     __kmp_release_bootstrap_lock(&__kmp_forkjoin_lock);
2652 
2653     // Update the t_nproc field in the threads that are still active.
2654     for (f = 0; f < new_nth; f++) {
2655       KMP_DEBUG_ASSERT(hot_team->t.t_threads[f] != NULL);
2656       hot_team->t.t_threads[f]->th.th_team_nproc = new_nth;
2657     }
2658     // Special flag in case omp_set_num_threads() call
2659     hot_team->t.t_size_changed = -1;
2660   }
2661 }
2662 
2663 /* Changes max_active_levels */
2664 void __kmp_set_max_active_levels(int gtid, int max_active_levels) {
2665   kmp_info_t *thread;
2666 
2667   KF_TRACE(10, ("__kmp_set_max_active_levels: new max_active_levels for thread "
2668                 "%d = (%d)\n",
2669                 gtid, max_active_levels));
2670   KMP_DEBUG_ASSERT(__kmp_init_serial);
2671 
2672   // validate max_active_levels
2673   if (max_active_levels < 0) {
2674     KMP_WARNING(ActiveLevelsNegative, max_active_levels);
2675     // We ignore this call if the user has specified a negative value.
2676     // The current setting won't be changed. The last valid setting will be
2677     // used. A warning will be issued (if warnings are allowed as controlled by
2678     // the KMP_WARNINGS env var).
2679     KF_TRACE(10, ("__kmp_set_max_active_levels: the call is ignored: new "
2680                   "max_active_levels for thread %d = (%d)\n",
2681                   gtid, max_active_levels));
2682     return;
2683   }
2684   if (max_active_levels <= KMP_MAX_ACTIVE_LEVELS_LIMIT) {
2685     // it's OK, the max_active_levels is within the valid range: [ 0;
2686     // KMP_MAX_ACTIVE_LEVELS_LIMIT ]
2687     // We allow a zero value. (implementation defined behavior)
2688   } else {
2689     KMP_WARNING(ActiveLevelsExceedLimit, max_active_levels,
2690                 KMP_MAX_ACTIVE_LEVELS_LIMIT);
2691     max_active_levels = KMP_MAX_ACTIVE_LEVELS_LIMIT;
2692     // Current upper limit is MAX_INT. (implementation defined behavior)
2693     // If the input exceeds the upper limit, we correct the input to be the
2694     // upper limit. (implementation defined behavior)
2695     // Actually, the flow should never get here until we use MAX_INT limit.
2696   }
2697   KF_TRACE(10, ("__kmp_set_max_active_levels: after validation: new "
2698                 "max_active_levels for thread %d = (%d)\n",
2699                 gtid, max_active_levels));
2700 
2701   thread = __kmp_threads[gtid];
2702 
2703   __kmp_save_internal_controls(thread);
2704 
2705   set__max_active_levels(thread, max_active_levels);
2706 }
2707 
2708 /* Gets max_active_levels */
2709 int __kmp_get_max_active_levels(int gtid) {
2710   kmp_info_t *thread;
2711 
2712   KF_TRACE(10, ("__kmp_get_max_active_levels: thread %d\n", gtid));
2713   KMP_DEBUG_ASSERT(__kmp_init_serial);
2714 
2715   thread = __kmp_threads[gtid];
2716   KMP_DEBUG_ASSERT(thread->th.th_current_task);
2717   KF_TRACE(10, ("__kmp_get_max_active_levels: thread %d, curtask=%p, "
2718                 "curtask_maxaclevel=%d\n",
2719                 gtid, thread->th.th_current_task,
2720                 thread->th.th_current_task->td_icvs.max_active_levels));
2721   return thread->th.th_current_task->td_icvs.max_active_levels;
2722 }
2723 
2724 // nteams-var per-device ICV
2725 void __kmp_set_num_teams(int num_teams) {
2726   if (num_teams > 0)
2727     __kmp_nteams = num_teams;
2728 }
2729 int __kmp_get_max_teams(void) { return __kmp_nteams; }
2730 // teams-thread-limit-var per-device ICV
2731 void __kmp_set_teams_thread_limit(int limit) {
2732   if (limit > 0)
2733     __kmp_teams_thread_limit = limit;
2734 }
2735 int __kmp_get_teams_thread_limit(void) { return __kmp_teams_thread_limit; }
2736 
2737 KMP_BUILD_ASSERT(sizeof(kmp_sched_t) == sizeof(int));
2738 KMP_BUILD_ASSERT(sizeof(enum sched_type) == sizeof(int));
2739 
2740 /* Changes def_sched_var ICV values (run-time schedule kind and chunk) */
2741 void __kmp_set_schedule(int gtid, kmp_sched_t kind, int chunk) {
2742   kmp_info_t *thread;
2743   kmp_sched_t orig_kind;
2744   //    kmp_team_t *team;
2745 
2746   KF_TRACE(10, ("__kmp_set_schedule: new schedule for thread %d = (%d, %d)\n",
2747                 gtid, (int)kind, chunk));
2748   KMP_DEBUG_ASSERT(__kmp_init_serial);
2749 
2750   // Check if the kind parameter is valid, correct if needed.
2751   // Valid parameters should fit in one of two intervals - standard or extended:
2752   //       <lower>, <valid>, <upper_std>, <lower_ext>, <valid>, <upper>
2753   // 2008-01-25: 0,  1 - 4,       5,         100,     101 - 102, 103
2754   orig_kind = kind;
2755   kind = __kmp_sched_without_mods(kind);
2756 
2757   if (kind <= kmp_sched_lower || kind >= kmp_sched_upper ||
2758       (kind <= kmp_sched_lower_ext && kind >= kmp_sched_upper_std)) {
2759     // TODO: Hint needs attention in case we change the default schedule.
2760     __kmp_msg(kmp_ms_warning, KMP_MSG(ScheduleKindOutOfRange, kind),
2761               KMP_HNT(DefaultScheduleKindUsed, "static, no chunk"),
2762               __kmp_msg_null);
2763     kind = kmp_sched_default;
2764     chunk = 0; // ignore chunk value in case of bad kind
2765   }
2766 
2767   thread = __kmp_threads[gtid];
2768 
2769   __kmp_save_internal_controls(thread);
2770 
2771   if (kind < kmp_sched_upper_std) {
2772     if (kind == kmp_sched_static && chunk < KMP_DEFAULT_CHUNK) {
2773       // differ static chunked vs. unchunked:  chunk should be invalid to
2774       // indicate unchunked schedule (which is the default)
2775       thread->th.th_current_task->td_icvs.sched.r_sched_type = kmp_sch_static;
2776     } else {
2777       thread->th.th_current_task->td_icvs.sched.r_sched_type =
2778           __kmp_sch_map[kind - kmp_sched_lower - 1];
2779     }
2780   } else {
2781     //    __kmp_sch_map[ kind - kmp_sched_lower_ext + kmp_sched_upper_std -
2782     //    kmp_sched_lower - 2 ];
2783     thread->th.th_current_task->td_icvs.sched.r_sched_type =
2784         __kmp_sch_map[kind - kmp_sched_lower_ext + kmp_sched_upper_std -
2785                       kmp_sched_lower - 2];
2786   }
2787   __kmp_sched_apply_mods_intkind(
2788       orig_kind, &(thread->th.th_current_task->td_icvs.sched.r_sched_type));
2789   if (kind == kmp_sched_auto || chunk < 1) {
2790     // ignore parameter chunk for schedule auto
2791     thread->th.th_current_task->td_icvs.sched.chunk = KMP_DEFAULT_CHUNK;
2792   } else {
2793     thread->th.th_current_task->td_icvs.sched.chunk = chunk;
2794   }
2795 }
2796 
2797 /* Gets def_sched_var ICV values */
2798 void __kmp_get_schedule(int gtid, kmp_sched_t *kind, int *chunk) {
2799   kmp_info_t *thread;
2800   enum sched_type th_type;
2801 
2802   KF_TRACE(10, ("__kmp_get_schedule: thread %d\n", gtid));
2803   KMP_DEBUG_ASSERT(__kmp_init_serial);
2804 
2805   thread = __kmp_threads[gtid];
2806 
2807   th_type = thread->th.th_current_task->td_icvs.sched.r_sched_type;
2808   switch (SCHEDULE_WITHOUT_MODIFIERS(th_type)) {
2809   case kmp_sch_static:
2810   case kmp_sch_static_greedy:
2811   case kmp_sch_static_balanced:
2812     *kind = kmp_sched_static;
2813     __kmp_sched_apply_mods_stdkind(kind, th_type);
2814     *chunk = 0; // chunk was not set, try to show this fact via zero value
2815     return;
2816   case kmp_sch_static_chunked:
2817     *kind = kmp_sched_static;
2818     break;
2819   case kmp_sch_dynamic_chunked:
2820     *kind = kmp_sched_dynamic;
2821     break;
2822   case kmp_sch_guided_chunked:
2823   case kmp_sch_guided_iterative_chunked:
2824   case kmp_sch_guided_analytical_chunked:
2825     *kind = kmp_sched_guided;
2826     break;
2827   case kmp_sch_auto:
2828     *kind = kmp_sched_auto;
2829     break;
2830   case kmp_sch_trapezoidal:
2831     *kind = kmp_sched_trapezoidal;
2832     break;
2833 #if KMP_STATIC_STEAL_ENABLED
2834   case kmp_sch_static_steal:
2835     *kind = kmp_sched_static_steal;
2836     break;
2837 #endif
2838   default:
2839     KMP_FATAL(UnknownSchedulingType, th_type);
2840   }
2841 
2842   __kmp_sched_apply_mods_stdkind(kind, th_type);
2843   *chunk = thread->th.th_current_task->td_icvs.sched.chunk;
2844 }
2845 
2846 int __kmp_get_ancestor_thread_num(int gtid, int level) {
2847 
2848   int ii, dd;
2849   kmp_team_t *team;
2850   kmp_info_t *thr;
2851 
2852   KF_TRACE(10, ("__kmp_get_ancestor_thread_num: thread %d %d\n", gtid, level));
2853   KMP_DEBUG_ASSERT(__kmp_init_serial);
2854 
2855   // validate level
2856   if (level == 0)
2857     return 0;
2858   if (level < 0)
2859     return -1;
2860   thr = __kmp_threads[gtid];
2861   team = thr->th.th_team;
2862   ii = team->t.t_level;
2863   if (level > ii)
2864     return -1;
2865 
2866   if (thr->th.th_teams_microtask) {
2867     // AC: we are in teams region where multiple nested teams have same level
2868     int tlevel = thr->th.th_teams_level; // the level of the teams construct
2869     if (level <=
2870         tlevel) { // otherwise usual algorithm works (will not touch the teams)
2871       KMP_DEBUG_ASSERT(ii >= tlevel);
2872       // AC: As we need to pass by the teams league, we need to artificially
2873       // increase ii
2874       if (ii == tlevel) {
2875         ii += 2; // three teams have same level
2876       } else {
2877         ii++; // two teams have same level
2878       }
2879     }
2880   }
2881 
2882   if (ii == level)
2883     return __kmp_tid_from_gtid(gtid);
2884 
2885   dd = team->t.t_serialized;
2886   level++;
2887   while (ii > level) {
2888     for (dd = team->t.t_serialized; (dd > 0) && (ii > level); dd--, ii--) {
2889     }
2890     if ((team->t.t_serialized) && (!dd)) {
2891       team = team->t.t_parent;
2892       continue;
2893     }
2894     if (ii > level) {
2895       team = team->t.t_parent;
2896       dd = team->t.t_serialized;
2897       ii--;
2898     }
2899   }
2900 
2901   return (dd > 1) ? (0) : (team->t.t_master_tid);
2902 }
2903 
2904 int __kmp_get_team_size(int gtid, int level) {
2905 
2906   int ii, dd;
2907   kmp_team_t *team;
2908   kmp_info_t *thr;
2909 
2910   KF_TRACE(10, ("__kmp_get_team_size: thread %d %d\n", gtid, level));
2911   KMP_DEBUG_ASSERT(__kmp_init_serial);
2912 
2913   // validate level
2914   if (level == 0)
2915     return 1;
2916   if (level < 0)
2917     return -1;
2918   thr = __kmp_threads[gtid];
2919   team = thr->th.th_team;
2920   ii = team->t.t_level;
2921   if (level > ii)
2922     return -1;
2923 
2924   if (thr->th.th_teams_microtask) {
2925     // AC: we are in teams region where multiple nested teams have same level
2926     int tlevel = thr->th.th_teams_level; // the level of the teams construct
2927     if (level <=
2928         tlevel) { // otherwise usual algorithm works (will not touch the teams)
2929       KMP_DEBUG_ASSERT(ii >= tlevel);
2930       // AC: As we need to pass by the teams league, we need to artificially
2931       // increase ii
2932       if (ii == tlevel) {
2933         ii += 2; // three teams have same level
2934       } else {
2935         ii++; // two teams have same level
2936       }
2937     }
2938   }
2939 
2940   while (ii > level) {
2941     for (dd = team->t.t_serialized; (dd > 0) && (ii > level); dd--, ii--) {
2942     }
2943     if (team->t.t_serialized && (!dd)) {
2944       team = team->t.t_parent;
2945       continue;
2946     }
2947     if (ii > level) {
2948       team = team->t.t_parent;
2949       ii--;
2950     }
2951   }
2952 
2953   return team->t.t_nproc;
2954 }
2955 
2956 kmp_r_sched_t __kmp_get_schedule_global() {
2957   // This routine created because pairs (__kmp_sched, __kmp_chunk) and
2958   // (__kmp_static, __kmp_guided) may be changed by kmp_set_defaults
2959   // independently. So one can get the updated schedule here.
2960 
2961   kmp_r_sched_t r_sched;
2962 
2963   // create schedule from 4 globals: __kmp_sched, __kmp_chunk, __kmp_static,
2964   // __kmp_guided. __kmp_sched should keep original value, so that user can set
2965   // KMP_SCHEDULE multiple times, and thus have different run-time schedules in
2966   // different roots (even in OMP 2.5)
2967   enum sched_type s = SCHEDULE_WITHOUT_MODIFIERS(__kmp_sched);
2968   enum sched_type sched_modifiers = SCHEDULE_GET_MODIFIERS(__kmp_sched);
2969   if (s == kmp_sch_static) {
2970     // replace STATIC with more detailed schedule (balanced or greedy)
2971     r_sched.r_sched_type = __kmp_static;
2972   } else if (s == kmp_sch_guided_chunked) {
2973     // replace GUIDED with more detailed schedule (iterative or analytical)
2974     r_sched.r_sched_type = __kmp_guided;
2975   } else { // (STATIC_CHUNKED), or (DYNAMIC_CHUNKED), or other
2976     r_sched.r_sched_type = __kmp_sched;
2977   }
2978   SCHEDULE_SET_MODIFIERS(r_sched.r_sched_type, sched_modifiers);
2979 
2980   if (__kmp_chunk < KMP_DEFAULT_CHUNK) {
2981     // __kmp_chunk may be wrong here (if it was not ever set)
2982     r_sched.chunk = KMP_DEFAULT_CHUNK;
2983   } else {
2984     r_sched.chunk = __kmp_chunk;
2985   }
2986 
2987   return r_sched;
2988 }
2989 
2990 /* Allocate (realloc == FALSE) * or reallocate (realloc == TRUE)
2991    at least argc number of *t_argv entries for the requested team. */
2992 static void __kmp_alloc_argv_entries(int argc, kmp_team_t *team, int realloc) {
2993 
2994   KMP_DEBUG_ASSERT(team);
2995   if (!realloc || argc > team->t.t_max_argc) {
2996 
2997     KA_TRACE(100, ("__kmp_alloc_argv_entries: team %d: needed entries=%d, "
2998                    "current entries=%d\n",
2999                    team->t.t_id, argc, (realloc) ? team->t.t_max_argc : 0));
3000     /* if previously allocated heap space for args, free them */
3001     if (realloc && team->t.t_argv != &team->t.t_inline_argv[0])
3002       __kmp_free((void *)team->t.t_argv);
3003 
3004     if (argc <= KMP_INLINE_ARGV_ENTRIES) {
3005       /* use unused space in the cache line for arguments */
3006       team->t.t_max_argc = KMP_INLINE_ARGV_ENTRIES;
3007       KA_TRACE(100, ("__kmp_alloc_argv_entries: team %d: inline allocate %d "
3008                      "argv entries\n",
3009                      team->t.t_id, team->t.t_max_argc));
3010       team->t.t_argv = &team->t.t_inline_argv[0];
3011       if (__kmp_storage_map) {
3012         __kmp_print_storage_map_gtid(
3013             -1, &team->t.t_inline_argv[0],
3014             &team->t.t_inline_argv[KMP_INLINE_ARGV_ENTRIES],
3015             (sizeof(void *) * KMP_INLINE_ARGV_ENTRIES), "team_%d.t_inline_argv",
3016             team->t.t_id);
3017       }
3018     } else {
3019       /* allocate space for arguments in the heap */
3020       team->t.t_max_argc = (argc <= (KMP_MIN_MALLOC_ARGV_ENTRIES >> 1))
3021                                ? KMP_MIN_MALLOC_ARGV_ENTRIES
3022                                : 2 * argc;
3023       KA_TRACE(100, ("__kmp_alloc_argv_entries: team %d: dynamic allocate %d "
3024                      "argv entries\n",
3025                      team->t.t_id, team->t.t_max_argc));
3026       team->t.t_argv =
3027           (void **)__kmp_page_allocate(sizeof(void *) * team->t.t_max_argc);
3028       if (__kmp_storage_map) {
3029         __kmp_print_storage_map_gtid(-1, &team->t.t_argv[0],
3030                                      &team->t.t_argv[team->t.t_max_argc],
3031                                      sizeof(void *) * team->t.t_max_argc,
3032                                      "team_%d.t_argv", team->t.t_id);
3033       }
3034     }
3035   }
3036 }
3037 
3038 static void __kmp_allocate_team_arrays(kmp_team_t *team, int max_nth) {
3039   int i;
3040   int num_disp_buff = max_nth > 1 ? __kmp_dispatch_num_buffers : 2;
3041   team->t.t_threads =
3042       (kmp_info_t **)__kmp_allocate(sizeof(kmp_info_t *) * max_nth);
3043   team->t.t_disp_buffer = (dispatch_shared_info_t *)__kmp_allocate(
3044       sizeof(dispatch_shared_info_t) * num_disp_buff);
3045   team->t.t_dispatch =
3046       (kmp_disp_t *)__kmp_allocate(sizeof(kmp_disp_t) * max_nth);
3047   team->t.t_implicit_task_taskdata =
3048       (kmp_taskdata_t *)__kmp_allocate(sizeof(kmp_taskdata_t) * max_nth);
3049   team->t.t_max_nproc = max_nth;
3050 
3051   /* setup dispatch buffers */
3052   for (i = 0; i < num_disp_buff; ++i) {
3053     team->t.t_disp_buffer[i].buffer_index = i;
3054     team->t.t_disp_buffer[i].doacross_buf_idx = i;
3055   }
3056 }
3057 
3058 static void __kmp_free_team_arrays(kmp_team_t *team) {
3059   /* Note: this does not free the threads in t_threads (__kmp_free_threads) */
3060   int i;
3061   for (i = 0; i < team->t.t_max_nproc; ++i) {
3062     if (team->t.t_dispatch[i].th_disp_buffer != NULL) {
3063       __kmp_free(team->t.t_dispatch[i].th_disp_buffer);
3064       team->t.t_dispatch[i].th_disp_buffer = NULL;
3065     }
3066   }
3067 #if KMP_USE_HIER_SCHED
3068   __kmp_dispatch_free_hierarchies(team);
3069 #endif
3070   __kmp_free(team->t.t_threads);
3071   __kmp_free(team->t.t_disp_buffer);
3072   __kmp_free(team->t.t_dispatch);
3073   __kmp_free(team->t.t_implicit_task_taskdata);
3074   team->t.t_threads = NULL;
3075   team->t.t_disp_buffer = NULL;
3076   team->t.t_dispatch = NULL;
3077   team->t.t_implicit_task_taskdata = 0;
3078 }
3079 
3080 static void __kmp_reallocate_team_arrays(kmp_team_t *team, int max_nth) {
3081   kmp_info_t **oldThreads = team->t.t_threads;
3082 
3083   __kmp_free(team->t.t_disp_buffer);
3084   __kmp_free(team->t.t_dispatch);
3085   __kmp_free(team->t.t_implicit_task_taskdata);
3086   __kmp_allocate_team_arrays(team, max_nth);
3087 
3088   KMP_MEMCPY(team->t.t_threads, oldThreads,
3089              team->t.t_nproc * sizeof(kmp_info_t *));
3090 
3091   __kmp_free(oldThreads);
3092 }
3093 
3094 static kmp_internal_control_t __kmp_get_global_icvs(void) {
3095 
3096   kmp_r_sched_t r_sched =
3097       __kmp_get_schedule_global(); // get current state of scheduling globals
3098 
3099   KMP_DEBUG_ASSERT(__kmp_nested_proc_bind.used > 0);
3100 
3101   kmp_internal_control_t g_icvs = {
3102     0, // int serial_nesting_level; //corresponds to value of th_team_serialized
3103     (kmp_int8)__kmp_global.g.g_dynamic, // internal control for dynamic
3104     // adjustment of threads (per thread)
3105     (kmp_int8)__kmp_env_blocktime, // int bt_set; //internal control for
3106     // whether blocktime is explicitly set
3107     __kmp_dflt_blocktime, // int blocktime; //internal control for blocktime
3108 #if KMP_USE_MONITOR
3109     __kmp_bt_intervals, // int bt_intervals; //internal control for blocktime
3110 // intervals
3111 #endif
3112     __kmp_dflt_team_nth, // int nproc; //internal control for # of threads for
3113     // next parallel region (per thread)
3114     // (use a max ub on value if __kmp_parallel_initialize not called yet)
3115     __kmp_cg_max_nth, // int thread_limit;
3116     __kmp_dflt_max_active_levels, // int max_active_levels; //internal control
3117     // for max_active_levels
3118     r_sched, // kmp_r_sched_t sched; //internal control for runtime schedule
3119     // {sched,chunk} pair
3120     __kmp_nested_proc_bind.bind_types[0],
3121     __kmp_default_device,
3122     NULL // struct kmp_internal_control *next;
3123   };
3124 
3125   return g_icvs;
3126 }
3127 
3128 static kmp_internal_control_t __kmp_get_x_global_icvs(const kmp_team_t *team) {
3129 
3130   kmp_internal_control_t gx_icvs;
3131   gx_icvs.serial_nesting_level =
3132       0; // probably =team->t.t_serial like in save_inter_controls
3133   copy_icvs(&gx_icvs, &team->t.t_threads[0]->th.th_current_task->td_icvs);
3134   gx_icvs.next = NULL;
3135 
3136   return gx_icvs;
3137 }
3138 
3139 static void __kmp_initialize_root(kmp_root_t *root) {
3140   int f;
3141   kmp_team_t *root_team;
3142   kmp_team_t *hot_team;
3143   int hot_team_max_nth;
3144   kmp_r_sched_t r_sched =
3145       __kmp_get_schedule_global(); // get current state of scheduling globals
3146   kmp_internal_control_t r_icvs = __kmp_get_global_icvs();
3147   KMP_DEBUG_ASSERT(root);
3148   KMP_ASSERT(!root->r.r_begin);
3149 
3150   /* setup the root state structure */
3151   __kmp_init_lock(&root->r.r_begin_lock);
3152   root->r.r_begin = FALSE;
3153   root->r.r_active = FALSE;
3154   root->r.r_in_parallel = 0;
3155   root->r.r_blocktime = __kmp_dflt_blocktime;
3156 
3157   /* setup the root team for this task */
3158   /* allocate the root team structure */
3159   KF_TRACE(10, ("__kmp_initialize_root: before root_team\n"));
3160 
3161   root_team =
3162       __kmp_allocate_team(root,
3163                           1, // new_nproc
3164                           1, // max_nproc
3165 #if OMPT_SUPPORT
3166                           ompt_data_none, // root parallel id
3167 #endif
3168                           __kmp_nested_proc_bind.bind_types[0], &r_icvs,
3169                           0 // argc
3170                           USE_NESTED_HOT_ARG(NULL) // primary thread is unknown
3171                           );
3172 #if USE_DEBUGGER
3173   // Non-NULL value should be assigned to make the debugger display the root
3174   // team.
3175   TCW_SYNC_PTR(root_team->t.t_pkfn, (microtask_t)(~0));
3176 #endif
3177 
3178   KF_TRACE(10, ("__kmp_initialize_root: after root_team = %p\n", root_team));
3179 
3180   root->r.r_root_team = root_team;
3181   root_team->t.t_control_stack_top = NULL;
3182 
3183   /* initialize root team */
3184   root_team->t.t_threads[0] = NULL;
3185   root_team->t.t_nproc = 1;
3186   root_team->t.t_serialized = 1;
3187   // TODO???: root_team->t.t_max_active_levels = __kmp_dflt_max_active_levels;
3188   root_team->t.t_sched.sched = r_sched.sched;
3189   KA_TRACE(
3190       20,
3191       ("__kmp_initialize_root: init root team %d arrived: join=%u, plain=%u\n",
3192        root_team->t.t_id, KMP_INIT_BARRIER_STATE, KMP_INIT_BARRIER_STATE));
3193 
3194   /* setup the  hot team for this task */
3195   /* allocate the hot team structure */
3196   KF_TRACE(10, ("__kmp_initialize_root: before hot_team\n"));
3197 
3198   hot_team =
3199       __kmp_allocate_team(root,
3200                           1, // new_nproc
3201                           __kmp_dflt_team_nth_ub * 2, // max_nproc
3202 #if OMPT_SUPPORT
3203                           ompt_data_none, // root parallel id
3204 #endif
3205                           __kmp_nested_proc_bind.bind_types[0], &r_icvs,
3206                           0 // argc
3207                           USE_NESTED_HOT_ARG(NULL) // primary thread is unknown
3208                           );
3209   KF_TRACE(10, ("__kmp_initialize_root: after hot_team = %p\n", hot_team));
3210 
3211   root->r.r_hot_team = hot_team;
3212   root_team->t.t_control_stack_top = NULL;
3213 
3214   /* first-time initialization */
3215   hot_team->t.t_parent = root_team;
3216 
3217   /* initialize hot team */
3218   hot_team_max_nth = hot_team->t.t_max_nproc;
3219   for (f = 0; f < hot_team_max_nth; ++f) {
3220     hot_team->t.t_threads[f] = NULL;
3221   }
3222   hot_team->t.t_nproc = 1;
3223   // TODO???: hot_team->t.t_max_active_levels = __kmp_dflt_max_active_levels;
3224   hot_team->t.t_sched.sched = r_sched.sched;
3225   hot_team->t.t_size_changed = 0;
3226 }
3227 
3228 #ifdef KMP_DEBUG
3229 
3230 typedef struct kmp_team_list_item {
3231   kmp_team_p const *entry;
3232   struct kmp_team_list_item *next;
3233 } kmp_team_list_item_t;
3234 typedef kmp_team_list_item_t *kmp_team_list_t;
3235 
3236 static void __kmp_print_structure_team_accum( // Add team to list of teams.
3237     kmp_team_list_t list, // List of teams.
3238     kmp_team_p const *team // Team to add.
3239 ) {
3240 
3241   // List must terminate with item where both entry and next are NULL.
3242   // Team is added to the list only once.
3243   // List is sorted in ascending order by team id.
3244   // Team id is *not* a key.
3245 
3246   kmp_team_list_t l;
3247 
3248   KMP_DEBUG_ASSERT(list != NULL);
3249   if (team == NULL) {
3250     return;
3251   }
3252 
3253   __kmp_print_structure_team_accum(list, team->t.t_parent);
3254   __kmp_print_structure_team_accum(list, team->t.t_next_pool);
3255 
3256   // Search list for the team.
3257   l = list;
3258   while (l->next != NULL && l->entry != team) {
3259     l = l->next;
3260   }
3261   if (l->next != NULL) {
3262     return; // Team has been added before, exit.
3263   }
3264 
3265   // Team is not found. Search list again for insertion point.
3266   l = list;
3267   while (l->next != NULL && l->entry->t.t_id <= team->t.t_id) {
3268     l = l->next;
3269   }
3270 
3271   // Insert team.
3272   {
3273     kmp_team_list_item_t *item = (kmp_team_list_item_t *)KMP_INTERNAL_MALLOC(
3274         sizeof(kmp_team_list_item_t));
3275     *item = *l;
3276     l->entry = team;
3277     l->next = item;
3278   }
3279 }
3280 
3281 static void __kmp_print_structure_team(char const *title, kmp_team_p const *team
3282 
3283 ) {
3284   __kmp_printf("%s", title);
3285   if (team != NULL) {
3286     __kmp_printf("%2x %p\n", team->t.t_id, team);
3287   } else {
3288     __kmp_printf(" - (nil)\n");
3289   }
3290 }
3291 
3292 static void __kmp_print_structure_thread(char const *title,
3293                                          kmp_info_p const *thread) {
3294   __kmp_printf("%s", title);
3295   if (thread != NULL) {
3296     __kmp_printf("%2d %p\n", thread->th.th_info.ds.ds_gtid, thread);
3297   } else {
3298     __kmp_printf(" - (nil)\n");
3299   }
3300 }
3301 
3302 void __kmp_print_structure(void) {
3303 
3304   kmp_team_list_t list;
3305 
3306   // Initialize list of teams.
3307   list =
3308       (kmp_team_list_item_t *)KMP_INTERNAL_MALLOC(sizeof(kmp_team_list_item_t));
3309   list->entry = NULL;
3310   list->next = NULL;
3311 
3312   __kmp_printf("\n------------------------------\nGlobal Thread "
3313                "Table\n------------------------------\n");
3314   {
3315     int gtid;
3316     for (gtid = 0; gtid < __kmp_threads_capacity; ++gtid) {
3317       __kmp_printf("%2d", gtid);
3318       if (__kmp_threads != NULL) {
3319         __kmp_printf(" %p", __kmp_threads[gtid]);
3320       }
3321       if (__kmp_root != NULL) {
3322         __kmp_printf(" %p", __kmp_root[gtid]);
3323       }
3324       __kmp_printf("\n");
3325     }
3326   }
3327 
3328   // Print out __kmp_threads array.
3329   __kmp_printf("\n------------------------------\nThreads\n--------------------"
3330                "----------\n");
3331   if (__kmp_threads != NULL) {
3332     int gtid;
3333     for (gtid = 0; gtid < __kmp_threads_capacity; ++gtid) {
3334       kmp_info_t const *thread = __kmp_threads[gtid];
3335       if (thread != NULL) {
3336         __kmp_printf("GTID %2d %p:\n", gtid, thread);
3337         __kmp_printf("    Our Root:        %p\n", thread->th.th_root);
3338         __kmp_print_structure_team("    Our Team:     ", thread->th.th_team);
3339         __kmp_print_structure_team("    Serial Team:  ",
3340                                    thread->th.th_serial_team);
3341         __kmp_printf("    Threads:      %2d\n", thread->th.th_team_nproc);
3342         __kmp_print_structure_thread("    Primary:      ",
3343                                      thread->th.th_team_master);
3344         __kmp_printf("    Serialized?:  %2d\n", thread->th.th_team_serialized);
3345         __kmp_printf("    Set NProc:    %2d\n", thread->th.th_set_nproc);
3346         __kmp_printf("    Set Proc Bind: %2d\n", thread->th.th_set_proc_bind);
3347         __kmp_print_structure_thread("    Next in pool: ",
3348                                      thread->th.th_next_pool);
3349         __kmp_printf("\n");
3350         __kmp_print_structure_team_accum(list, thread->th.th_team);
3351         __kmp_print_structure_team_accum(list, thread->th.th_serial_team);
3352       }
3353     }
3354   } else {
3355     __kmp_printf("Threads array is not allocated.\n");
3356   }
3357 
3358   // Print out __kmp_root array.
3359   __kmp_printf("\n------------------------------\nUbers\n----------------------"
3360                "--------\n");
3361   if (__kmp_root != NULL) {
3362     int gtid;
3363     for (gtid = 0; gtid < __kmp_threads_capacity; ++gtid) {
3364       kmp_root_t const *root = __kmp_root[gtid];
3365       if (root != NULL) {
3366         __kmp_printf("GTID %2d %p:\n", gtid, root);
3367         __kmp_print_structure_team("    Root Team:    ", root->r.r_root_team);
3368         __kmp_print_structure_team("    Hot Team:     ", root->r.r_hot_team);
3369         __kmp_print_structure_thread("    Uber Thread:  ",
3370                                      root->r.r_uber_thread);
3371         __kmp_printf("    Active?:      %2d\n", root->r.r_active);
3372         __kmp_printf("    In Parallel:  %2d\n",
3373                      KMP_ATOMIC_LD_RLX(&root->r.r_in_parallel));
3374         __kmp_printf("\n");
3375         __kmp_print_structure_team_accum(list, root->r.r_root_team);
3376         __kmp_print_structure_team_accum(list, root->r.r_hot_team);
3377       }
3378     }
3379   } else {
3380     __kmp_printf("Ubers array is not allocated.\n");
3381   }
3382 
3383   __kmp_printf("\n------------------------------\nTeams\n----------------------"
3384                "--------\n");
3385   while (list->next != NULL) {
3386     kmp_team_p const *team = list->entry;
3387     int i;
3388     __kmp_printf("Team %2x %p:\n", team->t.t_id, team);
3389     __kmp_print_structure_team("    Parent Team:      ", team->t.t_parent);
3390     __kmp_printf("    Primary TID:      %2d\n", team->t.t_master_tid);
3391     __kmp_printf("    Max threads:      %2d\n", team->t.t_max_nproc);
3392     __kmp_printf("    Levels of serial: %2d\n", team->t.t_serialized);
3393     __kmp_printf("    Number threads:   %2d\n", team->t.t_nproc);
3394     for (i = 0; i < team->t.t_nproc; ++i) {
3395       __kmp_printf("    Thread %2d:      ", i);
3396       __kmp_print_structure_thread("", team->t.t_threads[i]);
3397     }
3398     __kmp_print_structure_team("    Next in pool:     ", team->t.t_next_pool);
3399     __kmp_printf("\n");
3400     list = list->next;
3401   }
3402 
3403   // Print out __kmp_thread_pool and __kmp_team_pool.
3404   __kmp_printf("\n------------------------------\nPools\n----------------------"
3405                "--------\n");
3406   __kmp_print_structure_thread("Thread pool:          ",
3407                                CCAST(kmp_info_t *, __kmp_thread_pool));
3408   __kmp_print_structure_team("Team pool:            ",
3409                              CCAST(kmp_team_t *, __kmp_team_pool));
3410   __kmp_printf("\n");
3411 
3412   // Free team list.
3413   while (list != NULL) {
3414     kmp_team_list_item_t *item = list;
3415     list = list->next;
3416     KMP_INTERNAL_FREE(item);
3417   }
3418 }
3419 
3420 #endif
3421 
3422 //---------------------------------------------------------------------------
3423 //  Stuff for per-thread fast random number generator
3424 //  Table of primes
3425 static const unsigned __kmp_primes[] = {
3426     0x9e3779b1, 0xffe6cc59, 0x2109f6dd, 0x43977ab5, 0xba5703f5, 0xb495a877,
3427     0xe1626741, 0x79695e6b, 0xbc98c09f, 0xd5bee2b3, 0x287488f9, 0x3af18231,
3428     0x9677cd4d, 0xbe3a6929, 0xadc6a877, 0xdcf0674b, 0xbe4d6fe9, 0x5f15e201,
3429     0x99afc3fd, 0xf3f16801, 0xe222cfff, 0x24ba5fdb, 0x0620452d, 0x79f149e3,
3430     0xc8b93f49, 0x972702cd, 0xb07dd827, 0x6c97d5ed, 0x085a3d61, 0x46eb5ea7,
3431     0x3d9910ed, 0x2e687b5b, 0x29609227, 0x6eb081f1, 0x0954c4e1, 0x9d114db9,
3432     0x542acfa9, 0xb3e6bd7b, 0x0742d917, 0xe9f3ffa7, 0x54581edb, 0xf2480f45,
3433     0x0bb9288f, 0xef1affc7, 0x85fa0ca7, 0x3ccc14db, 0xe6baf34b, 0x343377f7,
3434     0x5ca19031, 0xe6d9293b, 0xf0a9f391, 0x5d2e980b, 0xfc411073, 0xc3749363,
3435     0xb892d829, 0x3549366b, 0x629750ad, 0xb98294e5, 0x892d9483, 0xc235baf3,
3436     0x3d2402a3, 0x6bdef3c9, 0xbec333cd, 0x40c9520f};
3437 
3438 //---------------------------------------------------------------------------
3439 //  __kmp_get_random: Get a random number using a linear congruential method.
3440 unsigned short __kmp_get_random(kmp_info_t *thread) {
3441   unsigned x = thread->th.th_x;
3442   unsigned short r = (unsigned short)(x >> 16);
3443 
3444   thread->th.th_x = x * thread->th.th_a + 1;
3445 
3446   KA_TRACE(30, ("__kmp_get_random: THREAD: %d, RETURN: %u\n",
3447                 thread->th.th_info.ds.ds_tid, r));
3448 
3449   return r;
3450 }
3451 //--------------------------------------------------------
3452 // __kmp_init_random: Initialize a random number generator
3453 void __kmp_init_random(kmp_info_t *thread) {
3454   unsigned seed = thread->th.th_info.ds.ds_tid;
3455 
3456   thread->th.th_a =
3457       __kmp_primes[seed % (sizeof(__kmp_primes) / sizeof(__kmp_primes[0]))];
3458   thread->th.th_x = (seed + 1) * thread->th.th_a + 1;
3459   KA_TRACE(30,
3460            ("__kmp_init_random: THREAD: %u; A: %u\n", seed, thread->th.th_a));
3461 }
3462 
3463 #if KMP_OS_WINDOWS
3464 /* reclaim array entries for root threads that are already dead, returns number
3465  * reclaimed */
3466 static int __kmp_reclaim_dead_roots(void) {
3467   int i, r = 0;
3468 
3469   for (i = 0; i < __kmp_threads_capacity; ++i) {
3470     if (KMP_UBER_GTID(i) &&
3471         !__kmp_still_running((kmp_info_t *)TCR_SYNC_PTR(__kmp_threads[i])) &&
3472         !__kmp_root[i]
3473              ->r.r_active) { // AC: reclaim only roots died in non-active state
3474       r += __kmp_unregister_root_other_thread(i);
3475     }
3476   }
3477   return r;
3478 }
3479 #endif
3480 
3481 /* This function attempts to create free entries in __kmp_threads and
3482    __kmp_root, and returns the number of free entries generated.
3483 
3484    For Windows* OS static library, the first mechanism used is to reclaim array
3485    entries for root threads that are already dead.
3486 
3487    On all platforms, expansion is attempted on the arrays __kmp_threads_ and
3488    __kmp_root, with appropriate update to __kmp_threads_capacity. Array
3489    capacity is increased by doubling with clipping to __kmp_tp_capacity, if
3490    threadprivate cache array has been created. Synchronization with
3491    __kmpc_threadprivate_cached is done using __kmp_tp_cached_lock.
3492 
3493    After any dead root reclamation, if the clipping value allows array expansion
3494    to result in the generation of a total of nNeed free slots, the function does
3495    that expansion. If not, nothing is done beyond the possible initial root
3496    thread reclamation.
3497 
3498    If any argument is negative, the behavior is undefined. */
3499 static int __kmp_expand_threads(int nNeed) {
3500   int added = 0;
3501   int minimumRequiredCapacity;
3502   int newCapacity;
3503   kmp_info_t **newThreads;
3504   kmp_root_t **newRoot;
3505 
3506   // All calls to __kmp_expand_threads should be under __kmp_forkjoin_lock, so
3507   // resizing __kmp_threads does not need additional protection if foreign
3508   // threads are present
3509 
3510 #if KMP_OS_WINDOWS && !KMP_DYNAMIC_LIB
3511   /* only for Windows static library */
3512   /* reclaim array entries for root threads that are already dead */
3513   added = __kmp_reclaim_dead_roots();
3514 
3515   if (nNeed) {
3516     nNeed -= added;
3517     if (nNeed < 0)
3518       nNeed = 0;
3519   }
3520 #endif
3521   if (nNeed <= 0)
3522     return added;
3523 
3524   // Note that __kmp_threads_capacity is not bounded by __kmp_max_nth. If
3525   // __kmp_max_nth is set to some value less than __kmp_sys_max_nth by the
3526   // user via KMP_DEVICE_THREAD_LIMIT, then __kmp_threads_capacity may become
3527   // > __kmp_max_nth in one of two ways:
3528   //
3529   // 1) The initialization thread (gtid = 0) exits.  __kmp_threads[0]
3530   //    may not be reused by another thread, so we may need to increase
3531   //    __kmp_threads_capacity to __kmp_max_nth + 1.
3532   //
3533   // 2) New foreign root(s) are encountered.  We always register new foreign
3534   //    roots. This may cause a smaller # of threads to be allocated at
3535   //    subsequent parallel regions, but the worker threads hang around (and
3536   //    eventually go to sleep) and need slots in the __kmp_threads[] array.
3537   //
3538   // Anyway, that is the reason for moving the check to see if
3539   // __kmp_max_nth was exceeded into __kmp_reserve_threads()
3540   // instead of having it performed here. -BB
3541 
3542   KMP_DEBUG_ASSERT(__kmp_sys_max_nth >= __kmp_threads_capacity);
3543 
3544   /* compute expansion headroom to check if we can expand */
3545   if (__kmp_sys_max_nth - __kmp_threads_capacity < nNeed) {
3546     /* possible expansion too small -- give up */
3547     return added;
3548   }
3549   minimumRequiredCapacity = __kmp_threads_capacity + nNeed;
3550 
3551   newCapacity = __kmp_threads_capacity;
3552   do {
3553     newCapacity = newCapacity <= (__kmp_sys_max_nth >> 1) ? (newCapacity << 1)
3554                                                           : __kmp_sys_max_nth;
3555   } while (newCapacity < minimumRequiredCapacity);
3556   newThreads = (kmp_info_t **)__kmp_allocate(
3557       (sizeof(kmp_info_t *) + sizeof(kmp_root_t *)) * newCapacity + CACHE_LINE);
3558   newRoot =
3559       (kmp_root_t **)((char *)newThreads + sizeof(kmp_info_t *) * newCapacity);
3560   KMP_MEMCPY(newThreads, __kmp_threads,
3561              __kmp_threads_capacity * sizeof(kmp_info_t *));
3562   KMP_MEMCPY(newRoot, __kmp_root,
3563              __kmp_threads_capacity * sizeof(kmp_root_t *));
3564 
3565   kmp_info_t **temp_threads = __kmp_threads;
3566   *(kmp_info_t * *volatile *)&__kmp_threads = newThreads;
3567   *(kmp_root_t * *volatile *)&__kmp_root = newRoot;
3568   __kmp_free(temp_threads);
3569   added += newCapacity - __kmp_threads_capacity;
3570   *(volatile int *)&__kmp_threads_capacity = newCapacity;
3571 
3572   if (newCapacity > __kmp_tp_capacity) {
3573     __kmp_acquire_bootstrap_lock(&__kmp_tp_cached_lock);
3574     if (__kmp_tp_cached && newCapacity > __kmp_tp_capacity) {
3575       __kmp_threadprivate_resize_cache(newCapacity);
3576     } else { // increase __kmp_tp_capacity to correspond with kmp_threads size
3577       *(volatile int *)&__kmp_tp_capacity = newCapacity;
3578     }
3579     __kmp_release_bootstrap_lock(&__kmp_tp_cached_lock);
3580   }
3581 
3582   return added;
3583 }
3584 
3585 /* Register the current thread as a root thread and obtain our gtid. We must
3586    have the __kmp_initz_lock held at this point. Argument TRUE only if are the
3587    thread that calls from __kmp_do_serial_initialize() */
3588 int __kmp_register_root(int initial_thread) {
3589   kmp_info_t *root_thread;
3590   kmp_root_t *root;
3591   int gtid;
3592   int capacity;
3593   __kmp_acquire_bootstrap_lock(&__kmp_forkjoin_lock);
3594   KA_TRACE(20, ("__kmp_register_root: entered\n"));
3595   KMP_MB();
3596 
3597   /* 2007-03-02:
3598      If initial thread did not invoke OpenMP RTL yet, and this thread is not an
3599      initial one, "__kmp_all_nth >= __kmp_threads_capacity" condition does not
3600      work as expected -- it may return false (that means there is at least one
3601      empty slot in __kmp_threads array), but it is possible the only free slot
3602      is #0, which is reserved for initial thread and so cannot be used for this
3603      one. Following code workarounds this bug.
3604 
3605      However, right solution seems to be not reserving slot #0 for initial
3606      thread because:
3607      (1) there is no magic in slot #0,
3608      (2) we cannot detect initial thread reliably (the first thread which does
3609         serial initialization may be not a real initial thread).
3610   */
3611   capacity = __kmp_threads_capacity;
3612   if (!initial_thread && TCR_PTR(__kmp_threads[0]) == NULL) {
3613     --capacity;
3614   }
3615 
3616   // If it is not for initializing the hidden helper team, we need to take
3617   // __kmp_hidden_helper_threads_num out of the capacity because it is included
3618   // in __kmp_threads_capacity.
3619   if (__kmp_enable_hidden_helper && !TCR_4(__kmp_init_hidden_helper_threads)) {
3620     capacity -= __kmp_hidden_helper_threads_num;
3621   }
3622 
3623   /* see if there are too many threads */
3624   if (__kmp_all_nth >= capacity && !__kmp_expand_threads(1)) {
3625     if (__kmp_tp_cached) {
3626       __kmp_fatal(KMP_MSG(CantRegisterNewThread),
3627                   KMP_HNT(Set_ALL_THREADPRIVATE, __kmp_tp_capacity),
3628                   KMP_HNT(PossibleSystemLimitOnThreads), __kmp_msg_null);
3629     } else {
3630       __kmp_fatal(KMP_MSG(CantRegisterNewThread), KMP_HNT(SystemLimitOnThreads),
3631                   __kmp_msg_null);
3632     }
3633   }
3634 
3635   // When hidden helper task is enabled, __kmp_threads is organized as follows:
3636   // 0: initial thread, also a regular OpenMP thread.
3637   // [1, __kmp_hidden_helper_threads_num]: slots for hidden helper threads.
3638   // [__kmp_hidden_helper_threads_num + 1, __kmp_threads_capacity): slots for
3639   // regular OpenMP threads.
3640   if (TCR_4(__kmp_init_hidden_helper_threads)) {
3641     // Find an available thread slot for hidden helper thread. Slots for hidden
3642     // helper threads start from 1 to __kmp_hidden_helper_threads_num.
3643     for (gtid = 1; TCR_PTR(__kmp_threads[gtid]) != NULL &&
3644                    gtid <= __kmp_hidden_helper_threads_num;
3645          gtid++)
3646       ;
3647     KMP_ASSERT(gtid <= __kmp_hidden_helper_threads_num);
3648     KA_TRACE(1, ("__kmp_register_root: found slot in threads array for "
3649                  "hidden helper thread: T#%d\n",
3650                  gtid));
3651   } else {
3652     /* find an available thread slot */
3653     // Don't reassign the zero slot since we need that to only be used by
3654     // initial thread. Slots for hidden helper threads should also be skipped.
3655     if (initial_thread && TCR_PTR(__kmp_threads[0]) == NULL) {
3656       gtid = 0;
3657     } else {
3658       for (gtid = __kmp_hidden_helper_threads_num + 1;
3659            TCR_PTR(__kmp_threads[gtid]) != NULL; gtid++)
3660         ;
3661     }
3662     KA_TRACE(
3663         1, ("__kmp_register_root: found slot in threads array: T#%d\n", gtid));
3664     KMP_ASSERT(gtid < __kmp_threads_capacity);
3665   }
3666 
3667   /* update global accounting */
3668   __kmp_all_nth++;
3669   TCW_4(__kmp_nth, __kmp_nth + 1);
3670 
3671   // if __kmp_adjust_gtid_mode is set, then we use method #1 (sp search) for low
3672   // numbers of procs, and method #2 (keyed API call) for higher numbers.
3673   if (__kmp_adjust_gtid_mode) {
3674     if (__kmp_all_nth >= __kmp_tls_gtid_min) {
3675       if (TCR_4(__kmp_gtid_mode) != 2) {
3676         TCW_4(__kmp_gtid_mode, 2);
3677       }
3678     } else {
3679       if (TCR_4(__kmp_gtid_mode) != 1) {
3680         TCW_4(__kmp_gtid_mode, 1);
3681       }
3682     }
3683   }
3684 
3685 #ifdef KMP_ADJUST_BLOCKTIME
3686   /* Adjust blocktime to zero if necessary            */
3687   /* Middle initialization might not have occurred yet */
3688   if (!__kmp_env_blocktime && (__kmp_avail_proc > 0)) {
3689     if (__kmp_nth > __kmp_avail_proc) {
3690       __kmp_zero_bt = TRUE;
3691     }
3692   }
3693 #endif /* KMP_ADJUST_BLOCKTIME */
3694 
3695   /* setup this new hierarchy */
3696   if (!(root = __kmp_root[gtid])) {
3697     root = __kmp_root[gtid] = (kmp_root_t *)__kmp_allocate(sizeof(kmp_root_t));
3698     KMP_DEBUG_ASSERT(!root->r.r_root_team);
3699   }
3700 
3701 #if KMP_STATS_ENABLED
3702   // Initialize stats as soon as possible (right after gtid assignment).
3703   __kmp_stats_thread_ptr = __kmp_stats_list->push_back(gtid);
3704   __kmp_stats_thread_ptr->startLife();
3705   KMP_SET_THREAD_STATE(SERIAL_REGION);
3706   KMP_INIT_PARTITIONED_TIMERS(OMP_serial);
3707 #endif
3708   __kmp_initialize_root(root);
3709 
3710   /* setup new root thread structure */
3711   if (root->r.r_uber_thread) {
3712     root_thread = root->r.r_uber_thread;
3713   } else {
3714     root_thread = (kmp_info_t *)__kmp_allocate(sizeof(kmp_info_t));
3715     if (__kmp_storage_map) {
3716       __kmp_print_thread_storage_map(root_thread, gtid);
3717     }
3718     root_thread->th.th_info.ds.ds_gtid = gtid;
3719 #if OMPT_SUPPORT
3720     root_thread->th.ompt_thread_info.thread_data = ompt_data_none;
3721 #endif
3722     root_thread->th.th_root = root;
3723     if (__kmp_env_consistency_check) {
3724       root_thread->th.th_cons = __kmp_allocate_cons_stack(gtid);
3725     }
3726 #if USE_FAST_MEMORY
3727     __kmp_initialize_fast_memory(root_thread);
3728 #endif /* USE_FAST_MEMORY */
3729 
3730 #if KMP_USE_BGET
3731     KMP_DEBUG_ASSERT(root_thread->th.th_local.bget_data == NULL);
3732     __kmp_initialize_bget(root_thread);
3733 #endif
3734     __kmp_init_random(root_thread); // Initialize random number generator
3735   }
3736 
3737   /* setup the serial team held in reserve by the root thread */
3738   if (!root_thread->th.th_serial_team) {
3739     kmp_internal_control_t r_icvs = __kmp_get_global_icvs();
3740     KF_TRACE(10, ("__kmp_register_root: before serial_team\n"));
3741     root_thread->th.th_serial_team = __kmp_allocate_team(
3742         root, 1, 1,
3743 #if OMPT_SUPPORT
3744         ompt_data_none, // root parallel id
3745 #endif
3746         proc_bind_default, &r_icvs, 0 USE_NESTED_HOT_ARG(NULL));
3747   }
3748   KMP_ASSERT(root_thread->th.th_serial_team);
3749   KF_TRACE(10, ("__kmp_register_root: after serial_team = %p\n",
3750                 root_thread->th.th_serial_team));
3751 
3752   /* drop root_thread into place */
3753   TCW_SYNC_PTR(__kmp_threads[gtid], root_thread);
3754 
3755   root->r.r_root_team->t.t_threads[0] = root_thread;
3756   root->r.r_hot_team->t.t_threads[0] = root_thread;
3757   root_thread->th.th_serial_team->t.t_threads[0] = root_thread;
3758   // AC: the team created in reserve, not for execution (it is unused for now).
3759   root_thread->th.th_serial_team->t.t_serialized = 0;
3760   root->r.r_uber_thread = root_thread;
3761 
3762   /* initialize the thread, get it ready to go */
3763   __kmp_initialize_info(root_thread, root->r.r_root_team, 0, gtid);
3764   TCW_4(__kmp_init_gtid, TRUE);
3765 
3766   /* prepare the primary thread for get_gtid() */
3767   __kmp_gtid_set_specific(gtid);
3768 
3769 #if USE_ITT_BUILD
3770   __kmp_itt_thread_name(gtid);
3771 #endif /* USE_ITT_BUILD */
3772 
3773 #ifdef KMP_TDATA_GTID
3774   __kmp_gtid = gtid;
3775 #endif
3776   __kmp_create_worker(gtid, root_thread, __kmp_stksize);
3777   KMP_DEBUG_ASSERT(__kmp_gtid_get_specific() == gtid);
3778 
3779   KA_TRACE(20, ("__kmp_register_root: T#%d init T#%d(%d:%d) arrived: join=%u, "
3780                 "plain=%u\n",
3781                 gtid, __kmp_gtid_from_tid(0, root->r.r_hot_team),
3782                 root->r.r_hot_team->t.t_id, 0, KMP_INIT_BARRIER_STATE,
3783                 KMP_INIT_BARRIER_STATE));
3784   { // Initialize barrier data.
3785     int b;
3786     for (b = 0; b < bs_last_barrier; ++b) {
3787       root_thread->th.th_bar[b].bb.b_arrived = KMP_INIT_BARRIER_STATE;
3788 #if USE_DEBUGGER
3789       root_thread->th.th_bar[b].bb.b_worker_arrived = 0;
3790 #endif
3791     }
3792   }
3793   KMP_DEBUG_ASSERT(root->r.r_hot_team->t.t_bar[bs_forkjoin_barrier].b_arrived ==
3794                    KMP_INIT_BARRIER_STATE);
3795 
3796 #if KMP_AFFINITY_SUPPORTED
3797   root_thread->th.th_current_place = KMP_PLACE_UNDEFINED;
3798   root_thread->th.th_new_place = KMP_PLACE_UNDEFINED;
3799   root_thread->th.th_first_place = KMP_PLACE_UNDEFINED;
3800   root_thread->th.th_last_place = KMP_PLACE_UNDEFINED;
3801   if (TCR_4(__kmp_init_middle)) {
3802     __kmp_affinity_set_init_mask(gtid, TRUE);
3803   }
3804 #endif /* KMP_AFFINITY_SUPPORTED */
3805   root_thread->th.th_def_allocator = __kmp_def_allocator;
3806   root_thread->th.th_prev_level = 0;
3807   root_thread->th.th_prev_num_threads = 1;
3808 
3809   kmp_cg_root_t *tmp = (kmp_cg_root_t *)__kmp_allocate(sizeof(kmp_cg_root_t));
3810   tmp->cg_root = root_thread;
3811   tmp->cg_thread_limit = __kmp_cg_max_nth;
3812   tmp->cg_nthreads = 1;
3813   KA_TRACE(100, ("__kmp_register_root: Thread %p created node %p with"
3814                  " cg_nthreads init to 1\n",
3815                  root_thread, tmp));
3816   tmp->up = NULL;
3817   root_thread->th.th_cg_roots = tmp;
3818 
3819   __kmp_root_counter++;
3820 
3821 #if OMPT_SUPPORT
3822   if (!initial_thread && ompt_enabled.enabled) {
3823 
3824     kmp_info_t *root_thread = ompt_get_thread();
3825 
3826     ompt_set_thread_state(root_thread, ompt_state_overhead);
3827 
3828     if (ompt_enabled.ompt_callback_thread_begin) {
3829       ompt_callbacks.ompt_callback(ompt_callback_thread_begin)(
3830           ompt_thread_initial, __ompt_get_thread_data_internal());
3831     }
3832     ompt_data_t *task_data;
3833     ompt_data_t *parallel_data;
3834     __ompt_get_task_info_internal(0, NULL, &task_data, NULL, &parallel_data,
3835                                   NULL);
3836     if (ompt_enabled.ompt_callback_implicit_task) {
3837       ompt_callbacks.ompt_callback(ompt_callback_implicit_task)(
3838           ompt_scope_begin, parallel_data, task_data, 1, 1, ompt_task_initial);
3839     }
3840 
3841     ompt_set_thread_state(root_thread, ompt_state_work_serial);
3842   }
3843 #endif
3844 
3845   KMP_MB();
3846   __kmp_release_bootstrap_lock(&__kmp_forkjoin_lock);
3847 
3848   return gtid;
3849 }
3850 
3851 #if KMP_NESTED_HOT_TEAMS
3852 static int __kmp_free_hot_teams(kmp_root_t *root, kmp_info_t *thr, int level,
3853                                 const int max_level) {
3854   int i, n, nth;
3855   kmp_hot_team_ptr_t *hot_teams = thr->th.th_hot_teams;
3856   if (!hot_teams || !hot_teams[level].hot_team) {
3857     return 0;
3858   }
3859   KMP_DEBUG_ASSERT(level < max_level);
3860   kmp_team_t *team = hot_teams[level].hot_team;
3861   nth = hot_teams[level].hot_team_nth;
3862   n = nth - 1; // primary thread is not freed
3863   if (level < max_level - 1) {
3864     for (i = 0; i < nth; ++i) {
3865       kmp_info_t *th = team->t.t_threads[i];
3866       n += __kmp_free_hot_teams(root, th, level + 1, max_level);
3867       if (i > 0 && th->th.th_hot_teams) {
3868         __kmp_free(th->th.th_hot_teams);
3869         th->th.th_hot_teams = NULL;
3870       }
3871     }
3872   }
3873   __kmp_free_team(root, team, NULL);
3874   return n;
3875 }
3876 #endif
3877 
3878 // Resets a root thread and clear its root and hot teams.
3879 // Returns the number of __kmp_threads entries directly and indirectly freed.
3880 static int __kmp_reset_root(int gtid, kmp_root_t *root) {
3881   kmp_team_t *root_team = root->r.r_root_team;
3882   kmp_team_t *hot_team = root->r.r_hot_team;
3883   int n = hot_team->t.t_nproc;
3884   int i;
3885 
3886   KMP_DEBUG_ASSERT(!root->r.r_active);
3887 
3888   root->r.r_root_team = NULL;
3889   root->r.r_hot_team = NULL;
3890   // __kmp_free_team() does not free hot teams, so we have to clear r_hot_team
3891   // before call to __kmp_free_team().
3892   __kmp_free_team(root, root_team USE_NESTED_HOT_ARG(NULL));
3893 #if KMP_NESTED_HOT_TEAMS
3894   if (__kmp_hot_teams_max_level >
3895       0) { // need to free nested hot teams and their threads if any
3896     for (i = 0; i < hot_team->t.t_nproc; ++i) {
3897       kmp_info_t *th = hot_team->t.t_threads[i];
3898       if (__kmp_hot_teams_max_level > 1) {
3899         n += __kmp_free_hot_teams(root, th, 1, __kmp_hot_teams_max_level);
3900       }
3901       if (th->th.th_hot_teams) {
3902         __kmp_free(th->th.th_hot_teams);
3903         th->th.th_hot_teams = NULL;
3904       }
3905     }
3906   }
3907 #endif
3908   __kmp_free_team(root, hot_team USE_NESTED_HOT_ARG(NULL));
3909 
3910   // Before we can reap the thread, we need to make certain that all other
3911   // threads in the teams that had this root as ancestor have stopped trying to
3912   // steal tasks.
3913   if (__kmp_tasking_mode != tskm_immediate_exec) {
3914     __kmp_wait_to_unref_task_teams();
3915   }
3916 
3917 #if KMP_OS_WINDOWS
3918   /* Close Handle of root duplicated in __kmp_create_worker (tr #62919) */
3919   KA_TRACE(
3920       10, ("__kmp_reset_root: free handle, th = %p, handle = %" KMP_UINTPTR_SPEC
3921            "\n",
3922            (LPVOID) & (root->r.r_uber_thread->th),
3923            root->r.r_uber_thread->th.th_info.ds.ds_thread));
3924   __kmp_free_handle(root->r.r_uber_thread->th.th_info.ds.ds_thread);
3925 #endif /* KMP_OS_WINDOWS */
3926 
3927 #if OMPT_SUPPORT
3928   ompt_data_t *task_data;
3929   ompt_data_t *parallel_data;
3930   __ompt_get_task_info_internal(0, NULL, &task_data, NULL, &parallel_data,
3931                                 NULL);
3932   if (ompt_enabled.ompt_callback_implicit_task) {
3933     ompt_callbacks.ompt_callback(ompt_callback_implicit_task)(
3934         ompt_scope_end, parallel_data, task_data, 0, 1, ompt_task_initial);
3935   }
3936   if (ompt_enabled.ompt_callback_thread_end) {
3937     ompt_callbacks.ompt_callback(ompt_callback_thread_end)(
3938         &(root->r.r_uber_thread->th.ompt_thread_info.thread_data));
3939   }
3940 #endif
3941 
3942   TCW_4(__kmp_nth,
3943         __kmp_nth - 1); // __kmp_reap_thread will decrement __kmp_all_nth.
3944   i = root->r.r_uber_thread->th.th_cg_roots->cg_nthreads--;
3945   KA_TRACE(100, ("__kmp_reset_root: Thread %p decrement cg_nthreads on node %p"
3946                  " to %d\n",
3947                  root->r.r_uber_thread, root->r.r_uber_thread->th.th_cg_roots,
3948                  root->r.r_uber_thread->th.th_cg_roots->cg_nthreads));
3949   if (i == 1) {
3950     // need to free contention group structure
3951     KMP_DEBUG_ASSERT(root->r.r_uber_thread ==
3952                      root->r.r_uber_thread->th.th_cg_roots->cg_root);
3953     KMP_DEBUG_ASSERT(root->r.r_uber_thread->th.th_cg_roots->up == NULL);
3954     __kmp_free(root->r.r_uber_thread->th.th_cg_roots);
3955     root->r.r_uber_thread->th.th_cg_roots = NULL;
3956   }
3957   __kmp_reap_thread(root->r.r_uber_thread, 1);
3958 
3959   // We canot put root thread to __kmp_thread_pool, so we have to reap it
3960   // instead of freeing.
3961   root->r.r_uber_thread = NULL;
3962   /* mark root as no longer in use */
3963   root->r.r_begin = FALSE;
3964 
3965   return n;
3966 }
3967 
3968 void __kmp_unregister_root_current_thread(int gtid) {
3969   KA_TRACE(1, ("__kmp_unregister_root_current_thread: enter T#%d\n", gtid));
3970   /* this lock should be ok, since unregister_root_current_thread is never
3971      called during an abort, only during a normal close. furthermore, if you
3972      have the forkjoin lock, you should never try to get the initz lock */
3973   __kmp_acquire_bootstrap_lock(&__kmp_forkjoin_lock);
3974   if (TCR_4(__kmp_global.g.g_done) || !__kmp_init_serial) {
3975     KC_TRACE(10, ("__kmp_unregister_root_current_thread: already finished, "
3976                   "exiting T#%d\n",
3977                   gtid));
3978     __kmp_release_bootstrap_lock(&__kmp_forkjoin_lock);
3979     return;
3980   }
3981   kmp_root_t *root = __kmp_root[gtid];
3982 
3983   KMP_DEBUG_ASSERT(__kmp_threads && __kmp_threads[gtid]);
3984   KMP_ASSERT(KMP_UBER_GTID(gtid));
3985   KMP_ASSERT(root == __kmp_threads[gtid]->th.th_root);
3986   KMP_ASSERT(root->r.r_active == FALSE);
3987 
3988   KMP_MB();
3989 
3990   kmp_info_t *thread = __kmp_threads[gtid];
3991   kmp_team_t *team = thread->th.th_team;
3992   kmp_task_team_t *task_team = thread->th.th_task_team;
3993 
3994   // we need to wait for the proxy tasks before finishing the thread
3995   if (task_team != NULL && task_team->tt.tt_found_proxy_tasks) {
3996 #if OMPT_SUPPORT
3997     // the runtime is shutting down so we won't report any events
3998     thread->th.ompt_thread_info.state = ompt_state_undefined;
3999 #endif
4000     __kmp_task_team_wait(thread, team USE_ITT_BUILD_ARG(NULL));
4001   }
4002 
4003   __kmp_reset_root(gtid, root);
4004 
4005   KMP_MB();
4006   KC_TRACE(10,
4007            ("__kmp_unregister_root_current_thread: T#%d unregistered\n", gtid));
4008 
4009   __kmp_release_bootstrap_lock(&__kmp_forkjoin_lock);
4010 }
4011 
4012 #if KMP_OS_WINDOWS
4013 /* __kmp_forkjoin_lock must be already held
4014    Unregisters a root thread that is not the current thread.  Returns the number
4015    of __kmp_threads entries freed as a result. */
4016 static int __kmp_unregister_root_other_thread(int gtid) {
4017   kmp_root_t *root = __kmp_root[gtid];
4018   int r;
4019 
4020   KA_TRACE(1, ("__kmp_unregister_root_other_thread: enter T#%d\n", gtid));
4021   KMP_DEBUG_ASSERT(__kmp_threads && __kmp_threads[gtid]);
4022   KMP_ASSERT(KMP_UBER_GTID(gtid));
4023   KMP_ASSERT(root == __kmp_threads[gtid]->th.th_root);
4024   KMP_ASSERT(root->r.r_active == FALSE);
4025 
4026   r = __kmp_reset_root(gtid, root);
4027   KC_TRACE(10,
4028            ("__kmp_unregister_root_other_thread: T#%d unregistered\n", gtid));
4029   return r;
4030 }
4031 #endif
4032 
4033 #if KMP_DEBUG
4034 void __kmp_task_info() {
4035 
4036   kmp_int32 gtid = __kmp_entry_gtid();
4037   kmp_int32 tid = __kmp_tid_from_gtid(gtid);
4038   kmp_info_t *this_thr = __kmp_threads[gtid];
4039   kmp_team_t *steam = this_thr->th.th_serial_team;
4040   kmp_team_t *team = this_thr->th.th_team;
4041 
4042   __kmp_printf(
4043       "__kmp_task_info: gtid=%d tid=%d t_thread=%p team=%p steam=%p curtask=%p "
4044       "ptask=%p\n",
4045       gtid, tid, this_thr, team, steam, this_thr->th.th_current_task,
4046       team->t.t_implicit_task_taskdata[tid].td_parent);
4047 }
4048 #endif // KMP_DEBUG
4049 
4050 /* TODO optimize with one big memclr, take out what isn't needed, split
4051    responsibility to workers as much as possible, and delay initialization of
4052    features as much as possible  */
4053 static void __kmp_initialize_info(kmp_info_t *this_thr, kmp_team_t *team,
4054                                   int tid, int gtid) {
4055   /* this_thr->th.th_info.ds.ds_gtid is setup in
4056      kmp_allocate_thread/create_worker.
4057      this_thr->th.th_serial_team is setup in __kmp_allocate_thread */
4058   KMP_DEBUG_ASSERT(this_thr != NULL);
4059   KMP_DEBUG_ASSERT(this_thr->th.th_serial_team);
4060   KMP_DEBUG_ASSERT(team);
4061   KMP_DEBUG_ASSERT(team->t.t_threads);
4062   KMP_DEBUG_ASSERT(team->t.t_dispatch);
4063   kmp_info_t *master = team->t.t_threads[0];
4064   KMP_DEBUG_ASSERT(master);
4065   KMP_DEBUG_ASSERT(master->th.th_root);
4066 
4067   KMP_MB();
4068 
4069   TCW_SYNC_PTR(this_thr->th.th_team, team);
4070 
4071   this_thr->th.th_info.ds.ds_tid = tid;
4072   this_thr->th.th_set_nproc = 0;
4073   if (__kmp_tasking_mode != tskm_immediate_exec)
4074     // When tasking is possible, threads are not safe to reap until they are
4075     // done tasking; this will be set when tasking code is exited in wait
4076     this_thr->th.th_reap_state = KMP_NOT_SAFE_TO_REAP;
4077   else // no tasking --> always safe to reap
4078     this_thr->th.th_reap_state = KMP_SAFE_TO_REAP;
4079   this_thr->th.th_set_proc_bind = proc_bind_default;
4080 #if KMP_AFFINITY_SUPPORTED
4081   this_thr->th.th_new_place = this_thr->th.th_current_place;
4082 #endif
4083   this_thr->th.th_root = master->th.th_root;
4084 
4085   /* setup the thread's cache of the team structure */
4086   this_thr->th.th_team_nproc = team->t.t_nproc;
4087   this_thr->th.th_team_master = master;
4088   this_thr->th.th_team_serialized = team->t.t_serialized;
4089   TCW_PTR(this_thr->th.th_sleep_loc, NULL);
4090 
4091   KMP_DEBUG_ASSERT(team->t.t_implicit_task_taskdata);
4092 
4093   KF_TRACE(10, ("__kmp_initialize_info1: T#%d:%d this_thread=%p curtask=%p\n",
4094                 tid, gtid, this_thr, this_thr->th.th_current_task));
4095 
4096   __kmp_init_implicit_task(this_thr->th.th_team_master->th.th_ident, this_thr,
4097                            team, tid, TRUE);
4098 
4099   KF_TRACE(10, ("__kmp_initialize_info2: T#%d:%d this_thread=%p curtask=%p\n",
4100                 tid, gtid, this_thr, this_thr->th.th_current_task));
4101   // TODO: Initialize ICVs from parent; GEH - isn't that already done in
4102   // __kmp_initialize_team()?
4103 
4104   /* TODO no worksharing in speculative threads */
4105   this_thr->th.th_dispatch = &team->t.t_dispatch[tid];
4106 
4107   this_thr->th.th_local.this_construct = 0;
4108 
4109   if (!this_thr->th.th_pri_common) {
4110     this_thr->th.th_pri_common =
4111         (struct common_table *)__kmp_allocate(sizeof(struct common_table));
4112     if (__kmp_storage_map) {
4113       __kmp_print_storage_map_gtid(
4114           gtid, this_thr->th.th_pri_common, this_thr->th.th_pri_common + 1,
4115           sizeof(struct common_table), "th_%d.th_pri_common\n", gtid);
4116     }
4117     this_thr->th.th_pri_head = NULL;
4118   }
4119 
4120   if (this_thr != master && // Primary thread's CG root is initialized elsewhere
4121       this_thr->th.th_cg_roots != master->th.th_cg_roots) { // CG root not set
4122     // Make new thread's CG root same as primary thread's
4123     KMP_DEBUG_ASSERT(master->th.th_cg_roots);
4124     kmp_cg_root_t *tmp = this_thr->th.th_cg_roots;
4125     if (tmp) {
4126       // worker changes CG, need to check if old CG should be freed
4127       int i = tmp->cg_nthreads--;
4128       KA_TRACE(100, ("__kmp_initialize_info: Thread %p decrement cg_nthreads"
4129                      " on node %p of thread %p to %d\n",
4130                      this_thr, tmp, tmp->cg_root, tmp->cg_nthreads));
4131       if (i == 1) {
4132         __kmp_free(tmp); // last thread left CG --> free it
4133       }
4134     }
4135     this_thr->th.th_cg_roots = master->th.th_cg_roots;
4136     // Increment new thread's CG root's counter to add the new thread
4137     this_thr->th.th_cg_roots->cg_nthreads++;
4138     KA_TRACE(100, ("__kmp_initialize_info: Thread %p increment cg_nthreads on"
4139                    " node %p of thread %p to %d\n",
4140                    this_thr, this_thr->th.th_cg_roots,
4141                    this_thr->th.th_cg_roots->cg_root,
4142                    this_thr->th.th_cg_roots->cg_nthreads));
4143     this_thr->th.th_current_task->td_icvs.thread_limit =
4144         this_thr->th.th_cg_roots->cg_thread_limit;
4145   }
4146 
4147   /* Initialize dynamic dispatch */
4148   {
4149     volatile kmp_disp_t *dispatch = this_thr->th.th_dispatch;
4150     // Use team max_nproc since this will never change for the team.
4151     size_t disp_size =
4152         sizeof(dispatch_private_info_t) *
4153         (team->t.t_max_nproc == 1 ? 1 : __kmp_dispatch_num_buffers);
4154     KD_TRACE(10, ("__kmp_initialize_info: T#%d max_nproc: %d\n", gtid,
4155                   team->t.t_max_nproc));
4156     KMP_ASSERT(dispatch);
4157     KMP_DEBUG_ASSERT(team->t.t_dispatch);
4158     KMP_DEBUG_ASSERT(dispatch == &team->t.t_dispatch[tid]);
4159 
4160     dispatch->th_disp_index = 0;
4161     dispatch->th_doacross_buf_idx = 0;
4162     if (!dispatch->th_disp_buffer) {
4163       dispatch->th_disp_buffer =
4164           (dispatch_private_info_t *)__kmp_allocate(disp_size);
4165 
4166       if (__kmp_storage_map) {
4167         __kmp_print_storage_map_gtid(
4168             gtid, &dispatch->th_disp_buffer[0],
4169             &dispatch->th_disp_buffer[team->t.t_max_nproc == 1
4170                                           ? 1
4171                                           : __kmp_dispatch_num_buffers],
4172             disp_size,
4173             "th_%d.th_dispatch.th_disp_buffer "
4174             "(team_%d.t_dispatch[%d].th_disp_buffer)",
4175             gtid, team->t.t_id, gtid);
4176       }
4177     } else {
4178       memset(&dispatch->th_disp_buffer[0], '\0', disp_size);
4179     }
4180 
4181     dispatch->th_dispatch_pr_current = 0;
4182     dispatch->th_dispatch_sh_current = 0;
4183 
4184     dispatch->th_deo_fcn = 0; /* ORDERED     */
4185     dispatch->th_dxo_fcn = 0; /* END ORDERED */
4186   }
4187 
4188   this_thr->th.th_next_pool = NULL;
4189 
4190   if (!this_thr->th.th_task_state_memo_stack) {
4191     size_t i;
4192     this_thr->th.th_task_state_memo_stack =
4193         (kmp_uint8 *)__kmp_allocate(4 * sizeof(kmp_uint8));
4194     this_thr->th.th_task_state_top = 0;
4195     this_thr->th.th_task_state_stack_sz = 4;
4196     for (i = 0; i < this_thr->th.th_task_state_stack_sz;
4197          ++i) // zero init the stack
4198       this_thr->th.th_task_state_memo_stack[i] = 0;
4199   }
4200 
4201   KMP_DEBUG_ASSERT(!this_thr->th.th_spin_here);
4202   KMP_DEBUG_ASSERT(this_thr->th.th_next_waiting == 0);
4203 
4204   KMP_MB();
4205 }
4206 
4207 /* allocate a new thread for the requesting team. this is only called from
4208    within a forkjoin critical section. we will first try to get an available
4209    thread from the thread pool. if none is available, we will fork a new one
4210    assuming we are able to create a new one. this should be assured, as the
4211    caller should check on this first. */
4212 kmp_info_t *__kmp_allocate_thread(kmp_root_t *root, kmp_team_t *team,
4213                                   int new_tid) {
4214   kmp_team_t *serial_team;
4215   kmp_info_t *new_thr;
4216   int new_gtid;
4217 
4218   KA_TRACE(20, ("__kmp_allocate_thread: T#%d\n", __kmp_get_gtid()));
4219   KMP_DEBUG_ASSERT(root && team);
4220 #if !KMP_NESTED_HOT_TEAMS
4221   KMP_DEBUG_ASSERT(KMP_MASTER_GTID(__kmp_get_gtid()));
4222 #endif
4223   KMP_MB();
4224 
4225   /* first, try to get one from the thread pool */
4226   if (__kmp_thread_pool) {
4227     new_thr = CCAST(kmp_info_t *, __kmp_thread_pool);
4228     __kmp_thread_pool = (volatile kmp_info_t *)new_thr->th.th_next_pool;
4229     if (new_thr == __kmp_thread_pool_insert_pt) {
4230       __kmp_thread_pool_insert_pt = NULL;
4231     }
4232     TCW_4(new_thr->th.th_in_pool, FALSE);
4233     __kmp_suspend_initialize_thread(new_thr);
4234     __kmp_lock_suspend_mx(new_thr);
4235     if (new_thr->th.th_active_in_pool == TRUE) {
4236       KMP_DEBUG_ASSERT(new_thr->th.th_active == TRUE);
4237       KMP_ATOMIC_DEC(&__kmp_thread_pool_active_nth);
4238       new_thr->th.th_active_in_pool = FALSE;
4239     }
4240     __kmp_unlock_suspend_mx(new_thr);
4241 
4242     KA_TRACE(20, ("__kmp_allocate_thread: T#%d using thread T#%d\n",
4243                   __kmp_get_gtid(), new_thr->th.th_info.ds.ds_gtid));
4244     KMP_ASSERT(!new_thr->th.th_team);
4245     KMP_DEBUG_ASSERT(__kmp_nth < __kmp_threads_capacity);
4246 
4247     /* setup the thread structure */
4248     __kmp_initialize_info(new_thr, team, new_tid,
4249                           new_thr->th.th_info.ds.ds_gtid);
4250     KMP_DEBUG_ASSERT(new_thr->th.th_serial_team);
4251 
4252     TCW_4(__kmp_nth, __kmp_nth + 1);
4253 
4254     new_thr->th.th_task_state = 0;
4255     new_thr->th.th_task_state_top = 0;
4256     new_thr->th.th_task_state_stack_sz = 4;
4257 
4258 #ifdef KMP_ADJUST_BLOCKTIME
4259     /* Adjust blocktime back to zero if necessary */
4260     /* Middle initialization might not have occurred yet */
4261     if (!__kmp_env_blocktime && (__kmp_avail_proc > 0)) {
4262       if (__kmp_nth > __kmp_avail_proc) {
4263         __kmp_zero_bt = TRUE;
4264       }
4265     }
4266 #endif /* KMP_ADJUST_BLOCKTIME */
4267 
4268 #if KMP_DEBUG
4269     // If thread entered pool via __kmp_free_thread, wait_flag should !=
4270     // KMP_BARRIER_PARENT_FLAG.
4271     int b;
4272     kmp_balign_t *balign = new_thr->th.th_bar;
4273     for (b = 0; b < bs_last_barrier; ++b)
4274       KMP_DEBUG_ASSERT(balign[b].bb.wait_flag != KMP_BARRIER_PARENT_FLAG);
4275 #endif
4276 
4277     KF_TRACE(10, ("__kmp_allocate_thread: T#%d using thread %p T#%d\n",
4278                   __kmp_get_gtid(), new_thr, new_thr->th.th_info.ds.ds_gtid));
4279 
4280     KMP_MB();
4281     return new_thr;
4282   }
4283 
4284   /* no, well fork a new one */
4285   KMP_ASSERT(__kmp_nth == __kmp_all_nth);
4286   KMP_ASSERT(__kmp_all_nth < __kmp_threads_capacity);
4287 
4288 #if KMP_USE_MONITOR
4289   // If this is the first worker thread the RTL is creating, then also
4290   // launch the monitor thread.  We try to do this as early as possible.
4291   if (!TCR_4(__kmp_init_monitor)) {
4292     __kmp_acquire_bootstrap_lock(&__kmp_monitor_lock);
4293     if (!TCR_4(__kmp_init_monitor)) {
4294       KF_TRACE(10, ("before __kmp_create_monitor\n"));
4295       TCW_4(__kmp_init_monitor, 1);
4296       __kmp_create_monitor(&__kmp_monitor);
4297       KF_TRACE(10, ("after __kmp_create_monitor\n"));
4298 #if KMP_OS_WINDOWS
4299       // AC: wait until monitor has started. This is a fix for CQ232808.
4300       // The reason is that if the library is loaded/unloaded in a loop with
4301       // small (parallel) work in between, then there is high probability that
4302       // monitor thread started after the library shutdown. At shutdown it is
4303       // too late to cope with the problem, because when the primary thread is
4304       // in DllMain (process detach) the monitor has no chances to start (it is
4305       // blocked), and primary thread has no means to inform the monitor that
4306       // the library has gone, because all the memory which the monitor can
4307       // access is going to be released/reset.
4308       while (TCR_4(__kmp_init_monitor) < 2) {
4309         KMP_YIELD(TRUE);
4310       }
4311       KF_TRACE(10, ("after monitor thread has started\n"));
4312 #endif
4313     }
4314     __kmp_release_bootstrap_lock(&__kmp_monitor_lock);
4315   }
4316 #endif
4317 
4318   KMP_MB();
4319 
4320   {
4321     int new_start_gtid = TCR_4(__kmp_init_hidden_helper_threads)
4322                              ? 1
4323                              : __kmp_hidden_helper_threads_num + 1;
4324 
4325     for (new_gtid = new_start_gtid; TCR_PTR(__kmp_threads[new_gtid]) != NULL;
4326          ++new_gtid) {
4327       KMP_DEBUG_ASSERT(new_gtid < __kmp_threads_capacity);
4328     }
4329 
4330     if (TCR_4(__kmp_init_hidden_helper_threads)) {
4331       KMP_DEBUG_ASSERT(new_gtid <= __kmp_hidden_helper_threads_num);
4332     }
4333   }
4334 
4335   /* allocate space for it. */
4336   new_thr = (kmp_info_t *)__kmp_allocate(sizeof(kmp_info_t));
4337 
4338   TCW_SYNC_PTR(__kmp_threads[new_gtid], new_thr);
4339 
4340 #if USE_ITT_BUILD && USE_ITT_NOTIFY && KMP_DEBUG
4341   // suppress race conditions detection on synchronization flags in debug mode
4342   // this helps to analyze library internals eliminating false positives
4343   __itt_suppress_mark_range(
4344       __itt_suppress_range, __itt_suppress_threading_errors,
4345       &new_thr->th.th_sleep_loc, sizeof(new_thr->th.th_sleep_loc));
4346   __itt_suppress_mark_range(
4347       __itt_suppress_range, __itt_suppress_threading_errors,
4348       &new_thr->th.th_reap_state, sizeof(new_thr->th.th_reap_state));
4349 #if KMP_OS_WINDOWS
4350   __itt_suppress_mark_range(
4351       __itt_suppress_range, __itt_suppress_threading_errors,
4352       &new_thr->th.th_suspend_init, sizeof(new_thr->th.th_suspend_init));
4353 #else
4354   __itt_suppress_mark_range(__itt_suppress_range,
4355                             __itt_suppress_threading_errors,
4356                             &new_thr->th.th_suspend_init_count,
4357                             sizeof(new_thr->th.th_suspend_init_count));
4358 #endif
4359   // TODO: check if we need to also suppress b_arrived flags
4360   __itt_suppress_mark_range(__itt_suppress_range,
4361                             __itt_suppress_threading_errors,
4362                             CCAST(kmp_uint64 *, &new_thr->th.th_bar[0].bb.b_go),
4363                             sizeof(new_thr->th.th_bar[0].bb.b_go));
4364   __itt_suppress_mark_range(__itt_suppress_range,
4365                             __itt_suppress_threading_errors,
4366                             CCAST(kmp_uint64 *, &new_thr->th.th_bar[1].bb.b_go),
4367                             sizeof(new_thr->th.th_bar[1].bb.b_go));
4368   __itt_suppress_mark_range(__itt_suppress_range,
4369                             __itt_suppress_threading_errors,
4370                             CCAST(kmp_uint64 *, &new_thr->th.th_bar[2].bb.b_go),
4371                             sizeof(new_thr->th.th_bar[2].bb.b_go));
4372 #endif /* USE_ITT_BUILD && USE_ITT_NOTIFY && KMP_DEBUG */
4373   if (__kmp_storage_map) {
4374     __kmp_print_thread_storage_map(new_thr, new_gtid);
4375   }
4376 
4377   // add the reserve serialized team, initialized from the team's primary thread
4378   {
4379     kmp_internal_control_t r_icvs = __kmp_get_x_global_icvs(team);
4380     KF_TRACE(10, ("__kmp_allocate_thread: before th_serial/serial_team\n"));
4381     new_thr->th.th_serial_team = serial_team =
4382         (kmp_team_t *)__kmp_allocate_team(root, 1, 1,
4383 #if OMPT_SUPPORT
4384                                           ompt_data_none, // root parallel id
4385 #endif
4386                                           proc_bind_default, &r_icvs,
4387                                           0 USE_NESTED_HOT_ARG(NULL));
4388   }
4389   KMP_ASSERT(serial_team);
4390   serial_team->t.t_serialized = 0; // AC: the team created in reserve, not for
4391   // execution (it is unused for now).
4392   serial_team->t.t_threads[0] = new_thr;
4393   KF_TRACE(10,
4394            ("__kmp_allocate_thread: after th_serial/serial_team : new_thr=%p\n",
4395             new_thr));
4396 
4397   /* setup the thread structures */
4398   __kmp_initialize_info(new_thr, team, new_tid, new_gtid);
4399 
4400 #if USE_FAST_MEMORY
4401   __kmp_initialize_fast_memory(new_thr);
4402 #endif /* USE_FAST_MEMORY */
4403 
4404 #if KMP_USE_BGET
4405   KMP_DEBUG_ASSERT(new_thr->th.th_local.bget_data == NULL);
4406   __kmp_initialize_bget(new_thr);
4407 #endif
4408 
4409   __kmp_init_random(new_thr); // Initialize random number generator
4410 
4411   /* Initialize these only once when thread is grabbed for a team allocation */
4412   KA_TRACE(20,
4413            ("__kmp_allocate_thread: T#%d init go fork=%u, plain=%u\n",
4414             __kmp_get_gtid(), KMP_INIT_BARRIER_STATE, KMP_INIT_BARRIER_STATE));
4415 
4416   int b;
4417   kmp_balign_t *balign = new_thr->th.th_bar;
4418   for (b = 0; b < bs_last_barrier; ++b) {
4419     balign[b].bb.b_go = KMP_INIT_BARRIER_STATE;
4420     balign[b].bb.team = NULL;
4421     balign[b].bb.wait_flag = KMP_BARRIER_NOT_WAITING;
4422     balign[b].bb.use_oncore_barrier = 0;
4423   }
4424 
4425   new_thr->th.th_spin_here = FALSE;
4426   new_thr->th.th_next_waiting = 0;
4427 #if KMP_OS_UNIX
4428   new_thr->th.th_blocking = false;
4429 #endif
4430 
4431 #if KMP_AFFINITY_SUPPORTED
4432   new_thr->th.th_current_place = KMP_PLACE_UNDEFINED;
4433   new_thr->th.th_new_place = KMP_PLACE_UNDEFINED;
4434   new_thr->th.th_first_place = KMP_PLACE_UNDEFINED;
4435   new_thr->th.th_last_place = KMP_PLACE_UNDEFINED;
4436 #endif
4437   new_thr->th.th_def_allocator = __kmp_def_allocator;
4438   new_thr->th.th_prev_level = 0;
4439   new_thr->th.th_prev_num_threads = 1;
4440 
4441   TCW_4(new_thr->th.th_in_pool, FALSE);
4442   new_thr->th.th_active_in_pool = FALSE;
4443   TCW_4(new_thr->th.th_active, TRUE);
4444 
4445   /* adjust the global counters */
4446   __kmp_all_nth++;
4447   __kmp_nth++;
4448 
4449   // if __kmp_adjust_gtid_mode is set, then we use method #1 (sp search) for low
4450   // numbers of procs, and method #2 (keyed API call) for higher numbers.
4451   if (__kmp_adjust_gtid_mode) {
4452     if (__kmp_all_nth >= __kmp_tls_gtid_min) {
4453       if (TCR_4(__kmp_gtid_mode) != 2) {
4454         TCW_4(__kmp_gtid_mode, 2);
4455       }
4456     } else {
4457       if (TCR_4(__kmp_gtid_mode) != 1) {
4458         TCW_4(__kmp_gtid_mode, 1);
4459       }
4460     }
4461   }
4462 
4463 #ifdef KMP_ADJUST_BLOCKTIME
4464   /* Adjust blocktime back to zero if necessary       */
4465   /* Middle initialization might not have occurred yet */
4466   if (!__kmp_env_blocktime && (__kmp_avail_proc > 0)) {
4467     if (__kmp_nth > __kmp_avail_proc) {
4468       __kmp_zero_bt = TRUE;
4469     }
4470   }
4471 #endif /* KMP_ADJUST_BLOCKTIME */
4472 
4473   /* actually fork it and create the new worker thread */
4474   KF_TRACE(
4475       10, ("__kmp_allocate_thread: before __kmp_create_worker: %p\n", new_thr));
4476   __kmp_create_worker(new_gtid, new_thr, __kmp_stksize);
4477   KF_TRACE(10,
4478            ("__kmp_allocate_thread: after __kmp_create_worker: %p\n", new_thr));
4479 
4480   KA_TRACE(20, ("__kmp_allocate_thread: T#%d forked T#%d\n", __kmp_get_gtid(),
4481                 new_gtid));
4482   KMP_MB();
4483   return new_thr;
4484 }
4485 
4486 /* Reinitialize team for reuse.
4487    The hot team code calls this case at every fork barrier, so EPCC barrier
4488    test are extremely sensitive to changes in it, esp. writes to the team
4489    struct, which cause a cache invalidation in all threads.
4490    IF YOU TOUCH THIS ROUTINE, RUN EPCC C SYNCBENCH ON A BIG-IRON MACHINE!!! */
4491 static void __kmp_reinitialize_team(kmp_team_t *team,
4492                                     kmp_internal_control_t *new_icvs,
4493                                     ident_t *loc) {
4494   KF_TRACE(10, ("__kmp_reinitialize_team: enter this_thread=%p team=%p\n",
4495                 team->t.t_threads[0], team));
4496   KMP_DEBUG_ASSERT(team && new_icvs);
4497   KMP_DEBUG_ASSERT((!TCR_4(__kmp_init_parallel)) || new_icvs->nproc);
4498   KMP_CHECK_UPDATE(team->t.t_ident, loc);
4499 
4500   KMP_CHECK_UPDATE(team->t.t_id, KMP_GEN_TEAM_ID());
4501   // Copy ICVs to the primary thread's implicit taskdata
4502   __kmp_init_implicit_task(loc, team->t.t_threads[0], team, 0, FALSE);
4503   copy_icvs(&team->t.t_implicit_task_taskdata[0].td_icvs, new_icvs);
4504 
4505   KF_TRACE(10, ("__kmp_reinitialize_team: exit this_thread=%p team=%p\n",
4506                 team->t.t_threads[0], team));
4507 }
4508 
4509 /* Initialize the team data structure.
4510    This assumes the t_threads and t_max_nproc are already set.
4511    Also, we don't touch the arguments */
4512 static void __kmp_initialize_team(kmp_team_t *team, int new_nproc,
4513                                   kmp_internal_control_t *new_icvs,
4514                                   ident_t *loc) {
4515   KF_TRACE(10, ("__kmp_initialize_team: enter: team=%p\n", team));
4516 
4517   /* verify */
4518   KMP_DEBUG_ASSERT(team);
4519   KMP_DEBUG_ASSERT(new_nproc <= team->t.t_max_nproc);
4520   KMP_DEBUG_ASSERT(team->t.t_threads);
4521   KMP_MB();
4522 
4523   team->t.t_master_tid = 0; /* not needed */
4524   /* team->t.t_master_bar;        not needed */
4525   team->t.t_serialized = new_nproc > 1 ? 0 : 1;
4526   team->t.t_nproc = new_nproc;
4527 
4528   /* team->t.t_parent     = NULL; TODO not needed & would mess up hot team */
4529   team->t.t_next_pool = NULL;
4530   /* memset( team->t.t_threads, 0, sizeof(kmp_info_t*)*new_nproc ); would mess
4531    * up hot team */
4532 
4533   TCW_SYNC_PTR(team->t.t_pkfn, NULL); /* not needed */
4534   team->t.t_invoke = NULL; /* not needed */
4535 
4536   // TODO???: team->t.t_max_active_levels       = new_max_active_levels;
4537   team->t.t_sched.sched = new_icvs->sched.sched;
4538 
4539 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
4540   team->t.t_fp_control_saved = FALSE; /* not needed */
4541   team->t.t_x87_fpu_control_word = 0; /* not needed */
4542   team->t.t_mxcsr = 0; /* not needed */
4543 #endif /* KMP_ARCH_X86 || KMP_ARCH_X86_64 */
4544 
4545   team->t.t_construct = 0;
4546 
4547   team->t.t_ordered.dt.t_value = 0;
4548   team->t.t_master_active = FALSE;
4549 
4550 #ifdef KMP_DEBUG
4551   team->t.t_copypriv_data = NULL; /* not necessary, but nice for debugging */
4552 #endif
4553 #if KMP_OS_WINDOWS
4554   team->t.t_copyin_counter = 0; /* for barrier-free copyin implementation */
4555 #endif
4556 
4557   team->t.t_control_stack_top = NULL;
4558 
4559   __kmp_reinitialize_team(team, new_icvs, loc);
4560 
4561   KMP_MB();
4562   KF_TRACE(10, ("__kmp_initialize_team: exit: team=%p\n", team));
4563 }
4564 
4565 #if (KMP_OS_LINUX || KMP_OS_FREEBSD) && KMP_AFFINITY_SUPPORTED
4566 /* Sets full mask for thread and returns old mask, no changes to structures. */
4567 static void
4568 __kmp_set_thread_affinity_mask_full_tmp(kmp_affin_mask_t *old_mask) {
4569   if (KMP_AFFINITY_CAPABLE()) {
4570     int status;
4571     if (old_mask != NULL) {
4572       status = __kmp_get_system_affinity(old_mask, TRUE);
4573       int error = errno;
4574       if (status != 0) {
4575         __kmp_fatal(KMP_MSG(ChangeThreadAffMaskError), KMP_ERR(error),
4576                     __kmp_msg_null);
4577       }
4578     }
4579     __kmp_set_system_affinity(__kmp_affin_fullMask, TRUE);
4580   }
4581 }
4582 #endif
4583 
4584 #if KMP_AFFINITY_SUPPORTED
4585 
4586 // __kmp_partition_places() is the heart of the OpenMP 4.0 affinity mechanism.
4587 // It calculates the worker + primary thread's partition based upon the parent
4588 // thread's partition, and binds each worker to a thread in their partition.
4589 // The primary thread's partition should already include its current binding.
4590 static void __kmp_partition_places(kmp_team_t *team, int update_master_only) {
4591   // Do not partition places for the hidden helper team
4592   if (KMP_HIDDEN_HELPER_TEAM(team))
4593     return;
4594   // Copy the primary thread's place partition to the team struct
4595   kmp_info_t *master_th = team->t.t_threads[0];
4596   KMP_DEBUG_ASSERT(master_th != NULL);
4597   kmp_proc_bind_t proc_bind = team->t.t_proc_bind;
4598   int first_place = master_th->th.th_first_place;
4599   int last_place = master_th->th.th_last_place;
4600   int masters_place = master_th->th.th_current_place;
4601   team->t.t_first_place = first_place;
4602   team->t.t_last_place = last_place;
4603 
4604   KA_TRACE(20, ("__kmp_partition_places: enter: proc_bind = %d T#%d(%d:0) "
4605                 "bound to place %d partition = [%d,%d]\n",
4606                 proc_bind, __kmp_gtid_from_thread(team->t.t_threads[0]),
4607                 team->t.t_id, masters_place, first_place, last_place));
4608 
4609   switch (proc_bind) {
4610 
4611   case proc_bind_default:
4612     // Serial teams might have the proc_bind policy set to proc_bind_default.
4613     // Not an issue -- we don't rebind primary thread for any proc_bind policy.
4614     KMP_DEBUG_ASSERT(team->t.t_nproc == 1);
4615     break;
4616 
4617   case proc_bind_primary: {
4618     int f;
4619     int n_th = team->t.t_nproc;
4620     for (f = 1; f < n_th; f++) {
4621       kmp_info_t *th = team->t.t_threads[f];
4622       KMP_DEBUG_ASSERT(th != NULL);
4623       th->th.th_first_place = first_place;
4624       th->th.th_last_place = last_place;
4625       th->th.th_new_place = masters_place;
4626       if (__kmp_display_affinity && masters_place != th->th.th_current_place &&
4627           team->t.t_display_affinity != 1) {
4628         team->t.t_display_affinity = 1;
4629       }
4630 
4631       KA_TRACE(100, ("__kmp_partition_places: primary: T#%d(%d:%d) place %d "
4632                      "partition = [%d,%d]\n",
4633                      __kmp_gtid_from_thread(team->t.t_threads[f]), team->t.t_id,
4634                      f, masters_place, first_place, last_place));
4635     }
4636   } break;
4637 
4638   case proc_bind_close: {
4639     int f;
4640     int n_th = team->t.t_nproc;
4641     int n_places;
4642     if (first_place <= last_place) {
4643       n_places = last_place - first_place + 1;
4644     } else {
4645       n_places = __kmp_affinity_num_masks - first_place + last_place + 1;
4646     }
4647     if (n_th <= n_places) {
4648       int place = masters_place;
4649       for (f = 1; f < n_th; f++) {
4650         kmp_info_t *th = team->t.t_threads[f];
4651         KMP_DEBUG_ASSERT(th != NULL);
4652 
4653         if (place == last_place) {
4654           place = first_place;
4655         } else if (place == (int)(__kmp_affinity_num_masks - 1)) {
4656           place = 0;
4657         } else {
4658           place++;
4659         }
4660         th->th.th_first_place = first_place;
4661         th->th.th_last_place = last_place;
4662         th->th.th_new_place = place;
4663         if (__kmp_display_affinity && place != th->th.th_current_place &&
4664             team->t.t_display_affinity != 1) {
4665           team->t.t_display_affinity = 1;
4666         }
4667 
4668         KA_TRACE(100, ("__kmp_partition_places: close: T#%d(%d:%d) place %d "
4669                        "partition = [%d,%d]\n",
4670                        __kmp_gtid_from_thread(team->t.t_threads[f]),
4671                        team->t.t_id, f, place, first_place, last_place));
4672       }
4673     } else {
4674       int S, rem, gap, s_count;
4675       S = n_th / n_places;
4676       s_count = 0;
4677       rem = n_th - (S * n_places);
4678       gap = rem > 0 ? n_places / rem : n_places;
4679       int place = masters_place;
4680       int gap_ct = gap;
4681       for (f = 0; f < n_th; f++) {
4682         kmp_info_t *th = team->t.t_threads[f];
4683         KMP_DEBUG_ASSERT(th != NULL);
4684 
4685         th->th.th_first_place = first_place;
4686         th->th.th_last_place = last_place;
4687         th->th.th_new_place = place;
4688         if (__kmp_display_affinity && place != th->th.th_current_place &&
4689             team->t.t_display_affinity != 1) {
4690           team->t.t_display_affinity = 1;
4691         }
4692         s_count++;
4693 
4694         if ((s_count == S) && rem && (gap_ct == gap)) {
4695           // do nothing, add an extra thread to place on next iteration
4696         } else if ((s_count == S + 1) && rem && (gap_ct == gap)) {
4697           // we added an extra thread to this place; move to next place
4698           if (place == last_place) {
4699             place = first_place;
4700           } else if (place == (int)(__kmp_affinity_num_masks - 1)) {
4701             place = 0;
4702           } else {
4703             place++;
4704           }
4705           s_count = 0;
4706           gap_ct = 1;
4707           rem--;
4708         } else if (s_count == S) { // place full; don't add extra
4709           if (place == last_place) {
4710             place = first_place;
4711           } else if (place == (int)(__kmp_affinity_num_masks - 1)) {
4712             place = 0;
4713           } else {
4714             place++;
4715           }
4716           gap_ct++;
4717           s_count = 0;
4718         }
4719 
4720         KA_TRACE(100,
4721                  ("__kmp_partition_places: close: T#%d(%d:%d) place %d "
4722                   "partition = [%d,%d]\n",
4723                   __kmp_gtid_from_thread(team->t.t_threads[f]), team->t.t_id, f,
4724                   th->th.th_new_place, first_place, last_place));
4725       }
4726       KMP_DEBUG_ASSERT(place == masters_place);
4727     }
4728   } break;
4729 
4730   case proc_bind_spread: {
4731     int f;
4732     int n_th = team->t.t_nproc;
4733     int n_places;
4734     int thidx;
4735     if (first_place <= last_place) {
4736       n_places = last_place - first_place + 1;
4737     } else {
4738       n_places = __kmp_affinity_num_masks - first_place + last_place + 1;
4739     }
4740     if (n_th <= n_places) {
4741       int place = -1;
4742 
4743       if (n_places != static_cast<int>(__kmp_affinity_num_masks)) {
4744         int S = n_places / n_th;
4745         int s_count, rem, gap, gap_ct;
4746 
4747         place = masters_place;
4748         rem = n_places - n_th * S;
4749         gap = rem ? n_th / rem : 1;
4750         gap_ct = gap;
4751         thidx = n_th;
4752         if (update_master_only == 1)
4753           thidx = 1;
4754         for (f = 0; f < thidx; f++) {
4755           kmp_info_t *th = team->t.t_threads[f];
4756           KMP_DEBUG_ASSERT(th != NULL);
4757 
4758           th->th.th_first_place = place;
4759           th->th.th_new_place = place;
4760           if (__kmp_display_affinity && place != th->th.th_current_place &&
4761               team->t.t_display_affinity != 1) {
4762             team->t.t_display_affinity = 1;
4763           }
4764           s_count = 1;
4765           while (s_count < S) {
4766             if (place == last_place) {
4767               place = first_place;
4768             } else if (place == (int)(__kmp_affinity_num_masks - 1)) {
4769               place = 0;
4770             } else {
4771               place++;
4772             }
4773             s_count++;
4774           }
4775           if (rem && (gap_ct == gap)) {
4776             if (place == last_place) {
4777               place = first_place;
4778             } else if (place == (int)(__kmp_affinity_num_masks - 1)) {
4779               place = 0;
4780             } else {
4781               place++;
4782             }
4783             rem--;
4784             gap_ct = 0;
4785           }
4786           th->th.th_last_place = place;
4787           gap_ct++;
4788 
4789           if (place == last_place) {
4790             place = first_place;
4791           } else if (place == (int)(__kmp_affinity_num_masks - 1)) {
4792             place = 0;
4793           } else {
4794             place++;
4795           }
4796 
4797           KA_TRACE(100,
4798                    ("__kmp_partition_places: spread: T#%d(%d:%d) place %d "
4799                     "partition = [%d,%d], __kmp_affinity_num_masks: %u\n",
4800                     __kmp_gtid_from_thread(team->t.t_threads[f]), team->t.t_id,
4801                     f, th->th.th_new_place, th->th.th_first_place,
4802                     th->th.th_last_place, __kmp_affinity_num_masks));
4803         }
4804       } else {
4805         /* Having uniform space of available computation places I can create
4806            T partitions of round(P/T) size and put threads into the first
4807            place of each partition. */
4808         double current = static_cast<double>(masters_place);
4809         double spacing =
4810             (static_cast<double>(n_places + 1) / static_cast<double>(n_th));
4811         int first, last;
4812         kmp_info_t *th;
4813 
4814         thidx = n_th + 1;
4815         if (update_master_only == 1)
4816           thidx = 1;
4817         for (f = 0; f < thidx; f++) {
4818           first = static_cast<int>(current);
4819           last = static_cast<int>(current + spacing) - 1;
4820           KMP_DEBUG_ASSERT(last >= first);
4821           if (first >= n_places) {
4822             if (masters_place) {
4823               first -= n_places;
4824               last -= n_places;
4825               if (first == (masters_place + 1)) {
4826                 KMP_DEBUG_ASSERT(f == n_th);
4827                 first--;
4828               }
4829               if (last == masters_place) {
4830                 KMP_DEBUG_ASSERT(f == (n_th - 1));
4831                 last--;
4832               }
4833             } else {
4834               KMP_DEBUG_ASSERT(f == n_th);
4835               first = 0;
4836               last = 0;
4837             }
4838           }
4839           if (last >= n_places) {
4840             last = (n_places - 1);
4841           }
4842           place = first;
4843           current += spacing;
4844           if (f < n_th) {
4845             KMP_DEBUG_ASSERT(0 <= first);
4846             KMP_DEBUG_ASSERT(n_places > first);
4847             KMP_DEBUG_ASSERT(0 <= last);
4848             KMP_DEBUG_ASSERT(n_places > last);
4849             KMP_DEBUG_ASSERT(last_place >= first_place);
4850             th = team->t.t_threads[f];
4851             KMP_DEBUG_ASSERT(th);
4852             th->th.th_first_place = first;
4853             th->th.th_new_place = place;
4854             th->th.th_last_place = last;
4855             if (__kmp_display_affinity && place != th->th.th_current_place &&
4856                 team->t.t_display_affinity != 1) {
4857               team->t.t_display_affinity = 1;
4858             }
4859             KA_TRACE(100,
4860                      ("__kmp_partition_places: spread: T#%d(%d:%d) place %d "
4861                       "partition = [%d,%d], spacing = %.4f\n",
4862                       __kmp_gtid_from_thread(team->t.t_threads[f]),
4863                       team->t.t_id, f, th->th.th_new_place,
4864                       th->th.th_first_place, th->th.th_last_place, spacing));
4865           }
4866         }
4867       }
4868       KMP_DEBUG_ASSERT(update_master_only || place == masters_place);
4869     } else {
4870       int S, rem, gap, s_count;
4871       S = n_th / n_places;
4872       s_count = 0;
4873       rem = n_th - (S * n_places);
4874       gap = rem > 0 ? n_places / rem : n_places;
4875       int place = masters_place;
4876       int gap_ct = gap;
4877       thidx = n_th;
4878       if (update_master_only == 1)
4879         thidx = 1;
4880       for (f = 0; f < thidx; f++) {
4881         kmp_info_t *th = team->t.t_threads[f];
4882         KMP_DEBUG_ASSERT(th != NULL);
4883 
4884         th->th.th_first_place = place;
4885         th->th.th_last_place = place;
4886         th->th.th_new_place = place;
4887         if (__kmp_display_affinity && place != th->th.th_current_place &&
4888             team->t.t_display_affinity != 1) {
4889           team->t.t_display_affinity = 1;
4890         }
4891         s_count++;
4892 
4893         if ((s_count == S) && rem && (gap_ct == gap)) {
4894           // do nothing, add an extra thread to place on next iteration
4895         } else if ((s_count == S + 1) && rem && (gap_ct == gap)) {
4896           // we added an extra thread to this place; move on to next place
4897           if (place == last_place) {
4898             place = first_place;
4899           } else if (place == (int)(__kmp_affinity_num_masks - 1)) {
4900             place = 0;
4901           } else {
4902             place++;
4903           }
4904           s_count = 0;
4905           gap_ct = 1;
4906           rem--;
4907         } else if (s_count == S) { // place is full; don't add extra thread
4908           if (place == last_place) {
4909             place = first_place;
4910           } else if (place == (int)(__kmp_affinity_num_masks - 1)) {
4911             place = 0;
4912           } else {
4913             place++;
4914           }
4915           gap_ct++;
4916           s_count = 0;
4917         }
4918 
4919         KA_TRACE(100, ("__kmp_partition_places: spread: T#%d(%d:%d) place %d "
4920                        "partition = [%d,%d]\n",
4921                        __kmp_gtid_from_thread(team->t.t_threads[f]),
4922                        team->t.t_id, f, th->th.th_new_place,
4923                        th->th.th_first_place, th->th.th_last_place));
4924       }
4925       KMP_DEBUG_ASSERT(update_master_only || place == masters_place);
4926     }
4927   } break;
4928 
4929   default:
4930     break;
4931   }
4932 
4933   KA_TRACE(20, ("__kmp_partition_places: exit T#%d\n", team->t.t_id));
4934 }
4935 
4936 #endif // KMP_AFFINITY_SUPPORTED
4937 
4938 /* allocate a new team data structure to use.  take one off of the free pool if
4939    available */
4940 kmp_team_t *
4941 __kmp_allocate_team(kmp_root_t *root, int new_nproc, int max_nproc,
4942 #if OMPT_SUPPORT
4943                     ompt_data_t ompt_parallel_data,
4944 #endif
4945                     kmp_proc_bind_t new_proc_bind,
4946                     kmp_internal_control_t *new_icvs,
4947                     int argc USE_NESTED_HOT_ARG(kmp_info_t *master)) {
4948   KMP_TIME_DEVELOPER_PARTITIONED_BLOCK(KMP_allocate_team);
4949   int f;
4950   kmp_team_t *team;
4951   int use_hot_team = !root->r.r_active;
4952   int level = 0;
4953 
4954   KA_TRACE(20, ("__kmp_allocate_team: called\n"));
4955   KMP_DEBUG_ASSERT(new_nproc >= 1 && argc >= 0);
4956   KMP_DEBUG_ASSERT(max_nproc >= new_nproc);
4957   KMP_MB();
4958 
4959 #if KMP_NESTED_HOT_TEAMS
4960   kmp_hot_team_ptr_t *hot_teams;
4961   if (master) {
4962     team = master->th.th_team;
4963     level = team->t.t_active_level;
4964     if (master->th.th_teams_microtask) { // in teams construct?
4965       if (master->th.th_teams_size.nteams > 1 &&
4966           ( // #teams > 1
4967               team->t.t_pkfn ==
4968                   (microtask_t)__kmp_teams_master || // inner fork of the teams
4969               master->th.th_teams_level <
4970                   team->t.t_level)) { // or nested parallel inside the teams
4971         ++level; // not increment if #teams==1, or for outer fork of the teams;
4972         // increment otherwise
4973       }
4974     }
4975     hot_teams = master->th.th_hot_teams;
4976     if (level < __kmp_hot_teams_max_level && hot_teams &&
4977         hot_teams[level].hot_team) {
4978       // hot team has already been allocated for given level
4979       use_hot_team = 1;
4980     } else {
4981       use_hot_team = 0;
4982     }
4983   } else {
4984     // check we won't access uninitialized hot_teams, just in case
4985     KMP_DEBUG_ASSERT(new_nproc == 1);
4986   }
4987 #endif
4988   // Optimization to use a "hot" team
4989   if (use_hot_team && new_nproc > 1) {
4990     KMP_DEBUG_ASSERT(new_nproc <= max_nproc);
4991 #if KMP_NESTED_HOT_TEAMS
4992     team = hot_teams[level].hot_team;
4993 #else
4994     team = root->r.r_hot_team;
4995 #endif
4996 #if KMP_DEBUG
4997     if (__kmp_tasking_mode != tskm_immediate_exec) {
4998       KA_TRACE(20, ("__kmp_allocate_team: hot team task_team[0] = %p "
4999                     "task_team[1] = %p before reinit\n",
5000                     team->t.t_task_team[0], team->t.t_task_team[1]));
5001     }
5002 #endif
5003 
5004     // Has the number of threads changed?
5005     /* Let's assume the most common case is that the number of threads is
5006        unchanged, and put that case first. */
5007     if (team->t.t_nproc == new_nproc) { // Check changes in number of threads
5008       KA_TRACE(20, ("__kmp_allocate_team: reusing hot team\n"));
5009       // This case can mean that omp_set_num_threads() was called and the hot
5010       // team size was already reduced, so we check the special flag
5011       if (team->t.t_size_changed == -1) {
5012         team->t.t_size_changed = 1;
5013       } else {
5014         KMP_CHECK_UPDATE(team->t.t_size_changed, 0);
5015       }
5016 
5017       // TODO???: team->t.t_max_active_levels = new_max_active_levels;
5018       kmp_r_sched_t new_sched = new_icvs->sched;
5019       // set primary thread's schedule as new run-time schedule
5020       KMP_CHECK_UPDATE(team->t.t_sched.sched, new_sched.sched);
5021 
5022       __kmp_reinitialize_team(team, new_icvs,
5023                               root->r.r_uber_thread->th.th_ident);
5024 
5025       KF_TRACE(10, ("__kmp_allocate_team2: T#%d, this_thread=%p team=%p\n", 0,
5026                     team->t.t_threads[0], team));
5027       __kmp_push_current_task_to_thread(team->t.t_threads[0], team, 0);
5028 
5029 #if KMP_AFFINITY_SUPPORTED
5030       if ((team->t.t_size_changed == 0) &&
5031           (team->t.t_proc_bind == new_proc_bind)) {
5032         if (new_proc_bind == proc_bind_spread) {
5033           __kmp_partition_places(
5034               team, 1); // add flag to update only master for spread
5035         }
5036         KA_TRACE(200, ("__kmp_allocate_team: reusing hot team #%d bindings: "
5037                        "proc_bind = %d, partition = [%d,%d]\n",
5038                        team->t.t_id, new_proc_bind, team->t.t_first_place,
5039                        team->t.t_last_place));
5040       } else {
5041         KMP_CHECK_UPDATE(team->t.t_proc_bind, new_proc_bind);
5042         __kmp_partition_places(team);
5043       }
5044 #else
5045       KMP_CHECK_UPDATE(team->t.t_proc_bind, new_proc_bind);
5046 #endif /* KMP_AFFINITY_SUPPORTED */
5047     } else if (team->t.t_nproc > new_nproc) {
5048       KA_TRACE(20,
5049                ("__kmp_allocate_team: decreasing hot team thread count to %d\n",
5050                 new_nproc));
5051 
5052       team->t.t_size_changed = 1;
5053 #if KMP_NESTED_HOT_TEAMS
5054       if (__kmp_hot_teams_mode == 0) {
5055         // AC: saved number of threads should correspond to team's value in this
5056         // mode, can be bigger in mode 1, when hot team has threads in reserve
5057         KMP_DEBUG_ASSERT(hot_teams[level].hot_team_nth == team->t.t_nproc);
5058         hot_teams[level].hot_team_nth = new_nproc;
5059 #endif // KMP_NESTED_HOT_TEAMS
5060         /* release the extra threads we don't need any more */
5061         for (f = new_nproc; f < team->t.t_nproc; f++) {
5062           KMP_DEBUG_ASSERT(team->t.t_threads[f]);
5063           if (__kmp_tasking_mode != tskm_immediate_exec) {
5064             // When decreasing team size, threads no longer in the team should
5065             // unref task team.
5066             team->t.t_threads[f]->th.th_task_team = NULL;
5067           }
5068           __kmp_free_thread(team->t.t_threads[f]);
5069           team->t.t_threads[f] = NULL;
5070         }
5071 #if KMP_NESTED_HOT_TEAMS
5072       } // (__kmp_hot_teams_mode == 0)
5073       else {
5074         // When keeping extra threads in team, switch threads to wait on own
5075         // b_go flag
5076         for (f = new_nproc; f < team->t.t_nproc; ++f) {
5077           KMP_DEBUG_ASSERT(team->t.t_threads[f]);
5078           kmp_balign_t *balign = team->t.t_threads[f]->th.th_bar;
5079           for (int b = 0; b < bs_last_barrier; ++b) {
5080             if (balign[b].bb.wait_flag == KMP_BARRIER_PARENT_FLAG) {
5081               balign[b].bb.wait_flag = KMP_BARRIER_SWITCH_TO_OWN_FLAG;
5082             }
5083             KMP_CHECK_UPDATE(balign[b].bb.leaf_kids, 0);
5084           }
5085         }
5086       }
5087 #endif // KMP_NESTED_HOT_TEAMS
5088       team->t.t_nproc = new_nproc;
5089       // TODO???: team->t.t_max_active_levels = new_max_active_levels;
5090       KMP_CHECK_UPDATE(team->t.t_sched.sched, new_icvs->sched.sched);
5091       __kmp_reinitialize_team(team, new_icvs,
5092                               root->r.r_uber_thread->th.th_ident);
5093 
5094       // Update remaining threads
5095       for (f = 0; f < new_nproc; ++f) {
5096         team->t.t_threads[f]->th.th_team_nproc = new_nproc;
5097       }
5098 
5099       // restore the current task state of the primary thread: should be the
5100       // implicit task
5101       KF_TRACE(10, ("__kmp_allocate_team: T#%d, this_thread=%p team=%p\n", 0,
5102                     team->t.t_threads[0], team));
5103 
5104       __kmp_push_current_task_to_thread(team->t.t_threads[0], team, 0);
5105 
5106 #ifdef KMP_DEBUG
5107       for (f = 0; f < team->t.t_nproc; f++) {
5108         KMP_DEBUG_ASSERT(team->t.t_threads[f] &&
5109                          team->t.t_threads[f]->th.th_team_nproc ==
5110                              team->t.t_nproc);
5111       }
5112 #endif
5113 
5114       KMP_CHECK_UPDATE(team->t.t_proc_bind, new_proc_bind);
5115 #if KMP_AFFINITY_SUPPORTED
5116       __kmp_partition_places(team);
5117 #endif
5118     } else { // team->t.t_nproc < new_nproc
5119 #if (KMP_OS_LINUX || KMP_OS_FREEBSD) && KMP_AFFINITY_SUPPORTED
5120       kmp_affin_mask_t *old_mask;
5121       if (KMP_AFFINITY_CAPABLE()) {
5122         KMP_CPU_ALLOC(old_mask);
5123       }
5124 #endif
5125 
5126       KA_TRACE(20,
5127                ("__kmp_allocate_team: increasing hot team thread count to %d\n",
5128                 new_nproc));
5129 
5130       team->t.t_size_changed = 1;
5131 
5132 #if KMP_NESTED_HOT_TEAMS
5133       int avail_threads = hot_teams[level].hot_team_nth;
5134       if (new_nproc < avail_threads)
5135         avail_threads = new_nproc;
5136       kmp_info_t **other_threads = team->t.t_threads;
5137       for (f = team->t.t_nproc; f < avail_threads; ++f) {
5138         // Adjust barrier data of reserved threads (if any) of the team
5139         // Other data will be set in __kmp_initialize_info() below.
5140         int b;
5141         kmp_balign_t *balign = other_threads[f]->th.th_bar;
5142         for (b = 0; b < bs_last_barrier; ++b) {
5143           balign[b].bb.b_arrived = team->t.t_bar[b].b_arrived;
5144           KMP_DEBUG_ASSERT(balign[b].bb.wait_flag != KMP_BARRIER_PARENT_FLAG);
5145 #if USE_DEBUGGER
5146           balign[b].bb.b_worker_arrived = team->t.t_bar[b].b_team_arrived;
5147 #endif
5148         }
5149       }
5150       if (hot_teams[level].hot_team_nth >= new_nproc) {
5151         // we have all needed threads in reserve, no need to allocate any
5152         // this only possible in mode 1, cannot have reserved threads in mode 0
5153         KMP_DEBUG_ASSERT(__kmp_hot_teams_mode == 1);
5154         team->t.t_nproc = new_nproc; // just get reserved threads involved
5155       } else {
5156         // we may have some threads in reserve, but not enough
5157         team->t.t_nproc =
5158             hot_teams[level]
5159                 .hot_team_nth; // get reserved threads involved if any
5160         hot_teams[level].hot_team_nth = new_nproc; // adjust hot team max size
5161 #endif // KMP_NESTED_HOT_TEAMS
5162         if (team->t.t_max_nproc < new_nproc) {
5163           /* reallocate larger arrays */
5164           __kmp_reallocate_team_arrays(team, new_nproc);
5165           __kmp_reinitialize_team(team, new_icvs, NULL);
5166         }
5167 
5168 #if (KMP_OS_LINUX || KMP_OS_FREEBSD) && KMP_AFFINITY_SUPPORTED
5169         /* Temporarily set full mask for primary thread before creation of
5170            workers. The reason is that workers inherit the affinity from the
5171            primary thread, so if a lot of workers are created on the single
5172            core quickly, they don't get a chance to set their own affinity for
5173            a long time. */
5174         __kmp_set_thread_affinity_mask_full_tmp(old_mask);
5175 #endif
5176 
5177         /* allocate new threads for the hot team */
5178         for (f = team->t.t_nproc; f < new_nproc; f++) {
5179           kmp_info_t *new_worker = __kmp_allocate_thread(root, team, f);
5180           KMP_DEBUG_ASSERT(new_worker);
5181           team->t.t_threads[f] = new_worker;
5182 
5183           KA_TRACE(20,
5184                    ("__kmp_allocate_team: team %d init T#%d arrived: "
5185                     "join=%llu, plain=%llu\n",
5186                     team->t.t_id, __kmp_gtid_from_tid(f, team), team->t.t_id, f,
5187                     team->t.t_bar[bs_forkjoin_barrier].b_arrived,
5188                     team->t.t_bar[bs_plain_barrier].b_arrived));
5189 
5190           { // Initialize barrier data for new threads.
5191             int b;
5192             kmp_balign_t *balign = new_worker->th.th_bar;
5193             for (b = 0; b < bs_last_barrier; ++b) {
5194               balign[b].bb.b_arrived = team->t.t_bar[b].b_arrived;
5195               KMP_DEBUG_ASSERT(balign[b].bb.wait_flag !=
5196                                KMP_BARRIER_PARENT_FLAG);
5197 #if USE_DEBUGGER
5198               balign[b].bb.b_worker_arrived = team->t.t_bar[b].b_team_arrived;
5199 #endif
5200             }
5201           }
5202         }
5203 
5204 #if (KMP_OS_LINUX || KMP_OS_FREEBSD) && KMP_AFFINITY_SUPPORTED
5205         if (KMP_AFFINITY_CAPABLE()) {
5206           /* Restore initial primary thread's affinity mask */
5207           __kmp_set_system_affinity(old_mask, TRUE);
5208           KMP_CPU_FREE(old_mask);
5209         }
5210 #endif
5211 #if KMP_NESTED_HOT_TEAMS
5212       } // end of check of t_nproc vs. new_nproc vs. hot_team_nth
5213 #endif // KMP_NESTED_HOT_TEAMS
5214       /* make sure everyone is syncronized */
5215       int old_nproc = team->t.t_nproc; // save old value and use to update only
5216       // new threads below
5217       __kmp_initialize_team(team, new_nproc, new_icvs,
5218                             root->r.r_uber_thread->th.th_ident);
5219 
5220       /* reinitialize the threads */
5221       KMP_DEBUG_ASSERT(team->t.t_nproc == new_nproc);
5222       for (f = 0; f < team->t.t_nproc; ++f)
5223         __kmp_initialize_info(team->t.t_threads[f], team, f,
5224                               __kmp_gtid_from_tid(f, team));
5225 
5226       if (level) { // set th_task_state for new threads in nested hot team
5227         // __kmp_initialize_info() no longer zeroes th_task_state, so we should
5228         // only need to set the th_task_state for the new threads. th_task_state
5229         // for primary thread will not be accurate until after this in
5230         // __kmp_fork_call(), so we look to the primary thread's memo_stack to
5231         // get the correct value.
5232         for (f = old_nproc; f < team->t.t_nproc; ++f)
5233           team->t.t_threads[f]->th.th_task_state =
5234               team->t.t_threads[0]->th.th_task_state_memo_stack[level];
5235       } else { // set th_task_state for new threads in non-nested hot team
5236         // copy primary thread's state
5237         kmp_uint8 old_state = team->t.t_threads[0]->th.th_task_state;
5238         for (f = old_nproc; f < team->t.t_nproc; ++f)
5239           team->t.t_threads[f]->th.th_task_state = old_state;
5240       }
5241 
5242 #ifdef KMP_DEBUG
5243       for (f = 0; f < team->t.t_nproc; ++f) {
5244         KMP_DEBUG_ASSERT(team->t.t_threads[f] &&
5245                          team->t.t_threads[f]->th.th_team_nproc ==
5246                              team->t.t_nproc);
5247       }
5248 #endif
5249 
5250       KMP_CHECK_UPDATE(team->t.t_proc_bind, new_proc_bind);
5251 #if KMP_AFFINITY_SUPPORTED
5252       __kmp_partition_places(team);
5253 #endif
5254     } // Check changes in number of threads
5255 
5256     kmp_info_t *master = team->t.t_threads[0];
5257     if (master->th.th_teams_microtask) {
5258       for (f = 1; f < new_nproc; ++f) {
5259         // propagate teams construct specific info to workers
5260         kmp_info_t *thr = team->t.t_threads[f];
5261         thr->th.th_teams_microtask = master->th.th_teams_microtask;
5262         thr->th.th_teams_level = master->th.th_teams_level;
5263         thr->th.th_teams_size = master->th.th_teams_size;
5264       }
5265     }
5266 #if KMP_NESTED_HOT_TEAMS
5267     if (level) {
5268       // Sync barrier state for nested hot teams, not needed for outermost hot
5269       // team.
5270       for (f = 1; f < new_nproc; ++f) {
5271         kmp_info_t *thr = team->t.t_threads[f];
5272         int b;
5273         kmp_balign_t *balign = thr->th.th_bar;
5274         for (b = 0; b < bs_last_barrier; ++b) {
5275           balign[b].bb.b_arrived = team->t.t_bar[b].b_arrived;
5276           KMP_DEBUG_ASSERT(balign[b].bb.wait_flag != KMP_BARRIER_PARENT_FLAG);
5277 #if USE_DEBUGGER
5278           balign[b].bb.b_worker_arrived = team->t.t_bar[b].b_team_arrived;
5279 #endif
5280         }
5281       }
5282     }
5283 #endif // KMP_NESTED_HOT_TEAMS
5284 
5285     /* reallocate space for arguments if necessary */
5286     __kmp_alloc_argv_entries(argc, team, TRUE);
5287     KMP_CHECK_UPDATE(team->t.t_argc, argc);
5288     // The hot team re-uses the previous task team,
5289     // if untouched during the previous release->gather phase.
5290 
5291     KF_TRACE(10, (" hot_team = %p\n", team));
5292 
5293 #if KMP_DEBUG
5294     if (__kmp_tasking_mode != tskm_immediate_exec) {
5295       KA_TRACE(20, ("__kmp_allocate_team: hot team task_team[0] = %p "
5296                     "task_team[1] = %p after reinit\n",
5297                     team->t.t_task_team[0], team->t.t_task_team[1]));
5298     }
5299 #endif
5300 
5301 #if OMPT_SUPPORT
5302     __ompt_team_assign_id(team, ompt_parallel_data);
5303 #endif
5304 
5305     KMP_MB();
5306 
5307     return team;
5308   }
5309 
5310   /* next, let's try to take one from the team pool */
5311   KMP_MB();
5312   for (team = CCAST(kmp_team_t *, __kmp_team_pool); (team);) {
5313     /* TODO: consider resizing undersized teams instead of reaping them, now
5314        that we have a resizing mechanism */
5315     if (team->t.t_max_nproc >= max_nproc) {
5316       /* take this team from the team pool */
5317       __kmp_team_pool = team->t.t_next_pool;
5318 
5319       /* setup the team for fresh use */
5320       __kmp_initialize_team(team, new_nproc, new_icvs, NULL);
5321 
5322       KA_TRACE(20, ("__kmp_allocate_team: setting task_team[0] %p and "
5323                     "task_team[1] %p to NULL\n",
5324                     &team->t.t_task_team[0], &team->t.t_task_team[1]));
5325       team->t.t_task_team[0] = NULL;
5326       team->t.t_task_team[1] = NULL;
5327 
5328       /* reallocate space for arguments if necessary */
5329       __kmp_alloc_argv_entries(argc, team, TRUE);
5330       KMP_CHECK_UPDATE(team->t.t_argc, argc);
5331 
5332       KA_TRACE(
5333           20, ("__kmp_allocate_team: team %d init arrived: join=%u, plain=%u\n",
5334                team->t.t_id, KMP_INIT_BARRIER_STATE, KMP_INIT_BARRIER_STATE));
5335       { // Initialize barrier data.
5336         int b;
5337         for (b = 0; b < bs_last_barrier; ++b) {
5338           team->t.t_bar[b].b_arrived = KMP_INIT_BARRIER_STATE;
5339 #if USE_DEBUGGER
5340           team->t.t_bar[b].b_master_arrived = 0;
5341           team->t.t_bar[b].b_team_arrived = 0;
5342 #endif
5343         }
5344       }
5345 
5346       team->t.t_proc_bind = new_proc_bind;
5347 
5348       KA_TRACE(20, ("__kmp_allocate_team: using team from pool %d.\n",
5349                     team->t.t_id));
5350 
5351 #if OMPT_SUPPORT
5352       __ompt_team_assign_id(team, ompt_parallel_data);
5353 #endif
5354 
5355       KMP_MB();
5356 
5357       return team;
5358     }
5359 
5360     /* reap team if it is too small, then loop back and check the next one */
5361     // not sure if this is wise, but, will be redone during the hot-teams
5362     // rewrite.
5363     /* TODO: Use technique to find the right size hot-team, don't reap them */
5364     team = __kmp_reap_team(team);
5365     __kmp_team_pool = team;
5366   }
5367 
5368   /* nothing available in the pool, no matter, make a new team! */
5369   KMP_MB();
5370   team = (kmp_team_t *)__kmp_allocate(sizeof(kmp_team_t));
5371 
5372   /* and set it up */
5373   team->t.t_max_nproc = max_nproc;
5374   /* NOTE well, for some reason allocating one big buffer and dividing it up
5375      seems to really hurt performance a lot on the P4, so, let's not use this */
5376   __kmp_allocate_team_arrays(team, max_nproc);
5377 
5378   KA_TRACE(20, ("__kmp_allocate_team: making a new team\n"));
5379   __kmp_initialize_team(team, new_nproc, new_icvs, NULL);
5380 
5381   KA_TRACE(20, ("__kmp_allocate_team: setting task_team[0] %p and task_team[1] "
5382                 "%p to NULL\n",
5383                 &team->t.t_task_team[0], &team->t.t_task_team[1]));
5384   team->t.t_task_team[0] = NULL; // to be removed, as __kmp_allocate zeroes
5385   // memory, no need to duplicate
5386   team->t.t_task_team[1] = NULL; // to be removed, as __kmp_allocate zeroes
5387   // memory, no need to duplicate
5388 
5389   if (__kmp_storage_map) {
5390     __kmp_print_team_storage_map("team", team, team->t.t_id, new_nproc);
5391   }
5392 
5393   /* allocate space for arguments */
5394   __kmp_alloc_argv_entries(argc, team, FALSE);
5395   team->t.t_argc = argc;
5396 
5397   KA_TRACE(20,
5398            ("__kmp_allocate_team: team %d init arrived: join=%u, plain=%u\n",
5399             team->t.t_id, KMP_INIT_BARRIER_STATE, KMP_INIT_BARRIER_STATE));
5400   { // Initialize barrier data.
5401     int b;
5402     for (b = 0; b < bs_last_barrier; ++b) {
5403       team->t.t_bar[b].b_arrived = KMP_INIT_BARRIER_STATE;
5404 #if USE_DEBUGGER
5405       team->t.t_bar[b].b_master_arrived = 0;
5406       team->t.t_bar[b].b_team_arrived = 0;
5407 #endif
5408     }
5409   }
5410 
5411   team->t.t_proc_bind = new_proc_bind;
5412 
5413 #if OMPT_SUPPORT
5414   __ompt_team_assign_id(team, ompt_parallel_data);
5415   team->t.ompt_serialized_team_info = NULL;
5416 #endif
5417 
5418   KMP_MB();
5419 
5420   KA_TRACE(20, ("__kmp_allocate_team: done creating a new team %d.\n",
5421                 team->t.t_id));
5422 
5423   return team;
5424 }
5425 
5426 /* TODO implement hot-teams at all levels */
5427 /* TODO implement lazy thread release on demand (disband request) */
5428 
5429 /* free the team.  return it to the team pool.  release all the threads
5430  * associated with it */
5431 void __kmp_free_team(kmp_root_t *root,
5432                      kmp_team_t *team USE_NESTED_HOT_ARG(kmp_info_t *master)) {
5433   int f;
5434   KA_TRACE(20, ("__kmp_free_team: T#%d freeing team %d\n", __kmp_get_gtid(),
5435                 team->t.t_id));
5436 
5437   /* verify state */
5438   KMP_DEBUG_ASSERT(root);
5439   KMP_DEBUG_ASSERT(team);
5440   KMP_DEBUG_ASSERT(team->t.t_nproc <= team->t.t_max_nproc);
5441   KMP_DEBUG_ASSERT(team->t.t_threads);
5442 
5443   int use_hot_team = team == root->r.r_hot_team;
5444 #if KMP_NESTED_HOT_TEAMS
5445   int level;
5446   kmp_hot_team_ptr_t *hot_teams;
5447   if (master) {
5448     level = team->t.t_active_level - 1;
5449     if (master->th.th_teams_microtask) { // in teams construct?
5450       if (master->th.th_teams_size.nteams > 1) {
5451         ++level; // level was not increased in teams construct for
5452         // team_of_masters
5453       }
5454       if (team->t.t_pkfn != (microtask_t)__kmp_teams_master &&
5455           master->th.th_teams_level == team->t.t_level) {
5456         ++level; // level was not increased in teams construct for
5457         // team_of_workers before the parallel
5458       } // team->t.t_level will be increased inside parallel
5459     }
5460     hot_teams = master->th.th_hot_teams;
5461     if (level < __kmp_hot_teams_max_level) {
5462       KMP_DEBUG_ASSERT(team == hot_teams[level].hot_team);
5463       use_hot_team = 1;
5464     }
5465   }
5466 #endif // KMP_NESTED_HOT_TEAMS
5467 
5468   /* team is done working */
5469   TCW_SYNC_PTR(team->t.t_pkfn,
5470                NULL); // Important for Debugging Support Library.
5471 #if KMP_OS_WINDOWS
5472   team->t.t_copyin_counter = 0; // init counter for possible reuse
5473 #endif
5474   // Do not reset pointer to parent team to NULL for hot teams.
5475 
5476   /* if we are non-hot team, release our threads */
5477   if (!use_hot_team) {
5478     if (__kmp_tasking_mode != tskm_immediate_exec) {
5479       // Wait for threads to reach reapable state
5480       for (f = 1; f < team->t.t_nproc; ++f) {
5481         KMP_DEBUG_ASSERT(team->t.t_threads[f]);
5482         kmp_info_t *th = team->t.t_threads[f];
5483         volatile kmp_uint32 *state = &th->th.th_reap_state;
5484         while (*state != KMP_SAFE_TO_REAP) {
5485 #if KMP_OS_WINDOWS
5486           // On Windows a thread can be killed at any time, check this
5487           DWORD ecode;
5488           if (!__kmp_is_thread_alive(th, &ecode)) {
5489             *state = KMP_SAFE_TO_REAP; // reset the flag for dead thread
5490             break;
5491           }
5492 #endif
5493           // first check if thread is sleeping
5494           kmp_flag_64<> fl(&th->th.th_bar[bs_forkjoin_barrier].bb.b_go, th);
5495           if (fl.is_sleeping())
5496             fl.resume(__kmp_gtid_from_thread(th));
5497           KMP_CPU_PAUSE();
5498         }
5499       }
5500 
5501       // Delete task teams
5502       int tt_idx;
5503       for (tt_idx = 0; tt_idx < 2; ++tt_idx) {
5504         kmp_task_team_t *task_team = team->t.t_task_team[tt_idx];
5505         if (task_team != NULL) {
5506           for (f = 0; f < team->t.t_nproc; ++f) { // threads unref task teams
5507             KMP_DEBUG_ASSERT(team->t.t_threads[f]);
5508             team->t.t_threads[f]->th.th_task_team = NULL;
5509           }
5510           KA_TRACE(
5511               20,
5512               ("__kmp_free_team: T#%d deactivating task_team %p on team %d\n",
5513                __kmp_get_gtid(), task_team, team->t.t_id));
5514 #if KMP_NESTED_HOT_TEAMS
5515           __kmp_free_task_team(master, task_team);
5516 #endif
5517           team->t.t_task_team[tt_idx] = NULL;
5518         }
5519       }
5520     }
5521 
5522     // Reset pointer to parent team only for non-hot teams.
5523     team->t.t_parent = NULL;
5524     team->t.t_level = 0;
5525     team->t.t_active_level = 0;
5526 
5527     /* free the worker threads */
5528     for (f = 1; f < team->t.t_nproc; ++f) {
5529       KMP_DEBUG_ASSERT(team->t.t_threads[f]);
5530       __kmp_free_thread(team->t.t_threads[f]);
5531       team->t.t_threads[f] = NULL;
5532     }
5533 
5534     /* put the team back in the team pool */
5535     /* TODO limit size of team pool, call reap_team if pool too large */
5536     team->t.t_next_pool = CCAST(kmp_team_t *, __kmp_team_pool);
5537     __kmp_team_pool = (volatile kmp_team_t *)team;
5538   } else { // Check if team was created for primary threads in teams construct
5539     // See if first worker is a CG root
5540     KMP_DEBUG_ASSERT(team->t.t_threads[1] &&
5541                      team->t.t_threads[1]->th.th_cg_roots);
5542     if (team->t.t_threads[1]->th.th_cg_roots->cg_root == team->t.t_threads[1]) {
5543       // Clean up the CG root nodes on workers so that this team can be re-used
5544       for (f = 1; f < team->t.t_nproc; ++f) {
5545         kmp_info_t *thr = team->t.t_threads[f];
5546         KMP_DEBUG_ASSERT(thr && thr->th.th_cg_roots &&
5547                          thr->th.th_cg_roots->cg_root == thr);
5548         // Pop current CG root off list
5549         kmp_cg_root_t *tmp = thr->th.th_cg_roots;
5550         thr->th.th_cg_roots = tmp->up;
5551         KA_TRACE(100, ("__kmp_free_team: Thread %p popping node %p and moving"
5552                        " up to node %p. cg_nthreads was %d\n",
5553                        thr, tmp, thr->th.th_cg_roots, tmp->cg_nthreads));
5554         int i = tmp->cg_nthreads--;
5555         if (i == 1) {
5556           __kmp_free(tmp); // free CG if we are the last thread in it
5557         }
5558         // Restore current task's thread_limit from CG root
5559         if (thr->th.th_cg_roots)
5560           thr->th.th_current_task->td_icvs.thread_limit =
5561               thr->th.th_cg_roots->cg_thread_limit;
5562       }
5563     }
5564   }
5565 
5566   KMP_MB();
5567 }
5568 
5569 /* reap the team.  destroy it, reclaim all its resources and free its memory */
5570 kmp_team_t *__kmp_reap_team(kmp_team_t *team) {
5571   kmp_team_t *next_pool = team->t.t_next_pool;
5572 
5573   KMP_DEBUG_ASSERT(team);
5574   KMP_DEBUG_ASSERT(team->t.t_dispatch);
5575   KMP_DEBUG_ASSERT(team->t.t_disp_buffer);
5576   KMP_DEBUG_ASSERT(team->t.t_threads);
5577   KMP_DEBUG_ASSERT(team->t.t_argv);
5578 
5579   /* TODO clean the threads that are a part of this? */
5580 
5581   /* free stuff */
5582   __kmp_free_team_arrays(team);
5583   if (team->t.t_argv != &team->t.t_inline_argv[0])
5584     __kmp_free((void *)team->t.t_argv);
5585   __kmp_free(team);
5586 
5587   KMP_MB();
5588   return next_pool;
5589 }
5590 
5591 // Free the thread.  Don't reap it, just place it on the pool of available
5592 // threads.
5593 //
5594 // Changes for Quad issue 527845: We need a predictable OMP tid <-> gtid
5595 // binding for the affinity mechanism to be useful.
5596 //
5597 // Now, we always keep the free list (__kmp_thread_pool) sorted by gtid.
5598 // However, we want to avoid a potential performance problem by always
5599 // scanning through the list to find the correct point at which to insert
5600 // the thread (potential N**2 behavior).  To do this we keep track of the
5601 // last place a thread struct was inserted (__kmp_thread_pool_insert_pt).
5602 // With single-level parallelism, threads will always be added to the tail
5603 // of the list, kept track of by __kmp_thread_pool_insert_pt.  With nested
5604 // parallelism, all bets are off and we may need to scan through the entire
5605 // free list.
5606 //
5607 // This change also has a potentially large performance benefit, for some
5608 // applications.  Previously, as threads were freed from the hot team, they
5609 // would be placed back on the free list in inverse order.  If the hot team
5610 // grew back to it's original size, then the freed thread would be placed
5611 // back on the hot team in reverse order.  This could cause bad cache
5612 // locality problems on programs where the size of the hot team regularly
5613 // grew and shrunk.
5614 //
5615 // Now, for single-level parallelism, the OMP tid is always == gtid.
5616 void __kmp_free_thread(kmp_info_t *this_th) {
5617   int gtid;
5618   kmp_info_t **scan;
5619 
5620   KA_TRACE(20, ("__kmp_free_thread: T#%d putting T#%d back on free pool.\n",
5621                 __kmp_get_gtid(), this_th->th.th_info.ds.ds_gtid));
5622 
5623   KMP_DEBUG_ASSERT(this_th);
5624 
5625   // When moving thread to pool, switch thread to wait on own b_go flag, and
5626   // uninitialized (NULL team).
5627   int b;
5628   kmp_balign_t *balign = this_th->th.th_bar;
5629   for (b = 0; b < bs_last_barrier; ++b) {
5630     if (balign[b].bb.wait_flag == KMP_BARRIER_PARENT_FLAG)
5631       balign[b].bb.wait_flag = KMP_BARRIER_SWITCH_TO_OWN_FLAG;
5632     balign[b].bb.team = NULL;
5633     balign[b].bb.leaf_kids = 0;
5634   }
5635   this_th->th.th_task_state = 0;
5636   this_th->th.th_reap_state = KMP_SAFE_TO_REAP;
5637 
5638   /* put thread back on the free pool */
5639   TCW_PTR(this_th->th.th_team, NULL);
5640   TCW_PTR(this_th->th.th_root, NULL);
5641   TCW_PTR(this_th->th.th_dispatch, NULL); /* NOT NEEDED */
5642 
5643   while (this_th->th.th_cg_roots) {
5644     this_th->th.th_cg_roots->cg_nthreads--;
5645     KA_TRACE(100, ("__kmp_free_thread: Thread %p decrement cg_nthreads on node"
5646                    " %p of thread  %p to %d\n",
5647                    this_th, this_th->th.th_cg_roots,
5648                    this_th->th.th_cg_roots->cg_root,
5649                    this_th->th.th_cg_roots->cg_nthreads));
5650     kmp_cg_root_t *tmp = this_th->th.th_cg_roots;
5651     if (tmp->cg_root == this_th) { // Thread is a cg_root
5652       KMP_DEBUG_ASSERT(tmp->cg_nthreads == 0);
5653       KA_TRACE(
5654           5, ("__kmp_free_thread: Thread %p freeing node %p\n", this_th, tmp));
5655       this_th->th.th_cg_roots = tmp->up;
5656       __kmp_free(tmp);
5657     } else { // Worker thread
5658       if (tmp->cg_nthreads == 0) { // last thread leaves contention group
5659         __kmp_free(tmp);
5660       }
5661       this_th->th.th_cg_roots = NULL;
5662       break;
5663     }
5664   }
5665 
5666   /* If the implicit task assigned to this thread can be used by other threads
5667    * -> multiple threads can share the data and try to free the task at
5668    * __kmp_reap_thread at exit. This duplicate use of the task data can happen
5669    * with higher probability when hot team is disabled but can occurs even when
5670    * the hot team is enabled */
5671   __kmp_free_implicit_task(this_th);
5672   this_th->th.th_current_task = NULL;
5673 
5674   // If the __kmp_thread_pool_insert_pt is already past the new insert
5675   // point, then we need to re-scan the entire list.
5676   gtid = this_th->th.th_info.ds.ds_gtid;
5677   if (__kmp_thread_pool_insert_pt != NULL) {
5678     KMP_DEBUG_ASSERT(__kmp_thread_pool != NULL);
5679     if (__kmp_thread_pool_insert_pt->th.th_info.ds.ds_gtid > gtid) {
5680       __kmp_thread_pool_insert_pt = NULL;
5681     }
5682   }
5683 
5684   // Scan down the list to find the place to insert the thread.
5685   // scan is the address of a link in the list, possibly the address of
5686   // __kmp_thread_pool itself.
5687   //
5688   // In the absence of nested parallelism, the for loop will have 0 iterations.
5689   if (__kmp_thread_pool_insert_pt != NULL) {
5690     scan = &(__kmp_thread_pool_insert_pt->th.th_next_pool);
5691   } else {
5692     scan = CCAST(kmp_info_t **, &__kmp_thread_pool);
5693   }
5694   for (; (*scan != NULL) && ((*scan)->th.th_info.ds.ds_gtid < gtid);
5695        scan = &((*scan)->th.th_next_pool))
5696     ;
5697 
5698   // Insert the new element on the list, and set __kmp_thread_pool_insert_pt
5699   // to its address.
5700   TCW_PTR(this_th->th.th_next_pool, *scan);
5701   __kmp_thread_pool_insert_pt = *scan = this_th;
5702   KMP_DEBUG_ASSERT((this_th->th.th_next_pool == NULL) ||
5703                    (this_th->th.th_info.ds.ds_gtid <
5704                     this_th->th.th_next_pool->th.th_info.ds.ds_gtid));
5705   TCW_4(this_th->th.th_in_pool, TRUE);
5706   __kmp_suspend_initialize_thread(this_th);
5707   __kmp_lock_suspend_mx(this_th);
5708   if (this_th->th.th_active == TRUE) {
5709     KMP_ATOMIC_INC(&__kmp_thread_pool_active_nth);
5710     this_th->th.th_active_in_pool = TRUE;
5711   }
5712 #if KMP_DEBUG
5713   else {
5714     KMP_DEBUG_ASSERT(this_th->th.th_active_in_pool == FALSE);
5715   }
5716 #endif
5717   __kmp_unlock_suspend_mx(this_th);
5718 
5719   TCW_4(__kmp_nth, __kmp_nth - 1);
5720 
5721 #ifdef KMP_ADJUST_BLOCKTIME
5722   /* Adjust blocktime back to user setting or default if necessary */
5723   /* Middle initialization might never have occurred                */
5724   if (!__kmp_env_blocktime && (__kmp_avail_proc > 0)) {
5725     KMP_DEBUG_ASSERT(__kmp_avail_proc > 0);
5726     if (__kmp_nth <= __kmp_avail_proc) {
5727       __kmp_zero_bt = FALSE;
5728     }
5729   }
5730 #endif /* KMP_ADJUST_BLOCKTIME */
5731 
5732   KMP_MB();
5733 }
5734 
5735 /* ------------------------------------------------------------------------ */
5736 
5737 void *__kmp_launch_thread(kmp_info_t *this_thr) {
5738 #if OMP_PROFILING_SUPPORT
5739   ProfileTraceFile = getenv("LIBOMPTARGET_PROFILE");
5740   // TODO: add a configuration option for time granularity
5741   if (ProfileTraceFile)
5742     llvm::timeTraceProfilerInitialize(500 /* us */, "libomptarget");
5743 #endif
5744 
5745   int gtid = this_thr->th.th_info.ds.ds_gtid;
5746   /*    void                 *stack_data;*/
5747   kmp_team_t **volatile pteam;
5748 
5749   KMP_MB();
5750   KA_TRACE(10, ("__kmp_launch_thread: T#%d start\n", gtid));
5751 
5752   if (__kmp_env_consistency_check) {
5753     this_thr->th.th_cons = __kmp_allocate_cons_stack(gtid); // ATT: Memory leak?
5754   }
5755 
5756 #if OMPT_SUPPORT
5757   ompt_data_t *thread_data = nullptr;
5758   if (ompt_enabled.enabled) {
5759     thread_data = &(this_thr->th.ompt_thread_info.thread_data);
5760     *thread_data = ompt_data_none;
5761 
5762     this_thr->th.ompt_thread_info.state = ompt_state_overhead;
5763     this_thr->th.ompt_thread_info.wait_id = 0;
5764     this_thr->th.ompt_thread_info.idle_frame = OMPT_GET_FRAME_ADDRESS(0);
5765     this_thr->th.ompt_thread_info.parallel_flags = 0;
5766     if (ompt_enabled.ompt_callback_thread_begin) {
5767       ompt_callbacks.ompt_callback(ompt_callback_thread_begin)(
5768           ompt_thread_worker, thread_data);
5769     }
5770     this_thr->th.ompt_thread_info.state = ompt_state_idle;
5771   }
5772 #endif
5773 
5774   /* This is the place where threads wait for work */
5775   while (!TCR_4(__kmp_global.g.g_done)) {
5776     KMP_DEBUG_ASSERT(this_thr == __kmp_threads[gtid]);
5777     KMP_MB();
5778 
5779     /* wait for work to do */
5780     KA_TRACE(20, ("__kmp_launch_thread: T#%d waiting for work\n", gtid));
5781 
5782     /* No tid yet since not part of a team */
5783     __kmp_fork_barrier(gtid, KMP_GTID_DNE);
5784 
5785 #if OMPT_SUPPORT
5786     if (ompt_enabled.enabled) {
5787       this_thr->th.ompt_thread_info.state = ompt_state_overhead;
5788     }
5789 #endif
5790 
5791     pteam = &this_thr->th.th_team;
5792 
5793     /* have we been allocated? */
5794     if (TCR_SYNC_PTR(*pteam) && !TCR_4(__kmp_global.g.g_done)) {
5795       /* we were just woken up, so run our new task */
5796       if (TCR_SYNC_PTR((*pteam)->t.t_pkfn) != NULL) {
5797         int rc;
5798         KA_TRACE(20,
5799                  ("__kmp_launch_thread: T#%d(%d:%d) invoke microtask = %p\n",
5800                   gtid, (*pteam)->t.t_id, __kmp_tid_from_gtid(gtid),
5801                   (*pteam)->t.t_pkfn));
5802 
5803         updateHWFPControl(*pteam);
5804 
5805 #if OMPT_SUPPORT
5806         if (ompt_enabled.enabled) {
5807           this_thr->th.ompt_thread_info.state = ompt_state_work_parallel;
5808         }
5809 #endif
5810 
5811         rc = (*pteam)->t.t_invoke(gtid);
5812         KMP_ASSERT(rc);
5813 
5814         KMP_MB();
5815         KA_TRACE(20, ("__kmp_launch_thread: T#%d(%d:%d) done microtask = %p\n",
5816                       gtid, (*pteam)->t.t_id, __kmp_tid_from_gtid(gtid),
5817                       (*pteam)->t.t_pkfn));
5818       }
5819 #if OMPT_SUPPORT
5820       if (ompt_enabled.enabled) {
5821         /* no frame set while outside task */
5822         __ompt_get_task_info_object(0)->frame.exit_frame = ompt_data_none;
5823 
5824         this_thr->th.ompt_thread_info.state = ompt_state_overhead;
5825       }
5826 #endif
5827       /* join barrier after parallel region */
5828       __kmp_join_barrier(gtid);
5829     }
5830   }
5831   TCR_SYNC_PTR((intptr_t)__kmp_global.g.g_done);
5832 
5833 #if OMPT_SUPPORT
5834   if (ompt_enabled.ompt_callback_thread_end) {
5835     ompt_callbacks.ompt_callback(ompt_callback_thread_end)(thread_data);
5836   }
5837 #endif
5838 
5839   this_thr->th.th_task_team = NULL;
5840   /* run the destructors for the threadprivate data for this thread */
5841   __kmp_common_destroy_gtid(gtid);
5842 
5843   KA_TRACE(10, ("__kmp_launch_thread: T#%d done\n", gtid));
5844   KMP_MB();
5845 
5846 #if OMP_PROFILING_SUPPORT
5847   llvm::timeTraceProfilerFinishThread();
5848 #endif
5849   return this_thr;
5850 }
5851 
5852 /* ------------------------------------------------------------------------ */
5853 
5854 void __kmp_internal_end_dest(void *specific_gtid) {
5855   // Make sure no significant bits are lost
5856   int gtid;
5857   __kmp_type_convert((kmp_intptr_t)specific_gtid - 1, &gtid);
5858 
5859   KA_TRACE(30, ("__kmp_internal_end_dest: T#%d\n", gtid));
5860   /* NOTE: the gtid is stored as gitd+1 in the thread-local-storage
5861    * this is because 0 is reserved for the nothing-stored case */
5862 
5863   __kmp_internal_end_thread(gtid);
5864 }
5865 
5866 #if KMP_OS_UNIX && KMP_DYNAMIC_LIB
5867 
5868 __attribute__((destructor)) void __kmp_internal_end_dtor(void) {
5869   __kmp_internal_end_atexit();
5870 }
5871 
5872 #endif
5873 
5874 /* [Windows] josh: when the atexit handler is called, there may still be more
5875    than one thread alive */
5876 void __kmp_internal_end_atexit(void) {
5877   KA_TRACE(30, ("__kmp_internal_end_atexit\n"));
5878   /* [Windows]
5879      josh: ideally, we want to completely shutdown the library in this atexit
5880      handler, but stat code that depends on thread specific data for gtid fails
5881      because that data becomes unavailable at some point during the shutdown, so
5882      we call __kmp_internal_end_thread instead. We should eventually remove the
5883      dependency on __kmp_get_specific_gtid in the stat code and use
5884      __kmp_internal_end_library to cleanly shutdown the library.
5885 
5886      // TODO: Can some of this comment about GVS be removed?
5887      I suspect that the offending stat code is executed when the calling thread
5888      tries to clean up a dead root thread's data structures, resulting in GVS
5889      code trying to close the GVS structures for that thread, but since the stat
5890      code uses __kmp_get_specific_gtid to get the gtid with the assumption that
5891      the calling thread is cleaning up itself instead of another thread, it get
5892      confused. This happens because allowing a thread to unregister and cleanup
5893      another thread is a recent modification for addressing an issue.
5894      Based on the current design (20050722), a thread may end up
5895      trying to unregister another thread only if thread death does not trigger
5896      the calling of __kmp_internal_end_thread.  For Linux* OS, there is the
5897      thread specific data destructor function to detect thread death. For
5898      Windows dynamic, there is DllMain(THREAD_DETACH). For Windows static, there
5899      is nothing.  Thus, the workaround is applicable only for Windows static
5900      stat library. */
5901   __kmp_internal_end_library(-1);
5902 #if KMP_OS_WINDOWS
5903   __kmp_close_console();
5904 #endif
5905 }
5906 
5907 static void __kmp_reap_thread(kmp_info_t *thread, int is_root) {
5908   // It is assumed __kmp_forkjoin_lock is acquired.
5909 
5910   int gtid;
5911 
5912   KMP_DEBUG_ASSERT(thread != NULL);
5913 
5914   gtid = thread->th.th_info.ds.ds_gtid;
5915 
5916   if (!is_root) {
5917     if (__kmp_dflt_blocktime != KMP_MAX_BLOCKTIME) {
5918       /* Assume the threads are at the fork barrier here */
5919       KA_TRACE(
5920           20, ("__kmp_reap_thread: releasing T#%d from fork barrier for reap\n",
5921                gtid));
5922       /* Need release fence here to prevent seg faults for tree forkjoin barrier
5923        * (GEH) */
5924       ANNOTATE_HAPPENS_BEFORE(thread);
5925       kmp_flag_64<> flag(&thread->th.th_bar[bs_forkjoin_barrier].bb.b_go,
5926                          thread);
5927       __kmp_release_64(&flag);
5928     }
5929 
5930     // Terminate OS thread.
5931     __kmp_reap_worker(thread);
5932 
5933     // The thread was killed asynchronously.  If it was actively
5934     // spinning in the thread pool, decrement the global count.
5935     //
5936     // There is a small timing hole here - if the worker thread was just waking
5937     // up after sleeping in the pool, had reset it's th_active_in_pool flag but
5938     // not decremented the global counter __kmp_thread_pool_active_nth yet, then
5939     // the global counter might not get updated.
5940     //
5941     // Currently, this can only happen as the library is unloaded,
5942     // so there are no harmful side effects.
5943     if (thread->th.th_active_in_pool) {
5944       thread->th.th_active_in_pool = FALSE;
5945       KMP_ATOMIC_DEC(&__kmp_thread_pool_active_nth);
5946       KMP_DEBUG_ASSERT(__kmp_thread_pool_active_nth >= 0);
5947     }
5948   }
5949 
5950   __kmp_free_implicit_task(thread);
5951 
5952 // Free the fast memory for tasking
5953 #if USE_FAST_MEMORY
5954   __kmp_free_fast_memory(thread);
5955 #endif /* USE_FAST_MEMORY */
5956 
5957   __kmp_suspend_uninitialize_thread(thread);
5958 
5959   KMP_DEBUG_ASSERT(__kmp_threads[gtid] == thread);
5960   TCW_SYNC_PTR(__kmp_threads[gtid], NULL);
5961 
5962   --__kmp_all_nth;
5963   // __kmp_nth was decremented when thread is added to the pool.
5964 
5965 #ifdef KMP_ADJUST_BLOCKTIME
5966   /* Adjust blocktime back to user setting or default if necessary */
5967   /* Middle initialization might never have occurred                */
5968   if (!__kmp_env_blocktime && (__kmp_avail_proc > 0)) {
5969     KMP_DEBUG_ASSERT(__kmp_avail_proc > 0);
5970     if (__kmp_nth <= __kmp_avail_proc) {
5971       __kmp_zero_bt = FALSE;
5972     }
5973   }
5974 #endif /* KMP_ADJUST_BLOCKTIME */
5975 
5976   /* free the memory being used */
5977   if (__kmp_env_consistency_check) {
5978     if (thread->th.th_cons) {
5979       __kmp_free_cons_stack(thread->th.th_cons);
5980       thread->th.th_cons = NULL;
5981     }
5982   }
5983 
5984   if (thread->th.th_pri_common != NULL) {
5985     __kmp_free(thread->th.th_pri_common);
5986     thread->th.th_pri_common = NULL;
5987   }
5988 
5989   if (thread->th.th_task_state_memo_stack != NULL) {
5990     __kmp_free(thread->th.th_task_state_memo_stack);
5991     thread->th.th_task_state_memo_stack = NULL;
5992   }
5993 
5994 #if KMP_USE_BGET
5995   if (thread->th.th_local.bget_data != NULL) {
5996     __kmp_finalize_bget(thread);
5997   }
5998 #endif
5999 
6000 #if KMP_AFFINITY_SUPPORTED
6001   if (thread->th.th_affin_mask != NULL) {
6002     KMP_CPU_FREE(thread->th.th_affin_mask);
6003     thread->th.th_affin_mask = NULL;
6004   }
6005 #endif /* KMP_AFFINITY_SUPPORTED */
6006 
6007 #if KMP_USE_HIER_SCHED
6008   if (thread->th.th_hier_bar_data != NULL) {
6009     __kmp_free(thread->th.th_hier_bar_data);
6010     thread->th.th_hier_bar_data = NULL;
6011   }
6012 #endif
6013 
6014   __kmp_reap_team(thread->th.th_serial_team);
6015   thread->th.th_serial_team = NULL;
6016   __kmp_free(thread);
6017 
6018   KMP_MB();
6019 
6020 } // __kmp_reap_thread
6021 
6022 static void __kmp_internal_end(void) {
6023   int i;
6024 
6025   /* First, unregister the library */
6026   __kmp_unregister_library();
6027 
6028 #if KMP_OS_WINDOWS
6029   /* In Win static library, we can't tell when a root actually dies, so we
6030      reclaim the data structures for any root threads that have died but not
6031      unregistered themselves, in order to shut down cleanly.
6032      In Win dynamic library we also can't tell when a thread dies.  */
6033   __kmp_reclaim_dead_roots(); // AC: moved here to always clean resources of
6034 // dead roots
6035 #endif
6036 
6037   for (i = 0; i < __kmp_threads_capacity; i++)
6038     if (__kmp_root[i])
6039       if (__kmp_root[i]->r.r_active)
6040         break;
6041   KMP_MB(); /* Flush all pending memory write invalidates.  */
6042   TCW_SYNC_4(__kmp_global.g.g_done, TRUE);
6043 
6044   if (i < __kmp_threads_capacity) {
6045 #if KMP_USE_MONITOR
6046     // 2009-09-08 (lev): Other alive roots found. Why do we kill the monitor??
6047     KMP_MB(); /* Flush all pending memory write invalidates.  */
6048 
6049     // Need to check that monitor was initialized before reaping it. If we are
6050     // called form __kmp_atfork_child (which sets __kmp_init_parallel = 0), then
6051     // __kmp_monitor will appear to contain valid data, but it is only valid in
6052     // the parent process, not the child.
6053     // New behavior (201008): instead of keying off of the flag
6054     // __kmp_init_parallel, the monitor thread creation is keyed off
6055     // of the new flag __kmp_init_monitor.
6056     __kmp_acquire_bootstrap_lock(&__kmp_monitor_lock);
6057     if (TCR_4(__kmp_init_monitor)) {
6058       __kmp_reap_monitor(&__kmp_monitor);
6059       TCW_4(__kmp_init_monitor, 0);
6060     }
6061     __kmp_release_bootstrap_lock(&__kmp_monitor_lock);
6062     KA_TRACE(10, ("__kmp_internal_end: monitor reaped\n"));
6063 #endif // KMP_USE_MONITOR
6064   } else {
6065 /* TODO move this to cleanup code */
6066 #ifdef KMP_DEBUG
6067     /* make sure that everything has properly ended */
6068     for (i = 0; i < __kmp_threads_capacity; i++) {
6069       if (__kmp_root[i]) {
6070         //                    KMP_ASSERT( ! KMP_UBER_GTID( i ) );         // AC:
6071         //                    there can be uber threads alive here
6072         KMP_ASSERT(!__kmp_root[i]->r.r_active); // TODO: can they be active?
6073       }
6074     }
6075 #endif
6076 
6077     KMP_MB();
6078 
6079     // Reap the worker threads.
6080     // This is valid for now, but be careful if threads are reaped sooner.
6081     while (__kmp_thread_pool != NULL) { // Loop thru all the thread in the pool.
6082       // Get the next thread from the pool.
6083       kmp_info_t *thread = CCAST(kmp_info_t *, __kmp_thread_pool);
6084       __kmp_thread_pool = thread->th.th_next_pool;
6085       // Reap it.
6086       KMP_DEBUG_ASSERT(thread->th.th_reap_state == KMP_SAFE_TO_REAP);
6087       thread->th.th_next_pool = NULL;
6088       thread->th.th_in_pool = FALSE;
6089       __kmp_reap_thread(thread, 0);
6090     }
6091     __kmp_thread_pool_insert_pt = NULL;
6092 
6093     // Reap teams.
6094     while (__kmp_team_pool != NULL) { // Loop thru all the teams in the pool.
6095       // Get the next team from the pool.
6096       kmp_team_t *team = CCAST(kmp_team_t *, __kmp_team_pool);
6097       __kmp_team_pool = team->t.t_next_pool;
6098       // Reap it.
6099       team->t.t_next_pool = NULL;
6100       __kmp_reap_team(team);
6101     }
6102 
6103     __kmp_reap_task_teams();
6104 
6105 #if KMP_OS_UNIX
6106     // Threads that are not reaped should not access any resources since they
6107     // are going to be deallocated soon, so the shutdown sequence should wait
6108     // until all threads either exit the final spin-waiting loop or begin
6109     // sleeping after the given blocktime.
6110     for (i = 0; i < __kmp_threads_capacity; i++) {
6111       kmp_info_t *thr = __kmp_threads[i];
6112       while (thr && KMP_ATOMIC_LD_ACQ(&thr->th.th_blocking))
6113         KMP_CPU_PAUSE();
6114     }
6115 #endif
6116 
6117     for (i = 0; i < __kmp_threads_capacity; ++i) {
6118       // TBD: Add some checking...
6119       // Something like KMP_DEBUG_ASSERT( __kmp_thread[ i ] == NULL );
6120     }
6121 
6122     /* Make sure all threadprivate destructors get run by joining with all
6123        worker threads before resetting this flag */
6124     TCW_SYNC_4(__kmp_init_common, FALSE);
6125 
6126     KA_TRACE(10, ("__kmp_internal_end: all workers reaped\n"));
6127     KMP_MB();
6128 
6129 #if KMP_USE_MONITOR
6130     // See note above: One of the possible fixes for CQ138434 / CQ140126
6131     //
6132     // FIXME: push both code fragments down and CSE them?
6133     // push them into __kmp_cleanup() ?
6134     __kmp_acquire_bootstrap_lock(&__kmp_monitor_lock);
6135     if (TCR_4(__kmp_init_monitor)) {
6136       __kmp_reap_monitor(&__kmp_monitor);
6137       TCW_4(__kmp_init_monitor, 0);
6138     }
6139     __kmp_release_bootstrap_lock(&__kmp_monitor_lock);
6140     KA_TRACE(10, ("__kmp_internal_end: monitor reaped\n"));
6141 #endif
6142   } /* else !__kmp_global.t_active */
6143   TCW_4(__kmp_init_gtid, FALSE);
6144   KMP_MB(); /* Flush all pending memory write invalidates.  */
6145 
6146   __kmp_cleanup();
6147 #if OMPT_SUPPORT
6148   ompt_fini();
6149 #endif
6150 }
6151 
6152 void __kmp_internal_end_library(int gtid_req) {
6153   /* if we have already cleaned up, don't try again, it wouldn't be pretty */
6154   /* this shouldn't be a race condition because __kmp_internal_end() is the
6155      only place to clear __kmp_serial_init */
6156   /* we'll check this later too, after we get the lock */
6157   // 2009-09-06: We do not set g_abort without setting g_done. This check looks
6158   // redundant, because the next check will work in any case.
6159   if (__kmp_global.g.g_abort) {
6160     KA_TRACE(11, ("__kmp_internal_end_library: abort, exiting\n"));
6161     /* TODO abort? */
6162     return;
6163   }
6164   if (TCR_4(__kmp_global.g.g_done) || !__kmp_init_serial) {
6165     KA_TRACE(10, ("__kmp_internal_end_library: already finished\n"));
6166     return;
6167   }
6168 
6169   KMP_MB(); /* Flush all pending memory write invalidates.  */
6170   /* find out who we are and what we should do */
6171   {
6172     int gtid = (gtid_req >= 0) ? gtid_req : __kmp_gtid_get_specific();
6173     KA_TRACE(
6174         10, ("__kmp_internal_end_library: enter T#%d  (%d)\n", gtid, gtid_req));
6175     if (gtid == KMP_GTID_SHUTDOWN) {
6176       KA_TRACE(10, ("__kmp_internal_end_library: !__kmp_init_runtime, system "
6177                     "already shutdown\n"));
6178       return;
6179     } else if (gtid == KMP_GTID_MONITOR) {
6180       KA_TRACE(10, ("__kmp_internal_end_library: monitor thread, gtid not "
6181                     "registered, or system shutdown\n"));
6182       return;
6183     } else if (gtid == KMP_GTID_DNE) {
6184       KA_TRACE(10, ("__kmp_internal_end_library: gtid not registered or system "
6185                     "shutdown\n"));
6186       /* we don't know who we are, but we may still shutdown the library */
6187     } else if (KMP_UBER_GTID(gtid)) {
6188       /* unregister ourselves as an uber thread.  gtid is no longer valid */
6189       if (__kmp_root[gtid]->r.r_active) {
6190         __kmp_global.g.g_abort = -1;
6191         TCW_SYNC_4(__kmp_global.g.g_done, TRUE);
6192         __kmp_unregister_library();
6193         KA_TRACE(10,
6194                  ("__kmp_internal_end_library: root still active, abort T#%d\n",
6195                   gtid));
6196         return;
6197       } else {
6198         KA_TRACE(
6199             10,
6200             ("__kmp_internal_end_library: unregistering sibling T#%d\n", gtid));
6201         __kmp_unregister_root_current_thread(gtid);
6202       }
6203     } else {
6204 /* worker threads may call this function through the atexit handler, if they
6205  * call exit() */
6206 /* For now, skip the usual subsequent processing and just dump the debug buffer.
6207    TODO: do a thorough shutdown instead */
6208 #ifdef DUMP_DEBUG_ON_EXIT
6209       if (__kmp_debug_buf)
6210         __kmp_dump_debug_buffer();
6211 #endif
6212       // added unregister library call here when we switch to shm linux
6213       // if we don't, it will leave lots of files in /dev/shm
6214       // cleanup shared memory file before exiting.
6215       __kmp_unregister_library();
6216       return;
6217     }
6218   }
6219   /* synchronize the termination process */
6220   __kmp_acquire_bootstrap_lock(&__kmp_initz_lock);
6221 
6222   /* have we already finished */
6223   if (__kmp_global.g.g_abort) {
6224     KA_TRACE(10, ("__kmp_internal_end_library: abort, exiting\n"));
6225     /* TODO abort? */
6226     __kmp_release_bootstrap_lock(&__kmp_initz_lock);
6227     return;
6228   }
6229   if (TCR_4(__kmp_global.g.g_done) || !__kmp_init_serial) {
6230     __kmp_release_bootstrap_lock(&__kmp_initz_lock);
6231     return;
6232   }
6233 
6234   /* We need this lock to enforce mutex between this reading of
6235      __kmp_threads_capacity and the writing by __kmp_register_root.
6236      Alternatively, we can use a counter of roots that is atomically updated by
6237      __kmp_get_global_thread_id_reg, __kmp_do_serial_initialize and
6238      __kmp_internal_end_*.  */
6239   __kmp_acquire_bootstrap_lock(&__kmp_forkjoin_lock);
6240 
6241   /* now we can safely conduct the actual termination */
6242   __kmp_internal_end();
6243 
6244   __kmp_release_bootstrap_lock(&__kmp_forkjoin_lock);
6245   __kmp_release_bootstrap_lock(&__kmp_initz_lock);
6246 
6247   KA_TRACE(10, ("__kmp_internal_end_library: exit\n"));
6248 
6249 #ifdef DUMP_DEBUG_ON_EXIT
6250   if (__kmp_debug_buf)
6251     __kmp_dump_debug_buffer();
6252 #endif
6253 
6254 #if KMP_OS_WINDOWS
6255   __kmp_close_console();
6256 #endif
6257 
6258   __kmp_fini_allocator();
6259 
6260 } // __kmp_internal_end_library
6261 
6262 void __kmp_internal_end_thread(int gtid_req) {
6263   int i;
6264 
6265   /* if we have already cleaned up, don't try again, it wouldn't be pretty */
6266   /* this shouldn't be a race condition because __kmp_internal_end() is the
6267    * only place to clear __kmp_serial_init */
6268   /* we'll check this later too, after we get the lock */
6269   // 2009-09-06: We do not set g_abort without setting g_done. This check looks
6270   // redundant, because the next check will work in any case.
6271   if (__kmp_global.g.g_abort) {
6272     KA_TRACE(11, ("__kmp_internal_end_thread: abort, exiting\n"));
6273     /* TODO abort? */
6274     return;
6275   }
6276   if (TCR_4(__kmp_global.g.g_done) || !__kmp_init_serial) {
6277     KA_TRACE(10, ("__kmp_internal_end_thread: already finished\n"));
6278     return;
6279   }
6280 
6281   // If hidden helper team has been initialized, we need to deinit it
6282   if (TCR_4(__kmp_init_hidden_helper)) {
6283     TCW_SYNC_4(__kmp_hidden_helper_team_done, TRUE);
6284     // First release the main thread to let it continue its work
6285     __kmp_hidden_helper_main_thread_release();
6286     // Wait until the hidden helper team has been destroyed
6287     __kmp_hidden_helper_threads_deinitz_wait();
6288   }
6289 
6290   KMP_MB(); /* Flush all pending memory write invalidates.  */
6291 
6292   /* find out who we are and what we should do */
6293   {
6294     int gtid = (gtid_req >= 0) ? gtid_req : __kmp_gtid_get_specific();
6295     KA_TRACE(10,
6296              ("__kmp_internal_end_thread: enter T#%d  (%d)\n", gtid, gtid_req));
6297     if (gtid == KMP_GTID_SHUTDOWN) {
6298       KA_TRACE(10, ("__kmp_internal_end_thread: !__kmp_init_runtime, system "
6299                     "already shutdown\n"));
6300       return;
6301     } else if (gtid == KMP_GTID_MONITOR) {
6302       KA_TRACE(10, ("__kmp_internal_end_thread: monitor thread, gtid not "
6303                     "registered, or system shutdown\n"));
6304       return;
6305     } else if (gtid == KMP_GTID_DNE) {
6306       KA_TRACE(10, ("__kmp_internal_end_thread: gtid not registered or system "
6307                     "shutdown\n"));
6308       return;
6309       /* we don't know who we are */
6310     } else if (KMP_UBER_GTID(gtid)) {
6311       /* unregister ourselves as an uber thread.  gtid is no longer valid */
6312       if (__kmp_root[gtid]->r.r_active) {
6313         __kmp_global.g.g_abort = -1;
6314         TCW_SYNC_4(__kmp_global.g.g_done, TRUE);
6315         KA_TRACE(10,
6316                  ("__kmp_internal_end_thread: root still active, abort T#%d\n",
6317                   gtid));
6318         return;
6319       } else {
6320         KA_TRACE(10, ("__kmp_internal_end_thread: unregistering sibling T#%d\n",
6321                       gtid));
6322         __kmp_unregister_root_current_thread(gtid);
6323       }
6324     } else {
6325       /* just a worker thread, let's leave */
6326       KA_TRACE(10, ("__kmp_internal_end_thread: worker thread T#%d\n", gtid));
6327 
6328       if (gtid >= 0) {
6329         __kmp_threads[gtid]->th.th_task_team = NULL;
6330       }
6331 
6332       KA_TRACE(10,
6333                ("__kmp_internal_end_thread: worker thread done, exiting T#%d\n",
6334                 gtid));
6335       return;
6336     }
6337   }
6338 #if KMP_DYNAMIC_LIB
6339   if (__kmp_pause_status != kmp_hard_paused)
6340   // AC: lets not shutdown the dynamic library at the exit of uber thread,
6341   // because we will better shutdown later in the library destructor.
6342   {
6343     KA_TRACE(10, ("__kmp_internal_end_thread: exiting T#%d\n", gtid_req));
6344     return;
6345   }
6346 #endif
6347   /* synchronize the termination process */
6348   __kmp_acquire_bootstrap_lock(&__kmp_initz_lock);
6349 
6350   /* have we already finished */
6351   if (__kmp_global.g.g_abort) {
6352     KA_TRACE(10, ("__kmp_internal_end_thread: abort, exiting\n"));
6353     /* TODO abort? */
6354     __kmp_release_bootstrap_lock(&__kmp_initz_lock);
6355     return;
6356   }
6357   if (TCR_4(__kmp_global.g.g_done) || !__kmp_init_serial) {
6358     __kmp_release_bootstrap_lock(&__kmp_initz_lock);
6359     return;
6360   }
6361 
6362   /* We need this lock to enforce mutex between this reading of
6363      __kmp_threads_capacity and the writing by __kmp_register_root.
6364      Alternatively, we can use a counter of roots that is atomically updated by
6365      __kmp_get_global_thread_id_reg, __kmp_do_serial_initialize and
6366      __kmp_internal_end_*.  */
6367 
6368   /* should we finish the run-time?  are all siblings done? */
6369   __kmp_acquire_bootstrap_lock(&__kmp_forkjoin_lock);
6370 
6371   for (i = 0; i < __kmp_threads_capacity; ++i) {
6372     if (KMP_UBER_GTID(i)) {
6373       KA_TRACE(
6374           10,
6375           ("__kmp_internal_end_thread: remaining sibling task: gtid==%d\n", i));
6376       __kmp_release_bootstrap_lock(&__kmp_forkjoin_lock);
6377       __kmp_release_bootstrap_lock(&__kmp_initz_lock);
6378       return;
6379     }
6380   }
6381 
6382   /* now we can safely conduct the actual termination */
6383 
6384   __kmp_internal_end();
6385 
6386   __kmp_release_bootstrap_lock(&__kmp_forkjoin_lock);
6387   __kmp_release_bootstrap_lock(&__kmp_initz_lock);
6388 
6389   KA_TRACE(10, ("__kmp_internal_end_thread: exit T#%d\n", gtid_req));
6390 
6391 #ifdef DUMP_DEBUG_ON_EXIT
6392   if (__kmp_debug_buf)
6393     __kmp_dump_debug_buffer();
6394 #endif
6395 } // __kmp_internal_end_thread
6396 
6397 // -----------------------------------------------------------------------------
6398 // Library registration stuff.
6399 
6400 static long __kmp_registration_flag = 0;
6401 // Random value used to indicate library initialization.
6402 static char *__kmp_registration_str = NULL;
6403 // Value to be saved in env var __KMP_REGISTERED_LIB_<pid>.
6404 
6405 static inline char *__kmp_reg_status_name() {
6406 /* On RHEL 3u5 if linked statically, getpid() returns different values in
6407    each thread. If registration and unregistration go in different threads
6408    (omp_misc_other_root_exit.cpp test case), the name of registered_lib_env
6409    env var can not be found, because the name will contain different pid. */
6410 // macOS* complains about name being too long with additional getuid()
6411 #if KMP_OS_UNIX && !KMP_OS_DARWIN && KMP_DYNAMIC_LIB
6412   return __kmp_str_format("__KMP_REGISTERED_LIB_%d_%d", (int)getpid(),
6413                           (int)getuid());
6414 #else
6415   return __kmp_str_format("__KMP_REGISTERED_LIB_%d", (int)getpid());
6416 #endif
6417 } // __kmp_reg_status_get
6418 
6419 void __kmp_register_library_startup(void) {
6420 
6421   char *name = __kmp_reg_status_name(); // Name of the environment variable.
6422   int done = 0;
6423   union {
6424     double dtime;
6425     long ltime;
6426   } time;
6427 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
6428   __kmp_initialize_system_tick();
6429 #endif
6430   __kmp_read_system_time(&time.dtime);
6431   __kmp_registration_flag = 0xCAFE0000L | (time.ltime & 0x0000FFFFL);
6432   __kmp_registration_str =
6433       __kmp_str_format("%p-%lx-%s", &__kmp_registration_flag,
6434                        __kmp_registration_flag, KMP_LIBRARY_FILE);
6435 
6436   KA_TRACE(50, ("__kmp_register_library_startup: %s=\"%s\"\n", name,
6437                 __kmp_registration_str));
6438 
6439   while (!done) {
6440 
6441     char *value = NULL; // Actual value of the environment variable.
6442 
6443 #if KMP_OS_UNIX && KMP_DYNAMIC_LIB // shared memory is with dynamic library
6444     char *shm_name = __kmp_str_format("/%s", name);
6445     int shm_preexist = 0;
6446     char *data1;
6447     int fd1 = shm_open(shm_name, O_CREAT | O_EXCL | O_RDWR, 0666);
6448     if ((fd1 == -1) && (errno == EEXIST)) {
6449       // file didn't open because it already exists.
6450       // try opening existing file
6451       fd1 = shm_open(shm_name, O_RDWR, 0666);
6452       if (fd1 == -1) { // file didn't open
6453         // error out here
6454         __kmp_fatal(KMP_MSG(FunctionError, "Can't open SHM"), KMP_ERR(0),
6455                     __kmp_msg_null);
6456       } else {
6457         // able to open existing file
6458         shm_preexist = 1;
6459       }
6460     } else if (fd1 == -1) { // SHM didn't open; it was due to error other than
6461       // already exists.
6462       // error out here.
6463       __kmp_fatal(KMP_MSG(FunctionError, "Can't open SHM2"), KMP_ERR(errno),
6464                   __kmp_msg_null);
6465     }
6466     if (shm_preexist == 0) {
6467       // we created SHM now set size
6468       if (ftruncate(fd1, SHM_SIZE) == -1) {
6469         // error occured setting size;
6470         __kmp_fatal(KMP_MSG(FunctionError, "Can't set size of SHM"),
6471                     KMP_ERR(errno), __kmp_msg_null);
6472       }
6473     }
6474     data1 =
6475         (char *)mmap(0, SHM_SIZE, PROT_READ | PROT_WRITE, MAP_SHARED, fd1, 0);
6476     if (data1 == MAP_FAILED) {
6477       // failed to map shared memory
6478       __kmp_fatal(KMP_MSG(FunctionError, "Can't map SHM"), KMP_ERR(errno),
6479                   __kmp_msg_null);
6480     }
6481     if (shm_preexist == 0) { // set data to SHM, set value
6482       KMP_STRCPY_S(data1, SHM_SIZE, __kmp_registration_str);
6483     }
6484     // Read value from either what we just wrote or existing file.
6485     value = __kmp_str_format("%s", data1); // read value from SHM
6486     munmap(data1, SHM_SIZE);
6487     close(fd1);
6488 #else // Windows and unix with static library
6489     // Set environment variable, but do not overwrite if it is exist.
6490     __kmp_env_set(name, __kmp_registration_str, 0);
6491     // read value to see if it got set
6492     value = __kmp_env_get(name);
6493 #endif
6494 
6495     if (value != NULL && strcmp(value, __kmp_registration_str) == 0) {
6496       done = 1; // Ok, environment variable set successfully, exit the loop.
6497     } else {
6498       // Oops. Write failed. Another copy of OpenMP RTL is in memory.
6499       // Check whether it alive or dead.
6500       int neighbor = 0; // 0 -- unknown status, 1 -- alive, 2 -- dead.
6501       char *tail = value;
6502       char *flag_addr_str = NULL;
6503       char *flag_val_str = NULL;
6504       char const *file_name = NULL;
6505       __kmp_str_split(tail, '-', &flag_addr_str, &tail);
6506       __kmp_str_split(tail, '-', &flag_val_str, &tail);
6507       file_name = tail;
6508       if (tail != NULL) {
6509         long *flag_addr = 0;
6510         unsigned long flag_val = 0;
6511         KMP_SSCANF(flag_addr_str, "%p", RCAST(void **, &flag_addr));
6512         KMP_SSCANF(flag_val_str, "%lx", &flag_val);
6513         if (flag_addr != 0 && flag_val != 0 && strcmp(file_name, "") != 0) {
6514           // First, check whether environment-encoded address is mapped into
6515           // addr space.
6516           // If so, dereference it to see if it still has the right value.
6517           if (__kmp_is_address_mapped(flag_addr) && *flag_addr == flag_val) {
6518             neighbor = 1;
6519           } else {
6520             // If not, then we know the other copy of the library is no longer
6521             // running.
6522             neighbor = 2;
6523           }
6524         }
6525       }
6526       switch (neighbor) {
6527       case 0: // Cannot parse environment variable -- neighbor status unknown.
6528         // Assume it is the incompatible format of future version of the
6529         // library. Assume the other library is alive.
6530         // WARN( ... ); // TODO: Issue a warning.
6531         file_name = "unknown library";
6532         KMP_FALLTHROUGH();
6533       // Attention! Falling to the next case. That's intentional.
6534       case 1: { // Neighbor is alive.
6535         // Check it is allowed.
6536         char *duplicate_ok = __kmp_env_get("KMP_DUPLICATE_LIB_OK");
6537         if (!__kmp_str_match_true(duplicate_ok)) {
6538           // That's not allowed. Issue fatal error.
6539           __kmp_fatal(KMP_MSG(DuplicateLibrary, KMP_LIBRARY_FILE, file_name),
6540                       KMP_HNT(DuplicateLibrary), __kmp_msg_null);
6541         }
6542         KMP_INTERNAL_FREE(duplicate_ok);
6543         __kmp_duplicate_library_ok = 1;
6544         done = 1; // Exit the loop.
6545       } break;
6546       case 2: { // Neighbor is dead.
6547 
6548 #if KMP_OS_UNIX && KMP_DYNAMIC_LIB // shared memory is with dynamic library
6549         // close shared memory.
6550         shm_unlink(shm_name); // this removes file in /dev/shm
6551 #else
6552         // Clear the variable and try to register library again.
6553         __kmp_env_unset(name);
6554 #endif
6555       } break;
6556       default: {
6557         KMP_DEBUG_ASSERT(0);
6558       } break;
6559       }
6560     }
6561     KMP_INTERNAL_FREE((void *)value);
6562 #if KMP_OS_UNIX && KMP_DYNAMIC_LIB // shared memory is with dynamic library
6563     KMP_INTERNAL_FREE((void *)shm_name);
6564 #endif
6565   } // while
6566   KMP_INTERNAL_FREE((void *)name);
6567 
6568 } // func __kmp_register_library_startup
6569 
6570 void __kmp_unregister_library(void) {
6571 
6572   char *name = __kmp_reg_status_name();
6573   char *value = NULL;
6574 
6575 #if KMP_OS_UNIX && KMP_DYNAMIC_LIB // shared memory is with dynamic library
6576   char *shm_name = __kmp_str_format("/%s", name);
6577   int fd1 = shm_open(shm_name, O_RDONLY, 0666);
6578   if (fd1 == -1) {
6579     // file did not open. return.
6580     return;
6581   }
6582   char *data1 = (char *)mmap(0, SHM_SIZE, PROT_READ, MAP_SHARED, fd1, 0);
6583   if (data1 != MAP_FAILED) {
6584     value = __kmp_str_format("%s", data1); // read value from SHM
6585     munmap(data1, SHM_SIZE);
6586   }
6587   close(fd1);
6588 #else
6589   value = __kmp_env_get(name);
6590 #endif
6591 
6592   KMP_DEBUG_ASSERT(__kmp_registration_flag != 0);
6593   KMP_DEBUG_ASSERT(__kmp_registration_str != NULL);
6594   if (value != NULL && strcmp(value, __kmp_registration_str) == 0) {
6595 //  Ok, this is our variable. Delete it.
6596 #if KMP_OS_UNIX && KMP_DYNAMIC_LIB // shared memory is with dynamic library
6597     shm_unlink(shm_name); // this removes file in /dev/shm
6598 #else
6599     __kmp_env_unset(name);
6600 #endif
6601   }
6602 
6603 #if KMP_OS_UNIX && KMP_DYNAMIC_LIB // shared memory is with dynamic library
6604   KMP_INTERNAL_FREE(shm_name);
6605 #endif
6606 
6607   KMP_INTERNAL_FREE(__kmp_registration_str);
6608   KMP_INTERNAL_FREE(value);
6609   KMP_INTERNAL_FREE(name);
6610 
6611   __kmp_registration_flag = 0;
6612   __kmp_registration_str = NULL;
6613 
6614 } // __kmp_unregister_library
6615 
6616 // End of Library registration stuff.
6617 // -----------------------------------------------------------------------------
6618 
6619 #if KMP_MIC_SUPPORTED
6620 
6621 static void __kmp_check_mic_type() {
6622   kmp_cpuid_t cpuid_state = {0};
6623   kmp_cpuid_t *cs_p = &cpuid_state;
6624   __kmp_x86_cpuid(1, 0, cs_p);
6625   // We don't support mic1 at the moment
6626   if ((cs_p->eax & 0xff0) == 0xB10) {
6627     __kmp_mic_type = mic2;
6628   } else if ((cs_p->eax & 0xf0ff0) == 0x50670) {
6629     __kmp_mic_type = mic3;
6630   } else {
6631     __kmp_mic_type = non_mic;
6632   }
6633 }
6634 
6635 #endif /* KMP_MIC_SUPPORTED */
6636 
6637 #if KMP_HAVE_UMWAIT
6638 static void __kmp_user_level_mwait_init() {
6639   struct kmp_cpuid buf;
6640   __kmp_x86_cpuid(7, 0, &buf);
6641   __kmp_umwait_enabled = ((buf.ecx >> 5) & 1) && __kmp_user_level_mwait;
6642   KF_TRACE(30, ("__kmp_user_level_mwait_init: __kmp_umwait_enabled = %d\n",
6643                 __kmp_umwait_enabled));
6644 }
6645 #elif KMP_HAVE_MWAIT
6646 #ifndef AT_INTELPHIUSERMWAIT
6647 // Spurious, non-existent value that should always fail to return anything.
6648 // Will be replaced with the correct value when we know that.
6649 #define AT_INTELPHIUSERMWAIT 10000
6650 #endif
6651 // getauxval() function is available in RHEL7 and SLES12. If a system with an
6652 // earlier OS is used to build the RTL, we'll use the following internal
6653 // function when the entry is not found.
6654 unsigned long getauxval(unsigned long) KMP_WEAK_ATTRIBUTE_EXTERNAL;
6655 unsigned long getauxval(unsigned long) { return 0; }
6656 
6657 static void __kmp_user_level_mwait_init() {
6658   // When getauxval() and correct value of AT_INTELPHIUSERMWAIT are available
6659   // use them to find if the user-level mwait is enabled. Otherwise, forcibly
6660   // set __kmp_mwait_enabled=TRUE on Intel MIC if the environment variable
6661   // KMP_USER_LEVEL_MWAIT was set to TRUE.
6662   if (__kmp_mic_type == mic3) {
6663     unsigned long res = getauxval(AT_INTELPHIUSERMWAIT);
6664     if ((res & 0x1) || __kmp_user_level_mwait) {
6665       __kmp_mwait_enabled = TRUE;
6666       if (__kmp_user_level_mwait) {
6667         KMP_INFORM(EnvMwaitWarn);
6668       }
6669     } else {
6670       __kmp_mwait_enabled = FALSE;
6671     }
6672   }
6673   KF_TRACE(30, ("__kmp_user_level_mwait_init: __kmp_mic_type = %d, "
6674                 "__kmp_mwait_enabled = %d\n",
6675                 __kmp_mic_type, __kmp_mwait_enabled));
6676 }
6677 #endif /* KMP_HAVE_UMWAIT */
6678 
6679 static void __kmp_do_serial_initialize(void) {
6680   int i, gtid;
6681   size_t size;
6682 
6683   KA_TRACE(10, ("__kmp_do_serial_initialize: enter\n"));
6684 
6685   KMP_DEBUG_ASSERT(sizeof(kmp_int32) == 4);
6686   KMP_DEBUG_ASSERT(sizeof(kmp_uint32) == 4);
6687   KMP_DEBUG_ASSERT(sizeof(kmp_int64) == 8);
6688   KMP_DEBUG_ASSERT(sizeof(kmp_uint64) == 8);
6689   KMP_DEBUG_ASSERT(sizeof(kmp_intptr_t) == sizeof(void *));
6690 
6691 #if OMPT_SUPPORT
6692   ompt_pre_init();
6693 #endif
6694 
6695   __kmp_validate_locks();
6696 
6697   /* Initialize internal memory allocator */
6698   __kmp_init_allocator();
6699 
6700   /* Register the library startup via an environment variable and check to see
6701      whether another copy of the library is already registered. */
6702 
6703   __kmp_register_library_startup();
6704 
6705   /* TODO reinitialization of library */
6706   if (TCR_4(__kmp_global.g.g_done)) {
6707     KA_TRACE(10, ("__kmp_do_serial_initialize: reinitialization of library\n"));
6708   }
6709 
6710   __kmp_global.g.g_abort = 0;
6711   TCW_SYNC_4(__kmp_global.g.g_done, FALSE);
6712 
6713 /* initialize the locks */
6714 #if KMP_USE_ADAPTIVE_LOCKS
6715 #if KMP_DEBUG_ADAPTIVE_LOCKS
6716   __kmp_init_speculative_stats();
6717 #endif
6718 #endif
6719 #if KMP_STATS_ENABLED
6720   __kmp_stats_init();
6721 #endif
6722   __kmp_init_lock(&__kmp_global_lock);
6723   __kmp_init_queuing_lock(&__kmp_dispatch_lock);
6724   __kmp_init_lock(&__kmp_debug_lock);
6725   __kmp_init_atomic_lock(&__kmp_atomic_lock);
6726   __kmp_init_atomic_lock(&__kmp_atomic_lock_1i);
6727   __kmp_init_atomic_lock(&__kmp_atomic_lock_2i);
6728   __kmp_init_atomic_lock(&__kmp_atomic_lock_4i);
6729   __kmp_init_atomic_lock(&__kmp_atomic_lock_4r);
6730   __kmp_init_atomic_lock(&__kmp_atomic_lock_8i);
6731   __kmp_init_atomic_lock(&__kmp_atomic_lock_8r);
6732   __kmp_init_atomic_lock(&__kmp_atomic_lock_8c);
6733   __kmp_init_atomic_lock(&__kmp_atomic_lock_10r);
6734   __kmp_init_atomic_lock(&__kmp_atomic_lock_16r);
6735   __kmp_init_atomic_lock(&__kmp_atomic_lock_16c);
6736   __kmp_init_atomic_lock(&__kmp_atomic_lock_20c);
6737   __kmp_init_atomic_lock(&__kmp_atomic_lock_32c);
6738   __kmp_init_bootstrap_lock(&__kmp_forkjoin_lock);
6739   __kmp_init_bootstrap_lock(&__kmp_exit_lock);
6740 #if KMP_USE_MONITOR
6741   __kmp_init_bootstrap_lock(&__kmp_monitor_lock);
6742 #endif
6743   __kmp_init_bootstrap_lock(&__kmp_tp_cached_lock);
6744 
6745   /* conduct initialization and initial setup of configuration */
6746 
6747   __kmp_runtime_initialize();
6748 
6749 #if KMP_MIC_SUPPORTED
6750   __kmp_check_mic_type();
6751 #endif
6752 
6753 // Some global variable initialization moved here from kmp_env_initialize()
6754 #ifdef KMP_DEBUG
6755   kmp_diag = 0;
6756 #endif
6757   __kmp_abort_delay = 0;
6758 
6759   // From __kmp_init_dflt_team_nth()
6760   /* assume the entire machine will be used */
6761   __kmp_dflt_team_nth_ub = __kmp_xproc;
6762   if (__kmp_dflt_team_nth_ub < KMP_MIN_NTH) {
6763     __kmp_dflt_team_nth_ub = KMP_MIN_NTH;
6764   }
6765   if (__kmp_dflt_team_nth_ub > __kmp_sys_max_nth) {
6766     __kmp_dflt_team_nth_ub = __kmp_sys_max_nth;
6767   }
6768   __kmp_max_nth = __kmp_sys_max_nth;
6769   __kmp_cg_max_nth = __kmp_sys_max_nth;
6770   __kmp_teams_max_nth = __kmp_xproc; // set a "reasonable" default
6771   if (__kmp_teams_max_nth > __kmp_sys_max_nth) {
6772     __kmp_teams_max_nth = __kmp_sys_max_nth;
6773   }
6774 
6775   // Three vars below moved here from __kmp_env_initialize() "KMP_BLOCKTIME"
6776   // part
6777   __kmp_dflt_blocktime = KMP_DEFAULT_BLOCKTIME;
6778 #if KMP_USE_MONITOR
6779   __kmp_monitor_wakeups =
6780       KMP_WAKEUPS_FROM_BLOCKTIME(__kmp_dflt_blocktime, __kmp_monitor_wakeups);
6781   __kmp_bt_intervals =
6782       KMP_INTERVALS_FROM_BLOCKTIME(__kmp_dflt_blocktime, __kmp_monitor_wakeups);
6783 #endif
6784   // From "KMP_LIBRARY" part of __kmp_env_initialize()
6785   __kmp_library = library_throughput;
6786   // From KMP_SCHEDULE initialization
6787   __kmp_static = kmp_sch_static_balanced;
6788 // AC: do not use analytical here, because it is non-monotonous
6789 //__kmp_guided = kmp_sch_guided_iterative_chunked;
6790 //__kmp_auto = kmp_sch_guided_analytical_chunked; // AC: it is the default, no
6791 // need to repeat assignment
6792 // Barrier initialization. Moved here from __kmp_env_initialize() Barrier branch
6793 // bit control and barrier method control parts
6794 #if KMP_FAST_REDUCTION_BARRIER
6795 #define kmp_reduction_barrier_gather_bb ((int)1)
6796 #define kmp_reduction_barrier_release_bb ((int)1)
6797 #define kmp_reduction_barrier_gather_pat bp_hyper_bar
6798 #define kmp_reduction_barrier_release_pat bp_hyper_bar
6799 #endif // KMP_FAST_REDUCTION_BARRIER
6800   for (i = bs_plain_barrier; i < bs_last_barrier; i++) {
6801     __kmp_barrier_gather_branch_bits[i] = __kmp_barrier_gather_bb_dflt;
6802     __kmp_barrier_release_branch_bits[i] = __kmp_barrier_release_bb_dflt;
6803     __kmp_barrier_gather_pattern[i] = __kmp_barrier_gather_pat_dflt;
6804     __kmp_barrier_release_pattern[i] = __kmp_barrier_release_pat_dflt;
6805 #if KMP_FAST_REDUCTION_BARRIER
6806     if (i == bs_reduction_barrier) { // tested and confirmed on ALTIX only (
6807       // lin_64 ): hyper,1
6808       __kmp_barrier_gather_branch_bits[i] = kmp_reduction_barrier_gather_bb;
6809       __kmp_barrier_release_branch_bits[i] = kmp_reduction_barrier_release_bb;
6810       __kmp_barrier_gather_pattern[i] = kmp_reduction_barrier_gather_pat;
6811       __kmp_barrier_release_pattern[i] = kmp_reduction_barrier_release_pat;
6812     }
6813 #endif // KMP_FAST_REDUCTION_BARRIER
6814   }
6815 #if KMP_FAST_REDUCTION_BARRIER
6816 #undef kmp_reduction_barrier_release_pat
6817 #undef kmp_reduction_barrier_gather_pat
6818 #undef kmp_reduction_barrier_release_bb
6819 #undef kmp_reduction_barrier_gather_bb
6820 #endif // KMP_FAST_REDUCTION_BARRIER
6821 #if KMP_MIC_SUPPORTED
6822   if (__kmp_mic_type == mic2) { // KNC
6823     // AC: plane=3,2, forkjoin=2,1 are optimal for 240 threads on KNC
6824     __kmp_barrier_gather_branch_bits[bs_plain_barrier] = 3; // plain gather
6825     __kmp_barrier_release_branch_bits[bs_forkjoin_barrier] =
6826         1; // forkjoin release
6827     __kmp_barrier_gather_pattern[bs_forkjoin_barrier] = bp_hierarchical_bar;
6828     __kmp_barrier_release_pattern[bs_forkjoin_barrier] = bp_hierarchical_bar;
6829   }
6830 #if KMP_FAST_REDUCTION_BARRIER
6831   if (__kmp_mic_type == mic2) { // KNC
6832     __kmp_barrier_gather_pattern[bs_reduction_barrier] = bp_hierarchical_bar;
6833     __kmp_barrier_release_pattern[bs_reduction_barrier] = bp_hierarchical_bar;
6834   }
6835 #endif // KMP_FAST_REDUCTION_BARRIER
6836 #endif // KMP_MIC_SUPPORTED
6837 
6838 // From KMP_CHECKS initialization
6839 #ifdef KMP_DEBUG
6840   __kmp_env_checks = TRUE; /* development versions have the extra checks */
6841 #else
6842   __kmp_env_checks = FALSE; /* port versions do not have the extra checks */
6843 #endif
6844 
6845   // From "KMP_FOREIGN_THREADS_THREADPRIVATE" initialization
6846   __kmp_foreign_tp = TRUE;
6847 
6848   __kmp_global.g.g_dynamic = FALSE;
6849   __kmp_global.g.g_dynamic_mode = dynamic_default;
6850 
6851   __kmp_env_initialize(NULL);
6852 
6853 #if KMP_HAVE_MWAIT || KMP_HAVE_UMWAIT
6854   __kmp_user_level_mwait_init();
6855 #endif
6856 // Print all messages in message catalog for testing purposes.
6857 #ifdef KMP_DEBUG
6858   char const *val = __kmp_env_get("KMP_DUMP_CATALOG");
6859   if (__kmp_str_match_true(val)) {
6860     kmp_str_buf_t buffer;
6861     __kmp_str_buf_init(&buffer);
6862     __kmp_i18n_dump_catalog(&buffer);
6863     __kmp_printf("%s", buffer.str);
6864     __kmp_str_buf_free(&buffer);
6865   }
6866   __kmp_env_free(&val);
6867 #endif
6868 
6869   __kmp_threads_capacity =
6870       __kmp_initial_threads_capacity(__kmp_dflt_team_nth_ub);
6871   // Moved here from __kmp_env_initialize() "KMP_ALL_THREADPRIVATE" part
6872   __kmp_tp_capacity = __kmp_default_tp_capacity(
6873       __kmp_dflt_team_nth_ub, __kmp_max_nth, __kmp_allThreadsSpecified);
6874 
6875   // If the library is shut down properly, both pools must be NULL. Just in
6876   // case, set them to NULL -- some memory may leak, but subsequent code will
6877   // work even if pools are not freed.
6878   KMP_DEBUG_ASSERT(__kmp_thread_pool == NULL);
6879   KMP_DEBUG_ASSERT(__kmp_thread_pool_insert_pt == NULL);
6880   KMP_DEBUG_ASSERT(__kmp_team_pool == NULL);
6881   __kmp_thread_pool = NULL;
6882   __kmp_thread_pool_insert_pt = NULL;
6883   __kmp_team_pool = NULL;
6884 
6885   /* Allocate all of the variable sized records */
6886   /* NOTE: __kmp_threads_capacity entries are allocated, but the arrays are
6887    * expandable */
6888   /* Since allocation is cache-aligned, just add extra padding at the end */
6889   size =
6890       (sizeof(kmp_info_t *) + sizeof(kmp_root_t *)) * __kmp_threads_capacity +
6891       CACHE_LINE;
6892   __kmp_threads = (kmp_info_t **)__kmp_allocate(size);
6893   __kmp_root = (kmp_root_t **)((char *)__kmp_threads +
6894                                sizeof(kmp_info_t *) * __kmp_threads_capacity);
6895 
6896   /* init thread counts */
6897   KMP_DEBUG_ASSERT(__kmp_all_nth ==
6898                    0); // Asserts fail if the library is reinitializing and
6899   KMP_DEBUG_ASSERT(__kmp_nth == 0); // something was wrong in termination.
6900   __kmp_all_nth = 0;
6901   __kmp_nth = 0;
6902 
6903   /* setup the uber master thread and hierarchy */
6904   gtid = __kmp_register_root(TRUE);
6905   KA_TRACE(10, ("__kmp_do_serial_initialize  T#%d\n", gtid));
6906   KMP_ASSERT(KMP_UBER_GTID(gtid));
6907   KMP_ASSERT(KMP_INITIAL_GTID(gtid));
6908 
6909   KMP_MB(); /* Flush all pending memory write invalidates.  */
6910 
6911   __kmp_common_initialize();
6912 
6913 #if KMP_OS_UNIX
6914   /* invoke the child fork handler */
6915   __kmp_register_atfork();
6916 #endif
6917 
6918 #if !KMP_DYNAMIC_LIB
6919   {
6920     /* Invoke the exit handler when the program finishes, only for static
6921        library. For dynamic library, we already have _fini and DllMain. */
6922     int rc = atexit(__kmp_internal_end_atexit);
6923     if (rc != 0) {
6924       __kmp_fatal(KMP_MSG(FunctionError, "atexit()"), KMP_ERR(rc),
6925                   __kmp_msg_null);
6926     }
6927   }
6928 #endif
6929 
6930 #if KMP_HANDLE_SIGNALS
6931 #if KMP_OS_UNIX
6932   /* NOTE: make sure that this is called before the user installs their own
6933      signal handlers so that the user handlers are called first. this way they
6934      can return false, not call our handler, avoid terminating the library, and
6935      continue execution where they left off. */
6936   __kmp_install_signals(FALSE);
6937 #endif /* KMP_OS_UNIX */
6938 #if KMP_OS_WINDOWS
6939   __kmp_install_signals(TRUE);
6940 #endif /* KMP_OS_WINDOWS */
6941 #endif
6942 
6943   /* we have finished the serial initialization */
6944   __kmp_init_counter++;
6945 
6946   __kmp_init_serial = TRUE;
6947 
6948   if (__kmp_settings) {
6949     __kmp_env_print();
6950   }
6951 
6952   if (__kmp_display_env || __kmp_display_env_verbose) {
6953     __kmp_env_print_2();
6954   }
6955 
6956 #if OMPT_SUPPORT
6957   ompt_post_init();
6958 #endif
6959 
6960   KMP_MB();
6961 
6962   KA_TRACE(10, ("__kmp_do_serial_initialize: exit\n"));
6963 }
6964 
6965 void __kmp_serial_initialize(void) {
6966   if (__kmp_init_serial) {
6967     return;
6968   }
6969   __kmp_acquire_bootstrap_lock(&__kmp_initz_lock);
6970   if (__kmp_init_serial) {
6971     __kmp_release_bootstrap_lock(&__kmp_initz_lock);
6972     return;
6973   }
6974   __kmp_do_serial_initialize();
6975   __kmp_release_bootstrap_lock(&__kmp_initz_lock);
6976 }
6977 
6978 static void __kmp_do_middle_initialize(void) {
6979   int i, j;
6980   int prev_dflt_team_nth;
6981 
6982   if (!__kmp_init_serial) {
6983     __kmp_do_serial_initialize();
6984   }
6985 
6986   KA_TRACE(10, ("__kmp_middle_initialize: enter\n"));
6987 
6988   // Save the previous value for the __kmp_dflt_team_nth so that
6989   // we can avoid some reinitialization if it hasn't changed.
6990   prev_dflt_team_nth = __kmp_dflt_team_nth;
6991 
6992 #if KMP_AFFINITY_SUPPORTED
6993   // __kmp_affinity_initialize() will try to set __kmp_ncores to the
6994   // number of cores on the machine.
6995   __kmp_affinity_initialize();
6996 
6997   // Run through the __kmp_threads array and set the affinity mask
6998   // for each root thread that is currently registered with the RTL.
6999   for (i = 0; i < __kmp_threads_capacity; i++) {
7000     if (TCR_PTR(__kmp_threads[i]) != NULL) {
7001       __kmp_affinity_set_init_mask(i, TRUE);
7002     }
7003   }
7004 #endif /* KMP_AFFINITY_SUPPORTED */
7005 
7006   KMP_ASSERT(__kmp_xproc > 0);
7007   if (__kmp_avail_proc == 0) {
7008     __kmp_avail_proc = __kmp_xproc;
7009   }
7010 
7011   // If there were empty places in num_threads list (OMP_NUM_THREADS=,,2,3),
7012   // correct them now
7013   j = 0;
7014   while ((j < __kmp_nested_nth.used) && !__kmp_nested_nth.nth[j]) {
7015     __kmp_nested_nth.nth[j] = __kmp_dflt_team_nth = __kmp_dflt_team_nth_ub =
7016         __kmp_avail_proc;
7017     j++;
7018   }
7019 
7020   if (__kmp_dflt_team_nth == 0) {
7021 #ifdef KMP_DFLT_NTH_CORES
7022     // Default #threads = #cores
7023     __kmp_dflt_team_nth = __kmp_ncores;
7024     KA_TRACE(20, ("__kmp_middle_initialize: setting __kmp_dflt_team_nth = "
7025                   "__kmp_ncores (%d)\n",
7026                   __kmp_dflt_team_nth));
7027 #else
7028     // Default #threads = #available OS procs
7029     __kmp_dflt_team_nth = __kmp_avail_proc;
7030     KA_TRACE(20, ("__kmp_middle_initialize: setting __kmp_dflt_team_nth = "
7031                   "__kmp_avail_proc(%d)\n",
7032                   __kmp_dflt_team_nth));
7033 #endif /* KMP_DFLT_NTH_CORES */
7034   }
7035 
7036   if (__kmp_dflt_team_nth < KMP_MIN_NTH) {
7037     __kmp_dflt_team_nth = KMP_MIN_NTH;
7038   }
7039   if (__kmp_dflt_team_nth > __kmp_sys_max_nth) {
7040     __kmp_dflt_team_nth = __kmp_sys_max_nth;
7041   }
7042 
7043   // There's no harm in continuing if the following check fails,
7044   // but it indicates an error in the previous logic.
7045   KMP_DEBUG_ASSERT(__kmp_dflt_team_nth <= __kmp_dflt_team_nth_ub);
7046 
7047   if (__kmp_dflt_team_nth != prev_dflt_team_nth) {
7048     // Run through the __kmp_threads array and set the num threads icv for each
7049     // root thread that is currently registered with the RTL (which has not
7050     // already explicitly set its nthreads-var with a call to
7051     // omp_set_num_threads()).
7052     for (i = 0; i < __kmp_threads_capacity; i++) {
7053       kmp_info_t *thread = __kmp_threads[i];
7054       if (thread == NULL)
7055         continue;
7056       if (thread->th.th_current_task->td_icvs.nproc != 0)
7057         continue;
7058 
7059       set__nproc(__kmp_threads[i], __kmp_dflt_team_nth);
7060     }
7061   }
7062   KA_TRACE(
7063       20,
7064       ("__kmp_middle_initialize: final value for __kmp_dflt_team_nth = %d\n",
7065        __kmp_dflt_team_nth));
7066 
7067 #ifdef KMP_ADJUST_BLOCKTIME
7068   /* Adjust blocktime to zero if necessary  now that __kmp_avail_proc is set */
7069   if (!__kmp_env_blocktime && (__kmp_avail_proc > 0)) {
7070     KMP_DEBUG_ASSERT(__kmp_avail_proc > 0);
7071     if (__kmp_nth > __kmp_avail_proc) {
7072       __kmp_zero_bt = TRUE;
7073     }
7074   }
7075 #endif /* KMP_ADJUST_BLOCKTIME */
7076 
7077   /* we have finished middle initialization */
7078   TCW_SYNC_4(__kmp_init_middle, TRUE);
7079 
7080   KA_TRACE(10, ("__kmp_do_middle_initialize: exit\n"));
7081 }
7082 
7083 void __kmp_middle_initialize(void) {
7084   if (__kmp_init_middle) {
7085     return;
7086   }
7087   __kmp_acquire_bootstrap_lock(&__kmp_initz_lock);
7088   if (__kmp_init_middle) {
7089     __kmp_release_bootstrap_lock(&__kmp_initz_lock);
7090     return;
7091   }
7092   __kmp_do_middle_initialize();
7093   __kmp_release_bootstrap_lock(&__kmp_initz_lock);
7094 }
7095 
7096 void __kmp_parallel_initialize(void) {
7097   int gtid = __kmp_entry_gtid(); // this might be a new root
7098 
7099   /* synchronize parallel initialization (for sibling) */
7100   if (TCR_4(__kmp_init_parallel))
7101     return;
7102   __kmp_acquire_bootstrap_lock(&__kmp_initz_lock);
7103   if (TCR_4(__kmp_init_parallel)) {
7104     __kmp_release_bootstrap_lock(&__kmp_initz_lock);
7105     return;
7106   }
7107 
7108   /* TODO reinitialization after we have already shut down */
7109   if (TCR_4(__kmp_global.g.g_done)) {
7110     KA_TRACE(
7111         10,
7112         ("__kmp_parallel_initialize: attempt to init while shutting down\n"));
7113     __kmp_infinite_loop();
7114   }
7115 
7116   /* jc: The lock __kmp_initz_lock is already held, so calling
7117      __kmp_serial_initialize would cause a deadlock.  So we call
7118      __kmp_do_serial_initialize directly. */
7119   if (!__kmp_init_middle) {
7120     __kmp_do_middle_initialize();
7121   }
7122   __kmp_resume_if_hard_paused();
7123 
7124   /* begin initialization */
7125   KA_TRACE(10, ("__kmp_parallel_initialize: enter\n"));
7126   KMP_ASSERT(KMP_UBER_GTID(gtid));
7127 
7128 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
7129   // Save the FP control regs.
7130   // Worker threads will set theirs to these values at thread startup.
7131   __kmp_store_x87_fpu_control_word(&__kmp_init_x87_fpu_control_word);
7132   __kmp_store_mxcsr(&__kmp_init_mxcsr);
7133   __kmp_init_mxcsr &= KMP_X86_MXCSR_MASK;
7134 #endif /* KMP_ARCH_X86 || KMP_ARCH_X86_64 */
7135 
7136 #if KMP_OS_UNIX
7137 #if KMP_HANDLE_SIGNALS
7138   /*  must be after __kmp_serial_initialize  */
7139   __kmp_install_signals(TRUE);
7140 #endif
7141 #endif
7142 
7143   __kmp_suspend_initialize();
7144 
7145 #if defined(USE_LOAD_BALANCE)
7146   if (__kmp_global.g.g_dynamic_mode == dynamic_default) {
7147     __kmp_global.g.g_dynamic_mode = dynamic_load_balance;
7148   }
7149 #else
7150   if (__kmp_global.g.g_dynamic_mode == dynamic_default) {
7151     __kmp_global.g.g_dynamic_mode = dynamic_thread_limit;
7152   }
7153 #endif
7154 
7155   if (__kmp_version) {
7156     __kmp_print_version_2();
7157   }
7158 
7159   /* we have finished parallel initialization */
7160   TCW_SYNC_4(__kmp_init_parallel, TRUE);
7161 
7162   KMP_MB();
7163   KA_TRACE(10, ("__kmp_parallel_initialize: exit\n"));
7164 
7165   __kmp_release_bootstrap_lock(&__kmp_initz_lock);
7166 }
7167 
7168 void __kmp_hidden_helper_initialize() {
7169   if (TCR_4(__kmp_init_hidden_helper))
7170     return;
7171 
7172   // __kmp_parallel_initialize is required before we initialize hidden helper
7173   if (!TCR_4(__kmp_init_parallel))
7174     __kmp_parallel_initialize();
7175 
7176   // Double check. Note that this double check should not be placed before
7177   // __kmp_parallel_initialize as it will cause dead lock.
7178   __kmp_acquire_bootstrap_lock(&__kmp_initz_lock);
7179   if (TCR_4(__kmp_init_hidden_helper)) {
7180     __kmp_release_bootstrap_lock(&__kmp_initz_lock);
7181     return;
7182   }
7183 
7184   // Set the count of hidden helper tasks to be executed to zero
7185   KMP_ATOMIC_ST_REL(&__kmp_unexecuted_hidden_helper_tasks, 0);
7186 
7187   // Set the global variable indicating that we're initializing hidden helper
7188   // team/threads
7189   TCW_SYNC_4(__kmp_init_hidden_helper_threads, TRUE);
7190 
7191   // Platform independent initialization
7192   __kmp_do_initialize_hidden_helper_threads();
7193 
7194   // Wait here for the finish of initialization of hidden helper teams
7195   __kmp_hidden_helper_threads_initz_wait();
7196 
7197   // We have finished hidden helper initialization
7198   TCW_SYNC_4(__kmp_init_hidden_helper, TRUE);
7199 
7200   __kmp_release_bootstrap_lock(&__kmp_initz_lock);
7201 }
7202 
7203 /* ------------------------------------------------------------------------ */
7204 
7205 void __kmp_run_before_invoked_task(int gtid, int tid, kmp_info_t *this_thr,
7206                                    kmp_team_t *team) {
7207   kmp_disp_t *dispatch;
7208 
7209   KMP_MB();
7210 
7211   /* none of the threads have encountered any constructs, yet. */
7212   this_thr->th.th_local.this_construct = 0;
7213 #if KMP_CACHE_MANAGE
7214   KMP_CACHE_PREFETCH(&this_thr->th.th_bar[bs_forkjoin_barrier].bb.b_arrived);
7215 #endif /* KMP_CACHE_MANAGE */
7216   dispatch = (kmp_disp_t *)TCR_PTR(this_thr->th.th_dispatch);
7217   KMP_DEBUG_ASSERT(dispatch);
7218   KMP_DEBUG_ASSERT(team->t.t_dispatch);
7219   // KMP_DEBUG_ASSERT( this_thr->th.th_dispatch == &team->t.t_dispatch[
7220   // this_thr->th.th_info.ds.ds_tid ] );
7221 
7222   dispatch->th_disp_index = 0; /* reset the dispatch buffer counter */
7223   dispatch->th_doacross_buf_idx = 0; // reset doacross dispatch buffer counter
7224   if (__kmp_env_consistency_check)
7225     __kmp_push_parallel(gtid, team->t.t_ident);
7226 
7227   KMP_MB(); /* Flush all pending memory write invalidates.  */
7228 }
7229 
7230 void __kmp_run_after_invoked_task(int gtid, int tid, kmp_info_t *this_thr,
7231                                   kmp_team_t *team) {
7232   if (__kmp_env_consistency_check)
7233     __kmp_pop_parallel(gtid, team->t.t_ident);
7234 
7235   __kmp_finish_implicit_task(this_thr);
7236 }
7237 
7238 int __kmp_invoke_task_func(int gtid) {
7239   int rc;
7240   int tid = __kmp_tid_from_gtid(gtid);
7241   kmp_info_t *this_thr = __kmp_threads[gtid];
7242   kmp_team_t *team = this_thr->th.th_team;
7243 
7244   __kmp_run_before_invoked_task(gtid, tid, this_thr, team);
7245 #if USE_ITT_BUILD
7246   if (__itt_stack_caller_create_ptr) {
7247     // inform ittnotify about entering user's code
7248     if (team->t.t_stack_id != NULL) {
7249       __kmp_itt_stack_callee_enter((__itt_caller)team->t.t_stack_id);
7250     } else {
7251       KMP_DEBUG_ASSERT(team->t.t_parent->t.t_stack_id != NULL);
7252       __kmp_itt_stack_callee_enter(
7253           (__itt_caller)team->t.t_parent->t.t_stack_id);
7254     }
7255   }
7256 #endif /* USE_ITT_BUILD */
7257 #if INCLUDE_SSC_MARKS
7258   SSC_MARK_INVOKING();
7259 #endif
7260 
7261 #if OMPT_SUPPORT
7262   void *dummy;
7263   void **exit_frame_p;
7264   ompt_data_t *my_task_data;
7265   ompt_data_t *my_parallel_data;
7266   int ompt_team_size;
7267 
7268   if (ompt_enabled.enabled) {
7269     exit_frame_p = &(team->t.t_implicit_task_taskdata[tid]
7270                          .ompt_task_info.frame.exit_frame.ptr);
7271   } else {
7272     exit_frame_p = &dummy;
7273   }
7274 
7275   my_task_data =
7276       &(team->t.t_implicit_task_taskdata[tid].ompt_task_info.task_data);
7277   my_parallel_data = &(team->t.ompt_team_info.parallel_data);
7278   if (ompt_enabled.ompt_callback_implicit_task) {
7279     ompt_team_size = team->t.t_nproc;
7280     ompt_callbacks.ompt_callback(ompt_callback_implicit_task)(
7281         ompt_scope_begin, my_parallel_data, my_task_data, ompt_team_size,
7282         __kmp_tid_from_gtid(gtid), ompt_task_implicit);
7283     OMPT_CUR_TASK_INFO(this_thr)->thread_num = __kmp_tid_from_gtid(gtid);
7284   }
7285 #endif
7286 
7287 #if KMP_STATS_ENABLED
7288   stats_state_e previous_state = KMP_GET_THREAD_STATE();
7289   if (previous_state == stats_state_e::TEAMS_REGION) {
7290     KMP_PUSH_PARTITIONED_TIMER(OMP_teams);
7291   } else {
7292     KMP_PUSH_PARTITIONED_TIMER(OMP_parallel);
7293   }
7294   KMP_SET_THREAD_STATE(IMPLICIT_TASK);
7295 #endif
7296 
7297   rc = __kmp_invoke_microtask((microtask_t)TCR_SYNC_PTR(team->t.t_pkfn), gtid,
7298                               tid, (int)team->t.t_argc, (void **)team->t.t_argv
7299 #if OMPT_SUPPORT
7300                               ,
7301                               exit_frame_p
7302 #endif
7303   );
7304 #if OMPT_SUPPORT
7305   *exit_frame_p = NULL;
7306   this_thr->th.ompt_thread_info.parallel_flags |= ompt_parallel_team;
7307 #endif
7308 
7309 #if KMP_STATS_ENABLED
7310   if (previous_state == stats_state_e::TEAMS_REGION) {
7311     KMP_SET_THREAD_STATE(previous_state);
7312   }
7313   KMP_POP_PARTITIONED_TIMER();
7314 #endif
7315 
7316 #if USE_ITT_BUILD
7317   if (__itt_stack_caller_create_ptr) {
7318     // inform ittnotify about leaving user's code
7319     if (team->t.t_stack_id != NULL) {
7320       __kmp_itt_stack_callee_leave((__itt_caller)team->t.t_stack_id);
7321     } else {
7322       KMP_DEBUG_ASSERT(team->t.t_parent->t.t_stack_id != NULL);
7323       __kmp_itt_stack_callee_leave(
7324           (__itt_caller)team->t.t_parent->t.t_stack_id);
7325     }
7326   }
7327 #endif /* USE_ITT_BUILD */
7328   __kmp_run_after_invoked_task(gtid, tid, this_thr, team);
7329 
7330   return rc;
7331 }
7332 
7333 void __kmp_teams_master(int gtid) {
7334   // This routine is called by all primary threads in teams construct
7335   kmp_info_t *thr = __kmp_threads[gtid];
7336   kmp_team_t *team = thr->th.th_team;
7337   ident_t *loc = team->t.t_ident;
7338   thr->th.th_set_nproc = thr->th.th_teams_size.nth;
7339   KMP_DEBUG_ASSERT(thr->th.th_teams_microtask);
7340   KMP_DEBUG_ASSERT(thr->th.th_set_nproc);
7341   KA_TRACE(20, ("__kmp_teams_master: T#%d, Tid %d, microtask %p\n", gtid,
7342                 __kmp_tid_from_gtid(gtid), thr->th.th_teams_microtask));
7343 
7344   // This thread is a new CG root.  Set up the proper variables.
7345   kmp_cg_root_t *tmp = (kmp_cg_root_t *)__kmp_allocate(sizeof(kmp_cg_root_t));
7346   tmp->cg_root = thr; // Make thr the CG root
7347   // Init to thread limit stored when league primary threads were forked
7348   tmp->cg_thread_limit = thr->th.th_current_task->td_icvs.thread_limit;
7349   tmp->cg_nthreads = 1; // Init counter to one active thread, this one
7350   KA_TRACE(100, ("__kmp_teams_master: Thread %p created node %p and init"
7351                  " cg_nthreads to 1\n",
7352                  thr, tmp));
7353   tmp->up = thr->th.th_cg_roots;
7354   thr->th.th_cg_roots = tmp;
7355 
7356 // Launch league of teams now, but not let workers execute
7357 // (they hang on fork barrier until next parallel)
7358 #if INCLUDE_SSC_MARKS
7359   SSC_MARK_FORKING();
7360 #endif
7361   __kmp_fork_call(loc, gtid, fork_context_intel, team->t.t_argc,
7362                   (microtask_t)thr->th.th_teams_microtask, // "wrapped" task
7363                   VOLATILE_CAST(launch_t) __kmp_invoke_task_func, NULL);
7364 #if INCLUDE_SSC_MARKS
7365   SSC_MARK_JOINING();
7366 #endif
7367   // If the team size was reduced from the limit, set it to the new size
7368   if (thr->th.th_team_nproc < thr->th.th_teams_size.nth)
7369     thr->th.th_teams_size.nth = thr->th.th_team_nproc;
7370   // AC: last parameter "1" eliminates join barrier which won't work because
7371   // worker threads are in a fork barrier waiting for more parallel regions
7372   __kmp_join_call(loc, gtid
7373 #if OMPT_SUPPORT
7374                   ,
7375                   fork_context_intel
7376 #endif
7377                   ,
7378                   1);
7379 }
7380 
7381 int __kmp_invoke_teams_master(int gtid) {
7382   kmp_info_t *this_thr = __kmp_threads[gtid];
7383   kmp_team_t *team = this_thr->th.th_team;
7384 #if KMP_DEBUG
7385   if (!__kmp_threads[gtid]->th.th_team->t.t_serialized)
7386     KMP_DEBUG_ASSERT((void *)__kmp_threads[gtid]->th.th_team->t.t_pkfn ==
7387                      (void *)__kmp_teams_master);
7388 #endif
7389   __kmp_run_before_invoked_task(gtid, 0, this_thr, team);
7390 #if OMPT_SUPPORT
7391   int tid = __kmp_tid_from_gtid(gtid);
7392   ompt_data_t *task_data =
7393       &team->t.t_implicit_task_taskdata[tid].ompt_task_info.task_data;
7394   ompt_data_t *parallel_data = &team->t.ompt_team_info.parallel_data;
7395   if (ompt_enabled.ompt_callback_implicit_task) {
7396     ompt_callbacks.ompt_callback(ompt_callback_implicit_task)(
7397         ompt_scope_begin, parallel_data, task_data, team->t.t_nproc, tid,
7398         ompt_task_initial);
7399     OMPT_CUR_TASK_INFO(this_thr)->thread_num = tid;
7400   }
7401 #endif
7402   __kmp_teams_master(gtid);
7403 #if OMPT_SUPPORT
7404   this_thr->th.ompt_thread_info.parallel_flags |= ompt_parallel_league;
7405 #endif
7406   __kmp_run_after_invoked_task(gtid, 0, this_thr, team);
7407   return 1;
7408 }
7409 
7410 /* this sets the requested number of threads for the next parallel region
7411    encountered by this team. since this should be enclosed in the forkjoin
7412    critical section it should avoid race conditions with asymmetrical nested
7413    parallelism */
7414 
7415 void __kmp_push_num_threads(ident_t *id, int gtid, int num_threads) {
7416   kmp_info_t *thr = __kmp_threads[gtid];
7417 
7418   if (num_threads > 0)
7419     thr->th.th_set_nproc = num_threads;
7420 }
7421 
7422 static void __kmp_push_thread_limit(kmp_info_t *thr, int num_teams,
7423                                     int num_threads) {
7424   KMP_DEBUG_ASSERT(thr);
7425   // Remember the number of threads for inner parallel regions
7426   if (!TCR_4(__kmp_init_middle))
7427     __kmp_middle_initialize(); // get internal globals calculated
7428   KMP_DEBUG_ASSERT(__kmp_avail_proc);
7429   KMP_DEBUG_ASSERT(__kmp_dflt_team_nth);
7430 
7431   if (num_threads == 0) {
7432     if (__kmp_teams_thread_limit > 0) {
7433       num_threads = __kmp_teams_thread_limit;
7434     } else {
7435       num_threads = __kmp_avail_proc / num_teams;
7436     }
7437     // adjust num_threads w/o warning as it is not user setting
7438     // num_threads = min(num_threads, nthreads-var, thread-limit-var)
7439     // no thread_limit clause specified -  do not change thread-limit-var ICV
7440     if (num_threads > __kmp_dflt_team_nth) {
7441       num_threads = __kmp_dflt_team_nth; // honor nthreads-var ICV
7442     }
7443     if (num_threads > thr->th.th_current_task->td_icvs.thread_limit) {
7444       num_threads = thr->th.th_current_task->td_icvs.thread_limit;
7445     } // prevent team size to exceed thread-limit-var
7446     if (num_teams * num_threads > __kmp_teams_max_nth) {
7447       num_threads = __kmp_teams_max_nth / num_teams;
7448     }
7449     if (num_threads == 0) {
7450       num_threads = 1;
7451     }
7452   } else {
7453     // This thread will be the primary thread of the league primary threads
7454     // Store new thread limit; old limit is saved in th_cg_roots list
7455     thr->th.th_current_task->td_icvs.thread_limit = num_threads;
7456     // num_threads = min(num_threads, nthreads-var)
7457     if (num_threads > __kmp_dflt_team_nth) {
7458       num_threads = __kmp_dflt_team_nth; // honor nthreads-var ICV
7459     }
7460     if (num_teams * num_threads > __kmp_teams_max_nth) {
7461       int new_threads = __kmp_teams_max_nth / num_teams;
7462       if (new_threads == 0) {
7463         new_threads = 1;
7464       }
7465       if (new_threads != num_threads) {
7466         if (!__kmp_reserve_warn) { // user asked for too many threads
7467           __kmp_reserve_warn = 1; // conflicts with KMP_TEAMS_THREAD_LIMIT
7468           __kmp_msg(kmp_ms_warning,
7469                     KMP_MSG(CantFormThrTeam, num_threads, new_threads),
7470                     KMP_HNT(Unset_ALL_THREADS), __kmp_msg_null);
7471         }
7472       }
7473       num_threads = new_threads;
7474     }
7475   }
7476   thr->th.th_teams_size.nth = num_threads;
7477 }
7478 
7479 /* this sets the requested number of teams for the teams region and/or
7480    the number of threads for the next parallel region encountered  */
7481 void __kmp_push_num_teams(ident_t *id, int gtid, int num_teams,
7482                           int num_threads) {
7483   kmp_info_t *thr = __kmp_threads[gtid];
7484   KMP_DEBUG_ASSERT(num_teams >= 0);
7485   KMP_DEBUG_ASSERT(num_threads >= 0);
7486 
7487   if (num_teams == 0) {
7488     if (__kmp_nteams > 0) {
7489       num_teams = __kmp_nteams;
7490     } else {
7491       num_teams = 1; // default number of teams is 1.
7492     }
7493   }
7494   if (num_teams > __kmp_teams_max_nth) { // if too many teams requested?
7495     if (!__kmp_reserve_warn) {
7496       __kmp_reserve_warn = 1;
7497       __kmp_msg(kmp_ms_warning,
7498                 KMP_MSG(CantFormThrTeam, num_teams, __kmp_teams_max_nth),
7499                 KMP_HNT(Unset_ALL_THREADS), __kmp_msg_null);
7500     }
7501     num_teams = __kmp_teams_max_nth;
7502   }
7503   // Set number of teams (number of threads in the outer "parallel" of the
7504   // teams)
7505   thr->th.th_set_nproc = thr->th.th_teams_size.nteams = num_teams;
7506 
7507   __kmp_push_thread_limit(thr, num_teams, num_threads);
7508 }
7509 
7510 /* This sets the requested number of teams for the teams region and/or
7511    the number of threads for the next parallel region encountered  */
7512 void __kmp_push_num_teams_51(ident_t *id, int gtid, int num_teams_lb,
7513                              int num_teams_ub, int num_threads) {
7514   kmp_info_t *thr = __kmp_threads[gtid];
7515   KMP_DEBUG_ASSERT(num_teams_lb >= 0 && num_teams_ub >= 0);
7516   KMP_DEBUG_ASSERT(num_teams_ub >= num_teams_lb);
7517   KMP_DEBUG_ASSERT(num_threads >= 0);
7518 
7519   if (num_teams_lb > num_teams_ub) {
7520     __kmp_fatal(KMP_MSG(FailedToCreateTeam, num_teams_lb, num_teams_ub),
7521                 KMP_HNT(SetNewBound, __kmp_teams_max_nth), __kmp_msg_null);
7522   }
7523 
7524   int num_teams = 1; // defalt number of teams is 1.
7525 
7526   if (num_teams_lb == 0 && num_teams_ub > 0)
7527     num_teams_lb = num_teams_ub;
7528 
7529   if (num_teams_lb == 0 && num_teams_ub == 0) { // no num_teams clause
7530     num_teams = (__kmp_nteams > 0) ? __kmp_nteams : num_teams;
7531     if (num_teams > __kmp_teams_max_nth) {
7532       if (!__kmp_reserve_warn) {
7533         __kmp_reserve_warn = 1;
7534         __kmp_msg(kmp_ms_warning,
7535                   KMP_MSG(CantFormThrTeam, num_teams, __kmp_teams_max_nth),
7536                   KMP_HNT(Unset_ALL_THREADS), __kmp_msg_null);
7537       }
7538       num_teams = __kmp_teams_max_nth;
7539     }
7540   } else if (num_teams_lb == num_teams_ub) { // requires exact number of teams
7541     num_teams = num_teams_ub;
7542   } else { // num_teams_lb <= num_teams <= num_teams_ub
7543     if (num_threads == 0) {
7544       if (num_teams_ub > __kmp_teams_max_nth) {
7545         num_teams = num_teams_lb;
7546       } else {
7547         num_teams = num_teams_ub;
7548       }
7549     } else {
7550       num_teams = (num_threads > __kmp_teams_max_nth)
7551                       ? num_teams
7552                       : __kmp_teams_max_nth / num_threads;
7553       if (num_teams < num_teams_lb) {
7554         num_teams = num_teams_lb;
7555       } else if (num_teams > num_teams_ub) {
7556         num_teams = num_teams_ub;
7557       }
7558     }
7559   }
7560   // Set number of teams (number of threads in the outer "parallel" of the
7561   // teams)
7562   thr->th.th_set_nproc = thr->th.th_teams_size.nteams = num_teams;
7563 
7564   __kmp_push_thread_limit(thr, num_teams, num_threads);
7565 }
7566 
7567 // Set the proc_bind var to use in the following parallel region.
7568 void __kmp_push_proc_bind(ident_t *id, int gtid, kmp_proc_bind_t proc_bind) {
7569   kmp_info_t *thr = __kmp_threads[gtid];
7570   thr->th.th_set_proc_bind = proc_bind;
7571 }
7572 
7573 /* Launch the worker threads into the microtask. */
7574 
7575 void __kmp_internal_fork(ident_t *id, int gtid, kmp_team_t *team) {
7576   kmp_info_t *this_thr = __kmp_threads[gtid];
7577 
7578 #ifdef KMP_DEBUG
7579   int f;
7580 #endif /* KMP_DEBUG */
7581 
7582   KMP_DEBUG_ASSERT(team);
7583   KMP_DEBUG_ASSERT(this_thr->th.th_team == team);
7584   KMP_ASSERT(KMP_MASTER_GTID(gtid));
7585   KMP_MB(); /* Flush all pending memory write invalidates.  */
7586 
7587   team->t.t_construct = 0; /* no single directives seen yet */
7588   team->t.t_ordered.dt.t_value =
7589       0; /* thread 0 enters the ordered section first */
7590 
7591   /* Reset the identifiers on the dispatch buffer */
7592   KMP_DEBUG_ASSERT(team->t.t_disp_buffer);
7593   if (team->t.t_max_nproc > 1) {
7594     int i;
7595     for (i = 0; i < __kmp_dispatch_num_buffers; ++i) {
7596       team->t.t_disp_buffer[i].buffer_index = i;
7597       team->t.t_disp_buffer[i].doacross_buf_idx = i;
7598     }
7599   } else {
7600     team->t.t_disp_buffer[0].buffer_index = 0;
7601     team->t.t_disp_buffer[0].doacross_buf_idx = 0;
7602   }
7603 
7604   KMP_MB(); /* Flush all pending memory write invalidates.  */
7605   KMP_ASSERT(this_thr->th.th_team == team);
7606 
7607 #ifdef KMP_DEBUG
7608   for (f = 0; f < team->t.t_nproc; f++) {
7609     KMP_DEBUG_ASSERT(team->t.t_threads[f] &&
7610                      team->t.t_threads[f]->th.th_team_nproc == team->t.t_nproc);
7611   }
7612 #endif /* KMP_DEBUG */
7613 
7614   /* release the worker threads so they may begin working */
7615   __kmp_fork_barrier(gtid, 0);
7616 }
7617 
7618 void __kmp_internal_join(ident_t *id, int gtid, kmp_team_t *team) {
7619   kmp_info_t *this_thr = __kmp_threads[gtid];
7620 
7621   KMP_DEBUG_ASSERT(team);
7622   KMP_DEBUG_ASSERT(this_thr->th.th_team == team);
7623   KMP_ASSERT(KMP_MASTER_GTID(gtid));
7624   KMP_MB(); /* Flush all pending memory write invalidates.  */
7625 
7626   /* Join barrier after fork */
7627 
7628 #ifdef KMP_DEBUG
7629   if (__kmp_threads[gtid] &&
7630       __kmp_threads[gtid]->th.th_team_nproc != team->t.t_nproc) {
7631     __kmp_printf("GTID: %d, __kmp_threads[%d]=%p\n", gtid, gtid,
7632                  __kmp_threads[gtid]);
7633     __kmp_printf("__kmp_threads[%d]->th.th_team_nproc=%d, TEAM: %p, "
7634                  "team->t.t_nproc=%d\n",
7635                  gtid, __kmp_threads[gtid]->th.th_team_nproc, team,
7636                  team->t.t_nproc);
7637     __kmp_print_structure();
7638   }
7639   KMP_DEBUG_ASSERT(__kmp_threads[gtid] &&
7640                    __kmp_threads[gtid]->th.th_team_nproc == team->t.t_nproc);
7641 #endif /* KMP_DEBUG */
7642 
7643   __kmp_join_barrier(gtid); /* wait for everyone */
7644 #if OMPT_SUPPORT
7645   if (ompt_enabled.enabled &&
7646       this_thr->th.ompt_thread_info.state == ompt_state_wait_barrier_implicit) {
7647     int ds_tid = this_thr->th.th_info.ds.ds_tid;
7648     ompt_data_t *task_data = OMPT_CUR_TASK_DATA(this_thr);
7649     this_thr->th.ompt_thread_info.state = ompt_state_overhead;
7650 #if OMPT_OPTIONAL
7651     void *codeptr = NULL;
7652     if (KMP_MASTER_TID(ds_tid) &&
7653         (ompt_callbacks.ompt_callback(ompt_callback_sync_region_wait) ||
7654          ompt_callbacks.ompt_callback(ompt_callback_sync_region)))
7655       codeptr = OMPT_CUR_TEAM_INFO(this_thr)->master_return_address;
7656 
7657     if (ompt_enabled.ompt_callback_sync_region_wait) {
7658       ompt_callbacks.ompt_callback(ompt_callback_sync_region_wait)(
7659           ompt_sync_region_barrier_implicit, ompt_scope_end, NULL, task_data,
7660           codeptr);
7661     }
7662     if (ompt_enabled.ompt_callback_sync_region) {
7663       ompt_callbacks.ompt_callback(ompt_callback_sync_region)(
7664           ompt_sync_region_barrier_implicit, ompt_scope_end, NULL, task_data,
7665           codeptr);
7666     }
7667 #endif
7668     if (!KMP_MASTER_TID(ds_tid) && ompt_enabled.ompt_callback_implicit_task) {
7669       ompt_callbacks.ompt_callback(ompt_callback_implicit_task)(
7670           ompt_scope_end, NULL, task_data, 0, ds_tid,
7671           ompt_task_implicit); // TODO: Can this be ompt_task_initial?
7672     }
7673   }
7674 #endif
7675 
7676   KMP_MB(); /* Flush all pending memory write invalidates.  */
7677   KMP_ASSERT(this_thr->th.th_team == team);
7678 }
7679 
7680 /* ------------------------------------------------------------------------ */
7681 
7682 #ifdef USE_LOAD_BALANCE
7683 
7684 // Return the worker threads actively spinning in the hot team, if we
7685 // are at the outermost level of parallelism.  Otherwise, return 0.
7686 static int __kmp_active_hot_team_nproc(kmp_root_t *root) {
7687   int i;
7688   int retval;
7689   kmp_team_t *hot_team;
7690 
7691   if (root->r.r_active) {
7692     return 0;
7693   }
7694   hot_team = root->r.r_hot_team;
7695   if (__kmp_dflt_blocktime == KMP_MAX_BLOCKTIME) {
7696     return hot_team->t.t_nproc - 1; // Don't count primary thread
7697   }
7698 
7699   // Skip the primary thread - it is accounted for elsewhere.
7700   retval = 0;
7701   for (i = 1; i < hot_team->t.t_nproc; i++) {
7702     if (hot_team->t.t_threads[i]->th.th_active) {
7703       retval++;
7704     }
7705   }
7706   return retval;
7707 }
7708 
7709 // Perform an automatic adjustment to the number of
7710 // threads used by the next parallel region.
7711 static int __kmp_load_balance_nproc(kmp_root_t *root, int set_nproc) {
7712   int retval;
7713   int pool_active;
7714   int hot_team_active;
7715   int team_curr_active;
7716   int system_active;
7717 
7718   KB_TRACE(20, ("__kmp_load_balance_nproc: called root:%p set_nproc:%d\n", root,
7719                 set_nproc));
7720   KMP_DEBUG_ASSERT(root);
7721   KMP_DEBUG_ASSERT(root->r.r_root_team->t.t_threads[0]
7722                        ->th.th_current_task->td_icvs.dynamic == TRUE);
7723   KMP_DEBUG_ASSERT(set_nproc > 1);
7724 
7725   if (set_nproc == 1) {
7726     KB_TRACE(20, ("__kmp_load_balance_nproc: serial execution.\n"));
7727     return 1;
7728   }
7729 
7730   // Threads that are active in the thread pool, active in the hot team for this
7731   // particular root (if we are at the outer par level), and the currently
7732   // executing thread (to become the primary thread) are available to add to the
7733   // new team, but are currently contributing to the system load, and must be
7734   // accounted for.
7735   pool_active = __kmp_thread_pool_active_nth;
7736   hot_team_active = __kmp_active_hot_team_nproc(root);
7737   team_curr_active = pool_active + hot_team_active + 1;
7738 
7739   // Check the system load.
7740   system_active = __kmp_get_load_balance(__kmp_avail_proc + team_curr_active);
7741   KB_TRACE(30, ("__kmp_load_balance_nproc: system active = %d pool active = %d "
7742                 "hot team active = %d\n",
7743                 system_active, pool_active, hot_team_active));
7744 
7745   if (system_active < 0) {
7746     // There was an error reading the necessary info from /proc, so use the
7747     // thread limit algorithm instead. Once we set __kmp_global.g.g_dynamic_mode
7748     // = dynamic_thread_limit, we shouldn't wind up getting back here.
7749     __kmp_global.g.g_dynamic_mode = dynamic_thread_limit;
7750     KMP_WARNING(CantLoadBalUsing, "KMP_DYNAMIC_MODE=thread limit");
7751 
7752     // Make this call behave like the thread limit algorithm.
7753     retval = __kmp_avail_proc - __kmp_nth +
7754              (root->r.r_active ? 1 : root->r.r_hot_team->t.t_nproc);
7755     if (retval > set_nproc) {
7756       retval = set_nproc;
7757     }
7758     if (retval < KMP_MIN_NTH) {
7759       retval = KMP_MIN_NTH;
7760     }
7761 
7762     KB_TRACE(20, ("__kmp_load_balance_nproc: thread limit exit. retval:%d\n",
7763                   retval));
7764     return retval;
7765   }
7766 
7767   // There is a slight delay in the load balance algorithm in detecting new
7768   // running procs. The real system load at this instant should be at least as
7769   // large as the #active omp thread that are available to add to the team.
7770   if (system_active < team_curr_active) {
7771     system_active = team_curr_active;
7772   }
7773   retval = __kmp_avail_proc - system_active + team_curr_active;
7774   if (retval > set_nproc) {
7775     retval = set_nproc;
7776   }
7777   if (retval < KMP_MIN_NTH) {
7778     retval = KMP_MIN_NTH;
7779   }
7780 
7781   KB_TRACE(20, ("__kmp_load_balance_nproc: exit. retval:%d\n", retval));
7782   return retval;
7783 } // __kmp_load_balance_nproc()
7784 
7785 #endif /* USE_LOAD_BALANCE */
7786 
7787 /* ------------------------------------------------------------------------ */
7788 
7789 /* NOTE: this is called with the __kmp_init_lock held */
7790 void __kmp_cleanup(void) {
7791   int f;
7792 
7793   KA_TRACE(10, ("__kmp_cleanup: enter\n"));
7794 
7795   if (TCR_4(__kmp_init_parallel)) {
7796 #if KMP_HANDLE_SIGNALS
7797     __kmp_remove_signals();
7798 #endif
7799     TCW_4(__kmp_init_parallel, FALSE);
7800   }
7801 
7802   if (TCR_4(__kmp_init_middle)) {
7803 #if KMP_AFFINITY_SUPPORTED
7804     __kmp_affinity_uninitialize();
7805 #endif /* KMP_AFFINITY_SUPPORTED */
7806     __kmp_cleanup_hierarchy();
7807     TCW_4(__kmp_init_middle, FALSE);
7808   }
7809 
7810   KA_TRACE(10, ("__kmp_cleanup: go serial cleanup\n"));
7811 
7812   if (__kmp_init_serial) {
7813     __kmp_runtime_destroy();
7814     __kmp_init_serial = FALSE;
7815   }
7816 
7817   __kmp_cleanup_threadprivate_caches();
7818 
7819   for (f = 0; f < __kmp_threads_capacity; f++) {
7820     if (__kmp_root[f] != NULL) {
7821       __kmp_free(__kmp_root[f]);
7822       __kmp_root[f] = NULL;
7823     }
7824   }
7825   __kmp_free(__kmp_threads);
7826   // __kmp_threads and __kmp_root were allocated at once, as single block, so
7827   // there is no need in freeing __kmp_root.
7828   __kmp_threads = NULL;
7829   __kmp_root = NULL;
7830   __kmp_threads_capacity = 0;
7831 
7832 #if KMP_USE_DYNAMIC_LOCK
7833   __kmp_cleanup_indirect_user_locks();
7834 #else
7835   __kmp_cleanup_user_locks();
7836 #endif
7837 
7838 #if KMP_AFFINITY_SUPPORTED
7839   KMP_INTERNAL_FREE(CCAST(char *, __kmp_cpuinfo_file));
7840   __kmp_cpuinfo_file = NULL;
7841 #endif /* KMP_AFFINITY_SUPPORTED */
7842 
7843 #if KMP_USE_ADAPTIVE_LOCKS
7844 #if KMP_DEBUG_ADAPTIVE_LOCKS
7845   __kmp_print_speculative_stats();
7846 #endif
7847 #endif
7848   KMP_INTERNAL_FREE(__kmp_nested_nth.nth);
7849   __kmp_nested_nth.nth = NULL;
7850   __kmp_nested_nth.size = 0;
7851   __kmp_nested_nth.used = 0;
7852   KMP_INTERNAL_FREE(__kmp_nested_proc_bind.bind_types);
7853   __kmp_nested_proc_bind.bind_types = NULL;
7854   __kmp_nested_proc_bind.size = 0;
7855   __kmp_nested_proc_bind.used = 0;
7856   if (__kmp_affinity_format) {
7857     KMP_INTERNAL_FREE(__kmp_affinity_format);
7858     __kmp_affinity_format = NULL;
7859   }
7860 
7861   __kmp_i18n_catclose();
7862 
7863 #if KMP_USE_HIER_SCHED
7864   __kmp_hier_scheds.deallocate();
7865 #endif
7866 
7867 #if KMP_STATS_ENABLED
7868   __kmp_stats_fini();
7869 #endif
7870 
7871   KA_TRACE(10, ("__kmp_cleanup: exit\n"));
7872 }
7873 
7874 /* ------------------------------------------------------------------------ */
7875 
7876 int __kmp_ignore_mppbeg(void) {
7877   char *env;
7878 
7879   if ((env = getenv("KMP_IGNORE_MPPBEG")) != NULL) {
7880     if (__kmp_str_match_false(env))
7881       return FALSE;
7882   }
7883   // By default __kmpc_begin() is no-op.
7884   return TRUE;
7885 }
7886 
7887 int __kmp_ignore_mppend(void) {
7888   char *env;
7889 
7890   if ((env = getenv("KMP_IGNORE_MPPEND")) != NULL) {
7891     if (__kmp_str_match_false(env))
7892       return FALSE;
7893   }
7894   // By default __kmpc_end() is no-op.
7895   return TRUE;
7896 }
7897 
7898 void __kmp_internal_begin(void) {
7899   int gtid;
7900   kmp_root_t *root;
7901 
7902   /* this is a very important step as it will register new sibling threads
7903      and assign these new uber threads a new gtid */
7904   gtid = __kmp_entry_gtid();
7905   root = __kmp_threads[gtid]->th.th_root;
7906   KMP_ASSERT(KMP_UBER_GTID(gtid));
7907 
7908   if (root->r.r_begin)
7909     return;
7910   __kmp_acquire_lock(&root->r.r_begin_lock, gtid);
7911   if (root->r.r_begin) {
7912     __kmp_release_lock(&root->r.r_begin_lock, gtid);
7913     return;
7914   }
7915 
7916   root->r.r_begin = TRUE;
7917 
7918   __kmp_release_lock(&root->r.r_begin_lock, gtid);
7919 }
7920 
7921 /* ------------------------------------------------------------------------ */
7922 
7923 void __kmp_user_set_library(enum library_type arg) {
7924   int gtid;
7925   kmp_root_t *root;
7926   kmp_info_t *thread;
7927 
7928   /* first, make sure we are initialized so we can get our gtid */
7929 
7930   gtid = __kmp_entry_gtid();
7931   thread = __kmp_threads[gtid];
7932 
7933   root = thread->th.th_root;
7934 
7935   KA_TRACE(20, ("__kmp_user_set_library: enter T#%d, arg: %d, %d\n", gtid, arg,
7936                 library_serial));
7937   if (root->r.r_in_parallel) { /* Must be called in serial section of top-level
7938                                   thread */
7939     KMP_WARNING(SetLibraryIncorrectCall);
7940     return;
7941   }
7942 
7943   switch (arg) {
7944   case library_serial:
7945     thread->th.th_set_nproc = 0;
7946     set__nproc(thread, 1);
7947     break;
7948   case library_turnaround:
7949     thread->th.th_set_nproc = 0;
7950     set__nproc(thread, __kmp_dflt_team_nth ? __kmp_dflt_team_nth
7951                                            : __kmp_dflt_team_nth_ub);
7952     break;
7953   case library_throughput:
7954     thread->th.th_set_nproc = 0;
7955     set__nproc(thread, __kmp_dflt_team_nth ? __kmp_dflt_team_nth
7956                                            : __kmp_dflt_team_nth_ub);
7957     break;
7958   default:
7959     KMP_FATAL(UnknownLibraryType, arg);
7960   }
7961 
7962   __kmp_aux_set_library(arg);
7963 }
7964 
7965 void __kmp_aux_set_stacksize(size_t arg) {
7966   if (!__kmp_init_serial)
7967     __kmp_serial_initialize();
7968 
7969 #if KMP_OS_DARWIN
7970   if (arg & (0x1000 - 1)) {
7971     arg &= ~(0x1000 - 1);
7972     if (arg + 0x1000) /* check for overflow if we round up */
7973       arg += 0x1000;
7974   }
7975 #endif
7976   __kmp_acquire_bootstrap_lock(&__kmp_initz_lock);
7977 
7978   /* only change the default stacksize before the first parallel region */
7979   if (!TCR_4(__kmp_init_parallel)) {
7980     size_t value = arg; /* argument is in bytes */
7981 
7982     if (value < __kmp_sys_min_stksize)
7983       value = __kmp_sys_min_stksize;
7984     else if (value > KMP_MAX_STKSIZE)
7985       value = KMP_MAX_STKSIZE;
7986 
7987     __kmp_stksize = value;
7988 
7989     __kmp_env_stksize = TRUE; /* was KMP_STACKSIZE specified? */
7990   }
7991 
7992   __kmp_release_bootstrap_lock(&__kmp_initz_lock);
7993 }
7994 
7995 /* set the behaviour of the runtime library */
7996 /* TODO this can cause some odd behaviour with sibling parallelism... */
7997 void __kmp_aux_set_library(enum library_type arg) {
7998   __kmp_library = arg;
7999 
8000   switch (__kmp_library) {
8001   case library_serial: {
8002     KMP_INFORM(LibraryIsSerial);
8003   } break;
8004   case library_turnaround:
8005     if (__kmp_use_yield == 1 && !__kmp_use_yield_exp_set)
8006       __kmp_use_yield = 2; // only yield when oversubscribed
8007     break;
8008   case library_throughput:
8009     if (__kmp_dflt_blocktime == KMP_MAX_BLOCKTIME)
8010       __kmp_dflt_blocktime = 200;
8011     break;
8012   default:
8013     KMP_FATAL(UnknownLibraryType, arg);
8014   }
8015 }
8016 
8017 /* Getting team information common for all team API */
8018 // Returns NULL if not in teams construct
8019 static kmp_team_t *__kmp_aux_get_team_info(int &teams_serialized) {
8020   kmp_info_t *thr = __kmp_entry_thread();
8021   teams_serialized = 0;
8022   if (thr->th.th_teams_microtask) {
8023     kmp_team_t *team = thr->th.th_team;
8024     int tlevel = thr->th.th_teams_level; // the level of the teams construct
8025     int ii = team->t.t_level;
8026     teams_serialized = team->t.t_serialized;
8027     int level = tlevel + 1;
8028     KMP_DEBUG_ASSERT(ii >= tlevel);
8029     while (ii > level) {
8030       for (teams_serialized = team->t.t_serialized;
8031            (teams_serialized > 0) && (ii > level); teams_serialized--, ii--) {
8032       }
8033       if (team->t.t_serialized && (!teams_serialized)) {
8034         team = team->t.t_parent;
8035         continue;
8036       }
8037       if (ii > level) {
8038         team = team->t.t_parent;
8039         ii--;
8040       }
8041     }
8042     return team;
8043   }
8044   return NULL;
8045 }
8046 
8047 int __kmp_aux_get_team_num() {
8048   int serialized;
8049   kmp_team_t *team = __kmp_aux_get_team_info(serialized);
8050   if (team) {
8051     if (serialized > 1) {
8052       return 0; // teams region is serialized ( 1 team of 1 thread ).
8053     } else {
8054       return team->t.t_master_tid;
8055     }
8056   }
8057   return 0;
8058 }
8059 
8060 int __kmp_aux_get_num_teams() {
8061   int serialized;
8062   kmp_team_t *team = __kmp_aux_get_team_info(serialized);
8063   if (team) {
8064     if (serialized > 1) {
8065       return 1;
8066     } else {
8067       return team->t.t_parent->t.t_nproc;
8068     }
8069   }
8070   return 1;
8071 }
8072 
8073 /* ------------------------------------------------------------------------ */
8074 
8075 /*
8076  * Affinity Format Parser
8077  *
8078  * Field is in form of: %[[[0].]size]type
8079  * % and type are required (%% means print a literal '%')
8080  * type is either single char or long name surrounded by {},
8081  * e.g., N or {num_threads}
8082  * 0 => leading zeros
8083  * . => right justified when size is specified
8084  * by default output is left justified
8085  * size is the *minimum* field length
8086  * All other characters are printed as is
8087  *
8088  * Available field types:
8089  * L {thread_level}      - omp_get_level()
8090  * n {thread_num}        - omp_get_thread_num()
8091  * h {host}              - name of host machine
8092  * P {process_id}        - process id (integer)
8093  * T {thread_identifier} - native thread identifier (integer)
8094  * N {num_threads}       - omp_get_num_threads()
8095  * A {ancestor_tnum}     - omp_get_ancestor_thread_num(omp_get_level()-1)
8096  * a {thread_affinity}   - comma separated list of integers or integer ranges
8097  *                         (values of affinity mask)
8098  *
8099  * Implementation-specific field types can be added
8100  * If a type is unknown, print "undefined"
8101  */
8102 
8103 // Structure holding the short name, long name, and corresponding data type
8104 // for snprintf.  A table of these will represent the entire valid keyword
8105 // field types.
8106 typedef struct kmp_affinity_format_field_t {
8107   char short_name; // from spec e.g., L -> thread level
8108   const char *long_name; // from spec thread_level -> thread level
8109   char field_format; // data type for snprintf (typically 'd' or 's'
8110   // for integer or string)
8111 } kmp_affinity_format_field_t;
8112 
8113 static const kmp_affinity_format_field_t __kmp_affinity_format_table[] = {
8114 #if KMP_AFFINITY_SUPPORTED
8115     {'A', "thread_affinity", 's'},
8116 #endif
8117     {'t', "team_num", 'd'},
8118     {'T', "num_teams", 'd'},
8119     {'L', "nesting_level", 'd'},
8120     {'n', "thread_num", 'd'},
8121     {'N', "num_threads", 'd'},
8122     {'a', "ancestor_tnum", 'd'},
8123     {'H', "host", 's'},
8124     {'P', "process_id", 'd'},
8125     {'i', "native_thread_id", 'd'}};
8126 
8127 // Return the number of characters it takes to hold field
8128 static int __kmp_aux_capture_affinity_field(int gtid, const kmp_info_t *th,
8129                                             const char **ptr,
8130                                             kmp_str_buf_t *field_buffer) {
8131   int rc, format_index, field_value;
8132   const char *width_left, *width_right;
8133   bool pad_zeros, right_justify, parse_long_name, found_valid_name;
8134   static const int FORMAT_SIZE = 20;
8135   char format[FORMAT_SIZE] = {0};
8136   char absolute_short_name = 0;
8137 
8138   KMP_DEBUG_ASSERT(gtid >= 0);
8139   KMP_DEBUG_ASSERT(th);
8140   KMP_DEBUG_ASSERT(**ptr == '%');
8141   KMP_DEBUG_ASSERT(field_buffer);
8142 
8143   __kmp_str_buf_clear(field_buffer);
8144 
8145   // Skip the initial %
8146   (*ptr)++;
8147 
8148   // Check for %% first
8149   if (**ptr == '%') {
8150     __kmp_str_buf_cat(field_buffer, "%", 1);
8151     (*ptr)++; // skip over the second %
8152     return 1;
8153   }
8154 
8155   // Parse field modifiers if they are present
8156   pad_zeros = false;
8157   if (**ptr == '0') {
8158     pad_zeros = true;
8159     (*ptr)++; // skip over 0
8160   }
8161   right_justify = false;
8162   if (**ptr == '.') {
8163     right_justify = true;
8164     (*ptr)++; // skip over .
8165   }
8166   // Parse width of field: [width_left, width_right)
8167   width_left = width_right = NULL;
8168   if (**ptr >= '0' && **ptr <= '9') {
8169     width_left = *ptr;
8170     SKIP_DIGITS(*ptr);
8171     width_right = *ptr;
8172   }
8173 
8174   // Create the format for KMP_SNPRINTF based on flags parsed above
8175   format_index = 0;
8176   format[format_index++] = '%';
8177   if (!right_justify)
8178     format[format_index++] = '-';
8179   if (pad_zeros)
8180     format[format_index++] = '0';
8181   if (width_left && width_right) {
8182     int i = 0;
8183     // Only allow 8 digit number widths.
8184     // This also prevents overflowing format variable
8185     while (i < 8 && width_left < width_right) {
8186       format[format_index++] = *width_left;
8187       width_left++;
8188       i++;
8189     }
8190   }
8191 
8192   // Parse a name (long or short)
8193   // Canonicalize the name into absolute_short_name
8194   found_valid_name = false;
8195   parse_long_name = (**ptr == '{');
8196   if (parse_long_name)
8197     (*ptr)++; // skip initial left brace
8198   for (size_t i = 0; i < sizeof(__kmp_affinity_format_table) /
8199                              sizeof(__kmp_affinity_format_table[0]);
8200        ++i) {
8201     char short_name = __kmp_affinity_format_table[i].short_name;
8202     const char *long_name = __kmp_affinity_format_table[i].long_name;
8203     char field_format = __kmp_affinity_format_table[i].field_format;
8204     if (parse_long_name) {
8205       size_t length = KMP_STRLEN(long_name);
8206       if (strncmp(*ptr, long_name, length) == 0) {
8207         found_valid_name = true;
8208         (*ptr) += length; // skip the long name
8209       }
8210     } else if (**ptr == short_name) {
8211       found_valid_name = true;
8212       (*ptr)++; // skip the short name
8213     }
8214     if (found_valid_name) {
8215       format[format_index++] = field_format;
8216       format[format_index++] = '\0';
8217       absolute_short_name = short_name;
8218       break;
8219     }
8220   }
8221   if (parse_long_name) {
8222     if (**ptr != '}') {
8223       absolute_short_name = 0;
8224     } else {
8225       (*ptr)++; // skip over the right brace
8226     }
8227   }
8228 
8229   // Attempt to fill the buffer with the requested
8230   // value using snprintf within __kmp_str_buf_print()
8231   switch (absolute_short_name) {
8232   case 't':
8233     rc = __kmp_str_buf_print(field_buffer, format, __kmp_aux_get_team_num());
8234     break;
8235   case 'T':
8236     rc = __kmp_str_buf_print(field_buffer, format, __kmp_aux_get_num_teams());
8237     break;
8238   case 'L':
8239     rc = __kmp_str_buf_print(field_buffer, format, th->th.th_team->t.t_level);
8240     break;
8241   case 'n':
8242     rc = __kmp_str_buf_print(field_buffer, format, __kmp_tid_from_gtid(gtid));
8243     break;
8244   case 'H': {
8245     static const int BUFFER_SIZE = 256;
8246     char buf[BUFFER_SIZE];
8247     __kmp_expand_host_name(buf, BUFFER_SIZE);
8248     rc = __kmp_str_buf_print(field_buffer, format, buf);
8249   } break;
8250   case 'P':
8251     rc = __kmp_str_buf_print(field_buffer, format, getpid());
8252     break;
8253   case 'i':
8254     rc = __kmp_str_buf_print(field_buffer, format, __kmp_gettid());
8255     break;
8256   case 'N':
8257     rc = __kmp_str_buf_print(field_buffer, format, th->th.th_team->t.t_nproc);
8258     break;
8259   case 'a':
8260     field_value =
8261         __kmp_get_ancestor_thread_num(gtid, th->th.th_team->t.t_level - 1);
8262     rc = __kmp_str_buf_print(field_buffer, format, field_value);
8263     break;
8264 #if KMP_AFFINITY_SUPPORTED
8265   case 'A': {
8266     kmp_str_buf_t buf;
8267     __kmp_str_buf_init(&buf);
8268     __kmp_affinity_str_buf_mask(&buf, th->th.th_affin_mask);
8269     rc = __kmp_str_buf_print(field_buffer, format, buf.str);
8270     __kmp_str_buf_free(&buf);
8271   } break;
8272 #endif
8273   default:
8274     // According to spec, If an implementation does not have info for field
8275     // type, then "undefined" is printed
8276     rc = __kmp_str_buf_print(field_buffer, "%s", "undefined");
8277     // Skip the field
8278     if (parse_long_name) {
8279       SKIP_TOKEN(*ptr);
8280       if (**ptr == '}')
8281         (*ptr)++;
8282     } else {
8283       (*ptr)++;
8284     }
8285   }
8286 
8287   KMP_ASSERT(format_index <= FORMAT_SIZE);
8288   return rc;
8289 }
8290 
8291 /*
8292  * Return number of characters needed to hold the affinity string
8293  * (not including null byte character)
8294  * The resultant string is printed to buffer, which the caller can then
8295  * handle afterwards
8296  */
8297 size_t __kmp_aux_capture_affinity(int gtid, const char *format,
8298                                   kmp_str_buf_t *buffer) {
8299   const char *parse_ptr;
8300   size_t retval;
8301   const kmp_info_t *th;
8302   kmp_str_buf_t field;
8303 
8304   KMP_DEBUG_ASSERT(buffer);
8305   KMP_DEBUG_ASSERT(gtid >= 0);
8306 
8307   __kmp_str_buf_init(&field);
8308   __kmp_str_buf_clear(buffer);
8309 
8310   th = __kmp_threads[gtid];
8311   retval = 0;
8312 
8313   // If format is NULL or zero-length string, then we use
8314   // affinity-format-var ICV
8315   parse_ptr = format;
8316   if (parse_ptr == NULL || *parse_ptr == '\0') {
8317     parse_ptr = __kmp_affinity_format;
8318   }
8319   KMP_DEBUG_ASSERT(parse_ptr);
8320 
8321   while (*parse_ptr != '\0') {
8322     // Parse a field
8323     if (*parse_ptr == '%') {
8324       // Put field in the buffer
8325       int rc = __kmp_aux_capture_affinity_field(gtid, th, &parse_ptr, &field);
8326       __kmp_str_buf_catbuf(buffer, &field);
8327       retval += rc;
8328     } else {
8329       // Put literal character in buffer
8330       __kmp_str_buf_cat(buffer, parse_ptr, 1);
8331       retval++;
8332       parse_ptr++;
8333     }
8334   }
8335   __kmp_str_buf_free(&field);
8336   return retval;
8337 }
8338 
8339 // Displays the affinity string to stdout
8340 void __kmp_aux_display_affinity(int gtid, const char *format) {
8341   kmp_str_buf_t buf;
8342   __kmp_str_buf_init(&buf);
8343   __kmp_aux_capture_affinity(gtid, format, &buf);
8344   __kmp_fprintf(kmp_out, "%s" KMP_END_OF_LINE, buf.str);
8345   __kmp_str_buf_free(&buf);
8346 }
8347 
8348 /* ------------------------------------------------------------------------ */
8349 
8350 void __kmp_aux_set_blocktime(int arg, kmp_info_t *thread, int tid) {
8351   int blocktime = arg; /* argument is in milliseconds */
8352 #if KMP_USE_MONITOR
8353   int bt_intervals;
8354 #endif
8355   kmp_int8 bt_set;
8356 
8357   __kmp_save_internal_controls(thread);
8358 
8359   /* Normalize and set blocktime for the teams */
8360   if (blocktime < KMP_MIN_BLOCKTIME)
8361     blocktime = KMP_MIN_BLOCKTIME;
8362   else if (blocktime > KMP_MAX_BLOCKTIME)
8363     blocktime = KMP_MAX_BLOCKTIME;
8364 
8365   set__blocktime_team(thread->th.th_team, tid, blocktime);
8366   set__blocktime_team(thread->th.th_serial_team, 0, blocktime);
8367 
8368 #if KMP_USE_MONITOR
8369   /* Calculate and set blocktime intervals for the teams */
8370   bt_intervals = KMP_INTERVALS_FROM_BLOCKTIME(blocktime, __kmp_monitor_wakeups);
8371 
8372   set__bt_intervals_team(thread->th.th_team, tid, bt_intervals);
8373   set__bt_intervals_team(thread->th.th_serial_team, 0, bt_intervals);
8374 #endif
8375 
8376   /* Set whether blocktime has been set to "TRUE" */
8377   bt_set = TRUE;
8378 
8379   set__bt_set_team(thread->th.th_team, tid, bt_set);
8380   set__bt_set_team(thread->th.th_serial_team, 0, bt_set);
8381 #if KMP_USE_MONITOR
8382   KF_TRACE(10, ("kmp_set_blocktime: T#%d(%d:%d), blocktime=%d, "
8383                 "bt_intervals=%d, monitor_updates=%d\n",
8384                 __kmp_gtid_from_tid(tid, thread->th.th_team),
8385                 thread->th.th_team->t.t_id, tid, blocktime, bt_intervals,
8386                 __kmp_monitor_wakeups));
8387 #else
8388   KF_TRACE(10, ("kmp_set_blocktime: T#%d(%d:%d), blocktime=%d\n",
8389                 __kmp_gtid_from_tid(tid, thread->th.th_team),
8390                 thread->th.th_team->t.t_id, tid, blocktime));
8391 #endif
8392 }
8393 
8394 void __kmp_aux_set_defaults(char const *str, size_t len) {
8395   if (!__kmp_init_serial) {
8396     __kmp_serial_initialize();
8397   }
8398   __kmp_env_initialize(str);
8399 
8400   if (__kmp_settings || __kmp_display_env || __kmp_display_env_verbose) {
8401     __kmp_env_print();
8402   }
8403 } // __kmp_aux_set_defaults
8404 
8405 /* ------------------------------------------------------------------------ */
8406 /* internal fast reduction routines */
8407 
8408 PACKED_REDUCTION_METHOD_T
8409 __kmp_determine_reduction_method(
8410     ident_t *loc, kmp_int32 global_tid, kmp_int32 num_vars, size_t reduce_size,
8411     void *reduce_data, void (*reduce_func)(void *lhs_data, void *rhs_data),
8412     kmp_critical_name *lck) {
8413 
8414   // Default reduction method: critical construct ( lck != NULL, like in current
8415   // PAROPT )
8416   // If ( reduce_data!=NULL && reduce_func!=NULL ): the tree-reduction method
8417   // can be selected by RTL
8418   // If loc->flags contains KMP_IDENT_ATOMIC_REDUCE, the atomic reduce method
8419   // can be selected by RTL
8420   // Finally, it's up to OpenMP RTL to make a decision on which method to select
8421   // among generated by PAROPT.
8422 
8423   PACKED_REDUCTION_METHOD_T retval;
8424 
8425   int team_size;
8426 
8427   KMP_DEBUG_ASSERT(loc); // it would be nice to test ( loc != 0 )
8428   KMP_DEBUG_ASSERT(lck); // it would be nice to test ( lck != 0 )
8429 
8430 #define FAST_REDUCTION_ATOMIC_METHOD_GENERATED                                 \
8431   ((loc->flags & (KMP_IDENT_ATOMIC_REDUCE)) == (KMP_IDENT_ATOMIC_REDUCE))
8432 #define FAST_REDUCTION_TREE_METHOD_GENERATED ((reduce_data) && (reduce_func))
8433 
8434   retval = critical_reduce_block;
8435 
8436   // another choice of getting a team size (with 1 dynamic deference) is slower
8437   team_size = __kmp_get_team_num_threads(global_tid);
8438   if (team_size == 1) {
8439 
8440     retval = empty_reduce_block;
8441 
8442   } else {
8443 
8444     int atomic_available = FAST_REDUCTION_ATOMIC_METHOD_GENERATED;
8445 
8446 #if KMP_ARCH_X86_64 || KMP_ARCH_PPC64 || KMP_ARCH_AARCH64 ||                   \
8447     KMP_ARCH_MIPS64 || KMP_ARCH_RISCV64
8448 
8449 #if KMP_OS_LINUX || KMP_OS_DRAGONFLY || KMP_OS_FREEBSD || KMP_OS_NETBSD ||     \
8450     KMP_OS_OPENBSD || KMP_OS_WINDOWS || KMP_OS_DARWIN || KMP_OS_HURD
8451 
8452     int teamsize_cutoff = 4;
8453 
8454 #if KMP_MIC_SUPPORTED
8455     if (__kmp_mic_type != non_mic) {
8456       teamsize_cutoff = 8;
8457     }
8458 #endif
8459     int tree_available = FAST_REDUCTION_TREE_METHOD_GENERATED;
8460     if (tree_available) {
8461       if (team_size <= teamsize_cutoff) {
8462         if (atomic_available) {
8463           retval = atomic_reduce_block;
8464         }
8465       } else {
8466         retval = TREE_REDUCE_BLOCK_WITH_REDUCTION_BARRIER;
8467       }
8468     } else if (atomic_available) {
8469       retval = atomic_reduce_block;
8470     }
8471 #else
8472 #error "Unknown or unsupported OS"
8473 #endif // KMP_OS_LINUX || KMP_OS_DRAGONFLY || KMP_OS_FREEBSD || KMP_OS_NETBSD ||
8474        // KMP_OS_OPENBSD || KMP_OS_WINDOWS || KMP_OS_DARWIN || KMP_OS_HURD
8475 
8476 #elif KMP_ARCH_X86 || KMP_ARCH_ARM || KMP_ARCH_AARCH || KMP_ARCH_MIPS
8477 
8478 #if KMP_OS_LINUX || KMP_OS_FREEBSD || KMP_OS_WINDOWS || KMP_OS_HURD
8479 
8480     // basic tuning
8481 
8482     if (atomic_available) {
8483       if (num_vars <= 2) { // && ( team_size <= 8 ) due to false-sharing ???
8484         retval = atomic_reduce_block;
8485       }
8486     } // otherwise: use critical section
8487 
8488 #elif KMP_OS_DARWIN
8489 
8490     int tree_available = FAST_REDUCTION_TREE_METHOD_GENERATED;
8491     if (atomic_available && (num_vars <= 3)) {
8492       retval = atomic_reduce_block;
8493     } else if (tree_available) {
8494       if ((reduce_size > (9 * sizeof(kmp_real64))) &&
8495           (reduce_size < (2000 * sizeof(kmp_real64)))) {
8496         retval = TREE_REDUCE_BLOCK_WITH_PLAIN_BARRIER;
8497       }
8498     } // otherwise: use critical section
8499 
8500 #else
8501 #error "Unknown or unsupported OS"
8502 #endif
8503 
8504 #else
8505 #error "Unknown or unsupported architecture"
8506 #endif
8507   }
8508 
8509   // KMP_FORCE_REDUCTION
8510 
8511   // If the team is serialized (team_size == 1), ignore the forced reduction
8512   // method and stay with the unsynchronized method (empty_reduce_block)
8513   if (__kmp_force_reduction_method != reduction_method_not_defined &&
8514       team_size != 1) {
8515 
8516     PACKED_REDUCTION_METHOD_T forced_retval = critical_reduce_block;
8517 
8518     int atomic_available, tree_available;
8519 
8520     switch ((forced_retval = __kmp_force_reduction_method)) {
8521     case critical_reduce_block:
8522       KMP_ASSERT(lck); // lck should be != 0
8523       break;
8524 
8525     case atomic_reduce_block:
8526       atomic_available = FAST_REDUCTION_ATOMIC_METHOD_GENERATED;
8527       if (!atomic_available) {
8528         KMP_WARNING(RedMethodNotSupported, "atomic");
8529         forced_retval = critical_reduce_block;
8530       }
8531       break;
8532 
8533     case tree_reduce_block:
8534       tree_available = FAST_REDUCTION_TREE_METHOD_GENERATED;
8535       if (!tree_available) {
8536         KMP_WARNING(RedMethodNotSupported, "tree");
8537         forced_retval = critical_reduce_block;
8538       } else {
8539 #if KMP_FAST_REDUCTION_BARRIER
8540         forced_retval = TREE_REDUCE_BLOCK_WITH_REDUCTION_BARRIER;
8541 #endif
8542       }
8543       break;
8544 
8545     default:
8546       KMP_ASSERT(0); // "unsupported method specified"
8547     }
8548 
8549     retval = forced_retval;
8550   }
8551 
8552   KA_TRACE(10, ("reduction method selected=%08x\n", retval));
8553 
8554 #undef FAST_REDUCTION_TREE_METHOD_GENERATED
8555 #undef FAST_REDUCTION_ATOMIC_METHOD_GENERATED
8556 
8557   return (retval);
8558 }
8559 // this function is for testing set/get/determine reduce method
8560 kmp_int32 __kmp_get_reduce_method(void) {
8561   return ((__kmp_entry_thread()->th.th_local.packed_reduction_method) >> 8);
8562 }
8563 
8564 // Soft pause sets up threads to ignore blocktime and just go to sleep.
8565 // Spin-wait code checks __kmp_pause_status and reacts accordingly.
8566 void __kmp_soft_pause() { __kmp_pause_status = kmp_soft_paused; }
8567 
8568 // Hard pause shuts down the runtime completely.  Resume happens naturally when
8569 // OpenMP is used subsequently.
8570 void __kmp_hard_pause() {
8571   __kmp_pause_status = kmp_hard_paused;
8572   __kmp_internal_end_thread(-1);
8573 }
8574 
8575 // Soft resume sets __kmp_pause_status, and wakes up all threads.
8576 void __kmp_resume_if_soft_paused() {
8577   if (__kmp_pause_status == kmp_soft_paused) {
8578     __kmp_pause_status = kmp_not_paused;
8579 
8580     for (int gtid = 1; gtid < __kmp_threads_capacity; ++gtid) {
8581       kmp_info_t *thread = __kmp_threads[gtid];
8582       if (thread) { // Wake it if sleeping
8583         kmp_flag_64<> fl(&thread->th.th_bar[bs_forkjoin_barrier].bb.b_go,
8584                          thread);
8585         if (fl.is_sleeping())
8586           fl.resume(gtid);
8587         else if (__kmp_try_suspend_mx(thread)) { // got suspend lock
8588           __kmp_unlock_suspend_mx(thread); // unlock it; it won't sleep
8589         } else { // thread holds the lock and may sleep soon
8590           do { // until either the thread sleeps, or we can get the lock
8591             if (fl.is_sleeping()) {
8592               fl.resume(gtid);
8593               break;
8594             } else if (__kmp_try_suspend_mx(thread)) {
8595               __kmp_unlock_suspend_mx(thread);
8596               break;
8597             }
8598           } while (1);
8599         }
8600       }
8601     }
8602   }
8603 }
8604 
8605 // This function is called via __kmpc_pause_resource. Returns 0 if successful.
8606 // TODO: add warning messages
8607 int __kmp_pause_resource(kmp_pause_status_t level) {
8608   if (level == kmp_not_paused) { // requesting resume
8609     if (__kmp_pause_status == kmp_not_paused) {
8610       // error message about runtime not being paused, so can't resume
8611       return 1;
8612     } else {
8613       KMP_DEBUG_ASSERT(__kmp_pause_status == kmp_soft_paused ||
8614                        __kmp_pause_status == kmp_hard_paused);
8615       __kmp_pause_status = kmp_not_paused;
8616       return 0;
8617     }
8618   } else if (level == kmp_soft_paused) { // requesting soft pause
8619     if (__kmp_pause_status != kmp_not_paused) {
8620       // error message about already being paused
8621       return 1;
8622     } else {
8623       __kmp_soft_pause();
8624       return 0;
8625     }
8626   } else if (level == kmp_hard_paused) { // requesting hard pause
8627     if (__kmp_pause_status != kmp_not_paused) {
8628       // error message about already being paused
8629       return 1;
8630     } else {
8631       __kmp_hard_pause();
8632       return 0;
8633     }
8634   } else {
8635     // error message about invalid level
8636     return 1;
8637   }
8638 }
8639 
8640 void __kmp_omp_display_env(int verbose) {
8641   __kmp_acquire_bootstrap_lock(&__kmp_initz_lock);
8642   if (__kmp_init_serial == 0)
8643     __kmp_do_serial_initialize();
8644   __kmp_display_env_impl(!verbose, verbose);
8645   __kmp_release_bootstrap_lock(&__kmp_initz_lock);
8646 }
8647 
8648 // Globals and functions for hidden helper task
8649 kmp_info_t **__kmp_hidden_helper_threads;
8650 kmp_info_t *__kmp_hidden_helper_main_thread;
8651 kmp_int32 __kmp_hidden_helper_threads_num = 8;
8652 std::atomic<kmp_int32> __kmp_unexecuted_hidden_helper_tasks;
8653 #if KMP_OS_LINUX
8654 kmp_int32 __kmp_enable_hidden_helper = TRUE;
8655 #else
8656 kmp_int32 __kmp_enable_hidden_helper = FALSE;
8657 #endif
8658 
8659 namespace {
8660 std::atomic<kmp_int32> __kmp_hit_hidden_helper_threads_num;
8661 
8662 void __kmp_hidden_helper_wrapper_fn(int *gtid, int *, ...) {
8663   // This is an explicit synchronization on all hidden helper threads in case
8664   // that when a regular thread pushes a hidden helper task to one hidden
8665   // helper thread, the thread has not been awaken once since they're released
8666   // by the main thread after creating the team.
8667   KMP_ATOMIC_INC(&__kmp_hit_hidden_helper_threads_num);
8668   while (KMP_ATOMIC_LD_ACQ(&__kmp_hit_hidden_helper_threads_num) !=
8669          __kmp_hidden_helper_threads_num)
8670     ;
8671 
8672   // If main thread, then wait for signal
8673   if (__kmpc_master(nullptr, *gtid)) {
8674     // First, unset the initial state and release the initial thread
8675     TCW_4(__kmp_init_hidden_helper_threads, FALSE);
8676     __kmp_hidden_helper_initz_release();
8677     __kmp_hidden_helper_main_thread_wait();
8678     // Now wake up all worker threads
8679     for (int i = 1; i < __kmp_hit_hidden_helper_threads_num; ++i) {
8680       __kmp_hidden_helper_worker_thread_signal();
8681     }
8682   }
8683 }
8684 } // namespace
8685 
8686 void __kmp_hidden_helper_threads_initz_routine() {
8687   // Create a new root for hidden helper team/threads
8688   const int gtid = __kmp_register_root(TRUE);
8689   __kmp_hidden_helper_main_thread = __kmp_threads[gtid];
8690   __kmp_hidden_helper_threads = &__kmp_threads[gtid];
8691   __kmp_hidden_helper_main_thread->th.th_set_nproc =
8692       __kmp_hidden_helper_threads_num;
8693 
8694   KMP_ATOMIC_ST_REL(&__kmp_hit_hidden_helper_threads_num, 0);
8695 
8696   __kmpc_fork_call(nullptr, 0, __kmp_hidden_helper_wrapper_fn);
8697 
8698   // Set the initialization flag to FALSE
8699   TCW_SYNC_4(__kmp_init_hidden_helper, FALSE);
8700 
8701   __kmp_hidden_helper_threads_deinitz_release();
8702 }
8703