1 /*
2  * kmp_runtime.cpp -- KPTS runtime support library
3  */
4 
5 //===----------------------------------------------------------------------===//
6 //
7 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
8 // See https://llvm.org/LICENSE.txt for license information.
9 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
10 //
11 //===----------------------------------------------------------------------===//
12 
13 #include "kmp.h"
14 #include "kmp_affinity.h"
15 #include "kmp_atomic.h"
16 #include "kmp_environment.h"
17 #include "kmp_error.h"
18 #include "kmp_i18n.h"
19 #include "kmp_io.h"
20 #include "kmp_itt.h"
21 #include "kmp_settings.h"
22 #include "kmp_stats.h"
23 #include "kmp_str.h"
24 #include "kmp_wait_release.h"
25 #include "kmp_wrapper_getpid.h"
26 #include "kmp_dispatch.h"
27 #if KMP_USE_HIER_SCHED
28 #include "kmp_dispatch_hier.h"
29 #endif
30 
31 #if OMPT_SUPPORT
32 #include "ompt-specific.h"
33 #endif
34 
35 /* these are temporary issues to be dealt with */
36 #define KMP_USE_PRCTL 0
37 
38 #if KMP_OS_WINDOWS
39 #include <process.h>
40 #endif
41 
42 #include "tsan_annotations.h"
43 
44 #if defined(KMP_GOMP_COMPAT)
45 char const __kmp_version_alt_comp[] =
46     KMP_VERSION_PREFIX "alternative compiler support: yes";
47 #endif /* defined(KMP_GOMP_COMPAT) */
48 
49 char const __kmp_version_omp_api[] = KMP_VERSION_PREFIX "API version: "
50 #if OMP_50_ENABLED
51                                                         "5.0 (201611)";
52 #elif OMP_45_ENABLED
53                                                         "4.5 (201511)";
54 #elif OMP_40_ENABLED
55                                                         "4.0 (201307)";
56 #else
57                                                         "3.1 (201107)";
58 #endif
59 
60 #ifdef KMP_DEBUG
61 char const __kmp_version_lock[] =
62     KMP_VERSION_PREFIX "lock type: run time selectable";
63 #endif /* KMP_DEBUG */
64 
65 #define KMP_MIN(x, y) ((x) < (y) ? (x) : (y))
66 
67 /* ------------------------------------------------------------------------ */
68 
69 #if KMP_USE_MONITOR
70 kmp_info_t __kmp_monitor;
71 #endif
72 
73 /* Forward declarations */
74 
75 void __kmp_cleanup(void);
76 
77 static void __kmp_initialize_info(kmp_info_t *, kmp_team_t *, int tid,
78                                   int gtid);
79 static void __kmp_initialize_team(kmp_team_t *team, int new_nproc,
80                                   kmp_internal_control_t *new_icvs,
81                                   ident_t *loc);
82 #if OMP_40_ENABLED && KMP_AFFINITY_SUPPORTED
83 static void __kmp_partition_places(kmp_team_t *team,
84                                    int update_master_only = 0);
85 #endif
86 static void __kmp_do_serial_initialize(void);
87 void __kmp_fork_barrier(int gtid, int tid);
88 void __kmp_join_barrier(int gtid);
89 void __kmp_setup_icv_copy(kmp_team_t *team, int new_nproc,
90                           kmp_internal_control_t *new_icvs, ident_t *loc);
91 
92 #ifdef USE_LOAD_BALANCE
93 static int __kmp_load_balance_nproc(kmp_root_t *root, int set_nproc);
94 #endif
95 
96 static int __kmp_expand_threads(int nNeed);
97 #if KMP_OS_WINDOWS
98 static int __kmp_unregister_root_other_thread(int gtid);
99 #endif
100 static void __kmp_unregister_library(void); // called by __kmp_internal_end()
101 static void __kmp_reap_thread(kmp_info_t *thread, int is_root);
102 kmp_info_t *__kmp_thread_pool_insert_pt = NULL;
103 
104 /* Calculate the identifier of the current thread */
105 /* fast (and somewhat portable) way to get unique identifier of executing
106    thread. Returns KMP_GTID_DNE if we haven't been assigned a gtid. */
107 int __kmp_get_global_thread_id() {
108   int i;
109   kmp_info_t **other_threads;
110   size_t stack_data;
111   char *stack_addr;
112   size_t stack_size;
113   char *stack_base;
114 
115   KA_TRACE(
116       1000,
117       ("*** __kmp_get_global_thread_id: entering, nproc=%d  all_nproc=%d\n",
118        __kmp_nth, __kmp_all_nth));
119 
120   /* JPH - to handle the case where __kmpc_end(0) is called immediately prior to
121      a parallel region, made it return KMP_GTID_DNE to force serial_initialize
122      by caller. Had to handle KMP_GTID_DNE at all call-sites, or else guarantee
123      __kmp_init_gtid for this to work. */
124 
125   if (!TCR_4(__kmp_init_gtid))
126     return KMP_GTID_DNE;
127 
128 #ifdef KMP_TDATA_GTID
129   if (TCR_4(__kmp_gtid_mode) >= 3) {
130     KA_TRACE(1000, ("*** __kmp_get_global_thread_id: using TDATA\n"));
131     return __kmp_gtid;
132   }
133 #endif
134   if (TCR_4(__kmp_gtid_mode) >= 2) {
135     KA_TRACE(1000, ("*** __kmp_get_global_thread_id: using keyed TLS\n"));
136     return __kmp_gtid_get_specific();
137   }
138   KA_TRACE(1000, ("*** __kmp_get_global_thread_id: using internal alg.\n"));
139 
140   stack_addr = (char *)&stack_data;
141   other_threads = __kmp_threads;
142 
143   /* ATT: The code below is a source of potential bugs due to unsynchronized
144      access to __kmp_threads array. For example:
145      1. Current thread loads other_threads[i] to thr and checks it, it is
146         non-NULL.
147      2. Current thread is suspended by OS.
148      3. Another thread unregisters and finishes (debug versions of free()
149         may fill memory with something like 0xEF).
150      4. Current thread is resumed.
151      5. Current thread reads junk from *thr.
152      TODO: Fix it.  --ln  */
153 
154   for (i = 0; i < __kmp_threads_capacity; i++) {
155 
156     kmp_info_t *thr = (kmp_info_t *)TCR_SYNC_PTR(other_threads[i]);
157     if (!thr)
158       continue;
159 
160     stack_size = (size_t)TCR_PTR(thr->th.th_info.ds.ds_stacksize);
161     stack_base = (char *)TCR_PTR(thr->th.th_info.ds.ds_stackbase);
162 
163     /* stack grows down -- search through all of the active threads */
164 
165     if (stack_addr <= stack_base) {
166       size_t stack_diff = stack_base - stack_addr;
167 
168       if (stack_diff <= stack_size) {
169         /* The only way we can be closer than the allocated */
170         /* stack size is if we are running on this thread. */
171         KMP_DEBUG_ASSERT(__kmp_gtid_get_specific() == i);
172         return i;
173       }
174     }
175   }
176 
177   /* get specific to try and determine our gtid */
178   KA_TRACE(1000,
179            ("*** __kmp_get_global_thread_id: internal alg. failed to find "
180             "thread, using TLS\n"));
181   i = __kmp_gtid_get_specific();
182 
183   /*fprintf( stderr, "=== %d\n", i );  */ /* GROO */
184 
185   /* if we havn't been assigned a gtid, then return code */
186   if (i < 0)
187     return i;
188 
189   /* dynamically updated stack window for uber threads to avoid get_specific
190      call */
191   if (!TCR_4(other_threads[i]->th.th_info.ds.ds_stackgrow)) {
192     KMP_FATAL(StackOverflow, i);
193   }
194 
195   stack_base = (char *)other_threads[i]->th.th_info.ds.ds_stackbase;
196   if (stack_addr > stack_base) {
197     TCW_PTR(other_threads[i]->th.th_info.ds.ds_stackbase, stack_addr);
198     TCW_PTR(other_threads[i]->th.th_info.ds.ds_stacksize,
199             other_threads[i]->th.th_info.ds.ds_stacksize + stack_addr -
200                 stack_base);
201   } else {
202     TCW_PTR(other_threads[i]->th.th_info.ds.ds_stacksize,
203             stack_base - stack_addr);
204   }
205 
206   /* Reprint stack bounds for ubermaster since they have been refined */
207   if (__kmp_storage_map) {
208     char *stack_end = (char *)other_threads[i]->th.th_info.ds.ds_stackbase;
209     char *stack_beg = stack_end - other_threads[i]->th.th_info.ds.ds_stacksize;
210     __kmp_print_storage_map_gtid(i, stack_beg, stack_end,
211                                  other_threads[i]->th.th_info.ds.ds_stacksize,
212                                  "th_%d stack (refinement)", i);
213   }
214   return i;
215 }
216 
217 int __kmp_get_global_thread_id_reg() {
218   int gtid;
219 
220   if (!__kmp_init_serial) {
221     gtid = KMP_GTID_DNE;
222   } else
223 #ifdef KMP_TDATA_GTID
224       if (TCR_4(__kmp_gtid_mode) >= 3) {
225     KA_TRACE(1000, ("*** __kmp_get_global_thread_id_reg: using TDATA\n"));
226     gtid = __kmp_gtid;
227   } else
228 #endif
229       if (TCR_4(__kmp_gtid_mode) >= 2) {
230     KA_TRACE(1000, ("*** __kmp_get_global_thread_id_reg: using keyed TLS\n"));
231     gtid = __kmp_gtid_get_specific();
232   } else {
233     KA_TRACE(1000,
234              ("*** __kmp_get_global_thread_id_reg: using internal alg.\n"));
235     gtid = __kmp_get_global_thread_id();
236   }
237 
238   /* we must be a new uber master sibling thread */
239   if (gtid == KMP_GTID_DNE) {
240     KA_TRACE(10,
241              ("__kmp_get_global_thread_id_reg: Encountered new root thread. "
242               "Registering a new gtid.\n"));
243     __kmp_acquire_bootstrap_lock(&__kmp_initz_lock);
244     if (!__kmp_init_serial) {
245       __kmp_do_serial_initialize();
246       gtid = __kmp_gtid_get_specific();
247     } else {
248       gtid = __kmp_register_root(FALSE);
249     }
250     __kmp_release_bootstrap_lock(&__kmp_initz_lock);
251     /*__kmp_printf( "+++ %d\n", gtid ); */ /* GROO */
252   }
253 
254   KMP_DEBUG_ASSERT(gtid >= 0);
255 
256   return gtid;
257 }
258 
259 /* caller must hold forkjoin_lock */
260 void __kmp_check_stack_overlap(kmp_info_t *th) {
261   int f;
262   char *stack_beg = NULL;
263   char *stack_end = NULL;
264   int gtid;
265 
266   KA_TRACE(10, ("__kmp_check_stack_overlap: called\n"));
267   if (__kmp_storage_map) {
268     stack_end = (char *)th->th.th_info.ds.ds_stackbase;
269     stack_beg = stack_end - th->th.th_info.ds.ds_stacksize;
270 
271     gtid = __kmp_gtid_from_thread(th);
272 
273     if (gtid == KMP_GTID_MONITOR) {
274       __kmp_print_storage_map_gtid(
275           gtid, stack_beg, stack_end, th->th.th_info.ds.ds_stacksize,
276           "th_%s stack (%s)", "mon",
277           (th->th.th_info.ds.ds_stackgrow) ? "initial" : "actual");
278     } else {
279       __kmp_print_storage_map_gtid(
280           gtid, stack_beg, stack_end, th->th.th_info.ds.ds_stacksize,
281           "th_%d stack (%s)", gtid,
282           (th->th.th_info.ds.ds_stackgrow) ? "initial" : "actual");
283     }
284   }
285 
286   /* No point in checking ubermaster threads since they use refinement and
287    * cannot overlap */
288   gtid = __kmp_gtid_from_thread(th);
289   if (__kmp_env_checks == TRUE && !KMP_UBER_GTID(gtid)) {
290     KA_TRACE(10,
291              ("__kmp_check_stack_overlap: performing extensive checking\n"));
292     if (stack_beg == NULL) {
293       stack_end = (char *)th->th.th_info.ds.ds_stackbase;
294       stack_beg = stack_end - th->th.th_info.ds.ds_stacksize;
295     }
296 
297     for (f = 0; f < __kmp_threads_capacity; f++) {
298       kmp_info_t *f_th = (kmp_info_t *)TCR_SYNC_PTR(__kmp_threads[f]);
299 
300       if (f_th && f_th != th) {
301         char *other_stack_end =
302             (char *)TCR_PTR(f_th->th.th_info.ds.ds_stackbase);
303         char *other_stack_beg =
304             other_stack_end - (size_t)TCR_PTR(f_th->th.th_info.ds.ds_stacksize);
305         if ((stack_beg > other_stack_beg && stack_beg < other_stack_end) ||
306             (stack_end > other_stack_beg && stack_end < other_stack_end)) {
307 
308           /* Print the other stack values before the abort */
309           if (__kmp_storage_map)
310             __kmp_print_storage_map_gtid(
311                 -1, other_stack_beg, other_stack_end,
312                 (size_t)TCR_PTR(f_th->th.th_info.ds.ds_stacksize),
313                 "th_%d stack (overlapped)", __kmp_gtid_from_thread(f_th));
314 
315           __kmp_fatal(KMP_MSG(StackOverlap), KMP_HNT(ChangeStackLimit),
316                       __kmp_msg_null);
317         }
318       }
319     }
320   }
321   KA_TRACE(10, ("__kmp_check_stack_overlap: returning\n"));
322 }
323 
324 /* ------------------------------------------------------------------------ */
325 
326 void __kmp_infinite_loop(void) {
327   static int done = FALSE;
328 
329   while (!done) {
330     KMP_YIELD(TRUE);
331   }
332 }
333 
334 #define MAX_MESSAGE 512
335 
336 void __kmp_print_storage_map_gtid(int gtid, void *p1, void *p2, size_t size,
337                                   char const *format, ...) {
338   char buffer[MAX_MESSAGE];
339   va_list ap;
340 
341   va_start(ap, format);
342   KMP_SNPRINTF(buffer, sizeof(buffer), "OMP storage map: %p %p%8lu %s\n", p1,
343                p2, (unsigned long)size, format);
344   __kmp_acquire_bootstrap_lock(&__kmp_stdio_lock);
345   __kmp_vprintf(kmp_err, buffer, ap);
346 #if KMP_PRINT_DATA_PLACEMENT
347   int node;
348   if (gtid >= 0) {
349     if (p1 <= p2 && (char *)p2 - (char *)p1 == size) {
350       if (__kmp_storage_map_verbose) {
351         node = __kmp_get_host_node(p1);
352         if (node < 0) /* doesn't work, so don't try this next time */
353           __kmp_storage_map_verbose = FALSE;
354         else {
355           char *last;
356           int lastNode;
357           int localProc = __kmp_get_cpu_from_gtid(gtid);
358 
359           const int page_size = KMP_GET_PAGE_SIZE();
360 
361           p1 = (void *)((size_t)p1 & ~((size_t)page_size - 1));
362           p2 = (void *)(((size_t)p2 - 1) & ~((size_t)page_size - 1));
363           if (localProc >= 0)
364             __kmp_printf_no_lock("  GTID %d localNode %d\n", gtid,
365                                  localProc >> 1);
366           else
367             __kmp_printf_no_lock("  GTID %d\n", gtid);
368 #if KMP_USE_PRCTL
369           /* The more elaborate format is disabled for now because of the prctl
370            * hanging bug. */
371           do {
372             last = p1;
373             lastNode = node;
374             /* This loop collates adjacent pages with the same host node. */
375             do {
376               (char *)p1 += page_size;
377             } while (p1 <= p2 && (node = __kmp_get_host_node(p1)) == lastNode);
378             __kmp_printf_no_lock("    %p-%p memNode %d\n", last, (char *)p1 - 1,
379                                  lastNode);
380           } while (p1 <= p2);
381 #else
382           __kmp_printf_no_lock("    %p-%p memNode %d\n", p1,
383                                (char *)p1 + (page_size - 1),
384                                __kmp_get_host_node(p1));
385           if (p1 < p2) {
386             __kmp_printf_no_lock("    %p-%p memNode %d\n", p2,
387                                  (char *)p2 + (page_size - 1),
388                                  __kmp_get_host_node(p2));
389           }
390 #endif
391         }
392       }
393     } else
394       __kmp_printf_no_lock("  %s\n", KMP_I18N_STR(StorageMapWarning));
395   }
396 #endif /* KMP_PRINT_DATA_PLACEMENT */
397   __kmp_release_bootstrap_lock(&__kmp_stdio_lock);
398 }
399 
400 void __kmp_warn(char const *format, ...) {
401   char buffer[MAX_MESSAGE];
402   va_list ap;
403 
404   if (__kmp_generate_warnings == kmp_warnings_off) {
405     return;
406   }
407 
408   va_start(ap, format);
409 
410   KMP_SNPRINTF(buffer, sizeof(buffer), "OMP warning: %s\n", format);
411   __kmp_acquire_bootstrap_lock(&__kmp_stdio_lock);
412   __kmp_vprintf(kmp_err, buffer, ap);
413   __kmp_release_bootstrap_lock(&__kmp_stdio_lock);
414 
415   va_end(ap);
416 }
417 
418 void __kmp_abort_process() {
419   // Later threads may stall here, but that's ok because abort() will kill them.
420   __kmp_acquire_bootstrap_lock(&__kmp_exit_lock);
421 
422   if (__kmp_debug_buf) {
423     __kmp_dump_debug_buffer();
424   }
425 
426   if (KMP_OS_WINDOWS) {
427     // Let other threads know of abnormal termination and prevent deadlock
428     // if abort happened during library initialization or shutdown
429     __kmp_global.g.g_abort = SIGABRT;
430 
431     /* On Windows* OS by default abort() causes pop-up error box, which stalls
432        nightly testing. Unfortunately, we cannot reliably suppress pop-up error
433        boxes. _set_abort_behavior() works well, but this function is not
434        available in VS7 (this is not problem for DLL, but it is a problem for
435        static OpenMP RTL). SetErrorMode (and so, timelimit utility) does not
436        help, at least in some versions of MS C RTL.
437 
438        It seems following sequence is the only way to simulate abort() and
439        avoid pop-up error box. */
440     raise(SIGABRT);
441     _exit(3); // Just in case, if signal ignored, exit anyway.
442   } else {
443     abort();
444   }
445 
446   __kmp_infinite_loop();
447   __kmp_release_bootstrap_lock(&__kmp_exit_lock);
448 
449 } // __kmp_abort_process
450 
451 void __kmp_abort_thread(void) {
452   // TODO: Eliminate g_abort global variable and this function.
453   // In case of abort just call abort(), it will kill all the threads.
454   __kmp_infinite_loop();
455 } // __kmp_abort_thread
456 
457 /* Print out the storage map for the major kmp_info_t thread data structures
458    that are allocated together. */
459 
460 static void __kmp_print_thread_storage_map(kmp_info_t *thr, int gtid) {
461   __kmp_print_storage_map_gtid(gtid, thr, thr + 1, sizeof(kmp_info_t), "th_%d",
462                                gtid);
463 
464   __kmp_print_storage_map_gtid(gtid, &thr->th.th_info, &thr->th.th_team,
465                                sizeof(kmp_desc_t), "th_%d.th_info", gtid);
466 
467   __kmp_print_storage_map_gtid(gtid, &thr->th.th_local, &thr->th.th_pri_head,
468                                sizeof(kmp_local_t), "th_%d.th_local", gtid);
469 
470   __kmp_print_storage_map_gtid(
471       gtid, &thr->th.th_bar[0], &thr->th.th_bar[bs_last_barrier],
472       sizeof(kmp_balign_t) * bs_last_barrier, "th_%d.th_bar", gtid);
473 
474   __kmp_print_storage_map_gtid(gtid, &thr->th.th_bar[bs_plain_barrier],
475                                &thr->th.th_bar[bs_plain_barrier + 1],
476                                sizeof(kmp_balign_t), "th_%d.th_bar[plain]",
477                                gtid);
478 
479   __kmp_print_storage_map_gtid(gtid, &thr->th.th_bar[bs_forkjoin_barrier],
480                                &thr->th.th_bar[bs_forkjoin_barrier + 1],
481                                sizeof(kmp_balign_t), "th_%d.th_bar[forkjoin]",
482                                gtid);
483 
484 #if KMP_FAST_REDUCTION_BARRIER
485   __kmp_print_storage_map_gtid(gtid, &thr->th.th_bar[bs_reduction_barrier],
486                                &thr->th.th_bar[bs_reduction_barrier + 1],
487                                sizeof(kmp_balign_t), "th_%d.th_bar[reduction]",
488                                gtid);
489 #endif // KMP_FAST_REDUCTION_BARRIER
490 }
491 
492 /* Print out the storage map for the major kmp_team_t team data structures
493    that are allocated together. */
494 
495 static void __kmp_print_team_storage_map(const char *header, kmp_team_t *team,
496                                          int team_id, int num_thr) {
497   int num_disp_buff = team->t.t_max_nproc > 1 ? __kmp_dispatch_num_buffers : 2;
498   __kmp_print_storage_map_gtid(-1, team, team + 1, sizeof(kmp_team_t), "%s_%d",
499                                header, team_id);
500 
501   __kmp_print_storage_map_gtid(-1, &team->t.t_bar[0],
502                                &team->t.t_bar[bs_last_barrier],
503                                sizeof(kmp_balign_team_t) * bs_last_barrier,
504                                "%s_%d.t_bar", header, team_id);
505 
506   __kmp_print_storage_map_gtid(-1, &team->t.t_bar[bs_plain_barrier],
507                                &team->t.t_bar[bs_plain_barrier + 1],
508                                sizeof(kmp_balign_team_t), "%s_%d.t_bar[plain]",
509                                header, team_id);
510 
511   __kmp_print_storage_map_gtid(-1, &team->t.t_bar[bs_forkjoin_barrier],
512                                &team->t.t_bar[bs_forkjoin_barrier + 1],
513                                sizeof(kmp_balign_team_t),
514                                "%s_%d.t_bar[forkjoin]", header, team_id);
515 
516 #if KMP_FAST_REDUCTION_BARRIER
517   __kmp_print_storage_map_gtid(-1, &team->t.t_bar[bs_reduction_barrier],
518                                &team->t.t_bar[bs_reduction_barrier + 1],
519                                sizeof(kmp_balign_team_t),
520                                "%s_%d.t_bar[reduction]", header, team_id);
521 #endif // KMP_FAST_REDUCTION_BARRIER
522 
523   __kmp_print_storage_map_gtid(
524       -1, &team->t.t_dispatch[0], &team->t.t_dispatch[num_thr],
525       sizeof(kmp_disp_t) * num_thr, "%s_%d.t_dispatch", header, team_id);
526 
527   __kmp_print_storage_map_gtid(
528       -1, &team->t.t_threads[0], &team->t.t_threads[num_thr],
529       sizeof(kmp_info_t *) * num_thr, "%s_%d.t_threads", header, team_id);
530 
531   __kmp_print_storage_map_gtid(-1, &team->t.t_disp_buffer[0],
532                                &team->t.t_disp_buffer[num_disp_buff],
533                                sizeof(dispatch_shared_info_t) * num_disp_buff,
534                                "%s_%d.t_disp_buffer", header, team_id);
535 }
536 
537 static void __kmp_init_allocator() {
538 #if OMP_50_ENABLED
539   __kmp_init_memkind();
540 #endif
541 }
542 static void __kmp_fini_allocator() {
543 #if OMP_50_ENABLED
544   __kmp_fini_memkind();
545 #endif
546 }
547 
548 /* ------------------------------------------------------------------------ */
549 
550 #if KMP_DYNAMIC_LIB
551 #if KMP_OS_WINDOWS
552 
553 static void __kmp_reset_lock(kmp_bootstrap_lock_t *lck) {
554   // TODO: Change to __kmp_break_bootstrap_lock().
555   __kmp_init_bootstrap_lock(lck); // make the lock released
556 }
557 
558 static void __kmp_reset_locks_on_process_detach(int gtid_req) {
559   int i;
560   int thread_count;
561 
562   // PROCESS_DETACH is expected to be called by a thread that executes
563   // ProcessExit() or FreeLibrary(). OS terminates other threads (except the one
564   // calling ProcessExit or FreeLibrary). So, it might be safe to access the
565   // __kmp_threads[] without taking the forkjoin_lock. However, in fact, some
566   // threads can be still alive here, although being about to be terminated. The
567   // threads in the array with ds_thread==0 are most suspicious. Actually, it
568   // can be not safe to access the __kmp_threads[].
569 
570   // TODO: does it make sense to check __kmp_roots[] ?
571 
572   // Let's check that there are no other alive threads registered with the OMP
573   // lib.
574   while (1) {
575     thread_count = 0;
576     for (i = 0; i < __kmp_threads_capacity; ++i) {
577       if (!__kmp_threads)
578         continue;
579       kmp_info_t *th = __kmp_threads[i];
580       if (th == NULL)
581         continue;
582       int gtid = th->th.th_info.ds.ds_gtid;
583       if (gtid == gtid_req)
584         continue;
585       if (gtid < 0)
586         continue;
587       DWORD exit_val;
588       int alive = __kmp_is_thread_alive(th, &exit_val);
589       if (alive) {
590         ++thread_count;
591       }
592     }
593     if (thread_count == 0)
594       break; // success
595   }
596 
597   // Assume that I'm alone. Now it might be safe to check and reset locks.
598   // __kmp_forkjoin_lock and __kmp_stdio_lock are expected to be reset.
599   __kmp_reset_lock(&__kmp_forkjoin_lock);
600 #ifdef KMP_DEBUG
601   __kmp_reset_lock(&__kmp_stdio_lock);
602 #endif // KMP_DEBUG
603 }
604 
605 BOOL WINAPI DllMain(HINSTANCE hInstDLL, DWORD fdwReason, LPVOID lpReserved) {
606   //__kmp_acquire_bootstrap_lock( &__kmp_initz_lock );
607 
608   switch (fdwReason) {
609 
610   case DLL_PROCESS_ATTACH:
611     KA_TRACE(10, ("DllMain: PROCESS_ATTACH\n"));
612 
613     return TRUE;
614 
615   case DLL_PROCESS_DETACH:
616     KA_TRACE(10, ("DllMain: PROCESS_DETACH T#%d\n", __kmp_gtid_get_specific()));
617 
618     if (lpReserved != NULL) {
619       // lpReserved is used for telling the difference:
620       //   lpReserved == NULL when FreeLibrary() was called,
621       //   lpReserved != NULL when the process terminates.
622       // When FreeLibrary() is called, worker threads remain alive. So they will
623       // release the forkjoin lock by themselves. When the process terminates,
624       // worker threads disappear triggering the problem of unreleased forkjoin
625       // lock as described below.
626 
627       // A worker thread can take the forkjoin lock. The problem comes up if
628       // that worker thread becomes dead before it releases the forkjoin lock.
629       // The forkjoin lock remains taken, while the thread executing
630       // DllMain()->PROCESS_DETACH->__kmp_internal_end_library() below will try
631       // to take the forkjoin lock and will always fail, so that the application
632       // will never finish [normally]. This scenario is possible if
633       // __kmpc_end() has not been executed. It looks like it's not a corner
634       // case, but common cases:
635       // - the main function was compiled by an alternative compiler;
636       // - the main function was compiled by icl but without /Qopenmp
637       //   (application with plugins);
638       // - application terminates by calling C exit(), Fortran CALL EXIT() or
639       //   Fortran STOP.
640       // - alive foreign thread prevented __kmpc_end from doing cleanup.
641       //
642       // This is a hack to work around the problem.
643       // TODO: !!! figure out something better.
644       __kmp_reset_locks_on_process_detach(__kmp_gtid_get_specific());
645     }
646 
647     __kmp_internal_end_library(__kmp_gtid_get_specific());
648 
649     return TRUE;
650 
651   case DLL_THREAD_ATTACH:
652     KA_TRACE(10, ("DllMain: THREAD_ATTACH\n"));
653 
654     /* if we want to register new siblings all the time here call
655      * __kmp_get_gtid(); */
656     return TRUE;
657 
658   case DLL_THREAD_DETACH:
659     KA_TRACE(10, ("DllMain: THREAD_DETACH T#%d\n", __kmp_gtid_get_specific()));
660 
661     __kmp_internal_end_thread(__kmp_gtid_get_specific());
662     return TRUE;
663   }
664 
665   return TRUE;
666 }
667 
668 #endif /* KMP_OS_WINDOWS */
669 #endif /* KMP_DYNAMIC_LIB */
670 
671 /* __kmp_parallel_deo -- Wait until it's our turn. */
672 void __kmp_parallel_deo(int *gtid_ref, int *cid_ref, ident_t *loc_ref) {
673   int gtid = *gtid_ref;
674 #ifdef BUILD_PARALLEL_ORDERED
675   kmp_team_t *team = __kmp_team_from_gtid(gtid);
676 #endif /* BUILD_PARALLEL_ORDERED */
677 
678   if (__kmp_env_consistency_check) {
679     if (__kmp_threads[gtid]->th.th_root->r.r_active)
680 #if KMP_USE_DYNAMIC_LOCK
681       __kmp_push_sync(gtid, ct_ordered_in_parallel, loc_ref, NULL, 0);
682 #else
683       __kmp_push_sync(gtid, ct_ordered_in_parallel, loc_ref, NULL);
684 #endif
685   }
686 #ifdef BUILD_PARALLEL_ORDERED
687   if (!team->t.t_serialized) {
688     KMP_MB();
689     KMP_WAIT(&team->t.t_ordered.dt.t_value, __kmp_tid_from_gtid(gtid), KMP_EQ,
690              NULL);
691     KMP_MB();
692   }
693 #endif /* BUILD_PARALLEL_ORDERED */
694 }
695 
696 /* __kmp_parallel_dxo -- Signal the next task. */
697 void __kmp_parallel_dxo(int *gtid_ref, int *cid_ref, ident_t *loc_ref) {
698   int gtid = *gtid_ref;
699 #ifdef BUILD_PARALLEL_ORDERED
700   int tid = __kmp_tid_from_gtid(gtid);
701   kmp_team_t *team = __kmp_team_from_gtid(gtid);
702 #endif /* BUILD_PARALLEL_ORDERED */
703 
704   if (__kmp_env_consistency_check) {
705     if (__kmp_threads[gtid]->th.th_root->r.r_active)
706       __kmp_pop_sync(gtid, ct_ordered_in_parallel, loc_ref);
707   }
708 #ifdef BUILD_PARALLEL_ORDERED
709   if (!team->t.t_serialized) {
710     KMP_MB(); /* Flush all pending memory write invalidates.  */
711 
712     /* use the tid of the next thread in this team */
713     /* TODO replace with general release procedure */
714     team->t.t_ordered.dt.t_value = ((tid + 1) % team->t.t_nproc);
715 
716     KMP_MB(); /* Flush all pending memory write invalidates.  */
717   }
718 #endif /* BUILD_PARALLEL_ORDERED */
719 }
720 
721 /* ------------------------------------------------------------------------ */
722 /* The BARRIER for a SINGLE process section is always explicit   */
723 
724 int __kmp_enter_single(int gtid, ident_t *id_ref, int push_ws) {
725   int status;
726   kmp_info_t *th;
727   kmp_team_t *team;
728 
729   if (!TCR_4(__kmp_init_parallel))
730     __kmp_parallel_initialize();
731 
732 #if OMP_50_ENABLED
733   __kmp_resume_if_soft_paused();
734 #endif
735 
736   th = __kmp_threads[gtid];
737   team = th->th.th_team;
738   status = 0;
739 
740   th->th.th_ident = id_ref;
741 
742   if (team->t.t_serialized) {
743     status = 1;
744   } else {
745     kmp_int32 old_this = th->th.th_local.this_construct;
746 
747     ++th->th.th_local.this_construct;
748     /* try to set team count to thread count--success means thread got the
749        single block */
750     /* TODO: Should this be acquire or release? */
751     if (team->t.t_construct == old_this) {
752       status = __kmp_atomic_compare_store_acq(&team->t.t_construct, old_this,
753                                               th->th.th_local.this_construct);
754     }
755 #if USE_ITT_BUILD
756     if (__itt_metadata_add_ptr && __kmp_forkjoin_frames_mode == 3 &&
757         KMP_MASTER_GTID(gtid) &&
758 #if OMP_40_ENABLED
759         th->th.th_teams_microtask == NULL &&
760 #endif
761         team->t.t_active_level ==
762             1) { // Only report metadata by master of active team at level 1
763       __kmp_itt_metadata_single(id_ref);
764     }
765 #endif /* USE_ITT_BUILD */
766   }
767 
768   if (__kmp_env_consistency_check) {
769     if (status && push_ws) {
770       __kmp_push_workshare(gtid, ct_psingle, id_ref);
771     } else {
772       __kmp_check_workshare(gtid, ct_psingle, id_ref);
773     }
774   }
775 #if USE_ITT_BUILD
776   if (status) {
777     __kmp_itt_single_start(gtid);
778   }
779 #endif /* USE_ITT_BUILD */
780   return status;
781 }
782 
783 void __kmp_exit_single(int gtid) {
784 #if USE_ITT_BUILD
785   __kmp_itt_single_end(gtid);
786 #endif /* USE_ITT_BUILD */
787   if (__kmp_env_consistency_check)
788     __kmp_pop_workshare(gtid, ct_psingle, NULL);
789 }
790 
791 /* determine if we can go parallel or must use a serialized parallel region and
792  * how many threads we can use
793  * set_nproc is the number of threads requested for the team
794  * returns 0 if we should serialize or only use one thread,
795  * otherwise the number of threads to use
796  * The forkjoin lock is held by the caller. */
797 static int __kmp_reserve_threads(kmp_root_t *root, kmp_team_t *parent_team,
798                                  int master_tid, int set_nthreads
799 #if OMP_40_ENABLED
800                                  ,
801                                  int enter_teams
802 #endif /* OMP_40_ENABLED */
803                                  ) {
804   int capacity;
805   int new_nthreads;
806   KMP_DEBUG_ASSERT(__kmp_init_serial);
807   KMP_DEBUG_ASSERT(root && parent_team);
808   kmp_info_t *this_thr = parent_team->t.t_threads[master_tid];
809 
810   // If dyn-var is set, dynamically adjust the number of desired threads,
811   // according to the method specified by dynamic_mode.
812   new_nthreads = set_nthreads;
813   if (!get__dynamic_2(parent_team, master_tid)) {
814     ;
815   }
816 #ifdef USE_LOAD_BALANCE
817   else if (__kmp_global.g.g_dynamic_mode == dynamic_load_balance) {
818     new_nthreads = __kmp_load_balance_nproc(root, set_nthreads);
819     if (new_nthreads == 1) {
820       KC_TRACE(10, ("__kmp_reserve_threads: T#%d load balance reduced "
821                     "reservation to 1 thread\n",
822                     master_tid));
823       return 1;
824     }
825     if (new_nthreads < set_nthreads) {
826       KC_TRACE(10, ("__kmp_reserve_threads: T#%d load balance reduced "
827                     "reservation to %d threads\n",
828                     master_tid, new_nthreads));
829     }
830   }
831 #endif /* USE_LOAD_BALANCE */
832   else if (__kmp_global.g.g_dynamic_mode == dynamic_thread_limit) {
833     new_nthreads = __kmp_avail_proc - __kmp_nth +
834                    (root->r.r_active ? 1 : root->r.r_hot_team->t.t_nproc);
835     if (new_nthreads <= 1) {
836       KC_TRACE(10, ("__kmp_reserve_threads: T#%d thread limit reduced "
837                     "reservation to 1 thread\n",
838                     master_tid));
839       return 1;
840     }
841     if (new_nthreads < set_nthreads) {
842       KC_TRACE(10, ("__kmp_reserve_threads: T#%d thread limit reduced "
843                     "reservation to %d threads\n",
844                     master_tid, new_nthreads));
845     } else {
846       new_nthreads = set_nthreads;
847     }
848   } else if (__kmp_global.g.g_dynamic_mode == dynamic_random) {
849     if (set_nthreads > 2) {
850       new_nthreads = __kmp_get_random(parent_team->t.t_threads[master_tid]);
851       new_nthreads = (new_nthreads % set_nthreads) + 1;
852       if (new_nthreads == 1) {
853         KC_TRACE(10, ("__kmp_reserve_threads: T#%d dynamic random reduced "
854                       "reservation to 1 thread\n",
855                       master_tid));
856         return 1;
857       }
858       if (new_nthreads < set_nthreads) {
859         KC_TRACE(10, ("__kmp_reserve_threads: T#%d dynamic random reduced "
860                       "reservation to %d threads\n",
861                       master_tid, new_nthreads));
862       }
863     }
864   } else {
865     KMP_ASSERT(0);
866   }
867 
868   // Respect KMP_ALL_THREADS/KMP_DEVICE_THREAD_LIMIT.
869   if (__kmp_nth + new_nthreads -
870           (root->r.r_active ? 1 : root->r.r_hot_team->t.t_nproc) >
871       __kmp_max_nth) {
872     int tl_nthreads = __kmp_max_nth - __kmp_nth +
873                       (root->r.r_active ? 1 : root->r.r_hot_team->t.t_nproc);
874     if (tl_nthreads <= 0) {
875       tl_nthreads = 1;
876     }
877 
878     // If dyn-var is false, emit a 1-time warning.
879     if (!get__dynamic_2(parent_team, master_tid) && (!__kmp_reserve_warn)) {
880       __kmp_reserve_warn = 1;
881       __kmp_msg(kmp_ms_warning,
882                 KMP_MSG(CantFormThrTeam, set_nthreads, tl_nthreads),
883                 KMP_HNT(Unset_ALL_THREADS), __kmp_msg_null);
884     }
885     if (tl_nthreads == 1) {
886       KC_TRACE(10, ("__kmp_reserve_threads: T#%d KMP_DEVICE_THREAD_LIMIT "
887                     "reduced reservation to 1 thread\n",
888                     master_tid));
889       return 1;
890     }
891     KC_TRACE(10, ("__kmp_reserve_threads: T#%d KMP_DEVICE_THREAD_LIMIT reduced "
892                   "reservation to %d threads\n",
893                   master_tid, tl_nthreads));
894     new_nthreads = tl_nthreads;
895   }
896 
897   // Respect OMP_THREAD_LIMIT
898   int cg_nthreads = this_thr->th.th_cg_roots->cg_nthreads;
899   int max_cg_threads = this_thr->th.th_cg_roots->cg_thread_limit;
900   if (cg_nthreads + new_nthreads -
901           (root->r.r_active ? 1 : root->r.r_hot_team->t.t_nproc) >
902       max_cg_threads) {
903     int tl_nthreads = max_cg_threads - cg_nthreads +
904                       (root->r.r_active ? 1 : root->r.r_hot_team->t.t_nproc);
905     if (tl_nthreads <= 0) {
906       tl_nthreads = 1;
907     }
908 
909     // If dyn-var is false, emit a 1-time warning.
910     if (!get__dynamic_2(parent_team, master_tid) && (!__kmp_reserve_warn)) {
911       __kmp_reserve_warn = 1;
912       __kmp_msg(kmp_ms_warning,
913                 KMP_MSG(CantFormThrTeam, set_nthreads, tl_nthreads),
914                 KMP_HNT(Unset_ALL_THREADS), __kmp_msg_null);
915     }
916     if (tl_nthreads == 1) {
917       KC_TRACE(10, ("__kmp_reserve_threads: T#%d OMP_THREAD_LIMIT "
918                     "reduced reservation to 1 thread\n",
919                     master_tid));
920       return 1;
921     }
922     KC_TRACE(10, ("__kmp_reserve_threads: T#%d OMP_THREAD_LIMIT reduced "
923                   "reservation to %d threads\n",
924                   master_tid, tl_nthreads));
925     new_nthreads = tl_nthreads;
926   }
927 
928   // Check if the threads array is large enough, or needs expanding.
929   // See comment in __kmp_register_root() about the adjustment if
930   // __kmp_threads[0] == NULL.
931   capacity = __kmp_threads_capacity;
932   if (TCR_PTR(__kmp_threads[0]) == NULL) {
933     --capacity;
934   }
935   if (__kmp_nth + new_nthreads -
936           (root->r.r_active ? 1 : root->r.r_hot_team->t.t_nproc) >
937       capacity) {
938     // Expand the threads array.
939     int slotsRequired = __kmp_nth + new_nthreads -
940                         (root->r.r_active ? 1 : root->r.r_hot_team->t.t_nproc) -
941                         capacity;
942     int slotsAdded = __kmp_expand_threads(slotsRequired);
943     if (slotsAdded < slotsRequired) {
944       // The threads array was not expanded enough.
945       new_nthreads -= (slotsRequired - slotsAdded);
946       KMP_ASSERT(new_nthreads >= 1);
947 
948       // If dyn-var is false, emit a 1-time warning.
949       if (!get__dynamic_2(parent_team, master_tid) && (!__kmp_reserve_warn)) {
950         __kmp_reserve_warn = 1;
951         if (__kmp_tp_cached) {
952           __kmp_msg(kmp_ms_warning,
953                     KMP_MSG(CantFormThrTeam, set_nthreads, new_nthreads),
954                     KMP_HNT(Set_ALL_THREADPRIVATE, __kmp_tp_capacity),
955                     KMP_HNT(PossibleSystemLimitOnThreads), __kmp_msg_null);
956         } else {
957           __kmp_msg(kmp_ms_warning,
958                     KMP_MSG(CantFormThrTeam, set_nthreads, new_nthreads),
959                     KMP_HNT(SystemLimitOnThreads), __kmp_msg_null);
960         }
961       }
962     }
963   }
964 
965 #ifdef KMP_DEBUG
966   if (new_nthreads == 1) {
967     KC_TRACE(10,
968              ("__kmp_reserve_threads: T#%d serializing team after reclaiming "
969               "dead roots and rechecking; requested %d threads\n",
970               __kmp_get_gtid(), set_nthreads));
971   } else {
972     KC_TRACE(10, ("__kmp_reserve_threads: T#%d allocating %d threads; requested"
973                   " %d threads\n",
974                   __kmp_get_gtid(), new_nthreads, set_nthreads));
975   }
976 #endif // KMP_DEBUG
977   return new_nthreads;
978 }
979 
980 /* Allocate threads from the thread pool and assign them to the new team. We are
981    assured that there are enough threads available, because we checked on that
982    earlier within critical section forkjoin */
983 static void __kmp_fork_team_threads(kmp_root_t *root, kmp_team_t *team,
984                                     kmp_info_t *master_th, int master_gtid) {
985   int i;
986   int use_hot_team;
987 
988   KA_TRACE(10, ("__kmp_fork_team_threads: new_nprocs = %d\n", team->t.t_nproc));
989   KMP_DEBUG_ASSERT(master_gtid == __kmp_get_gtid());
990   KMP_MB();
991 
992   /* first, let's setup the master thread */
993   master_th->th.th_info.ds.ds_tid = 0;
994   master_th->th.th_team = team;
995   master_th->th.th_team_nproc = team->t.t_nproc;
996   master_th->th.th_team_master = master_th;
997   master_th->th.th_team_serialized = FALSE;
998   master_th->th.th_dispatch = &team->t.t_dispatch[0];
999 
1000 /* make sure we are not the optimized hot team */
1001 #if KMP_NESTED_HOT_TEAMS
1002   use_hot_team = 0;
1003   kmp_hot_team_ptr_t *hot_teams = master_th->th.th_hot_teams;
1004   if (hot_teams) { // hot teams array is not allocated if
1005     // KMP_HOT_TEAMS_MAX_LEVEL=0
1006     int level = team->t.t_active_level - 1; // index in array of hot teams
1007     if (master_th->th.th_teams_microtask) { // are we inside the teams?
1008       if (master_th->th.th_teams_size.nteams > 1) {
1009         ++level; // level was not increased in teams construct for
1010         // team_of_masters
1011       }
1012       if (team->t.t_pkfn != (microtask_t)__kmp_teams_master &&
1013           master_th->th.th_teams_level == team->t.t_level) {
1014         ++level; // level was not increased in teams construct for
1015         // team_of_workers before the parallel
1016       } // team->t.t_level will be increased inside parallel
1017     }
1018     if (level < __kmp_hot_teams_max_level) {
1019       if (hot_teams[level].hot_team) {
1020         // hot team has already been allocated for given level
1021         KMP_DEBUG_ASSERT(hot_teams[level].hot_team == team);
1022         use_hot_team = 1; // the team is ready to use
1023       } else {
1024         use_hot_team = 0; // AC: threads are not allocated yet
1025         hot_teams[level].hot_team = team; // remember new hot team
1026         hot_teams[level].hot_team_nth = team->t.t_nproc;
1027       }
1028     } else {
1029       use_hot_team = 0;
1030     }
1031   }
1032 #else
1033   use_hot_team = team == root->r.r_hot_team;
1034 #endif
1035   if (!use_hot_team) {
1036 
1037     /* install the master thread */
1038     team->t.t_threads[0] = master_th;
1039     __kmp_initialize_info(master_th, team, 0, master_gtid);
1040 
1041     /* now, install the worker threads */
1042     for (i = 1; i < team->t.t_nproc; i++) {
1043 
1044       /* fork or reallocate a new thread and install it in team */
1045       kmp_info_t *thr = __kmp_allocate_thread(root, team, i);
1046       team->t.t_threads[i] = thr;
1047       KMP_DEBUG_ASSERT(thr);
1048       KMP_DEBUG_ASSERT(thr->th.th_team == team);
1049       /* align team and thread arrived states */
1050       KA_TRACE(20, ("__kmp_fork_team_threads: T#%d(%d:%d) init arrived "
1051                     "T#%d(%d:%d) join =%llu, plain=%llu\n",
1052                     __kmp_gtid_from_tid(0, team), team->t.t_id, 0,
1053                     __kmp_gtid_from_tid(i, team), team->t.t_id, i,
1054                     team->t.t_bar[bs_forkjoin_barrier].b_arrived,
1055                     team->t.t_bar[bs_plain_barrier].b_arrived));
1056 #if OMP_40_ENABLED
1057       thr->th.th_teams_microtask = master_th->th.th_teams_microtask;
1058       thr->th.th_teams_level = master_th->th.th_teams_level;
1059       thr->th.th_teams_size = master_th->th.th_teams_size;
1060 #endif
1061       { // Initialize threads' barrier data.
1062         int b;
1063         kmp_balign_t *balign = team->t.t_threads[i]->th.th_bar;
1064         for (b = 0; b < bs_last_barrier; ++b) {
1065           balign[b].bb.b_arrived = team->t.t_bar[b].b_arrived;
1066           KMP_DEBUG_ASSERT(balign[b].bb.wait_flag != KMP_BARRIER_PARENT_FLAG);
1067 #if USE_DEBUGGER
1068           balign[b].bb.b_worker_arrived = team->t.t_bar[b].b_team_arrived;
1069 #endif
1070         }
1071       }
1072     }
1073 
1074 #if OMP_40_ENABLED && KMP_AFFINITY_SUPPORTED
1075     __kmp_partition_places(team);
1076 #endif
1077   }
1078 
1079 #if OMP_50_ENABLED
1080   if (__kmp_display_affinity && team->t.t_display_affinity != 1) {
1081     for (i = 0; i < team->t.t_nproc; i++) {
1082       kmp_info_t *thr = team->t.t_threads[i];
1083       if (thr->th.th_prev_num_threads != team->t.t_nproc ||
1084           thr->th.th_prev_level != team->t.t_level) {
1085         team->t.t_display_affinity = 1;
1086         break;
1087       }
1088     }
1089   }
1090 #endif
1091 
1092   KMP_MB();
1093 }
1094 
1095 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
1096 // Propagate any changes to the floating point control registers out to the team
1097 // We try to avoid unnecessary writes to the relevant cache line in the team
1098 // structure, so we don't make changes unless they are needed.
1099 inline static void propagateFPControl(kmp_team_t *team) {
1100   if (__kmp_inherit_fp_control) {
1101     kmp_int16 x87_fpu_control_word;
1102     kmp_uint32 mxcsr;
1103 
1104     // Get master values of FPU control flags (both X87 and vector)
1105     __kmp_store_x87_fpu_control_word(&x87_fpu_control_word);
1106     __kmp_store_mxcsr(&mxcsr);
1107     mxcsr &= KMP_X86_MXCSR_MASK;
1108 
1109     // There is no point looking at t_fp_control_saved here.
1110     // If it is TRUE, we still have to update the values if they are different
1111     // from those we now have. If it is FALSE we didn't save anything yet, but
1112     // our objective is the same. We have to ensure that the values in the team
1113     // are the same as those we have.
1114     // So, this code achieves what we need whether or not t_fp_control_saved is
1115     // true. By checking whether the value needs updating we avoid unnecessary
1116     // writes that would put the cache-line into a written state, causing all
1117     // threads in the team to have to read it again.
1118     KMP_CHECK_UPDATE(team->t.t_x87_fpu_control_word, x87_fpu_control_word);
1119     KMP_CHECK_UPDATE(team->t.t_mxcsr, mxcsr);
1120     // Although we don't use this value, other code in the runtime wants to know
1121     // whether it should restore them. So we must ensure it is correct.
1122     KMP_CHECK_UPDATE(team->t.t_fp_control_saved, TRUE);
1123   } else {
1124     // Similarly here. Don't write to this cache-line in the team structure
1125     // unless we have to.
1126     KMP_CHECK_UPDATE(team->t.t_fp_control_saved, FALSE);
1127   }
1128 }
1129 
1130 // Do the opposite, setting the hardware registers to the updated values from
1131 // the team.
1132 inline static void updateHWFPControl(kmp_team_t *team) {
1133   if (__kmp_inherit_fp_control && team->t.t_fp_control_saved) {
1134     // Only reset the fp control regs if they have been changed in the team.
1135     // the parallel region that we are exiting.
1136     kmp_int16 x87_fpu_control_word;
1137     kmp_uint32 mxcsr;
1138     __kmp_store_x87_fpu_control_word(&x87_fpu_control_word);
1139     __kmp_store_mxcsr(&mxcsr);
1140     mxcsr &= KMP_X86_MXCSR_MASK;
1141 
1142     if (team->t.t_x87_fpu_control_word != x87_fpu_control_word) {
1143       __kmp_clear_x87_fpu_status_word();
1144       __kmp_load_x87_fpu_control_word(&team->t.t_x87_fpu_control_word);
1145     }
1146 
1147     if (team->t.t_mxcsr != mxcsr) {
1148       __kmp_load_mxcsr(&team->t.t_mxcsr);
1149     }
1150   }
1151 }
1152 #else
1153 #define propagateFPControl(x) ((void)0)
1154 #define updateHWFPControl(x) ((void)0)
1155 #endif /* KMP_ARCH_X86 || KMP_ARCH_X86_64 */
1156 
1157 static void __kmp_alloc_argv_entries(int argc, kmp_team_t *team,
1158                                      int realloc); // forward declaration
1159 
1160 /* Run a parallel region that has been serialized, so runs only in a team of the
1161    single master thread. */
1162 void __kmp_serialized_parallel(ident_t *loc, kmp_int32 global_tid) {
1163   kmp_info_t *this_thr;
1164   kmp_team_t *serial_team;
1165 
1166   KC_TRACE(10, ("__kmpc_serialized_parallel: called by T#%d\n", global_tid));
1167 
1168   /* Skip all this code for autopar serialized loops since it results in
1169      unacceptable overhead */
1170   if (loc != NULL && (loc->flags & KMP_IDENT_AUTOPAR))
1171     return;
1172 
1173   if (!TCR_4(__kmp_init_parallel))
1174     __kmp_parallel_initialize();
1175 
1176 #if OMP_50_ENABLED
1177   __kmp_resume_if_soft_paused();
1178 #endif
1179 
1180   this_thr = __kmp_threads[global_tid];
1181   serial_team = this_thr->th.th_serial_team;
1182 
1183   /* utilize the serialized team held by this thread */
1184   KMP_DEBUG_ASSERT(serial_team);
1185   KMP_MB();
1186 
1187   if (__kmp_tasking_mode != tskm_immediate_exec) {
1188     KMP_DEBUG_ASSERT(
1189         this_thr->th.th_task_team ==
1190         this_thr->th.th_team->t.t_task_team[this_thr->th.th_task_state]);
1191     KMP_DEBUG_ASSERT(serial_team->t.t_task_team[this_thr->th.th_task_state] ==
1192                      NULL);
1193     KA_TRACE(20, ("__kmpc_serialized_parallel: T#%d pushing task_team %p / "
1194                   "team %p, new task_team = NULL\n",
1195                   global_tid, this_thr->th.th_task_team, this_thr->th.th_team));
1196     this_thr->th.th_task_team = NULL;
1197   }
1198 
1199 #if OMP_40_ENABLED
1200   kmp_proc_bind_t proc_bind = this_thr->th.th_set_proc_bind;
1201   if (this_thr->th.th_current_task->td_icvs.proc_bind == proc_bind_false) {
1202     proc_bind = proc_bind_false;
1203   } else if (proc_bind == proc_bind_default) {
1204     // No proc_bind clause was specified, so use the current value
1205     // of proc-bind-var for this parallel region.
1206     proc_bind = this_thr->th.th_current_task->td_icvs.proc_bind;
1207   }
1208   // Reset for next parallel region
1209   this_thr->th.th_set_proc_bind = proc_bind_default;
1210 #endif /* OMP_40_ENABLED */
1211 
1212 #if OMPT_SUPPORT
1213   ompt_data_t ompt_parallel_data = ompt_data_none;
1214   ompt_data_t *implicit_task_data;
1215   void *codeptr = OMPT_LOAD_RETURN_ADDRESS(global_tid);
1216   if (ompt_enabled.enabled &&
1217       this_thr->th.ompt_thread_info.state != ompt_state_overhead) {
1218 
1219     ompt_task_info_t *parent_task_info;
1220     parent_task_info = OMPT_CUR_TASK_INFO(this_thr);
1221 
1222     parent_task_info->frame.enter_frame.ptr = OMPT_GET_FRAME_ADDRESS(0);
1223     if (ompt_enabled.ompt_callback_parallel_begin) {
1224       int team_size = 1;
1225 
1226       ompt_callbacks.ompt_callback(ompt_callback_parallel_begin)(
1227           &(parent_task_info->task_data), &(parent_task_info->frame),
1228           &ompt_parallel_data, team_size, ompt_parallel_invoker_program,
1229           codeptr);
1230     }
1231   }
1232 #endif // OMPT_SUPPORT
1233 
1234   if (this_thr->th.th_team != serial_team) {
1235     // Nested level will be an index in the nested nthreads array
1236     int level = this_thr->th.th_team->t.t_level;
1237 
1238     if (serial_team->t.t_serialized) {
1239       /* this serial team was already used
1240          TODO increase performance by making this locks more specific */
1241       kmp_team_t *new_team;
1242 
1243       __kmp_acquire_bootstrap_lock(&__kmp_forkjoin_lock);
1244 
1245       new_team = __kmp_allocate_team(this_thr->th.th_root, 1, 1,
1246 #if OMPT_SUPPORT
1247                                      ompt_parallel_data,
1248 #endif
1249 #if OMP_40_ENABLED
1250                                      proc_bind,
1251 #endif
1252                                      &this_thr->th.th_current_task->td_icvs,
1253                                      0 USE_NESTED_HOT_ARG(NULL));
1254       __kmp_release_bootstrap_lock(&__kmp_forkjoin_lock);
1255       KMP_ASSERT(new_team);
1256 
1257       /* setup new serialized team and install it */
1258       new_team->t.t_threads[0] = this_thr;
1259       new_team->t.t_parent = this_thr->th.th_team;
1260       serial_team = new_team;
1261       this_thr->th.th_serial_team = serial_team;
1262 
1263       KF_TRACE(
1264           10,
1265           ("__kmpc_serialized_parallel: T#%d allocated new serial team %p\n",
1266            global_tid, serial_team));
1267 
1268       /* TODO the above breaks the requirement that if we run out of resources,
1269          then we can still guarantee that serialized teams are ok, since we may
1270          need to allocate a new one */
1271     } else {
1272       KF_TRACE(
1273           10,
1274           ("__kmpc_serialized_parallel: T#%d reusing cached serial team %p\n",
1275            global_tid, serial_team));
1276     }
1277 
1278     /* we have to initialize this serial team */
1279     KMP_DEBUG_ASSERT(serial_team->t.t_threads);
1280     KMP_DEBUG_ASSERT(serial_team->t.t_threads[0] == this_thr);
1281     KMP_DEBUG_ASSERT(this_thr->th.th_team != serial_team);
1282     serial_team->t.t_ident = loc;
1283     serial_team->t.t_serialized = 1;
1284     serial_team->t.t_nproc = 1;
1285     serial_team->t.t_parent = this_thr->th.th_team;
1286     serial_team->t.t_sched.sched = this_thr->th.th_team->t.t_sched.sched;
1287     this_thr->th.th_team = serial_team;
1288     serial_team->t.t_master_tid = this_thr->th.th_info.ds.ds_tid;
1289 
1290     KF_TRACE(10, ("__kmpc_serialized_parallel: T#d curtask=%p\n", global_tid,
1291                   this_thr->th.th_current_task));
1292     KMP_ASSERT(this_thr->th.th_current_task->td_flags.executing == 1);
1293     this_thr->th.th_current_task->td_flags.executing = 0;
1294 
1295     __kmp_push_current_task_to_thread(this_thr, serial_team, 0);
1296 
1297     /* TODO: GEH: do ICVs work for nested serialized teams? Don't we need an
1298        implicit task for each serialized task represented by
1299        team->t.t_serialized? */
1300     copy_icvs(&this_thr->th.th_current_task->td_icvs,
1301               &this_thr->th.th_current_task->td_parent->td_icvs);
1302 
1303     // Thread value exists in the nested nthreads array for the next nested
1304     // level
1305     if (__kmp_nested_nth.used && (level + 1 < __kmp_nested_nth.used)) {
1306       this_thr->th.th_current_task->td_icvs.nproc =
1307           __kmp_nested_nth.nth[level + 1];
1308     }
1309 
1310 #if OMP_40_ENABLED
1311     if (__kmp_nested_proc_bind.used &&
1312         (level + 1 < __kmp_nested_proc_bind.used)) {
1313       this_thr->th.th_current_task->td_icvs.proc_bind =
1314           __kmp_nested_proc_bind.bind_types[level + 1];
1315     }
1316 #endif /* OMP_40_ENABLED */
1317 
1318 #if USE_DEBUGGER
1319     serial_team->t.t_pkfn = (microtask_t)(~0); // For the debugger.
1320 #endif
1321     this_thr->th.th_info.ds.ds_tid = 0;
1322 
1323     /* set thread cache values */
1324     this_thr->th.th_team_nproc = 1;
1325     this_thr->th.th_team_master = this_thr;
1326     this_thr->th.th_team_serialized = 1;
1327 
1328     serial_team->t.t_level = serial_team->t.t_parent->t.t_level + 1;
1329     serial_team->t.t_active_level = serial_team->t.t_parent->t.t_active_level;
1330 #if OMP_50_ENABLED
1331     serial_team->t.t_def_allocator = this_thr->th.th_def_allocator; // save
1332 #endif
1333 
1334     propagateFPControl(serial_team);
1335 
1336     /* check if we need to allocate dispatch buffers stack */
1337     KMP_DEBUG_ASSERT(serial_team->t.t_dispatch);
1338     if (!serial_team->t.t_dispatch->th_disp_buffer) {
1339       serial_team->t.t_dispatch->th_disp_buffer =
1340           (dispatch_private_info_t *)__kmp_allocate(
1341               sizeof(dispatch_private_info_t));
1342     }
1343     this_thr->th.th_dispatch = serial_team->t.t_dispatch;
1344 
1345     KMP_MB();
1346 
1347   } else {
1348     /* this serialized team is already being used,
1349      * that's fine, just add another nested level */
1350     KMP_DEBUG_ASSERT(this_thr->th.th_team == serial_team);
1351     KMP_DEBUG_ASSERT(serial_team->t.t_threads);
1352     KMP_DEBUG_ASSERT(serial_team->t.t_threads[0] == this_thr);
1353     ++serial_team->t.t_serialized;
1354     this_thr->th.th_team_serialized = serial_team->t.t_serialized;
1355 
1356     // Nested level will be an index in the nested nthreads array
1357     int level = this_thr->th.th_team->t.t_level;
1358     // Thread value exists in the nested nthreads array for the next nested
1359     // level
1360     if (__kmp_nested_nth.used && (level + 1 < __kmp_nested_nth.used)) {
1361       this_thr->th.th_current_task->td_icvs.nproc =
1362           __kmp_nested_nth.nth[level + 1];
1363     }
1364     serial_team->t.t_level++;
1365     KF_TRACE(10, ("__kmpc_serialized_parallel: T#%d increasing nesting level "
1366                   "of serial team %p to %d\n",
1367                   global_tid, serial_team, serial_team->t.t_level));
1368 
1369     /* allocate/push dispatch buffers stack */
1370     KMP_DEBUG_ASSERT(serial_team->t.t_dispatch);
1371     {
1372       dispatch_private_info_t *disp_buffer =
1373           (dispatch_private_info_t *)__kmp_allocate(
1374               sizeof(dispatch_private_info_t));
1375       disp_buffer->next = serial_team->t.t_dispatch->th_disp_buffer;
1376       serial_team->t.t_dispatch->th_disp_buffer = disp_buffer;
1377     }
1378     this_thr->th.th_dispatch = serial_team->t.t_dispatch;
1379 
1380     KMP_MB();
1381   }
1382 #if OMP_40_ENABLED
1383   KMP_CHECK_UPDATE(serial_team->t.t_cancel_request, cancel_noreq);
1384 #endif
1385 
1386 #if OMP_50_ENABLED
1387   // Perform the display affinity functionality for
1388   // serialized parallel regions
1389   if (__kmp_display_affinity) {
1390     if (this_thr->th.th_prev_level != serial_team->t.t_level ||
1391         this_thr->th.th_prev_num_threads != 1) {
1392       // NULL means use the affinity-format-var ICV
1393       __kmp_aux_display_affinity(global_tid, NULL);
1394       this_thr->th.th_prev_level = serial_team->t.t_level;
1395       this_thr->th.th_prev_num_threads = 1;
1396     }
1397   }
1398 #endif
1399 
1400   if (__kmp_env_consistency_check)
1401     __kmp_push_parallel(global_tid, NULL);
1402 #if OMPT_SUPPORT
1403   serial_team->t.ompt_team_info.master_return_address = codeptr;
1404   if (ompt_enabled.enabled &&
1405       this_thr->th.ompt_thread_info.state != ompt_state_overhead) {
1406     OMPT_CUR_TASK_INFO(this_thr)->frame.exit_frame.ptr = OMPT_GET_FRAME_ADDRESS(0);
1407 
1408     ompt_lw_taskteam_t lw_taskteam;
1409     __ompt_lw_taskteam_init(&lw_taskteam, this_thr, global_tid,
1410                             &ompt_parallel_data, codeptr);
1411 
1412     __ompt_lw_taskteam_link(&lw_taskteam, this_thr, 1);
1413     // don't use lw_taskteam after linking. content was swaped
1414 
1415     /* OMPT implicit task begin */
1416     implicit_task_data = OMPT_CUR_TASK_DATA(this_thr);
1417     if (ompt_enabled.ompt_callback_implicit_task) {
1418       ompt_callbacks.ompt_callback(ompt_callback_implicit_task)(
1419           ompt_scope_begin, OMPT_CUR_TEAM_DATA(this_thr),
1420           OMPT_CUR_TASK_DATA(this_thr), 1, __kmp_tid_from_gtid(global_tid), ompt_task_implicit); // TODO: Can this be ompt_task_initial?
1421       OMPT_CUR_TASK_INFO(this_thr)
1422           ->thread_num = __kmp_tid_from_gtid(global_tid);
1423     }
1424 
1425     /* OMPT state */
1426     this_thr->th.ompt_thread_info.state = ompt_state_work_parallel;
1427     OMPT_CUR_TASK_INFO(this_thr)->frame.exit_frame.ptr = OMPT_GET_FRAME_ADDRESS(0);
1428   }
1429 #endif
1430 }
1431 
1432 /* most of the work for a fork */
1433 /* return true if we really went parallel, false if serialized */
1434 int __kmp_fork_call(ident_t *loc, int gtid,
1435                     enum fork_context_e call_context, // Intel, GNU, ...
1436                     kmp_int32 argc, microtask_t microtask, launch_t invoker,
1437 /* TODO: revert workaround for Intel(R) 64 tracker #96 */
1438 #if (KMP_ARCH_X86_64 || KMP_ARCH_ARM || KMP_ARCH_AARCH64) && KMP_OS_LINUX
1439                     va_list *ap
1440 #else
1441                     va_list ap
1442 #endif
1443                     ) {
1444   void **argv;
1445   int i;
1446   int master_tid;
1447   int master_this_cons;
1448   kmp_team_t *team;
1449   kmp_team_t *parent_team;
1450   kmp_info_t *master_th;
1451   kmp_root_t *root;
1452   int nthreads;
1453   int master_active;
1454   int master_set_numthreads;
1455   int level;
1456 #if OMP_40_ENABLED
1457   int active_level;
1458   int teams_level;
1459 #endif
1460 #if KMP_NESTED_HOT_TEAMS
1461   kmp_hot_team_ptr_t **p_hot_teams;
1462 #endif
1463   { // KMP_TIME_BLOCK
1464     KMP_TIME_DEVELOPER_PARTITIONED_BLOCK(KMP_fork_call);
1465     KMP_COUNT_VALUE(OMP_PARALLEL_args, argc);
1466 
1467     KA_TRACE(20, ("__kmp_fork_call: enter T#%d\n", gtid));
1468     if (__kmp_stkpadding > 0 && __kmp_root[gtid] != NULL) {
1469       /* Some systems prefer the stack for the root thread(s) to start with */
1470       /* some gap from the parent stack to prevent false sharing. */
1471       void *dummy = KMP_ALLOCA(__kmp_stkpadding);
1472       /* These 2 lines below are so this does not get optimized out */
1473       if (__kmp_stkpadding > KMP_MAX_STKPADDING)
1474         __kmp_stkpadding += (short)((kmp_int64)dummy);
1475     }
1476 
1477     /* initialize if needed */
1478     KMP_DEBUG_ASSERT(
1479         __kmp_init_serial); // AC: potentially unsafe, not in sync with shutdown
1480     if (!TCR_4(__kmp_init_parallel))
1481       __kmp_parallel_initialize();
1482 
1483 #if OMP_50_ENABLED
1484     __kmp_resume_if_soft_paused();
1485 #endif
1486 
1487     /* setup current data */
1488     master_th = __kmp_threads[gtid]; // AC: potentially unsafe, not in sync with
1489     // shutdown
1490     parent_team = master_th->th.th_team;
1491     master_tid = master_th->th.th_info.ds.ds_tid;
1492     master_this_cons = master_th->th.th_local.this_construct;
1493     root = master_th->th.th_root;
1494     master_active = root->r.r_active;
1495     master_set_numthreads = master_th->th.th_set_nproc;
1496 
1497 #if OMPT_SUPPORT
1498     ompt_data_t ompt_parallel_data = ompt_data_none;
1499     ompt_data_t *parent_task_data;
1500     ompt_frame_t *ompt_frame;
1501     ompt_data_t *implicit_task_data;
1502     void *return_address = NULL;
1503 
1504     if (ompt_enabled.enabled) {
1505       __ompt_get_task_info_internal(0, NULL, &parent_task_data, &ompt_frame,
1506                                     NULL, NULL);
1507       return_address = OMPT_LOAD_RETURN_ADDRESS(gtid);
1508     }
1509 #endif
1510 
1511     // Nested level will be an index in the nested nthreads array
1512     level = parent_team->t.t_level;
1513     // used to launch non-serial teams even if nested is not allowed
1514     active_level = parent_team->t.t_active_level;
1515 #if OMP_40_ENABLED
1516     // needed to check nesting inside the teams
1517     teams_level = master_th->th.th_teams_level;
1518 #endif
1519 #if KMP_NESTED_HOT_TEAMS
1520     p_hot_teams = &master_th->th.th_hot_teams;
1521     if (*p_hot_teams == NULL && __kmp_hot_teams_max_level > 0) {
1522       *p_hot_teams = (kmp_hot_team_ptr_t *)__kmp_allocate(
1523           sizeof(kmp_hot_team_ptr_t) * __kmp_hot_teams_max_level);
1524       (*p_hot_teams)[0].hot_team = root->r.r_hot_team;
1525       // it is either actual or not needed (when active_level > 0)
1526       (*p_hot_teams)[0].hot_team_nth = 1;
1527     }
1528 #endif
1529 
1530 #if OMPT_SUPPORT
1531     if (ompt_enabled.enabled) {
1532       if (ompt_enabled.ompt_callback_parallel_begin) {
1533         int team_size = master_set_numthreads
1534                             ? master_set_numthreads
1535                             : get__nproc_2(parent_team, master_tid);
1536         ompt_callbacks.ompt_callback(ompt_callback_parallel_begin)(
1537             parent_task_data, ompt_frame, &ompt_parallel_data, team_size,
1538             OMPT_INVOKER(call_context), return_address);
1539       }
1540       master_th->th.ompt_thread_info.state = ompt_state_overhead;
1541     }
1542 #endif
1543 
1544     master_th->th.th_ident = loc;
1545 
1546 #if OMP_40_ENABLED
1547     if (master_th->th.th_teams_microtask && ap &&
1548         microtask != (microtask_t)__kmp_teams_master && level == teams_level) {
1549       // AC: This is start of parallel that is nested inside teams construct.
1550       // The team is actual (hot), all workers are ready at the fork barrier.
1551       // No lock needed to initialize the team a bit, then free workers.
1552       parent_team->t.t_ident = loc;
1553       __kmp_alloc_argv_entries(argc, parent_team, TRUE);
1554       parent_team->t.t_argc = argc;
1555       argv = (void **)parent_team->t.t_argv;
1556       for (i = argc - 1; i >= 0; --i)
1557 /* TODO: revert workaround for Intel(R) 64 tracker #96 */
1558 #if (KMP_ARCH_X86_64 || KMP_ARCH_ARM || KMP_ARCH_AARCH64) && KMP_OS_LINUX
1559         *argv++ = va_arg(*ap, void *);
1560 #else
1561         *argv++ = va_arg(ap, void *);
1562 #endif
1563       // Increment our nested depth levels, but not increase the serialization
1564       if (parent_team == master_th->th.th_serial_team) {
1565         // AC: we are in serialized parallel
1566         __kmpc_serialized_parallel(loc, gtid);
1567         KMP_DEBUG_ASSERT(parent_team->t.t_serialized > 1);
1568         // AC: need this in order enquiry functions work
1569         // correctly, will restore at join time
1570         parent_team->t.t_serialized--;
1571 #if OMPT_SUPPORT
1572         void *dummy;
1573         void **exit_runtime_p;
1574 
1575         ompt_lw_taskteam_t lw_taskteam;
1576 
1577         if (ompt_enabled.enabled) {
1578           __ompt_lw_taskteam_init(&lw_taskteam, master_th, gtid,
1579                                   &ompt_parallel_data, return_address);
1580           exit_runtime_p = &(lw_taskteam.ompt_task_info.frame.exit_frame.ptr);
1581 
1582           __ompt_lw_taskteam_link(&lw_taskteam, master_th, 0);
1583           // don't use lw_taskteam after linking. content was swaped
1584 
1585           /* OMPT implicit task begin */
1586           implicit_task_data = OMPT_CUR_TASK_DATA(master_th);
1587           if (ompt_enabled.ompt_callback_implicit_task) {
1588             ompt_callbacks.ompt_callback(ompt_callback_implicit_task)(
1589                 ompt_scope_begin, OMPT_CUR_TEAM_DATA(master_th),
1590                 implicit_task_data, 1, __kmp_tid_from_gtid(gtid), ompt_task_implicit); // TODO: Can this be ompt_task_initial?
1591             OMPT_CUR_TASK_INFO(master_th)
1592                 ->thread_num = __kmp_tid_from_gtid(gtid);
1593           }
1594 
1595           /* OMPT state */
1596           master_th->th.ompt_thread_info.state = ompt_state_work_parallel;
1597         } else {
1598           exit_runtime_p = &dummy;
1599         }
1600 #endif
1601 
1602         {
1603           KMP_TIME_PARTITIONED_BLOCK(OMP_parallel);
1604           KMP_SET_THREAD_STATE_BLOCK(IMPLICIT_TASK);
1605           __kmp_invoke_microtask(microtask, gtid, 0, argc, parent_team->t.t_argv
1606 #if OMPT_SUPPORT
1607                                  ,
1608                                  exit_runtime_p
1609 #endif
1610                                  );
1611         }
1612 
1613 #if OMPT_SUPPORT
1614         *exit_runtime_p = NULL;
1615         if (ompt_enabled.enabled) {
1616           OMPT_CUR_TASK_INFO(master_th)->frame.exit_frame = ompt_data_none;
1617           if (ompt_enabled.ompt_callback_implicit_task) {
1618             ompt_callbacks.ompt_callback(ompt_callback_implicit_task)(
1619                 ompt_scope_end, NULL, implicit_task_data, 1,
1620                 OMPT_CUR_TASK_INFO(master_th)->thread_num, ompt_task_implicit); // TODO: Can this be ompt_task_initial?
1621           }
1622           __ompt_lw_taskteam_unlink(master_th);
1623 
1624           if (ompt_enabled.ompt_callback_parallel_end) {
1625             ompt_callbacks.ompt_callback(ompt_callback_parallel_end)(
1626                 OMPT_CUR_TEAM_DATA(master_th), OMPT_CUR_TASK_DATA(master_th),
1627                 OMPT_INVOKER(call_context), return_address);
1628           }
1629           master_th->th.ompt_thread_info.state = ompt_state_overhead;
1630         }
1631 #endif
1632         return TRUE;
1633       }
1634 
1635       parent_team->t.t_pkfn = microtask;
1636       parent_team->t.t_invoke = invoker;
1637       KMP_ATOMIC_INC(&root->r.r_in_parallel);
1638       parent_team->t.t_active_level++;
1639       parent_team->t.t_level++;
1640 #if OMP_50_ENABLED
1641       parent_team->t.t_def_allocator = master_th->th.th_def_allocator; // save
1642 #endif
1643 
1644       /* Change number of threads in the team if requested */
1645       if (master_set_numthreads) { // The parallel has num_threads clause
1646         if (master_set_numthreads < master_th->th.th_teams_size.nth) {
1647           // AC: only can reduce number of threads dynamically, can't increase
1648           kmp_info_t **other_threads = parent_team->t.t_threads;
1649           parent_team->t.t_nproc = master_set_numthreads;
1650           for (i = 0; i < master_set_numthreads; ++i) {
1651             other_threads[i]->th.th_team_nproc = master_set_numthreads;
1652           }
1653           // Keep extra threads hot in the team for possible next parallels
1654         }
1655         master_th->th.th_set_nproc = 0;
1656       }
1657 
1658 #if USE_DEBUGGER
1659       if (__kmp_debugging) { // Let debugger override number of threads.
1660         int nth = __kmp_omp_num_threads(loc);
1661         if (nth > 0) { // 0 means debugger doesn't want to change num threads
1662           master_set_numthreads = nth;
1663         }
1664       }
1665 #endif
1666 
1667       KF_TRACE(10, ("__kmp_fork_call: before internal fork: root=%p, team=%p, "
1668                     "master_th=%p, gtid=%d\n",
1669                     root, parent_team, master_th, gtid));
1670       __kmp_internal_fork(loc, gtid, parent_team);
1671       KF_TRACE(10, ("__kmp_fork_call: after internal fork: root=%p, team=%p, "
1672                     "master_th=%p, gtid=%d\n",
1673                     root, parent_team, master_th, gtid));
1674 
1675       /* Invoke microtask for MASTER thread */
1676       KA_TRACE(20, ("__kmp_fork_call: T#%d(%d:0) invoke microtask = %p\n", gtid,
1677                     parent_team->t.t_id, parent_team->t.t_pkfn));
1678 
1679       if (!parent_team->t.t_invoke(gtid)) {
1680         KMP_ASSERT2(0, "cannot invoke microtask for MASTER thread");
1681       }
1682       KA_TRACE(20, ("__kmp_fork_call: T#%d(%d:0) done microtask = %p\n", gtid,
1683                     parent_team->t.t_id, parent_team->t.t_pkfn));
1684       KMP_MB(); /* Flush all pending memory write invalidates.  */
1685 
1686       KA_TRACE(20, ("__kmp_fork_call: parallel exit T#%d\n", gtid));
1687 
1688       return TRUE;
1689     } // Parallel closely nested in teams construct
1690 #endif /* OMP_40_ENABLED */
1691 
1692 #if KMP_DEBUG
1693     if (__kmp_tasking_mode != tskm_immediate_exec) {
1694       KMP_DEBUG_ASSERT(master_th->th.th_task_team ==
1695                        parent_team->t.t_task_team[master_th->th.th_task_state]);
1696     }
1697 #endif
1698 
1699     if (parent_team->t.t_active_level >=
1700         master_th->th.th_current_task->td_icvs.max_active_levels) {
1701       nthreads = 1;
1702     } else {
1703 #if OMP_40_ENABLED
1704       int enter_teams = ((ap == NULL && active_level == 0) ||
1705                          (ap && teams_level > 0 && teams_level == level));
1706 #endif
1707       nthreads =
1708           master_set_numthreads
1709               ? master_set_numthreads
1710               : get__nproc_2(
1711                     parent_team,
1712                     master_tid); // TODO: get nproc directly from current task
1713 
1714       // Check if we need to take forkjoin lock? (no need for serialized
1715       // parallel out of teams construct). This code moved here from
1716       // __kmp_reserve_threads() to speedup nested serialized parallels.
1717       if (nthreads > 1) {
1718         if ((get__max_active_levels(master_th) == 1 && (root->r.r_in_parallel
1719 #if OMP_40_ENABLED
1720                                                         && !enter_teams
1721 #endif /* OMP_40_ENABLED */
1722                                                         )) ||
1723             (__kmp_library == library_serial)) {
1724           KC_TRACE(10, ("__kmp_fork_call: T#%d serializing team; requested %d"
1725                         " threads\n",
1726                         gtid, nthreads));
1727           nthreads = 1;
1728         }
1729       }
1730       if (nthreads > 1) {
1731         /* determine how many new threads we can use */
1732         __kmp_acquire_bootstrap_lock(&__kmp_forkjoin_lock);
1733         nthreads = __kmp_reserve_threads(
1734             root, parent_team, master_tid, nthreads
1735 #if OMP_40_ENABLED
1736             /* AC: If we execute teams from parallel region (on host), then
1737                teams should be created but each can only have 1 thread if
1738                nesting is disabled. If teams called from serial region, then
1739                teams and their threads should be created regardless of the
1740                nesting setting. */
1741             ,
1742             enter_teams
1743 #endif /* OMP_40_ENABLED */
1744             );
1745         if (nthreads == 1) {
1746           // Free lock for single thread execution here; for multi-thread
1747           // execution it will be freed later after team of threads created
1748           // and initialized
1749           __kmp_release_bootstrap_lock(&__kmp_forkjoin_lock);
1750         }
1751       }
1752     }
1753     KMP_DEBUG_ASSERT(nthreads > 0);
1754 
1755     // If we temporarily changed the set number of threads then restore it now
1756     master_th->th.th_set_nproc = 0;
1757 
1758     /* create a serialized parallel region? */
1759     if (nthreads == 1) {
1760 /* josh todo: hypothetical question: what do we do for OS X*? */
1761 #if KMP_OS_LINUX &&                                                            \
1762     (KMP_ARCH_X86 || KMP_ARCH_X86_64 || KMP_ARCH_ARM || KMP_ARCH_AARCH64)
1763       void *args[argc];
1764 #else
1765       void **args = (void **)KMP_ALLOCA(argc * sizeof(void *));
1766 #endif /* KMP_OS_LINUX && ( KMP_ARCH_X86 || KMP_ARCH_X86_64 || KMP_ARCH_ARM || \
1767           KMP_ARCH_AARCH64) */
1768 
1769       KA_TRACE(20,
1770                ("__kmp_fork_call: T#%d serializing parallel region\n", gtid));
1771 
1772       __kmpc_serialized_parallel(loc, gtid);
1773 
1774       if (call_context == fork_context_intel) {
1775         /* TODO this sucks, use the compiler itself to pass args! :) */
1776         master_th->th.th_serial_team->t.t_ident = loc;
1777 #if OMP_40_ENABLED
1778         if (!ap) {
1779           // revert change made in __kmpc_serialized_parallel()
1780           master_th->th.th_serial_team->t.t_level--;
1781 // Get args from parent team for teams construct
1782 
1783 #if OMPT_SUPPORT
1784           void *dummy;
1785           void **exit_runtime_p;
1786           ompt_task_info_t *task_info;
1787 
1788           ompt_lw_taskteam_t lw_taskteam;
1789 
1790           if (ompt_enabled.enabled) {
1791             __ompt_lw_taskteam_init(&lw_taskteam, master_th, gtid,
1792                                     &ompt_parallel_data, return_address);
1793 
1794             __ompt_lw_taskteam_link(&lw_taskteam, master_th, 0);
1795             // don't use lw_taskteam after linking. content was swaped
1796 
1797             task_info = OMPT_CUR_TASK_INFO(master_th);
1798             exit_runtime_p = &(task_info->frame.exit_frame.ptr);
1799             if (ompt_enabled.ompt_callback_implicit_task) {
1800               ompt_callbacks.ompt_callback(ompt_callback_implicit_task)(
1801                   ompt_scope_begin, OMPT_CUR_TEAM_DATA(master_th),
1802                   &(task_info->task_data), 1, __kmp_tid_from_gtid(gtid), ompt_task_implicit); // TODO: Can this be ompt_task_initial?
1803               OMPT_CUR_TASK_INFO(master_th)
1804                   ->thread_num = __kmp_tid_from_gtid(gtid);
1805             }
1806 
1807             /* OMPT state */
1808             master_th->th.ompt_thread_info.state = ompt_state_work_parallel;
1809           } else {
1810             exit_runtime_p = &dummy;
1811           }
1812 #endif
1813 
1814           {
1815             KMP_TIME_PARTITIONED_BLOCK(OMP_parallel);
1816             KMP_SET_THREAD_STATE_BLOCK(IMPLICIT_TASK);
1817             __kmp_invoke_microtask(microtask, gtid, 0, argc,
1818                                    parent_team->t.t_argv
1819 #if OMPT_SUPPORT
1820                                    ,
1821                                    exit_runtime_p
1822 #endif
1823                                    );
1824           }
1825 
1826 #if OMPT_SUPPORT
1827           if (ompt_enabled.enabled) {
1828             exit_runtime_p = NULL;
1829             if (ompt_enabled.ompt_callback_implicit_task) {
1830               ompt_callbacks.ompt_callback(ompt_callback_implicit_task)(
1831                   ompt_scope_end, NULL, &(task_info->task_data), 1,
1832                   OMPT_CUR_TASK_INFO(master_th)->thread_num, ompt_task_implicit); // TODO: Can this be ompt_task_initial?
1833             }
1834 
1835             __ompt_lw_taskteam_unlink(master_th);
1836             if (ompt_enabled.ompt_callback_parallel_end) {
1837               ompt_callbacks.ompt_callback(ompt_callback_parallel_end)(
1838                   OMPT_CUR_TEAM_DATA(master_th), parent_task_data,
1839                   OMPT_INVOKER(call_context), return_address);
1840             }
1841             master_th->th.ompt_thread_info.state = ompt_state_overhead;
1842           }
1843 #endif
1844         } else if (microtask == (microtask_t)__kmp_teams_master) {
1845           KMP_DEBUG_ASSERT(master_th->th.th_team ==
1846                            master_th->th.th_serial_team);
1847           team = master_th->th.th_team;
1848           // team->t.t_pkfn = microtask;
1849           team->t.t_invoke = invoker;
1850           __kmp_alloc_argv_entries(argc, team, TRUE);
1851           team->t.t_argc = argc;
1852           argv = (void **)team->t.t_argv;
1853           if (ap) {
1854             for (i = argc - 1; i >= 0; --i)
1855 // TODO: revert workaround for Intel(R) 64 tracker #96
1856 #if (KMP_ARCH_X86_64 || KMP_ARCH_ARM || KMP_ARCH_AARCH64) && KMP_OS_LINUX
1857               *argv++ = va_arg(*ap, void *);
1858 #else
1859               *argv++ = va_arg(ap, void *);
1860 #endif
1861           } else {
1862             for (i = 0; i < argc; ++i)
1863               // Get args from parent team for teams construct
1864               argv[i] = parent_team->t.t_argv[i];
1865           }
1866           // AC: revert change made in __kmpc_serialized_parallel()
1867           //     because initial code in teams should have level=0
1868           team->t.t_level--;
1869           // AC: call special invoker for outer "parallel" of teams construct
1870           invoker(gtid);
1871         } else {
1872 #endif /* OMP_40_ENABLED */
1873           argv = args;
1874           for (i = argc - 1; i >= 0; --i)
1875 // TODO: revert workaround for Intel(R) 64 tracker #96
1876 #if (KMP_ARCH_X86_64 || KMP_ARCH_ARM || KMP_ARCH_AARCH64) && KMP_OS_LINUX
1877             *argv++ = va_arg(*ap, void *);
1878 #else
1879           *argv++ = va_arg(ap, void *);
1880 #endif
1881           KMP_MB();
1882 
1883 #if OMPT_SUPPORT
1884           void *dummy;
1885           void **exit_runtime_p;
1886           ompt_task_info_t *task_info;
1887 
1888           ompt_lw_taskteam_t lw_taskteam;
1889 
1890           if (ompt_enabled.enabled) {
1891             __ompt_lw_taskteam_init(&lw_taskteam, master_th, gtid,
1892                                     &ompt_parallel_data, return_address);
1893             __ompt_lw_taskteam_link(&lw_taskteam, master_th, 0);
1894             // don't use lw_taskteam after linking. content was swaped
1895             task_info = OMPT_CUR_TASK_INFO(master_th);
1896             exit_runtime_p = &(task_info->frame.exit_frame.ptr);
1897 
1898             /* OMPT implicit task begin */
1899             implicit_task_data = OMPT_CUR_TASK_DATA(master_th);
1900             if (ompt_enabled.ompt_callback_implicit_task) {
1901               ompt_callbacks.ompt_callback(ompt_callback_implicit_task)(
1902                   ompt_scope_begin, OMPT_CUR_TEAM_DATA(master_th),
1903                   implicit_task_data, 1, __kmp_tid_from_gtid(gtid), ompt_task_implicit); // TODO: Can this be ompt_task_initial?
1904               OMPT_CUR_TASK_INFO(master_th)
1905                   ->thread_num = __kmp_tid_from_gtid(gtid);
1906             }
1907 
1908             /* OMPT state */
1909             master_th->th.ompt_thread_info.state = ompt_state_work_parallel;
1910           } else {
1911             exit_runtime_p = &dummy;
1912           }
1913 #endif
1914 
1915           {
1916             KMP_TIME_PARTITIONED_BLOCK(OMP_parallel);
1917             KMP_SET_THREAD_STATE_BLOCK(IMPLICIT_TASK);
1918             __kmp_invoke_microtask(microtask, gtid, 0, argc, args
1919 #if OMPT_SUPPORT
1920                                    ,
1921                                    exit_runtime_p
1922 #endif
1923                                    );
1924           }
1925 
1926 #if OMPT_SUPPORT
1927           if (ompt_enabled.enabled) {
1928             *exit_runtime_p = NULL;
1929             if (ompt_enabled.ompt_callback_implicit_task) {
1930               ompt_callbacks.ompt_callback(ompt_callback_implicit_task)(
1931                   ompt_scope_end, NULL, &(task_info->task_data), 1,
1932                   OMPT_CUR_TASK_INFO(master_th)->thread_num, ompt_task_implicit); // TODO: Can this be ompt_task_initial?
1933             }
1934 
1935             ompt_parallel_data = *OMPT_CUR_TEAM_DATA(master_th);
1936             __ompt_lw_taskteam_unlink(master_th);
1937             if (ompt_enabled.ompt_callback_parallel_end) {
1938               ompt_callbacks.ompt_callback(ompt_callback_parallel_end)(
1939                   &ompt_parallel_data, parent_task_data,
1940                   OMPT_INVOKER(call_context), return_address);
1941             }
1942             master_th->th.ompt_thread_info.state = ompt_state_overhead;
1943           }
1944 #endif
1945 #if OMP_40_ENABLED
1946         }
1947 #endif /* OMP_40_ENABLED */
1948       } else if (call_context == fork_context_gnu) {
1949 #if OMPT_SUPPORT
1950         ompt_lw_taskteam_t lwt;
1951         __ompt_lw_taskteam_init(&lwt, master_th, gtid, &ompt_parallel_data,
1952                                 return_address);
1953 
1954         lwt.ompt_task_info.frame.exit_frame = ompt_data_none;
1955         __ompt_lw_taskteam_link(&lwt, master_th, 1);
1956 // don't use lw_taskteam after linking. content was swaped
1957 #endif
1958 
1959         // we were called from GNU native code
1960         KA_TRACE(20, ("__kmp_fork_call: T#%d serial exit\n", gtid));
1961         return FALSE;
1962       } else {
1963         KMP_ASSERT2(call_context < fork_context_last,
1964                     "__kmp_fork_call: unknown fork_context parameter");
1965       }
1966 
1967       KA_TRACE(20, ("__kmp_fork_call: T#%d serial exit\n", gtid));
1968       KMP_MB();
1969       return FALSE;
1970     } // if (nthreads == 1)
1971 
1972     // GEH: only modify the executing flag in the case when not serialized
1973     //      serialized case is handled in kmpc_serialized_parallel
1974     KF_TRACE(10, ("__kmp_fork_call: parent_team_aclevel=%d, master_th=%p, "
1975                   "curtask=%p, curtask_max_aclevel=%d\n",
1976                   parent_team->t.t_active_level, master_th,
1977                   master_th->th.th_current_task,
1978                   master_th->th.th_current_task->td_icvs.max_active_levels));
1979     // TODO: GEH - cannot do this assertion because root thread not set up as
1980     // executing
1981     // KMP_ASSERT( master_th->th.th_current_task->td_flags.executing == 1 );
1982     master_th->th.th_current_task->td_flags.executing = 0;
1983 
1984 #if OMP_40_ENABLED
1985     if (!master_th->th.th_teams_microtask || level > teams_level)
1986 #endif /* OMP_40_ENABLED */
1987     {
1988       /* Increment our nested depth level */
1989       KMP_ATOMIC_INC(&root->r.r_in_parallel);
1990     }
1991 
1992     // See if we need to make a copy of the ICVs.
1993     int nthreads_icv = master_th->th.th_current_task->td_icvs.nproc;
1994     if ((level + 1 < __kmp_nested_nth.used) &&
1995         (__kmp_nested_nth.nth[level + 1] != nthreads_icv)) {
1996       nthreads_icv = __kmp_nested_nth.nth[level + 1];
1997     } else {
1998       nthreads_icv = 0; // don't update
1999     }
2000 
2001 #if OMP_40_ENABLED
2002     // Figure out the proc_bind_policy for the new team.
2003     kmp_proc_bind_t proc_bind = master_th->th.th_set_proc_bind;
2004     kmp_proc_bind_t proc_bind_icv =
2005         proc_bind_default; // proc_bind_default means don't update
2006     if (master_th->th.th_current_task->td_icvs.proc_bind == proc_bind_false) {
2007       proc_bind = proc_bind_false;
2008     } else {
2009       if (proc_bind == proc_bind_default) {
2010         // No proc_bind clause specified; use current proc-bind-var for this
2011         // parallel region
2012         proc_bind = master_th->th.th_current_task->td_icvs.proc_bind;
2013       }
2014       /* else: The proc_bind policy was specified explicitly on parallel clause.
2015          This overrides proc-bind-var for this parallel region, but does not
2016          change proc-bind-var. */
2017       // Figure the value of proc-bind-var for the child threads.
2018       if ((level + 1 < __kmp_nested_proc_bind.used) &&
2019           (__kmp_nested_proc_bind.bind_types[level + 1] !=
2020            master_th->th.th_current_task->td_icvs.proc_bind)) {
2021         proc_bind_icv = __kmp_nested_proc_bind.bind_types[level + 1];
2022       }
2023     }
2024 
2025     // Reset for next parallel region
2026     master_th->th.th_set_proc_bind = proc_bind_default;
2027 #endif /* OMP_40_ENABLED */
2028 
2029     if ((nthreads_icv > 0)
2030 #if OMP_40_ENABLED
2031         || (proc_bind_icv != proc_bind_default)
2032 #endif /* OMP_40_ENABLED */
2033             ) {
2034       kmp_internal_control_t new_icvs;
2035       copy_icvs(&new_icvs, &master_th->th.th_current_task->td_icvs);
2036       new_icvs.next = NULL;
2037       if (nthreads_icv > 0) {
2038         new_icvs.nproc = nthreads_icv;
2039       }
2040 
2041 #if OMP_40_ENABLED
2042       if (proc_bind_icv != proc_bind_default) {
2043         new_icvs.proc_bind = proc_bind_icv;
2044       }
2045 #endif /* OMP_40_ENABLED */
2046 
2047       /* allocate a new parallel team */
2048       KF_TRACE(10, ("__kmp_fork_call: before __kmp_allocate_team\n"));
2049       team = __kmp_allocate_team(root, nthreads, nthreads,
2050 #if OMPT_SUPPORT
2051                                  ompt_parallel_data,
2052 #endif
2053 #if OMP_40_ENABLED
2054                                  proc_bind,
2055 #endif
2056                                  &new_icvs, argc USE_NESTED_HOT_ARG(master_th));
2057     } else {
2058       /* allocate a new parallel team */
2059       KF_TRACE(10, ("__kmp_fork_call: before __kmp_allocate_team\n"));
2060       team = __kmp_allocate_team(root, nthreads, nthreads,
2061 #if OMPT_SUPPORT
2062                                  ompt_parallel_data,
2063 #endif
2064 #if OMP_40_ENABLED
2065                                  proc_bind,
2066 #endif
2067                                  &master_th->th.th_current_task->td_icvs,
2068                                  argc USE_NESTED_HOT_ARG(master_th));
2069     }
2070     KF_TRACE(
2071         10, ("__kmp_fork_call: after __kmp_allocate_team - team = %p\n", team));
2072 
2073     /* setup the new team */
2074     KMP_CHECK_UPDATE(team->t.t_master_tid, master_tid);
2075     KMP_CHECK_UPDATE(team->t.t_master_this_cons, master_this_cons);
2076     KMP_CHECK_UPDATE(team->t.t_ident, loc);
2077     KMP_CHECK_UPDATE(team->t.t_parent, parent_team);
2078     KMP_CHECK_UPDATE_SYNC(team->t.t_pkfn, microtask);
2079 #if OMPT_SUPPORT
2080     KMP_CHECK_UPDATE_SYNC(team->t.ompt_team_info.master_return_address,
2081                           return_address);
2082 #endif
2083     KMP_CHECK_UPDATE(team->t.t_invoke, invoker); // TODO move to root, maybe
2084 // TODO: parent_team->t.t_level == INT_MAX ???
2085 #if OMP_40_ENABLED
2086     if (!master_th->th.th_teams_microtask || level > teams_level) {
2087 #endif /* OMP_40_ENABLED */
2088       int new_level = parent_team->t.t_level + 1;
2089       KMP_CHECK_UPDATE(team->t.t_level, new_level);
2090       new_level = parent_team->t.t_active_level + 1;
2091       KMP_CHECK_UPDATE(team->t.t_active_level, new_level);
2092 #if OMP_40_ENABLED
2093     } else {
2094       // AC: Do not increase parallel level at start of the teams construct
2095       int new_level = parent_team->t.t_level;
2096       KMP_CHECK_UPDATE(team->t.t_level, new_level);
2097       new_level = parent_team->t.t_active_level;
2098       KMP_CHECK_UPDATE(team->t.t_active_level, new_level);
2099     }
2100 #endif /* OMP_40_ENABLED */
2101     kmp_r_sched_t new_sched = get__sched_2(parent_team, master_tid);
2102     // set master's schedule as new run-time schedule
2103     KMP_CHECK_UPDATE(team->t.t_sched.sched, new_sched.sched);
2104 
2105 #if OMP_40_ENABLED
2106     KMP_CHECK_UPDATE(team->t.t_cancel_request, cancel_noreq);
2107 #endif
2108 #if OMP_50_ENABLED
2109     KMP_CHECK_UPDATE(team->t.t_def_allocator, master_th->th.th_def_allocator);
2110 #endif
2111 
2112     // Update the floating point rounding in the team if required.
2113     propagateFPControl(team);
2114 
2115     if (__kmp_tasking_mode != tskm_immediate_exec) {
2116       // Set master's task team to team's task team. Unless this is hot team, it
2117       // should be NULL.
2118       KMP_DEBUG_ASSERT(master_th->th.th_task_team ==
2119                        parent_team->t.t_task_team[master_th->th.th_task_state]);
2120       KA_TRACE(20, ("__kmp_fork_call: Master T#%d pushing task_team %p / team "
2121                     "%p, new task_team %p / team %p\n",
2122                     __kmp_gtid_from_thread(master_th),
2123                     master_th->th.th_task_team, parent_team,
2124                     team->t.t_task_team[master_th->th.th_task_state], team));
2125 
2126       if (active_level || master_th->th.th_task_team) {
2127         // Take a memo of master's task_state
2128         KMP_DEBUG_ASSERT(master_th->th.th_task_state_memo_stack);
2129         if (master_th->th.th_task_state_top >=
2130             master_th->th.th_task_state_stack_sz) { // increase size
2131           kmp_uint32 new_size = 2 * master_th->th.th_task_state_stack_sz;
2132           kmp_uint8 *old_stack, *new_stack;
2133           kmp_uint32 i;
2134           new_stack = (kmp_uint8 *)__kmp_allocate(new_size);
2135           for (i = 0; i < master_th->th.th_task_state_stack_sz; ++i) {
2136             new_stack[i] = master_th->th.th_task_state_memo_stack[i];
2137           }
2138           for (i = master_th->th.th_task_state_stack_sz; i < new_size;
2139                ++i) { // zero-init rest of stack
2140             new_stack[i] = 0;
2141           }
2142           old_stack = master_th->th.th_task_state_memo_stack;
2143           master_th->th.th_task_state_memo_stack = new_stack;
2144           master_th->th.th_task_state_stack_sz = new_size;
2145           __kmp_free(old_stack);
2146         }
2147         // Store master's task_state on stack
2148         master_th->th
2149             .th_task_state_memo_stack[master_th->th.th_task_state_top] =
2150             master_th->th.th_task_state;
2151         master_th->th.th_task_state_top++;
2152 #if KMP_NESTED_HOT_TEAMS
2153         if (master_th->th.th_hot_teams &&
2154             active_level < __kmp_hot_teams_max_level &&
2155             team == master_th->th.th_hot_teams[active_level].hot_team) {
2156           // Restore master's nested state if nested hot team
2157           master_th->th.th_task_state =
2158               master_th->th
2159                   .th_task_state_memo_stack[master_th->th.th_task_state_top];
2160         } else {
2161 #endif
2162           master_th->th.th_task_state = 0;
2163 #if KMP_NESTED_HOT_TEAMS
2164         }
2165 #endif
2166       }
2167 #if !KMP_NESTED_HOT_TEAMS
2168       KMP_DEBUG_ASSERT((master_th->th.th_task_team == NULL) ||
2169                        (team == root->r.r_hot_team));
2170 #endif
2171     }
2172 
2173     KA_TRACE(
2174         20,
2175         ("__kmp_fork_call: T#%d(%d:%d)->(%d:0) created a team of %d threads\n",
2176          gtid, parent_team->t.t_id, team->t.t_master_tid, team->t.t_id,
2177          team->t.t_nproc));
2178     KMP_DEBUG_ASSERT(team != root->r.r_hot_team ||
2179                      (team->t.t_master_tid == 0 &&
2180                       (team->t.t_parent == root->r.r_root_team ||
2181                        team->t.t_parent->t.t_serialized)));
2182     KMP_MB();
2183 
2184     /* now, setup the arguments */
2185     argv = (void **)team->t.t_argv;
2186 #if OMP_40_ENABLED
2187     if (ap) {
2188 #endif /* OMP_40_ENABLED */
2189       for (i = argc - 1; i >= 0; --i) {
2190 // TODO: revert workaround for Intel(R) 64 tracker #96
2191 #if (KMP_ARCH_X86_64 || KMP_ARCH_ARM || KMP_ARCH_AARCH64) && KMP_OS_LINUX
2192         void *new_argv = va_arg(*ap, void *);
2193 #else
2194       void *new_argv = va_arg(ap, void *);
2195 #endif
2196         KMP_CHECK_UPDATE(*argv, new_argv);
2197         argv++;
2198       }
2199 #if OMP_40_ENABLED
2200     } else {
2201       for (i = 0; i < argc; ++i) {
2202         // Get args from parent team for teams construct
2203         KMP_CHECK_UPDATE(argv[i], team->t.t_parent->t.t_argv[i]);
2204       }
2205     }
2206 #endif /* OMP_40_ENABLED */
2207 
2208     /* now actually fork the threads */
2209     KMP_CHECK_UPDATE(team->t.t_master_active, master_active);
2210     if (!root->r.r_active) // Only do assignment if it prevents cache ping-pong
2211       root->r.r_active = TRUE;
2212 
2213     __kmp_fork_team_threads(root, team, master_th, gtid);
2214     __kmp_setup_icv_copy(team, nthreads,
2215                          &master_th->th.th_current_task->td_icvs, loc);
2216 
2217 #if OMPT_SUPPORT
2218     master_th->th.ompt_thread_info.state = ompt_state_work_parallel;
2219 #endif
2220 
2221     __kmp_release_bootstrap_lock(&__kmp_forkjoin_lock);
2222 
2223 #if USE_ITT_BUILD
2224     if (team->t.t_active_level == 1 // only report frames at level 1
2225 #if OMP_40_ENABLED
2226         && !master_th->th.th_teams_microtask // not in teams construct
2227 #endif /* OMP_40_ENABLED */
2228         ) {
2229 #if USE_ITT_NOTIFY
2230       if ((__itt_frame_submit_v3_ptr || KMP_ITT_DEBUG) &&
2231           (__kmp_forkjoin_frames_mode == 3 ||
2232            __kmp_forkjoin_frames_mode == 1)) {
2233         kmp_uint64 tmp_time = 0;
2234         if (__itt_get_timestamp_ptr)
2235           tmp_time = __itt_get_timestamp();
2236         // Internal fork - report frame begin
2237         master_th->th.th_frame_time = tmp_time;
2238         if (__kmp_forkjoin_frames_mode == 3)
2239           team->t.t_region_time = tmp_time;
2240       } else
2241 // only one notification scheme (either "submit" or "forking/joined", not both)
2242 #endif /* USE_ITT_NOTIFY */
2243           if ((__itt_frame_begin_v3_ptr || KMP_ITT_DEBUG) &&
2244               __kmp_forkjoin_frames && !__kmp_forkjoin_frames_mode) {
2245         // Mark start of "parallel" region for Intel(R) VTune(TM) analyzer.
2246         __kmp_itt_region_forking(gtid, team->t.t_nproc, 0);
2247       }
2248     }
2249 #endif /* USE_ITT_BUILD */
2250 
2251     /* now go on and do the work */
2252     KMP_DEBUG_ASSERT(team == __kmp_threads[gtid]->th.th_team);
2253     KMP_MB();
2254     KF_TRACE(10,
2255              ("__kmp_internal_fork : root=%p, team=%p, master_th=%p, gtid=%d\n",
2256               root, team, master_th, gtid));
2257 
2258 #if USE_ITT_BUILD
2259     if (__itt_stack_caller_create_ptr) {
2260       team->t.t_stack_id =
2261           __kmp_itt_stack_caller_create(); // create new stack stitching id
2262       // before entering fork barrier
2263     }
2264 #endif /* USE_ITT_BUILD */
2265 
2266 #if OMP_40_ENABLED
2267     // AC: skip __kmp_internal_fork at teams construct, let only master
2268     // threads execute
2269     if (ap)
2270 #endif /* OMP_40_ENABLED */
2271     {
2272       __kmp_internal_fork(loc, gtid, team);
2273       KF_TRACE(10, ("__kmp_internal_fork : after : root=%p, team=%p, "
2274                     "master_th=%p, gtid=%d\n",
2275                     root, team, master_th, gtid));
2276     }
2277 
2278     if (call_context == fork_context_gnu) {
2279       KA_TRACE(20, ("__kmp_fork_call: parallel exit T#%d\n", gtid));
2280       return TRUE;
2281     }
2282 
2283     /* Invoke microtask for MASTER thread */
2284     KA_TRACE(20, ("__kmp_fork_call: T#%d(%d:0) invoke microtask = %p\n", gtid,
2285                   team->t.t_id, team->t.t_pkfn));
2286   } // END of timer KMP_fork_call block
2287 
2288 #if KMP_STATS_ENABLED && OMP_40_ENABLED
2289   // If beginning a teams construct, then change thread state
2290   stats_state_e previous_state = KMP_GET_THREAD_STATE();
2291   if (!ap) {
2292     KMP_SET_THREAD_STATE(stats_state_e::TEAMS_REGION);
2293   }
2294 #endif
2295 
2296   if (!team->t.t_invoke(gtid)) {
2297     KMP_ASSERT2(0, "cannot invoke microtask for MASTER thread");
2298   }
2299 
2300 #if KMP_STATS_ENABLED && OMP_40_ENABLED
2301   // If was beginning of a teams construct, then reset thread state
2302   if (!ap) {
2303     KMP_SET_THREAD_STATE(previous_state);
2304   }
2305 #endif
2306 
2307   KA_TRACE(20, ("__kmp_fork_call: T#%d(%d:0) done microtask = %p\n", gtid,
2308                 team->t.t_id, team->t.t_pkfn));
2309   KMP_MB(); /* Flush all pending memory write invalidates.  */
2310 
2311   KA_TRACE(20, ("__kmp_fork_call: parallel exit T#%d\n", gtid));
2312 
2313 #if OMPT_SUPPORT
2314   if (ompt_enabled.enabled) {
2315     master_th->th.ompt_thread_info.state = ompt_state_overhead;
2316   }
2317 #endif
2318 
2319   return TRUE;
2320 }
2321 
2322 #if OMPT_SUPPORT
2323 static inline void __kmp_join_restore_state(kmp_info_t *thread,
2324                                             kmp_team_t *team) {
2325   // restore state outside the region
2326   thread->th.ompt_thread_info.state =
2327       ((team->t.t_serialized) ? ompt_state_work_serial
2328                               : ompt_state_work_parallel);
2329 }
2330 
2331 static inline void __kmp_join_ompt(int gtid, kmp_info_t *thread,
2332                                    kmp_team_t *team, ompt_data_t *parallel_data,
2333                                    fork_context_e fork_context, void *codeptr) {
2334   ompt_task_info_t *task_info = __ompt_get_task_info_object(0);
2335   if (ompt_enabled.ompt_callback_parallel_end) {
2336     ompt_callbacks.ompt_callback(ompt_callback_parallel_end)(
2337         parallel_data, &(task_info->task_data), OMPT_INVOKER(fork_context),
2338         codeptr);
2339   }
2340 
2341   task_info->frame.enter_frame = ompt_data_none;
2342   __kmp_join_restore_state(thread, team);
2343 }
2344 #endif
2345 
2346 void __kmp_join_call(ident_t *loc, int gtid
2347 #if OMPT_SUPPORT
2348                      ,
2349                      enum fork_context_e fork_context
2350 #endif
2351 #if OMP_40_ENABLED
2352                      ,
2353                      int exit_teams
2354 #endif /* OMP_40_ENABLED */
2355                      ) {
2356   KMP_TIME_DEVELOPER_PARTITIONED_BLOCK(KMP_join_call);
2357   kmp_team_t *team;
2358   kmp_team_t *parent_team;
2359   kmp_info_t *master_th;
2360   kmp_root_t *root;
2361   int master_active;
2362 
2363   KA_TRACE(20, ("__kmp_join_call: enter T#%d\n", gtid));
2364 
2365   /* setup current data */
2366   master_th = __kmp_threads[gtid];
2367   root = master_th->th.th_root;
2368   team = master_th->th.th_team;
2369   parent_team = team->t.t_parent;
2370 
2371   master_th->th.th_ident = loc;
2372 
2373 #if OMPT_SUPPORT
2374   if (ompt_enabled.enabled) {
2375     master_th->th.ompt_thread_info.state = ompt_state_overhead;
2376   }
2377 #endif
2378 
2379 #if KMP_DEBUG
2380   if (__kmp_tasking_mode != tskm_immediate_exec && !exit_teams) {
2381     KA_TRACE(20, ("__kmp_join_call: T#%d, old team = %p old task_team = %p, "
2382                   "th_task_team = %p\n",
2383                   __kmp_gtid_from_thread(master_th), team,
2384                   team->t.t_task_team[master_th->th.th_task_state],
2385                   master_th->th.th_task_team));
2386     KMP_DEBUG_ASSERT(master_th->th.th_task_team ==
2387                      team->t.t_task_team[master_th->th.th_task_state]);
2388   }
2389 #endif
2390 
2391   if (team->t.t_serialized) {
2392 #if OMP_40_ENABLED
2393     if (master_th->th.th_teams_microtask) {
2394       // We are in teams construct
2395       int level = team->t.t_level;
2396       int tlevel = master_th->th.th_teams_level;
2397       if (level == tlevel) {
2398         // AC: we haven't incremented it earlier at start of teams construct,
2399         //     so do it here - at the end of teams construct
2400         team->t.t_level++;
2401       } else if (level == tlevel + 1) {
2402         // AC: we are exiting parallel inside teams, need to increment
2403         // serialization in order to restore it in the next call to
2404         // __kmpc_end_serialized_parallel
2405         team->t.t_serialized++;
2406       }
2407     }
2408 #endif /* OMP_40_ENABLED */
2409     __kmpc_end_serialized_parallel(loc, gtid);
2410 
2411 #if OMPT_SUPPORT
2412     if (ompt_enabled.enabled) {
2413       __kmp_join_restore_state(master_th, parent_team);
2414     }
2415 #endif
2416 
2417     return;
2418   }
2419 
2420   master_active = team->t.t_master_active;
2421 
2422 #if OMP_40_ENABLED
2423   if (!exit_teams)
2424 #endif /* OMP_40_ENABLED */
2425   {
2426     // AC: No barrier for internal teams at exit from teams construct.
2427     //     But there is barrier for external team (league).
2428     __kmp_internal_join(loc, gtid, team);
2429   }
2430 #if OMP_40_ENABLED
2431   else {
2432     master_th->th.th_task_state =
2433         0; // AC: no tasking in teams (out of any parallel)
2434   }
2435 #endif /* OMP_40_ENABLED */
2436 
2437   KMP_MB();
2438 
2439 #if OMPT_SUPPORT
2440   ompt_data_t *parallel_data = &(team->t.ompt_team_info.parallel_data);
2441   void *codeptr = team->t.ompt_team_info.master_return_address;
2442 #endif
2443 
2444 #if USE_ITT_BUILD
2445   if (__itt_stack_caller_create_ptr) {
2446     __kmp_itt_stack_caller_destroy(
2447         (__itt_caller)team->t
2448             .t_stack_id); // destroy the stack stitching id after join barrier
2449   }
2450 
2451   // Mark end of "parallel" region for Intel(R) VTune(TM) analyzer.
2452   if (team->t.t_active_level == 1
2453 #if OMP_40_ENABLED
2454       && !master_th->th.th_teams_microtask /* not in teams construct */
2455 #endif /* OMP_40_ENABLED */
2456       ) {
2457     master_th->th.th_ident = loc;
2458     // only one notification scheme (either "submit" or "forking/joined", not
2459     // both)
2460     if ((__itt_frame_submit_v3_ptr || KMP_ITT_DEBUG) &&
2461         __kmp_forkjoin_frames_mode == 3)
2462       __kmp_itt_frame_submit(gtid, team->t.t_region_time,
2463                              master_th->th.th_frame_time, 0, loc,
2464                              master_th->th.th_team_nproc, 1);
2465     else if ((__itt_frame_end_v3_ptr || KMP_ITT_DEBUG) &&
2466              !__kmp_forkjoin_frames_mode && __kmp_forkjoin_frames)
2467       __kmp_itt_region_joined(gtid);
2468   } // active_level == 1
2469 #endif /* USE_ITT_BUILD */
2470 
2471 #if OMP_40_ENABLED
2472   if (master_th->th.th_teams_microtask && !exit_teams &&
2473       team->t.t_pkfn != (microtask_t)__kmp_teams_master &&
2474       team->t.t_level == master_th->th.th_teams_level + 1) {
2475     // AC: We need to leave the team structure intact at the end of parallel
2476     // inside the teams construct, so that at the next parallel same (hot) team
2477     // works, only adjust nesting levels
2478 
2479     /* Decrement our nested depth level */
2480     team->t.t_level--;
2481     team->t.t_active_level--;
2482     KMP_ATOMIC_DEC(&root->r.r_in_parallel);
2483 
2484     // Restore number of threads in the team if needed. This code relies on
2485     // the proper adjustment of th_teams_size.nth after the fork in
2486     // __kmp_teams_master on each teams master in the case that
2487     // __kmp_reserve_threads reduced it.
2488     if (master_th->th.th_team_nproc < master_th->th.th_teams_size.nth) {
2489       int old_num = master_th->th.th_team_nproc;
2490       int new_num = master_th->th.th_teams_size.nth;
2491       kmp_info_t **other_threads = team->t.t_threads;
2492       team->t.t_nproc = new_num;
2493       for (int i = 0; i < old_num; ++i) {
2494         other_threads[i]->th.th_team_nproc = new_num;
2495       }
2496       // Adjust states of non-used threads of the team
2497       for (int i = old_num; i < new_num; ++i) {
2498         // Re-initialize thread's barrier data.
2499         KMP_DEBUG_ASSERT(other_threads[i]);
2500         kmp_balign_t *balign = other_threads[i]->th.th_bar;
2501         for (int b = 0; b < bs_last_barrier; ++b) {
2502           balign[b].bb.b_arrived = team->t.t_bar[b].b_arrived;
2503           KMP_DEBUG_ASSERT(balign[b].bb.wait_flag != KMP_BARRIER_PARENT_FLAG);
2504 #if USE_DEBUGGER
2505           balign[b].bb.b_worker_arrived = team->t.t_bar[b].b_team_arrived;
2506 #endif
2507         }
2508         if (__kmp_tasking_mode != tskm_immediate_exec) {
2509           // Synchronize thread's task state
2510           other_threads[i]->th.th_task_state = master_th->th.th_task_state;
2511         }
2512       }
2513     }
2514 
2515 #if OMPT_SUPPORT
2516     if (ompt_enabled.enabled) {
2517       __kmp_join_ompt(gtid, master_th, parent_team, parallel_data, fork_context,
2518                       codeptr);
2519     }
2520 #endif
2521 
2522     return;
2523   }
2524 #endif /* OMP_40_ENABLED */
2525 
2526   /* do cleanup and restore the parent team */
2527   master_th->th.th_info.ds.ds_tid = team->t.t_master_tid;
2528   master_th->th.th_local.this_construct = team->t.t_master_this_cons;
2529 
2530   master_th->th.th_dispatch = &parent_team->t.t_dispatch[team->t.t_master_tid];
2531 
2532   /* jc: The following lock has instructions with REL and ACQ semantics,
2533      separating the parallel user code called in this parallel region
2534      from the serial user code called after this function returns. */
2535   __kmp_acquire_bootstrap_lock(&__kmp_forkjoin_lock);
2536 
2537 #if OMP_40_ENABLED
2538   if (!master_th->th.th_teams_microtask ||
2539       team->t.t_level > master_th->th.th_teams_level)
2540 #endif /* OMP_40_ENABLED */
2541   {
2542     /* Decrement our nested depth level */
2543     KMP_ATOMIC_DEC(&root->r.r_in_parallel);
2544   }
2545   KMP_DEBUG_ASSERT(root->r.r_in_parallel >= 0);
2546 
2547 #if OMPT_SUPPORT
2548   if (ompt_enabled.enabled) {
2549     ompt_task_info_t *task_info = __ompt_get_task_info_object(0);
2550     if (ompt_enabled.ompt_callback_implicit_task) {
2551       int ompt_team_size = team->t.t_nproc;
2552       ompt_callbacks.ompt_callback(ompt_callback_implicit_task)(
2553           ompt_scope_end, NULL, &(task_info->task_data), ompt_team_size,
2554           OMPT_CUR_TASK_INFO(master_th)->thread_num, ompt_task_implicit); // TODO: Can this be ompt_task_initial?
2555     }
2556 
2557     task_info->frame.exit_frame = ompt_data_none;
2558     task_info->task_data = ompt_data_none;
2559   }
2560 #endif
2561 
2562   KF_TRACE(10, ("__kmp_join_call1: T#%d, this_thread=%p team=%p\n", 0,
2563                 master_th, team));
2564   __kmp_pop_current_task_from_thread(master_th);
2565 
2566 #if OMP_40_ENABLED && KMP_AFFINITY_SUPPORTED
2567   // Restore master thread's partition.
2568   master_th->th.th_first_place = team->t.t_first_place;
2569   master_th->th.th_last_place = team->t.t_last_place;
2570 #endif /* OMP_40_ENABLED */
2571 #if OMP_50_ENABLED
2572   master_th->th.th_def_allocator = team->t.t_def_allocator;
2573 #endif
2574 
2575   updateHWFPControl(team);
2576 
2577   if (root->r.r_active != master_active)
2578     root->r.r_active = master_active;
2579 
2580   __kmp_free_team(root, team USE_NESTED_HOT_ARG(
2581                             master_th)); // this will free worker threads
2582 
2583   /* this race was fun to find. make sure the following is in the critical
2584      region otherwise assertions may fail occasionally since the old team may be
2585      reallocated and the hierarchy appears inconsistent. it is actually safe to
2586      run and won't cause any bugs, but will cause those assertion failures. it's
2587      only one deref&assign so might as well put this in the critical region */
2588   master_th->th.th_team = parent_team;
2589   master_th->th.th_team_nproc = parent_team->t.t_nproc;
2590   master_th->th.th_team_master = parent_team->t.t_threads[0];
2591   master_th->th.th_team_serialized = parent_team->t.t_serialized;
2592 
2593   /* restore serialized team, if need be */
2594   if (parent_team->t.t_serialized &&
2595       parent_team != master_th->th.th_serial_team &&
2596       parent_team != root->r.r_root_team) {
2597     __kmp_free_team(root,
2598                     master_th->th.th_serial_team USE_NESTED_HOT_ARG(NULL));
2599     master_th->th.th_serial_team = parent_team;
2600   }
2601 
2602   if (__kmp_tasking_mode != tskm_immediate_exec) {
2603     if (master_th->th.th_task_state_top >
2604         0) { // Restore task state from memo stack
2605       KMP_DEBUG_ASSERT(master_th->th.th_task_state_memo_stack);
2606       // Remember master's state if we re-use this nested hot team
2607       master_th->th.th_task_state_memo_stack[master_th->th.th_task_state_top] =
2608           master_th->th.th_task_state;
2609       --master_th->th.th_task_state_top; // pop
2610       // Now restore state at this level
2611       master_th->th.th_task_state =
2612           master_th->th
2613               .th_task_state_memo_stack[master_th->th.th_task_state_top];
2614     }
2615     // Copy the task team from the parent team to the master thread
2616     master_th->th.th_task_team =
2617         parent_team->t.t_task_team[master_th->th.th_task_state];
2618     KA_TRACE(20,
2619              ("__kmp_join_call: Master T#%d restoring task_team %p / team %p\n",
2620               __kmp_gtid_from_thread(master_th), master_th->th.th_task_team,
2621               parent_team));
2622   }
2623 
2624   // TODO: GEH - cannot do this assertion because root thread not set up as
2625   // executing
2626   // KMP_ASSERT( master_th->th.th_current_task->td_flags.executing == 0 );
2627   master_th->th.th_current_task->td_flags.executing = 1;
2628 
2629   __kmp_release_bootstrap_lock(&__kmp_forkjoin_lock);
2630 
2631 #if OMPT_SUPPORT
2632   if (ompt_enabled.enabled) {
2633     __kmp_join_ompt(gtid, master_th, parent_team, parallel_data, fork_context,
2634                     codeptr);
2635   }
2636 #endif
2637 
2638   KMP_MB();
2639   KA_TRACE(20, ("__kmp_join_call: exit T#%d\n", gtid));
2640 }
2641 
2642 /* Check whether we should push an internal control record onto the
2643    serial team stack.  If so, do it.  */
2644 void __kmp_save_internal_controls(kmp_info_t *thread) {
2645 
2646   if (thread->th.th_team != thread->th.th_serial_team) {
2647     return;
2648   }
2649   if (thread->th.th_team->t.t_serialized > 1) {
2650     int push = 0;
2651 
2652     if (thread->th.th_team->t.t_control_stack_top == NULL) {
2653       push = 1;
2654     } else {
2655       if (thread->th.th_team->t.t_control_stack_top->serial_nesting_level !=
2656           thread->th.th_team->t.t_serialized) {
2657         push = 1;
2658       }
2659     }
2660     if (push) { /* push a record on the serial team's stack */
2661       kmp_internal_control_t *control =
2662           (kmp_internal_control_t *)__kmp_allocate(
2663               sizeof(kmp_internal_control_t));
2664 
2665       copy_icvs(control, &thread->th.th_current_task->td_icvs);
2666 
2667       control->serial_nesting_level = thread->th.th_team->t.t_serialized;
2668 
2669       control->next = thread->th.th_team->t.t_control_stack_top;
2670       thread->th.th_team->t.t_control_stack_top = control;
2671     }
2672   }
2673 }
2674 
2675 /* Changes set_nproc */
2676 void __kmp_set_num_threads(int new_nth, int gtid) {
2677   kmp_info_t *thread;
2678   kmp_root_t *root;
2679 
2680   KF_TRACE(10, ("__kmp_set_num_threads: new __kmp_nth = %d\n", new_nth));
2681   KMP_DEBUG_ASSERT(__kmp_init_serial);
2682 
2683   if (new_nth < 1)
2684     new_nth = 1;
2685   else if (new_nth > __kmp_max_nth)
2686     new_nth = __kmp_max_nth;
2687 
2688   KMP_COUNT_VALUE(OMP_set_numthreads, new_nth);
2689   thread = __kmp_threads[gtid];
2690   if (thread->th.th_current_task->td_icvs.nproc == new_nth)
2691     return; // nothing to do
2692 
2693   __kmp_save_internal_controls(thread);
2694 
2695   set__nproc(thread, new_nth);
2696 
2697   // If this omp_set_num_threads() call will cause the hot team size to be
2698   // reduced (in the absence of a num_threads clause), then reduce it now,
2699   // rather than waiting for the next parallel region.
2700   root = thread->th.th_root;
2701   if (__kmp_init_parallel && (!root->r.r_active) &&
2702       (root->r.r_hot_team->t.t_nproc > new_nth)
2703 #if KMP_NESTED_HOT_TEAMS
2704       && __kmp_hot_teams_max_level && !__kmp_hot_teams_mode
2705 #endif
2706       ) {
2707     kmp_team_t *hot_team = root->r.r_hot_team;
2708     int f;
2709 
2710     __kmp_acquire_bootstrap_lock(&__kmp_forkjoin_lock);
2711 
2712     // Release the extra threads we don't need any more.
2713     for (f = new_nth; f < hot_team->t.t_nproc; f++) {
2714       KMP_DEBUG_ASSERT(hot_team->t.t_threads[f] != NULL);
2715       if (__kmp_tasking_mode != tskm_immediate_exec) {
2716         // When decreasing team size, threads no longer in the team should unref
2717         // task team.
2718         hot_team->t.t_threads[f]->th.th_task_team = NULL;
2719       }
2720       __kmp_free_thread(hot_team->t.t_threads[f]);
2721       hot_team->t.t_threads[f] = NULL;
2722     }
2723     hot_team->t.t_nproc = new_nth;
2724 #if KMP_NESTED_HOT_TEAMS
2725     if (thread->th.th_hot_teams) {
2726       KMP_DEBUG_ASSERT(hot_team == thread->th.th_hot_teams[0].hot_team);
2727       thread->th.th_hot_teams[0].hot_team_nth = new_nth;
2728     }
2729 #endif
2730 
2731     __kmp_release_bootstrap_lock(&__kmp_forkjoin_lock);
2732 
2733     // Update the t_nproc field in the threads that are still active.
2734     for (f = 0; f < new_nth; f++) {
2735       KMP_DEBUG_ASSERT(hot_team->t.t_threads[f] != NULL);
2736       hot_team->t.t_threads[f]->th.th_team_nproc = new_nth;
2737     }
2738     // Special flag in case omp_set_num_threads() call
2739     hot_team->t.t_size_changed = -1;
2740   }
2741 }
2742 
2743 /* Changes max_active_levels */
2744 void __kmp_set_max_active_levels(int gtid, int max_active_levels) {
2745   kmp_info_t *thread;
2746 
2747   KF_TRACE(10, ("__kmp_set_max_active_levels: new max_active_levels for thread "
2748                 "%d = (%d)\n",
2749                 gtid, max_active_levels));
2750   KMP_DEBUG_ASSERT(__kmp_init_serial);
2751 
2752   // validate max_active_levels
2753   if (max_active_levels < 0) {
2754     KMP_WARNING(ActiveLevelsNegative, max_active_levels);
2755     // We ignore this call if the user has specified a negative value.
2756     // The current setting won't be changed. The last valid setting will be
2757     // used. A warning will be issued (if warnings are allowed as controlled by
2758     // the KMP_WARNINGS env var).
2759     KF_TRACE(10, ("__kmp_set_max_active_levels: the call is ignored: new "
2760                   "max_active_levels for thread %d = (%d)\n",
2761                   gtid, max_active_levels));
2762     return;
2763   }
2764   if (max_active_levels <= KMP_MAX_ACTIVE_LEVELS_LIMIT) {
2765     // it's OK, the max_active_levels is within the valid range: [ 0;
2766     // KMP_MAX_ACTIVE_LEVELS_LIMIT ]
2767     // We allow a zero value. (implementation defined behavior)
2768   } else {
2769     KMP_WARNING(ActiveLevelsExceedLimit, max_active_levels,
2770                 KMP_MAX_ACTIVE_LEVELS_LIMIT);
2771     max_active_levels = KMP_MAX_ACTIVE_LEVELS_LIMIT;
2772     // Current upper limit is MAX_INT. (implementation defined behavior)
2773     // If the input exceeds the upper limit, we correct the input to be the
2774     // upper limit. (implementation defined behavior)
2775     // Actually, the flow should never get here until we use MAX_INT limit.
2776   }
2777   KF_TRACE(10, ("__kmp_set_max_active_levels: after validation: new "
2778                 "max_active_levels for thread %d = (%d)\n",
2779                 gtid, max_active_levels));
2780 
2781   thread = __kmp_threads[gtid];
2782 
2783   __kmp_save_internal_controls(thread);
2784 
2785   set__max_active_levels(thread, max_active_levels);
2786 }
2787 
2788 /* Gets max_active_levels */
2789 int __kmp_get_max_active_levels(int gtid) {
2790   kmp_info_t *thread;
2791 
2792   KF_TRACE(10, ("__kmp_get_max_active_levels: thread %d\n", gtid));
2793   KMP_DEBUG_ASSERT(__kmp_init_serial);
2794 
2795   thread = __kmp_threads[gtid];
2796   KMP_DEBUG_ASSERT(thread->th.th_current_task);
2797   KF_TRACE(10, ("__kmp_get_max_active_levels: thread %d, curtask=%p, "
2798                 "curtask_maxaclevel=%d\n",
2799                 gtid, thread->th.th_current_task,
2800                 thread->th.th_current_task->td_icvs.max_active_levels));
2801   return thread->th.th_current_task->td_icvs.max_active_levels;
2802 }
2803 
2804 /* Changes def_sched_var ICV values (run-time schedule kind and chunk) */
2805 void __kmp_set_schedule(int gtid, kmp_sched_t kind, int chunk) {
2806   kmp_info_t *thread;
2807   //    kmp_team_t *team;
2808 
2809   KF_TRACE(10, ("__kmp_set_schedule: new schedule for thread %d = (%d, %d)\n",
2810                 gtid, (int)kind, chunk));
2811   KMP_DEBUG_ASSERT(__kmp_init_serial);
2812 
2813   // Check if the kind parameter is valid, correct if needed.
2814   // Valid parameters should fit in one of two intervals - standard or extended:
2815   //       <lower>, <valid>, <upper_std>, <lower_ext>, <valid>, <upper>
2816   // 2008-01-25: 0,  1 - 4,       5,         100,     101 - 102, 103
2817   if (kind <= kmp_sched_lower || kind >= kmp_sched_upper ||
2818       (kind <= kmp_sched_lower_ext && kind >= kmp_sched_upper_std)) {
2819     // TODO: Hint needs attention in case we change the default schedule.
2820     __kmp_msg(kmp_ms_warning, KMP_MSG(ScheduleKindOutOfRange, kind),
2821               KMP_HNT(DefaultScheduleKindUsed, "static, no chunk"),
2822               __kmp_msg_null);
2823     kind = kmp_sched_default;
2824     chunk = 0; // ignore chunk value in case of bad kind
2825   }
2826 
2827   thread = __kmp_threads[gtid];
2828 
2829   __kmp_save_internal_controls(thread);
2830 
2831   if (kind < kmp_sched_upper_std) {
2832     if (kind == kmp_sched_static && chunk < KMP_DEFAULT_CHUNK) {
2833       // differ static chunked vs. unchunked:  chunk should be invalid to
2834       // indicate unchunked schedule (which is the default)
2835       thread->th.th_current_task->td_icvs.sched.r_sched_type = kmp_sch_static;
2836     } else {
2837       thread->th.th_current_task->td_icvs.sched.r_sched_type =
2838           __kmp_sch_map[kind - kmp_sched_lower - 1];
2839     }
2840   } else {
2841     //    __kmp_sch_map[ kind - kmp_sched_lower_ext + kmp_sched_upper_std -
2842     //    kmp_sched_lower - 2 ];
2843     thread->th.th_current_task->td_icvs.sched.r_sched_type =
2844         __kmp_sch_map[kind - kmp_sched_lower_ext + kmp_sched_upper_std -
2845                       kmp_sched_lower - 2];
2846   }
2847   if (kind == kmp_sched_auto || chunk < 1) {
2848     // ignore parameter chunk for schedule auto
2849     thread->th.th_current_task->td_icvs.sched.chunk = KMP_DEFAULT_CHUNK;
2850   } else {
2851     thread->th.th_current_task->td_icvs.sched.chunk = chunk;
2852   }
2853 }
2854 
2855 /* Gets def_sched_var ICV values */
2856 void __kmp_get_schedule(int gtid, kmp_sched_t *kind, int *chunk) {
2857   kmp_info_t *thread;
2858   enum sched_type th_type;
2859 
2860   KF_TRACE(10, ("__kmp_get_schedule: thread %d\n", gtid));
2861   KMP_DEBUG_ASSERT(__kmp_init_serial);
2862 
2863   thread = __kmp_threads[gtid];
2864 
2865   th_type = thread->th.th_current_task->td_icvs.sched.r_sched_type;
2866 
2867   switch (th_type) {
2868   case kmp_sch_static:
2869   case kmp_sch_static_greedy:
2870   case kmp_sch_static_balanced:
2871     *kind = kmp_sched_static;
2872     *chunk = 0; // chunk was not set, try to show this fact via zero value
2873     return;
2874   case kmp_sch_static_chunked:
2875     *kind = kmp_sched_static;
2876     break;
2877   case kmp_sch_dynamic_chunked:
2878     *kind = kmp_sched_dynamic;
2879     break;
2880   case kmp_sch_guided_chunked:
2881   case kmp_sch_guided_iterative_chunked:
2882   case kmp_sch_guided_analytical_chunked:
2883     *kind = kmp_sched_guided;
2884     break;
2885   case kmp_sch_auto:
2886     *kind = kmp_sched_auto;
2887     break;
2888   case kmp_sch_trapezoidal:
2889     *kind = kmp_sched_trapezoidal;
2890     break;
2891 #if KMP_STATIC_STEAL_ENABLED
2892   case kmp_sch_static_steal:
2893     *kind = kmp_sched_static_steal;
2894     break;
2895 #endif
2896   default:
2897     KMP_FATAL(UnknownSchedulingType, th_type);
2898   }
2899 
2900   *chunk = thread->th.th_current_task->td_icvs.sched.chunk;
2901 }
2902 
2903 int __kmp_get_ancestor_thread_num(int gtid, int level) {
2904 
2905   int ii, dd;
2906   kmp_team_t *team;
2907   kmp_info_t *thr;
2908 
2909   KF_TRACE(10, ("__kmp_get_ancestor_thread_num: thread %d %d\n", gtid, level));
2910   KMP_DEBUG_ASSERT(__kmp_init_serial);
2911 
2912   // validate level
2913   if (level == 0)
2914     return 0;
2915   if (level < 0)
2916     return -1;
2917   thr = __kmp_threads[gtid];
2918   team = thr->th.th_team;
2919   ii = team->t.t_level;
2920   if (level > ii)
2921     return -1;
2922 
2923 #if OMP_40_ENABLED
2924   if (thr->th.th_teams_microtask) {
2925     // AC: we are in teams region where multiple nested teams have same level
2926     int tlevel = thr->th.th_teams_level; // the level of the teams construct
2927     if (level <=
2928         tlevel) { // otherwise usual algorithm works (will not touch the teams)
2929       KMP_DEBUG_ASSERT(ii >= tlevel);
2930       // AC: As we need to pass by the teams league, we need to artificially
2931       // increase ii
2932       if (ii == tlevel) {
2933         ii += 2; // three teams have same level
2934       } else {
2935         ii++; // two teams have same level
2936       }
2937     }
2938   }
2939 #endif
2940 
2941   if (ii == level)
2942     return __kmp_tid_from_gtid(gtid);
2943 
2944   dd = team->t.t_serialized;
2945   level++;
2946   while (ii > level) {
2947     for (dd = team->t.t_serialized; (dd > 0) && (ii > level); dd--, ii--) {
2948     }
2949     if ((team->t.t_serialized) && (!dd)) {
2950       team = team->t.t_parent;
2951       continue;
2952     }
2953     if (ii > level) {
2954       team = team->t.t_parent;
2955       dd = team->t.t_serialized;
2956       ii--;
2957     }
2958   }
2959 
2960   return (dd > 1) ? (0) : (team->t.t_master_tid);
2961 }
2962 
2963 int __kmp_get_team_size(int gtid, int level) {
2964 
2965   int ii, dd;
2966   kmp_team_t *team;
2967   kmp_info_t *thr;
2968 
2969   KF_TRACE(10, ("__kmp_get_team_size: thread %d %d\n", gtid, level));
2970   KMP_DEBUG_ASSERT(__kmp_init_serial);
2971 
2972   // validate level
2973   if (level == 0)
2974     return 1;
2975   if (level < 0)
2976     return -1;
2977   thr = __kmp_threads[gtid];
2978   team = thr->th.th_team;
2979   ii = team->t.t_level;
2980   if (level > ii)
2981     return -1;
2982 
2983 #if OMP_40_ENABLED
2984   if (thr->th.th_teams_microtask) {
2985     // AC: we are in teams region where multiple nested teams have same level
2986     int tlevel = thr->th.th_teams_level; // the level of the teams construct
2987     if (level <=
2988         tlevel) { // otherwise usual algorithm works (will not touch the teams)
2989       KMP_DEBUG_ASSERT(ii >= tlevel);
2990       // AC: As we need to pass by the teams league, we need to artificially
2991       // increase ii
2992       if (ii == tlevel) {
2993         ii += 2; // three teams have same level
2994       } else {
2995         ii++; // two teams have same level
2996       }
2997     }
2998   }
2999 #endif
3000 
3001   while (ii > level) {
3002     for (dd = team->t.t_serialized; (dd > 0) && (ii > level); dd--, ii--) {
3003     }
3004     if (team->t.t_serialized && (!dd)) {
3005       team = team->t.t_parent;
3006       continue;
3007     }
3008     if (ii > level) {
3009       team = team->t.t_parent;
3010       ii--;
3011     }
3012   }
3013 
3014   return team->t.t_nproc;
3015 }
3016 
3017 kmp_r_sched_t __kmp_get_schedule_global() {
3018   // This routine created because pairs (__kmp_sched, __kmp_chunk) and
3019   // (__kmp_static, __kmp_guided) may be changed by kmp_set_defaults
3020   // independently. So one can get the updated schedule here.
3021 
3022   kmp_r_sched_t r_sched;
3023 
3024   // create schedule from 4 globals: __kmp_sched, __kmp_chunk, __kmp_static,
3025   // __kmp_guided. __kmp_sched should keep original value, so that user can set
3026   // KMP_SCHEDULE multiple times, and thus have different run-time schedules in
3027   // different roots (even in OMP 2.5)
3028   if (__kmp_sched == kmp_sch_static) {
3029     // replace STATIC with more detailed schedule (balanced or greedy)
3030     r_sched.r_sched_type = __kmp_static;
3031   } else if (__kmp_sched == kmp_sch_guided_chunked) {
3032     // replace GUIDED with more detailed schedule (iterative or analytical)
3033     r_sched.r_sched_type = __kmp_guided;
3034   } else { // (STATIC_CHUNKED), or (DYNAMIC_CHUNKED), or other
3035     r_sched.r_sched_type = __kmp_sched;
3036   }
3037 
3038   if (__kmp_chunk < KMP_DEFAULT_CHUNK) {
3039     // __kmp_chunk may be wrong here (if it was not ever set)
3040     r_sched.chunk = KMP_DEFAULT_CHUNK;
3041   } else {
3042     r_sched.chunk = __kmp_chunk;
3043   }
3044 
3045   return r_sched;
3046 }
3047 
3048 /* Allocate (realloc == FALSE) * or reallocate (realloc == TRUE)
3049    at least argc number of *t_argv entries for the requested team. */
3050 static void __kmp_alloc_argv_entries(int argc, kmp_team_t *team, int realloc) {
3051 
3052   KMP_DEBUG_ASSERT(team);
3053   if (!realloc || argc > team->t.t_max_argc) {
3054 
3055     KA_TRACE(100, ("__kmp_alloc_argv_entries: team %d: needed entries=%d, "
3056                    "current entries=%d\n",
3057                    team->t.t_id, argc, (realloc) ? team->t.t_max_argc : 0));
3058     /* if previously allocated heap space for args, free them */
3059     if (realloc && team->t.t_argv != &team->t.t_inline_argv[0])
3060       __kmp_free((void *)team->t.t_argv);
3061 
3062     if (argc <= KMP_INLINE_ARGV_ENTRIES) {
3063       /* use unused space in the cache line for arguments */
3064       team->t.t_max_argc = KMP_INLINE_ARGV_ENTRIES;
3065       KA_TRACE(100, ("__kmp_alloc_argv_entries: team %d: inline allocate %d "
3066                      "argv entries\n",
3067                      team->t.t_id, team->t.t_max_argc));
3068       team->t.t_argv = &team->t.t_inline_argv[0];
3069       if (__kmp_storage_map) {
3070         __kmp_print_storage_map_gtid(
3071             -1, &team->t.t_inline_argv[0],
3072             &team->t.t_inline_argv[KMP_INLINE_ARGV_ENTRIES],
3073             (sizeof(void *) * KMP_INLINE_ARGV_ENTRIES), "team_%d.t_inline_argv",
3074             team->t.t_id);
3075       }
3076     } else {
3077       /* allocate space for arguments in the heap */
3078       team->t.t_max_argc = (argc <= (KMP_MIN_MALLOC_ARGV_ENTRIES >> 1))
3079                                ? KMP_MIN_MALLOC_ARGV_ENTRIES
3080                                : 2 * argc;
3081       KA_TRACE(100, ("__kmp_alloc_argv_entries: team %d: dynamic allocate %d "
3082                      "argv entries\n",
3083                      team->t.t_id, team->t.t_max_argc));
3084       team->t.t_argv =
3085           (void **)__kmp_page_allocate(sizeof(void *) * team->t.t_max_argc);
3086       if (__kmp_storage_map) {
3087         __kmp_print_storage_map_gtid(-1, &team->t.t_argv[0],
3088                                      &team->t.t_argv[team->t.t_max_argc],
3089                                      sizeof(void *) * team->t.t_max_argc,
3090                                      "team_%d.t_argv", team->t.t_id);
3091       }
3092     }
3093   }
3094 }
3095 
3096 static void __kmp_allocate_team_arrays(kmp_team_t *team, int max_nth) {
3097   int i;
3098   int num_disp_buff = max_nth > 1 ? __kmp_dispatch_num_buffers : 2;
3099   team->t.t_threads =
3100       (kmp_info_t **)__kmp_allocate(sizeof(kmp_info_t *) * max_nth);
3101   team->t.t_disp_buffer = (dispatch_shared_info_t *)__kmp_allocate(
3102       sizeof(dispatch_shared_info_t) * num_disp_buff);
3103   team->t.t_dispatch =
3104       (kmp_disp_t *)__kmp_allocate(sizeof(kmp_disp_t) * max_nth);
3105   team->t.t_implicit_task_taskdata =
3106       (kmp_taskdata_t *)__kmp_allocate(sizeof(kmp_taskdata_t) * max_nth);
3107   team->t.t_max_nproc = max_nth;
3108 
3109   /* setup dispatch buffers */
3110   for (i = 0; i < num_disp_buff; ++i) {
3111     team->t.t_disp_buffer[i].buffer_index = i;
3112 #if OMP_45_ENABLED
3113     team->t.t_disp_buffer[i].doacross_buf_idx = i;
3114 #endif
3115   }
3116 }
3117 
3118 static void __kmp_free_team_arrays(kmp_team_t *team) {
3119   /* Note: this does not free the threads in t_threads (__kmp_free_threads) */
3120   int i;
3121   for (i = 0; i < team->t.t_max_nproc; ++i) {
3122     if (team->t.t_dispatch[i].th_disp_buffer != NULL) {
3123       __kmp_free(team->t.t_dispatch[i].th_disp_buffer);
3124       team->t.t_dispatch[i].th_disp_buffer = NULL;
3125     }
3126   }
3127 #if KMP_USE_HIER_SCHED
3128   __kmp_dispatch_free_hierarchies(team);
3129 #endif
3130   __kmp_free(team->t.t_threads);
3131   __kmp_free(team->t.t_disp_buffer);
3132   __kmp_free(team->t.t_dispatch);
3133   __kmp_free(team->t.t_implicit_task_taskdata);
3134   team->t.t_threads = NULL;
3135   team->t.t_disp_buffer = NULL;
3136   team->t.t_dispatch = NULL;
3137   team->t.t_implicit_task_taskdata = 0;
3138 }
3139 
3140 static void __kmp_reallocate_team_arrays(kmp_team_t *team, int max_nth) {
3141   kmp_info_t **oldThreads = team->t.t_threads;
3142 
3143   __kmp_free(team->t.t_disp_buffer);
3144   __kmp_free(team->t.t_dispatch);
3145   __kmp_free(team->t.t_implicit_task_taskdata);
3146   __kmp_allocate_team_arrays(team, max_nth);
3147 
3148   KMP_MEMCPY(team->t.t_threads, oldThreads,
3149              team->t.t_nproc * sizeof(kmp_info_t *));
3150 
3151   __kmp_free(oldThreads);
3152 }
3153 
3154 static kmp_internal_control_t __kmp_get_global_icvs(void) {
3155 
3156   kmp_r_sched_t r_sched =
3157       __kmp_get_schedule_global(); // get current state of scheduling globals
3158 
3159 #if OMP_40_ENABLED
3160   KMP_DEBUG_ASSERT(__kmp_nested_proc_bind.used > 0);
3161 #endif /* OMP_40_ENABLED */
3162 
3163   kmp_internal_control_t g_icvs = {
3164     0, // int serial_nesting_level; //corresponds to value of th_team_serialized
3165     (kmp_int8)__kmp_global.g.g_dynamic, // internal control for dynamic
3166     // adjustment of threads (per thread)
3167     (kmp_int8)__kmp_env_blocktime, // int bt_set; //internal control for
3168     // whether blocktime is explicitly set
3169     __kmp_dflt_blocktime, // int blocktime; //internal control for blocktime
3170 #if KMP_USE_MONITOR
3171     __kmp_bt_intervals, // int bt_intervals; //internal control for blocktime
3172 // intervals
3173 #endif
3174     __kmp_dflt_team_nth, // int nproc; //internal control for # of threads for
3175     // next parallel region (per thread)
3176     // (use a max ub on value if __kmp_parallel_initialize not called yet)
3177     __kmp_cg_max_nth, // int thread_limit;
3178     __kmp_dflt_max_active_levels, // int max_active_levels; //internal control
3179     // for max_active_levels
3180     r_sched, // kmp_r_sched_t sched; //internal control for runtime schedule
3181 // {sched,chunk} pair
3182 #if OMP_40_ENABLED
3183     __kmp_nested_proc_bind.bind_types[0],
3184     __kmp_default_device,
3185 #endif /* OMP_40_ENABLED */
3186     NULL // struct kmp_internal_control *next;
3187   };
3188 
3189   return g_icvs;
3190 }
3191 
3192 static kmp_internal_control_t __kmp_get_x_global_icvs(const kmp_team_t *team) {
3193 
3194   kmp_internal_control_t gx_icvs;
3195   gx_icvs.serial_nesting_level =
3196       0; // probably =team->t.t_serial like in save_inter_controls
3197   copy_icvs(&gx_icvs, &team->t.t_threads[0]->th.th_current_task->td_icvs);
3198   gx_icvs.next = NULL;
3199 
3200   return gx_icvs;
3201 }
3202 
3203 static void __kmp_initialize_root(kmp_root_t *root) {
3204   int f;
3205   kmp_team_t *root_team;
3206   kmp_team_t *hot_team;
3207   int hot_team_max_nth;
3208   kmp_r_sched_t r_sched =
3209       __kmp_get_schedule_global(); // get current state of scheduling globals
3210   kmp_internal_control_t r_icvs = __kmp_get_global_icvs();
3211   KMP_DEBUG_ASSERT(root);
3212   KMP_ASSERT(!root->r.r_begin);
3213 
3214   /* setup the root state structure */
3215   __kmp_init_lock(&root->r.r_begin_lock);
3216   root->r.r_begin = FALSE;
3217   root->r.r_active = FALSE;
3218   root->r.r_in_parallel = 0;
3219   root->r.r_blocktime = __kmp_dflt_blocktime;
3220 
3221   /* setup the root team for this task */
3222   /* allocate the root team structure */
3223   KF_TRACE(10, ("__kmp_initialize_root: before root_team\n"));
3224 
3225   root_team =
3226       __kmp_allocate_team(root,
3227                           1, // new_nproc
3228                           1, // max_nproc
3229 #if OMPT_SUPPORT
3230                           ompt_data_none, // root parallel id
3231 #endif
3232 #if OMP_40_ENABLED
3233                           __kmp_nested_proc_bind.bind_types[0],
3234 #endif
3235                           &r_icvs,
3236                           0 // argc
3237                           USE_NESTED_HOT_ARG(NULL) // master thread is unknown
3238                           );
3239 #if USE_DEBUGGER
3240   // Non-NULL value should be assigned to make the debugger display the root
3241   // team.
3242   TCW_SYNC_PTR(root_team->t.t_pkfn, (microtask_t)(~0));
3243 #endif
3244 
3245   KF_TRACE(10, ("__kmp_initialize_root: after root_team = %p\n", root_team));
3246 
3247   root->r.r_root_team = root_team;
3248   root_team->t.t_control_stack_top = NULL;
3249 
3250   /* initialize root team */
3251   root_team->t.t_threads[0] = NULL;
3252   root_team->t.t_nproc = 1;
3253   root_team->t.t_serialized = 1;
3254   // TODO???: root_team->t.t_max_active_levels = __kmp_dflt_max_active_levels;
3255   root_team->t.t_sched.sched = r_sched.sched;
3256   KA_TRACE(
3257       20,
3258       ("__kmp_initialize_root: init root team %d arrived: join=%u, plain=%u\n",
3259        root_team->t.t_id, KMP_INIT_BARRIER_STATE, KMP_INIT_BARRIER_STATE));
3260 
3261   /* setup the  hot team for this task */
3262   /* allocate the hot team structure */
3263   KF_TRACE(10, ("__kmp_initialize_root: before hot_team\n"));
3264 
3265   hot_team =
3266       __kmp_allocate_team(root,
3267                           1, // new_nproc
3268                           __kmp_dflt_team_nth_ub * 2, // max_nproc
3269 #if OMPT_SUPPORT
3270                           ompt_data_none, // root parallel id
3271 #endif
3272 #if OMP_40_ENABLED
3273                           __kmp_nested_proc_bind.bind_types[0],
3274 #endif
3275                           &r_icvs,
3276                           0 // argc
3277                           USE_NESTED_HOT_ARG(NULL) // master thread is unknown
3278                           );
3279   KF_TRACE(10, ("__kmp_initialize_root: after hot_team = %p\n", hot_team));
3280 
3281   root->r.r_hot_team = hot_team;
3282   root_team->t.t_control_stack_top = NULL;
3283 
3284   /* first-time initialization */
3285   hot_team->t.t_parent = root_team;
3286 
3287   /* initialize hot team */
3288   hot_team_max_nth = hot_team->t.t_max_nproc;
3289   for (f = 0; f < hot_team_max_nth; ++f) {
3290     hot_team->t.t_threads[f] = NULL;
3291   }
3292   hot_team->t.t_nproc = 1;
3293   // TODO???: hot_team->t.t_max_active_levels = __kmp_dflt_max_active_levels;
3294   hot_team->t.t_sched.sched = r_sched.sched;
3295   hot_team->t.t_size_changed = 0;
3296 }
3297 
3298 #ifdef KMP_DEBUG
3299 
3300 typedef struct kmp_team_list_item {
3301   kmp_team_p const *entry;
3302   struct kmp_team_list_item *next;
3303 } kmp_team_list_item_t;
3304 typedef kmp_team_list_item_t *kmp_team_list_t;
3305 
3306 static void __kmp_print_structure_team_accum( // Add team to list of teams.
3307     kmp_team_list_t list, // List of teams.
3308     kmp_team_p const *team // Team to add.
3309     ) {
3310 
3311   // List must terminate with item where both entry and next are NULL.
3312   // Team is added to the list only once.
3313   // List is sorted in ascending order by team id.
3314   // Team id is *not* a key.
3315 
3316   kmp_team_list_t l;
3317 
3318   KMP_DEBUG_ASSERT(list != NULL);
3319   if (team == NULL) {
3320     return;
3321   }
3322 
3323   __kmp_print_structure_team_accum(list, team->t.t_parent);
3324   __kmp_print_structure_team_accum(list, team->t.t_next_pool);
3325 
3326   // Search list for the team.
3327   l = list;
3328   while (l->next != NULL && l->entry != team) {
3329     l = l->next;
3330   }
3331   if (l->next != NULL) {
3332     return; // Team has been added before, exit.
3333   }
3334 
3335   // Team is not found. Search list again for insertion point.
3336   l = list;
3337   while (l->next != NULL && l->entry->t.t_id <= team->t.t_id) {
3338     l = l->next;
3339   }
3340 
3341   // Insert team.
3342   {
3343     kmp_team_list_item_t *item = (kmp_team_list_item_t *)KMP_INTERNAL_MALLOC(
3344         sizeof(kmp_team_list_item_t));
3345     *item = *l;
3346     l->entry = team;
3347     l->next = item;
3348   }
3349 }
3350 
3351 static void __kmp_print_structure_team(char const *title, kmp_team_p const *team
3352 
3353                                        ) {
3354   __kmp_printf("%s", title);
3355   if (team != NULL) {
3356     __kmp_printf("%2x %p\n", team->t.t_id, team);
3357   } else {
3358     __kmp_printf(" - (nil)\n");
3359   }
3360 }
3361 
3362 static void __kmp_print_structure_thread(char const *title,
3363                                          kmp_info_p const *thread) {
3364   __kmp_printf("%s", title);
3365   if (thread != NULL) {
3366     __kmp_printf("%2d %p\n", thread->th.th_info.ds.ds_gtid, thread);
3367   } else {
3368     __kmp_printf(" - (nil)\n");
3369   }
3370 }
3371 
3372 void __kmp_print_structure(void) {
3373 
3374   kmp_team_list_t list;
3375 
3376   // Initialize list of teams.
3377   list =
3378       (kmp_team_list_item_t *)KMP_INTERNAL_MALLOC(sizeof(kmp_team_list_item_t));
3379   list->entry = NULL;
3380   list->next = NULL;
3381 
3382   __kmp_printf("\n------------------------------\nGlobal Thread "
3383                "Table\n------------------------------\n");
3384   {
3385     int gtid;
3386     for (gtid = 0; gtid < __kmp_threads_capacity; ++gtid) {
3387       __kmp_printf("%2d", gtid);
3388       if (__kmp_threads != NULL) {
3389         __kmp_printf(" %p", __kmp_threads[gtid]);
3390       }
3391       if (__kmp_root != NULL) {
3392         __kmp_printf(" %p", __kmp_root[gtid]);
3393       }
3394       __kmp_printf("\n");
3395     }
3396   }
3397 
3398   // Print out __kmp_threads array.
3399   __kmp_printf("\n------------------------------\nThreads\n--------------------"
3400                "----------\n");
3401   if (__kmp_threads != NULL) {
3402     int gtid;
3403     for (gtid = 0; gtid < __kmp_threads_capacity; ++gtid) {
3404       kmp_info_t const *thread = __kmp_threads[gtid];
3405       if (thread != NULL) {
3406         __kmp_printf("GTID %2d %p:\n", gtid, thread);
3407         __kmp_printf("    Our Root:        %p\n", thread->th.th_root);
3408         __kmp_print_structure_team("    Our Team:     ", thread->th.th_team);
3409         __kmp_print_structure_team("    Serial Team:  ",
3410                                    thread->th.th_serial_team);
3411         __kmp_printf("    Threads:      %2d\n", thread->th.th_team_nproc);
3412         __kmp_print_structure_thread("    Master:       ",
3413                                      thread->th.th_team_master);
3414         __kmp_printf("    Serialized?:  %2d\n", thread->th.th_team_serialized);
3415         __kmp_printf("    Set NProc:    %2d\n", thread->th.th_set_nproc);
3416 #if OMP_40_ENABLED
3417         __kmp_printf("    Set Proc Bind: %2d\n", thread->th.th_set_proc_bind);
3418 #endif
3419         __kmp_print_structure_thread("    Next in pool: ",
3420                                      thread->th.th_next_pool);
3421         __kmp_printf("\n");
3422         __kmp_print_structure_team_accum(list, thread->th.th_team);
3423         __kmp_print_structure_team_accum(list, thread->th.th_serial_team);
3424       }
3425     }
3426   } else {
3427     __kmp_printf("Threads array is not allocated.\n");
3428   }
3429 
3430   // Print out __kmp_root array.
3431   __kmp_printf("\n------------------------------\nUbers\n----------------------"
3432                "--------\n");
3433   if (__kmp_root != NULL) {
3434     int gtid;
3435     for (gtid = 0; gtid < __kmp_threads_capacity; ++gtid) {
3436       kmp_root_t const *root = __kmp_root[gtid];
3437       if (root != NULL) {
3438         __kmp_printf("GTID %2d %p:\n", gtid, root);
3439         __kmp_print_structure_team("    Root Team:    ", root->r.r_root_team);
3440         __kmp_print_structure_team("    Hot Team:     ", root->r.r_hot_team);
3441         __kmp_print_structure_thread("    Uber Thread:  ",
3442                                      root->r.r_uber_thread);
3443         __kmp_printf("    Active?:      %2d\n", root->r.r_active);
3444         __kmp_printf("    In Parallel:  %2d\n",
3445                      KMP_ATOMIC_LD_RLX(&root->r.r_in_parallel));
3446         __kmp_printf("\n");
3447         __kmp_print_structure_team_accum(list, root->r.r_root_team);
3448         __kmp_print_structure_team_accum(list, root->r.r_hot_team);
3449       }
3450     }
3451   } else {
3452     __kmp_printf("Ubers array is not allocated.\n");
3453   }
3454 
3455   __kmp_printf("\n------------------------------\nTeams\n----------------------"
3456                "--------\n");
3457   while (list->next != NULL) {
3458     kmp_team_p const *team = list->entry;
3459     int i;
3460     __kmp_printf("Team %2x %p:\n", team->t.t_id, team);
3461     __kmp_print_structure_team("    Parent Team:      ", team->t.t_parent);
3462     __kmp_printf("    Master TID:       %2d\n", team->t.t_master_tid);
3463     __kmp_printf("    Max threads:      %2d\n", team->t.t_max_nproc);
3464     __kmp_printf("    Levels of serial: %2d\n", team->t.t_serialized);
3465     __kmp_printf("    Number threads:   %2d\n", team->t.t_nproc);
3466     for (i = 0; i < team->t.t_nproc; ++i) {
3467       __kmp_printf("    Thread %2d:      ", i);
3468       __kmp_print_structure_thread("", team->t.t_threads[i]);
3469     }
3470     __kmp_print_structure_team("    Next in pool:     ", team->t.t_next_pool);
3471     __kmp_printf("\n");
3472     list = list->next;
3473   }
3474 
3475   // Print out __kmp_thread_pool and __kmp_team_pool.
3476   __kmp_printf("\n------------------------------\nPools\n----------------------"
3477                "--------\n");
3478   __kmp_print_structure_thread("Thread pool:          ",
3479                                CCAST(kmp_info_t *, __kmp_thread_pool));
3480   __kmp_print_structure_team("Team pool:            ",
3481                              CCAST(kmp_team_t *, __kmp_team_pool));
3482   __kmp_printf("\n");
3483 
3484   // Free team list.
3485   while (list != NULL) {
3486     kmp_team_list_item_t *item = list;
3487     list = list->next;
3488     KMP_INTERNAL_FREE(item);
3489   }
3490 }
3491 
3492 #endif
3493 
3494 //---------------------------------------------------------------------------
3495 //  Stuff for per-thread fast random number generator
3496 //  Table of primes
3497 static const unsigned __kmp_primes[] = {
3498     0x9e3779b1, 0xffe6cc59, 0x2109f6dd, 0x43977ab5, 0xba5703f5, 0xb495a877,
3499     0xe1626741, 0x79695e6b, 0xbc98c09f, 0xd5bee2b3, 0x287488f9, 0x3af18231,
3500     0x9677cd4d, 0xbe3a6929, 0xadc6a877, 0xdcf0674b, 0xbe4d6fe9, 0x5f15e201,
3501     0x99afc3fd, 0xf3f16801, 0xe222cfff, 0x24ba5fdb, 0x0620452d, 0x79f149e3,
3502     0xc8b93f49, 0x972702cd, 0xb07dd827, 0x6c97d5ed, 0x085a3d61, 0x46eb5ea7,
3503     0x3d9910ed, 0x2e687b5b, 0x29609227, 0x6eb081f1, 0x0954c4e1, 0x9d114db9,
3504     0x542acfa9, 0xb3e6bd7b, 0x0742d917, 0xe9f3ffa7, 0x54581edb, 0xf2480f45,
3505     0x0bb9288f, 0xef1affc7, 0x85fa0ca7, 0x3ccc14db, 0xe6baf34b, 0x343377f7,
3506     0x5ca19031, 0xe6d9293b, 0xf0a9f391, 0x5d2e980b, 0xfc411073, 0xc3749363,
3507     0xb892d829, 0x3549366b, 0x629750ad, 0xb98294e5, 0x892d9483, 0xc235baf3,
3508     0x3d2402a3, 0x6bdef3c9, 0xbec333cd, 0x40c9520f};
3509 
3510 //---------------------------------------------------------------------------
3511 //  __kmp_get_random: Get a random number using a linear congruential method.
3512 unsigned short __kmp_get_random(kmp_info_t *thread) {
3513   unsigned x = thread->th.th_x;
3514   unsigned short r = x >> 16;
3515 
3516   thread->th.th_x = x * thread->th.th_a + 1;
3517 
3518   KA_TRACE(30, ("__kmp_get_random: THREAD: %d, RETURN: %u\n",
3519                 thread->th.th_info.ds.ds_tid, r));
3520 
3521   return r;
3522 }
3523 //--------------------------------------------------------
3524 // __kmp_init_random: Initialize a random number generator
3525 void __kmp_init_random(kmp_info_t *thread) {
3526   unsigned seed = thread->th.th_info.ds.ds_tid;
3527 
3528   thread->th.th_a =
3529       __kmp_primes[seed % (sizeof(__kmp_primes) / sizeof(__kmp_primes[0]))];
3530   thread->th.th_x = (seed + 1) * thread->th.th_a + 1;
3531   KA_TRACE(30,
3532            ("__kmp_init_random: THREAD: %u; A: %u\n", seed, thread->th.th_a));
3533 }
3534 
3535 #if KMP_OS_WINDOWS
3536 /* reclaim array entries for root threads that are already dead, returns number
3537  * reclaimed */
3538 static int __kmp_reclaim_dead_roots(void) {
3539   int i, r = 0;
3540 
3541   for (i = 0; i < __kmp_threads_capacity; ++i) {
3542     if (KMP_UBER_GTID(i) &&
3543         !__kmp_still_running((kmp_info_t *)TCR_SYNC_PTR(__kmp_threads[i])) &&
3544         !__kmp_root[i]
3545              ->r.r_active) { // AC: reclaim only roots died in non-active state
3546       r += __kmp_unregister_root_other_thread(i);
3547     }
3548   }
3549   return r;
3550 }
3551 #endif
3552 
3553 /* This function attempts to create free entries in __kmp_threads and
3554    __kmp_root, and returns the number of free entries generated.
3555 
3556    For Windows* OS static library, the first mechanism used is to reclaim array
3557    entries for root threads that are already dead.
3558 
3559    On all platforms, expansion is attempted on the arrays __kmp_threads_ and
3560    __kmp_root, with appropriate update to __kmp_threads_capacity. Array
3561    capacity is increased by doubling with clipping to __kmp_tp_capacity, if
3562    threadprivate cache array has been created. Synchronization with
3563    __kmpc_threadprivate_cached is done using __kmp_tp_cached_lock.
3564 
3565    After any dead root reclamation, if the clipping value allows array expansion
3566    to result in the generation of a total of nNeed free slots, the function does
3567    that expansion. If not, nothing is done beyond the possible initial root
3568    thread reclamation.
3569 
3570    If any argument is negative, the behavior is undefined. */
3571 static int __kmp_expand_threads(int nNeed) {
3572   int added = 0;
3573   int minimumRequiredCapacity;
3574   int newCapacity;
3575   kmp_info_t **newThreads;
3576   kmp_root_t **newRoot;
3577 
3578 // All calls to __kmp_expand_threads should be under __kmp_forkjoin_lock, so
3579 // resizing __kmp_threads does not need additional protection if foreign
3580 // threads are present
3581 
3582 #if KMP_OS_WINDOWS && !KMP_DYNAMIC_LIB
3583   /* only for Windows static library */
3584   /* reclaim array entries for root threads that are already dead */
3585   added = __kmp_reclaim_dead_roots();
3586 
3587   if (nNeed) {
3588     nNeed -= added;
3589     if (nNeed < 0)
3590       nNeed = 0;
3591   }
3592 #endif
3593   if (nNeed <= 0)
3594     return added;
3595 
3596   // Note that __kmp_threads_capacity is not bounded by __kmp_max_nth. If
3597   // __kmp_max_nth is set to some value less than __kmp_sys_max_nth by the
3598   // user via KMP_DEVICE_THREAD_LIMIT, then __kmp_threads_capacity may become
3599   // > __kmp_max_nth in one of two ways:
3600   //
3601   // 1) The initialization thread (gtid = 0) exits.  __kmp_threads[0]
3602   //    may not be resused by another thread, so we may need to increase
3603   //    __kmp_threads_capacity to __kmp_max_nth + 1.
3604   //
3605   // 2) New foreign root(s) are encountered.  We always register new foreign
3606   //    roots. This may cause a smaller # of threads to be allocated at
3607   //    subsequent parallel regions, but the worker threads hang around (and
3608   //    eventually go to sleep) and need slots in the __kmp_threads[] array.
3609   //
3610   // Anyway, that is the reason for moving the check to see if
3611   // __kmp_max_nth was exceeded into __kmp_reserve_threads()
3612   // instead of having it performed here. -BB
3613 
3614   KMP_DEBUG_ASSERT(__kmp_sys_max_nth >= __kmp_threads_capacity);
3615 
3616   /* compute expansion headroom to check if we can expand */
3617   if (__kmp_sys_max_nth - __kmp_threads_capacity < nNeed) {
3618     /* possible expansion too small -- give up */
3619     return added;
3620   }
3621   minimumRequiredCapacity = __kmp_threads_capacity + nNeed;
3622 
3623   newCapacity = __kmp_threads_capacity;
3624   do {
3625     newCapacity = newCapacity <= (__kmp_sys_max_nth >> 1) ? (newCapacity << 1)
3626                                                           : __kmp_sys_max_nth;
3627   } while (newCapacity < minimumRequiredCapacity);
3628   newThreads = (kmp_info_t **)__kmp_allocate(
3629       (sizeof(kmp_info_t *) + sizeof(kmp_root_t *)) * newCapacity + CACHE_LINE);
3630   newRoot =
3631       (kmp_root_t **)((char *)newThreads + sizeof(kmp_info_t *) * newCapacity);
3632   KMP_MEMCPY(newThreads, __kmp_threads,
3633              __kmp_threads_capacity * sizeof(kmp_info_t *));
3634   KMP_MEMCPY(newRoot, __kmp_root,
3635              __kmp_threads_capacity * sizeof(kmp_root_t *));
3636 
3637   kmp_info_t **temp_threads = __kmp_threads;
3638   *(kmp_info_t * *volatile *)&__kmp_threads = newThreads;
3639   *(kmp_root_t * *volatile *)&__kmp_root = newRoot;
3640   __kmp_free(temp_threads);
3641   added += newCapacity - __kmp_threads_capacity;
3642   *(volatile int *)&__kmp_threads_capacity = newCapacity;
3643 
3644   if (newCapacity > __kmp_tp_capacity) {
3645     __kmp_acquire_bootstrap_lock(&__kmp_tp_cached_lock);
3646     if (__kmp_tp_cached && newCapacity > __kmp_tp_capacity) {
3647       __kmp_threadprivate_resize_cache(newCapacity);
3648     } else { // increase __kmp_tp_capacity to correspond with kmp_threads size
3649       *(volatile int *)&__kmp_tp_capacity = newCapacity;
3650     }
3651     __kmp_release_bootstrap_lock(&__kmp_tp_cached_lock);
3652   }
3653 
3654   return added;
3655 }
3656 
3657 /* Register the current thread as a root thread and obtain our gtid. We must
3658    have the __kmp_initz_lock held at this point. Argument TRUE only if are the
3659    thread that calls from __kmp_do_serial_initialize() */
3660 int __kmp_register_root(int initial_thread) {
3661   kmp_info_t *root_thread;
3662   kmp_root_t *root;
3663   int gtid;
3664   int capacity;
3665   __kmp_acquire_bootstrap_lock(&__kmp_forkjoin_lock);
3666   KA_TRACE(20, ("__kmp_register_root: entered\n"));
3667   KMP_MB();
3668 
3669   /* 2007-03-02:
3670      If initial thread did not invoke OpenMP RTL yet, and this thread is not an
3671      initial one, "__kmp_all_nth >= __kmp_threads_capacity" condition does not
3672      work as expected -- it may return false (that means there is at least one
3673      empty slot in __kmp_threads array), but it is possible the only free slot
3674      is #0, which is reserved for initial thread and so cannot be used for this
3675      one. Following code workarounds this bug.
3676 
3677      However, right solution seems to be not reserving slot #0 for initial
3678      thread because:
3679      (1) there is no magic in slot #0,
3680      (2) we cannot detect initial thread reliably (the first thread which does
3681         serial initialization may be not a real initial thread).
3682   */
3683   capacity = __kmp_threads_capacity;
3684   if (!initial_thread && TCR_PTR(__kmp_threads[0]) == NULL) {
3685     --capacity;
3686   }
3687 
3688   /* see if there are too many threads */
3689   if (__kmp_all_nth >= capacity && !__kmp_expand_threads(1)) {
3690     if (__kmp_tp_cached) {
3691       __kmp_fatal(KMP_MSG(CantRegisterNewThread),
3692                   KMP_HNT(Set_ALL_THREADPRIVATE, __kmp_tp_capacity),
3693                   KMP_HNT(PossibleSystemLimitOnThreads), __kmp_msg_null);
3694     } else {
3695       __kmp_fatal(KMP_MSG(CantRegisterNewThread), KMP_HNT(SystemLimitOnThreads),
3696                   __kmp_msg_null);
3697     }
3698   }
3699 
3700   /* find an available thread slot */
3701   /* Don't reassign the zero slot since we need that to only be used by initial
3702      thread */
3703   for (gtid = (initial_thread ? 0 : 1); TCR_PTR(__kmp_threads[gtid]) != NULL;
3704        gtid++)
3705     ;
3706   KA_TRACE(1,
3707            ("__kmp_register_root: found slot in threads array: T#%d\n", gtid));
3708   KMP_ASSERT(gtid < __kmp_threads_capacity);
3709 
3710   /* update global accounting */
3711   __kmp_all_nth++;
3712   TCW_4(__kmp_nth, __kmp_nth + 1);
3713 
3714   // if __kmp_adjust_gtid_mode is set, then we use method #1 (sp search) for low
3715   // numbers of procs, and method #2 (keyed API call) for higher numbers.
3716   if (__kmp_adjust_gtid_mode) {
3717     if (__kmp_all_nth >= __kmp_tls_gtid_min) {
3718       if (TCR_4(__kmp_gtid_mode) != 2) {
3719         TCW_4(__kmp_gtid_mode, 2);
3720       }
3721     } else {
3722       if (TCR_4(__kmp_gtid_mode) != 1) {
3723         TCW_4(__kmp_gtid_mode, 1);
3724       }
3725     }
3726   }
3727 
3728 #ifdef KMP_ADJUST_BLOCKTIME
3729   /* Adjust blocktime to zero if necessary            */
3730   /* Middle initialization might not have occurred yet */
3731   if (!__kmp_env_blocktime && (__kmp_avail_proc > 0)) {
3732     if (__kmp_nth > __kmp_avail_proc) {
3733       __kmp_zero_bt = TRUE;
3734     }
3735   }
3736 #endif /* KMP_ADJUST_BLOCKTIME */
3737 
3738   /* setup this new hierarchy */
3739   if (!(root = __kmp_root[gtid])) {
3740     root = __kmp_root[gtid] = (kmp_root_t *)__kmp_allocate(sizeof(kmp_root_t));
3741     KMP_DEBUG_ASSERT(!root->r.r_root_team);
3742   }
3743 
3744 #if KMP_STATS_ENABLED
3745   // Initialize stats as soon as possible (right after gtid assignment).
3746   __kmp_stats_thread_ptr = __kmp_stats_list->push_back(gtid);
3747   __kmp_stats_thread_ptr->startLife();
3748   KMP_SET_THREAD_STATE(SERIAL_REGION);
3749   KMP_INIT_PARTITIONED_TIMERS(OMP_serial);
3750 #endif
3751   __kmp_initialize_root(root);
3752 
3753   /* setup new root thread structure */
3754   if (root->r.r_uber_thread) {
3755     root_thread = root->r.r_uber_thread;
3756   } else {
3757     root_thread = (kmp_info_t *)__kmp_allocate(sizeof(kmp_info_t));
3758     if (__kmp_storage_map) {
3759       __kmp_print_thread_storage_map(root_thread, gtid);
3760     }
3761     root_thread->th.th_info.ds.ds_gtid = gtid;
3762 #if OMPT_SUPPORT
3763     root_thread->th.ompt_thread_info.thread_data = ompt_data_none;
3764 #endif
3765     root_thread->th.th_root = root;
3766     if (__kmp_env_consistency_check) {
3767       root_thread->th.th_cons = __kmp_allocate_cons_stack(gtid);
3768     }
3769 #if USE_FAST_MEMORY
3770     __kmp_initialize_fast_memory(root_thread);
3771 #endif /* USE_FAST_MEMORY */
3772 
3773 #if KMP_USE_BGET
3774     KMP_DEBUG_ASSERT(root_thread->th.th_local.bget_data == NULL);
3775     __kmp_initialize_bget(root_thread);
3776 #endif
3777     __kmp_init_random(root_thread); // Initialize random number generator
3778   }
3779 
3780   /* setup the serial team held in reserve by the root thread */
3781   if (!root_thread->th.th_serial_team) {
3782     kmp_internal_control_t r_icvs = __kmp_get_global_icvs();
3783     KF_TRACE(10, ("__kmp_register_root: before serial_team\n"));
3784     root_thread->th.th_serial_team =
3785         __kmp_allocate_team(root, 1, 1,
3786 #if OMPT_SUPPORT
3787                             ompt_data_none, // root parallel id
3788 #endif
3789 #if OMP_40_ENABLED
3790                             proc_bind_default,
3791 #endif
3792                             &r_icvs, 0 USE_NESTED_HOT_ARG(NULL));
3793   }
3794   KMP_ASSERT(root_thread->th.th_serial_team);
3795   KF_TRACE(10, ("__kmp_register_root: after serial_team = %p\n",
3796                 root_thread->th.th_serial_team));
3797 
3798   /* drop root_thread into place */
3799   TCW_SYNC_PTR(__kmp_threads[gtid], root_thread);
3800 
3801   root->r.r_root_team->t.t_threads[0] = root_thread;
3802   root->r.r_hot_team->t.t_threads[0] = root_thread;
3803   root_thread->th.th_serial_team->t.t_threads[0] = root_thread;
3804   // AC: the team created in reserve, not for execution (it is unused for now).
3805   root_thread->th.th_serial_team->t.t_serialized = 0;
3806   root->r.r_uber_thread = root_thread;
3807 
3808   /* initialize the thread, get it ready to go */
3809   __kmp_initialize_info(root_thread, root->r.r_root_team, 0, gtid);
3810   TCW_4(__kmp_init_gtid, TRUE);
3811 
3812   /* prepare the master thread for get_gtid() */
3813   __kmp_gtid_set_specific(gtid);
3814 
3815 #if USE_ITT_BUILD
3816   __kmp_itt_thread_name(gtid);
3817 #endif /* USE_ITT_BUILD */
3818 
3819 #ifdef KMP_TDATA_GTID
3820   __kmp_gtid = gtid;
3821 #endif
3822   __kmp_create_worker(gtid, root_thread, __kmp_stksize);
3823   KMP_DEBUG_ASSERT(__kmp_gtid_get_specific() == gtid);
3824 
3825   KA_TRACE(20, ("__kmp_register_root: T#%d init T#%d(%d:%d) arrived: join=%u, "
3826                 "plain=%u\n",
3827                 gtid, __kmp_gtid_from_tid(0, root->r.r_hot_team),
3828                 root->r.r_hot_team->t.t_id, 0, KMP_INIT_BARRIER_STATE,
3829                 KMP_INIT_BARRIER_STATE));
3830   { // Initialize barrier data.
3831     int b;
3832     for (b = 0; b < bs_last_barrier; ++b) {
3833       root_thread->th.th_bar[b].bb.b_arrived = KMP_INIT_BARRIER_STATE;
3834 #if USE_DEBUGGER
3835       root_thread->th.th_bar[b].bb.b_worker_arrived = 0;
3836 #endif
3837     }
3838   }
3839   KMP_DEBUG_ASSERT(root->r.r_hot_team->t.t_bar[bs_forkjoin_barrier].b_arrived ==
3840                    KMP_INIT_BARRIER_STATE);
3841 
3842 #if KMP_AFFINITY_SUPPORTED
3843 #if OMP_40_ENABLED
3844   root_thread->th.th_current_place = KMP_PLACE_UNDEFINED;
3845   root_thread->th.th_new_place = KMP_PLACE_UNDEFINED;
3846   root_thread->th.th_first_place = KMP_PLACE_UNDEFINED;
3847   root_thread->th.th_last_place = KMP_PLACE_UNDEFINED;
3848 #endif
3849   if (TCR_4(__kmp_init_middle)) {
3850     __kmp_affinity_set_init_mask(gtid, TRUE);
3851   }
3852 #endif /* KMP_AFFINITY_SUPPORTED */
3853 #if OMP_50_ENABLED
3854   root_thread->th.th_def_allocator = __kmp_def_allocator;
3855   root_thread->th.th_prev_level = 0;
3856   root_thread->th.th_prev_num_threads = 1;
3857 #endif
3858 
3859   kmp_cg_root_t *tmp = (kmp_cg_root_t *)__kmp_allocate(sizeof(kmp_cg_root_t));
3860   tmp->cg_root = root_thread;
3861   tmp->cg_thread_limit = __kmp_cg_max_nth;
3862   tmp->cg_nthreads = 1;
3863   KA_TRACE(100, ("__kmp_register_root: Thread %p created node %p with"
3864                  " cg_nthreads init to 1\n",
3865                  root_thread, tmp));
3866   tmp->up = NULL;
3867   root_thread->th.th_cg_roots = tmp;
3868 
3869   __kmp_root_counter++;
3870 
3871 #if OMPT_SUPPORT
3872   if (!initial_thread && ompt_enabled.enabled) {
3873 
3874     kmp_info_t *root_thread = ompt_get_thread();
3875 
3876     ompt_set_thread_state(root_thread, ompt_state_overhead);
3877 
3878     if (ompt_enabled.ompt_callback_thread_begin) {
3879       ompt_callbacks.ompt_callback(ompt_callback_thread_begin)(
3880           ompt_thread_initial, __ompt_get_thread_data_internal());
3881     }
3882     ompt_data_t *task_data;
3883     __ompt_get_task_info_internal(0, NULL, &task_data, NULL, NULL, NULL);
3884     if (ompt_enabled.ompt_callback_task_create) {
3885       ompt_callbacks.ompt_callback(ompt_callback_task_create)(
3886           NULL, NULL, task_data, ompt_task_initial, 0, NULL);
3887       // initial task has nothing to return to
3888     }
3889 
3890     ompt_set_thread_state(root_thread, ompt_state_work_serial);
3891   }
3892 #endif
3893 
3894   KMP_MB();
3895   __kmp_release_bootstrap_lock(&__kmp_forkjoin_lock);
3896 
3897   return gtid;
3898 }
3899 
3900 #if KMP_NESTED_HOT_TEAMS
3901 static int __kmp_free_hot_teams(kmp_root_t *root, kmp_info_t *thr, int level,
3902                                 const int max_level) {
3903   int i, n, nth;
3904   kmp_hot_team_ptr_t *hot_teams = thr->th.th_hot_teams;
3905   if (!hot_teams || !hot_teams[level].hot_team) {
3906     return 0;
3907   }
3908   KMP_DEBUG_ASSERT(level < max_level);
3909   kmp_team_t *team = hot_teams[level].hot_team;
3910   nth = hot_teams[level].hot_team_nth;
3911   n = nth - 1; // master is not freed
3912   if (level < max_level - 1) {
3913     for (i = 0; i < nth; ++i) {
3914       kmp_info_t *th = team->t.t_threads[i];
3915       n += __kmp_free_hot_teams(root, th, level + 1, max_level);
3916       if (i > 0 && th->th.th_hot_teams) {
3917         __kmp_free(th->th.th_hot_teams);
3918         th->th.th_hot_teams = NULL;
3919       }
3920     }
3921   }
3922   __kmp_free_team(root, team, NULL);
3923   return n;
3924 }
3925 #endif
3926 
3927 // Resets a root thread and clear its root and hot teams.
3928 // Returns the number of __kmp_threads entries directly and indirectly freed.
3929 static int __kmp_reset_root(int gtid, kmp_root_t *root) {
3930   kmp_team_t *root_team = root->r.r_root_team;
3931   kmp_team_t *hot_team = root->r.r_hot_team;
3932   int n = hot_team->t.t_nproc;
3933   int i;
3934 
3935   KMP_DEBUG_ASSERT(!root->r.r_active);
3936 
3937   root->r.r_root_team = NULL;
3938   root->r.r_hot_team = NULL;
3939   // __kmp_free_team() does not free hot teams, so we have to clear r_hot_team
3940   // before call to __kmp_free_team().
3941   __kmp_free_team(root, root_team USE_NESTED_HOT_ARG(NULL));
3942 #if KMP_NESTED_HOT_TEAMS
3943   if (__kmp_hot_teams_max_level >
3944       0) { // need to free nested hot teams and their threads if any
3945     for (i = 0; i < hot_team->t.t_nproc; ++i) {
3946       kmp_info_t *th = hot_team->t.t_threads[i];
3947       if (__kmp_hot_teams_max_level > 1) {
3948         n += __kmp_free_hot_teams(root, th, 1, __kmp_hot_teams_max_level);
3949       }
3950       if (th->th.th_hot_teams) {
3951         __kmp_free(th->th.th_hot_teams);
3952         th->th.th_hot_teams = NULL;
3953       }
3954     }
3955   }
3956 #endif
3957   __kmp_free_team(root, hot_team USE_NESTED_HOT_ARG(NULL));
3958 
3959   // Before we can reap the thread, we need to make certain that all other
3960   // threads in the teams that had this root as ancestor have stopped trying to
3961   // steal tasks.
3962   if (__kmp_tasking_mode != tskm_immediate_exec) {
3963     __kmp_wait_to_unref_task_teams();
3964   }
3965 
3966 #if KMP_OS_WINDOWS
3967   /* Close Handle of root duplicated in __kmp_create_worker (tr #62919) */
3968   KA_TRACE(
3969       10, ("__kmp_reset_root: free handle, th = %p, handle = %" KMP_UINTPTR_SPEC
3970            "\n",
3971            (LPVOID) & (root->r.r_uber_thread->th),
3972            root->r.r_uber_thread->th.th_info.ds.ds_thread));
3973   __kmp_free_handle(root->r.r_uber_thread->th.th_info.ds.ds_thread);
3974 #endif /* KMP_OS_WINDOWS */
3975 
3976 #if OMPT_SUPPORT
3977   if (ompt_enabled.ompt_callback_thread_end) {
3978     ompt_callbacks.ompt_callback(ompt_callback_thread_end)(
3979         &(root->r.r_uber_thread->th.ompt_thread_info.thread_data));
3980   }
3981 #endif
3982 
3983   TCW_4(__kmp_nth,
3984         __kmp_nth - 1); // __kmp_reap_thread will decrement __kmp_all_nth.
3985   root->r.r_uber_thread->th.th_cg_roots->cg_nthreads--;
3986   KA_TRACE(100, ("__kmp_reset_root: Thread %p decrement cg_nthreads on node %p"
3987                  " to %d\n",
3988                  root->r.r_uber_thread, root->r.r_uber_thread->th.th_cg_roots,
3989                  root->r.r_uber_thread->th.th_cg_roots->cg_nthreads));
3990 
3991   __kmp_reap_thread(root->r.r_uber_thread, 1);
3992 
3993   // We canot put root thread to __kmp_thread_pool, so we have to reap it istead
3994   // of freeing.
3995   root->r.r_uber_thread = NULL;
3996   /* mark root as no longer in use */
3997   root->r.r_begin = FALSE;
3998 
3999   return n;
4000 }
4001 
4002 void __kmp_unregister_root_current_thread(int gtid) {
4003   KA_TRACE(1, ("__kmp_unregister_root_current_thread: enter T#%d\n", gtid));
4004   /* this lock should be ok, since unregister_root_current_thread is never
4005      called during an abort, only during a normal close. furthermore, if you
4006      have the forkjoin lock, you should never try to get the initz lock */
4007   __kmp_acquire_bootstrap_lock(&__kmp_forkjoin_lock);
4008   if (TCR_4(__kmp_global.g.g_done) || !__kmp_init_serial) {
4009     KC_TRACE(10, ("__kmp_unregister_root_current_thread: already finished, "
4010                   "exiting T#%d\n",
4011                   gtid));
4012     __kmp_release_bootstrap_lock(&__kmp_forkjoin_lock);
4013     return;
4014   }
4015   kmp_root_t *root = __kmp_root[gtid];
4016 
4017   KMP_DEBUG_ASSERT(__kmp_threads && __kmp_threads[gtid]);
4018   KMP_ASSERT(KMP_UBER_GTID(gtid));
4019   KMP_ASSERT(root == __kmp_threads[gtid]->th.th_root);
4020   KMP_ASSERT(root->r.r_active == FALSE);
4021 
4022   KMP_MB();
4023 
4024 #if OMP_45_ENABLED
4025   kmp_info_t *thread = __kmp_threads[gtid];
4026   kmp_team_t *team = thread->th.th_team;
4027   kmp_task_team_t *task_team = thread->th.th_task_team;
4028 
4029   // we need to wait for the proxy tasks before finishing the thread
4030   if (task_team != NULL && task_team->tt.tt_found_proxy_tasks) {
4031 #if OMPT_SUPPORT
4032     // the runtime is shutting down so we won't report any events
4033     thread->th.ompt_thread_info.state = ompt_state_undefined;
4034 #endif
4035     __kmp_task_team_wait(thread, team USE_ITT_BUILD_ARG(NULL));
4036   }
4037 #endif
4038 
4039   __kmp_reset_root(gtid, root);
4040 
4041   /* free up this thread slot */
4042   __kmp_gtid_set_specific(KMP_GTID_DNE);
4043 #ifdef KMP_TDATA_GTID
4044   __kmp_gtid = KMP_GTID_DNE;
4045 #endif
4046 
4047   KMP_MB();
4048   KC_TRACE(10,
4049            ("__kmp_unregister_root_current_thread: T#%d unregistered\n", gtid));
4050 
4051   __kmp_release_bootstrap_lock(&__kmp_forkjoin_lock);
4052 }
4053 
4054 #if KMP_OS_WINDOWS
4055 /* __kmp_forkjoin_lock must be already held
4056    Unregisters a root thread that is not the current thread.  Returns the number
4057    of __kmp_threads entries freed as a result. */
4058 static int __kmp_unregister_root_other_thread(int gtid) {
4059   kmp_root_t *root = __kmp_root[gtid];
4060   int r;
4061 
4062   KA_TRACE(1, ("__kmp_unregister_root_other_thread: enter T#%d\n", gtid));
4063   KMP_DEBUG_ASSERT(__kmp_threads && __kmp_threads[gtid]);
4064   KMP_ASSERT(KMP_UBER_GTID(gtid));
4065   KMP_ASSERT(root == __kmp_threads[gtid]->th.th_root);
4066   KMP_ASSERT(root->r.r_active == FALSE);
4067 
4068   r = __kmp_reset_root(gtid, root);
4069   KC_TRACE(10,
4070            ("__kmp_unregister_root_other_thread: T#%d unregistered\n", gtid));
4071   return r;
4072 }
4073 #endif
4074 
4075 #if KMP_DEBUG
4076 void __kmp_task_info() {
4077 
4078   kmp_int32 gtid = __kmp_entry_gtid();
4079   kmp_int32 tid = __kmp_tid_from_gtid(gtid);
4080   kmp_info_t *this_thr = __kmp_threads[gtid];
4081   kmp_team_t *steam = this_thr->th.th_serial_team;
4082   kmp_team_t *team = this_thr->th.th_team;
4083 
4084   __kmp_printf(
4085       "__kmp_task_info: gtid=%d tid=%d t_thread=%p team=%p steam=%p curtask=%p "
4086       "ptask=%p\n",
4087       gtid, tid, this_thr, team, steam, this_thr->th.th_current_task,
4088       team->t.t_implicit_task_taskdata[tid].td_parent);
4089 }
4090 #endif // KMP_DEBUG
4091 
4092 /* TODO optimize with one big memclr, take out what isn't needed, split
4093    responsibility to workers as much as possible, and delay initialization of
4094    features as much as possible  */
4095 static void __kmp_initialize_info(kmp_info_t *this_thr, kmp_team_t *team,
4096                                   int tid, int gtid) {
4097   /* this_thr->th.th_info.ds.ds_gtid is setup in
4098      kmp_allocate_thread/create_worker.
4099      this_thr->th.th_serial_team is setup in __kmp_allocate_thread */
4100   kmp_info_t *master = team->t.t_threads[0];
4101   KMP_DEBUG_ASSERT(this_thr != NULL);
4102   KMP_DEBUG_ASSERT(this_thr->th.th_serial_team);
4103   KMP_DEBUG_ASSERT(team);
4104   KMP_DEBUG_ASSERT(team->t.t_threads);
4105   KMP_DEBUG_ASSERT(team->t.t_dispatch);
4106   KMP_DEBUG_ASSERT(master);
4107   KMP_DEBUG_ASSERT(master->th.th_root);
4108 
4109   KMP_MB();
4110 
4111   TCW_SYNC_PTR(this_thr->th.th_team, team);
4112 
4113   this_thr->th.th_info.ds.ds_tid = tid;
4114   this_thr->th.th_set_nproc = 0;
4115   if (__kmp_tasking_mode != tskm_immediate_exec)
4116     // When tasking is possible, threads are not safe to reap until they are
4117     // done tasking; this will be set when tasking code is exited in wait
4118     this_thr->th.th_reap_state = KMP_NOT_SAFE_TO_REAP;
4119   else // no tasking --> always safe to reap
4120     this_thr->th.th_reap_state = KMP_SAFE_TO_REAP;
4121 #if OMP_40_ENABLED
4122   this_thr->th.th_set_proc_bind = proc_bind_default;
4123 #if KMP_AFFINITY_SUPPORTED
4124   this_thr->th.th_new_place = this_thr->th.th_current_place;
4125 #endif
4126 #endif
4127   this_thr->th.th_root = master->th.th_root;
4128 
4129   /* setup the thread's cache of the team structure */
4130   this_thr->th.th_team_nproc = team->t.t_nproc;
4131   this_thr->th.th_team_master = master;
4132   this_thr->th.th_team_serialized = team->t.t_serialized;
4133   TCW_PTR(this_thr->th.th_sleep_loc, NULL);
4134 
4135   KMP_DEBUG_ASSERT(team->t.t_implicit_task_taskdata);
4136 
4137   KF_TRACE(10, ("__kmp_initialize_info1: T#%d:%d this_thread=%p curtask=%p\n",
4138                 tid, gtid, this_thr, this_thr->th.th_current_task));
4139 
4140   __kmp_init_implicit_task(this_thr->th.th_team_master->th.th_ident, this_thr,
4141                            team, tid, TRUE);
4142 
4143   KF_TRACE(10, ("__kmp_initialize_info2: T#%d:%d this_thread=%p curtask=%p\n",
4144                 tid, gtid, this_thr, this_thr->th.th_current_task));
4145   // TODO: Initialize ICVs from parent; GEH - isn't that already done in
4146   // __kmp_initialize_team()?
4147 
4148   /* TODO no worksharing in speculative threads */
4149   this_thr->th.th_dispatch = &team->t.t_dispatch[tid];
4150 
4151   this_thr->th.th_local.this_construct = 0;
4152 
4153   if (!this_thr->th.th_pri_common) {
4154     this_thr->th.th_pri_common =
4155         (struct common_table *)__kmp_allocate(sizeof(struct common_table));
4156     if (__kmp_storage_map) {
4157       __kmp_print_storage_map_gtid(
4158           gtid, this_thr->th.th_pri_common, this_thr->th.th_pri_common + 1,
4159           sizeof(struct common_table), "th_%d.th_pri_common\n", gtid);
4160     }
4161     this_thr->th.th_pri_head = NULL;
4162   }
4163 
4164   if (this_thr != master && // Master's CG root is initialized elsewhere
4165       this_thr->th.th_cg_roots != master->th.th_cg_roots) { // CG root not set
4166     // Make new thread's CG root same as master's
4167     KMP_DEBUG_ASSERT(master->th.th_cg_roots);
4168     this_thr->th.th_cg_roots = master->th.th_cg_roots;
4169     // Increment new thread's CG root's counter to add the new thread
4170     this_thr->th.th_cg_roots->cg_nthreads++;
4171     KA_TRACE(100, ("__kmp_initialize_info: Thread %p increment cg_nthreads on"
4172                    " node %p of thread %p to %d\n",
4173                    this_thr, this_thr->th.th_cg_roots,
4174                    this_thr->th.th_cg_roots->cg_root,
4175                    this_thr->th.th_cg_roots->cg_nthreads));
4176     this_thr->th.th_current_task->td_icvs.thread_limit =
4177         this_thr->th.th_cg_roots->cg_thread_limit;
4178   }
4179 
4180   /* Initialize dynamic dispatch */
4181   {
4182     volatile kmp_disp_t *dispatch = this_thr->th.th_dispatch;
4183     // Use team max_nproc since this will never change for the team.
4184     size_t disp_size =
4185         sizeof(dispatch_private_info_t) *
4186         (team->t.t_max_nproc == 1 ? 1 : __kmp_dispatch_num_buffers);
4187     KD_TRACE(10, ("__kmp_initialize_info: T#%d max_nproc: %d\n", gtid,
4188                   team->t.t_max_nproc));
4189     KMP_ASSERT(dispatch);
4190     KMP_DEBUG_ASSERT(team->t.t_dispatch);
4191     KMP_DEBUG_ASSERT(dispatch == &team->t.t_dispatch[tid]);
4192 
4193     dispatch->th_disp_index = 0;
4194 #if OMP_45_ENABLED
4195     dispatch->th_doacross_buf_idx = 0;
4196 #endif
4197     if (!dispatch->th_disp_buffer) {
4198       dispatch->th_disp_buffer =
4199           (dispatch_private_info_t *)__kmp_allocate(disp_size);
4200 
4201       if (__kmp_storage_map) {
4202         __kmp_print_storage_map_gtid(
4203             gtid, &dispatch->th_disp_buffer[0],
4204             &dispatch->th_disp_buffer[team->t.t_max_nproc == 1
4205                                           ? 1
4206                                           : __kmp_dispatch_num_buffers],
4207             disp_size, "th_%d.th_dispatch.th_disp_buffer "
4208                        "(team_%d.t_dispatch[%d].th_disp_buffer)",
4209             gtid, team->t.t_id, gtid);
4210       }
4211     } else {
4212       memset(&dispatch->th_disp_buffer[0], '\0', disp_size);
4213     }
4214 
4215     dispatch->th_dispatch_pr_current = 0;
4216     dispatch->th_dispatch_sh_current = 0;
4217 
4218     dispatch->th_deo_fcn = 0; /* ORDERED     */
4219     dispatch->th_dxo_fcn = 0; /* END ORDERED */
4220   }
4221 
4222   this_thr->th.th_next_pool = NULL;
4223 
4224   if (!this_thr->th.th_task_state_memo_stack) {
4225     size_t i;
4226     this_thr->th.th_task_state_memo_stack =
4227         (kmp_uint8 *)__kmp_allocate(4 * sizeof(kmp_uint8));
4228     this_thr->th.th_task_state_top = 0;
4229     this_thr->th.th_task_state_stack_sz = 4;
4230     for (i = 0; i < this_thr->th.th_task_state_stack_sz;
4231          ++i) // zero init the stack
4232       this_thr->th.th_task_state_memo_stack[i] = 0;
4233   }
4234 
4235   KMP_DEBUG_ASSERT(!this_thr->th.th_spin_here);
4236   KMP_DEBUG_ASSERT(this_thr->th.th_next_waiting == 0);
4237 
4238   KMP_MB();
4239 }
4240 
4241 /* allocate a new thread for the requesting team. this is only called from
4242    within a forkjoin critical section. we will first try to get an available
4243    thread from the thread pool. if none is available, we will fork a new one
4244    assuming we are able to create a new one. this should be assured, as the
4245    caller should check on this first. */
4246 kmp_info_t *__kmp_allocate_thread(kmp_root_t *root, kmp_team_t *team,
4247                                   int new_tid) {
4248   kmp_team_t *serial_team;
4249   kmp_info_t *new_thr;
4250   int new_gtid;
4251 
4252   KA_TRACE(20, ("__kmp_allocate_thread: T#%d\n", __kmp_get_gtid()));
4253   KMP_DEBUG_ASSERT(root && team);
4254 #if !KMP_NESTED_HOT_TEAMS
4255   KMP_DEBUG_ASSERT(KMP_MASTER_GTID(__kmp_get_gtid()));
4256 #endif
4257   KMP_MB();
4258 
4259   /* first, try to get one from the thread pool */
4260   if (__kmp_thread_pool) {
4261     new_thr = CCAST(kmp_info_t *, __kmp_thread_pool);
4262     __kmp_thread_pool = (volatile kmp_info_t *)new_thr->th.th_next_pool;
4263     if (new_thr == __kmp_thread_pool_insert_pt) {
4264       __kmp_thread_pool_insert_pt = NULL;
4265     }
4266     TCW_4(new_thr->th.th_in_pool, FALSE);
4267     // Don't touch th_active_in_pool or th_active.
4268     // The worker thread adjusts those flags as it sleeps/awakens.
4269     __kmp_thread_pool_nth--;
4270 
4271     KA_TRACE(20, ("__kmp_allocate_thread: T#%d using thread T#%d\n",
4272                   __kmp_get_gtid(), new_thr->th.th_info.ds.ds_gtid));
4273     KMP_ASSERT(!new_thr->th.th_team);
4274     KMP_DEBUG_ASSERT(__kmp_nth < __kmp_threads_capacity);
4275     KMP_DEBUG_ASSERT(__kmp_thread_pool_nth >= 0);
4276 
4277     /* setup the thread structure */
4278     __kmp_initialize_info(new_thr, team, new_tid,
4279                           new_thr->th.th_info.ds.ds_gtid);
4280     KMP_DEBUG_ASSERT(new_thr->th.th_serial_team);
4281 
4282     TCW_4(__kmp_nth, __kmp_nth + 1);
4283 
4284     new_thr->th.th_task_state = 0;
4285     new_thr->th.th_task_state_top = 0;
4286     new_thr->th.th_task_state_stack_sz = 4;
4287 
4288 #ifdef KMP_ADJUST_BLOCKTIME
4289     /* Adjust blocktime back to zero if necessary */
4290     /* Middle initialization might not have occurred yet */
4291     if (!__kmp_env_blocktime && (__kmp_avail_proc > 0)) {
4292       if (__kmp_nth > __kmp_avail_proc) {
4293         __kmp_zero_bt = TRUE;
4294       }
4295     }
4296 #endif /* KMP_ADJUST_BLOCKTIME */
4297 
4298 #if KMP_DEBUG
4299     // If thread entered pool via __kmp_free_thread, wait_flag should !=
4300     // KMP_BARRIER_PARENT_FLAG.
4301     int b;
4302     kmp_balign_t *balign = new_thr->th.th_bar;
4303     for (b = 0; b < bs_last_barrier; ++b)
4304       KMP_DEBUG_ASSERT(balign[b].bb.wait_flag != KMP_BARRIER_PARENT_FLAG);
4305 #endif
4306 
4307     KF_TRACE(10, ("__kmp_allocate_thread: T#%d using thread %p T#%d\n",
4308                   __kmp_get_gtid(), new_thr, new_thr->th.th_info.ds.ds_gtid));
4309 
4310     KMP_MB();
4311     return new_thr;
4312   }
4313 
4314   /* no, well fork a new one */
4315   KMP_ASSERT(__kmp_nth == __kmp_all_nth);
4316   KMP_ASSERT(__kmp_all_nth < __kmp_threads_capacity);
4317 
4318 #if KMP_USE_MONITOR
4319   // If this is the first worker thread the RTL is creating, then also
4320   // launch the monitor thread.  We try to do this as early as possible.
4321   if (!TCR_4(__kmp_init_monitor)) {
4322     __kmp_acquire_bootstrap_lock(&__kmp_monitor_lock);
4323     if (!TCR_4(__kmp_init_monitor)) {
4324       KF_TRACE(10, ("before __kmp_create_monitor\n"));
4325       TCW_4(__kmp_init_monitor, 1);
4326       __kmp_create_monitor(&__kmp_monitor);
4327       KF_TRACE(10, ("after __kmp_create_monitor\n"));
4328 #if KMP_OS_WINDOWS
4329       // AC: wait until monitor has started. This is a fix for CQ232808.
4330       // The reason is that if the library is loaded/unloaded in a loop with
4331       // small (parallel) work in between, then there is high probability that
4332       // monitor thread started after the library shutdown. At shutdown it is
4333       // too late to cope with the problem, because when the master is in
4334       // DllMain (process detach) the monitor has no chances to start (it is
4335       // blocked), and master has no means to inform the monitor that the
4336       // library has gone, because all the memory which the monitor can access
4337       // is going to be released/reset.
4338       while (TCR_4(__kmp_init_monitor) < 2) {
4339         KMP_YIELD(TRUE);
4340       }
4341       KF_TRACE(10, ("after monitor thread has started\n"));
4342 #endif
4343     }
4344     __kmp_release_bootstrap_lock(&__kmp_monitor_lock);
4345   }
4346 #endif
4347 
4348   KMP_MB();
4349   for (new_gtid = 1; TCR_PTR(__kmp_threads[new_gtid]) != NULL; ++new_gtid) {
4350     KMP_DEBUG_ASSERT(new_gtid < __kmp_threads_capacity);
4351   }
4352 
4353   /* allocate space for it. */
4354   new_thr = (kmp_info_t *)__kmp_allocate(sizeof(kmp_info_t));
4355 
4356   TCW_SYNC_PTR(__kmp_threads[new_gtid], new_thr);
4357 
4358   if (__kmp_storage_map) {
4359     __kmp_print_thread_storage_map(new_thr, new_gtid);
4360   }
4361 
4362   // add the reserve serialized team, initialized from the team's master thread
4363   {
4364     kmp_internal_control_t r_icvs = __kmp_get_x_global_icvs(team);
4365     KF_TRACE(10, ("__kmp_allocate_thread: before th_serial/serial_team\n"));
4366     new_thr->th.th_serial_team = serial_team =
4367         (kmp_team_t *)__kmp_allocate_team(root, 1, 1,
4368 #if OMPT_SUPPORT
4369                                           ompt_data_none, // root parallel id
4370 #endif
4371 #if OMP_40_ENABLED
4372                                           proc_bind_default,
4373 #endif
4374                                           &r_icvs, 0 USE_NESTED_HOT_ARG(NULL));
4375   }
4376   KMP_ASSERT(serial_team);
4377   serial_team->t.t_serialized = 0; // AC: the team created in reserve, not for
4378   // execution (it is unused for now).
4379   serial_team->t.t_threads[0] = new_thr;
4380   KF_TRACE(10,
4381            ("__kmp_allocate_thread: after th_serial/serial_team : new_thr=%p\n",
4382             new_thr));
4383 
4384   /* setup the thread structures */
4385   __kmp_initialize_info(new_thr, team, new_tid, new_gtid);
4386 
4387 #if USE_FAST_MEMORY
4388   __kmp_initialize_fast_memory(new_thr);
4389 #endif /* USE_FAST_MEMORY */
4390 
4391 #if KMP_USE_BGET
4392   KMP_DEBUG_ASSERT(new_thr->th.th_local.bget_data == NULL);
4393   __kmp_initialize_bget(new_thr);
4394 #endif
4395 
4396   __kmp_init_random(new_thr); // Initialize random number generator
4397 
4398   /* Initialize these only once when thread is grabbed for a team allocation */
4399   KA_TRACE(20,
4400            ("__kmp_allocate_thread: T#%d init go fork=%u, plain=%u\n",
4401             __kmp_get_gtid(), KMP_INIT_BARRIER_STATE, KMP_INIT_BARRIER_STATE));
4402 
4403   int b;
4404   kmp_balign_t *balign = new_thr->th.th_bar;
4405   for (b = 0; b < bs_last_barrier; ++b) {
4406     balign[b].bb.b_go = KMP_INIT_BARRIER_STATE;
4407     balign[b].bb.team = NULL;
4408     balign[b].bb.wait_flag = KMP_BARRIER_NOT_WAITING;
4409     balign[b].bb.use_oncore_barrier = 0;
4410   }
4411 
4412   new_thr->th.th_spin_here = FALSE;
4413   new_thr->th.th_next_waiting = 0;
4414 #if KMP_OS_UNIX
4415   new_thr->th.th_blocking = false;
4416 #endif
4417 
4418 #if OMP_40_ENABLED && KMP_AFFINITY_SUPPORTED
4419   new_thr->th.th_current_place = KMP_PLACE_UNDEFINED;
4420   new_thr->th.th_new_place = KMP_PLACE_UNDEFINED;
4421   new_thr->th.th_first_place = KMP_PLACE_UNDEFINED;
4422   new_thr->th.th_last_place = KMP_PLACE_UNDEFINED;
4423 #endif
4424 #if OMP_50_ENABLED
4425   new_thr->th.th_def_allocator = __kmp_def_allocator;
4426   new_thr->th.th_prev_level = 0;
4427   new_thr->th.th_prev_num_threads = 1;
4428 #endif
4429 
4430   TCW_4(new_thr->th.th_in_pool, FALSE);
4431   new_thr->th.th_active_in_pool = FALSE;
4432   TCW_4(new_thr->th.th_active, TRUE);
4433 
4434   /* adjust the global counters */
4435   __kmp_all_nth++;
4436   __kmp_nth++;
4437 
4438   // if __kmp_adjust_gtid_mode is set, then we use method #1 (sp search) for low
4439   // numbers of procs, and method #2 (keyed API call) for higher numbers.
4440   if (__kmp_adjust_gtid_mode) {
4441     if (__kmp_all_nth >= __kmp_tls_gtid_min) {
4442       if (TCR_4(__kmp_gtid_mode) != 2) {
4443         TCW_4(__kmp_gtid_mode, 2);
4444       }
4445     } else {
4446       if (TCR_4(__kmp_gtid_mode) != 1) {
4447         TCW_4(__kmp_gtid_mode, 1);
4448       }
4449     }
4450   }
4451 
4452 #ifdef KMP_ADJUST_BLOCKTIME
4453   /* Adjust blocktime back to zero if necessary       */
4454   /* Middle initialization might not have occurred yet */
4455   if (!__kmp_env_blocktime && (__kmp_avail_proc > 0)) {
4456     if (__kmp_nth > __kmp_avail_proc) {
4457       __kmp_zero_bt = TRUE;
4458     }
4459   }
4460 #endif /* KMP_ADJUST_BLOCKTIME */
4461 
4462   /* actually fork it and create the new worker thread */
4463   KF_TRACE(
4464       10, ("__kmp_allocate_thread: before __kmp_create_worker: %p\n", new_thr));
4465   __kmp_create_worker(new_gtid, new_thr, __kmp_stksize);
4466   KF_TRACE(10,
4467            ("__kmp_allocate_thread: after __kmp_create_worker: %p\n", new_thr));
4468 
4469   KA_TRACE(20, ("__kmp_allocate_thread: T#%d forked T#%d\n", __kmp_get_gtid(),
4470                 new_gtid));
4471   KMP_MB();
4472   return new_thr;
4473 }
4474 
4475 /* Reinitialize team for reuse.
4476    The hot team code calls this case at every fork barrier, so EPCC barrier
4477    test are extremely sensitive to changes in it, esp. writes to the team
4478    struct, which cause a cache invalidation in all threads.
4479    IF YOU TOUCH THIS ROUTINE, RUN EPCC C SYNCBENCH ON A BIG-IRON MACHINE!!! */
4480 static void __kmp_reinitialize_team(kmp_team_t *team,
4481                                     kmp_internal_control_t *new_icvs,
4482                                     ident_t *loc) {
4483   KF_TRACE(10, ("__kmp_reinitialize_team: enter this_thread=%p team=%p\n",
4484                 team->t.t_threads[0], team));
4485   KMP_DEBUG_ASSERT(team && new_icvs);
4486   KMP_DEBUG_ASSERT((!TCR_4(__kmp_init_parallel)) || new_icvs->nproc);
4487   KMP_CHECK_UPDATE(team->t.t_ident, loc);
4488 
4489   KMP_CHECK_UPDATE(team->t.t_id, KMP_GEN_TEAM_ID());
4490   // Copy ICVs to the master thread's implicit taskdata
4491   __kmp_init_implicit_task(loc, team->t.t_threads[0], team, 0, FALSE);
4492   copy_icvs(&team->t.t_implicit_task_taskdata[0].td_icvs, new_icvs);
4493 
4494   KF_TRACE(10, ("__kmp_reinitialize_team: exit this_thread=%p team=%p\n",
4495                 team->t.t_threads[0], team));
4496 }
4497 
4498 /* Initialize the team data structure.
4499    This assumes the t_threads and t_max_nproc are already set.
4500    Also, we don't touch the arguments */
4501 static void __kmp_initialize_team(kmp_team_t *team, int new_nproc,
4502                                   kmp_internal_control_t *new_icvs,
4503                                   ident_t *loc) {
4504   KF_TRACE(10, ("__kmp_initialize_team: enter: team=%p\n", team));
4505 
4506   /* verify */
4507   KMP_DEBUG_ASSERT(team);
4508   KMP_DEBUG_ASSERT(new_nproc <= team->t.t_max_nproc);
4509   KMP_DEBUG_ASSERT(team->t.t_threads);
4510   KMP_MB();
4511 
4512   team->t.t_master_tid = 0; /* not needed */
4513   /* team->t.t_master_bar;        not needed */
4514   team->t.t_serialized = new_nproc > 1 ? 0 : 1;
4515   team->t.t_nproc = new_nproc;
4516 
4517   /* team->t.t_parent     = NULL; TODO not needed & would mess up hot team */
4518   team->t.t_next_pool = NULL;
4519   /* memset( team->t.t_threads, 0, sizeof(kmp_info_t*)*new_nproc ); would mess
4520    * up hot team */
4521 
4522   TCW_SYNC_PTR(team->t.t_pkfn, NULL); /* not needed */
4523   team->t.t_invoke = NULL; /* not needed */
4524 
4525   // TODO???: team->t.t_max_active_levels       = new_max_active_levels;
4526   team->t.t_sched.sched = new_icvs->sched.sched;
4527 
4528 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
4529   team->t.t_fp_control_saved = FALSE; /* not needed */
4530   team->t.t_x87_fpu_control_word = 0; /* not needed */
4531   team->t.t_mxcsr = 0; /* not needed */
4532 #endif /* KMP_ARCH_X86 || KMP_ARCH_X86_64 */
4533 
4534   team->t.t_construct = 0;
4535 
4536   team->t.t_ordered.dt.t_value = 0;
4537   team->t.t_master_active = FALSE;
4538 
4539 #ifdef KMP_DEBUG
4540   team->t.t_copypriv_data = NULL; /* not necessary, but nice for debugging */
4541 #endif
4542 #if KMP_OS_WINDOWS
4543   team->t.t_copyin_counter = 0; /* for barrier-free copyin implementation */
4544 #endif
4545 
4546   team->t.t_control_stack_top = NULL;
4547 
4548   __kmp_reinitialize_team(team, new_icvs, loc);
4549 
4550   KMP_MB();
4551   KF_TRACE(10, ("__kmp_initialize_team: exit: team=%p\n", team));
4552 }
4553 
4554 #if KMP_OS_LINUX && KMP_AFFINITY_SUPPORTED
4555 /* Sets full mask for thread and returns old mask, no changes to structures. */
4556 static void
4557 __kmp_set_thread_affinity_mask_full_tmp(kmp_affin_mask_t *old_mask) {
4558   if (KMP_AFFINITY_CAPABLE()) {
4559     int status;
4560     if (old_mask != NULL) {
4561       status = __kmp_get_system_affinity(old_mask, TRUE);
4562       int error = errno;
4563       if (status != 0) {
4564         __kmp_fatal(KMP_MSG(ChangeThreadAffMaskError), KMP_ERR(error),
4565                     __kmp_msg_null);
4566       }
4567     }
4568     __kmp_set_system_affinity(__kmp_affin_fullMask, TRUE);
4569   }
4570 }
4571 #endif
4572 
4573 #if OMP_40_ENABLED && KMP_AFFINITY_SUPPORTED
4574 
4575 // __kmp_partition_places() is the heart of the OpenMP 4.0 affinity mechanism.
4576 // It calculats the worker + master thread's partition based upon the parent
4577 // thread's partition, and binds each worker to a thread in their partition.
4578 // The master thread's partition should already include its current binding.
4579 static void __kmp_partition_places(kmp_team_t *team, int update_master_only) {
4580   // Copy the master thread's place partion to the team struct
4581   kmp_info_t *master_th = team->t.t_threads[0];
4582   KMP_DEBUG_ASSERT(master_th != NULL);
4583   kmp_proc_bind_t proc_bind = team->t.t_proc_bind;
4584   int first_place = master_th->th.th_first_place;
4585   int last_place = master_th->th.th_last_place;
4586   int masters_place = master_th->th.th_current_place;
4587   team->t.t_first_place = first_place;
4588   team->t.t_last_place = last_place;
4589 
4590   KA_TRACE(20, ("__kmp_partition_places: enter: proc_bind = %d T#%d(%d:0) "
4591                 "bound to place %d partition = [%d,%d]\n",
4592                 proc_bind, __kmp_gtid_from_thread(team->t.t_threads[0]),
4593                 team->t.t_id, masters_place, first_place, last_place));
4594 
4595   switch (proc_bind) {
4596 
4597   case proc_bind_default:
4598     // serial teams might have the proc_bind policy set to proc_bind_default. It
4599     // doesn't matter, as we don't rebind master thread for any proc_bind policy
4600     KMP_DEBUG_ASSERT(team->t.t_nproc == 1);
4601     break;
4602 
4603   case proc_bind_master: {
4604     int f;
4605     int n_th = team->t.t_nproc;
4606     for (f = 1; f < n_th; f++) {
4607       kmp_info_t *th = team->t.t_threads[f];
4608       KMP_DEBUG_ASSERT(th != NULL);
4609       th->th.th_first_place = first_place;
4610       th->th.th_last_place = last_place;
4611       th->th.th_new_place = masters_place;
4612 #if OMP_50_ENABLED
4613       if (__kmp_display_affinity && masters_place != th->th.th_current_place &&
4614           team->t.t_display_affinity != 1) {
4615         team->t.t_display_affinity = 1;
4616       }
4617 #endif
4618 
4619       KA_TRACE(100, ("__kmp_partition_places: master: T#%d(%d:%d) place %d "
4620                      "partition = [%d,%d]\n",
4621                      __kmp_gtid_from_thread(team->t.t_threads[f]), team->t.t_id,
4622                      f, masters_place, first_place, last_place));
4623     }
4624   } break;
4625 
4626   case proc_bind_close: {
4627     int f;
4628     int n_th = team->t.t_nproc;
4629     int n_places;
4630     if (first_place <= last_place) {
4631       n_places = last_place - first_place + 1;
4632     } else {
4633       n_places = __kmp_affinity_num_masks - first_place + last_place + 1;
4634     }
4635     if (n_th <= n_places) {
4636       int place = masters_place;
4637       for (f = 1; f < n_th; f++) {
4638         kmp_info_t *th = team->t.t_threads[f];
4639         KMP_DEBUG_ASSERT(th != NULL);
4640 
4641         if (place == last_place) {
4642           place = first_place;
4643         } else if (place == (int)(__kmp_affinity_num_masks - 1)) {
4644           place = 0;
4645         } else {
4646           place++;
4647         }
4648         th->th.th_first_place = first_place;
4649         th->th.th_last_place = last_place;
4650         th->th.th_new_place = place;
4651 #if OMP_50_ENABLED
4652         if (__kmp_display_affinity && place != th->th.th_current_place &&
4653             team->t.t_display_affinity != 1) {
4654           team->t.t_display_affinity = 1;
4655         }
4656 #endif
4657 
4658         KA_TRACE(100, ("__kmp_partition_places: close: T#%d(%d:%d) place %d "
4659                        "partition = [%d,%d]\n",
4660                        __kmp_gtid_from_thread(team->t.t_threads[f]),
4661                        team->t.t_id, f, place, first_place, last_place));
4662       }
4663     } else {
4664       int S, rem, gap, s_count;
4665       S = n_th / n_places;
4666       s_count = 0;
4667       rem = n_th - (S * n_places);
4668       gap = rem > 0 ? n_places / rem : n_places;
4669       int place = masters_place;
4670       int gap_ct = gap;
4671       for (f = 0; f < n_th; f++) {
4672         kmp_info_t *th = team->t.t_threads[f];
4673         KMP_DEBUG_ASSERT(th != NULL);
4674 
4675         th->th.th_first_place = first_place;
4676         th->th.th_last_place = last_place;
4677         th->th.th_new_place = place;
4678 #if OMP_50_ENABLED
4679         if (__kmp_display_affinity && place != th->th.th_current_place &&
4680             team->t.t_display_affinity != 1) {
4681           team->t.t_display_affinity = 1;
4682         }
4683 #endif
4684         s_count++;
4685 
4686         if ((s_count == S) && rem && (gap_ct == gap)) {
4687           // do nothing, add an extra thread to place on next iteration
4688         } else if ((s_count == S + 1) && rem && (gap_ct == gap)) {
4689           // we added an extra thread to this place; move to next place
4690           if (place == last_place) {
4691             place = first_place;
4692           } else if (place == (int)(__kmp_affinity_num_masks - 1)) {
4693             place = 0;
4694           } else {
4695             place++;
4696           }
4697           s_count = 0;
4698           gap_ct = 1;
4699           rem--;
4700         } else if (s_count == S) { // place full; don't add extra
4701           if (place == last_place) {
4702             place = first_place;
4703           } else if (place == (int)(__kmp_affinity_num_masks - 1)) {
4704             place = 0;
4705           } else {
4706             place++;
4707           }
4708           gap_ct++;
4709           s_count = 0;
4710         }
4711 
4712         KA_TRACE(100,
4713                  ("__kmp_partition_places: close: T#%d(%d:%d) place %d "
4714                   "partition = [%d,%d]\n",
4715                   __kmp_gtid_from_thread(team->t.t_threads[f]), team->t.t_id, f,
4716                   th->th.th_new_place, first_place, last_place));
4717       }
4718       KMP_DEBUG_ASSERT(place == masters_place);
4719     }
4720   } break;
4721 
4722   case proc_bind_spread: {
4723     int f;
4724     int n_th = team->t.t_nproc;
4725     int n_places;
4726     int thidx;
4727     if (first_place <= last_place) {
4728       n_places = last_place - first_place + 1;
4729     } else {
4730       n_places = __kmp_affinity_num_masks - first_place + last_place + 1;
4731     }
4732     if (n_th <= n_places) {
4733       int place = -1;
4734 
4735       if (n_places != static_cast<int>(__kmp_affinity_num_masks)) {
4736         int S = n_places / n_th;
4737         int s_count, rem, gap, gap_ct;
4738 
4739         place = masters_place;
4740         rem = n_places - n_th * S;
4741         gap = rem ? n_th / rem : 1;
4742         gap_ct = gap;
4743         thidx = n_th;
4744         if (update_master_only == 1)
4745           thidx = 1;
4746         for (f = 0; f < thidx; f++) {
4747           kmp_info_t *th = team->t.t_threads[f];
4748           KMP_DEBUG_ASSERT(th != NULL);
4749 
4750           th->th.th_first_place = place;
4751           th->th.th_new_place = place;
4752 #if OMP_50_ENABLED
4753           if (__kmp_display_affinity && place != th->th.th_current_place &&
4754               team->t.t_display_affinity != 1) {
4755             team->t.t_display_affinity = 1;
4756           }
4757 #endif
4758           s_count = 1;
4759           while (s_count < S) {
4760             if (place == last_place) {
4761               place = first_place;
4762             } else if (place == (int)(__kmp_affinity_num_masks - 1)) {
4763               place = 0;
4764             } else {
4765               place++;
4766             }
4767             s_count++;
4768           }
4769           if (rem && (gap_ct == gap)) {
4770             if (place == last_place) {
4771               place = first_place;
4772             } else if (place == (int)(__kmp_affinity_num_masks - 1)) {
4773               place = 0;
4774             } else {
4775               place++;
4776             }
4777             rem--;
4778             gap_ct = 0;
4779           }
4780           th->th.th_last_place = place;
4781           gap_ct++;
4782 
4783           if (place == last_place) {
4784             place = first_place;
4785           } else if (place == (int)(__kmp_affinity_num_masks - 1)) {
4786             place = 0;
4787           } else {
4788             place++;
4789           }
4790 
4791           KA_TRACE(100,
4792                    ("__kmp_partition_places: spread: T#%d(%d:%d) place %d "
4793                     "partition = [%d,%d], __kmp_affinity_num_masks: %u\n",
4794                     __kmp_gtid_from_thread(team->t.t_threads[f]), team->t.t_id,
4795                     f, th->th.th_new_place, th->th.th_first_place,
4796                     th->th.th_last_place, __kmp_affinity_num_masks));
4797         }
4798       } else {
4799         /* Having uniform space of available computation places I can create
4800            T partitions of round(P/T) size and put threads into the first
4801            place of each partition. */
4802         double current = static_cast<double>(masters_place);
4803         double spacing =
4804             (static_cast<double>(n_places + 1) / static_cast<double>(n_th));
4805         int first, last;
4806         kmp_info_t *th;
4807 
4808         thidx = n_th + 1;
4809         if (update_master_only == 1)
4810           thidx = 1;
4811         for (f = 0; f < thidx; f++) {
4812           first = static_cast<int>(current);
4813           last = static_cast<int>(current + spacing) - 1;
4814           KMP_DEBUG_ASSERT(last >= first);
4815           if (first >= n_places) {
4816             if (masters_place) {
4817               first -= n_places;
4818               last -= n_places;
4819               if (first == (masters_place + 1)) {
4820                 KMP_DEBUG_ASSERT(f == n_th);
4821                 first--;
4822               }
4823               if (last == masters_place) {
4824                 KMP_DEBUG_ASSERT(f == (n_th - 1));
4825                 last--;
4826               }
4827             } else {
4828               KMP_DEBUG_ASSERT(f == n_th);
4829               first = 0;
4830               last = 0;
4831             }
4832           }
4833           if (last >= n_places) {
4834             last = (n_places - 1);
4835           }
4836           place = first;
4837           current += spacing;
4838           if (f < n_th) {
4839             KMP_DEBUG_ASSERT(0 <= first);
4840             KMP_DEBUG_ASSERT(n_places > first);
4841             KMP_DEBUG_ASSERT(0 <= last);
4842             KMP_DEBUG_ASSERT(n_places > last);
4843             KMP_DEBUG_ASSERT(last_place >= first_place);
4844             th = team->t.t_threads[f];
4845             KMP_DEBUG_ASSERT(th);
4846             th->th.th_first_place = first;
4847             th->th.th_new_place = place;
4848             th->th.th_last_place = last;
4849 #if OMP_50_ENABLED
4850             if (__kmp_display_affinity && place != th->th.th_current_place &&
4851                 team->t.t_display_affinity != 1) {
4852               team->t.t_display_affinity = 1;
4853             }
4854 #endif
4855             KA_TRACE(100,
4856                      ("__kmp_partition_places: spread: T#%d(%d:%d) place %d "
4857                       "partition = [%d,%d], spacing = %.4f\n",
4858                       __kmp_gtid_from_thread(team->t.t_threads[f]),
4859                       team->t.t_id, f, th->th.th_new_place,
4860                       th->th.th_first_place, th->th.th_last_place, spacing));
4861           }
4862         }
4863       }
4864       KMP_DEBUG_ASSERT(update_master_only || place == masters_place);
4865     } else {
4866       int S, rem, gap, s_count;
4867       S = n_th / n_places;
4868       s_count = 0;
4869       rem = n_th - (S * n_places);
4870       gap = rem > 0 ? n_places / rem : n_places;
4871       int place = masters_place;
4872       int gap_ct = gap;
4873       thidx = n_th;
4874       if (update_master_only == 1)
4875         thidx = 1;
4876       for (f = 0; f < thidx; f++) {
4877         kmp_info_t *th = team->t.t_threads[f];
4878         KMP_DEBUG_ASSERT(th != NULL);
4879 
4880         th->th.th_first_place = place;
4881         th->th.th_last_place = place;
4882         th->th.th_new_place = place;
4883 #if OMP_50_ENABLED
4884         if (__kmp_display_affinity && place != th->th.th_current_place &&
4885             team->t.t_display_affinity != 1) {
4886           team->t.t_display_affinity = 1;
4887         }
4888 #endif
4889         s_count++;
4890 
4891         if ((s_count == S) && rem && (gap_ct == gap)) {
4892           // do nothing, add an extra thread to place on next iteration
4893         } else if ((s_count == S + 1) && rem && (gap_ct == gap)) {
4894           // we added an extra thread to this place; move on to next place
4895           if (place == last_place) {
4896             place = first_place;
4897           } else if (place == (int)(__kmp_affinity_num_masks - 1)) {
4898             place = 0;
4899           } else {
4900             place++;
4901           }
4902           s_count = 0;
4903           gap_ct = 1;
4904           rem--;
4905         } else if (s_count == S) { // place is full; don't add extra thread
4906           if (place == last_place) {
4907             place = first_place;
4908           } else if (place == (int)(__kmp_affinity_num_masks - 1)) {
4909             place = 0;
4910           } else {
4911             place++;
4912           }
4913           gap_ct++;
4914           s_count = 0;
4915         }
4916 
4917         KA_TRACE(100, ("__kmp_partition_places: spread: T#%d(%d:%d) place %d "
4918                        "partition = [%d,%d]\n",
4919                        __kmp_gtid_from_thread(team->t.t_threads[f]),
4920                        team->t.t_id, f, th->th.th_new_place,
4921                        th->th.th_first_place, th->th.th_last_place));
4922       }
4923       KMP_DEBUG_ASSERT(update_master_only || place == masters_place);
4924     }
4925   } break;
4926 
4927   default:
4928     break;
4929   }
4930 
4931   KA_TRACE(20, ("__kmp_partition_places: exit T#%d\n", team->t.t_id));
4932 }
4933 
4934 #endif /* OMP_40_ENABLED && KMP_AFFINITY_SUPPORTED */
4935 
4936 /* allocate a new team data structure to use.  take one off of the free pool if
4937    available */
4938 kmp_team_t *
4939 __kmp_allocate_team(kmp_root_t *root, int new_nproc, int max_nproc,
4940 #if OMPT_SUPPORT
4941                     ompt_data_t ompt_parallel_data,
4942 #endif
4943 #if OMP_40_ENABLED
4944                     kmp_proc_bind_t new_proc_bind,
4945 #endif
4946                     kmp_internal_control_t *new_icvs,
4947                     int argc USE_NESTED_HOT_ARG(kmp_info_t *master)) {
4948   KMP_TIME_DEVELOPER_PARTITIONED_BLOCK(KMP_allocate_team);
4949   int f;
4950   kmp_team_t *team;
4951   int use_hot_team = !root->r.r_active;
4952   int level = 0;
4953 
4954   KA_TRACE(20, ("__kmp_allocate_team: called\n"));
4955   KMP_DEBUG_ASSERT(new_nproc >= 1 && argc >= 0);
4956   KMP_DEBUG_ASSERT(max_nproc >= new_nproc);
4957   KMP_MB();
4958 
4959 #if KMP_NESTED_HOT_TEAMS
4960   kmp_hot_team_ptr_t *hot_teams;
4961   if (master) {
4962     team = master->th.th_team;
4963     level = team->t.t_active_level;
4964     if (master->th.th_teams_microtask) { // in teams construct?
4965       if (master->th.th_teams_size.nteams > 1 &&
4966           ( // #teams > 1
4967               team->t.t_pkfn ==
4968                   (microtask_t)__kmp_teams_master || // inner fork of the teams
4969               master->th.th_teams_level <
4970                   team->t.t_level)) { // or nested parallel inside the teams
4971         ++level; // not increment if #teams==1, or for outer fork of the teams;
4972         // increment otherwise
4973       }
4974     }
4975     hot_teams = master->th.th_hot_teams;
4976     if (level < __kmp_hot_teams_max_level && hot_teams &&
4977         hot_teams[level]
4978             .hot_team) { // hot team has already been allocated for given level
4979       use_hot_team = 1;
4980     } else {
4981       use_hot_team = 0;
4982     }
4983   }
4984 #endif
4985   // Optimization to use a "hot" team
4986   if (use_hot_team && new_nproc > 1) {
4987     KMP_DEBUG_ASSERT(new_nproc <= max_nproc);
4988 #if KMP_NESTED_HOT_TEAMS
4989     team = hot_teams[level].hot_team;
4990 #else
4991     team = root->r.r_hot_team;
4992 #endif
4993 #if KMP_DEBUG
4994     if (__kmp_tasking_mode != tskm_immediate_exec) {
4995       KA_TRACE(20, ("__kmp_allocate_team: hot team task_team[0] = %p "
4996                     "task_team[1] = %p before reinit\n",
4997                     team->t.t_task_team[0], team->t.t_task_team[1]));
4998     }
4999 #endif
5000 
5001     // Has the number of threads changed?
5002     /* Let's assume the most common case is that the number of threads is
5003        unchanged, and put that case first. */
5004     if (team->t.t_nproc == new_nproc) { // Check changes in number of threads
5005       KA_TRACE(20, ("__kmp_allocate_team: reusing hot team\n"));
5006       // This case can mean that omp_set_num_threads() was called and the hot
5007       // team size was already reduced, so we check the special flag
5008       if (team->t.t_size_changed == -1) {
5009         team->t.t_size_changed = 1;
5010       } else {
5011         KMP_CHECK_UPDATE(team->t.t_size_changed, 0);
5012       }
5013 
5014       // TODO???: team->t.t_max_active_levels = new_max_active_levels;
5015       kmp_r_sched_t new_sched = new_icvs->sched;
5016       // set master's schedule as new run-time schedule
5017       KMP_CHECK_UPDATE(team->t.t_sched.sched, new_sched.sched);
5018 
5019       __kmp_reinitialize_team(team, new_icvs,
5020                               root->r.r_uber_thread->th.th_ident);
5021 
5022       KF_TRACE(10, ("__kmp_allocate_team2: T#%d, this_thread=%p team=%p\n", 0,
5023                     team->t.t_threads[0], team));
5024       __kmp_push_current_task_to_thread(team->t.t_threads[0], team, 0);
5025 
5026 #if OMP_40_ENABLED
5027 #if KMP_AFFINITY_SUPPORTED
5028       if ((team->t.t_size_changed == 0) &&
5029           (team->t.t_proc_bind == new_proc_bind)) {
5030         if (new_proc_bind == proc_bind_spread) {
5031           __kmp_partition_places(
5032               team, 1); // add flag to update only master for spread
5033         }
5034         KA_TRACE(200, ("__kmp_allocate_team: reusing hot team #%d bindings: "
5035                        "proc_bind = %d, partition = [%d,%d]\n",
5036                        team->t.t_id, new_proc_bind, team->t.t_first_place,
5037                        team->t.t_last_place));
5038       } else {
5039         KMP_CHECK_UPDATE(team->t.t_proc_bind, new_proc_bind);
5040         __kmp_partition_places(team);
5041       }
5042 #else
5043       KMP_CHECK_UPDATE(team->t.t_proc_bind, new_proc_bind);
5044 #endif /* KMP_AFFINITY_SUPPORTED */
5045 #endif /* OMP_40_ENABLED */
5046     } else if (team->t.t_nproc > new_nproc) {
5047       KA_TRACE(20,
5048                ("__kmp_allocate_team: decreasing hot team thread count to %d\n",
5049                 new_nproc));
5050 
5051       team->t.t_size_changed = 1;
5052 #if KMP_NESTED_HOT_TEAMS
5053       if (__kmp_hot_teams_mode == 0) {
5054         // AC: saved number of threads should correspond to team's value in this
5055         // mode, can be bigger in mode 1, when hot team has threads in reserve
5056         KMP_DEBUG_ASSERT(hot_teams[level].hot_team_nth == team->t.t_nproc);
5057         hot_teams[level].hot_team_nth = new_nproc;
5058 #endif // KMP_NESTED_HOT_TEAMS
5059         /* release the extra threads we don't need any more */
5060         for (f = new_nproc; f < team->t.t_nproc; f++) {
5061           KMP_DEBUG_ASSERT(team->t.t_threads[f]);
5062           if (__kmp_tasking_mode != tskm_immediate_exec) {
5063             // When decreasing team size, threads no longer in the team should
5064             // unref task team.
5065             team->t.t_threads[f]->th.th_task_team = NULL;
5066           }
5067           __kmp_free_thread(team->t.t_threads[f]);
5068           team->t.t_threads[f] = NULL;
5069         }
5070 #if KMP_NESTED_HOT_TEAMS
5071       } // (__kmp_hot_teams_mode == 0)
5072       else {
5073         // When keeping extra threads in team, switch threads to wait on own
5074         // b_go flag
5075         for (f = new_nproc; f < team->t.t_nproc; ++f) {
5076           KMP_DEBUG_ASSERT(team->t.t_threads[f]);
5077           kmp_balign_t *balign = team->t.t_threads[f]->th.th_bar;
5078           for (int b = 0; b < bs_last_barrier; ++b) {
5079             if (balign[b].bb.wait_flag == KMP_BARRIER_PARENT_FLAG) {
5080               balign[b].bb.wait_flag = KMP_BARRIER_SWITCH_TO_OWN_FLAG;
5081             }
5082             KMP_CHECK_UPDATE(balign[b].bb.leaf_kids, 0);
5083           }
5084         }
5085       }
5086 #endif // KMP_NESTED_HOT_TEAMS
5087       team->t.t_nproc = new_nproc;
5088       // TODO???: team->t.t_max_active_levels = new_max_active_levels;
5089       KMP_CHECK_UPDATE(team->t.t_sched.sched, new_icvs->sched.sched);
5090       __kmp_reinitialize_team(team, new_icvs,
5091                               root->r.r_uber_thread->th.th_ident);
5092 
5093       // Update remaining threads
5094       for (f = 0; f < new_nproc; ++f) {
5095         team->t.t_threads[f]->th.th_team_nproc = new_nproc;
5096       }
5097 
5098       // restore the current task state of the master thread: should be the
5099       // implicit task
5100       KF_TRACE(10, ("__kmp_allocate_team: T#%d, this_thread=%p team=%p\n", 0,
5101                     team->t.t_threads[0], team));
5102 
5103       __kmp_push_current_task_to_thread(team->t.t_threads[0], team, 0);
5104 
5105 #ifdef KMP_DEBUG
5106       for (f = 0; f < team->t.t_nproc; f++) {
5107         KMP_DEBUG_ASSERT(team->t.t_threads[f] &&
5108                          team->t.t_threads[f]->th.th_team_nproc ==
5109                              team->t.t_nproc);
5110       }
5111 #endif
5112 
5113 #if OMP_40_ENABLED
5114       KMP_CHECK_UPDATE(team->t.t_proc_bind, new_proc_bind);
5115 #if KMP_AFFINITY_SUPPORTED
5116       __kmp_partition_places(team);
5117 #endif
5118 #endif
5119     } else { // team->t.t_nproc < new_nproc
5120 #if KMP_OS_LINUX && KMP_AFFINITY_SUPPORTED
5121       kmp_affin_mask_t *old_mask;
5122       if (KMP_AFFINITY_CAPABLE()) {
5123         KMP_CPU_ALLOC(old_mask);
5124       }
5125 #endif
5126 
5127       KA_TRACE(20,
5128                ("__kmp_allocate_team: increasing hot team thread count to %d\n",
5129                 new_nproc));
5130 
5131       team->t.t_size_changed = 1;
5132 
5133 #if KMP_NESTED_HOT_TEAMS
5134       int avail_threads = hot_teams[level].hot_team_nth;
5135       if (new_nproc < avail_threads)
5136         avail_threads = new_nproc;
5137       kmp_info_t **other_threads = team->t.t_threads;
5138       for (f = team->t.t_nproc; f < avail_threads; ++f) {
5139         // Adjust barrier data of reserved threads (if any) of the team
5140         // Other data will be set in __kmp_initialize_info() below.
5141         int b;
5142         kmp_balign_t *balign = other_threads[f]->th.th_bar;
5143         for (b = 0; b < bs_last_barrier; ++b) {
5144           balign[b].bb.b_arrived = team->t.t_bar[b].b_arrived;
5145           KMP_DEBUG_ASSERT(balign[b].bb.wait_flag != KMP_BARRIER_PARENT_FLAG);
5146 #if USE_DEBUGGER
5147           balign[b].bb.b_worker_arrived = team->t.t_bar[b].b_team_arrived;
5148 #endif
5149         }
5150       }
5151       if (hot_teams[level].hot_team_nth >= new_nproc) {
5152         // we have all needed threads in reserve, no need to allocate any
5153         // this only possible in mode 1, cannot have reserved threads in mode 0
5154         KMP_DEBUG_ASSERT(__kmp_hot_teams_mode == 1);
5155         team->t.t_nproc = new_nproc; // just get reserved threads involved
5156       } else {
5157         // we may have some threads in reserve, but not enough
5158         team->t.t_nproc =
5159             hot_teams[level]
5160                 .hot_team_nth; // get reserved threads involved if any
5161         hot_teams[level].hot_team_nth = new_nproc; // adjust hot team max size
5162 #endif // KMP_NESTED_HOT_TEAMS
5163         if (team->t.t_max_nproc < new_nproc) {
5164           /* reallocate larger arrays */
5165           __kmp_reallocate_team_arrays(team, new_nproc);
5166           __kmp_reinitialize_team(team, new_icvs, NULL);
5167         }
5168 
5169 #if KMP_OS_LINUX && KMP_AFFINITY_SUPPORTED
5170         /* Temporarily set full mask for master thread before creation of
5171            workers. The reason is that workers inherit the affinity from master,
5172            so if a lot of workers are created on the single core quickly, they
5173            don't get a chance to set their own affinity for a long time. */
5174         __kmp_set_thread_affinity_mask_full_tmp(old_mask);
5175 #endif
5176 
5177         /* allocate new threads for the hot team */
5178         for (f = team->t.t_nproc; f < new_nproc; f++) {
5179           kmp_info_t *new_worker = __kmp_allocate_thread(root, team, f);
5180           KMP_DEBUG_ASSERT(new_worker);
5181           team->t.t_threads[f] = new_worker;
5182 
5183           KA_TRACE(20,
5184                    ("__kmp_allocate_team: team %d init T#%d arrived: "
5185                     "join=%llu, plain=%llu\n",
5186                     team->t.t_id, __kmp_gtid_from_tid(f, team), team->t.t_id, f,
5187                     team->t.t_bar[bs_forkjoin_barrier].b_arrived,
5188                     team->t.t_bar[bs_plain_barrier].b_arrived));
5189 
5190           { // Initialize barrier data for new threads.
5191             int b;
5192             kmp_balign_t *balign = new_worker->th.th_bar;
5193             for (b = 0; b < bs_last_barrier; ++b) {
5194               balign[b].bb.b_arrived = team->t.t_bar[b].b_arrived;
5195               KMP_DEBUG_ASSERT(balign[b].bb.wait_flag !=
5196                                KMP_BARRIER_PARENT_FLAG);
5197 #if USE_DEBUGGER
5198               balign[b].bb.b_worker_arrived = team->t.t_bar[b].b_team_arrived;
5199 #endif
5200             }
5201           }
5202         }
5203 
5204 #if KMP_OS_LINUX && KMP_AFFINITY_SUPPORTED
5205         if (KMP_AFFINITY_CAPABLE()) {
5206           /* Restore initial master thread's affinity mask */
5207           __kmp_set_system_affinity(old_mask, TRUE);
5208           KMP_CPU_FREE(old_mask);
5209         }
5210 #endif
5211 #if KMP_NESTED_HOT_TEAMS
5212       } // end of check of t_nproc vs. new_nproc vs. hot_team_nth
5213 #endif // KMP_NESTED_HOT_TEAMS
5214       /* make sure everyone is syncronized */
5215       int old_nproc = team->t.t_nproc; // save old value and use to update only
5216       // new threads below
5217       __kmp_initialize_team(team, new_nproc, new_icvs,
5218                             root->r.r_uber_thread->th.th_ident);
5219 
5220       /* reinitialize the threads */
5221       KMP_DEBUG_ASSERT(team->t.t_nproc == new_nproc);
5222       for (f = 0; f < team->t.t_nproc; ++f)
5223         __kmp_initialize_info(team->t.t_threads[f], team, f,
5224                               __kmp_gtid_from_tid(f, team));
5225 
5226       if (level) { // set th_task_state for new threads in nested hot team
5227         // __kmp_initialize_info() no longer zeroes th_task_state, so we should
5228         // only need to set the th_task_state for the new threads. th_task_state
5229         // for master thread will not be accurate until after this in
5230         // __kmp_fork_call(), so we look to the master's memo_stack to get the
5231         // correct value.
5232         for (f = old_nproc; f < team->t.t_nproc; ++f)
5233           team->t.t_threads[f]->th.th_task_state =
5234               team->t.t_threads[0]->th.th_task_state_memo_stack[level];
5235       } else { // set th_task_state for new threads in non-nested hot team
5236         int old_state =
5237             team->t.t_threads[0]->th.th_task_state; // copy master's state
5238         for (f = old_nproc; f < team->t.t_nproc; ++f)
5239           team->t.t_threads[f]->th.th_task_state = old_state;
5240       }
5241 
5242 #ifdef KMP_DEBUG
5243       for (f = 0; f < team->t.t_nproc; ++f) {
5244         KMP_DEBUG_ASSERT(team->t.t_threads[f] &&
5245                          team->t.t_threads[f]->th.th_team_nproc ==
5246                              team->t.t_nproc);
5247       }
5248 #endif
5249 
5250 #if OMP_40_ENABLED
5251       KMP_CHECK_UPDATE(team->t.t_proc_bind, new_proc_bind);
5252 #if KMP_AFFINITY_SUPPORTED
5253       __kmp_partition_places(team);
5254 #endif
5255 #endif
5256     } // Check changes in number of threads
5257 
5258 #if OMP_40_ENABLED
5259     kmp_info_t *master = team->t.t_threads[0];
5260     if (master->th.th_teams_microtask) {
5261       for (f = 1; f < new_nproc; ++f) {
5262         // propagate teams construct specific info to workers
5263         kmp_info_t *thr = team->t.t_threads[f];
5264         thr->th.th_teams_microtask = master->th.th_teams_microtask;
5265         thr->th.th_teams_level = master->th.th_teams_level;
5266         thr->th.th_teams_size = master->th.th_teams_size;
5267       }
5268     }
5269 #endif /* OMP_40_ENABLED */
5270 #if KMP_NESTED_HOT_TEAMS
5271     if (level) {
5272       // Sync barrier state for nested hot teams, not needed for outermost hot
5273       // team.
5274       for (f = 1; f < new_nproc; ++f) {
5275         kmp_info_t *thr = team->t.t_threads[f];
5276         int b;
5277         kmp_balign_t *balign = thr->th.th_bar;
5278         for (b = 0; b < bs_last_barrier; ++b) {
5279           balign[b].bb.b_arrived = team->t.t_bar[b].b_arrived;
5280           KMP_DEBUG_ASSERT(balign[b].bb.wait_flag != KMP_BARRIER_PARENT_FLAG);
5281 #if USE_DEBUGGER
5282           balign[b].bb.b_worker_arrived = team->t.t_bar[b].b_team_arrived;
5283 #endif
5284         }
5285       }
5286     }
5287 #endif // KMP_NESTED_HOT_TEAMS
5288 
5289     /* reallocate space for arguments if necessary */
5290     __kmp_alloc_argv_entries(argc, team, TRUE);
5291     KMP_CHECK_UPDATE(team->t.t_argc, argc);
5292     // The hot team re-uses the previous task team,
5293     // if untouched during the previous release->gather phase.
5294 
5295     KF_TRACE(10, (" hot_team = %p\n", team));
5296 
5297 #if KMP_DEBUG
5298     if (__kmp_tasking_mode != tskm_immediate_exec) {
5299       KA_TRACE(20, ("__kmp_allocate_team: hot team task_team[0] = %p "
5300                     "task_team[1] = %p after reinit\n",
5301                     team->t.t_task_team[0], team->t.t_task_team[1]));
5302     }
5303 #endif
5304 
5305 #if OMPT_SUPPORT
5306     __ompt_team_assign_id(team, ompt_parallel_data);
5307 #endif
5308 
5309     KMP_MB();
5310 
5311     return team;
5312   }
5313 
5314   /* next, let's try to take one from the team pool */
5315   KMP_MB();
5316   for (team = CCAST(kmp_team_t *, __kmp_team_pool); (team);) {
5317     /* TODO: consider resizing undersized teams instead of reaping them, now
5318        that we have a resizing mechanism */
5319     if (team->t.t_max_nproc >= max_nproc) {
5320       /* take this team from the team pool */
5321       __kmp_team_pool = team->t.t_next_pool;
5322 
5323       /* setup the team for fresh use */
5324       __kmp_initialize_team(team, new_nproc, new_icvs, NULL);
5325 
5326       KA_TRACE(20, ("__kmp_allocate_team: setting task_team[0] %p and "
5327                     "task_team[1] %p to NULL\n",
5328                     &team->t.t_task_team[0], &team->t.t_task_team[1]));
5329       team->t.t_task_team[0] = NULL;
5330       team->t.t_task_team[1] = NULL;
5331 
5332       /* reallocate space for arguments if necessary */
5333       __kmp_alloc_argv_entries(argc, team, TRUE);
5334       KMP_CHECK_UPDATE(team->t.t_argc, argc);
5335 
5336       KA_TRACE(
5337           20, ("__kmp_allocate_team: team %d init arrived: join=%u, plain=%u\n",
5338                team->t.t_id, KMP_INIT_BARRIER_STATE, KMP_INIT_BARRIER_STATE));
5339       { // Initialize barrier data.
5340         int b;
5341         for (b = 0; b < bs_last_barrier; ++b) {
5342           team->t.t_bar[b].b_arrived = KMP_INIT_BARRIER_STATE;
5343 #if USE_DEBUGGER
5344           team->t.t_bar[b].b_master_arrived = 0;
5345           team->t.t_bar[b].b_team_arrived = 0;
5346 #endif
5347         }
5348       }
5349 
5350 #if OMP_40_ENABLED
5351       team->t.t_proc_bind = new_proc_bind;
5352 #endif
5353 
5354       KA_TRACE(20, ("__kmp_allocate_team: using team from pool %d.\n",
5355                     team->t.t_id));
5356 
5357 #if OMPT_SUPPORT
5358       __ompt_team_assign_id(team, ompt_parallel_data);
5359 #endif
5360 
5361       KMP_MB();
5362 
5363       return team;
5364     }
5365 
5366     /* reap team if it is too small, then loop back and check the next one */
5367     // not sure if this is wise, but, will be redone during the hot-teams
5368     // rewrite.
5369     /* TODO: Use technique to find the right size hot-team, don't reap them */
5370     team = __kmp_reap_team(team);
5371     __kmp_team_pool = team;
5372   }
5373 
5374   /* nothing available in the pool, no matter, make a new team! */
5375   KMP_MB();
5376   team = (kmp_team_t *)__kmp_allocate(sizeof(kmp_team_t));
5377 
5378   /* and set it up */
5379   team->t.t_max_nproc = max_nproc;
5380   /* NOTE well, for some reason allocating one big buffer and dividing it up
5381      seems to really hurt performance a lot on the P4, so, let's not use this */
5382   __kmp_allocate_team_arrays(team, max_nproc);
5383 
5384   KA_TRACE(20, ("__kmp_allocate_team: making a new team\n"));
5385   __kmp_initialize_team(team, new_nproc, new_icvs, NULL);
5386 
5387   KA_TRACE(20, ("__kmp_allocate_team: setting task_team[0] %p and task_team[1] "
5388                 "%p to NULL\n",
5389                 &team->t.t_task_team[0], &team->t.t_task_team[1]));
5390   team->t.t_task_team[0] = NULL; // to be removed, as __kmp_allocate zeroes
5391   // memory, no need to duplicate
5392   team->t.t_task_team[1] = NULL; // to be removed, as __kmp_allocate zeroes
5393   // memory, no need to duplicate
5394 
5395   if (__kmp_storage_map) {
5396     __kmp_print_team_storage_map("team", team, team->t.t_id, new_nproc);
5397   }
5398 
5399   /* allocate space for arguments */
5400   __kmp_alloc_argv_entries(argc, team, FALSE);
5401   team->t.t_argc = argc;
5402 
5403   KA_TRACE(20,
5404            ("__kmp_allocate_team: team %d init arrived: join=%u, plain=%u\n",
5405             team->t.t_id, KMP_INIT_BARRIER_STATE, KMP_INIT_BARRIER_STATE));
5406   { // Initialize barrier data.
5407     int b;
5408     for (b = 0; b < bs_last_barrier; ++b) {
5409       team->t.t_bar[b].b_arrived = KMP_INIT_BARRIER_STATE;
5410 #if USE_DEBUGGER
5411       team->t.t_bar[b].b_master_arrived = 0;
5412       team->t.t_bar[b].b_team_arrived = 0;
5413 #endif
5414     }
5415   }
5416 
5417 #if OMP_40_ENABLED
5418   team->t.t_proc_bind = new_proc_bind;
5419 #endif
5420 
5421 #if OMPT_SUPPORT
5422   __ompt_team_assign_id(team, ompt_parallel_data);
5423   team->t.ompt_serialized_team_info = NULL;
5424 #endif
5425 
5426   KMP_MB();
5427 
5428   KA_TRACE(20, ("__kmp_allocate_team: done creating a new team %d.\n",
5429                 team->t.t_id));
5430 
5431   return team;
5432 }
5433 
5434 /* TODO implement hot-teams at all levels */
5435 /* TODO implement lazy thread release on demand (disband request) */
5436 
5437 /* free the team.  return it to the team pool.  release all the threads
5438  * associated with it */
5439 void __kmp_free_team(kmp_root_t *root,
5440                      kmp_team_t *team USE_NESTED_HOT_ARG(kmp_info_t *master)) {
5441   int f;
5442   KA_TRACE(20, ("__kmp_free_team: T#%d freeing team %d\n", __kmp_get_gtid(),
5443                 team->t.t_id));
5444 
5445   /* verify state */
5446   KMP_DEBUG_ASSERT(root);
5447   KMP_DEBUG_ASSERT(team);
5448   KMP_DEBUG_ASSERT(team->t.t_nproc <= team->t.t_max_nproc);
5449   KMP_DEBUG_ASSERT(team->t.t_threads);
5450 
5451   int use_hot_team = team == root->r.r_hot_team;
5452 #if KMP_NESTED_HOT_TEAMS
5453   int level;
5454   kmp_hot_team_ptr_t *hot_teams;
5455   if (master) {
5456     level = team->t.t_active_level - 1;
5457     if (master->th.th_teams_microtask) { // in teams construct?
5458       if (master->th.th_teams_size.nteams > 1) {
5459         ++level; // level was not increased in teams construct for
5460         // team_of_masters
5461       }
5462       if (team->t.t_pkfn != (microtask_t)__kmp_teams_master &&
5463           master->th.th_teams_level == team->t.t_level) {
5464         ++level; // level was not increased in teams construct for
5465         // team_of_workers before the parallel
5466       } // team->t.t_level will be increased inside parallel
5467     }
5468     hot_teams = master->th.th_hot_teams;
5469     if (level < __kmp_hot_teams_max_level) {
5470       KMP_DEBUG_ASSERT(team == hot_teams[level].hot_team);
5471       use_hot_team = 1;
5472     }
5473   }
5474 #endif // KMP_NESTED_HOT_TEAMS
5475 
5476   /* team is done working */
5477   TCW_SYNC_PTR(team->t.t_pkfn,
5478                NULL); // Important for Debugging Support Library.
5479 #if KMP_OS_WINDOWS
5480   team->t.t_copyin_counter = 0; // init counter for possible reuse
5481 #endif
5482   // Do not reset pointer to parent team to NULL for hot teams.
5483 
5484   /* if we are non-hot team, release our threads */
5485   if (!use_hot_team) {
5486     if (__kmp_tasking_mode != tskm_immediate_exec) {
5487       // Wait for threads to reach reapable state
5488       for (f = 1; f < team->t.t_nproc; ++f) {
5489         KMP_DEBUG_ASSERT(team->t.t_threads[f]);
5490         kmp_info_t *th = team->t.t_threads[f];
5491         volatile kmp_uint32 *state = &th->th.th_reap_state;
5492         while (*state != KMP_SAFE_TO_REAP) {
5493 #if KMP_OS_WINDOWS
5494           // On Windows a thread can be killed at any time, check this
5495           DWORD ecode;
5496           if (!__kmp_is_thread_alive(th, &ecode)) {
5497             *state = KMP_SAFE_TO_REAP; // reset the flag for dead thread
5498             break;
5499           }
5500 #endif
5501           // first check if thread is sleeping
5502           kmp_flag_64 fl(&th->th.th_bar[bs_forkjoin_barrier].bb.b_go, th);
5503           if (fl.is_sleeping())
5504             fl.resume(__kmp_gtid_from_thread(th));
5505           KMP_CPU_PAUSE();
5506         }
5507       }
5508 
5509       // Delete task teams
5510       int tt_idx;
5511       for (tt_idx = 0; tt_idx < 2; ++tt_idx) {
5512         kmp_task_team_t *task_team = team->t.t_task_team[tt_idx];
5513         if (task_team != NULL) {
5514           for (f = 0; f < team->t.t_nproc; ++f) { // threads unref task teams
5515             KMP_DEBUG_ASSERT(team->t.t_threads[f]);
5516             team->t.t_threads[f]->th.th_task_team = NULL;
5517           }
5518           KA_TRACE(
5519               20,
5520               ("__kmp_free_team: T#%d deactivating task_team %p on team %d\n",
5521                __kmp_get_gtid(), task_team, team->t.t_id));
5522 #if KMP_NESTED_HOT_TEAMS
5523           __kmp_free_task_team(master, task_team);
5524 #endif
5525           team->t.t_task_team[tt_idx] = NULL;
5526         }
5527       }
5528     }
5529 
5530     // Reset pointer to parent team only for non-hot teams.
5531     team->t.t_parent = NULL;
5532     team->t.t_level = 0;
5533     team->t.t_active_level = 0;
5534 
5535     /* free the worker threads */
5536     for (f = 1; f < team->t.t_nproc; ++f) {
5537       KMP_DEBUG_ASSERT(team->t.t_threads[f]);
5538       __kmp_free_thread(team->t.t_threads[f]);
5539       team->t.t_threads[f] = NULL;
5540     }
5541 
5542     /* put the team back in the team pool */
5543     /* TODO limit size of team pool, call reap_team if pool too large */
5544     team->t.t_next_pool = CCAST(kmp_team_t *, __kmp_team_pool);
5545     __kmp_team_pool = (volatile kmp_team_t *)team;
5546   } else { // Check if team was created for the masters in a teams construct
5547     // See if first worker is a CG root
5548     KMP_DEBUG_ASSERT(team->t.t_threads[1] &&
5549                      team->t.t_threads[1]->th.th_cg_roots);
5550     if (team->t.t_threads[1]->th.th_cg_roots->cg_root == team->t.t_threads[1]) {
5551       // Clean up the CG root nodes on workers so that this team can be re-used
5552       for (f = 1; f < team->t.t_nproc; ++f) {
5553         kmp_info_t *thr = team->t.t_threads[f];
5554         KMP_DEBUG_ASSERT(thr && thr->th.th_cg_roots &&
5555                          thr->th.th_cg_roots->cg_root == thr);
5556         // Pop current CG root off list
5557         kmp_cg_root_t *tmp = thr->th.th_cg_roots;
5558         thr->th.th_cg_roots = tmp->up;
5559         KA_TRACE(100, ("__kmp_free_team: Thread %p popping node %p and moving"
5560                        " up to node %p. cg_nthreads was %d\n",
5561                        thr, tmp, thr->th.th_cg_roots, tmp->cg_nthreads));
5562         __kmp_free(tmp);
5563         // Restore current task's thread_limit from CG root
5564         if (thr->th.th_cg_roots)
5565           thr->th.th_current_task->td_icvs.thread_limit =
5566               thr->th.th_cg_roots->cg_thread_limit;
5567       }
5568     }
5569   }
5570 
5571   KMP_MB();
5572 }
5573 
5574 /* reap the team.  destroy it, reclaim all its resources and free its memory */
5575 kmp_team_t *__kmp_reap_team(kmp_team_t *team) {
5576   kmp_team_t *next_pool = team->t.t_next_pool;
5577 
5578   KMP_DEBUG_ASSERT(team);
5579   KMP_DEBUG_ASSERT(team->t.t_dispatch);
5580   KMP_DEBUG_ASSERT(team->t.t_disp_buffer);
5581   KMP_DEBUG_ASSERT(team->t.t_threads);
5582   KMP_DEBUG_ASSERT(team->t.t_argv);
5583 
5584   /* TODO clean the threads that are a part of this? */
5585 
5586   /* free stuff */
5587   __kmp_free_team_arrays(team);
5588   if (team->t.t_argv != &team->t.t_inline_argv[0])
5589     __kmp_free((void *)team->t.t_argv);
5590   __kmp_free(team);
5591 
5592   KMP_MB();
5593   return next_pool;
5594 }
5595 
5596 // Free the thread.  Don't reap it, just place it on the pool of available
5597 // threads.
5598 //
5599 // Changes for Quad issue 527845: We need a predictable OMP tid <-> gtid
5600 // binding for the affinity mechanism to be useful.
5601 //
5602 // Now, we always keep the free list (__kmp_thread_pool) sorted by gtid.
5603 // However, we want to avoid a potential performance problem by always
5604 // scanning through the list to find the correct point at which to insert
5605 // the thread (potential N**2 behavior).  To do this we keep track of the
5606 // last place a thread struct was inserted (__kmp_thread_pool_insert_pt).
5607 // With single-level parallelism, threads will always be added to the tail
5608 // of the list, kept track of by __kmp_thread_pool_insert_pt.  With nested
5609 // parallelism, all bets are off and we may need to scan through the entire
5610 // free list.
5611 //
5612 // This change also has a potentially large performance benefit, for some
5613 // applications.  Previously, as threads were freed from the hot team, they
5614 // would be placed back on the free list in inverse order.  If the hot team
5615 // grew back to it's original size, then the freed thread would be placed
5616 // back on the hot team in reverse order.  This could cause bad cache
5617 // locality problems on programs where the size of the hot team regularly
5618 // grew and shrunk.
5619 //
5620 // Now, for single-level parallelism, the OMP tid is alway == gtid.
5621 void __kmp_free_thread(kmp_info_t *this_th) {
5622   int gtid;
5623   kmp_info_t **scan;
5624 
5625   KA_TRACE(20, ("__kmp_free_thread: T#%d putting T#%d back on free pool.\n",
5626                 __kmp_get_gtid(), this_th->th.th_info.ds.ds_gtid));
5627 
5628   KMP_DEBUG_ASSERT(this_th);
5629 
5630   // When moving thread to pool, switch thread to wait on own b_go flag, and
5631   // uninitialized (NULL team).
5632   int b;
5633   kmp_balign_t *balign = this_th->th.th_bar;
5634   for (b = 0; b < bs_last_barrier; ++b) {
5635     if (balign[b].bb.wait_flag == KMP_BARRIER_PARENT_FLAG)
5636       balign[b].bb.wait_flag = KMP_BARRIER_SWITCH_TO_OWN_FLAG;
5637     balign[b].bb.team = NULL;
5638     balign[b].bb.leaf_kids = 0;
5639   }
5640   this_th->th.th_task_state = 0;
5641   this_th->th.th_reap_state = KMP_SAFE_TO_REAP;
5642 
5643   /* put thread back on the free pool */
5644   TCW_PTR(this_th->th.th_team, NULL);
5645   TCW_PTR(this_th->th.th_root, NULL);
5646   TCW_PTR(this_th->th.th_dispatch, NULL); /* NOT NEEDED */
5647 
5648   while (this_th->th.th_cg_roots) {
5649     this_th->th.th_cg_roots->cg_nthreads--;
5650     KA_TRACE(100, ("__kmp_free_thread: Thread %p decrement cg_nthreads on node"
5651                    " %p of thread  %p to %d\n",
5652                    this_th, this_th->th.th_cg_roots,
5653                    this_th->th.th_cg_roots->cg_root,
5654                    this_th->th.th_cg_roots->cg_nthreads));
5655     kmp_cg_root_t *tmp = this_th->th.th_cg_roots;
5656     if (tmp->cg_root == this_th) { // Thread is a cg_root
5657       KMP_DEBUG_ASSERT(tmp->cg_nthreads == 0);
5658       KA_TRACE(
5659           5, ("__kmp_free_thread: Thread %p freeing node %p\n", this_th, tmp));
5660       this_th->th.th_cg_roots = tmp->up;
5661       __kmp_free(tmp);
5662     } else { // Worker thread
5663       this_th->th.th_cg_roots = NULL;
5664       break;
5665     }
5666   }
5667 
5668   /* If the implicit task assigned to this thread can be used by other threads
5669    * -> multiple threads can share the data and try to free the task at
5670    * __kmp_reap_thread at exit. This duplicate use of the task data can happen
5671    * with higher probability when hot team is disabled but can occurs even when
5672    * the hot team is enabled */
5673   __kmp_free_implicit_task(this_th);
5674   this_th->th.th_current_task = NULL;
5675 
5676   // If the __kmp_thread_pool_insert_pt is already past the new insert
5677   // point, then we need to re-scan the entire list.
5678   gtid = this_th->th.th_info.ds.ds_gtid;
5679   if (__kmp_thread_pool_insert_pt != NULL) {
5680     KMP_DEBUG_ASSERT(__kmp_thread_pool != NULL);
5681     if (__kmp_thread_pool_insert_pt->th.th_info.ds.ds_gtid > gtid) {
5682       __kmp_thread_pool_insert_pt = NULL;
5683     }
5684   }
5685 
5686   // Scan down the list to find the place to insert the thread.
5687   // scan is the address of a link in the list, possibly the address of
5688   // __kmp_thread_pool itself.
5689   //
5690   // In the absence of nested parallism, the for loop will have 0 iterations.
5691   if (__kmp_thread_pool_insert_pt != NULL) {
5692     scan = &(__kmp_thread_pool_insert_pt->th.th_next_pool);
5693   } else {
5694     scan = CCAST(kmp_info_t **, &__kmp_thread_pool);
5695   }
5696   for (; (*scan != NULL) && ((*scan)->th.th_info.ds.ds_gtid < gtid);
5697        scan = &((*scan)->th.th_next_pool))
5698     ;
5699 
5700   // Insert the new element on the list, and set __kmp_thread_pool_insert_pt
5701   // to its address.
5702   TCW_PTR(this_th->th.th_next_pool, *scan);
5703   __kmp_thread_pool_insert_pt = *scan = this_th;
5704   KMP_DEBUG_ASSERT((this_th->th.th_next_pool == NULL) ||
5705                    (this_th->th.th_info.ds.ds_gtid <
5706                     this_th->th.th_next_pool->th.th_info.ds.ds_gtid));
5707   TCW_4(this_th->th.th_in_pool, TRUE);
5708   __kmp_thread_pool_nth++;
5709 
5710   TCW_4(__kmp_nth, __kmp_nth - 1);
5711 
5712 #ifdef KMP_ADJUST_BLOCKTIME
5713   /* Adjust blocktime back to user setting or default if necessary */
5714   /* Middle initialization might never have occurred                */
5715   if (!__kmp_env_blocktime && (__kmp_avail_proc > 0)) {
5716     KMP_DEBUG_ASSERT(__kmp_avail_proc > 0);
5717     if (__kmp_nth <= __kmp_avail_proc) {
5718       __kmp_zero_bt = FALSE;
5719     }
5720   }
5721 #endif /* KMP_ADJUST_BLOCKTIME */
5722 
5723   KMP_MB();
5724 }
5725 
5726 /* ------------------------------------------------------------------------ */
5727 
5728 void *__kmp_launch_thread(kmp_info_t *this_thr) {
5729   int gtid = this_thr->th.th_info.ds.ds_gtid;
5730   /*    void                 *stack_data;*/
5731   kmp_team_t *(*volatile pteam);
5732 
5733   KMP_MB();
5734   KA_TRACE(10, ("__kmp_launch_thread: T#%d start\n", gtid));
5735 
5736   if (__kmp_env_consistency_check) {
5737     this_thr->th.th_cons = __kmp_allocate_cons_stack(gtid); // ATT: Memory leak?
5738   }
5739 
5740 #if OMPT_SUPPORT
5741   ompt_data_t *thread_data;
5742   if (ompt_enabled.enabled) {
5743     thread_data = &(this_thr->th.ompt_thread_info.thread_data);
5744     *thread_data = ompt_data_none;
5745 
5746     this_thr->th.ompt_thread_info.state = ompt_state_overhead;
5747     this_thr->th.ompt_thread_info.wait_id = 0;
5748     this_thr->th.ompt_thread_info.idle_frame = OMPT_GET_FRAME_ADDRESS(0);
5749     if (ompt_enabled.ompt_callback_thread_begin) {
5750       ompt_callbacks.ompt_callback(ompt_callback_thread_begin)(
5751           ompt_thread_worker, thread_data);
5752     }
5753   }
5754 #endif
5755 
5756 #if OMPT_SUPPORT
5757   if (ompt_enabled.enabled) {
5758     this_thr->th.ompt_thread_info.state = ompt_state_idle;
5759   }
5760 #endif
5761   /* This is the place where threads wait for work */
5762   while (!TCR_4(__kmp_global.g.g_done)) {
5763     KMP_DEBUG_ASSERT(this_thr == __kmp_threads[gtid]);
5764     KMP_MB();
5765 
5766     /* wait for work to do */
5767     KA_TRACE(20, ("__kmp_launch_thread: T#%d waiting for work\n", gtid));
5768 
5769     /* No tid yet since not part of a team */
5770     __kmp_fork_barrier(gtid, KMP_GTID_DNE);
5771 
5772 #if OMPT_SUPPORT
5773     if (ompt_enabled.enabled) {
5774       this_thr->th.ompt_thread_info.state = ompt_state_overhead;
5775     }
5776 #endif
5777 
5778     pteam = (kmp_team_t * (*))(&this_thr->th.th_team);
5779 
5780     /* have we been allocated? */
5781     if (TCR_SYNC_PTR(*pteam) && !TCR_4(__kmp_global.g.g_done)) {
5782       /* we were just woken up, so run our new task */
5783       if (TCR_SYNC_PTR((*pteam)->t.t_pkfn) != NULL) {
5784         int rc;
5785         KA_TRACE(20,
5786                  ("__kmp_launch_thread: T#%d(%d:%d) invoke microtask = %p\n",
5787                   gtid, (*pteam)->t.t_id, __kmp_tid_from_gtid(gtid),
5788                   (*pteam)->t.t_pkfn));
5789 
5790         updateHWFPControl(*pteam);
5791 
5792 #if OMPT_SUPPORT
5793         if (ompt_enabled.enabled) {
5794           this_thr->th.ompt_thread_info.state = ompt_state_work_parallel;
5795         }
5796 #endif
5797 
5798         rc = (*pteam)->t.t_invoke(gtid);
5799         KMP_ASSERT(rc);
5800 
5801         KMP_MB();
5802         KA_TRACE(20, ("__kmp_launch_thread: T#%d(%d:%d) done microtask = %p\n",
5803                       gtid, (*pteam)->t.t_id, __kmp_tid_from_gtid(gtid),
5804                       (*pteam)->t.t_pkfn));
5805       }
5806 #if OMPT_SUPPORT
5807       if (ompt_enabled.enabled) {
5808         /* no frame set while outside task */
5809         __ompt_get_task_info_object(0)->frame.exit_frame = ompt_data_none;
5810 
5811         this_thr->th.ompt_thread_info.state = ompt_state_overhead;
5812       }
5813 #endif
5814       /* join barrier after parallel region */
5815       __kmp_join_barrier(gtid);
5816     }
5817   }
5818   TCR_SYNC_PTR((intptr_t)__kmp_global.g.g_done);
5819 
5820 #if OMPT_SUPPORT
5821   if (ompt_enabled.ompt_callback_thread_end) {
5822     ompt_callbacks.ompt_callback(ompt_callback_thread_end)(thread_data);
5823   }
5824 #endif
5825 
5826   this_thr->th.th_task_team = NULL;
5827   /* run the destructors for the threadprivate data for this thread */
5828   __kmp_common_destroy_gtid(gtid);
5829 
5830   KA_TRACE(10, ("__kmp_launch_thread: T#%d done\n", gtid));
5831   KMP_MB();
5832   return this_thr;
5833 }
5834 
5835 /* ------------------------------------------------------------------------ */
5836 
5837 void __kmp_internal_end_dest(void *specific_gtid) {
5838 #if KMP_COMPILER_ICC
5839 #pragma warning(push)
5840 #pragma warning(disable : 810) // conversion from "void *" to "int" may lose
5841 // significant bits
5842 #endif
5843   // Make sure no significant bits are lost
5844   int gtid = (kmp_intptr_t)specific_gtid - 1;
5845 #if KMP_COMPILER_ICC
5846 #pragma warning(pop)
5847 #endif
5848 
5849   KA_TRACE(30, ("__kmp_internal_end_dest: T#%d\n", gtid));
5850   /* NOTE: the gtid is stored as gitd+1 in the thread-local-storage
5851    * this is because 0 is reserved for the nothing-stored case */
5852 
5853   /* josh: One reason for setting the gtid specific data even when it is being
5854      destroyed by pthread is to allow gtid lookup through thread specific data
5855      (__kmp_gtid_get_specific).  Some of the code, especially stat code,
5856      that gets executed in the call to __kmp_internal_end_thread, actually
5857      gets the gtid through the thread specific data.  Setting it here seems
5858      rather inelegant and perhaps wrong, but allows __kmp_internal_end_thread
5859      to run smoothly.
5860      todo: get rid of this after we remove the dependence on
5861      __kmp_gtid_get_specific  */
5862   if (gtid >= 0 && KMP_UBER_GTID(gtid))
5863     __kmp_gtid_set_specific(gtid);
5864 #ifdef KMP_TDATA_GTID
5865   __kmp_gtid = gtid;
5866 #endif
5867   __kmp_internal_end_thread(gtid);
5868 }
5869 
5870 #if KMP_OS_UNIX && KMP_DYNAMIC_LIB
5871 
5872 // 2009-09-08 (lev): It looks the destructor does not work. In simple test cases
5873 // destructors work perfectly, but in real libomp.so I have no evidence it is
5874 // ever called. However, -fini linker option in makefile.mk works fine.
5875 
5876 __attribute__((destructor)) void __kmp_internal_end_dtor(void) {
5877   __kmp_internal_end_atexit();
5878 }
5879 
5880 void __kmp_internal_end_fini(void) { __kmp_internal_end_atexit(); }
5881 
5882 #endif
5883 
5884 /* [Windows] josh: when the atexit handler is called, there may still be more
5885    than one thread alive */
5886 void __kmp_internal_end_atexit(void) {
5887   KA_TRACE(30, ("__kmp_internal_end_atexit\n"));
5888   /* [Windows]
5889      josh: ideally, we want to completely shutdown the library in this atexit
5890      handler, but stat code that depends on thread specific data for gtid fails
5891      because that data becomes unavailable at some point during the shutdown, so
5892      we call __kmp_internal_end_thread instead. We should eventually remove the
5893      dependency on __kmp_get_specific_gtid in the stat code and use
5894      __kmp_internal_end_library to cleanly shutdown the library.
5895 
5896      // TODO: Can some of this comment about GVS be removed?
5897      I suspect that the offending stat code is executed when the calling thread
5898      tries to clean up a dead root thread's data structures, resulting in GVS
5899      code trying to close the GVS structures for that thread, but since the stat
5900      code uses __kmp_get_specific_gtid to get the gtid with the assumption that
5901      the calling thread is cleaning up itself instead of another thread, it get
5902      confused. This happens because allowing a thread to unregister and cleanup
5903      another thread is a recent modification for addressing an issue.
5904      Based on the current design (20050722), a thread may end up
5905      trying to unregister another thread only if thread death does not trigger
5906      the calling of __kmp_internal_end_thread.  For Linux* OS, there is the
5907      thread specific data destructor function to detect thread death. For
5908      Windows dynamic, there is DllMain(THREAD_DETACH). For Windows static, there
5909      is nothing.  Thus, the workaround is applicable only for Windows static
5910      stat library. */
5911   __kmp_internal_end_library(-1);
5912 #if KMP_OS_WINDOWS
5913   __kmp_close_console();
5914 #endif
5915 }
5916 
5917 static void __kmp_reap_thread(kmp_info_t *thread, int is_root) {
5918   // It is assumed __kmp_forkjoin_lock is acquired.
5919 
5920   int gtid;
5921 
5922   KMP_DEBUG_ASSERT(thread != NULL);
5923 
5924   gtid = thread->th.th_info.ds.ds_gtid;
5925 
5926   if (!is_root) {
5927     if (__kmp_dflt_blocktime != KMP_MAX_BLOCKTIME) {
5928       /* Assume the threads are at the fork barrier here */
5929       KA_TRACE(
5930           20, ("__kmp_reap_thread: releasing T#%d from fork barrier for reap\n",
5931                gtid));
5932       /* Need release fence here to prevent seg faults for tree forkjoin barrier
5933        * (GEH) */
5934       ANNOTATE_HAPPENS_BEFORE(thread);
5935       kmp_flag_64 flag(&thread->th.th_bar[bs_forkjoin_barrier].bb.b_go, thread);
5936       __kmp_release_64(&flag);
5937     }
5938 
5939     // Terminate OS thread.
5940     __kmp_reap_worker(thread);
5941 
5942     // The thread was killed asynchronously.  If it was actively
5943     // spinning in the thread pool, decrement the global count.
5944     //
5945     // There is a small timing hole here - if the worker thread was just waking
5946     // up after sleeping in the pool, had reset it's th_active_in_pool flag but
5947     // not decremented the global counter __kmp_thread_pool_active_nth yet, then
5948     // the global counter might not get updated.
5949     //
5950     // Currently, this can only happen as the library is unloaded,
5951     // so there are no harmful side effects.
5952     if (thread->th.th_active_in_pool) {
5953       thread->th.th_active_in_pool = FALSE;
5954       KMP_ATOMIC_DEC(&__kmp_thread_pool_active_nth);
5955       KMP_DEBUG_ASSERT(__kmp_thread_pool_active_nth >= 0);
5956     }
5957 
5958     // Decrement # of [worker] threads in the pool.
5959     KMP_DEBUG_ASSERT(__kmp_thread_pool_nth > 0);
5960     --__kmp_thread_pool_nth;
5961   }
5962 
5963   __kmp_free_implicit_task(thread);
5964 
5965 // Free the fast memory for tasking
5966 #if USE_FAST_MEMORY
5967   __kmp_free_fast_memory(thread);
5968 #endif /* USE_FAST_MEMORY */
5969 
5970   __kmp_suspend_uninitialize_thread(thread);
5971 
5972   KMP_DEBUG_ASSERT(__kmp_threads[gtid] == thread);
5973   TCW_SYNC_PTR(__kmp_threads[gtid], NULL);
5974 
5975   --__kmp_all_nth;
5976 // __kmp_nth was decremented when thread is added to the pool.
5977 
5978 #ifdef KMP_ADJUST_BLOCKTIME
5979   /* Adjust blocktime back to user setting or default if necessary */
5980   /* Middle initialization might never have occurred                */
5981   if (!__kmp_env_blocktime && (__kmp_avail_proc > 0)) {
5982     KMP_DEBUG_ASSERT(__kmp_avail_proc > 0);
5983     if (__kmp_nth <= __kmp_avail_proc) {
5984       __kmp_zero_bt = FALSE;
5985     }
5986   }
5987 #endif /* KMP_ADJUST_BLOCKTIME */
5988 
5989   /* free the memory being used */
5990   if (__kmp_env_consistency_check) {
5991     if (thread->th.th_cons) {
5992       __kmp_free_cons_stack(thread->th.th_cons);
5993       thread->th.th_cons = NULL;
5994     }
5995   }
5996 
5997   if (thread->th.th_pri_common != NULL) {
5998     __kmp_free(thread->th.th_pri_common);
5999     thread->th.th_pri_common = NULL;
6000   }
6001 
6002   if (thread->th.th_task_state_memo_stack != NULL) {
6003     __kmp_free(thread->th.th_task_state_memo_stack);
6004     thread->th.th_task_state_memo_stack = NULL;
6005   }
6006 
6007 #if KMP_USE_BGET
6008   if (thread->th.th_local.bget_data != NULL) {
6009     __kmp_finalize_bget(thread);
6010   }
6011 #endif
6012 
6013 #if KMP_AFFINITY_SUPPORTED
6014   if (thread->th.th_affin_mask != NULL) {
6015     KMP_CPU_FREE(thread->th.th_affin_mask);
6016     thread->th.th_affin_mask = NULL;
6017   }
6018 #endif /* KMP_AFFINITY_SUPPORTED */
6019 
6020 #if KMP_USE_HIER_SCHED
6021   if (thread->th.th_hier_bar_data != NULL) {
6022     __kmp_free(thread->th.th_hier_bar_data);
6023     thread->th.th_hier_bar_data = NULL;
6024   }
6025 #endif
6026 
6027   __kmp_reap_team(thread->th.th_serial_team);
6028   thread->th.th_serial_team = NULL;
6029   __kmp_free(thread);
6030 
6031   KMP_MB();
6032 
6033 } // __kmp_reap_thread
6034 
6035 static void __kmp_internal_end(void) {
6036   int i;
6037 
6038   /* First, unregister the library */
6039   __kmp_unregister_library();
6040 
6041 #if KMP_OS_WINDOWS
6042   /* In Win static library, we can't tell when a root actually dies, so we
6043      reclaim the data structures for any root threads that have died but not
6044      unregistered themselves, in order to shut down cleanly.
6045      In Win dynamic library we also can't tell when a thread dies.  */
6046   __kmp_reclaim_dead_roots(); // AC: moved here to always clean resources of
6047 // dead roots
6048 #endif
6049 
6050   for (i = 0; i < __kmp_threads_capacity; i++)
6051     if (__kmp_root[i])
6052       if (__kmp_root[i]->r.r_active)
6053         break;
6054   KMP_MB(); /* Flush all pending memory write invalidates.  */
6055   TCW_SYNC_4(__kmp_global.g.g_done, TRUE);
6056 
6057   if (i < __kmp_threads_capacity) {
6058 #if KMP_USE_MONITOR
6059     // 2009-09-08 (lev): Other alive roots found. Why do we kill the monitor??
6060     KMP_MB(); /* Flush all pending memory write invalidates.  */
6061 
6062     // Need to check that monitor was initialized before reaping it. If we are
6063     // called form __kmp_atfork_child (which sets __kmp_init_parallel = 0), then
6064     // __kmp_monitor will appear to contain valid data, but it is only valid in
6065     // the parent process, not the child.
6066     // New behavior (201008): instead of keying off of the flag
6067     // __kmp_init_parallel, the monitor thread creation is keyed off
6068     // of the new flag __kmp_init_monitor.
6069     __kmp_acquire_bootstrap_lock(&__kmp_monitor_lock);
6070     if (TCR_4(__kmp_init_monitor)) {
6071       __kmp_reap_monitor(&__kmp_monitor);
6072       TCW_4(__kmp_init_monitor, 0);
6073     }
6074     __kmp_release_bootstrap_lock(&__kmp_monitor_lock);
6075     KA_TRACE(10, ("__kmp_internal_end: monitor reaped\n"));
6076 #endif // KMP_USE_MONITOR
6077   } else {
6078 /* TODO move this to cleanup code */
6079 #ifdef KMP_DEBUG
6080     /* make sure that everything has properly ended */
6081     for (i = 0; i < __kmp_threads_capacity; i++) {
6082       if (__kmp_root[i]) {
6083         //                    KMP_ASSERT( ! KMP_UBER_GTID( i ) );         // AC:
6084         //                    there can be uber threads alive here
6085         KMP_ASSERT(!__kmp_root[i]->r.r_active); // TODO: can they be active?
6086       }
6087     }
6088 #endif
6089 
6090     KMP_MB();
6091 
6092     // Reap the worker threads.
6093     // This is valid for now, but be careful if threads are reaped sooner.
6094     while (__kmp_thread_pool != NULL) { // Loop thru all the thread in the pool.
6095       // Get the next thread from the pool.
6096       kmp_info_t *thread = CCAST(kmp_info_t *, __kmp_thread_pool);
6097       __kmp_thread_pool = thread->th.th_next_pool;
6098       // Reap it.
6099       KMP_DEBUG_ASSERT(thread->th.th_reap_state == KMP_SAFE_TO_REAP);
6100       thread->th.th_next_pool = NULL;
6101       thread->th.th_in_pool = FALSE;
6102       __kmp_reap_thread(thread, 0);
6103     }
6104     __kmp_thread_pool_insert_pt = NULL;
6105 
6106     // Reap teams.
6107     while (__kmp_team_pool != NULL) { // Loop thru all the teams in the pool.
6108       // Get the next team from the pool.
6109       kmp_team_t *team = CCAST(kmp_team_t *, __kmp_team_pool);
6110       __kmp_team_pool = team->t.t_next_pool;
6111       // Reap it.
6112       team->t.t_next_pool = NULL;
6113       __kmp_reap_team(team);
6114     }
6115 
6116     __kmp_reap_task_teams();
6117 
6118 #if KMP_OS_UNIX
6119     // Threads that are not reaped should not access any resources since they
6120     // are going to be deallocated soon, so the shutdown sequence should wait
6121     // until all threads either exit the final spin-waiting loop or begin
6122     // sleeping after the given blocktime.
6123     for (i = 0; i < __kmp_threads_capacity; i++) {
6124       kmp_info_t *thr = __kmp_threads[i];
6125       while (thr && KMP_ATOMIC_LD_ACQ(&thr->th.th_blocking))
6126         KMP_CPU_PAUSE();
6127     }
6128 #endif
6129 
6130     for (i = 0; i < __kmp_threads_capacity; ++i) {
6131       // TBD: Add some checking...
6132       // Something like KMP_DEBUG_ASSERT( __kmp_thread[ i ] == NULL );
6133     }
6134 
6135     /* Make sure all threadprivate destructors get run by joining with all
6136        worker threads before resetting this flag */
6137     TCW_SYNC_4(__kmp_init_common, FALSE);
6138 
6139     KA_TRACE(10, ("__kmp_internal_end: all workers reaped\n"));
6140     KMP_MB();
6141 
6142 #if KMP_USE_MONITOR
6143     // See note above: One of the possible fixes for CQ138434 / CQ140126
6144     //
6145     // FIXME: push both code fragments down and CSE them?
6146     // push them into __kmp_cleanup() ?
6147     __kmp_acquire_bootstrap_lock(&__kmp_monitor_lock);
6148     if (TCR_4(__kmp_init_monitor)) {
6149       __kmp_reap_monitor(&__kmp_monitor);
6150       TCW_4(__kmp_init_monitor, 0);
6151     }
6152     __kmp_release_bootstrap_lock(&__kmp_monitor_lock);
6153     KA_TRACE(10, ("__kmp_internal_end: monitor reaped\n"));
6154 #endif
6155   } /* else !__kmp_global.t_active */
6156   TCW_4(__kmp_init_gtid, FALSE);
6157   KMP_MB(); /* Flush all pending memory write invalidates.  */
6158 
6159   __kmp_cleanup();
6160 #if OMPT_SUPPORT
6161   ompt_fini();
6162 #endif
6163 }
6164 
6165 void __kmp_internal_end_library(int gtid_req) {
6166   /* if we have already cleaned up, don't try again, it wouldn't be pretty */
6167   /* this shouldn't be a race condition because __kmp_internal_end() is the
6168      only place to clear __kmp_serial_init */
6169   /* we'll check this later too, after we get the lock */
6170   // 2009-09-06: We do not set g_abort without setting g_done. This check looks
6171   // redundaant, because the next check will work in any case.
6172   if (__kmp_global.g.g_abort) {
6173     KA_TRACE(11, ("__kmp_internal_end_library: abort, exiting\n"));
6174     /* TODO abort? */
6175     return;
6176   }
6177   if (TCR_4(__kmp_global.g.g_done) || !__kmp_init_serial) {
6178     KA_TRACE(10, ("__kmp_internal_end_library: already finished\n"));
6179     return;
6180   }
6181 
6182   KMP_MB(); /* Flush all pending memory write invalidates.  */
6183 
6184   /* find out who we are and what we should do */
6185   {
6186     int gtid = (gtid_req >= 0) ? gtid_req : __kmp_gtid_get_specific();
6187     KA_TRACE(
6188         10, ("__kmp_internal_end_library: enter T#%d  (%d)\n", gtid, gtid_req));
6189     if (gtid == KMP_GTID_SHUTDOWN) {
6190       KA_TRACE(10, ("__kmp_internal_end_library: !__kmp_init_runtime, system "
6191                     "already shutdown\n"));
6192       return;
6193     } else if (gtid == KMP_GTID_MONITOR) {
6194       KA_TRACE(10, ("__kmp_internal_end_library: monitor thread, gtid not "
6195                     "registered, or system shutdown\n"));
6196       return;
6197     } else if (gtid == KMP_GTID_DNE) {
6198       KA_TRACE(10, ("__kmp_internal_end_library: gtid not registered or system "
6199                     "shutdown\n"));
6200       /* we don't know who we are, but we may still shutdown the library */
6201     } else if (KMP_UBER_GTID(gtid)) {
6202       /* unregister ourselves as an uber thread.  gtid is no longer valid */
6203       if (__kmp_root[gtid]->r.r_active) {
6204         __kmp_global.g.g_abort = -1;
6205         TCW_SYNC_4(__kmp_global.g.g_done, TRUE);
6206         KA_TRACE(10,
6207                  ("__kmp_internal_end_library: root still active, abort T#%d\n",
6208                   gtid));
6209         return;
6210       } else {
6211         KA_TRACE(
6212             10,
6213             ("__kmp_internal_end_library: unregistering sibling T#%d\n", gtid));
6214         __kmp_unregister_root_current_thread(gtid);
6215       }
6216     } else {
6217 /* worker threads may call this function through the atexit handler, if they
6218  * call exit() */
6219 /* For now, skip the usual subsequent processing and just dump the debug buffer.
6220    TODO: do a thorough shutdown instead */
6221 #ifdef DUMP_DEBUG_ON_EXIT
6222       if (__kmp_debug_buf)
6223         __kmp_dump_debug_buffer();
6224 #endif
6225       return;
6226     }
6227   }
6228   /* synchronize the termination process */
6229   __kmp_acquire_bootstrap_lock(&__kmp_initz_lock);
6230 
6231   /* have we already finished */
6232   if (__kmp_global.g.g_abort) {
6233     KA_TRACE(10, ("__kmp_internal_end_library: abort, exiting\n"));
6234     /* TODO abort? */
6235     __kmp_release_bootstrap_lock(&__kmp_initz_lock);
6236     return;
6237   }
6238   if (TCR_4(__kmp_global.g.g_done) || !__kmp_init_serial) {
6239     __kmp_release_bootstrap_lock(&__kmp_initz_lock);
6240     return;
6241   }
6242 
6243   /* We need this lock to enforce mutex between this reading of
6244      __kmp_threads_capacity and the writing by __kmp_register_root.
6245      Alternatively, we can use a counter of roots that is atomically updated by
6246      __kmp_get_global_thread_id_reg, __kmp_do_serial_initialize and
6247      __kmp_internal_end_*.  */
6248   __kmp_acquire_bootstrap_lock(&__kmp_forkjoin_lock);
6249 
6250   /* now we can safely conduct the actual termination */
6251   __kmp_internal_end();
6252 
6253   __kmp_release_bootstrap_lock(&__kmp_forkjoin_lock);
6254   __kmp_release_bootstrap_lock(&__kmp_initz_lock);
6255 
6256   KA_TRACE(10, ("__kmp_internal_end_library: exit\n"));
6257 
6258 #ifdef DUMP_DEBUG_ON_EXIT
6259   if (__kmp_debug_buf)
6260     __kmp_dump_debug_buffer();
6261 #endif
6262 
6263 #if KMP_OS_WINDOWS
6264   __kmp_close_console();
6265 #endif
6266 
6267   __kmp_fini_allocator();
6268 
6269 } // __kmp_internal_end_library
6270 
6271 void __kmp_internal_end_thread(int gtid_req) {
6272   int i;
6273 
6274   /* if we have already cleaned up, don't try again, it wouldn't be pretty */
6275   /* this shouldn't be a race condition because __kmp_internal_end() is the
6276    * only place to clear __kmp_serial_init */
6277   /* we'll check this later too, after we get the lock */
6278   // 2009-09-06: We do not set g_abort without setting g_done. This check looks
6279   // redundant, because the next check will work in any case.
6280   if (__kmp_global.g.g_abort) {
6281     KA_TRACE(11, ("__kmp_internal_end_thread: abort, exiting\n"));
6282     /* TODO abort? */
6283     return;
6284   }
6285   if (TCR_4(__kmp_global.g.g_done) || !__kmp_init_serial) {
6286     KA_TRACE(10, ("__kmp_internal_end_thread: already finished\n"));
6287     return;
6288   }
6289 
6290   KMP_MB(); /* Flush all pending memory write invalidates.  */
6291 
6292   /* find out who we are and what we should do */
6293   {
6294     int gtid = (gtid_req >= 0) ? gtid_req : __kmp_gtid_get_specific();
6295     KA_TRACE(10,
6296              ("__kmp_internal_end_thread: enter T#%d  (%d)\n", gtid, gtid_req));
6297     if (gtid == KMP_GTID_SHUTDOWN) {
6298       KA_TRACE(10, ("__kmp_internal_end_thread: !__kmp_init_runtime, system "
6299                     "already shutdown\n"));
6300       return;
6301     } else if (gtid == KMP_GTID_MONITOR) {
6302       KA_TRACE(10, ("__kmp_internal_end_thread: monitor thread, gtid not "
6303                     "registered, or system shutdown\n"));
6304       return;
6305     } else if (gtid == KMP_GTID_DNE) {
6306       KA_TRACE(10, ("__kmp_internal_end_thread: gtid not registered or system "
6307                     "shutdown\n"));
6308       return;
6309       /* we don't know who we are */
6310     } else if (KMP_UBER_GTID(gtid)) {
6311       /* unregister ourselves as an uber thread.  gtid is no longer valid */
6312       if (__kmp_root[gtid]->r.r_active) {
6313         __kmp_global.g.g_abort = -1;
6314         TCW_SYNC_4(__kmp_global.g.g_done, TRUE);
6315         KA_TRACE(10,
6316                  ("__kmp_internal_end_thread: root still active, abort T#%d\n",
6317                   gtid));
6318         return;
6319       } else {
6320         KA_TRACE(10, ("__kmp_internal_end_thread: unregistering sibling T#%d\n",
6321                       gtid));
6322         __kmp_unregister_root_current_thread(gtid);
6323       }
6324     } else {
6325       /* just a worker thread, let's leave */
6326       KA_TRACE(10, ("__kmp_internal_end_thread: worker thread T#%d\n", gtid));
6327 
6328       if (gtid >= 0) {
6329         __kmp_threads[gtid]->th.th_task_team = NULL;
6330       }
6331 
6332       KA_TRACE(10,
6333                ("__kmp_internal_end_thread: worker thread done, exiting T#%d\n",
6334                 gtid));
6335       return;
6336     }
6337   }
6338 #if KMP_DYNAMIC_LIB
6339 #if OMP_50_ENABLED
6340   if (__kmp_pause_status != kmp_hard_paused)
6341 #endif
6342   // AC: lets not shutdown the dynamic library at the exit of uber thread,
6343   // because we will better shutdown later in the library destructor.
6344   {
6345     KA_TRACE(10, ("__kmp_internal_end_thread: exiting T#%d\n", gtid_req));
6346     return;
6347   }
6348 #endif
6349   /* synchronize the termination process */
6350   __kmp_acquire_bootstrap_lock(&__kmp_initz_lock);
6351 
6352   /* have we already finished */
6353   if (__kmp_global.g.g_abort) {
6354     KA_TRACE(10, ("__kmp_internal_end_thread: abort, exiting\n"));
6355     /* TODO abort? */
6356     __kmp_release_bootstrap_lock(&__kmp_initz_lock);
6357     return;
6358   }
6359   if (TCR_4(__kmp_global.g.g_done) || !__kmp_init_serial) {
6360     __kmp_release_bootstrap_lock(&__kmp_initz_lock);
6361     return;
6362   }
6363 
6364   /* We need this lock to enforce mutex between this reading of
6365      __kmp_threads_capacity and the writing by __kmp_register_root.
6366      Alternatively, we can use a counter of roots that is atomically updated by
6367      __kmp_get_global_thread_id_reg, __kmp_do_serial_initialize and
6368      __kmp_internal_end_*.  */
6369 
6370   /* should we finish the run-time?  are all siblings done? */
6371   __kmp_acquire_bootstrap_lock(&__kmp_forkjoin_lock);
6372 
6373   for (i = 0; i < __kmp_threads_capacity; ++i) {
6374     if (KMP_UBER_GTID(i)) {
6375       KA_TRACE(
6376           10,
6377           ("__kmp_internal_end_thread: remaining sibling task: gtid==%d\n", i));
6378       __kmp_release_bootstrap_lock(&__kmp_forkjoin_lock);
6379       __kmp_release_bootstrap_lock(&__kmp_initz_lock);
6380       return;
6381     }
6382   }
6383 
6384   /* now we can safely conduct the actual termination */
6385 
6386   __kmp_internal_end();
6387 
6388   __kmp_release_bootstrap_lock(&__kmp_forkjoin_lock);
6389   __kmp_release_bootstrap_lock(&__kmp_initz_lock);
6390 
6391   KA_TRACE(10, ("__kmp_internal_end_thread: exit T#%d\n", gtid_req));
6392 
6393 #ifdef DUMP_DEBUG_ON_EXIT
6394   if (__kmp_debug_buf)
6395     __kmp_dump_debug_buffer();
6396 #endif
6397 } // __kmp_internal_end_thread
6398 
6399 // -----------------------------------------------------------------------------
6400 // Library registration stuff.
6401 
6402 static long __kmp_registration_flag = 0;
6403 // Random value used to indicate library initialization.
6404 static char *__kmp_registration_str = NULL;
6405 // Value to be saved in env var __KMP_REGISTERED_LIB_<pid>.
6406 
6407 static inline char *__kmp_reg_status_name() {
6408   /* On RHEL 3u5 if linked statically, getpid() returns different values in
6409      each thread. If registration and unregistration go in different threads
6410      (omp_misc_other_root_exit.cpp test case), the name of registered_lib_env
6411      env var can not be found, because the name will contain different pid. */
6412   return __kmp_str_format("__KMP_REGISTERED_LIB_%d", (int)getpid());
6413 } // __kmp_reg_status_get
6414 
6415 void __kmp_register_library_startup(void) {
6416 
6417   char *name = __kmp_reg_status_name(); // Name of the environment variable.
6418   int done = 0;
6419   union {
6420     double dtime;
6421     long ltime;
6422   } time;
6423 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
6424   __kmp_initialize_system_tick();
6425 #endif
6426   __kmp_read_system_time(&time.dtime);
6427   __kmp_registration_flag = 0xCAFE0000L | (time.ltime & 0x0000FFFFL);
6428   __kmp_registration_str =
6429       __kmp_str_format("%p-%lx-%s", &__kmp_registration_flag,
6430                        __kmp_registration_flag, KMP_LIBRARY_FILE);
6431 
6432   KA_TRACE(50, ("__kmp_register_library_startup: %s=\"%s\"\n", name,
6433                 __kmp_registration_str));
6434 
6435   while (!done) {
6436 
6437     char *value = NULL; // Actual value of the environment variable.
6438 
6439     // Set environment variable, but do not overwrite if it is exist.
6440     __kmp_env_set(name, __kmp_registration_str, 0);
6441     // Check the variable is written.
6442     value = __kmp_env_get(name);
6443     if (value != NULL && strcmp(value, __kmp_registration_str) == 0) {
6444 
6445       done = 1; // Ok, environment variable set successfully, exit the loop.
6446 
6447     } else {
6448 
6449       // Oops. Write failed. Another copy of OpenMP RTL is in memory.
6450       // Check whether it alive or dead.
6451       int neighbor = 0; // 0 -- unknown status, 1 -- alive, 2 -- dead.
6452       char *tail = value;
6453       char *flag_addr_str = NULL;
6454       char *flag_val_str = NULL;
6455       char const *file_name = NULL;
6456       __kmp_str_split(tail, '-', &flag_addr_str, &tail);
6457       __kmp_str_split(tail, '-', &flag_val_str, &tail);
6458       file_name = tail;
6459       if (tail != NULL) {
6460         long *flag_addr = 0;
6461         long flag_val = 0;
6462         KMP_SSCANF(flag_addr_str, "%p", RCAST(void**, &flag_addr));
6463         KMP_SSCANF(flag_val_str, "%lx", &flag_val);
6464         if (flag_addr != 0 && flag_val != 0 && strcmp(file_name, "") != 0) {
6465           // First, check whether environment-encoded address is mapped into
6466           // addr space.
6467           // If so, dereference it to see if it still has the right value.
6468           if (__kmp_is_address_mapped(flag_addr) && *flag_addr == flag_val) {
6469             neighbor = 1;
6470           } else {
6471             // If not, then we know the other copy of the library is no longer
6472             // running.
6473             neighbor = 2;
6474           }
6475         }
6476       }
6477       switch (neighbor) {
6478       case 0: // Cannot parse environment variable -- neighbor status unknown.
6479         // Assume it is the incompatible format of future version of the
6480         // library. Assume the other library is alive.
6481         // WARN( ... ); // TODO: Issue a warning.
6482         file_name = "unknown library";
6483         KMP_FALLTHROUGH();
6484       // Attention! Falling to the next case. That's intentional.
6485       case 1: { // Neighbor is alive.
6486         // Check it is allowed.
6487         char *duplicate_ok = __kmp_env_get("KMP_DUPLICATE_LIB_OK");
6488         if (!__kmp_str_match_true(duplicate_ok)) {
6489           // That's not allowed. Issue fatal error.
6490           __kmp_fatal(KMP_MSG(DuplicateLibrary, KMP_LIBRARY_FILE, file_name),
6491                       KMP_HNT(DuplicateLibrary), __kmp_msg_null);
6492         }
6493         KMP_INTERNAL_FREE(duplicate_ok);
6494         __kmp_duplicate_library_ok = 1;
6495         done = 1; // Exit the loop.
6496       } break;
6497       case 2: { // Neighbor is dead.
6498         // Clear the variable and try to register library again.
6499         __kmp_env_unset(name);
6500       } break;
6501       default: { KMP_DEBUG_ASSERT(0); } break;
6502       }
6503     }
6504     KMP_INTERNAL_FREE((void *)value);
6505   }
6506   KMP_INTERNAL_FREE((void *)name);
6507 
6508 } // func __kmp_register_library_startup
6509 
6510 void __kmp_unregister_library(void) {
6511 
6512   char *name = __kmp_reg_status_name();
6513   char *value = __kmp_env_get(name);
6514 
6515   KMP_DEBUG_ASSERT(__kmp_registration_flag != 0);
6516   KMP_DEBUG_ASSERT(__kmp_registration_str != NULL);
6517   if (value != NULL && strcmp(value, __kmp_registration_str) == 0) {
6518     // Ok, this is our variable. Delete it.
6519     __kmp_env_unset(name);
6520   }
6521 
6522   KMP_INTERNAL_FREE(__kmp_registration_str);
6523   KMP_INTERNAL_FREE(value);
6524   KMP_INTERNAL_FREE(name);
6525 
6526   __kmp_registration_flag = 0;
6527   __kmp_registration_str = NULL;
6528 
6529 } // __kmp_unregister_library
6530 
6531 // End of Library registration stuff.
6532 // -----------------------------------------------------------------------------
6533 
6534 #if KMP_MIC_SUPPORTED
6535 
6536 static void __kmp_check_mic_type() {
6537   kmp_cpuid_t cpuid_state = {0};
6538   kmp_cpuid_t *cs_p = &cpuid_state;
6539   __kmp_x86_cpuid(1, 0, cs_p);
6540   // We don't support mic1 at the moment
6541   if ((cs_p->eax & 0xff0) == 0xB10) {
6542     __kmp_mic_type = mic2;
6543   } else if ((cs_p->eax & 0xf0ff0) == 0x50670) {
6544     __kmp_mic_type = mic3;
6545   } else {
6546     __kmp_mic_type = non_mic;
6547   }
6548 }
6549 
6550 #endif /* KMP_MIC_SUPPORTED */
6551 
6552 static void __kmp_do_serial_initialize(void) {
6553   int i, gtid;
6554   int size;
6555 
6556   KA_TRACE(10, ("__kmp_do_serial_initialize: enter\n"));
6557 
6558   KMP_DEBUG_ASSERT(sizeof(kmp_int32) == 4);
6559   KMP_DEBUG_ASSERT(sizeof(kmp_uint32) == 4);
6560   KMP_DEBUG_ASSERT(sizeof(kmp_int64) == 8);
6561   KMP_DEBUG_ASSERT(sizeof(kmp_uint64) == 8);
6562   KMP_DEBUG_ASSERT(sizeof(kmp_intptr_t) == sizeof(void *));
6563 
6564 #if OMPT_SUPPORT
6565   ompt_pre_init();
6566 #endif
6567 
6568   __kmp_validate_locks();
6569 
6570   /* Initialize internal memory allocator */
6571   __kmp_init_allocator();
6572 
6573   /* Register the library startup via an environment variable and check to see
6574      whether another copy of the library is already registered. */
6575 
6576   __kmp_register_library_startup();
6577 
6578   /* TODO reinitialization of library */
6579   if (TCR_4(__kmp_global.g.g_done)) {
6580     KA_TRACE(10, ("__kmp_do_serial_initialize: reinitialization of library\n"));
6581   }
6582 
6583   __kmp_global.g.g_abort = 0;
6584   TCW_SYNC_4(__kmp_global.g.g_done, FALSE);
6585 
6586 /* initialize the locks */
6587 #if KMP_USE_ADAPTIVE_LOCKS
6588 #if KMP_DEBUG_ADAPTIVE_LOCKS
6589   __kmp_init_speculative_stats();
6590 #endif
6591 #endif
6592 #if KMP_STATS_ENABLED
6593   __kmp_stats_init();
6594 #endif
6595   __kmp_init_lock(&__kmp_global_lock);
6596   __kmp_init_queuing_lock(&__kmp_dispatch_lock);
6597   __kmp_init_lock(&__kmp_debug_lock);
6598   __kmp_init_atomic_lock(&__kmp_atomic_lock);
6599   __kmp_init_atomic_lock(&__kmp_atomic_lock_1i);
6600   __kmp_init_atomic_lock(&__kmp_atomic_lock_2i);
6601   __kmp_init_atomic_lock(&__kmp_atomic_lock_4i);
6602   __kmp_init_atomic_lock(&__kmp_atomic_lock_4r);
6603   __kmp_init_atomic_lock(&__kmp_atomic_lock_8i);
6604   __kmp_init_atomic_lock(&__kmp_atomic_lock_8r);
6605   __kmp_init_atomic_lock(&__kmp_atomic_lock_8c);
6606   __kmp_init_atomic_lock(&__kmp_atomic_lock_10r);
6607   __kmp_init_atomic_lock(&__kmp_atomic_lock_16r);
6608   __kmp_init_atomic_lock(&__kmp_atomic_lock_16c);
6609   __kmp_init_atomic_lock(&__kmp_atomic_lock_20c);
6610   __kmp_init_atomic_lock(&__kmp_atomic_lock_32c);
6611   __kmp_init_bootstrap_lock(&__kmp_forkjoin_lock);
6612   __kmp_init_bootstrap_lock(&__kmp_exit_lock);
6613 #if KMP_USE_MONITOR
6614   __kmp_init_bootstrap_lock(&__kmp_monitor_lock);
6615 #endif
6616   __kmp_init_bootstrap_lock(&__kmp_tp_cached_lock);
6617 
6618   /* conduct initialization and initial setup of configuration */
6619 
6620   __kmp_runtime_initialize();
6621 
6622 #if KMP_MIC_SUPPORTED
6623   __kmp_check_mic_type();
6624 #endif
6625 
6626 // Some global variable initialization moved here from kmp_env_initialize()
6627 #ifdef KMP_DEBUG
6628   kmp_diag = 0;
6629 #endif
6630   __kmp_abort_delay = 0;
6631 
6632   // From __kmp_init_dflt_team_nth()
6633   /* assume the entire machine will be used */
6634   __kmp_dflt_team_nth_ub = __kmp_xproc;
6635   if (__kmp_dflt_team_nth_ub < KMP_MIN_NTH) {
6636     __kmp_dflt_team_nth_ub = KMP_MIN_NTH;
6637   }
6638   if (__kmp_dflt_team_nth_ub > __kmp_sys_max_nth) {
6639     __kmp_dflt_team_nth_ub = __kmp_sys_max_nth;
6640   }
6641   __kmp_max_nth = __kmp_sys_max_nth;
6642   __kmp_cg_max_nth = __kmp_sys_max_nth;
6643   __kmp_teams_max_nth = __kmp_xproc; // set a "reasonable" default
6644   if (__kmp_teams_max_nth > __kmp_sys_max_nth) {
6645     __kmp_teams_max_nth = __kmp_sys_max_nth;
6646   }
6647 
6648   // Three vars below moved here from __kmp_env_initialize() "KMP_BLOCKTIME"
6649   // part
6650   __kmp_dflt_blocktime = KMP_DEFAULT_BLOCKTIME;
6651 #if KMP_USE_MONITOR
6652   __kmp_monitor_wakeups =
6653       KMP_WAKEUPS_FROM_BLOCKTIME(__kmp_dflt_blocktime, __kmp_monitor_wakeups);
6654   __kmp_bt_intervals =
6655       KMP_INTERVALS_FROM_BLOCKTIME(__kmp_dflt_blocktime, __kmp_monitor_wakeups);
6656 #endif
6657   // From "KMP_LIBRARY" part of __kmp_env_initialize()
6658   __kmp_library = library_throughput;
6659   // From KMP_SCHEDULE initialization
6660   __kmp_static = kmp_sch_static_balanced;
6661 // AC: do not use analytical here, because it is non-monotonous
6662 //__kmp_guided = kmp_sch_guided_iterative_chunked;
6663 //__kmp_auto = kmp_sch_guided_analytical_chunked; // AC: it is the default, no
6664 // need to repeat assignment
6665 // Barrier initialization. Moved here from __kmp_env_initialize() Barrier branch
6666 // bit control and barrier method control parts
6667 #if KMP_FAST_REDUCTION_BARRIER
6668 #define kmp_reduction_barrier_gather_bb ((int)1)
6669 #define kmp_reduction_barrier_release_bb ((int)1)
6670 #define kmp_reduction_barrier_gather_pat bp_hyper_bar
6671 #define kmp_reduction_barrier_release_pat bp_hyper_bar
6672 #endif // KMP_FAST_REDUCTION_BARRIER
6673   for (i = bs_plain_barrier; i < bs_last_barrier; i++) {
6674     __kmp_barrier_gather_branch_bits[i] = __kmp_barrier_gather_bb_dflt;
6675     __kmp_barrier_release_branch_bits[i] = __kmp_barrier_release_bb_dflt;
6676     __kmp_barrier_gather_pattern[i] = __kmp_barrier_gather_pat_dflt;
6677     __kmp_barrier_release_pattern[i] = __kmp_barrier_release_pat_dflt;
6678 #if KMP_FAST_REDUCTION_BARRIER
6679     if (i == bs_reduction_barrier) { // tested and confirmed on ALTIX only (
6680       // lin_64 ): hyper,1
6681       __kmp_barrier_gather_branch_bits[i] = kmp_reduction_barrier_gather_bb;
6682       __kmp_barrier_release_branch_bits[i] = kmp_reduction_barrier_release_bb;
6683       __kmp_barrier_gather_pattern[i] = kmp_reduction_barrier_gather_pat;
6684       __kmp_barrier_release_pattern[i] = kmp_reduction_barrier_release_pat;
6685     }
6686 #endif // KMP_FAST_REDUCTION_BARRIER
6687   }
6688 #if KMP_FAST_REDUCTION_BARRIER
6689 #undef kmp_reduction_barrier_release_pat
6690 #undef kmp_reduction_barrier_gather_pat
6691 #undef kmp_reduction_barrier_release_bb
6692 #undef kmp_reduction_barrier_gather_bb
6693 #endif // KMP_FAST_REDUCTION_BARRIER
6694 #if KMP_MIC_SUPPORTED
6695   if (__kmp_mic_type == mic2) { // KNC
6696     // AC: plane=3,2, forkjoin=2,1 are optimal for 240 threads on KNC
6697     __kmp_barrier_gather_branch_bits[bs_plain_barrier] = 3; // plain gather
6698     __kmp_barrier_release_branch_bits[bs_forkjoin_barrier] =
6699         1; // forkjoin release
6700     __kmp_barrier_gather_pattern[bs_forkjoin_barrier] = bp_hierarchical_bar;
6701     __kmp_barrier_release_pattern[bs_forkjoin_barrier] = bp_hierarchical_bar;
6702   }
6703 #if KMP_FAST_REDUCTION_BARRIER
6704   if (__kmp_mic_type == mic2) { // KNC
6705     __kmp_barrier_gather_pattern[bs_reduction_barrier] = bp_hierarchical_bar;
6706     __kmp_barrier_release_pattern[bs_reduction_barrier] = bp_hierarchical_bar;
6707   }
6708 #endif // KMP_FAST_REDUCTION_BARRIER
6709 #endif // KMP_MIC_SUPPORTED
6710 
6711 // From KMP_CHECKS initialization
6712 #ifdef KMP_DEBUG
6713   __kmp_env_checks = TRUE; /* development versions have the extra checks */
6714 #else
6715   __kmp_env_checks = FALSE; /* port versions do not have the extra checks */
6716 #endif
6717 
6718   // From "KMP_FOREIGN_THREADS_THREADPRIVATE" initialization
6719   __kmp_foreign_tp = TRUE;
6720 
6721   __kmp_global.g.g_dynamic = FALSE;
6722   __kmp_global.g.g_dynamic_mode = dynamic_default;
6723 
6724   __kmp_env_initialize(NULL);
6725 
6726 // Print all messages in message catalog for testing purposes.
6727 #ifdef KMP_DEBUG
6728   char const *val = __kmp_env_get("KMP_DUMP_CATALOG");
6729   if (__kmp_str_match_true(val)) {
6730     kmp_str_buf_t buffer;
6731     __kmp_str_buf_init(&buffer);
6732     __kmp_i18n_dump_catalog(&buffer);
6733     __kmp_printf("%s", buffer.str);
6734     __kmp_str_buf_free(&buffer);
6735   }
6736   __kmp_env_free(&val);
6737 #endif
6738 
6739   __kmp_threads_capacity =
6740       __kmp_initial_threads_capacity(__kmp_dflt_team_nth_ub);
6741   // Moved here from __kmp_env_initialize() "KMP_ALL_THREADPRIVATE" part
6742   __kmp_tp_capacity = __kmp_default_tp_capacity(
6743       __kmp_dflt_team_nth_ub, __kmp_max_nth, __kmp_allThreadsSpecified);
6744 
6745   // If the library is shut down properly, both pools must be NULL. Just in
6746   // case, set them to NULL -- some memory may leak, but subsequent code will
6747   // work even if pools are not freed.
6748   KMP_DEBUG_ASSERT(__kmp_thread_pool == NULL);
6749   KMP_DEBUG_ASSERT(__kmp_thread_pool_insert_pt == NULL);
6750   KMP_DEBUG_ASSERT(__kmp_team_pool == NULL);
6751   __kmp_thread_pool = NULL;
6752   __kmp_thread_pool_insert_pt = NULL;
6753   __kmp_team_pool = NULL;
6754 
6755   /* Allocate all of the variable sized records */
6756   /* NOTE: __kmp_threads_capacity entries are allocated, but the arrays are
6757    * expandable */
6758   /* Since allocation is cache-aligned, just add extra padding at the end */
6759   size =
6760       (sizeof(kmp_info_t *) + sizeof(kmp_root_t *)) * __kmp_threads_capacity +
6761       CACHE_LINE;
6762   __kmp_threads = (kmp_info_t **)__kmp_allocate(size);
6763   __kmp_root = (kmp_root_t **)((char *)__kmp_threads +
6764                                sizeof(kmp_info_t *) * __kmp_threads_capacity);
6765 
6766   /* init thread counts */
6767   KMP_DEBUG_ASSERT(__kmp_all_nth ==
6768                    0); // Asserts fail if the library is reinitializing and
6769   KMP_DEBUG_ASSERT(__kmp_nth == 0); // something was wrong in termination.
6770   __kmp_all_nth = 0;
6771   __kmp_nth = 0;
6772 
6773   /* setup the uber master thread and hierarchy */
6774   gtid = __kmp_register_root(TRUE);
6775   KA_TRACE(10, ("__kmp_do_serial_initialize  T#%d\n", gtid));
6776   KMP_ASSERT(KMP_UBER_GTID(gtid));
6777   KMP_ASSERT(KMP_INITIAL_GTID(gtid));
6778 
6779   KMP_MB(); /* Flush all pending memory write invalidates.  */
6780 
6781   __kmp_common_initialize();
6782 
6783 #if KMP_OS_UNIX
6784   /* invoke the child fork handler */
6785   __kmp_register_atfork();
6786 #endif
6787 
6788 #if !KMP_DYNAMIC_LIB
6789   {
6790     /* Invoke the exit handler when the program finishes, only for static
6791        library. For dynamic library, we already have _fini and DllMain. */
6792     int rc = atexit(__kmp_internal_end_atexit);
6793     if (rc != 0) {
6794       __kmp_fatal(KMP_MSG(FunctionError, "atexit()"), KMP_ERR(rc),
6795                   __kmp_msg_null);
6796     }
6797   }
6798 #endif
6799 
6800 #if KMP_HANDLE_SIGNALS
6801 #if KMP_OS_UNIX
6802   /* NOTE: make sure that this is called before the user installs their own
6803      signal handlers so that the user handlers are called first. this way they
6804      can return false, not call our handler, avoid terminating the library, and
6805      continue execution where they left off. */
6806   __kmp_install_signals(FALSE);
6807 #endif /* KMP_OS_UNIX */
6808 #if KMP_OS_WINDOWS
6809   __kmp_install_signals(TRUE);
6810 #endif /* KMP_OS_WINDOWS */
6811 #endif
6812 
6813   /* we have finished the serial initialization */
6814   __kmp_init_counter++;
6815 
6816   __kmp_init_serial = TRUE;
6817 
6818   if (__kmp_settings) {
6819     __kmp_env_print();
6820   }
6821 
6822 #if OMP_40_ENABLED
6823   if (__kmp_display_env || __kmp_display_env_verbose) {
6824     __kmp_env_print_2();
6825   }
6826 #endif // OMP_40_ENABLED
6827 
6828 #if OMPT_SUPPORT
6829   ompt_post_init();
6830 #endif
6831 
6832   KMP_MB();
6833 
6834   KA_TRACE(10, ("__kmp_do_serial_initialize: exit\n"));
6835 }
6836 
6837 void __kmp_serial_initialize(void) {
6838   if (__kmp_init_serial) {
6839     return;
6840   }
6841   __kmp_acquire_bootstrap_lock(&__kmp_initz_lock);
6842   if (__kmp_init_serial) {
6843     __kmp_release_bootstrap_lock(&__kmp_initz_lock);
6844     return;
6845   }
6846   __kmp_do_serial_initialize();
6847   __kmp_release_bootstrap_lock(&__kmp_initz_lock);
6848 }
6849 
6850 static void __kmp_do_middle_initialize(void) {
6851   int i, j;
6852   int prev_dflt_team_nth;
6853 
6854   if (!__kmp_init_serial) {
6855     __kmp_do_serial_initialize();
6856   }
6857 
6858   KA_TRACE(10, ("__kmp_middle_initialize: enter\n"));
6859 
6860   // Save the previous value for the __kmp_dflt_team_nth so that
6861   // we can avoid some reinitialization if it hasn't changed.
6862   prev_dflt_team_nth = __kmp_dflt_team_nth;
6863 
6864 #if KMP_AFFINITY_SUPPORTED
6865   // __kmp_affinity_initialize() will try to set __kmp_ncores to the
6866   // number of cores on the machine.
6867   __kmp_affinity_initialize();
6868 
6869   // Run through the __kmp_threads array and set the affinity mask
6870   // for each root thread that is currently registered with the RTL.
6871   for (i = 0; i < __kmp_threads_capacity; i++) {
6872     if (TCR_PTR(__kmp_threads[i]) != NULL) {
6873       __kmp_affinity_set_init_mask(i, TRUE);
6874     }
6875   }
6876 #endif /* KMP_AFFINITY_SUPPORTED */
6877 
6878   KMP_ASSERT(__kmp_xproc > 0);
6879   if (__kmp_avail_proc == 0) {
6880     __kmp_avail_proc = __kmp_xproc;
6881   }
6882 
6883   // If there were empty places in num_threads list (OMP_NUM_THREADS=,,2,3),
6884   // correct them now
6885   j = 0;
6886   while ((j < __kmp_nested_nth.used) && !__kmp_nested_nth.nth[j]) {
6887     __kmp_nested_nth.nth[j] = __kmp_dflt_team_nth = __kmp_dflt_team_nth_ub =
6888         __kmp_avail_proc;
6889     j++;
6890   }
6891 
6892   if (__kmp_dflt_team_nth == 0) {
6893 #ifdef KMP_DFLT_NTH_CORES
6894     // Default #threads = #cores
6895     __kmp_dflt_team_nth = __kmp_ncores;
6896     KA_TRACE(20, ("__kmp_middle_initialize: setting __kmp_dflt_team_nth = "
6897                   "__kmp_ncores (%d)\n",
6898                   __kmp_dflt_team_nth));
6899 #else
6900     // Default #threads = #available OS procs
6901     __kmp_dflt_team_nth = __kmp_avail_proc;
6902     KA_TRACE(20, ("__kmp_middle_initialize: setting __kmp_dflt_team_nth = "
6903                   "__kmp_avail_proc(%d)\n",
6904                   __kmp_dflt_team_nth));
6905 #endif /* KMP_DFLT_NTH_CORES */
6906   }
6907 
6908   if (__kmp_dflt_team_nth < KMP_MIN_NTH) {
6909     __kmp_dflt_team_nth = KMP_MIN_NTH;
6910   }
6911   if (__kmp_dflt_team_nth > __kmp_sys_max_nth) {
6912     __kmp_dflt_team_nth = __kmp_sys_max_nth;
6913   }
6914 
6915   // There's no harm in continuing if the following check fails,
6916   // but it indicates an error in the previous logic.
6917   KMP_DEBUG_ASSERT(__kmp_dflt_team_nth <= __kmp_dflt_team_nth_ub);
6918 
6919   if (__kmp_dflt_team_nth != prev_dflt_team_nth) {
6920     // Run through the __kmp_threads array and set the num threads icv for each
6921     // root thread that is currently registered with the RTL (which has not
6922     // already explicitly set its nthreads-var with a call to
6923     // omp_set_num_threads()).
6924     for (i = 0; i < __kmp_threads_capacity; i++) {
6925       kmp_info_t *thread = __kmp_threads[i];
6926       if (thread == NULL)
6927         continue;
6928       if (thread->th.th_current_task->td_icvs.nproc != 0)
6929         continue;
6930 
6931       set__nproc(__kmp_threads[i], __kmp_dflt_team_nth);
6932     }
6933   }
6934   KA_TRACE(
6935       20,
6936       ("__kmp_middle_initialize: final value for __kmp_dflt_team_nth = %d\n",
6937        __kmp_dflt_team_nth));
6938 
6939 #ifdef KMP_ADJUST_BLOCKTIME
6940   /* Adjust blocktime to zero if necessary  now that __kmp_avail_proc is set */
6941   if (!__kmp_env_blocktime && (__kmp_avail_proc > 0)) {
6942     KMP_DEBUG_ASSERT(__kmp_avail_proc > 0);
6943     if (__kmp_nth > __kmp_avail_proc) {
6944       __kmp_zero_bt = TRUE;
6945     }
6946   }
6947 #endif /* KMP_ADJUST_BLOCKTIME */
6948 
6949   /* we have finished middle initialization */
6950   TCW_SYNC_4(__kmp_init_middle, TRUE);
6951 
6952   KA_TRACE(10, ("__kmp_do_middle_initialize: exit\n"));
6953 }
6954 
6955 void __kmp_middle_initialize(void) {
6956   if (__kmp_init_middle) {
6957     return;
6958   }
6959   __kmp_acquire_bootstrap_lock(&__kmp_initz_lock);
6960   if (__kmp_init_middle) {
6961     __kmp_release_bootstrap_lock(&__kmp_initz_lock);
6962     return;
6963   }
6964   __kmp_do_middle_initialize();
6965   __kmp_release_bootstrap_lock(&__kmp_initz_lock);
6966 }
6967 
6968 void __kmp_parallel_initialize(void) {
6969   int gtid = __kmp_entry_gtid(); // this might be a new root
6970 
6971   /* synchronize parallel initialization (for sibling) */
6972   if (TCR_4(__kmp_init_parallel))
6973     return;
6974   __kmp_acquire_bootstrap_lock(&__kmp_initz_lock);
6975   if (TCR_4(__kmp_init_parallel)) {
6976     __kmp_release_bootstrap_lock(&__kmp_initz_lock);
6977     return;
6978   }
6979 
6980   /* TODO reinitialization after we have already shut down */
6981   if (TCR_4(__kmp_global.g.g_done)) {
6982     KA_TRACE(
6983         10,
6984         ("__kmp_parallel_initialize: attempt to init while shutting down\n"));
6985     __kmp_infinite_loop();
6986   }
6987 
6988   /* jc: The lock __kmp_initz_lock is already held, so calling
6989      __kmp_serial_initialize would cause a deadlock.  So we call
6990      __kmp_do_serial_initialize directly. */
6991   if (!__kmp_init_middle) {
6992     __kmp_do_middle_initialize();
6993   }
6994 
6995 #if OMP_50_ENABLED
6996   __kmp_resume_if_hard_paused();
6997 #endif
6998 
6999   /* begin initialization */
7000   KA_TRACE(10, ("__kmp_parallel_initialize: enter\n"));
7001   KMP_ASSERT(KMP_UBER_GTID(gtid));
7002 
7003 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
7004   // Save the FP control regs.
7005   // Worker threads will set theirs to these values at thread startup.
7006   __kmp_store_x87_fpu_control_word(&__kmp_init_x87_fpu_control_word);
7007   __kmp_store_mxcsr(&__kmp_init_mxcsr);
7008   __kmp_init_mxcsr &= KMP_X86_MXCSR_MASK;
7009 #endif /* KMP_ARCH_X86 || KMP_ARCH_X86_64 */
7010 
7011 #if KMP_OS_UNIX
7012 #if KMP_HANDLE_SIGNALS
7013   /*  must be after __kmp_serial_initialize  */
7014   __kmp_install_signals(TRUE);
7015 #endif
7016 #endif
7017 
7018   __kmp_suspend_initialize();
7019 
7020 #if defined(USE_LOAD_BALANCE)
7021   if (__kmp_global.g.g_dynamic_mode == dynamic_default) {
7022     __kmp_global.g.g_dynamic_mode = dynamic_load_balance;
7023   }
7024 #else
7025   if (__kmp_global.g.g_dynamic_mode == dynamic_default) {
7026     __kmp_global.g.g_dynamic_mode = dynamic_thread_limit;
7027   }
7028 #endif
7029 
7030   if (__kmp_version) {
7031     __kmp_print_version_2();
7032   }
7033 
7034   /* we have finished parallel initialization */
7035   TCW_SYNC_4(__kmp_init_parallel, TRUE);
7036 
7037   KMP_MB();
7038   KA_TRACE(10, ("__kmp_parallel_initialize: exit\n"));
7039 
7040   __kmp_release_bootstrap_lock(&__kmp_initz_lock);
7041 }
7042 
7043 /* ------------------------------------------------------------------------ */
7044 
7045 void __kmp_run_before_invoked_task(int gtid, int tid, kmp_info_t *this_thr,
7046                                    kmp_team_t *team) {
7047   kmp_disp_t *dispatch;
7048 
7049   KMP_MB();
7050 
7051   /* none of the threads have encountered any constructs, yet. */
7052   this_thr->th.th_local.this_construct = 0;
7053 #if KMP_CACHE_MANAGE
7054   KMP_CACHE_PREFETCH(&this_thr->th.th_bar[bs_forkjoin_barrier].bb.b_arrived);
7055 #endif /* KMP_CACHE_MANAGE */
7056   dispatch = (kmp_disp_t *)TCR_PTR(this_thr->th.th_dispatch);
7057   KMP_DEBUG_ASSERT(dispatch);
7058   KMP_DEBUG_ASSERT(team->t.t_dispatch);
7059   // KMP_DEBUG_ASSERT( this_thr->th.th_dispatch == &team->t.t_dispatch[
7060   // this_thr->th.th_info.ds.ds_tid ] );
7061 
7062   dispatch->th_disp_index = 0; /* reset the dispatch buffer counter */
7063 #if OMP_45_ENABLED
7064   dispatch->th_doacross_buf_idx =
7065       0; /* reset the doacross dispatch buffer counter */
7066 #endif
7067   if (__kmp_env_consistency_check)
7068     __kmp_push_parallel(gtid, team->t.t_ident);
7069 
7070   KMP_MB(); /* Flush all pending memory write invalidates.  */
7071 }
7072 
7073 void __kmp_run_after_invoked_task(int gtid, int tid, kmp_info_t *this_thr,
7074                                   kmp_team_t *team) {
7075   if (__kmp_env_consistency_check)
7076     __kmp_pop_parallel(gtid, team->t.t_ident);
7077 
7078   __kmp_finish_implicit_task(this_thr);
7079 }
7080 
7081 int __kmp_invoke_task_func(int gtid) {
7082   int rc;
7083   int tid = __kmp_tid_from_gtid(gtid);
7084   kmp_info_t *this_thr = __kmp_threads[gtid];
7085   kmp_team_t *team = this_thr->th.th_team;
7086 
7087   __kmp_run_before_invoked_task(gtid, tid, this_thr, team);
7088 #if USE_ITT_BUILD
7089   if (__itt_stack_caller_create_ptr) {
7090     __kmp_itt_stack_callee_enter(
7091         (__itt_caller)
7092             team->t.t_stack_id); // inform ittnotify about entering user's code
7093   }
7094 #endif /* USE_ITT_BUILD */
7095 #if INCLUDE_SSC_MARKS
7096   SSC_MARK_INVOKING();
7097 #endif
7098 
7099 #if OMPT_SUPPORT
7100   void *dummy;
7101   void **exit_runtime_p;
7102   ompt_data_t *my_task_data;
7103   ompt_data_t *my_parallel_data;
7104   int ompt_team_size;
7105 
7106   if (ompt_enabled.enabled) {
7107     exit_runtime_p = &(
7108         team->t.t_implicit_task_taskdata[tid].ompt_task_info.frame.exit_frame.ptr);
7109   } else {
7110     exit_runtime_p = &dummy;
7111   }
7112 
7113   my_task_data =
7114       &(team->t.t_implicit_task_taskdata[tid].ompt_task_info.task_data);
7115   my_parallel_data = &(team->t.ompt_team_info.parallel_data);
7116   if (ompt_enabled.ompt_callback_implicit_task) {
7117     ompt_team_size = team->t.t_nproc;
7118     ompt_callbacks.ompt_callback(ompt_callback_implicit_task)(
7119         ompt_scope_begin, my_parallel_data, my_task_data, ompt_team_size,
7120         __kmp_tid_from_gtid(gtid), ompt_task_implicit); // TODO: Can this be ompt_task_initial?
7121     OMPT_CUR_TASK_INFO(this_thr)->thread_num = __kmp_tid_from_gtid(gtid);
7122   }
7123 #endif
7124 
7125 #if KMP_STATS_ENABLED
7126   stats_state_e previous_state = KMP_GET_THREAD_STATE();
7127   if (previous_state == stats_state_e::TEAMS_REGION) {
7128     KMP_PUSH_PARTITIONED_TIMER(OMP_teams);
7129   } else {
7130     KMP_PUSH_PARTITIONED_TIMER(OMP_parallel);
7131   }
7132   KMP_SET_THREAD_STATE(IMPLICIT_TASK);
7133 #endif
7134 
7135   rc = __kmp_invoke_microtask((microtask_t)TCR_SYNC_PTR(team->t.t_pkfn), gtid,
7136                               tid, (int)team->t.t_argc, (void **)team->t.t_argv
7137 #if OMPT_SUPPORT
7138                               ,
7139                               exit_runtime_p
7140 #endif
7141                               );
7142 #if OMPT_SUPPORT
7143   *exit_runtime_p = NULL;
7144 #endif
7145 
7146 #if KMP_STATS_ENABLED
7147   if (previous_state == stats_state_e::TEAMS_REGION) {
7148     KMP_SET_THREAD_STATE(previous_state);
7149   }
7150   KMP_POP_PARTITIONED_TIMER();
7151 #endif
7152 
7153 #if USE_ITT_BUILD
7154   if (__itt_stack_caller_create_ptr) {
7155     __kmp_itt_stack_callee_leave(
7156         (__itt_caller)
7157             team->t.t_stack_id); // inform ittnotify about leaving user's code
7158   }
7159 #endif /* USE_ITT_BUILD */
7160   __kmp_run_after_invoked_task(gtid, tid, this_thr, team);
7161 
7162   return rc;
7163 }
7164 
7165 #if OMP_40_ENABLED
7166 void __kmp_teams_master(int gtid) {
7167   // This routine is called by all master threads in teams construct
7168   kmp_info_t *thr = __kmp_threads[gtid];
7169   kmp_team_t *team = thr->th.th_team;
7170   ident_t *loc = team->t.t_ident;
7171   thr->th.th_set_nproc = thr->th.th_teams_size.nth;
7172   KMP_DEBUG_ASSERT(thr->th.th_teams_microtask);
7173   KMP_DEBUG_ASSERT(thr->th.th_set_nproc);
7174   KA_TRACE(20, ("__kmp_teams_master: T#%d, Tid %d, microtask %p\n", gtid,
7175                 __kmp_tid_from_gtid(gtid), thr->th.th_teams_microtask));
7176 
7177   // This thread is a new CG root.  Set up the proper variables.
7178   kmp_cg_root_t *tmp = (kmp_cg_root_t *)__kmp_allocate(sizeof(kmp_cg_root_t));
7179   tmp->cg_root = thr; // Make thr the CG root
7180   // Init to thread limit that was stored when league masters were forked
7181   tmp->cg_thread_limit = thr->th.th_current_task->td_icvs.thread_limit;
7182   tmp->cg_nthreads = 1; // Init counter to one active thread, this one
7183   KA_TRACE(100, ("__kmp_teams_master: Thread %p created node %p and init"
7184                  " cg_threads to 1\n",
7185                  thr, tmp));
7186   tmp->up = thr->th.th_cg_roots;
7187   thr->th.th_cg_roots = tmp;
7188 
7189 // Launch league of teams now, but not let workers execute
7190 // (they hang on fork barrier until next parallel)
7191 #if INCLUDE_SSC_MARKS
7192   SSC_MARK_FORKING();
7193 #endif
7194   __kmp_fork_call(loc, gtid, fork_context_intel, team->t.t_argc,
7195                   (microtask_t)thr->th.th_teams_microtask, // "wrapped" task
7196                   VOLATILE_CAST(launch_t) __kmp_invoke_task_func, NULL);
7197 #if INCLUDE_SSC_MARKS
7198   SSC_MARK_JOINING();
7199 #endif
7200   // If the team size was reduced from the limit, set it to the new size
7201   if (thr->th.th_team_nproc < thr->th.th_teams_size.nth)
7202     thr->th.th_teams_size.nth = thr->th.th_team_nproc;
7203   // AC: last parameter "1" eliminates join barrier which won't work because
7204   // worker threads are in a fork barrier waiting for more parallel regions
7205   __kmp_join_call(loc, gtid
7206 #if OMPT_SUPPORT
7207                   ,
7208                   fork_context_intel
7209 #endif
7210                   ,
7211                   1);
7212 }
7213 
7214 int __kmp_invoke_teams_master(int gtid) {
7215   kmp_info_t *this_thr = __kmp_threads[gtid];
7216   kmp_team_t *team = this_thr->th.th_team;
7217 #if KMP_DEBUG
7218   if (!__kmp_threads[gtid]->th.th_team->t.t_serialized)
7219     KMP_DEBUG_ASSERT((void *)__kmp_threads[gtid]->th.th_team->t.t_pkfn ==
7220                      (void *)__kmp_teams_master);
7221 #endif
7222   __kmp_run_before_invoked_task(gtid, 0, this_thr, team);
7223   __kmp_teams_master(gtid);
7224   __kmp_run_after_invoked_task(gtid, 0, this_thr, team);
7225   return 1;
7226 }
7227 #endif /* OMP_40_ENABLED */
7228 
7229 /* this sets the requested number of threads for the next parallel region
7230    encountered by this team. since this should be enclosed in the forkjoin
7231    critical section it should avoid race conditions with assymmetrical nested
7232    parallelism */
7233 
7234 void __kmp_push_num_threads(ident_t *id, int gtid, int num_threads) {
7235   kmp_info_t *thr = __kmp_threads[gtid];
7236 
7237   if (num_threads > 0)
7238     thr->th.th_set_nproc = num_threads;
7239 }
7240 
7241 #if OMP_40_ENABLED
7242 
7243 /* this sets the requested number of teams for the teams region and/or
7244    the number of threads for the next parallel region encountered  */
7245 void __kmp_push_num_teams(ident_t *id, int gtid, int num_teams,
7246                           int num_threads) {
7247   kmp_info_t *thr = __kmp_threads[gtid];
7248   KMP_DEBUG_ASSERT(num_teams >= 0);
7249   KMP_DEBUG_ASSERT(num_threads >= 0);
7250 
7251   if (num_teams == 0)
7252     num_teams = 1; // default number of teams is 1.
7253   if (num_teams > __kmp_teams_max_nth) { // if too many teams requested?
7254     if (!__kmp_reserve_warn) {
7255       __kmp_reserve_warn = 1;
7256       __kmp_msg(kmp_ms_warning,
7257                 KMP_MSG(CantFormThrTeam, num_teams, __kmp_teams_max_nth),
7258                 KMP_HNT(Unset_ALL_THREADS), __kmp_msg_null);
7259     }
7260     num_teams = __kmp_teams_max_nth;
7261   }
7262   // Set number of teams (number of threads in the outer "parallel" of the
7263   // teams)
7264   thr->th.th_set_nproc = thr->th.th_teams_size.nteams = num_teams;
7265 
7266   // Remember the number of threads for inner parallel regions
7267   if (num_threads == 0) {
7268     if (!TCR_4(__kmp_init_middle))
7269       __kmp_middle_initialize(); // get __kmp_avail_proc calculated
7270     num_threads = __kmp_avail_proc / num_teams;
7271     if (num_teams * num_threads > __kmp_teams_max_nth) {
7272       // adjust num_threads w/o warning as it is not user setting
7273       num_threads = __kmp_teams_max_nth / num_teams;
7274     }
7275   } else {
7276     // This thread will be the master of the league masters
7277     // Store new thread limit; old limit is saved in th_cg_roots list
7278     thr->th.th_current_task->td_icvs.thread_limit = num_threads;
7279 
7280     if (num_teams * num_threads > __kmp_teams_max_nth) {
7281       int new_threads = __kmp_teams_max_nth / num_teams;
7282       if (!__kmp_reserve_warn) { // user asked for too many threads
7283         __kmp_reserve_warn = 1; // conflicts with KMP_TEAMS_THREAD_LIMIT
7284         __kmp_msg(kmp_ms_warning,
7285                   KMP_MSG(CantFormThrTeam, num_threads, new_threads),
7286                   KMP_HNT(Unset_ALL_THREADS), __kmp_msg_null);
7287       }
7288       num_threads = new_threads;
7289     }
7290   }
7291   thr->th.th_teams_size.nth = num_threads;
7292 }
7293 
7294 // Set the proc_bind var to use in the following parallel region.
7295 void __kmp_push_proc_bind(ident_t *id, int gtid, kmp_proc_bind_t proc_bind) {
7296   kmp_info_t *thr = __kmp_threads[gtid];
7297   thr->th.th_set_proc_bind = proc_bind;
7298 }
7299 
7300 #endif /* OMP_40_ENABLED */
7301 
7302 /* Launch the worker threads into the microtask. */
7303 
7304 void __kmp_internal_fork(ident_t *id, int gtid, kmp_team_t *team) {
7305   kmp_info_t *this_thr = __kmp_threads[gtid];
7306 
7307 #ifdef KMP_DEBUG
7308   int f;
7309 #endif /* KMP_DEBUG */
7310 
7311   KMP_DEBUG_ASSERT(team);
7312   KMP_DEBUG_ASSERT(this_thr->th.th_team == team);
7313   KMP_ASSERT(KMP_MASTER_GTID(gtid));
7314   KMP_MB(); /* Flush all pending memory write invalidates.  */
7315 
7316   team->t.t_construct = 0; /* no single directives seen yet */
7317   team->t.t_ordered.dt.t_value =
7318       0; /* thread 0 enters the ordered section first */
7319 
7320   /* Reset the identifiers on the dispatch buffer */
7321   KMP_DEBUG_ASSERT(team->t.t_disp_buffer);
7322   if (team->t.t_max_nproc > 1) {
7323     int i;
7324     for (i = 0; i < __kmp_dispatch_num_buffers; ++i) {
7325       team->t.t_disp_buffer[i].buffer_index = i;
7326 #if OMP_45_ENABLED
7327       team->t.t_disp_buffer[i].doacross_buf_idx = i;
7328 #endif
7329     }
7330   } else {
7331     team->t.t_disp_buffer[0].buffer_index = 0;
7332 #if OMP_45_ENABLED
7333     team->t.t_disp_buffer[0].doacross_buf_idx = 0;
7334 #endif
7335   }
7336 
7337   KMP_MB(); /* Flush all pending memory write invalidates.  */
7338   KMP_ASSERT(this_thr->th.th_team == team);
7339 
7340 #ifdef KMP_DEBUG
7341   for (f = 0; f < team->t.t_nproc; f++) {
7342     KMP_DEBUG_ASSERT(team->t.t_threads[f] &&
7343                      team->t.t_threads[f]->th.th_team_nproc == team->t.t_nproc);
7344   }
7345 #endif /* KMP_DEBUG */
7346 
7347   /* release the worker threads so they may begin working */
7348   __kmp_fork_barrier(gtid, 0);
7349 }
7350 
7351 void __kmp_internal_join(ident_t *id, int gtid, kmp_team_t *team) {
7352   kmp_info_t *this_thr = __kmp_threads[gtid];
7353 
7354   KMP_DEBUG_ASSERT(team);
7355   KMP_DEBUG_ASSERT(this_thr->th.th_team == team);
7356   KMP_ASSERT(KMP_MASTER_GTID(gtid));
7357   KMP_MB(); /* Flush all pending memory write invalidates.  */
7358 
7359 /* Join barrier after fork */
7360 
7361 #ifdef KMP_DEBUG
7362   if (__kmp_threads[gtid] &&
7363       __kmp_threads[gtid]->th.th_team_nproc != team->t.t_nproc) {
7364     __kmp_printf("GTID: %d, __kmp_threads[%d]=%p\n", gtid, gtid,
7365                  __kmp_threads[gtid]);
7366     __kmp_printf("__kmp_threads[%d]->th.th_team_nproc=%d, TEAM: %p, "
7367                  "team->t.t_nproc=%d\n",
7368                  gtid, __kmp_threads[gtid]->th.th_team_nproc, team,
7369                  team->t.t_nproc);
7370     __kmp_print_structure();
7371   }
7372   KMP_DEBUG_ASSERT(__kmp_threads[gtid] &&
7373                    __kmp_threads[gtid]->th.th_team_nproc == team->t.t_nproc);
7374 #endif /* KMP_DEBUG */
7375 
7376   __kmp_join_barrier(gtid); /* wait for everyone */
7377 #if OMPT_SUPPORT
7378   if (ompt_enabled.enabled &&
7379       this_thr->th.ompt_thread_info.state == ompt_state_wait_barrier_implicit) {
7380     int ds_tid = this_thr->th.th_info.ds.ds_tid;
7381     ompt_data_t *task_data = OMPT_CUR_TASK_DATA(this_thr);
7382     this_thr->th.ompt_thread_info.state = ompt_state_overhead;
7383 #if OMPT_OPTIONAL
7384     void *codeptr = NULL;
7385     if (KMP_MASTER_TID(ds_tid) &&
7386         (ompt_callbacks.ompt_callback(ompt_callback_sync_region_wait) ||
7387          ompt_callbacks.ompt_callback(ompt_callback_sync_region)))
7388       codeptr = OMPT_CUR_TEAM_INFO(this_thr)->master_return_address;
7389 
7390     if (ompt_enabled.ompt_callback_sync_region_wait) {
7391       ompt_callbacks.ompt_callback(ompt_callback_sync_region_wait)(
7392           ompt_sync_region_barrier_implicit, ompt_scope_end, NULL, task_data,
7393           codeptr);
7394     }
7395     if (ompt_enabled.ompt_callback_sync_region) {
7396       ompt_callbacks.ompt_callback(ompt_callback_sync_region)(
7397           ompt_sync_region_barrier_implicit, ompt_scope_end, NULL, task_data,
7398           codeptr);
7399     }
7400 #endif
7401     if (!KMP_MASTER_TID(ds_tid) && ompt_enabled.ompt_callback_implicit_task) {
7402       ompt_callbacks.ompt_callback(ompt_callback_implicit_task)(
7403           ompt_scope_end, NULL, task_data, 0, ds_tid, ompt_task_implicit); // TODO: Can this be ompt_task_initial?
7404     }
7405   }
7406 #endif
7407 
7408   KMP_MB(); /* Flush all pending memory write invalidates.  */
7409   KMP_ASSERT(this_thr->th.th_team == team);
7410 }
7411 
7412 /* ------------------------------------------------------------------------ */
7413 
7414 #ifdef USE_LOAD_BALANCE
7415 
7416 // Return the worker threads actively spinning in the hot team, if we
7417 // are at the outermost level of parallelism.  Otherwise, return 0.
7418 static int __kmp_active_hot_team_nproc(kmp_root_t *root) {
7419   int i;
7420   int retval;
7421   kmp_team_t *hot_team;
7422 
7423   if (root->r.r_active) {
7424     return 0;
7425   }
7426   hot_team = root->r.r_hot_team;
7427   if (__kmp_dflt_blocktime == KMP_MAX_BLOCKTIME) {
7428     return hot_team->t.t_nproc - 1; // Don't count master thread
7429   }
7430 
7431   // Skip the master thread - it is accounted for elsewhere.
7432   retval = 0;
7433   for (i = 1; i < hot_team->t.t_nproc; i++) {
7434     if (hot_team->t.t_threads[i]->th.th_active) {
7435       retval++;
7436     }
7437   }
7438   return retval;
7439 }
7440 
7441 // Perform an automatic adjustment to the number of
7442 // threads used by the next parallel region.
7443 static int __kmp_load_balance_nproc(kmp_root_t *root, int set_nproc) {
7444   int retval;
7445   int pool_active;
7446   int hot_team_active;
7447   int team_curr_active;
7448   int system_active;
7449 
7450   KB_TRACE(20, ("__kmp_load_balance_nproc: called root:%p set_nproc:%d\n", root,
7451                 set_nproc));
7452   KMP_DEBUG_ASSERT(root);
7453   KMP_DEBUG_ASSERT(root->r.r_root_team->t.t_threads[0]
7454                        ->th.th_current_task->td_icvs.dynamic == TRUE);
7455   KMP_DEBUG_ASSERT(set_nproc > 1);
7456 
7457   if (set_nproc == 1) {
7458     KB_TRACE(20, ("__kmp_load_balance_nproc: serial execution.\n"));
7459     return 1;
7460   }
7461 
7462   // Threads that are active in the thread pool, active in the hot team for this
7463   // particular root (if we are at the outer par level), and the currently
7464   // executing thread (to become the master) are available to add to the new
7465   // team, but are currently contributing to the system load, and must be
7466   // accounted for.
7467   pool_active = __kmp_thread_pool_active_nth;
7468   hot_team_active = __kmp_active_hot_team_nproc(root);
7469   team_curr_active = pool_active + hot_team_active + 1;
7470 
7471   // Check the system load.
7472   system_active = __kmp_get_load_balance(__kmp_avail_proc + team_curr_active);
7473   KB_TRACE(30, ("__kmp_load_balance_nproc: system active = %d pool active = %d "
7474                 "hot team active = %d\n",
7475                 system_active, pool_active, hot_team_active));
7476 
7477   if (system_active < 0) {
7478     // There was an error reading the necessary info from /proc, so use the
7479     // thread limit algorithm instead. Once we set __kmp_global.g.g_dynamic_mode
7480     // = dynamic_thread_limit, we shouldn't wind up getting back here.
7481     __kmp_global.g.g_dynamic_mode = dynamic_thread_limit;
7482     KMP_WARNING(CantLoadBalUsing, "KMP_DYNAMIC_MODE=thread limit");
7483 
7484     // Make this call behave like the thread limit algorithm.
7485     retval = __kmp_avail_proc - __kmp_nth +
7486              (root->r.r_active ? 1 : root->r.r_hot_team->t.t_nproc);
7487     if (retval > set_nproc) {
7488       retval = set_nproc;
7489     }
7490     if (retval < KMP_MIN_NTH) {
7491       retval = KMP_MIN_NTH;
7492     }
7493 
7494     KB_TRACE(20, ("__kmp_load_balance_nproc: thread limit exit. retval:%d\n",
7495                   retval));
7496     return retval;
7497   }
7498 
7499   // There is a slight delay in the load balance algorithm in detecting new
7500   // running procs. The real system load at this instant should be at least as
7501   // large as the #active omp thread that are available to add to the team.
7502   if (system_active < team_curr_active) {
7503     system_active = team_curr_active;
7504   }
7505   retval = __kmp_avail_proc - system_active + team_curr_active;
7506   if (retval > set_nproc) {
7507     retval = set_nproc;
7508   }
7509   if (retval < KMP_MIN_NTH) {
7510     retval = KMP_MIN_NTH;
7511   }
7512 
7513   KB_TRACE(20, ("__kmp_load_balance_nproc: exit. retval:%d\n", retval));
7514   return retval;
7515 } // __kmp_load_balance_nproc()
7516 
7517 #endif /* USE_LOAD_BALANCE */
7518 
7519 /* ------------------------------------------------------------------------ */
7520 
7521 /* NOTE: this is called with the __kmp_init_lock held */
7522 void __kmp_cleanup(void) {
7523   int f;
7524 
7525   KA_TRACE(10, ("__kmp_cleanup: enter\n"));
7526 
7527   if (TCR_4(__kmp_init_parallel)) {
7528 #if KMP_HANDLE_SIGNALS
7529     __kmp_remove_signals();
7530 #endif
7531     TCW_4(__kmp_init_parallel, FALSE);
7532   }
7533 
7534   if (TCR_4(__kmp_init_middle)) {
7535 #if KMP_AFFINITY_SUPPORTED
7536     __kmp_affinity_uninitialize();
7537 #endif /* KMP_AFFINITY_SUPPORTED */
7538     __kmp_cleanup_hierarchy();
7539     TCW_4(__kmp_init_middle, FALSE);
7540   }
7541 
7542   KA_TRACE(10, ("__kmp_cleanup: go serial cleanup\n"));
7543 
7544   if (__kmp_init_serial) {
7545     __kmp_runtime_destroy();
7546     __kmp_init_serial = FALSE;
7547   }
7548 
7549   __kmp_cleanup_threadprivate_caches();
7550 
7551   for (f = 0; f < __kmp_threads_capacity; f++) {
7552     if (__kmp_root[f] != NULL) {
7553       __kmp_free(__kmp_root[f]);
7554       __kmp_root[f] = NULL;
7555     }
7556   }
7557   __kmp_free(__kmp_threads);
7558   // __kmp_threads and __kmp_root were allocated at once, as single block, so
7559   // there is no need in freeing __kmp_root.
7560   __kmp_threads = NULL;
7561   __kmp_root = NULL;
7562   __kmp_threads_capacity = 0;
7563 
7564 #if KMP_USE_DYNAMIC_LOCK
7565   __kmp_cleanup_indirect_user_locks();
7566 #else
7567   __kmp_cleanup_user_locks();
7568 #endif
7569 
7570 #if KMP_AFFINITY_SUPPORTED
7571   KMP_INTERNAL_FREE(CCAST(char *, __kmp_cpuinfo_file));
7572   __kmp_cpuinfo_file = NULL;
7573 #endif /* KMP_AFFINITY_SUPPORTED */
7574 
7575 #if KMP_USE_ADAPTIVE_LOCKS
7576 #if KMP_DEBUG_ADAPTIVE_LOCKS
7577   __kmp_print_speculative_stats();
7578 #endif
7579 #endif
7580   KMP_INTERNAL_FREE(__kmp_nested_nth.nth);
7581   __kmp_nested_nth.nth = NULL;
7582   __kmp_nested_nth.size = 0;
7583   __kmp_nested_nth.used = 0;
7584   KMP_INTERNAL_FREE(__kmp_nested_proc_bind.bind_types);
7585   __kmp_nested_proc_bind.bind_types = NULL;
7586   __kmp_nested_proc_bind.size = 0;
7587   __kmp_nested_proc_bind.used = 0;
7588 #if OMP_50_ENABLED
7589   if (__kmp_affinity_format) {
7590     KMP_INTERNAL_FREE(__kmp_affinity_format);
7591     __kmp_affinity_format = NULL;
7592   }
7593 #endif
7594 
7595   __kmp_i18n_catclose();
7596 
7597 #if KMP_USE_HIER_SCHED
7598   __kmp_hier_scheds.deallocate();
7599 #endif
7600 
7601 #if KMP_STATS_ENABLED
7602   __kmp_stats_fini();
7603 #endif
7604 
7605   KA_TRACE(10, ("__kmp_cleanup: exit\n"));
7606 }
7607 
7608 /* ------------------------------------------------------------------------ */
7609 
7610 int __kmp_ignore_mppbeg(void) {
7611   char *env;
7612 
7613   if ((env = getenv("KMP_IGNORE_MPPBEG")) != NULL) {
7614     if (__kmp_str_match_false(env))
7615       return FALSE;
7616   }
7617   // By default __kmpc_begin() is no-op.
7618   return TRUE;
7619 }
7620 
7621 int __kmp_ignore_mppend(void) {
7622   char *env;
7623 
7624   if ((env = getenv("KMP_IGNORE_MPPEND")) != NULL) {
7625     if (__kmp_str_match_false(env))
7626       return FALSE;
7627   }
7628   // By default __kmpc_end() is no-op.
7629   return TRUE;
7630 }
7631 
7632 void __kmp_internal_begin(void) {
7633   int gtid;
7634   kmp_root_t *root;
7635 
7636   /* this is a very important step as it will register new sibling threads
7637      and assign these new uber threads a new gtid */
7638   gtid = __kmp_entry_gtid();
7639   root = __kmp_threads[gtid]->th.th_root;
7640   KMP_ASSERT(KMP_UBER_GTID(gtid));
7641 
7642   if (root->r.r_begin)
7643     return;
7644   __kmp_acquire_lock(&root->r.r_begin_lock, gtid);
7645   if (root->r.r_begin) {
7646     __kmp_release_lock(&root->r.r_begin_lock, gtid);
7647     return;
7648   }
7649 
7650   root->r.r_begin = TRUE;
7651 
7652   __kmp_release_lock(&root->r.r_begin_lock, gtid);
7653 }
7654 
7655 /* ------------------------------------------------------------------------ */
7656 
7657 void __kmp_user_set_library(enum library_type arg) {
7658   int gtid;
7659   kmp_root_t *root;
7660   kmp_info_t *thread;
7661 
7662   /* first, make sure we are initialized so we can get our gtid */
7663 
7664   gtid = __kmp_entry_gtid();
7665   thread = __kmp_threads[gtid];
7666 
7667   root = thread->th.th_root;
7668 
7669   KA_TRACE(20, ("__kmp_user_set_library: enter T#%d, arg: %d, %d\n", gtid, arg,
7670                 library_serial));
7671   if (root->r.r_in_parallel) { /* Must be called in serial section of top-level
7672                                   thread */
7673     KMP_WARNING(SetLibraryIncorrectCall);
7674     return;
7675   }
7676 
7677   switch (arg) {
7678   case library_serial:
7679     thread->th.th_set_nproc = 0;
7680     set__nproc(thread, 1);
7681     break;
7682   case library_turnaround:
7683     thread->th.th_set_nproc = 0;
7684     set__nproc(thread, __kmp_dflt_team_nth ? __kmp_dflt_team_nth
7685                                            : __kmp_dflt_team_nth_ub);
7686     break;
7687   case library_throughput:
7688     thread->th.th_set_nproc = 0;
7689     set__nproc(thread, __kmp_dflt_team_nth ? __kmp_dflt_team_nth
7690                                            : __kmp_dflt_team_nth_ub);
7691     break;
7692   default:
7693     KMP_FATAL(UnknownLibraryType, arg);
7694   }
7695 
7696   __kmp_aux_set_library(arg);
7697 }
7698 
7699 void __kmp_aux_set_stacksize(size_t arg) {
7700   if (!__kmp_init_serial)
7701     __kmp_serial_initialize();
7702 
7703 #if KMP_OS_DARWIN
7704   if (arg & (0x1000 - 1)) {
7705     arg &= ~(0x1000 - 1);
7706     if (arg + 0x1000) /* check for overflow if we round up */
7707       arg += 0x1000;
7708   }
7709 #endif
7710   __kmp_acquire_bootstrap_lock(&__kmp_initz_lock);
7711 
7712   /* only change the default stacksize before the first parallel region */
7713   if (!TCR_4(__kmp_init_parallel)) {
7714     size_t value = arg; /* argument is in bytes */
7715 
7716     if (value < __kmp_sys_min_stksize)
7717       value = __kmp_sys_min_stksize;
7718     else if (value > KMP_MAX_STKSIZE)
7719       value = KMP_MAX_STKSIZE;
7720 
7721     __kmp_stksize = value;
7722 
7723     __kmp_env_stksize = TRUE; /* was KMP_STACKSIZE specified? */
7724   }
7725 
7726   __kmp_release_bootstrap_lock(&__kmp_initz_lock);
7727 }
7728 
7729 /* set the behaviour of the runtime library */
7730 /* TODO this can cause some odd behaviour with sibling parallelism... */
7731 void __kmp_aux_set_library(enum library_type arg) {
7732   __kmp_library = arg;
7733 
7734   switch (__kmp_library) {
7735   case library_serial: {
7736     KMP_INFORM(LibraryIsSerial);
7737   } break;
7738   case library_turnaround:
7739     if (__kmp_use_yield == 1 && !__kmp_use_yield_exp_set)
7740       __kmp_use_yield = 2; // only yield when oversubscribed
7741     break;
7742   case library_throughput:
7743     if (__kmp_dflt_blocktime == KMP_MAX_BLOCKTIME)
7744       __kmp_dflt_blocktime = 200;
7745     break;
7746   default:
7747     KMP_FATAL(UnknownLibraryType, arg);
7748   }
7749 }
7750 
7751 /* Getting team information common for all team API */
7752 // Returns NULL if not in teams construct
7753 static kmp_team_t *__kmp_aux_get_team_info(int &teams_serialized) {
7754   kmp_info_t *thr = __kmp_entry_thread();
7755   teams_serialized = 0;
7756   if (thr->th.th_teams_microtask) {
7757     kmp_team_t *team = thr->th.th_team;
7758     int tlevel = thr->th.th_teams_level; // the level of the teams construct
7759     int ii = team->t.t_level;
7760     teams_serialized = team->t.t_serialized;
7761     int level = tlevel + 1;
7762     KMP_DEBUG_ASSERT(ii >= tlevel);
7763     while (ii > level) {
7764       for (teams_serialized = team->t.t_serialized;
7765            (teams_serialized > 0) && (ii > level); teams_serialized--, ii--) {
7766       }
7767       if (team->t.t_serialized && (!teams_serialized)) {
7768         team = team->t.t_parent;
7769         continue;
7770       }
7771       if (ii > level) {
7772         team = team->t.t_parent;
7773         ii--;
7774       }
7775     }
7776     return team;
7777   }
7778   return NULL;
7779 }
7780 
7781 int __kmp_aux_get_team_num() {
7782   int serialized;
7783   kmp_team_t *team = __kmp_aux_get_team_info(serialized);
7784   if (team) {
7785     if (serialized > 1) {
7786       return 0; // teams region is serialized ( 1 team of 1 thread ).
7787     } else {
7788       return team->t.t_master_tid;
7789     }
7790   }
7791   return 0;
7792 }
7793 
7794 int __kmp_aux_get_num_teams() {
7795   int serialized;
7796   kmp_team_t *team = __kmp_aux_get_team_info(serialized);
7797   if (team) {
7798     if (serialized > 1) {
7799       return 1;
7800     } else {
7801       return team->t.t_parent->t.t_nproc;
7802     }
7803   }
7804   return 1;
7805 }
7806 
7807 /* ------------------------------------------------------------------------ */
7808 
7809 #if OMP_50_ENABLED
7810 /*
7811  * Affinity Format Parser
7812  *
7813  * Field is in form of: %[[[0].]size]type
7814  * % and type are required (%% means print a literal '%')
7815  * type is either single char or long name surrounded by {},
7816  * e.g., N or {num_threads}
7817  * 0 => leading zeros
7818  * . => right justified when size is specified
7819  * by default output is left justified
7820  * size is the *minimum* field length
7821  * All other characters are printed as is
7822  *
7823  * Available field types:
7824  * L {thread_level}      - omp_get_level()
7825  * n {thread_num}        - omp_get_thread_num()
7826  * h {host}              - name of host machine
7827  * P {process_id}        - process id (integer)
7828  * T {thread_identifier} - native thread identifier (integer)
7829  * N {num_threads}       - omp_get_num_threads()
7830  * A {ancestor_tnum}     - omp_get_ancestor_thread_num(omp_get_level()-1)
7831  * a {thread_affinity}   - comma separated list of integers or integer ranges
7832  *                         (values of affinity mask)
7833  *
7834  * Implementation-specific field types can be added
7835  * If a type is unknown, print "undefined"
7836 */
7837 
7838 // Structure holding the short name, long name, and corresponding data type
7839 // for snprintf.  A table of these will represent the entire valid keyword
7840 // field types.
7841 typedef struct kmp_affinity_format_field_t {
7842   char short_name; // from spec e.g., L -> thread level
7843   const char *long_name; // from spec thread_level -> thread level
7844   char field_format; // data type for snprintf (typically 'd' or 's'
7845   // for integer or string)
7846 } kmp_affinity_format_field_t;
7847 
7848 static const kmp_affinity_format_field_t __kmp_affinity_format_table[] = {
7849 #if KMP_AFFINITY_SUPPORTED
7850     {'A', "thread_affinity", 's'},
7851 #endif
7852     {'t', "team_num", 'd'},
7853     {'T', "num_teams", 'd'},
7854     {'L', "nesting_level", 'd'},
7855     {'n', "thread_num", 'd'},
7856     {'N', "num_threads", 'd'},
7857     {'a', "ancestor_tnum", 'd'},
7858     {'H', "host", 's'},
7859     {'P', "process_id", 'd'},
7860     {'i', "native_thread_id", 'd'}};
7861 
7862 // Return the number of characters it takes to hold field
7863 static int __kmp_aux_capture_affinity_field(int gtid, const kmp_info_t *th,
7864                                             const char **ptr,
7865                                             kmp_str_buf_t *field_buffer) {
7866   int rc, format_index, field_value;
7867   const char *width_left, *width_right;
7868   bool pad_zeros, right_justify, parse_long_name, found_valid_name;
7869   static const int FORMAT_SIZE = 20;
7870   char format[FORMAT_SIZE] = {0};
7871   char absolute_short_name = 0;
7872 
7873   KMP_DEBUG_ASSERT(gtid >= 0);
7874   KMP_DEBUG_ASSERT(th);
7875   KMP_DEBUG_ASSERT(**ptr == '%');
7876   KMP_DEBUG_ASSERT(field_buffer);
7877 
7878   __kmp_str_buf_clear(field_buffer);
7879 
7880   // Skip the initial %
7881   (*ptr)++;
7882 
7883   // Check for %% first
7884   if (**ptr == '%') {
7885     __kmp_str_buf_cat(field_buffer, "%", 1);
7886     (*ptr)++; // skip over the second %
7887     return 1;
7888   }
7889 
7890   // Parse field modifiers if they are present
7891   pad_zeros = false;
7892   if (**ptr == '0') {
7893     pad_zeros = true;
7894     (*ptr)++; // skip over 0
7895   }
7896   right_justify = false;
7897   if (**ptr == '.') {
7898     right_justify = true;
7899     (*ptr)++; // skip over .
7900   }
7901   // Parse width of field: [width_left, width_right)
7902   width_left = width_right = NULL;
7903   if (**ptr >= '0' && **ptr <= '9') {
7904     width_left = *ptr;
7905     SKIP_DIGITS(*ptr);
7906     width_right = *ptr;
7907   }
7908 
7909   // Create the format for KMP_SNPRINTF based on flags parsed above
7910   format_index = 0;
7911   format[format_index++] = '%';
7912   if (!right_justify)
7913     format[format_index++] = '-';
7914   if (pad_zeros)
7915     format[format_index++] = '0';
7916   if (width_left && width_right) {
7917     int i = 0;
7918     // Only allow 8 digit number widths.
7919     // This also prevents overflowing format variable
7920     while (i < 8 && width_left < width_right) {
7921       format[format_index++] = *width_left;
7922       width_left++;
7923       i++;
7924     }
7925   }
7926 
7927   // Parse a name (long or short)
7928   // Canonicalize the name into absolute_short_name
7929   found_valid_name = false;
7930   parse_long_name = (**ptr == '{');
7931   if (parse_long_name)
7932     (*ptr)++; // skip initial left brace
7933   for (size_t i = 0; i < sizeof(__kmp_affinity_format_table) /
7934                              sizeof(__kmp_affinity_format_table[0]);
7935        ++i) {
7936     char short_name = __kmp_affinity_format_table[i].short_name;
7937     const char *long_name = __kmp_affinity_format_table[i].long_name;
7938     char field_format = __kmp_affinity_format_table[i].field_format;
7939     if (parse_long_name) {
7940       int length = KMP_STRLEN(long_name);
7941       if (strncmp(*ptr, long_name, length) == 0) {
7942         found_valid_name = true;
7943         (*ptr) += length; // skip the long name
7944       }
7945     } else if (**ptr == short_name) {
7946       found_valid_name = true;
7947       (*ptr)++; // skip the short name
7948     }
7949     if (found_valid_name) {
7950       format[format_index++] = field_format;
7951       format[format_index++] = '\0';
7952       absolute_short_name = short_name;
7953       break;
7954     }
7955   }
7956   if (parse_long_name) {
7957     if (**ptr != '}') {
7958       absolute_short_name = 0;
7959     } else {
7960       (*ptr)++; // skip over the right brace
7961     }
7962   }
7963 
7964   // Attempt to fill the buffer with the requested
7965   // value using snprintf within __kmp_str_buf_print()
7966   switch (absolute_short_name) {
7967   case 't':
7968     rc = __kmp_str_buf_print(field_buffer, format, __kmp_aux_get_team_num());
7969     break;
7970   case 'T':
7971     rc = __kmp_str_buf_print(field_buffer, format, __kmp_aux_get_num_teams());
7972     break;
7973   case 'L':
7974     rc = __kmp_str_buf_print(field_buffer, format, th->th.th_team->t.t_level);
7975     break;
7976   case 'n':
7977     rc = __kmp_str_buf_print(field_buffer, format, __kmp_tid_from_gtid(gtid));
7978     break;
7979   case 'H': {
7980     static const int BUFFER_SIZE = 256;
7981     char buf[BUFFER_SIZE];
7982     __kmp_expand_host_name(buf, BUFFER_SIZE);
7983     rc = __kmp_str_buf_print(field_buffer, format, buf);
7984   } break;
7985   case 'P':
7986     rc = __kmp_str_buf_print(field_buffer, format, getpid());
7987     break;
7988   case 'i':
7989     rc = __kmp_str_buf_print(field_buffer, format, __kmp_gettid());
7990     break;
7991   case 'N':
7992     rc = __kmp_str_buf_print(field_buffer, format, th->th.th_team->t.t_nproc);
7993     break;
7994   case 'a':
7995     field_value =
7996         __kmp_get_ancestor_thread_num(gtid, th->th.th_team->t.t_level - 1);
7997     rc = __kmp_str_buf_print(field_buffer, format, field_value);
7998     break;
7999 #if KMP_AFFINITY_SUPPORTED
8000   case 'A': {
8001     kmp_str_buf_t buf;
8002     __kmp_str_buf_init(&buf);
8003     __kmp_affinity_str_buf_mask(&buf, th->th.th_affin_mask);
8004     rc = __kmp_str_buf_print(field_buffer, format, buf.str);
8005     __kmp_str_buf_free(&buf);
8006   } break;
8007 #endif
8008   default:
8009     // According to spec, If an implementation does not have info for field
8010     // type, then "undefined" is printed
8011     rc = __kmp_str_buf_print(field_buffer, "%s", "undefined");
8012     // Skip the field
8013     if (parse_long_name) {
8014       SKIP_TOKEN(*ptr);
8015       if (**ptr == '}')
8016         (*ptr)++;
8017     } else {
8018       (*ptr)++;
8019     }
8020   }
8021 
8022   KMP_ASSERT(format_index <= FORMAT_SIZE);
8023   return rc;
8024 }
8025 
8026 /*
8027  * Return number of characters needed to hold the affinity string
8028  * (not including null byte character)
8029  * The resultant string is printed to buffer, which the caller can then
8030  * handle afterwards
8031 */
8032 size_t __kmp_aux_capture_affinity(int gtid, const char *format,
8033                                   kmp_str_buf_t *buffer) {
8034   const char *parse_ptr;
8035   size_t retval;
8036   const kmp_info_t *th;
8037   kmp_str_buf_t field;
8038 
8039   KMP_DEBUG_ASSERT(buffer);
8040   KMP_DEBUG_ASSERT(gtid >= 0);
8041 
8042   __kmp_str_buf_init(&field);
8043   __kmp_str_buf_clear(buffer);
8044 
8045   th = __kmp_threads[gtid];
8046   retval = 0;
8047 
8048   // If format is NULL or zero-length string, then we use
8049   // affinity-format-var ICV
8050   parse_ptr = format;
8051   if (parse_ptr == NULL || *parse_ptr == '\0') {
8052     parse_ptr = __kmp_affinity_format;
8053   }
8054   KMP_DEBUG_ASSERT(parse_ptr);
8055 
8056   while (*parse_ptr != '\0') {
8057     // Parse a field
8058     if (*parse_ptr == '%') {
8059       // Put field in the buffer
8060       int rc = __kmp_aux_capture_affinity_field(gtid, th, &parse_ptr, &field);
8061       __kmp_str_buf_catbuf(buffer, &field);
8062       retval += rc;
8063     } else {
8064       // Put literal character in buffer
8065       __kmp_str_buf_cat(buffer, parse_ptr, 1);
8066       retval++;
8067       parse_ptr++;
8068     }
8069   }
8070   __kmp_str_buf_free(&field);
8071   return retval;
8072 }
8073 
8074 // Displays the affinity string to stdout
8075 void __kmp_aux_display_affinity(int gtid, const char *format) {
8076   kmp_str_buf_t buf;
8077   __kmp_str_buf_init(&buf);
8078   __kmp_aux_capture_affinity(gtid, format, &buf);
8079   __kmp_fprintf(kmp_out, "%s" KMP_END_OF_LINE, buf.str);
8080   __kmp_str_buf_free(&buf);
8081 }
8082 #endif // OMP_50_ENABLED
8083 
8084 /* ------------------------------------------------------------------------ */
8085 
8086 void __kmp_aux_set_blocktime(int arg, kmp_info_t *thread, int tid) {
8087   int blocktime = arg; /* argument is in milliseconds */
8088 #if KMP_USE_MONITOR
8089   int bt_intervals;
8090 #endif
8091   int bt_set;
8092 
8093   __kmp_save_internal_controls(thread);
8094 
8095   /* Normalize and set blocktime for the teams */
8096   if (blocktime < KMP_MIN_BLOCKTIME)
8097     blocktime = KMP_MIN_BLOCKTIME;
8098   else if (blocktime > KMP_MAX_BLOCKTIME)
8099     blocktime = KMP_MAX_BLOCKTIME;
8100 
8101   set__blocktime_team(thread->th.th_team, tid, blocktime);
8102   set__blocktime_team(thread->th.th_serial_team, 0, blocktime);
8103 
8104 #if KMP_USE_MONITOR
8105   /* Calculate and set blocktime intervals for the teams */
8106   bt_intervals = KMP_INTERVALS_FROM_BLOCKTIME(blocktime, __kmp_monitor_wakeups);
8107 
8108   set__bt_intervals_team(thread->th.th_team, tid, bt_intervals);
8109   set__bt_intervals_team(thread->th.th_serial_team, 0, bt_intervals);
8110 #endif
8111 
8112   /* Set whether blocktime has been set to "TRUE" */
8113   bt_set = TRUE;
8114 
8115   set__bt_set_team(thread->th.th_team, tid, bt_set);
8116   set__bt_set_team(thread->th.th_serial_team, 0, bt_set);
8117 #if KMP_USE_MONITOR
8118   KF_TRACE(10, ("kmp_set_blocktime: T#%d(%d:%d), blocktime=%d, "
8119                 "bt_intervals=%d, monitor_updates=%d\n",
8120                 __kmp_gtid_from_tid(tid, thread->th.th_team),
8121                 thread->th.th_team->t.t_id, tid, blocktime, bt_intervals,
8122                 __kmp_monitor_wakeups));
8123 #else
8124   KF_TRACE(10, ("kmp_set_blocktime: T#%d(%d:%d), blocktime=%d\n",
8125                 __kmp_gtid_from_tid(tid, thread->th.th_team),
8126                 thread->th.th_team->t.t_id, tid, blocktime));
8127 #endif
8128 }
8129 
8130 void __kmp_aux_set_defaults(char const *str, int len) {
8131   if (!__kmp_init_serial) {
8132     __kmp_serial_initialize();
8133   }
8134   __kmp_env_initialize(str);
8135 
8136   if (__kmp_settings
8137 #if OMP_40_ENABLED
8138       || __kmp_display_env || __kmp_display_env_verbose
8139 #endif // OMP_40_ENABLED
8140       ) {
8141     __kmp_env_print();
8142   }
8143 } // __kmp_aux_set_defaults
8144 
8145 /* ------------------------------------------------------------------------ */
8146 /* internal fast reduction routines */
8147 
8148 PACKED_REDUCTION_METHOD_T
8149 __kmp_determine_reduction_method(
8150     ident_t *loc, kmp_int32 global_tid, kmp_int32 num_vars, size_t reduce_size,
8151     void *reduce_data, void (*reduce_func)(void *lhs_data, void *rhs_data),
8152     kmp_critical_name *lck) {
8153 
8154   // Default reduction method: critical construct ( lck != NULL, like in current
8155   // PAROPT )
8156   // If ( reduce_data!=NULL && reduce_func!=NULL ): the tree-reduction method
8157   // can be selected by RTL
8158   // If loc->flags contains KMP_IDENT_ATOMIC_REDUCE, the atomic reduce method
8159   // can be selected by RTL
8160   // Finally, it's up to OpenMP RTL to make a decision on which method to select
8161   // among generated by PAROPT.
8162 
8163   PACKED_REDUCTION_METHOD_T retval;
8164 
8165   int team_size;
8166 
8167   KMP_DEBUG_ASSERT(loc); // it would be nice to test ( loc != 0 )
8168   KMP_DEBUG_ASSERT(lck); // it would be nice to test ( lck != 0 )
8169 
8170 #define FAST_REDUCTION_ATOMIC_METHOD_GENERATED                                 \
8171   ((loc->flags & (KMP_IDENT_ATOMIC_REDUCE)) == (KMP_IDENT_ATOMIC_REDUCE))
8172 #define FAST_REDUCTION_TREE_METHOD_GENERATED ((reduce_data) && (reduce_func))
8173 
8174   retval = critical_reduce_block;
8175 
8176   // another choice of getting a team size (with 1 dynamic deference) is slower
8177   team_size = __kmp_get_team_num_threads(global_tid);
8178   if (team_size == 1) {
8179 
8180     retval = empty_reduce_block;
8181 
8182   } else {
8183 
8184     int atomic_available = FAST_REDUCTION_ATOMIC_METHOD_GENERATED;
8185 
8186 #if KMP_ARCH_X86_64 || KMP_ARCH_PPC64 || KMP_ARCH_AARCH64 || KMP_ARCH_MIPS64
8187 
8188 #if KMP_OS_LINUX || KMP_OS_DRAGONFLY || KMP_OS_FREEBSD || KMP_OS_NETBSD ||     \
8189     KMP_OS_OPENBSD || KMP_OS_WINDOWS || KMP_OS_DARWIN || KMP_OS_HURD
8190 
8191     int teamsize_cutoff = 4;
8192 
8193 #if KMP_MIC_SUPPORTED
8194     if (__kmp_mic_type != non_mic) {
8195       teamsize_cutoff = 8;
8196     }
8197 #endif
8198     int tree_available = FAST_REDUCTION_TREE_METHOD_GENERATED;
8199     if (tree_available) {
8200       if (team_size <= teamsize_cutoff) {
8201         if (atomic_available) {
8202           retval = atomic_reduce_block;
8203         }
8204       } else {
8205         retval = TREE_REDUCE_BLOCK_WITH_REDUCTION_BARRIER;
8206       }
8207     } else if (atomic_available) {
8208       retval = atomic_reduce_block;
8209     }
8210 #else
8211 #error "Unknown or unsupported OS"
8212 #endif // KMP_OS_LINUX || KMP_OS_DRAGONFLY || KMP_OS_FREEBSD || KMP_OS_NETBSD ||
8213        // KMP_OS_OPENBSD || KMP_OS_WINDOWS || KMP_OS_DARWIN || KMP_OS_HURD
8214 
8215 #elif KMP_ARCH_X86 || KMP_ARCH_ARM || KMP_ARCH_AARCH || KMP_ARCH_MIPS
8216 
8217 #if KMP_OS_LINUX || KMP_OS_WINDOWS || KMP_OS_HURD
8218 
8219     // basic tuning
8220 
8221     if (atomic_available) {
8222       if (num_vars <= 2) { // && ( team_size <= 8 ) due to false-sharing ???
8223         retval = atomic_reduce_block;
8224       }
8225     } // otherwise: use critical section
8226 
8227 #elif KMP_OS_DARWIN
8228 
8229     int tree_available = FAST_REDUCTION_TREE_METHOD_GENERATED;
8230     if (atomic_available && (num_vars <= 3)) {
8231       retval = atomic_reduce_block;
8232     } else if (tree_available) {
8233       if ((reduce_size > (9 * sizeof(kmp_real64))) &&
8234           (reduce_size < (2000 * sizeof(kmp_real64)))) {
8235         retval = TREE_REDUCE_BLOCK_WITH_PLAIN_BARRIER;
8236       }
8237     } // otherwise: use critical section
8238 
8239 #else
8240 #error "Unknown or unsupported OS"
8241 #endif
8242 
8243 #else
8244 #error "Unknown or unsupported architecture"
8245 #endif
8246   }
8247 
8248   // KMP_FORCE_REDUCTION
8249 
8250   // If the team is serialized (team_size == 1), ignore the forced reduction
8251   // method and stay with the unsynchronized method (empty_reduce_block)
8252   if (__kmp_force_reduction_method != reduction_method_not_defined &&
8253       team_size != 1) {
8254 
8255     PACKED_REDUCTION_METHOD_T forced_retval = critical_reduce_block;
8256 
8257     int atomic_available, tree_available;
8258 
8259     switch ((forced_retval = __kmp_force_reduction_method)) {
8260     case critical_reduce_block:
8261       KMP_ASSERT(lck); // lck should be != 0
8262       break;
8263 
8264     case atomic_reduce_block:
8265       atomic_available = FAST_REDUCTION_ATOMIC_METHOD_GENERATED;
8266       if (!atomic_available) {
8267         KMP_WARNING(RedMethodNotSupported, "atomic");
8268         forced_retval = critical_reduce_block;
8269       }
8270       break;
8271 
8272     case tree_reduce_block:
8273       tree_available = FAST_REDUCTION_TREE_METHOD_GENERATED;
8274       if (!tree_available) {
8275         KMP_WARNING(RedMethodNotSupported, "tree");
8276         forced_retval = critical_reduce_block;
8277       } else {
8278 #if KMP_FAST_REDUCTION_BARRIER
8279         forced_retval = TREE_REDUCE_BLOCK_WITH_REDUCTION_BARRIER;
8280 #endif
8281       }
8282       break;
8283 
8284     default:
8285       KMP_ASSERT(0); // "unsupported method specified"
8286     }
8287 
8288     retval = forced_retval;
8289   }
8290 
8291   KA_TRACE(10, ("reduction method selected=%08x\n", retval));
8292 
8293 #undef FAST_REDUCTION_TREE_METHOD_GENERATED
8294 #undef FAST_REDUCTION_ATOMIC_METHOD_GENERATED
8295 
8296   return (retval);
8297 }
8298 
8299 // this function is for testing set/get/determine reduce method
8300 kmp_int32 __kmp_get_reduce_method(void) {
8301   return ((__kmp_entry_thread()->th.th_local.packed_reduction_method) >> 8);
8302 }
8303 
8304 #if OMP_50_ENABLED
8305 
8306 // Soft pause sets up threads to ignore blocktime and just go to sleep.
8307 // Spin-wait code checks __kmp_pause_status and reacts accordingly.
8308 void __kmp_soft_pause() { __kmp_pause_status = kmp_soft_paused; }
8309 
8310 // Hard pause shuts down the runtime completely.  Resume happens naturally when
8311 // OpenMP is used subsequently.
8312 void __kmp_hard_pause() {
8313   __kmp_pause_status = kmp_hard_paused;
8314   __kmp_internal_end_thread(-1);
8315 }
8316 
8317 // Soft resume sets __kmp_pause_status, and wakes up all threads.
8318 void __kmp_resume_if_soft_paused() {
8319   if (__kmp_pause_status == kmp_soft_paused) {
8320     __kmp_pause_status = kmp_not_paused;
8321 
8322     for (int gtid = 1; gtid < __kmp_threads_capacity; ++gtid) {
8323       kmp_info_t *thread = __kmp_threads[gtid];
8324       if (thread) { // Wake it if sleeping
8325         kmp_flag_64 fl(&thread->th.th_bar[bs_forkjoin_barrier].bb.b_go, thread);
8326         if (fl.is_sleeping())
8327           fl.resume(gtid);
8328         else if (__kmp_try_suspend_mx(thread)) { // got suspend lock
8329           __kmp_unlock_suspend_mx(thread); // unlock it; it won't sleep
8330         } else { // thread holds the lock and may sleep soon
8331           do { // until either the thread sleeps, or we can get the lock
8332             if (fl.is_sleeping()) {
8333               fl.resume(gtid);
8334               break;
8335             } else if (__kmp_try_suspend_mx(thread)) {
8336               __kmp_unlock_suspend_mx(thread);
8337               break;
8338             }
8339           } while (1);
8340         }
8341       }
8342     }
8343   }
8344 }
8345 
8346 // This function is called via __kmpc_pause_resource. Returns 0 if successful.
8347 // TODO: add warning messages
8348 int __kmp_pause_resource(kmp_pause_status_t level) {
8349   if (level == kmp_not_paused) { // requesting resume
8350     if (__kmp_pause_status == kmp_not_paused) {
8351       // error message about runtime not being paused, so can't resume
8352       return 1;
8353     } else {
8354       KMP_DEBUG_ASSERT(__kmp_pause_status == kmp_soft_paused ||
8355                        __kmp_pause_status == kmp_hard_paused);
8356       __kmp_pause_status = kmp_not_paused;
8357       return 0;
8358     }
8359   } else if (level == kmp_soft_paused) { // requesting soft pause
8360     if (__kmp_pause_status != kmp_not_paused) {
8361       // error message about already being paused
8362       return 1;
8363     } else {
8364       __kmp_soft_pause();
8365       return 0;
8366     }
8367   } else if (level == kmp_hard_paused) { // requesting hard pause
8368     if (__kmp_pause_status != kmp_not_paused) {
8369       // error message about already being paused
8370       return 1;
8371     } else {
8372       __kmp_hard_pause();
8373       return 0;
8374     }
8375   } else {
8376     // error message about invalid level
8377     return 1;
8378   }
8379 }
8380 
8381 #endif // OMP_50_ENABLED
8382