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 /* these are temporary issues to be dealt with */
36 #define KMP_USE_PRCTL 0
37 
38 #if KMP_OS_WINDOWS
39 #include <process.h>
40 #endif
41 
42 #include "tsan_annotations.h"
43 
44 #if defined(KMP_GOMP_COMPAT)
45 char const __kmp_version_alt_comp[] =
46     KMP_VERSION_PREFIX "alternative compiler support: yes";
47 #endif /* defined(KMP_GOMP_COMPAT) */
48 
49 char const __kmp_version_omp_api[] = KMP_VERSION_PREFIX "API version: "
50 #if OMP_50_ENABLED
51                                                         "5.0 (201611)";
52 #elif OMP_45_ENABLED
53                                                         "4.5 (201511)";
54 #elif OMP_40_ENABLED
55                                                         "4.0 (201307)";
56 #else
57                                                         "3.1 (201107)";
58 #endif
59 
60 #ifdef KMP_DEBUG
61 char const __kmp_version_lock[] =
62     KMP_VERSION_PREFIX "lock type: run time selectable";
63 #endif /* KMP_DEBUG */
64 
65 #define KMP_MIN(x, y) ((x) < (y) ? (x) : (y))
66 
67 /* ------------------------------------------------------------------------ */
68 
69 #if KMP_USE_MONITOR
70 kmp_info_t __kmp_monitor;
71 #endif
72 
73 /* Forward declarations */
74 
75 void __kmp_cleanup(void);
76 
77 static void __kmp_initialize_info(kmp_info_t *, kmp_team_t *, int tid,
78                                   int gtid);
79 static void __kmp_initialize_team(kmp_team_t *team, int new_nproc,
80                                   kmp_internal_control_t *new_icvs,
81                                   ident_t *loc);
82 #if OMP_40_ENABLED && KMP_AFFINITY_SUPPORTED
83 static void __kmp_partition_places(kmp_team_t *team,
84                                    int update_master_only = 0);
85 #endif
86 static void __kmp_do_serial_initialize(void);
87 void __kmp_fork_barrier(int gtid, int tid);
88 void __kmp_join_barrier(int gtid);
89 void __kmp_setup_icv_copy(kmp_team_t *team, int new_nproc,
90                           kmp_internal_control_t *new_icvs, ident_t *loc);
91 
92 #ifdef USE_LOAD_BALANCE
93 static int __kmp_load_balance_nproc(kmp_root_t *root, int set_nproc);
94 #endif
95 
96 static int __kmp_expand_threads(int nNeed);
97 #if KMP_OS_WINDOWS
98 static int __kmp_unregister_root_other_thread(int gtid);
99 #endif
100 static void __kmp_unregister_library(void); // called by __kmp_internal_end()
101 static void __kmp_reap_thread(kmp_info_t *thread, int is_root);
102 kmp_info_t *__kmp_thread_pool_insert_pt = NULL;
103 
104 /* Calculate the identifier of the current thread */
105 /* fast (and somewhat portable) way to get unique identifier of executing
106    thread. Returns KMP_GTID_DNE if we haven't been assigned a gtid. */
107 int __kmp_get_global_thread_id() {
108   int i;
109   kmp_info_t **other_threads;
110   size_t stack_data;
111   char *stack_addr;
112   size_t stack_size;
113   char *stack_base;
114 
115   KA_TRACE(
116       1000,
117       ("*** __kmp_get_global_thread_id: entering, nproc=%d  all_nproc=%d\n",
118        __kmp_nth, __kmp_all_nth));
119 
120   /* JPH - to handle the case where __kmpc_end(0) is called immediately prior to
121      a parallel region, made it return KMP_GTID_DNE to force serial_initialize
122      by caller. Had to handle KMP_GTID_DNE at all call-sites, or else guarantee
123      __kmp_init_gtid for this to work. */
124 
125   if (!TCR_4(__kmp_init_gtid))
126     return KMP_GTID_DNE;
127 
128 #ifdef KMP_TDATA_GTID
129   if (TCR_4(__kmp_gtid_mode) >= 3) {
130     KA_TRACE(1000, ("*** __kmp_get_global_thread_id: using TDATA\n"));
131     return __kmp_gtid;
132   }
133 #endif
134   if (TCR_4(__kmp_gtid_mode) >= 2) {
135     KA_TRACE(1000, ("*** __kmp_get_global_thread_id: using keyed TLS\n"));
136     return __kmp_gtid_get_specific();
137   }
138   KA_TRACE(1000, ("*** __kmp_get_global_thread_id: using internal alg.\n"));
139 
140   stack_addr = (char *)&stack_data;
141   other_threads = __kmp_threads;
142 
143   /* ATT: The code below is a source of potential bugs due to unsynchronized
144      access to __kmp_threads array. For example:
145      1. Current thread loads other_threads[i] to thr and checks it, it is
146         non-NULL.
147      2. Current thread is suspended by OS.
148      3. Another thread unregisters and finishes (debug versions of free()
149         may fill memory with something like 0xEF).
150      4. Current thread is resumed.
151      5. Current thread reads junk from *thr.
152      TODO: Fix it.  --ln  */
153 
154   for (i = 0; i < __kmp_threads_capacity; i++) {
155 
156     kmp_info_t *thr = (kmp_info_t *)TCR_SYNC_PTR(other_threads[i]);
157     if (!thr)
158       continue;
159 
160     stack_size = (size_t)TCR_PTR(thr->th.th_info.ds.ds_stacksize);
161     stack_base = (char *)TCR_PTR(thr->th.th_info.ds.ds_stackbase);
162 
163     /* stack grows down -- search through all of the active threads */
164 
165     if (stack_addr <= stack_base) {
166       size_t stack_diff = stack_base - stack_addr;
167 
168       if (stack_diff <= stack_size) {
169         /* The only way we can be closer than the allocated */
170         /* stack size is if we are running on this thread. */
171         KMP_DEBUG_ASSERT(__kmp_gtid_get_specific() == i);
172         return i;
173       }
174     }
175   }
176 
177   /* get specific to try and determine our gtid */
178   KA_TRACE(1000,
179            ("*** __kmp_get_global_thread_id: internal alg. failed to find "
180             "thread, using TLS\n"));
181   i = __kmp_gtid_get_specific();
182 
183   /*fprintf( stderr, "=== %d\n", i );  */ /* GROO */
184 
185   /* if we havn't been assigned a gtid, then return code */
186   if (i < 0)
187     return i;
188 
189   /* dynamically updated stack window for uber threads to avoid get_specific
190      call */
191   if (!TCR_4(other_threads[i]->th.th_info.ds.ds_stackgrow)) {
192     KMP_FATAL(StackOverflow, i);
193   }
194 
195   stack_base = (char *)other_threads[i]->th.th_info.ds.ds_stackbase;
196   if (stack_addr > stack_base) {
197     TCW_PTR(other_threads[i]->th.th_info.ds.ds_stackbase, stack_addr);
198     TCW_PTR(other_threads[i]->th.th_info.ds.ds_stacksize,
199             other_threads[i]->th.th_info.ds.ds_stacksize + stack_addr -
200                 stack_base);
201   } else {
202     TCW_PTR(other_threads[i]->th.th_info.ds.ds_stacksize,
203             stack_base - stack_addr);
204   }
205 
206   /* Reprint stack bounds for ubermaster since they have been refined */
207   if (__kmp_storage_map) {
208     char *stack_end = (char *)other_threads[i]->th.th_info.ds.ds_stackbase;
209     char *stack_beg = stack_end - other_threads[i]->th.th_info.ds.ds_stacksize;
210     __kmp_print_storage_map_gtid(i, stack_beg, stack_end,
211                                  other_threads[i]->th.th_info.ds.ds_stacksize,
212                                  "th_%d stack (refinement)", i);
213   }
214   return i;
215 }
216 
217 int __kmp_get_global_thread_id_reg() {
218   int gtid;
219 
220   if (!__kmp_init_serial) {
221     gtid = KMP_GTID_DNE;
222   } else
223 #ifdef KMP_TDATA_GTID
224       if (TCR_4(__kmp_gtid_mode) >= 3) {
225     KA_TRACE(1000, ("*** __kmp_get_global_thread_id_reg: using TDATA\n"));
226     gtid = __kmp_gtid;
227   } else
228 #endif
229       if (TCR_4(__kmp_gtid_mode) >= 2) {
230     KA_TRACE(1000, ("*** __kmp_get_global_thread_id_reg: using keyed TLS\n"));
231     gtid = __kmp_gtid_get_specific();
232   } else {
233     KA_TRACE(1000,
234              ("*** __kmp_get_global_thread_id_reg: using internal alg.\n"));
235     gtid = __kmp_get_global_thread_id();
236   }
237 
238   /* we must be a new uber master sibling thread */
239   if (gtid == KMP_GTID_DNE) {
240     KA_TRACE(10,
241              ("__kmp_get_global_thread_id_reg: Encountered new root thread. "
242               "Registering a new gtid.\n"));
243     __kmp_acquire_bootstrap_lock(&__kmp_initz_lock);
244     if (!__kmp_init_serial) {
245       __kmp_do_serial_initialize();
246       gtid = __kmp_gtid_get_specific();
247     } else {
248       gtid = __kmp_register_root(FALSE);
249     }
250     __kmp_release_bootstrap_lock(&__kmp_initz_lock);
251     /*__kmp_printf( "+++ %d\n", gtid ); */ /* GROO */
252   }
253 
254   KMP_DEBUG_ASSERT(gtid >= 0);
255 
256   return gtid;
257 }
258 
259 /* caller must hold forkjoin_lock */
260 void __kmp_check_stack_overlap(kmp_info_t *th) {
261   int f;
262   char *stack_beg = NULL;
263   char *stack_end = NULL;
264   int gtid;
265 
266   KA_TRACE(10, ("__kmp_check_stack_overlap: called\n"));
267   if (__kmp_storage_map) {
268     stack_end = (char *)th->th.th_info.ds.ds_stackbase;
269     stack_beg = stack_end - th->th.th_info.ds.ds_stacksize;
270 
271     gtid = __kmp_gtid_from_thread(th);
272 
273     if (gtid == KMP_GTID_MONITOR) {
274       __kmp_print_storage_map_gtid(
275           gtid, stack_beg, stack_end, th->th.th_info.ds.ds_stacksize,
276           "th_%s stack (%s)", "mon",
277           (th->th.th_info.ds.ds_stackgrow) ? "initial" : "actual");
278     } else {
279       __kmp_print_storage_map_gtid(
280           gtid, stack_beg, stack_end, th->th.th_info.ds.ds_stacksize,
281           "th_%d stack (%s)", gtid,
282           (th->th.th_info.ds.ds_stackgrow) ? "initial" : "actual");
283     }
284   }
285 
286   /* No point in checking ubermaster threads since they use refinement and
287    * cannot overlap */
288   gtid = __kmp_gtid_from_thread(th);
289   if (__kmp_env_checks == TRUE && !KMP_UBER_GTID(gtid)) {
290     KA_TRACE(10,
291              ("__kmp_check_stack_overlap: performing extensive checking\n"));
292     if (stack_beg == NULL) {
293       stack_end = (char *)th->th.th_info.ds.ds_stackbase;
294       stack_beg = stack_end - th->th.th_info.ds.ds_stacksize;
295     }
296 
297     for (f = 0; f < __kmp_threads_capacity; f++) {
298       kmp_info_t *f_th = (kmp_info_t *)TCR_SYNC_PTR(__kmp_threads[f]);
299 
300       if (f_th && f_th != th) {
301         char *other_stack_end =
302             (char *)TCR_PTR(f_th->th.th_info.ds.ds_stackbase);
303         char *other_stack_beg =
304             other_stack_end - (size_t)TCR_PTR(f_th->th.th_info.ds.ds_stacksize);
305         if ((stack_beg > other_stack_beg && stack_beg < other_stack_end) ||
306             (stack_end > other_stack_beg && stack_end < other_stack_end)) {
307 
308           /* Print the other stack values before the abort */
309           if (__kmp_storage_map)
310             __kmp_print_storage_map_gtid(
311                 -1, other_stack_beg, other_stack_end,
312                 (size_t)TCR_PTR(f_th->th.th_info.ds.ds_stacksize),
313                 "th_%d stack (overlapped)", __kmp_gtid_from_thread(f_th));
314 
315           __kmp_fatal(KMP_MSG(StackOverlap), KMP_HNT(ChangeStackLimit),
316                       __kmp_msg_null);
317         }
318       }
319     }
320   }
321   KA_TRACE(10, ("__kmp_check_stack_overlap: returning\n"));
322 }
323 
324 /* ------------------------------------------------------------------------ */
325 
326 void __kmp_infinite_loop(void) {
327   static int done = FALSE;
328 
329   while (!done) {
330     KMP_YIELD(TRUE);
331   }
332 }
333 
334 #define MAX_MESSAGE 512
335 
336 void __kmp_print_storage_map_gtid(int gtid, void *p1, void *p2, size_t size,
337                                   char const *format, ...) {
338   char buffer[MAX_MESSAGE];
339   va_list ap;
340 
341   va_start(ap, format);
342   KMP_SNPRINTF(buffer, sizeof(buffer), "OMP storage map: %p %p%8lu %s\n", p1,
343                p2, (unsigned long)size, format);
344   __kmp_acquire_bootstrap_lock(&__kmp_stdio_lock);
345   __kmp_vprintf(kmp_err, buffer, ap);
346 #if KMP_PRINT_DATA_PLACEMENT
347   int node;
348   if (gtid >= 0) {
349     if (p1 <= p2 && (char *)p2 - (char *)p1 == size) {
350       if (__kmp_storage_map_verbose) {
351         node = __kmp_get_host_node(p1);
352         if (node < 0) /* doesn't work, so don't try this next time */
353           __kmp_storage_map_verbose = FALSE;
354         else {
355           char *last;
356           int lastNode;
357           int localProc = __kmp_get_cpu_from_gtid(gtid);
358 
359           const int page_size = KMP_GET_PAGE_SIZE();
360 
361           p1 = (void *)((size_t)p1 & ~((size_t)page_size - 1));
362           p2 = (void *)(((size_t)p2 - 1) & ~((size_t)page_size - 1));
363           if (localProc >= 0)
364             __kmp_printf_no_lock("  GTID %d localNode %d\n", gtid,
365                                  localProc >> 1);
366           else
367             __kmp_printf_no_lock("  GTID %d\n", gtid);
368 #if KMP_USE_PRCTL
369           /* The more elaborate format is disabled for now because of the prctl
370            * hanging bug. */
371           do {
372             last = p1;
373             lastNode = node;
374             /* This loop collates adjacent pages with the same host node. */
375             do {
376               (char *)p1 += page_size;
377             } while (p1 <= p2 && (node = __kmp_get_host_node(p1)) == lastNode);
378             __kmp_printf_no_lock("    %p-%p memNode %d\n", last, (char *)p1 - 1,
379                                  lastNode);
380           } while (p1 <= p2);
381 #else
382           __kmp_printf_no_lock("    %p-%p memNode %d\n", p1,
383                                (char *)p1 + (page_size - 1),
384                                __kmp_get_host_node(p1));
385           if (p1 < p2) {
386             __kmp_printf_no_lock("    %p-%p memNode %d\n", p2,
387                                  (char *)p2 + (page_size - 1),
388                                  __kmp_get_host_node(p2));
389           }
390 #endif
391         }
392       }
393     } else
394       __kmp_printf_no_lock("  %s\n", KMP_I18N_STR(StorageMapWarning));
395   }
396 #endif /* KMP_PRINT_DATA_PLACEMENT */
397   __kmp_release_bootstrap_lock(&__kmp_stdio_lock);
398 }
399 
400 void __kmp_warn(char const *format, ...) {
401   char buffer[MAX_MESSAGE];
402   va_list ap;
403 
404   if (__kmp_generate_warnings == kmp_warnings_off) {
405     return;
406   }
407 
408   va_start(ap, format);
409 
410   KMP_SNPRINTF(buffer, sizeof(buffer), "OMP warning: %s\n", format);
411   __kmp_acquire_bootstrap_lock(&__kmp_stdio_lock);
412   __kmp_vprintf(kmp_err, buffer, ap);
413   __kmp_release_bootstrap_lock(&__kmp_stdio_lock);
414 
415   va_end(ap);
416 }
417 
418 void __kmp_abort_process() {
419   // Later threads may stall here, but that's ok because abort() will kill them.
420   __kmp_acquire_bootstrap_lock(&__kmp_exit_lock);
421 
422   if (__kmp_debug_buf) {
423     __kmp_dump_debug_buffer();
424   }
425 
426   if (KMP_OS_WINDOWS) {
427     // Let other threads know of abnormal termination and prevent deadlock
428     // if abort happened during library initialization or shutdown
429     __kmp_global.g.g_abort = SIGABRT;
430 
431     /* On Windows* OS by default abort() causes pop-up error box, which stalls
432        nightly testing. Unfortunately, we cannot reliably suppress pop-up error
433        boxes. _set_abort_behavior() works well, but this function is not
434        available in VS7 (this is not problem for DLL, but it is a problem for
435        static OpenMP RTL). SetErrorMode (and so, timelimit utility) does not
436        help, at least in some versions of MS C RTL.
437 
438        It seems following sequence is the only way to simulate abort() and
439        avoid pop-up error box. */
440     raise(SIGABRT);
441     _exit(3); // Just in case, if signal ignored, exit anyway.
442   } else {
443     abort();
444   }
445 
446   __kmp_infinite_loop();
447   __kmp_release_bootstrap_lock(&__kmp_exit_lock);
448 
449 } // __kmp_abort_process
450 
451 void __kmp_abort_thread(void) {
452   // TODO: Eliminate g_abort global variable and this function.
453   // In case of abort just call abort(), it will kill all the threads.
454   __kmp_infinite_loop();
455 } // __kmp_abort_thread
456 
457 /* Print out the storage map for the major kmp_info_t thread data structures
458    that are allocated together. */
459 
460 static void __kmp_print_thread_storage_map(kmp_info_t *thr, int gtid) {
461   __kmp_print_storage_map_gtid(gtid, thr, thr + 1, sizeof(kmp_info_t), "th_%d",
462                                gtid);
463 
464   __kmp_print_storage_map_gtid(gtid, &thr->th.th_info, &thr->th.th_team,
465                                sizeof(kmp_desc_t), "th_%d.th_info", gtid);
466 
467   __kmp_print_storage_map_gtid(gtid, &thr->th.th_local, &thr->th.th_pri_head,
468                                sizeof(kmp_local_t), "th_%d.th_local", gtid);
469 
470   __kmp_print_storage_map_gtid(
471       gtid, &thr->th.th_bar[0], &thr->th.th_bar[bs_last_barrier],
472       sizeof(kmp_balign_t) * bs_last_barrier, "th_%d.th_bar", gtid);
473 
474   __kmp_print_storage_map_gtid(gtid, &thr->th.th_bar[bs_plain_barrier],
475                                &thr->th.th_bar[bs_plain_barrier + 1],
476                                sizeof(kmp_balign_t), "th_%d.th_bar[plain]",
477                                gtid);
478 
479   __kmp_print_storage_map_gtid(gtid, &thr->th.th_bar[bs_forkjoin_barrier],
480                                &thr->th.th_bar[bs_forkjoin_barrier + 1],
481                                sizeof(kmp_balign_t), "th_%d.th_bar[forkjoin]",
482                                gtid);
483 
484 #if KMP_FAST_REDUCTION_BARRIER
485   __kmp_print_storage_map_gtid(gtid, &thr->th.th_bar[bs_reduction_barrier],
486                                &thr->th.th_bar[bs_reduction_barrier + 1],
487                                sizeof(kmp_balign_t), "th_%d.th_bar[reduction]",
488                                gtid);
489 #endif // KMP_FAST_REDUCTION_BARRIER
490 }
491 
492 /* Print out the storage map for the major kmp_team_t team data structures
493    that are allocated together. */
494 
495 static void __kmp_print_team_storage_map(const char *header, kmp_team_t *team,
496                                          int team_id, int num_thr) {
497   int num_disp_buff = team->t.t_max_nproc > 1 ? __kmp_dispatch_num_buffers : 2;
498   __kmp_print_storage_map_gtid(-1, team, team + 1, sizeof(kmp_team_t), "%s_%d",
499                                header, team_id);
500 
501   __kmp_print_storage_map_gtid(-1, &team->t.t_bar[0],
502                                &team->t.t_bar[bs_last_barrier],
503                                sizeof(kmp_balign_team_t) * bs_last_barrier,
504                                "%s_%d.t_bar", header, team_id);
505 
506   __kmp_print_storage_map_gtid(-1, &team->t.t_bar[bs_plain_barrier],
507                                &team->t.t_bar[bs_plain_barrier + 1],
508                                sizeof(kmp_balign_team_t), "%s_%d.t_bar[plain]",
509                                header, team_id);
510 
511   __kmp_print_storage_map_gtid(-1, &team->t.t_bar[bs_forkjoin_barrier],
512                                &team->t.t_bar[bs_forkjoin_barrier + 1],
513                                sizeof(kmp_balign_team_t),
514                                "%s_%d.t_bar[forkjoin]", header, team_id);
515 
516 #if KMP_FAST_REDUCTION_BARRIER
517   __kmp_print_storage_map_gtid(-1, &team->t.t_bar[bs_reduction_barrier],
518                                &team->t.t_bar[bs_reduction_barrier + 1],
519                                sizeof(kmp_balign_team_t),
520                                "%s_%d.t_bar[reduction]", header, team_id);
521 #endif // KMP_FAST_REDUCTION_BARRIER
522 
523   __kmp_print_storage_map_gtid(
524       -1, &team->t.t_dispatch[0], &team->t.t_dispatch[num_thr],
525       sizeof(kmp_disp_t) * num_thr, "%s_%d.t_dispatch", header, team_id);
526 
527   __kmp_print_storage_map_gtid(
528       -1, &team->t.t_threads[0], &team->t.t_threads[num_thr],
529       sizeof(kmp_info_t *) * num_thr, "%s_%d.t_threads", header, team_id);
530 
531   __kmp_print_storage_map_gtid(-1, &team->t.t_disp_buffer[0],
532                                &team->t.t_disp_buffer[num_disp_buff],
533                                sizeof(dispatch_shared_info_t) * num_disp_buff,
534                                "%s_%d.t_disp_buffer", header, team_id);
535 }
536 
537 static void __kmp_init_allocator() {
538 #if OMP_50_ENABLED
539   __kmp_init_memkind();
540 #endif
541 }
542 static void __kmp_fini_allocator() {
543 #if OMP_50_ENABLED
544   __kmp_fini_memkind();
545 #endif
546 }
547 
548 /* ------------------------------------------------------------------------ */
549 
550 #if KMP_DYNAMIC_LIB
551 #if KMP_OS_WINDOWS
552 
553 static void __kmp_reset_lock(kmp_bootstrap_lock_t *lck) {
554   // TODO: Change to __kmp_break_bootstrap_lock().
555   __kmp_init_bootstrap_lock(lck); // make the lock released
556 }
557 
558 static void __kmp_reset_locks_on_process_detach(int gtid_req) {
559   int i;
560   int thread_count;
561 
562   // PROCESS_DETACH is expected to be called by a thread that executes
563   // ProcessExit() or FreeLibrary(). OS terminates other threads (except the one
564   // calling ProcessExit or FreeLibrary). So, it might be safe to access the
565   // __kmp_threads[] without taking the forkjoin_lock. However, in fact, some
566   // threads can be still alive here, although being about to be terminated. The
567   // threads in the array with ds_thread==0 are most suspicious. Actually, it
568   // can be not safe to access the __kmp_threads[].
569 
570   // TODO: does it make sense to check __kmp_roots[] ?
571 
572   // Let's check that there are no other alive threads registered with the OMP
573   // lib.
574   while (1) {
575     thread_count = 0;
576     for (i = 0; i < __kmp_threads_capacity; ++i) {
577       if (!__kmp_threads)
578         continue;
579       kmp_info_t *th = __kmp_threads[i];
580       if (th == NULL)
581         continue;
582       int gtid = th->th.th_info.ds.ds_gtid;
583       if (gtid == gtid_req)
584         continue;
585       if (gtid < 0)
586         continue;
587       DWORD exit_val;
588       int alive = __kmp_is_thread_alive(th, &exit_val);
589       if (alive) {
590         ++thread_count;
591       }
592     }
593     if (thread_count == 0)
594       break; // success
595   }
596 
597   // Assume that I'm alone. Now it might be safe to check and reset locks.
598   // __kmp_forkjoin_lock and __kmp_stdio_lock are expected to be reset.
599   __kmp_reset_lock(&__kmp_forkjoin_lock);
600 #ifdef KMP_DEBUG
601   __kmp_reset_lock(&__kmp_stdio_lock);
602 #endif // KMP_DEBUG
603 }
604 
605 BOOL WINAPI DllMain(HINSTANCE hInstDLL, DWORD fdwReason, LPVOID lpReserved) {
606   //__kmp_acquire_bootstrap_lock( &__kmp_initz_lock );
607 
608   switch (fdwReason) {
609 
610   case DLL_PROCESS_ATTACH:
611     KA_TRACE(10, ("DllMain: PROCESS_ATTACH\n"));
612 
613     return TRUE;
614 
615   case DLL_PROCESS_DETACH:
616     KA_TRACE(10, ("DllMain: PROCESS_DETACH T#%d\n", __kmp_gtid_get_specific()));
617 
618     if (lpReserved != NULL) {
619       // lpReserved is used for telling the difference:
620       //   lpReserved == NULL when FreeLibrary() was called,
621       //   lpReserved != NULL when the process terminates.
622       // When FreeLibrary() is called, worker threads remain alive. So they will
623       // release the forkjoin lock by themselves. When the process terminates,
624       // worker threads disappear triggering the problem of unreleased forkjoin
625       // lock as described below.
626 
627       // A worker thread can take the forkjoin lock. The problem comes up if
628       // that worker thread becomes dead before it releases the forkjoin lock.
629       // The forkjoin lock remains taken, while the thread executing
630       // DllMain()->PROCESS_DETACH->__kmp_internal_end_library() below will try
631       // to take the forkjoin lock and will always fail, so that the application
632       // will never finish [normally]. This scenario is possible if
633       // __kmpc_end() has not been executed. It looks like it's not a corner
634       // case, but common cases:
635       // - the main function was compiled by an alternative compiler;
636       // - the main function was compiled by icl but without /Qopenmp
637       //   (application with plugins);
638       // - application terminates by calling C exit(), Fortran CALL EXIT() or
639       //   Fortran STOP.
640       // - alive foreign thread prevented __kmpc_end from doing cleanup.
641       //
642       // This is a hack to work around the problem.
643       // TODO: !!! figure out something better.
644       __kmp_reset_locks_on_process_detach(__kmp_gtid_get_specific());
645     }
646 
647     __kmp_internal_end_library(__kmp_gtid_get_specific());
648 
649     return TRUE;
650 
651   case DLL_THREAD_ATTACH:
652     KA_TRACE(10, ("DllMain: THREAD_ATTACH\n"));
653 
654     /* if we want to register new siblings all the time here call
655      * __kmp_get_gtid(); */
656     return TRUE;
657 
658   case DLL_THREAD_DETACH:
659     KA_TRACE(10, ("DllMain: THREAD_DETACH T#%d\n", __kmp_gtid_get_specific()));
660 
661     __kmp_internal_end_thread(__kmp_gtid_get_specific());
662     return TRUE;
663   }
664 
665   return TRUE;
666 }
667 
668 #endif /* KMP_OS_WINDOWS */
669 #endif /* KMP_DYNAMIC_LIB */
670 
671 /* __kmp_parallel_deo -- Wait until it's our turn. */
672 void __kmp_parallel_deo(int *gtid_ref, int *cid_ref, ident_t *loc_ref) {
673   int gtid = *gtid_ref;
674 #ifdef BUILD_PARALLEL_ORDERED
675   kmp_team_t *team = __kmp_team_from_gtid(gtid);
676 #endif /* BUILD_PARALLEL_ORDERED */
677 
678   if (__kmp_env_consistency_check) {
679     if (__kmp_threads[gtid]->th.th_root->r.r_active)
680 #if KMP_USE_DYNAMIC_LOCK
681       __kmp_push_sync(gtid, ct_ordered_in_parallel, loc_ref, NULL, 0);
682 #else
683       __kmp_push_sync(gtid, ct_ordered_in_parallel, loc_ref, NULL);
684 #endif
685   }
686 #ifdef BUILD_PARALLEL_ORDERED
687   if (!team->t.t_serialized) {
688     KMP_MB();
689     KMP_WAIT(&team->t.t_ordered.dt.t_value, __kmp_tid_from_gtid(gtid), KMP_EQ,
690              NULL);
691     KMP_MB();
692   }
693 #endif /* BUILD_PARALLEL_ORDERED */
694 }
695 
696 /* __kmp_parallel_dxo -- Signal the next task. */
697 void __kmp_parallel_dxo(int *gtid_ref, int *cid_ref, ident_t *loc_ref) {
698   int gtid = *gtid_ref;
699 #ifdef BUILD_PARALLEL_ORDERED
700   int tid = __kmp_tid_from_gtid(gtid);
701   kmp_team_t *team = __kmp_team_from_gtid(gtid);
702 #endif /* BUILD_PARALLEL_ORDERED */
703 
704   if (__kmp_env_consistency_check) {
705     if (__kmp_threads[gtid]->th.th_root->r.r_active)
706       __kmp_pop_sync(gtid, ct_ordered_in_parallel, loc_ref);
707   }
708 #ifdef BUILD_PARALLEL_ORDERED
709   if (!team->t.t_serialized) {
710     KMP_MB(); /* Flush all pending memory write invalidates.  */
711 
712     /* use the tid of the next thread in this team */
713     /* TODO replace with general release procedure */
714     team->t.t_ordered.dt.t_value = ((tid + 1) % team->t.t_nproc);
715 
716     KMP_MB(); /* Flush all pending memory write invalidates.  */
717   }
718 #endif /* BUILD_PARALLEL_ORDERED */
719 }
720 
721 /* ------------------------------------------------------------------------ */
722 /* The BARRIER for a SINGLE process section is always explicit   */
723 
724 int __kmp_enter_single(int gtid, ident_t *id_ref, int push_ws) {
725   int status;
726   kmp_info_t *th;
727   kmp_team_t *team;
728 
729   if (!TCR_4(__kmp_init_parallel))
730     __kmp_parallel_initialize();
731 
732 #if OMP_50_ENABLED
733   __kmp_resume_if_soft_paused();
734 #endif
735 
736   th = __kmp_threads[gtid];
737   team = th->th.th_team;
738   status = 0;
739 
740   th->th.th_ident = id_ref;
741 
742   if (team->t.t_serialized) {
743     status = 1;
744   } else {
745     kmp_int32 old_this = th->th.th_local.this_construct;
746 
747     ++th->th.th_local.this_construct;
748     /* try to set team count to thread count--success means thread got the
749        single block */
750     /* TODO: Should this be acquire or release? */
751     if (team->t.t_construct == old_this) {
752       status = __kmp_atomic_compare_store_acq(&team->t.t_construct, old_this,
753                                               th->th.th_local.this_construct);
754     }
755 #if USE_ITT_BUILD
756     if (__itt_metadata_add_ptr && __kmp_forkjoin_frames_mode == 3 &&
757         KMP_MASTER_GTID(gtid) &&
758 #if OMP_40_ENABLED
759         th->th.th_teams_microtask == NULL &&
760 #endif
761         team->t.t_active_level ==
762             1) { // Only report metadata by master of active team at level 1
763       __kmp_itt_metadata_single(id_ref);
764     }
765 #endif /* USE_ITT_BUILD */
766   }
767 
768   if (__kmp_env_consistency_check) {
769     if (status && push_ws) {
770       __kmp_push_workshare(gtid, ct_psingle, id_ref);
771     } else {
772       __kmp_check_workshare(gtid, ct_psingle, id_ref);
773     }
774   }
775 #if USE_ITT_BUILD
776   if (status) {
777     __kmp_itt_single_start(gtid);
778   }
779 #endif /* USE_ITT_BUILD */
780   return status;
781 }
782 
783 void __kmp_exit_single(int gtid) {
784 #if USE_ITT_BUILD
785   __kmp_itt_single_end(gtid);
786 #endif /* USE_ITT_BUILD */
787   if (__kmp_env_consistency_check)
788     __kmp_pop_workshare(gtid, ct_psingle, NULL);
789 }
790 
791 /* determine if we can go parallel or must use a serialized parallel region and
792  * how many threads we can use
793  * set_nproc is the number of threads requested for the team
794  * returns 0 if we should serialize or only use one thread,
795  * otherwise the number of threads to use
796  * The forkjoin lock is held by the caller. */
797 static int __kmp_reserve_threads(kmp_root_t *root, kmp_team_t *parent_team,
798                                  int master_tid, int set_nthreads
799 #if OMP_40_ENABLED
800                                  ,
801                                  int enter_teams
802 #endif /* OMP_40_ENABLED */
803                                  ) {
804   int capacity;
805   int new_nthreads;
806   KMP_DEBUG_ASSERT(__kmp_init_serial);
807   KMP_DEBUG_ASSERT(root && parent_team);
808   kmp_info_t *this_thr = parent_team->t.t_threads[master_tid];
809 
810   // If dyn-var is set, dynamically adjust the number of desired threads,
811   // according to the method specified by dynamic_mode.
812   new_nthreads = set_nthreads;
813   if (!get__dynamic_2(parent_team, master_tid)) {
814     ;
815   }
816 #ifdef USE_LOAD_BALANCE
817   else if (__kmp_global.g.g_dynamic_mode == dynamic_load_balance) {
818     new_nthreads = __kmp_load_balance_nproc(root, set_nthreads);
819     if (new_nthreads == 1) {
820       KC_TRACE(10, ("__kmp_reserve_threads: T#%d load balance reduced "
821                     "reservation to 1 thread\n",
822                     master_tid));
823       return 1;
824     }
825     if (new_nthreads < set_nthreads) {
826       KC_TRACE(10, ("__kmp_reserve_threads: T#%d load balance reduced "
827                     "reservation to %d threads\n",
828                     master_tid, new_nthreads));
829     }
830   }
831 #endif /* USE_LOAD_BALANCE */
832   else if (__kmp_global.g.g_dynamic_mode == dynamic_thread_limit) {
833     new_nthreads = __kmp_avail_proc - __kmp_nth +
834                    (root->r.r_active ? 1 : root->r.r_hot_team->t.t_nproc);
835     if (new_nthreads <= 1) {
836       KC_TRACE(10, ("__kmp_reserve_threads: T#%d thread limit reduced "
837                     "reservation to 1 thread\n",
838                     master_tid));
839       return 1;
840     }
841     if (new_nthreads < set_nthreads) {
842       KC_TRACE(10, ("__kmp_reserve_threads: T#%d thread limit reduced "
843                     "reservation to %d threads\n",
844                     master_tid, new_nthreads));
845     } else {
846       new_nthreads = set_nthreads;
847     }
848   } else if (__kmp_global.g.g_dynamic_mode == dynamic_random) {
849     if (set_nthreads > 2) {
850       new_nthreads = __kmp_get_random(parent_team->t.t_threads[master_tid]);
851       new_nthreads = (new_nthreads % set_nthreads) + 1;
852       if (new_nthreads == 1) {
853         KC_TRACE(10, ("__kmp_reserve_threads: T#%d dynamic random reduced "
854                       "reservation to 1 thread\n",
855                       master_tid));
856         return 1;
857       }
858       if (new_nthreads < set_nthreads) {
859         KC_TRACE(10, ("__kmp_reserve_threads: T#%d dynamic random reduced "
860                       "reservation to %d threads\n",
861                       master_tid, new_nthreads));
862       }
863     }
864   } else {
865     KMP_ASSERT(0);
866   }
867 
868   // Respect KMP_ALL_THREADS/KMP_DEVICE_THREAD_LIMIT.
869   if (__kmp_nth + new_nthreads -
870           (root->r.r_active ? 1 : root->r.r_hot_team->t.t_nproc) >
871       __kmp_max_nth) {
872     int tl_nthreads = __kmp_max_nth - __kmp_nth +
873                       (root->r.r_active ? 1 : root->r.r_hot_team->t.t_nproc);
874     if (tl_nthreads <= 0) {
875       tl_nthreads = 1;
876     }
877 
878     // If dyn-var is false, emit a 1-time warning.
879     if (!get__dynamic_2(parent_team, master_tid) && (!__kmp_reserve_warn)) {
880       __kmp_reserve_warn = 1;
881       __kmp_msg(kmp_ms_warning,
882                 KMP_MSG(CantFormThrTeam, set_nthreads, tl_nthreads),
883                 KMP_HNT(Unset_ALL_THREADS), __kmp_msg_null);
884     }
885     if (tl_nthreads == 1) {
886       KC_TRACE(10, ("__kmp_reserve_threads: T#%d KMP_DEVICE_THREAD_LIMIT "
887                     "reduced reservation to 1 thread\n",
888                     master_tid));
889       return 1;
890     }
891     KC_TRACE(10, ("__kmp_reserve_threads: T#%d KMP_DEVICE_THREAD_LIMIT reduced "
892                   "reservation to %d threads\n",
893                   master_tid, tl_nthreads));
894     new_nthreads = tl_nthreads;
895   }
896 
897   // Respect OMP_THREAD_LIMIT
898   int cg_nthreads = this_thr->th.th_cg_roots->cg_nthreads;
899   int max_cg_threads = this_thr->th.th_cg_roots->cg_thread_limit;
900   if (cg_nthreads + new_nthreads -
901           (root->r.r_active ? 1 : root->r.r_hot_team->t.t_nproc) >
902       max_cg_threads) {
903     int tl_nthreads = max_cg_threads - cg_nthreads +
904                       (root->r.r_active ? 1 : root->r.r_hot_team->t.t_nproc);
905     if (tl_nthreads <= 0) {
906       tl_nthreads = 1;
907     }
908 
909     // If dyn-var is false, emit a 1-time warning.
910     if (!get__dynamic_2(parent_team, master_tid) && (!__kmp_reserve_warn)) {
911       __kmp_reserve_warn = 1;
912       __kmp_msg(kmp_ms_warning,
913                 KMP_MSG(CantFormThrTeam, set_nthreads, tl_nthreads),
914                 KMP_HNT(Unset_ALL_THREADS), __kmp_msg_null);
915     }
916     if (tl_nthreads == 1) {
917       KC_TRACE(10, ("__kmp_reserve_threads: T#%d OMP_THREAD_LIMIT "
918                     "reduced reservation to 1 thread\n",
919                     master_tid));
920       return 1;
921     }
922     KC_TRACE(10, ("__kmp_reserve_threads: T#%d OMP_THREAD_LIMIT reduced "
923                   "reservation to %d threads\n",
924                   master_tid, tl_nthreads));
925     new_nthreads = tl_nthreads;
926   }
927 
928   // Check if the threads array is large enough, or needs expanding.
929   // See comment in __kmp_register_root() about the adjustment if
930   // __kmp_threads[0] == NULL.
931   capacity = __kmp_threads_capacity;
932   if (TCR_PTR(__kmp_threads[0]) == NULL) {
933     --capacity;
934   }
935   if (__kmp_nth + new_nthreads -
936           (root->r.r_active ? 1 : root->r.r_hot_team->t.t_nproc) >
937       capacity) {
938     // Expand the threads array.
939     int slotsRequired = __kmp_nth + new_nthreads -
940                         (root->r.r_active ? 1 : root->r.r_hot_team->t.t_nproc) -
941                         capacity;
942     int slotsAdded = __kmp_expand_threads(slotsRequired);
943     if (slotsAdded < slotsRequired) {
944       // The threads array was not expanded enough.
945       new_nthreads -= (slotsRequired - slotsAdded);
946       KMP_ASSERT(new_nthreads >= 1);
947 
948       // If dyn-var is false, emit a 1-time warning.
949       if (!get__dynamic_2(parent_team, master_tid) && (!__kmp_reserve_warn)) {
950         __kmp_reserve_warn = 1;
951         if (__kmp_tp_cached) {
952           __kmp_msg(kmp_ms_warning,
953                     KMP_MSG(CantFormThrTeam, set_nthreads, new_nthreads),
954                     KMP_HNT(Set_ALL_THREADPRIVATE, __kmp_tp_capacity),
955                     KMP_HNT(PossibleSystemLimitOnThreads), __kmp_msg_null);
956         } else {
957           __kmp_msg(kmp_ms_warning,
958                     KMP_MSG(CantFormThrTeam, set_nthreads, new_nthreads),
959                     KMP_HNT(SystemLimitOnThreads), __kmp_msg_null);
960         }
961       }
962     }
963   }
964 
965 #ifdef KMP_DEBUG
966   if (new_nthreads == 1) {
967     KC_TRACE(10,
968              ("__kmp_reserve_threads: T#%d serializing team after reclaiming "
969               "dead roots and rechecking; requested %d threads\n",
970               __kmp_get_gtid(), set_nthreads));
971   } else {
972     KC_TRACE(10, ("__kmp_reserve_threads: T#%d allocating %d threads; requested"
973                   " %d threads\n",
974                   __kmp_get_gtid(), new_nthreads, set_nthreads));
975   }
976 #endif // KMP_DEBUG
977   return new_nthreads;
978 }
979 
980 /* Allocate threads from the thread pool and assign them to the new team. We are
981    assured that there are enough threads available, because we checked on that
982    earlier within critical section forkjoin */
983 static void __kmp_fork_team_threads(kmp_root_t *root, kmp_team_t *team,
984                                     kmp_info_t *master_th, int master_gtid) {
985   int i;
986   int use_hot_team;
987 
988   KA_TRACE(10, ("__kmp_fork_team_threads: new_nprocs = %d\n", team->t.t_nproc));
989   KMP_DEBUG_ASSERT(master_gtid == __kmp_get_gtid());
990   KMP_MB();
991 
992   /* first, let's setup the master thread */
993   master_th->th.th_info.ds.ds_tid = 0;
994   master_th->th.th_team = team;
995   master_th->th.th_team_nproc = team->t.t_nproc;
996   master_th->th.th_team_master = master_th;
997   master_th->th.th_team_serialized = FALSE;
998   master_th->th.th_dispatch = &team->t.t_dispatch[0];
999 
1000 /* make sure we are not the optimized hot team */
1001 #if KMP_NESTED_HOT_TEAMS
1002   use_hot_team = 0;
1003   kmp_hot_team_ptr_t *hot_teams = master_th->th.th_hot_teams;
1004   if (hot_teams) { // hot teams array is not allocated if
1005     // KMP_HOT_TEAMS_MAX_LEVEL=0
1006     int level = team->t.t_active_level - 1; // index in array of hot teams
1007     if (master_th->th.th_teams_microtask) { // are we inside the teams?
1008       if (master_th->th.th_teams_size.nteams > 1) {
1009         ++level; // level was not increased in teams construct for
1010         // team_of_masters
1011       }
1012       if (team->t.t_pkfn != (microtask_t)__kmp_teams_master &&
1013           master_th->th.th_teams_level == team->t.t_level) {
1014         ++level; // level was not increased in teams construct for
1015         // team_of_workers before the parallel
1016       } // team->t.t_level will be increased inside parallel
1017     }
1018     if (level < __kmp_hot_teams_max_level) {
1019       if (hot_teams[level].hot_team) {
1020         // hot team has already been allocated for given level
1021         KMP_DEBUG_ASSERT(hot_teams[level].hot_team == team);
1022         use_hot_team = 1; // the team is ready to use
1023       } else {
1024         use_hot_team = 0; // AC: threads are not allocated yet
1025         hot_teams[level].hot_team = team; // remember new hot team
1026         hot_teams[level].hot_team_nth = team->t.t_nproc;
1027       }
1028     } else {
1029       use_hot_team = 0;
1030     }
1031   }
1032 #else
1033   use_hot_team = team == root->r.r_hot_team;
1034 #endif
1035   if (!use_hot_team) {
1036 
1037     /* install the master thread */
1038     team->t.t_threads[0] = master_th;
1039     __kmp_initialize_info(master_th, team, 0, master_gtid);
1040 
1041     /* now, install the worker threads */
1042     for (i = 1; i < team->t.t_nproc; i++) {
1043 
1044       /* fork or reallocate a new thread and install it in team */
1045       kmp_info_t *thr = __kmp_allocate_thread(root, team, i);
1046       team->t.t_threads[i] = thr;
1047       KMP_DEBUG_ASSERT(thr);
1048       KMP_DEBUG_ASSERT(thr->th.th_team == team);
1049       /* align team and thread arrived states */
1050       KA_TRACE(20, ("__kmp_fork_team_threads: T#%d(%d:%d) init arrived "
1051                     "T#%d(%d:%d) join =%llu, plain=%llu\n",
1052                     __kmp_gtid_from_tid(0, team), team->t.t_id, 0,
1053                     __kmp_gtid_from_tid(i, team), team->t.t_id, i,
1054                     team->t.t_bar[bs_forkjoin_barrier].b_arrived,
1055                     team->t.t_bar[bs_plain_barrier].b_arrived));
1056 #if OMP_40_ENABLED
1057       thr->th.th_teams_microtask = master_th->th.th_teams_microtask;
1058       thr->th.th_teams_level = master_th->th.th_teams_level;
1059       thr->th.th_teams_size = master_th->th.th_teams_size;
1060 #endif
1061       { // Initialize threads' barrier data.
1062         int b;
1063         kmp_balign_t *balign = team->t.t_threads[i]->th.th_bar;
1064         for (b = 0; b < bs_last_barrier; ++b) {
1065           balign[b].bb.b_arrived = team->t.t_bar[b].b_arrived;
1066           KMP_DEBUG_ASSERT(balign[b].bb.wait_flag != KMP_BARRIER_PARENT_FLAG);
1067 #if USE_DEBUGGER
1068           balign[b].bb.b_worker_arrived = team->t.t_bar[b].b_team_arrived;
1069 #endif
1070         }
1071       }
1072     }
1073 
1074 #if OMP_40_ENABLED && KMP_AFFINITY_SUPPORTED
1075     __kmp_partition_places(team);
1076 #endif
1077   }
1078 
1079 #if OMP_50_ENABLED
1080   if (__kmp_display_affinity && team->t.t_display_affinity != 1) {
1081     for (i = 0; i < team->t.t_nproc; i++) {
1082       kmp_info_t *thr = team->t.t_threads[i];
1083       if (thr->th.th_prev_num_threads != team->t.t_nproc ||
1084           thr->th.th_prev_level != team->t.t_level) {
1085         team->t.t_display_affinity = 1;
1086         break;
1087       }
1088     }
1089   }
1090 #endif
1091 
1092   KMP_MB();
1093 }
1094 
1095 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
1096 // Propagate any changes to the floating point control registers out to the team
1097 // We try to avoid unnecessary writes to the relevant cache line in the team
1098 // structure, so we don't make changes unless they are needed.
1099 inline static void propagateFPControl(kmp_team_t *team) {
1100   if (__kmp_inherit_fp_control) {
1101     kmp_int16 x87_fpu_control_word;
1102     kmp_uint32 mxcsr;
1103 
1104     // Get master values of FPU control flags (both X87 and vector)
1105     __kmp_store_x87_fpu_control_word(&x87_fpu_control_word);
1106     __kmp_store_mxcsr(&mxcsr);
1107     mxcsr &= KMP_X86_MXCSR_MASK;
1108 
1109     // There is no point looking at t_fp_control_saved here.
1110     // If it is TRUE, we still have to update the values if they are different
1111     // from those we now have. If it is FALSE we didn't save anything yet, but
1112     // our objective is the same. We have to ensure that the values in the team
1113     // are the same as those we have.
1114     // So, this code achieves what we need whether or not t_fp_control_saved is
1115     // true. By checking whether the value needs updating we avoid unnecessary
1116     // writes that would put the cache-line into a written state, causing all
1117     // threads in the team to have to read it again.
1118     KMP_CHECK_UPDATE(team->t.t_x87_fpu_control_word, x87_fpu_control_word);
1119     KMP_CHECK_UPDATE(team->t.t_mxcsr, mxcsr);
1120     // Although we don't use this value, other code in the runtime wants to know
1121     // whether it should restore them. So we must ensure it is correct.
1122     KMP_CHECK_UPDATE(team->t.t_fp_control_saved, TRUE);
1123   } else {
1124     // Similarly here. Don't write to this cache-line in the team structure
1125     // unless we have to.
1126     KMP_CHECK_UPDATE(team->t.t_fp_control_saved, FALSE);
1127   }
1128 }
1129 
1130 // Do the opposite, setting the hardware registers to the updated values from
1131 // the team.
1132 inline static void updateHWFPControl(kmp_team_t *team) {
1133   if (__kmp_inherit_fp_control && team->t.t_fp_control_saved) {
1134     // Only reset the fp control regs if they have been changed in the team.
1135     // the parallel region that we are exiting.
1136     kmp_int16 x87_fpu_control_word;
1137     kmp_uint32 mxcsr;
1138     __kmp_store_x87_fpu_control_word(&x87_fpu_control_word);
1139     __kmp_store_mxcsr(&mxcsr);
1140     mxcsr &= KMP_X86_MXCSR_MASK;
1141 
1142     if (team->t.t_x87_fpu_control_word != x87_fpu_control_word) {
1143       __kmp_clear_x87_fpu_status_word();
1144       __kmp_load_x87_fpu_control_word(&team->t.t_x87_fpu_control_word);
1145     }
1146 
1147     if (team->t.t_mxcsr != mxcsr) {
1148       __kmp_load_mxcsr(&team->t.t_mxcsr);
1149     }
1150   }
1151 }
1152 #else
1153 #define propagateFPControl(x) ((void)0)
1154 #define updateHWFPControl(x) ((void)0)
1155 #endif /* KMP_ARCH_X86 || KMP_ARCH_X86_64 */
1156 
1157 static void __kmp_alloc_argv_entries(int argc, kmp_team_t *team,
1158                                      int realloc); // forward declaration
1159 
1160 /* Run a parallel region that has been serialized, so runs only in a team of the
1161    single master thread. */
1162 void __kmp_serialized_parallel(ident_t *loc, kmp_int32 global_tid) {
1163   kmp_info_t *this_thr;
1164   kmp_team_t *serial_team;
1165 
1166   KC_TRACE(10, ("__kmpc_serialized_parallel: called by T#%d\n", global_tid));
1167 
1168   /* Skip all this code for autopar serialized loops since it results in
1169      unacceptable overhead */
1170   if (loc != NULL && (loc->flags & KMP_IDENT_AUTOPAR))
1171     return;
1172 
1173   if (!TCR_4(__kmp_init_parallel))
1174     __kmp_parallel_initialize();
1175 
1176 #if OMP_50_ENABLED
1177   __kmp_resume_if_soft_paused();
1178 #endif
1179 
1180   this_thr = __kmp_threads[global_tid];
1181   serial_team = this_thr->th.th_serial_team;
1182 
1183   /* utilize the serialized team held by this thread */
1184   KMP_DEBUG_ASSERT(serial_team);
1185   KMP_MB();
1186 
1187   if (__kmp_tasking_mode != tskm_immediate_exec) {
1188     KMP_DEBUG_ASSERT(
1189         this_thr->th.th_task_team ==
1190         this_thr->th.th_team->t.t_task_team[this_thr->th.th_task_state]);
1191     KMP_DEBUG_ASSERT(serial_team->t.t_task_team[this_thr->th.th_task_state] ==
1192                      NULL);
1193     KA_TRACE(20, ("__kmpc_serialized_parallel: T#%d pushing task_team %p / "
1194                   "team %p, new task_team = NULL\n",
1195                   global_tid, this_thr->th.th_task_team, this_thr->th.th_team));
1196     this_thr->th.th_task_team = NULL;
1197   }
1198 
1199 #if OMP_40_ENABLED
1200   kmp_proc_bind_t proc_bind = this_thr->th.th_set_proc_bind;
1201   if (this_thr->th.th_current_task->td_icvs.proc_bind == proc_bind_false) {
1202     proc_bind = proc_bind_false;
1203   } else if (proc_bind == proc_bind_default) {
1204     // No proc_bind clause was specified, so use the current value
1205     // of proc-bind-var for this parallel region.
1206     proc_bind = this_thr->th.th_current_task->td_icvs.proc_bind;
1207   }
1208   // Reset for next parallel region
1209   this_thr->th.th_set_proc_bind = proc_bind_default;
1210 #endif /* OMP_40_ENABLED */
1211 
1212 #if OMPT_SUPPORT
1213   ompt_data_t ompt_parallel_data = ompt_data_none;
1214   ompt_data_t *implicit_task_data;
1215   void *codeptr = OMPT_LOAD_RETURN_ADDRESS(global_tid);
1216   if (ompt_enabled.enabled &&
1217       this_thr->th.ompt_thread_info.state != ompt_state_overhead) {
1218 
1219     ompt_task_info_t *parent_task_info;
1220     parent_task_info = OMPT_CUR_TASK_INFO(this_thr);
1221 
1222     parent_task_info->frame.enter_frame.ptr = OMPT_GET_FRAME_ADDRESS(0);
1223     if (ompt_enabled.ompt_callback_parallel_begin) {
1224       int team_size = 1;
1225 
1226       ompt_callbacks.ompt_callback(ompt_callback_parallel_begin)(
1227           &(parent_task_info->task_data), &(parent_task_info->frame),
1228           &ompt_parallel_data, team_size, ompt_parallel_invoker_program,
1229           codeptr);
1230     }
1231   }
1232 #endif // OMPT_SUPPORT
1233 
1234   if (this_thr->th.th_team != serial_team) {
1235     // Nested level will be an index in the nested nthreads array
1236     int level = this_thr->th.th_team->t.t_level;
1237 
1238     if (serial_team->t.t_serialized) {
1239       /* this serial team was already used
1240          TODO increase performance by making this locks more specific */
1241       kmp_team_t *new_team;
1242 
1243       __kmp_acquire_bootstrap_lock(&__kmp_forkjoin_lock);
1244 
1245       new_team = __kmp_allocate_team(this_thr->th.th_root, 1, 1,
1246 #if OMPT_SUPPORT
1247                                      ompt_parallel_data,
1248 #endif
1249 #if OMP_40_ENABLED
1250                                      proc_bind,
1251 #endif
1252                                      &this_thr->th.th_current_task->td_icvs,
1253                                      0 USE_NESTED_HOT_ARG(NULL));
1254       __kmp_release_bootstrap_lock(&__kmp_forkjoin_lock);
1255       KMP_ASSERT(new_team);
1256 
1257       /* setup new serialized team and install it */
1258       new_team->t.t_threads[0] = this_thr;
1259       new_team->t.t_parent = this_thr->th.th_team;
1260       serial_team = new_team;
1261       this_thr->th.th_serial_team = serial_team;
1262 
1263       KF_TRACE(
1264           10,
1265           ("__kmpc_serialized_parallel: T#%d allocated new serial team %p\n",
1266            global_tid, serial_team));
1267 
1268       /* TODO the above breaks the requirement that if we run out of resources,
1269          then we can still guarantee that serialized teams are ok, since we may
1270          need to allocate a new one */
1271     } else {
1272       KF_TRACE(
1273           10,
1274           ("__kmpc_serialized_parallel: T#%d reusing cached serial team %p\n",
1275            global_tid, serial_team));
1276     }
1277 
1278     /* we have to initialize this serial team */
1279     KMP_DEBUG_ASSERT(serial_team->t.t_threads);
1280     KMP_DEBUG_ASSERT(serial_team->t.t_threads[0] == this_thr);
1281     KMP_DEBUG_ASSERT(this_thr->th.th_team != serial_team);
1282     serial_team->t.t_ident = loc;
1283     serial_team->t.t_serialized = 1;
1284     serial_team->t.t_nproc = 1;
1285     serial_team->t.t_parent = this_thr->th.th_team;
1286     serial_team->t.t_sched.sched = this_thr->th.th_team->t.t_sched.sched;
1287     this_thr->th.th_team = serial_team;
1288     serial_team->t.t_master_tid = this_thr->th.th_info.ds.ds_tid;
1289 
1290     KF_TRACE(10, ("__kmpc_serialized_parallel: T#d curtask=%p\n", global_tid,
1291                   this_thr->th.th_current_task));
1292     KMP_ASSERT(this_thr->th.th_current_task->td_flags.executing == 1);
1293     this_thr->th.th_current_task->td_flags.executing = 0;
1294 
1295     __kmp_push_current_task_to_thread(this_thr, serial_team, 0);
1296 
1297     /* TODO: GEH: do ICVs work for nested serialized teams? Don't we need an
1298        implicit task for each serialized task represented by
1299        team->t.t_serialized? */
1300     copy_icvs(&this_thr->th.th_current_task->td_icvs,
1301               &this_thr->th.th_current_task->td_parent->td_icvs);
1302 
1303     // Thread value exists in the nested nthreads array for the next nested
1304     // level
1305     if (__kmp_nested_nth.used && (level + 1 < __kmp_nested_nth.used)) {
1306       this_thr->th.th_current_task->td_icvs.nproc =
1307           __kmp_nested_nth.nth[level + 1];
1308     }
1309 
1310 #if OMP_40_ENABLED
1311     if (__kmp_nested_proc_bind.used &&
1312         (level + 1 < __kmp_nested_proc_bind.used)) {
1313       this_thr->th.th_current_task->td_icvs.proc_bind =
1314           __kmp_nested_proc_bind.bind_types[level + 1];
1315     }
1316 #endif /* OMP_40_ENABLED */
1317 
1318 #if USE_DEBUGGER
1319     serial_team->t.t_pkfn = (microtask_t)(~0); // For the debugger.
1320 #endif
1321     this_thr->th.th_info.ds.ds_tid = 0;
1322 
1323     /* set thread cache values */
1324     this_thr->th.th_team_nproc = 1;
1325     this_thr->th.th_team_master = this_thr;
1326     this_thr->th.th_team_serialized = 1;
1327 
1328     serial_team->t.t_level = serial_team->t.t_parent->t.t_level + 1;
1329     serial_team->t.t_active_level = serial_team->t.t_parent->t.t_active_level;
1330 #if OMP_50_ENABLED
1331     serial_team->t.t_def_allocator = this_thr->th.th_def_allocator; // save
1332 #endif
1333 
1334     propagateFPControl(serial_team);
1335 
1336     /* check if we need to allocate dispatch buffers stack */
1337     KMP_DEBUG_ASSERT(serial_team->t.t_dispatch);
1338     if (!serial_team->t.t_dispatch->th_disp_buffer) {
1339       serial_team->t.t_dispatch->th_disp_buffer =
1340           (dispatch_private_info_t *)__kmp_allocate(
1341               sizeof(dispatch_private_info_t));
1342     }
1343     this_thr->th.th_dispatch = serial_team->t.t_dispatch;
1344 
1345     KMP_MB();
1346 
1347   } else {
1348     /* this serialized team is already being used,
1349      * that's fine, just add another nested level */
1350     KMP_DEBUG_ASSERT(this_thr->th.th_team == serial_team);
1351     KMP_DEBUG_ASSERT(serial_team->t.t_threads);
1352     KMP_DEBUG_ASSERT(serial_team->t.t_threads[0] == this_thr);
1353     ++serial_team->t.t_serialized;
1354     this_thr->th.th_team_serialized = serial_team->t.t_serialized;
1355 
1356     // Nested level will be an index in the nested nthreads array
1357     int level = this_thr->th.th_team->t.t_level;
1358     // Thread value exists in the nested nthreads array for the next nested
1359     // level
1360     if (__kmp_nested_nth.used && (level + 1 < __kmp_nested_nth.used)) {
1361       this_thr->th.th_current_task->td_icvs.nproc =
1362           __kmp_nested_nth.nth[level + 1];
1363     }
1364     serial_team->t.t_level++;
1365     KF_TRACE(10, ("__kmpc_serialized_parallel: T#%d increasing nesting level "
1366                   "of serial team %p to %d\n",
1367                   global_tid, serial_team, serial_team->t.t_level));
1368 
1369     /* allocate/push dispatch buffers stack */
1370     KMP_DEBUG_ASSERT(serial_team->t.t_dispatch);
1371     {
1372       dispatch_private_info_t *disp_buffer =
1373           (dispatch_private_info_t *)__kmp_allocate(
1374               sizeof(dispatch_private_info_t));
1375       disp_buffer->next = serial_team->t.t_dispatch->th_disp_buffer;
1376       serial_team->t.t_dispatch->th_disp_buffer = disp_buffer;
1377     }
1378     this_thr->th.th_dispatch = serial_team->t.t_dispatch;
1379 
1380     KMP_MB();
1381   }
1382 #if OMP_40_ENABLED
1383   KMP_CHECK_UPDATE(serial_team->t.t_cancel_request, cancel_noreq);
1384 #endif
1385 
1386 #if OMP_50_ENABLED
1387   // Perform the display affinity functionality for
1388   // serialized parallel regions
1389   if (__kmp_display_affinity) {
1390     if (this_thr->th.th_prev_level != serial_team->t.t_level ||
1391         this_thr->th.th_prev_num_threads != 1) {
1392       // NULL means use the affinity-format-var ICV
1393       __kmp_aux_display_affinity(global_tid, NULL);
1394       this_thr->th.th_prev_level = serial_team->t.t_level;
1395       this_thr->th.th_prev_num_threads = 1;
1396     }
1397   }
1398 #endif
1399 
1400   if (__kmp_env_consistency_check)
1401     __kmp_push_parallel(global_tid, NULL);
1402 #if OMPT_SUPPORT
1403   serial_team->t.ompt_team_info.master_return_address = codeptr;
1404   if (ompt_enabled.enabled &&
1405       this_thr->th.ompt_thread_info.state != ompt_state_overhead) {
1406     OMPT_CUR_TASK_INFO(this_thr)->frame.exit_frame.ptr = OMPT_GET_FRAME_ADDRESS(0);
1407 
1408     ompt_lw_taskteam_t lw_taskteam;
1409     __ompt_lw_taskteam_init(&lw_taskteam, this_thr, global_tid,
1410                             &ompt_parallel_data, codeptr);
1411 
1412     __ompt_lw_taskteam_link(&lw_taskteam, this_thr, 1);
1413     // don't use lw_taskteam after linking. content was swaped
1414 
1415     /* OMPT implicit task begin */
1416     implicit_task_data = OMPT_CUR_TASK_DATA(this_thr);
1417     if (ompt_enabled.ompt_callback_implicit_task) {
1418       ompt_callbacks.ompt_callback(ompt_callback_implicit_task)(
1419           ompt_scope_begin, OMPT_CUR_TEAM_DATA(this_thr),
1420           OMPT_CUR_TASK_DATA(this_thr), 1, __kmp_tid_from_gtid(global_tid), ompt_task_implicit); // TODO: Can this be ompt_task_initial?
1421       OMPT_CUR_TASK_INFO(this_thr)
1422           ->thread_num = __kmp_tid_from_gtid(global_tid);
1423     }
1424 
1425     /* OMPT state */
1426     this_thr->th.ompt_thread_info.state = ompt_state_work_parallel;
1427     OMPT_CUR_TASK_INFO(this_thr)->frame.exit_frame.ptr = OMPT_GET_FRAME_ADDRESS(0);
1428   }
1429 #endif
1430 }
1431 
1432 /* most of the work for a fork */
1433 /* return true if we really went parallel, false if serialized */
1434 int __kmp_fork_call(ident_t *loc, int gtid,
1435                     enum fork_context_e call_context, // Intel, GNU, ...
1436                     kmp_int32 argc, microtask_t microtask, launch_t invoker,
1437 /* TODO: revert workaround for Intel(R) 64 tracker #96 */
1438 #if (KMP_ARCH_X86_64 || KMP_ARCH_ARM || KMP_ARCH_AARCH64) && KMP_OS_LINUX
1439                     va_list *ap
1440 #else
1441                     va_list ap
1442 #endif
1443                     ) {
1444   void **argv;
1445   int i;
1446   int master_tid;
1447   int master_this_cons;
1448   kmp_team_t *team;
1449   kmp_team_t *parent_team;
1450   kmp_info_t *master_th;
1451   kmp_root_t *root;
1452   int nthreads;
1453   int master_active;
1454   int master_set_numthreads;
1455   int level;
1456 #if OMP_40_ENABLED
1457   int active_level;
1458   int teams_level;
1459 #endif
1460 #if KMP_NESTED_HOT_TEAMS
1461   kmp_hot_team_ptr_t **p_hot_teams;
1462 #endif
1463   { // KMP_TIME_BLOCK
1464     KMP_TIME_DEVELOPER_PARTITIONED_BLOCK(KMP_fork_call);
1465     KMP_COUNT_VALUE(OMP_PARALLEL_args, argc);
1466 
1467     KA_TRACE(20, ("__kmp_fork_call: enter T#%d\n", gtid));
1468     if (__kmp_stkpadding > 0 && __kmp_root[gtid] != NULL) {
1469       /* Some systems prefer the stack for the root thread(s) to start with */
1470       /* some gap from the parent stack to prevent false sharing. */
1471       void *dummy = KMP_ALLOCA(__kmp_stkpadding);
1472       /* These 2 lines below are so this does not get optimized out */
1473       if (__kmp_stkpadding > KMP_MAX_STKPADDING)
1474         __kmp_stkpadding += (short)((kmp_int64)dummy);
1475     }
1476 
1477     /* initialize if needed */
1478     KMP_DEBUG_ASSERT(
1479         __kmp_init_serial); // AC: potentially unsafe, not in sync with shutdown
1480     if (!TCR_4(__kmp_init_parallel))
1481       __kmp_parallel_initialize();
1482 
1483 #if OMP_50_ENABLED
1484     __kmp_resume_if_soft_paused();
1485 #endif
1486 
1487     /* setup current data */
1488     master_th = __kmp_threads[gtid]; // AC: potentially unsafe, not in sync with
1489     // shutdown
1490     parent_team = master_th->th.th_team;
1491     master_tid = master_th->th.th_info.ds.ds_tid;
1492     master_this_cons = master_th->th.th_local.this_construct;
1493     root = master_th->th.th_root;
1494     master_active = root->r.r_active;
1495     master_set_numthreads = master_th->th.th_set_nproc;
1496 
1497 #if OMPT_SUPPORT
1498     ompt_data_t ompt_parallel_data = ompt_data_none;
1499     ompt_data_t *parent_task_data;
1500     ompt_frame_t *ompt_frame;
1501     ompt_data_t *implicit_task_data;
1502     void *return_address = NULL;
1503 
1504     if (ompt_enabled.enabled) {
1505       __ompt_get_task_info_internal(0, NULL, &parent_task_data, &ompt_frame,
1506                                     NULL, NULL);
1507       return_address = OMPT_LOAD_RETURN_ADDRESS(gtid);
1508     }
1509 #endif
1510 
1511     // Nested level will be an index in the nested nthreads array
1512     level = parent_team->t.t_level;
1513     // used to launch non-serial teams even if nested is not allowed
1514     active_level = parent_team->t.t_active_level;
1515 #if OMP_40_ENABLED
1516     // needed to check nesting inside the teams
1517     teams_level = master_th->th.th_teams_level;
1518 #endif
1519 #if KMP_NESTED_HOT_TEAMS
1520     p_hot_teams = &master_th->th.th_hot_teams;
1521     if (*p_hot_teams == NULL && __kmp_hot_teams_max_level > 0) {
1522       *p_hot_teams = (kmp_hot_team_ptr_t *)__kmp_allocate(
1523           sizeof(kmp_hot_team_ptr_t) * __kmp_hot_teams_max_level);
1524       (*p_hot_teams)[0].hot_team = root->r.r_hot_team;
1525       // it is either actual or not needed (when active_level > 0)
1526       (*p_hot_teams)[0].hot_team_nth = 1;
1527     }
1528 #endif
1529 
1530 #if OMPT_SUPPORT
1531     if (ompt_enabled.enabled) {
1532       if (ompt_enabled.ompt_callback_parallel_begin) {
1533         int team_size = master_set_numthreads
1534                             ? master_set_numthreads
1535                             : get__nproc_2(parent_team, master_tid);
1536         ompt_callbacks.ompt_callback(ompt_callback_parallel_begin)(
1537             parent_task_data, ompt_frame, &ompt_parallel_data, team_size,
1538             OMPT_INVOKER(call_context), return_address);
1539       }
1540       master_th->th.ompt_thread_info.state = ompt_state_overhead;
1541     }
1542 #endif
1543 
1544     master_th->th.th_ident = loc;
1545 
1546 #if OMP_40_ENABLED
1547     if (master_th->th.th_teams_microtask && ap &&
1548         microtask != (microtask_t)__kmp_teams_master && level == teams_level) {
1549       // AC: This is start of parallel that is nested inside teams construct.
1550       // The team is actual (hot), all workers are ready at the fork barrier.
1551       // No lock needed to initialize the team a bit, then free workers.
1552       parent_team->t.t_ident = loc;
1553       __kmp_alloc_argv_entries(argc, parent_team, TRUE);
1554       parent_team->t.t_argc = argc;
1555       argv = (void **)parent_team->t.t_argv;
1556       for (i = argc - 1; i >= 0; --i)
1557 /* TODO: revert workaround for Intel(R) 64 tracker #96 */
1558 #if (KMP_ARCH_X86_64 || KMP_ARCH_ARM || KMP_ARCH_AARCH64) && KMP_OS_LINUX
1559         *argv++ = va_arg(*ap, void *);
1560 #else
1561         *argv++ = va_arg(ap, void *);
1562 #endif
1563       // Increment our nested depth levels, but not increase the serialization
1564       if (parent_team == master_th->th.th_serial_team) {
1565         // AC: we are in serialized parallel
1566         __kmpc_serialized_parallel(loc, gtid);
1567         KMP_DEBUG_ASSERT(parent_team->t.t_serialized > 1);
1568         // AC: need this in order enquiry functions work
1569         // correctly, will restore at join time
1570         parent_team->t.t_serialized--;
1571 #if OMPT_SUPPORT
1572         void *dummy;
1573         void **exit_runtime_p;
1574 
1575         ompt_lw_taskteam_t lw_taskteam;
1576 
1577         if (ompt_enabled.enabled) {
1578           __ompt_lw_taskteam_init(&lw_taskteam, master_th, gtid,
1579                                   &ompt_parallel_data, return_address);
1580           exit_runtime_p = &(lw_taskteam.ompt_task_info.frame.exit_frame.ptr);
1581 
1582           __ompt_lw_taskteam_link(&lw_taskteam, master_th, 0);
1583           // don't use lw_taskteam after linking. content was swaped
1584 
1585           /* OMPT implicit task begin */
1586           implicit_task_data = OMPT_CUR_TASK_DATA(master_th);
1587           if (ompt_enabled.ompt_callback_implicit_task) {
1588             ompt_callbacks.ompt_callback(ompt_callback_implicit_task)(
1589                 ompt_scope_begin, OMPT_CUR_TEAM_DATA(master_th),
1590                 implicit_task_data, 1, __kmp_tid_from_gtid(gtid), ompt_task_implicit); // TODO: Can this be ompt_task_initial?
1591             OMPT_CUR_TASK_INFO(master_th)
1592                 ->thread_num = __kmp_tid_from_gtid(gtid);
1593           }
1594 
1595           /* OMPT state */
1596           master_th->th.ompt_thread_info.state = ompt_state_work_parallel;
1597         } else {
1598           exit_runtime_p = &dummy;
1599         }
1600 #endif
1601 
1602         {
1603           KMP_TIME_PARTITIONED_BLOCK(OMP_parallel);
1604           KMP_SET_THREAD_STATE_BLOCK(IMPLICIT_TASK);
1605           __kmp_invoke_microtask(microtask, gtid, 0, argc, parent_team->t.t_argv
1606 #if OMPT_SUPPORT
1607                                  ,
1608                                  exit_runtime_p
1609 #endif
1610                                  );
1611         }
1612 
1613 #if OMPT_SUPPORT
1614         *exit_runtime_p = NULL;
1615         if (ompt_enabled.enabled) {
1616           OMPT_CUR_TASK_INFO(master_th)->frame.exit_frame = ompt_data_none;
1617           if (ompt_enabled.ompt_callback_implicit_task) {
1618             ompt_callbacks.ompt_callback(ompt_callback_implicit_task)(
1619                 ompt_scope_end, NULL, implicit_task_data, 1,
1620                 OMPT_CUR_TASK_INFO(master_th)->thread_num, ompt_task_implicit); // TODO: Can this be ompt_task_initial?
1621           }
1622           __ompt_lw_taskteam_unlink(master_th);
1623 
1624           if (ompt_enabled.ompt_callback_parallel_end) {
1625             ompt_callbacks.ompt_callback(ompt_callback_parallel_end)(
1626                 OMPT_CUR_TEAM_DATA(master_th), OMPT_CUR_TASK_DATA(master_th),
1627                 OMPT_INVOKER(call_context), return_address);
1628           }
1629           master_th->th.ompt_thread_info.state = ompt_state_overhead;
1630         }
1631 #endif
1632         return TRUE;
1633       }
1634 
1635       parent_team->t.t_pkfn = microtask;
1636       parent_team->t.t_invoke = invoker;
1637       KMP_ATOMIC_INC(&root->r.r_in_parallel);
1638       parent_team->t.t_active_level++;
1639       parent_team->t.t_level++;
1640 #if OMP_50_ENABLED
1641       parent_team->t.t_def_allocator = master_th->th.th_def_allocator; // save
1642 #endif
1643 
1644       /* Change number of threads in the team if requested */
1645       if (master_set_numthreads) { // The parallel has num_threads clause
1646         if (master_set_numthreads < master_th->th.th_teams_size.nth) {
1647           // AC: only can reduce number of threads dynamically, can't increase
1648           kmp_info_t **other_threads = parent_team->t.t_threads;
1649           parent_team->t.t_nproc = master_set_numthreads;
1650           for (i = 0; i < master_set_numthreads; ++i) {
1651             other_threads[i]->th.th_team_nproc = master_set_numthreads;
1652           }
1653           // Keep extra threads hot in the team for possible next parallels
1654         }
1655         master_th->th.th_set_nproc = 0;
1656       }
1657 
1658 #if USE_DEBUGGER
1659       if (__kmp_debugging) { // Let debugger override number of threads.
1660         int nth = __kmp_omp_num_threads(loc);
1661         if (nth > 0) { // 0 means debugger doesn't want to change num threads
1662           master_set_numthreads = nth;
1663         }
1664       }
1665 #endif
1666 
1667       KF_TRACE(10, ("__kmp_fork_call: before internal fork: root=%p, team=%p, "
1668                     "master_th=%p, gtid=%d\n",
1669                     root, parent_team, master_th, gtid));
1670       __kmp_internal_fork(loc, gtid, parent_team);
1671       KF_TRACE(10, ("__kmp_fork_call: after internal fork: root=%p, team=%p, "
1672                     "master_th=%p, gtid=%d\n",
1673                     root, parent_team, master_th, gtid));
1674 
1675       /* Invoke microtask for MASTER thread */
1676       KA_TRACE(20, ("__kmp_fork_call: T#%d(%d:0) invoke microtask = %p\n", gtid,
1677                     parent_team->t.t_id, parent_team->t.t_pkfn));
1678 
1679       if (!parent_team->t.t_invoke(gtid)) {
1680         KMP_ASSERT2(0, "cannot invoke microtask for MASTER thread");
1681       }
1682       KA_TRACE(20, ("__kmp_fork_call: T#%d(%d:0) done microtask = %p\n", gtid,
1683                     parent_team->t.t_id, parent_team->t.t_pkfn));
1684       KMP_MB(); /* Flush all pending memory write invalidates.  */
1685 
1686       KA_TRACE(20, ("__kmp_fork_call: parallel exit T#%d\n", gtid));
1687 
1688       return TRUE;
1689     } // Parallel closely nested in teams construct
1690 #endif /* OMP_40_ENABLED */
1691 
1692 #if KMP_DEBUG
1693     if (__kmp_tasking_mode != tskm_immediate_exec) {
1694       KMP_DEBUG_ASSERT(master_th->th.th_task_team ==
1695                        parent_team->t.t_task_team[master_th->th.th_task_state]);
1696     }
1697 #endif
1698 
1699     if (parent_team->t.t_active_level >=
1700         master_th->th.th_current_task->td_icvs.max_active_levels) {
1701       nthreads = 1;
1702     } else {
1703 #if OMP_40_ENABLED
1704       int enter_teams = ((ap == NULL && active_level == 0) ||
1705                          (ap && teams_level > 0 && teams_level == level));
1706 #endif
1707       nthreads =
1708           master_set_numthreads
1709               ? master_set_numthreads
1710               : get__nproc_2(
1711                     parent_team,
1712                     master_tid); // TODO: get nproc directly from current task
1713 
1714       // Check if we need to take forkjoin lock? (no need for serialized
1715       // parallel out of teams construct). This code moved here from
1716       // __kmp_reserve_threads() to speedup nested serialized parallels.
1717       if (nthreads > 1) {
1718         if ((get__max_active_levels(master_th) == 1 && (root->r.r_in_parallel
1719 #if OMP_40_ENABLED
1720                                                         && !enter_teams
1721 #endif /* OMP_40_ENABLED */
1722                                                         )) ||
1723             (__kmp_library == library_serial)) {
1724           KC_TRACE(10, ("__kmp_fork_call: T#%d serializing team; requested %d"
1725                         " threads\n",
1726                         gtid, nthreads));
1727           nthreads = 1;
1728         }
1729       }
1730       if (nthreads > 1) {
1731         /* determine how many new threads we can use */
1732         __kmp_acquire_bootstrap_lock(&__kmp_forkjoin_lock);
1733         nthreads = __kmp_reserve_threads(
1734             root, parent_team, master_tid, nthreads
1735 #if OMP_40_ENABLED
1736             /* AC: If we execute teams from parallel region (on host), then
1737                teams should be created but each can only have 1 thread if
1738                nesting is disabled. If teams called from serial region, then
1739                teams and their threads should be created regardless of the
1740                nesting setting. */
1741             ,
1742             enter_teams
1743 #endif /* OMP_40_ENABLED */
1744             );
1745         if (nthreads == 1) {
1746           // Free lock for single thread execution here; for multi-thread
1747           // execution it will be freed later after team of threads created
1748           // and initialized
1749           __kmp_release_bootstrap_lock(&__kmp_forkjoin_lock);
1750         }
1751       }
1752     }
1753     KMP_DEBUG_ASSERT(nthreads > 0);
1754 
1755     // If we temporarily changed the set number of threads then restore it now
1756     master_th->th.th_set_nproc = 0;
1757 
1758     /* create a serialized parallel region? */
1759     if (nthreads == 1) {
1760 /* josh todo: hypothetical question: what do we do for OS X*? */
1761 #if KMP_OS_LINUX &&                                                            \
1762     (KMP_ARCH_X86 || KMP_ARCH_X86_64 || KMP_ARCH_ARM || KMP_ARCH_AARCH64)
1763       void *args[argc];
1764 #else
1765       void **args = (void **)KMP_ALLOCA(argc * sizeof(void *));
1766 #endif /* KMP_OS_LINUX && ( KMP_ARCH_X86 || KMP_ARCH_X86_64 || KMP_ARCH_ARM || \
1767           KMP_ARCH_AARCH64) */
1768 
1769       KA_TRACE(20,
1770                ("__kmp_fork_call: T#%d serializing parallel region\n", gtid));
1771 
1772       __kmpc_serialized_parallel(loc, gtid);
1773 
1774       if (call_context == fork_context_intel) {
1775         /* TODO this sucks, use the compiler itself to pass args! :) */
1776         master_th->th.th_serial_team->t.t_ident = loc;
1777 #if OMP_40_ENABLED
1778         if (!ap) {
1779           // revert change made in __kmpc_serialized_parallel()
1780           master_th->th.th_serial_team->t.t_level--;
1781 // Get args from parent team for teams construct
1782 
1783 #if OMPT_SUPPORT
1784           void *dummy;
1785           void **exit_runtime_p;
1786           ompt_task_info_t *task_info;
1787 
1788           ompt_lw_taskteam_t lw_taskteam;
1789 
1790           if (ompt_enabled.enabled) {
1791             __ompt_lw_taskteam_init(&lw_taskteam, master_th, gtid,
1792                                     &ompt_parallel_data, return_address);
1793 
1794             __ompt_lw_taskteam_link(&lw_taskteam, master_th, 0);
1795             // don't use lw_taskteam after linking. content was swaped
1796 
1797             task_info = OMPT_CUR_TASK_INFO(master_th);
1798             exit_runtime_p = &(task_info->frame.exit_frame.ptr);
1799             if (ompt_enabled.ompt_callback_implicit_task) {
1800               ompt_callbacks.ompt_callback(ompt_callback_implicit_task)(
1801                   ompt_scope_begin, OMPT_CUR_TEAM_DATA(master_th),
1802                   &(task_info->task_data), 1, __kmp_tid_from_gtid(gtid), ompt_task_implicit); // TODO: Can this be ompt_task_initial?
1803               OMPT_CUR_TASK_INFO(master_th)
1804                   ->thread_num = __kmp_tid_from_gtid(gtid);
1805             }
1806 
1807             /* OMPT state */
1808             master_th->th.ompt_thread_info.state = ompt_state_work_parallel;
1809           } else {
1810             exit_runtime_p = &dummy;
1811           }
1812 #endif
1813 
1814           {
1815             KMP_TIME_PARTITIONED_BLOCK(OMP_parallel);
1816             KMP_SET_THREAD_STATE_BLOCK(IMPLICIT_TASK);
1817             __kmp_invoke_microtask(microtask, gtid, 0, argc,
1818                                    parent_team->t.t_argv
1819 #if OMPT_SUPPORT
1820                                    ,
1821                                    exit_runtime_p
1822 #endif
1823                                    );
1824           }
1825 
1826 #if OMPT_SUPPORT
1827           if (ompt_enabled.enabled) {
1828             exit_runtime_p = NULL;
1829             if (ompt_enabled.ompt_callback_implicit_task) {
1830               ompt_callbacks.ompt_callback(ompt_callback_implicit_task)(
1831                   ompt_scope_end, NULL, &(task_info->task_data), 1,
1832                   OMPT_CUR_TASK_INFO(master_th)->thread_num, ompt_task_implicit); // TODO: Can this be ompt_task_initial?
1833             }
1834 
1835             __ompt_lw_taskteam_unlink(master_th);
1836             if (ompt_enabled.ompt_callback_parallel_end) {
1837               ompt_callbacks.ompt_callback(ompt_callback_parallel_end)(
1838                   OMPT_CUR_TEAM_DATA(master_th), parent_task_data,
1839                   OMPT_INVOKER(call_context), return_address);
1840             }
1841             master_th->th.ompt_thread_info.state = ompt_state_overhead;
1842           }
1843 #endif
1844         } else if (microtask == (microtask_t)__kmp_teams_master) {
1845           KMP_DEBUG_ASSERT(master_th->th.th_team ==
1846                            master_th->th.th_serial_team);
1847           team = master_th->th.th_team;
1848           // team->t.t_pkfn = microtask;
1849           team->t.t_invoke = invoker;
1850           __kmp_alloc_argv_entries(argc, team, TRUE);
1851           team->t.t_argc = argc;
1852           argv = (void **)team->t.t_argv;
1853           if (ap) {
1854             for (i = argc - 1; i >= 0; --i)
1855 // TODO: revert workaround for Intel(R) 64 tracker #96
1856 #if (KMP_ARCH_X86_64 || KMP_ARCH_ARM || KMP_ARCH_AARCH64) && KMP_OS_LINUX
1857               *argv++ = va_arg(*ap, void *);
1858 #else
1859               *argv++ = va_arg(ap, void *);
1860 #endif
1861           } else {
1862             for (i = 0; i < argc; ++i)
1863               // Get args from parent team for teams construct
1864               argv[i] = parent_team->t.t_argv[i];
1865           }
1866           // AC: revert change made in __kmpc_serialized_parallel()
1867           //     because initial code in teams should have level=0
1868           team->t.t_level--;
1869           // AC: call special invoker for outer "parallel" of teams construct
1870           invoker(gtid);
1871         } else {
1872 #endif /* OMP_40_ENABLED */
1873           argv = args;
1874           for (i = argc - 1; i >= 0; --i)
1875 // TODO: revert workaround for Intel(R) 64 tracker #96
1876 #if (KMP_ARCH_X86_64 || KMP_ARCH_ARM || KMP_ARCH_AARCH64) && KMP_OS_LINUX
1877             *argv++ = va_arg(*ap, void *);
1878 #else
1879           *argv++ = va_arg(ap, void *);
1880 #endif
1881           KMP_MB();
1882 
1883 #if OMPT_SUPPORT
1884           void *dummy;
1885           void **exit_runtime_p;
1886           ompt_task_info_t *task_info;
1887 
1888           ompt_lw_taskteam_t lw_taskteam;
1889 
1890           if (ompt_enabled.enabled) {
1891             __ompt_lw_taskteam_init(&lw_taskteam, master_th, gtid,
1892                                     &ompt_parallel_data, return_address);
1893             __ompt_lw_taskteam_link(&lw_taskteam, master_th, 0);
1894             // don't use lw_taskteam after linking. content was swaped
1895             task_info = OMPT_CUR_TASK_INFO(master_th);
1896             exit_runtime_p = &(task_info->frame.exit_frame.ptr);
1897 
1898             /* OMPT implicit task begin */
1899             implicit_task_data = OMPT_CUR_TASK_DATA(master_th);
1900             if (ompt_enabled.ompt_callback_implicit_task) {
1901               ompt_callbacks.ompt_callback(ompt_callback_implicit_task)(
1902                   ompt_scope_begin, OMPT_CUR_TEAM_DATA(master_th),
1903                   implicit_task_data, 1, __kmp_tid_from_gtid(gtid), ompt_task_implicit); // TODO: Can this be ompt_task_initial?
1904               OMPT_CUR_TASK_INFO(master_th)
1905                   ->thread_num = __kmp_tid_from_gtid(gtid);
1906             }
1907 
1908             /* OMPT state */
1909             master_th->th.ompt_thread_info.state = ompt_state_work_parallel;
1910           } else {
1911             exit_runtime_p = &dummy;
1912           }
1913 #endif
1914 
1915           {
1916             KMP_TIME_PARTITIONED_BLOCK(OMP_parallel);
1917             KMP_SET_THREAD_STATE_BLOCK(IMPLICIT_TASK);
1918             __kmp_invoke_microtask(microtask, gtid, 0, argc, args
1919 #if OMPT_SUPPORT
1920                                    ,
1921                                    exit_runtime_p
1922 #endif
1923                                    );
1924           }
1925 
1926 #if OMPT_SUPPORT
1927           if (ompt_enabled.enabled) {
1928             *exit_runtime_p = NULL;
1929             if (ompt_enabled.ompt_callback_implicit_task) {
1930               ompt_callbacks.ompt_callback(ompt_callback_implicit_task)(
1931                   ompt_scope_end, NULL, &(task_info->task_data), 1,
1932                   OMPT_CUR_TASK_INFO(master_th)->thread_num, ompt_task_implicit); // TODO: Can this be ompt_task_initial?
1933             }
1934 
1935             ompt_parallel_data = *OMPT_CUR_TEAM_DATA(master_th);
1936             __ompt_lw_taskteam_unlink(master_th);
1937             if (ompt_enabled.ompt_callback_parallel_end) {
1938               ompt_callbacks.ompt_callback(ompt_callback_parallel_end)(
1939                   &ompt_parallel_data, parent_task_data,
1940                   OMPT_INVOKER(call_context), return_address);
1941             }
1942             master_th->th.ompt_thread_info.state = ompt_state_overhead;
1943           }
1944 #endif
1945 #if OMP_40_ENABLED
1946         }
1947 #endif /* OMP_40_ENABLED */
1948       } else if (call_context == fork_context_gnu) {
1949 #if OMPT_SUPPORT
1950         ompt_lw_taskteam_t lwt;
1951         __ompt_lw_taskteam_init(&lwt, master_th, gtid, &ompt_parallel_data,
1952                                 return_address);
1953 
1954         lwt.ompt_task_info.frame.exit_frame = ompt_data_none;
1955         __ompt_lw_taskteam_link(&lwt, master_th, 1);
1956 // don't use lw_taskteam after linking. content was swaped
1957 #endif
1958 
1959         // we were called from GNU native code
1960         KA_TRACE(20, ("__kmp_fork_call: T#%d serial exit\n", gtid));
1961         return FALSE;
1962       } else {
1963         KMP_ASSERT2(call_context < fork_context_last,
1964                     "__kmp_fork_call: unknown fork_context parameter");
1965       }
1966 
1967       KA_TRACE(20, ("__kmp_fork_call: T#%d serial exit\n", gtid));
1968       KMP_MB();
1969       return FALSE;
1970     } // if (nthreads == 1)
1971 
1972     // GEH: only modify the executing flag in the case when not serialized
1973     //      serialized case is handled in kmpc_serialized_parallel
1974     KF_TRACE(10, ("__kmp_fork_call: parent_team_aclevel=%d, master_th=%p, "
1975                   "curtask=%p, curtask_max_aclevel=%d\n",
1976                   parent_team->t.t_active_level, master_th,
1977                   master_th->th.th_current_task,
1978                   master_th->th.th_current_task->td_icvs.max_active_levels));
1979     // TODO: GEH - cannot do this assertion because root thread not set up as
1980     // executing
1981     // KMP_ASSERT( master_th->th.th_current_task->td_flags.executing == 1 );
1982     master_th->th.th_current_task->td_flags.executing = 0;
1983 
1984 #if OMP_40_ENABLED
1985     if (!master_th->th.th_teams_microtask || level > teams_level)
1986 #endif /* OMP_40_ENABLED */
1987     {
1988       /* Increment our nested depth level */
1989       KMP_ATOMIC_INC(&root->r.r_in_parallel);
1990     }
1991 
1992     // See if we need to make a copy of the ICVs.
1993     int nthreads_icv = master_th->th.th_current_task->td_icvs.nproc;
1994     if ((level + 1 < __kmp_nested_nth.used) &&
1995         (__kmp_nested_nth.nth[level + 1] != nthreads_icv)) {
1996       nthreads_icv = __kmp_nested_nth.nth[level + 1];
1997     } else {
1998       nthreads_icv = 0; // don't update
1999     }
2000 
2001 #if OMP_40_ENABLED
2002     // Figure out the proc_bind_policy for the new team.
2003     kmp_proc_bind_t proc_bind = master_th->th.th_set_proc_bind;
2004     kmp_proc_bind_t proc_bind_icv =
2005         proc_bind_default; // proc_bind_default means don't update
2006     if (master_th->th.th_current_task->td_icvs.proc_bind == proc_bind_false) {
2007       proc_bind = proc_bind_false;
2008     } else {
2009       if (proc_bind == proc_bind_default) {
2010         // No proc_bind clause specified; use current proc-bind-var for this
2011         // parallel region
2012         proc_bind = master_th->th.th_current_task->td_icvs.proc_bind;
2013       }
2014       /* else: The proc_bind policy was specified explicitly on parallel clause.
2015          This overrides proc-bind-var for this parallel region, but does not
2016          change proc-bind-var. */
2017       // Figure the value of proc-bind-var for the child threads.
2018       if ((level + 1 < __kmp_nested_proc_bind.used) &&
2019           (__kmp_nested_proc_bind.bind_types[level + 1] !=
2020            master_th->th.th_current_task->td_icvs.proc_bind)) {
2021         proc_bind_icv = __kmp_nested_proc_bind.bind_types[level + 1];
2022       }
2023     }
2024 
2025     // Reset for next parallel region
2026     master_th->th.th_set_proc_bind = proc_bind_default;
2027 #endif /* OMP_40_ENABLED */
2028 
2029     if ((nthreads_icv > 0)
2030 #if OMP_40_ENABLED
2031         || (proc_bind_icv != proc_bind_default)
2032 #endif /* OMP_40_ENABLED */
2033             ) {
2034       kmp_internal_control_t new_icvs;
2035       copy_icvs(&new_icvs, &master_th->th.th_current_task->td_icvs);
2036       new_icvs.next = NULL;
2037       if (nthreads_icv > 0) {
2038         new_icvs.nproc = nthreads_icv;
2039       }
2040 
2041 #if OMP_40_ENABLED
2042       if (proc_bind_icv != proc_bind_default) {
2043         new_icvs.proc_bind = proc_bind_icv;
2044       }
2045 #endif /* OMP_40_ENABLED */
2046 
2047       /* allocate a new parallel team */
2048       KF_TRACE(10, ("__kmp_fork_call: before __kmp_allocate_team\n"));
2049       team = __kmp_allocate_team(root, nthreads, nthreads,
2050 #if OMPT_SUPPORT
2051                                  ompt_parallel_data,
2052 #endif
2053 #if OMP_40_ENABLED
2054                                  proc_bind,
2055 #endif
2056                                  &new_icvs, argc USE_NESTED_HOT_ARG(master_th));
2057     } else {
2058       /* allocate a new parallel team */
2059       KF_TRACE(10, ("__kmp_fork_call: before __kmp_allocate_team\n"));
2060       team = __kmp_allocate_team(root, nthreads, nthreads,
2061 #if OMPT_SUPPORT
2062                                  ompt_parallel_data,
2063 #endif
2064 #if OMP_40_ENABLED
2065                                  proc_bind,
2066 #endif
2067                                  &master_th->th.th_current_task->td_icvs,
2068                                  argc USE_NESTED_HOT_ARG(master_th));
2069     }
2070     KF_TRACE(
2071         10, ("__kmp_fork_call: after __kmp_allocate_team - team = %p\n", team));
2072 
2073     /* setup the new team */
2074     KMP_CHECK_UPDATE(team->t.t_master_tid, master_tid);
2075     KMP_CHECK_UPDATE(team->t.t_master_this_cons, master_this_cons);
2076     KMP_CHECK_UPDATE(team->t.t_ident, loc);
2077     KMP_CHECK_UPDATE(team->t.t_parent, parent_team);
2078     KMP_CHECK_UPDATE_SYNC(team->t.t_pkfn, microtask);
2079 #if OMPT_SUPPORT
2080     KMP_CHECK_UPDATE_SYNC(team->t.ompt_team_info.master_return_address,
2081                           return_address);
2082 #endif
2083     KMP_CHECK_UPDATE(team->t.t_invoke, invoker); // TODO move to root, maybe
2084 // TODO: parent_team->t.t_level == INT_MAX ???
2085 #if OMP_40_ENABLED
2086     if (!master_th->th.th_teams_microtask || level > teams_level) {
2087 #endif /* OMP_40_ENABLED */
2088       int new_level = parent_team->t.t_level + 1;
2089       KMP_CHECK_UPDATE(team->t.t_level, new_level);
2090       new_level = parent_team->t.t_active_level + 1;
2091       KMP_CHECK_UPDATE(team->t.t_active_level, new_level);
2092 #if OMP_40_ENABLED
2093     } else {
2094       // AC: Do not increase parallel level at start of the teams construct
2095       int new_level = parent_team->t.t_level;
2096       KMP_CHECK_UPDATE(team->t.t_level, new_level);
2097       new_level = parent_team->t.t_active_level;
2098       KMP_CHECK_UPDATE(team->t.t_active_level, new_level);
2099     }
2100 #endif /* OMP_40_ENABLED */
2101     kmp_r_sched_t new_sched = get__sched_2(parent_team, master_tid);
2102     // set master's schedule as new run-time schedule
2103     KMP_CHECK_UPDATE(team->t.t_sched.sched, new_sched.sched);
2104 
2105 #if OMP_40_ENABLED
2106     KMP_CHECK_UPDATE(team->t.t_cancel_request, cancel_noreq);
2107 #endif
2108 #if OMP_50_ENABLED
2109     KMP_CHECK_UPDATE(team->t.t_def_allocator, master_th->th.th_def_allocator);
2110 #endif
2111 
2112     // Update the floating point rounding in the team if required.
2113     propagateFPControl(team);
2114 
2115     if (__kmp_tasking_mode != tskm_immediate_exec) {
2116       // Set master's task team to team's task team. Unless this is hot team, it
2117       // should be NULL.
2118       KMP_DEBUG_ASSERT(master_th->th.th_task_team ==
2119                        parent_team->t.t_task_team[master_th->th.th_task_state]);
2120       KA_TRACE(20, ("__kmp_fork_call: Master T#%d pushing task_team %p / team "
2121                     "%p, new task_team %p / team %p\n",
2122                     __kmp_gtid_from_thread(master_th),
2123                     master_th->th.th_task_team, parent_team,
2124                     team->t.t_task_team[master_th->th.th_task_state], team));
2125 
2126       if (active_level || master_th->th.th_task_team) {
2127         // Take a memo of master's task_state
2128         KMP_DEBUG_ASSERT(master_th->th.th_task_state_memo_stack);
2129         if (master_th->th.th_task_state_top >=
2130             master_th->th.th_task_state_stack_sz) { // increase size
2131           kmp_uint32 new_size = 2 * master_th->th.th_task_state_stack_sz;
2132           kmp_uint8 *old_stack, *new_stack;
2133           kmp_uint32 i;
2134           new_stack = (kmp_uint8 *)__kmp_allocate(new_size);
2135           for (i = 0; i < master_th->th.th_task_state_stack_sz; ++i) {
2136             new_stack[i] = master_th->th.th_task_state_memo_stack[i];
2137           }
2138           for (i = master_th->th.th_task_state_stack_sz; i < new_size;
2139                ++i) { // zero-init rest of stack
2140             new_stack[i] = 0;
2141           }
2142           old_stack = master_th->th.th_task_state_memo_stack;
2143           master_th->th.th_task_state_memo_stack = new_stack;
2144           master_th->th.th_task_state_stack_sz = new_size;
2145           __kmp_free(old_stack);
2146         }
2147         // Store master's task_state on stack
2148         master_th->th
2149             .th_task_state_memo_stack[master_th->th.th_task_state_top] =
2150             master_th->th.th_task_state;
2151         master_th->th.th_task_state_top++;
2152 #if KMP_NESTED_HOT_TEAMS
2153         if (master_th->th.th_hot_teams &&
2154             active_level < __kmp_hot_teams_max_level &&
2155             team == master_th->th.th_hot_teams[active_level].hot_team) {
2156           // Restore master's nested state if nested hot team
2157           master_th->th.th_task_state =
2158               master_th->th
2159                   .th_task_state_memo_stack[master_th->th.th_task_state_top];
2160         } else {
2161 #endif
2162           master_th->th.th_task_state = 0;
2163 #if KMP_NESTED_HOT_TEAMS
2164         }
2165 #endif
2166       }
2167 #if !KMP_NESTED_HOT_TEAMS
2168       KMP_DEBUG_ASSERT((master_th->th.th_task_team == NULL) ||
2169                        (team == root->r.r_hot_team));
2170 #endif
2171     }
2172 
2173     KA_TRACE(
2174         20,
2175         ("__kmp_fork_call: T#%d(%d:%d)->(%d:0) created a team of %d threads\n",
2176          gtid, parent_team->t.t_id, team->t.t_master_tid, team->t.t_id,
2177          team->t.t_nproc));
2178     KMP_DEBUG_ASSERT(team != root->r.r_hot_team ||
2179                      (team->t.t_master_tid == 0 &&
2180                       (team->t.t_parent == root->r.r_root_team ||
2181                        team->t.t_parent->t.t_serialized)));
2182     KMP_MB();
2183 
2184     /* now, setup the arguments */
2185     argv = (void **)team->t.t_argv;
2186 #if OMP_40_ENABLED
2187     if (ap) {
2188 #endif /* OMP_40_ENABLED */
2189       for (i = argc - 1; i >= 0; --i) {
2190 // TODO: revert workaround for Intel(R) 64 tracker #96
2191 #if (KMP_ARCH_X86_64 || KMP_ARCH_ARM || KMP_ARCH_AARCH64) && KMP_OS_LINUX
2192         void *new_argv = va_arg(*ap, void *);
2193 #else
2194       void *new_argv = va_arg(ap, void *);
2195 #endif
2196         KMP_CHECK_UPDATE(*argv, new_argv);
2197         argv++;
2198       }
2199 #if OMP_40_ENABLED
2200     } else {
2201       for (i = 0; i < argc; ++i) {
2202         // Get args from parent team for teams construct
2203         KMP_CHECK_UPDATE(argv[i], team->t.t_parent->t.t_argv[i]);
2204       }
2205     }
2206 #endif /* OMP_40_ENABLED */
2207 
2208     /* now actually fork the threads */
2209     KMP_CHECK_UPDATE(team->t.t_master_active, master_active);
2210     if (!root->r.r_active) // Only do assignment if it prevents cache ping-pong
2211       root->r.r_active = TRUE;
2212 
2213     __kmp_fork_team_threads(root, team, master_th, gtid);
2214     __kmp_setup_icv_copy(team, nthreads,
2215                          &master_th->th.th_current_task->td_icvs, loc);
2216 
2217 #if OMPT_SUPPORT
2218     master_th->th.ompt_thread_info.state = ompt_state_work_parallel;
2219 #endif
2220 
2221     __kmp_release_bootstrap_lock(&__kmp_forkjoin_lock);
2222 
2223 #if USE_ITT_BUILD
2224     if (team->t.t_active_level == 1 // only report frames at level 1
2225 #if OMP_40_ENABLED
2226         && !master_th->th.th_teams_microtask // not in teams construct
2227 #endif /* OMP_40_ENABLED */
2228         ) {
2229 #if USE_ITT_NOTIFY
2230       if ((__itt_frame_submit_v3_ptr || KMP_ITT_DEBUG) &&
2231           (__kmp_forkjoin_frames_mode == 3 ||
2232            __kmp_forkjoin_frames_mode == 1)) {
2233         kmp_uint64 tmp_time = 0;
2234         if (__itt_get_timestamp_ptr)
2235           tmp_time = __itt_get_timestamp();
2236         // Internal fork - report frame begin
2237         master_th->th.th_frame_time = tmp_time;
2238         if (__kmp_forkjoin_frames_mode == 3)
2239           team->t.t_region_time = tmp_time;
2240       } else
2241 // only one notification scheme (either "submit" or "forking/joined", not both)
2242 #endif /* USE_ITT_NOTIFY */
2243           if ((__itt_frame_begin_v3_ptr || KMP_ITT_DEBUG) &&
2244               __kmp_forkjoin_frames && !__kmp_forkjoin_frames_mode) {
2245         // Mark start of "parallel" region for Intel(R) VTune(TM) analyzer.
2246         __kmp_itt_region_forking(gtid, team->t.t_nproc, 0);
2247       }
2248     }
2249 #endif /* USE_ITT_BUILD */
2250 
2251     /* now go on and do the work */
2252     KMP_DEBUG_ASSERT(team == __kmp_threads[gtid]->th.th_team);
2253     KMP_MB();
2254     KF_TRACE(10,
2255              ("__kmp_internal_fork : root=%p, team=%p, master_th=%p, gtid=%d\n",
2256               root, team, master_th, gtid));
2257 
2258 #if USE_ITT_BUILD
2259     if (__itt_stack_caller_create_ptr) {
2260       team->t.t_stack_id =
2261           __kmp_itt_stack_caller_create(); // create new stack stitching id
2262       // before entering fork barrier
2263     }
2264 #endif /* USE_ITT_BUILD */
2265 
2266 #if OMP_40_ENABLED
2267     // AC: skip __kmp_internal_fork at teams construct, let only master
2268     // threads execute
2269     if (ap)
2270 #endif /* OMP_40_ENABLED */
2271     {
2272       __kmp_internal_fork(loc, gtid, team);
2273       KF_TRACE(10, ("__kmp_internal_fork : after : root=%p, team=%p, "
2274                     "master_th=%p, gtid=%d\n",
2275                     root, team, master_th, gtid));
2276     }
2277 
2278     if (call_context == fork_context_gnu) {
2279       KA_TRACE(20, ("__kmp_fork_call: parallel exit T#%d\n", gtid));
2280       return TRUE;
2281     }
2282 
2283     /* Invoke microtask for MASTER thread */
2284     KA_TRACE(20, ("__kmp_fork_call: T#%d(%d:0) invoke microtask = %p\n", gtid,
2285                   team->t.t_id, team->t.t_pkfn));
2286   } // END of timer KMP_fork_call block
2287 
2288 #if KMP_STATS_ENABLED && OMP_40_ENABLED
2289   // If beginning a teams construct, then change thread state
2290   stats_state_e previous_state = KMP_GET_THREAD_STATE();
2291   if (!ap) {
2292     KMP_SET_THREAD_STATE(stats_state_e::TEAMS_REGION);
2293   }
2294 #endif
2295 
2296   if (!team->t.t_invoke(gtid)) {
2297     KMP_ASSERT2(0, "cannot invoke microtask for MASTER thread");
2298   }
2299 
2300 #if KMP_STATS_ENABLED && OMP_40_ENABLED
2301   // If was beginning of a teams construct, then reset thread state
2302   if (!ap) {
2303     KMP_SET_THREAD_STATE(previous_state);
2304   }
2305 #endif
2306 
2307   KA_TRACE(20, ("__kmp_fork_call: T#%d(%d:0) done microtask = %p\n", gtid,
2308                 team->t.t_id, team->t.t_pkfn));
2309   KMP_MB(); /* Flush all pending memory write invalidates.  */
2310 
2311   KA_TRACE(20, ("__kmp_fork_call: parallel exit T#%d\n", gtid));
2312 
2313 #if OMPT_SUPPORT
2314   if (ompt_enabled.enabled) {
2315     master_th->th.ompt_thread_info.state = ompt_state_overhead;
2316   }
2317 #endif
2318 
2319   return TRUE;
2320 }
2321 
2322 #if OMPT_SUPPORT
2323 static inline void __kmp_join_restore_state(kmp_info_t *thread,
2324                                             kmp_team_t *team) {
2325   // restore state outside the region
2326   thread->th.ompt_thread_info.state =
2327       ((team->t.t_serialized) ? ompt_state_work_serial
2328                               : ompt_state_work_parallel);
2329 }
2330 
2331 static inline void __kmp_join_ompt(int gtid, kmp_info_t *thread,
2332                                    kmp_team_t *team, ompt_data_t *parallel_data,
2333                                    fork_context_e fork_context, void *codeptr) {
2334   ompt_task_info_t *task_info = __ompt_get_task_info_object(0);
2335   if (ompt_enabled.ompt_callback_parallel_end) {
2336     ompt_callbacks.ompt_callback(ompt_callback_parallel_end)(
2337         parallel_data, &(task_info->task_data), OMPT_INVOKER(fork_context),
2338         codeptr);
2339   }
2340 
2341   task_info->frame.enter_frame = ompt_data_none;
2342   __kmp_join_restore_state(thread, team);
2343 }
2344 #endif
2345 
2346 void __kmp_join_call(ident_t *loc, int gtid
2347 #if OMPT_SUPPORT
2348                      ,
2349                      enum fork_context_e fork_context
2350 #endif
2351 #if OMP_40_ENABLED
2352                      ,
2353                      int exit_teams
2354 #endif /* OMP_40_ENABLED */
2355                      ) {
2356   KMP_TIME_DEVELOPER_PARTITIONED_BLOCK(KMP_join_call);
2357   kmp_team_t *team;
2358   kmp_team_t *parent_team;
2359   kmp_info_t *master_th;
2360   kmp_root_t *root;
2361   int master_active;
2362 
2363   KA_TRACE(20, ("__kmp_join_call: enter T#%d\n", gtid));
2364 
2365   /* setup current data */
2366   master_th = __kmp_threads[gtid];
2367   root = master_th->th.th_root;
2368   team = master_th->th.th_team;
2369   parent_team = team->t.t_parent;
2370 
2371   master_th->th.th_ident = loc;
2372 
2373 #if OMPT_SUPPORT
2374   if (ompt_enabled.enabled) {
2375     master_th->th.ompt_thread_info.state = ompt_state_overhead;
2376   }
2377 #endif
2378 
2379 #if KMP_DEBUG
2380   if (__kmp_tasking_mode != tskm_immediate_exec && !exit_teams) {
2381     KA_TRACE(20, ("__kmp_join_call: T#%d, old team = %p old task_team = %p, "
2382                   "th_task_team = %p\n",
2383                   __kmp_gtid_from_thread(master_th), team,
2384                   team->t.t_task_team[master_th->th.th_task_state],
2385                   master_th->th.th_task_team));
2386     KMP_DEBUG_ASSERT(master_th->th.th_task_team ==
2387                      team->t.t_task_team[master_th->th.th_task_state]);
2388   }
2389 #endif
2390 
2391   if (team->t.t_serialized) {
2392 #if OMP_40_ENABLED
2393     if (master_th->th.th_teams_microtask) {
2394       // We are in teams construct
2395       int level = team->t.t_level;
2396       int tlevel = master_th->th.th_teams_level;
2397       if (level == tlevel) {
2398         // AC: we haven't incremented it earlier at start of teams construct,
2399         //     so do it here - at the end of teams construct
2400         team->t.t_level++;
2401       } else if (level == tlevel + 1) {
2402         // AC: we are exiting parallel inside teams, need to increment
2403         // serialization in order to restore it in the next call to
2404         // __kmpc_end_serialized_parallel
2405         team->t.t_serialized++;
2406       }
2407     }
2408 #endif /* OMP_40_ENABLED */
2409     __kmpc_end_serialized_parallel(loc, gtid);
2410 
2411 #if OMPT_SUPPORT
2412     if (ompt_enabled.enabled) {
2413       __kmp_join_restore_state(master_th, parent_team);
2414     }
2415 #endif
2416 
2417     return;
2418   }
2419 
2420   master_active = team->t.t_master_active;
2421 
2422 #if OMP_40_ENABLED
2423   if (!exit_teams)
2424 #endif /* OMP_40_ENABLED */
2425   {
2426     // AC: No barrier for internal teams at exit from teams construct.
2427     //     But there is barrier for external team (league).
2428     __kmp_internal_join(loc, gtid, team);
2429   }
2430 #if OMP_40_ENABLED
2431   else {
2432     master_th->th.th_task_state =
2433         0; // AC: no tasking in teams (out of any parallel)
2434   }
2435 #endif /* OMP_40_ENABLED */
2436 
2437   KMP_MB();
2438 
2439 #if OMPT_SUPPORT
2440   ompt_data_t *parallel_data = &(team->t.ompt_team_info.parallel_data);
2441   void *codeptr = team->t.ompt_team_info.master_return_address;
2442 #endif
2443 
2444 #if USE_ITT_BUILD
2445   if (__itt_stack_caller_create_ptr) {
2446     __kmp_itt_stack_caller_destroy(
2447         (__itt_caller)team->t
2448             .t_stack_id); // destroy the stack stitching id after join barrier
2449   }
2450 
2451   // Mark end of "parallel" region for Intel(R) VTune(TM) analyzer.
2452   if (team->t.t_active_level == 1
2453 #if OMP_40_ENABLED
2454       && !master_th->th.th_teams_microtask /* not in teams construct */
2455 #endif /* OMP_40_ENABLED */
2456       ) {
2457     master_th->th.th_ident = loc;
2458     // only one notification scheme (either "submit" or "forking/joined", not
2459     // both)
2460     if ((__itt_frame_submit_v3_ptr || KMP_ITT_DEBUG) &&
2461         __kmp_forkjoin_frames_mode == 3)
2462       __kmp_itt_frame_submit(gtid, team->t.t_region_time,
2463                              master_th->th.th_frame_time, 0, loc,
2464                              master_th->th.th_team_nproc, 1);
2465     else if ((__itt_frame_end_v3_ptr || KMP_ITT_DEBUG) &&
2466              !__kmp_forkjoin_frames_mode && __kmp_forkjoin_frames)
2467       __kmp_itt_region_joined(gtid);
2468   } // active_level == 1
2469 #endif /* USE_ITT_BUILD */
2470 
2471 #if OMP_40_ENABLED
2472   if (master_th->th.th_teams_microtask && !exit_teams &&
2473       team->t.t_pkfn != (microtask_t)__kmp_teams_master &&
2474       team->t.t_level == master_th->th.th_teams_level + 1) {
2475     // AC: We need to leave the team structure intact at the end of parallel
2476     // inside the teams construct, so that at the next parallel same (hot) team
2477     // works, only adjust nesting levels
2478 
2479     /* Decrement our nested depth level */
2480     team->t.t_level--;
2481     team->t.t_active_level--;
2482     KMP_ATOMIC_DEC(&root->r.r_in_parallel);
2483 
2484     // Restore number of threads in the team if needed. This code relies on
2485     // the proper adjustment of th_teams_size.nth after the fork in
2486     // __kmp_teams_master on each teams master in the case that
2487     // __kmp_reserve_threads reduced it.
2488     if (master_th->th.th_team_nproc < master_th->th.th_teams_size.nth) {
2489       int old_num = master_th->th.th_team_nproc;
2490       int new_num = master_th->th.th_teams_size.nth;
2491       kmp_info_t **other_threads = team->t.t_threads;
2492       team->t.t_nproc = new_num;
2493       for (int i = 0; i < old_num; ++i) {
2494         other_threads[i]->th.th_team_nproc = new_num;
2495       }
2496       // Adjust states of non-used threads of the team
2497       for (int i = old_num; i < new_num; ++i) {
2498         // Re-initialize thread's barrier data.
2499         KMP_DEBUG_ASSERT(other_threads[i]);
2500         kmp_balign_t *balign = other_threads[i]->th.th_bar;
2501         for (int b = 0; b < bs_last_barrier; ++b) {
2502           balign[b].bb.b_arrived = team->t.t_bar[b].b_arrived;
2503           KMP_DEBUG_ASSERT(balign[b].bb.wait_flag != KMP_BARRIER_PARENT_FLAG);
2504 #if USE_DEBUGGER
2505           balign[b].bb.b_worker_arrived = team->t.t_bar[b].b_team_arrived;
2506 #endif
2507         }
2508         if (__kmp_tasking_mode != tskm_immediate_exec) {
2509           // Synchronize thread's task state
2510           other_threads[i]->th.th_task_state = master_th->th.th_task_state;
2511         }
2512       }
2513     }
2514 
2515 #if OMPT_SUPPORT
2516     if (ompt_enabled.enabled) {
2517       __kmp_join_ompt(gtid, master_th, parent_team, parallel_data, fork_context,
2518                       codeptr);
2519     }
2520 #endif
2521 
2522     return;
2523   }
2524 #endif /* OMP_40_ENABLED */
2525 
2526   /* do cleanup and restore the parent team */
2527   master_th->th.th_info.ds.ds_tid = team->t.t_master_tid;
2528   master_th->th.th_local.this_construct = team->t.t_master_this_cons;
2529 
2530   master_th->th.th_dispatch = &parent_team->t.t_dispatch[team->t.t_master_tid];
2531 
2532   /* jc: The following lock has instructions with REL and ACQ semantics,
2533      separating the parallel user code called in this parallel region
2534      from the serial user code called after this function returns. */
2535   __kmp_acquire_bootstrap_lock(&__kmp_forkjoin_lock);
2536 
2537 #if OMP_40_ENABLED
2538   if (!master_th->th.th_teams_microtask ||
2539       team->t.t_level > master_th->th.th_teams_level)
2540 #endif /* OMP_40_ENABLED */
2541   {
2542     /* Decrement our nested depth level */
2543     KMP_ATOMIC_DEC(&root->r.r_in_parallel);
2544   }
2545   KMP_DEBUG_ASSERT(root->r.r_in_parallel >= 0);
2546 
2547 #if OMPT_SUPPORT
2548   if (ompt_enabled.enabled) {
2549     ompt_task_info_t *task_info = __ompt_get_task_info_object(0);
2550     if (ompt_enabled.ompt_callback_implicit_task) {
2551       int ompt_team_size = team->t.t_nproc;
2552       ompt_callbacks.ompt_callback(ompt_callback_implicit_task)(
2553           ompt_scope_end, NULL, &(task_info->task_data), ompt_team_size,
2554           OMPT_CUR_TASK_INFO(master_th)->thread_num, ompt_task_implicit); // TODO: Can this be ompt_task_initial?
2555     }
2556 
2557     task_info->frame.exit_frame = ompt_data_none;
2558     task_info->task_data = ompt_data_none;
2559   }
2560 #endif
2561 
2562   KF_TRACE(10, ("__kmp_join_call1: T#%d, this_thread=%p team=%p\n", 0,
2563                 master_th, team));
2564   __kmp_pop_current_task_from_thread(master_th);
2565 
2566 #if OMP_40_ENABLED && KMP_AFFINITY_SUPPORTED
2567   // Restore master thread's partition.
2568   master_th->th.th_first_place = team->t.t_first_place;
2569   master_th->th.th_last_place = team->t.t_last_place;
2570 #endif /* OMP_40_ENABLED */
2571 #if OMP_50_ENABLED
2572   master_th->th.th_def_allocator = team->t.t_def_allocator;
2573 #endif
2574 
2575   updateHWFPControl(team);
2576 
2577   if (root->r.r_active != master_active)
2578     root->r.r_active = master_active;
2579 
2580   __kmp_free_team(root, team USE_NESTED_HOT_ARG(
2581                             master_th)); // this will free worker threads
2582 
2583   /* this race was fun to find. make sure the following is in the critical
2584      region otherwise assertions may fail occasionally since the old team may be
2585      reallocated and the hierarchy appears inconsistent. it is actually safe to
2586      run and won't cause any bugs, but will cause those assertion failures. it's
2587      only one deref&assign so might as well put this in the critical region */
2588   master_th->th.th_team = parent_team;
2589   master_th->th.th_team_nproc = parent_team->t.t_nproc;
2590   master_th->th.th_team_master = parent_team->t.t_threads[0];
2591   master_th->th.th_team_serialized = parent_team->t.t_serialized;
2592 
2593   /* restore serialized team, if need be */
2594   if (parent_team->t.t_serialized &&
2595       parent_team != master_th->th.th_serial_team &&
2596       parent_team != root->r.r_root_team) {
2597     __kmp_free_team(root,
2598                     master_th->th.th_serial_team USE_NESTED_HOT_ARG(NULL));
2599     master_th->th.th_serial_team = parent_team;
2600   }
2601 
2602   if (__kmp_tasking_mode != tskm_immediate_exec) {
2603     if (master_th->th.th_task_state_top >
2604         0) { // Restore task state from memo stack
2605       KMP_DEBUG_ASSERT(master_th->th.th_task_state_memo_stack);
2606       // Remember master's state if we re-use this nested hot team
2607       master_th->th.th_task_state_memo_stack[master_th->th.th_task_state_top] =
2608           master_th->th.th_task_state;
2609       --master_th->th.th_task_state_top; // pop
2610       // Now restore state at this level
2611       master_th->th.th_task_state =
2612           master_th->th
2613               .th_task_state_memo_stack[master_th->th.th_task_state_top];
2614     }
2615     // Copy the task team from the parent team to the master thread
2616     master_th->th.th_task_team =
2617         parent_team->t.t_task_team[master_th->th.th_task_state];
2618     KA_TRACE(20,
2619              ("__kmp_join_call: Master T#%d restoring task_team %p / team %p\n",
2620               __kmp_gtid_from_thread(master_th), master_th->th.th_task_team,
2621               parent_team));
2622   }
2623 
2624   // TODO: GEH - cannot do this assertion because root thread not set up as
2625   // executing
2626   // KMP_ASSERT( master_th->th.th_current_task->td_flags.executing == 0 );
2627   master_th->th.th_current_task->td_flags.executing = 1;
2628 
2629   __kmp_release_bootstrap_lock(&__kmp_forkjoin_lock);
2630 
2631 #if OMPT_SUPPORT
2632   if (ompt_enabled.enabled) {
2633     __kmp_join_ompt(gtid, master_th, parent_team, parallel_data, fork_context,
2634                     codeptr);
2635   }
2636 #endif
2637 
2638   KMP_MB();
2639   KA_TRACE(20, ("__kmp_join_call: exit T#%d\n", gtid));
2640 }
2641 
2642 /* Check whether we should push an internal control record onto the
2643    serial team stack.  If so, do it.  */
2644 void __kmp_save_internal_controls(kmp_info_t *thread) {
2645 
2646   if (thread->th.th_team != thread->th.th_serial_team) {
2647     return;
2648   }
2649   if (thread->th.th_team->t.t_serialized > 1) {
2650     int push = 0;
2651 
2652     if (thread->th.th_team->t.t_control_stack_top == NULL) {
2653       push = 1;
2654     } else {
2655       if (thread->th.th_team->t.t_control_stack_top->serial_nesting_level !=
2656           thread->th.th_team->t.t_serialized) {
2657         push = 1;
2658       }
2659     }
2660     if (push) { /* push a record on the serial team's stack */
2661       kmp_internal_control_t *control =
2662           (kmp_internal_control_t *)__kmp_allocate(
2663               sizeof(kmp_internal_control_t));
2664 
2665       copy_icvs(control, &thread->th.th_current_task->td_icvs);
2666 
2667       control->serial_nesting_level = thread->th.th_team->t.t_serialized;
2668 
2669       control->next = thread->th.th_team->t.t_control_stack_top;
2670       thread->th.th_team->t.t_control_stack_top = control;
2671     }
2672   }
2673 }
2674 
2675 /* Changes set_nproc */
2676 void __kmp_set_num_threads(int new_nth, int gtid) {
2677   kmp_info_t *thread;
2678   kmp_root_t *root;
2679 
2680   KF_TRACE(10, ("__kmp_set_num_threads: new __kmp_nth = %d\n", new_nth));
2681   KMP_DEBUG_ASSERT(__kmp_init_serial);
2682 
2683   if (new_nth < 1)
2684     new_nth = 1;
2685   else if (new_nth > __kmp_max_nth)
2686     new_nth = __kmp_max_nth;
2687 
2688   KMP_COUNT_VALUE(OMP_set_numthreads, new_nth);
2689   thread = __kmp_threads[gtid];
2690   if (thread->th.th_current_task->td_icvs.nproc == new_nth)
2691     return; // nothing to do
2692 
2693   __kmp_save_internal_controls(thread);
2694 
2695   set__nproc(thread, new_nth);
2696 
2697   // If this omp_set_num_threads() call will cause the hot team size to be
2698   // reduced (in the absence of a num_threads clause), then reduce it now,
2699   // rather than waiting for the next parallel region.
2700   root = thread->th.th_root;
2701   if (__kmp_init_parallel && (!root->r.r_active) &&
2702       (root->r.r_hot_team->t.t_nproc > new_nth)
2703 #if KMP_NESTED_HOT_TEAMS
2704       && __kmp_hot_teams_max_level && !__kmp_hot_teams_mode
2705 #endif
2706       ) {
2707     kmp_team_t *hot_team = root->r.r_hot_team;
2708     int f;
2709 
2710     __kmp_acquire_bootstrap_lock(&__kmp_forkjoin_lock);
2711 
2712     // Release the extra threads we don't need any more.
2713     for (f = new_nth; f < hot_team->t.t_nproc; f++) {
2714       KMP_DEBUG_ASSERT(hot_team->t.t_threads[f] != NULL);
2715       if (__kmp_tasking_mode != tskm_immediate_exec) {
2716         // When decreasing team size, threads no longer in the team should unref
2717         // task team.
2718         hot_team->t.t_threads[f]->th.th_task_team = NULL;
2719       }
2720       __kmp_free_thread(hot_team->t.t_threads[f]);
2721       hot_team->t.t_threads[f] = NULL;
2722     }
2723     hot_team->t.t_nproc = new_nth;
2724 #if KMP_NESTED_HOT_TEAMS
2725     if (thread->th.th_hot_teams) {
2726       KMP_DEBUG_ASSERT(hot_team == thread->th.th_hot_teams[0].hot_team);
2727       thread->th.th_hot_teams[0].hot_team_nth = new_nth;
2728     }
2729 #endif
2730 
2731     __kmp_release_bootstrap_lock(&__kmp_forkjoin_lock);
2732 
2733     // Update the t_nproc field in the threads that are still active.
2734     for (f = 0; f < new_nth; f++) {
2735       KMP_DEBUG_ASSERT(hot_team->t.t_threads[f] != NULL);
2736       hot_team->t.t_threads[f]->th.th_team_nproc = new_nth;
2737     }
2738     // Special flag in case omp_set_num_threads() call
2739     hot_team->t.t_size_changed = -1;
2740   }
2741 }
2742 
2743 /* Changes max_active_levels */
2744 void __kmp_set_max_active_levels(int gtid, int max_active_levels) {
2745   kmp_info_t *thread;
2746 
2747   KF_TRACE(10, ("__kmp_set_max_active_levels: new max_active_levels for thread "
2748                 "%d = (%d)\n",
2749                 gtid, max_active_levels));
2750   KMP_DEBUG_ASSERT(__kmp_init_serial);
2751 
2752   // validate max_active_levels
2753   if (max_active_levels < 0) {
2754     KMP_WARNING(ActiveLevelsNegative, max_active_levels);
2755     // We ignore this call if the user has specified a negative value.
2756     // The current setting won't be changed. The last valid setting will be
2757     // used. A warning will be issued (if warnings are allowed as controlled by
2758     // the KMP_WARNINGS env var).
2759     KF_TRACE(10, ("__kmp_set_max_active_levels: the call is ignored: new "
2760                   "max_active_levels for thread %d = (%d)\n",
2761                   gtid, max_active_levels));
2762     return;
2763   }
2764   if (max_active_levels <= KMP_MAX_ACTIVE_LEVELS_LIMIT) {
2765     // it's OK, the max_active_levels is within the valid range: [ 0;
2766     // KMP_MAX_ACTIVE_LEVELS_LIMIT ]
2767     // We allow a zero value. (implementation defined behavior)
2768   } else {
2769     KMP_WARNING(ActiveLevelsExceedLimit, max_active_levels,
2770                 KMP_MAX_ACTIVE_LEVELS_LIMIT);
2771     max_active_levels = KMP_MAX_ACTIVE_LEVELS_LIMIT;
2772     // Current upper limit is MAX_INT. (implementation defined behavior)
2773     // If the input exceeds the upper limit, we correct the input to be the
2774     // upper limit. (implementation defined behavior)
2775     // Actually, the flow should never get here until we use MAX_INT limit.
2776   }
2777   KF_TRACE(10, ("__kmp_set_max_active_levels: after validation: new "
2778                 "max_active_levels for thread %d = (%d)\n",
2779                 gtid, max_active_levels));
2780 
2781   thread = __kmp_threads[gtid];
2782 
2783   __kmp_save_internal_controls(thread);
2784 
2785   set__max_active_levels(thread, max_active_levels);
2786 }
2787 
2788 /* Gets max_active_levels */
2789 int __kmp_get_max_active_levels(int gtid) {
2790   kmp_info_t *thread;
2791 
2792   KF_TRACE(10, ("__kmp_get_max_active_levels: thread %d\n", gtid));
2793   KMP_DEBUG_ASSERT(__kmp_init_serial);
2794 
2795   thread = __kmp_threads[gtid];
2796   KMP_DEBUG_ASSERT(thread->th.th_current_task);
2797   KF_TRACE(10, ("__kmp_get_max_active_levels: thread %d, curtask=%p, "
2798                 "curtask_maxaclevel=%d\n",
2799                 gtid, thread->th.th_current_task,
2800                 thread->th.th_current_task->td_icvs.max_active_levels));
2801   return thread->th.th_current_task->td_icvs.max_active_levels;
2802 }
2803 
2804 /* Changes def_sched_var ICV values (run-time schedule kind and chunk) */
2805 void __kmp_set_schedule(int gtid, kmp_sched_t kind, int chunk) {
2806   kmp_info_t *thread;
2807   //    kmp_team_t *team;
2808 
2809   KF_TRACE(10, ("__kmp_set_schedule: new schedule for thread %d = (%d, %d)\n",
2810                 gtid, (int)kind, chunk));
2811   KMP_DEBUG_ASSERT(__kmp_init_serial);
2812 
2813   // Check if the kind parameter is valid, correct if needed.
2814   // Valid parameters should fit in one of two intervals - standard or extended:
2815   //       <lower>, <valid>, <upper_std>, <lower_ext>, <valid>, <upper>
2816   // 2008-01-25: 0,  1 - 4,       5,         100,     101 - 102, 103
2817   if (kind <= kmp_sched_lower || kind >= kmp_sched_upper ||
2818       (kind <= kmp_sched_lower_ext && kind >= kmp_sched_upper_std)) {
2819     // TODO: Hint needs attention in case we change the default schedule.
2820     __kmp_msg(kmp_ms_warning, KMP_MSG(ScheduleKindOutOfRange, kind),
2821               KMP_HNT(DefaultScheduleKindUsed, "static, no chunk"),
2822               __kmp_msg_null);
2823     kind = kmp_sched_default;
2824     chunk = 0; // ignore chunk value in case of bad kind
2825   }
2826 
2827   thread = __kmp_threads[gtid];
2828 
2829   __kmp_save_internal_controls(thread);
2830 
2831   if (kind < kmp_sched_upper_std) {
2832     if (kind == kmp_sched_static && chunk < KMP_DEFAULT_CHUNK) {
2833       // differ static chunked vs. unchunked:  chunk should be invalid to
2834       // indicate unchunked schedule (which is the default)
2835       thread->th.th_current_task->td_icvs.sched.r_sched_type = kmp_sch_static;
2836     } else {
2837       thread->th.th_current_task->td_icvs.sched.r_sched_type =
2838           __kmp_sch_map[kind - kmp_sched_lower - 1];
2839     }
2840   } else {
2841     //    __kmp_sch_map[ kind - kmp_sched_lower_ext + kmp_sched_upper_std -
2842     //    kmp_sched_lower - 2 ];
2843     thread->th.th_current_task->td_icvs.sched.r_sched_type =
2844         __kmp_sch_map[kind - kmp_sched_lower_ext + kmp_sched_upper_std -
2845                       kmp_sched_lower - 2];
2846   }
2847   if (kind == kmp_sched_auto || chunk < 1) {
2848     // ignore parameter chunk for schedule auto
2849     thread->th.th_current_task->td_icvs.sched.chunk = KMP_DEFAULT_CHUNK;
2850   } else {
2851     thread->th.th_current_task->td_icvs.sched.chunk = chunk;
2852   }
2853 }
2854 
2855 /* Gets def_sched_var ICV values */
2856 void __kmp_get_schedule(int gtid, kmp_sched_t *kind, int *chunk) {
2857   kmp_info_t *thread;
2858   enum sched_type th_type;
2859 
2860   KF_TRACE(10, ("__kmp_get_schedule: thread %d\n", gtid));
2861   KMP_DEBUG_ASSERT(__kmp_init_serial);
2862 
2863   thread = __kmp_threads[gtid];
2864 
2865   th_type = thread->th.th_current_task->td_icvs.sched.r_sched_type;
2866 
2867   switch (th_type) {
2868   case kmp_sch_static:
2869   case kmp_sch_static_greedy:
2870   case kmp_sch_static_balanced:
2871     *kind = kmp_sched_static;
2872     *chunk = 0; // chunk was not set, try to show this fact via zero value
2873     return;
2874   case kmp_sch_static_chunked:
2875     *kind = kmp_sched_static;
2876     break;
2877   case kmp_sch_dynamic_chunked:
2878     *kind = kmp_sched_dynamic;
2879     break;
2880   case kmp_sch_guided_chunked:
2881   case kmp_sch_guided_iterative_chunked:
2882   case kmp_sch_guided_analytical_chunked:
2883     *kind = kmp_sched_guided;
2884     break;
2885   case kmp_sch_auto:
2886     *kind = kmp_sched_auto;
2887     break;
2888   case kmp_sch_trapezoidal:
2889     *kind = kmp_sched_trapezoidal;
2890     break;
2891 #if KMP_STATIC_STEAL_ENABLED
2892   case kmp_sch_static_steal:
2893     *kind = kmp_sched_static_steal;
2894     break;
2895 #endif
2896   default:
2897     KMP_FATAL(UnknownSchedulingType, th_type);
2898   }
2899 
2900   *chunk = thread->th.th_current_task->td_icvs.sched.chunk;
2901 }
2902 
2903 int __kmp_get_ancestor_thread_num(int gtid, int level) {
2904 
2905   int ii, dd;
2906   kmp_team_t *team;
2907   kmp_info_t *thr;
2908 
2909   KF_TRACE(10, ("__kmp_get_ancestor_thread_num: thread %d %d\n", gtid, level));
2910   KMP_DEBUG_ASSERT(__kmp_init_serial);
2911 
2912   // validate level
2913   if (level == 0)
2914     return 0;
2915   if (level < 0)
2916     return -1;
2917   thr = __kmp_threads[gtid];
2918   team = thr->th.th_team;
2919   ii = team->t.t_level;
2920   if (level > ii)
2921     return -1;
2922 
2923 #if OMP_40_ENABLED
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 #endif
2940 
2941   if (ii == level)
2942     return __kmp_tid_from_gtid(gtid);
2943 
2944   dd = team->t.t_serialized;
2945   level++;
2946   while (ii > level) {
2947     for (dd = team->t.t_serialized; (dd > 0) && (ii > level); dd--, ii--) {
2948     }
2949     if ((team->t.t_serialized) && (!dd)) {
2950       team = team->t.t_parent;
2951       continue;
2952     }
2953     if (ii > level) {
2954       team = team->t.t_parent;
2955       dd = team->t.t_serialized;
2956       ii--;
2957     }
2958   }
2959 
2960   return (dd > 1) ? (0) : (team->t.t_master_tid);
2961 }
2962 
2963 int __kmp_get_team_size(int gtid, int level) {
2964 
2965   int ii, dd;
2966   kmp_team_t *team;
2967   kmp_info_t *thr;
2968 
2969   KF_TRACE(10, ("__kmp_get_team_size: thread %d %d\n", gtid, level));
2970   KMP_DEBUG_ASSERT(__kmp_init_serial);
2971 
2972   // validate level
2973   if (level == 0)
2974     return 1;
2975   if (level < 0)
2976     return -1;
2977   thr = __kmp_threads[gtid];
2978   team = thr->th.th_team;
2979   ii = team->t.t_level;
2980   if (level > ii)
2981     return -1;
2982 
2983 #if OMP_40_ENABLED
2984   if (thr->th.th_teams_microtask) {
2985     // AC: we are in teams region where multiple nested teams have same level
2986     int tlevel = thr->th.th_teams_level; // the level of the teams construct
2987     if (level <=
2988         tlevel) { // otherwise usual algorithm works (will not touch the teams)
2989       KMP_DEBUG_ASSERT(ii >= tlevel);
2990       // AC: As we need to pass by the teams league, we need to artificially
2991       // increase ii
2992       if (ii == tlevel) {
2993         ii += 2; // three teams have same level
2994       } else {
2995         ii++; // two teams have same level
2996       }
2997     }
2998   }
2999 #endif
3000 
3001   while (ii > level) {
3002     for (dd = team->t.t_serialized; (dd > 0) && (ii > level); dd--, ii--) {
3003     }
3004     if (team->t.t_serialized && (!dd)) {
3005       team = team->t.t_parent;
3006       continue;
3007     }
3008     if (ii > level) {
3009       team = team->t.t_parent;
3010       ii--;
3011     }
3012   }
3013 
3014   return team->t.t_nproc;
3015 }
3016 
3017 kmp_r_sched_t __kmp_get_schedule_global() {
3018   // This routine created because pairs (__kmp_sched, __kmp_chunk) and
3019   // (__kmp_static, __kmp_guided) may be changed by kmp_set_defaults
3020   // independently. So one can get the updated schedule here.
3021 
3022   kmp_r_sched_t r_sched;
3023 
3024   // create schedule from 4 globals: __kmp_sched, __kmp_chunk, __kmp_static,
3025   // __kmp_guided. __kmp_sched should keep original value, so that user can set
3026   // KMP_SCHEDULE multiple times, and thus have different run-time schedules in
3027   // different roots (even in OMP 2.5)
3028   if (__kmp_sched == kmp_sch_static) {
3029     // replace STATIC with more detailed schedule (balanced or greedy)
3030     r_sched.r_sched_type = __kmp_static;
3031   } else if (__kmp_sched == kmp_sch_guided_chunked) {
3032     // replace GUIDED with more detailed schedule (iterative or analytical)
3033     r_sched.r_sched_type = __kmp_guided;
3034   } else { // (STATIC_CHUNKED), or (DYNAMIC_CHUNKED), or other
3035     r_sched.r_sched_type = __kmp_sched;
3036   }
3037 
3038   if (__kmp_chunk < KMP_DEFAULT_CHUNK) {
3039     // __kmp_chunk may be wrong here (if it was not ever set)
3040     r_sched.chunk = KMP_DEFAULT_CHUNK;
3041   } else {
3042     r_sched.chunk = __kmp_chunk;
3043   }
3044 
3045   return r_sched;
3046 }
3047 
3048 /* Allocate (realloc == FALSE) * or reallocate (realloc == TRUE)
3049    at least argc number of *t_argv entries for the requested team. */
3050 static void __kmp_alloc_argv_entries(int argc, kmp_team_t *team, int realloc) {
3051 
3052   KMP_DEBUG_ASSERT(team);
3053   if (!realloc || argc > team->t.t_max_argc) {
3054 
3055     KA_TRACE(100, ("__kmp_alloc_argv_entries: team %d: needed entries=%d, "
3056                    "current entries=%d\n",
3057                    team->t.t_id, argc, (realloc) ? team->t.t_max_argc : 0));
3058     /* if previously allocated heap space for args, free them */
3059     if (realloc && team->t.t_argv != &team->t.t_inline_argv[0])
3060       __kmp_free((void *)team->t.t_argv);
3061 
3062     if (argc <= KMP_INLINE_ARGV_ENTRIES) {
3063       /* use unused space in the cache line for arguments */
3064       team->t.t_max_argc = KMP_INLINE_ARGV_ENTRIES;
3065       KA_TRACE(100, ("__kmp_alloc_argv_entries: team %d: inline allocate %d "
3066                      "argv entries\n",
3067                      team->t.t_id, team->t.t_max_argc));
3068       team->t.t_argv = &team->t.t_inline_argv[0];
3069       if (__kmp_storage_map) {
3070         __kmp_print_storage_map_gtid(
3071             -1, &team->t.t_inline_argv[0],
3072             &team->t.t_inline_argv[KMP_INLINE_ARGV_ENTRIES],
3073             (sizeof(void *) * KMP_INLINE_ARGV_ENTRIES), "team_%d.t_inline_argv",
3074             team->t.t_id);
3075       }
3076     } else {
3077       /* allocate space for arguments in the heap */
3078       team->t.t_max_argc = (argc <= (KMP_MIN_MALLOC_ARGV_ENTRIES >> 1))
3079                                ? KMP_MIN_MALLOC_ARGV_ENTRIES
3080                                : 2 * argc;
3081       KA_TRACE(100, ("__kmp_alloc_argv_entries: team %d: dynamic allocate %d "
3082                      "argv entries\n",
3083                      team->t.t_id, team->t.t_max_argc));
3084       team->t.t_argv =
3085           (void **)__kmp_page_allocate(sizeof(void *) * team->t.t_max_argc);
3086       if (__kmp_storage_map) {
3087         __kmp_print_storage_map_gtid(-1, &team->t.t_argv[0],
3088                                      &team->t.t_argv[team->t.t_max_argc],
3089                                      sizeof(void *) * team->t.t_max_argc,
3090                                      "team_%d.t_argv", team->t.t_id);
3091       }
3092     }
3093   }
3094 }
3095 
3096 static void __kmp_allocate_team_arrays(kmp_team_t *team, int max_nth) {
3097   int i;
3098   int num_disp_buff = max_nth > 1 ? __kmp_dispatch_num_buffers : 2;
3099   team->t.t_threads =
3100       (kmp_info_t **)__kmp_allocate(sizeof(kmp_info_t *) * max_nth);
3101   team->t.t_disp_buffer = (dispatch_shared_info_t *)__kmp_allocate(
3102       sizeof(dispatch_shared_info_t) * num_disp_buff);
3103   team->t.t_dispatch =
3104       (kmp_disp_t *)__kmp_allocate(sizeof(kmp_disp_t) * max_nth);
3105   team->t.t_implicit_task_taskdata =
3106       (kmp_taskdata_t *)__kmp_allocate(sizeof(kmp_taskdata_t) * max_nth);
3107   team->t.t_max_nproc = max_nth;
3108 
3109   /* setup dispatch buffers */
3110   for (i = 0; i < num_disp_buff; ++i) {
3111     team->t.t_disp_buffer[i].buffer_index = i;
3112 #if OMP_45_ENABLED
3113     team->t.t_disp_buffer[i].doacross_buf_idx = i;
3114 #endif
3115   }
3116 }
3117 
3118 static void __kmp_free_team_arrays(kmp_team_t *team) {
3119   /* Note: this does not free the threads in t_threads (__kmp_free_threads) */
3120   int i;
3121   for (i = 0; i < team->t.t_max_nproc; ++i) {
3122     if (team->t.t_dispatch[i].th_disp_buffer != NULL) {
3123       __kmp_free(team->t.t_dispatch[i].th_disp_buffer);
3124       team->t.t_dispatch[i].th_disp_buffer = NULL;
3125     }
3126   }
3127 #if KMP_USE_HIER_SCHED
3128   __kmp_dispatch_free_hierarchies(team);
3129 #endif
3130   __kmp_free(team->t.t_threads);
3131   __kmp_free(team->t.t_disp_buffer);
3132   __kmp_free(team->t.t_dispatch);
3133   __kmp_free(team->t.t_implicit_task_taskdata);
3134   team->t.t_threads = NULL;
3135   team->t.t_disp_buffer = NULL;
3136   team->t.t_dispatch = NULL;
3137   team->t.t_implicit_task_taskdata = 0;
3138 }
3139 
3140 static void __kmp_reallocate_team_arrays(kmp_team_t *team, int max_nth) {
3141   kmp_info_t **oldThreads = team->t.t_threads;
3142 
3143   __kmp_free(team->t.t_disp_buffer);
3144   __kmp_free(team->t.t_dispatch);
3145   __kmp_free(team->t.t_implicit_task_taskdata);
3146   __kmp_allocate_team_arrays(team, max_nth);
3147 
3148   KMP_MEMCPY(team->t.t_threads, oldThreads,
3149              team->t.t_nproc * sizeof(kmp_info_t *));
3150 
3151   __kmp_free(oldThreads);
3152 }
3153 
3154 static kmp_internal_control_t __kmp_get_global_icvs(void) {
3155 
3156   kmp_r_sched_t r_sched =
3157       __kmp_get_schedule_global(); // get current state of scheduling globals
3158 
3159 #if OMP_40_ENABLED
3160   KMP_DEBUG_ASSERT(__kmp_nested_proc_bind.used > 0);
3161 #endif /* OMP_40_ENABLED */
3162 
3163   kmp_internal_control_t g_icvs = {
3164     0, // int serial_nesting_level; //corresponds to value of th_team_serialized
3165     (kmp_int8)__kmp_global.g.g_dynamic, // internal control for dynamic
3166     // adjustment of threads (per thread)
3167     (kmp_int8)__kmp_env_blocktime, // int bt_set; //internal control for
3168     // whether blocktime is explicitly set
3169     __kmp_dflt_blocktime, // int blocktime; //internal control for blocktime
3170 #if KMP_USE_MONITOR
3171     __kmp_bt_intervals, // int bt_intervals; //internal control for blocktime
3172 // intervals
3173 #endif
3174     __kmp_dflt_team_nth, // int nproc; //internal control for # of threads for
3175     // next parallel region (per thread)
3176     // (use a max ub on value if __kmp_parallel_initialize not called yet)
3177     __kmp_cg_max_nth, // int thread_limit;
3178     __kmp_dflt_max_active_levels, // int max_active_levels; //internal control
3179     // for max_active_levels
3180     r_sched, // kmp_r_sched_t sched; //internal control for runtime schedule
3181 // {sched,chunk} pair
3182 #if OMP_40_ENABLED
3183     __kmp_nested_proc_bind.bind_types[0],
3184     __kmp_default_device,
3185 #endif /* OMP_40_ENABLED */
3186     NULL // struct kmp_internal_control *next;
3187   };
3188 
3189   return g_icvs;
3190 }
3191 
3192 static kmp_internal_control_t __kmp_get_x_global_icvs(const kmp_team_t *team) {
3193 
3194   kmp_internal_control_t gx_icvs;
3195   gx_icvs.serial_nesting_level =
3196       0; // probably =team->t.t_serial like in save_inter_controls
3197   copy_icvs(&gx_icvs, &team->t.t_threads[0]->th.th_current_task->td_icvs);
3198   gx_icvs.next = NULL;
3199 
3200   return gx_icvs;
3201 }
3202 
3203 static void __kmp_initialize_root(kmp_root_t *root) {
3204   int f;
3205   kmp_team_t *root_team;
3206   kmp_team_t *hot_team;
3207   int hot_team_max_nth;
3208   kmp_r_sched_t r_sched =
3209       __kmp_get_schedule_global(); // get current state of scheduling globals
3210   kmp_internal_control_t r_icvs = __kmp_get_global_icvs();
3211   KMP_DEBUG_ASSERT(root);
3212   KMP_ASSERT(!root->r.r_begin);
3213 
3214   /* setup the root state structure */
3215   __kmp_init_lock(&root->r.r_begin_lock);
3216   root->r.r_begin = FALSE;
3217   root->r.r_active = FALSE;
3218   root->r.r_in_parallel = 0;
3219   root->r.r_blocktime = __kmp_dflt_blocktime;
3220 
3221   /* setup the root team for this task */
3222   /* allocate the root team structure */
3223   KF_TRACE(10, ("__kmp_initialize_root: before root_team\n"));
3224 
3225   root_team =
3226       __kmp_allocate_team(root,
3227                           1, // new_nproc
3228                           1, // max_nproc
3229 #if OMPT_SUPPORT
3230                           ompt_data_none, // root parallel id
3231 #endif
3232 #if OMP_40_ENABLED
3233                           __kmp_nested_proc_bind.bind_types[0],
3234 #endif
3235                           &r_icvs,
3236                           0 // argc
3237                           USE_NESTED_HOT_ARG(NULL) // master thread is unknown
3238                           );
3239 #if USE_DEBUGGER
3240   // Non-NULL value should be assigned to make the debugger display the root
3241   // team.
3242   TCW_SYNC_PTR(root_team->t.t_pkfn, (microtask_t)(~0));
3243 #endif
3244 
3245   KF_TRACE(10, ("__kmp_initialize_root: after root_team = %p\n", root_team));
3246 
3247   root->r.r_root_team = root_team;
3248   root_team->t.t_control_stack_top = NULL;
3249 
3250   /* initialize root team */
3251   root_team->t.t_threads[0] = NULL;
3252   root_team->t.t_nproc = 1;
3253   root_team->t.t_serialized = 1;
3254   // TODO???: root_team->t.t_max_active_levels = __kmp_dflt_max_active_levels;
3255   root_team->t.t_sched.sched = r_sched.sched;
3256   KA_TRACE(
3257       20,
3258       ("__kmp_initialize_root: init root team %d arrived: join=%u, plain=%u\n",
3259        root_team->t.t_id, KMP_INIT_BARRIER_STATE, KMP_INIT_BARRIER_STATE));
3260 
3261   /* setup the  hot team for this task */
3262   /* allocate the hot team structure */
3263   KF_TRACE(10, ("__kmp_initialize_root: before hot_team\n"));
3264 
3265   hot_team =
3266       __kmp_allocate_team(root,
3267                           1, // new_nproc
3268                           __kmp_dflt_team_nth_ub * 2, // max_nproc
3269 #if OMPT_SUPPORT
3270                           ompt_data_none, // root parallel id
3271 #endif
3272 #if OMP_40_ENABLED
3273                           __kmp_nested_proc_bind.bind_types[0],
3274 #endif
3275                           &r_icvs,
3276                           0 // argc
3277                           USE_NESTED_HOT_ARG(NULL) // master thread is unknown
3278                           );
3279   KF_TRACE(10, ("__kmp_initialize_root: after hot_team = %p\n", hot_team));
3280 
3281   root->r.r_hot_team = hot_team;
3282   root_team->t.t_control_stack_top = NULL;
3283 
3284   /* first-time initialization */
3285   hot_team->t.t_parent = root_team;
3286 
3287   /* initialize hot team */
3288   hot_team_max_nth = hot_team->t.t_max_nproc;
3289   for (f = 0; f < hot_team_max_nth; ++f) {
3290     hot_team->t.t_threads[f] = NULL;
3291   }
3292   hot_team->t.t_nproc = 1;
3293   // TODO???: hot_team->t.t_max_active_levels = __kmp_dflt_max_active_levels;
3294   hot_team->t.t_sched.sched = r_sched.sched;
3295   hot_team->t.t_size_changed = 0;
3296 }
3297 
3298 #ifdef KMP_DEBUG
3299 
3300 typedef struct kmp_team_list_item {
3301   kmp_team_p const *entry;
3302   struct kmp_team_list_item *next;
3303 } kmp_team_list_item_t;
3304 typedef kmp_team_list_item_t *kmp_team_list_t;
3305 
3306 static void __kmp_print_structure_team_accum( // Add team to list of teams.
3307     kmp_team_list_t list, // List of teams.
3308     kmp_team_p const *team // Team to add.
3309     ) {
3310 
3311   // List must terminate with item where both entry and next are NULL.
3312   // Team is added to the list only once.
3313   // List is sorted in ascending order by team id.
3314   // Team id is *not* a key.
3315 
3316   kmp_team_list_t l;
3317 
3318   KMP_DEBUG_ASSERT(list != NULL);
3319   if (team == NULL) {
3320     return;
3321   }
3322 
3323   __kmp_print_structure_team_accum(list, team->t.t_parent);
3324   __kmp_print_structure_team_accum(list, team->t.t_next_pool);
3325 
3326   // Search list for the team.
3327   l = list;
3328   while (l->next != NULL && l->entry != team) {
3329     l = l->next;
3330   }
3331   if (l->next != NULL) {
3332     return; // Team has been added before, exit.
3333   }
3334 
3335   // Team is not found. Search list again for insertion point.
3336   l = list;
3337   while (l->next != NULL && l->entry->t.t_id <= team->t.t_id) {
3338     l = l->next;
3339   }
3340 
3341   // Insert team.
3342   {
3343     kmp_team_list_item_t *item = (kmp_team_list_item_t *)KMP_INTERNAL_MALLOC(
3344         sizeof(kmp_team_list_item_t));
3345     *item = *l;
3346     l->entry = team;
3347     l->next = item;
3348   }
3349 }
3350 
3351 static void __kmp_print_structure_team(char const *title, kmp_team_p const *team
3352 
3353                                        ) {
3354   __kmp_printf("%s", title);
3355   if (team != NULL) {
3356     __kmp_printf("%2x %p\n", team->t.t_id, team);
3357   } else {
3358     __kmp_printf(" - (nil)\n");
3359   }
3360 }
3361 
3362 static void __kmp_print_structure_thread(char const *title,
3363                                          kmp_info_p const *thread) {
3364   __kmp_printf("%s", title);
3365   if (thread != NULL) {
3366     __kmp_printf("%2d %p\n", thread->th.th_info.ds.ds_gtid, thread);
3367   } else {
3368     __kmp_printf(" - (nil)\n");
3369   }
3370 }
3371 
3372 void __kmp_print_structure(void) {
3373 
3374   kmp_team_list_t list;
3375 
3376   // Initialize list of teams.
3377   list =
3378       (kmp_team_list_item_t *)KMP_INTERNAL_MALLOC(sizeof(kmp_team_list_item_t));
3379   list->entry = NULL;
3380   list->next = NULL;
3381 
3382   __kmp_printf("\n------------------------------\nGlobal Thread "
3383                "Table\n------------------------------\n");
3384   {
3385     int gtid;
3386     for (gtid = 0; gtid < __kmp_threads_capacity; ++gtid) {
3387       __kmp_printf("%2d", gtid);
3388       if (__kmp_threads != NULL) {
3389         __kmp_printf(" %p", __kmp_threads[gtid]);
3390       }
3391       if (__kmp_root != NULL) {
3392         __kmp_printf(" %p", __kmp_root[gtid]);
3393       }
3394       __kmp_printf("\n");
3395     }
3396   }
3397 
3398   // Print out __kmp_threads array.
3399   __kmp_printf("\n------------------------------\nThreads\n--------------------"
3400                "----------\n");
3401   if (__kmp_threads != NULL) {
3402     int gtid;
3403     for (gtid = 0; gtid < __kmp_threads_capacity; ++gtid) {
3404       kmp_info_t const *thread = __kmp_threads[gtid];
3405       if (thread != NULL) {
3406         __kmp_printf("GTID %2d %p:\n", gtid, thread);
3407         __kmp_printf("    Our Root:        %p\n", thread->th.th_root);
3408         __kmp_print_structure_team("    Our Team:     ", thread->th.th_team);
3409         __kmp_print_structure_team("    Serial Team:  ",
3410                                    thread->th.th_serial_team);
3411         __kmp_printf("    Threads:      %2d\n", thread->th.th_team_nproc);
3412         __kmp_print_structure_thread("    Master:       ",
3413                                      thread->th.th_team_master);
3414         __kmp_printf("    Serialized?:  %2d\n", thread->th.th_team_serialized);
3415         __kmp_printf("    Set NProc:    %2d\n", thread->th.th_set_nproc);
3416 #if OMP_40_ENABLED
3417         __kmp_printf("    Set Proc Bind: %2d\n", thread->th.th_set_proc_bind);
3418 #endif
3419         __kmp_print_structure_thread("    Next in pool: ",
3420                                      thread->th.th_next_pool);
3421         __kmp_printf("\n");
3422         __kmp_print_structure_team_accum(list, thread->th.th_team);
3423         __kmp_print_structure_team_accum(list, thread->th.th_serial_team);
3424       }
3425     }
3426   } else {
3427     __kmp_printf("Threads array is not allocated.\n");
3428   }
3429 
3430   // Print out __kmp_root array.
3431   __kmp_printf("\n------------------------------\nUbers\n----------------------"
3432                "--------\n");
3433   if (__kmp_root != NULL) {
3434     int gtid;
3435     for (gtid = 0; gtid < __kmp_threads_capacity; ++gtid) {
3436       kmp_root_t const *root = __kmp_root[gtid];
3437       if (root != NULL) {
3438         __kmp_printf("GTID %2d %p:\n", gtid, root);
3439         __kmp_print_structure_team("    Root Team:    ", root->r.r_root_team);
3440         __kmp_print_structure_team("    Hot Team:     ", root->r.r_hot_team);
3441         __kmp_print_structure_thread("    Uber Thread:  ",
3442                                      root->r.r_uber_thread);
3443         __kmp_printf("    Active?:      %2d\n", root->r.r_active);
3444         __kmp_printf("    In Parallel:  %2d\n",
3445                      KMP_ATOMIC_LD_RLX(&root->r.r_in_parallel));
3446         __kmp_printf("\n");
3447         __kmp_print_structure_team_accum(list, root->r.r_root_team);
3448         __kmp_print_structure_team_accum(list, root->r.r_hot_team);
3449       }
3450     }
3451   } else {
3452     __kmp_printf("Ubers array is not allocated.\n");
3453   }
3454 
3455   __kmp_printf("\n------------------------------\nTeams\n----------------------"
3456                "--------\n");
3457   while (list->next != NULL) {
3458     kmp_team_p const *team = list->entry;
3459     int i;
3460     __kmp_printf("Team %2x %p:\n", team->t.t_id, team);
3461     __kmp_print_structure_team("    Parent Team:      ", team->t.t_parent);
3462     __kmp_printf("    Master TID:       %2d\n", team->t.t_master_tid);
3463     __kmp_printf("    Max threads:      %2d\n", team->t.t_max_nproc);
3464     __kmp_printf("    Levels of serial: %2d\n", team->t.t_serialized);
3465     __kmp_printf("    Number threads:   %2d\n", team->t.t_nproc);
3466     for (i = 0; i < team->t.t_nproc; ++i) {
3467       __kmp_printf("    Thread %2d:      ", i);
3468       __kmp_print_structure_thread("", team->t.t_threads[i]);
3469     }
3470     __kmp_print_structure_team("    Next in pool:     ", team->t.t_next_pool);
3471     __kmp_printf("\n");
3472     list = list->next;
3473   }
3474 
3475   // Print out __kmp_thread_pool and __kmp_team_pool.
3476   __kmp_printf("\n------------------------------\nPools\n----------------------"
3477                "--------\n");
3478   __kmp_print_structure_thread("Thread pool:          ",
3479                                CCAST(kmp_info_t *, __kmp_thread_pool));
3480   __kmp_print_structure_team("Team pool:            ",
3481                              CCAST(kmp_team_t *, __kmp_team_pool));
3482   __kmp_printf("\n");
3483 
3484   // Free team list.
3485   while (list != NULL) {
3486     kmp_team_list_item_t *item = list;
3487     list = list->next;
3488     KMP_INTERNAL_FREE(item);
3489   }
3490 }
3491 
3492 #endif
3493 
3494 //---------------------------------------------------------------------------
3495 //  Stuff for per-thread fast random number generator
3496 //  Table of primes
3497 static const unsigned __kmp_primes[] = {
3498     0x9e3779b1, 0xffe6cc59, 0x2109f6dd, 0x43977ab5, 0xba5703f5, 0xb495a877,
3499     0xe1626741, 0x79695e6b, 0xbc98c09f, 0xd5bee2b3, 0x287488f9, 0x3af18231,
3500     0x9677cd4d, 0xbe3a6929, 0xadc6a877, 0xdcf0674b, 0xbe4d6fe9, 0x5f15e201,
3501     0x99afc3fd, 0xf3f16801, 0xe222cfff, 0x24ba5fdb, 0x0620452d, 0x79f149e3,
3502     0xc8b93f49, 0x972702cd, 0xb07dd827, 0x6c97d5ed, 0x085a3d61, 0x46eb5ea7,
3503     0x3d9910ed, 0x2e687b5b, 0x29609227, 0x6eb081f1, 0x0954c4e1, 0x9d114db9,
3504     0x542acfa9, 0xb3e6bd7b, 0x0742d917, 0xe9f3ffa7, 0x54581edb, 0xf2480f45,
3505     0x0bb9288f, 0xef1affc7, 0x85fa0ca7, 0x3ccc14db, 0xe6baf34b, 0x343377f7,
3506     0x5ca19031, 0xe6d9293b, 0xf0a9f391, 0x5d2e980b, 0xfc411073, 0xc3749363,
3507     0xb892d829, 0x3549366b, 0x629750ad, 0xb98294e5, 0x892d9483, 0xc235baf3,
3508     0x3d2402a3, 0x6bdef3c9, 0xbec333cd, 0x40c9520f};
3509 
3510 //---------------------------------------------------------------------------
3511 //  __kmp_get_random: Get a random number using a linear congruential method.
3512 unsigned short __kmp_get_random(kmp_info_t *thread) {
3513   unsigned x = thread->th.th_x;
3514   unsigned short r = x >> 16;
3515 
3516   thread->th.th_x = x * thread->th.th_a + 1;
3517 
3518   KA_TRACE(30, ("__kmp_get_random: THREAD: %d, RETURN: %u\n",
3519                 thread->th.th_info.ds.ds_tid, r));
3520 
3521   return r;
3522 }
3523 //--------------------------------------------------------
3524 // __kmp_init_random: Initialize a random number generator
3525 void __kmp_init_random(kmp_info_t *thread) {
3526   unsigned seed = thread->th.th_info.ds.ds_tid;
3527 
3528   thread->th.th_a =
3529       __kmp_primes[seed % (sizeof(__kmp_primes) / sizeof(__kmp_primes[0]))];
3530   thread->th.th_x = (seed + 1) * thread->th.th_a + 1;
3531   KA_TRACE(30,
3532            ("__kmp_init_random: THREAD: %u; A: %u\n", seed, thread->th.th_a));
3533 }
3534 
3535 #if KMP_OS_WINDOWS
3536 /* reclaim array entries for root threads that are already dead, returns number
3537  * reclaimed */
3538 static int __kmp_reclaim_dead_roots(void) {
3539   int i, r = 0;
3540 
3541   for (i = 0; i < __kmp_threads_capacity; ++i) {
3542     if (KMP_UBER_GTID(i) &&
3543         !__kmp_still_running((kmp_info_t *)TCR_SYNC_PTR(__kmp_threads[i])) &&
3544         !__kmp_root[i]
3545              ->r.r_active) { // AC: reclaim only roots died in non-active state
3546       r += __kmp_unregister_root_other_thread(i);
3547     }
3548   }
3549   return r;
3550 }
3551 #endif
3552 
3553 /* This function attempts to create free entries in __kmp_threads and
3554    __kmp_root, and returns the number of free entries generated.
3555 
3556    For Windows* OS static library, the first mechanism used is to reclaim array
3557    entries for root threads that are already dead.
3558 
3559    On all platforms, expansion is attempted on the arrays __kmp_threads_ and
3560    __kmp_root, with appropriate update to __kmp_threads_capacity. Array
3561    capacity is increased by doubling with clipping to __kmp_tp_capacity, if
3562    threadprivate cache array has been created. Synchronization with
3563    __kmpc_threadprivate_cached is done using __kmp_tp_cached_lock.
3564 
3565    After any dead root reclamation, if the clipping value allows array expansion
3566    to result in the generation of a total of nNeed free slots, the function does
3567    that expansion. If not, nothing is done beyond the possible initial root
3568    thread reclamation.
3569 
3570    If any argument is negative, the behavior is undefined. */
3571 static int __kmp_expand_threads(int nNeed) {
3572   int added = 0;
3573   int minimumRequiredCapacity;
3574   int newCapacity;
3575   kmp_info_t **newThreads;
3576   kmp_root_t **newRoot;
3577 
3578 // All calls to __kmp_expand_threads should be under __kmp_forkjoin_lock, so
3579 // resizing __kmp_threads does not need additional protection if foreign
3580 // threads are present
3581 
3582 #if KMP_OS_WINDOWS && !KMP_DYNAMIC_LIB
3583   /* only for Windows static library */
3584   /* reclaim array entries for root threads that are already dead */
3585   added = __kmp_reclaim_dead_roots();
3586 
3587   if (nNeed) {
3588     nNeed -= added;
3589     if (nNeed < 0)
3590       nNeed = 0;
3591   }
3592 #endif
3593   if (nNeed <= 0)
3594     return added;
3595 
3596   // Note that __kmp_threads_capacity is not bounded by __kmp_max_nth. If
3597   // __kmp_max_nth is set to some value less than __kmp_sys_max_nth by the
3598   // user via KMP_DEVICE_THREAD_LIMIT, then __kmp_threads_capacity may become
3599   // > __kmp_max_nth in one of two ways:
3600   //
3601   // 1) The initialization thread (gtid = 0) exits.  __kmp_threads[0]
3602   //    may not be resused by another thread, so we may need to increase
3603   //    __kmp_threads_capacity to __kmp_max_nth + 1.
3604   //
3605   // 2) New foreign root(s) are encountered.  We always register new foreign
3606   //    roots. This may cause a smaller # of threads to be allocated at
3607   //    subsequent parallel regions, but the worker threads hang around (and
3608   //    eventually go to sleep) and need slots in the __kmp_threads[] array.
3609   //
3610   // Anyway, that is the reason for moving the check to see if
3611   // __kmp_max_nth was exceeded into __kmp_reserve_threads()
3612   // instead of having it performed here. -BB
3613 
3614   KMP_DEBUG_ASSERT(__kmp_sys_max_nth >= __kmp_threads_capacity);
3615 
3616   /* compute expansion headroom to check if we can expand */
3617   if (__kmp_sys_max_nth - __kmp_threads_capacity < nNeed) {
3618     /* possible expansion too small -- give up */
3619     return added;
3620   }
3621   minimumRequiredCapacity = __kmp_threads_capacity + nNeed;
3622 
3623   newCapacity = __kmp_threads_capacity;
3624   do {
3625     newCapacity = newCapacity <= (__kmp_sys_max_nth >> 1) ? (newCapacity << 1)
3626                                                           : __kmp_sys_max_nth;
3627   } while (newCapacity < minimumRequiredCapacity);
3628   newThreads = (kmp_info_t **)__kmp_allocate(
3629       (sizeof(kmp_info_t *) + sizeof(kmp_root_t *)) * newCapacity + CACHE_LINE);
3630   newRoot =
3631       (kmp_root_t **)((char *)newThreads + sizeof(kmp_info_t *) * newCapacity);
3632   KMP_MEMCPY(newThreads, __kmp_threads,
3633              __kmp_threads_capacity * sizeof(kmp_info_t *));
3634   KMP_MEMCPY(newRoot, __kmp_root,
3635              __kmp_threads_capacity * sizeof(kmp_root_t *));
3636 
3637   kmp_info_t **temp_threads = __kmp_threads;
3638   *(kmp_info_t * *volatile *)&__kmp_threads = newThreads;
3639   *(kmp_root_t * *volatile *)&__kmp_root = newRoot;
3640   __kmp_free(temp_threads);
3641   added += newCapacity - __kmp_threads_capacity;
3642   *(volatile int *)&__kmp_threads_capacity = newCapacity;
3643 
3644   if (newCapacity > __kmp_tp_capacity) {
3645     __kmp_acquire_bootstrap_lock(&__kmp_tp_cached_lock);
3646     if (__kmp_tp_cached && newCapacity > __kmp_tp_capacity) {
3647       __kmp_threadprivate_resize_cache(newCapacity);
3648     } else { // increase __kmp_tp_capacity to correspond with kmp_threads size
3649       *(volatile int *)&__kmp_tp_capacity = newCapacity;
3650     }
3651     __kmp_release_bootstrap_lock(&__kmp_tp_cached_lock);
3652   }
3653 
3654   return added;
3655 }
3656 
3657 /* Register the current thread as a root thread and obtain our gtid. We must
3658    have the __kmp_initz_lock held at this point. Argument TRUE only if are the
3659    thread that calls from __kmp_do_serial_initialize() */
3660 int __kmp_register_root(int initial_thread) {
3661   kmp_info_t *root_thread;
3662   kmp_root_t *root;
3663   int gtid;
3664   int capacity;
3665   __kmp_acquire_bootstrap_lock(&__kmp_forkjoin_lock);
3666   KA_TRACE(20, ("__kmp_register_root: entered\n"));
3667   KMP_MB();
3668 
3669   /* 2007-03-02:
3670      If initial thread did not invoke OpenMP RTL yet, and this thread is not an
3671      initial one, "__kmp_all_nth >= __kmp_threads_capacity" condition does not
3672      work as expected -- it may return false (that means there is at least one
3673      empty slot in __kmp_threads array), but it is possible the only free slot
3674      is #0, which is reserved for initial thread and so cannot be used for this
3675      one. Following code workarounds this bug.
3676 
3677      However, right solution seems to be not reserving slot #0 for initial
3678      thread because:
3679      (1) there is no magic in slot #0,
3680      (2) we cannot detect initial thread reliably (the first thread which does
3681         serial initialization may be not a real initial thread).
3682   */
3683   capacity = __kmp_threads_capacity;
3684   if (!initial_thread && TCR_PTR(__kmp_threads[0]) == NULL) {
3685     --capacity;
3686   }
3687 
3688   /* see if there are too many threads */
3689   if (__kmp_all_nth >= capacity && !__kmp_expand_threads(1)) {
3690     if (__kmp_tp_cached) {
3691       __kmp_fatal(KMP_MSG(CantRegisterNewThread),
3692                   KMP_HNT(Set_ALL_THREADPRIVATE, __kmp_tp_capacity),
3693                   KMP_HNT(PossibleSystemLimitOnThreads), __kmp_msg_null);
3694     } else {
3695       __kmp_fatal(KMP_MSG(CantRegisterNewThread), KMP_HNT(SystemLimitOnThreads),
3696                   __kmp_msg_null);
3697     }
3698   }
3699 
3700   /* find an available thread slot */
3701   /* Don't reassign the zero slot since we need that to only be used by initial
3702      thread */
3703   for (gtid = (initial_thread ? 0 : 1); TCR_PTR(__kmp_threads[gtid]) != NULL;
3704        gtid++)
3705     ;
3706   KA_TRACE(1,
3707            ("__kmp_register_root: found slot in threads array: T#%d\n", gtid));
3708   KMP_ASSERT(gtid < __kmp_threads_capacity);
3709 
3710   /* update global accounting */
3711   __kmp_all_nth++;
3712   TCW_4(__kmp_nth, __kmp_nth + 1);
3713 
3714   // if __kmp_adjust_gtid_mode is set, then we use method #1 (sp search) for low
3715   // numbers of procs, and method #2 (keyed API call) for higher numbers.
3716   if (__kmp_adjust_gtid_mode) {
3717     if (__kmp_all_nth >= __kmp_tls_gtid_min) {
3718       if (TCR_4(__kmp_gtid_mode) != 2) {
3719         TCW_4(__kmp_gtid_mode, 2);
3720       }
3721     } else {
3722       if (TCR_4(__kmp_gtid_mode) != 1) {
3723         TCW_4(__kmp_gtid_mode, 1);
3724       }
3725     }
3726   }
3727 
3728 #ifdef KMP_ADJUST_BLOCKTIME
3729   /* Adjust blocktime to zero if necessary            */
3730   /* Middle initialization might not have occurred yet */
3731   if (!__kmp_env_blocktime && (__kmp_avail_proc > 0)) {
3732     if (__kmp_nth > __kmp_avail_proc) {
3733       __kmp_zero_bt = TRUE;
3734     }
3735   }
3736 #endif /* KMP_ADJUST_BLOCKTIME */
3737 
3738   /* setup this new hierarchy */
3739   if (!(root = __kmp_root[gtid])) {
3740     root = __kmp_root[gtid] = (kmp_root_t *)__kmp_allocate(sizeof(kmp_root_t));
3741     KMP_DEBUG_ASSERT(!root->r.r_root_team);
3742   }
3743 
3744 #if KMP_STATS_ENABLED
3745   // Initialize stats as soon as possible (right after gtid assignment).
3746   __kmp_stats_thread_ptr = __kmp_stats_list->push_back(gtid);
3747   __kmp_stats_thread_ptr->startLife();
3748   KMP_SET_THREAD_STATE(SERIAL_REGION);
3749   KMP_INIT_PARTITIONED_TIMERS(OMP_serial);
3750 #endif
3751   __kmp_initialize_root(root);
3752 
3753   /* setup new root thread structure */
3754   if (root->r.r_uber_thread) {
3755     root_thread = root->r.r_uber_thread;
3756   } else {
3757     root_thread = (kmp_info_t *)__kmp_allocate(sizeof(kmp_info_t));
3758     if (__kmp_storage_map) {
3759       __kmp_print_thread_storage_map(root_thread, gtid);
3760     }
3761     root_thread->th.th_info.ds.ds_gtid = gtid;
3762 #if OMPT_SUPPORT
3763     root_thread->th.ompt_thread_info.thread_data = ompt_data_none;
3764 #endif
3765     root_thread->th.th_root = root;
3766     if (__kmp_env_consistency_check) {
3767       root_thread->th.th_cons = __kmp_allocate_cons_stack(gtid);
3768     }
3769 #if USE_FAST_MEMORY
3770     __kmp_initialize_fast_memory(root_thread);
3771 #endif /* USE_FAST_MEMORY */
3772 
3773 #if KMP_USE_BGET
3774     KMP_DEBUG_ASSERT(root_thread->th.th_local.bget_data == NULL);
3775     __kmp_initialize_bget(root_thread);
3776 #endif
3777     __kmp_init_random(root_thread); // Initialize random number generator
3778   }
3779 
3780   /* setup the serial team held in reserve by the root thread */
3781   if (!root_thread->th.th_serial_team) {
3782     kmp_internal_control_t r_icvs = __kmp_get_global_icvs();
3783     KF_TRACE(10, ("__kmp_register_root: before serial_team\n"));
3784     root_thread->th.th_serial_team =
3785         __kmp_allocate_team(root, 1, 1,
3786 #if OMPT_SUPPORT
3787                             ompt_data_none, // root parallel id
3788 #endif
3789 #if OMP_40_ENABLED
3790                             proc_bind_default,
3791 #endif
3792                             &r_icvs, 0 USE_NESTED_HOT_ARG(NULL));
3793   }
3794   KMP_ASSERT(root_thread->th.th_serial_team);
3795   KF_TRACE(10, ("__kmp_register_root: after serial_team = %p\n",
3796                 root_thread->th.th_serial_team));
3797 
3798   /* drop root_thread into place */
3799   TCW_SYNC_PTR(__kmp_threads[gtid], root_thread);
3800 
3801   root->r.r_root_team->t.t_threads[0] = root_thread;
3802   root->r.r_hot_team->t.t_threads[0] = root_thread;
3803   root_thread->th.th_serial_team->t.t_threads[0] = root_thread;
3804   // AC: the team created in reserve, not for execution (it is unused for now).
3805   root_thread->th.th_serial_team->t.t_serialized = 0;
3806   root->r.r_uber_thread = root_thread;
3807 
3808   /* initialize the thread, get it ready to go */
3809   __kmp_initialize_info(root_thread, root->r.r_root_team, 0, gtid);
3810   TCW_4(__kmp_init_gtid, TRUE);
3811 
3812   /* prepare the master thread for get_gtid() */
3813   __kmp_gtid_set_specific(gtid);
3814 
3815 #if USE_ITT_BUILD
3816   __kmp_itt_thread_name(gtid);
3817 #endif /* USE_ITT_BUILD */
3818 
3819 #ifdef KMP_TDATA_GTID
3820   __kmp_gtid = gtid;
3821 #endif
3822   __kmp_create_worker(gtid, root_thread, __kmp_stksize);
3823   KMP_DEBUG_ASSERT(__kmp_gtid_get_specific() == gtid);
3824 
3825   KA_TRACE(20, ("__kmp_register_root: T#%d init T#%d(%d:%d) arrived: join=%u, "
3826                 "plain=%u\n",
3827                 gtid, __kmp_gtid_from_tid(0, root->r.r_hot_team),
3828                 root->r.r_hot_team->t.t_id, 0, KMP_INIT_BARRIER_STATE,
3829                 KMP_INIT_BARRIER_STATE));
3830   { // Initialize barrier data.
3831     int b;
3832     for (b = 0; b < bs_last_barrier; ++b) {
3833       root_thread->th.th_bar[b].bb.b_arrived = KMP_INIT_BARRIER_STATE;
3834 #if USE_DEBUGGER
3835       root_thread->th.th_bar[b].bb.b_worker_arrived = 0;
3836 #endif
3837     }
3838   }
3839   KMP_DEBUG_ASSERT(root->r.r_hot_team->t.t_bar[bs_forkjoin_barrier].b_arrived ==
3840                    KMP_INIT_BARRIER_STATE);
3841 
3842 #if KMP_AFFINITY_SUPPORTED
3843 #if OMP_40_ENABLED
3844   root_thread->th.th_current_place = KMP_PLACE_UNDEFINED;
3845   root_thread->th.th_new_place = KMP_PLACE_UNDEFINED;
3846   root_thread->th.th_first_place = KMP_PLACE_UNDEFINED;
3847   root_thread->th.th_last_place = KMP_PLACE_UNDEFINED;
3848 #endif
3849   if (TCR_4(__kmp_init_middle)) {
3850     __kmp_affinity_set_init_mask(gtid, TRUE);
3851   }
3852 #endif /* KMP_AFFINITY_SUPPORTED */
3853 #if OMP_50_ENABLED
3854   root_thread->th.th_def_allocator = __kmp_def_allocator;
3855   root_thread->th.th_prev_level = 0;
3856   root_thread->th.th_prev_num_threads = 1;
3857 #endif
3858 
3859   kmp_cg_root_t *tmp = (kmp_cg_root_t *)__kmp_allocate(sizeof(kmp_cg_root_t));
3860   tmp->cg_root = root_thread;
3861   tmp->cg_thread_limit = __kmp_cg_max_nth;
3862   tmp->cg_nthreads = 1;
3863   KA_TRACE(100, ("__kmp_register_root: Thread %p created node %p with"
3864                  " cg_nthreads init to 1\n",
3865                  root_thread, tmp));
3866   tmp->up = NULL;
3867   root_thread->th.th_cg_roots = tmp;
3868 
3869   __kmp_root_counter++;
3870 
3871 #if OMPT_SUPPORT
3872   if (!initial_thread && ompt_enabled.enabled) {
3873 
3874     kmp_info_t *root_thread = ompt_get_thread();
3875 
3876     ompt_set_thread_state(root_thread, ompt_state_overhead);
3877 
3878     if (ompt_enabled.ompt_callback_thread_begin) {
3879       ompt_callbacks.ompt_callback(ompt_callback_thread_begin)(
3880           ompt_thread_initial, __ompt_get_thread_data_internal());
3881     }
3882     ompt_data_t *task_data;
3883     __ompt_get_task_info_internal(0, NULL, &task_data, NULL, NULL, NULL);
3884     if (ompt_enabled.ompt_callback_task_create) {
3885       ompt_callbacks.ompt_callback(ompt_callback_task_create)(
3886           NULL, NULL, task_data, ompt_task_initial, 0, NULL);
3887       // initial task has nothing to return to
3888     }
3889 
3890     ompt_set_thread_state(root_thread, ompt_state_work_serial);
3891   }
3892 #endif
3893 
3894   KMP_MB();
3895   __kmp_release_bootstrap_lock(&__kmp_forkjoin_lock);
3896 
3897   return gtid;
3898 }
3899 
3900 #if KMP_NESTED_HOT_TEAMS
3901 static int __kmp_free_hot_teams(kmp_root_t *root, kmp_info_t *thr, int level,
3902                                 const int max_level) {
3903   int i, n, nth;
3904   kmp_hot_team_ptr_t *hot_teams = thr->th.th_hot_teams;
3905   if (!hot_teams || !hot_teams[level].hot_team) {
3906     return 0;
3907   }
3908   KMP_DEBUG_ASSERT(level < max_level);
3909   kmp_team_t *team = hot_teams[level].hot_team;
3910   nth = hot_teams[level].hot_team_nth;
3911   n = nth - 1; // master is not freed
3912   if (level < max_level - 1) {
3913     for (i = 0; i < nth; ++i) {
3914       kmp_info_t *th = team->t.t_threads[i];
3915       n += __kmp_free_hot_teams(root, th, level + 1, max_level);
3916       if (i > 0 && th->th.th_hot_teams) {
3917         __kmp_free(th->th.th_hot_teams);
3918         th->th.th_hot_teams = NULL;
3919       }
3920     }
3921   }
3922   __kmp_free_team(root, team, NULL);
3923   return n;
3924 }
3925 #endif
3926 
3927 // Resets a root thread and clear its root and hot teams.
3928 // Returns the number of __kmp_threads entries directly and indirectly freed.
3929 static int __kmp_reset_root(int gtid, kmp_root_t *root) {
3930   kmp_team_t *root_team = root->r.r_root_team;
3931   kmp_team_t *hot_team = root->r.r_hot_team;
3932   int n = hot_team->t.t_nproc;
3933   int i;
3934 
3935   KMP_DEBUG_ASSERT(!root->r.r_active);
3936 
3937   root->r.r_root_team = NULL;
3938   root->r.r_hot_team = NULL;
3939   // __kmp_free_team() does not free hot teams, so we have to clear r_hot_team
3940   // before call to __kmp_free_team().
3941   __kmp_free_team(root, root_team USE_NESTED_HOT_ARG(NULL));
3942 #if KMP_NESTED_HOT_TEAMS
3943   if (__kmp_hot_teams_max_level >
3944       0) { // need to free nested hot teams and their threads if any
3945     for (i = 0; i < hot_team->t.t_nproc; ++i) {
3946       kmp_info_t *th = hot_team->t.t_threads[i];
3947       if (__kmp_hot_teams_max_level > 1) {
3948         n += __kmp_free_hot_teams(root, th, 1, __kmp_hot_teams_max_level);
3949       }
3950       if (th->th.th_hot_teams) {
3951         __kmp_free(th->th.th_hot_teams);
3952         th->th.th_hot_teams = NULL;
3953       }
3954     }
3955   }
3956 #endif
3957   __kmp_free_team(root, hot_team USE_NESTED_HOT_ARG(NULL));
3958 
3959   // Before we can reap the thread, we need to make certain that all other
3960   // threads in the teams that had this root as ancestor have stopped trying to
3961   // steal tasks.
3962   if (__kmp_tasking_mode != tskm_immediate_exec) {
3963     __kmp_wait_to_unref_task_teams();
3964   }
3965 
3966 #if KMP_OS_WINDOWS
3967   /* Close Handle of root duplicated in __kmp_create_worker (tr #62919) */
3968   KA_TRACE(
3969       10, ("__kmp_reset_root: free handle, th = %p, handle = %" KMP_UINTPTR_SPEC
3970            "\n",
3971            (LPVOID) & (root->r.r_uber_thread->th),
3972            root->r.r_uber_thread->th.th_info.ds.ds_thread));
3973   __kmp_free_handle(root->r.r_uber_thread->th.th_info.ds.ds_thread);
3974 #endif /* KMP_OS_WINDOWS */
3975 
3976 #if OMPT_SUPPORT
3977   if (ompt_enabled.ompt_callback_thread_end) {
3978     ompt_callbacks.ompt_callback(ompt_callback_thread_end)(
3979         &(root->r.r_uber_thread->th.ompt_thread_info.thread_data));
3980   }
3981 #endif
3982 
3983   TCW_4(__kmp_nth,
3984         __kmp_nth - 1); // __kmp_reap_thread will decrement __kmp_all_nth.
3985   i = root->r.r_uber_thread->th.th_cg_roots->cg_nthreads--;
3986   KA_TRACE(100, ("__kmp_reset_root: Thread %p decrement cg_nthreads on node %p"
3987                  " to %d\n",
3988                  root->r.r_uber_thread, root->r.r_uber_thread->th.th_cg_roots,
3989                  root->r.r_uber_thread->th.th_cg_roots->cg_nthreads));
3990   if (i == 1) {
3991     // need to free contention group structure
3992     KMP_DEBUG_ASSERT(root->r.r_uber_thread ==
3993                      root->r.r_uber_thread->th.th_cg_roots->cg_root);
3994     KMP_DEBUG_ASSERT(root->r.r_uber_thread->th.th_cg_roots->up == NULL);
3995     __kmp_free(root->r.r_uber_thread->th.th_cg_roots);
3996     root->r.r_uber_thread->th.th_cg_roots = NULL;
3997   }
3998   __kmp_reap_thread(root->r.r_uber_thread, 1);
3999 
4000   // We canot put root thread to __kmp_thread_pool, so we have to reap it istead
4001   // of freeing.
4002   root->r.r_uber_thread = NULL;
4003   /* mark root as no longer in use */
4004   root->r.r_begin = FALSE;
4005 
4006   return n;
4007 }
4008 
4009 void __kmp_unregister_root_current_thread(int gtid) {
4010   KA_TRACE(1, ("__kmp_unregister_root_current_thread: enter T#%d\n", gtid));
4011   /* this lock should be ok, since unregister_root_current_thread is never
4012      called during an abort, only during a normal close. furthermore, if you
4013      have the forkjoin lock, you should never try to get the initz lock */
4014   __kmp_acquire_bootstrap_lock(&__kmp_forkjoin_lock);
4015   if (TCR_4(__kmp_global.g.g_done) || !__kmp_init_serial) {
4016     KC_TRACE(10, ("__kmp_unregister_root_current_thread: already finished, "
4017                   "exiting T#%d\n",
4018                   gtid));
4019     __kmp_release_bootstrap_lock(&__kmp_forkjoin_lock);
4020     return;
4021   }
4022   kmp_root_t *root = __kmp_root[gtid];
4023 
4024   KMP_DEBUG_ASSERT(__kmp_threads && __kmp_threads[gtid]);
4025   KMP_ASSERT(KMP_UBER_GTID(gtid));
4026   KMP_ASSERT(root == __kmp_threads[gtid]->th.th_root);
4027   KMP_ASSERT(root->r.r_active == FALSE);
4028 
4029   KMP_MB();
4030 
4031 #if OMP_45_ENABLED
4032   kmp_info_t *thread = __kmp_threads[gtid];
4033   kmp_team_t *team = thread->th.th_team;
4034   kmp_task_team_t *task_team = thread->th.th_task_team;
4035 
4036   // we need to wait for the proxy tasks before finishing the thread
4037   if (task_team != NULL && task_team->tt.tt_found_proxy_tasks) {
4038 #if OMPT_SUPPORT
4039     // the runtime is shutting down so we won't report any events
4040     thread->th.ompt_thread_info.state = ompt_state_undefined;
4041 #endif
4042     __kmp_task_team_wait(thread, team USE_ITT_BUILD_ARG(NULL));
4043   }
4044 #endif
4045 
4046   __kmp_reset_root(gtid, root);
4047 
4048   /* free up this thread slot */
4049   __kmp_gtid_set_specific(KMP_GTID_DNE);
4050 #ifdef KMP_TDATA_GTID
4051   __kmp_gtid = KMP_GTID_DNE;
4052 #endif
4053 
4054   KMP_MB();
4055   KC_TRACE(10,
4056            ("__kmp_unregister_root_current_thread: T#%d unregistered\n", gtid));
4057 
4058   __kmp_release_bootstrap_lock(&__kmp_forkjoin_lock);
4059 }
4060 
4061 #if KMP_OS_WINDOWS
4062 /* __kmp_forkjoin_lock must be already held
4063    Unregisters a root thread that is not the current thread.  Returns the number
4064    of __kmp_threads entries freed as a result. */
4065 static int __kmp_unregister_root_other_thread(int gtid) {
4066   kmp_root_t *root = __kmp_root[gtid];
4067   int r;
4068 
4069   KA_TRACE(1, ("__kmp_unregister_root_other_thread: enter T#%d\n", gtid));
4070   KMP_DEBUG_ASSERT(__kmp_threads && __kmp_threads[gtid]);
4071   KMP_ASSERT(KMP_UBER_GTID(gtid));
4072   KMP_ASSERT(root == __kmp_threads[gtid]->th.th_root);
4073   KMP_ASSERT(root->r.r_active == FALSE);
4074 
4075   r = __kmp_reset_root(gtid, root);
4076   KC_TRACE(10,
4077            ("__kmp_unregister_root_other_thread: T#%d unregistered\n", gtid));
4078   return r;
4079 }
4080 #endif
4081 
4082 #if KMP_DEBUG
4083 void __kmp_task_info() {
4084 
4085   kmp_int32 gtid = __kmp_entry_gtid();
4086   kmp_int32 tid = __kmp_tid_from_gtid(gtid);
4087   kmp_info_t *this_thr = __kmp_threads[gtid];
4088   kmp_team_t *steam = this_thr->th.th_serial_team;
4089   kmp_team_t *team = this_thr->th.th_team;
4090 
4091   __kmp_printf(
4092       "__kmp_task_info: gtid=%d tid=%d t_thread=%p team=%p steam=%p curtask=%p "
4093       "ptask=%p\n",
4094       gtid, tid, this_thr, team, steam, this_thr->th.th_current_task,
4095       team->t.t_implicit_task_taskdata[tid].td_parent);
4096 }
4097 #endif // KMP_DEBUG
4098 
4099 /* TODO optimize with one big memclr, take out what isn't needed, split
4100    responsibility to workers as much as possible, and delay initialization of
4101    features as much as possible  */
4102 static void __kmp_initialize_info(kmp_info_t *this_thr, kmp_team_t *team,
4103                                   int tid, int gtid) {
4104   /* this_thr->th.th_info.ds.ds_gtid is setup in
4105      kmp_allocate_thread/create_worker.
4106      this_thr->th.th_serial_team is setup in __kmp_allocate_thread */
4107   kmp_info_t *master = team->t.t_threads[0];
4108   KMP_DEBUG_ASSERT(this_thr != NULL);
4109   KMP_DEBUG_ASSERT(this_thr->th.th_serial_team);
4110   KMP_DEBUG_ASSERT(team);
4111   KMP_DEBUG_ASSERT(team->t.t_threads);
4112   KMP_DEBUG_ASSERT(team->t.t_dispatch);
4113   KMP_DEBUG_ASSERT(master);
4114   KMP_DEBUG_ASSERT(master->th.th_root);
4115 
4116   KMP_MB();
4117 
4118   TCW_SYNC_PTR(this_thr->th.th_team, team);
4119 
4120   this_thr->th.th_info.ds.ds_tid = tid;
4121   this_thr->th.th_set_nproc = 0;
4122   if (__kmp_tasking_mode != tskm_immediate_exec)
4123     // When tasking is possible, threads are not safe to reap until they are
4124     // done tasking; this will be set when tasking code is exited in wait
4125     this_thr->th.th_reap_state = KMP_NOT_SAFE_TO_REAP;
4126   else // no tasking --> always safe to reap
4127     this_thr->th.th_reap_state = KMP_SAFE_TO_REAP;
4128 #if OMP_40_ENABLED
4129   this_thr->th.th_set_proc_bind = proc_bind_default;
4130 #if KMP_AFFINITY_SUPPORTED
4131   this_thr->th.th_new_place = this_thr->th.th_current_place;
4132 #endif
4133 #endif
4134   this_thr->th.th_root = master->th.th_root;
4135 
4136   /* setup the thread's cache of the team structure */
4137   this_thr->th.th_team_nproc = team->t.t_nproc;
4138   this_thr->th.th_team_master = master;
4139   this_thr->th.th_team_serialized = team->t.t_serialized;
4140   TCW_PTR(this_thr->th.th_sleep_loc, NULL);
4141 
4142   KMP_DEBUG_ASSERT(team->t.t_implicit_task_taskdata);
4143 
4144   KF_TRACE(10, ("__kmp_initialize_info1: T#%d:%d this_thread=%p curtask=%p\n",
4145                 tid, gtid, this_thr, this_thr->th.th_current_task));
4146 
4147   __kmp_init_implicit_task(this_thr->th.th_team_master->th.th_ident, this_thr,
4148                            team, tid, TRUE);
4149 
4150   KF_TRACE(10, ("__kmp_initialize_info2: T#%d:%d this_thread=%p curtask=%p\n",
4151                 tid, gtid, this_thr, this_thr->th.th_current_task));
4152   // TODO: Initialize ICVs from parent; GEH - isn't that already done in
4153   // __kmp_initialize_team()?
4154 
4155   /* TODO no worksharing in speculative threads */
4156   this_thr->th.th_dispatch = &team->t.t_dispatch[tid];
4157 
4158   this_thr->th.th_local.this_construct = 0;
4159 
4160   if (!this_thr->th.th_pri_common) {
4161     this_thr->th.th_pri_common =
4162         (struct common_table *)__kmp_allocate(sizeof(struct common_table));
4163     if (__kmp_storage_map) {
4164       __kmp_print_storage_map_gtid(
4165           gtid, this_thr->th.th_pri_common, this_thr->th.th_pri_common + 1,
4166           sizeof(struct common_table), "th_%d.th_pri_common\n", gtid);
4167     }
4168     this_thr->th.th_pri_head = NULL;
4169   }
4170 
4171   if (this_thr != master && // Master's CG root is initialized elsewhere
4172       this_thr->th.th_cg_roots != master->th.th_cg_roots) { // CG root not set
4173     // Make new thread's CG root same as master's
4174     KMP_DEBUG_ASSERT(master->th.th_cg_roots);
4175     this_thr->th.th_cg_roots = master->th.th_cg_roots;
4176     // Increment new thread's CG root's counter to add the new thread
4177     this_thr->th.th_cg_roots->cg_nthreads++;
4178     KA_TRACE(100, ("__kmp_initialize_info: Thread %p increment cg_nthreads on"
4179                    " node %p of thread %p to %d\n",
4180                    this_thr, this_thr->th.th_cg_roots,
4181                    this_thr->th.th_cg_roots->cg_root,
4182                    this_thr->th.th_cg_roots->cg_nthreads));
4183     this_thr->th.th_current_task->td_icvs.thread_limit =
4184         this_thr->th.th_cg_roots->cg_thread_limit;
4185   }
4186 
4187   /* Initialize dynamic dispatch */
4188   {
4189     volatile kmp_disp_t *dispatch = this_thr->th.th_dispatch;
4190     // Use team max_nproc since this will never change for the team.
4191     size_t disp_size =
4192         sizeof(dispatch_private_info_t) *
4193         (team->t.t_max_nproc == 1 ? 1 : __kmp_dispatch_num_buffers);
4194     KD_TRACE(10, ("__kmp_initialize_info: T#%d max_nproc: %d\n", gtid,
4195                   team->t.t_max_nproc));
4196     KMP_ASSERT(dispatch);
4197     KMP_DEBUG_ASSERT(team->t.t_dispatch);
4198     KMP_DEBUG_ASSERT(dispatch == &team->t.t_dispatch[tid]);
4199 
4200     dispatch->th_disp_index = 0;
4201 #if OMP_45_ENABLED
4202     dispatch->th_doacross_buf_idx = 0;
4203 #endif
4204     if (!dispatch->th_disp_buffer) {
4205       dispatch->th_disp_buffer =
4206           (dispatch_private_info_t *)__kmp_allocate(disp_size);
4207 
4208       if (__kmp_storage_map) {
4209         __kmp_print_storage_map_gtid(
4210             gtid, &dispatch->th_disp_buffer[0],
4211             &dispatch->th_disp_buffer[team->t.t_max_nproc == 1
4212                                           ? 1
4213                                           : __kmp_dispatch_num_buffers],
4214             disp_size, "th_%d.th_dispatch.th_disp_buffer "
4215                        "(team_%d.t_dispatch[%d].th_disp_buffer)",
4216             gtid, team->t.t_id, gtid);
4217       }
4218     } else {
4219       memset(&dispatch->th_disp_buffer[0], '\0', disp_size);
4220     }
4221 
4222     dispatch->th_dispatch_pr_current = 0;
4223     dispatch->th_dispatch_sh_current = 0;
4224 
4225     dispatch->th_deo_fcn = 0; /* ORDERED     */
4226     dispatch->th_dxo_fcn = 0; /* END ORDERED */
4227   }
4228 
4229   this_thr->th.th_next_pool = NULL;
4230 
4231   if (!this_thr->th.th_task_state_memo_stack) {
4232     size_t i;
4233     this_thr->th.th_task_state_memo_stack =
4234         (kmp_uint8 *)__kmp_allocate(4 * sizeof(kmp_uint8));
4235     this_thr->th.th_task_state_top = 0;
4236     this_thr->th.th_task_state_stack_sz = 4;
4237     for (i = 0; i < this_thr->th.th_task_state_stack_sz;
4238          ++i) // zero init the stack
4239       this_thr->th.th_task_state_memo_stack[i] = 0;
4240   }
4241 
4242   KMP_DEBUG_ASSERT(!this_thr->th.th_spin_here);
4243   KMP_DEBUG_ASSERT(this_thr->th.th_next_waiting == 0);
4244 
4245   KMP_MB();
4246 }
4247 
4248 /* allocate a new thread for the requesting team. this is only called from
4249    within a forkjoin critical section. we will first try to get an available
4250    thread from the thread pool. if none is available, we will fork a new one
4251    assuming we are able to create a new one. this should be assured, as the
4252    caller should check on this first. */
4253 kmp_info_t *__kmp_allocate_thread(kmp_root_t *root, kmp_team_t *team,
4254                                   int new_tid) {
4255   kmp_team_t *serial_team;
4256   kmp_info_t *new_thr;
4257   int new_gtid;
4258 
4259   KA_TRACE(20, ("__kmp_allocate_thread: T#%d\n", __kmp_get_gtid()));
4260   KMP_DEBUG_ASSERT(root && team);
4261 #if !KMP_NESTED_HOT_TEAMS
4262   KMP_DEBUG_ASSERT(KMP_MASTER_GTID(__kmp_get_gtid()));
4263 #endif
4264   KMP_MB();
4265 
4266   /* first, try to get one from the thread pool */
4267   if (__kmp_thread_pool) {
4268     new_thr = CCAST(kmp_info_t *, __kmp_thread_pool);
4269     __kmp_thread_pool = (volatile kmp_info_t *)new_thr->th.th_next_pool;
4270     if (new_thr == __kmp_thread_pool_insert_pt) {
4271       __kmp_thread_pool_insert_pt = NULL;
4272     }
4273     TCW_4(new_thr->th.th_in_pool, FALSE);
4274     __kmp_suspend_initialize_thread(new_thr);
4275     __kmp_lock_suspend_mx(new_thr);
4276     if (new_thr->th.th_active_in_pool == TRUE) {
4277       KMP_DEBUG_ASSERT(new_thr->th.th_active == TRUE);
4278       KMP_ATOMIC_DEC(&__kmp_thread_pool_active_nth);
4279       new_thr->th.th_active_in_pool = FALSE;
4280     }
4281 #if KMP_DEBUG
4282     else {
4283       KMP_DEBUG_ASSERT(new_thr->th.th_active == FALSE);
4284     }
4285 #endif
4286     __kmp_unlock_suspend_mx(new_thr);
4287 
4288     KA_TRACE(20, ("__kmp_allocate_thread: T#%d using thread T#%d\n",
4289                   __kmp_get_gtid(), new_thr->th.th_info.ds.ds_gtid));
4290     KMP_ASSERT(!new_thr->th.th_team);
4291     KMP_DEBUG_ASSERT(__kmp_nth < __kmp_threads_capacity);
4292 
4293     /* setup the thread structure */
4294     __kmp_initialize_info(new_thr, team, new_tid,
4295                           new_thr->th.th_info.ds.ds_gtid);
4296     KMP_DEBUG_ASSERT(new_thr->th.th_serial_team);
4297 
4298     TCW_4(__kmp_nth, __kmp_nth + 1);
4299 
4300     new_thr->th.th_task_state = 0;
4301     new_thr->th.th_task_state_top = 0;
4302     new_thr->th.th_task_state_stack_sz = 4;
4303 
4304 #ifdef KMP_ADJUST_BLOCKTIME
4305     /* Adjust blocktime back to zero if necessary */
4306     /* Middle initialization might not have occurred yet */
4307     if (!__kmp_env_blocktime && (__kmp_avail_proc > 0)) {
4308       if (__kmp_nth > __kmp_avail_proc) {
4309         __kmp_zero_bt = TRUE;
4310       }
4311     }
4312 #endif /* KMP_ADJUST_BLOCKTIME */
4313 
4314 #if KMP_DEBUG
4315     // If thread entered pool via __kmp_free_thread, wait_flag should !=
4316     // KMP_BARRIER_PARENT_FLAG.
4317     int b;
4318     kmp_balign_t *balign = new_thr->th.th_bar;
4319     for (b = 0; b < bs_last_barrier; ++b)
4320       KMP_DEBUG_ASSERT(balign[b].bb.wait_flag != KMP_BARRIER_PARENT_FLAG);
4321 #endif
4322 
4323     KF_TRACE(10, ("__kmp_allocate_thread: T#%d using thread %p T#%d\n",
4324                   __kmp_get_gtid(), new_thr, new_thr->th.th_info.ds.ds_gtid));
4325 
4326     KMP_MB();
4327     return new_thr;
4328   }
4329 
4330   /* no, well fork a new one */
4331   KMP_ASSERT(__kmp_nth == __kmp_all_nth);
4332   KMP_ASSERT(__kmp_all_nth < __kmp_threads_capacity);
4333 
4334 #if KMP_USE_MONITOR
4335   // If this is the first worker thread the RTL is creating, then also
4336   // launch the monitor thread.  We try to do this as early as possible.
4337   if (!TCR_4(__kmp_init_monitor)) {
4338     __kmp_acquire_bootstrap_lock(&__kmp_monitor_lock);
4339     if (!TCR_4(__kmp_init_monitor)) {
4340       KF_TRACE(10, ("before __kmp_create_monitor\n"));
4341       TCW_4(__kmp_init_monitor, 1);
4342       __kmp_create_monitor(&__kmp_monitor);
4343       KF_TRACE(10, ("after __kmp_create_monitor\n"));
4344 #if KMP_OS_WINDOWS
4345       // AC: wait until monitor has started. This is a fix for CQ232808.
4346       // The reason is that if the library is loaded/unloaded in a loop with
4347       // small (parallel) work in between, then there is high probability that
4348       // monitor thread started after the library shutdown. At shutdown it is
4349       // too late to cope with the problem, because when the master is in
4350       // DllMain (process detach) the monitor has no chances to start (it is
4351       // blocked), and master has no means to inform the monitor that the
4352       // library has gone, because all the memory which the monitor can access
4353       // is going to be released/reset.
4354       while (TCR_4(__kmp_init_monitor) < 2) {
4355         KMP_YIELD(TRUE);
4356       }
4357       KF_TRACE(10, ("after monitor thread has started\n"));
4358 #endif
4359     }
4360     __kmp_release_bootstrap_lock(&__kmp_monitor_lock);
4361   }
4362 #endif
4363 
4364   KMP_MB();
4365   for (new_gtid = 1; TCR_PTR(__kmp_threads[new_gtid]) != NULL; ++new_gtid) {
4366     KMP_DEBUG_ASSERT(new_gtid < __kmp_threads_capacity);
4367   }
4368 
4369   /* allocate space for it. */
4370   new_thr = (kmp_info_t *)__kmp_allocate(sizeof(kmp_info_t));
4371 
4372   TCW_SYNC_PTR(__kmp_threads[new_gtid], new_thr);
4373 
4374   if (__kmp_storage_map) {
4375     __kmp_print_thread_storage_map(new_thr, new_gtid);
4376   }
4377 
4378   // add the reserve serialized team, initialized from the team's master thread
4379   {
4380     kmp_internal_control_t r_icvs = __kmp_get_x_global_icvs(team);
4381     KF_TRACE(10, ("__kmp_allocate_thread: before th_serial/serial_team\n"));
4382     new_thr->th.th_serial_team = serial_team =
4383         (kmp_team_t *)__kmp_allocate_team(root, 1, 1,
4384 #if OMPT_SUPPORT
4385                                           ompt_data_none, // root parallel id
4386 #endif
4387 #if OMP_40_ENABLED
4388                                           proc_bind_default,
4389 #endif
4390                                           &r_icvs, 0 USE_NESTED_HOT_ARG(NULL));
4391   }
4392   KMP_ASSERT(serial_team);
4393   serial_team->t.t_serialized = 0; // AC: the team created in reserve, not for
4394   // execution (it is unused for now).
4395   serial_team->t.t_threads[0] = new_thr;
4396   KF_TRACE(10,
4397            ("__kmp_allocate_thread: after th_serial/serial_team : new_thr=%p\n",
4398             new_thr));
4399 
4400   /* setup the thread structures */
4401   __kmp_initialize_info(new_thr, team, new_tid, new_gtid);
4402 
4403 #if USE_FAST_MEMORY
4404   __kmp_initialize_fast_memory(new_thr);
4405 #endif /* USE_FAST_MEMORY */
4406 
4407 #if KMP_USE_BGET
4408   KMP_DEBUG_ASSERT(new_thr->th.th_local.bget_data == NULL);
4409   __kmp_initialize_bget(new_thr);
4410 #endif
4411 
4412   __kmp_init_random(new_thr); // Initialize random number generator
4413 
4414   /* Initialize these only once when thread is grabbed for a team allocation */
4415   KA_TRACE(20,
4416            ("__kmp_allocate_thread: T#%d init go fork=%u, plain=%u\n",
4417             __kmp_get_gtid(), KMP_INIT_BARRIER_STATE, KMP_INIT_BARRIER_STATE));
4418 
4419   int b;
4420   kmp_balign_t *balign = new_thr->th.th_bar;
4421   for (b = 0; b < bs_last_barrier; ++b) {
4422     balign[b].bb.b_go = KMP_INIT_BARRIER_STATE;
4423     balign[b].bb.team = NULL;
4424     balign[b].bb.wait_flag = KMP_BARRIER_NOT_WAITING;
4425     balign[b].bb.use_oncore_barrier = 0;
4426   }
4427 
4428   new_thr->th.th_spin_here = FALSE;
4429   new_thr->th.th_next_waiting = 0;
4430 #if KMP_OS_UNIX
4431   new_thr->th.th_blocking = false;
4432 #endif
4433 
4434 #if OMP_40_ENABLED && KMP_AFFINITY_SUPPORTED
4435   new_thr->th.th_current_place = KMP_PLACE_UNDEFINED;
4436   new_thr->th.th_new_place = KMP_PLACE_UNDEFINED;
4437   new_thr->th.th_first_place = KMP_PLACE_UNDEFINED;
4438   new_thr->th.th_last_place = KMP_PLACE_UNDEFINED;
4439 #endif
4440 #if OMP_50_ENABLED
4441   new_thr->th.th_def_allocator = __kmp_def_allocator;
4442   new_thr->th.th_prev_level = 0;
4443   new_thr->th.th_prev_num_threads = 1;
4444 #endif
4445 
4446   TCW_4(new_thr->th.th_in_pool, FALSE);
4447   new_thr->th.th_active_in_pool = FALSE;
4448   TCW_4(new_thr->th.th_active, TRUE);
4449 
4450   /* adjust the global counters */
4451   __kmp_all_nth++;
4452   __kmp_nth++;
4453 
4454   // if __kmp_adjust_gtid_mode is set, then we use method #1 (sp search) for low
4455   // numbers of procs, and method #2 (keyed API call) for higher numbers.
4456   if (__kmp_adjust_gtid_mode) {
4457     if (__kmp_all_nth >= __kmp_tls_gtid_min) {
4458       if (TCR_4(__kmp_gtid_mode) != 2) {
4459         TCW_4(__kmp_gtid_mode, 2);
4460       }
4461     } else {
4462       if (TCR_4(__kmp_gtid_mode) != 1) {
4463         TCW_4(__kmp_gtid_mode, 1);
4464       }
4465     }
4466   }
4467 
4468 #ifdef KMP_ADJUST_BLOCKTIME
4469   /* Adjust blocktime back to zero if necessary       */
4470   /* Middle initialization might not have occurred yet */
4471   if (!__kmp_env_blocktime && (__kmp_avail_proc > 0)) {
4472     if (__kmp_nth > __kmp_avail_proc) {
4473       __kmp_zero_bt = TRUE;
4474     }
4475   }
4476 #endif /* KMP_ADJUST_BLOCKTIME */
4477 
4478   /* actually fork it and create the new worker thread */
4479   KF_TRACE(
4480       10, ("__kmp_allocate_thread: before __kmp_create_worker: %p\n", new_thr));
4481   __kmp_create_worker(new_gtid, new_thr, __kmp_stksize);
4482   KF_TRACE(10,
4483            ("__kmp_allocate_thread: after __kmp_create_worker: %p\n", new_thr));
4484 
4485   KA_TRACE(20, ("__kmp_allocate_thread: T#%d forked T#%d\n", __kmp_get_gtid(),
4486                 new_gtid));
4487   KMP_MB();
4488   return new_thr;
4489 }
4490 
4491 /* Reinitialize team for reuse.
4492    The hot team code calls this case at every fork barrier, so EPCC barrier
4493    test are extremely sensitive to changes in it, esp. writes to the team
4494    struct, which cause a cache invalidation in all threads.
4495    IF YOU TOUCH THIS ROUTINE, RUN EPCC C SYNCBENCH ON A BIG-IRON MACHINE!!! */
4496 static void __kmp_reinitialize_team(kmp_team_t *team,
4497                                     kmp_internal_control_t *new_icvs,
4498                                     ident_t *loc) {
4499   KF_TRACE(10, ("__kmp_reinitialize_team: enter this_thread=%p team=%p\n",
4500                 team->t.t_threads[0], team));
4501   KMP_DEBUG_ASSERT(team && new_icvs);
4502   KMP_DEBUG_ASSERT((!TCR_4(__kmp_init_parallel)) || new_icvs->nproc);
4503   KMP_CHECK_UPDATE(team->t.t_ident, loc);
4504 
4505   KMP_CHECK_UPDATE(team->t.t_id, KMP_GEN_TEAM_ID());
4506   // Copy ICVs to the master thread's implicit taskdata
4507   __kmp_init_implicit_task(loc, team->t.t_threads[0], team, 0, FALSE);
4508   copy_icvs(&team->t.t_implicit_task_taskdata[0].td_icvs, new_icvs);
4509 
4510   KF_TRACE(10, ("__kmp_reinitialize_team: exit this_thread=%p team=%p\n",
4511                 team->t.t_threads[0], team));
4512 }
4513 
4514 /* Initialize the team data structure.
4515    This assumes the t_threads and t_max_nproc are already set.
4516    Also, we don't touch the arguments */
4517 static void __kmp_initialize_team(kmp_team_t *team, int new_nproc,
4518                                   kmp_internal_control_t *new_icvs,
4519                                   ident_t *loc) {
4520   KF_TRACE(10, ("__kmp_initialize_team: enter: team=%p\n", team));
4521 
4522   /* verify */
4523   KMP_DEBUG_ASSERT(team);
4524   KMP_DEBUG_ASSERT(new_nproc <= team->t.t_max_nproc);
4525   KMP_DEBUG_ASSERT(team->t.t_threads);
4526   KMP_MB();
4527 
4528   team->t.t_master_tid = 0; /* not needed */
4529   /* team->t.t_master_bar;        not needed */
4530   team->t.t_serialized = new_nproc > 1 ? 0 : 1;
4531   team->t.t_nproc = new_nproc;
4532 
4533   /* team->t.t_parent     = NULL; TODO not needed & would mess up hot team */
4534   team->t.t_next_pool = NULL;
4535   /* memset( team->t.t_threads, 0, sizeof(kmp_info_t*)*new_nproc ); would mess
4536    * up hot team */
4537 
4538   TCW_SYNC_PTR(team->t.t_pkfn, NULL); /* not needed */
4539   team->t.t_invoke = NULL; /* not needed */
4540 
4541   // TODO???: team->t.t_max_active_levels       = new_max_active_levels;
4542   team->t.t_sched.sched = new_icvs->sched.sched;
4543 
4544 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
4545   team->t.t_fp_control_saved = FALSE; /* not needed */
4546   team->t.t_x87_fpu_control_word = 0; /* not needed */
4547   team->t.t_mxcsr = 0; /* not needed */
4548 #endif /* KMP_ARCH_X86 || KMP_ARCH_X86_64 */
4549 
4550   team->t.t_construct = 0;
4551 
4552   team->t.t_ordered.dt.t_value = 0;
4553   team->t.t_master_active = FALSE;
4554 
4555 #ifdef KMP_DEBUG
4556   team->t.t_copypriv_data = NULL; /* not necessary, but nice for debugging */
4557 #endif
4558 #if KMP_OS_WINDOWS
4559   team->t.t_copyin_counter = 0; /* for barrier-free copyin implementation */
4560 #endif
4561 
4562   team->t.t_control_stack_top = NULL;
4563 
4564   __kmp_reinitialize_team(team, new_icvs, loc);
4565 
4566   KMP_MB();
4567   KF_TRACE(10, ("__kmp_initialize_team: exit: team=%p\n", team));
4568 }
4569 
4570 #if KMP_OS_LINUX && KMP_AFFINITY_SUPPORTED
4571 /* Sets full mask for thread and returns old mask, no changes to structures. */
4572 static void
4573 __kmp_set_thread_affinity_mask_full_tmp(kmp_affin_mask_t *old_mask) {
4574   if (KMP_AFFINITY_CAPABLE()) {
4575     int status;
4576     if (old_mask != NULL) {
4577       status = __kmp_get_system_affinity(old_mask, TRUE);
4578       int error = errno;
4579       if (status != 0) {
4580         __kmp_fatal(KMP_MSG(ChangeThreadAffMaskError), KMP_ERR(error),
4581                     __kmp_msg_null);
4582       }
4583     }
4584     __kmp_set_system_affinity(__kmp_affin_fullMask, TRUE);
4585   }
4586 }
4587 #endif
4588 
4589 #if OMP_40_ENABLED && KMP_AFFINITY_SUPPORTED
4590 
4591 // __kmp_partition_places() is the heart of the OpenMP 4.0 affinity mechanism.
4592 // It calculats the worker + master thread's partition based upon the parent
4593 // thread's partition, and binds each worker to a thread in their partition.
4594 // The master thread's partition should already include its current binding.
4595 static void __kmp_partition_places(kmp_team_t *team, int update_master_only) {
4596   // Copy the master thread's place partion to the team struct
4597   kmp_info_t *master_th = team->t.t_threads[0];
4598   KMP_DEBUG_ASSERT(master_th != NULL);
4599   kmp_proc_bind_t proc_bind = team->t.t_proc_bind;
4600   int first_place = master_th->th.th_first_place;
4601   int last_place = master_th->th.th_last_place;
4602   int masters_place = master_th->th.th_current_place;
4603   team->t.t_first_place = first_place;
4604   team->t.t_last_place = last_place;
4605 
4606   KA_TRACE(20, ("__kmp_partition_places: enter: proc_bind = %d T#%d(%d:0) "
4607                 "bound to place %d partition = [%d,%d]\n",
4608                 proc_bind, __kmp_gtid_from_thread(team->t.t_threads[0]),
4609                 team->t.t_id, masters_place, first_place, last_place));
4610 
4611   switch (proc_bind) {
4612 
4613   case proc_bind_default:
4614     // serial teams might have the proc_bind policy set to proc_bind_default. It
4615     // doesn't matter, as we don't rebind master thread for any proc_bind policy
4616     KMP_DEBUG_ASSERT(team->t.t_nproc == 1);
4617     break;
4618 
4619   case proc_bind_master: {
4620     int f;
4621     int n_th = team->t.t_nproc;
4622     for (f = 1; f < n_th; f++) {
4623       kmp_info_t *th = team->t.t_threads[f];
4624       KMP_DEBUG_ASSERT(th != NULL);
4625       th->th.th_first_place = first_place;
4626       th->th.th_last_place = last_place;
4627       th->th.th_new_place = masters_place;
4628 #if OMP_50_ENABLED
4629       if (__kmp_display_affinity && masters_place != th->th.th_current_place &&
4630           team->t.t_display_affinity != 1) {
4631         team->t.t_display_affinity = 1;
4632       }
4633 #endif
4634 
4635       KA_TRACE(100, ("__kmp_partition_places: master: T#%d(%d:%d) place %d "
4636                      "partition = [%d,%d]\n",
4637                      __kmp_gtid_from_thread(team->t.t_threads[f]), team->t.t_id,
4638                      f, masters_place, first_place, last_place));
4639     }
4640   } break;
4641 
4642   case proc_bind_close: {
4643     int f;
4644     int n_th = team->t.t_nproc;
4645     int n_places;
4646     if (first_place <= last_place) {
4647       n_places = last_place - first_place + 1;
4648     } else {
4649       n_places = __kmp_affinity_num_masks - first_place + last_place + 1;
4650     }
4651     if (n_th <= n_places) {
4652       int place = masters_place;
4653       for (f = 1; f < n_th; f++) {
4654         kmp_info_t *th = team->t.t_threads[f];
4655         KMP_DEBUG_ASSERT(th != NULL);
4656 
4657         if (place == last_place) {
4658           place = first_place;
4659         } else if (place == (int)(__kmp_affinity_num_masks - 1)) {
4660           place = 0;
4661         } else {
4662           place++;
4663         }
4664         th->th.th_first_place = first_place;
4665         th->th.th_last_place = last_place;
4666         th->th.th_new_place = place;
4667 #if OMP_50_ENABLED
4668         if (__kmp_display_affinity && place != th->th.th_current_place &&
4669             team->t.t_display_affinity != 1) {
4670           team->t.t_display_affinity = 1;
4671         }
4672 #endif
4673 
4674         KA_TRACE(100, ("__kmp_partition_places: close: T#%d(%d:%d) place %d "
4675                        "partition = [%d,%d]\n",
4676                        __kmp_gtid_from_thread(team->t.t_threads[f]),
4677                        team->t.t_id, f, place, first_place, last_place));
4678       }
4679     } else {
4680       int S, rem, gap, s_count;
4681       S = n_th / n_places;
4682       s_count = 0;
4683       rem = n_th - (S * n_places);
4684       gap = rem > 0 ? n_places / rem : n_places;
4685       int place = masters_place;
4686       int gap_ct = gap;
4687       for (f = 0; f < n_th; f++) {
4688         kmp_info_t *th = team->t.t_threads[f];
4689         KMP_DEBUG_ASSERT(th != NULL);
4690 
4691         th->th.th_first_place = first_place;
4692         th->th.th_last_place = last_place;
4693         th->th.th_new_place = place;
4694 #if OMP_50_ENABLED
4695         if (__kmp_display_affinity && place != th->th.th_current_place &&
4696             team->t.t_display_affinity != 1) {
4697           team->t.t_display_affinity = 1;
4698         }
4699 #endif
4700         s_count++;
4701 
4702         if ((s_count == S) && rem && (gap_ct == gap)) {
4703           // do nothing, add an extra thread to place on next iteration
4704         } else if ((s_count == S + 1) && rem && (gap_ct == gap)) {
4705           // we added an extra thread to this place; move to next place
4706           if (place == last_place) {
4707             place = first_place;
4708           } else if (place == (int)(__kmp_affinity_num_masks - 1)) {
4709             place = 0;
4710           } else {
4711             place++;
4712           }
4713           s_count = 0;
4714           gap_ct = 1;
4715           rem--;
4716         } else if (s_count == S) { // place full; don't add extra
4717           if (place == last_place) {
4718             place = first_place;
4719           } else if (place == (int)(__kmp_affinity_num_masks - 1)) {
4720             place = 0;
4721           } else {
4722             place++;
4723           }
4724           gap_ct++;
4725           s_count = 0;
4726         }
4727 
4728         KA_TRACE(100,
4729                  ("__kmp_partition_places: close: T#%d(%d:%d) place %d "
4730                   "partition = [%d,%d]\n",
4731                   __kmp_gtid_from_thread(team->t.t_threads[f]), team->t.t_id, f,
4732                   th->th.th_new_place, first_place, last_place));
4733       }
4734       KMP_DEBUG_ASSERT(place == masters_place);
4735     }
4736   } break;
4737 
4738   case proc_bind_spread: {
4739     int f;
4740     int n_th = team->t.t_nproc;
4741     int n_places;
4742     int thidx;
4743     if (first_place <= last_place) {
4744       n_places = last_place - first_place + 1;
4745     } else {
4746       n_places = __kmp_affinity_num_masks - first_place + last_place + 1;
4747     }
4748     if (n_th <= n_places) {
4749       int place = -1;
4750 
4751       if (n_places != static_cast<int>(__kmp_affinity_num_masks)) {
4752         int S = n_places / n_th;
4753         int s_count, rem, gap, gap_ct;
4754 
4755         place = masters_place;
4756         rem = n_places - n_th * S;
4757         gap = rem ? n_th / rem : 1;
4758         gap_ct = gap;
4759         thidx = n_th;
4760         if (update_master_only == 1)
4761           thidx = 1;
4762         for (f = 0; f < thidx; f++) {
4763           kmp_info_t *th = team->t.t_threads[f];
4764           KMP_DEBUG_ASSERT(th != NULL);
4765 
4766           th->th.th_first_place = place;
4767           th->th.th_new_place = place;
4768 #if OMP_50_ENABLED
4769           if (__kmp_display_affinity && place != th->th.th_current_place &&
4770               team->t.t_display_affinity != 1) {
4771             team->t.t_display_affinity = 1;
4772           }
4773 #endif
4774           s_count = 1;
4775           while (s_count < S) {
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             s_count++;
4784           }
4785           if (rem && (gap_ct == gap)) {
4786             if (place == last_place) {
4787               place = first_place;
4788             } else if (place == (int)(__kmp_affinity_num_masks - 1)) {
4789               place = 0;
4790             } else {
4791               place++;
4792             }
4793             rem--;
4794             gap_ct = 0;
4795           }
4796           th->th.th_last_place = place;
4797           gap_ct++;
4798 
4799           if (place == last_place) {
4800             place = first_place;
4801           } else if (place == (int)(__kmp_affinity_num_masks - 1)) {
4802             place = 0;
4803           } else {
4804             place++;
4805           }
4806 
4807           KA_TRACE(100,
4808                    ("__kmp_partition_places: spread: T#%d(%d:%d) place %d "
4809                     "partition = [%d,%d], __kmp_affinity_num_masks: %u\n",
4810                     __kmp_gtid_from_thread(team->t.t_threads[f]), team->t.t_id,
4811                     f, th->th.th_new_place, th->th.th_first_place,
4812                     th->th.th_last_place, __kmp_affinity_num_masks));
4813         }
4814       } else {
4815         /* Having uniform space of available computation places I can create
4816            T partitions of round(P/T) size and put threads into the first
4817            place of each partition. */
4818         double current = static_cast<double>(masters_place);
4819         double spacing =
4820             (static_cast<double>(n_places + 1) / static_cast<double>(n_th));
4821         int first, last;
4822         kmp_info_t *th;
4823 
4824         thidx = n_th + 1;
4825         if (update_master_only == 1)
4826           thidx = 1;
4827         for (f = 0; f < thidx; f++) {
4828           first = static_cast<int>(current);
4829           last = static_cast<int>(current + spacing) - 1;
4830           KMP_DEBUG_ASSERT(last >= first);
4831           if (first >= n_places) {
4832             if (masters_place) {
4833               first -= n_places;
4834               last -= n_places;
4835               if (first == (masters_place + 1)) {
4836                 KMP_DEBUG_ASSERT(f == n_th);
4837                 first--;
4838               }
4839               if (last == masters_place) {
4840                 KMP_DEBUG_ASSERT(f == (n_th - 1));
4841                 last--;
4842               }
4843             } else {
4844               KMP_DEBUG_ASSERT(f == n_th);
4845               first = 0;
4846               last = 0;
4847             }
4848           }
4849           if (last >= n_places) {
4850             last = (n_places - 1);
4851           }
4852           place = first;
4853           current += spacing;
4854           if (f < n_th) {
4855             KMP_DEBUG_ASSERT(0 <= first);
4856             KMP_DEBUG_ASSERT(n_places > first);
4857             KMP_DEBUG_ASSERT(0 <= last);
4858             KMP_DEBUG_ASSERT(n_places > last);
4859             KMP_DEBUG_ASSERT(last_place >= first_place);
4860             th = team->t.t_threads[f];
4861             KMP_DEBUG_ASSERT(th);
4862             th->th.th_first_place = first;
4863             th->th.th_new_place = place;
4864             th->th.th_last_place = last;
4865 #if OMP_50_ENABLED
4866             if (__kmp_display_affinity && place != th->th.th_current_place &&
4867                 team->t.t_display_affinity != 1) {
4868               team->t.t_display_affinity = 1;
4869             }
4870 #endif
4871             KA_TRACE(100,
4872                      ("__kmp_partition_places: spread: T#%d(%d:%d) place %d "
4873                       "partition = [%d,%d], spacing = %.4f\n",
4874                       __kmp_gtid_from_thread(team->t.t_threads[f]),
4875                       team->t.t_id, f, th->th.th_new_place,
4876                       th->th.th_first_place, th->th.th_last_place, spacing));
4877           }
4878         }
4879       }
4880       KMP_DEBUG_ASSERT(update_master_only || place == masters_place);
4881     } else {
4882       int S, rem, gap, s_count;
4883       S = n_th / n_places;
4884       s_count = 0;
4885       rem = n_th - (S * n_places);
4886       gap = rem > 0 ? n_places / rem : n_places;
4887       int place = masters_place;
4888       int gap_ct = gap;
4889       thidx = n_th;
4890       if (update_master_only == 1)
4891         thidx = 1;
4892       for (f = 0; f < thidx; f++) {
4893         kmp_info_t *th = team->t.t_threads[f];
4894         KMP_DEBUG_ASSERT(th != NULL);
4895 
4896         th->th.th_first_place = place;
4897         th->th.th_last_place = place;
4898         th->th.th_new_place = place;
4899 #if OMP_50_ENABLED
4900         if (__kmp_display_affinity && place != th->th.th_current_place &&
4901             team->t.t_display_affinity != 1) {
4902           team->t.t_display_affinity = 1;
4903         }
4904 #endif
4905         s_count++;
4906 
4907         if ((s_count == S) && rem && (gap_ct == gap)) {
4908           // do nothing, add an extra thread to place on next iteration
4909         } else if ((s_count == S + 1) && rem && (gap_ct == gap)) {
4910           // we added an extra thread to this place; move on to next place
4911           if (place == last_place) {
4912             place = first_place;
4913           } else if (place == (int)(__kmp_affinity_num_masks - 1)) {
4914             place = 0;
4915           } else {
4916             place++;
4917           }
4918           s_count = 0;
4919           gap_ct = 1;
4920           rem--;
4921         } else if (s_count == S) { // place is full; don't add extra thread
4922           if (place == last_place) {
4923             place = first_place;
4924           } else if (place == (int)(__kmp_affinity_num_masks - 1)) {
4925             place = 0;
4926           } else {
4927             place++;
4928           }
4929           gap_ct++;
4930           s_count = 0;
4931         }
4932 
4933         KA_TRACE(100, ("__kmp_partition_places: spread: T#%d(%d:%d) place %d "
4934                        "partition = [%d,%d]\n",
4935                        __kmp_gtid_from_thread(team->t.t_threads[f]),
4936                        team->t.t_id, f, th->th.th_new_place,
4937                        th->th.th_first_place, th->th.th_last_place));
4938       }
4939       KMP_DEBUG_ASSERT(update_master_only || place == masters_place);
4940     }
4941   } break;
4942 
4943   default:
4944     break;
4945   }
4946 
4947   KA_TRACE(20, ("__kmp_partition_places: exit T#%d\n", team->t.t_id));
4948 }
4949 
4950 #endif /* OMP_40_ENABLED && KMP_AFFINITY_SUPPORTED */
4951 
4952 /* allocate a new team data structure to use.  take one off of the free pool if
4953    available */
4954 kmp_team_t *
4955 __kmp_allocate_team(kmp_root_t *root, int new_nproc, int max_nproc,
4956 #if OMPT_SUPPORT
4957                     ompt_data_t ompt_parallel_data,
4958 #endif
4959 #if OMP_40_ENABLED
4960                     kmp_proc_bind_t new_proc_bind,
4961 #endif
4962                     kmp_internal_control_t *new_icvs,
4963                     int argc USE_NESTED_HOT_ARG(kmp_info_t *master)) {
4964   KMP_TIME_DEVELOPER_PARTITIONED_BLOCK(KMP_allocate_team);
4965   int f;
4966   kmp_team_t *team;
4967   int use_hot_team = !root->r.r_active;
4968   int level = 0;
4969 
4970   KA_TRACE(20, ("__kmp_allocate_team: called\n"));
4971   KMP_DEBUG_ASSERT(new_nproc >= 1 && argc >= 0);
4972   KMP_DEBUG_ASSERT(max_nproc >= new_nproc);
4973   KMP_MB();
4974 
4975 #if KMP_NESTED_HOT_TEAMS
4976   kmp_hot_team_ptr_t *hot_teams;
4977   if (master) {
4978     team = master->th.th_team;
4979     level = team->t.t_active_level;
4980     if (master->th.th_teams_microtask) { // in teams construct?
4981       if (master->th.th_teams_size.nteams > 1 &&
4982           ( // #teams > 1
4983               team->t.t_pkfn ==
4984                   (microtask_t)__kmp_teams_master || // inner fork of the teams
4985               master->th.th_teams_level <
4986                   team->t.t_level)) { // or nested parallel inside the teams
4987         ++level; // not increment if #teams==1, or for outer fork of the teams;
4988         // increment otherwise
4989       }
4990     }
4991     hot_teams = master->th.th_hot_teams;
4992     if (level < __kmp_hot_teams_max_level && hot_teams &&
4993         hot_teams[level]
4994             .hot_team) { // hot team has already been allocated for given level
4995       use_hot_team = 1;
4996     } else {
4997       use_hot_team = 0;
4998     }
4999   }
5000 #endif
5001   // Optimization to use a "hot" team
5002   if (use_hot_team && new_nproc > 1) {
5003     KMP_DEBUG_ASSERT(new_nproc <= max_nproc);
5004 #if KMP_NESTED_HOT_TEAMS
5005     team = hot_teams[level].hot_team;
5006 #else
5007     team = root->r.r_hot_team;
5008 #endif
5009 #if KMP_DEBUG
5010     if (__kmp_tasking_mode != tskm_immediate_exec) {
5011       KA_TRACE(20, ("__kmp_allocate_team: hot team task_team[0] = %p "
5012                     "task_team[1] = %p before reinit\n",
5013                     team->t.t_task_team[0], team->t.t_task_team[1]));
5014     }
5015 #endif
5016 
5017     // Has the number of threads changed?
5018     /* Let's assume the most common case is that the number of threads is
5019        unchanged, and put that case first. */
5020     if (team->t.t_nproc == new_nproc) { // Check changes in number of threads
5021       KA_TRACE(20, ("__kmp_allocate_team: reusing hot team\n"));
5022       // This case can mean that omp_set_num_threads() was called and the hot
5023       // team size was already reduced, so we check the special flag
5024       if (team->t.t_size_changed == -1) {
5025         team->t.t_size_changed = 1;
5026       } else {
5027         KMP_CHECK_UPDATE(team->t.t_size_changed, 0);
5028       }
5029 
5030       // TODO???: team->t.t_max_active_levels = new_max_active_levels;
5031       kmp_r_sched_t new_sched = new_icvs->sched;
5032       // set master's schedule as new run-time schedule
5033       KMP_CHECK_UPDATE(team->t.t_sched.sched, new_sched.sched);
5034 
5035       __kmp_reinitialize_team(team, new_icvs,
5036                               root->r.r_uber_thread->th.th_ident);
5037 
5038       KF_TRACE(10, ("__kmp_allocate_team2: T#%d, this_thread=%p team=%p\n", 0,
5039                     team->t.t_threads[0], team));
5040       __kmp_push_current_task_to_thread(team->t.t_threads[0], team, 0);
5041 
5042 #if OMP_40_ENABLED
5043 #if KMP_AFFINITY_SUPPORTED
5044       if ((team->t.t_size_changed == 0) &&
5045           (team->t.t_proc_bind == new_proc_bind)) {
5046         if (new_proc_bind == proc_bind_spread) {
5047           __kmp_partition_places(
5048               team, 1); // add flag to update only master for spread
5049         }
5050         KA_TRACE(200, ("__kmp_allocate_team: reusing hot team #%d bindings: "
5051                        "proc_bind = %d, partition = [%d,%d]\n",
5052                        team->t.t_id, new_proc_bind, team->t.t_first_place,
5053                        team->t.t_last_place));
5054       } else {
5055         KMP_CHECK_UPDATE(team->t.t_proc_bind, new_proc_bind);
5056         __kmp_partition_places(team);
5057       }
5058 #else
5059       KMP_CHECK_UPDATE(team->t.t_proc_bind, new_proc_bind);
5060 #endif /* KMP_AFFINITY_SUPPORTED */
5061 #endif /* OMP_40_ENABLED */
5062     } else if (team->t.t_nproc > new_nproc) {
5063       KA_TRACE(20,
5064                ("__kmp_allocate_team: decreasing hot team thread count to %d\n",
5065                 new_nproc));
5066 
5067       team->t.t_size_changed = 1;
5068 #if KMP_NESTED_HOT_TEAMS
5069       if (__kmp_hot_teams_mode == 0) {
5070         // AC: saved number of threads should correspond to team's value in this
5071         // mode, can be bigger in mode 1, when hot team has threads in reserve
5072         KMP_DEBUG_ASSERT(hot_teams[level].hot_team_nth == team->t.t_nproc);
5073         hot_teams[level].hot_team_nth = new_nproc;
5074 #endif // KMP_NESTED_HOT_TEAMS
5075         /* release the extra threads we don't need any more */
5076         for (f = new_nproc; f < team->t.t_nproc; f++) {
5077           KMP_DEBUG_ASSERT(team->t.t_threads[f]);
5078           if (__kmp_tasking_mode != tskm_immediate_exec) {
5079             // When decreasing team size, threads no longer in the team should
5080             // unref task team.
5081             team->t.t_threads[f]->th.th_task_team = NULL;
5082           }
5083           __kmp_free_thread(team->t.t_threads[f]);
5084           team->t.t_threads[f] = NULL;
5085         }
5086 #if KMP_NESTED_HOT_TEAMS
5087       } // (__kmp_hot_teams_mode == 0)
5088       else {
5089         // When keeping extra threads in team, switch threads to wait on own
5090         // b_go flag
5091         for (f = new_nproc; f < team->t.t_nproc; ++f) {
5092           KMP_DEBUG_ASSERT(team->t.t_threads[f]);
5093           kmp_balign_t *balign = team->t.t_threads[f]->th.th_bar;
5094           for (int b = 0; b < bs_last_barrier; ++b) {
5095             if (balign[b].bb.wait_flag == KMP_BARRIER_PARENT_FLAG) {
5096               balign[b].bb.wait_flag = KMP_BARRIER_SWITCH_TO_OWN_FLAG;
5097             }
5098             KMP_CHECK_UPDATE(balign[b].bb.leaf_kids, 0);
5099           }
5100         }
5101       }
5102 #endif // KMP_NESTED_HOT_TEAMS
5103       team->t.t_nproc = new_nproc;
5104       // TODO???: team->t.t_max_active_levels = new_max_active_levels;
5105       KMP_CHECK_UPDATE(team->t.t_sched.sched, new_icvs->sched.sched);
5106       __kmp_reinitialize_team(team, new_icvs,
5107                               root->r.r_uber_thread->th.th_ident);
5108 
5109       // Update remaining threads
5110       for (f = 0; f < new_nproc; ++f) {
5111         team->t.t_threads[f]->th.th_team_nproc = new_nproc;
5112       }
5113 
5114       // restore the current task state of the master thread: should be the
5115       // implicit task
5116       KF_TRACE(10, ("__kmp_allocate_team: T#%d, this_thread=%p team=%p\n", 0,
5117                     team->t.t_threads[0], team));
5118 
5119       __kmp_push_current_task_to_thread(team->t.t_threads[0], team, 0);
5120 
5121 #ifdef KMP_DEBUG
5122       for (f = 0; f < team->t.t_nproc; f++) {
5123         KMP_DEBUG_ASSERT(team->t.t_threads[f] &&
5124                          team->t.t_threads[f]->th.th_team_nproc ==
5125                              team->t.t_nproc);
5126       }
5127 #endif
5128 
5129 #if OMP_40_ENABLED
5130       KMP_CHECK_UPDATE(team->t.t_proc_bind, new_proc_bind);
5131 #if KMP_AFFINITY_SUPPORTED
5132       __kmp_partition_places(team);
5133 #endif
5134 #endif
5135     } else { // team->t.t_nproc < new_nproc
5136 #if KMP_OS_LINUX && KMP_AFFINITY_SUPPORTED
5137       kmp_affin_mask_t *old_mask;
5138       if (KMP_AFFINITY_CAPABLE()) {
5139         KMP_CPU_ALLOC(old_mask);
5140       }
5141 #endif
5142 
5143       KA_TRACE(20,
5144                ("__kmp_allocate_team: increasing hot team thread count to %d\n",
5145                 new_nproc));
5146 
5147       team->t.t_size_changed = 1;
5148 
5149 #if KMP_NESTED_HOT_TEAMS
5150       int avail_threads = hot_teams[level].hot_team_nth;
5151       if (new_nproc < avail_threads)
5152         avail_threads = new_nproc;
5153       kmp_info_t **other_threads = team->t.t_threads;
5154       for (f = team->t.t_nproc; f < avail_threads; ++f) {
5155         // Adjust barrier data of reserved threads (if any) of the team
5156         // Other data will be set in __kmp_initialize_info() below.
5157         int b;
5158         kmp_balign_t *balign = other_threads[f]->th.th_bar;
5159         for (b = 0; b < bs_last_barrier; ++b) {
5160           balign[b].bb.b_arrived = team->t.t_bar[b].b_arrived;
5161           KMP_DEBUG_ASSERT(balign[b].bb.wait_flag != KMP_BARRIER_PARENT_FLAG);
5162 #if USE_DEBUGGER
5163           balign[b].bb.b_worker_arrived = team->t.t_bar[b].b_team_arrived;
5164 #endif
5165         }
5166       }
5167       if (hot_teams[level].hot_team_nth >= new_nproc) {
5168         // we have all needed threads in reserve, no need to allocate any
5169         // this only possible in mode 1, cannot have reserved threads in mode 0
5170         KMP_DEBUG_ASSERT(__kmp_hot_teams_mode == 1);
5171         team->t.t_nproc = new_nproc; // just get reserved threads involved
5172       } else {
5173         // we may have some threads in reserve, but not enough
5174         team->t.t_nproc =
5175             hot_teams[level]
5176                 .hot_team_nth; // get reserved threads involved if any
5177         hot_teams[level].hot_team_nth = new_nproc; // adjust hot team max size
5178 #endif // KMP_NESTED_HOT_TEAMS
5179         if (team->t.t_max_nproc < new_nproc) {
5180           /* reallocate larger arrays */
5181           __kmp_reallocate_team_arrays(team, new_nproc);
5182           __kmp_reinitialize_team(team, new_icvs, NULL);
5183         }
5184 
5185 #if KMP_OS_LINUX && KMP_AFFINITY_SUPPORTED
5186         /* Temporarily set full mask for master thread before creation of
5187            workers. The reason is that workers inherit the affinity from master,
5188            so if a lot of workers are created on the single core quickly, they
5189            don't get a chance to set their own affinity for a long time. */
5190         __kmp_set_thread_affinity_mask_full_tmp(old_mask);
5191 #endif
5192 
5193         /* allocate new threads for the hot team */
5194         for (f = team->t.t_nproc; f < new_nproc; f++) {
5195           kmp_info_t *new_worker = __kmp_allocate_thread(root, team, f);
5196           KMP_DEBUG_ASSERT(new_worker);
5197           team->t.t_threads[f] = new_worker;
5198 
5199           KA_TRACE(20,
5200                    ("__kmp_allocate_team: team %d init T#%d arrived: "
5201                     "join=%llu, plain=%llu\n",
5202                     team->t.t_id, __kmp_gtid_from_tid(f, team), team->t.t_id, f,
5203                     team->t.t_bar[bs_forkjoin_barrier].b_arrived,
5204                     team->t.t_bar[bs_plain_barrier].b_arrived));
5205 
5206           { // Initialize barrier data for new threads.
5207             int b;
5208             kmp_balign_t *balign = new_worker->th.th_bar;
5209             for (b = 0; b < bs_last_barrier; ++b) {
5210               balign[b].bb.b_arrived = team->t.t_bar[b].b_arrived;
5211               KMP_DEBUG_ASSERT(balign[b].bb.wait_flag !=
5212                                KMP_BARRIER_PARENT_FLAG);
5213 #if USE_DEBUGGER
5214               balign[b].bb.b_worker_arrived = team->t.t_bar[b].b_team_arrived;
5215 #endif
5216             }
5217           }
5218         }
5219 
5220 #if KMP_OS_LINUX && KMP_AFFINITY_SUPPORTED
5221         if (KMP_AFFINITY_CAPABLE()) {
5222           /* Restore initial master thread's affinity mask */
5223           __kmp_set_system_affinity(old_mask, TRUE);
5224           KMP_CPU_FREE(old_mask);
5225         }
5226 #endif
5227 #if KMP_NESTED_HOT_TEAMS
5228       } // end of check of t_nproc vs. new_nproc vs. hot_team_nth
5229 #endif // KMP_NESTED_HOT_TEAMS
5230       /* make sure everyone is syncronized */
5231       int old_nproc = team->t.t_nproc; // save old value and use to update only
5232       // new threads below
5233       __kmp_initialize_team(team, new_nproc, new_icvs,
5234                             root->r.r_uber_thread->th.th_ident);
5235 
5236       /* reinitialize the threads */
5237       KMP_DEBUG_ASSERT(team->t.t_nproc == new_nproc);
5238       for (f = 0; f < team->t.t_nproc; ++f)
5239         __kmp_initialize_info(team->t.t_threads[f], team, f,
5240                               __kmp_gtid_from_tid(f, team));
5241 
5242       if (level) { // set th_task_state for new threads in nested hot team
5243         // __kmp_initialize_info() no longer zeroes th_task_state, so we should
5244         // only need to set the th_task_state for the new threads. th_task_state
5245         // for master thread will not be accurate until after this in
5246         // __kmp_fork_call(), so we look to the master's memo_stack to get the
5247         // correct value.
5248         for (f = old_nproc; f < team->t.t_nproc; ++f)
5249           team->t.t_threads[f]->th.th_task_state =
5250               team->t.t_threads[0]->th.th_task_state_memo_stack[level];
5251       } else { // set th_task_state for new threads in non-nested hot team
5252         int old_state =
5253             team->t.t_threads[0]->th.th_task_state; // copy master's state
5254         for (f = old_nproc; f < team->t.t_nproc; ++f)
5255           team->t.t_threads[f]->th.th_task_state = old_state;
5256       }
5257 
5258 #ifdef KMP_DEBUG
5259       for (f = 0; f < team->t.t_nproc; ++f) {
5260         KMP_DEBUG_ASSERT(team->t.t_threads[f] &&
5261                          team->t.t_threads[f]->th.th_team_nproc ==
5262                              team->t.t_nproc);
5263       }
5264 #endif
5265 
5266 #if OMP_40_ENABLED
5267       KMP_CHECK_UPDATE(team->t.t_proc_bind, new_proc_bind);
5268 #if KMP_AFFINITY_SUPPORTED
5269       __kmp_partition_places(team);
5270 #endif
5271 #endif
5272     } // Check changes in number of threads
5273 
5274 #if OMP_40_ENABLED
5275     kmp_info_t *master = team->t.t_threads[0];
5276     if (master->th.th_teams_microtask) {
5277       for (f = 1; f < new_nproc; ++f) {
5278         // propagate teams construct specific info to workers
5279         kmp_info_t *thr = team->t.t_threads[f];
5280         thr->th.th_teams_microtask = master->th.th_teams_microtask;
5281         thr->th.th_teams_level = master->th.th_teams_level;
5282         thr->th.th_teams_size = master->th.th_teams_size;
5283       }
5284     }
5285 #endif /* OMP_40_ENABLED */
5286 #if KMP_NESTED_HOT_TEAMS
5287     if (level) {
5288       // Sync barrier state for nested hot teams, not needed for outermost hot
5289       // team.
5290       for (f = 1; f < new_nproc; ++f) {
5291         kmp_info_t *thr = team->t.t_threads[f];
5292         int b;
5293         kmp_balign_t *balign = thr->th.th_bar;
5294         for (b = 0; b < bs_last_barrier; ++b) {
5295           balign[b].bb.b_arrived = team->t.t_bar[b].b_arrived;
5296           KMP_DEBUG_ASSERT(balign[b].bb.wait_flag != KMP_BARRIER_PARENT_FLAG);
5297 #if USE_DEBUGGER
5298           balign[b].bb.b_worker_arrived = team->t.t_bar[b].b_team_arrived;
5299 #endif
5300         }
5301       }
5302     }
5303 #endif // KMP_NESTED_HOT_TEAMS
5304 
5305     /* reallocate space for arguments if necessary */
5306     __kmp_alloc_argv_entries(argc, team, TRUE);
5307     KMP_CHECK_UPDATE(team->t.t_argc, argc);
5308     // The hot team re-uses the previous task team,
5309     // if untouched during the previous release->gather phase.
5310 
5311     KF_TRACE(10, (" hot_team = %p\n", team));
5312 
5313 #if KMP_DEBUG
5314     if (__kmp_tasking_mode != tskm_immediate_exec) {
5315       KA_TRACE(20, ("__kmp_allocate_team: hot team task_team[0] = %p "
5316                     "task_team[1] = %p after reinit\n",
5317                     team->t.t_task_team[0], team->t.t_task_team[1]));
5318     }
5319 #endif
5320 
5321 #if OMPT_SUPPORT
5322     __ompt_team_assign_id(team, ompt_parallel_data);
5323 #endif
5324 
5325     KMP_MB();
5326 
5327     return team;
5328   }
5329 
5330   /* next, let's try to take one from the team pool */
5331   KMP_MB();
5332   for (team = CCAST(kmp_team_t *, __kmp_team_pool); (team);) {
5333     /* TODO: consider resizing undersized teams instead of reaping them, now
5334        that we have a resizing mechanism */
5335     if (team->t.t_max_nproc >= max_nproc) {
5336       /* take this team from the team pool */
5337       __kmp_team_pool = team->t.t_next_pool;
5338 
5339       /* setup the team for fresh use */
5340       __kmp_initialize_team(team, new_nproc, new_icvs, NULL);
5341 
5342       KA_TRACE(20, ("__kmp_allocate_team: setting task_team[0] %p and "
5343                     "task_team[1] %p to NULL\n",
5344                     &team->t.t_task_team[0], &team->t.t_task_team[1]));
5345       team->t.t_task_team[0] = NULL;
5346       team->t.t_task_team[1] = NULL;
5347 
5348       /* reallocate space for arguments if necessary */
5349       __kmp_alloc_argv_entries(argc, team, TRUE);
5350       KMP_CHECK_UPDATE(team->t.t_argc, argc);
5351 
5352       KA_TRACE(
5353           20, ("__kmp_allocate_team: team %d init arrived: join=%u, plain=%u\n",
5354                team->t.t_id, KMP_INIT_BARRIER_STATE, KMP_INIT_BARRIER_STATE));
5355       { // Initialize barrier data.
5356         int b;
5357         for (b = 0; b < bs_last_barrier; ++b) {
5358           team->t.t_bar[b].b_arrived = KMP_INIT_BARRIER_STATE;
5359 #if USE_DEBUGGER
5360           team->t.t_bar[b].b_master_arrived = 0;
5361           team->t.t_bar[b].b_team_arrived = 0;
5362 #endif
5363         }
5364       }
5365 
5366 #if OMP_40_ENABLED
5367       team->t.t_proc_bind = new_proc_bind;
5368 #endif
5369 
5370       KA_TRACE(20, ("__kmp_allocate_team: using team from pool %d.\n",
5371                     team->t.t_id));
5372 
5373 #if OMPT_SUPPORT
5374       __ompt_team_assign_id(team, ompt_parallel_data);
5375 #endif
5376 
5377       KMP_MB();
5378 
5379       return team;
5380     }
5381 
5382     /* reap team if it is too small, then loop back and check the next one */
5383     // not sure if this is wise, but, will be redone during the hot-teams
5384     // rewrite.
5385     /* TODO: Use technique to find the right size hot-team, don't reap them */
5386     team = __kmp_reap_team(team);
5387     __kmp_team_pool = team;
5388   }
5389 
5390   /* nothing available in the pool, no matter, make a new team! */
5391   KMP_MB();
5392   team = (kmp_team_t *)__kmp_allocate(sizeof(kmp_team_t));
5393 
5394   /* and set it up */
5395   team->t.t_max_nproc = max_nproc;
5396   /* NOTE well, for some reason allocating one big buffer and dividing it up
5397      seems to really hurt performance a lot on the P4, so, let's not use this */
5398   __kmp_allocate_team_arrays(team, max_nproc);
5399 
5400   KA_TRACE(20, ("__kmp_allocate_team: making a new team\n"));
5401   __kmp_initialize_team(team, new_nproc, new_icvs, NULL);
5402 
5403   KA_TRACE(20, ("__kmp_allocate_team: setting task_team[0] %p and task_team[1] "
5404                 "%p to NULL\n",
5405                 &team->t.t_task_team[0], &team->t.t_task_team[1]));
5406   team->t.t_task_team[0] = NULL; // to be removed, as __kmp_allocate zeroes
5407   // memory, no need to duplicate
5408   team->t.t_task_team[1] = NULL; // to be removed, as __kmp_allocate zeroes
5409   // memory, no need to duplicate
5410 
5411   if (__kmp_storage_map) {
5412     __kmp_print_team_storage_map("team", team, team->t.t_id, new_nproc);
5413   }
5414 
5415   /* allocate space for arguments */
5416   __kmp_alloc_argv_entries(argc, team, FALSE);
5417   team->t.t_argc = argc;
5418 
5419   KA_TRACE(20,
5420            ("__kmp_allocate_team: team %d init arrived: join=%u, plain=%u\n",
5421             team->t.t_id, KMP_INIT_BARRIER_STATE, KMP_INIT_BARRIER_STATE));
5422   { // Initialize barrier data.
5423     int b;
5424     for (b = 0; b < bs_last_barrier; ++b) {
5425       team->t.t_bar[b].b_arrived = KMP_INIT_BARRIER_STATE;
5426 #if USE_DEBUGGER
5427       team->t.t_bar[b].b_master_arrived = 0;
5428       team->t.t_bar[b].b_team_arrived = 0;
5429 #endif
5430     }
5431   }
5432 
5433 #if OMP_40_ENABLED
5434   team->t.t_proc_bind = new_proc_bind;
5435 #endif
5436 
5437 #if OMPT_SUPPORT
5438   __ompt_team_assign_id(team, ompt_parallel_data);
5439   team->t.ompt_serialized_team_info = NULL;
5440 #endif
5441 
5442   KMP_MB();
5443 
5444   KA_TRACE(20, ("__kmp_allocate_team: done creating a new team %d.\n",
5445                 team->t.t_id));
5446 
5447   return team;
5448 }
5449 
5450 /* TODO implement hot-teams at all levels */
5451 /* TODO implement lazy thread release on demand (disband request) */
5452 
5453 /* free the team.  return it to the team pool.  release all the threads
5454  * associated with it */
5455 void __kmp_free_team(kmp_root_t *root,
5456                      kmp_team_t *team USE_NESTED_HOT_ARG(kmp_info_t *master)) {
5457   int f;
5458   KA_TRACE(20, ("__kmp_free_team: T#%d freeing team %d\n", __kmp_get_gtid(),
5459                 team->t.t_id));
5460 
5461   /* verify state */
5462   KMP_DEBUG_ASSERT(root);
5463   KMP_DEBUG_ASSERT(team);
5464   KMP_DEBUG_ASSERT(team->t.t_nproc <= team->t.t_max_nproc);
5465   KMP_DEBUG_ASSERT(team->t.t_threads);
5466 
5467   int use_hot_team = team == root->r.r_hot_team;
5468 #if KMP_NESTED_HOT_TEAMS
5469   int level;
5470   kmp_hot_team_ptr_t *hot_teams;
5471   if (master) {
5472     level = team->t.t_active_level - 1;
5473     if (master->th.th_teams_microtask) { // in teams construct?
5474       if (master->th.th_teams_size.nteams > 1) {
5475         ++level; // level was not increased in teams construct for
5476         // team_of_masters
5477       }
5478       if (team->t.t_pkfn != (microtask_t)__kmp_teams_master &&
5479           master->th.th_teams_level == team->t.t_level) {
5480         ++level; // level was not increased in teams construct for
5481         // team_of_workers before the parallel
5482       } // team->t.t_level will be increased inside parallel
5483     }
5484     hot_teams = master->th.th_hot_teams;
5485     if (level < __kmp_hot_teams_max_level) {
5486       KMP_DEBUG_ASSERT(team == hot_teams[level].hot_team);
5487       use_hot_team = 1;
5488     }
5489   }
5490 #endif // KMP_NESTED_HOT_TEAMS
5491 
5492   /* team is done working */
5493   TCW_SYNC_PTR(team->t.t_pkfn,
5494                NULL); // Important for Debugging Support Library.
5495 #if KMP_OS_WINDOWS
5496   team->t.t_copyin_counter = 0; // init counter for possible reuse
5497 #endif
5498   // Do not reset pointer to parent team to NULL for hot teams.
5499 
5500   /* if we are non-hot team, release our threads */
5501   if (!use_hot_team) {
5502     if (__kmp_tasking_mode != tskm_immediate_exec) {
5503       // Wait for threads to reach reapable state
5504       for (f = 1; f < team->t.t_nproc; ++f) {
5505         KMP_DEBUG_ASSERT(team->t.t_threads[f]);
5506         kmp_info_t *th = team->t.t_threads[f];
5507         volatile kmp_uint32 *state = &th->th.th_reap_state;
5508         while (*state != KMP_SAFE_TO_REAP) {
5509 #if KMP_OS_WINDOWS
5510           // On Windows a thread can be killed at any time, check this
5511           DWORD ecode;
5512           if (!__kmp_is_thread_alive(th, &ecode)) {
5513             *state = KMP_SAFE_TO_REAP; // reset the flag for dead thread
5514             break;
5515           }
5516 #endif
5517           // first check if thread is sleeping
5518           kmp_flag_64 fl(&th->th.th_bar[bs_forkjoin_barrier].bb.b_go, th);
5519           if (fl.is_sleeping())
5520             fl.resume(__kmp_gtid_from_thread(th));
5521           KMP_CPU_PAUSE();
5522         }
5523       }
5524 
5525       // Delete task teams
5526       int tt_idx;
5527       for (tt_idx = 0; tt_idx < 2; ++tt_idx) {
5528         kmp_task_team_t *task_team = team->t.t_task_team[tt_idx];
5529         if (task_team != NULL) {
5530           for (f = 0; f < team->t.t_nproc; ++f) { // threads unref task teams
5531             KMP_DEBUG_ASSERT(team->t.t_threads[f]);
5532             team->t.t_threads[f]->th.th_task_team = NULL;
5533           }
5534           KA_TRACE(
5535               20,
5536               ("__kmp_free_team: T#%d deactivating task_team %p on team %d\n",
5537                __kmp_get_gtid(), task_team, team->t.t_id));
5538 #if KMP_NESTED_HOT_TEAMS
5539           __kmp_free_task_team(master, task_team);
5540 #endif
5541           team->t.t_task_team[tt_idx] = NULL;
5542         }
5543       }
5544     }
5545 
5546     // Reset pointer to parent team only for non-hot teams.
5547     team->t.t_parent = NULL;
5548     team->t.t_level = 0;
5549     team->t.t_active_level = 0;
5550 
5551     /* free the worker threads */
5552     for (f = 1; f < team->t.t_nproc; ++f) {
5553       KMP_DEBUG_ASSERT(team->t.t_threads[f]);
5554       __kmp_free_thread(team->t.t_threads[f]);
5555       team->t.t_threads[f] = NULL;
5556     }
5557 
5558     /* put the team back in the team pool */
5559     /* TODO limit size of team pool, call reap_team if pool too large */
5560     team->t.t_next_pool = CCAST(kmp_team_t *, __kmp_team_pool);
5561     __kmp_team_pool = (volatile kmp_team_t *)team;
5562   } else { // Check if team was created for the masters in a teams construct
5563     // See if first worker is a CG root
5564     KMP_DEBUG_ASSERT(team->t.t_threads[1] &&
5565                      team->t.t_threads[1]->th.th_cg_roots);
5566     if (team->t.t_threads[1]->th.th_cg_roots->cg_root == team->t.t_threads[1]) {
5567       // Clean up the CG root nodes on workers so that this team can be re-used
5568       for (f = 1; f < team->t.t_nproc; ++f) {
5569         kmp_info_t *thr = team->t.t_threads[f];
5570         KMP_DEBUG_ASSERT(thr && thr->th.th_cg_roots &&
5571                          thr->th.th_cg_roots->cg_root == thr);
5572         // Pop current CG root off list
5573         kmp_cg_root_t *tmp = thr->th.th_cg_roots;
5574         thr->th.th_cg_roots = tmp->up;
5575         KA_TRACE(100, ("__kmp_free_team: Thread %p popping node %p and moving"
5576                        " up to node %p. cg_nthreads was %d\n",
5577                        thr, tmp, thr->th.th_cg_roots, tmp->cg_nthreads));
5578         __kmp_free(tmp);
5579         // Restore current task's thread_limit from CG root
5580         if (thr->th.th_cg_roots)
5581           thr->th.th_current_task->td_icvs.thread_limit =
5582               thr->th.th_cg_roots->cg_thread_limit;
5583       }
5584     }
5585   }
5586 
5587   KMP_MB();
5588 }
5589 
5590 /* reap the team.  destroy it, reclaim all its resources and free its memory */
5591 kmp_team_t *__kmp_reap_team(kmp_team_t *team) {
5592   kmp_team_t *next_pool = team->t.t_next_pool;
5593 
5594   KMP_DEBUG_ASSERT(team);
5595   KMP_DEBUG_ASSERT(team->t.t_dispatch);
5596   KMP_DEBUG_ASSERT(team->t.t_disp_buffer);
5597   KMP_DEBUG_ASSERT(team->t.t_threads);
5598   KMP_DEBUG_ASSERT(team->t.t_argv);
5599 
5600   /* TODO clean the threads that are a part of this? */
5601 
5602   /* free stuff */
5603   __kmp_free_team_arrays(team);
5604   if (team->t.t_argv != &team->t.t_inline_argv[0])
5605     __kmp_free((void *)team->t.t_argv);
5606   __kmp_free(team);
5607 
5608   KMP_MB();
5609   return next_pool;
5610 }
5611 
5612 // Free the thread.  Don't reap it, just place it on the pool of available
5613 // threads.
5614 //
5615 // Changes for Quad issue 527845: We need a predictable OMP tid <-> gtid
5616 // binding for the affinity mechanism to be useful.
5617 //
5618 // Now, we always keep the free list (__kmp_thread_pool) sorted by gtid.
5619 // However, we want to avoid a potential performance problem by always
5620 // scanning through the list to find the correct point at which to insert
5621 // the thread (potential N**2 behavior).  To do this we keep track of the
5622 // last place a thread struct was inserted (__kmp_thread_pool_insert_pt).
5623 // With single-level parallelism, threads will always be added to the tail
5624 // of the list, kept track of by __kmp_thread_pool_insert_pt.  With nested
5625 // parallelism, all bets are off and we may need to scan through the entire
5626 // free list.
5627 //
5628 // This change also has a potentially large performance benefit, for some
5629 // applications.  Previously, as threads were freed from the hot team, they
5630 // would be placed back on the free list in inverse order.  If the hot team
5631 // grew back to it's original size, then the freed thread would be placed
5632 // back on the hot team in reverse order.  This could cause bad cache
5633 // locality problems on programs where the size of the hot team regularly
5634 // grew and shrunk.
5635 //
5636 // Now, for single-level parallelism, the OMP tid is alway == gtid.
5637 void __kmp_free_thread(kmp_info_t *this_th) {
5638   int gtid;
5639   kmp_info_t **scan;
5640 
5641   KA_TRACE(20, ("__kmp_free_thread: T#%d putting T#%d back on free pool.\n",
5642                 __kmp_get_gtid(), this_th->th.th_info.ds.ds_gtid));
5643 
5644   KMP_DEBUG_ASSERT(this_th);
5645 
5646   // When moving thread to pool, switch thread to wait on own b_go flag, and
5647   // uninitialized (NULL team).
5648   int b;
5649   kmp_balign_t *balign = this_th->th.th_bar;
5650   for (b = 0; b < bs_last_barrier; ++b) {
5651     if (balign[b].bb.wait_flag == KMP_BARRIER_PARENT_FLAG)
5652       balign[b].bb.wait_flag = KMP_BARRIER_SWITCH_TO_OWN_FLAG;
5653     balign[b].bb.team = NULL;
5654     balign[b].bb.leaf_kids = 0;
5655   }
5656   this_th->th.th_task_state = 0;
5657   this_th->th.th_reap_state = KMP_SAFE_TO_REAP;
5658 
5659   /* put thread back on the free pool */
5660   TCW_PTR(this_th->th.th_team, NULL);
5661   TCW_PTR(this_th->th.th_root, NULL);
5662   TCW_PTR(this_th->th.th_dispatch, NULL); /* NOT NEEDED */
5663 
5664   while (this_th->th.th_cg_roots) {
5665     this_th->th.th_cg_roots->cg_nthreads--;
5666     KA_TRACE(100, ("__kmp_free_thread: Thread %p decrement cg_nthreads on node"
5667                    " %p of thread  %p to %d\n",
5668                    this_th, this_th->th.th_cg_roots,
5669                    this_th->th.th_cg_roots->cg_root,
5670                    this_th->th.th_cg_roots->cg_nthreads));
5671     kmp_cg_root_t *tmp = this_th->th.th_cg_roots;
5672     if (tmp->cg_root == this_th) { // Thread is a cg_root
5673       KMP_DEBUG_ASSERT(tmp->cg_nthreads == 0);
5674       KA_TRACE(
5675           5, ("__kmp_free_thread: Thread %p freeing node %p\n", this_th, tmp));
5676       this_th->th.th_cg_roots = tmp->up;
5677       __kmp_free(tmp);
5678     } else { // Worker thread
5679       this_th->th.th_cg_roots = NULL;
5680       break;
5681     }
5682   }
5683 
5684   /* If the implicit task assigned to this thread can be used by other threads
5685    * -> multiple threads can share the data and try to free the task at
5686    * __kmp_reap_thread at exit. This duplicate use of the task data can happen
5687    * with higher probability when hot team is disabled but can occurs even when
5688    * the hot team is enabled */
5689   __kmp_free_implicit_task(this_th);
5690   this_th->th.th_current_task = NULL;
5691 
5692   // If the __kmp_thread_pool_insert_pt is already past the new insert
5693   // point, then we need to re-scan the entire list.
5694   gtid = this_th->th.th_info.ds.ds_gtid;
5695   if (__kmp_thread_pool_insert_pt != NULL) {
5696     KMP_DEBUG_ASSERT(__kmp_thread_pool != NULL);
5697     if (__kmp_thread_pool_insert_pt->th.th_info.ds.ds_gtid > gtid) {
5698       __kmp_thread_pool_insert_pt = NULL;
5699     }
5700   }
5701 
5702   // Scan down the list to find the place to insert the thread.
5703   // scan is the address of a link in the list, possibly the address of
5704   // __kmp_thread_pool itself.
5705   //
5706   // In the absence of nested parallism, the for loop will have 0 iterations.
5707   if (__kmp_thread_pool_insert_pt != NULL) {
5708     scan = &(__kmp_thread_pool_insert_pt->th.th_next_pool);
5709   } else {
5710     scan = CCAST(kmp_info_t **, &__kmp_thread_pool);
5711   }
5712   for (; (*scan != NULL) && ((*scan)->th.th_info.ds.ds_gtid < gtid);
5713        scan = &((*scan)->th.th_next_pool))
5714     ;
5715 
5716   // Insert the new element on the list, and set __kmp_thread_pool_insert_pt
5717   // to its address.
5718   TCW_PTR(this_th->th.th_next_pool, *scan);
5719   __kmp_thread_pool_insert_pt = *scan = this_th;
5720   KMP_DEBUG_ASSERT((this_th->th.th_next_pool == NULL) ||
5721                    (this_th->th.th_info.ds.ds_gtid <
5722                     this_th->th.th_next_pool->th.th_info.ds.ds_gtid));
5723   TCW_4(this_th->th.th_in_pool, TRUE);
5724   __kmp_suspend_initialize_thread(this_th);
5725   __kmp_lock_suspend_mx(this_th);
5726   if (this_th->th.th_active == TRUE) {
5727     KMP_ATOMIC_INC(&__kmp_thread_pool_active_nth);
5728     this_th->th.th_active_in_pool = TRUE;
5729   }
5730 #if KMP_DEBUG
5731   else {
5732     KMP_DEBUG_ASSERT(this_th->th.th_active_in_pool == FALSE);
5733   }
5734 #endif
5735   __kmp_unlock_suspend_mx(this_th);
5736 
5737   TCW_4(__kmp_nth, __kmp_nth - 1);
5738 
5739 #ifdef KMP_ADJUST_BLOCKTIME
5740   /* Adjust blocktime back to user setting or default if necessary */
5741   /* Middle initialization might never have occurred                */
5742   if (!__kmp_env_blocktime && (__kmp_avail_proc > 0)) {
5743     KMP_DEBUG_ASSERT(__kmp_avail_proc > 0);
5744     if (__kmp_nth <= __kmp_avail_proc) {
5745       __kmp_zero_bt = FALSE;
5746     }
5747   }
5748 #endif /* KMP_ADJUST_BLOCKTIME */
5749 
5750   KMP_MB();
5751 }
5752 
5753 /* ------------------------------------------------------------------------ */
5754 
5755 void *__kmp_launch_thread(kmp_info_t *this_thr) {
5756   int gtid = this_thr->th.th_info.ds.ds_gtid;
5757   /*    void                 *stack_data;*/
5758   kmp_team_t *(*volatile pteam);
5759 
5760   KMP_MB();
5761   KA_TRACE(10, ("__kmp_launch_thread: T#%d start\n", gtid));
5762 
5763   if (__kmp_env_consistency_check) {
5764     this_thr->th.th_cons = __kmp_allocate_cons_stack(gtid); // ATT: Memory leak?
5765   }
5766 
5767 #if OMPT_SUPPORT
5768   ompt_data_t *thread_data;
5769   if (ompt_enabled.enabled) {
5770     thread_data = &(this_thr->th.ompt_thread_info.thread_data);
5771     *thread_data = ompt_data_none;
5772 
5773     this_thr->th.ompt_thread_info.state = ompt_state_overhead;
5774     this_thr->th.ompt_thread_info.wait_id = 0;
5775     this_thr->th.ompt_thread_info.idle_frame = OMPT_GET_FRAME_ADDRESS(0);
5776     if (ompt_enabled.ompt_callback_thread_begin) {
5777       ompt_callbacks.ompt_callback(ompt_callback_thread_begin)(
5778           ompt_thread_worker, thread_data);
5779     }
5780   }
5781 #endif
5782 
5783 #if OMPT_SUPPORT
5784   if (ompt_enabled.enabled) {
5785     this_thr->th.ompt_thread_info.state = ompt_state_idle;
5786   }
5787 #endif
5788   /* This is the place where threads wait for work */
5789   while (!TCR_4(__kmp_global.g.g_done)) {
5790     KMP_DEBUG_ASSERT(this_thr == __kmp_threads[gtid]);
5791     KMP_MB();
5792 
5793     /* wait for work to do */
5794     KA_TRACE(20, ("__kmp_launch_thread: T#%d waiting for work\n", gtid));
5795 
5796     /* No tid yet since not part of a team */
5797     __kmp_fork_barrier(gtid, KMP_GTID_DNE);
5798 
5799 #if OMPT_SUPPORT
5800     if (ompt_enabled.enabled) {
5801       this_thr->th.ompt_thread_info.state = ompt_state_overhead;
5802     }
5803 #endif
5804 
5805     pteam = (kmp_team_t * (*))(&this_thr->th.th_team);
5806 
5807     /* have we been allocated? */
5808     if (TCR_SYNC_PTR(*pteam) && !TCR_4(__kmp_global.g.g_done)) {
5809       /* we were just woken up, so run our new task */
5810       if (TCR_SYNC_PTR((*pteam)->t.t_pkfn) != NULL) {
5811         int rc;
5812         KA_TRACE(20,
5813                  ("__kmp_launch_thread: T#%d(%d:%d) invoke microtask = %p\n",
5814                   gtid, (*pteam)->t.t_id, __kmp_tid_from_gtid(gtid),
5815                   (*pteam)->t.t_pkfn));
5816 
5817         updateHWFPControl(*pteam);
5818 
5819 #if OMPT_SUPPORT
5820         if (ompt_enabled.enabled) {
5821           this_thr->th.ompt_thread_info.state = ompt_state_work_parallel;
5822         }
5823 #endif
5824 
5825         rc = (*pteam)->t.t_invoke(gtid);
5826         KMP_ASSERT(rc);
5827 
5828         KMP_MB();
5829         KA_TRACE(20, ("__kmp_launch_thread: T#%d(%d:%d) done microtask = %p\n",
5830                       gtid, (*pteam)->t.t_id, __kmp_tid_from_gtid(gtid),
5831                       (*pteam)->t.t_pkfn));
5832       }
5833 #if OMPT_SUPPORT
5834       if (ompt_enabled.enabled) {
5835         /* no frame set while outside task */
5836         __ompt_get_task_info_object(0)->frame.exit_frame = ompt_data_none;
5837 
5838         this_thr->th.ompt_thread_info.state = ompt_state_overhead;
5839       }
5840 #endif
5841       /* join barrier after parallel region */
5842       __kmp_join_barrier(gtid);
5843     }
5844   }
5845   TCR_SYNC_PTR((intptr_t)__kmp_global.g.g_done);
5846 
5847 #if OMPT_SUPPORT
5848   if (ompt_enabled.ompt_callback_thread_end) {
5849     ompt_callbacks.ompt_callback(ompt_callback_thread_end)(thread_data);
5850   }
5851 #endif
5852 
5853   this_thr->th.th_task_team = NULL;
5854   /* run the destructors for the threadprivate data for this thread */
5855   __kmp_common_destroy_gtid(gtid);
5856 
5857   KA_TRACE(10, ("__kmp_launch_thread: T#%d done\n", gtid));
5858   KMP_MB();
5859   return this_thr;
5860 }
5861 
5862 /* ------------------------------------------------------------------------ */
5863 
5864 void __kmp_internal_end_dest(void *specific_gtid) {
5865 #if KMP_COMPILER_ICC
5866 #pragma warning(push)
5867 #pragma warning(disable : 810) // conversion from "void *" to "int" may lose
5868 // significant bits
5869 #endif
5870   // Make sure no significant bits are lost
5871   int gtid = (kmp_intptr_t)specific_gtid - 1;
5872 #if KMP_COMPILER_ICC
5873 #pragma warning(pop)
5874 #endif
5875 
5876   KA_TRACE(30, ("__kmp_internal_end_dest: T#%d\n", gtid));
5877   /* NOTE: the gtid is stored as gitd+1 in the thread-local-storage
5878    * this is because 0 is reserved for the nothing-stored case */
5879 
5880   /* josh: One reason for setting the gtid specific data even when it is being
5881      destroyed by pthread is to allow gtid lookup through thread specific data
5882      (__kmp_gtid_get_specific).  Some of the code, especially stat code,
5883      that gets executed in the call to __kmp_internal_end_thread, actually
5884      gets the gtid through the thread specific data.  Setting it here seems
5885      rather inelegant and perhaps wrong, but allows __kmp_internal_end_thread
5886      to run smoothly.
5887      todo: get rid of this after we remove the dependence on
5888      __kmp_gtid_get_specific  */
5889   if (gtid >= 0 && KMP_UBER_GTID(gtid))
5890     __kmp_gtid_set_specific(gtid);
5891 #ifdef KMP_TDATA_GTID
5892   __kmp_gtid = gtid;
5893 #endif
5894   __kmp_internal_end_thread(gtid);
5895 }
5896 
5897 #if KMP_OS_UNIX && KMP_DYNAMIC_LIB
5898 
5899 // 2009-09-08 (lev): It looks the destructor does not work. In simple test cases
5900 // destructors work perfectly, but in real libomp.so I have no evidence it is
5901 // ever called. However, -fini linker option in makefile.mk works fine.
5902 
5903 __attribute__((destructor)) void __kmp_internal_end_dtor(void) {
5904   __kmp_internal_end_atexit();
5905 }
5906 
5907 void __kmp_internal_end_fini(void) { __kmp_internal_end_atexit(); }
5908 
5909 #endif
5910 
5911 /* [Windows] josh: when the atexit handler is called, there may still be more
5912    than one thread alive */
5913 void __kmp_internal_end_atexit(void) {
5914   KA_TRACE(30, ("__kmp_internal_end_atexit\n"));
5915   /* [Windows]
5916      josh: ideally, we want to completely shutdown the library in this atexit
5917      handler, but stat code that depends on thread specific data for gtid fails
5918      because that data becomes unavailable at some point during the shutdown, so
5919      we call __kmp_internal_end_thread instead. We should eventually remove the
5920      dependency on __kmp_get_specific_gtid in the stat code and use
5921      __kmp_internal_end_library to cleanly shutdown the library.
5922 
5923      // TODO: Can some of this comment about GVS be removed?
5924      I suspect that the offending stat code is executed when the calling thread
5925      tries to clean up a dead root thread's data structures, resulting in GVS
5926      code trying to close the GVS structures for that thread, but since the stat
5927      code uses __kmp_get_specific_gtid to get the gtid with the assumption that
5928      the calling thread is cleaning up itself instead of another thread, it get
5929      confused. This happens because allowing a thread to unregister and cleanup
5930      another thread is a recent modification for addressing an issue.
5931      Based on the current design (20050722), a thread may end up
5932      trying to unregister another thread only if thread death does not trigger
5933      the calling of __kmp_internal_end_thread.  For Linux* OS, there is the
5934      thread specific data destructor function to detect thread death. For
5935      Windows dynamic, there is DllMain(THREAD_DETACH). For Windows static, there
5936      is nothing.  Thus, the workaround is applicable only for Windows static
5937      stat library. */
5938   __kmp_internal_end_library(-1);
5939 #if KMP_OS_WINDOWS
5940   __kmp_close_console();
5941 #endif
5942 }
5943 
5944 static void __kmp_reap_thread(kmp_info_t *thread, int is_root) {
5945   // It is assumed __kmp_forkjoin_lock is acquired.
5946 
5947   int gtid;
5948 
5949   KMP_DEBUG_ASSERT(thread != NULL);
5950 
5951   gtid = thread->th.th_info.ds.ds_gtid;
5952 
5953   if (!is_root) {
5954     if (__kmp_dflt_blocktime != KMP_MAX_BLOCKTIME) {
5955       /* Assume the threads are at the fork barrier here */
5956       KA_TRACE(
5957           20, ("__kmp_reap_thread: releasing T#%d from fork barrier for reap\n",
5958                gtid));
5959       /* Need release fence here to prevent seg faults for tree forkjoin barrier
5960        * (GEH) */
5961       ANNOTATE_HAPPENS_BEFORE(thread);
5962       kmp_flag_64 flag(&thread->th.th_bar[bs_forkjoin_barrier].bb.b_go, thread);
5963       __kmp_release_64(&flag);
5964     }
5965 
5966     // Terminate OS thread.
5967     __kmp_reap_worker(thread);
5968 
5969     // The thread was killed asynchronously.  If it was actively
5970     // spinning in the thread pool, decrement the global count.
5971     //
5972     // There is a small timing hole here - if the worker thread was just waking
5973     // up after sleeping in the pool, had reset it's th_active_in_pool flag but
5974     // not decremented the global counter __kmp_thread_pool_active_nth yet, then
5975     // the global counter might not get updated.
5976     //
5977     // Currently, this can only happen as the library is unloaded,
5978     // so there are no harmful side effects.
5979     if (thread->th.th_active_in_pool) {
5980       thread->th.th_active_in_pool = FALSE;
5981       KMP_ATOMIC_DEC(&__kmp_thread_pool_active_nth);
5982       KMP_DEBUG_ASSERT(__kmp_thread_pool_active_nth >= 0);
5983     }
5984   }
5985 
5986   __kmp_free_implicit_task(thread);
5987 
5988 // Free the fast memory for tasking
5989 #if USE_FAST_MEMORY
5990   __kmp_free_fast_memory(thread);
5991 #endif /* USE_FAST_MEMORY */
5992 
5993   __kmp_suspend_uninitialize_thread(thread);
5994 
5995   KMP_DEBUG_ASSERT(__kmp_threads[gtid] == thread);
5996   TCW_SYNC_PTR(__kmp_threads[gtid], NULL);
5997 
5998   --__kmp_all_nth;
5999 // __kmp_nth was decremented when thread is added to the pool.
6000 
6001 #ifdef KMP_ADJUST_BLOCKTIME
6002   /* Adjust blocktime back to user setting or default if necessary */
6003   /* Middle initialization might never have occurred                */
6004   if (!__kmp_env_blocktime && (__kmp_avail_proc > 0)) {
6005     KMP_DEBUG_ASSERT(__kmp_avail_proc > 0);
6006     if (__kmp_nth <= __kmp_avail_proc) {
6007       __kmp_zero_bt = FALSE;
6008     }
6009   }
6010 #endif /* KMP_ADJUST_BLOCKTIME */
6011 
6012   /* free the memory being used */
6013   if (__kmp_env_consistency_check) {
6014     if (thread->th.th_cons) {
6015       __kmp_free_cons_stack(thread->th.th_cons);
6016       thread->th.th_cons = NULL;
6017     }
6018   }
6019 
6020   if (thread->th.th_pri_common != NULL) {
6021     __kmp_free(thread->th.th_pri_common);
6022     thread->th.th_pri_common = NULL;
6023   }
6024 
6025   if (thread->th.th_task_state_memo_stack != NULL) {
6026     __kmp_free(thread->th.th_task_state_memo_stack);
6027     thread->th.th_task_state_memo_stack = NULL;
6028   }
6029 
6030 #if KMP_USE_BGET
6031   if (thread->th.th_local.bget_data != NULL) {
6032     __kmp_finalize_bget(thread);
6033   }
6034 #endif
6035 
6036 #if KMP_AFFINITY_SUPPORTED
6037   if (thread->th.th_affin_mask != NULL) {
6038     KMP_CPU_FREE(thread->th.th_affin_mask);
6039     thread->th.th_affin_mask = NULL;
6040   }
6041 #endif /* KMP_AFFINITY_SUPPORTED */
6042 
6043 #if KMP_USE_HIER_SCHED
6044   if (thread->th.th_hier_bar_data != NULL) {
6045     __kmp_free(thread->th.th_hier_bar_data);
6046     thread->th.th_hier_bar_data = NULL;
6047   }
6048 #endif
6049 
6050   __kmp_reap_team(thread->th.th_serial_team);
6051   thread->th.th_serial_team = NULL;
6052   __kmp_free(thread);
6053 
6054   KMP_MB();
6055 
6056 } // __kmp_reap_thread
6057 
6058 static void __kmp_internal_end(void) {
6059   int i;
6060 
6061   /* First, unregister the library */
6062   __kmp_unregister_library();
6063 
6064 #if KMP_OS_WINDOWS
6065   /* In Win static library, we can't tell when a root actually dies, so we
6066      reclaim the data structures for any root threads that have died but not
6067      unregistered themselves, in order to shut down cleanly.
6068      In Win dynamic library we also can't tell when a thread dies.  */
6069   __kmp_reclaim_dead_roots(); // AC: moved here to always clean resources of
6070 // dead roots
6071 #endif
6072 
6073   for (i = 0; i < __kmp_threads_capacity; i++)
6074     if (__kmp_root[i])
6075       if (__kmp_root[i]->r.r_active)
6076         break;
6077   KMP_MB(); /* Flush all pending memory write invalidates.  */
6078   TCW_SYNC_4(__kmp_global.g.g_done, TRUE);
6079 
6080   if (i < __kmp_threads_capacity) {
6081 #if KMP_USE_MONITOR
6082     // 2009-09-08 (lev): Other alive roots found. Why do we kill the monitor??
6083     KMP_MB(); /* Flush all pending memory write invalidates.  */
6084 
6085     // Need to check that monitor was initialized before reaping it. If we are
6086     // called form __kmp_atfork_child (which sets __kmp_init_parallel = 0), then
6087     // __kmp_monitor will appear to contain valid data, but it is only valid in
6088     // the parent process, not the child.
6089     // New behavior (201008): instead of keying off of the flag
6090     // __kmp_init_parallel, the monitor thread creation is keyed off
6091     // of the new flag __kmp_init_monitor.
6092     __kmp_acquire_bootstrap_lock(&__kmp_monitor_lock);
6093     if (TCR_4(__kmp_init_monitor)) {
6094       __kmp_reap_monitor(&__kmp_monitor);
6095       TCW_4(__kmp_init_monitor, 0);
6096     }
6097     __kmp_release_bootstrap_lock(&__kmp_monitor_lock);
6098     KA_TRACE(10, ("__kmp_internal_end: monitor reaped\n"));
6099 #endif // KMP_USE_MONITOR
6100   } else {
6101 /* TODO move this to cleanup code */
6102 #ifdef KMP_DEBUG
6103     /* make sure that everything has properly ended */
6104     for (i = 0; i < __kmp_threads_capacity; i++) {
6105       if (__kmp_root[i]) {
6106         //                    KMP_ASSERT( ! KMP_UBER_GTID( i ) );         // AC:
6107         //                    there can be uber threads alive here
6108         KMP_ASSERT(!__kmp_root[i]->r.r_active); // TODO: can they be active?
6109       }
6110     }
6111 #endif
6112 
6113     KMP_MB();
6114 
6115     // Reap the worker threads.
6116     // This is valid for now, but be careful if threads are reaped sooner.
6117     while (__kmp_thread_pool != NULL) { // Loop thru all the thread in the pool.
6118       // Get the next thread from the pool.
6119       kmp_info_t *thread = CCAST(kmp_info_t *, __kmp_thread_pool);
6120       __kmp_thread_pool = thread->th.th_next_pool;
6121       // Reap it.
6122       KMP_DEBUG_ASSERT(thread->th.th_reap_state == KMP_SAFE_TO_REAP);
6123       thread->th.th_next_pool = NULL;
6124       thread->th.th_in_pool = FALSE;
6125       thread->th.th_active_in_pool = FALSE;
6126       KMP_ATOMIC_DEC(&__kmp_thread_pool_active_nth);
6127       __kmp_reap_thread(thread, 0);
6128     }
6129     __kmp_thread_pool_insert_pt = NULL;
6130 
6131     // Reap teams.
6132     while (__kmp_team_pool != NULL) { // Loop thru all the teams in the pool.
6133       // Get the next team from the pool.
6134       kmp_team_t *team = CCAST(kmp_team_t *, __kmp_team_pool);
6135       __kmp_team_pool = team->t.t_next_pool;
6136       // Reap it.
6137       team->t.t_next_pool = NULL;
6138       __kmp_reap_team(team);
6139     }
6140 
6141     __kmp_reap_task_teams();
6142 
6143 #if KMP_OS_UNIX
6144     // Threads that are not reaped should not access any resources since they
6145     // are going to be deallocated soon, so the shutdown sequence should wait
6146     // until all threads either exit the final spin-waiting loop or begin
6147     // sleeping after the given blocktime.
6148     for (i = 0; i < __kmp_threads_capacity; i++) {
6149       kmp_info_t *thr = __kmp_threads[i];
6150       while (thr && KMP_ATOMIC_LD_ACQ(&thr->th.th_blocking))
6151         KMP_CPU_PAUSE();
6152     }
6153 #endif
6154 
6155     for (i = 0; i < __kmp_threads_capacity; ++i) {
6156       // TBD: Add some checking...
6157       // Something like KMP_DEBUG_ASSERT( __kmp_thread[ i ] == NULL );
6158     }
6159 
6160     /* Make sure all threadprivate destructors get run by joining with all
6161        worker threads before resetting this flag */
6162     TCW_SYNC_4(__kmp_init_common, FALSE);
6163 
6164     KA_TRACE(10, ("__kmp_internal_end: all workers reaped\n"));
6165     KMP_MB();
6166 
6167 #if KMP_USE_MONITOR
6168     // See note above: One of the possible fixes for CQ138434 / CQ140126
6169     //
6170     // FIXME: push both code fragments down and CSE them?
6171     // push them into __kmp_cleanup() ?
6172     __kmp_acquire_bootstrap_lock(&__kmp_monitor_lock);
6173     if (TCR_4(__kmp_init_monitor)) {
6174       __kmp_reap_monitor(&__kmp_monitor);
6175       TCW_4(__kmp_init_monitor, 0);
6176     }
6177     __kmp_release_bootstrap_lock(&__kmp_monitor_lock);
6178     KA_TRACE(10, ("__kmp_internal_end: monitor reaped\n"));
6179 #endif
6180   } /* else !__kmp_global.t_active */
6181   TCW_4(__kmp_init_gtid, FALSE);
6182   KMP_MB(); /* Flush all pending memory write invalidates.  */
6183 
6184   __kmp_cleanup();
6185 #if OMPT_SUPPORT
6186   ompt_fini();
6187 #endif
6188 }
6189 
6190 void __kmp_internal_end_library(int gtid_req) {
6191   /* if we have already cleaned up, don't try again, it wouldn't be pretty */
6192   /* this shouldn't be a race condition because __kmp_internal_end() is the
6193      only place to clear __kmp_serial_init */
6194   /* we'll check this later too, after we get the lock */
6195   // 2009-09-06: We do not set g_abort without setting g_done. This check looks
6196   // redundaant, because the next check will work in any case.
6197   if (__kmp_global.g.g_abort) {
6198     KA_TRACE(11, ("__kmp_internal_end_library: abort, exiting\n"));
6199     /* TODO abort? */
6200     return;
6201   }
6202   if (TCR_4(__kmp_global.g.g_done) || !__kmp_init_serial) {
6203     KA_TRACE(10, ("__kmp_internal_end_library: already finished\n"));
6204     return;
6205   }
6206 
6207   KMP_MB(); /* Flush all pending memory write invalidates.  */
6208 
6209   /* find out who we are and what we should do */
6210   {
6211     int gtid = (gtid_req >= 0) ? gtid_req : __kmp_gtid_get_specific();
6212     KA_TRACE(
6213         10, ("__kmp_internal_end_library: enter T#%d  (%d)\n", gtid, gtid_req));
6214     if (gtid == KMP_GTID_SHUTDOWN) {
6215       KA_TRACE(10, ("__kmp_internal_end_library: !__kmp_init_runtime, system "
6216                     "already shutdown\n"));
6217       return;
6218     } else if (gtid == KMP_GTID_MONITOR) {
6219       KA_TRACE(10, ("__kmp_internal_end_library: monitor thread, gtid not "
6220                     "registered, or system shutdown\n"));
6221       return;
6222     } else if (gtid == KMP_GTID_DNE) {
6223       KA_TRACE(10, ("__kmp_internal_end_library: gtid not registered or system "
6224                     "shutdown\n"));
6225       /* we don't know who we are, but we may still shutdown the library */
6226     } else if (KMP_UBER_GTID(gtid)) {
6227       /* unregister ourselves as an uber thread.  gtid is no longer valid */
6228       if (__kmp_root[gtid]->r.r_active) {
6229         __kmp_global.g.g_abort = -1;
6230         TCW_SYNC_4(__kmp_global.g.g_done, TRUE);
6231         KA_TRACE(10,
6232                  ("__kmp_internal_end_library: root still active, abort T#%d\n",
6233                   gtid));
6234         return;
6235       } else {
6236         KA_TRACE(
6237             10,
6238             ("__kmp_internal_end_library: unregistering sibling T#%d\n", gtid));
6239         __kmp_unregister_root_current_thread(gtid);
6240       }
6241     } else {
6242 /* worker threads may call this function through the atexit handler, if they
6243  * call exit() */
6244 /* For now, skip the usual subsequent processing and just dump the debug buffer.
6245    TODO: do a thorough shutdown instead */
6246 #ifdef DUMP_DEBUG_ON_EXIT
6247       if (__kmp_debug_buf)
6248         __kmp_dump_debug_buffer();
6249 #endif
6250       return;
6251     }
6252   }
6253   /* synchronize the termination process */
6254   __kmp_acquire_bootstrap_lock(&__kmp_initz_lock);
6255 
6256   /* have we already finished */
6257   if (__kmp_global.g.g_abort) {
6258     KA_TRACE(10, ("__kmp_internal_end_library: abort, exiting\n"));
6259     /* TODO abort? */
6260     __kmp_release_bootstrap_lock(&__kmp_initz_lock);
6261     return;
6262   }
6263   if (TCR_4(__kmp_global.g.g_done) || !__kmp_init_serial) {
6264     __kmp_release_bootstrap_lock(&__kmp_initz_lock);
6265     return;
6266   }
6267 
6268   /* We need this lock to enforce mutex between this reading of
6269      __kmp_threads_capacity and the writing by __kmp_register_root.
6270      Alternatively, we can use a counter of roots that is atomically updated by
6271      __kmp_get_global_thread_id_reg, __kmp_do_serial_initialize and
6272      __kmp_internal_end_*.  */
6273   __kmp_acquire_bootstrap_lock(&__kmp_forkjoin_lock);
6274 
6275   /* now we can safely conduct the actual termination */
6276   __kmp_internal_end();
6277 
6278   __kmp_release_bootstrap_lock(&__kmp_forkjoin_lock);
6279   __kmp_release_bootstrap_lock(&__kmp_initz_lock);
6280 
6281   KA_TRACE(10, ("__kmp_internal_end_library: exit\n"));
6282 
6283 #ifdef DUMP_DEBUG_ON_EXIT
6284   if (__kmp_debug_buf)
6285     __kmp_dump_debug_buffer();
6286 #endif
6287 
6288 #if KMP_OS_WINDOWS
6289   __kmp_close_console();
6290 #endif
6291 
6292   __kmp_fini_allocator();
6293 
6294 } // __kmp_internal_end_library
6295 
6296 void __kmp_internal_end_thread(int gtid_req) {
6297   int i;
6298 
6299   /* if we have already cleaned up, don't try again, it wouldn't be pretty */
6300   /* this shouldn't be a race condition because __kmp_internal_end() is the
6301    * only place to clear __kmp_serial_init */
6302   /* we'll check this later too, after we get the lock */
6303   // 2009-09-06: We do not set g_abort without setting g_done. This check looks
6304   // redundant, because the next check will work in any case.
6305   if (__kmp_global.g.g_abort) {
6306     KA_TRACE(11, ("__kmp_internal_end_thread: abort, exiting\n"));
6307     /* TODO abort? */
6308     return;
6309   }
6310   if (TCR_4(__kmp_global.g.g_done) || !__kmp_init_serial) {
6311     KA_TRACE(10, ("__kmp_internal_end_thread: already finished\n"));
6312     return;
6313   }
6314 
6315   KMP_MB(); /* Flush all pending memory write invalidates.  */
6316 
6317   /* find out who we are and what we should do */
6318   {
6319     int gtid = (gtid_req >= 0) ? gtid_req : __kmp_gtid_get_specific();
6320     KA_TRACE(10,
6321              ("__kmp_internal_end_thread: enter T#%d  (%d)\n", gtid, gtid_req));
6322     if (gtid == KMP_GTID_SHUTDOWN) {
6323       KA_TRACE(10, ("__kmp_internal_end_thread: !__kmp_init_runtime, system "
6324                     "already shutdown\n"));
6325       return;
6326     } else if (gtid == KMP_GTID_MONITOR) {
6327       KA_TRACE(10, ("__kmp_internal_end_thread: monitor thread, gtid not "
6328                     "registered, or system shutdown\n"));
6329       return;
6330     } else if (gtid == KMP_GTID_DNE) {
6331       KA_TRACE(10, ("__kmp_internal_end_thread: gtid not registered or system "
6332                     "shutdown\n"));
6333       return;
6334       /* we don't know who we are */
6335     } else if (KMP_UBER_GTID(gtid)) {
6336       /* unregister ourselves as an uber thread.  gtid is no longer valid */
6337       if (__kmp_root[gtid]->r.r_active) {
6338         __kmp_global.g.g_abort = -1;
6339         TCW_SYNC_4(__kmp_global.g.g_done, TRUE);
6340         KA_TRACE(10,
6341                  ("__kmp_internal_end_thread: root still active, abort T#%d\n",
6342                   gtid));
6343         return;
6344       } else {
6345         KA_TRACE(10, ("__kmp_internal_end_thread: unregistering sibling T#%d\n",
6346                       gtid));
6347         __kmp_unregister_root_current_thread(gtid);
6348       }
6349     } else {
6350       /* just a worker thread, let's leave */
6351       KA_TRACE(10, ("__kmp_internal_end_thread: worker thread T#%d\n", gtid));
6352 
6353       if (gtid >= 0) {
6354         __kmp_threads[gtid]->th.th_task_team = NULL;
6355       }
6356 
6357       KA_TRACE(10,
6358                ("__kmp_internal_end_thread: worker thread done, exiting T#%d\n",
6359                 gtid));
6360       return;
6361     }
6362   }
6363 #if KMP_DYNAMIC_LIB
6364 #if OMP_50_ENABLED
6365   if (__kmp_pause_status != kmp_hard_paused)
6366 #endif
6367   // AC: lets not shutdown the dynamic library at the exit of uber thread,
6368   // because we will better shutdown later in the library destructor.
6369   {
6370     KA_TRACE(10, ("__kmp_internal_end_thread: exiting T#%d\n", gtid_req));
6371     return;
6372   }
6373 #endif
6374   /* synchronize the termination process */
6375   __kmp_acquire_bootstrap_lock(&__kmp_initz_lock);
6376 
6377   /* have we already finished */
6378   if (__kmp_global.g.g_abort) {
6379     KA_TRACE(10, ("__kmp_internal_end_thread: abort, exiting\n"));
6380     /* TODO abort? */
6381     __kmp_release_bootstrap_lock(&__kmp_initz_lock);
6382     return;
6383   }
6384   if (TCR_4(__kmp_global.g.g_done) || !__kmp_init_serial) {
6385     __kmp_release_bootstrap_lock(&__kmp_initz_lock);
6386     return;
6387   }
6388 
6389   /* We need this lock to enforce mutex between this reading of
6390      __kmp_threads_capacity and the writing by __kmp_register_root.
6391      Alternatively, we can use a counter of roots that is atomically updated by
6392      __kmp_get_global_thread_id_reg, __kmp_do_serial_initialize and
6393      __kmp_internal_end_*.  */
6394 
6395   /* should we finish the run-time?  are all siblings done? */
6396   __kmp_acquire_bootstrap_lock(&__kmp_forkjoin_lock);
6397 
6398   for (i = 0; i < __kmp_threads_capacity; ++i) {
6399     if (KMP_UBER_GTID(i)) {
6400       KA_TRACE(
6401           10,
6402           ("__kmp_internal_end_thread: remaining sibling task: gtid==%d\n", i));
6403       __kmp_release_bootstrap_lock(&__kmp_forkjoin_lock);
6404       __kmp_release_bootstrap_lock(&__kmp_initz_lock);
6405       return;
6406     }
6407   }
6408 
6409   /* now we can safely conduct the actual termination */
6410 
6411   __kmp_internal_end();
6412 
6413   __kmp_release_bootstrap_lock(&__kmp_forkjoin_lock);
6414   __kmp_release_bootstrap_lock(&__kmp_initz_lock);
6415 
6416   KA_TRACE(10, ("__kmp_internal_end_thread: exit T#%d\n", gtid_req));
6417 
6418 #ifdef DUMP_DEBUG_ON_EXIT
6419   if (__kmp_debug_buf)
6420     __kmp_dump_debug_buffer();
6421 #endif
6422 } // __kmp_internal_end_thread
6423 
6424 // -----------------------------------------------------------------------------
6425 // Library registration stuff.
6426 
6427 static long __kmp_registration_flag = 0;
6428 // Random value used to indicate library initialization.
6429 static char *__kmp_registration_str = NULL;
6430 // Value to be saved in env var __KMP_REGISTERED_LIB_<pid>.
6431 
6432 static inline char *__kmp_reg_status_name() {
6433   /* On RHEL 3u5 if linked statically, getpid() returns different values in
6434      each thread. If registration and unregistration go in different threads
6435      (omp_misc_other_root_exit.cpp test case), the name of registered_lib_env
6436      env var can not be found, because the name will contain different pid. */
6437   return __kmp_str_format("__KMP_REGISTERED_LIB_%d", (int)getpid());
6438 } // __kmp_reg_status_get
6439 
6440 void __kmp_register_library_startup(void) {
6441 
6442   char *name = __kmp_reg_status_name(); // Name of the environment variable.
6443   int done = 0;
6444   union {
6445     double dtime;
6446     long ltime;
6447   } time;
6448 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
6449   __kmp_initialize_system_tick();
6450 #endif
6451   __kmp_read_system_time(&time.dtime);
6452   __kmp_registration_flag = 0xCAFE0000L | (time.ltime & 0x0000FFFFL);
6453   __kmp_registration_str =
6454       __kmp_str_format("%p-%lx-%s", &__kmp_registration_flag,
6455                        __kmp_registration_flag, KMP_LIBRARY_FILE);
6456 
6457   KA_TRACE(50, ("__kmp_register_library_startup: %s=\"%s\"\n", name,
6458                 __kmp_registration_str));
6459 
6460   while (!done) {
6461 
6462     char *value = NULL; // Actual value of the environment variable.
6463 
6464     // Set environment variable, but do not overwrite if it is exist.
6465     __kmp_env_set(name, __kmp_registration_str, 0);
6466     // Check the variable is written.
6467     value = __kmp_env_get(name);
6468     if (value != NULL && strcmp(value, __kmp_registration_str) == 0) {
6469 
6470       done = 1; // Ok, environment variable set successfully, exit the loop.
6471 
6472     } else {
6473 
6474       // Oops. Write failed. Another copy of OpenMP RTL is in memory.
6475       // Check whether it alive or dead.
6476       int neighbor = 0; // 0 -- unknown status, 1 -- alive, 2 -- dead.
6477       char *tail = value;
6478       char *flag_addr_str = NULL;
6479       char *flag_val_str = NULL;
6480       char const *file_name = NULL;
6481       __kmp_str_split(tail, '-', &flag_addr_str, &tail);
6482       __kmp_str_split(tail, '-', &flag_val_str, &tail);
6483       file_name = tail;
6484       if (tail != NULL) {
6485         long *flag_addr = 0;
6486         long flag_val = 0;
6487         KMP_SSCANF(flag_addr_str, "%p", RCAST(void**, &flag_addr));
6488         KMP_SSCANF(flag_val_str, "%lx", &flag_val);
6489         if (flag_addr != 0 && flag_val != 0 && strcmp(file_name, "") != 0) {
6490           // First, check whether environment-encoded address is mapped into
6491           // addr space.
6492           // If so, dereference it to see if it still has the right value.
6493           if (__kmp_is_address_mapped(flag_addr) && *flag_addr == flag_val) {
6494             neighbor = 1;
6495           } else {
6496             // If not, then we know the other copy of the library is no longer
6497             // running.
6498             neighbor = 2;
6499           }
6500         }
6501       }
6502       switch (neighbor) {
6503       case 0: // Cannot parse environment variable -- neighbor status unknown.
6504         // Assume it is the incompatible format of future version of the
6505         // library. Assume the other library is alive.
6506         // WARN( ... ); // TODO: Issue a warning.
6507         file_name = "unknown library";
6508         KMP_FALLTHROUGH();
6509       // Attention! Falling to the next case. That's intentional.
6510       case 1: { // Neighbor is alive.
6511         // Check it is allowed.
6512         char *duplicate_ok = __kmp_env_get("KMP_DUPLICATE_LIB_OK");
6513         if (!__kmp_str_match_true(duplicate_ok)) {
6514           // That's not allowed. Issue fatal error.
6515           __kmp_fatal(KMP_MSG(DuplicateLibrary, KMP_LIBRARY_FILE, file_name),
6516                       KMP_HNT(DuplicateLibrary), __kmp_msg_null);
6517         }
6518         KMP_INTERNAL_FREE(duplicate_ok);
6519         __kmp_duplicate_library_ok = 1;
6520         done = 1; // Exit the loop.
6521       } break;
6522       case 2: { // Neighbor is dead.
6523         // Clear the variable and try to register library again.
6524         __kmp_env_unset(name);
6525       } break;
6526       default: { KMP_DEBUG_ASSERT(0); } break;
6527       }
6528     }
6529     KMP_INTERNAL_FREE((void *)value);
6530   }
6531   KMP_INTERNAL_FREE((void *)name);
6532 
6533 } // func __kmp_register_library_startup
6534 
6535 void __kmp_unregister_library(void) {
6536 
6537   char *name = __kmp_reg_status_name();
6538   char *value = __kmp_env_get(name);
6539 
6540   KMP_DEBUG_ASSERT(__kmp_registration_flag != 0);
6541   KMP_DEBUG_ASSERT(__kmp_registration_str != NULL);
6542   if (value != NULL && strcmp(value, __kmp_registration_str) == 0) {
6543     // Ok, this is our variable. Delete it.
6544     __kmp_env_unset(name);
6545   }
6546 
6547   KMP_INTERNAL_FREE(__kmp_registration_str);
6548   KMP_INTERNAL_FREE(value);
6549   KMP_INTERNAL_FREE(name);
6550 
6551   __kmp_registration_flag = 0;
6552   __kmp_registration_str = NULL;
6553 
6554 } // __kmp_unregister_library
6555 
6556 // End of Library registration stuff.
6557 // -----------------------------------------------------------------------------
6558 
6559 #if KMP_MIC_SUPPORTED
6560 
6561 static void __kmp_check_mic_type() {
6562   kmp_cpuid_t cpuid_state = {0};
6563   kmp_cpuid_t *cs_p = &cpuid_state;
6564   __kmp_x86_cpuid(1, 0, cs_p);
6565   // We don't support mic1 at the moment
6566   if ((cs_p->eax & 0xff0) == 0xB10) {
6567     __kmp_mic_type = mic2;
6568   } else if ((cs_p->eax & 0xf0ff0) == 0x50670) {
6569     __kmp_mic_type = mic3;
6570   } else {
6571     __kmp_mic_type = non_mic;
6572   }
6573 }
6574 
6575 #endif /* KMP_MIC_SUPPORTED */
6576 
6577 static void __kmp_do_serial_initialize(void) {
6578   int i, gtid;
6579   int size;
6580 
6581   KA_TRACE(10, ("__kmp_do_serial_initialize: enter\n"));
6582 
6583   KMP_DEBUG_ASSERT(sizeof(kmp_int32) == 4);
6584   KMP_DEBUG_ASSERT(sizeof(kmp_uint32) == 4);
6585   KMP_DEBUG_ASSERT(sizeof(kmp_int64) == 8);
6586   KMP_DEBUG_ASSERT(sizeof(kmp_uint64) == 8);
6587   KMP_DEBUG_ASSERT(sizeof(kmp_intptr_t) == sizeof(void *));
6588 
6589 #if OMPT_SUPPORT
6590   ompt_pre_init();
6591 #endif
6592 
6593   __kmp_validate_locks();
6594 
6595   /* Initialize internal memory allocator */
6596   __kmp_init_allocator();
6597 
6598   /* Register the library startup via an environment variable and check to see
6599      whether another copy of the library is already registered. */
6600 
6601   __kmp_register_library_startup();
6602 
6603   /* TODO reinitialization of library */
6604   if (TCR_4(__kmp_global.g.g_done)) {
6605     KA_TRACE(10, ("__kmp_do_serial_initialize: reinitialization of library\n"));
6606   }
6607 
6608   __kmp_global.g.g_abort = 0;
6609   TCW_SYNC_4(__kmp_global.g.g_done, FALSE);
6610 
6611 /* initialize the locks */
6612 #if KMP_USE_ADAPTIVE_LOCKS
6613 #if KMP_DEBUG_ADAPTIVE_LOCKS
6614   __kmp_init_speculative_stats();
6615 #endif
6616 #endif
6617 #if KMP_STATS_ENABLED
6618   __kmp_stats_init();
6619 #endif
6620   __kmp_init_lock(&__kmp_global_lock);
6621   __kmp_init_queuing_lock(&__kmp_dispatch_lock);
6622   __kmp_init_lock(&__kmp_debug_lock);
6623   __kmp_init_atomic_lock(&__kmp_atomic_lock);
6624   __kmp_init_atomic_lock(&__kmp_atomic_lock_1i);
6625   __kmp_init_atomic_lock(&__kmp_atomic_lock_2i);
6626   __kmp_init_atomic_lock(&__kmp_atomic_lock_4i);
6627   __kmp_init_atomic_lock(&__kmp_atomic_lock_4r);
6628   __kmp_init_atomic_lock(&__kmp_atomic_lock_8i);
6629   __kmp_init_atomic_lock(&__kmp_atomic_lock_8r);
6630   __kmp_init_atomic_lock(&__kmp_atomic_lock_8c);
6631   __kmp_init_atomic_lock(&__kmp_atomic_lock_10r);
6632   __kmp_init_atomic_lock(&__kmp_atomic_lock_16r);
6633   __kmp_init_atomic_lock(&__kmp_atomic_lock_16c);
6634   __kmp_init_atomic_lock(&__kmp_atomic_lock_20c);
6635   __kmp_init_atomic_lock(&__kmp_atomic_lock_32c);
6636   __kmp_init_bootstrap_lock(&__kmp_forkjoin_lock);
6637   __kmp_init_bootstrap_lock(&__kmp_exit_lock);
6638 #if KMP_USE_MONITOR
6639   __kmp_init_bootstrap_lock(&__kmp_monitor_lock);
6640 #endif
6641   __kmp_init_bootstrap_lock(&__kmp_tp_cached_lock);
6642 
6643   /* conduct initialization and initial setup of configuration */
6644 
6645   __kmp_runtime_initialize();
6646 
6647 #if KMP_MIC_SUPPORTED
6648   __kmp_check_mic_type();
6649 #endif
6650 
6651 // Some global variable initialization moved here from kmp_env_initialize()
6652 #ifdef KMP_DEBUG
6653   kmp_diag = 0;
6654 #endif
6655   __kmp_abort_delay = 0;
6656 
6657   // From __kmp_init_dflt_team_nth()
6658   /* assume the entire machine will be used */
6659   __kmp_dflt_team_nth_ub = __kmp_xproc;
6660   if (__kmp_dflt_team_nth_ub < KMP_MIN_NTH) {
6661     __kmp_dflt_team_nth_ub = KMP_MIN_NTH;
6662   }
6663   if (__kmp_dflt_team_nth_ub > __kmp_sys_max_nth) {
6664     __kmp_dflt_team_nth_ub = __kmp_sys_max_nth;
6665   }
6666   __kmp_max_nth = __kmp_sys_max_nth;
6667   __kmp_cg_max_nth = __kmp_sys_max_nth;
6668   __kmp_teams_max_nth = __kmp_xproc; // set a "reasonable" default
6669   if (__kmp_teams_max_nth > __kmp_sys_max_nth) {
6670     __kmp_teams_max_nth = __kmp_sys_max_nth;
6671   }
6672 
6673   // Three vars below moved here from __kmp_env_initialize() "KMP_BLOCKTIME"
6674   // part
6675   __kmp_dflt_blocktime = KMP_DEFAULT_BLOCKTIME;
6676 #if KMP_USE_MONITOR
6677   __kmp_monitor_wakeups =
6678       KMP_WAKEUPS_FROM_BLOCKTIME(__kmp_dflt_blocktime, __kmp_monitor_wakeups);
6679   __kmp_bt_intervals =
6680       KMP_INTERVALS_FROM_BLOCKTIME(__kmp_dflt_blocktime, __kmp_monitor_wakeups);
6681 #endif
6682   // From "KMP_LIBRARY" part of __kmp_env_initialize()
6683   __kmp_library = library_throughput;
6684   // From KMP_SCHEDULE initialization
6685   __kmp_static = kmp_sch_static_balanced;
6686 // AC: do not use analytical here, because it is non-monotonous
6687 //__kmp_guided = kmp_sch_guided_iterative_chunked;
6688 //__kmp_auto = kmp_sch_guided_analytical_chunked; // AC: it is the default, no
6689 // need to repeat assignment
6690 // Barrier initialization. Moved here from __kmp_env_initialize() Barrier branch
6691 // bit control and barrier method control parts
6692 #if KMP_FAST_REDUCTION_BARRIER
6693 #define kmp_reduction_barrier_gather_bb ((int)1)
6694 #define kmp_reduction_barrier_release_bb ((int)1)
6695 #define kmp_reduction_barrier_gather_pat bp_hyper_bar
6696 #define kmp_reduction_barrier_release_pat bp_hyper_bar
6697 #endif // KMP_FAST_REDUCTION_BARRIER
6698   for (i = bs_plain_barrier; i < bs_last_barrier; i++) {
6699     __kmp_barrier_gather_branch_bits[i] = __kmp_barrier_gather_bb_dflt;
6700     __kmp_barrier_release_branch_bits[i] = __kmp_barrier_release_bb_dflt;
6701     __kmp_barrier_gather_pattern[i] = __kmp_barrier_gather_pat_dflt;
6702     __kmp_barrier_release_pattern[i] = __kmp_barrier_release_pat_dflt;
6703 #if KMP_FAST_REDUCTION_BARRIER
6704     if (i == bs_reduction_barrier) { // tested and confirmed on ALTIX only (
6705       // lin_64 ): hyper,1
6706       __kmp_barrier_gather_branch_bits[i] = kmp_reduction_barrier_gather_bb;
6707       __kmp_barrier_release_branch_bits[i] = kmp_reduction_barrier_release_bb;
6708       __kmp_barrier_gather_pattern[i] = kmp_reduction_barrier_gather_pat;
6709       __kmp_barrier_release_pattern[i] = kmp_reduction_barrier_release_pat;
6710     }
6711 #endif // KMP_FAST_REDUCTION_BARRIER
6712   }
6713 #if KMP_FAST_REDUCTION_BARRIER
6714 #undef kmp_reduction_barrier_release_pat
6715 #undef kmp_reduction_barrier_gather_pat
6716 #undef kmp_reduction_barrier_release_bb
6717 #undef kmp_reduction_barrier_gather_bb
6718 #endif // KMP_FAST_REDUCTION_BARRIER
6719 #if KMP_MIC_SUPPORTED
6720   if (__kmp_mic_type == mic2) { // KNC
6721     // AC: plane=3,2, forkjoin=2,1 are optimal for 240 threads on KNC
6722     __kmp_barrier_gather_branch_bits[bs_plain_barrier] = 3; // plain gather
6723     __kmp_barrier_release_branch_bits[bs_forkjoin_barrier] =
6724         1; // forkjoin release
6725     __kmp_barrier_gather_pattern[bs_forkjoin_barrier] = bp_hierarchical_bar;
6726     __kmp_barrier_release_pattern[bs_forkjoin_barrier] = bp_hierarchical_bar;
6727   }
6728 #if KMP_FAST_REDUCTION_BARRIER
6729   if (__kmp_mic_type == mic2) { // KNC
6730     __kmp_barrier_gather_pattern[bs_reduction_barrier] = bp_hierarchical_bar;
6731     __kmp_barrier_release_pattern[bs_reduction_barrier] = bp_hierarchical_bar;
6732   }
6733 #endif // KMP_FAST_REDUCTION_BARRIER
6734 #endif // KMP_MIC_SUPPORTED
6735 
6736 // From KMP_CHECKS initialization
6737 #ifdef KMP_DEBUG
6738   __kmp_env_checks = TRUE; /* development versions have the extra checks */
6739 #else
6740   __kmp_env_checks = FALSE; /* port versions do not have the extra checks */
6741 #endif
6742 
6743   // From "KMP_FOREIGN_THREADS_THREADPRIVATE" initialization
6744   __kmp_foreign_tp = TRUE;
6745 
6746   __kmp_global.g.g_dynamic = FALSE;
6747   __kmp_global.g.g_dynamic_mode = dynamic_default;
6748 
6749   __kmp_env_initialize(NULL);
6750 
6751 // Print all messages in message catalog for testing purposes.
6752 #ifdef KMP_DEBUG
6753   char const *val = __kmp_env_get("KMP_DUMP_CATALOG");
6754   if (__kmp_str_match_true(val)) {
6755     kmp_str_buf_t buffer;
6756     __kmp_str_buf_init(&buffer);
6757     __kmp_i18n_dump_catalog(&buffer);
6758     __kmp_printf("%s", buffer.str);
6759     __kmp_str_buf_free(&buffer);
6760   }
6761   __kmp_env_free(&val);
6762 #endif
6763 
6764   __kmp_threads_capacity =
6765       __kmp_initial_threads_capacity(__kmp_dflt_team_nth_ub);
6766   // Moved here from __kmp_env_initialize() "KMP_ALL_THREADPRIVATE" part
6767   __kmp_tp_capacity = __kmp_default_tp_capacity(
6768       __kmp_dflt_team_nth_ub, __kmp_max_nth, __kmp_allThreadsSpecified);
6769 
6770   // If the library is shut down properly, both pools must be NULL. Just in
6771   // case, set them to NULL -- some memory may leak, but subsequent code will
6772   // work even if pools are not freed.
6773   KMP_DEBUG_ASSERT(__kmp_thread_pool == NULL);
6774   KMP_DEBUG_ASSERT(__kmp_thread_pool_insert_pt == NULL);
6775   KMP_DEBUG_ASSERT(__kmp_team_pool == NULL);
6776   __kmp_thread_pool = NULL;
6777   __kmp_thread_pool_insert_pt = NULL;
6778   __kmp_team_pool = NULL;
6779 
6780   /* Allocate all of the variable sized records */
6781   /* NOTE: __kmp_threads_capacity entries are allocated, but the arrays are
6782    * expandable */
6783   /* Since allocation is cache-aligned, just add extra padding at the end */
6784   size =
6785       (sizeof(kmp_info_t *) + sizeof(kmp_root_t *)) * __kmp_threads_capacity +
6786       CACHE_LINE;
6787   __kmp_threads = (kmp_info_t **)__kmp_allocate(size);
6788   __kmp_root = (kmp_root_t **)((char *)__kmp_threads +
6789                                sizeof(kmp_info_t *) * __kmp_threads_capacity);
6790 
6791   /* init thread counts */
6792   KMP_DEBUG_ASSERT(__kmp_all_nth ==
6793                    0); // Asserts fail if the library is reinitializing and
6794   KMP_DEBUG_ASSERT(__kmp_nth == 0); // something was wrong in termination.
6795   __kmp_all_nth = 0;
6796   __kmp_nth = 0;
6797 
6798   /* setup the uber master thread and hierarchy */
6799   gtid = __kmp_register_root(TRUE);
6800   KA_TRACE(10, ("__kmp_do_serial_initialize  T#%d\n", gtid));
6801   KMP_ASSERT(KMP_UBER_GTID(gtid));
6802   KMP_ASSERT(KMP_INITIAL_GTID(gtid));
6803 
6804   KMP_MB(); /* Flush all pending memory write invalidates.  */
6805 
6806   __kmp_common_initialize();
6807 
6808 #if KMP_OS_UNIX
6809   /* invoke the child fork handler */
6810   __kmp_register_atfork();
6811 #endif
6812 
6813 #if !KMP_DYNAMIC_LIB
6814   {
6815     /* Invoke the exit handler when the program finishes, only for static
6816        library. For dynamic library, we already have _fini and DllMain. */
6817     int rc = atexit(__kmp_internal_end_atexit);
6818     if (rc != 0) {
6819       __kmp_fatal(KMP_MSG(FunctionError, "atexit()"), KMP_ERR(rc),
6820                   __kmp_msg_null);
6821     }
6822   }
6823 #endif
6824 
6825 #if KMP_HANDLE_SIGNALS
6826 #if KMP_OS_UNIX
6827   /* NOTE: make sure that this is called before the user installs their own
6828      signal handlers so that the user handlers are called first. this way they
6829      can return false, not call our handler, avoid terminating the library, and
6830      continue execution where they left off. */
6831   __kmp_install_signals(FALSE);
6832 #endif /* KMP_OS_UNIX */
6833 #if KMP_OS_WINDOWS
6834   __kmp_install_signals(TRUE);
6835 #endif /* KMP_OS_WINDOWS */
6836 #endif
6837 
6838   /* we have finished the serial initialization */
6839   __kmp_init_counter++;
6840 
6841   __kmp_init_serial = TRUE;
6842 
6843   if (__kmp_settings) {
6844     __kmp_env_print();
6845   }
6846 
6847 #if OMP_40_ENABLED
6848   if (__kmp_display_env || __kmp_display_env_verbose) {
6849     __kmp_env_print_2();
6850   }
6851 #endif // OMP_40_ENABLED
6852 
6853 #if OMPT_SUPPORT
6854   ompt_post_init();
6855 #endif
6856 
6857   KMP_MB();
6858 
6859   KA_TRACE(10, ("__kmp_do_serial_initialize: exit\n"));
6860 }
6861 
6862 void __kmp_serial_initialize(void) {
6863   if (__kmp_init_serial) {
6864     return;
6865   }
6866   __kmp_acquire_bootstrap_lock(&__kmp_initz_lock);
6867   if (__kmp_init_serial) {
6868     __kmp_release_bootstrap_lock(&__kmp_initz_lock);
6869     return;
6870   }
6871   __kmp_do_serial_initialize();
6872   __kmp_release_bootstrap_lock(&__kmp_initz_lock);
6873 }
6874 
6875 static void __kmp_do_middle_initialize(void) {
6876   int i, j;
6877   int prev_dflt_team_nth;
6878 
6879   if (!__kmp_init_serial) {
6880     __kmp_do_serial_initialize();
6881   }
6882 
6883   KA_TRACE(10, ("__kmp_middle_initialize: enter\n"));
6884 
6885   // Save the previous value for the __kmp_dflt_team_nth so that
6886   // we can avoid some reinitialization if it hasn't changed.
6887   prev_dflt_team_nth = __kmp_dflt_team_nth;
6888 
6889 #if KMP_AFFINITY_SUPPORTED
6890   // __kmp_affinity_initialize() will try to set __kmp_ncores to the
6891   // number of cores on the machine.
6892   __kmp_affinity_initialize();
6893 
6894   // Run through the __kmp_threads array and set the affinity mask
6895   // for each root thread that is currently registered with the RTL.
6896   for (i = 0; i < __kmp_threads_capacity; i++) {
6897     if (TCR_PTR(__kmp_threads[i]) != NULL) {
6898       __kmp_affinity_set_init_mask(i, TRUE);
6899     }
6900   }
6901 #endif /* KMP_AFFINITY_SUPPORTED */
6902 
6903   KMP_ASSERT(__kmp_xproc > 0);
6904   if (__kmp_avail_proc == 0) {
6905     __kmp_avail_proc = __kmp_xproc;
6906   }
6907 
6908   // If there were empty places in num_threads list (OMP_NUM_THREADS=,,2,3),
6909   // correct them now
6910   j = 0;
6911   while ((j < __kmp_nested_nth.used) && !__kmp_nested_nth.nth[j]) {
6912     __kmp_nested_nth.nth[j] = __kmp_dflt_team_nth = __kmp_dflt_team_nth_ub =
6913         __kmp_avail_proc;
6914     j++;
6915   }
6916 
6917   if (__kmp_dflt_team_nth == 0) {
6918 #ifdef KMP_DFLT_NTH_CORES
6919     // Default #threads = #cores
6920     __kmp_dflt_team_nth = __kmp_ncores;
6921     KA_TRACE(20, ("__kmp_middle_initialize: setting __kmp_dflt_team_nth = "
6922                   "__kmp_ncores (%d)\n",
6923                   __kmp_dflt_team_nth));
6924 #else
6925     // Default #threads = #available OS procs
6926     __kmp_dflt_team_nth = __kmp_avail_proc;
6927     KA_TRACE(20, ("__kmp_middle_initialize: setting __kmp_dflt_team_nth = "
6928                   "__kmp_avail_proc(%d)\n",
6929                   __kmp_dflt_team_nth));
6930 #endif /* KMP_DFLT_NTH_CORES */
6931   }
6932 
6933   if (__kmp_dflt_team_nth < KMP_MIN_NTH) {
6934     __kmp_dflt_team_nth = KMP_MIN_NTH;
6935   }
6936   if (__kmp_dflt_team_nth > __kmp_sys_max_nth) {
6937     __kmp_dflt_team_nth = __kmp_sys_max_nth;
6938   }
6939 
6940   // There's no harm in continuing if the following check fails,
6941   // but it indicates an error in the previous logic.
6942   KMP_DEBUG_ASSERT(__kmp_dflt_team_nth <= __kmp_dflt_team_nth_ub);
6943 
6944   if (__kmp_dflt_team_nth != prev_dflt_team_nth) {
6945     // Run through the __kmp_threads array and set the num threads icv for each
6946     // root thread that is currently registered with the RTL (which has not
6947     // already explicitly set its nthreads-var with a call to
6948     // omp_set_num_threads()).
6949     for (i = 0; i < __kmp_threads_capacity; i++) {
6950       kmp_info_t *thread = __kmp_threads[i];
6951       if (thread == NULL)
6952         continue;
6953       if (thread->th.th_current_task->td_icvs.nproc != 0)
6954         continue;
6955 
6956       set__nproc(__kmp_threads[i], __kmp_dflt_team_nth);
6957     }
6958   }
6959   KA_TRACE(
6960       20,
6961       ("__kmp_middle_initialize: final value for __kmp_dflt_team_nth = %d\n",
6962        __kmp_dflt_team_nth));
6963 
6964 #ifdef KMP_ADJUST_BLOCKTIME
6965   /* Adjust blocktime to zero if necessary  now that __kmp_avail_proc is set */
6966   if (!__kmp_env_blocktime && (__kmp_avail_proc > 0)) {
6967     KMP_DEBUG_ASSERT(__kmp_avail_proc > 0);
6968     if (__kmp_nth > __kmp_avail_proc) {
6969       __kmp_zero_bt = TRUE;
6970     }
6971   }
6972 #endif /* KMP_ADJUST_BLOCKTIME */
6973 
6974   /* we have finished middle initialization */
6975   TCW_SYNC_4(__kmp_init_middle, TRUE);
6976 
6977   KA_TRACE(10, ("__kmp_do_middle_initialize: exit\n"));
6978 }
6979 
6980 void __kmp_middle_initialize(void) {
6981   if (__kmp_init_middle) {
6982     return;
6983   }
6984   __kmp_acquire_bootstrap_lock(&__kmp_initz_lock);
6985   if (__kmp_init_middle) {
6986     __kmp_release_bootstrap_lock(&__kmp_initz_lock);
6987     return;
6988   }
6989   __kmp_do_middle_initialize();
6990   __kmp_release_bootstrap_lock(&__kmp_initz_lock);
6991 }
6992 
6993 void __kmp_parallel_initialize(void) {
6994   int gtid = __kmp_entry_gtid(); // this might be a new root
6995 
6996   /* synchronize parallel initialization (for sibling) */
6997   if (TCR_4(__kmp_init_parallel))
6998     return;
6999   __kmp_acquire_bootstrap_lock(&__kmp_initz_lock);
7000   if (TCR_4(__kmp_init_parallel)) {
7001     __kmp_release_bootstrap_lock(&__kmp_initz_lock);
7002     return;
7003   }
7004 
7005   /* TODO reinitialization after we have already shut down */
7006   if (TCR_4(__kmp_global.g.g_done)) {
7007     KA_TRACE(
7008         10,
7009         ("__kmp_parallel_initialize: attempt to init while shutting down\n"));
7010     __kmp_infinite_loop();
7011   }
7012 
7013   /* jc: The lock __kmp_initz_lock is already held, so calling
7014      __kmp_serial_initialize would cause a deadlock.  So we call
7015      __kmp_do_serial_initialize directly. */
7016   if (!__kmp_init_middle) {
7017     __kmp_do_middle_initialize();
7018   }
7019 
7020 #if OMP_50_ENABLED
7021   __kmp_resume_if_hard_paused();
7022 #endif
7023 
7024   /* begin initialization */
7025   KA_TRACE(10, ("__kmp_parallel_initialize: enter\n"));
7026   KMP_ASSERT(KMP_UBER_GTID(gtid));
7027 
7028 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
7029   // Save the FP control regs.
7030   // Worker threads will set theirs to these values at thread startup.
7031   __kmp_store_x87_fpu_control_word(&__kmp_init_x87_fpu_control_word);
7032   __kmp_store_mxcsr(&__kmp_init_mxcsr);
7033   __kmp_init_mxcsr &= KMP_X86_MXCSR_MASK;
7034 #endif /* KMP_ARCH_X86 || KMP_ARCH_X86_64 */
7035 
7036 #if KMP_OS_UNIX
7037 #if KMP_HANDLE_SIGNALS
7038   /*  must be after __kmp_serial_initialize  */
7039   __kmp_install_signals(TRUE);
7040 #endif
7041 #endif
7042 
7043   __kmp_suspend_initialize();
7044 
7045 #if defined(USE_LOAD_BALANCE)
7046   if (__kmp_global.g.g_dynamic_mode == dynamic_default) {
7047     __kmp_global.g.g_dynamic_mode = dynamic_load_balance;
7048   }
7049 #else
7050   if (__kmp_global.g.g_dynamic_mode == dynamic_default) {
7051     __kmp_global.g.g_dynamic_mode = dynamic_thread_limit;
7052   }
7053 #endif
7054 
7055   if (__kmp_version) {
7056     __kmp_print_version_2();
7057   }
7058 
7059   /* we have finished parallel initialization */
7060   TCW_SYNC_4(__kmp_init_parallel, TRUE);
7061 
7062   KMP_MB();
7063   KA_TRACE(10, ("__kmp_parallel_initialize: exit\n"));
7064 
7065   __kmp_release_bootstrap_lock(&__kmp_initz_lock);
7066 }
7067 
7068 /* ------------------------------------------------------------------------ */
7069 
7070 void __kmp_run_before_invoked_task(int gtid, int tid, kmp_info_t *this_thr,
7071                                    kmp_team_t *team) {
7072   kmp_disp_t *dispatch;
7073 
7074   KMP_MB();
7075 
7076   /* none of the threads have encountered any constructs, yet. */
7077   this_thr->th.th_local.this_construct = 0;
7078 #if KMP_CACHE_MANAGE
7079   KMP_CACHE_PREFETCH(&this_thr->th.th_bar[bs_forkjoin_barrier].bb.b_arrived);
7080 #endif /* KMP_CACHE_MANAGE */
7081   dispatch = (kmp_disp_t *)TCR_PTR(this_thr->th.th_dispatch);
7082   KMP_DEBUG_ASSERT(dispatch);
7083   KMP_DEBUG_ASSERT(team->t.t_dispatch);
7084   // KMP_DEBUG_ASSERT( this_thr->th.th_dispatch == &team->t.t_dispatch[
7085   // this_thr->th.th_info.ds.ds_tid ] );
7086 
7087   dispatch->th_disp_index = 0; /* reset the dispatch buffer counter */
7088 #if OMP_45_ENABLED
7089   dispatch->th_doacross_buf_idx =
7090       0; /* reset the doacross dispatch buffer counter */
7091 #endif
7092   if (__kmp_env_consistency_check)
7093     __kmp_push_parallel(gtid, team->t.t_ident);
7094 
7095   KMP_MB(); /* Flush all pending memory write invalidates.  */
7096 }
7097 
7098 void __kmp_run_after_invoked_task(int gtid, int tid, kmp_info_t *this_thr,
7099                                   kmp_team_t *team) {
7100   if (__kmp_env_consistency_check)
7101     __kmp_pop_parallel(gtid, team->t.t_ident);
7102 
7103   __kmp_finish_implicit_task(this_thr);
7104 }
7105 
7106 int __kmp_invoke_task_func(int gtid) {
7107   int rc;
7108   int tid = __kmp_tid_from_gtid(gtid);
7109   kmp_info_t *this_thr = __kmp_threads[gtid];
7110   kmp_team_t *team = this_thr->th.th_team;
7111 
7112   __kmp_run_before_invoked_task(gtid, tid, this_thr, team);
7113 #if USE_ITT_BUILD
7114   if (__itt_stack_caller_create_ptr) {
7115     __kmp_itt_stack_callee_enter(
7116         (__itt_caller)
7117             team->t.t_stack_id); // inform ittnotify about entering user's code
7118   }
7119 #endif /* USE_ITT_BUILD */
7120 #if INCLUDE_SSC_MARKS
7121   SSC_MARK_INVOKING();
7122 #endif
7123 
7124 #if OMPT_SUPPORT
7125   void *dummy;
7126   void **exit_runtime_p;
7127   ompt_data_t *my_task_data;
7128   ompt_data_t *my_parallel_data;
7129   int ompt_team_size;
7130 
7131   if (ompt_enabled.enabled) {
7132     exit_runtime_p = &(
7133         team->t.t_implicit_task_taskdata[tid].ompt_task_info.frame.exit_frame.ptr);
7134   } else {
7135     exit_runtime_p = &dummy;
7136   }
7137 
7138   my_task_data =
7139       &(team->t.t_implicit_task_taskdata[tid].ompt_task_info.task_data);
7140   my_parallel_data = &(team->t.ompt_team_info.parallel_data);
7141   if (ompt_enabled.ompt_callback_implicit_task) {
7142     ompt_team_size = team->t.t_nproc;
7143     ompt_callbacks.ompt_callback(ompt_callback_implicit_task)(
7144         ompt_scope_begin, my_parallel_data, my_task_data, ompt_team_size,
7145         __kmp_tid_from_gtid(gtid), ompt_task_implicit); // TODO: Can this be ompt_task_initial?
7146     OMPT_CUR_TASK_INFO(this_thr)->thread_num = __kmp_tid_from_gtid(gtid);
7147   }
7148 #endif
7149 
7150 #if KMP_STATS_ENABLED
7151   stats_state_e previous_state = KMP_GET_THREAD_STATE();
7152   if (previous_state == stats_state_e::TEAMS_REGION) {
7153     KMP_PUSH_PARTITIONED_TIMER(OMP_teams);
7154   } else {
7155     KMP_PUSH_PARTITIONED_TIMER(OMP_parallel);
7156   }
7157   KMP_SET_THREAD_STATE(IMPLICIT_TASK);
7158 #endif
7159 
7160   rc = __kmp_invoke_microtask((microtask_t)TCR_SYNC_PTR(team->t.t_pkfn), gtid,
7161                               tid, (int)team->t.t_argc, (void **)team->t.t_argv
7162 #if OMPT_SUPPORT
7163                               ,
7164                               exit_runtime_p
7165 #endif
7166                               );
7167 #if OMPT_SUPPORT
7168   *exit_runtime_p = NULL;
7169 #endif
7170 
7171 #if KMP_STATS_ENABLED
7172   if (previous_state == stats_state_e::TEAMS_REGION) {
7173     KMP_SET_THREAD_STATE(previous_state);
7174   }
7175   KMP_POP_PARTITIONED_TIMER();
7176 #endif
7177 
7178 #if USE_ITT_BUILD
7179   if (__itt_stack_caller_create_ptr) {
7180     __kmp_itt_stack_callee_leave(
7181         (__itt_caller)
7182             team->t.t_stack_id); // inform ittnotify about leaving user's code
7183   }
7184 #endif /* USE_ITT_BUILD */
7185   __kmp_run_after_invoked_task(gtid, tid, this_thr, team);
7186 
7187   return rc;
7188 }
7189 
7190 #if OMP_40_ENABLED
7191 void __kmp_teams_master(int gtid) {
7192   // This routine is called by all master threads in teams construct
7193   kmp_info_t *thr = __kmp_threads[gtid];
7194   kmp_team_t *team = thr->th.th_team;
7195   ident_t *loc = team->t.t_ident;
7196   thr->th.th_set_nproc = thr->th.th_teams_size.nth;
7197   KMP_DEBUG_ASSERT(thr->th.th_teams_microtask);
7198   KMP_DEBUG_ASSERT(thr->th.th_set_nproc);
7199   KA_TRACE(20, ("__kmp_teams_master: T#%d, Tid %d, microtask %p\n", gtid,
7200                 __kmp_tid_from_gtid(gtid), thr->th.th_teams_microtask));
7201 
7202   // This thread is a new CG root.  Set up the proper variables.
7203   kmp_cg_root_t *tmp = (kmp_cg_root_t *)__kmp_allocate(sizeof(kmp_cg_root_t));
7204   tmp->cg_root = thr; // Make thr the CG root
7205   // Init to thread limit that was stored when league masters were forked
7206   tmp->cg_thread_limit = thr->th.th_current_task->td_icvs.thread_limit;
7207   tmp->cg_nthreads = 1; // Init counter to one active thread, this one
7208   KA_TRACE(100, ("__kmp_teams_master: Thread %p created node %p and init"
7209                  " cg_threads to 1\n",
7210                  thr, tmp));
7211   tmp->up = thr->th.th_cg_roots;
7212   thr->th.th_cg_roots = tmp;
7213 
7214 // Launch league of teams now, but not let workers execute
7215 // (they hang on fork barrier until next parallel)
7216 #if INCLUDE_SSC_MARKS
7217   SSC_MARK_FORKING();
7218 #endif
7219   __kmp_fork_call(loc, gtid, fork_context_intel, team->t.t_argc,
7220                   (microtask_t)thr->th.th_teams_microtask, // "wrapped" task
7221                   VOLATILE_CAST(launch_t) __kmp_invoke_task_func, NULL);
7222 #if INCLUDE_SSC_MARKS
7223   SSC_MARK_JOINING();
7224 #endif
7225   // If the team size was reduced from the limit, set it to the new size
7226   if (thr->th.th_team_nproc < thr->th.th_teams_size.nth)
7227     thr->th.th_teams_size.nth = thr->th.th_team_nproc;
7228   // AC: last parameter "1" eliminates join barrier which won't work because
7229   // worker threads are in a fork barrier waiting for more parallel regions
7230   __kmp_join_call(loc, gtid
7231 #if OMPT_SUPPORT
7232                   ,
7233                   fork_context_intel
7234 #endif
7235                   ,
7236                   1);
7237 }
7238 
7239 int __kmp_invoke_teams_master(int gtid) {
7240   kmp_info_t *this_thr = __kmp_threads[gtid];
7241   kmp_team_t *team = this_thr->th.th_team;
7242 #if KMP_DEBUG
7243   if (!__kmp_threads[gtid]->th.th_team->t.t_serialized)
7244     KMP_DEBUG_ASSERT((void *)__kmp_threads[gtid]->th.th_team->t.t_pkfn ==
7245                      (void *)__kmp_teams_master);
7246 #endif
7247   __kmp_run_before_invoked_task(gtid, 0, this_thr, team);
7248   __kmp_teams_master(gtid);
7249   __kmp_run_after_invoked_task(gtid, 0, this_thr, team);
7250   return 1;
7251 }
7252 #endif /* OMP_40_ENABLED */
7253 
7254 /* this sets the requested number of threads for the next parallel region
7255    encountered by this team. since this should be enclosed in the forkjoin
7256    critical section it should avoid race conditions with assymmetrical nested
7257    parallelism */
7258 
7259 void __kmp_push_num_threads(ident_t *id, int gtid, int num_threads) {
7260   kmp_info_t *thr = __kmp_threads[gtid];
7261 
7262   if (num_threads > 0)
7263     thr->th.th_set_nproc = num_threads;
7264 }
7265 
7266 #if OMP_40_ENABLED
7267 
7268 /* this sets the requested number of teams for the teams region and/or
7269    the number of threads for the next parallel region encountered  */
7270 void __kmp_push_num_teams(ident_t *id, int gtid, int num_teams,
7271                           int num_threads) {
7272   kmp_info_t *thr = __kmp_threads[gtid];
7273   KMP_DEBUG_ASSERT(num_teams >= 0);
7274   KMP_DEBUG_ASSERT(num_threads >= 0);
7275 
7276   if (num_teams == 0)
7277     num_teams = 1; // default number of teams is 1.
7278   if (num_teams > __kmp_teams_max_nth) { // if too many teams requested?
7279     if (!__kmp_reserve_warn) {
7280       __kmp_reserve_warn = 1;
7281       __kmp_msg(kmp_ms_warning,
7282                 KMP_MSG(CantFormThrTeam, num_teams, __kmp_teams_max_nth),
7283                 KMP_HNT(Unset_ALL_THREADS), __kmp_msg_null);
7284     }
7285     num_teams = __kmp_teams_max_nth;
7286   }
7287   // Set number of teams (number of threads in the outer "parallel" of the
7288   // teams)
7289   thr->th.th_set_nproc = thr->th.th_teams_size.nteams = num_teams;
7290 
7291   // Remember the number of threads for inner parallel regions
7292   if (num_threads == 0) {
7293     if (!TCR_4(__kmp_init_middle))
7294       __kmp_middle_initialize(); // get __kmp_avail_proc calculated
7295     num_threads = __kmp_avail_proc / num_teams;
7296     if (num_teams * num_threads > __kmp_teams_max_nth) {
7297       // adjust num_threads w/o warning as it is not user setting
7298       num_threads = __kmp_teams_max_nth / num_teams;
7299     }
7300   } else {
7301     // This thread will be the master of the league masters
7302     // Store new thread limit; old limit is saved in th_cg_roots list
7303     thr->th.th_current_task->td_icvs.thread_limit = num_threads;
7304 
7305     if (num_teams * num_threads > __kmp_teams_max_nth) {
7306       int new_threads = __kmp_teams_max_nth / num_teams;
7307       if (!__kmp_reserve_warn) { // user asked for too many threads
7308         __kmp_reserve_warn = 1; // conflicts with KMP_TEAMS_THREAD_LIMIT
7309         __kmp_msg(kmp_ms_warning,
7310                   KMP_MSG(CantFormThrTeam, num_threads, new_threads),
7311                   KMP_HNT(Unset_ALL_THREADS), __kmp_msg_null);
7312       }
7313       num_threads = new_threads;
7314     }
7315   }
7316   thr->th.th_teams_size.nth = num_threads;
7317 }
7318 
7319 // Set the proc_bind var to use in the following parallel region.
7320 void __kmp_push_proc_bind(ident_t *id, int gtid, kmp_proc_bind_t proc_bind) {
7321   kmp_info_t *thr = __kmp_threads[gtid];
7322   thr->th.th_set_proc_bind = proc_bind;
7323 }
7324 
7325 #endif /* OMP_40_ENABLED */
7326 
7327 /* Launch the worker threads into the microtask. */
7328 
7329 void __kmp_internal_fork(ident_t *id, int gtid, kmp_team_t *team) {
7330   kmp_info_t *this_thr = __kmp_threads[gtid];
7331 
7332 #ifdef KMP_DEBUG
7333   int f;
7334 #endif /* KMP_DEBUG */
7335 
7336   KMP_DEBUG_ASSERT(team);
7337   KMP_DEBUG_ASSERT(this_thr->th.th_team == team);
7338   KMP_ASSERT(KMP_MASTER_GTID(gtid));
7339   KMP_MB(); /* Flush all pending memory write invalidates.  */
7340 
7341   team->t.t_construct = 0; /* no single directives seen yet */
7342   team->t.t_ordered.dt.t_value =
7343       0; /* thread 0 enters the ordered section first */
7344 
7345   /* Reset the identifiers on the dispatch buffer */
7346   KMP_DEBUG_ASSERT(team->t.t_disp_buffer);
7347   if (team->t.t_max_nproc > 1) {
7348     int i;
7349     for (i = 0; i < __kmp_dispatch_num_buffers; ++i) {
7350       team->t.t_disp_buffer[i].buffer_index = i;
7351 #if OMP_45_ENABLED
7352       team->t.t_disp_buffer[i].doacross_buf_idx = i;
7353 #endif
7354     }
7355   } else {
7356     team->t.t_disp_buffer[0].buffer_index = 0;
7357 #if OMP_45_ENABLED
7358     team->t.t_disp_buffer[0].doacross_buf_idx = 0;
7359 #endif
7360   }
7361 
7362   KMP_MB(); /* Flush all pending memory write invalidates.  */
7363   KMP_ASSERT(this_thr->th.th_team == team);
7364 
7365 #ifdef KMP_DEBUG
7366   for (f = 0; f < team->t.t_nproc; f++) {
7367     KMP_DEBUG_ASSERT(team->t.t_threads[f] &&
7368                      team->t.t_threads[f]->th.th_team_nproc == team->t.t_nproc);
7369   }
7370 #endif /* KMP_DEBUG */
7371 
7372   /* release the worker threads so they may begin working */
7373   __kmp_fork_barrier(gtid, 0);
7374 }
7375 
7376 void __kmp_internal_join(ident_t *id, int gtid, kmp_team_t *team) {
7377   kmp_info_t *this_thr = __kmp_threads[gtid];
7378 
7379   KMP_DEBUG_ASSERT(team);
7380   KMP_DEBUG_ASSERT(this_thr->th.th_team == team);
7381   KMP_ASSERT(KMP_MASTER_GTID(gtid));
7382   KMP_MB(); /* Flush all pending memory write invalidates.  */
7383 
7384 /* Join barrier after fork */
7385 
7386 #ifdef KMP_DEBUG
7387   if (__kmp_threads[gtid] &&
7388       __kmp_threads[gtid]->th.th_team_nproc != team->t.t_nproc) {
7389     __kmp_printf("GTID: %d, __kmp_threads[%d]=%p\n", gtid, gtid,
7390                  __kmp_threads[gtid]);
7391     __kmp_printf("__kmp_threads[%d]->th.th_team_nproc=%d, TEAM: %p, "
7392                  "team->t.t_nproc=%d\n",
7393                  gtid, __kmp_threads[gtid]->th.th_team_nproc, team,
7394                  team->t.t_nproc);
7395     __kmp_print_structure();
7396   }
7397   KMP_DEBUG_ASSERT(__kmp_threads[gtid] &&
7398                    __kmp_threads[gtid]->th.th_team_nproc == team->t.t_nproc);
7399 #endif /* KMP_DEBUG */
7400 
7401   __kmp_join_barrier(gtid); /* wait for everyone */
7402 #if OMPT_SUPPORT
7403   if (ompt_enabled.enabled &&
7404       this_thr->th.ompt_thread_info.state == ompt_state_wait_barrier_implicit) {
7405     int ds_tid = this_thr->th.th_info.ds.ds_tid;
7406     ompt_data_t *task_data = OMPT_CUR_TASK_DATA(this_thr);
7407     this_thr->th.ompt_thread_info.state = ompt_state_overhead;
7408 #if OMPT_OPTIONAL
7409     void *codeptr = NULL;
7410     if (KMP_MASTER_TID(ds_tid) &&
7411         (ompt_callbacks.ompt_callback(ompt_callback_sync_region_wait) ||
7412          ompt_callbacks.ompt_callback(ompt_callback_sync_region)))
7413       codeptr = OMPT_CUR_TEAM_INFO(this_thr)->master_return_address;
7414 
7415     if (ompt_enabled.ompt_callback_sync_region_wait) {
7416       ompt_callbacks.ompt_callback(ompt_callback_sync_region_wait)(
7417           ompt_sync_region_barrier_implicit, ompt_scope_end, NULL, task_data,
7418           codeptr);
7419     }
7420     if (ompt_enabled.ompt_callback_sync_region) {
7421       ompt_callbacks.ompt_callback(ompt_callback_sync_region)(
7422           ompt_sync_region_barrier_implicit, ompt_scope_end, NULL, task_data,
7423           codeptr);
7424     }
7425 #endif
7426     if (!KMP_MASTER_TID(ds_tid) && ompt_enabled.ompt_callback_implicit_task) {
7427       ompt_callbacks.ompt_callback(ompt_callback_implicit_task)(
7428           ompt_scope_end, NULL, task_data, 0, ds_tid, ompt_task_implicit); // TODO: Can this be ompt_task_initial?
7429     }
7430   }
7431 #endif
7432 
7433   KMP_MB(); /* Flush all pending memory write invalidates.  */
7434   KMP_ASSERT(this_thr->th.th_team == team);
7435 }
7436 
7437 /* ------------------------------------------------------------------------ */
7438 
7439 #ifdef USE_LOAD_BALANCE
7440 
7441 // Return the worker threads actively spinning in the hot team, if we
7442 // are at the outermost level of parallelism.  Otherwise, return 0.
7443 static int __kmp_active_hot_team_nproc(kmp_root_t *root) {
7444   int i;
7445   int retval;
7446   kmp_team_t *hot_team;
7447 
7448   if (root->r.r_active) {
7449     return 0;
7450   }
7451   hot_team = root->r.r_hot_team;
7452   if (__kmp_dflt_blocktime == KMP_MAX_BLOCKTIME) {
7453     return hot_team->t.t_nproc - 1; // Don't count master thread
7454   }
7455 
7456   // Skip the master thread - it is accounted for elsewhere.
7457   retval = 0;
7458   for (i = 1; i < hot_team->t.t_nproc; i++) {
7459     if (hot_team->t.t_threads[i]->th.th_active) {
7460       retval++;
7461     }
7462   }
7463   return retval;
7464 }
7465 
7466 // Perform an automatic adjustment to the number of
7467 // threads used by the next parallel region.
7468 static int __kmp_load_balance_nproc(kmp_root_t *root, int set_nproc) {
7469   int retval;
7470   int pool_active;
7471   int hot_team_active;
7472   int team_curr_active;
7473   int system_active;
7474 
7475   KB_TRACE(20, ("__kmp_load_balance_nproc: called root:%p set_nproc:%d\n", root,
7476                 set_nproc));
7477   KMP_DEBUG_ASSERT(root);
7478   KMP_DEBUG_ASSERT(root->r.r_root_team->t.t_threads[0]
7479                        ->th.th_current_task->td_icvs.dynamic == TRUE);
7480   KMP_DEBUG_ASSERT(set_nproc > 1);
7481 
7482   if (set_nproc == 1) {
7483     KB_TRACE(20, ("__kmp_load_balance_nproc: serial execution.\n"));
7484     return 1;
7485   }
7486 
7487   // Threads that are active in the thread pool, active in the hot team for this
7488   // particular root (if we are at the outer par level), and the currently
7489   // executing thread (to become the master) are available to add to the new
7490   // team, but are currently contributing to the system load, and must be
7491   // accounted for.
7492   pool_active = __kmp_thread_pool_active_nth;
7493   hot_team_active = __kmp_active_hot_team_nproc(root);
7494   team_curr_active = pool_active + hot_team_active + 1;
7495 
7496   // Check the system load.
7497   system_active = __kmp_get_load_balance(__kmp_avail_proc + team_curr_active);
7498   KB_TRACE(30, ("__kmp_load_balance_nproc: system active = %d pool active = %d "
7499                 "hot team active = %d\n",
7500                 system_active, pool_active, hot_team_active));
7501 
7502   if (system_active < 0) {
7503     // There was an error reading the necessary info from /proc, so use the
7504     // thread limit algorithm instead. Once we set __kmp_global.g.g_dynamic_mode
7505     // = dynamic_thread_limit, we shouldn't wind up getting back here.
7506     __kmp_global.g.g_dynamic_mode = dynamic_thread_limit;
7507     KMP_WARNING(CantLoadBalUsing, "KMP_DYNAMIC_MODE=thread limit");
7508 
7509     // Make this call behave like the thread limit algorithm.
7510     retval = __kmp_avail_proc - __kmp_nth +
7511              (root->r.r_active ? 1 : root->r.r_hot_team->t.t_nproc);
7512     if (retval > set_nproc) {
7513       retval = set_nproc;
7514     }
7515     if (retval < KMP_MIN_NTH) {
7516       retval = KMP_MIN_NTH;
7517     }
7518 
7519     KB_TRACE(20, ("__kmp_load_balance_nproc: thread limit exit. retval:%d\n",
7520                   retval));
7521     return retval;
7522   }
7523 
7524   // There is a slight delay in the load balance algorithm in detecting new
7525   // running procs. The real system load at this instant should be at least as
7526   // large as the #active omp thread that are available to add to the team.
7527   if (system_active < team_curr_active) {
7528     system_active = team_curr_active;
7529   }
7530   retval = __kmp_avail_proc - system_active + team_curr_active;
7531   if (retval > set_nproc) {
7532     retval = set_nproc;
7533   }
7534   if (retval < KMP_MIN_NTH) {
7535     retval = KMP_MIN_NTH;
7536   }
7537 
7538   KB_TRACE(20, ("__kmp_load_balance_nproc: exit. retval:%d\n", retval));
7539   return retval;
7540 } // __kmp_load_balance_nproc()
7541 
7542 #endif /* USE_LOAD_BALANCE */
7543 
7544 /* ------------------------------------------------------------------------ */
7545 
7546 /* NOTE: this is called with the __kmp_init_lock held */
7547 void __kmp_cleanup(void) {
7548   int f;
7549 
7550   KA_TRACE(10, ("__kmp_cleanup: enter\n"));
7551 
7552   if (TCR_4(__kmp_init_parallel)) {
7553 #if KMP_HANDLE_SIGNALS
7554     __kmp_remove_signals();
7555 #endif
7556     TCW_4(__kmp_init_parallel, FALSE);
7557   }
7558 
7559   if (TCR_4(__kmp_init_middle)) {
7560 #if KMP_AFFINITY_SUPPORTED
7561     __kmp_affinity_uninitialize();
7562 #endif /* KMP_AFFINITY_SUPPORTED */
7563     __kmp_cleanup_hierarchy();
7564     TCW_4(__kmp_init_middle, FALSE);
7565   }
7566 
7567   KA_TRACE(10, ("__kmp_cleanup: go serial cleanup\n"));
7568 
7569   if (__kmp_init_serial) {
7570     __kmp_runtime_destroy();
7571     __kmp_init_serial = FALSE;
7572   }
7573 
7574   __kmp_cleanup_threadprivate_caches();
7575 
7576   for (f = 0; f < __kmp_threads_capacity; f++) {
7577     if (__kmp_root[f] != NULL) {
7578       __kmp_free(__kmp_root[f]);
7579       __kmp_root[f] = NULL;
7580     }
7581   }
7582   __kmp_free(__kmp_threads);
7583   // __kmp_threads and __kmp_root were allocated at once, as single block, so
7584   // there is no need in freeing __kmp_root.
7585   __kmp_threads = NULL;
7586   __kmp_root = NULL;
7587   __kmp_threads_capacity = 0;
7588 
7589 #if KMP_USE_DYNAMIC_LOCK
7590   __kmp_cleanup_indirect_user_locks();
7591 #else
7592   __kmp_cleanup_user_locks();
7593 #endif
7594 
7595 #if KMP_AFFINITY_SUPPORTED
7596   KMP_INTERNAL_FREE(CCAST(char *, __kmp_cpuinfo_file));
7597   __kmp_cpuinfo_file = NULL;
7598 #endif /* KMP_AFFINITY_SUPPORTED */
7599 
7600 #if KMP_USE_ADAPTIVE_LOCKS
7601 #if KMP_DEBUG_ADAPTIVE_LOCKS
7602   __kmp_print_speculative_stats();
7603 #endif
7604 #endif
7605   KMP_INTERNAL_FREE(__kmp_nested_nth.nth);
7606   __kmp_nested_nth.nth = NULL;
7607   __kmp_nested_nth.size = 0;
7608   __kmp_nested_nth.used = 0;
7609   KMP_INTERNAL_FREE(__kmp_nested_proc_bind.bind_types);
7610   __kmp_nested_proc_bind.bind_types = NULL;
7611   __kmp_nested_proc_bind.size = 0;
7612   __kmp_nested_proc_bind.used = 0;
7613 #if OMP_50_ENABLED
7614   if (__kmp_affinity_format) {
7615     KMP_INTERNAL_FREE(__kmp_affinity_format);
7616     __kmp_affinity_format = NULL;
7617   }
7618 #endif
7619 
7620   __kmp_i18n_catclose();
7621 
7622 #if KMP_USE_HIER_SCHED
7623   __kmp_hier_scheds.deallocate();
7624 #endif
7625 
7626 #if KMP_STATS_ENABLED
7627   __kmp_stats_fini();
7628 #endif
7629 
7630   KA_TRACE(10, ("__kmp_cleanup: exit\n"));
7631 }
7632 
7633 /* ------------------------------------------------------------------------ */
7634 
7635 int __kmp_ignore_mppbeg(void) {
7636   char *env;
7637 
7638   if ((env = getenv("KMP_IGNORE_MPPBEG")) != NULL) {
7639     if (__kmp_str_match_false(env))
7640       return FALSE;
7641   }
7642   // By default __kmpc_begin() is no-op.
7643   return TRUE;
7644 }
7645 
7646 int __kmp_ignore_mppend(void) {
7647   char *env;
7648 
7649   if ((env = getenv("KMP_IGNORE_MPPEND")) != NULL) {
7650     if (__kmp_str_match_false(env))
7651       return FALSE;
7652   }
7653   // By default __kmpc_end() is no-op.
7654   return TRUE;
7655 }
7656 
7657 void __kmp_internal_begin(void) {
7658   int gtid;
7659   kmp_root_t *root;
7660 
7661   /* this is a very important step as it will register new sibling threads
7662      and assign these new uber threads a new gtid */
7663   gtid = __kmp_entry_gtid();
7664   root = __kmp_threads[gtid]->th.th_root;
7665   KMP_ASSERT(KMP_UBER_GTID(gtid));
7666 
7667   if (root->r.r_begin)
7668     return;
7669   __kmp_acquire_lock(&root->r.r_begin_lock, gtid);
7670   if (root->r.r_begin) {
7671     __kmp_release_lock(&root->r.r_begin_lock, gtid);
7672     return;
7673   }
7674 
7675   root->r.r_begin = TRUE;
7676 
7677   __kmp_release_lock(&root->r.r_begin_lock, gtid);
7678 }
7679 
7680 /* ------------------------------------------------------------------------ */
7681 
7682 void __kmp_user_set_library(enum library_type arg) {
7683   int gtid;
7684   kmp_root_t *root;
7685   kmp_info_t *thread;
7686 
7687   /* first, make sure we are initialized so we can get our gtid */
7688 
7689   gtid = __kmp_entry_gtid();
7690   thread = __kmp_threads[gtid];
7691 
7692   root = thread->th.th_root;
7693 
7694   KA_TRACE(20, ("__kmp_user_set_library: enter T#%d, arg: %d, %d\n", gtid, arg,
7695                 library_serial));
7696   if (root->r.r_in_parallel) { /* Must be called in serial section of top-level
7697                                   thread */
7698     KMP_WARNING(SetLibraryIncorrectCall);
7699     return;
7700   }
7701 
7702   switch (arg) {
7703   case library_serial:
7704     thread->th.th_set_nproc = 0;
7705     set__nproc(thread, 1);
7706     break;
7707   case library_turnaround:
7708     thread->th.th_set_nproc = 0;
7709     set__nproc(thread, __kmp_dflt_team_nth ? __kmp_dflt_team_nth
7710                                            : __kmp_dflt_team_nth_ub);
7711     break;
7712   case library_throughput:
7713     thread->th.th_set_nproc = 0;
7714     set__nproc(thread, __kmp_dflt_team_nth ? __kmp_dflt_team_nth
7715                                            : __kmp_dflt_team_nth_ub);
7716     break;
7717   default:
7718     KMP_FATAL(UnknownLibraryType, arg);
7719   }
7720 
7721   __kmp_aux_set_library(arg);
7722 }
7723 
7724 void __kmp_aux_set_stacksize(size_t arg) {
7725   if (!__kmp_init_serial)
7726     __kmp_serial_initialize();
7727 
7728 #if KMP_OS_DARWIN
7729   if (arg & (0x1000 - 1)) {
7730     arg &= ~(0x1000 - 1);
7731     if (arg + 0x1000) /* check for overflow if we round up */
7732       arg += 0x1000;
7733   }
7734 #endif
7735   __kmp_acquire_bootstrap_lock(&__kmp_initz_lock);
7736 
7737   /* only change the default stacksize before the first parallel region */
7738   if (!TCR_4(__kmp_init_parallel)) {
7739     size_t value = arg; /* argument is in bytes */
7740 
7741     if (value < __kmp_sys_min_stksize)
7742       value = __kmp_sys_min_stksize;
7743     else if (value > KMP_MAX_STKSIZE)
7744       value = KMP_MAX_STKSIZE;
7745 
7746     __kmp_stksize = value;
7747 
7748     __kmp_env_stksize = TRUE; /* was KMP_STACKSIZE specified? */
7749   }
7750 
7751   __kmp_release_bootstrap_lock(&__kmp_initz_lock);
7752 }
7753 
7754 /* set the behaviour of the runtime library */
7755 /* TODO this can cause some odd behaviour with sibling parallelism... */
7756 void __kmp_aux_set_library(enum library_type arg) {
7757   __kmp_library = arg;
7758 
7759   switch (__kmp_library) {
7760   case library_serial: {
7761     KMP_INFORM(LibraryIsSerial);
7762   } break;
7763   case library_turnaround:
7764     if (__kmp_use_yield == 1 && !__kmp_use_yield_exp_set)
7765       __kmp_use_yield = 2; // only yield when oversubscribed
7766     break;
7767   case library_throughput:
7768     if (__kmp_dflt_blocktime == KMP_MAX_BLOCKTIME)
7769       __kmp_dflt_blocktime = 200;
7770     break;
7771   default:
7772     KMP_FATAL(UnknownLibraryType, arg);
7773   }
7774 }
7775 
7776 /* Getting team information common for all team API */
7777 // Returns NULL if not in teams construct
7778 static kmp_team_t *__kmp_aux_get_team_info(int &teams_serialized) {
7779   kmp_info_t *thr = __kmp_entry_thread();
7780   teams_serialized = 0;
7781   if (thr->th.th_teams_microtask) {
7782     kmp_team_t *team = thr->th.th_team;
7783     int tlevel = thr->th.th_teams_level; // the level of the teams construct
7784     int ii = team->t.t_level;
7785     teams_serialized = team->t.t_serialized;
7786     int level = tlevel + 1;
7787     KMP_DEBUG_ASSERT(ii >= tlevel);
7788     while (ii > level) {
7789       for (teams_serialized = team->t.t_serialized;
7790            (teams_serialized > 0) && (ii > level); teams_serialized--, ii--) {
7791       }
7792       if (team->t.t_serialized && (!teams_serialized)) {
7793         team = team->t.t_parent;
7794         continue;
7795       }
7796       if (ii > level) {
7797         team = team->t.t_parent;
7798         ii--;
7799       }
7800     }
7801     return team;
7802   }
7803   return NULL;
7804 }
7805 
7806 int __kmp_aux_get_team_num() {
7807   int serialized;
7808   kmp_team_t *team = __kmp_aux_get_team_info(serialized);
7809   if (team) {
7810     if (serialized > 1) {
7811       return 0; // teams region is serialized ( 1 team of 1 thread ).
7812     } else {
7813       return team->t.t_master_tid;
7814     }
7815   }
7816   return 0;
7817 }
7818 
7819 int __kmp_aux_get_num_teams() {
7820   int serialized;
7821   kmp_team_t *team = __kmp_aux_get_team_info(serialized);
7822   if (team) {
7823     if (serialized > 1) {
7824       return 1;
7825     } else {
7826       return team->t.t_parent->t.t_nproc;
7827     }
7828   }
7829   return 1;
7830 }
7831 
7832 /* ------------------------------------------------------------------------ */
7833 
7834 #if OMP_50_ENABLED
7835 /*
7836  * Affinity Format Parser
7837  *
7838  * Field is in form of: %[[[0].]size]type
7839  * % and type are required (%% means print a literal '%')
7840  * type is either single char or long name surrounded by {},
7841  * e.g., N or {num_threads}
7842  * 0 => leading zeros
7843  * . => right justified when size is specified
7844  * by default output is left justified
7845  * size is the *minimum* field length
7846  * All other characters are printed as is
7847  *
7848  * Available field types:
7849  * L {thread_level}      - omp_get_level()
7850  * n {thread_num}        - omp_get_thread_num()
7851  * h {host}              - name of host machine
7852  * P {process_id}        - process id (integer)
7853  * T {thread_identifier} - native thread identifier (integer)
7854  * N {num_threads}       - omp_get_num_threads()
7855  * A {ancestor_tnum}     - omp_get_ancestor_thread_num(omp_get_level()-1)
7856  * a {thread_affinity}   - comma separated list of integers or integer ranges
7857  *                         (values of affinity mask)
7858  *
7859  * Implementation-specific field types can be added
7860  * If a type is unknown, print "undefined"
7861 */
7862 
7863 // Structure holding the short name, long name, and corresponding data type
7864 // for snprintf.  A table of these will represent the entire valid keyword
7865 // field types.
7866 typedef struct kmp_affinity_format_field_t {
7867   char short_name; // from spec e.g., L -> thread level
7868   const char *long_name; // from spec thread_level -> thread level
7869   char field_format; // data type for snprintf (typically 'd' or 's'
7870   // for integer or string)
7871 } kmp_affinity_format_field_t;
7872 
7873 static const kmp_affinity_format_field_t __kmp_affinity_format_table[] = {
7874 #if KMP_AFFINITY_SUPPORTED
7875     {'A', "thread_affinity", 's'},
7876 #endif
7877     {'t', "team_num", 'd'},
7878     {'T', "num_teams", 'd'},
7879     {'L', "nesting_level", 'd'},
7880     {'n', "thread_num", 'd'},
7881     {'N', "num_threads", 'd'},
7882     {'a', "ancestor_tnum", 'd'},
7883     {'H', "host", 's'},
7884     {'P', "process_id", 'd'},
7885     {'i', "native_thread_id", 'd'}};
7886 
7887 // Return the number of characters it takes to hold field
7888 static int __kmp_aux_capture_affinity_field(int gtid, const kmp_info_t *th,
7889                                             const char **ptr,
7890                                             kmp_str_buf_t *field_buffer) {
7891   int rc, format_index, field_value;
7892   const char *width_left, *width_right;
7893   bool pad_zeros, right_justify, parse_long_name, found_valid_name;
7894   static const int FORMAT_SIZE = 20;
7895   char format[FORMAT_SIZE] = {0};
7896   char absolute_short_name = 0;
7897 
7898   KMP_DEBUG_ASSERT(gtid >= 0);
7899   KMP_DEBUG_ASSERT(th);
7900   KMP_DEBUG_ASSERT(**ptr == '%');
7901   KMP_DEBUG_ASSERT(field_buffer);
7902 
7903   __kmp_str_buf_clear(field_buffer);
7904 
7905   // Skip the initial %
7906   (*ptr)++;
7907 
7908   // Check for %% first
7909   if (**ptr == '%') {
7910     __kmp_str_buf_cat(field_buffer, "%", 1);
7911     (*ptr)++; // skip over the second %
7912     return 1;
7913   }
7914 
7915   // Parse field modifiers if they are present
7916   pad_zeros = false;
7917   if (**ptr == '0') {
7918     pad_zeros = true;
7919     (*ptr)++; // skip over 0
7920   }
7921   right_justify = false;
7922   if (**ptr == '.') {
7923     right_justify = true;
7924     (*ptr)++; // skip over .
7925   }
7926   // Parse width of field: [width_left, width_right)
7927   width_left = width_right = NULL;
7928   if (**ptr >= '0' && **ptr <= '9') {
7929     width_left = *ptr;
7930     SKIP_DIGITS(*ptr);
7931     width_right = *ptr;
7932   }
7933 
7934   // Create the format for KMP_SNPRINTF based on flags parsed above
7935   format_index = 0;
7936   format[format_index++] = '%';
7937   if (!right_justify)
7938     format[format_index++] = '-';
7939   if (pad_zeros)
7940     format[format_index++] = '0';
7941   if (width_left && width_right) {
7942     int i = 0;
7943     // Only allow 8 digit number widths.
7944     // This also prevents overflowing format variable
7945     while (i < 8 && width_left < width_right) {
7946       format[format_index++] = *width_left;
7947       width_left++;
7948       i++;
7949     }
7950   }
7951 
7952   // Parse a name (long or short)
7953   // Canonicalize the name into absolute_short_name
7954   found_valid_name = false;
7955   parse_long_name = (**ptr == '{');
7956   if (parse_long_name)
7957     (*ptr)++; // skip initial left brace
7958   for (size_t i = 0; i < sizeof(__kmp_affinity_format_table) /
7959                              sizeof(__kmp_affinity_format_table[0]);
7960        ++i) {
7961     char short_name = __kmp_affinity_format_table[i].short_name;
7962     const char *long_name = __kmp_affinity_format_table[i].long_name;
7963     char field_format = __kmp_affinity_format_table[i].field_format;
7964     if (parse_long_name) {
7965       int length = KMP_STRLEN(long_name);
7966       if (strncmp(*ptr, long_name, length) == 0) {
7967         found_valid_name = true;
7968         (*ptr) += length; // skip the long name
7969       }
7970     } else if (**ptr == short_name) {
7971       found_valid_name = true;
7972       (*ptr)++; // skip the short name
7973     }
7974     if (found_valid_name) {
7975       format[format_index++] = field_format;
7976       format[format_index++] = '\0';
7977       absolute_short_name = short_name;
7978       break;
7979     }
7980   }
7981   if (parse_long_name) {
7982     if (**ptr != '}') {
7983       absolute_short_name = 0;
7984     } else {
7985       (*ptr)++; // skip over the right brace
7986     }
7987   }
7988 
7989   // Attempt to fill the buffer with the requested
7990   // value using snprintf within __kmp_str_buf_print()
7991   switch (absolute_short_name) {
7992   case 't':
7993     rc = __kmp_str_buf_print(field_buffer, format, __kmp_aux_get_team_num());
7994     break;
7995   case 'T':
7996     rc = __kmp_str_buf_print(field_buffer, format, __kmp_aux_get_num_teams());
7997     break;
7998   case 'L':
7999     rc = __kmp_str_buf_print(field_buffer, format, th->th.th_team->t.t_level);
8000     break;
8001   case 'n':
8002     rc = __kmp_str_buf_print(field_buffer, format, __kmp_tid_from_gtid(gtid));
8003     break;
8004   case 'H': {
8005     static const int BUFFER_SIZE = 256;
8006     char buf[BUFFER_SIZE];
8007     __kmp_expand_host_name(buf, BUFFER_SIZE);
8008     rc = __kmp_str_buf_print(field_buffer, format, buf);
8009   } break;
8010   case 'P':
8011     rc = __kmp_str_buf_print(field_buffer, format, getpid());
8012     break;
8013   case 'i':
8014     rc = __kmp_str_buf_print(field_buffer, format, __kmp_gettid());
8015     break;
8016   case 'N':
8017     rc = __kmp_str_buf_print(field_buffer, format, th->th.th_team->t.t_nproc);
8018     break;
8019   case 'a':
8020     field_value =
8021         __kmp_get_ancestor_thread_num(gtid, th->th.th_team->t.t_level - 1);
8022     rc = __kmp_str_buf_print(field_buffer, format, field_value);
8023     break;
8024 #if KMP_AFFINITY_SUPPORTED
8025   case 'A': {
8026     kmp_str_buf_t buf;
8027     __kmp_str_buf_init(&buf);
8028     __kmp_affinity_str_buf_mask(&buf, th->th.th_affin_mask);
8029     rc = __kmp_str_buf_print(field_buffer, format, buf.str);
8030     __kmp_str_buf_free(&buf);
8031   } break;
8032 #endif
8033   default:
8034     // According to spec, If an implementation does not have info for field
8035     // type, then "undefined" is printed
8036     rc = __kmp_str_buf_print(field_buffer, "%s", "undefined");
8037     // Skip the field
8038     if (parse_long_name) {
8039       SKIP_TOKEN(*ptr);
8040       if (**ptr == '}')
8041         (*ptr)++;
8042     } else {
8043       (*ptr)++;
8044     }
8045   }
8046 
8047   KMP_ASSERT(format_index <= FORMAT_SIZE);
8048   return rc;
8049 }
8050 
8051 /*
8052  * Return number of characters needed to hold the affinity string
8053  * (not including null byte character)
8054  * The resultant string is printed to buffer, which the caller can then
8055  * handle afterwards
8056 */
8057 size_t __kmp_aux_capture_affinity(int gtid, const char *format,
8058                                   kmp_str_buf_t *buffer) {
8059   const char *parse_ptr;
8060   size_t retval;
8061   const kmp_info_t *th;
8062   kmp_str_buf_t field;
8063 
8064   KMP_DEBUG_ASSERT(buffer);
8065   KMP_DEBUG_ASSERT(gtid >= 0);
8066 
8067   __kmp_str_buf_init(&field);
8068   __kmp_str_buf_clear(buffer);
8069 
8070   th = __kmp_threads[gtid];
8071   retval = 0;
8072 
8073   // If format is NULL or zero-length string, then we use
8074   // affinity-format-var ICV
8075   parse_ptr = format;
8076   if (parse_ptr == NULL || *parse_ptr == '\0') {
8077     parse_ptr = __kmp_affinity_format;
8078   }
8079   KMP_DEBUG_ASSERT(parse_ptr);
8080 
8081   while (*parse_ptr != '\0') {
8082     // Parse a field
8083     if (*parse_ptr == '%') {
8084       // Put field in the buffer
8085       int rc = __kmp_aux_capture_affinity_field(gtid, th, &parse_ptr, &field);
8086       __kmp_str_buf_catbuf(buffer, &field);
8087       retval += rc;
8088     } else {
8089       // Put literal character in buffer
8090       __kmp_str_buf_cat(buffer, parse_ptr, 1);
8091       retval++;
8092       parse_ptr++;
8093     }
8094   }
8095   __kmp_str_buf_free(&field);
8096   return retval;
8097 }
8098 
8099 // Displays the affinity string to stdout
8100 void __kmp_aux_display_affinity(int gtid, const char *format) {
8101   kmp_str_buf_t buf;
8102   __kmp_str_buf_init(&buf);
8103   __kmp_aux_capture_affinity(gtid, format, &buf);
8104   __kmp_fprintf(kmp_out, "%s" KMP_END_OF_LINE, buf.str);
8105   __kmp_str_buf_free(&buf);
8106 }
8107 #endif // OMP_50_ENABLED
8108 
8109 /* ------------------------------------------------------------------------ */
8110 
8111 void __kmp_aux_set_blocktime(int arg, kmp_info_t *thread, int tid) {
8112   int blocktime = arg; /* argument is in milliseconds */
8113 #if KMP_USE_MONITOR
8114   int bt_intervals;
8115 #endif
8116   int bt_set;
8117 
8118   __kmp_save_internal_controls(thread);
8119 
8120   /* Normalize and set blocktime for the teams */
8121   if (blocktime < KMP_MIN_BLOCKTIME)
8122     blocktime = KMP_MIN_BLOCKTIME;
8123   else if (blocktime > KMP_MAX_BLOCKTIME)
8124     blocktime = KMP_MAX_BLOCKTIME;
8125 
8126   set__blocktime_team(thread->th.th_team, tid, blocktime);
8127   set__blocktime_team(thread->th.th_serial_team, 0, blocktime);
8128 
8129 #if KMP_USE_MONITOR
8130   /* Calculate and set blocktime intervals for the teams */
8131   bt_intervals = KMP_INTERVALS_FROM_BLOCKTIME(blocktime, __kmp_monitor_wakeups);
8132 
8133   set__bt_intervals_team(thread->th.th_team, tid, bt_intervals);
8134   set__bt_intervals_team(thread->th.th_serial_team, 0, bt_intervals);
8135 #endif
8136 
8137   /* Set whether blocktime has been set to "TRUE" */
8138   bt_set = TRUE;
8139 
8140   set__bt_set_team(thread->th.th_team, tid, bt_set);
8141   set__bt_set_team(thread->th.th_serial_team, 0, bt_set);
8142 #if KMP_USE_MONITOR
8143   KF_TRACE(10, ("kmp_set_blocktime: T#%d(%d:%d), blocktime=%d, "
8144                 "bt_intervals=%d, monitor_updates=%d\n",
8145                 __kmp_gtid_from_tid(tid, thread->th.th_team),
8146                 thread->th.th_team->t.t_id, tid, blocktime, bt_intervals,
8147                 __kmp_monitor_wakeups));
8148 #else
8149   KF_TRACE(10, ("kmp_set_blocktime: T#%d(%d:%d), blocktime=%d\n",
8150                 __kmp_gtid_from_tid(tid, thread->th.th_team),
8151                 thread->th.th_team->t.t_id, tid, blocktime));
8152 #endif
8153 }
8154 
8155 void __kmp_aux_set_defaults(char const *str, int len) {
8156   if (!__kmp_init_serial) {
8157     __kmp_serial_initialize();
8158   }
8159   __kmp_env_initialize(str);
8160 
8161   if (__kmp_settings
8162 #if OMP_40_ENABLED
8163       || __kmp_display_env || __kmp_display_env_verbose
8164 #endif // OMP_40_ENABLED
8165       ) {
8166     __kmp_env_print();
8167   }
8168 } // __kmp_aux_set_defaults
8169 
8170 /* ------------------------------------------------------------------------ */
8171 /* internal fast reduction routines */
8172 
8173 PACKED_REDUCTION_METHOD_T
8174 __kmp_determine_reduction_method(
8175     ident_t *loc, kmp_int32 global_tid, kmp_int32 num_vars, size_t reduce_size,
8176     void *reduce_data, void (*reduce_func)(void *lhs_data, void *rhs_data),
8177     kmp_critical_name *lck) {
8178 
8179   // Default reduction method: critical construct ( lck != NULL, like in current
8180   // PAROPT )
8181   // If ( reduce_data!=NULL && reduce_func!=NULL ): the tree-reduction method
8182   // can be selected by RTL
8183   // If loc->flags contains KMP_IDENT_ATOMIC_REDUCE, the atomic reduce method
8184   // can be selected by RTL
8185   // Finally, it's up to OpenMP RTL to make a decision on which method to select
8186   // among generated by PAROPT.
8187 
8188   PACKED_REDUCTION_METHOD_T retval;
8189 
8190   int team_size;
8191 
8192   KMP_DEBUG_ASSERT(loc); // it would be nice to test ( loc != 0 )
8193   KMP_DEBUG_ASSERT(lck); // it would be nice to test ( lck != 0 )
8194 
8195 #define FAST_REDUCTION_ATOMIC_METHOD_GENERATED                                 \
8196   ((loc->flags & (KMP_IDENT_ATOMIC_REDUCE)) == (KMP_IDENT_ATOMIC_REDUCE))
8197 #define FAST_REDUCTION_TREE_METHOD_GENERATED ((reduce_data) && (reduce_func))
8198 
8199   retval = critical_reduce_block;
8200 
8201   // another choice of getting a team size (with 1 dynamic deference) is slower
8202   team_size = __kmp_get_team_num_threads(global_tid);
8203   if (team_size == 1) {
8204 
8205     retval = empty_reduce_block;
8206 
8207   } else {
8208 
8209     int atomic_available = FAST_REDUCTION_ATOMIC_METHOD_GENERATED;
8210 
8211 #if KMP_ARCH_X86_64 || KMP_ARCH_PPC64 || KMP_ARCH_AARCH64 || KMP_ARCH_MIPS64
8212 
8213 #if KMP_OS_LINUX || KMP_OS_DRAGONFLY || KMP_OS_FREEBSD || KMP_OS_NETBSD ||     \
8214     KMP_OS_OPENBSD || KMP_OS_WINDOWS || KMP_OS_DARWIN || KMP_OS_HURD
8215 
8216     int teamsize_cutoff = 4;
8217 
8218 #if KMP_MIC_SUPPORTED
8219     if (__kmp_mic_type != non_mic) {
8220       teamsize_cutoff = 8;
8221     }
8222 #endif
8223     int tree_available = FAST_REDUCTION_TREE_METHOD_GENERATED;
8224     if (tree_available) {
8225       if (team_size <= teamsize_cutoff) {
8226         if (atomic_available) {
8227           retval = atomic_reduce_block;
8228         }
8229       } else {
8230         retval = TREE_REDUCE_BLOCK_WITH_REDUCTION_BARRIER;
8231       }
8232     } else if (atomic_available) {
8233       retval = atomic_reduce_block;
8234     }
8235 #else
8236 #error "Unknown or unsupported OS"
8237 #endif // KMP_OS_LINUX || KMP_OS_DRAGONFLY || KMP_OS_FREEBSD || KMP_OS_NETBSD ||
8238        // KMP_OS_OPENBSD || KMP_OS_WINDOWS || KMP_OS_DARWIN || KMP_OS_HURD
8239 
8240 #elif KMP_ARCH_X86 || KMP_ARCH_ARM || KMP_ARCH_AARCH || KMP_ARCH_MIPS
8241 
8242 #if KMP_OS_LINUX || KMP_OS_WINDOWS || KMP_OS_HURD
8243 
8244     // basic tuning
8245 
8246     if (atomic_available) {
8247       if (num_vars <= 2) { // && ( team_size <= 8 ) due to false-sharing ???
8248         retval = atomic_reduce_block;
8249       }
8250     } // otherwise: use critical section
8251 
8252 #elif KMP_OS_DARWIN
8253 
8254     int tree_available = FAST_REDUCTION_TREE_METHOD_GENERATED;
8255     if (atomic_available && (num_vars <= 3)) {
8256       retval = atomic_reduce_block;
8257     } else if (tree_available) {
8258       if ((reduce_size > (9 * sizeof(kmp_real64))) &&
8259           (reduce_size < (2000 * sizeof(kmp_real64)))) {
8260         retval = TREE_REDUCE_BLOCK_WITH_PLAIN_BARRIER;
8261       }
8262     } // otherwise: use critical section
8263 
8264 #else
8265 #error "Unknown or unsupported OS"
8266 #endif
8267 
8268 #else
8269 #error "Unknown or unsupported architecture"
8270 #endif
8271   }
8272 
8273   // KMP_FORCE_REDUCTION
8274 
8275   // If the team is serialized (team_size == 1), ignore the forced reduction
8276   // method and stay with the unsynchronized method (empty_reduce_block)
8277   if (__kmp_force_reduction_method != reduction_method_not_defined &&
8278       team_size != 1) {
8279 
8280     PACKED_REDUCTION_METHOD_T forced_retval = critical_reduce_block;
8281 
8282     int atomic_available, tree_available;
8283 
8284     switch ((forced_retval = __kmp_force_reduction_method)) {
8285     case critical_reduce_block:
8286       KMP_ASSERT(lck); // lck should be != 0
8287       break;
8288 
8289     case atomic_reduce_block:
8290       atomic_available = FAST_REDUCTION_ATOMIC_METHOD_GENERATED;
8291       if (!atomic_available) {
8292         KMP_WARNING(RedMethodNotSupported, "atomic");
8293         forced_retval = critical_reduce_block;
8294       }
8295       break;
8296 
8297     case tree_reduce_block:
8298       tree_available = FAST_REDUCTION_TREE_METHOD_GENERATED;
8299       if (!tree_available) {
8300         KMP_WARNING(RedMethodNotSupported, "tree");
8301         forced_retval = critical_reduce_block;
8302       } else {
8303 #if KMP_FAST_REDUCTION_BARRIER
8304         forced_retval = TREE_REDUCE_BLOCK_WITH_REDUCTION_BARRIER;
8305 #endif
8306       }
8307       break;
8308 
8309     default:
8310       KMP_ASSERT(0); // "unsupported method specified"
8311     }
8312 
8313     retval = forced_retval;
8314   }
8315 
8316   KA_TRACE(10, ("reduction method selected=%08x\n", retval));
8317 
8318 #undef FAST_REDUCTION_TREE_METHOD_GENERATED
8319 #undef FAST_REDUCTION_ATOMIC_METHOD_GENERATED
8320 
8321   return (retval);
8322 }
8323 
8324 // this function is for testing set/get/determine reduce method
8325 kmp_int32 __kmp_get_reduce_method(void) {
8326   return ((__kmp_entry_thread()->th.th_local.packed_reduction_method) >> 8);
8327 }
8328 
8329 #if OMP_50_ENABLED
8330 
8331 // Soft pause sets up threads to ignore blocktime and just go to sleep.
8332 // Spin-wait code checks __kmp_pause_status and reacts accordingly.
8333 void __kmp_soft_pause() { __kmp_pause_status = kmp_soft_paused; }
8334 
8335 // Hard pause shuts down the runtime completely.  Resume happens naturally when
8336 // OpenMP is used subsequently.
8337 void __kmp_hard_pause() {
8338   __kmp_pause_status = kmp_hard_paused;
8339   __kmp_internal_end_thread(-1);
8340 }
8341 
8342 // Soft resume sets __kmp_pause_status, and wakes up all threads.
8343 void __kmp_resume_if_soft_paused() {
8344   if (__kmp_pause_status == kmp_soft_paused) {
8345     __kmp_pause_status = kmp_not_paused;
8346 
8347     for (int gtid = 1; gtid < __kmp_threads_capacity; ++gtid) {
8348       kmp_info_t *thread = __kmp_threads[gtid];
8349       if (thread) { // Wake it if sleeping
8350         kmp_flag_64 fl(&thread->th.th_bar[bs_forkjoin_barrier].bb.b_go, thread);
8351         if (fl.is_sleeping())
8352           fl.resume(gtid);
8353         else if (__kmp_try_suspend_mx(thread)) { // got suspend lock
8354           __kmp_unlock_suspend_mx(thread); // unlock it; it won't sleep
8355         } else { // thread holds the lock and may sleep soon
8356           do { // until either the thread sleeps, or we can get the lock
8357             if (fl.is_sleeping()) {
8358               fl.resume(gtid);
8359               break;
8360             } else if (__kmp_try_suspend_mx(thread)) {
8361               __kmp_unlock_suspend_mx(thread);
8362               break;
8363             }
8364           } while (1);
8365         }
8366       }
8367     }
8368   }
8369 }
8370 
8371 // This function is called via __kmpc_pause_resource. Returns 0 if successful.
8372 // TODO: add warning messages
8373 int __kmp_pause_resource(kmp_pause_status_t level) {
8374   if (level == kmp_not_paused) { // requesting resume
8375     if (__kmp_pause_status == kmp_not_paused) {
8376       // error message about runtime not being paused, so can't resume
8377       return 1;
8378     } else {
8379       KMP_DEBUG_ASSERT(__kmp_pause_status == kmp_soft_paused ||
8380                        __kmp_pause_status == kmp_hard_paused);
8381       __kmp_pause_status = kmp_not_paused;
8382       return 0;
8383     }
8384   } else if (level == kmp_soft_paused) { // requesting soft pause
8385     if (__kmp_pause_status != kmp_not_paused) {
8386       // error message about already being paused
8387       return 1;
8388     } else {
8389       __kmp_soft_pause();
8390       return 0;
8391     }
8392   } else if (level == kmp_hard_paused) { // requesting hard pause
8393     if (__kmp_pause_status != kmp_not_paused) {
8394       // error message about already being paused
8395       return 1;
8396     } else {
8397       __kmp_hard_pause();
8398       return 0;
8399     }
8400   } else {
8401     // error message about invalid level
8402     return 1;
8403   }
8404 }
8405 
8406 #endif // OMP_50_ENABLED
8407