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 (!team->t.t_invoke(gtid)) {
2289     KMP_ASSERT2(0, "cannot invoke microtask for MASTER thread");
2290   }
2291   KA_TRACE(20, ("__kmp_fork_call: T#%d(%d:0) done microtask = %p\n", gtid,
2292                 team->t.t_id, team->t.t_pkfn));
2293   KMP_MB(); /* Flush all pending memory write invalidates.  */
2294 
2295   KA_TRACE(20, ("__kmp_fork_call: parallel exit T#%d\n", gtid));
2296 
2297 #if OMPT_SUPPORT
2298   if (ompt_enabled.enabled) {
2299     master_th->th.ompt_thread_info.state = ompt_state_overhead;
2300   }
2301 #endif
2302 
2303   return TRUE;
2304 }
2305 
2306 #if OMPT_SUPPORT
2307 static inline void __kmp_join_restore_state(kmp_info_t *thread,
2308                                             kmp_team_t *team) {
2309   // restore state outside the region
2310   thread->th.ompt_thread_info.state =
2311       ((team->t.t_serialized) ? ompt_state_work_serial
2312                               : ompt_state_work_parallel);
2313 }
2314 
2315 static inline void __kmp_join_ompt(int gtid, kmp_info_t *thread,
2316                                    kmp_team_t *team, ompt_data_t *parallel_data,
2317                                    fork_context_e fork_context, void *codeptr) {
2318   ompt_task_info_t *task_info = __ompt_get_task_info_object(0);
2319   if (ompt_enabled.ompt_callback_parallel_end) {
2320     ompt_callbacks.ompt_callback(ompt_callback_parallel_end)(
2321         parallel_data, &(task_info->task_data), OMPT_INVOKER(fork_context),
2322         codeptr);
2323   }
2324 
2325   task_info->frame.enter_frame = ompt_data_none;
2326   __kmp_join_restore_state(thread, team);
2327 }
2328 #endif
2329 
2330 void __kmp_join_call(ident_t *loc, int gtid
2331 #if OMPT_SUPPORT
2332                      ,
2333                      enum fork_context_e fork_context
2334 #endif
2335 #if OMP_40_ENABLED
2336                      ,
2337                      int exit_teams
2338 #endif /* OMP_40_ENABLED */
2339                      ) {
2340   KMP_TIME_DEVELOPER_PARTITIONED_BLOCK(KMP_join_call);
2341   kmp_team_t *team;
2342   kmp_team_t *parent_team;
2343   kmp_info_t *master_th;
2344   kmp_root_t *root;
2345   int master_active;
2346 
2347   KA_TRACE(20, ("__kmp_join_call: enter T#%d\n", gtid));
2348 
2349   /* setup current data */
2350   master_th = __kmp_threads[gtid];
2351   root = master_th->th.th_root;
2352   team = master_th->th.th_team;
2353   parent_team = team->t.t_parent;
2354 
2355   master_th->th.th_ident = loc;
2356 
2357 #if OMPT_SUPPORT
2358   if (ompt_enabled.enabled) {
2359     master_th->th.ompt_thread_info.state = ompt_state_overhead;
2360   }
2361 #endif
2362 
2363 #if KMP_DEBUG
2364   if (__kmp_tasking_mode != tskm_immediate_exec && !exit_teams) {
2365     KA_TRACE(20, ("__kmp_join_call: T#%d, old team = %p old task_team = %p, "
2366                   "th_task_team = %p\n",
2367                   __kmp_gtid_from_thread(master_th), team,
2368                   team->t.t_task_team[master_th->th.th_task_state],
2369                   master_th->th.th_task_team));
2370     KMP_DEBUG_ASSERT(master_th->th.th_task_team ==
2371                      team->t.t_task_team[master_th->th.th_task_state]);
2372   }
2373 #endif
2374 
2375   if (team->t.t_serialized) {
2376 #if OMP_40_ENABLED
2377     if (master_th->th.th_teams_microtask) {
2378       // We are in teams construct
2379       int level = team->t.t_level;
2380       int tlevel = master_th->th.th_teams_level;
2381       if (level == tlevel) {
2382         // AC: we haven't incremented it earlier at start of teams construct,
2383         //     so do it here - at the end of teams construct
2384         team->t.t_level++;
2385       } else if (level == tlevel + 1) {
2386         // AC: we are exiting parallel inside teams, need to increment
2387         // serialization in order to restore it in the next call to
2388         // __kmpc_end_serialized_parallel
2389         team->t.t_serialized++;
2390       }
2391     }
2392 #endif /* OMP_40_ENABLED */
2393     __kmpc_end_serialized_parallel(loc, gtid);
2394 
2395 #if OMPT_SUPPORT
2396     if (ompt_enabled.enabled) {
2397       __kmp_join_restore_state(master_th, parent_team);
2398     }
2399 #endif
2400 
2401     return;
2402   }
2403 
2404   master_active = team->t.t_master_active;
2405 
2406 #if OMP_40_ENABLED
2407   if (!exit_teams)
2408 #endif /* OMP_40_ENABLED */
2409   {
2410     // AC: No barrier for internal teams at exit from teams construct.
2411     //     But there is barrier for external team (league).
2412     __kmp_internal_join(loc, gtid, team);
2413   }
2414 #if OMP_40_ENABLED
2415   else {
2416     master_th->th.th_task_state =
2417         0; // AC: no tasking in teams (out of any parallel)
2418   }
2419 #endif /* OMP_40_ENABLED */
2420 
2421   KMP_MB();
2422 
2423 #if OMPT_SUPPORT
2424   ompt_data_t *parallel_data = &(team->t.ompt_team_info.parallel_data);
2425   void *codeptr = team->t.ompt_team_info.master_return_address;
2426 #endif
2427 
2428 #if USE_ITT_BUILD
2429   if (__itt_stack_caller_create_ptr) {
2430     __kmp_itt_stack_caller_destroy(
2431         (__itt_caller)team->t
2432             .t_stack_id); // destroy the stack stitching id after join barrier
2433   }
2434 
2435   // Mark end of "parallel" region for Intel(R) VTune(TM) analyzer.
2436   if (team->t.t_active_level == 1
2437 #if OMP_40_ENABLED
2438       && !master_th->th.th_teams_microtask /* not in teams construct */
2439 #endif /* OMP_40_ENABLED */
2440       ) {
2441     master_th->th.th_ident = loc;
2442     // only one notification scheme (either "submit" or "forking/joined", not
2443     // both)
2444     if ((__itt_frame_submit_v3_ptr || KMP_ITT_DEBUG) &&
2445         __kmp_forkjoin_frames_mode == 3)
2446       __kmp_itt_frame_submit(gtid, team->t.t_region_time,
2447                              master_th->th.th_frame_time, 0, loc,
2448                              master_th->th.th_team_nproc, 1);
2449     else if ((__itt_frame_end_v3_ptr || KMP_ITT_DEBUG) &&
2450              !__kmp_forkjoin_frames_mode && __kmp_forkjoin_frames)
2451       __kmp_itt_region_joined(gtid);
2452   } // active_level == 1
2453 #endif /* USE_ITT_BUILD */
2454 
2455 #if OMP_40_ENABLED
2456   if (master_th->th.th_teams_microtask && !exit_teams &&
2457       team->t.t_pkfn != (microtask_t)__kmp_teams_master &&
2458       team->t.t_level == master_th->th.th_teams_level + 1) {
2459     // AC: We need to leave the team structure intact at the end of parallel
2460     // inside the teams construct, so that at the next parallel same (hot) team
2461     // works, only adjust nesting levels
2462 
2463     /* Decrement our nested depth level */
2464     team->t.t_level--;
2465     team->t.t_active_level--;
2466     KMP_ATOMIC_DEC(&root->r.r_in_parallel);
2467 
2468     // Restore number of threads in the team if needed. This code relies on
2469     // the proper adjustment of th_teams_size.nth after the fork in
2470     // __kmp_teams_master on each teams master in the case that
2471     // __kmp_reserve_threads reduced it.
2472     if (master_th->th.th_team_nproc < master_th->th.th_teams_size.nth) {
2473       int old_num = master_th->th.th_team_nproc;
2474       int new_num = master_th->th.th_teams_size.nth;
2475       kmp_info_t **other_threads = team->t.t_threads;
2476       team->t.t_nproc = new_num;
2477       for (int i = 0; i < old_num; ++i) {
2478         other_threads[i]->th.th_team_nproc = new_num;
2479       }
2480       // Adjust states of non-used threads of the team
2481       for (int i = old_num; i < new_num; ++i) {
2482         // Re-initialize thread's barrier data.
2483         KMP_DEBUG_ASSERT(other_threads[i]);
2484         kmp_balign_t *balign = other_threads[i]->th.th_bar;
2485         for (int b = 0; b < bs_last_barrier; ++b) {
2486           balign[b].bb.b_arrived = team->t.t_bar[b].b_arrived;
2487           KMP_DEBUG_ASSERT(balign[b].bb.wait_flag != KMP_BARRIER_PARENT_FLAG);
2488 #if USE_DEBUGGER
2489           balign[b].bb.b_worker_arrived = team->t.t_bar[b].b_team_arrived;
2490 #endif
2491         }
2492         if (__kmp_tasking_mode != tskm_immediate_exec) {
2493           // Synchronize thread's task state
2494           other_threads[i]->th.th_task_state = master_th->th.th_task_state;
2495         }
2496       }
2497     }
2498 
2499 #if OMPT_SUPPORT
2500     if (ompt_enabled.enabled) {
2501       __kmp_join_ompt(gtid, master_th, parent_team, parallel_data, fork_context,
2502                       codeptr);
2503     }
2504 #endif
2505 
2506     return;
2507   }
2508 #endif /* OMP_40_ENABLED */
2509 
2510   /* do cleanup and restore the parent team */
2511   master_th->th.th_info.ds.ds_tid = team->t.t_master_tid;
2512   master_th->th.th_local.this_construct = team->t.t_master_this_cons;
2513 
2514   master_th->th.th_dispatch = &parent_team->t.t_dispatch[team->t.t_master_tid];
2515 
2516   /* jc: The following lock has instructions with REL and ACQ semantics,
2517      separating the parallel user code called in this parallel region
2518      from the serial user code called after this function returns. */
2519   __kmp_acquire_bootstrap_lock(&__kmp_forkjoin_lock);
2520 
2521 #if OMP_40_ENABLED
2522   if (!master_th->th.th_teams_microtask ||
2523       team->t.t_level > master_th->th.th_teams_level)
2524 #endif /* OMP_40_ENABLED */
2525   {
2526     /* Decrement our nested depth level */
2527     KMP_ATOMIC_DEC(&root->r.r_in_parallel);
2528   }
2529   KMP_DEBUG_ASSERT(root->r.r_in_parallel >= 0);
2530 
2531 #if OMPT_SUPPORT
2532   if (ompt_enabled.enabled) {
2533     ompt_task_info_t *task_info = __ompt_get_task_info_object(0);
2534     if (ompt_enabled.ompt_callback_implicit_task) {
2535       int ompt_team_size = team->t.t_nproc;
2536       ompt_callbacks.ompt_callback(ompt_callback_implicit_task)(
2537           ompt_scope_end, NULL, &(task_info->task_data), ompt_team_size,
2538           OMPT_CUR_TASK_INFO(master_th)->thread_num, ompt_task_implicit); // TODO: Can this be ompt_task_initial?
2539     }
2540 
2541     task_info->frame.exit_frame = ompt_data_none;
2542     task_info->task_data = ompt_data_none;
2543   }
2544 #endif
2545 
2546   KF_TRACE(10, ("__kmp_join_call1: T#%d, this_thread=%p team=%p\n", 0,
2547                 master_th, team));
2548   __kmp_pop_current_task_from_thread(master_th);
2549 
2550 #if OMP_40_ENABLED && KMP_AFFINITY_SUPPORTED
2551   // Restore master thread's partition.
2552   master_th->th.th_first_place = team->t.t_first_place;
2553   master_th->th.th_last_place = team->t.t_last_place;
2554 #endif /* OMP_40_ENABLED */
2555 #if OMP_50_ENABLED
2556   master_th->th.th_def_allocator = team->t.t_def_allocator;
2557 #endif
2558 
2559   updateHWFPControl(team);
2560 
2561   if (root->r.r_active != master_active)
2562     root->r.r_active = master_active;
2563 
2564   __kmp_free_team(root, team USE_NESTED_HOT_ARG(
2565                             master_th)); // this will free worker threads
2566 
2567   /* this race was fun to find. make sure the following is in the critical
2568      region otherwise assertions may fail occasionally since the old team may be
2569      reallocated and the hierarchy appears inconsistent. it is actually safe to
2570      run and won't cause any bugs, but will cause those assertion failures. it's
2571      only one deref&assign so might as well put this in the critical region */
2572   master_th->th.th_team = parent_team;
2573   master_th->th.th_team_nproc = parent_team->t.t_nproc;
2574   master_th->th.th_team_master = parent_team->t.t_threads[0];
2575   master_th->th.th_team_serialized = parent_team->t.t_serialized;
2576 
2577   /* restore serialized team, if need be */
2578   if (parent_team->t.t_serialized &&
2579       parent_team != master_th->th.th_serial_team &&
2580       parent_team != root->r.r_root_team) {
2581     __kmp_free_team(root,
2582                     master_th->th.th_serial_team USE_NESTED_HOT_ARG(NULL));
2583     master_th->th.th_serial_team = parent_team;
2584   }
2585 
2586   if (__kmp_tasking_mode != tskm_immediate_exec) {
2587     if (master_th->th.th_task_state_top >
2588         0) { // Restore task state from memo stack
2589       KMP_DEBUG_ASSERT(master_th->th.th_task_state_memo_stack);
2590       // Remember master's state if we re-use this nested hot team
2591       master_th->th.th_task_state_memo_stack[master_th->th.th_task_state_top] =
2592           master_th->th.th_task_state;
2593       --master_th->th.th_task_state_top; // pop
2594       // Now restore state at this level
2595       master_th->th.th_task_state =
2596           master_th->th
2597               .th_task_state_memo_stack[master_th->th.th_task_state_top];
2598     }
2599     // Copy the task team from the parent team to the master thread
2600     master_th->th.th_task_team =
2601         parent_team->t.t_task_team[master_th->th.th_task_state];
2602     KA_TRACE(20,
2603              ("__kmp_join_call: Master T#%d restoring task_team %p / team %p\n",
2604               __kmp_gtid_from_thread(master_th), master_th->th.th_task_team,
2605               parent_team));
2606   }
2607 
2608   // TODO: GEH - cannot do this assertion because root thread not set up as
2609   // executing
2610   // KMP_ASSERT( master_th->th.th_current_task->td_flags.executing == 0 );
2611   master_th->th.th_current_task->td_flags.executing = 1;
2612 
2613   __kmp_release_bootstrap_lock(&__kmp_forkjoin_lock);
2614 
2615 #if OMPT_SUPPORT
2616   if (ompt_enabled.enabled) {
2617     __kmp_join_ompt(gtid, master_th, parent_team, parallel_data, fork_context,
2618                     codeptr);
2619   }
2620 #endif
2621 
2622   KMP_MB();
2623   KA_TRACE(20, ("__kmp_join_call: exit T#%d\n", gtid));
2624 }
2625 
2626 /* Check whether we should push an internal control record onto the
2627    serial team stack.  If so, do it.  */
2628 void __kmp_save_internal_controls(kmp_info_t *thread) {
2629 
2630   if (thread->th.th_team != thread->th.th_serial_team) {
2631     return;
2632   }
2633   if (thread->th.th_team->t.t_serialized > 1) {
2634     int push = 0;
2635 
2636     if (thread->th.th_team->t.t_control_stack_top == NULL) {
2637       push = 1;
2638     } else {
2639       if (thread->th.th_team->t.t_control_stack_top->serial_nesting_level !=
2640           thread->th.th_team->t.t_serialized) {
2641         push = 1;
2642       }
2643     }
2644     if (push) { /* push a record on the serial team's stack */
2645       kmp_internal_control_t *control =
2646           (kmp_internal_control_t *)__kmp_allocate(
2647               sizeof(kmp_internal_control_t));
2648 
2649       copy_icvs(control, &thread->th.th_current_task->td_icvs);
2650 
2651       control->serial_nesting_level = thread->th.th_team->t.t_serialized;
2652 
2653       control->next = thread->th.th_team->t.t_control_stack_top;
2654       thread->th.th_team->t.t_control_stack_top = control;
2655     }
2656   }
2657 }
2658 
2659 /* Changes set_nproc */
2660 void __kmp_set_num_threads(int new_nth, int gtid) {
2661   kmp_info_t *thread;
2662   kmp_root_t *root;
2663 
2664   KF_TRACE(10, ("__kmp_set_num_threads: new __kmp_nth = %d\n", new_nth));
2665   KMP_DEBUG_ASSERT(__kmp_init_serial);
2666 
2667   if (new_nth < 1)
2668     new_nth = 1;
2669   else if (new_nth > __kmp_max_nth)
2670     new_nth = __kmp_max_nth;
2671 
2672   KMP_COUNT_VALUE(OMP_set_numthreads, new_nth);
2673   thread = __kmp_threads[gtid];
2674   if (thread->th.th_current_task->td_icvs.nproc == new_nth)
2675     return; // nothing to do
2676 
2677   __kmp_save_internal_controls(thread);
2678 
2679   set__nproc(thread, new_nth);
2680 
2681   // If this omp_set_num_threads() call will cause the hot team size to be
2682   // reduced (in the absence of a num_threads clause), then reduce it now,
2683   // rather than waiting for the next parallel region.
2684   root = thread->th.th_root;
2685   if (__kmp_init_parallel && (!root->r.r_active) &&
2686       (root->r.r_hot_team->t.t_nproc > new_nth)
2687 #if KMP_NESTED_HOT_TEAMS
2688       && __kmp_hot_teams_max_level && !__kmp_hot_teams_mode
2689 #endif
2690       ) {
2691     kmp_team_t *hot_team = root->r.r_hot_team;
2692     int f;
2693 
2694     __kmp_acquire_bootstrap_lock(&__kmp_forkjoin_lock);
2695 
2696     // Release the extra threads we don't need any more.
2697     for (f = new_nth; f < hot_team->t.t_nproc; f++) {
2698       KMP_DEBUG_ASSERT(hot_team->t.t_threads[f] != NULL);
2699       if (__kmp_tasking_mode != tskm_immediate_exec) {
2700         // When decreasing team size, threads no longer in the team should unref
2701         // task team.
2702         hot_team->t.t_threads[f]->th.th_task_team = NULL;
2703       }
2704       __kmp_free_thread(hot_team->t.t_threads[f]);
2705       hot_team->t.t_threads[f] = NULL;
2706     }
2707     hot_team->t.t_nproc = new_nth;
2708 #if KMP_NESTED_HOT_TEAMS
2709     if (thread->th.th_hot_teams) {
2710       KMP_DEBUG_ASSERT(hot_team == thread->th.th_hot_teams[0].hot_team);
2711       thread->th.th_hot_teams[0].hot_team_nth = new_nth;
2712     }
2713 #endif
2714 
2715     __kmp_release_bootstrap_lock(&__kmp_forkjoin_lock);
2716 
2717     // Update the t_nproc field in the threads that are still active.
2718     for (f = 0; f < new_nth; f++) {
2719       KMP_DEBUG_ASSERT(hot_team->t.t_threads[f] != NULL);
2720       hot_team->t.t_threads[f]->th.th_team_nproc = new_nth;
2721     }
2722     // Special flag in case omp_set_num_threads() call
2723     hot_team->t.t_size_changed = -1;
2724   }
2725 }
2726 
2727 /* Changes max_active_levels */
2728 void __kmp_set_max_active_levels(int gtid, int max_active_levels) {
2729   kmp_info_t *thread;
2730 
2731   KF_TRACE(10, ("__kmp_set_max_active_levels: new max_active_levels for thread "
2732                 "%d = (%d)\n",
2733                 gtid, max_active_levels));
2734   KMP_DEBUG_ASSERT(__kmp_init_serial);
2735 
2736   // validate max_active_levels
2737   if (max_active_levels < 0) {
2738     KMP_WARNING(ActiveLevelsNegative, max_active_levels);
2739     // We ignore this call if the user has specified a negative value.
2740     // The current setting won't be changed. The last valid setting will be
2741     // used. A warning will be issued (if warnings are allowed as controlled by
2742     // the KMP_WARNINGS env var).
2743     KF_TRACE(10, ("__kmp_set_max_active_levels: the call is ignored: new "
2744                   "max_active_levels for thread %d = (%d)\n",
2745                   gtid, max_active_levels));
2746     return;
2747   }
2748   if (max_active_levels <= KMP_MAX_ACTIVE_LEVELS_LIMIT) {
2749     // it's OK, the max_active_levels is within the valid range: [ 0;
2750     // KMP_MAX_ACTIVE_LEVELS_LIMIT ]
2751     // We allow a zero value. (implementation defined behavior)
2752   } else {
2753     KMP_WARNING(ActiveLevelsExceedLimit, max_active_levels,
2754                 KMP_MAX_ACTIVE_LEVELS_LIMIT);
2755     max_active_levels = KMP_MAX_ACTIVE_LEVELS_LIMIT;
2756     // Current upper limit is MAX_INT. (implementation defined behavior)
2757     // If the input exceeds the upper limit, we correct the input to be the
2758     // upper limit. (implementation defined behavior)
2759     // Actually, the flow should never get here until we use MAX_INT limit.
2760   }
2761   KF_TRACE(10, ("__kmp_set_max_active_levels: after validation: new "
2762                 "max_active_levels for thread %d = (%d)\n",
2763                 gtid, max_active_levels));
2764 
2765   thread = __kmp_threads[gtid];
2766 
2767   __kmp_save_internal_controls(thread);
2768 
2769   set__max_active_levels(thread, max_active_levels);
2770 }
2771 
2772 /* Gets max_active_levels */
2773 int __kmp_get_max_active_levels(int gtid) {
2774   kmp_info_t *thread;
2775 
2776   KF_TRACE(10, ("__kmp_get_max_active_levels: thread %d\n", gtid));
2777   KMP_DEBUG_ASSERT(__kmp_init_serial);
2778 
2779   thread = __kmp_threads[gtid];
2780   KMP_DEBUG_ASSERT(thread->th.th_current_task);
2781   KF_TRACE(10, ("__kmp_get_max_active_levels: thread %d, curtask=%p, "
2782                 "curtask_maxaclevel=%d\n",
2783                 gtid, thread->th.th_current_task,
2784                 thread->th.th_current_task->td_icvs.max_active_levels));
2785   return thread->th.th_current_task->td_icvs.max_active_levels;
2786 }
2787 
2788 /* Changes def_sched_var ICV values (run-time schedule kind and chunk) */
2789 void __kmp_set_schedule(int gtid, kmp_sched_t kind, int chunk) {
2790   kmp_info_t *thread;
2791   //    kmp_team_t *team;
2792 
2793   KF_TRACE(10, ("__kmp_set_schedule: new schedule for thread %d = (%d, %d)\n",
2794                 gtid, (int)kind, chunk));
2795   KMP_DEBUG_ASSERT(__kmp_init_serial);
2796 
2797   // Check if the kind parameter is valid, correct if needed.
2798   // Valid parameters should fit in one of two intervals - standard or extended:
2799   //       <lower>, <valid>, <upper_std>, <lower_ext>, <valid>, <upper>
2800   // 2008-01-25: 0,  1 - 4,       5,         100,     101 - 102, 103
2801   if (kind <= kmp_sched_lower || kind >= kmp_sched_upper ||
2802       (kind <= kmp_sched_lower_ext && kind >= kmp_sched_upper_std)) {
2803     // TODO: Hint needs attention in case we change the default schedule.
2804     __kmp_msg(kmp_ms_warning, KMP_MSG(ScheduleKindOutOfRange, kind),
2805               KMP_HNT(DefaultScheduleKindUsed, "static, no chunk"),
2806               __kmp_msg_null);
2807     kind = kmp_sched_default;
2808     chunk = 0; // ignore chunk value in case of bad kind
2809   }
2810 
2811   thread = __kmp_threads[gtid];
2812 
2813   __kmp_save_internal_controls(thread);
2814 
2815   if (kind < kmp_sched_upper_std) {
2816     if (kind == kmp_sched_static && chunk < KMP_DEFAULT_CHUNK) {
2817       // differ static chunked vs. unchunked:  chunk should be invalid to
2818       // indicate unchunked schedule (which is the default)
2819       thread->th.th_current_task->td_icvs.sched.r_sched_type = kmp_sch_static;
2820     } else {
2821       thread->th.th_current_task->td_icvs.sched.r_sched_type =
2822           __kmp_sch_map[kind - kmp_sched_lower - 1];
2823     }
2824   } else {
2825     //    __kmp_sch_map[ kind - kmp_sched_lower_ext + kmp_sched_upper_std -
2826     //    kmp_sched_lower - 2 ];
2827     thread->th.th_current_task->td_icvs.sched.r_sched_type =
2828         __kmp_sch_map[kind - kmp_sched_lower_ext + kmp_sched_upper_std -
2829                       kmp_sched_lower - 2];
2830   }
2831   if (kind == kmp_sched_auto || chunk < 1) {
2832     // ignore parameter chunk for schedule auto
2833     thread->th.th_current_task->td_icvs.sched.chunk = KMP_DEFAULT_CHUNK;
2834   } else {
2835     thread->th.th_current_task->td_icvs.sched.chunk = chunk;
2836   }
2837 }
2838 
2839 /* Gets def_sched_var ICV values */
2840 void __kmp_get_schedule(int gtid, kmp_sched_t *kind, int *chunk) {
2841   kmp_info_t *thread;
2842   enum sched_type th_type;
2843 
2844   KF_TRACE(10, ("__kmp_get_schedule: thread %d\n", gtid));
2845   KMP_DEBUG_ASSERT(__kmp_init_serial);
2846 
2847   thread = __kmp_threads[gtid];
2848 
2849   th_type = thread->th.th_current_task->td_icvs.sched.r_sched_type;
2850 
2851   switch (th_type) {
2852   case kmp_sch_static:
2853   case kmp_sch_static_greedy:
2854   case kmp_sch_static_balanced:
2855     *kind = kmp_sched_static;
2856     *chunk = 0; // chunk was not set, try to show this fact via zero value
2857     return;
2858   case kmp_sch_static_chunked:
2859     *kind = kmp_sched_static;
2860     break;
2861   case kmp_sch_dynamic_chunked:
2862     *kind = kmp_sched_dynamic;
2863     break;
2864   case kmp_sch_guided_chunked:
2865   case kmp_sch_guided_iterative_chunked:
2866   case kmp_sch_guided_analytical_chunked:
2867     *kind = kmp_sched_guided;
2868     break;
2869   case kmp_sch_auto:
2870     *kind = kmp_sched_auto;
2871     break;
2872   case kmp_sch_trapezoidal:
2873     *kind = kmp_sched_trapezoidal;
2874     break;
2875 #if KMP_STATIC_STEAL_ENABLED
2876   case kmp_sch_static_steal:
2877     *kind = kmp_sched_static_steal;
2878     break;
2879 #endif
2880   default:
2881     KMP_FATAL(UnknownSchedulingType, th_type);
2882   }
2883 
2884   *chunk = thread->th.th_current_task->td_icvs.sched.chunk;
2885 }
2886 
2887 int __kmp_get_ancestor_thread_num(int gtid, int level) {
2888 
2889   int ii, dd;
2890   kmp_team_t *team;
2891   kmp_info_t *thr;
2892 
2893   KF_TRACE(10, ("__kmp_get_ancestor_thread_num: thread %d %d\n", gtid, level));
2894   KMP_DEBUG_ASSERT(__kmp_init_serial);
2895 
2896   // validate level
2897   if (level == 0)
2898     return 0;
2899   if (level < 0)
2900     return -1;
2901   thr = __kmp_threads[gtid];
2902   team = thr->th.th_team;
2903   ii = team->t.t_level;
2904   if (level > ii)
2905     return -1;
2906 
2907 #if OMP_40_ENABLED
2908   if (thr->th.th_teams_microtask) {
2909     // AC: we are in teams region where multiple nested teams have same level
2910     int tlevel = thr->th.th_teams_level; // the level of the teams construct
2911     if (level <=
2912         tlevel) { // otherwise usual algorithm works (will not touch the teams)
2913       KMP_DEBUG_ASSERT(ii >= tlevel);
2914       // AC: As we need to pass by the teams league, we need to artificially
2915       // increase ii
2916       if (ii == tlevel) {
2917         ii += 2; // three teams have same level
2918       } else {
2919         ii++; // two teams have same level
2920       }
2921     }
2922   }
2923 #endif
2924 
2925   if (ii == level)
2926     return __kmp_tid_from_gtid(gtid);
2927 
2928   dd = team->t.t_serialized;
2929   level++;
2930   while (ii > level) {
2931     for (dd = team->t.t_serialized; (dd > 0) && (ii > level); dd--, ii--) {
2932     }
2933     if ((team->t.t_serialized) && (!dd)) {
2934       team = team->t.t_parent;
2935       continue;
2936     }
2937     if (ii > level) {
2938       team = team->t.t_parent;
2939       dd = team->t.t_serialized;
2940       ii--;
2941     }
2942   }
2943 
2944   return (dd > 1) ? (0) : (team->t.t_master_tid);
2945 }
2946 
2947 int __kmp_get_team_size(int gtid, int level) {
2948 
2949   int ii, dd;
2950   kmp_team_t *team;
2951   kmp_info_t *thr;
2952 
2953   KF_TRACE(10, ("__kmp_get_team_size: thread %d %d\n", gtid, level));
2954   KMP_DEBUG_ASSERT(__kmp_init_serial);
2955 
2956   // validate level
2957   if (level == 0)
2958     return 1;
2959   if (level < 0)
2960     return -1;
2961   thr = __kmp_threads[gtid];
2962   team = thr->th.th_team;
2963   ii = team->t.t_level;
2964   if (level > ii)
2965     return -1;
2966 
2967 #if OMP_40_ENABLED
2968   if (thr->th.th_teams_microtask) {
2969     // AC: we are in teams region where multiple nested teams have same level
2970     int tlevel = thr->th.th_teams_level; // the level of the teams construct
2971     if (level <=
2972         tlevel) { // otherwise usual algorithm works (will not touch the teams)
2973       KMP_DEBUG_ASSERT(ii >= tlevel);
2974       // AC: As we need to pass by the teams league, we need to artificially
2975       // increase ii
2976       if (ii == tlevel) {
2977         ii += 2; // three teams have same level
2978       } else {
2979         ii++; // two teams have same level
2980       }
2981     }
2982   }
2983 #endif
2984 
2985   while (ii > level) {
2986     for (dd = team->t.t_serialized; (dd > 0) && (ii > level); dd--, ii--) {
2987     }
2988     if (team->t.t_serialized && (!dd)) {
2989       team = team->t.t_parent;
2990       continue;
2991     }
2992     if (ii > level) {
2993       team = team->t.t_parent;
2994       ii--;
2995     }
2996   }
2997 
2998   return team->t.t_nproc;
2999 }
3000 
3001 kmp_r_sched_t __kmp_get_schedule_global() {
3002   // This routine created because pairs (__kmp_sched, __kmp_chunk) and
3003   // (__kmp_static, __kmp_guided) may be changed by kmp_set_defaults
3004   // independently. So one can get the updated schedule here.
3005 
3006   kmp_r_sched_t r_sched;
3007 
3008   // create schedule from 4 globals: __kmp_sched, __kmp_chunk, __kmp_static,
3009   // __kmp_guided. __kmp_sched should keep original value, so that user can set
3010   // KMP_SCHEDULE multiple times, and thus have different run-time schedules in
3011   // different roots (even in OMP 2.5)
3012   if (__kmp_sched == kmp_sch_static) {
3013     // replace STATIC with more detailed schedule (balanced or greedy)
3014     r_sched.r_sched_type = __kmp_static;
3015   } else if (__kmp_sched == kmp_sch_guided_chunked) {
3016     // replace GUIDED with more detailed schedule (iterative or analytical)
3017     r_sched.r_sched_type = __kmp_guided;
3018   } else { // (STATIC_CHUNKED), or (DYNAMIC_CHUNKED), or other
3019     r_sched.r_sched_type = __kmp_sched;
3020   }
3021 
3022   if (__kmp_chunk < KMP_DEFAULT_CHUNK) {
3023     // __kmp_chunk may be wrong here (if it was not ever set)
3024     r_sched.chunk = KMP_DEFAULT_CHUNK;
3025   } else {
3026     r_sched.chunk = __kmp_chunk;
3027   }
3028 
3029   return r_sched;
3030 }
3031 
3032 /* Allocate (realloc == FALSE) * or reallocate (realloc == TRUE)
3033    at least argc number of *t_argv entries for the requested team. */
3034 static void __kmp_alloc_argv_entries(int argc, kmp_team_t *team, int realloc) {
3035 
3036   KMP_DEBUG_ASSERT(team);
3037   if (!realloc || argc > team->t.t_max_argc) {
3038 
3039     KA_TRACE(100, ("__kmp_alloc_argv_entries: team %d: needed entries=%d, "
3040                    "current entries=%d\n",
3041                    team->t.t_id, argc, (realloc) ? team->t.t_max_argc : 0));
3042     /* if previously allocated heap space for args, free them */
3043     if (realloc && team->t.t_argv != &team->t.t_inline_argv[0])
3044       __kmp_free((void *)team->t.t_argv);
3045 
3046     if (argc <= KMP_INLINE_ARGV_ENTRIES) {
3047       /* use unused space in the cache line for arguments */
3048       team->t.t_max_argc = KMP_INLINE_ARGV_ENTRIES;
3049       KA_TRACE(100, ("__kmp_alloc_argv_entries: team %d: inline allocate %d "
3050                      "argv entries\n",
3051                      team->t.t_id, team->t.t_max_argc));
3052       team->t.t_argv = &team->t.t_inline_argv[0];
3053       if (__kmp_storage_map) {
3054         __kmp_print_storage_map_gtid(
3055             -1, &team->t.t_inline_argv[0],
3056             &team->t.t_inline_argv[KMP_INLINE_ARGV_ENTRIES],
3057             (sizeof(void *) * KMP_INLINE_ARGV_ENTRIES), "team_%d.t_inline_argv",
3058             team->t.t_id);
3059       }
3060     } else {
3061       /* allocate space for arguments in the heap */
3062       team->t.t_max_argc = (argc <= (KMP_MIN_MALLOC_ARGV_ENTRIES >> 1))
3063                                ? KMP_MIN_MALLOC_ARGV_ENTRIES
3064                                : 2 * argc;
3065       KA_TRACE(100, ("__kmp_alloc_argv_entries: team %d: dynamic allocate %d "
3066                      "argv entries\n",
3067                      team->t.t_id, team->t.t_max_argc));
3068       team->t.t_argv =
3069           (void **)__kmp_page_allocate(sizeof(void *) * team->t.t_max_argc);
3070       if (__kmp_storage_map) {
3071         __kmp_print_storage_map_gtid(-1, &team->t.t_argv[0],
3072                                      &team->t.t_argv[team->t.t_max_argc],
3073                                      sizeof(void *) * team->t.t_max_argc,
3074                                      "team_%d.t_argv", team->t.t_id);
3075       }
3076     }
3077   }
3078 }
3079 
3080 static void __kmp_allocate_team_arrays(kmp_team_t *team, int max_nth) {
3081   int i;
3082   int num_disp_buff = max_nth > 1 ? __kmp_dispatch_num_buffers : 2;
3083   team->t.t_threads =
3084       (kmp_info_t **)__kmp_allocate(sizeof(kmp_info_t *) * max_nth);
3085   team->t.t_disp_buffer = (dispatch_shared_info_t *)__kmp_allocate(
3086       sizeof(dispatch_shared_info_t) * num_disp_buff);
3087   team->t.t_dispatch =
3088       (kmp_disp_t *)__kmp_allocate(sizeof(kmp_disp_t) * max_nth);
3089   team->t.t_implicit_task_taskdata =
3090       (kmp_taskdata_t *)__kmp_allocate(sizeof(kmp_taskdata_t) * max_nth);
3091   team->t.t_max_nproc = max_nth;
3092 
3093   /* setup dispatch buffers */
3094   for (i = 0; i < num_disp_buff; ++i) {
3095     team->t.t_disp_buffer[i].buffer_index = i;
3096 #if OMP_45_ENABLED
3097     team->t.t_disp_buffer[i].doacross_buf_idx = i;
3098 #endif
3099   }
3100 }
3101 
3102 static void __kmp_free_team_arrays(kmp_team_t *team) {
3103   /* Note: this does not free the threads in t_threads (__kmp_free_threads) */
3104   int i;
3105   for (i = 0; i < team->t.t_max_nproc; ++i) {
3106     if (team->t.t_dispatch[i].th_disp_buffer != NULL) {
3107       __kmp_free(team->t.t_dispatch[i].th_disp_buffer);
3108       team->t.t_dispatch[i].th_disp_buffer = NULL;
3109     }
3110   }
3111 #if KMP_USE_HIER_SCHED
3112   __kmp_dispatch_free_hierarchies(team);
3113 #endif
3114   __kmp_free(team->t.t_threads);
3115   __kmp_free(team->t.t_disp_buffer);
3116   __kmp_free(team->t.t_dispatch);
3117   __kmp_free(team->t.t_implicit_task_taskdata);
3118   team->t.t_threads = NULL;
3119   team->t.t_disp_buffer = NULL;
3120   team->t.t_dispatch = NULL;
3121   team->t.t_implicit_task_taskdata = 0;
3122 }
3123 
3124 static void __kmp_reallocate_team_arrays(kmp_team_t *team, int max_nth) {
3125   kmp_info_t **oldThreads = team->t.t_threads;
3126 
3127   __kmp_free(team->t.t_disp_buffer);
3128   __kmp_free(team->t.t_dispatch);
3129   __kmp_free(team->t.t_implicit_task_taskdata);
3130   __kmp_allocate_team_arrays(team, max_nth);
3131 
3132   KMP_MEMCPY(team->t.t_threads, oldThreads,
3133              team->t.t_nproc * sizeof(kmp_info_t *));
3134 
3135   __kmp_free(oldThreads);
3136 }
3137 
3138 static kmp_internal_control_t __kmp_get_global_icvs(void) {
3139 
3140   kmp_r_sched_t r_sched =
3141       __kmp_get_schedule_global(); // get current state of scheduling globals
3142 
3143 #if OMP_40_ENABLED
3144   KMP_DEBUG_ASSERT(__kmp_nested_proc_bind.used > 0);
3145 #endif /* OMP_40_ENABLED */
3146 
3147   kmp_internal_control_t g_icvs = {
3148     0, // int serial_nesting_level; //corresponds to value of th_team_serialized
3149     (kmp_int8)__kmp_global.g.g_dynamic, // internal control for dynamic
3150     // adjustment of threads (per thread)
3151     (kmp_int8)__kmp_env_blocktime, // int bt_set; //internal control for
3152     // whether blocktime is explicitly set
3153     __kmp_dflt_blocktime, // int blocktime; //internal control for blocktime
3154 #if KMP_USE_MONITOR
3155     __kmp_bt_intervals, // int bt_intervals; //internal control for blocktime
3156 // intervals
3157 #endif
3158     __kmp_dflt_team_nth, // int nproc; //internal control for # of threads for
3159     // next parallel region (per thread)
3160     // (use a max ub on value if __kmp_parallel_initialize not called yet)
3161     __kmp_cg_max_nth, // int thread_limit;
3162     __kmp_dflt_max_active_levels, // int max_active_levels; //internal control
3163     // for max_active_levels
3164     r_sched, // kmp_r_sched_t sched; //internal control for runtime schedule
3165 // {sched,chunk} pair
3166 #if OMP_40_ENABLED
3167     __kmp_nested_proc_bind.bind_types[0],
3168     __kmp_default_device,
3169 #endif /* OMP_40_ENABLED */
3170     NULL // struct kmp_internal_control *next;
3171   };
3172 
3173   return g_icvs;
3174 }
3175 
3176 static kmp_internal_control_t __kmp_get_x_global_icvs(const kmp_team_t *team) {
3177 
3178   kmp_internal_control_t gx_icvs;
3179   gx_icvs.serial_nesting_level =
3180       0; // probably =team->t.t_serial like in save_inter_controls
3181   copy_icvs(&gx_icvs, &team->t.t_threads[0]->th.th_current_task->td_icvs);
3182   gx_icvs.next = NULL;
3183 
3184   return gx_icvs;
3185 }
3186 
3187 static void __kmp_initialize_root(kmp_root_t *root) {
3188   int f;
3189   kmp_team_t *root_team;
3190   kmp_team_t *hot_team;
3191   int hot_team_max_nth;
3192   kmp_r_sched_t r_sched =
3193       __kmp_get_schedule_global(); // get current state of scheduling globals
3194   kmp_internal_control_t r_icvs = __kmp_get_global_icvs();
3195   KMP_DEBUG_ASSERT(root);
3196   KMP_ASSERT(!root->r.r_begin);
3197 
3198   /* setup the root state structure */
3199   __kmp_init_lock(&root->r.r_begin_lock);
3200   root->r.r_begin = FALSE;
3201   root->r.r_active = FALSE;
3202   root->r.r_in_parallel = 0;
3203   root->r.r_blocktime = __kmp_dflt_blocktime;
3204 
3205   /* setup the root team for this task */
3206   /* allocate the root team structure */
3207   KF_TRACE(10, ("__kmp_initialize_root: before root_team\n"));
3208 
3209   root_team =
3210       __kmp_allocate_team(root,
3211                           1, // new_nproc
3212                           1, // max_nproc
3213 #if OMPT_SUPPORT
3214                           ompt_data_none, // root parallel id
3215 #endif
3216 #if OMP_40_ENABLED
3217                           __kmp_nested_proc_bind.bind_types[0],
3218 #endif
3219                           &r_icvs,
3220                           0 // argc
3221                           USE_NESTED_HOT_ARG(NULL) // master thread is unknown
3222                           );
3223 #if USE_DEBUGGER
3224   // Non-NULL value should be assigned to make the debugger display the root
3225   // team.
3226   TCW_SYNC_PTR(root_team->t.t_pkfn, (microtask_t)(~0));
3227 #endif
3228 
3229   KF_TRACE(10, ("__kmp_initialize_root: after root_team = %p\n", root_team));
3230 
3231   root->r.r_root_team = root_team;
3232   root_team->t.t_control_stack_top = NULL;
3233 
3234   /* initialize root team */
3235   root_team->t.t_threads[0] = NULL;
3236   root_team->t.t_nproc = 1;
3237   root_team->t.t_serialized = 1;
3238   // TODO???: root_team->t.t_max_active_levels = __kmp_dflt_max_active_levels;
3239   root_team->t.t_sched.sched = r_sched.sched;
3240   KA_TRACE(
3241       20,
3242       ("__kmp_initialize_root: init root team %d arrived: join=%u, plain=%u\n",
3243        root_team->t.t_id, KMP_INIT_BARRIER_STATE, KMP_INIT_BARRIER_STATE));
3244 
3245   /* setup the  hot team for this task */
3246   /* allocate the hot team structure */
3247   KF_TRACE(10, ("__kmp_initialize_root: before hot_team\n"));
3248 
3249   hot_team =
3250       __kmp_allocate_team(root,
3251                           1, // new_nproc
3252                           __kmp_dflt_team_nth_ub * 2, // max_nproc
3253 #if OMPT_SUPPORT
3254                           ompt_data_none, // root parallel id
3255 #endif
3256 #if OMP_40_ENABLED
3257                           __kmp_nested_proc_bind.bind_types[0],
3258 #endif
3259                           &r_icvs,
3260                           0 // argc
3261                           USE_NESTED_HOT_ARG(NULL) // master thread is unknown
3262                           );
3263   KF_TRACE(10, ("__kmp_initialize_root: after hot_team = %p\n", hot_team));
3264 
3265   root->r.r_hot_team = hot_team;
3266   root_team->t.t_control_stack_top = NULL;
3267 
3268   /* first-time initialization */
3269   hot_team->t.t_parent = root_team;
3270 
3271   /* initialize hot team */
3272   hot_team_max_nth = hot_team->t.t_max_nproc;
3273   for (f = 0; f < hot_team_max_nth; ++f) {
3274     hot_team->t.t_threads[f] = NULL;
3275   }
3276   hot_team->t.t_nproc = 1;
3277   // TODO???: hot_team->t.t_max_active_levels = __kmp_dflt_max_active_levels;
3278   hot_team->t.t_sched.sched = r_sched.sched;
3279   hot_team->t.t_size_changed = 0;
3280 }
3281 
3282 #ifdef KMP_DEBUG
3283 
3284 typedef struct kmp_team_list_item {
3285   kmp_team_p const *entry;
3286   struct kmp_team_list_item *next;
3287 } kmp_team_list_item_t;
3288 typedef kmp_team_list_item_t *kmp_team_list_t;
3289 
3290 static void __kmp_print_structure_team_accum( // Add team to list of teams.
3291     kmp_team_list_t list, // List of teams.
3292     kmp_team_p const *team // Team to add.
3293     ) {
3294 
3295   // List must terminate with item where both entry and next are NULL.
3296   // Team is added to the list only once.
3297   // List is sorted in ascending order by team id.
3298   // Team id is *not* a key.
3299 
3300   kmp_team_list_t l;
3301 
3302   KMP_DEBUG_ASSERT(list != NULL);
3303   if (team == NULL) {
3304     return;
3305   }
3306 
3307   __kmp_print_structure_team_accum(list, team->t.t_parent);
3308   __kmp_print_structure_team_accum(list, team->t.t_next_pool);
3309 
3310   // Search list for the team.
3311   l = list;
3312   while (l->next != NULL && l->entry != team) {
3313     l = l->next;
3314   }
3315   if (l->next != NULL) {
3316     return; // Team has been added before, exit.
3317   }
3318 
3319   // Team is not found. Search list again for insertion point.
3320   l = list;
3321   while (l->next != NULL && l->entry->t.t_id <= team->t.t_id) {
3322     l = l->next;
3323   }
3324 
3325   // Insert team.
3326   {
3327     kmp_team_list_item_t *item = (kmp_team_list_item_t *)KMP_INTERNAL_MALLOC(
3328         sizeof(kmp_team_list_item_t));
3329     *item = *l;
3330     l->entry = team;
3331     l->next = item;
3332   }
3333 }
3334 
3335 static void __kmp_print_structure_team(char const *title, kmp_team_p const *team
3336 
3337                                        ) {
3338   __kmp_printf("%s", title);
3339   if (team != NULL) {
3340     __kmp_printf("%2x %p\n", team->t.t_id, team);
3341   } else {
3342     __kmp_printf(" - (nil)\n");
3343   }
3344 }
3345 
3346 static void __kmp_print_structure_thread(char const *title,
3347                                          kmp_info_p const *thread) {
3348   __kmp_printf("%s", title);
3349   if (thread != NULL) {
3350     __kmp_printf("%2d %p\n", thread->th.th_info.ds.ds_gtid, thread);
3351   } else {
3352     __kmp_printf(" - (nil)\n");
3353   }
3354 }
3355 
3356 void __kmp_print_structure(void) {
3357 
3358   kmp_team_list_t list;
3359 
3360   // Initialize list of teams.
3361   list =
3362       (kmp_team_list_item_t *)KMP_INTERNAL_MALLOC(sizeof(kmp_team_list_item_t));
3363   list->entry = NULL;
3364   list->next = NULL;
3365 
3366   __kmp_printf("\n------------------------------\nGlobal Thread "
3367                "Table\n------------------------------\n");
3368   {
3369     int gtid;
3370     for (gtid = 0; gtid < __kmp_threads_capacity; ++gtid) {
3371       __kmp_printf("%2d", gtid);
3372       if (__kmp_threads != NULL) {
3373         __kmp_printf(" %p", __kmp_threads[gtid]);
3374       }
3375       if (__kmp_root != NULL) {
3376         __kmp_printf(" %p", __kmp_root[gtid]);
3377       }
3378       __kmp_printf("\n");
3379     }
3380   }
3381 
3382   // Print out __kmp_threads array.
3383   __kmp_printf("\n------------------------------\nThreads\n--------------------"
3384                "----------\n");
3385   if (__kmp_threads != NULL) {
3386     int gtid;
3387     for (gtid = 0; gtid < __kmp_threads_capacity; ++gtid) {
3388       kmp_info_t const *thread = __kmp_threads[gtid];
3389       if (thread != NULL) {
3390         __kmp_printf("GTID %2d %p:\n", gtid, thread);
3391         __kmp_printf("    Our Root:        %p\n", thread->th.th_root);
3392         __kmp_print_structure_team("    Our Team:     ", thread->th.th_team);
3393         __kmp_print_structure_team("    Serial Team:  ",
3394                                    thread->th.th_serial_team);
3395         __kmp_printf("    Threads:      %2d\n", thread->th.th_team_nproc);
3396         __kmp_print_structure_thread("    Master:       ",
3397                                      thread->th.th_team_master);
3398         __kmp_printf("    Serialized?:  %2d\n", thread->th.th_team_serialized);
3399         __kmp_printf("    Set NProc:    %2d\n", thread->th.th_set_nproc);
3400 #if OMP_40_ENABLED
3401         __kmp_printf("    Set Proc Bind: %2d\n", thread->th.th_set_proc_bind);
3402 #endif
3403         __kmp_print_structure_thread("    Next in pool: ",
3404                                      thread->th.th_next_pool);
3405         __kmp_printf("\n");
3406         __kmp_print_structure_team_accum(list, thread->th.th_team);
3407         __kmp_print_structure_team_accum(list, thread->th.th_serial_team);
3408       }
3409     }
3410   } else {
3411     __kmp_printf("Threads array is not allocated.\n");
3412   }
3413 
3414   // Print out __kmp_root array.
3415   __kmp_printf("\n------------------------------\nUbers\n----------------------"
3416                "--------\n");
3417   if (__kmp_root != NULL) {
3418     int gtid;
3419     for (gtid = 0; gtid < __kmp_threads_capacity; ++gtid) {
3420       kmp_root_t const *root = __kmp_root[gtid];
3421       if (root != NULL) {
3422         __kmp_printf("GTID %2d %p:\n", gtid, root);
3423         __kmp_print_structure_team("    Root Team:    ", root->r.r_root_team);
3424         __kmp_print_structure_team("    Hot Team:     ", root->r.r_hot_team);
3425         __kmp_print_structure_thread("    Uber Thread:  ",
3426                                      root->r.r_uber_thread);
3427         __kmp_printf("    Active?:      %2d\n", root->r.r_active);
3428         __kmp_printf("    In Parallel:  %2d\n",
3429                      KMP_ATOMIC_LD_RLX(&root->r.r_in_parallel));
3430         __kmp_printf("\n");
3431         __kmp_print_structure_team_accum(list, root->r.r_root_team);
3432         __kmp_print_structure_team_accum(list, root->r.r_hot_team);
3433       }
3434     }
3435   } else {
3436     __kmp_printf("Ubers array is not allocated.\n");
3437   }
3438 
3439   __kmp_printf("\n------------------------------\nTeams\n----------------------"
3440                "--------\n");
3441   while (list->next != NULL) {
3442     kmp_team_p const *team = list->entry;
3443     int i;
3444     __kmp_printf("Team %2x %p:\n", team->t.t_id, team);
3445     __kmp_print_structure_team("    Parent Team:      ", team->t.t_parent);
3446     __kmp_printf("    Master TID:       %2d\n", team->t.t_master_tid);
3447     __kmp_printf("    Max threads:      %2d\n", team->t.t_max_nproc);
3448     __kmp_printf("    Levels of serial: %2d\n", team->t.t_serialized);
3449     __kmp_printf("    Number threads:   %2d\n", team->t.t_nproc);
3450     for (i = 0; i < team->t.t_nproc; ++i) {
3451       __kmp_printf("    Thread %2d:      ", i);
3452       __kmp_print_structure_thread("", team->t.t_threads[i]);
3453     }
3454     __kmp_print_structure_team("    Next in pool:     ", team->t.t_next_pool);
3455     __kmp_printf("\n");
3456     list = list->next;
3457   }
3458 
3459   // Print out __kmp_thread_pool and __kmp_team_pool.
3460   __kmp_printf("\n------------------------------\nPools\n----------------------"
3461                "--------\n");
3462   __kmp_print_structure_thread("Thread pool:          ",
3463                                CCAST(kmp_info_t *, __kmp_thread_pool));
3464   __kmp_print_structure_team("Team pool:            ",
3465                              CCAST(kmp_team_t *, __kmp_team_pool));
3466   __kmp_printf("\n");
3467 
3468   // Free team list.
3469   while (list != NULL) {
3470     kmp_team_list_item_t *item = list;
3471     list = list->next;
3472     KMP_INTERNAL_FREE(item);
3473   }
3474 }
3475 
3476 #endif
3477 
3478 //---------------------------------------------------------------------------
3479 //  Stuff for per-thread fast random number generator
3480 //  Table of primes
3481 static const unsigned __kmp_primes[] = {
3482     0x9e3779b1, 0xffe6cc59, 0x2109f6dd, 0x43977ab5, 0xba5703f5, 0xb495a877,
3483     0xe1626741, 0x79695e6b, 0xbc98c09f, 0xd5bee2b3, 0x287488f9, 0x3af18231,
3484     0x9677cd4d, 0xbe3a6929, 0xadc6a877, 0xdcf0674b, 0xbe4d6fe9, 0x5f15e201,
3485     0x99afc3fd, 0xf3f16801, 0xe222cfff, 0x24ba5fdb, 0x0620452d, 0x79f149e3,
3486     0xc8b93f49, 0x972702cd, 0xb07dd827, 0x6c97d5ed, 0x085a3d61, 0x46eb5ea7,
3487     0x3d9910ed, 0x2e687b5b, 0x29609227, 0x6eb081f1, 0x0954c4e1, 0x9d114db9,
3488     0x542acfa9, 0xb3e6bd7b, 0x0742d917, 0xe9f3ffa7, 0x54581edb, 0xf2480f45,
3489     0x0bb9288f, 0xef1affc7, 0x85fa0ca7, 0x3ccc14db, 0xe6baf34b, 0x343377f7,
3490     0x5ca19031, 0xe6d9293b, 0xf0a9f391, 0x5d2e980b, 0xfc411073, 0xc3749363,
3491     0xb892d829, 0x3549366b, 0x629750ad, 0xb98294e5, 0x892d9483, 0xc235baf3,
3492     0x3d2402a3, 0x6bdef3c9, 0xbec333cd, 0x40c9520f};
3493 
3494 //---------------------------------------------------------------------------
3495 //  __kmp_get_random: Get a random number using a linear congruential method.
3496 unsigned short __kmp_get_random(kmp_info_t *thread) {
3497   unsigned x = thread->th.th_x;
3498   unsigned short r = x >> 16;
3499 
3500   thread->th.th_x = x * thread->th.th_a + 1;
3501 
3502   KA_TRACE(30, ("__kmp_get_random: THREAD: %d, RETURN: %u\n",
3503                 thread->th.th_info.ds.ds_tid, r));
3504 
3505   return r;
3506 }
3507 //--------------------------------------------------------
3508 // __kmp_init_random: Initialize a random number generator
3509 void __kmp_init_random(kmp_info_t *thread) {
3510   unsigned seed = thread->th.th_info.ds.ds_tid;
3511 
3512   thread->th.th_a =
3513       __kmp_primes[seed % (sizeof(__kmp_primes) / sizeof(__kmp_primes[0]))];
3514   thread->th.th_x = (seed + 1) * thread->th.th_a + 1;
3515   KA_TRACE(30,
3516            ("__kmp_init_random: THREAD: %u; A: %u\n", seed, thread->th.th_a));
3517 }
3518 
3519 #if KMP_OS_WINDOWS
3520 /* reclaim array entries for root threads that are already dead, returns number
3521  * reclaimed */
3522 static int __kmp_reclaim_dead_roots(void) {
3523   int i, r = 0;
3524 
3525   for (i = 0; i < __kmp_threads_capacity; ++i) {
3526     if (KMP_UBER_GTID(i) &&
3527         !__kmp_still_running((kmp_info_t *)TCR_SYNC_PTR(__kmp_threads[i])) &&
3528         !__kmp_root[i]
3529              ->r.r_active) { // AC: reclaim only roots died in non-active state
3530       r += __kmp_unregister_root_other_thread(i);
3531     }
3532   }
3533   return r;
3534 }
3535 #endif
3536 
3537 /* This function attempts to create free entries in __kmp_threads and
3538    __kmp_root, and returns the number of free entries generated.
3539 
3540    For Windows* OS static library, the first mechanism used is to reclaim array
3541    entries for root threads that are already dead.
3542 
3543    On all platforms, expansion is attempted on the arrays __kmp_threads_ and
3544    __kmp_root, with appropriate update to __kmp_threads_capacity. Array
3545    capacity is increased by doubling with clipping to __kmp_tp_capacity, if
3546    threadprivate cache array has been created. Synchronization with
3547    __kmpc_threadprivate_cached is done using __kmp_tp_cached_lock.
3548 
3549    After any dead root reclamation, if the clipping value allows array expansion
3550    to result in the generation of a total of nNeed free slots, the function does
3551    that expansion. If not, nothing is done beyond the possible initial root
3552    thread reclamation.
3553 
3554    If any argument is negative, the behavior is undefined. */
3555 static int __kmp_expand_threads(int nNeed) {
3556   int added = 0;
3557   int minimumRequiredCapacity;
3558   int newCapacity;
3559   kmp_info_t **newThreads;
3560   kmp_root_t **newRoot;
3561 
3562 // All calls to __kmp_expand_threads should be under __kmp_forkjoin_lock, so
3563 // resizing __kmp_threads does not need additional protection if foreign
3564 // threads are present
3565 
3566 #if KMP_OS_WINDOWS && !KMP_DYNAMIC_LIB
3567   /* only for Windows static library */
3568   /* reclaim array entries for root threads that are already dead */
3569   added = __kmp_reclaim_dead_roots();
3570 
3571   if (nNeed) {
3572     nNeed -= added;
3573     if (nNeed < 0)
3574       nNeed = 0;
3575   }
3576 #endif
3577   if (nNeed <= 0)
3578     return added;
3579 
3580   // Note that __kmp_threads_capacity is not bounded by __kmp_max_nth. If
3581   // __kmp_max_nth is set to some value less than __kmp_sys_max_nth by the
3582   // user via KMP_DEVICE_THREAD_LIMIT, then __kmp_threads_capacity may become
3583   // > __kmp_max_nth in one of two ways:
3584   //
3585   // 1) The initialization thread (gtid = 0) exits.  __kmp_threads[0]
3586   //    may not be resused by another thread, so we may need to increase
3587   //    __kmp_threads_capacity to __kmp_max_nth + 1.
3588   //
3589   // 2) New foreign root(s) are encountered.  We always register new foreign
3590   //    roots. This may cause a smaller # of threads to be allocated at
3591   //    subsequent parallel regions, but the worker threads hang around (and
3592   //    eventually go to sleep) and need slots in the __kmp_threads[] array.
3593   //
3594   // Anyway, that is the reason for moving the check to see if
3595   // __kmp_max_nth was exceeded into __kmp_reserve_threads()
3596   // instead of having it performed here. -BB
3597 
3598   KMP_DEBUG_ASSERT(__kmp_sys_max_nth >= __kmp_threads_capacity);
3599 
3600   /* compute expansion headroom to check if we can expand */
3601   if (__kmp_sys_max_nth - __kmp_threads_capacity < nNeed) {
3602     /* possible expansion too small -- give up */
3603     return added;
3604   }
3605   minimumRequiredCapacity = __kmp_threads_capacity + nNeed;
3606 
3607   newCapacity = __kmp_threads_capacity;
3608   do {
3609     newCapacity = newCapacity <= (__kmp_sys_max_nth >> 1) ? (newCapacity << 1)
3610                                                           : __kmp_sys_max_nth;
3611   } while (newCapacity < minimumRequiredCapacity);
3612   newThreads = (kmp_info_t **)__kmp_allocate(
3613       (sizeof(kmp_info_t *) + sizeof(kmp_root_t *)) * newCapacity + CACHE_LINE);
3614   newRoot =
3615       (kmp_root_t **)((char *)newThreads + sizeof(kmp_info_t *) * newCapacity);
3616   KMP_MEMCPY(newThreads, __kmp_threads,
3617              __kmp_threads_capacity * sizeof(kmp_info_t *));
3618   KMP_MEMCPY(newRoot, __kmp_root,
3619              __kmp_threads_capacity * sizeof(kmp_root_t *));
3620 
3621   kmp_info_t **temp_threads = __kmp_threads;
3622   *(kmp_info_t * *volatile *)&__kmp_threads = newThreads;
3623   *(kmp_root_t * *volatile *)&__kmp_root = newRoot;
3624   __kmp_free(temp_threads);
3625   added += newCapacity - __kmp_threads_capacity;
3626   *(volatile int *)&__kmp_threads_capacity = newCapacity;
3627 
3628   if (newCapacity > __kmp_tp_capacity) {
3629     __kmp_acquire_bootstrap_lock(&__kmp_tp_cached_lock);
3630     if (__kmp_tp_cached && newCapacity > __kmp_tp_capacity) {
3631       __kmp_threadprivate_resize_cache(newCapacity);
3632     } else { // increase __kmp_tp_capacity to correspond with kmp_threads size
3633       *(volatile int *)&__kmp_tp_capacity = newCapacity;
3634     }
3635     __kmp_release_bootstrap_lock(&__kmp_tp_cached_lock);
3636   }
3637 
3638   return added;
3639 }
3640 
3641 /* Register the current thread as a root thread and obtain our gtid. We must
3642    have the __kmp_initz_lock held at this point. Argument TRUE only if are the
3643    thread that calls from __kmp_do_serial_initialize() */
3644 int __kmp_register_root(int initial_thread) {
3645   kmp_info_t *root_thread;
3646   kmp_root_t *root;
3647   int gtid;
3648   int capacity;
3649   __kmp_acquire_bootstrap_lock(&__kmp_forkjoin_lock);
3650   KA_TRACE(20, ("__kmp_register_root: entered\n"));
3651   KMP_MB();
3652 
3653   /* 2007-03-02:
3654      If initial thread did not invoke OpenMP RTL yet, and this thread is not an
3655      initial one, "__kmp_all_nth >= __kmp_threads_capacity" condition does not
3656      work as expected -- it may return false (that means there is at least one
3657      empty slot in __kmp_threads array), but it is possible the only free slot
3658      is #0, which is reserved for initial thread and so cannot be used for this
3659      one. Following code workarounds this bug.
3660 
3661      However, right solution seems to be not reserving slot #0 for initial
3662      thread because:
3663      (1) there is no magic in slot #0,
3664      (2) we cannot detect initial thread reliably (the first thread which does
3665         serial initialization may be not a real initial thread).
3666   */
3667   capacity = __kmp_threads_capacity;
3668   if (!initial_thread && TCR_PTR(__kmp_threads[0]) == NULL) {
3669     --capacity;
3670   }
3671 
3672   /* see if there are too many threads */
3673   if (__kmp_all_nth >= capacity && !__kmp_expand_threads(1)) {
3674     if (__kmp_tp_cached) {
3675       __kmp_fatal(KMP_MSG(CantRegisterNewThread),
3676                   KMP_HNT(Set_ALL_THREADPRIVATE, __kmp_tp_capacity),
3677                   KMP_HNT(PossibleSystemLimitOnThreads), __kmp_msg_null);
3678     } else {
3679       __kmp_fatal(KMP_MSG(CantRegisterNewThread), KMP_HNT(SystemLimitOnThreads),
3680                   __kmp_msg_null);
3681     }
3682   }
3683 
3684   /* find an available thread slot */
3685   /* Don't reassign the zero slot since we need that to only be used by initial
3686      thread */
3687   for (gtid = (initial_thread ? 0 : 1); TCR_PTR(__kmp_threads[gtid]) != NULL;
3688        gtid++)
3689     ;
3690   KA_TRACE(1,
3691            ("__kmp_register_root: found slot in threads array: T#%d\n", gtid));
3692   KMP_ASSERT(gtid < __kmp_threads_capacity);
3693 
3694   /* update global accounting */
3695   __kmp_all_nth++;
3696   TCW_4(__kmp_nth, __kmp_nth + 1);
3697 
3698   // if __kmp_adjust_gtid_mode is set, then we use method #1 (sp search) for low
3699   // numbers of procs, and method #2 (keyed API call) for higher numbers.
3700   if (__kmp_adjust_gtid_mode) {
3701     if (__kmp_all_nth >= __kmp_tls_gtid_min) {
3702       if (TCR_4(__kmp_gtid_mode) != 2) {
3703         TCW_4(__kmp_gtid_mode, 2);
3704       }
3705     } else {
3706       if (TCR_4(__kmp_gtid_mode) != 1) {
3707         TCW_4(__kmp_gtid_mode, 1);
3708       }
3709     }
3710   }
3711 
3712 #ifdef KMP_ADJUST_BLOCKTIME
3713   /* Adjust blocktime to zero if necessary            */
3714   /* Middle initialization might not have occurred yet */
3715   if (!__kmp_env_blocktime && (__kmp_avail_proc > 0)) {
3716     if (__kmp_nth > __kmp_avail_proc) {
3717       __kmp_zero_bt = TRUE;
3718     }
3719   }
3720 #endif /* KMP_ADJUST_BLOCKTIME */
3721 
3722   /* setup this new hierarchy */
3723   if (!(root = __kmp_root[gtid])) {
3724     root = __kmp_root[gtid] = (kmp_root_t *)__kmp_allocate(sizeof(kmp_root_t));
3725     KMP_DEBUG_ASSERT(!root->r.r_root_team);
3726   }
3727 
3728 #if KMP_STATS_ENABLED
3729   // Initialize stats as soon as possible (right after gtid assignment).
3730   __kmp_stats_thread_ptr = __kmp_stats_list->push_back(gtid);
3731   __kmp_stats_thread_ptr->startLife();
3732   KMP_SET_THREAD_STATE(SERIAL_REGION);
3733   KMP_INIT_PARTITIONED_TIMERS(OMP_serial);
3734 #endif
3735   __kmp_initialize_root(root);
3736 
3737   /* setup new root thread structure */
3738   if (root->r.r_uber_thread) {
3739     root_thread = root->r.r_uber_thread;
3740   } else {
3741     root_thread = (kmp_info_t *)__kmp_allocate(sizeof(kmp_info_t));
3742     if (__kmp_storage_map) {
3743       __kmp_print_thread_storage_map(root_thread, gtid);
3744     }
3745     root_thread->th.th_info.ds.ds_gtid = gtid;
3746 #if OMPT_SUPPORT
3747     root_thread->th.ompt_thread_info.thread_data = ompt_data_none;
3748 #endif
3749     root_thread->th.th_root = root;
3750     if (__kmp_env_consistency_check) {
3751       root_thread->th.th_cons = __kmp_allocate_cons_stack(gtid);
3752     }
3753 #if USE_FAST_MEMORY
3754     __kmp_initialize_fast_memory(root_thread);
3755 #endif /* USE_FAST_MEMORY */
3756 
3757 #if KMP_USE_BGET
3758     KMP_DEBUG_ASSERT(root_thread->th.th_local.bget_data == NULL);
3759     __kmp_initialize_bget(root_thread);
3760 #endif
3761     __kmp_init_random(root_thread); // Initialize random number generator
3762   }
3763 
3764   /* setup the serial team held in reserve by the root thread */
3765   if (!root_thread->th.th_serial_team) {
3766     kmp_internal_control_t r_icvs = __kmp_get_global_icvs();
3767     KF_TRACE(10, ("__kmp_register_root: before serial_team\n"));
3768     root_thread->th.th_serial_team =
3769         __kmp_allocate_team(root, 1, 1,
3770 #if OMPT_SUPPORT
3771                             ompt_data_none, // root parallel id
3772 #endif
3773 #if OMP_40_ENABLED
3774                             proc_bind_default,
3775 #endif
3776                             &r_icvs, 0 USE_NESTED_HOT_ARG(NULL));
3777   }
3778   KMP_ASSERT(root_thread->th.th_serial_team);
3779   KF_TRACE(10, ("__kmp_register_root: after serial_team = %p\n",
3780                 root_thread->th.th_serial_team));
3781 
3782   /* drop root_thread into place */
3783   TCW_SYNC_PTR(__kmp_threads[gtid], root_thread);
3784 
3785   root->r.r_root_team->t.t_threads[0] = root_thread;
3786   root->r.r_hot_team->t.t_threads[0] = root_thread;
3787   root_thread->th.th_serial_team->t.t_threads[0] = root_thread;
3788   // AC: the team created in reserve, not for execution (it is unused for now).
3789   root_thread->th.th_serial_team->t.t_serialized = 0;
3790   root->r.r_uber_thread = root_thread;
3791 
3792   /* initialize the thread, get it ready to go */
3793   __kmp_initialize_info(root_thread, root->r.r_root_team, 0, gtid);
3794   TCW_4(__kmp_init_gtid, TRUE);
3795 
3796   /* prepare the master thread for get_gtid() */
3797   __kmp_gtid_set_specific(gtid);
3798 
3799 #if USE_ITT_BUILD
3800   __kmp_itt_thread_name(gtid);
3801 #endif /* USE_ITT_BUILD */
3802 
3803 #ifdef KMP_TDATA_GTID
3804   __kmp_gtid = gtid;
3805 #endif
3806   __kmp_create_worker(gtid, root_thread, __kmp_stksize);
3807   KMP_DEBUG_ASSERT(__kmp_gtid_get_specific() == gtid);
3808 
3809   KA_TRACE(20, ("__kmp_register_root: T#%d init T#%d(%d:%d) arrived: join=%u, "
3810                 "plain=%u\n",
3811                 gtid, __kmp_gtid_from_tid(0, root->r.r_hot_team),
3812                 root->r.r_hot_team->t.t_id, 0, KMP_INIT_BARRIER_STATE,
3813                 KMP_INIT_BARRIER_STATE));
3814   { // Initialize barrier data.
3815     int b;
3816     for (b = 0; b < bs_last_barrier; ++b) {
3817       root_thread->th.th_bar[b].bb.b_arrived = KMP_INIT_BARRIER_STATE;
3818 #if USE_DEBUGGER
3819       root_thread->th.th_bar[b].bb.b_worker_arrived = 0;
3820 #endif
3821     }
3822   }
3823   KMP_DEBUG_ASSERT(root->r.r_hot_team->t.t_bar[bs_forkjoin_barrier].b_arrived ==
3824                    KMP_INIT_BARRIER_STATE);
3825 
3826 #if KMP_AFFINITY_SUPPORTED
3827 #if OMP_40_ENABLED
3828   root_thread->th.th_current_place = KMP_PLACE_UNDEFINED;
3829   root_thread->th.th_new_place = KMP_PLACE_UNDEFINED;
3830   root_thread->th.th_first_place = KMP_PLACE_UNDEFINED;
3831   root_thread->th.th_last_place = KMP_PLACE_UNDEFINED;
3832 #endif
3833   if (TCR_4(__kmp_init_middle)) {
3834     __kmp_affinity_set_init_mask(gtid, TRUE);
3835   }
3836 #endif /* KMP_AFFINITY_SUPPORTED */
3837 #if OMP_50_ENABLED
3838   root_thread->th.th_def_allocator = __kmp_def_allocator;
3839   root_thread->th.th_prev_level = 0;
3840   root_thread->th.th_prev_num_threads = 1;
3841 #endif
3842 
3843   kmp_cg_root_t *tmp = (kmp_cg_root_t *)__kmp_allocate(sizeof(kmp_cg_root_t));
3844   tmp->cg_root = root_thread;
3845   tmp->cg_thread_limit = __kmp_cg_max_nth;
3846   tmp->cg_nthreads = 1;
3847   KA_TRACE(100, ("__kmp_register_root: Thread %p created node %p with"
3848                  " cg_nthreads init to 1\n",
3849                  root_thread, tmp));
3850   tmp->up = NULL;
3851   root_thread->th.th_cg_roots = tmp;
3852 
3853   __kmp_root_counter++;
3854 
3855 #if OMPT_SUPPORT
3856   if (!initial_thread && ompt_enabled.enabled) {
3857 
3858     kmp_info_t *root_thread = ompt_get_thread();
3859 
3860     ompt_set_thread_state(root_thread, ompt_state_overhead);
3861 
3862     if (ompt_enabled.ompt_callback_thread_begin) {
3863       ompt_callbacks.ompt_callback(ompt_callback_thread_begin)(
3864           ompt_thread_initial, __ompt_get_thread_data_internal());
3865     }
3866     ompt_data_t *task_data;
3867     __ompt_get_task_info_internal(0, NULL, &task_data, NULL, NULL, NULL);
3868     if (ompt_enabled.ompt_callback_task_create) {
3869       ompt_callbacks.ompt_callback(ompt_callback_task_create)(
3870           NULL, NULL, task_data, ompt_task_initial, 0, NULL);
3871       // initial task has nothing to return to
3872     }
3873 
3874     ompt_set_thread_state(root_thread, ompt_state_work_serial);
3875   }
3876 #endif
3877 
3878   KMP_MB();
3879   __kmp_release_bootstrap_lock(&__kmp_forkjoin_lock);
3880 
3881   return gtid;
3882 }
3883 
3884 #if KMP_NESTED_HOT_TEAMS
3885 static int __kmp_free_hot_teams(kmp_root_t *root, kmp_info_t *thr, int level,
3886                                 const int max_level) {
3887   int i, n, nth;
3888   kmp_hot_team_ptr_t *hot_teams = thr->th.th_hot_teams;
3889   if (!hot_teams || !hot_teams[level].hot_team) {
3890     return 0;
3891   }
3892   KMP_DEBUG_ASSERT(level < max_level);
3893   kmp_team_t *team = hot_teams[level].hot_team;
3894   nth = hot_teams[level].hot_team_nth;
3895   n = nth - 1; // master is not freed
3896   if (level < max_level - 1) {
3897     for (i = 0; i < nth; ++i) {
3898       kmp_info_t *th = team->t.t_threads[i];
3899       n += __kmp_free_hot_teams(root, th, level + 1, max_level);
3900       if (i > 0 && th->th.th_hot_teams) {
3901         __kmp_free(th->th.th_hot_teams);
3902         th->th.th_hot_teams = NULL;
3903       }
3904     }
3905   }
3906   __kmp_free_team(root, team, NULL);
3907   return n;
3908 }
3909 #endif
3910 
3911 // Resets a root thread and clear its root and hot teams.
3912 // Returns the number of __kmp_threads entries directly and indirectly freed.
3913 static int __kmp_reset_root(int gtid, kmp_root_t *root) {
3914   kmp_team_t *root_team = root->r.r_root_team;
3915   kmp_team_t *hot_team = root->r.r_hot_team;
3916   int n = hot_team->t.t_nproc;
3917   int i;
3918 
3919   KMP_DEBUG_ASSERT(!root->r.r_active);
3920 
3921   root->r.r_root_team = NULL;
3922   root->r.r_hot_team = NULL;
3923   // __kmp_free_team() does not free hot teams, so we have to clear r_hot_team
3924   // before call to __kmp_free_team().
3925   __kmp_free_team(root, root_team USE_NESTED_HOT_ARG(NULL));
3926 #if KMP_NESTED_HOT_TEAMS
3927   if (__kmp_hot_teams_max_level >
3928       0) { // need to free nested hot teams and their threads if any
3929     for (i = 0; i < hot_team->t.t_nproc; ++i) {
3930       kmp_info_t *th = hot_team->t.t_threads[i];
3931       if (__kmp_hot_teams_max_level > 1) {
3932         n += __kmp_free_hot_teams(root, th, 1, __kmp_hot_teams_max_level);
3933       }
3934       if (th->th.th_hot_teams) {
3935         __kmp_free(th->th.th_hot_teams);
3936         th->th.th_hot_teams = NULL;
3937       }
3938     }
3939   }
3940 #endif
3941   __kmp_free_team(root, hot_team USE_NESTED_HOT_ARG(NULL));
3942 
3943   // Before we can reap the thread, we need to make certain that all other
3944   // threads in the teams that had this root as ancestor have stopped trying to
3945   // steal tasks.
3946   if (__kmp_tasking_mode != tskm_immediate_exec) {
3947     __kmp_wait_to_unref_task_teams();
3948   }
3949 
3950 #if KMP_OS_WINDOWS
3951   /* Close Handle of root duplicated in __kmp_create_worker (tr #62919) */
3952   KA_TRACE(
3953       10, ("__kmp_reset_root: free handle, th = %p, handle = %" KMP_UINTPTR_SPEC
3954            "\n",
3955            (LPVOID) & (root->r.r_uber_thread->th),
3956            root->r.r_uber_thread->th.th_info.ds.ds_thread));
3957   __kmp_free_handle(root->r.r_uber_thread->th.th_info.ds.ds_thread);
3958 #endif /* KMP_OS_WINDOWS */
3959 
3960 #if OMPT_SUPPORT
3961   if (ompt_enabled.ompt_callback_thread_end) {
3962     ompt_callbacks.ompt_callback(ompt_callback_thread_end)(
3963         &(root->r.r_uber_thread->th.ompt_thread_info.thread_data));
3964   }
3965 #endif
3966 
3967   TCW_4(__kmp_nth,
3968         __kmp_nth - 1); // __kmp_reap_thread will decrement __kmp_all_nth.
3969   root->r.r_uber_thread->th.th_cg_roots->cg_nthreads--;
3970   KA_TRACE(100, ("__kmp_reset_root: Thread %p decrement cg_nthreads on node %p"
3971                  " to %d\n",
3972                  root->r.r_uber_thread, root->r.r_uber_thread->th.th_cg_roots,
3973                  root->r.r_uber_thread->th.th_cg_roots->cg_nthreads));
3974 
3975   __kmp_reap_thread(root->r.r_uber_thread, 1);
3976 
3977   // We canot put root thread to __kmp_thread_pool, so we have to reap it istead
3978   // of freeing.
3979   root->r.r_uber_thread = NULL;
3980   /* mark root as no longer in use */
3981   root->r.r_begin = FALSE;
3982 
3983   return n;
3984 }
3985 
3986 void __kmp_unregister_root_current_thread(int gtid) {
3987   KA_TRACE(1, ("__kmp_unregister_root_current_thread: enter T#%d\n", gtid));
3988   /* this lock should be ok, since unregister_root_current_thread is never
3989      called during an abort, only during a normal close. furthermore, if you
3990      have the forkjoin lock, you should never try to get the initz lock */
3991   __kmp_acquire_bootstrap_lock(&__kmp_forkjoin_lock);
3992   if (TCR_4(__kmp_global.g.g_done) || !__kmp_init_serial) {
3993     KC_TRACE(10, ("__kmp_unregister_root_current_thread: already finished, "
3994                   "exiting T#%d\n",
3995                   gtid));
3996     __kmp_release_bootstrap_lock(&__kmp_forkjoin_lock);
3997     return;
3998   }
3999   kmp_root_t *root = __kmp_root[gtid];
4000 
4001   KMP_DEBUG_ASSERT(__kmp_threads && __kmp_threads[gtid]);
4002   KMP_ASSERT(KMP_UBER_GTID(gtid));
4003   KMP_ASSERT(root == __kmp_threads[gtid]->th.th_root);
4004   KMP_ASSERT(root->r.r_active == FALSE);
4005 
4006   KMP_MB();
4007 
4008 #if OMP_45_ENABLED
4009   kmp_info_t *thread = __kmp_threads[gtid];
4010   kmp_team_t *team = thread->th.th_team;
4011   kmp_task_team_t *task_team = thread->th.th_task_team;
4012 
4013   // we need to wait for the proxy tasks before finishing the thread
4014   if (task_team != NULL && task_team->tt.tt_found_proxy_tasks) {
4015 #if OMPT_SUPPORT
4016     // the runtime is shutting down so we won't report any events
4017     thread->th.ompt_thread_info.state = ompt_state_undefined;
4018 #endif
4019     __kmp_task_team_wait(thread, team USE_ITT_BUILD_ARG(NULL));
4020   }
4021 #endif
4022 
4023   __kmp_reset_root(gtid, root);
4024 
4025   /* free up this thread slot */
4026   __kmp_gtid_set_specific(KMP_GTID_DNE);
4027 #ifdef KMP_TDATA_GTID
4028   __kmp_gtid = KMP_GTID_DNE;
4029 #endif
4030 
4031   KMP_MB();
4032   KC_TRACE(10,
4033            ("__kmp_unregister_root_current_thread: T#%d unregistered\n", gtid));
4034 
4035   __kmp_release_bootstrap_lock(&__kmp_forkjoin_lock);
4036 }
4037 
4038 #if KMP_OS_WINDOWS
4039 /* __kmp_forkjoin_lock must be already held
4040    Unregisters a root thread that is not the current thread.  Returns the number
4041    of __kmp_threads entries freed as a result. */
4042 static int __kmp_unregister_root_other_thread(int gtid) {
4043   kmp_root_t *root = __kmp_root[gtid];
4044   int r;
4045 
4046   KA_TRACE(1, ("__kmp_unregister_root_other_thread: enter T#%d\n", gtid));
4047   KMP_DEBUG_ASSERT(__kmp_threads && __kmp_threads[gtid]);
4048   KMP_ASSERT(KMP_UBER_GTID(gtid));
4049   KMP_ASSERT(root == __kmp_threads[gtid]->th.th_root);
4050   KMP_ASSERT(root->r.r_active == FALSE);
4051 
4052   r = __kmp_reset_root(gtid, root);
4053   KC_TRACE(10,
4054            ("__kmp_unregister_root_other_thread: T#%d unregistered\n", gtid));
4055   return r;
4056 }
4057 #endif
4058 
4059 #if KMP_DEBUG
4060 void __kmp_task_info() {
4061 
4062   kmp_int32 gtid = __kmp_entry_gtid();
4063   kmp_int32 tid = __kmp_tid_from_gtid(gtid);
4064   kmp_info_t *this_thr = __kmp_threads[gtid];
4065   kmp_team_t *steam = this_thr->th.th_serial_team;
4066   kmp_team_t *team = this_thr->th.th_team;
4067 
4068   __kmp_printf(
4069       "__kmp_task_info: gtid=%d tid=%d t_thread=%p team=%p steam=%p curtask=%p "
4070       "ptask=%p\n",
4071       gtid, tid, this_thr, team, steam, this_thr->th.th_current_task,
4072       team->t.t_implicit_task_taskdata[tid].td_parent);
4073 }
4074 #endif // KMP_DEBUG
4075 
4076 /* TODO optimize with one big memclr, take out what isn't needed, split
4077    responsibility to workers as much as possible, and delay initialization of
4078    features as much as possible  */
4079 static void __kmp_initialize_info(kmp_info_t *this_thr, kmp_team_t *team,
4080                                   int tid, int gtid) {
4081   /* this_thr->th.th_info.ds.ds_gtid is setup in
4082      kmp_allocate_thread/create_worker.
4083      this_thr->th.th_serial_team is setup in __kmp_allocate_thread */
4084   kmp_info_t *master = team->t.t_threads[0];
4085   KMP_DEBUG_ASSERT(this_thr != NULL);
4086   KMP_DEBUG_ASSERT(this_thr->th.th_serial_team);
4087   KMP_DEBUG_ASSERT(team);
4088   KMP_DEBUG_ASSERT(team->t.t_threads);
4089   KMP_DEBUG_ASSERT(team->t.t_dispatch);
4090   KMP_DEBUG_ASSERT(master);
4091   KMP_DEBUG_ASSERT(master->th.th_root);
4092 
4093   KMP_MB();
4094 
4095   TCW_SYNC_PTR(this_thr->th.th_team, team);
4096 
4097   this_thr->th.th_info.ds.ds_tid = tid;
4098   this_thr->th.th_set_nproc = 0;
4099   if (__kmp_tasking_mode != tskm_immediate_exec)
4100     // When tasking is possible, threads are not safe to reap until they are
4101     // done tasking; this will be set when tasking code is exited in wait
4102     this_thr->th.th_reap_state = KMP_NOT_SAFE_TO_REAP;
4103   else // no tasking --> always safe to reap
4104     this_thr->th.th_reap_state = KMP_SAFE_TO_REAP;
4105 #if OMP_40_ENABLED
4106   this_thr->th.th_set_proc_bind = proc_bind_default;
4107 #if KMP_AFFINITY_SUPPORTED
4108   this_thr->th.th_new_place = this_thr->th.th_current_place;
4109 #endif
4110 #endif
4111   this_thr->th.th_root = master->th.th_root;
4112 
4113   /* setup the thread's cache of the team structure */
4114   this_thr->th.th_team_nproc = team->t.t_nproc;
4115   this_thr->th.th_team_master = master;
4116   this_thr->th.th_team_serialized = team->t.t_serialized;
4117   TCW_PTR(this_thr->th.th_sleep_loc, NULL);
4118 
4119   KMP_DEBUG_ASSERT(team->t.t_implicit_task_taskdata);
4120 
4121   KF_TRACE(10, ("__kmp_initialize_info1: T#%d:%d this_thread=%p curtask=%p\n",
4122                 tid, gtid, this_thr, this_thr->th.th_current_task));
4123 
4124   __kmp_init_implicit_task(this_thr->th.th_team_master->th.th_ident, this_thr,
4125                            team, tid, TRUE);
4126 
4127   KF_TRACE(10, ("__kmp_initialize_info2: T#%d:%d this_thread=%p curtask=%p\n",
4128                 tid, gtid, this_thr, this_thr->th.th_current_task));
4129   // TODO: Initialize ICVs from parent; GEH - isn't that already done in
4130   // __kmp_initialize_team()?
4131 
4132   /* TODO no worksharing in speculative threads */
4133   this_thr->th.th_dispatch = &team->t.t_dispatch[tid];
4134 
4135   this_thr->th.th_local.this_construct = 0;
4136 
4137   if (!this_thr->th.th_pri_common) {
4138     this_thr->th.th_pri_common =
4139         (struct common_table *)__kmp_allocate(sizeof(struct common_table));
4140     if (__kmp_storage_map) {
4141       __kmp_print_storage_map_gtid(
4142           gtid, this_thr->th.th_pri_common, this_thr->th.th_pri_common + 1,
4143           sizeof(struct common_table), "th_%d.th_pri_common\n", gtid);
4144     }
4145     this_thr->th.th_pri_head = NULL;
4146   }
4147 
4148   if (this_thr != master && // Master's CG root is initialized elsewhere
4149       this_thr->th.th_cg_roots != master->th.th_cg_roots) { // CG root not set
4150     // Make new thread's CG root same as master's
4151     KMP_DEBUG_ASSERT(master->th.th_cg_roots);
4152     this_thr->th.th_cg_roots = master->th.th_cg_roots;
4153     // Increment new thread's CG root's counter to add the new thread
4154     this_thr->th.th_cg_roots->cg_nthreads++;
4155     KA_TRACE(100, ("__kmp_initialize_info: Thread %p increment cg_nthreads on"
4156                    " node %p of thread %p to %d\n",
4157                    this_thr, this_thr->th.th_cg_roots,
4158                    this_thr->th.th_cg_roots->cg_root,
4159                    this_thr->th.th_cg_roots->cg_nthreads));
4160     this_thr->th.th_current_task->td_icvs.thread_limit =
4161         this_thr->th.th_cg_roots->cg_thread_limit;
4162   }
4163 
4164   /* Initialize dynamic dispatch */
4165   {
4166     volatile kmp_disp_t *dispatch = this_thr->th.th_dispatch;
4167     // Use team max_nproc since this will never change for the team.
4168     size_t disp_size =
4169         sizeof(dispatch_private_info_t) *
4170         (team->t.t_max_nproc == 1 ? 1 : __kmp_dispatch_num_buffers);
4171     KD_TRACE(10, ("__kmp_initialize_info: T#%d max_nproc: %d\n", gtid,
4172                   team->t.t_max_nproc));
4173     KMP_ASSERT(dispatch);
4174     KMP_DEBUG_ASSERT(team->t.t_dispatch);
4175     KMP_DEBUG_ASSERT(dispatch == &team->t.t_dispatch[tid]);
4176 
4177     dispatch->th_disp_index = 0;
4178 #if OMP_45_ENABLED
4179     dispatch->th_doacross_buf_idx = 0;
4180 #endif
4181     if (!dispatch->th_disp_buffer) {
4182       dispatch->th_disp_buffer =
4183           (dispatch_private_info_t *)__kmp_allocate(disp_size);
4184 
4185       if (__kmp_storage_map) {
4186         __kmp_print_storage_map_gtid(
4187             gtid, &dispatch->th_disp_buffer[0],
4188             &dispatch->th_disp_buffer[team->t.t_max_nproc == 1
4189                                           ? 1
4190                                           : __kmp_dispatch_num_buffers],
4191             disp_size, "th_%d.th_dispatch.th_disp_buffer "
4192                        "(team_%d.t_dispatch[%d].th_disp_buffer)",
4193             gtid, team->t.t_id, gtid);
4194       }
4195     } else {
4196       memset(&dispatch->th_disp_buffer[0], '\0', disp_size);
4197     }
4198 
4199     dispatch->th_dispatch_pr_current = 0;
4200     dispatch->th_dispatch_sh_current = 0;
4201 
4202     dispatch->th_deo_fcn = 0; /* ORDERED     */
4203     dispatch->th_dxo_fcn = 0; /* END ORDERED */
4204   }
4205 
4206   this_thr->th.th_next_pool = NULL;
4207 
4208   if (!this_thr->th.th_task_state_memo_stack) {
4209     size_t i;
4210     this_thr->th.th_task_state_memo_stack =
4211         (kmp_uint8 *)__kmp_allocate(4 * sizeof(kmp_uint8));
4212     this_thr->th.th_task_state_top = 0;
4213     this_thr->th.th_task_state_stack_sz = 4;
4214     for (i = 0; i < this_thr->th.th_task_state_stack_sz;
4215          ++i) // zero init the stack
4216       this_thr->th.th_task_state_memo_stack[i] = 0;
4217   }
4218 
4219   KMP_DEBUG_ASSERT(!this_thr->th.th_spin_here);
4220   KMP_DEBUG_ASSERT(this_thr->th.th_next_waiting == 0);
4221 
4222   KMP_MB();
4223 }
4224 
4225 /* allocate a new thread for the requesting team. this is only called from
4226    within a forkjoin critical section. we will first try to get an available
4227    thread from the thread pool. if none is available, we will fork a new one
4228    assuming we are able to create a new one. this should be assured, as the
4229    caller should check on this first. */
4230 kmp_info_t *__kmp_allocate_thread(kmp_root_t *root, kmp_team_t *team,
4231                                   int new_tid) {
4232   kmp_team_t *serial_team;
4233   kmp_info_t *new_thr;
4234   int new_gtid;
4235 
4236   KA_TRACE(20, ("__kmp_allocate_thread: T#%d\n", __kmp_get_gtid()));
4237   KMP_DEBUG_ASSERT(root && team);
4238 #if !KMP_NESTED_HOT_TEAMS
4239   KMP_DEBUG_ASSERT(KMP_MASTER_GTID(__kmp_get_gtid()));
4240 #endif
4241   KMP_MB();
4242 
4243   /* first, try to get one from the thread pool */
4244   if (__kmp_thread_pool) {
4245     new_thr = CCAST(kmp_info_t *, __kmp_thread_pool);
4246     __kmp_thread_pool = (volatile kmp_info_t *)new_thr->th.th_next_pool;
4247     if (new_thr == __kmp_thread_pool_insert_pt) {
4248       __kmp_thread_pool_insert_pt = NULL;
4249     }
4250     TCW_4(new_thr->th.th_in_pool, FALSE);
4251     // Don't touch th_active_in_pool or th_active.
4252     // The worker thread adjusts those flags as it sleeps/awakens.
4253     __kmp_thread_pool_nth--;
4254 
4255     KA_TRACE(20, ("__kmp_allocate_thread: T#%d using thread T#%d\n",
4256                   __kmp_get_gtid(), new_thr->th.th_info.ds.ds_gtid));
4257     KMP_ASSERT(!new_thr->th.th_team);
4258     KMP_DEBUG_ASSERT(__kmp_nth < __kmp_threads_capacity);
4259     KMP_DEBUG_ASSERT(__kmp_thread_pool_nth >= 0);
4260 
4261     /* setup the thread structure */
4262     __kmp_initialize_info(new_thr, team, new_tid,
4263                           new_thr->th.th_info.ds.ds_gtid);
4264     KMP_DEBUG_ASSERT(new_thr->th.th_serial_team);
4265 
4266     TCW_4(__kmp_nth, __kmp_nth + 1);
4267 
4268     new_thr->th.th_task_state = 0;
4269     new_thr->th.th_task_state_top = 0;
4270     new_thr->th.th_task_state_stack_sz = 4;
4271 
4272 #ifdef KMP_ADJUST_BLOCKTIME
4273     /* Adjust blocktime back to zero if necessary */
4274     /* Middle initialization might not have occurred yet */
4275     if (!__kmp_env_blocktime && (__kmp_avail_proc > 0)) {
4276       if (__kmp_nth > __kmp_avail_proc) {
4277         __kmp_zero_bt = TRUE;
4278       }
4279     }
4280 #endif /* KMP_ADJUST_BLOCKTIME */
4281 
4282 #if KMP_DEBUG
4283     // If thread entered pool via __kmp_free_thread, wait_flag should !=
4284     // KMP_BARRIER_PARENT_FLAG.
4285     int b;
4286     kmp_balign_t *balign = new_thr->th.th_bar;
4287     for (b = 0; b < bs_last_barrier; ++b)
4288       KMP_DEBUG_ASSERT(balign[b].bb.wait_flag != KMP_BARRIER_PARENT_FLAG);
4289 #endif
4290 
4291     KF_TRACE(10, ("__kmp_allocate_thread: T#%d using thread %p T#%d\n",
4292                   __kmp_get_gtid(), new_thr, new_thr->th.th_info.ds.ds_gtid));
4293 
4294     KMP_MB();
4295     return new_thr;
4296   }
4297 
4298   /* no, well fork a new one */
4299   KMP_ASSERT(__kmp_nth == __kmp_all_nth);
4300   KMP_ASSERT(__kmp_all_nth < __kmp_threads_capacity);
4301 
4302 #if KMP_USE_MONITOR
4303   // If this is the first worker thread the RTL is creating, then also
4304   // launch the monitor thread.  We try to do this as early as possible.
4305   if (!TCR_4(__kmp_init_monitor)) {
4306     __kmp_acquire_bootstrap_lock(&__kmp_monitor_lock);
4307     if (!TCR_4(__kmp_init_monitor)) {
4308       KF_TRACE(10, ("before __kmp_create_monitor\n"));
4309       TCW_4(__kmp_init_monitor, 1);
4310       __kmp_create_monitor(&__kmp_monitor);
4311       KF_TRACE(10, ("after __kmp_create_monitor\n"));
4312 #if KMP_OS_WINDOWS
4313       // AC: wait until monitor has started. This is a fix for CQ232808.
4314       // The reason is that if the library is loaded/unloaded in a loop with
4315       // small (parallel) work in between, then there is high probability that
4316       // monitor thread started after the library shutdown. At shutdown it is
4317       // too late to cope with the problem, because when the master is in
4318       // DllMain (process detach) the monitor has no chances to start (it is
4319       // blocked), and master has no means to inform the monitor that the
4320       // library has gone, because all the memory which the monitor can access
4321       // is going to be released/reset.
4322       while (TCR_4(__kmp_init_monitor) < 2) {
4323         KMP_YIELD(TRUE);
4324       }
4325       KF_TRACE(10, ("after monitor thread has started\n"));
4326 #endif
4327     }
4328     __kmp_release_bootstrap_lock(&__kmp_monitor_lock);
4329   }
4330 #endif
4331 
4332   KMP_MB();
4333   for (new_gtid = 1; TCR_PTR(__kmp_threads[new_gtid]) != NULL; ++new_gtid) {
4334     KMP_DEBUG_ASSERT(new_gtid < __kmp_threads_capacity);
4335   }
4336 
4337   /* allocate space for it. */
4338   new_thr = (kmp_info_t *)__kmp_allocate(sizeof(kmp_info_t));
4339 
4340   TCW_SYNC_PTR(__kmp_threads[new_gtid], new_thr);
4341 
4342   if (__kmp_storage_map) {
4343     __kmp_print_thread_storage_map(new_thr, new_gtid);
4344   }
4345 
4346   // add the reserve serialized team, initialized from the team's master thread
4347   {
4348     kmp_internal_control_t r_icvs = __kmp_get_x_global_icvs(team);
4349     KF_TRACE(10, ("__kmp_allocate_thread: before th_serial/serial_team\n"));
4350     new_thr->th.th_serial_team = serial_team =
4351         (kmp_team_t *)__kmp_allocate_team(root, 1, 1,
4352 #if OMPT_SUPPORT
4353                                           ompt_data_none, // root parallel id
4354 #endif
4355 #if OMP_40_ENABLED
4356                                           proc_bind_default,
4357 #endif
4358                                           &r_icvs, 0 USE_NESTED_HOT_ARG(NULL));
4359   }
4360   KMP_ASSERT(serial_team);
4361   serial_team->t.t_serialized = 0; // AC: the team created in reserve, not for
4362   // execution (it is unused for now).
4363   serial_team->t.t_threads[0] = new_thr;
4364   KF_TRACE(10,
4365            ("__kmp_allocate_thread: after th_serial/serial_team : new_thr=%p\n",
4366             new_thr));
4367 
4368   /* setup the thread structures */
4369   __kmp_initialize_info(new_thr, team, new_tid, new_gtid);
4370 
4371 #if USE_FAST_MEMORY
4372   __kmp_initialize_fast_memory(new_thr);
4373 #endif /* USE_FAST_MEMORY */
4374 
4375 #if KMP_USE_BGET
4376   KMP_DEBUG_ASSERT(new_thr->th.th_local.bget_data == NULL);
4377   __kmp_initialize_bget(new_thr);
4378 #endif
4379 
4380   __kmp_init_random(new_thr); // Initialize random number generator
4381 
4382   /* Initialize these only once when thread is grabbed for a team allocation */
4383   KA_TRACE(20,
4384            ("__kmp_allocate_thread: T#%d init go fork=%u, plain=%u\n",
4385             __kmp_get_gtid(), KMP_INIT_BARRIER_STATE, KMP_INIT_BARRIER_STATE));
4386 
4387   int b;
4388   kmp_balign_t *balign = new_thr->th.th_bar;
4389   for (b = 0; b < bs_last_barrier; ++b) {
4390     balign[b].bb.b_go = KMP_INIT_BARRIER_STATE;
4391     balign[b].bb.team = NULL;
4392     balign[b].bb.wait_flag = KMP_BARRIER_NOT_WAITING;
4393     balign[b].bb.use_oncore_barrier = 0;
4394   }
4395 
4396   new_thr->th.th_spin_here = FALSE;
4397   new_thr->th.th_next_waiting = 0;
4398 #if KMP_OS_UNIX
4399   new_thr->th.th_blocking = false;
4400 #endif
4401 
4402 #if OMP_40_ENABLED && KMP_AFFINITY_SUPPORTED
4403   new_thr->th.th_current_place = KMP_PLACE_UNDEFINED;
4404   new_thr->th.th_new_place = KMP_PLACE_UNDEFINED;
4405   new_thr->th.th_first_place = KMP_PLACE_UNDEFINED;
4406   new_thr->th.th_last_place = KMP_PLACE_UNDEFINED;
4407 #endif
4408 #if OMP_50_ENABLED
4409   new_thr->th.th_def_allocator = __kmp_def_allocator;
4410   new_thr->th.th_prev_level = 0;
4411   new_thr->th.th_prev_num_threads = 1;
4412 #endif
4413 
4414   TCW_4(new_thr->th.th_in_pool, FALSE);
4415   new_thr->th.th_active_in_pool = FALSE;
4416   TCW_4(new_thr->th.th_active, TRUE);
4417 
4418   /* adjust the global counters */
4419   __kmp_all_nth++;
4420   __kmp_nth++;
4421 
4422   // if __kmp_adjust_gtid_mode is set, then we use method #1 (sp search) for low
4423   // numbers of procs, and method #2 (keyed API call) for higher numbers.
4424   if (__kmp_adjust_gtid_mode) {
4425     if (__kmp_all_nth >= __kmp_tls_gtid_min) {
4426       if (TCR_4(__kmp_gtid_mode) != 2) {
4427         TCW_4(__kmp_gtid_mode, 2);
4428       }
4429     } else {
4430       if (TCR_4(__kmp_gtid_mode) != 1) {
4431         TCW_4(__kmp_gtid_mode, 1);
4432       }
4433     }
4434   }
4435 
4436 #ifdef KMP_ADJUST_BLOCKTIME
4437   /* Adjust blocktime back to zero if necessary       */
4438   /* Middle initialization might not have occurred yet */
4439   if (!__kmp_env_blocktime && (__kmp_avail_proc > 0)) {
4440     if (__kmp_nth > __kmp_avail_proc) {
4441       __kmp_zero_bt = TRUE;
4442     }
4443   }
4444 #endif /* KMP_ADJUST_BLOCKTIME */
4445 
4446   /* actually fork it and create the new worker thread */
4447   KF_TRACE(
4448       10, ("__kmp_allocate_thread: before __kmp_create_worker: %p\n", new_thr));
4449   __kmp_create_worker(new_gtid, new_thr, __kmp_stksize);
4450   KF_TRACE(10,
4451            ("__kmp_allocate_thread: after __kmp_create_worker: %p\n", new_thr));
4452 
4453   KA_TRACE(20, ("__kmp_allocate_thread: T#%d forked T#%d\n", __kmp_get_gtid(),
4454                 new_gtid));
4455   KMP_MB();
4456   return new_thr;
4457 }
4458 
4459 /* Reinitialize team for reuse.
4460    The hot team code calls this case at every fork barrier, so EPCC barrier
4461    test are extremely sensitive to changes in it, esp. writes to the team
4462    struct, which cause a cache invalidation in all threads.
4463    IF YOU TOUCH THIS ROUTINE, RUN EPCC C SYNCBENCH ON A BIG-IRON MACHINE!!! */
4464 static void __kmp_reinitialize_team(kmp_team_t *team,
4465                                     kmp_internal_control_t *new_icvs,
4466                                     ident_t *loc) {
4467   KF_TRACE(10, ("__kmp_reinitialize_team: enter this_thread=%p team=%p\n",
4468                 team->t.t_threads[0], team));
4469   KMP_DEBUG_ASSERT(team && new_icvs);
4470   KMP_DEBUG_ASSERT((!TCR_4(__kmp_init_parallel)) || new_icvs->nproc);
4471   KMP_CHECK_UPDATE(team->t.t_ident, loc);
4472 
4473   KMP_CHECK_UPDATE(team->t.t_id, KMP_GEN_TEAM_ID());
4474   // Copy ICVs to the master thread's implicit taskdata
4475   __kmp_init_implicit_task(loc, team->t.t_threads[0], team, 0, FALSE);
4476   copy_icvs(&team->t.t_implicit_task_taskdata[0].td_icvs, new_icvs);
4477 
4478   KF_TRACE(10, ("__kmp_reinitialize_team: exit this_thread=%p team=%p\n",
4479                 team->t.t_threads[0], team));
4480 }
4481 
4482 /* Initialize the team data structure.
4483    This assumes the t_threads and t_max_nproc are already set.
4484    Also, we don't touch the arguments */
4485 static void __kmp_initialize_team(kmp_team_t *team, int new_nproc,
4486                                   kmp_internal_control_t *new_icvs,
4487                                   ident_t *loc) {
4488   KF_TRACE(10, ("__kmp_initialize_team: enter: team=%p\n", team));
4489 
4490   /* verify */
4491   KMP_DEBUG_ASSERT(team);
4492   KMP_DEBUG_ASSERT(new_nproc <= team->t.t_max_nproc);
4493   KMP_DEBUG_ASSERT(team->t.t_threads);
4494   KMP_MB();
4495 
4496   team->t.t_master_tid = 0; /* not needed */
4497   /* team->t.t_master_bar;        not needed */
4498   team->t.t_serialized = new_nproc > 1 ? 0 : 1;
4499   team->t.t_nproc = new_nproc;
4500 
4501   /* team->t.t_parent     = NULL; TODO not needed & would mess up hot team */
4502   team->t.t_next_pool = NULL;
4503   /* memset( team->t.t_threads, 0, sizeof(kmp_info_t*)*new_nproc ); would mess
4504    * up hot team */
4505 
4506   TCW_SYNC_PTR(team->t.t_pkfn, NULL); /* not needed */
4507   team->t.t_invoke = NULL; /* not needed */
4508 
4509   // TODO???: team->t.t_max_active_levels       = new_max_active_levels;
4510   team->t.t_sched.sched = new_icvs->sched.sched;
4511 
4512 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
4513   team->t.t_fp_control_saved = FALSE; /* not needed */
4514   team->t.t_x87_fpu_control_word = 0; /* not needed */
4515   team->t.t_mxcsr = 0; /* not needed */
4516 #endif /* KMP_ARCH_X86 || KMP_ARCH_X86_64 */
4517 
4518   team->t.t_construct = 0;
4519 
4520   team->t.t_ordered.dt.t_value = 0;
4521   team->t.t_master_active = FALSE;
4522 
4523 #ifdef KMP_DEBUG
4524   team->t.t_copypriv_data = NULL; /* not necessary, but nice for debugging */
4525 #endif
4526 #if KMP_OS_WINDOWS
4527   team->t.t_copyin_counter = 0; /* for barrier-free copyin implementation */
4528 #endif
4529 
4530   team->t.t_control_stack_top = NULL;
4531 
4532   __kmp_reinitialize_team(team, new_icvs, loc);
4533 
4534   KMP_MB();
4535   KF_TRACE(10, ("__kmp_initialize_team: exit: team=%p\n", team));
4536 }
4537 
4538 #if KMP_OS_LINUX && KMP_AFFINITY_SUPPORTED
4539 /* Sets full mask for thread and returns old mask, no changes to structures. */
4540 static void
4541 __kmp_set_thread_affinity_mask_full_tmp(kmp_affin_mask_t *old_mask) {
4542   if (KMP_AFFINITY_CAPABLE()) {
4543     int status;
4544     if (old_mask != NULL) {
4545       status = __kmp_get_system_affinity(old_mask, TRUE);
4546       int error = errno;
4547       if (status != 0) {
4548         __kmp_fatal(KMP_MSG(ChangeThreadAffMaskError), KMP_ERR(error),
4549                     __kmp_msg_null);
4550       }
4551     }
4552     __kmp_set_system_affinity(__kmp_affin_fullMask, TRUE);
4553   }
4554 }
4555 #endif
4556 
4557 #if OMP_40_ENABLED && KMP_AFFINITY_SUPPORTED
4558 
4559 // __kmp_partition_places() is the heart of the OpenMP 4.0 affinity mechanism.
4560 // It calculats the worker + master thread's partition based upon the parent
4561 // thread's partition, and binds each worker to a thread in their partition.
4562 // The master thread's partition should already include its current binding.
4563 static void __kmp_partition_places(kmp_team_t *team, int update_master_only) {
4564   // Copy the master thread's place partion to the team struct
4565   kmp_info_t *master_th = team->t.t_threads[0];
4566   KMP_DEBUG_ASSERT(master_th != NULL);
4567   kmp_proc_bind_t proc_bind = team->t.t_proc_bind;
4568   int first_place = master_th->th.th_first_place;
4569   int last_place = master_th->th.th_last_place;
4570   int masters_place = master_th->th.th_current_place;
4571   team->t.t_first_place = first_place;
4572   team->t.t_last_place = last_place;
4573 
4574   KA_TRACE(20, ("__kmp_partition_places: enter: proc_bind = %d T#%d(%d:0) "
4575                 "bound to place %d partition = [%d,%d]\n",
4576                 proc_bind, __kmp_gtid_from_thread(team->t.t_threads[0]),
4577                 team->t.t_id, masters_place, first_place, last_place));
4578 
4579   switch (proc_bind) {
4580 
4581   case proc_bind_default:
4582     // serial teams might have the proc_bind policy set to proc_bind_default. It
4583     // doesn't matter, as we don't rebind master thread for any proc_bind policy
4584     KMP_DEBUG_ASSERT(team->t.t_nproc == 1);
4585     break;
4586 
4587   case proc_bind_master: {
4588     int f;
4589     int n_th = team->t.t_nproc;
4590     for (f = 1; f < n_th; f++) {
4591       kmp_info_t *th = team->t.t_threads[f];
4592       KMP_DEBUG_ASSERT(th != NULL);
4593       th->th.th_first_place = first_place;
4594       th->th.th_last_place = last_place;
4595       th->th.th_new_place = masters_place;
4596 #if OMP_50_ENABLED
4597       if (__kmp_display_affinity && masters_place != th->th.th_current_place &&
4598           team->t.t_display_affinity != 1) {
4599         team->t.t_display_affinity = 1;
4600       }
4601 #endif
4602 
4603       KA_TRACE(100, ("__kmp_partition_places: master: T#%d(%d:%d) place %d "
4604                      "partition = [%d,%d]\n",
4605                      __kmp_gtid_from_thread(team->t.t_threads[f]), team->t.t_id,
4606                      f, masters_place, first_place, last_place));
4607     }
4608   } break;
4609 
4610   case proc_bind_close: {
4611     int f;
4612     int n_th = team->t.t_nproc;
4613     int n_places;
4614     if (first_place <= last_place) {
4615       n_places = last_place - first_place + 1;
4616     } else {
4617       n_places = __kmp_affinity_num_masks - first_place + last_place + 1;
4618     }
4619     if (n_th <= n_places) {
4620       int place = masters_place;
4621       for (f = 1; f < n_th; f++) {
4622         kmp_info_t *th = team->t.t_threads[f];
4623         KMP_DEBUG_ASSERT(th != NULL);
4624 
4625         if (place == last_place) {
4626           place = first_place;
4627         } else if (place == (int)(__kmp_affinity_num_masks - 1)) {
4628           place = 0;
4629         } else {
4630           place++;
4631         }
4632         th->th.th_first_place = first_place;
4633         th->th.th_last_place = last_place;
4634         th->th.th_new_place = place;
4635 #if OMP_50_ENABLED
4636         if (__kmp_display_affinity && place != th->th.th_current_place &&
4637             team->t.t_display_affinity != 1) {
4638           team->t.t_display_affinity = 1;
4639         }
4640 #endif
4641 
4642         KA_TRACE(100, ("__kmp_partition_places: close: T#%d(%d:%d) place %d "
4643                        "partition = [%d,%d]\n",
4644                        __kmp_gtid_from_thread(team->t.t_threads[f]),
4645                        team->t.t_id, f, place, first_place, last_place));
4646       }
4647     } else {
4648       int S, rem, gap, s_count;
4649       S = n_th / n_places;
4650       s_count = 0;
4651       rem = n_th - (S * n_places);
4652       gap = rem > 0 ? n_places / rem : n_places;
4653       int place = masters_place;
4654       int gap_ct = gap;
4655       for (f = 0; f < n_th; f++) {
4656         kmp_info_t *th = team->t.t_threads[f];
4657         KMP_DEBUG_ASSERT(th != NULL);
4658 
4659         th->th.th_first_place = first_place;
4660         th->th.th_last_place = last_place;
4661         th->th.th_new_place = place;
4662 #if OMP_50_ENABLED
4663         if (__kmp_display_affinity && place != th->th.th_current_place &&
4664             team->t.t_display_affinity != 1) {
4665           team->t.t_display_affinity = 1;
4666         }
4667 #endif
4668         s_count++;
4669 
4670         if ((s_count == S) && rem && (gap_ct == gap)) {
4671           // do nothing, add an extra thread to place on next iteration
4672         } else if ((s_count == S + 1) && rem && (gap_ct == gap)) {
4673           // we added an extra thread to this place; move to next place
4674           if (place == last_place) {
4675             place = first_place;
4676           } else if (place == (int)(__kmp_affinity_num_masks - 1)) {
4677             place = 0;
4678           } else {
4679             place++;
4680           }
4681           s_count = 0;
4682           gap_ct = 1;
4683           rem--;
4684         } else if (s_count == S) { // place full; don't add extra
4685           if (place == last_place) {
4686             place = first_place;
4687           } else if (place == (int)(__kmp_affinity_num_masks - 1)) {
4688             place = 0;
4689           } else {
4690             place++;
4691           }
4692           gap_ct++;
4693           s_count = 0;
4694         }
4695 
4696         KA_TRACE(100,
4697                  ("__kmp_partition_places: close: T#%d(%d:%d) place %d "
4698                   "partition = [%d,%d]\n",
4699                   __kmp_gtid_from_thread(team->t.t_threads[f]), team->t.t_id, f,
4700                   th->th.th_new_place, first_place, last_place));
4701       }
4702       KMP_DEBUG_ASSERT(place == masters_place);
4703     }
4704   } break;
4705 
4706   case proc_bind_spread: {
4707     int f;
4708     int n_th = team->t.t_nproc;
4709     int n_places;
4710     int thidx;
4711     if (first_place <= last_place) {
4712       n_places = last_place - first_place + 1;
4713     } else {
4714       n_places = __kmp_affinity_num_masks - first_place + last_place + 1;
4715     }
4716     if (n_th <= n_places) {
4717       int place = -1;
4718 
4719       if (n_places != static_cast<int>(__kmp_affinity_num_masks)) {
4720         int S = n_places / n_th;
4721         int s_count, rem, gap, gap_ct;
4722 
4723         place = masters_place;
4724         rem = n_places - n_th * S;
4725         gap = rem ? n_th / rem : 1;
4726         gap_ct = gap;
4727         thidx = n_th;
4728         if (update_master_only == 1)
4729           thidx = 1;
4730         for (f = 0; f < thidx; f++) {
4731           kmp_info_t *th = team->t.t_threads[f];
4732           KMP_DEBUG_ASSERT(th != NULL);
4733 
4734           th->th.th_first_place = place;
4735           th->th.th_new_place = place;
4736 #if OMP_50_ENABLED
4737           if (__kmp_display_affinity && place != th->th.th_current_place &&
4738               team->t.t_display_affinity != 1) {
4739             team->t.t_display_affinity = 1;
4740           }
4741 #endif
4742           s_count = 1;
4743           while (s_count < S) {
4744             if (place == last_place) {
4745               place = first_place;
4746             } else if (place == (int)(__kmp_affinity_num_masks - 1)) {
4747               place = 0;
4748             } else {
4749               place++;
4750             }
4751             s_count++;
4752           }
4753           if (rem && (gap_ct == gap)) {
4754             if (place == last_place) {
4755               place = first_place;
4756             } else if (place == (int)(__kmp_affinity_num_masks - 1)) {
4757               place = 0;
4758             } else {
4759               place++;
4760             }
4761             rem--;
4762             gap_ct = 0;
4763           }
4764           th->th.th_last_place = place;
4765           gap_ct++;
4766 
4767           if (place == last_place) {
4768             place = first_place;
4769           } else if (place == (int)(__kmp_affinity_num_masks - 1)) {
4770             place = 0;
4771           } else {
4772             place++;
4773           }
4774 
4775           KA_TRACE(100,
4776                    ("__kmp_partition_places: spread: T#%d(%d:%d) place %d "
4777                     "partition = [%d,%d], __kmp_affinity_num_masks: %u\n",
4778                     __kmp_gtid_from_thread(team->t.t_threads[f]), team->t.t_id,
4779                     f, th->th.th_new_place, th->th.th_first_place,
4780                     th->th.th_last_place, __kmp_affinity_num_masks));
4781         }
4782       } else {
4783         /* Having uniform space of available computation places I can create
4784            T partitions of round(P/T) size and put threads into the first
4785            place of each partition. */
4786         double current = static_cast<double>(masters_place);
4787         double spacing =
4788             (static_cast<double>(n_places + 1) / static_cast<double>(n_th));
4789         int first, last;
4790         kmp_info_t *th;
4791 
4792         thidx = n_th + 1;
4793         if (update_master_only == 1)
4794           thidx = 1;
4795         for (f = 0; f < thidx; f++) {
4796           first = static_cast<int>(current);
4797           last = static_cast<int>(current + spacing) - 1;
4798           KMP_DEBUG_ASSERT(last >= first);
4799           if (first >= n_places) {
4800             if (masters_place) {
4801               first -= n_places;
4802               last -= n_places;
4803               if (first == (masters_place + 1)) {
4804                 KMP_DEBUG_ASSERT(f == n_th);
4805                 first--;
4806               }
4807               if (last == masters_place) {
4808                 KMP_DEBUG_ASSERT(f == (n_th - 1));
4809                 last--;
4810               }
4811             } else {
4812               KMP_DEBUG_ASSERT(f == n_th);
4813               first = 0;
4814               last = 0;
4815             }
4816           }
4817           if (last >= n_places) {
4818             last = (n_places - 1);
4819           }
4820           place = first;
4821           current += spacing;
4822           if (f < n_th) {
4823             KMP_DEBUG_ASSERT(0 <= first);
4824             KMP_DEBUG_ASSERT(n_places > first);
4825             KMP_DEBUG_ASSERT(0 <= last);
4826             KMP_DEBUG_ASSERT(n_places > last);
4827             KMP_DEBUG_ASSERT(last_place >= first_place);
4828             th = team->t.t_threads[f];
4829             KMP_DEBUG_ASSERT(th);
4830             th->th.th_first_place = first;
4831             th->th.th_new_place = place;
4832             th->th.th_last_place = last;
4833 #if OMP_50_ENABLED
4834             if (__kmp_display_affinity && place != th->th.th_current_place &&
4835                 team->t.t_display_affinity != 1) {
4836               team->t.t_display_affinity = 1;
4837             }
4838 #endif
4839             KA_TRACE(100,
4840                      ("__kmp_partition_places: spread: T#%d(%d:%d) place %d "
4841                       "partition = [%d,%d], spacing = %.4f\n",
4842                       __kmp_gtid_from_thread(team->t.t_threads[f]),
4843                       team->t.t_id, f, th->th.th_new_place,
4844                       th->th.th_first_place, th->th.th_last_place, spacing));
4845           }
4846         }
4847       }
4848       KMP_DEBUG_ASSERT(update_master_only || place == masters_place);
4849     } else {
4850       int S, rem, gap, s_count;
4851       S = n_th / n_places;
4852       s_count = 0;
4853       rem = n_th - (S * n_places);
4854       gap = rem > 0 ? n_places / rem : n_places;
4855       int place = masters_place;
4856       int gap_ct = gap;
4857       thidx = n_th;
4858       if (update_master_only == 1)
4859         thidx = 1;
4860       for (f = 0; f < thidx; f++) {
4861         kmp_info_t *th = team->t.t_threads[f];
4862         KMP_DEBUG_ASSERT(th != NULL);
4863 
4864         th->th.th_first_place = place;
4865         th->th.th_last_place = place;
4866         th->th.th_new_place = place;
4867 #if OMP_50_ENABLED
4868         if (__kmp_display_affinity && place != th->th.th_current_place &&
4869             team->t.t_display_affinity != 1) {
4870           team->t.t_display_affinity = 1;
4871         }
4872 #endif
4873         s_count++;
4874 
4875         if ((s_count == S) && rem && (gap_ct == gap)) {
4876           // do nothing, add an extra thread to place on next iteration
4877         } else if ((s_count == S + 1) && rem && (gap_ct == gap)) {
4878           // we added an extra thread to this place; move on to next place
4879           if (place == last_place) {
4880             place = first_place;
4881           } else if (place == (int)(__kmp_affinity_num_masks - 1)) {
4882             place = 0;
4883           } else {
4884             place++;
4885           }
4886           s_count = 0;
4887           gap_ct = 1;
4888           rem--;
4889         } else if (s_count == S) { // place is full; don't add extra thread
4890           if (place == last_place) {
4891             place = first_place;
4892           } else if (place == (int)(__kmp_affinity_num_masks - 1)) {
4893             place = 0;
4894           } else {
4895             place++;
4896           }
4897           gap_ct++;
4898           s_count = 0;
4899         }
4900 
4901         KA_TRACE(100, ("__kmp_partition_places: spread: T#%d(%d:%d) place %d "
4902                        "partition = [%d,%d]\n",
4903                        __kmp_gtid_from_thread(team->t.t_threads[f]),
4904                        team->t.t_id, f, th->th.th_new_place,
4905                        th->th.th_first_place, th->th.th_last_place));
4906       }
4907       KMP_DEBUG_ASSERT(update_master_only || place == masters_place);
4908     }
4909   } break;
4910 
4911   default:
4912     break;
4913   }
4914 
4915   KA_TRACE(20, ("__kmp_partition_places: exit T#%d\n", team->t.t_id));
4916 }
4917 
4918 #endif /* OMP_40_ENABLED && KMP_AFFINITY_SUPPORTED */
4919 
4920 /* allocate a new team data structure to use.  take one off of the free pool if
4921    available */
4922 kmp_team_t *
4923 __kmp_allocate_team(kmp_root_t *root, int new_nproc, int max_nproc,
4924 #if OMPT_SUPPORT
4925                     ompt_data_t ompt_parallel_data,
4926 #endif
4927 #if OMP_40_ENABLED
4928                     kmp_proc_bind_t new_proc_bind,
4929 #endif
4930                     kmp_internal_control_t *new_icvs,
4931                     int argc USE_NESTED_HOT_ARG(kmp_info_t *master)) {
4932   KMP_TIME_DEVELOPER_PARTITIONED_BLOCK(KMP_allocate_team);
4933   int f;
4934   kmp_team_t *team;
4935   int use_hot_team = !root->r.r_active;
4936   int level = 0;
4937 
4938   KA_TRACE(20, ("__kmp_allocate_team: called\n"));
4939   KMP_DEBUG_ASSERT(new_nproc >= 1 && argc >= 0);
4940   KMP_DEBUG_ASSERT(max_nproc >= new_nproc);
4941   KMP_MB();
4942 
4943 #if KMP_NESTED_HOT_TEAMS
4944   kmp_hot_team_ptr_t *hot_teams;
4945   if (master) {
4946     team = master->th.th_team;
4947     level = team->t.t_active_level;
4948     if (master->th.th_teams_microtask) { // in teams construct?
4949       if (master->th.th_teams_size.nteams > 1 &&
4950           ( // #teams > 1
4951               team->t.t_pkfn ==
4952                   (microtask_t)__kmp_teams_master || // inner fork of the teams
4953               master->th.th_teams_level <
4954                   team->t.t_level)) { // or nested parallel inside the teams
4955         ++level; // not increment if #teams==1, or for outer fork of the teams;
4956         // increment otherwise
4957       }
4958     }
4959     hot_teams = master->th.th_hot_teams;
4960     if (level < __kmp_hot_teams_max_level && hot_teams &&
4961         hot_teams[level]
4962             .hot_team) { // hot team has already been allocated for given level
4963       use_hot_team = 1;
4964     } else {
4965       use_hot_team = 0;
4966     }
4967   }
4968 #endif
4969   // Optimization to use a "hot" team
4970   if (use_hot_team && new_nproc > 1) {
4971     KMP_DEBUG_ASSERT(new_nproc <= max_nproc);
4972 #if KMP_NESTED_HOT_TEAMS
4973     team = hot_teams[level].hot_team;
4974 #else
4975     team = root->r.r_hot_team;
4976 #endif
4977 #if KMP_DEBUG
4978     if (__kmp_tasking_mode != tskm_immediate_exec) {
4979       KA_TRACE(20, ("__kmp_allocate_team: hot team task_team[0] = %p "
4980                     "task_team[1] = %p before reinit\n",
4981                     team->t.t_task_team[0], team->t.t_task_team[1]));
4982     }
4983 #endif
4984 
4985     // Has the number of threads changed?
4986     /* Let's assume the most common case is that the number of threads is
4987        unchanged, and put that case first. */
4988     if (team->t.t_nproc == new_nproc) { // Check changes in number of threads
4989       KA_TRACE(20, ("__kmp_allocate_team: reusing hot team\n"));
4990       // This case can mean that omp_set_num_threads() was called and the hot
4991       // team size was already reduced, so we check the special flag
4992       if (team->t.t_size_changed == -1) {
4993         team->t.t_size_changed = 1;
4994       } else {
4995         KMP_CHECK_UPDATE(team->t.t_size_changed, 0);
4996       }
4997 
4998       // TODO???: team->t.t_max_active_levels = new_max_active_levels;
4999       kmp_r_sched_t new_sched = new_icvs->sched;
5000       // set master's schedule as new run-time schedule
5001       KMP_CHECK_UPDATE(team->t.t_sched.sched, new_sched.sched);
5002 
5003       __kmp_reinitialize_team(team, new_icvs,
5004                               root->r.r_uber_thread->th.th_ident);
5005 
5006       KF_TRACE(10, ("__kmp_allocate_team2: T#%d, this_thread=%p team=%p\n", 0,
5007                     team->t.t_threads[0], team));
5008       __kmp_push_current_task_to_thread(team->t.t_threads[0], team, 0);
5009 
5010 #if OMP_40_ENABLED
5011 #if KMP_AFFINITY_SUPPORTED
5012       if ((team->t.t_size_changed == 0) &&
5013           (team->t.t_proc_bind == new_proc_bind)) {
5014         if (new_proc_bind == proc_bind_spread) {
5015           __kmp_partition_places(
5016               team, 1); // add flag to update only master for spread
5017         }
5018         KA_TRACE(200, ("__kmp_allocate_team: reusing hot team #%d bindings: "
5019                        "proc_bind = %d, partition = [%d,%d]\n",
5020                        team->t.t_id, new_proc_bind, team->t.t_first_place,
5021                        team->t.t_last_place));
5022       } else {
5023         KMP_CHECK_UPDATE(team->t.t_proc_bind, new_proc_bind);
5024         __kmp_partition_places(team);
5025       }
5026 #else
5027       KMP_CHECK_UPDATE(team->t.t_proc_bind, new_proc_bind);
5028 #endif /* KMP_AFFINITY_SUPPORTED */
5029 #endif /* OMP_40_ENABLED */
5030     } else if (team->t.t_nproc > new_nproc) {
5031       KA_TRACE(20,
5032                ("__kmp_allocate_team: decreasing hot team thread count to %d\n",
5033                 new_nproc));
5034 
5035       team->t.t_size_changed = 1;
5036 #if KMP_NESTED_HOT_TEAMS
5037       if (__kmp_hot_teams_mode == 0) {
5038         // AC: saved number of threads should correspond to team's value in this
5039         // mode, can be bigger in mode 1, when hot team has threads in reserve
5040         KMP_DEBUG_ASSERT(hot_teams[level].hot_team_nth == team->t.t_nproc);
5041         hot_teams[level].hot_team_nth = new_nproc;
5042 #endif // KMP_NESTED_HOT_TEAMS
5043         /* release the extra threads we don't need any more */
5044         for (f = new_nproc; f < team->t.t_nproc; f++) {
5045           KMP_DEBUG_ASSERT(team->t.t_threads[f]);
5046           if (__kmp_tasking_mode != tskm_immediate_exec) {
5047             // When decreasing team size, threads no longer in the team should
5048             // unref task team.
5049             team->t.t_threads[f]->th.th_task_team = NULL;
5050           }
5051           __kmp_free_thread(team->t.t_threads[f]);
5052           team->t.t_threads[f] = NULL;
5053         }
5054 #if KMP_NESTED_HOT_TEAMS
5055       } // (__kmp_hot_teams_mode == 0)
5056       else {
5057         // When keeping extra threads in team, switch threads to wait on own
5058         // b_go flag
5059         for (f = new_nproc; f < team->t.t_nproc; ++f) {
5060           KMP_DEBUG_ASSERT(team->t.t_threads[f]);
5061           kmp_balign_t *balign = team->t.t_threads[f]->th.th_bar;
5062           for (int b = 0; b < bs_last_barrier; ++b) {
5063             if (balign[b].bb.wait_flag == KMP_BARRIER_PARENT_FLAG) {
5064               balign[b].bb.wait_flag = KMP_BARRIER_SWITCH_TO_OWN_FLAG;
5065             }
5066             KMP_CHECK_UPDATE(balign[b].bb.leaf_kids, 0);
5067           }
5068         }
5069       }
5070 #endif // KMP_NESTED_HOT_TEAMS
5071       team->t.t_nproc = new_nproc;
5072       // TODO???: team->t.t_max_active_levels = new_max_active_levels;
5073       KMP_CHECK_UPDATE(team->t.t_sched.sched, new_icvs->sched.sched);
5074       __kmp_reinitialize_team(team, new_icvs,
5075                               root->r.r_uber_thread->th.th_ident);
5076 
5077       // Update remaining threads
5078       for (f = 0; f < new_nproc; ++f) {
5079         team->t.t_threads[f]->th.th_team_nproc = new_nproc;
5080       }
5081 
5082       // restore the current task state of the master thread: should be the
5083       // implicit task
5084       KF_TRACE(10, ("__kmp_allocate_team: T#%d, this_thread=%p team=%p\n", 0,
5085                     team->t.t_threads[0], team));
5086 
5087       __kmp_push_current_task_to_thread(team->t.t_threads[0], team, 0);
5088 
5089 #ifdef KMP_DEBUG
5090       for (f = 0; f < team->t.t_nproc; f++) {
5091         KMP_DEBUG_ASSERT(team->t.t_threads[f] &&
5092                          team->t.t_threads[f]->th.th_team_nproc ==
5093                              team->t.t_nproc);
5094       }
5095 #endif
5096 
5097 #if OMP_40_ENABLED
5098       KMP_CHECK_UPDATE(team->t.t_proc_bind, new_proc_bind);
5099 #if KMP_AFFINITY_SUPPORTED
5100       __kmp_partition_places(team);
5101 #endif
5102 #endif
5103     } else { // team->t.t_nproc < new_nproc
5104 #if KMP_OS_LINUX && KMP_AFFINITY_SUPPORTED
5105       kmp_affin_mask_t *old_mask;
5106       if (KMP_AFFINITY_CAPABLE()) {
5107         KMP_CPU_ALLOC(old_mask);
5108       }
5109 #endif
5110 
5111       KA_TRACE(20,
5112                ("__kmp_allocate_team: increasing hot team thread count to %d\n",
5113                 new_nproc));
5114 
5115       team->t.t_size_changed = 1;
5116 
5117 #if KMP_NESTED_HOT_TEAMS
5118       int avail_threads = hot_teams[level].hot_team_nth;
5119       if (new_nproc < avail_threads)
5120         avail_threads = new_nproc;
5121       kmp_info_t **other_threads = team->t.t_threads;
5122       for (f = team->t.t_nproc; f < avail_threads; ++f) {
5123         // Adjust barrier data of reserved threads (if any) of the team
5124         // Other data will be set in __kmp_initialize_info() below.
5125         int b;
5126         kmp_balign_t *balign = other_threads[f]->th.th_bar;
5127         for (b = 0; b < bs_last_barrier; ++b) {
5128           balign[b].bb.b_arrived = team->t.t_bar[b].b_arrived;
5129           KMP_DEBUG_ASSERT(balign[b].bb.wait_flag != KMP_BARRIER_PARENT_FLAG);
5130 #if USE_DEBUGGER
5131           balign[b].bb.b_worker_arrived = team->t.t_bar[b].b_team_arrived;
5132 #endif
5133         }
5134       }
5135       if (hot_teams[level].hot_team_nth >= new_nproc) {
5136         // we have all needed threads in reserve, no need to allocate any
5137         // this only possible in mode 1, cannot have reserved threads in mode 0
5138         KMP_DEBUG_ASSERT(__kmp_hot_teams_mode == 1);
5139         team->t.t_nproc = new_nproc; // just get reserved threads involved
5140       } else {
5141         // we may have some threads in reserve, but not enough
5142         team->t.t_nproc =
5143             hot_teams[level]
5144                 .hot_team_nth; // get reserved threads involved if any
5145         hot_teams[level].hot_team_nth = new_nproc; // adjust hot team max size
5146 #endif // KMP_NESTED_HOT_TEAMS
5147         if (team->t.t_max_nproc < new_nproc) {
5148           /* reallocate larger arrays */
5149           __kmp_reallocate_team_arrays(team, new_nproc);
5150           __kmp_reinitialize_team(team, new_icvs, NULL);
5151         }
5152 
5153 #if KMP_OS_LINUX && KMP_AFFINITY_SUPPORTED
5154         /* Temporarily set full mask for master thread before creation of
5155            workers. The reason is that workers inherit the affinity from master,
5156            so if a lot of workers are created on the single core quickly, they
5157            don't get a chance to set their own affinity for a long time. */
5158         __kmp_set_thread_affinity_mask_full_tmp(old_mask);
5159 #endif
5160 
5161         /* allocate new threads for the hot team */
5162         for (f = team->t.t_nproc; f < new_nproc; f++) {
5163           kmp_info_t *new_worker = __kmp_allocate_thread(root, team, f);
5164           KMP_DEBUG_ASSERT(new_worker);
5165           team->t.t_threads[f] = new_worker;
5166 
5167           KA_TRACE(20,
5168                    ("__kmp_allocate_team: team %d init T#%d arrived: "
5169                     "join=%llu, plain=%llu\n",
5170                     team->t.t_id, __kmp_gtid_from_tid(f, team), team->t.t_id, f,
5171                     team->t.t_bar[bs_forkjoin_barrier].b_arrived,
5172                     team->t.t_bar[bs_plain_barrier].b_arrived));
5173 
5174           { // Initialize barrier data for new threads.
5175             int b;
5176             kmp_balign_t *balign = new_worker->th.th_bar;
5177             for (b = 0; b < bs_last_barrier; ++b) {
5178               balign[b].bb.b_arrived = team->t.t_bar[b].b_arrived;
5179               KMP_DEBUG_ASSERT(balign[b].bb.wait_flag !=
5180                                KMP_BARRIER_PARENT_FLAG);
5181 #if USE_DEBUGGER
5182               balign[b].bb.b_worker_arrived = team->t.t_bar[b].b_team_arrived;
5183 #endif
5184             }
5185           }
5186         }
5187 
5188 #if KMP_OS_LINUX && KMP_AFFINITY_SUPPORTED
5189         if (KMP_AFFINITY_CAPABLE()) {
5190           /* Restore initial master thread's affinity mask */
5191           __kmp_set_system_affinity(old_mask, TRUE);
5192           KMP_CPU_FREE(old_mask);
5193         }
5194 #endif
5195 #if KMP_NESTED_HOT_TEAMS
5196       } // end of check of t_nproc vs. new_nproc vs. hot_team_nth
5197 #endif // KMP_NESTED_HOT_TEAMS
5198       /* make sure everyone is syncronized */
5199       int old_nproc = team->t.t_nproc; // save old value and use to update only
5200       // new threads below
5201       __kmp_initialize_team(team, new_nproc, new_icvs,
5202                             root->r.r_uber_thread->th.th_ident);
5203 
5204       /* reinitialize the threads */
5205       KMP_DEBUG_ASSERT(team->t.t_nproc == new_nproc);
5206       for (f = 0; f < team->t.t_nproc; ++f)
5207         __kmp_initialize_info(team->t.t_threads[f], team, f,
5208                               __kmp_gtid_from_tid(f, team));
5209 
5210       if (level) { // set th_task_state for new threads in nested hot team
5211         // __kmp_initialize_info() no longer zeroes th_task_state, so we should
5212         // only need to set the th_task_state for the new threads. th_task_state
5213         // for master thread will not be accurate until after this in
5214         // __kmp_fork_call(), so we look to the master's memo_stack to get the
5215         // correct value.
5216         for (f = old_nproc; f < team->t.t_nproc; ++f)
5217           team->t.t_threads[f]->th.th_task_state =
5218               team->t.t_threads[0]->th.th_task_state_memo_stack[level];
5219       } else { // set th_task_state for new threads in non-nested hot team
5220         int old_state =
5221             team->t.t_threads[0]->th.th_task_state; // copy master's state
5222         for (f = old_nproc; f < team->t.t_nproc; ++f)
5223           team->t.t_threads[f]->th.th_task_state = old_state;
5224       }
5225 
5226 #ifdef KMP_DEBUG
5227       for (f = 0; f < team->t.t_nproc; ++f) {
5228         KMP_DEBUG_ASSERT(team->t.t_threads[f] &&
5229                          team->t.t_threads[f]->th.th_team_nproc ==
5230                              team->t.t_nproc);
5231       }
5232 #endif
5233 
5234 #if OMP_40_ENABLED
5235       KMP_CHECK_UPDATE(team->t.t_proc_bind, new_proc_bind);
5236 #if KMP_AFFINITY_SUPPORTED
5237       __kmp_partition_places(team);
5238 #endif
5239 #endif
5240     } // Check changes in number of threads
5241 
5242 #if OMP_40_ENABLED
5243     kmp_info_t *master = team->t.t_threads[0];
5244     if (master->th.th_teams_microtask) {
5245       for (f = 1; f < new_nproc; ++f) {
5246         // propagate teams construct specific info to workers
5247         kmp_info_t *thr = team->t.t_threads[f];
5248         thr->th.th_teams_microtask = master->th.th_teams_microtask;
5249         thr->th.th_teams_level = master->th.th_teams_level;
5250         thr->th.th_teams_size = master->th.th_teams_size;
5251       }
5252     }
5253 #endif /* OMP_40_ENABLED */
5254 #if KMP_NESTED_HOT_TEAMS
5255     if (level) {
5256       // Sync barrier state for nested hot teams, not needed for outermost hot
5257       // team.
5258       for (f = 1; f < new_nproc; ++f) {
5259         kmp_info_t *thr = team->t.t_threads[f];
5260         int b;
5261         kmp_balign_t *balign = thr->th.th_bar;
5262         for (b = 0; b < bs_last_barrier; ++b) {
5263           balign[b].bb.b_arrived = team->t.t_bar[b].b_arrived;
5264           KMP_DEBUG_ASSERT(balign[b].bb.wait_flag != KMP_BARRIER_PARENT_FLAG);
5265 #if USE_DEBUGGER
5266           balign[b].bb.b_worker_arrived = team->t.t_bar[b].b_team_arrived;
5267 #endif
5268         }
5269       }
5270     }
5271 #endif // KMP_NESTED_HOT_TEAMS
5272 
5273     /* reallocate space for arguments if necessary */
5274     __kmp_alloc_argv_entries(argc, team, TRUE);
5275     KMP_CHECK_UPDATE(team->t.t_argc, argc);
5276     // The hot team re-uses the previous task team,
5277     // if untouched during the previous release->gather phase.
5278 
5279     KF_TRACE(10, (" hot_team = %p\n", team));
5280 
5281 #if KMP_DEBUG
5282     if (__kmp_tasking_mode != tskm_immediate_exec) {
5283       KA_TRACE(20, ("__kmp_allocate_team: hot team task_team[0] = %p "
5284                     "task_team[1] = %p after reinit\n",
5285                     team->t.t_task_team[0], team->t.t_task_team[1]));
5286     }
5287 #endif
5288 
5289 #if OMPT_SUPPORT
5290     __ompt_team_assign_id(team, ompt_parallel_data);
5291 #endif
5292 
5293     KMP_MB();
5294 
5295     return team;
5296   }
5297 
5298   /* next, let's try to take one from the team pool */
5299   KMP_MB();
5300   for (team = CCAST(kmp_team_t *, __kmp_team_pool); (team);) {
5301     /* TODO: consider resizing undersized teams instead of reaping them, now
5302        that we have a resizing mechanism */
5303     if (team->t.t_max_nproc >= max_nproc) {
5304       /* take this team from the team pool */
5305       __kmp_team_pool = team->t.t_next_pool;
5306 
5307       /* setup the team for fresh use */
5308       __kmp_initialize_team(team, new_nproc, new_icvs, NULL);
5309 
5310       KA_TRACE(20, ("__kmp_allocate_team: setting task_team[0] %p and "
5311                     "task_team[1] %p to NULL\n",
5312                     &team->t.t_task_team[0], &team->t.t_task_team[1]));
5313       team->t.t_task_team[0] = NULL;
5314       team->t.t_task_team[1] = NULL;
5315 
5316       /* reallocate space for arguments if necessary */
5317       __kmp_alloc_argv_entries(argc, team, TRUE);
5318       KMP_CHECK_UPDATE(team->t.t_argc, argc);
5319 
5320       KA_TRACE(
5321           20, ("__kmp_allocate_team: team %d init arrived: join=%u, plain=%u\n",
5322                team->t.t_id, KMP_INIT_BARRIER_STATE, KMP_INIT_BARRIER_STATE));
5323       { // Initialize barrier data.
5324         int b;
5325         for (b = 0; b < bs_last_barrier; ++b) {
5326           team->t.t_bar[b].b_arrived = KMP_INIT_BARRIER_STATE;
5327 #if USE_DEBUGGER
5328           team->t.t_bar[b].b_master_arrived = 0;
5329           team->t.t_bar[b].b_team_arrived = 0;
5330 #endif
5331         }
5332       }
5333 
5334 #if OMP_40_ENABLED
5335       team->t.t_proc_bind = new_proc_bind;
5336 #endif
5337 
5338       KA_TRACE(20, ("__kmp_allocate_team: using team from pool %d.\n",
5339                     team->t.t_id));
5340 
5341 #if OMPT_SUPPORT
5342       __ompt_team_assign_id(team, ompt_parallel_data);
5343 #endif
5344 
5345       KMP_MB();
5346 
5347       return team;
5348     }
5349 
5350     /* reap team if it is too small, then loop back and check the next one */
5351     // not sure if this is wise, but, will be redone during the hot-teams
5352     // rewrite.
5353     /* TODO: Use technique to find the right size hot-team, don't reap them */
5354     team = __kmp_reap_team(team);
5355     __kmp_team_pool = team;
5356   }
5357 
5358   /* nothing available in the pool, no matter, make a new team! */
5359   KMP_MB();
5360   team = (kmp_team_t *)__kmp_allocate(sizeof(kmp_team_t));
5361 
5362   /* and set it up */
5363   team->t.t_max_nproc = max_nproc;
5364   /* NOTE well, for some reason allocating one big buffer and dividing it up
5365      seems to really hurt performance a lot on the P4, so, let's not use this */
5366   __kmp_allocate_team_arrays(team, max_nproc);
5367 
5368   KA_TRACE(20, ("__kmp_allocate_team: making a new team\n"));
5369   __kmp_initialize_team(team, new_nproc, new_icvs, NULL);
5370 
5371   KA_TRACE(20, ("__kmp_allocate_team: setting task_team[0] %p and task_team[1] "
5372                 "%p to NULL\n",
5373                 &team->t.t_task_team[0], &team->t.t_task_team[1]));
5374   team->t.t_task_team[0] = NULL; // to be removed, as __kmp_allocate zeroes
5375   // memory, no need to duplicate
5376   team->t.t_task_team[1] = NULL; // to be removed, as __kmp_allocate zeroes
5377   // memory, no need to duplicate
5378 
5379   if (__kmp_storage_map) {
5380     __kmp_print_team_storage_map("team", team, team->t.t_id, new_nproc);
5381   }
5382 
5383   /* allocate space for arguments */
5384   __kmp_alloc_argv_entries(argc, team, FALSE);
5385   team->t.t_argc = argc;
5386 
5387   KA_TRACE(20,
5388            ("__kmp_allocate_team: team %d init arrived: join=%u, plain=%u\n",
5389             team->t.t_id, KMP_INIT_BARRIER_STATE, KMP_INIT_BARRIER_STATE));
5390   { // Initialize barrier data.
5391     int b;
5392     for (b = 0; b < bs_last_barrier; ++b) {
5393       team->t.t_bar[b].b_arrived = KMP_INIT_BARRIER_STATE;
5394 #if USE_DEBUGGER
5395       team->t.t_bar[b].b_master_arrived = 0;
5396       team->t.t_bar[b].b_team_arrived = 0;
5397 #endif
5398     }
5399   }
5400 
5401 #if OMP_40_ENABLED
5402   team->t.t_proc_bind = new_proc_bind;
5403 #endif
5404 
5405 #if OMPT_SUPPORT
5406   __ompt_team_assign_id(team, ompt_parallel_data);
5407   team->t.ompt_serialized_team_info = NULL;
5408 #endif
5409 
5410   KMP_MB();
5411 
5412   KA_TRACE(20, ("__kmp_allocate_team: done creating a new team %d.\n",
5413                 team->t.t_id));
5414 
5415   return team;
5416 }
5417 
5418 /* TODO implement hot-teams at all levels */
5419 /* TODO implement lazy thread release on demand (disband request) */
5420 
5421 /* free the team.  return it to the team pool.  release all the threads
5422  * associated with it */
5423 void __kmp_free_team(kmp_root_t *root,
5424                      kmp_team_t *team USE_NESTED_HOT_ARG(kmp_info_t *master)) {
5425   int f;
5426   KA_TRACE(20, ("__kmp_free_team: T#%d freeing team %d\n", __kmp_get_gtid(),
5427                 team->t.t_id));
5428 
5429   /* verify state */
5430   KMP_DEBUG_ASSERT(root);
5431   KMP_DEBUG_ASSERT(team);
5432   KMP_DEBUG_ASSERT(team->t.t_nproc <= team->t.t_max_nproc);
5433   KMP_DEBUG_ASSERT(team->t.t_threads);
5434 
5435   int use_hot_team = team == root->r.r_hot_team;
5436 #if KMP_NESTED_HOT_TEAMS
5437   int level;
5438   kmp_hot_team_ptr_t *hot_teams;
5439   if (master) {
5440     level = team->t.t_active_level - 1;
5441     if (master->th.th_teams_microtask) { // in teams construct?
5442       if (master->th.th_teams_size.nteams > 1) {
5443         ++level; // level was not increased in teams construct for
5444         // team_of_masters
5445       }
5446       if (team->t.t_pkfn != (microtask_t)__kmp_teams_master &&
5447           master->th.th_teams_level == team->t.t_level) {
5448         ++level; // level was not increased in teams construct for
5449         // team_of_workers before the parallel
5450       } // team->t.t_level will be increased inside parallel
5451     }
5452     hot_teams = master->th.th_hot_teams;
5453     if (level < __kmp_hot_teams_max_level) {
5454       KMP_DEBUG_ASSERT(team == hot_teams[level].hot_team);
5455       use_hot_team = 1;
5456     }
5457   }
5458 #endif // KMP_NESTED_HOT_TEAMS
5459 
5460   /* team is done working */
5461   TCW_SYNC_PTR(team->t.t_pkfn,
5462                NULL); // Important for Debugging Support Library.
5463 #if KMP_OS_WINDOWS
5464   team->t.t_copyin_counter = 0; // init counter for possible reuse
5465 #endif
5466   // Do not reset pointer to parent team to NULL for hot teams.
5467 
5468   /* if we are non-hot team, release our threads */
5469   if (!use_hot_team) {
5470     if (__kmp_tasking_mode != tskm_immediate_exec) {
5471       // Wait for threads to reach reapable state
5472       for (f = 1; f < team->t.t_nproc; ++f) {
5473         KMP_DEBUG_ASSERT(team->t.t_threads[f]);
5474         kmp_info_t *th = team->t.t_threads[f];
5475         volatile kmp_uint32 *state = &th->th.th_reap_state;
5476         while (*state != KMP_SAFE_TO_REAP) {
5477 #if KMP_OS_WINDOWS
5478           // On Windows a thread can be killed at any time, check this
5479           DWORD ecode;
5480           if (!__kmp_is_thread_alive(th, &ecode)) {
5481             *state = KMP_SAFE_TO_REAP; // reset the flag for dead thread
5482             break;
5483           }
5484 #endif
5485           // first check if thread is sleeping
5486           kmp_flag_64 fl(&th->th.th_bar[bs_forkjoin_barrier].bb.b_go, th);
5487           if (fl.is_sleeping())
5488             fl.resume(__kmp_gtid_from_thread(th));
5489           KMP_CPU_PAUSE();
5490         }
5491       }
5492 
5493       // Delete task teams
5494       int tt_idx;
5495       for (tt_idx = 0; tt_idx < 2; ++tt_idx) {
5496         kmp_task_team_t *task_team = team->t.t_task_team[tt_idx];
5497         if (task_team != NULL) {
5498           for (f = 0; f < team->t.t_nproc; ++f) { // threads unref task teams
5499             KMP_DEBUG_ASSERT(team->t.t_threads[f]);
5500             team->t.t_threads[f]->th.th_task_team = NULL;
5501           }
5502           KA_TRACE(
5503               20,
5504               ("__kmp_free_team: T#%d deactivating task_team %p on team %d\n",
5505                __kmp_get_gtid(), task_team, team->t.t_id));
5506 #if KMP_NESTED_HOT_TEAMS
5507           __kmp_free_task_team(master, task_team);
5508 #endif
5509           team->t.t_task_team[tt_idx] = NULL;
5510         }
5511       }
5512     }
5513 
5514     // Reset pointer to parent team only for non-hot teams.
5515     team->t.t_parent = NULL;
5516     team->t.t_level = 0;
5517     team->t.t_active_level = 0;
5518 
5519     /* free the worker threads */
5520     for (f = 1; f < team->t.t_nproc; ++f) {
5521       KMP_DEBUG_ASSERT(team->t.t_threads[f]);
5522       __kmp_free_thread(team->t.t_threads[f]);
5523       team->t.t_threads[f] = NULL;
5524     }
5525 
5526     /* put the team back in the team pool */
5527     /* TODO limit size of team pool, call reap_team if pool too large */
5528     team->t.t_next_pool = CCAST(kmp_team_t *, __kmp_team_pool);
5529     __kmp_team_pool = (volatile kmp_team_t *)team;
5530   } else { // Check if team was created for the masters in a teams construct
5531     // See if first worker is a CG root
5532     KMP_DEBUG_ASSERT(team->t.t_threads[1] &&
5533                      team->t.t_threads[1]->th.th_cg_roots);
5534     if (team->t.t_threads[1]->th.th_cg_roots->cg_root == team->t.t_threads[1]) {
5535       // Clean up the CG root nodes on workers so that this team can be re-used
5536       for (f = 1; f < team->t.t_nproc; ++f) {
5537         kmp_info_t *thr = team->t.t_threads[f];
5538         KMP_DEBUG_ASSERT(thr && thr->th.th_cg_roots &&
5539                          thr->th.th_cg_roots->cg_root == thr);
5540         // Pop current CG root off list
5541         kmp_cg_root_t *tmp = thr->th.th_cg_roots;
5542         thr->th.th_cg_roots = tmp->up;
5543         KA_TRACE(100, ("__kmp_free_team: Thread %p popping node %p and moving"
5544                        " up to node %p. cg_nthreads was %d\n",
5545                        thr, tmp, thr->th.th_cg_roots, tmp->cg_nthreads));
5546         __kmp_free(tmp);
5547         // Restore current task's thread_limit from CG root
5548         if (thr->th.th_cg_roots)
5549           thr->th.th_current_task->td_icvs.thread_limit =
5550               thr->th.th_cg_roots->cg_thread_limit;
5551       }
5552     }
5553   }
5554 
5555   KMP_MB();
5556 }
5557 
5558 /* reap the team.  destroy it, reclaim all its resources and free its memory */
5559 kmp_team_t *__kmp_reap_team(kmp_team_t *team) {
5560   kmp_team_t *next_pool = team->t.t_next_pool;
5561 
5562   KMP_DEBUG_ASSERT(team);
5563   KMP_DEBUG_ASSERT(team->t.t_dispatch);
5564   KMP_DEBUG_ASSERT(team->t.t_disp_buffer);
5565   KMP_DEBUG_ASSERT(team->t.t_threads);
5566   KMP_DEBUG_ASSERT(team->t.t_argv);
5567 
5568   /* TODO clean the threads that are a part of this? */
5569 
5570   /* free stuff */
5571   __kmp_free_team_arrays(team);
5572   if (team->t.t_argv != &team->t.t_inline_argv[0])
5573     __kmp_free((void *)team->t.t_argv);
5574   __kmp_free(team);
5575 
5576   KMP_MB();
5577   return next_pool;
5578 }
5579 
5580 // Free the thread.  Don't reap it, just place it on the pool of available
5581 // threads.
5582 //
5583 // Changes for Quad issue 527845: We need a predictable OMP tid <-> gtid
5584 // binding for the affinity mechanism to be useful.
5585 //
5586 // Now, we always keep the free list (__kmp_thread_pool) sorted by gtid.
5587 // However, we want to avoid a potential performance problem by always
5588 // scanning through the list to find the correct point at which to insert
5589 // the thread (potential N**2 behavior).  To do this we keep track of the
5590 // last place a thread struct was inserted (__kmp_thread_pool_insert_pt).
5591 // With single-level parallelism, threads will always be added to the tail
5592 // of the list, kept track of by __kmp_thread_pool_insert_pt.  With nested
5593 // parallelism, all bets are off and we may need to scan through the entire
5594 // free list.
5595 //
5596 // This change also has a potentially large performance benefit, for some
5597 // applications.  Previously, as threads were freed from the hot team, they
5598 // would be placed back on the free list in inverse order.  If the hot team
5599 // grew back to it's original size, then the freed thread would be placed
5600 // back on the hot team in reverse order.  This could cause bad cache
5601 // locality problems on programs where the size of the hot team regularly
5602 // grew and shrunk.
5603 //
5604 // Now, for single-level parallelism, the OMP tid is alway == gtid.
5605 void __kmp_free_thread(kmp_info_t *this_th) {
5606   int gtid;
5607   kmp_info_t **scan;
5608 
5609   KA_TRACE(20, ("__kmp_free_thread: T#%d putting T#%d back on free pool.\n",
5610                 __kmp_get_gtid(), this_th->th.th_info.ds.ds_gtid));
5611 
5612   KMP_DEBUG_ASSERT(this_th);
5613 
5614   // When moving thread to pool, switch thread to wait on own b_go flag, and
5615   // uninitialized (NULL team).
5616   int b;
5617   kmp_balign_t *balign = this_th->th.th_bar;
5618   for (b = 0; b < bs_last_barrier; ++b) {
5619     if (balign[b].bb.wait_flag == KMP_BARRIER_PARENT_FLAG)
5620       balign[b].bb.wait_flag = KMP_BARRIER_SWITCH_TO_OWN_FLAG;
5621     balign[b].bb.team = NULL;
5622     balign[b].bb.leaf_kids = 0;
5623   }
5624   this_th->th.th_task_state = 0;
5625   this_th->th.th_reap_state = KMP_SAFE_TO_REAP;
5626 
5627   /* put thread back on the free pool */
5628   TCW_PTR(this_th->th.th_team, NULL);
5629   TCW_PTR(this_th->th.th_root, NULL);
5630   TCW_PTR(this_th->th.th_dispatch, NULL); /* NOT NEEDED */
5631 
5632   while (this_th->th.th_cg_roots) {
5633     this_th->th.th_cg_roots->cg_nthreads--;
5634     KA_TRACE(100, ("__kmp_free_thread: Thread %p decrement cg_nthreads on node"
5635                    " %p of thread  %p to %d\n",
5636                    this_th, this_th->th.th_cg_roots,
5637                    this_th->th.th_cg_roots->cg_root,
5638                    this_th->th.th_cg_roots->cg_nthreads));
5639     kmp_cg_root_t *tmp = this_th->th.th_cg_roots;
5640     if (tmp->cg_root == this_th) { // Thread is a cg_root
5641       KMP_DEBUG_ASSERT(tmp->cg_nthreads == 0);
5642       KA_TRACE(
5643           5, ("__kmp_free_thread: Thread %p freeing node %p\n", this_th, tmp));
5644       this_th->th.th_cg_roots = tmp->up;
5645       __kmp_free(tmp);
5646     } else { // Worker thread
5647       this_th->th.th_cg_roots = NULL;
5648       break;
5649     }
5650   }
5651 
5652   /* If the implicit task assigned to this thread can be used by other threads
5653    * -> multiple threads can share the data and try to free the task at
5654    * __kmp_reap_thread at exit. This duplicate use of the task data can happen
5655    * with higher probability when hot team is disabled but can occurs even when
5656    * the hot team is enabled */
5657   __kmp_free_implicit_task(this_th);
5658   this_th->th.th_current_task = NULL;
5659 
5660   // If the __kmp_thread_pool_insert_pt is already past the new insert
5661   // point, then we need to re-scan the entire list.
5662   gtid = this_th->th.th_info.ds.ds_gtid;
5663   if (__kmp_thread_pool_insert_pt != NULL) {
5664     KMP_DEBUG_ASSERT(__kmp_thread_pool != NULL);
5665     if (__kmp_thread_pool_insert_pt->th.th_info.ds.ds_gtid > gtid) {
5666       __kmp_thread_pool_insert_pt = NULL;
5667     }
5668   }
5669 
5670   // Scan down the list to find the place to insert the thread.
5671   // scan is the address of a link in the list, possibly the address of
5672   // __kmp_thread_pool itself.
5673   //
5674   // In the absence of nested parallism, the for loop will have 0 iterations.
5675   if (__kmp_thread_pool_insert_pt != NULL) {
5676     scan = &(__kmp_thread_pool_insert_pt->th.th_next_pool);
5677   } else {
5678     scan = CCAST(kmp_info_t **, &__kmp_thread_pool);
5679   }
5680   for (; (*scan != NULL) && ((*scan)->th.th_info.ds.ds_gtid < gtid);
5681        scan = &((*scan)->th.th_next_pool))
5682     ;
5683 
5684   // Insert the new element on the list, and set __kmp_thread_pool_insert_pt
5685   // to its address.
5686   TCW_PTR(this_th->th.th_next_pool, *scan);
5687   __kmp_thread_pool_insert_pt = *scan = this_th;
5688   KMP_DEBUG_ASSERT((this_th->th.th_next_pool == NULL) ||
5689                    (this_th->th.th_info.ds.ds_gtid <
5690                     this_th->th.th_next_pool->th.th_info.ds.ds_gtid));
5691   TCW_4(this_th->th.th_in_pool, TRUE);
5692   __kmp_thread_pool_nth++;
5693 
5694   TCW_4(__kmp_nth, __kmp_nth - 1);
5695 
5696 #ifdef KMP_ADJUST_BLOCKTIME
5697   /* Adjust blocktime back to user setting or default if necessary */
5698   /* Middle initialization might never have occurred                */
5699   if (!__kmp_env_blocktime && (__kmp_avail_proc > 0)) {
5700     KMP_DEBUG_ASSERT(__kmp_avail_proc > 0);
5701     if (__kmp_nth <= __kmp_avail_proc) {
5702       __kmp_zero_bt = FALSE;
5703     }
5704   }
5705 #endif /* KMP_ADJUST_BLOCKTIME */
5706 
5707   KMP_MB();
5708 }
5709 
5710 /* ------------------------------------------------------------------------ */
5711 
5712 void *__kmp_launch_thread(kmp_info_t *this_thr) {
5713   int gtid = this_thr->th.th_info.ds.ds_gtid;
5714   /*    void                 *stack_data;*/
5715   kmp_team_t *(*volatile pteam);
5716 
5717   KMP_MB();
5718   KA_TRACE(10, ("__kmp_launch_thread: T#%d start\n", gtid));
5719 
5720   if (__kmp_env_consistency_check) {
5721     this_thr->th.th_cons = __kmp_allocate_cons_stack(gtid); // ATT: Memory leak?
5722   }
5723 
5724 #if OMPT_SUPPORT
5725   ompt_data_t *thread_data;
5726   if (ompt_enabled.enabled) {
5727     thread_data = &(this_thr->th.ompt_thread_info.thread_data);
5728     *thread_data = ompt_data_none;
5729 
5730     this_thr->th.ompt_thread_info.state = ompt_state_overhead;
5731     this_thr->th.ompt_thread_info.wait_id = 0;
5732     this_thr->th.ompt_thread_info.idle_frame = OMPT_GET_FRAME_ADDRESS(0);
5733     if (ompt_enabled.ompt_callback_thread_begin) {
5734       ompt_callbacks.ompt_callback(ompt_callback_thread_begin)(
5735           ompt_thread_worker, thread_data);
5736     }
5737   }
5738 #endif
5739 
5740 #if OMPT_SUPPORT
5741   if (ompt_enabled.enabled) {
5742     this_thr->th.ompt_thread_info.state = ompt_state_idle;
5743   }
5744 #endif
5745   /* This is the place where threads wait for work */
5746   while (!TCR_4(__kmp_global.g.g_done)) {
5747     KMP_DEBUG_ASSERT(this_thr == __kmp_threads[gtid]);
5748     KMP_MB();
5749 
5750     /* wait for work to do */
5751     KA_TRACE(20, ("__kmp_launch_thread: T#%d waiting for work\n", gtid));
5752 
5753     /* No tid yet since not part of a team */
5754     __kmp_fork_barrier(gtid, KMP_GTID_DNE);
5755 
5756 #if OMPT_SUPPORT
5757     if (ompt_enabled.enabled) {
5758       this_thr->th.ompt_thread_info.state = ompt_state_overhead;
5759     }
5760 #endif
5761 
5762     pteam = (kmp_team_t * (*))(&this_thr->th.th_team);
5763 
5764     /* have we been allocated? */
5765     if (TCR_SYNC_PTR(*pteam) && !TCR_4(__kmp_global.g.g_done)) {
5766       /* we were just woken up, so run our new task */
5767       if (TCR_SYNC_PTR((*pteam)->t.t_pkfn) != NULL) {
5768         int rc;
5769         KA_TRACE(20,
5770                  ("__kmp_launch_thread: T#%d(%d:%d) invoke microtask = %p\n",
5771                   gtid, (*pteam)->t.t_id, __kmp_tid_from_gtid(gtid),
5772                   (*pteam)->t.t_pkfn));
5773 
5774         updateHWFPControl(*pteam);
5775 
5776 #if OMPT_SUPPORT
5777         if (ompt_enabled.enabled) {
5778           this_thr->th.ompt_thread_info.state = ompt_state_work_parallel;
5779         }
5780 #endif
5781 
5782         rc = (*pteam)->t.t_invoke(gtid);
5783         KMP_ASSERT(rc);
5784 
5785         KMP_MB();
5786         KA_TRACE(20, ("__kmp_launch_thread: T#%d(%d:%d) done microtask = %p\n",
5787                       gtid, (*pteam)->t.t_id, __kmp_tid_from_gtid(gtid),
5788                       (*pteam)->t.t_pkfn));
5789       }
5790 #if OMPT_SUPPORT
5791       if (ompt_enabled.enabled) {
5792         /* no frame set while outside task */
5793         __ompt_get_task_info_object(0)->frame.exit_frame = ompt_data_none;
5794 
5795         this_thr->th.ompt_thread_info.state = ompt_state_overhead;
5796       }
5797 #endif
5798       /* join barrier after parallel region */
5799       __kmp_join_barrier(gtid);
5800     }
5801   }
5802   TCR_SYNC_PTR((intptr_t)__kmp_global.g.g_done);
5803 
5804 #if OMPT_SUPPORT
5805   if (ompt_enabled.ompt_callback_thread_end) {
5806     ompt_callbacks.ompt_callback(ompt_callback_thread_end)(thread_data);
5807   }
5808 #endif
5809 
5810   this_thr->th.th_task_team = NULL;
5811   /* run the destructors for the threadprivate data for this thread */
5812   __kmp_common_destroy_gtid(gtid);
5813 
5814   KA_TRACE(10, ("__kmp_launch_thread: T#%d done\n", gtid));
5815   KMP_MB();
5816   return this_thr;
5817 }
5818 
5819 /* ------------------------------------------------------------------------ */
5820 
5821 void __kmp_internal_end_dest(void *specific_gtid) {
5822 #if KMP_COMPILER_ICC
5823 #pragma warning(push)
5824 #pragma warning(disable : 810) // conversion from "void *" to "int" may lose
5825 // significant bits
5826 #endif
5827   // Make sure no significant bits are lost
5828   int gtid = (kmp_intptr_t)specific_gtid - 1;
5829 #if KMP_COMPILER_ICC
5830 #pragma warning(pop)
5831 #endif
5832 
5833   KA_TRACE(30, ("__kmp_internal_end_dest: T#%d\n", gtid));
5834   /* NOTE: the gtid is stored as gitd+1 in the thread-local-storage
5835    * this is because 0 is reserved for the nothing-stored case */
5836 
5837   /* josh: One reason for setting the gtid specific data even when it is being
5838      destroyed by pthread is to allow gtid lookup through thread specific data
5839      (__kmp_gtid_get_specific).  Some of the code, especially stat code,
5840      that gets executed in the call to __kmp_internal_end_thread, actually
5841      gets the gtid through the thread specific data.  Setting it here seems
5842      rather inelegant and perhaps wrong, but allows __kmp_internal_end_thread
5843      to run smoothly.
5844      todo: get rid of this after we remove the dependence on
5845      __kmp_gtid_get_specific  */
5846   if (gtid >= 0 && KMP_UBER_GTID(gtid))
5847     __kmp_gtid_set_specific(gtid);
5848 #ifdef KMP_TDATA_GTID
5849   __kmp_gtid = gtid;
5850 #endif
5851   __kmp_internal_end_thread(gtid);
5852 }
5853 
5854 #if KMP_OS_UNIX && KMP_DYNAMIC_LIB
5855 
5856 // 2009-09-08 (lev): It looks the destructor does not work. In simple test cases
5857 // destructors work perfectly, but in real libomp.so I have no evidence it is
5858 // ever called. However, -fini linker option in makefile.mk works fine.
5859 
5860 __attribute__((destructor)) void __kmp_internal_end_dtor(void) {
5861   __kmp_internal_end_atexit();
5862 }
5863 
5864 void __kmp_internal_end_fini(void) { __kmp_internal_end_atexit(); }
5865 
5866 #endif
5867 
5868 /* [Windows] josh: when the atexit handler is called, there may still be more
5869    than one thread alive */
5870 void __kmp_internal_end_atexit(void) {
5871   KA_TRACE(30, ("__kmp_internal_end_atexit\n"));
5872   /* [Windows]
5873      josh: ideally, we want to completely shutdown the library in this atexit
5874      handler, but stat code that depends on thread specific data for gtid fails
5875      because that data becomes unavailable at some point during the shutdown, so
5876      we call __kmp_internal_end_thread instead. We should eventually remove the
5877      dependency on __kmp_get_specific_gtid in the stat code and use
5878      __kmp_internal_end_library to cleanly shutdown the library.
5879 
5880      // TODO: Can some of this comment about GVS be removed?
5881      I suspect that the offending stat code is executed when the calling thread
5882      tries to clean up a dead root thread's data structures, resulting in GVS
5883      code trying to close the GVS structures for that thread, but since the stat
5884      code uses __kmp_get_specific_gtid to get the gtid with the assumption that
5885      the calling thread is cleaning up itself instead of another thread, it get
5886      confused. This happens because allowing a thread to unregister and cleanup
5887      another thread is a recent modification for addressing an issue.
5888      Based on the current design (20050722), a thread may end up
5889      trying to unregister another thread only if thread death does not trigger
5890      the calling of __kmp_internal_end_thread.  For Linux* OS, there is the
5891      thread specific data destructor function to detect thread death. For
5892      Windows dynamic, there is DllMain(THREAD_DETACH). For Windows static, there
5893      is nothing.  Thus, the workaround is applicable only for Windows static
5894      stat library. */
5895   __kmp_internal_end_library(-1);
5896 #if KMP_OS_WINDOWS
5897   __kmp_close_console();
5898 #endif
5899 }
5900 
5901 static void __kmp_reap_thread(kmp_info_t *thread, int is_root) {
5902   // It is assumed __kmp_forkjoin_lock is acquired.
5903 
5904   int gtid;
5905 
5906   KMP_DEBUG_ASSERT(thread != NULL);
5907 
5908   gtid = thread->th.th_info.ds.ds_gtid;
5909 
5910   if (!is_root) {
5911     if (__kmp_dflt_blocktime != KMP_MAX_BLOCKTIME) {
5912       /* Assume the threads are at the fork barrier here */
5913       KA_TRACE(
5914           20, ("__kmp_reap_thread: releasing T#%d from fork barrier for reap\n",
5915                gtid));
5916       /* Need release fence here to prevent seg faults for tree forkjoin barrier
5917        * (GEH) */
5918       ANNOTATE_HAPPENS_BEFORE(thread);
5919       kmp_flag_64 flag(&thread->th.th_bar[bs_forkjoin_barrier].bb.b_go, thread);
5920       __kmp_release_64(&flag);
5921     }
5922 
5923     // Terminate OS thread.
5924     __kmp_reap_worker(thread);
5925 
5926     // The thread was killed asynchronously.  If it was actively
5927     // spinning in the thread pool, decrement the global count.
5928     //
5929     // There is a small timing hole here - if the worker thread was just waking
5930     // up after sleeping in the pool, had reset it's th_active_in_pool flag but
5931     // not decremented the global counter __kmp_thread_pool_active_nth yet, then
5932     // the global counter might not get updated.
5933     //
5934     // Currently, this can only happen as the library is unloaded,
5935     // so there are no harmful side effects.
5936     if (thread->th.th_active_in_pool) {
5937       thread->th.th_active_in_pool = FALSE;
5938       KMP_ATOMIC_DEC(&__kmp_thread_pool_active_nth);
5939       KMP_DEBUG_ASSERT(__kmp_thread_pool_active_nth >= 0);
5940     }
5941 
5942     // Decrement # of [worker] threads in the pool.
5943     KMP_DEBUG_ASSERT(__kmp_thread_pool_nth > 0);
5944     --__kmp_thread_pool_nth;
5945   }
5946 
5947   __kmp_free_implicit_task(thread);
5948 
5949 // Free the fast memory for tasking
5950 #if USE_FAST_MEMORY
5951   __kmp_free_fast_memory(thread);
5952 #endif /* USE_FAST_MEMORY */
5953 
5954   __kmp_suspend_uninitialize_thread(thread);
5955 
5956   KMP_DEBUG_ASSERT(__kmp_threads[gtid] == thread);
5957   TCW_SYNC_PTR(__kmp_threads[gtid], NULL);
5958 
5959   --__kmp_all_nth;
5960 // __kmp_nth was decremented when thread is added to the pool.
5961 
5962 #ifdef KMP_ADJUST_BLOCKTIME
5963   /* Adjust blocktime back to user setting or default if necessary */
5964   /* Middle initialization might never have occurred                */
5965   if (!__kmp_env_blocktime && (__kmp_avail_proc > 0)) {
5966     KMP_DEBUG_ASSERT(__kmp_avail_proc > 0);
5967     if (__kmp_nth <= __kmp_avail_proc) {
5968       __kmp_zero_bt = FALSE;
5969     }
5970   }
5971 #endif /* KMP_ADJUST_BLOCKTIME */
5972 
5973   /* free the memory being used */
5974   if (__kmp_env_consistency_check) {
5975     if (thread->th.th_cons) {
5976       __kmp_free_cons_stack(thread->th.th_cons);
5977       thread->th.th_cons = NULL;
5978     }
5979   }
5980 
5981   if (thread->th.th_pri_common != NULL) {
5982     __kmp_free(thread->th.th_pri_common);
5983     thread->th.th_pri_common = NULL;
5984   }
5985 
5986   if (thread->th.th_task_state_memo_stack != NULL) {
5987     __kmp_free(thread->th.th_task_state_memo_stack);
5988     thread->th.th_task_state_memo_stack = NULL;
5989   }
5990 
5991 #if KMP_USE_BGET
5992   if (thread->th.th_local.bget_data != NULL) {
5993     __kmp_finalize_bget(thread);
5994   }
5995 #endif
5996 
5997 #if KMP_AFFINITY_SUPPORTED
5998   if (thread->th.th_affin_mask != NULL) {
5999     KMP_CPU_FREE(thread->th.th_affin_mask);
6000     thread->th.th_affin_mask = NULL;
6001   }
6002 #endif /* KMP_AFFINITY_SUPPORTED */
6003 
6004 #if KMP_USE_HIER_SCHED
6005   if (thread->th.th_hier_bar_data != NULL) {
6006     __kmp_free(thread->th.th_hier_bar_data);
6007     thread->th.th_hier_bar_data = NULL;
6008   }
6009 #endif
6010 
6011   __kmp_reap_team(thread->th.th_serial_team);
6012   thread->th.th_serial_team = NULL;
6013   __kmp_free(thread);
6014 
6015   KMP_MB();
6016 
6017 } // __kmp_reap_thread
6018 
6019 static void __kmp_internal_end(void) {
6020   int i;
6021 
6022   /* First, unregister the library */
6023   __kmp_unregister_library();
6024 
6025 #if KMP_OS_WINDOWS
6026   /* In Win static library, we can't tell when a root actually dies, so we
6027      reclaim the data structures for any root threads that have died but not
6028      unregistered themselves, in order to shut down cleanly.
6029      In Win dynamic library we also can't tell when a thread dies.  */
6030   __kmp_reclaim_dead_roots(); // AC: moved here to always clean resources of
6031 // dead roots
6032 #endif
6033 
6034   for (i = 0; i < __kmp_threads_capacity; i++)
6035     if (__kmp_root[i])
6036       if (__kmp_root[i]->r.r_active)
6037         break;
6038   KMP_MB(); /* Flush all pending memory write invalidates.  */
6039   TCW_SYNC_4(__kmp_global.g.g_done, TRUE);
6040 
6041   if (i < __kmp_threads_capacity) {
6042 #if KMP_USE_MONITOR
6043     // 2009-09-08 (lev): Other alive roots found. Why do we kill the monitor??
6044     KMP_MB(); /* Flush all pending memory write invalidates.  */
6045 
6046     // Need to check that monitor was initialized before reaping it. If we are
6047     // called form __kmp_atfork_child (which sets __kmp_init_parallel = 0), then
6048     // __kmp_monitor will appear to contain valid data, but it is only valid in
6049     // the parent process, not the child.
6050     // New behavior (201008): instead of keying off of the flag
6051     // __kmp_init_parallel, the monitor thread creation is keyed off
6052     // of the new flag __kmp_init_monitor.
6053     __kmp_acquire_bootstrap_lock(&__kmp_monitor_lock);
6054     if (TCR_4(__kmp_init_monitor)) {
6055       __kmp_reap_monitor(&__kmp_monitor);
6056       TCW_4(__kmp_init_monitor, 0);
6057     }
6058     __kmp_release_bootstrap_lock(&__kmp_monitor_lock);
6059     KA_TRACE(10, ("__kmp_internal_end: monitor reaped\n"));
6060 #endif // KMP_USE_MONITOR
6061   } else {
6062 /* TODO move this to cleanup code */
6063 #ifdef KMP_DEBUG
6064     /* make sure that everything has properly ended */
6065     for (i = 0; i < __kmp_threads_capacity; i++) {
6066       if (__kmp_root[i]) {
6067         //                    KMP_ASSERT( ! KMP_UBER_GTID( i ) );         // AC:
6068         //                    there can be uber threads alive here
6069         KMP_ASSERT(!__kmp_root[i]->r.r_active); // TODO: can they be active?
6070       }
6071     }
6072 #endif
6073 
6074     KMP_MB();
6075 
6076     // Reap the worker threads.
6077     // This is valid for now, but be careful if threads are reaped sooner.
6078     while (__kmp_thread_pool != NULL) { // Loop thru all the thread in the pool.
6079       // Get the next thread from the pool.
6080       kmp_info_t *thread = CCAST(kmp_info_t *, __kmp_thread_pool);
6081       __kmp_thread_pool = thread->th.th_next_pool;
6082       // Reap it.
6083       KMP_DEBUG_ASSERT(thread->th.th_reap_state == KMP_SAFE_TO_REAP);
6084       thread->th.th_next_pool = NULL;
6085       thread->th.th_in_pool = FALSE;
6086       __kmp_reap_thread(thread, 0);
6087     }
6088     __kmp_thread_pool_insert_pt = NULL;
6089 
6090     // Reap teams.
6091     while (__kmp_team_pool != NULL) { // Loop thru all the teams in the pool.
6092       // Get the next team from the pool.
6093       kmp_team_t *team = CCAST(kmp_team_t *, __kmp_team_pool);
6094       __kmp_team_pool = team->t.t_next_pool;
6095       // Reap it.
6096       team->t.t_next_pool = NULL;
6097       __kmp_reap_team(team);
6098     }
6099 
6100     __kmp_reap_task_teams();
6101 
6102 #if KMP_OS_UNIX
6103     // Threads that are not reaped should not access any resources since they
6104     // are going to be deallocated soon, so the shutdown sequence should wait
6105     // until all threads either exit the final spin-waiting loop or begin
6106     // sleeping after the given blocktime.
6107     for (i = 0; i < __kmp_threads_capacity; i++) {
6108       kmp_info_t *thr = __kmp_threads[i];
6109       while (thr && KMP_ATOMIC_LD_ACQ(&thr->th.th_blocking))
6110         KMP_CPU_PAUSE();
6111     }
6112 #endif
6113 
6114     for (i = 0; i < __kmp_threads_capacity; ++i) {
6115       // TBD: Add some checking...
6116       // Something like KMP_DEBUG_ASSERT( __kmp_thread[ i ] == NULL );
6117     }
6118 
6119     /* Make sure all threadprivate destructors get run by joining with all
6120        worker threads before resetting this flag */
6121     TCW_SYNC_4(__kmp_init_common, FALSE);
6122 
6123     KA_TRACE(10, ("__kmp_internal_end: all workers reaped\n"));
6124     KMP_MB();
6125 
6126 #if KMP_USE_MONITOR
6127     // See note above: One of the possible fixes for CQ138434 / CQ140126
6128     //
6129     // FIXME: push both code fragments down and CSE them?
6130     // push them into __kmp_cleanup() ?
6131     __kmp_acquire_bootstrap_lock(&__kmp_monitor_lock);
6132     if (TCR_4(__kmp_init_monitor)) {
6133       __kmp_reap_monitor(&__kmp_monitor);
6134       TCW_4(__kmp_init_monitor, 0);
6135     }
6136     __kmp_release_bootstrap_lock(&__kmp_monitor_lock);
6137     KA_TRACE(10, ("__kmp_internal_end: monitor reaped\n"));
6138 #endif
6139   } /* else !__kmp_global.t_active */
6140   TCW_4(__kmp_init_gtid, FALSE);
6141   KMP_MB(); /* Flush all pending memory write invalidates.  */
6142 
6143   __kmp_cleanup();
6144 #if OMPT_SUPPORT
6145   ompt_fini();
6146 #endif
6147 }
6148 
6149 void __kmp_internal_end_library(int gtid_req) {
6150   /* if we have already cleaned up, don't try again, it wouldn't be pretty */
6151   /* this shouldn't be a race condition because __kmp_internal_end() is the
6152      only place to clear __kmp_serial_init */
6153   /* we'll check this later too, after we get the lock */
6154   // 2009-09-06: We do not set g_abort without setting g_done. This check looks
6155   // redundaant, because the next check will work in any case.
6156   if (__kmp_global.g.g_abort) {
6157     KA_TRACE(11, ("__kmp_internal_end_library: abort, exiting\n"));
6158     /* TODO abort? */
6159     return;
6160   }
6161   if (TCR_4(__kmp_global.g.g_done) || !__kmp_init_serial) {
6162     KA_TRACE(10, ("__kmp_internal_end_library: already finished\n"));
6163     return;
6164   }
6165 
6166   KMP_MB(); /* Flush all pending memory write invalidates.  */
6167 
6168   /* find out who we are and what we should do */
6169   {
6170     int gtid = (gtid_req >= 0) ? gtid_req : __kmp_gtid_get_specific();
6171     KA_TRACE(
6172         10, ("__kmp_internal_end_library: enter T#%d  (%d)\n", gtid, gtid_req));
6173     if (gtid == KMP_GTID_SHUTDOWN) {
6174       KA_TRACE(10, ("__kmp_internal_end_library: !__kmp_init_runtime, system "
6175                     "already shutdown\n"));
6176       return;
6177     } else if (gtid == KMP_GTID_MONITOR) {
6178       KA_TRACE(10, ("__kmp_internal_end_library: monitor thread, gtid not "
6179                     "registered, or system shutdown\n"));
6180       return;
6181     } else if (gtid == KMP_GTID_DNE) {
6182       KA_TRACE(10, ("__kmp_internal_end_library: gtid not registered or system "
6183                     "shutdown\n"));
6184       /* we don't know who we are, but we may still shutdown the library */
6185     } else if (KMP_UBER_GTID(gtid)) {
6186       /* unregister ourselves as an uber thread.  gtid is no longer valid */
6187       if (__kmp_root[gtid]->r.r_active) {
6188         __kmp_global.g.g_abort = -1;
6189         TCW_SYNC_4(__kmp_global.g.g_done, TRUE);
6190         KA_TRACE(10,
6191                  ("__kmp_internal_end_library: root still active, abort T#%d\n",
6192                   gtid));
6193         return;
6194       } else {
6195         KA_TRACE(
6196             10,
6197             ("__kmp_internal_end_library: unregistering sibling T#%d\n", gtid));
6198         __kmp_unregister_root_current_thread(gtid);
6199       }
6200     } else {
6201 /* worker threads may call this function through the atexit handler, if they
6202  * call exit() */
6203 /* For now, skip the usual subsequent processing and just dump the debug buffer.
6204    TODO: do a thorough shutdown instead */
6205 #ifdef DUMP_DEBUG_ON_EXIT
6206       if (__kmp_debug_buf)
6207         __kmp_dump_debug_buffer();
6208 #endif
6209       return;
6210     }
6211   }
6212   /* synchronize the termination process */
6213   __kmp_acquire_bootstrap_lock(&__kmp_initz_lock);
6214 
6215   /* have we already finished */
6216   if (__kmp_global.g.g_abort) {
6217     KA_TRACE(10, ("__kmp_internal_end_library: abort, exiting\n"));
6218     /* TODO abort? */
6219     __kmp_release_bootstrap_lock(&__kmp_initz_lock);
6220     return;
6221   }
6222   if (TCR_4(__kmp_global.g.g_done) || !__kmp_init_serial) {
6223     __kmp_release_bootstrap_lock(&__kmp_initz_lock);
6224     return;
6225   }
6226 
6227   /* We need this lock to enforce mutex between this reading of
6228      __kmp_threads_capacity and the writing by __kmp_register_root.
6229      Alternatively, we can use a counter of roots that is atomically updated by
6230      __kmp_get_global_thread_id_reg, __kmp_do_serial_initialize and
6231      __kmp_internal_end_*.  */
6232   __kmp_acquire_bootstrap_lock(&__kmp_forkjoin_lock);
6233 
6234   /* now we can safely conduct the actual termination */
6235   __kmp_internal_end();
6236 
6237   __kmp_release_bootstrap_lock(&__kmp_forkjoin_lock);
6238   __kmp_release_bootstrap_lock(&__kmp_initz_lock);
6239 
6240   KA_TRACE(10, ("__kmp_internal_end_library: exit\n"));
6241 
6242 #ifdef DUMP_DEBUG_ON_EXIT
6243   if (__kmp_debug_buf)
6244     __kmp_dump_debug_buffer();
6245 #endif
6246 
6247 #if KMP_OS_WINDOWS
6248   __kmp_close_console();
6249 #endif
6250 
6251   __kmp_fini_allocator();
6252 
6253 } // __kmp_internal_end_library
6254 
6255 void __kmp_internal_end_thread(int gtid_req) {
6256   int i;
6257 
6258   /* if we have already cleaned up, don't try again, it wouldn't be pretty */
6259   /* this shouldn't be a race condition because __kmp_internal_end() is the
6260    * only place to clear __kmp_serial_init */
6261   /* we'll check this later too, after we get the lock */
6262   // 2009-09-06: We do not set g_abort without setting g_done. This check looks
6263   // redundant, because the next check will work in any case.
6264   if (__kmp_global.g.g_abort) {
6265     KA_TRACE(11, ("__kmp_internal_end_thread: abort, exiting\n"));
6266     /* TODO abort? */
6267     return;
6268   }
6269   if (TCR_4(__kmp_global.g.g_done) || !__kmp_init_serial) {
6270     KA_TRACE(10, ("__kmp_internal_end_thread: already finished\n"));
6271     return;
6272   }
6273 
6274   KMP_MB(); /* Flush all pending memory write invalidates.  */
6275 
6276   /* find out who we are and what we should do */
6277   {
6278     int gtid = (gtid_req >= 0) ? gtid_req : __kmp_gtid_get_specific();
6279     KA_TRACE(10,
6280              ("__kmp_internal_end_thread: enter T#%d  (%d)\n", gtid, gtid_req));
6281     if (gtid == KMP_GTID_SHUTDOWN) {
6282       KA_TRACE(10, ("__kmp_internal_end_thread: !__kmp_init_runtime, system "
6283                     "already shutdown\n"));
6284       return;
6285     } else if (gtid == KMP_GTID_MONITOR) {
6286       KA_TRACE(10, ("__kmp_internal_end_thread: monitor thread, gtid not "
6287                     "registered, or system shutdown\n"));
6288       return;
6289     } else if (gtid == KMP_GTID_DNE) {
6290       KA_TRACE(10, ("__kmp_internal_end_thread: gtid not registered or system "
6291                     "shutdown\n"));
6292       return;
6293       /* we don't know who we are */
6294     } else if (KMP_UBER_GTID(gtid)) {
6295       /* unregister ourselves as an uber thread.  gtid is no longer valid */
6296       if (__kmp_root[gtid]->r.r_active) {
6297         __kmp_global.g.g_abort = -1;
6298         TCW_SYNC_4(__kmp_global.g.g_done, TRUE);
6299         KA_TRACE(10,
6300                  ("__kmp_internal_end_thread: root still active, abort T#%d\n",
6301                   gtid));
6302         return;
6303       } else {
6304         KA_TRACE(10, ("__kmp_internal_end_thread: unregistering sibling T#%d\n",
6305                       gtid));
6306         __kmp_unregister_root_current_thread(gtid);
6307       }
6308     } else {
6309       /* just a worker thread, let's leave */
6310       KA_TRACE(10, ("__kmp_internal_end_thread: worker thread T#%d\n", gtid));
6311 
6312       if (gtid >= 0) {
6313         __kmp_threads[gtid]->th.th_task_team = NULL;
6314       }
6315 
6316       KA_TRACE(10,
6317                ("__kmp_internal_end_thread: worker thread done, exiting T#%d\n",
6318                 gtid));
6319       return;
6320     }
6321   }
6322 #if KMP_DYNAMIC_LIB
6323 #if OMP_50_ENABLED
6324   if (__kmp_pause_status != kmp_hard_paused)
6325 #endif
6326   // AC: lets not shutdown the dynamic library at the exit of uber thread,
6327   // because we will better shutdown later in the library destructor.
6328   {
6329     KA_TRACE(10, ("__kmp_internal_end_thread: exiting T#%d\n", gtid_req));
6330     return;
6331   }
6332 #endif
6333   /* synchronize the termination process */
6334   __kmp_acquire_bootstrap_lock(&__kmp_initz_lock);
6335 
6336   /* have we already finished */
6337   if (__kmp_global.g.g_abort) {
6338     KA_TRACE(10, ("__kmp_internal_end_thread: abort, exiting\n"));
6339     /* TODO abort? */
6340     __kmp_release_bootstrap_lock(&__kmp_initz_lock);
6341     return;
6342   }
6343   if (TCR_4(__kmp_global.g.g_done) || !__kmp_init_serial) {
6344     __kmp_release_bootstrap_lock(&__kmp_initz_lock);
6345     return;
6346   }
6347 
6348   /* We need this lock to enforce mutex between this reading of
6349      __kmp_threads_capacity and the writing by __kmp_register_root.
6350      Alternatively, we can use a counter of roots that is atomically updated by
6351      __kmp_get_global_thread_id_reg, __kmp_do_serial_initialize and
6352      __kmp_internal_end_*.  */
6353 
6354   /* should we finish the run-time?  are all siblings done? */
6355   __kmp_acquire_bootstrap_lock(&__kmp_forkjoin_lock);
6356 
6357   for (i = 0; i < __kmp_threads_capacity; ++i) {
6358     if (KMP_UBER_GTID(i)) {
6359       KA_TRACE(
6360           10,
6361           ("__kmp_internal_end_thread: remaining sibling task: gtid==%d\n", i));
6362       __kmp_release_bootstrap_lock(&__kmp_forkjoin_lock);
6363       __kmp_release_bootstrap_lock(&__kmp_initz_lock);
6364       return;
6365     }
6366   }
6367 
6368   /* now we can safely conduct the actual termination */
6369 
6370   __kmp_internal_end();
6371 
6372   __kmp_release_bootstrap_lock(&__kmp_forkjoin_lock);
6373   __kmp_release_bootstrap_lock(&__kmp_initz_lock);
6374 
6375   KA_TRACE(10, ("__kmp_internal_end_thread: exit T#%d\n", gtid_req));
6376 
6377 #ifdef DUMP_DEBUG_ON_EXIT
6378   if (__kmp_debug_buf)
6379     __kmp_dump_debug_buffer();
6380 #endif
6381 } // __kmp_internal_end_thread
6382 
6383 // -----------------------------------------------------------------------------
6384 // Library registration stuff.
6385 
6386 static long __kmp_registration_flag = 0;
6387 // Random value used to indicate library initialization.
6388 static char *__kmp_registration_str = NULL;
6389 // Value to be saved in env var __KMP_REGISTERED_LIB_<pid>.
6390 
6391 static inline char *__kmp_reg_status_name() {
6392   /* On RHEL 3u5 if linked statically, getpid() returns different values in
6393      each thread. If registration and unregistration go in different threads
6394      (omp_misc_other_root_exit.cpp test case), the name of registered_lib_env
6395      env var can not be found, because the name will contain different pid. */
6396   return __kmp_str_format("__KMP_REGISTERED_LIB_%d", (int)getpid());
6397 } // __kmp_reg_status_get
6398 
6399 void __kmp_register_library_startup(void) {
6400 
6401   char *name = __kmp_reg_status_name(); // Name of the environment variable.
6402   int done = 0;
6403   union {
6404     double dtime;
6405     long ltime;
6406   } time;
6407 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
6408   __kmp_initialize_system_tick();
6409 #endif
6410   __kmp_read_system_time(&time.dtime);
6411   __kmp_registration_flag = 0xCAFE0000L | (time.ltime & 0x0000FFFFL);
6412   __kmp_registration_str =
6413       __kmp_str_format("%p-%lx-%s", &__kmp_registration_flag,
6414                        __kmp_registration_flag, KMP_LIBRARY_FILE);
6415 
6416   KA_TRACE(50, ("__kmp_register_library_startup: %s=\"%s\"\n", name,
6417                 __kmp_registration_str));
6418 
6419   while (!done) {
6420 
6421     char *value = NULL; // Actual value of the environment variable.
6422 
6423     // Set environment variable, but do not overwrite if it is exist.
6424     __kmp_env_set(name, __kmp_registration_str, 0);
6425     // Check the variable is written.
6426     value = __kmp_env_get(name);
6427     if (value != NULL && strcmp(value, __kmp_registration_str) == 0) {
6428 
6429       done = 1; // Ok, environment variable set successfully, exit the loop.
6430 
6431     } else {
6432 
6433       // Oops. Write failed. Another copy of OpenMP RTL is in memory.
6434       // Check whether it alive or dead.
6435       int neighbor = 0; // 0 -- unknown status, 1 -- alive, 2 -- dead.
6436       char *tail = value;
6437       char *flag_addr_str = NULL;
6438       char *flag_val_str = NULL;
6439       char const *file_name = NULL;
6440       __kmp_str_split(tail, '-', &flag_addr_str, &tail);
6441       __kmp_str_split(tail, '-', &flag_val_str, &tail);
6442       file_name = tail;
6443       if (tail != NULL) {
6444         long *flag_addr = 0;
6445         long flag_val = 0;
6446         KMP_SSCANF(flag_addr_str, "%p", RCAST(void**, &flag_addr));
6447         KMP_SSCANF(flag_val_str, "%lx", &flag_val);
6448         if (flag_addr != 0 && flag_val != 0 && strcmp(file_name, "") != 0) {
6449           // First, check whether environment-encoded address is mapped into
6450           // addr space.
6451           // If so, dereference it to see if it still has the right value.
6452           if (__kmp_is_address_mapped(flag_addr) && *flag_addr == flag_val) {
6453             neighbor = 1;
6454           } else {
6455             // If not, then we know the other copy of the library is no longer
6456             // running.
6457             neighbor = 2;
6458           }
6459         }
6460       }
6461       switch (neighbor) {
6462       case 0: // Cannot parse environment variable -- neighbor status unknown.
6463         // Assume it is the incompatible format of future version of the
6464         // library. Assume the other library is alive.
6465         // WARN( ... ); // TODO: Issue a warning.
6466         file_name = "unknown library";
6467         KMP_FALLTHROUGH();
6468       // Attention! Falling to the next case. That's intentional.
6469       case 1: { // Neighbor is alive.
6470         // Check it is allowed.
6471         char *duplicate_ok = __kmp_env_get("KMP_DUPLICATE_LIB_OK");
6472         if (!__kmp_str_match_true(duplicate_ok)) {
6473           // That's not allowed. Issue fatal error.
6474           __kmp_fatal(KMP_MSG(DuplicateLibrary, KMP_LIBRARY_FILE, file_name),
6475                       KMP_HNT(DuplicateLibrary), __kmp_msg_null);
6476         }
6477         KMP_INTERNAL_FREE(duplicate_ok);
6478         __kmp_duplicate_library_ok = 1;
6479         done = 1; // Exit the loop.
6480       } break;
6481       case 2: { // Neighbor is dead.
6482         // Clear the variable and try to register library again.
6483         __kmp_env_unset(name);
6484       } break;
6485       default: { KMP_DEBUG_ASSERT(0); } break;
6486       }
6487     }
6488     KMP_INTERNAL_FREE((void *)value);
6489   }
6490   KMP_INTERNAL_FREE((void *)name);
6491 
6492 } // func __kmp_register_library_startup
6493 
6494 void __kmp_unregister_library(void) {
6495 
6496   char *name = __kmp_reg_status_name();
6497   char *value = __kmp_env_get(name);
6498 
6499   KMP_DEBUG_ASSERT(__kmp_registration_flag != 0);
6500   KMP_DEBUG_ASSERT(__kmp_registration_str != NULL);
6501   if (value != NULL && strcmp(value, __kmp_registration_str) == 0) {
6502     // Ok, this is our variable. Delete it.
6503     __kmp_env_unset(name);
6504   }
6505 
6506   KMP_INTERNAL_FREE(__kmp_registration_str);
6507   KMP_INTERNAL_FREE(value);
6508   KMP_INTERNAL_FREE(name);
6509 
6510   __kmp_registration_flag = 0;
6511   __kmp_registration_str = NULL;
6512 
6513 } // __kmp_unregister_library
6514 
6515 // End of Library registration stuff.
6516 // -----------------------------------------------------------------------------
6517 
6518 #if KMP_MIC_SUPPORTED
6519 
6520 static void __kmp_check_mic_type() {
6521   kmp_cpuid_t cpuid_state = {0};
6522   kmp_cpuid_t *cs_p = &cpuid_state;
6523   __kmp_x86_cpuid(1, 0, cs_p);
6524   // We don't support mic1 at the moment
6525   if ((cs_p->eax & 0xff0) == 0xB10) {
6526     __kmp_mic_type = mic2;
6527   } else if ((cs_p->eax & 0xf0ff0) == 0x50670) {
6528     __kmp_mic_type = mic3;
6529   } else {
6530     __kmp_mic_type = non_mic;
6531   }
6532 }
6533 
6534 #endif /* KMP_MIC_SUPPORTED */
6535 
6536 static void __kmp_do_serial_initialize(void) {
6537   int i, gtid;
6538   int size;
6539 
6540   KA_TRACE(10, ("__kmp_do_serial_initialize: enter\n"));
6541 
6542   KMP_DEBUG_ASSERT(sizeof(kmp_int32) == 4);
6543   KMP_DEBUG_ASSERT(sizeof(kmp_uint32) == 4);
6544   KMP_DEBUG_ASSERT(sizeof(kmp_int64) == 8);
6545   KMP_DEBUG_ASSERT(sizeof(kmp_uint64) == 8);
6546   KMP_DEBUG_ASSERT(sizeof(kmp_intptr_t) == sizeof(void *));
6547 
6548 #if OMPT_SUPPORT
6549   ompt_pre_init();
6550 #endif
6551 
6552   __kmp_validate_locks();
6553 
6554   /* Initialize internal memory allocator */
6555   __kmp_init_allocator();
6556 
6557   /* Register the library startup via an environment variable and check to see
6558      whether another copy of the library is already registered. */
6559 
6560   __kmp_register_library_startup();
6561 
6562   /* TODO reinitialization of library */
6563   if (TCR_4(__kmp_global.g.g_done)) {
6564     KA_TRACE(10, ("__kmp_do_serial_initialize: reinitialization of library\n"));
6565   }
6566 
6567   __kmp_global.g.g_abort = 0;
6568   TCW_SYNC_4(__kmp_global.g.g_done, FALSE);
6569 
6570 /* initialize the locks */
6571 #if KMP_USE_ADAPTIVE_LOCKS
6572 #if KMP_DEBUG_ADAPTIVE_LOCKS
6573   __kmp_init_speculative_stats();
6574 #endif
6575 #endif
6576 #if KMP_STATS_ENABLED
6577   __kmp_stats_init();
6578 #endif
6579   __kmp_init_lock(&__kmp_global_lock);
6580   __kmp_init_queuing_lock(&__kmp_dispatch_lock);
6581   __kmp_init_lock(&__kmp_debug_lock);
6582   __kmp_init_atomic_lock(&__kmp_atomic_lock);
6583   __kmp_init_atomic_lock(&__kmp_atomic_lock_1i);
6584   __kmp_init_atomic_lock(&__kmp_atomic_lock_2i);
6585   __kmp_init_atomic_lock(&__kmp_atomic_lock_4i);
6586   __kmp_init_atomic_lock(&__kmp_atomic_lock_4r);
6587   __kmp_init_atomic_lock(&__kmp_atomic_lock_8i);
6588   __kmp_init_atomic_lock(&__kmp_atomic_lock_8r);
6589   __kmp_init_atomic_lock(&__kmp_atomic_lock_8c);
6590   __kmp_init_atomic_lock(&__kmp_atomic_lock_10r);
6591   __kmp_init_atomic_lock(&__kmp_atomic_lock_16r);
6592   __kmp_init_atomic_lock(&__kmp_atomic_lock_16c);
6593   __kmp_init_atomic_lock(&__kmp_atomic_lock_20c);
6594   __kmp_init_atomic_lock(&__kmp_atomic_lock_32c);
6595   __kmp_init_bootstrap_lock(&__kmp_forkjoin_lock);
6596   __kmp_init_bootstrap_lock(&__kmp_exit_lock);
6597 #if KMP_USE_MONITOR
6598   __kmp_init_bootstrap_lock(&__kmp_monitor_lock);
6599 #endif
6600   __kmp_init_bootstrap_lock(&__kmp_tp_cached_lock);
6601 
6602   /* conduct initialization and initial setup of configuration */
6603 
6604   __kmp_runtime_initialize();
6605 
6606 #if KMP_MIC_SUPPORTED
6607   __kmp_check_mic_type();
6608 #endif
6609 
6610 // Some global variable initialization moved here from kmp_env_initialize()
6611 #ifdef KMP_DEBUG
6612   kmp_diag = 0;
6613 #endif
6614   __kmp_abort_delay = 0;
6615 
6616   // From __kmp_init_dflt_team_nth()
6617   /* assume the entire machine will be used */
6618   __kmp_dflt_team_nth_ub = __kmp_xproc;
6619   if (__kmp_dflt_team_nth_ub < KMP_MIN_NTH) {
6620     __kmp_dflt_team_nth_ub = KMP_MIN_NTH;
6621   }
6622   if (__kmp_dflt_team_nth_ub > __kmp_sys_max_nth) {
6623     __kmp_dflt_team_nth_ub = __kmp_sys_max_nth;
6624   }
6625   __kmp_max_nth = __kmp_sys_max_nth;
6626   __kmp_cg_max_nth = __kmp_sys_max_nth;
6627   __kmp_teams_max_nth = __kmp_xproc; // set a "reasonable" default
6628   if (__kmp_teams_max_nth > __kmp_sys_max_nth) {
6629     __kmp_teams_max_nth = __kmp_sys_max_nth;
6630   }
6631 
6632   // Three vars below moved here from __kmp_env_initialize() "KMP_BLOCKTIME"
6633   // part
6634   __kmp_dflt_blocktime = KMP_DEFAULT_BLOCKTIME;
6635 #if KMP_USE_MONITOR
6636   __kmp_monitor_wakeups =
6637       KMP_WAKEUPS_FROM_BLOCKTIME(__kmp_dflt_blocktime, __kmp_monitor_wakeups);
6638   __kmp_bt_intervals =
6639       KMP_INTERVALS_FROM_BLOCKTIME(__kmp_dflt_blocktime, __kmp_monitor_wakeups);
6640 #endif
6641   // From "KMP_LIBRARY" part of __kmp_env_initialize()
6642   __kmp_library = library_throughput;
6643   // From KMP_SCHEDULE initialization
6644   __kmp_static = kmp_sch_static_balanced;
6645 // AC: do not use analytical here, because it is non-monotonous
6646 //__kmp_guided = kmp_sch_guided_iterative_chunked;
6647 //__kmp_auto = kmp_sch_guided_analytical_chunked; // AC: it is the default, no
6648 // need to repeat assignment
6649 // Barrier initialization. Moved here from __kmp_env_initialize() Barrier branch
6650 // bit control and barrier method control parts
6651 #if KMP_FAST_REDUCTION_BARRIER
6652 #define kmp_reduction_barrier_gather_bb ((int)1)
6653 #define kmp_reduction_barrier_release_bb ((int)1)
6654 #define kmp_reduction_barrier_gather_pat bp_hyper_bar
6655 #define kmp_reduction_barrier_release_pat bp_hyper_bar
6656 #endif // KMP_FAST_REDUCTION_BARRIER
6657   for (i = bs_plain_barrier; i < bs_last_barrier; i++) {
6658     __kmp_barrier_gather_branch_bits[i] = __kmp_barrier_gather_bb_dflt;
6659     __kmp_barrier_release_branch_bits[i] = __kmp_barrier_release_bb_dflt;
6660     __kmp_barrier_gather_pattern[i] = __kmp_barrier_gather_pat_dflt;
6661     __kmp_barrier_release_pattern[i] = __kmp_barrier_release_pat_dflt;
6662 #if KMP_FAST_REDUCTION_BARRIER
6663     if (i == bs_reduction_barrier) { // tested and confirmed on ALTIX only (
6664       // lin_64 ): hyper,1
6665       __kmp_barrier_gather_branch_bits[i] = kmp_reduction_barrier_gather_bb;
6666       __kmp_barrier_release_branch_bits[i] = kmp_reduction_barrier_release_bb;
6667       __kmp_barrier_gather_pattern[i] = kmp_reduction_barrier_gather_pat;
6668       __kmp_barrier_release_pattern[i] = kmp_reduction_barrier_release_pat;
6669     }
6670 #endif // KMP_FAST_REDUCTION_BARRIER
6671   }
6672 #if KMP_FAST_REDUCTION_BARRIER
6673 #undef kmp_reduction_barrier_release_pat
6674 #undef kmp_reduction_barrier_gather_pat
6675 #undef kmp_reduction_barrier_release_bb
6676 #undef kmp_reduction_barrier_gather_bb
6677 #endif // KMP_FAST_REDUCTION_BARRIER
6678 #if KMP_MIC_SUPPORTED
6679   if (__kmp_mic_type == mic2) { // KNC
6680     // AC: plane=3,2, forkjoin=2,1 are optimal for 240 threads on KNC
6681     __kmp_barrier_gather_branch_bits[bs_plain_barrier] = 3; // plain gather
6682     __kmp_barrier_release_branch_bits[bs_forkjoin_barrier] =
6683         1; // forkjoin release
6684     __kmp_barrier_gather_pattern[bs_forkjoin_barrier] = bp_hierarchical_bar;
6685     __kmp_barrier_release_pattern[bs_forkjoin_barrier] = bp_hierarchical_bar;
6686   }
6687 #if KMP_FAST_REDUCTION_BARRIER
6688   if (__kmp_mic_type == mic2) { // KNC
6689     __kmp_barrier_gather_pattern[bs_reduction_barrier] = bp_hierarchical_bar;
6690     __kmp_barrier_release_pattern[bs_reduction_barrier] = bp_hierarchical_bar;
6691   }
6692 #endif // KMP_FAST_REDUCTION_BARRIER
6693 #endif // KMP_MIC_SUPPORTED
6694 
6695 // From KMP_CHECKS initialization
6696 #ifdef KMP_DEBUG
6697   __kmp_env_checks = TRUE; /* development versions have the extra checks */
6698 #else
6699   __kmp_env_checks = FALSE; /* port versions do not have the extra checks */
6700 #endif
6701 
6702   // From "KMP_FOREIGN_THREADS_THREADPRIVATE" initialization
6703   __kmp_foreign_tp = TRUE;
6704 
6705   __kmp_global.g.g_dynamic = FALSE;
6706   __kmp_global.g.g_dynamic_mode = dynamic_default;
6707 
6708   __kmp_env_initialize(NULL);
6709 
6710 // Print all messages in message catalog for testing purposes.
6711 #ifdef KMP_DEBUG
6712   char const *val = __kmp_env_get("KMP_DUMP_CATALOG");
6713   if (__kmp_str_match_true(val)) {
6714     kmp_str_buf_t buffer;
6715     __kmp_str_buf_init(&buffer);
6716     __kmp_i18n_dump_catalog(&buffer);
6717     __kmp_printf("%s", buffer.str);
6718     __kmp_str_buf_free(&buffer);
6719   }
6720   __kmp_env_free(&val);
6721 #endif
6722 
6723   __kmp_threads_capacity =
6724       __kmp_initial_threads_capacity(__kmp_dflt_team_nth_ub);
6725   // Moved here from __kmp_env_initialize() "KMP_ALL_THREADPRIVATE" part
6726   __kmp_tp_capacity = __kmp_default_tp_capacity(
6727       __kmp_dflt_team_nth_ub, __kmp_max_nth, __kmp_allThreadsSpecified);
6728 
6729   // If the library is shut down properly, both pools must be NULL. Just in
6730   // case, set them to NULL -- some memory may leak, but subsequent code will
6731   // work even if pools are not freed.
6732   KMP_DEBUG_ASSERT(__kmp_thread_pool == NULL);
6733   KMP_DEBUG_ASSERT(__kmp_thread_pool_insert_pt == NULL);
6734   KMP_DEBUG_ASSERT(__kmp_team_pool == NULL);
6735   __kmp_thread_pool = NULL;
6736   __kmp_thread_pool_insert_pt = NULL;
6737   __kmp_team_pool = NULL;
6738 
6739   /* Allocate all of the variable sized records */
6740   /* NOTE: __kmp_threads_capacity entries are allocated, but the arrays are
6741    * expandable */
6742   /* Since allocation is cache-aligned, just add extra padding at the end */
6743   size =
6744       (sizeof(kmp_info_t *) + sizeof(kmp_root_t *)) * __kmp_threads_capacity +
6745       CACHE_LINE;
6746   __kmp_threads = (kmp_info_t **)__kmp_allocate(size);
6747   __kmp_root = (kmp_root_t **)((char *)__kmp_threads +
6748                                sizeof(kmp_info_t *) * __kmp_threads_capacity);
6749 
6750   /* init thread counts */
6751   KMP_DEBUG_ASSERT(__kmp_all_nth ==
6752                    0); // Asserts fail if the library is reinitializing and
6753   KMP_DEBUG_ASSERT(__kmp_nth == 0); // something was wrong in termination.
6754   __kmp_all_nth = 0;
6755   __kmp_nth = 0;
6756 
6757   /* setup the uber master thread and hierarchy */
6758   gtid = __kmp_register_root(TRUE);
6759   KA_TRACE(10, ("__kmp_do_serial_initialize  T#%d\n", gtid));
6760   KMP_ASSERT(KMP_UBER_GTID(gtid));
6761   KMP_ASSERT(KMP_INITIAL_GTID(gtid));
6762 
6763   KMP_MB(); /* Flush all pending memory write invalidates.  */
6764 
6765   __kmp_common_initialize();
6766 
6767 #if KMP_OS_UNIX
6768   /* invoke the child fork handler */
6769   __kmp_register_atfork();
6770 #endif
6771 
6772 #if !KMP_DYNAMIC_LIB
6773   {
6774     /* Invoke the exit handler when the program finishes, only for static
6775        library. For dynamic library, we already have _fini and DllMain. */
6776     int rc = atexit(__kmp_internal_end_atexit);
6777     if (rc != 0) {
6778       __kmp_fatal(KMP_MSG(FunctionError, "atexit()"), KMP_ERR(rc),
6779                   __kmp_msg_null);
6780     }
6781   }
6782 #endif
6783 
6784 #if KMP_HANDLE_SIGNALS
6785 #if KMP_OS_UNIX
6786   /* NOTE: make sure that this is called before the user installs their own
6787      signal handlers so that the user handlers are called first. this way they
6788      can return false, not call our handler, avoid terminating the library, and
6789      continue execution where they left off. */
6790   __kmp_install_signals(FALSE);
6791 #endif /* KMP_OS_UNIX */
6792 #if KMP_OS_WINDOWS
6793   __kmp_install_signals(TRUE);
6794 #endif /* KMP_OS_WINDOWS */
6795 #endif
6796 
6797   /* we have finished the serial initialization */
6798   __kmp_init_counter++;
6799 
6800   __kmp_init_serial = TRUE;
6801 
6802   if (__kmp_settings) {
6803     __kmp_env_print();
6804   }
6805 
6806 #if OMP_40_ENABLED
6807   if (__kmp_display_env || __kmp_display_env_verbose) {
6808     __kmp_env_print_2();
6809   }
6810 #endif // OMP_40_ENABLED
6811 
6812 #if OMPT_SUPPORT
6813   ompt_post_init();
6814 #endif
6815 
6816   KMP_MB();
6817 
6818   KA_TRACE(10, ("__kmp_do_serial_initialize: exit\n"));
6819 }
6820 
6821 void __kmp_serial_initialize(void) {
6822   if (__kmp_init_serial) {
6823     return;
6824   }
6825   __kmp_acquire_bootstrap_lock(&__kmp_initz_lock);
6826   if (__kmp_init_serial) {
6827     __kmp_release_bootstrap_lock(&__kmp_initz_lock);
6828     return;
6829   }
6830   __kmp_do_serial_initialize();
6831   __kmp_release_bootstrap_lock(&__kmp_initz_lock);
6832 }
6833 
6834 static void __kmp_do_middle_initialize(void) {
6835   int i, j;
6836   int prev_dflt_team_nth;
6837 
6838   if (!__kmp_init_serial) {
6839     __kmp_do_serial_initialize();
6840   }
6841 
6842   KA_TRACE(10, ("__kmp_middle_initialize: enter\n"));
6843 
6844   // Save the previous value for the __kmp_dflt_team_nth so that
6845   // we can avoid some reinitialization if it hasn't changed.
6846   prev_dflt_team_nth = __kmp_dflt_team_nth;
6847 
6848 #if KMP_AFFINITY_SUPPORTED
6849   // __kmp_affinity_initialize() will try to set __kmp_ncores to the
6850   // number of cores on the machine.
6851   __kmp_affinity_initialize();
6852 
6853   // Run through the __kmp_threads array and set the affinity mask
6854   // for each root thread that is currently registered with the RTL.
6855   for (i = 0; i < __kmp_threads_capacity; i++) {
6856     if (TCR_PTR(__kmp_threads[i]) != NULL) {
6857       __kmp_affinity_set_init_mask(i, TRUE);
6858     }
6859   }
6860 #endif /* KMP_AFFINITY_SUPPORTED */
6861 
6862   KMP_ASSERT(__kmp_xproc > 0);
6863   if (__kmp_avail_proc == 0) {
6864     __kmp_avail_proc = __kmp_xproc;
6865   }
6866 
6867   // If there were empty places in num_threads list (OMP_NUM_THREADS=,,2,3),
6868   // correct them now
6869   j = 0;
6870   while ((j < __kmp_nested_nth.used) && !__kmp_nested_nth.nth[j]) {
6871     __kmp_nested_nth.nth[j] = __kmp_dflt_team_nth = __kmp_dflt_team_nth_ub =
6872         __kmp_avail_proc;
6873     j++;
6874   }
6875 
6876   if (__kmp_dflt_team_nth == 0) {
6877 #ifdef KMP_DFLT_NTH_CORES
6878     // Default #threads = #cores
6879     __kmp_dflt_team_nth = __kmp_ncores;
6880     KA_TRACE(20, ("__kmp_middle_initialize: setting __kmp_dflt_team_nth = "
6881                   "__kmp_ncores (%d)\n",
6882                   __kmp_dflt_team_nth));
6883 #else
6884     // Default #threads = #available OS procs
6885     __kmp_dflt_team_nth = __kmp_avail_proc;
6886     KA_TRACE(20, ("__kmp_middle_initialize: setting __kmp_dflt_team_nth = "
6887                   "__kmp_avail_proc(%d)\n",
6888                   __kmp_dflt_team_nth));
6889 #endif /* KMP_DFLT_NTH_CORES */
6890   }
6891 
6892   if (__kmp_dflt_team_nth < KMP_MIN_NTH) {
6893     __kmp_dflt_team_nth = KMP_MIN_NTH;
6894   }
6895   if (__kmp_dflt_team_nth > __kmp_sys_max_nth) {
6896     __kmp_dflt_team_nth = __kmp_sys_max_nth;
6897   }
6898 
6899   // There's no harm in continuing if the following check fails,
6900   // but it indicates an error in the previous logic.
6901   KMP_DEBUG_ASSERT(__kmp_dflt_team_nth <= __kmp_dflt_team_nth_ub);
6902 
6903   if (__kmp_dflt_team_nth != prev_dflt_team_nth) {
6904     // Run through the __kmp_threads array and set the num threads icv for each
6905     // root thread that is currently registered with the RTL (which has not
6906     // already explicitly set its nthreads-var with a call to
6907     // omp_set_num_threads()).
6908     for (i = 0; i < __kmp_threads_capacity; i++) {
6909       kmp_info_t *thread = __kmp_threads[i];
6910       if (thread == NULL)
6911         continue;
6912       if (thread->th.th_current_task->td_icvs.nproc != 0)
6913         continue;
6914 
6915       set__nproc(__kmp_threads[i], __kmp_dflt_team_nth);
6916     }
6917   }
6918   KA_TRACE(
6919       20,
6920       ("__kmp_middle_initialize: final value for __kmp_dflt_team_nth = %d\n",
6921        __kmp_dflt_team_nth));
6922 
6923 #ifdef KMP_ADJUST_BLOCKTIME
6924   /* Adjust blocktime to zero if necessary  now that __kmp_avail_proc is set */
6925   if (!__kmp_env_blocktime && (__kmp_avail_proc > 0)) {
6926     KMP_DEBUG_ASSERT(__kmp_avail_proc > 0);
6927     if (__kmp_nth > __kmp_avail_proc) {
6928       __kmp_zero_bt = TRUE;
6929     }
6930   }
6931 #endif /* KMP_ADJUST_BLOCKTIME */
6932 
6933   /* we have finished middle initialization */
6934   TCW_SYNC_4(__kmp_init_middle, TRUE);
6935 
6936   KA_TRACE(10, ("__kmp_do_middle_initialize: exit\n"));
6937 }
6938 
6939 void __kmp_middle_initialize(void) {
6940   if (__kmp_init_middle) {
6941     return;
6942   }
6943   __kmp_acquire_bootstrap_lock(&__kmp_initz_lock);
6944   if (__kmp_init_middle) {
6945     __kmp_release_bootstrap_lock(&__kmp_initz_lock);
6946     return;
6947   }
6948   __kmp_do_middle_initialize();
6949   __kmp_release_bootstrap_lock(&__kmp_initz_lock);
6950 }
6951 
6952 void __kmp_parallel_initialize(void) {
6953   int gtid = __kmp_entry_gtid(); // this might be a new root
6954 
6955   /* synchronize parallel initialization (for sibling) */
6956   if (TCR_4(__kmp_init_parallel))
6957     return;
6958   __kmp_acquire_bootstrap_lock(&__kmp_initz_lock);
6959   if (TCR_4(__kmp_init_parallel)) {
6960     __kmp_release_bootstrap_lock(&__kmp_initz_lock);
6961     return;
6962   }
6963 
6964   /* TODO reinitialization after we have already shut down */
6965   if (TCR_4(__kmp_global.g.g_done)) {
6966     KA_TRACE(
6967         10,
6968         ("__kmp_parallel_initialize: attempt to init while shutting down\n"));
6969     __kmp_infinite_loop();
6970   }
6971 
6972   /* jc: The lock __kmp_initz_lock is already held, so calling
6973      __kmp_serial_initialize would cause a deadlock.  So we call
6974      __kmp_do_serial_initialize directly. */
6975   if (!__kmp_init_middle) {
6976     __kmp_do_middle_initialize();
6977   }
6978 
6979 #if OMP_50_ENABLED
6980   __kmp_resume_if_hard_paused();
6981 #endif
6982 
6983   /* begin initialization */
6984   KA_TRACE(10, ("__kmp_parallel_initialize: enter\n"));
6985   KMP_ASSERT(KMP_UBER_GTID(gtid));
6986 
6987 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
6988   // Save the FP control regs.
6989   // Worker threads will set theirs to these values at thread startup.
6990   __kmp_store_x87_fpu_control_word(&__kmp_init_x87_fpu_control_word);
6991   __kmp_store_mxcsr(&__kmp_init_mxcsr);
6992   __kmp_init_mxcsr &= KMP_X86_MXCSR_MASK;
6993 #endif /* KMP_ARCH_X86 || KMP_ARCH_X86_64 */
6994 
6995 #if KMP_OS_UNIX
6996 #if KMP_HANDLE_SIGNALS
6997   /*  must be after __kmp_serial_initialize  */
6998   __kmp_install_signals(TRUE);
6999 #endif
7000 #endif
7001 
7002   __kmp_suspend_initialize();
7003 
7004 #if defined(USE_LOAD_BALANCE)
7005   if (__kmp_global.g.g_dynamic_mode == dynamic_default) {
7006     __kmp_global.g.g_dynamic_mode = dynamic_load_balance;
7007   }
7008 #else
7009   if (__kmp_global.g.g_dynamic_mode == dynamic_default) {
7010     __kmp_global.g.g_dynamic_mode = dynamic_thread_limit;
7011   }
7012 #endif
7013 
7014   if (__kmp_version) {
7015     __kmp_print_version_2();
7016   }
7017 
7018   /* we have finished parallel initialization */
7019   TCW_SYNC_4(__kmp_init_parallel, TRUE);
7020 
7021   KMP_MB();
7022   KA_TRACE(10, ("__kmp_parallel_initialize: exit\n"));
7023 
7024   __kmp_release_bootstrap_lock(&__kmp_initz_lock);
7025 }
7026 
7027 /* ------------------------------------------------------------------------ */
7028 
7029 void __kmp_run_before_invoked_task(int gtid, int tid, kmp_info_t *this_thr,
7030                                    kmp_team_t *team) {
7031   kmp_disp_t *dispatch;
7032 
7033   KMP_MB();
7034 
7035   /* none of the threads have encountered any constructs, yet. */
7036   this_thr->th.th_local.this_construct = 0;
7037 #if KMP_CACHE_MANAGE
7038   KMP_CACHE_PREFETCH(&this_thr->th.th_bar[bs_forkjoin_barrier].bb.b_arrived);
7039 #endif /* KMP_CACHE_MANAGE */
7040   dispatch = (kmp_disp_t *)TCR_PTR(this_thr->th.th_dispatch);
7041   KMP_DEBUG_ASSERT(dispatch);
7042   KMP_DEBUG_ASSERT(team->t.t_dispatch);
7043   // KMP_DEBUG_ASSERT( this_thr->th.th_dispatch == &team->t.t_dispatch[
7044   // this_thr->th.th_info.ds.ds_tid ] );
7045 
7046   dispatch->th_disp_index = 0; /* reset the dispatch buffer counter */
7047 #if OMP_45_ENABLED
7048   dispatch->th_doacross_buf_idx =
7049       0; /* reset the doacross dispatch buffer counter */
7050 #endif
7051   if (__kmp_env_consistency_check)
7052     __kmp_push_parallel(gtid, team->t.t_ident);
7053 
7054   KMP_MB(); /* Flush all pending memory write invalidates.  */
7055 }
7056 
7057 void __kmp_run_after_invoked_task(int gtid, int tid, kmp_info_t *this_thr,
7058                                   kmp_team_t *team) {
7059   if (__kmp_env_consistency_check)
7060     __kmp_pop_parallel(gtid, team->t.t_ident);
7061 
7062   __kmp_finish_implicit_task(this_thr);
7063 }
7064 
7065 int __kmp_invoke_task_func(int gtid) {
7066   int rc;
7067   int tid = __kmp_tid_from_gtid(gtid);
7068   kmp_info_t *this_thr = __kmp_threads[gtid];
7069   kmp_team_t *team = this_thr->th.th_team;
7070 
7071   __kmp_run_before_invoked_task(gtid, tid, this_thr, team);
7072 #if USE_ITT_BUILD
7073   if (__itt_stack_caller_create_ptr) {
7074     __kmp_itt_stack_callee_enter(
7075         (__itt_caller)
7076             team->t.t_stack_id); // inform ittnotify about entering user's code
7077   }
7078 #endif /* USE_ITT_BUILD */
7079 #if INCLUDE_SSC_MARKS
7080   SSC_MARK_INVOKING();
7081 #endif
7082 
7083 #if OMPT_SUPPORT
7084   void *dummy;
7085   void **exit_runtime_p;
7086   ompt_data_t *my_task_data;
7087   ompt_data_t *my_parallel_data;
7088   int ompt_team_size;
7089 
7090   if (ompt_enabled.enabled) {
7091     exit_runtime_p = &(
7092         team->t.t_implicit_task_taskdata[tid].ompt_task_info.frame.exit_frame.ptr);
7093   } else {
7094     exit_runtime_p = &dummy;
7095   }
7096 
7097   my_task_data =
7098       &(team->t.t_implicit_task_taskdata[tid].ompt_task_info.task_data);
7099   my_parallel_data = &(team->t.ompt_team_info.parallel_data);
7100   if (ompt_enabled.ompt_callback_implicit_task) {
7101     ompt_team_size = team->t.t_nproc;
7102     ompt_callbacks.ompt_callback(ompt_callback_implicit_task)(
7103         ompt_scope_begin, my_parallel_data, my_task_data, ompt_team_size,
7104         __kmp_tid_from_gtid(gtid), ompt_task_implicit); // TODO: Can this be ompt_task_initial?
7105     OMPT_CUR_TASK_INFO(this_thr)->thread_num = __kmp_tid_from_gtid(gtid);
7106   }
7107 #endif
7108 
7109   {
7110     KMP_TIME_PARTITIONED_BLOCK(OMP_parallel);
7111     KMP_SET_THREAD_STATE_BLOCK(IMPLICIT_TASK);
7112     rc =
7113         __kmp_invoke_microtask((microtask_t)TCR_SYNC_PTR(team->t.t_pkfn), gtid,
7114                                tid, (int)team->t.t_argc, (void **)team->t.t_argv
7115 #if OMPT_SUPPORT
7116                                ,
7117                                exit_runtime_p
7118 #endif
7119                                );
7120 #if OMPT_SUPPORT
7121     *exit_runtime_p = NULL;
7122 #endif
7123   }
7124 
7125 #if USE_ITT_BUILD
7126   if (__itt_stack_caller_create_ptr) {
7127     __kmp_itt_stack_callee_leave(
7128         (__itt_caller)
7129             team->t.t_stack_id); // inform ittnotify about leaving user's code
7130   }
7131 #endif /* USE_ITT_BUILD */
7132   __kmp_run_after_invoked_task(gtid, tid, this_thr, team);
7133 
7134   return rc;
7135 }
7136 
7137 #if OMP_40_ENABLED
7138 void __kmp_teams_master(int gtid) {
7139   // This routine is called by all master threads in teams construct
7140   kmp_info_t *thr = __kmp_threads[gtid];
7141   kmp_team_t *team = thr->th.th_team;
7142   ident_t *loc = team->t.t_ident;
7143   thr->th.th_set_nproc = thr->th.th_teams_size.nth;
7144   KMP_DEBUG_ASSERT(thr->th.th_teams_microtask);
7145   KMP_DEBUG_ASSERT(thr->th.th_set_nproc);
7146   KA_TRACE(20, ("__kmp_teams_master: T#%d, Tid %d, microtask %p\n", gtid,
7147                 __kmp_tid_from_gtid(gtid), thr->th.th_teams_microtask));
7148 
7149   // This thread is a new CG root.  Set up the proper variables.
7150   kmp_cg_root_t *tmp = (kmp_cg_root_t *)__kmp_allocate(sizeof(kmp_cg_root_t));
7151   tmp->cg_root = thr; // Make thr the CG root
7152   // Init to thread limit that was stored when league masters were forked
7153   tmp->cg_thread_limit = thr->th.th_current_task->td_icvs.thread_limit;
7154   tmp->cg_nthreads = 1; // Init counter to one active thread, this one
7155   KA_TRACE(100, ("__kmp_teams_master: Thread %p created node %p and init"
7156                  " cg_threads to 1\n",
7157                  thr, tmp));
7158   tmp->up = thr->th.th_cg_roots;
7159   thr->th.th_cg_roots = tmp;
7160 
7161 // Launch league of teams now, but not let workers execute
7162 // (they hang on fork barrier until next parallel)
7163 #if INCLUDE_SSC_MARKS
7164   SSC_MARK_FORKING();
7165 #endif
7166   __kmp_fork_call(loc, gtid, fork_context_intel, team->t.t_argc,
7167                   (microtask_t)thr->th.th_teams_microtask, // "wrapped" task
7168                   VOLATILE_CAST(launch_t) __kmp_invoke_task_func, NULL);
7169 #if INCLUDE_SSC_MARKS
7170   SSC_MARK_JOINING();
7171 #endif
7172   // If the team size was reduced from the limit, set it to the new size
7173   if (thr->th.th_team_nproc < thr->th.th_teams_size.nth)
7174     thr->th.th_teams_size.nth = thr->th.th_team_nproc;
7175   // AC: last parameter "1" eliminates join barrier which won't work because
7176   // worker threads are in a fork barrier waiting for more parallel regions
7177   __kmp_join_call(loc, gtid
7178 #if OMPT_SUPPORT
7179                   ,
7180                   fork_context_intel
7181 #endif
7182                   ,
7183                   1);
7184 }
7185 
7186 int __kmp_invoke_teams_master(int gtid) {
7187   kmp_info_t *this_thr = __kmp_threads[gtid];
7188   kmp_team_t *team = this_thr->th.th_team;
7189 #if KMP_DEBUG
7190   if (!__kmp_threads[gtid]->th.th_team->t.t_serialized)
7191     KMP_DEBUG_ASSERT((void *)__kmp_threads[gtid]->th.th_team->t.t_pkfn ==
7192                      (void *)__kmp_teams_master);
7193 #endif
7194   __kmp_run_before_invoked_task(gtid, 0, this_thr, team);
7195   __kmp_teams_master(gtid);
7196   __kmp_run_after_invoked_task(gtid, 0, this_thr, team);
7197   return 1;
7198 }
7199 #endif /* OMP_40_ENABLED */
7200 
7201 /* this sets the requested number of threads for the next parallel region
7202    encountered by this team. since this should be enclosed in the forkjoin
7203    critical section it should avoid race conditions with assymmetrical nested
7204    parallelism */
7205 
7206 void __kmp_push_num_threads(ident_t *id, int gtid, int num_threads) {
7207   kmp_info_t *thr = __kmp_threads[gtid];
7208 
7209   if (num_threads > 0)
7210     thr->th.th_set_nproc = num_threads;
7211 }
7212 
7213 #if OMP_40_ENABLED
7214 
7215 /* this sets the requested number of teams for the teams region and/or
7216    the number of threads for the next parallel region encountered  */
7217 void __kmp_push_num_teams(ident_t *id, int gtid, int num_teams,
7218                           int num_threads) {
7219   kmp_info_t *thr = __kmp_threads[gtid];
7220   KMP_DEBUG_ASSERT(num_teams >= 0);
7221   KMP_DEBUG_ASSERT(num_threads >= 0);
7222 
7223   if (num_teams == 0)
7224     num_teams = 1; // default number of teams is 1.
7225   if (num_teams > __kmp_teams_max_nth) { // if too many teams requested?
7226     if (!__kmp_reserve_warn) {
7227       __kmp_reserve_warn = 1;
7228       __kmp_msg(kmp_ms_warning,
7229                 KMP_MSG(CantFormThrTeam, num_teams, __kmp_teams_max_nth),
7230                 KMP_HNT(Unset_ALL_THREADS), __kmp_msg_null);
7231     }
7232     num_teams = __kmp_teams_max_nth;
7233   }
7234   // Set number of teams (number of threads in the outer "parallel" of the
7235   // teams)
7236   thr->th.th_set_nproc = thr->th.th_teams_size.nteams = num_teams;
7237 
7238   // Remember the number of threads for inner parallel regions
7239   if (num_threads == 0) {
7240     if (!TCR_4(__kmp_init_middle))
7241       __kmp_middle_initialize(); // get __kmp_avail_proc calculated
7242     num_threads = __kmp_avail_proc / num_teams;
7243     if (num_teams * num_threads > __kmp_teams_max_nth) {
7244       // adjust num_threads w/o warning as it is not user setting
7245       num_threads = __kmp_teams_max_nth / num_teams;
7246     }
7247   } else {
7248     // This thread will be the master of the league masters
7249     // Store new thread limit; old limit is saved in th_cg_roots list
7250     thr->th.th_current_task->td_icvs.thread_limit = num_threads;
7251 
7252     if (num_teams * num_threads > __kmp_teams_max_nth) {
7253       int new_threads = __kmp_teams_max_nth / num_teams;
7254       if (!__kmp_reserve_warn) { // user asked for too many threads
7255         __kmp_reserve_warn = 1; // conflicts with KMP_TEAMS_THREAD_LIMIT
7256         __kmp_msg(kmp_ms_warning,
7257                   KMP_MSG(CantFormThrTeam, num_threads, new_threads),
7258                   KMP_HNT(Unset_ALL_THREADS), __kmp_msg_null);
7259       }
7260       num_threads = new_threads;
7261     }
7262   }
7263   thr->th.th_teams_size.nth = num_threads;
7264 }
7265 
7266 // Set the proc_bind var to use in the following parallel region.
7267 void __kmp_push_proc_bind(ident_t *id, int gtid, kmp_proc_bind_t proc_bind) {
7268   kmp_info_t *thr = __kmp_threads[gtid];
7269   thr->th.th_set_proc_bind = proc_bind;
7270 }
7271 
7272 #endif /* OMP_40_ENABLED */
7273 
7274 /* Launch the worker threads into the microtask. */
7275 
7276 void __kmp_internal_fork(ident_t *id, int gtid, kmp_team_t *team) {
7277   kmp_info_t *this_thr = __kmp_threads[gtid];
7278 
7279 #ifdef KMP_DEBUG
7280   int f;
7281 #endif /* KMP_DEBUG */
7282 
7283   KMP_DEBUG_ASSERT(team);
7284   KMP_DEBUG_ASSERT(this_thr->th.th_team == team);
7285   KMP_ASSERT(KMP_MASTER_GTID(gtid));
7286   KMP_MB(); /* Flush all pending memory write invalidates.  */
7287 
7288   team->t.t_construct = 0; /* no single directives seen yet */
7289   team->t.t_ordered.dt.t_value =
7290       0; /* thread 0 enters the ordered section first */
7291 
7292   /* Reset the identifiers on the dispatch buffer */
7293   KMP_DEBUG_ASSERT(team->t.t_disp_buffer);
7294   if (team->t.t_max_nproc > 1) {
7295     int i;
7296     for (i = 0; i < __kmp_dispatch_num_buffers; ++i) {
7297       team->t.t_disp_buffer[i].buffer_index = i;
7298 #if OMP_45_ENABLED
7299       team->t.t_disp_buffer[i].doacross_buf_idx = i;
7300 #endif
7301     }
7302   } else {
7303     team->t.t_disp_buffer[0].buffer_index = 0;
7304 #if OMP_45_ENABLED
7305     team->t.t_disp_buffer[0].doacross_buf_idx = 0;
7306 #endif
7307   }
7308 
7309   KMP_MB(); /* Flush all pending memory write invalidates.  */
7310   KMP_ASSERT(this_thr->th.th_team == team);
7311 
7312 #ifdef KMP_DEBUG
7313   for (f = 0; f < team->t.t_nproc; f++) {
7314     KMP_DEBUG_ASSERT(team->t.t_threads[f] &&
7315                      team->t.t_threads[f]->th.th_team_nproc == team->t.t_nproc);
7316   }
7317 #endif /* KMP_DEBUG */
7318 
7319   /* release the worker threads so they may begin working */
7320   __kmp_fork_barrier(gtid, 0);
7321 }
7322 
7323 void __kmp_internal_join(ident_t *id, int gtid, kmp_team_t *team) {
7324   kmp_info_t *this_thr = __kmp_threads[gtid];
7325 
7326   KMP_DEBUG_ASSERT(team);
7327   KMP_DEBUG_ASSERT(this_thr->th.th_team == team);
7328   KMP_ASSERT(KMP_MASTER_GTID(gtid));
7329   KMP_MB(); /* Flush all pending memory write invalidates.  */
7330 
7331 /* Join barrier after fork */
7332 
7333 #ifdef KMP_DEBUG
7334   if (__kmp_threads[gtid] &&
7335       __kmp_threads[gtid]->th.th_team_nproc != team->t.t_nproc) {
7336     __kmp_printf("GTID: %d, __kmp_threads[%d]=%p\n", gtid, gtid,
7337                  __kmp_threads[gtid]);
7338     __kmp_printf("__kmp_threads[%d]->th.th_team_nproc=%d, TEAM: %p, "
7339                  "team->t.t_nproc=%d\n",
7340                  gtid, __kmp_threads[gtid]->th.th_team_nproc, team,
7341                  team->t.t_nproc);
7342     __kmp_print_structure();
7343   }
7344   KMP_DEBUG_ASSERT(__kmp_threads[gtid] &&
7345                    __kmp_threads[gtid]->th.th_team_nproc == team->t.t_nproc);
7346 #endif /* KMP_DEBUG */
7347 
7348   __kmp_join_barrier(gtid); /* wait for everyone */
7349 #if OMPT_SUPPORT
7350   if (ompt_enabled.enabled &&
7351       this_thr->th.ompt_thread_info.state == ompt_state_wait_barrier_implicit) {
7352     int ds_tid = this_thr->th.th_info.ds.ds_tid;
7353     ompt_data_t *task_data = OMPT_CUR_TASK_DATA(this_thr);
7354     this_thr->th.ompt_thread_info.state = ompt_state_overhead;
7355 #if OMPT_OPTIONAL
7356     void *codeptr = NULL;
7357     if (KMP_MASTER_TID(ds_tid) &&
7358         (ompt_callbacks.ompt_callback(ompt_callback_sync_region_wait) ||
7359          ompt_callbacks.ompt_callback(ompt_callback_sync_region)))
7360       codeptr = OMPT_CUR_TEAM_INFO(this_thr)->master_return_address;
7361 
7362     if (ompt_enabled.ompt_callback_sync_region_wait) {
7363       ompt_callbacks.ompt_callback(ompt_callback_sync_region_wait)(
7364           ompt_sync_region_barrier_implicit, ompt_scope_end, NULL, task_data,
7365           codeptr);
7366     }
7367     if (ompt_enabled.ompt_callback_sync_region) {
7368       ompt_callbacks.ompt_callback(ompt_callback_sync_region)(
7369           ompt_sync_region_barrier_implicit, ompt_scope_end, NULL, task_data,
7370           codeptr);
7371     }
7372 #endif
7373     if (!KMP_MASTER_TID(ds_tid) && ompt_enabled.ompt_callback_implicit_task) {
7374       ompt_callbacks.ompt_callback(ompt_callback_implicit_task)(
7375           ompt_scope_end, NULL, task_data, 0, ds_tid, ompt_task_implicit); // TODO: Can this be ompt_task_initial?
7376     }
7377   }
7378 #endif
7379 
7380   KMP_MB(); /* Flush all pending memory write invalidates.  */
7381   KMP_ASSERT(this_thr->th.th_team == team);
7382 }
7383 
7384 /* ------------------------------------------------------------------------ */
7385 
7386 #ifdef USE_LOAD_BALANCE
7387 
7388 // Return the worker threads actively spinning in the hot team, if we
7389 // are at the outermost level of parallelism.  Otherwise, return 0.
7390 static int __kmp_active_hot_team_nproc(kmp_root_t *root) {
7391   int i;
7392   int retval;
7393   kmp_team_t *hot_team;
7394 
7395   if (root->r.r_active) {
7396     return 0;
7397   }
7398   hot_team = root->r.r_hot_team;
7399   if (__kmp_dflt_blocktime == KMP_MAX_BLOCKTIME) {
7400     return hot_team->t.t_nproc - 1; // Don't count master thread
7401   }
7402 
7403   // Skip the master thread - it is accounted for elsewhere.
7404   retval = 0;
7405   for (i = 1; i < hot_team->t.t_nproc; i++) {
7406     if (hot_team->t.t_threads[i]->th.th_active) {
7407       retval++;
7408     }
7409   }
7410   return retval;
7411 }
7412 
7413 // Perform an automatic adjustment to the number of
7414 // threads used by the next parallel region.
7415 static int __kmp_load_balance_nproc(kmp_root_t *root, int set_nproc) {
7416   int retval;
7417   int pool_active;
7418   int hot_team_active;
7419   int team_curr_active;
7420   int system_active;
7421 
7422   KB_TRACE(20, ("__kmp_load_balance_nproc: called root:%p set_nproc:%d\n", root,
7423                 set_nproc));
7424   KMP_DEBUG_ASSERT(root);
7425   KMP_DEBUG_ASSERT(root->r.r_root_team->t.t_threads[0]
7426                        ->th.th_current_task->td_icvs.dynamic == TRUE);
7427   KMP_DEBUG_ASSERT(set_nproc > 1);
7428 
7429   if (set_nproc == 1) {
7430     KB_TRACE(20, ("__kmp_load_balance_nproc: serial execution.\n"));
7431     return 1;
7432   }
7433 
7434   // Threads that are active in the thread pool, active in the hot team for this
7435   // particular root (if we are at the outer par level), and the currently
7436   // executing thread (to become the master) are available to add to the new
7437   // team, but are currently contributing to the system load, and must be
7438   // accounted for.
7439   pool_active = __kmp_thread_pool_active_nth;
7440   hot_team_active = __kmp_active_hot_team_nproc(root);
7441   team_curr_active = pool_active + hot_team_active + 1;
7442 
7443   // Check the system load.
7444   system_active = __kmp_get_load_balance(__kmp_avail_proc + team_curr_active);
7445   KB_TRACE(30, ("__kmp_load_balance_nproc: system active = %d pool active = %d "
7446                 "hot team active = %d\n",
7447                 system_active, pool_active, hot_team_active));
7448 
7449   if (system_active < 0) {
7450     // There was an error reading the necessary info from /proc, so use the
7451     // thread limit algorithm instead. Once we set __kmp_global.g.g_dynamic_mode
7452     // = dynamic_thread_limit, we shouldn't wind up getting back here.
7453     __kmp_global.g.g_dynamic_mode = dynamic_thread_limit;
7454     KMP_WARNING(CantLoadBalUsing, "KMP_DYNAMIC_MODE=thread limit");
7455 
7456     // Make this call behave like the thread limit algorithm.
7457     retval = __kmp_avail_proc - __kmp_nth +
7458              (root->r.r_active ? 1 : root->r.r_hot_team->t.t_nproc);
7459     if (retval > set_nproc) {
7460       retval = set_nproc;
7461     }
7462     if (retval < KMP_MIN_NTH) {
7463       retval = KMP_MIN_NTH;
7464     }
7465 
7466     KB_TRACE(20, ("__kmp_load_balance_nproc: thread limit exit. retval:%d\n",
7467                   retval));
7468     return retval;
7469   }
7470 
7471   // There is a slight delay in the load balance algorithm in detecting new
7472   // running procs. The real system load at this instant should be at least as
7473   // large as the #active omp thread that are available to add to the team.
7474   if (system_active < team_curr_active) {
7475     system_active = team_curr_active;
7476   }
7477   retval = __kmp_avail_proc - system_active + team_curr_active;
7478   if (retval > set_nproc) {
7479     retval = set_nproc;
7480   }
7481   if (retval < KMP_MIN_NTH) {
7482     retval = KMP_MIN_NTH;
7483   }
7484 
7485   KB_TRACE(20, ("__kmp_load_balance_nproc: exit. retval:%d\n", retval));
7486   return retval;
7487 } // __kmp_load_balance_nproc()
7488 
7489 #endif /* USE_LOAD_BALANCE */
7490 
7491 /* ------------------------------------------------------------------------ */
7492 
7493 /* NOTE: this is called with the __kmp_init_lock held */
7494 void __kmp_cleanup(void) {
7495   int f;
7496 
7497   KA_TRACE(10, ("__kmp_cleanup: enter\n"));
7498 
7499   if (TCR_4(__kmp_init_parallel)) {
7500 #if KMP_HANDLE_SIGNALS
7501     __kmp_remove_signals();
7502 #endif
7503     TCW_4(__kmp_init_parallel, FALSE);
7504   }
7505 
7506   if (TCR_4(__kmp_init_middle)) {
7507 #if KMP_AFFINITY_SUPPORTED
7508     __kmp_affinity_uninitialize();
7509 #endif /* KMP_AFFINITY_SUPPORTED */
7510     __kmp_cleanup_hierarchy();
7511     TCW_4(__kmp_init_middle, FALSE);
7512   }
7513 
7514   KA_TRACE(10, ("__kmp_cleanup: go serial cleanup\n"));
7515 
7516   if (__kmp_init_serial) {
7517     __kmp_runtime_destroy();
7518     __kmp_init_serial = FALSE;
7519   }
7520 
7521   __kmp_cleanup_threadprivate_caches();
7522 
7523   for (f = 0; f < __kmp_threads_capacity; f++) {
7524     if (__kmp_root[f] != NULL) {
7525       __kmp_free(__kmp_root[f]);
7526       __kmp_root[f] = NULL;
7527     }
7528   }
7529   __kmp_free(__kmp_threads);
7530   // __kmp_threads and __kmp_root were allocated at once, as single block, so
7531   // there is no need in freeing __kmp_root.
7532   __kmp_threads = NULL;
7533   __kmp_root = NULL;
7534   __kmp_threads_capacity = 0;
7535 
7536 #if KMP_USE_DYNAMIC_LOCK
7537   __kmp_cleanup_indirect_user_locks();
7538 #else
7539   __kmp_cleanup_user_locks();
7540 #endif
7541 
7542 #if KMP_AFFINITY_SUPPORTED
7543   KMP_INTERNAL_FREE(CCAST(char *, __kmp_cpuinfo_file));
7544   __kmp_cpuinfo_file = NULL;
7545 #endif /* KMP_AFFINITY_SUPPORTED */
7546 
7547 #if KMP_USE_ADAPTIVE_LOCKS
7548 #if KMP_DEBUG_ADAPTIVE_LOCKS
7549   __kmp_print_speculative_stats();
7550 #endif
7551 #endif
7552   KMP_INTERNAL_FREE(__kmp_nested_nth.nth);
7553   __kmp_nested_nth.nth = NULL;
7554   __kmp_nested_nth.size = 0;
7555   __kmp_nested_nth.used = 0;
7556   KMP_INTERNAL_FREE(__kmp_nested_proc_bind.bind_types);
7557   __kmp_nested_proc_bind.bind_types = NULL;
7558   __kmp_nested_proc_bind.size = 0;
7559   __kmp_nested_proc_bind.used = 0;
7560 #if OMP_50_ENABLED
7561   if (__kmp_affinity_format) {
7562     KMP_INTERNAL_FREE(__kmp_affinity_format);
7563     __kmp_affinity_format = NULL;
7564   }
7565 #endif
7566 
7567   __kmp_i18n_catclose();
7568 
7569 #if KMP_USE_HIER_SCHED
7570   __kmp_hier_scheds.deallocate();
7571 #endif
7572 
7573 #if KMP_STATS_ENABLED
7574   __kmp_stats_fini();
7575 #endif
7576 
7577   KA_TRACE(10, ("__kmp_cleanup: exit\n"));
7578 }
7579 
7580 /* ------------------------------------------------------------------------ */
7581 
7582 int __kmp_ignore_mppbeg(void) {
7583   char *env;
7584 
7585   if ((env = getenv("KMP_IGNORE_MPPBEG")) != NULL) {
7586     if (__kmp_str_match_false(env))
7587       return FALSE;
7588   }
7589   // By default __kmpc_begin() is no-op.
7590   return TRUE;
7591 }
7592 
7593 int __kmp_ignore_mppend(void) {
7594   char *env;
7595 
7596   if ((env = getenv("KMP_IGNORE_MPPEND")) != NULL) {
7597     if (__kmp_str_match_false(env))
7598       return FALSE;
7599   }
7600   // By default __kmpc_end() is no-op.
7601   return TRUE;
7602 }
7603 
7604 void __kmp_internal_begin(void) {
7605   int gtid;
7606   kmp_root_t *root;
7607 
7608   /* this is a very important step as it will register new sibling threads
7609      and assign these new uber threads a new gtid */
7610   gtid = __kmp_entry_gtid();
7611   root = __kmp_threads[gtid]->th.th_root;
7612   KMP_ASSERT(KMP_UBER_GTID(gtid));
7613 
7614   if (root->r.r_begin)
7615     return;
7616   __kmp_acquire_lock(&root->r.r_begin_lock, gtid);
7617   if (root->r.r_begin) {
7618     __kmp_release_lock(&root->r.r_begin_lock, gtid);
7619     return;
7620   }
7621 
7622   root->r.r_begin = TRUE;
7623 
7624   __kmp_release_lock(&root->r.r_begin_lock, gtid);
7625 }
7626 
7627 /* ------------------------------------------------------------------------ */
7628 
7629 void __kmp_user_set_library(enum library_type arg) {
7630   int gtid;
7631   kmp_root_t *root;
7632   kmp_info_t *thread;
7633 
7634   /* first, make sure we are initialized so we can get our gtid */
7635 
7636   gtid = __kmp_entry_gtid();
7637   thread = __kmp_threads[gtid];
7638 
7639   root = thread->th.th_root;
7640 
7641   KA_TRACE(20, ("__kmp_user_set_library: enter T#%d, arg: %d, %d\n", gtid, arg,
7642                 library_serial));
7643   if (root->r.r_in_parallel) { /* Must be called in serial section of top-level
7644                                   thread */
7645     KMP_WARNING(SetLibraryIncorrectCall);
7646     return;
7647   }
7648 
7649   switch (arg) {
7650   case library_serial:
7651     thread->th.th_set_nproc = 0;
7652     set__nproc(thread, 1);
7653     break;
7654   case library_turnaround:
7655     thread->th.th_set_nproc = 0;
7656     set__nproc(thread, __kmp_dflt_team_nth ? __kmp_dflt_team_nth
7657                                            : __kmp_dflt_team_nth_ub);
7658     break;
7659   case library_throughput:
7660     thread->th.th_set_nproc = 0;
7661     set__nproc(thread, __kmp_dflt_team_nth ? __kmp_dflt_team_nth
7662                                            : __kmp_dflt_team_nth_ub);
7663     break;
7664   default:
7665     KMP_FATAL(UnknownLibraryType, arg);
7666   }
7667 
7668   __kmp_aux_set_library(arg);
7669 }
7670 
7671 void __kmp_aux_set_stacksize(size_t arg) {
7672   if (!__kmp_init_serial)
7673     __kmp_serial_initialize();
7674 
7675 #if KMP_OS_DARWIN
7676   if (arg & (0x1000 - 1)) {
7677     arg &= ~(0x1000 - 1);
7678     if (arg + 0x1000) /* check for overflow if we round up */
7679       arg += 0x1000;
7680   }
7681 #endif
7682   __kmp_acquire_bootstrap_lock(&__kmp_initz_lock);
7683 
7684   /* only change the default stacksize before the first parallel region */
7685   if (!TCR_4(__kmp_init_parallel)) {
7686     size_t value = arg; /* argument is in bytes */
7687 
7688     if (value < __kmp_sys_min_stksize)
7689       value = __kmp_sys_min_stksize;
7690     else if (value > KMP_MAX_STKSIZE)
7691       value = KMP_MAX_STKSIZE;
7692 
7693     __kmp_stksize = value;
7694 
7695     __kmp_env_stksize = TRUE; /* was KMP_STACKSIZE specified? */
7696   }
7697 
7698   __kmp_release_bootstrap_lock(&__kmp_initz_lock);
7699 }
7700 
7701 /* set the behaviour of the runtime library */
7702 /* TODO this can cause some odd behaviour with sibling parallelism... */
7703 void __kmp_aux_set_library(enum library_type arg) {
7704   __kmp_library = arg;
7705 
7706   switch (__kmp_library) {
7707   case library_serial: {
7708     KMP_INFORM(LibraryIsSerial);
7709   } break;
7710   case library_turnaround:
7711     if (__kmp_use_yield == 1 && !__kmp_use_yield_exp_set)
7712       __kmp_use_yield = 2; // only yield when oversubscribed
7713     break;
7714   case library_throughput:
7715     if (__kmp_dflt_blocktime == KMP_MAX_BLOCKTIME)
7716       __kmp_dflt_blocktime = 200;
7717     break;
7718   default:
7719     KMP_FATAL(UnknownLibraryType, arg);
7720   }
7721 }
7722 
7723 /* Getting team information common for all team API */
7724 // Returns NULL if not in teams construct
7725 static kmp_team_t *__kmp_aux_get_team_info(int &teams_serialized) {
7726   kmp_info_t *thr = __kmp_entry_thread();
7727   teams_serialized = 0;
7728   if (thr->th.th_teams_microtask) {
7729     kmp_team_t *team = thr->th.th_team;
7730     int tlevel = thr->th.th_teams_level; // the level of the teams construct
7731     int ii = team->t.t_level;
7732     teams_serialized = team->t.t_serialized;
7733     int level = tlevel + 1;
7734     KMP_DEBUG_ASSERT(ii >= tlevel);
7735     while (ii > level) {
7736       for (teams_serialized = team->t.t_serialized;
7737            (teams_serialized > 0) && (ii > level); teams_serialized--, ii--) {
7738       }
7739       if (team->t.t_serialized && (!teams_serialized)) {
7740         team = team->t.t_parent;
7741         continue;
7742       }
7743       if (ii > level) {
7744         team = team->t.t_parent;
7745         ii--;
7746       }
7747     }
7748     return team;
7749   }
7750   return NULL;
7751 }
7752 
7753 int __kmp_aux_get_team_num() {
7754   int serialized;
7755   kmp_team_t *team = __kmp_aux_get_team_info(serialized);
7756   if (team) {
7757     if (serialized > 1) {
7758       return 0; // teams region is serialized ( 1 team of 1 thread ).
7759     } else {
7760       return team->t.t_master_tid;
7761     }
7762   }
7763   return 0;
7764 }
7765 
7766 int __kmp_aux_get_num_teams() {
7767   int serialized;
7768   kmp_team_t *team = __kmp_aux_get_team_info(serialized);
7769   if (team) {
7770     if (serialized > 1) {
7771       return 1;
7772     } else {
7773       return team->t.t_parent->t.t_nproc;
7774     }
7775   }
7776   return 1;
7777 }
7778 
7779 /* ------------------------------------------------------------------------ */
7780 
7781 #if OMP_50_ENABLED
7782 /*
7783  * Affinity Format Parser
7784  *
7785  * Field is in form of: %[[[0].]size]type
7786  * % and type are required (%% means print a literal '%')
7787  * type is either single char or long name surrounded by {},
7788  * e.g., N or {num_threads}
7789  * 0 => leading zeros
7790  * . => right justified when size is specified
7791  * by default output is left justified
7792  * size is the *minimum* field length
7793  * All other characters are printed as is
7794  *
7795  * Available field types:
7796  * L {thread_level}      - omp_get_level()
7797  * n {thread_num}        - omp_get_thread_num()
7798  * h {host}              - name of host machine
7799  * P {process_id}        - process id (integer)
7800  * T {thread_identifier} - native thread identifier (integer)
7801  * N {num_threads}       - omp_get_num_threads()
7802  * A {ancestor_tnum}     - omp_get_ancestor_thread_num(omp_get_level()-1)
7803  * a {thread_affinity}   - comma separated list of integers or integer ranges
7804  *                         (values of affinity mask)
7805  *
7806  * Implementation-specific field types can be added
7807  * If a type is unknown, print "undefined"
7808 */
7809 
7810 // Structure holding the short name, long name, and corresponding data type
7811 // for snprintf.  A table of these will represent the entire valid keyword
7812 // field types.
7813 typedef struct kmp_affinity_format_field_t {
7814   char short_name; // from spec e.g., L -> thread level
7815   const char *long_name; // from spec thread_level -> thread level
7816   char field_format; // data type for snprintf (typically 'd' or 's'
7817   // for integer or string)
7818 } kmp_affinity_format_field_t;
7819 
7820 static const kmp_affinity_format_field_t __kmp_affinity_format_table[] = {
7821 #if KMP_AFFINITY_SUPPORTED
7822     {'A', "thread_affinity", 's'},
7823 #endif
7824     {'t', "team_num", 'd'},
7825     {'T', "num_teams", 'd'},
7826     {'L', "nesting_level", 'd'},
7827     {'n', "thread_num", 'd'},
7828     {'N', "num_threads", 'd'},
7829     {'a', "ancestor_tnum", 'd'},
7830     {'H', "host", 's'},
7831     {'P', "process_id", 'd'},
7832     {'i', "native_thread_id", 'd'}};
7833 
7834 // Return the number of characters it takes to hold field
7835 static int __kmp_aux_capture_affinity_field(int gtid, const kmp_info_t *th,
7836                                             const char **ptr,
7837                                             kmp_str_buf_t *field_buffer) {
7838   int rc, format_index, field_value;
7839   const char *width_left, *width_right;
7840   bool pad_zeros, right_justify, parse_long_name, found_valid_name;
7841   static const int FORMAT_SIZE = 20;
7842   char format[FORMAT_SIZE] = {0};
7843   char absolute_short_name = 0;
7844 
7845   KMP_DEBUG_ASSERT(gtid >= 0);
7846   KMP_DEBUG_ASSERT(th);
7847   KMP_DEBUG_ASSERT(**ptr == '%');
7848   KMP_DEBUG_ASSERT(field_buffer);
7849 
7850   __kmp_str_buf_clear(field_buffer);
7851 
7852   // Skip the initial %
7853   (*ptr)++;
7854 
7855   // Check for %% first
7856   if (**ptr == '%') {
7857     __kmp_str_buf_cat(field_buffer, "%", 1);
7858     (*ptr)++; // skip over the second %
7859     return 1;
7860   }
7861 
7862   // Parse field modifiers if they are present
7863   pad_zeros = false;
7864   if (**ptr == '0') {
7865     pad_zeros = true;
7866     (*ptr)++; // skip over 0
7867   }
7868   right_justify = false;
7869   if (**ptr == '.') {
7870     right_justify = true;
7871     (*ptr)++; // skip over .
7872   }
7873   // Parse width of field: [width_left, width_right)
7874   width_left = width_right = NULL;
7875   if (**ptr >= '0' && **ptr <= '9') {
7876     width_left = *ptr;
7877     SKIP_DIGITS(*ptr);
7878     width_right = *ptr;
7879   }
7880 
7881   // Create the format for KMP_SNPRINTF based on flags parsed above
7882   format_index = 0;
7883   format[format_index++] = '%';
7884   if (!right_justify)
7885     format[format_index++] = '-';
7886   if (pad_zeros)
7887     format[format_index++] = '0';
7888   if (width_left && width_right) {
7889     int i = 0;
7890     // Only allow 8 digit number widths.
7891     // This also prevents overflowing format variable
7892     while (i < 8 && width_left < width_right) {
7893       format[format_index++] = *width_left;
7894       width_left++;
7895       i++;
7896     }
7897   }
7898 
7899   // Parse a name (long or short)
7900   // Canonicalize the name into absolute_short_name
7901   found_valid_name = false;
7902   parse_long_name = (**ptr == '{');
7903   if (parse_long_name)
7904     (*ptr)++; // skip initial left brace
7905   for (size_t i = 0; i < sizeof(__kmp_affinity_format_table) /
7906                              sizeof(__kmp_affinity_format_table[0]);
7907        ++i) {
7908     char short_name = __kmp_affinity_format_table[i].short_name;
7909     const char *long_name = __kmp_affinity_format_table[i].long_name;
7910     char field_format = __kmp_affinity_format_table[i].field_format;
7911     if (parse_long_name) {
7912       int length = KMP_STRLEN(long_name);
7913       if (strncmp(*ptr, long_name, length) == 0) {
7914         found_valid_name = true;
7915         (*ptr) += length; // skip the long name
7916       }
7917     } else if (**ptr == short_name) {
7918       found_valid_name = true;
7919       (*ptr)++; // skip the short name
7920     }
7921     if (found_valid_name) {
7922       format[format_index++] = field_format;
7923       format[format_index++] = '\0';
7924       absolute_short_name = short_name;
7925       break;
7926     }
7927   }
7928   if (parse_long_name) {
7929     if (**ptr != '}') {
7930       absolute_short_name = 0;
7931     } else {
7932       (*ptr)++; // skip over the right brace
7933     }
7934   }
7935 
7936   // Attempt to fill the buffer with the requested
7937   // value using snprintf within __kmp_str_buf_print()
7938   switch (absolute_short_name) {
7939   case 't':
7940     rc = __kmp_str_buf_print(field_buffer, format, __kmp_aux_get_team_num());
7941     break;
7942   case 'T':
7943     rc = __kmp_str_buf_print(field_buffer, format, __kmp_aux_get_num_teams());
7944     break;
7945   case 'L':
7946     rc = __kmp_str_buf_print(field_buffer, format, th->th.th_team->t.t_level);
7947     break;
7948   case 'n':
7949     rc = __kmp_str_buf_print(field_buffer, format, __kmp_tid_from_gtid(gtid));
7950     break;
7951   case 'H': {
7952     static const int BUFFER_SIZE = 256;
7953     char buf[BUFFER_SIZE];
7954     __kmp_expand_host_name(buf, BUFFER_SIZE);
7955     rc = __kmp_str_buf_print(field_buffer, format, buf);
7956   } break;
7957   case 'P':
7958     rc = __kmp_str_buf_print(field_buffer, format, getpid());
7959     break;
7960   case 'i':
7961     rc = __kmp_str_buf_print(field_buffer, format, __kmp_gettid());
7962     break;
7963   case 'N':
7964     rc = __kmp_str_buf_print(field_buffer, format, th->th.th_team->t.t_nproc);
7965     break;
7966   case 'a':
7967     field_value =
7968         __kmp_get_ancestor_thread_num(gtid, th->th.th_team->t.t_level - 1);
7969     rc = __kmp_str_buf_print(field_buffer, format, field_value);
7970     break;
7971 #if KMP_AFFINITY_SUPPORTED
7972   case 'A': {
7973     kmp_str_buf_t buf;
7974     __kmp_str_buf_init(&buf);
7975     __kmp_affinity_str_buf_mask(&buf, th->th.th_affin_mask);
7976     rc = __kmp_str_buf_print(field_buffer, format, buf.str);
7977     __kmp_str_buf_free(&buf);
7978   } break;
7979 #endif
7980   default:
7981     // According to spec, If an implementation does not have info for field
7982     // type, then "undefined" is printed
7983     rc = __kmp_str_buf_print(field_buffer, "%s", "undefined");
7984     // Skip the field
7985     if (parse_long_name) {
7986       SKIP_TOKEN(*ptr);
7987       if (**ptr == '}')
7988         (*ptr)++;
7989     } else {
7990       (*ptr)++;
7991     }
7992   }
7993 
7994   KMP_ASSERT(format_index <= FORMAT_SIZE);
7995   return rc;
7996 }
7997 
7998 /*
7999  * Return number of characters needed to hold the affinity string
8000  * (not including null byte character)
8001  * The resultant string is printed to buffer, which the caller can then
8002  * handle afterwards
8003 */
8004 size_t __kmp_aux_capture_affinity(int gtid, const char *format,
8005                                   kmp_str_buf_t *buffer) {
8006   const char *parse_ptr;
8007   size_t retval;
8008   const kmp_info_t *th;
8009   kmp_str_buf_t field;
8010 
8011   KMP_DEBUG_ASSERT(buffer);
8012   KMP_DEBUG_ASSERT(gtid >= 0);
8013 
8014   __kmp_str_buf_init(&field);
8015   __kmp_str_buf_clear(buffer);
8016 
8017   th = __kmp_threads[gtid];
8018   retval = 0;
8019 
8020   // If format is NULL or zero-length string, then we use
8021   // affinity-format-var ICV
8022   parse_ptr = format;
8023   if (parse_ptr == NULL || *parse_ptr == '\0') {
8024     parse_ptr = __kmp_affinity_format;
8025   }
8026   KMP_DEBUG_ASSERT(parse_ptr);
8027 
8028   while (*parse_ptr != '\0') {
8029     // Parse a field
8030     if (*parse_ptr == '%') {
8031       // Put field in the buffer
8032       int rc = __kmp_aux_capture_affinity_field(gtid, th, &parse_ptr, &field);
8033       __kmp_str_buf_catbuf(buffer, &field);
8034       retval += rc;
8035     } else {
8036       // Put literal character in buffer
8037       __kmp_str_buf_cat(buffer, parse_ptr, 1);
8038       retval++;
8039       parse_ptr++;
8040     }
8041   }
8042   __kmp_str_buf_free(&field);
8043   return retval;
8044 }
8045 
8046 // Displays the affinity string to stdout
8047 void __kmp_aux_display_affinity(int gtid, const char *format) {
8048   kmp_str_buf_t buf;
8049   __kmp_str_buf_init(&buf);
8050   __kmp_aux_capture_affinity(gtid, format, &buf);
8051   __kmp_fprintf(kmp_out, "%s" KMP_END_OF_LINE, buf.str);
8052   __kmp_str_buf_free(&buf);
8053 }
8054 #endif // OMP_50_ENABLED
8055 
8056 /* ------------------------------------------------------------------------ */
8057 
8058 void __kmp_aux_set_blocktime(int arg, kmp_info_t *thread, int tid) {
8059   int blocktime = arg; /* argument is in milliseconds */
8060 #if KMP_USE_MONITOR
8061   int bt_intervals;
8062 #endif
8063   int bt_set;
8064 
8065   __kmp_save_internal_controls(thread);
8066 
8067   /* Normalize and set blocktime for the teams */
8068   if (blocktime < KMP_MIN_BLOCKTIME)
8069     blocktime = KMP_MIN_BLOCKTIME;
8070   else if (blocktime > KMP_MAX_BLOCKTIME)
8071     blocktime = KMP_MAX_BLOCKTIME;
8072 
8073   set__blocktime_team(thread->th.th_team, tid, blocktime);
8074   set__blocktime_team(thread->th.th_serial_team, 0, blocktime);
8075 
8076 #if KMP_USE_MONITOR
8077   /* Calculate and set blocktime intervals for the teams */
8078   bt_intervals = KMP_INTERVALS_FROM_BLOCKTIME(blocktime, __kmp_monitor_wakeups);
8079 
8080   set__bt_intervals_team(thread->th.th_team, tid, bt_intervals);
8081   set__bt_intervals_team(thread->th.th_serial_team, 0, bt_intervals);
8082 #endif
8083 
8084   /* Set whether blocktime has been set to "TRUE" */
8085   bt_set = TRUE;
8086 
8087   set__bt_set_team(thread->th.th_team, tid, bt_set);
8088   set__bt_set_team(thread->th.th_serial_team, 0, bt_set);
8089 #if KMP_USE_MONITOR
8090   KF_TRACE(10, ("kmp_set_blocktime: T#%d(%d:%d), blocktime=%d, "
8091                 "bt_intervals=%d, monitor_updates=%d\n",
8092                 __kmp_gtid_from_tid(tid, thread->th.th_team),
8093                 thread->th.th_team->t.t_id, tid, blocktime, bt_intervals,
8094                 __kmp_monitor_wakeups));
8095 #else
8096   KF_TRACE(10, ("kmp_set_blocktime: T#%d(%d:%d), blocktime=%d\n",
8097                 __kmp_gtid_from_tid(tid, thread->th.th_team),
8098                 thread->th.th_team->t.t_id, tid, blocktime));
8099 #endif
8100 }
8101 
8102 void __kmp_aux_set_defaults(char const *str, int len) {
8103   if (!__kmp_init_serial) {
8104     __kmp_serial_initialize();
8105   }
8106   __kmp_env_initialize(str);
8107 
8108   if (__kmp_settings
8109 #if OMP_40_ENABLED
8110       || __kmp_display_env || __kmp_display_env_verbose
8111 #endif // OMP_40_ENABLED
8112       ) {
8113     __kmp_env_print();
8114   }
8115 } // __kmp_aux_set_defaults
8116 
8117 /* ------------------------------------------------------------------------ */
8118 /* internal fast reduction routines */
8119 
8120 PACKED_REDUCTION_METHOD_T
8121 __kmp_determine_reduction_method(
8122     ident_t *loc, kmp_int32 global_tid, kmp_int32 num_vars, size_t reduce_size,
8123     void *reduce_data, void (*reduce_func)(void *lhs_data, void *rhs_data),
8124     kmp_critical_name *lck) {
8125 
8126   // Default reduction method: critical construct ( lck != NULL, like in current
8127   // PAROPT )
8128   // If ( reduce_data!=NULL && reduce_func!=NULL ): the tree-reduction method
8129   // can be selected by RTL
8130   // If loc->flags contains KMP_IDENT_ATOMIC_REDUCE, the atomic reduce method
8131   // can be selected by RTL
8132   // Finally, it's up to OpenMP RTL to make a decision on which method to select
8133   // among generated by PAROPT.
8134 
8135   PACKED_REDUCTION_METHOD_T retval;
8136 
8137   int team_size;
8138 
8139   KMP_DEBUG_ASSERT(loc); // it would be nice to test ( loc != 0 )
8140   KMP_DEBUG_ASSERT(lck); // it would be nice to test ( lck != 0 )
8141 
8142 #define FAST_REDUCTION_ATOMIC_METHOD_GENERATED                                 \
8143   ((loc->flags & (KMP_IDENT_ATOMIC_REDUCE)) == (KMP_IDENT_ATOMIC_REDUCE))
8144 #define FAST_REDUCTION_TREE_METHOD_GENERATED ((reduce_data) && (reduce_func))
8145 
8146   retval = critical_reduce_block;
8147 
8148   // another choice of getting a team size (with 1 dynamic deference) is slower
8149   team_size = __kmp_get_team_num_threads(global_tid);
8150   if (team_size == 1) {
8151 
8152     retval = empty_reduce_block;
8153 
8154   } else {
8155 
8156     int atomic_available = FAST_REDUCTION_ATOMIC_METHOD_GENERATED;
8157 
8158 #if KMP_ARCH_X86_64 || KMP_ARCH_PPC64 || KMP_ARCH_AARCH64 || KMP_ARCH_MIPS64
8159 
8160 #if KMP_OS_LINUX || KMP_OS_DRAGONFLY || KMP_OS_FREEBSD || KMP_OS_NETBSD ||     \
8161     KMP_OS_OPENBSD || KMP_OS_WINDOWS || KMP_OS_DARWIN || KMP_OS_HURD
8162 
8163     int teamsize_cutoff = 4;
8164 
8165 #if KMP_MIC_SUPPORTED
8166     if (__kmp_mic_type != non_mic) {
8167       teamsize_cutoff = 8;
8168     }
8169 #endif
8170     int tree_available = FAST_REDUCTION_TREE_METHOD_GENERATED;
8171     if (tree_available) {
8172       if (team_size <= teamsize_cutoff) {
8173         if (atomic_available) {
8174           retval = atomic_reduce_block;
8175         }
8176       } else {
8177         retval = TREE_REDUCE_BLOCK_WITH_REDUCTION_BARRIER;
8178       }
8179     } else if (atomic_available) {
8180       retval = atomic_reduce_block;
8181     }
8182 #else
8183 #error "Unknown or unsupported OS"
8184 #endif // KMP_OS_LINUX || KMP_OS_DRAGONFLY || KMP_OS_FREEBSD || KMP_OS_NETBSD ||
8185        // KMP_OS_OPENBSD || KMP_OS_WINDOWS || KMP_OS_DARWIN || KMP_OS_HURD
8186 
8187 #elif KMP_ARCH_X86 || KMP_ARCH_ARM || KMP_ARCH_AARCH || KMP_ARCH_MIPS
8188 
8189 #if KMP_OS_LINUX || KMP_OS_WINDOWS || KMP_OS_HURD
8190 
8191     // basic tuning
8192 
8193     if (atomic_available) {
8194       if (num_vars <= 2) { // && ( team_size <= 8 ) due to false-sharing ???
8195         retval = atomic_reduce_block;
8196       }
8197     } // otherwise: use critical section
8198 
8199 #elif KMP_OS_DARWIN
8200 
8201     int tree_available = FAST_REDUCTION_TREE_METHOD_GENERATED;
8202     if (atomic_available && (num_vars <= 3)) {
8203       retval = atomic_reduce_block;
8204     } else if (tree_available) {
8205       if ((reduce_size > (9 * sizeof(kmp_real64))) &&
8206           (reduce_size < (2000 * sizeof(kmp_real64)))) {
8207         retval = TREE_REDUCE_BLOCK_WITH_PLAIN_BARRIER;
8208       }
8209     } // otherwise: use critical section
8210 
8211 #else
8212 #error "Unknown or unsupported OS"
8213 #endif
8214 
8215 #else
8216 #error "Unknown or unsupported architecture"
8217 #endif
8218   }
8219 
8220   // KMP_FORCE_REDUCTION
8221 
8222   // If the team is serialized (team_size == 1), ignore the forced reduction
8223   // method and stay with the unsynchronized method (empty_reduce_block)
8224   if (__kmp_force_reduction_method != reduction_method_not_defined &&
8225       team_size != 1) {
8226 
8227     PACKED_REDUCTION_METHOD_T forced_retval = critical_reduce_block;
8228 
8229     int atomic_available, tree_available;
8230 
8231     switch ((forced_retval = __kmp_force_reduction_method)) {
8232     case critical_reduce_block:
8233       KMP_ASSERT(lck); // lck should be != 0
8234       break;
8235 
8236     case atomic_reduce_block:
8237       atomic_available = FAST_REDUCTION_ATOMIC_METHOD_GENERATED;
8238       if (!atomic_available) {
8239         KMP_WARNING(RedMethodNotSupported, "atomic");
8240         forced_retval = critical_reduce_block;
8241       }
8242       break;
8243 
8244     case tree_reduce_block:
8245       tree_available = FAST_REDUCTION_TREE_METHOD_GENERATED;
8246       if (!tree_available) {
8247         KMP_WARNING(RedMethodNotSupported, "tree");
8248         forced_retval = critical_reduce_block;
8249       } else {
8250 #if KMP_FAST_REDUCTION_BARRIER
8251         forced_retval = TREE_REDUCE_BLOCK_WITH_REDUCTION_BARRIER;
8252 #endif
8253       }
8254       break;
8255 
8256     default:
8257       KMP_ASSERT(0); // "unsupported method specified"
8258     }
8259 
8260     retval = forced_retval;
8261   }
8262 
8263   KA_TRACE(10, ("reduction method selected=%08x\n", retval));
8264 
8265 #undef FAST_REDUCTION_TREE_METHOD_GENERATED
8266 #undef FAST_REDUCTION_ATOMIC_METHOD_GENERATED
8267 
8268   return (retval);
8269 }
8270 
8271 // this function is for testing set/get/determine reduce method
8272 kmp_int32 __kmp_get_reduce_method(void) {
8273   return ((__kmp_entry_thread()->th.th_local.packed_reduction_method) >> 8);
8274 }
8275 
8276 #if OMP_50_ENABLED
8277 
8278 // Soft pause sets up threads to ignore blocktime and just go to sleep.
8279 // Spin-wait code checks __kmp_pause_status and reacts accordingly.
8280 void __kmp_soft_pause() { __kmp_pause_status = kmp_soft_paused; }
8281 
8282 // Hard pause shuts down the runtime completely.  Resume happens naturally when
8283 // OpenMP is used subsequently.
8284 void __kmp_hard_pause() {
8285   __kmp_pause_status = kmp_hard_paused;
8286   __kmp_internal_end_thread(-1);
8287 }
8288 
8289 // Soft resume sets __kmp_pause_status, and wakes up all threads.
8290 void __kmp_resume_if_soft_paused() {
8291   if (__kmp_pause_status == kmp_soft_paused) {
8292     __kmp_pause_status = kmp_not_paused;
8293 
8294     for (int gtid = 1; gtid < __kmp_threads_capacity; ++gtid) {
8295       kmp_info_t *thread = __kmp_threads[gtid];
8296       if (thread) { // Wake it if sleeping
8297         kmp_flag_64 fl(&thread->th.th_bar[bs_forkjoin_barrier].bb.b_go, thread);
8298         if (fl.is_sleeping())
8299           fl.resume(gtid);
8300         else if (__kmp_try_suspend_mx(thread)) { // got suspend lock
8301           __kmp_unlock_suspend_mx(thread); // unlock it; it won't sleep
8302         } else { // thread holds the lock and may sleep soon
8303           do { // until either the thread sleeps, or we can get the lock
8304             if (fl.is_sleeping()) {
8305               fl.resume(gtid);
8306               break;
8307             } else if (__kmp_try_suspend_mx(thread)) {
8308               __kmp_unlock_suspend_mx(thread);
8309               break;
8310             }
8311           } while (1);
8312         }
8313       }
8314     }
8315   }
8316 }
8317 
8318 // This function is called via __kmpc_pause_resource. Returns 0 if successful.
8319 // TODO: add warning messages
8320 int __kmp_pause_resource(kmp_pause_status_t level) {
8321   if (level == kmp_not_paused) { // requesting resume
8322     if (__kmp_pause_status == kmp_not_paused) {
8323       // error message about runtime not being paused, so can't resume
8324       return 1;
8325     } else {
8326       KMP_DEBUG_ASSERT(__kmp_pause_status == kmp_soft_paused ||
8327                        __kmp_pause_status == kmp_hard_paused);
8328       __kmp_pause_status = kmp_not_paused;
8329       return 0;
8330     }
8331   } else if (level == kmp_soft_paused) { // requesting soft pause
8332     if (__kmp_pause_status != kmp_not_paused) {
8333       // error message about already being paused
8334       return 1;
8335     } else {
8336       __kmp_soft_pause();
8337       return 0;
8338     }
8339   } else if (level == kmp_hard_paused) { // requesting hard pause
8340     if (__kmp_pause_status != kmp_not_paused) {
8341       // error message about already being paused
8342       return 1;
8343     } else {
8344       __kmp_hard_pause();
8345       return 0;
8346     }
8347   } else {
8348     // error message about invalid level
8349     return 1;
8350   }
8351 }
8352 
8353 #endif // OMP_50_ENABLED
8354