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 KMP_BUILD_ASSERT(sizeof(kmp_sched_t) == sizeof(int));
2805 KMP_BUILD_ASSERT(sizeof(enum sched_type) == sizeof(int));
2806 
2807 /* Changes def_sched_var ICV values (run-time schedule kind and chunk) */
2808 void __kmp_set_schedule(int gtid, kmp_sched_t kind, int chunk) {
2809   kmp_info_t *thread;
2810   kmp_sched_t orig_kind;
2811   //    kmp_team_t *team;
2812 
2813   KF_TRACE(10, ("__kmp_set_schedule: new schedule for thread %d = (%d, %d)\n",
2814                 gtid, (int)kind, chunk));
2815   KMP_DEBUG_ASSERT(__kmp_init_serial);
2816 
2817   // Check if the kind parameter is valid, correct if needed.
2818   // Valid parameters should fit in one of two intervals - standard or extended:
2819   //       <lower>, <valid>, <upper_std>, <lower_ext>, <valid>, <upper>
2820   // 2008-01-25: 0,  1 - 4,       5,         100,     101 - 102, 103
2821   orig_kind = kind;
2822   kind = __kmp_sched_without_mods(kind);
2823 
2824   if (kind <= kmp_sched_lower || kind >= kmp_sched_upper ||
2825       (kind <= kmp_sched_lower_ext && kind >= kmp_sched_upper_std)) {
2826     // TODO: Hint needs attention in case we change the default schedule.
2827     __kmp_msg(kmp_ms_warning, KMP_MSG(ScheduleKindOutOfRange, kind),
2828               KMP_HNT(DefaultScheduleKindUsed, "static, no chunk"),
2829               __kmp_msg_null);
2830     kind = kmp_sched_default;
2831     chunk = 0; // ignore chunk value in case of bad kind
2832   }
2833 
2834   thread = __kmp_threads[gtid];
2835 
2836   __kmp_save_internal_controls(thread);
2837 
2838   if (kind < kmp_sched_upper_std) {
2839     if (kind == kmp_sched_static && chunk < KMP_DEFAULT_CHUNK) {
2840       // differ static chunked vs. unchunked:  chunk should be invalid to
2841       // indicate unchunked schedule (which is the default)
2842       thread->th.th_current_task->td_icvs.sched.r_sched_type = kmp_sch_static;
2843     } else {
2844       thread->th.th_current_task->td_icvs.sched.r_sched_type =
2845           __kmp_sch_map[kind - kmp_sched_lower - 1];
2846     }
2847   } else {
2848     //    __kmp_sch_map[ kind - kmp_sched_lower_ext + kmp_sched_upper_std -
2849     //    kmp_sched_lower - 2 ];
2850     thread->th.th_current_task->td_icvs.sched.r_sched_type =
2851         __kmp_sch_map[kind - kmp_sched_lower_ext + kmp_sched_upper_std -
2852                       kmp_sched_lower - 2];
2853   }
2854   __kmp_sched_apply_mods_intkind(
2855       orig_kind, &(thread->th.th_current_task->td_icvs.sched.r_sched_type));
2856   if (kind == kmp_sched_auto || chunk < 1) {
2857     // ignore parameter chunk for schedule auto
2858     thread->th.th_current_task->td_icvs.sched.chunk = KMP_DEFAULT_CHUNK;
2859   } else {
2860     thread->th.th_current_task->td_icvs.sched.chunk = chunk;
2861   }
2862 }
2863 
2864 /* Gets def_sched_var ICV values */
2865 void __kmp_get_schedule(int gtid, kmp_sched_t *kind, int *chunk) {
2866   kmp_info_t *thread;
2867   enum sched_type th_type;
2868 
2869   KF_TRACE(10, ("__kmp_get_schedule: thread %d\n", gtid));
2870   KMP_DEBUG_ASSERT(__kmp_init_serial);
2871 
2872   thread = __kmp_threads[gtid];
2873 
2874   th_type = thread->th.th_current_task->td_icvs.sched.r_sched_type;
2875   switch (SCHEDULE_WITHOUT_MODIFIERS(th_type)) {
2876   case kmp_sch_static:
2877   case kmp_sch_static_greedy:
2878   case kmp_sch_static_balanced:
2879     *kind = kmp_sched_static;
2880     __kmp_sched_apply_mods_stdkind(kind, th_type);
2881     *chunk = 0; // chunk was not set, try to show this fact via zero value
2882     return;
2883   case kmp_sch_static_chunked:
2884     *kind = kmp_sched_static;
2885     break;
2886   case kmp_sch_dynamic_chunked:
2887     *kind = kmp_sched_dynamic;
2888     break;
2889   case kmp_sch_guided_chunked:
2890   case kmp_sch_guided_iterative_chunked:
2891   case kmp_sch_guided_analytical_chunked:
2892     *kind = kmp_sched_guided;
2893     break;
2894   case kmp_sch_auto:
2895     *kind = kmp_sched_auto;
2896     break;
2897   case kmp_sch_trapezoidal:
2898     *kind = kmp_sched_trapezoidal;
2899     break;
2900 #if KMP_STATIC_STEAL_ENABLED
2901   case kmp_sch_static_steal:
2902     *kind = kmp_sched_static_steal;
2903     break;
2904 #endif
2905   default:
2906     KMP_FATAL(UnknownSchedulingType, th_type);
2907   }
2908 
2909   __kmp_sched_apply_mods_stdkind(kind, th_type);
2910   *chunk = thread->th.th_current_task->td_icvs.sched.chunk;
2911 }
2912 
2913 int __kmp_get_ancestor_thread_num(int gtid, int level) {
2914 
2915   int ii, dd;
2916   kmp_team_t *team;
2917   kmp_info_t *thr;
2918 
2919   KF_TRACE(10, ("__kmp_get_ancestor_thread_num: thread %d %d\n", gtid, level));
2920   KMP_DEBUG_ASSERT(__kmp_init_serial);
2921 
2922   // validate level
2923   if (level == 0)
2924     return 0;
2925   if (level < 0)
2926     return -1;
2927   thr = __kmp_threads[gtid];
2928   team = thr->th.th_team;
2929   ii = team->t.t_level;
2930   if (level > ii)
2931     return -1;
2932 
2933 #if OMP_40_ENABLED
2934   if (thr->th.th_teams_microtask) {
2935     // AC: we are in teams region where multiple nested teams have same level
2936     int tlevel = thr->th.th_teams_level; // the level of the teams construct
2937     if (level <=
2938         tlevel) { // otherwise usual algorithm works (will not touch the teams)
2939       KMP_DEBUG_ASSERT(ii >= tlevel);
2940       // AC: As we need to pass by the teams league, we need to artificially
2941       // increase ii
2942       if (ii == tlevel) {
2943         ii += 2; // three teams have same level
2944       } else {
2945         ii++; // two teams have same level
2946       }
2947     }
2948   }
2949 #endif
2950 
2951   if (ii == level)
2952     return __kmp_tid_from_gtid(gtid);
2953 
2954   dd = team->t.t_serialized;
2955   level++;
2956   while (ii > level) {
2957     for (dd = team->t.t_serialized; (dd > 0) && (ii > level); dd--, ii--) {
2958     }
2959     if ((team->t.t_serialized) && (!dd)) {
2960       team = team->t.t_parent;
2961       continue;
2962     }
2963     if (ii > level) {
2964       team = team->t.t_parent;
2965       dd = team->t.t_serialized;
2966       ii--;
2967     }
2968   }
2969 
2970   return (dd > 1) ? (0) : (team->t.t_master_tid);
2971 }
2972 
2973 int __kmp_get_team_size(int gtid, int level) {
2974 
2975   int ii, dd;
2976   kmp_team_t *team;
2977   kmp_info_t *thr;
2978 
2979   KF_TRACE(10, ("__kmp_get_team_size: thread %d %d\n", gtid, level));
2980   KMP_DEBUG_ASSERT(__kmp_init_serial);
2981 
2982   // validate level
2983   if (level == 0)
2984     return 1;
2985   if (level < 0)
2986     return -1;
2987   thr = __kmp_threads[gtid];
2988   team = thr->th.th_team;
2989   ii = team->t.t_level;
2990   if (level > ii)
2991     return -1;
2992 
2993 #if OMP_40_ENABLED
2994   if (thr->th.th_teams_microtask) {
2995     // AC: we are in teams region where multiple nested teams have same level
2996     int tlevel = thr->th.th_teams_level; // the level of the teams construct
2997     if (level <=
2998         tlevel) { // otherwise usual algorithm works (will not touch the teams)
2999       KMP_DEBUG_ASSERT(ii >= tlevel);
3000       // AC: As we need to pass by the teams league, we need to artificially
3001       // increase ii
3002       if (ii == tlevel) {
3003         ii += 2; // three teams have same level
3004       } else {
3005         ii++; // two teams have same level
3006       }
3007     }
3008   }
3009 #endif
3010 
3011   while (ii > level) {
3012     for (dd = team->t.t_serialized; (dd > 0) && (ii > level); dd--, ii--) {
3013     }
3014     if (team->t.t_serialized && (!dd)) {
3015       team = team->t.t_parent;
3016       continue;
3017     }
3018     if (ii > level) {
3019       team = team->t.t_parent;
3020       ii--;
3021     }
3022   }
3023 
3024   return team->t.t_nproc;
3025 }
3026 
3027 kmp_r_sched_t __kmp_get_schedule_global() {
3028   // This routine created because pairs (__kmp_sched, __kmp_chunk) and
3029   // (__kmp_static, __kmp_guided) may be changed by kmp_set_defaults
3030   // independently. So one can get the updated schedule here.
3031 
3032   kmp_r_sched_t r_sched;
3033 
3034   // create schedule from 4 globals: __kmp_sched, __kmp_chunk, __kmp_static,
3035   // __kmp_guided. __kmp_sched should keep original value, so that user can set
3036   // KMP_SCHEDULE multiple times, and thus have different run-time schedules in
3037   // different roots (even in OMP 2.5)
3038   enum sched_type s = SCHEDULE_WITHOUT_MODIFIERS(__kmp_sched);
3039 #if OMP_45_ENABLED
3040   enum sched_type sched_modifiers = SCHEDULE_GET_MODIFIERS(__kmp_sched);
3041 #endif
3042   if (s == kmp_sch_static) {
3043     // replace STATIC with more detailed schedule (balanced or greedy)
3044     r_sched.r_sched_type = __kmp_static;
3045   } else if (s == kmp_sch_guided_chunked) {
3046     // replace GUIDED with more detailed schedule (iterative or analytical)
3047     r_sched.r_sched_type = __kmp_guided;
3048   } else { // (STATIC_CHUNKED), or (DYNAMIC_CHUNKED), or other
3049     r_sched.r_sched_type = __kmp_sched;
3050   }
3051 #if OMP_45_ENABLED
3052   SCHEDULE_SET_MODIFIERS(r_sched.r_sched_type, sched_modifiers);
3053 #endif
3054 
3055   if (__kmp_chunk < KMP_DEFAULT_CHUNK) {
3056     // __kmp_chunk may be wrong here (if it was not ever set)
3057     r_sched.chunk = KMP_DEFAULT_CHUNK;
3058   } else {
3059     r_sched.chunk = __kmp_chunk;
3060   }
3061 
3062   return r_sched;
3063 }
3064 
3065 /* Allocate (realloc == FALSE) * or reallocate (realloc == TRUE)
3066    at least argc number of *t_argv entries for the requested team. */
3067 static void __kmp_alloc_argv_entries(int argc, kmp_team_t *team, int realloc) {
3068 
3069   KMP_DEBUG_ASSERT(team);
3070   if (!realloc || argc > team->t.t_max_argc) {
3071 
3072     KA_TRACE(100, ("__kmp_alloc_argv_entries: team %d: needed entries=%d, "
3073                    "current entries=%d\n",
3074                    team->t.t_id, argc, (realloc) ? team->t.t_max_argc : 0));
3075     /* if previously allocated heap space for args, free them */
3076     if (realloc && team->t.t_argv != &team->t.t_inline_argv[0])
3077       __kmp_free((void *)team->t.t_argv);
3078 
3079     if (argc <= KMP_INLINE_ARGV_ENTRIES) {
3080       /* use unused space in the cache line for arguments */
3081       team->t.t_max_argc = KMP_INLINE_ARGV_ENTRIES;
3082       KA_TRACE(100, ("__kmp_alloc_argv_entries: team %d: inline allocate %d "
3083                      "argv entries\n",
3084                      team->t.t_id, team->t.t_max_argc));
3085       team->t.t_argv = &team->t.t_inline_argv[0];
3086       if (__kmp_storage_map) {
3087         __kmp_print_storage_map_gtid(
3088             -1, &team->t.t_inline_argv[0],
3089             &team->t.t_inline_argv[KMP_INLINE_ARGV_ENTRIES],
3090             (sizeof(void *) * KMP_INLINE_ARGV_ENTRIES), "team_%d.t_inline_argv",
3091             team->t.t_id);
3092       }
3093     } else {
3094       /* allocate space for arguments in the heap */
3095       team->t.t_max_argc = (argc <= (KMP_MIN_MALLOC_ARGV_ENTRIES >> 1))
3096                                ? KMP_MIN_MALLOC_ARGV_ENTRIES
3097                                : 2 * argc;
3098       KA_TRACE(100, ("__kmp_alloc_argv_entries: team %d: dynamic allocate %d "
3099                      "argv entries\n",
3100                      team->t.t_id, team->t.t_max_argc));
3101       team->t.t_argv =
3102           (void **)__kmp_page_allocate(sizeof(void *) * team->t.t_max_argc);
3103       if (__kmp_storage_map) {
3104         __kmp_print_storage_map_gtid(-1, &team->t.t_argv[0],
3105                                      &team->t.t_argv[team->t.t_max_argc],
3106                                      sizeof(void *) * team->t.t_max_argc,
3107                                      "team_%d.t_argv", team->t.t_id);
3108       }
3109     }
3110   }
3111 }
3112 
3113 static void __kmp_allocate_team_arrays(kmp_team_t *team, int max_nth) {
3114   int i;
3115   int num_disp_buff = max_nth > 1 ? __kmp_dispatch_num_buffers : 2;
3116   team->t.t_threads =
3117       (kmp_info_t **)__kmp_allocate(sizeof(kmp_info_t *) * max_nth);
3118   team->t.t_disp_buffer = (dispatch_shared_info_t *)__kmp_allocate(
3119       sizeof(dispatch_shared_info_t) * num_disp_buff);
3120   team->t.t_dispatch =
3121       (kmp_disp_t *)__kmp_allocate(sizeof(kmp_disp_t) * max_nth);
3122   team->t.t_implicit_task_taskdata =
3123       (kmp_taskdata_t *)__kmp_allocate(sizeof(kmp_taskdata_t) * max_nth);
3124   team->t.t_max_nproc = max_nth;
3125 
3126   /* setup dispatch buffers */
3127   for (i = 0; i < num_disp_buff; ++i) {
3128     team->t.t_disp_buffer[i].buffer_index = i;
3129 #if OMP_45_ENABLED
3130     team->t.t_disp_buffer[i].doacross_buf_idx = i;
3131 #endif
3132   }
3133 }
3134 
3135 static void __kmp_free_team_arrays(kmp_team_t *team) {
3136   /* Note: this does not free the threads in t_threads (__kmp_free_threads) */
3137   int i;
3138   for (i = 0; i < team->t.t_max_nproc; ++i) {
3139     if (team->t.t_dispatch[i].th_disp_buffer != NULL) {
3140       __kmp_free(team->t.t_dispatch[i].th_disp_buffer);
3141       team->t.t_dispatch[i].th_disp_buffer = NULL;
3142     }
3143   }
3144 #if KMP_USE_HIER_SCHED
3145   __kmp_dispatch_free_hierarchies(team);
3146 #endif
3147   __kmp_free(team->t.t_threads);
3148   __kmp_free(team->t.t_disp_buffer);
3149   __kmp_free(team->t.t_dispatch);
3150   __kmp_free(team->t.t_implicit_task_taskdata);
3151   team->t.t_threads = NULL;
3152   team->t.t_disp_buffer = NULL;
3153   team->t.t_dispatch = NULL;
3154   team->t.t_implicit_task_taskdata = 0;
3155 }
3156 
3157 static void __kmp_reallocate_team_arrays(kmp_team_t *team, int max_nth) {
3158   kmp_info_t **oldThreads = team->t.t_threads;
3159 
3160   __kmp_free(team->t.t_disp_buffer);
3161   __kmp_free(team->t.t_dispatch);
3162   __kmp_free(team->t.t_implicit_task_taskdata);
3163   __kmp_allocate_team_arrays(team, max_nth);
3164 
3165   KMP_MEMCPY(team->t.t_threads, oldThreads,
3166              team->t.t_nproc * sizeof(kmp_info_t *));
3167 
3168   __kmp_free(oldThreads);
3169 }
3170 
3171 static kmp_internal_control_t __kmp_get_global_icvs(void) {
3172 
3173   kmp_r_sched_t r_sched =
3174       __kmp_get_schedule_global(); // get current state of scheduling globals
3175 
3176 #if OMP_40_ENABLED
3177   KMP_DEBUG_ASSERT(__kmp_nested_proc_bind.used > 0);
3178 #endif /* OMP_40_ENABLED */
3179 
3180   kmp_internal_control_t g_icvs = {
3181     0, // int serial_nesting_level; //corresponds to value of th_team_serialized
3182     (kmp_int8)__kmp_global.g.g_dynamic, // internal control for dynamic
3183     // adjustment of threads (per thread)
3184     (kmp_int8)__kmp_env_blocktime, // int bt_set; //internal control for
3185     // whether blocktime is explicitly set
3186     __kmp_dflt_blocktime, // int blocktime; //internal control for blocktime
3187 #if KMP_USE_MONITOR
3188     __kmp_bt_intervals, // int bt_intervals; //internal control for blocktime
3189 // intervals
3190 #endif
3191     __kmp_dflt_team_nth, // int nproc; //internal control for # of threads for
3192     // next parallel region (per thread)
3193     // (use a max ub on value if __kmp_parallel_initialize not called yet)
3194     __kmp_cg_max_nth, // int thread_limit;
3195     __kmp_dflt_max_active_levels, // int max_active_levels; //internal control
3196     // for max_active_levels
3197     r_sched, // kmp_r_sched_t sched; //internal control for runtime schedule
3198 // {sched,chunk} pair
3199 #if OMP_40_ENABLED
3200     __kmp_nested_proc_bind.bind_types[0],
3201     __kmp_default_device,
3202 #endif /* OMP_40_ENABLED */
3203     NULL // struct kmp_internal_control *next;
3204   };
3205 
3206   return g_icvs;
3207 }
3208 
3209 static kmp_internal_control_t __kmp_get_x_global_icvs(const kmp_team_t *team) {
3210 
3211   kmp_internal_control_t gx_icvs;
3212   gx_icvs.serial_nesting_level =
3213       0; // probably =team->t.t_serial like in save_inter_controls
3214   copy_icvs(&gx_icvs, &team->t.t_threads[0]->th.th_current_task->td_icvs);
3215   gx_icvs.next = NULL;
3216 
3217   return gx_icvs;
3218 }
3219 
3220 static void __kmp_initialize_root(kmp_root_t *root) {
3221   int f;
3222   kmp_team_t *root_team;
3223   kmp_team_t *hot_team;
3224   int hot_team_max_nth;
3225   kmp_r_sched_t r_sched =
3226       __kmp_get_schedule_global(); // get current state of scheduling globals
3227   kmp_internal_control_t r_icvs = __kmp_get_global_icvs();
3228   KMP_DEBUG_ASSERT(root);
3229   KMP_ASSERT(!root->r.r_begin);
3230 
3231   /* setup the root state structure */
3232   __kmp_init_lock(&root->r.r_begin_lock);
3233   root->r.r_begin = FALSE;
3234   root->r.r_active = FALSE;
3235   root->r.r_in_parallel = 0;
3236   root->r.r_blocktime = __kmp_dflt_blocktime;
3237 
3238   /* setup the root team for this task */
3239   /* allocate the root team structure */
3240   KF_TRACE(10, ("__kmp_initialize_root: before root_team\n"));
3241 
3242   root_team =
3243       __kmp_allocate_team(root,
3244                           1, // new_nproc
3245                           1, // max_nproc
3246 #if OMPT_SUPPORT
3247                           ompt_data_none, // root parallel id
3248 #endif
3249 #if OMP_40_ENABLED
3250                           __kmp_nested_proc_bind.bind_types[0],
3251 #endif
3252                           &r_icvs,
3253                           0 // argc
3254                           USE_NESTED_HOT_ARG(NULL) // master thread is unknown
3255                           );
3256 #if USE_DEBUGGER
3257   // Non-NULL value should be assigned to make the debugger display the root
3258   // team.
3259   TCW_SYNC_PTR(root_team->t.t_pkfn, (microtask_t)(~0));
3260 #endif
3261 
3262   KF_TRACE(10, ("__kmp_initialize_root: after root_team = %p\n", root_team));
3263 
3264   root->r.r_root_team = root_team;
3265   root_team->t.t_control_stack_top = NULL;
3266 
3267   /* initialize root team */
3268   root_team->t.t_threads[0] = NULL;
3269   root_team->t.t_nproc = 1;
3270   root_team->t.t_serialized = 1;
3271   // TODO???: root_team->t.t_max_active_levels = __kmp_dflt_max_active_levels;
3272   root_team->t.t_sched.sched = r_sched.sched;
3273   KA_TRACE(
3274       20,
3275       ("__kmp_initialize_root: init root team %d arrived: join=%u, plain=%u\n",
3276        root_team->t.t_id, KMP_INIT_BARRIER_STATE, KMP_INIT_BARRIER_STATE));
3277 
3278   /* setup the  hot team for this task */
3279   /* allocate the hot team structure */
3280   KF_TRACE(10, ("__kmp_initialize_root: before hot_team\n"));
3281 
3282   hot_team =
3283       __kmp_allocate_team(root,
3284                           1, // new_nproc
3285                           __kmp_dflt_team_nth_ub * 2, // max_nproc
3286 #if OMPT_SUPPORT
3287                           ompt_data_none, // root parallel id
3288 #endif
3289 #if OMP_40_ENABLED
3290                           __kmp_nested_proc_bind.bind_types[0],
3291 #endif
3292                           &r_icvs,
3293                           0 // argc
3294                           USE_NESTED_HOT_ARG(NULL) // master thread is unknown
3295                           );
3296   KF_TRACE(10, ("__kmp_initialize_root: after hot_team = %p\n", hot_team));
3297 
3298   root->r.r_hot_team = hot_team;
3299   root_team->t.t_control_stack_top = NULL;
3300 
3301   /* first-time initialization */
3302   hot_team->t.t_parent = root_team;
3303 
3304   /* initialize hot team */
3305   hot_team_max_nth = hot_team->t.t_max_nproc;
3306   for (f = 0; f < hot_team_max_nth; ++f) {
3307     hot_team->t.t_threads[f] = NULL;
3308   }
3309   hot_team->t.t_nproc = 1;
3310   // TODO???: hot_team->t.t_max_active_levels = __kmp_dflt_max_active_levels;
3311   hot_team->t.t_sched.sched = r_sched.sched;
3312   hot_team->t.t_size_changed = 0;
3313 }
3314 
3315 #ifdef KMP_DEBUG
3316 
3317 typedef struct kmp_team_list_item {
3318   kmp_team_p const *entry;
3319   struct kmp_team_list_item *next;
3320 } kmp_team_list_item_t;
3321 typedef kmp_team_list_item_t *kmp_team_list_t;
3322 
3323 static void __kmp_print_structure_team_accum( // Add team to list of teams.
3324     kmp_team_list_t list, // List of teams.
3325     kmp_team_p const *team // Team to add.
3326     ) {
3327 
3328   // List must terminate with item where both entry and next are NULL.
3329   // Team is added to the list only once.
3330   // List is sorted in ascending order by team id.
3331   // Team id is *not* a key.
3332 
3333   kmp_team_list_t l;
3334 
3335   KMP_DEBUG_ASSERT(list != NULL);
3336   if (team == NULL) {
3337     return;
3338   }
3339 
3340   __kmp_print_structure_team_accum(list, team->t.t_parent);
3341   __kmp_print_structure_team_accum(list, team->t.t_next_pool);
3342 
3343   // Search list for the team.
3344   l = list;
3345   while (l->next != NULL && l->entry != team) {
3346     l = l->next;
3347   }
3348   if (l->next != NULL) {
3349     return; // Team has been added before, exit.
3350   }
3351 
3352   // Team is not found. Search list again for insertion point.
3353   l = list;
3354   while (l->next != NULL && l->entry->t.t_id <= team->t.t_id) {
3355     l = l->next;
3356   }
3357 
3358   // Insert team.
3359   {
3360     kmp_team_list_item_t *item = (kmp_team_list_item_t *)KMP_INTERNAL_MALLOC(
3361         sizeof(kmp_team_list_item_t));
3362     *item = *l;
3363     l->entry = team;
3364     l->next = item;
3365   }
3366 }
3367 
3368 static void __kmp_print_structure_team(char const *title, kmp_team_p const *team
3369 
3370                                        ) {
3371   __kmp_printf("%s", title);
3372   if (team != NULL) {
3373     __kmp_printf("%2x %p\n", team->t.t_id, team);
3374   } else {
3375     __kmp_printf(" - (nil)\n");
3376   }
3377 }
3378 
3379 static void __kmp_print_structure_thread(char const *title,
3380                                          kmp_info_p const *thread) {
3381   __kmp_printf("%s", title);
3382   if (thread != NULL) {
3383     __kmp_printf("%2d %p\n", thread->th.th_info.ds.ds_gtid, thread);
3384   } else {
3385     __kmp_printf(" - (nil)\n");
3386   }
3387 }
3388 
3389 void __kmp_print_structure(void) {
3390 
3391   kmp_team_list_t list;
3392 
3393   // Initialize list of teams.
3394   list =
3395       (kmp_team_list_item_t *)KMP_INTERNAL_MALLOC(sizeof(kmp_team_list_item_t));
3396   list->entry = NULL;
3397   list->next = NULL;
3398 
3399   __kmp_printf("\n------------------------------\nGlobal Thread "
3400                "Table\n------------------------------\n");
3401   {
3402     int gtid;
3403     for (gtid = 0; gtid < __kmp_threads_capacity; ++gtid) {
3404       __kmp_printf("%2d", gtid);
3405       if (__kmp_threads != NULL) {
3406         __kmp_printf(" %p", __kmp_threads[gtid]);
3407       }
3408       if (__kmp_root != NULL) {
3409         __kmp_printf(" %p", __kmp_root[gtid]);
3410       }
3411       __kmp_printf("\n");
3412     }
3413   }
3414 
3415   // Print out __kmp_threads array.
3416   __kmp_printf("\n------------------------------\nThreads\n--------------------"
3417                "----------\n");
3418   if (__kmp_threads != NULL) {
3419     int gtid;
3420     for (gtid = 0; gtid < __kmp_threads_capacity; ++gtid) {
3421       kmp_info_t const *thread = __kmp_threads[gtid];
3422       if (thread != NULL) {
3423         __kmp_printf("GTID %2d %p:\n", gtid, thread);
3424         __kmp_printf("    Our Root:        %p\n", thread->th.th_root);
3425         __kmp_print_structure_team("    Our Team:     ", thread->th.th_team);
3426         __kmp_print_structure_team("    Serial Team:  ",
3427                                    thread->th.th_serial_team);
3428         __kmp_printf("    Threads:      %2d\n", thread->th.th_team_nproc);
3429         __kmp_print_structure_thread("    Master:       ",
3430                                      thread->th.th_team_master);
3431         __kmp_printf("    Serialized?:  %2d\n", thread->th.th_team_serialized);
3432         __kmp_printf("    Set NProc:    %2d\n", thread->th.th_set_nproc);
3433 #if OMP_40_ENABLED
3434         __kmp_printf("    Set Proc Bind: %2d\n", thread->th.th_set_proc_bind);
3435 #endif
3436         __kmp_print_structure_thread("    Next in pool: ",
3437                                      thread->th.th_next_pool);
3438         __kmp_printf("\n");
3439         __kmp_print_structure_team_accum(list, thread->th.th_team);
3440         __kmp_print_structure_team_accum(list, thread->th.th_serial_team);
3441       }
3442     }
3443   } else {
3444     __kmp_printf("Threads array is not allocated.\n");
3445   }
3446 
3447   // Print out __kmp_root array.
3448   __kmp_printf("\n------------------------------\nUbers\n----------------------"
3449                "--------\n");
3450   if (__kmp_root != NULL) {
3451     int gtid;
3452     for (gtid = 0; gtid < __kmp_threads_capacity; ++gtid) {
3453       kmp_root_t const *root = __kmp_root[gtid];
3454       if (root != NULL) {
3455         __kmp_printf("GTID %2d %p:\n", gtid, root);
3456         __kmp_print_structure_team("    Root Team:    ", root->r.r_root_team);
3457         __kmp_print_structure_team("    Hot Team:     ", root->r.r_hot_team);
3458         __kmp_print_structure_thread("    Uber Thread:  ",
3459                                      root->r.r_uber_thread);
3460         __kmp_printf("    Active?:      %2d\n", root->r.r_active);
3461         __kmp_printf("    In Parallel:  %2d\n",
3462                      KMP_ATOMIC_LD_RLX(&root->r.r_in_parallel));
3463         __kmp_printf("\n");
3464         __kmp_print_structure_team_accum(list, root->r.r_root_team);
3465         __kmp_print_structure_team_accum(list, root->r.r_hot_team);
3466       }
3467     }
3468   } else {
3469     __kmp_printf("Ubers array is not allocated.\n");
3470   }
3471 
3472   __kmp_printf("\n------------------------------\nTeams\n----------------------"
3473                "--------\n");
3474   while (list->next != NULL) {
3475     kmp_team_p const *team = list->entry;
3476     int i;
3477     __kmp_printf("Team %2x %p:\n", team->t.t_id, team);
3478     __kmp_print_structure_team("    Parent Team:      ", team->t.t_parent);
3479     __kmp_printf("    Master TID:       %2d\n", team->t.t_master_tid);
3480     __kmp_printf("    Max threads:      %2d\n", team->t.t_max_nproc);
3481     __kmp_printf("    Levels of serial: %2d\n", team->t.t_serialized);
3482     __kmp_printf("    Number threads:   %2d\n", team->t.t_nproc);
3483     for (i = 0; i < team->t.t_nproc; ++i) {
3484       __kmp_printf("    Thread %2d:      ", i);
3485       __kmp_print_structure_thread("", team->t.t_threads[i]);
3486     }
3487     __kmp_print_structure_team("    Next in pool:     ", team->t.t_next_pool);
3488     __kmp_printf("\n");
3489     list = list->next;
3490   }
3491 
3492   // Print out __kmp_thread_pool and __kmp_team_pool.
3493   __kmp_printf("\n------------------------------\nPools\n----------------------"
3494                "--------\n");
3495   __kmp_print_structure_thread("Thread pool:          ",
3496                                CCAST(kmp_info_t *, __kmp_thread_pool));
3497   __kmp_print_structure_team("Team pool:            ",
3498                              CCAST(kmp_team_t *, __kmp_team_pool));
3499   __kmp_printf("\n");
3500 
3501   // Free team list.
3502   while (list != NULL) {
3503     kmp_team_list_item_t *item = list;
3504     list = list->next;
3505     KMP_INTERNAL_FREE(item);
3506   }
3507 }
3508 
3509 #endif
3510 
3511 //---------------------------------------------------------------------------
3512 //  Stuff for per-thread fast random number generator
3513 //  Table of primes
3514 static const unsigned __kmp_primes[] = {
3515     0x9e3779b1, 0xffe6cc59, 0x2109f6dd, 0x43977ab5, 0xba5703f5, 0xb495a877,
3516     0xe1626741, 0x79695e6b, 0xbc98c09f, 0xd5bee2b3, 0x287488f9, 0x3af18231,
3517     0x9677cd4d, 0xbe3a6929, 0xadc6a877, 0xdcf0674b, 0xbe4d6fe9, 0x5f15e201,
3518     0x99afc3fd, 0xf3f16801, 0xe222cfff, 0x24ba5fdb, 0x0620452d, 0x79f149e3,
3519     0xc8b93f49, 0x972702cd, 0xb07dd827, 0x6c97d5ed, 0x085a3d61, 0x46eb5ea7,
3520     0x3d9910ed, 0x2e687b5b, 0x29609227, 0x6eb081f1, 0x0954c4e1, 0x9d114db9,
3521     0x542acfa9, 0xb3e6bd7b, 0x0742d917, 0xe9f3ffa7, 0x54581edb, 0xf2480f45,
3522     0x0bb9288f, 0xef1affc7, 0x85fa0ca7, 0x3ccc14db, 0xe6baf34b, 0x343377f7,
3523     0x5ca19031, 0xe6d9293b, 0xf0a9f391, 0x5d2e980b, 0xfc411073, 0xc3749363,
3524     0xb892d829, 0x3549366b, 0x629750ad, 0xb98294e5, 0x892d9483, 0xc235baf3,
3525     0x3d2402a3, 0x6bdef3c9, 0xbec333cd, 0x40c9520f};
3526 
3527 //---------------------------------------------------------------------------
3528 //  __kmp_get_random: Get a random number using a linear congruential method.
3529 unsigned short __kmp_get_random(kmp_info_t *thread) {
3530   unsigned x = thread->th.th_x;
3531   unsigned short r = x >> 16;
3532 
3533   thread->th.th_x = x * thread->th.th_a + 1;
3534 
3535   KA_TRACE(30, ("__kmp_get_random: THREAD: %d, RETURN: %u\n",
3536                 thread->th.th_info.ds.ds_tid, r));
3537 
3538   return r;
3539 }
3540 //--------------------------------------------------------
3541 // __kmp_init_random: Initialize a random number generator
3542 void __kmp_init_random(kmp_info_t *thread) {
3543   unsigned seed = thread->th.th_info.ds.ds_tid;
3544 
3545   thread->th.th_a =
3546       __kmp_primes[seed % (sizeof(__kmp_primes) / sizeof(__kmp_primes[0]))];
3547   thread->th.th_x = (seed + 1) * thread->th.th_a + 1;
3548   KA_TRACE(30,
3549            ("__kmp_init_random: THREAD: %u; A: %u\n", seed, thread->th.th_a));
3550 }
3551 
3552 #if KMP_OS_WINDOWS
3553 /* reclaim array entries for root threads that are already dead, returns number
3554  * reclaimed */
3555 static int __kmp_reclaim_dead_roots(void) {
3556   int i, r = 0;
3557 
3558   for (i = 0; i < __kmp_threads_capacity; ++i) {
3559     if (KMP_UBER_GTID(i) &&
3560         !__kmp_still_running((kmp_info_t *)TCR_SYNC_PTR(__kmp_threads[i])) &&
3561         !__kmp_root[i]
3562              ->r.r_active) { // AC: reclaim only roots died in non-active state
3563       r += __kmp_unregister_root_other_thread(i);
3564     }
3565   }
3566   return r;
3567 }
3568 #endif
3569 
3570 /* This function attempts to create free entries in __kmp_threads and
3571    __kmp_root, and returns the number of free entries generated.
3572 
3573    For Windows* OS static library, the first mechanism used is to reclaim array
3574    entries for root threads that are already dead.
3575 
3576    On all platforms, expansion is attempted on the arrays __kmp_threads_ and
3577    __kmp_root, with appropriate update to __kmp_threads_capacity. Array
3578    capacity is increased by doubling with clipping to __kmp_tp_capacity, if
3579    threadprivate cache array has been created. Synchronization with
3580    __kmpc_threadprivate_cached is done using __kmp_tp_cached_lock.
3581 
3582    After any dead root reclamation, if the clipping value allows array expansion
3583    to result in the generation of a total of nNeed free slots, the function does
3584    that expansion. If not, nothing is done beyond the possible initial root
3585    thread reclamation.
3586 
3587    If any argument is negative, the behavior is undefined. */
3588 static int __kmp_expand_threads(int nNeed) {
3589   int added = 0;
3590   int minimumRequiredCapacity;
3591   int newCapacity;
3592   kmp_info_t **newThreads;
3593   kmp_root_t **newRoot;
3594 
3595 // All calls to __kmp_expand_threads should be under __kmp_forkjoin_lock, so
3596 // resizing __kmp_threads does not need additional protection if foreign
3597 // threads are present
3598 
3599 #if KMP_OS_WINDOWS && !KMP_DYNAMIC_LIB
3600   /* only for Windows static library */
3601   /* reclaim array entries for root threads that are already dead */
3602   added = __kmp_reclaim_dead_roots();
3603 
3604   if (nNeed) {
3605     nNeed -= added;
3606     if (nNeed < 0)
3607       nNeed = 0;
3608   }
3609 #endif
3610   if (nNeed <= 0)
3611     return added;
3612 
3613   // Note that __kmp_threads_capacity is not bounded by __kmp_max_nth. If
3614   // __kmp_max_nth is set to some value less than __kmp_sys_max_nth by the
3615   // user via KMP_DEVICE_THREAD_LIMIT, then __kmp_threads_capacity may become
3616   // > __kmp_max_nth in one of two ways:
3617   //
3618   // 1) The initialization thread (gtid = 0) exits.  __kmp_threads[0]
3619   //    may not be resused by another thread, so we may need to increase
3620   //    __kmp_threads_capacity to __kmp_max_nth + 1.
3621   //
3622   // 2) New foreign root(s) are encountered.  We always register new foreign
3623   //    roots. This may cause a smaller # of threads to be allocated at
3624   //    subsequent parallel regions, but the worker threads hang around (and
3625   //    eventually go to sleep) and need slots in the __kmp_threads[] array.
3626   //
3627   // Anyway, that is the reason for moving the check to see if
3628   // __kmp_max_nth was exceeded into __kmp_reserve_threads()
3629   // instead of having it performed here. -BB
3630 
3631   KMP_DEBUG_ASSERT(__kmp_sys_max_nth >= __kmp_threads_capacity);
3632 
3633   /* compute expansion headroom to check if we can expand */
3634   if (__kmp_sys_max_nth - __kmp_threads_capacity < nNeed) {
3635     /* possible expansion too small -- give up */
3636     return added;
3637   }
3638   minimumRequiredCapacity = __kmp_threads_capacity + nNeed;
3639 
3640   newCapacity = __kmp_threads_capacity;
3641   do {
3642     newCapacity = newCapacity <= (__kmp_sys_max_nth >> 1) ? (newCapacity << 1)
3643                                                           : __kmp_sys_max_nth;
3644   } while (newCapacity < minimumRequiredCapacity);
3645   newThreads = (kmp_info_t **)__kmp_allocate(
3646       (sizeof(kmp_info_t *) + sizeof(kmp_root_t *)) * newCapacity + CACHE_LINE);
3647   newRoot =
3648       (kmp_root_t **)((char *)newThreads + sizeof(kmp_info_t *) * newCapacity);
3649   KMP_MEMCPY(newThreads, __kmp_threads,
3650              __kmp_threads_capacity * sizeof(kmp_info_t *));
3651   KMP_MEMCPY(newRoot, __kmp_root,
3652              __kmp_threads_capacity * sizeof(kmp_root_t *));
3653 
3654   kmp_info_t **temp_threads = __kmp_threads;
3655   *(kmp_info_t * *volatile *)&__kmp_threads = newThreads;
3656   *(kmp_root_t * *volatile *)&__kmp_root = newRoot;
3657   __kmp_free(temp_threads);
3658   added += newCapacity - __kmp_threads_capacity;
3659   *(volatile int *)&__kmp_threads_capacity = newCapacity;
3660 
3661   if (newCapacity > __kmp_tp_capacity) {
3662     __kmp_acquire_bootstrap_lock(&__kmp_tp_cached_lock);
3663     if (__kmp_tp_cached && newCapacity > __kmp_tp_capacity) {
3664       __kmp_threadprivate_resize_cache(newCapacity);
3665     } else { // increase __kmp_tp_capacity to correspond with kmp_threads size
3666       *(volatile int *)&__kmp_tp_capacity = newCapacity;
3667     }
3668     __kmp_release_bootstrap_lock(&__kmp_tp_cached_lock);
3669   }
3670 
3671   return added;
3672 }
3673 
3674 /* Register the current thread as a root thread and obtain our gtid. We must
3675    have the __kmp_initz_lock held at this point. Argument TRUE only if are the
3676    thread that calls from __kmp_do_serial_initialize() */
3677 int __kmp_register_root(int initial_thread) {
3678   kmp_info_t *root_thread;
3679   kmp_root_t *root;
3680   int gtid;
3681   int capacity;
3682   __kmp_acquire_bootstrap_lock(&__kmp_forkjoin_lock);
3683   KA_TRACE(20, ("__kmp_register_root: entered\n"));
3684   KMP_MB();
3685 
3686   /* 2007-03-02:
3687      If initial thread did not invoke OpenMP RTL yet, and this thread is not an
3688      initial one, "__kmp_all_nth >= __kmp_threads_capacity" condition does not
3689      work as expected -- it may return false (that means there is at least one
3690      empty slot in __kmp_threads array), but it is possible the only free slot
3691      is #0, which is reserved for initial thread and so cannot be used for this
3692      one. Following code workarounds this bug.
3693 
3694      However, right solution seems to be not reserving slot #0 for initial
3695      thread because:
3696      (1) there is no magic in slot #0,
3697      (2) we cannot detect initial thread reliably (the first thread which does
3698         serial initialization may be not a real initial thread).
3699   */
3700   capacity = __kmp_threads_capacity;
3701   if (!initial_thread && TCR_PTR(__kmp_threads[0]) == NULL) {
3702     --capacity;
3703   }
3704 
3705   /* see if there are too many threads */
3706   if (__kmp_all_nth >= capacity && !__kmp_expand_threads(1)) {
3707     if (__kmp_tp_cached) {
3708       __kmp_fatal(KMP_MSG(CantRegisterNewThread),
3709                   KMP_HNT(Set_ALL_THREADPRIVATE, __kmp_tp_capacity),
3710                   KMP_HNT(PossibleSystemLimitOnThreads), __kmp_msg_null);
3711     } else {
3712       __kmp_fatal(KMP_MSG(CantRegisterNewThread), KMP_HNT(SystemLimitOnThreads),
3713                   __kmp_msg_null);
3714     }
3715   }
3716 
3717   /* find an available thread slot */
3718   /* Don't reassign the zero slot since we need that to only be used by initial
3719      thread */
3720   for (gtid = (initial_thread ? 0 : 1); TCR_PTR(__kmp_threads[gtid]) != NULL;
3721        gtid++)
3722     ;
3723   KA_TRACE(1,
3724            ("__kmp_register_root: found slot in threads array: T#%d\n", gtid));
3725   KMP_ASSERT(gtid < __kmp_threads_capacity);
3726 
3727   /* update global accounting */
3728   __kmp_all_nth++;
3729   TCW_4(__kmp_nth, __kmp_nth + 1);
3730 
3731   // if __kmp_adjust_gtid_mode is set, then we use method #1 (sp search) for low
3732   // numbers of procs, and method #2 (keyed API call) for higher numbers.
3733   if (__kmp_adjust_gtid_mode) {
3734     if (__kmp_all_nth >= __kmp_tls_gtid_min) {
3735       if (TCR_4(__kmp_gtid_mode) != 2) {
3736         TCW_4(__kmp_gtid_mode, 2);
3737       }
3738     } else {
3739       if (TCR_4(__kmp_gtid_mode) != 1) {
3740         TCW_4(__kmp_gtid_mode, 1);
3741       }
3742     }
3743   }
3744 
3745 #ifdef KMP_ADJUST_BLOCKTIME
3746   /* Adjust blocktime to zero if necessary            */
3747   /* Middle initialization might not have occurred yet */
3748   if (!__kmp_env_blocktime && (__kmp_avail_proc > 0)) {
3749     if (__kmp_nth > __kmp_avail_proc) {
3750       __kmp_zero_bt = TRUE;
3751     }
3752   }
3753 #endif /* KMP_ADJUST_BLOCKTIME */
3754 
3755   /* setup this new hierarchy */
3756   if (!(root = __kmp_root[gtid])) {
3757     root = __kmp_root[gtid] = (kmp_root_t *)__kmp_allocate(sizeof(kmp_root_t));
3758     KMP_DEBUG_ASSERT(!root->r.r_root_team);
3759   }
3760 
3761 #if KMP_STATS_ENABLED
3762   // Initialize stats as soon as possible (right after gtid assignment).
3763   __kmp_stats_thread_ptr = __kmp_stats_list->push_back(gtid);
3764   __kmp_stats_thread_ptr->startLife();
3765   KMP_SET_THREAD_STATE(SERIAL_REGION);
3766   KMP_INIT_PARTITIONED_TIMERS(OMP_serial);
3767 #endif
3768   __kmp_initialize_root(root);
3769 
3770   /* setup new root thread structure */
3771   if (root->r.r_uber_thread) {
3772     root_thread = root->r.r_uber_thread;
3773   } else {
3774     root_thread = (kmp_info_t *)__kmp_allocate(sizeof(kmp_info_t));
3775     if (__kmp_storage_map) {
3776       __kmp_print_thread_storage_map(root_thread, gtid);
3777     }
3778     root_thread->th.th_info.ds.ds_gtid = gtid;
3779 #if OMPT_SUPPORT
3780     root_thread->th.ompt_thread_info.thread_data = ompt_data_none;
3781 #endif
3782     root_thread->th.th_root = root;
3783     if (__kmp_env_consistency_check) {
3784       root_thread->th.th_cons = __kmp_allocate_cons_stack(gtid);
3785     }
3786 #if USE_FAST_MEMORY
3787     __kmp_initialize_fast_memory(root_thread);
3788 #endif /* USE_FAST_MEMORY */
3789 
3790 #if KMP_USE_BGET
3791     KMP_DEBUG_ASSERT(root_thread->th.th_local.bget_data == NULL);
3792     __kmp_initialize_bget(root_thread);
3793 #endif
3794     __kmp_init_random(root_thread); // Initialize random number generator
3795   }
3796 
3797   /* setup the serial team held in reserve by the root thread */
3798   if (!root_thread->th.th_serial_team) {
3799     kmp_internal_control_t r_icvs = __kmp_get_global_icvs();
3800     KF_TRACE(10, ("__kmp_register_root: before serial_team\n"));
3801     root_thread->th.th_serial_team =
3802         __kmp_allocate_team(root, 1, 1,
3803 #if OMPT_SUPPORT
3804                             ompt_data_none, // root parallel id
3805 #endif
3806 #if OMP_40_ENABLED
3807                             proc_bind_default,
3808 #endif
3809                             &r_icvs, 0 USE_NESTED_HOT_ARG(NULL));
3810   }
3811   KMP_ASSERT(root_thread->th.th_serial_team);
3812   KF_TRACE(10, ("__kmp_register_root: after serial_team = %p\n",
3813                 root_thread->th.th_serial_team));
3814 
3815   /* drop root_thread into place */
3816   TCW_SYNC_PTR(__kmp_threads[gtid], root_thread);
3817 
3818   root->r.r_root_team->t.t_threads[0] = root_thread;
3819   root->r.r_hot_team->t.t_threads[0] = root_thread;
3820   root_thread->th.th_serial_team->t.t_threads[0] = root_thread;
3821   // AC: the team created in reserve, not for execution (it is unused for now).
3822   root_thread->th.th_serial_team->t.t_serialized = 0;
3823   root->r.r_uber_thread = root_thread;
3824 
3825   /* initialize the thread, get it ready to go */
3826   __kmp_initialize_info(root_thread, root->r.r_root_team, 0, gtid);
3827   TCW_4(__kmp_init_gtid, TRUE);
3828 
3829   /* prepare the master thread for get_gtid() */
3830   __kmp_gtid_set_specific(gtid);
3831 
3832 #if USE_ITT_BUILD
3833   __kmp_itt_thread_name(gtid);
3834 #endif /* USE_ITT_BUILD */
3835 
3836 #ifdef KMP_TDATA_GTID
3837   __kmp_gtid = gtid;
3838 #endif
3839   __kmp_create_worker(gtid, root_thread, __kmp_stksize);
3840   KMP_DEBUG_ASSERT(__kmp_gtid_get_specific() == gtid);
3841 
3842   KA_TRACE(20, ("__kmp_register_root: T#%d init T#%d(%d:%d) arrived: join=%u, "
3843                 "plain=%u\n",
3844                 gtid, __kmp_gtid_from_tid(0, root->r.r_hot_team),
3845                 root->r.r_hot_team->t.t_id, 0, KMP_INIT_BARRIER_STATE,
3846                 KMP_INIT_BARRIER_STATE));
3847   { // Initialize barrier data.
3848     int b;
3849     for (b = 0; b < bs_last_barrier; ++b) {
3850       root_thread->th.th_bar[b].bb.b_arrived = KMP_INIT_BARRIER_STATE;
3851 #if USE_DEBUGGER
3852       root_thread->th.th_bar[b].bb.b_worker_arrived = 0;
3853 #endif
3854     }
3855   }
3856   KMP_DEBUG_ASSERT(root->r.r_hot_team->t.t_bar[bs_forkjoin_barrier].b_arrived ==
3857                    KMP_INIT_BARRIER_STATE);
3858 
3859 #if KMP_AFFINITY_SUPPORTED
3860 #if OMP_40_ENABLED
3861   root_thread->th.th_current_place = KMP_PLACE_UNDEFINED;
3862   root_thread->th.th_new_place = KMP_PLACE_UNDEFINED;
3863   root_thread->th.th_first_place = KMP_PLACE_UNDEFINED;
3864   root_thread->th.th_last_place = KMP_PLACE_UNDEFINED;
3865 #endif
3866   if (TCR_4(__kmp_init_middle)) {
3867     __kmp_affinity_set_init_mask(gtid, TRUE);
3868   }
3869 #endif /* KMP_AFFINITY_SUPPORTED */
3870 #if OMP_50_ENABLED
3871   root_thread->th.th_def_allocator = __kmp_def_allocator;
3872   root_thread->th.th_prev_level = 0;
3873   root_thread->th.th_prev_num_threads = 1;
3874 #endif
3875 
3876   kmp_cg_root_t *tmp = (kmp_cg_root_t *)__kmp_allocate(sizeof(kmp_cg_root_t));
3877   tmp->cg_root = root_thread;
3878   tmp->cg_thread_limit = __kmp_cg_max_nth;
3879   tmp->cg_nthreads = 1;
3880   KA_TRACE(100, ("__kmp_register_root: Thread %p created node %p with"
3881                  " cg_nthreads init to 1\n",
3882                  root_thread, tmp));
3883   tmp->up = NULL;
3884   root_thread->th.th_cg_roots = tmp;
3885 
3886   __kmp_root_counter++;
3887 
3888 #if OMPT_SUPPORT
3889   if (!initial_thread && ompt_enabled.enabled) {
3890 
3891     kmp_info_t *root_thread = ompt_get_thread();
3892 
3893     ompt_set_thread_state(root_thread, ompt_state_overhead);
3894 
3895     if (ompt_enabled.ompt_callback_thread_begin) {
3896       ompt_callbacks.ompt_callback(ompt_callback_thread_begin)(
3897           ompt_thread_initial, __ompt_get_thread_data_internal());
3898     }
3899     ompt_data_t *task_data;
3900     __ompt_get_task_info_internal(0, NULL, &task_data, NULL, NULL, NULL);
3901     if (ompt_enabled.ompt_callback_task_create) {
3902       ompt_callbacks.ompt_callback(ompt_callback_task_create)(
3903           NULL, NULL, task_data, ompt_task_initial, 0, NULL);
3904       // initial task has nothing to return to
3905     }
3906 
3907     ompt_set_thread_state(root_thread, ompt_state_work_serial);
3908   }
3909 #endif
3910 
3911   KMP_MB();
3912   __kmp_release_bootstrap_lock(&__kmp_forkjoin_lock);
3913 
3914   return gtid;
3915 }
3916 
3917 #if KMP_NESTED_HOT_TEAMS
3918 static int __kmp_free_hot_teams(kmp_root_t *root, kmp_info_t *thr, int level,
3919                                 const int max_level) {
3920   int i, n, nth;
3921   kmp_hot_team_ptr_t *hot_teams = thr->th.th_hot_teams;
3922   if (!hot_teams || !hot_teams[level].hot_team) {
3923     return 0;
3924   }
3925   KMP_DEBUG_ASSERT(level < max_level);
3926   kmp_team_t *team = hot_teams[level].hot_team;
3927   nth = hot_teams[level].hot_team_nth;
3928   n = nth - 1; // master is not freed
3929   if (level < max_level - 1) {
3930     for (i = 0; i < nth; ++i) {
3931       kmp_info_t *th = team->t.t_threads[i];
3932       n += __kmp_free_hot_teams(root, th, level + 1, max_level);
3933       if (i > 0 && th->th.th_hot_teams) {
3934         __kmp_free(th->th.th_hot_teams);
3935         th->th.th_hot_teams = NULL;
3936       }
3937     }
3938   }
3939   __kmp_free_team(root, team, NULL);
3940   return n;
3941 }
3942 #endif
3943 
3944 // Resets a root thread and clear its root and hot teams.
3945 // Returns the number of __kmp_threads entries directly and indirectly freed.
3946 static int __kmp_reset_root(int gtid, kmp_root_t *root) {
3947   kmp_team_t *root_team = root->r.r_root_team;
3948   kmp_team_t *hot_team = root->r.r_hot_team;
3949   int n = hot_team->t.t_nproc;
3950   int i;
3951 
3952   KMP_DEBUG_ASSERT(!root->r.r_active);
3953 
3954   root->r.r_root_team = NULL;
3955   root->r.r_hot_team = NULL;
3956   // __kmp_free_team() does not free hot teams, so we have to clear r_hot_team
3957   // before call to __kmp_free_team().
3958   __kmp_free_team(root, root_team USE_NESTED_HOT_ARG(NULL));
3959 #if KMP_NESTED_HOT_TEAMS
3960   if (__kmp_hot_teams_max_level >
3961       0) { // need to free nested hot teams and their threads if any
3962     for (i = 0; i < hot_team->t.t_nproc; ++i) {
3963       kmp_info_t *th = hot_team->t.t_threads[i];
3964       if (__kmp_hot_teams_max_level > 1) {
3965         n += __kmp_free_hot_teams(root, th, 1, __kmp_hot_teams_max_level);
3966       }
3967       if (th->th.th_hot_teams) {
3968         __kmp_free(th->th.th_hot_teams);
3969         th->th.th_hot_teams = NULL;
3970       }
3971     }
3972   }
3973 #endif
3974   __kmp_free_team(root, hot_team USE_NESTED_HOT_ARG(NULL));
3975 
3976   // Before we can reap the thread, we need to make certain that all other
3977   // threads in the teams that had this root as ancestor have stopped trying to
3978   // steal tasks.
3979   if (__kmp_tasking_mode != tskm_immediate_exec) {
3980     __kmp_wait_to_unref_task_teams();
3981   }
3982 
3983 #if KMP_OS_WINDOWS
3984   /* Close Handle of root duplicated in __kmp_create_worker (tr #62919) */
3985   KA_TRACE(
3986       10, ("__kmp_reset_root: free handle, th = %p, handle = %" KMP_UINTPTR_SPEC
3987            "\n",
3988            (LPVOID) & (root->r.r_uber_thread->th),
3989            root->r.r_uber_thread->th.th_info.ds.ds_thread));
3990   __kmp_free_handle(root->r.r_uber_thread->th.th_info.ds.ds_thread);
3991 #endif /* KMP_OS_WINDOWS */
3992 
3993 #if OMPT_SUPPORT
3994   if (ompt_enabled.ompt_callback_thread_end) {
3995     ompt_callbacks.ompt_callback(ompt_callback_thread_end)(
3996         &(root->r.r_uber_thread->th.ompt_thread_info.thread_data));
3997   }
3998 #endif
3999 
4000   TCW_4(__kmp_nth,
4001         __kmp_nth - 1); // __kmp_reap_thread will decrement __kmp_all_nth.
4002   i = root->r.r_uber_thread->th.th_cg_roots->cg_nthreads--;
4003   KA_TRACE(100, ("__kmp_reset_root: Thread %p decrement cg_nthreads on node %p"
4004                  " to %d\n",
4005                  root->r.r_uber_thread, root->r.r_uber_thread->th.th_cg_roots,
4006                  root->r.r_uber_thread->th.th_cg_roots->cg_nthreads));
4007   if (i == 1) {
4008     // need to free contention group structure
4009     KMP_DEBUG_ASSERT(root->r.r_uber_thread ==
4010                      root->r.r_uber_thread->th.th_cg_roots->cg_root);
4011     KMP_DEBUG_ASSERT(root->r.r_uber_thread->th.th_cg_roots->up == NULL);
4012     __kmp_free(root->r.r_uber_thread->th.th_cg_roots);
4013     root->r.r_uber_thread->th.th_cg_roots = NULL;
4014   }
4015   __kmp_reap_thread(root->r.r_uber_thread, 1);
4016 
4017   // We canot put root thread to __kmp_thread_pool, so we have to reap it istead
4018   // of freeing.
4019   root->r.r_uber_thread = NULL;
4020   /* mark root as no longer in use */
4021   root->r.r_begin = FALSE;
4022 
4023   return n;
4024 }
4025 
4026 void __kmp_unregister_root_current_thread(int gtid) {
4027   KA_TRACE(1, ("__kmp_unregister_root_current_thread: enter T#%d\n", gtid));
4028   /* this lock should be ok, since unregister_root_current_thread is never
4029      called during an abort, only during a normal close. furthermore, if you
4030      have the forkjoin lock, you should never try to get the initz lock */
4031   __kmp_acquire_bootstrap_lock(&__kmp_forkjoin_lock);
4032   if (TCR_4(__kmp_global.g.g_done) || !__kmp_init_serial) {
4033     KC_TRACE(10, ("__kmp_unregister_root_current_thread: already finished, "
4034                   "exiting T#%d\n",
4035                   gtid));
4036     __kmp_release_bootstrap_lock(&__kmp_forkjoin_lock);
4037     return;
4038   }
4039   kmp_root_t *root = __kmp_root[gtid];
4040 
4041   KMP_DEBUG_ASSERT(__kmp_threads && __kmp_threads[gtid]);
4042   KMP_ASSERT(KMP_UBER_GTID(gtid));
4043   KMP_ASSERT(root == __kmp_threads[gtid]->th.th_root);
4044   KMP_ASSERT(root->r.r_active == FALSE);
4045 
4046   KMP_MB();
4047 
4048 #if OMP_45_ENABLED
4049   kmp_info_t *thread = __kmp_threads[gtid];
4050   kmp_team_t *team = thread->th.th_team;
4051   kmp_task_team_t *task_team = thread->th.th_task_team;
4052 
4053   // we need to wait for the proxy tasks before finishing the thread
4054   if (task_team != NULL && task_team->tt.tt_found_proxy_tasks) {
4055 #if OMPT_SUPPORT
4056     // the runtime is shutting down so we won't report any events
4057     thread->th.ompt_thread_info.state = ompt_state_undefined;
4058 #endif
4059     __kmp_task_team_wait(thread, team USE_ITT_BUILD_ARG(NULL));
4060   }
4061 #endif
4062 
4063   __kmp_reset_root(gtid, root);
4064 
4065   /* free up this thread slot */
4066   __kmp_gtid_set_specific(KMP_GTID_DNE);
4067 #ifdef KMP_TDATA_GTID
4068   __kmp_gtid = KMP_GTID_DNE;
4069 #endif
4070 
4071   KMP_MB();
4072   KC_TRACE(10,
4073            ("__kmp_unregister_root_current_thread: T#%d unregistered\n", gtid));
4074 
4075   __kmp_release_bootstrap_lock(&__kmp_forkjoin_lock);
4076 }
4077 
4078 #if KMP_OS_WINDOWS
4079 /* __kmp_forkjoin_lock must be already held
4080    Unregisters a root thread that is not the current thread.  Returns the number
4081    of __kmp_threads entries freed as a result. */
4082 static int __kmp_unregister_root_other_thread(int gtid) {
4083   kmp_root_t *root = __kmp_root[gtid];
4084   int r;
4085 
4086   KA_TRACE(1, ("__kmp_unregister_root_other_thread: enter T#%d\n", gtid));
4087   KMP_DEBUG_ASSERT(__kmp_threads && __kmp_threads[gtid]);
4088   KMP_ASSERT(KMP_UBER_GTID(gtid));
4089   KMP_ASSERT(root == __kmp_threads[gtid]->th.th_root);
4090   KMP_ASSERT(root->r.r_active == FALSE);
4091 
4092   r = __kmp_reset_root(gtid, root);
4093   KC_TRACE(10,
4094            ("__kmp_unregister_root_other_thread: T#%d unregistered\n", gtid));
4095   return r;
4096 }
4097 #endif
4098 
4099 #if KMP_DEBUG
4100 void __kmp_task_info() {
4101 
4102   kmp_int32 gtid = __kmp_entry_gtid();
4103   kmp_int32 tid = __kmp_tid_from_gtid(gtid);
4104   kmp_info_t *this_thr = __kmp_threads[gtid];
4105   kmp_team_t *steam = this_thr->th.th_serial_team;
4106   kmp_team_t *team = this_thr->th.th_team;
4107 
4108   __kmp_printf(
4109       "__kmp_task_info: gtid=%d tid=%d t_thread=%p team=%p steam=%p curtask=%p "
4110       "ptask=%p\n",
4111       gtid, tid, this_thr, team, steam, this_thr->th.th_current_task,
4112       team->t.t_implicit_task_taskdata[tid].td_parent);
4113 }
4114 #endif // KMP_DEBUG
4115 
4116 /* TODO optimize with one big memclr, take out what isn't needed, split
4117    responsibility to workers as much as possible, and delay initialization of
4118    features as much as possible  */
4119 static void __kmp_initialize_info(kmp_info_t *this_thr, kmp_team_t *team,
4120                                   int tid, int gtid) {
4121   /* this_thr->th.th_info.ds.ds_gtid is setup in
4122      kmp_allocate_thread/create_worker.
4123      this_thr->th.th_serial_team is setup in __kmp_allocate_thread */
4124   kmp_info_t *master = team->t.t_threads[0];
4125   KMP_DEBUG_ASSERT(this_thr != NULL);
4126   KMP_DEBUG_ASSERT(this_thr->th.th_serial_team);
4127   KMP_DEBUG_ASSERT(team);
4128   KMP_DEBUG_ASSERT(team->t.t_threads);
4129   KMP_DEBUG_ASSERT(team->t.t_dispatch);
4130   KMP_DEBUG_ASSERT(master);
4131   KMP_DEBUG_ASSERT(master->th.th_root);
4132 
4133   KMP_MB();
4134 
4135   TCW_SYNC_PTR(this_thr->th.th_team, team);
4136 
4137   this_thr->th.th_info.ds.ds_tid = tid;
4138   this_thr->th.th_set_nproc = 0;
4139   if (__kmp_tasking_mode != tskm_immediate_exec)
4140     // When tasking is possible, threads are not safe to reap until they are
4141     // done tasking; this will be set when tasking code is exited in wait
4142     this_thr->th.th_reap_state = KMP_NOT_SAFE_TO_REAP;
4143   else // no tasking --> always safe to reap
4144     this_thr->th.th_reap_state = KMP_SAFE_TO_REAP;
4145 #if OMP_40_ENABLED
4146   this_thr->th.th_set_proc_bind = proc_bind_default;
4147 #if KMP_AFFINITY_SUPPORTED
4148   this_thr->th.th_new_place = this_thr->th.th_current_place;
4149 #endif
4150 #endif
4151   this_thr->th.th_root = master->th.th_root;
4152 
4153   /* setup the thread's cache of the team structure */
4154   this_thr->th.th_team_nproc = team->t.t_nproc;
4155   this_thr->th.th_team_master = master;
4156   this_thr->th.th_team_serialized = team->t.t_serialized;
4157   TCW_PTR(this_thr->th.th_sleep_loc, NULL);
4158 
4159   KMP_DEBUG_ASSERT(team->t.t_implicit_task_taskdata);
4160 
4161   KF_TRACE(10, ("__kmp_initialize_info1: T#%d:%d this_thread=%p curtask=%p\n",
4162                 tid, gtid, this_thr, this_thr->th.th_current_task));
4163 
4164   __kmp_init_implicit_task(this_thr->th.th_team_master->th.th_ident, this_thr,
4165                            team, tid, TRUE);
4166 
4167   KF_TRACE(10, ("__kmp_initialize_info2: T#%d:%d this_thread=%p curtask=%p\n",
4168                 tid, gtid, this_thr, this_thr->th.th_current_task));
4169   // TODO: Initialize ICVs from parent; GEH - isn't that already done in
4170   // __kmp_initialize_team()?
4171 
4172   /* TODO no worksharing in speculative threads */
4173   this_thr->th.th_dispatch = &team->t.t_dispatch[tid];
4174 
4175   this_thr->th.th_local.this_construct = 0;
4176 
4177   if (!this_thr->th.th_pri_common) {
4178     this_thr->th.th_pri_common =
4179         (struct common_table *)__kmp_allocate(sizeof(struct common_table));
4180     if (__kmp_storage_map) {
4181       __kmp_print_storage_map_gtid(
4182           gtid, this_thr->th.th_pri_common, this_thr->th.th_pri_common + 1,
4183           sizeof(struct common_table), "th_%d.th_pri_common\n", gtid);
4184     }
4185     this_thr->th.th_pri_head = NULL;
4186   }
4187 
4188   if (this_thr != master && // Master's CG root is initialized elsewhere
4189       this_thr->th.th_cg_roots != master->th.th_cg_roots) { // CG root not set
4190     // Make new thread's CG root same as master's
4191     KMP_DEBUG_ASSERT(master->th.th_cg_roots);
4192     this_thr->th.th_cg_roots = master->th.th_cg_roots;
4193     // Increment new thread's CG root's counter to add the new thread
4194     this_thr->th.th_cg_roots->cg_nthreads++;
4195     KA_TRACE(100, ("__kmp_initialize_info: Thread %p increment cg_nthreads on"
4196                    " node %p of thread %p to %d\n",
4197                    this_thr, this_thr->th.th_cg_roots,
4198                    this_thr->th.th_cg_roots->cg_root,
4199                    this_thr->th.th_cg_roots->cg_nthreads));
4200     this_thr->th.th_current_task->td_icvs.thread_limit =
4201         this_thr->th.th_cg_roots->cg_thread_limit;
4202   }
4203 
4204   /* Initialize dynamic dispatch */
4205   {
4206     volatile kmp_disp_t *dispatch = this_thr->th.th_dispatch;
4207     // Use team max_nproc since this will never change for the team.
4208     size_t disp_size =
4209         sizeof(dispatch_private_info_t) *
4210         (team->t.t_max_nproc == 1 ? 1 : __kmp_dispatch_num_buffers);
4211     KD_TRACE(10, ("__kmp_initialize_info: T#%d max_nproc: %d\n", gtid,
4212                   team->t.t_max_nproc));
4213     KMP_ASSERT(dispatch);
4214     KMP_DEBUG_ASSERT(team->t.t_dispatch);
4215     KMP_DEBUG_ASSERT(dispatch == &team->t.t_dispatch[tid]);
4216 
4217     dispatch->th_disp_index = 0;
4218 #if OMP_45_ENABLED
4219     dispatch->th_doacross_buf_idx = 0;
4220 #endif
4221     if (!dispatch->th_disp_buffer) {
4222       dispatch->th_disp_buffer =
4223           (dispatch_private_info_t *)__kmp_allocate(disp_size);
4224 
4225       if (__kmp_storage_map) {
4226         __kmp_print_storage_map_gtid(
4227             gtid, &dispatch->th_disp_buffer[0],
4228             &dispatch->th_disp_buffer[team->t.t_max_nproc == 1
4229                                           ? 1
4230                                           : __kmp_dispatch_num_buffers],
4231             disp_size, "th_%d.th_dispatch.th_disp_buffer "
4232                        "(team_%d.t_dispatch[%d].th_disp_buffer)",
4233             gtid, team->t.t_id, gtid);
4234       }
4235     } else {
4236       memset(&dispatch->th_disp_buffer[0], '\0', disp_size);
4237     }
4238 
4239     dispatch->th_dispatch_pr_current = 0;
4240     dispatch->th_dispatch_sh_current = 0;
4241 
4242     dispatch->th_deo_fcn = 0; /* ORDERED     */
4243     dispatch->th_dxo_fcn = 0; /* END ORDERED */
4244   }
4245 
4246   this_thr->th.th_next_pool = NULL;
4247 
4248   if (!this_thr->th.th_task_state_memo_stack) {
4249     size_t i;
4250     this_thr->th.th_task_state_memo_stack =
4251         (kmp_uint8 *)__kmp_allocate(4 * sizeof(kmp_uint8));
4252     this_thr->th.th_task_state_top = 0;
4253     this_thr->th.th_task_state_stack_sz = 4;
4254     for (i = 0; i < this_thr->th.th_task_state_stack_sz;
4255          ++i) // zero init the stack
4256       this_thr->th.th_task_state_memo_stack[i] = 0;
4257   }
4258 
4259   KMP_DEBUG_ASSERT(!this_thr->th.th_spin_here);
4260   KMP_DEBUG_ASSERT(this_thr->th.th_next_waiting == 0);
4261 
4262   KMP_MB();
4263 }
4264 
4265 /* allocate a new thread for the requesting team. this is only called from
4266    within a forkjoin critical section. we will first try to get an available
4267    thread from the thread pool. if none is available, we will fork a new one
4268    assuming we are able to create a new one. this should be assured, as the
4269    caller should check on this first. */
4270 kmp_info_t *__kmp_allocate_thread(kmp_root_t *root, kmp_team_t *team,
4271                                   int new_tid) {
4272   kmp_team_t *serial_team;
4273   kmp_info_t *new_thr;
4274   int new_gtid;
4275 
4276   KA_TRACE(20, ("__kmp_allocate_thread: T#%d\n", __kmp_get_gtid()));
4277   KMP_DEBUG_ASSERT(root && team);
4278 #if !KMP_NESTED_HOT_TEAMS
4279   KMP_DEBUG_ASSERT(KMP_MASTER_GTID(__kmp_get_gtid()));
4280 #endif
4281   KMP_MB();
4282 
4283   /* first, try to get one from the thread pool */
4284   if (__kmp_thread_pool) {
4285     new_thr = CCAST(kmp_info_t *, __kmp_thread_pool);
4286     __kmp_thread_pool = (volatile kmp_info_t *)new_thr->th.th_next_pool;
4287     if (new_thr == __kmp_thread_pool_insert_pt) {
4288       __kmp_thread_pool_insert_pt = NULL;
4289     }
4290     TCW_4(new_thr->th.th_in_pool, FALSE);
4291     __kmp_suspend_initialize_thread(new_thr);
4292     __kmp_lock_suspend_mx(new_thr);
4293     if (new_thr->th.th_active_in_pool == TRUE) {
4294       KMP_DEBUG_ASSERT(new_thr->th.th_active == TRUE);
4295       KMP_ATOMIC_DEC(&__kmp_thread_pool_active_nth);
4296       new_thr->th.th_active_in_pool = FALSE;
4297     }
4298 #if KMP_DEBUG
4299     else {
4300       KMP_DEBUG_ASSERT(new_thr->th.th_active == FALSE);
4301     }
4302 #endif
4303     __kmp_unlock_suspend_mx(new_thr);
4304 
4305     KA_TRACE(20, ("__kmp_allocate_thread: T#%d using thread T#%d\n",
4306                   __kmp_get_gtid(), new_thr->th.th_info.ds.ds_gtid));
4307     KMP_ASSERT(!new_thr->th.th_team);
4308     KMP_DEBUG_ASSERT(__kmp_nth < __kmp_threads_capacity);
4309 
4310     /* setup the thread structure */
4311     __kmp_initialize_info(new_thr, team, new_tid,
4312                           new_thr->th.th_info.ds.ds_gtid);
4313     KMP_DEBUG_ASSERT(new_thr->th.th_serial_team);
4314 
4315     TCW_4(__kmp_nth, __kmp_nth + 1);
4316 
4317     new_thr->th.th_task_state = 0;
4318     new_thr->th.th_task_state_top = 0;
4319     new_thr->th.th_task_state_stack_sz = 4;
4320 
4321 #ifdef KMP_ADJUST_BLOCKTIME
4322     /* Adjust blocktime back to zero if necessary */
4323     /* Middle initialization might not have occurred yet */
4324     if (!__kmp_env_blocktime && (__kmp_avail_proc > 0)) {
4325       if (__kmp_nth > __kmp_avail_proc) {
4326         __kmp_zero_bt = TRUE;
4327       }
4328     }
4329 #endif /* KMP_ADJUST_BLOCKTIME */
4330 
4331 #if KMP_DEBUG
4332     // If thread entered pool via __kmp_free_thread, wait_flag should !=
4333     // KMP_BARRIER_PARENT_FLAG.
4334     int b;
4335     kmp_balign_t *balign = new_thr->th.th_bar;
4336     for (b = 0; b < bs_last_barrier; ++b)
4337       KMP_DEBUG_ASSERT(balign[b].bb.wait_flag != KMP_BARRIER_PARENT_FLAG);
4338 #endif
4339 
4340     KF_TRACE(10, ("__kmp_allocate_thread: T#%d using thread %p T#%d\n",
4341                   __kmp_get_gtid(), new_thr, new_thr->th.th_info.ds.ds_gtid));
4342 
4343     KMP_MB();
4344     return new_thr;
4345   }
4346 
4347   /* no, well fork a new one */
4348   KMP_ASSERT(__kmp_nth == __kmp_all_nth);
4349   KMP_ASSERT(__kmp_all_nth < __kmp_threads_capacity);
4350 
4351 #if KMP_USE_MONITOR
4352   // If this is the first worker thread the RTL is creating, then also
4353   // launch the monitor thread.  We try to do this as early as possible.
4354   if (!TCR_4(__kmp_init_monitor)) {
4355     __kmp_acquire_bootstrap_lock(&__kmp_monitor_lock);
4356     if (!TCR_4(__kmp_init_monitor)) {
4357       KF_TRACE(10, ("before __kmp_create_monitor\n"));
4358       TCW_4(__kmp_init_monitor, 1);
4359       __kmp_create_monitor(&__kmp_monitor);
4360       KF_TRACE(10, ("after __kmp_create_monitor\n"));
4361 #if KMP_OS_WINDOWS
4362       // AC: wait until monitor has started. This is a fix for CQ232808.
4363       // The reason is that if the library is loaded/unloaded in a loop with
4364       // small (parallel) work in between, then there is high probability that
4365       // monitor thread started after the library shutdown. At shutdown it is
4366       // too late to cope with the problem, because when the master is in
4367       // DllMain (process detach) the monitor has no chances to start (it is
4368       // blocked), and master has no means to inform the monitor that the
4369       // library has gone, because all the memory which the monitor can access
4370       // is going to be released/reset.
4371       while (TCR_4(__kmp_init_monitor) < 2) {
4372         KMP_YIELD(TRUE);
4373       }
4374       KF_TRACE(10, ("after monitor thread has started\n"));
4375 #endif
4376     }
4377     __kmp_release_bootstrap_lock(&__kmp_monitor_lock);
4378   }
4379 #endif
4380 
4381   KMP_MB();
4382   for (new_gtid = 1; TCR_PTR(__kmp_threads[new_gtid]) != NULL; ++new_gtid) {
4383     KMP_DEBUG_ASSERT(new_gtid < __kmp_threads_capacity);
4384   }
4385 
4386   /* allocate space for it. */
4387   new_thr = (kmp_info_t *)__kmp_allocate(sizeof(kmp_info_t));
4388 
4389   TCW_SYNC_PTR(__kmp_threads[new_gtid], new_thr);
4390 
4391   if (__kmp_storage_map) {
4392     __kmp_print_thread_storage_map(new_thr, new_gtid);
4393   }
4394 
4395   // add the reserve serialized team, initialized from the team's master thread
4396   {
4397     kmp_internal_control_t r_icvs = __kmp_get_x_global_icvs(team);
4398     KF_TRACE(10, ("__kmp_allocate_thread: before th_serial/serial_team\n"));
4399     new_thr->th.th_serial_team = serial_team =
4400         (kmp_team_t *)__kmp_allocate_team(root, 1, 1,
4401 #if OMPT_SUPPORT
4402                                           ompt_data_none, // root parallel id
4403 #endif
4404 #if OMP_40_ENABLED
4405                                           proc_bind_default,
4406 #endif
4407                                           &r_icvs, 0 USE_NESTED_HOT_ARG(NULL));
4408   }
4409   KMP_ASSERT(serial_team);
4410   serial_team->t.t_serialized = 0; // AC: the team created in reserve, not for
4411   // execution (it is unused for now).
4412   serial_team->t.t_threads[0] = new_thr;
4413   KF_TRACE(10,
4414            ("__kmp_allocate_thread: after th_serial/serial_team : new_thr=%p\n",
4415             new_thr));
4416 
4417   /* setup the thread structures */
4418   __kmp_initialize_info(new_thr, team, new_tid, new_gtid);
4419 
4420 #if USE_FAST_MEMORY
4421   __kmp_initialize_fast_memory(new_thr);
4422 #endif /* USE_FAST_MEMORY */
4423 
4424 #if KMP_USE_BGET
4425   KMP_DEBUG_ASSERT(new_thr->th.th_local.bget_data == NULL);
4426   __kmp_initialize_bget(new_thr);
4427 #endif
4428 
4429   __kmp_init_random(new_thr); // Initialize random number generator
4430 
4431   /* Initialize these only once when thread is grabbed for a team allocation */
4432   KA_TRACE(20,
4433            ("__kmp_allocate_thread: T#%d init go fork=%u, plain=%u\n",
4434             __kmp_get_gtid(), KMP_INIT_BARRIER_STATE, KMP_INIT_BARRIER_STATE));
4435 
4436   int b;
4437   kmp_balign_t *balign = new_thr->th.th_bar;
4438   for (b = 0; b < bs_last_barrier; ++b) {
4439     balign[b].bb.b_go = KMP_INIT_BARRIER_STATE;
4440     balign[b].bb.team = NULL;
4441     balign[b].bb.wait_flag = KMP_BARRIER_NOT_WAITING;
4442     balign[b].bb.use_oncore_barrier = 0;
4443   }
4444 
4445   new_thr->th.th_spin_here = FALSE;
4446   new_thr->th.th_next_waiting = 0;
4447 #if KMP_OS_UNIX
4448   new_thr->th.th_blocking = false;
4449 #endif
4450 
4451 #if OMP_40_ENABLED && KMP_AFFINITY_SUPPORTED
4452   new_thr->th.th_current_place = KMP_PLACE_UNDEFINED;
4453   new_thr->th.th_new_place = KMP_PLACE_UNDEFINED;
4454   new_thr->th.th_first_place = KMP_PLACE_UNDEFINED;
4455   new_thr->th.th_last_place = KMP_PLACE_UNDEFINED;
4456 #endif
4457 #if OMP_50_ENABLED
4458   new_thr->th.th_def_allocator = __kmp_def_allocator;
4459   new_thr->th.th_prev_level = 0;
4460   new_thr->th.th_prev_num_threads = 1;
4461 #endif
4462 
4463   TCW_4(new_thr->th.th_in_pool, FALSE);
4464   new_thr->th.th_active_in_pool = FALSE;
4465   TCW_4(new_thr->th.th_active, TRUE);
4466 
4467   /* adjust the global counters */
4468   __kmp_all_nth++;
4469   __kmp_nth++;
4470 
4471   // if __kmp_adjust_gtid_mode is set, then we use method #1 (sp search) for low
4472   // numbers of procs, and method #2 (keyed API call) for higher numbers.
4473   if (__kmp_adjust_gtid_mode) {
4474     if (__kmp_all_nth >= __kmp_tls_gtid_min) {
4475       if (TCR_4(__kmp_gtid_mode) != 2) {
4476         TCW_4(__kmp_gtid_mode, 2);
4477       }
4478     } else {
4479       if (TCR_4(__kmp_gtid_mode) != 1) {
4480         TCW_4(__kmp_gtid_mode, 1);
4481       }
4482     }
4483   }
4484 
4485 #ifdef KMP_ADJUST_BLOCKTIME
4486   /* Adjust blocktime back to zero if necessary       */
4487   /* Middle initialization might not have occurred yet */
4488   if (!__kmp_env_blocktime && (__kmp_avail_proc > 0)) {
4489     if (__kmp_nth > __kmp_avail_proc) {
4490       __kmp_zero_bt = TRUE;
4491     }
4492   }
4493 #endif /* KMP_ADJUST_BLOCKTIME */
4494 
4495   /* actually fork it and create the new worker thread */
4496   KF_TRACE(
4497       10, ("__kmp_allocate_thread: before __kmp_create_worker: %p\n", new_thr));
4498   __kmp_create_worker(new_gtid, new_thr, __kmp_stksize);
4499   KF_TRACE(10,
4500            ("__kmp_allocate_thread: after __kmp_create_worker: %p\n", new_thr));
4501 
4502   KA_TRACE(20, ("__kmp_allocate_thread: T#%d forked T#%d\n", __kmp_get_gtid(),
4503                 new_gtid));
4504   KMP_MB();
4505   return new_thr;
4506 }
4507 
4508 /* Reinitialize team for reuse.
4509    The hot team code calls this case at every fork barrier, so EPCC barrier
4510    test are extremely sensitive to changes in it, esp. writes to the team
4511    struct, which cause a cache invalidation in all threads.
4512    IF YOU TOUCH THIS ROUTINE, RUN EPCC C SYNCBENCH ON A BIG-IRON MACHINE!!! */
4513 static void __kmp_reinitialize_team(kmp_team_t *team,
4514                                     kmp_internal_control_t *new_icvs,
4515                                     ident_t *loc) {
4516   KF_TRACE(10, ("__kmp_reinitialize_team: enter this_thread=%p team=%p\n",
4517                 team->t.t_threads[0], team));
4518   KMP_DEBUG_ASSERT(team && new_icvs);
4519   KMP_DEBUG_ASSERT((!TCR_4(__kmp_init_parallel)) || new_icvs->nproc);
4520   KMP_CHECK_UPDATE(team->t.t_ident, loc);
4521 
4522   KMP_CHECK_UPDATE(team->t.t_id, KMP_GEN_TEAM_ID());
4523   // Copy ICVs to the master thread's implicit taskdata
4524   __kmp_init_implicit_task(loc, team->t.t_threads[0], team, 0, FALSE);
4525   copy_icvs(&team->t.t_implicit_task_taskdata[0].td_icvs, new_icvs);
4526 
4527   KF_TRACE(10, ("__kmp_reinitialize_team: exit this_thread=%p team=%p\n",
4528                 team->t.t_threads[0], team));
4529 }
4530 
4531 /* Initialize the team data structure.
4532    This assumes the t_threads and t_max_nproc are already set.
4533    Also, we don't touch the arguments */
4534 static void __kmp_initialize_team(kmp_team_t *team, int new_nproc,
4535                                   kmp_internal_control_t *new_icvs,
4536                                   ident_t *loc) {
4537   KF_TRACE(10, ("__kmp_initialize_team: enter: team=%p\n", team));
4538 
4539   /* verify */
4540   KMP_DEBUG_ASSERT(team);
4541   KMP_DEBUG_ASSERT(new_nproc <= team->t.t_max_nproc);
4542   KMP_DEBUG_ASSERT(team->t.t_threads);
4543   KMP_MB();
4544 
4545   team->t.t_master_tid = 0; /* not needed */
4546   /* team->t.t_master_bar;        not needed */
4547   team->t.t_serialized = new_nproc > 1 ? 0 : 1;
4548   team->t.t_nproc = new_nproc;
4549 
4550   /* team->t.t_parent     = NULL; TODO not needed & would mess up hot team */
4551   team->t.t_next_pool = NULL;
4552   /* memset( team->t.t_threads, 0, sizeof(kmp_info_t*)*new_nproc ); would mess
4553    * up hot team */
4554 
4555   TCW_SYNC_PTR(team->t.t_pkfn, NULL); /* not needed */
4556   team->t.t_invoke = NULL; /* not needed */
4557 
4558   // TODO???: team->t.t_max_active_levels       = new_max_active_levels;
4559   team->t.t_sched.sched = new_icvs->sched.sched;
4560 
4561 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
4562   team->t.t_fp_control_saved = FALSE; /* not needed */
4563   team->t.t_x87_fpu_control_word = 0; /* not needed */
4564   team->t.t_mxcsr = 0; /* not needed */
4565 #endif /* KMP_ARCH_X86 || KMP_ARCH_X86_64 */
4566 
4567   team->t.t_construct = 0;
4568 
4569   team->t.t_ordered.dt.t_value = 0;
4570   team->t.t_master_active = FALSE;
4571 
4572 #ifdef KMP_DEBUG
4573   team->t.t_copypriv_data = NULL; /* not necessary, but nice for debugging */
4574 #endif
4575 #if KMP_OS_WINDOWS
4576   team->t.t_copyin_counter = 0; /* for barrier-free copyin implementation */
4577 #endif
4578 
4579   team->t.t_control_stack_top = NULL;
4580 
4581   __kmp_reinitialize_team(team, new_icvs, loc);
4582 
4583   KMP_MB();
4584   KF_TRACE(10, ("__kmp_initialize_team: exit: team=%p\n", team));
4585 }
4586 
4587 #if KMP_OS_LINUX && KMP_AFFINITY_SUPPORTED
4588 /* Sets full mask for thread and returns old mask, no changes to structures. */
4589 static void
4590 __kmp_set_thread_affinity_mask_full_tmp(kmp_affin_mask_t *old_mask) {
4591   if (KMP_AFFINITY_CAPABLE()) {
4592     int status;
4593     if (old_mask != NULL) {
4594       status = __kmp_get_system_affinity(old_mask, TRUE);
4595       int error = errno;
4596       if (status != 0) {
4597         __kmp_fatal(KMP_MSG(ChangeThreadAffMaskError), KMP_ERR(error),
4598                     __kmp_msg_null);
4599       }
4600     }
4601     __kmp_set_system_affinity(__kmp_affin_fullMask, TRUE);
4602   }
4603 }
4604 #endif
4605 
4606 #if OMP_40_ENABLED && KMP_AFFINITY_SUPPORTED
4607 
4608 // __kmp_partition_places() is the heart of the OpenMP 4.0 affinity mechanism.
4609 // It calculats the worker + master thread's partition based upon the parent
4610 // thread's partition, and binds each worker to a thread in their partition.
4611 // The master thread's partition should already include its current binding.
4612 static void __kmp_partition_places(kmp_team_t *team, int update_master_only) {
4613   // Copy the master thread's place partion to the team struct
4614   kmp_info_t *master_th = team->t.t_threads[0];
4615   KMP_DEBUG_ASSERT(master_th != NULL);
4616   kmp_proc_bind_t proc_bind = team->t.t_proc_bind;
4617   int first_place = master_th->th.th_first_place;
4618   int last_place = master_th->th.th_last_place;
4619   int masters_place = master_th->th.th_current_place;
4620   team->t.t_first_place = first_place;
4621   team->t.t_last_place = last_place;
4622 
4623   KA_TRACE(20, ("__kmp_partition_places: enter: proc_bind = %d T#%d(%d:0) "
4624                 "bound to place %d partition = [%d,%d]\n",
4625                 proc_bind, __kmp_gtid_from_thread(team->t.t_threads[0]),
4626                 team->t.t_id, masters_place, first_place, last_place));
4627 
4628   switch (proc_bind) {
4629 
4630   case proc_bind_default:
4631     // serial teams might have the proc_bind policy set to proc_bind_default. It
4632     // doesn't matter, as we don't rebind master thread for any proc_bind policy
4633     KMP_DEBUG_ASSERT(team->t.t_nproc == 1);
4634     break;
4635 
4636   case proc_bind_master: {
4637     int f;
4638     int n_th = team->t.t_nproc;
4639     for (f = 1; f < n_th; f++) {
4640       kmp_info_t *th = team->t.t_threads[f];
4641       KMP_DEBUG_ASSERT(th != NULL);
4642       th->th.th_first_place = first_place;
4643       th->th.th_last_place = last_place;
4644       th->th.th_new_place = masters_place;
4645 #if OMP_50_ENABLED
4646       if (__kmp_display_affinity && masters_place != th->th.th_current_place &&
4647           team->t.t_display_affinity != 1) {
4648         team->t.t_display_affinity = 1;
4649       }
4650 #endif
4651 
4652       KA_TRACE(100, ("__kmp_partition_places: master: T#%d(%d:%d) place %d "
4653                      "partition = [%d,%d]\n",
4654                      __kmp_gtid_from_thread(team->t.t_threads[f]), team->t.t_id,
4655                      f, masters_place, first_place, last_place));
4656     }
4657   } break;
4658 
4659   case proc_bind_close: {
4660     int f;
4661     int n_th = team->t.t_nproc;
4662     int n_places;
4663     if (first_place <= last_place) {
4664       n_places = last_place - first_place + 1;
4665     } else {
4666       n_places = __kmp_affinity_num_masks - first_place + last_place + 1;
4667     }
4668     if (n_th <= n_places) {
4669       int place = masters_place;
4670       for (f = 1; f < n_th; f++) {
4671         kmp_info_t *th = team->t.t_threads[f];
4672         KMP_DEBUG_ASSERT(th != NULL);
4673 
4674         if (place == last_place) {
4675           place = first_place;
4676         } else if (place == (int)(__kmp_affinity_num_masks - 1)) {
4677           place = 0;
4678         } else {
4679           place++;
4680         }
4681         th->th.th_first_place = first_place;
4682         th->th.th_last_place = last_place;
4683         th->th.th_new_place = place;
4684 #if OMP_50_ENABLED
4685         if (__kmp_display_affinity && place != th->th.th_current_place &&
4686             team->t.t_display_affinity != 1) {
4687           team->t.t_display_affinity = 1;
4688         }
4689 #endif
4690 
4691         KA_TRACE(100, ("__kmp_partition_places: close: T#%d(%d:%d) place %d "
4692                        "partition = [%d,%d]\n",
4693                        __kmp_gtid_from_thread(team->t.t_threads[f]),
4694                        team->t.t_id, f, place, first_place, last_place));
4695       }
4696     } else {
4697       int S, rem, gap, s_count;
4698       S = n_th / n_places;
4699       s_count = 0;
4700       rem = n_th - (S * n_places);
4701       gap = rem > 0 ? n_places / rem : n_places;
4702       int place = masters_place;
4703       int gap_ct = gap;
4704       for (f = 0; f < n_th; f++) {
4705         kmp_info_t *th = team->t.t_threads[f];
4706         KMP_DEBUG_ASSERT(th != NULL);
4707 
4708         th->th.th_first_place = first_place;
4709         th->th.th_last_place = last_place;
4710         th->th.th_new_place = place;
4711 #if OMP_50_ENABLED
4712         if (__kmp_display_affinity && place != th->th.th_current_place &&
4713             team->t.t_display_affinity != 1) {
4714           team->t.t_display_affinity = 1;
4715         }
4716 #endif
4717         s_count++;
4718 
4719         if ((s_count == S) && rem && (gap_ct == gap)) {
4720           // do nothing, add an extra thread to place on next iteration
4721         } else if ((s_count == S + 1) && rem && (gap_ct == gap)) {
4722           // we added an extra thread to this place; move to next place
4723           if (place == last_place) {
4724             place = first_place;
4725           } else if (place == (int)(__kmp_affinity_num_masks - 1)) {
4726             place = 0;
4727           } else {
4728             place++;
4729           }
4730           s_count = 0;
4731           gap_ct = 1;
4732           rem--;
4733         } else if (s_count == S) { // place full; don't add extra
4734           if (place == last_place) {
4735             place = first_place;
4736           } else if (place == (int)(__kmp_affinity_num_masks - 1)) {
4737             place = 0;
4738           } else {
4739             place++;
4740           }
4741           gap_ct++;
4742           s_count = 0;
4743         }
4744 
4745         KA_TRACE(100,
4746                  ("__kmp_partition_places: close: T#%d(%d:%d) place %d "
4747                   "partition = [%d,%d]\n",
4748                   __kmp_gtid_from_thread(team->t.t_threads[f]), team->t.t_id, f,
4749                   th->th.th_new_place, first_place, last_place));
4750       }
4751       KMP_DEBUG_ASSERT(place == masters_place);
4752     }
4753   } break;
4754 
4755   case proc_bind_spread: {
4756     int f;
4757     int n_th = team->t.t_nproc;
4758     int n_places;
4759     int thidx;
4760     if (first_place <= last_place) {
4761       n_places = last_place - first_place + 1;
4762     } else {
4763       n_places = __kmp_affinity_num_masks - first_place + last_place + 1;
4764     }
4765     if (n_th <= n_places) {
4766       int place = -1;
4767 
4768       if (n_places != static_cast<int>(__kmp_affinity_num_masks)) {
4769         int S = n_places / n_th;
4770         int s_count, rem, gap, gap_ct;
4771 
4772         place = masters_place;
4773         rem = n_places - n_th * S;
4774         gap = rem ? n_th / rem : 1;
4775         gap_ct = gap;
4776         thidx = n_th;
4777         if (update_master_only == 1)
4778           thidx = 1;
4779         for (f = 0; f < thidx; f++) {
4780           kmp_info_t *th = team->t.t_threads[f];
4781           KMP_DEBUG_ASSERT(th != NULL);
4782 
4783           th->th.th_first_place = place;
4784           th->th.th_new_place = place;
4785 #if OMP_50_ENABLED
4786           if (__kmp_display_affinity && place != th->th.th_current_place &&
4787               team->t.t_display_affinity != 1) {
4788             team->t.t_display_affinity = 1;
4789           }
4790 #endif
4791           s_count = 1;
4792           while (s_count < S) {
4793             if (place == last_place) {
4794               place = first_place;
4795             } else if (place == (int)(__kmp_affinity_num_masks - 1)) {
4796               place = 0;
4797             } else {
4798               place++;
4799             }
4800             s_count++;
4801           }
4802           if (rem && (gap_ct == gap)) {
4803             if (place == last_place) {
4804               place = first_place;
4805             } else if (place == (int)(__kmp_affinity_num_masks - 1)) {
4806               place = 0;
4807             } else {
4808               place++;
4809             }
4810             rem--;
4811             gap_ct = 0;
4812           }
4813           th->th.th_last_place = place;
4814           gap_ct++;
4815 
4816           if (place == last_place) {
4817             place = first_place;
4818           } else if (place == (int)(__kmp_affinity_num_masks - 1)) {
4819             place = 0;
4820           } else {
4821             place++;
4822           }
4823 
4824           KA_TRACE(100,
4825                    ("__kmp_partition_places: spread: T#%d(%d:%d) place %d "
4826                     "partition = [%d,%d], __kmp_affinity_num_masks: %u\n",
4827                     __kmp_gtid_from_thread(team->t.t_threads[f]), team->t.t_id,
4828                     f, th->th.th_new_place, th->th.th_first_place,
4829                     th->th.th_last_place, __kmp_affinity_num_masks));
4830         }
4831       } else {
4832         /* Having uniform space of available computation places I can create
4833            T partitions of round(P/T) size and put threads into the first
4834            place of each partition. */
4835         double current = static_cast<double>(masters_place);
4836         double spacing =
4837             (static_cast<double>(n_places + 1) / static_cast<double>(n_th));
4838         int first, last;
4839         kmp_info_t *th;
4840 
4841         thidx = n_th + 1;
4842         if (update_master_only == 1)
4843           thidx = 1;
4844         for (f = 0; f < thidx; f++) {
4845           first = static_cast<int>(current);
4846           last = static_cast<int>(current + spacing) - 1;
4847           KMP_DEBUG_ASSERT(last >= first);
4848           if (first >= n_places) {
4849             if (masters_place) {
4850               first -= n_places;
4851               last -= n_places;
4852               if (first == (masters_place + 1)) {
4853                 KMP_DEBUG_ASSERT(f == n_th);
4854                 first--;
4855               }
4856               if (last == masters_place) {
4857                 KMP_DEBUG_ASSERT(f == (n_th - 1));
4858                 last--;
4859               }
4860             } else {
4861               KMP_DEBUG_ASSERT(f == n_th);
4862               first = 0;
4863               last = 0;
4864             }
4865           }
4866           if (last >= n_places) {
4867             last = (n_places - 1);
4868           }
4869           place = first;
4870           current += spacing;
4871           if (f < n_th) {
4872             KMP_DEBUG_ASSERT(0 <= first);
4873             KMP_DEBUG_ASSERT(n_places > first);
4874             KMP_DEBUG_ASSERT(0 <= last);
4875             KMP_DEBUG_ASSERT(n_places > last);
4876             KMP_DEBUG_ASSERT(last_place >= first_place);
4877             th = team->t.t_threads[f];
4878             KMP_DEBUG_ASSERT(th);
4879             th->th.th_first_place = first;
4880             th->th.th_new_place = place;
4881             th->th.th_last_place = last;
4882 #if OMP_50_ENABLED
4883             if (__kmp_display_affinity && place != th->th.th_current_place &&
4884                 team->t.t_display_affinity != 1) {
4885               team->t.t_display_affinity = 1;
4886             }
4887 #endif
4888             KA_TRACE(100,
4889                      ("__kmp_partition_places: spread: T#%d(%d:%d) place %d "
4890                       "partition = [%d,%d], spacing = %.4f\n",
4891                       __kmp_gtid_from_thread(team->t.t_threads[f]),
4892                       team->t.t_id, f, th->th.th_new_place,
4893                       th->th.th_first_place, th->th.th_last_place, spacing));
4894           }
4895         }
4896       }
4897       KMP_DEBUG_ASSERT(update_master_only || place == masters_place);
4898     } else {
4899       int S, rem, gap, s_count;
4900       S = n_th / n_places;
4901       s_count = 0;
4902       rem = n_th - (S * n_places);
4903       gap = rem > 0 ? n_places / rem : n_places;
4904       int place = masters_place;
4905       int gap_ct = gap;
4906       thidx = n_th;
4907       if (update_master_only == 1)
4908         thidx = 1;
4909       for (f = 0; f < thidx; f++) {
4910         kmp_info_t *th = team->t.t_threads[f];
4911         KMP_DEBUG_ASSERT(th != NULL);
4912 
4913         th->th.th_first_place = place;
4914         th->th.th_last_place = place;
4915         th->th.th_new_place = place;
4916 #if OMP_50_ENABLED
4917         if (__kmp_display_affinity && place != th->th.th_current_place &&
4918             team->t.t_display_affinity != 1) {
4919           team->t.t_display_affinity = 1;
4920         }
4921 #endif
4922         s_count++;
4923 
4924         if ((s_count == S) && rem && (gap_ct == gap)) {
4925           // do nothing, add an extra thread to place on next iteration
4926         } else if ((s_count == S + 1) && rem && (gap_ct == gap)) {
4927           // we added an extra thread to this place; move on to next place
4928           if (place == last_place) {
4929             place = first_place;
4930           } else if (place == (int)(__kmp_affinity_num_masks - 1)) {
4931             place = 0;
4932           } else {
4933             place++;
4934           }
4935           s_count = 0;
4936           gap_ct = 1;
4937           rem--;
4938         } else if (s_count == S) { // place is full; don't add extra thread
4939           if (place == last_place) {
4940             place = first_place;
4941           } else if (place == (int)(__kmp_affinity_num_masks - 1)) {
4942             place = 0;
4943           } else {
4944             place++;
4945           }
4946           gap_ct++;
4947           s_count = 0;
4948         }
4949 
4950         KA_TRACE(100, ("__kmp_partition_places: spread: T#%d(%d:%d) place %d "
4951                        "partition = [%d,%d]\n",
4952                        __kmp_gtid_from_thread(team->t.t_threads[f]),
4953                        team->t.t_id, f, th->th.th_new_place,
4954                        th->th.th_first_place, th->th.th_last_place));
4955       }
4956       KMP_DEBUG_ASSERT(update_master_only || place == masters_place);
4957     }
4958   } break;
4959 
4960   default:
4961     break;
4962   }
4963 
4964   KA_TRACE(20, ("__kmp_partition_places: exit T#%d\n", team->t.t_id));
4965 }
4966 
4967 #endif /* OMP_40_ENABLED && KMP_AFFINITY_SUPPORTED */
4968 
4969 /* allocate a new team data structure to use.  take one off of the free pool if
4970    available */
4971 kmp_team_t *
4972 __kmp_allocate_team(kmp_root_t *root, int new_nproc, int max_nproc,
4973 #if OMPT_SUPPORT
4974                     ompt_data_t ompt_parallel_data,
4975 #endif
4976 #if OMP_40_ENABLED
4977                     kmp_proc_bind_t new_proc_bind,
4978 #endif
4979                     kmp_internal_control_t *new_icvs,
4980                     int argc USE_NESTED_HOT_ARG(kmp_info_t *master)) {
4981   KMP_TIME_DEVELOPER_PARTITIONED_BLOCK(KMP_allocate_team);
4982   int f;
4983   kmp_team_t *team;
4984   int use_hot_team = !root->r.r_active;
4985   int level = 0;
4986 
4987   KA_TRACE(20, ("__kmp_allocate_team: called\n"));
4988   KMP_DEBUG_ASSERT(new_nproc >= 1 && argc >= 0);
4989   KMP_DEBUG_ASSERT(max_nproc >= new_nproc);
4990   KMP_MB();
4991 
4992 #if KMP_NESTED_HOT_TEAMS
4993   kmp_hot_team_ptr_t *hot_teams;
4994   if (master) {
4995     team = master->th.th_team;
4996     level = team->t.t_active_level;
4997     if (master->th.th_teams_microtask) { // in teams construct?
4998       if (master->th.th_teams_size.nteams > 1 &&
4999           ( // #teams > 1
5000               team->t.t_pkfn ==
5001                   (microtask_t)__kmp_teams_master || // inner fork of the teams
5002               master->th.th_teams_level <
5003                   team->t.t_level)) { // or nested parallel inside the teams
5004         ++level; // not increment if #teams==1, or for outer fork of the teams;
5005         // increment otherwise
5006       }
5007     }
5008     hot_teams = master->th.th_hot_teams;
5009     if (level < __kmp_hot_teams_max_level && hot_teams &&
5010         hot_teams[level]
5011             .hot_team) { // hot team has already been allocated for given level
5012       use_hot_team = 1;
5013     } else {
5014       use_hot_team = 0;
5015     }
5016   }
5017 #endif
5018   // Optimization to use a "hot" team
5019   if (use_hot_team && new_nproc > 1) {
5020     KMP_DEBUG_ASSERT(new_nproc <= max_nproc);
5021 #if KMP_NESTED_HOT_TEAMS
5022     team = hot_teams[level].hot_team;
5023 #else
5024     team = root->r.r_hot_team;
5025 #endif
5026 #if KMP_DEBUG
5027     if (__kmp_tasking_mode != tskm_immediate_exec) {
5028       KA_TRACE(20, ("__kmp_allocate_team: hot team task_team[0] = %p "
5029                     "task_team[1] = %p before reinit\n",
5030                     team->t.t_task_team[0], team->t.t_task_team[1]));
5031     }
5032 #endif
5033 
5034     // Has the number of threads changed?
5035     /* Let's assume the most common case is that the number of threads is
5036        unchanged, and put that case first. */
5037     if (team->t.t_nproc == new_nproc) { // Check changes in number of threads
5038       KA_TRACE(20, ("__kmp_allocate_team: reusing hot team\n"));
5039       // This case can mean that omp_set_num_threads() was called and the hot
5040       // team size was already reduced, so we check the special flag
5041       if (team->t.t_size_changed == -1) {
5042         team->t.t_size_changed = 1;
5043       } else {
5044         KMP_CHECK_UPDATE(team->t.t_size_changed, 0);
5045       }
5046 
5047       // TODO???: team->t.t_max_active_levels = new_max_active_levels;
5048       kmp_r_sched_t new_sched = new_icvs->sched;
5049       // set master's schedule as new run-time schedule
5050       KMP_CHECK_UPDATE(team->t.t_sched.sched, new_sched.sched);
5051 
5052       __kmp_reinitialize_team(team, new_icvs,
5053                               root->r.r_uber_thread->th.th_ident);
5054 
5055       KF_TRACE(10, ("__kmp_allocate_team2: T#%d, this_thread=%p team=%p\n", 0,
5056                     team->t.t_threads[0], team));
5057       __kmp_push_current_task_to_thread(team->t.t_threads[0], team, 0);
5058 
5059 #if OMP_40_ENABLED
5060 #if KMP_AFFINITY_SUPPORTED
5061       if ((team->t.t_size_changed == 0) &&
5062           (team->t.t_proc_bind == new_proc_bind)) {
5063         if (new_proc_bind == proc_bind_spread) {
5064           __kmp_partition_places(
5065               team, 1); // add flag to update only master for spread
5066         }
5067         KA_TRACE(200, ("__kmp_allocate_team: reusing hot team #%d bindings: "
5068                        "proc_bind = %d, partition = [%d,%d]\n",
5069                        team->t.t_id, new_proc_bind, team->t.t_first_place,
5070                        team->t.t_last_place));
5071       } else {
5072         KMP_CHECK_UPDATE(team->t.t_proc_bind, new_proc_bind);
5073         __kmp_partition_places(team);
5074       }
5075 #else
5076       KMP_CHECK_UPDATE(team->t.t_proc_bind, new_proc_bind);
5077 #endif /* KMP_AFFINITY_SUPPORTED */
5078 #endif /* OMP_40_ENABLED */
5079     } else if (team->t.t_nproc > new_nproc) {
5080       KA_TRACE(20,
5081                ("__kmp_allocate_team: decreasing hot team thread count to %d\n",
5082                 new_nproc));
5083 
5084       team->t.t_size_changed = 1;
5085 #if KMP_NESTED_HOT_TEAMS
5086       if (__kmp_hot_teams_mode == 0) {
5087         // AC: saved number of threads should correspond to team's value in this
5088         // mode, can be bigger in mode 1, when hot team has threads in reserve
5089         KMP_DEBUG_ASSERT(hot_teams[level].hot_team_nth == team->t.t_nproc);
5090         hot_teams[level].hot_team_nth = new_nproc;
5091 #endif // KMP_NESTED_HOT_TEAMS
5092         /* release the extra threads we don't need any more */
5093         for (f = new_nproc; f < team->t.t_nproc; f++) {
5094           KMP_DEBUG_ASSERT(team->t.t_threads[f]);
5095           if (__kmp_tasking_mode != tskm_immediate_exec) {
5096             // When decreasing team size, threads no longer in the team should
5097             // unref task team.
5098             team->t.t_threads[f]->th.th_task_team = NULL;
5099           }
5100           __kmp_free_thread(team->t.t_threads[f]);
5101           team->t.t_threads[f] = NULL;
5102         }
5103 #if KMP_NESTED_HOT_TEAMS
5104       } // (__kmp_hot_teams_mode == 0)
5105       else {
5106         // When keeping extra threads in team, switch threads to wait on own
5107         // b_go flag
5108         for (f = new_nproc; f < team->t.t_nproc; ++f) {
5109           KMP_DEBUG_ASSERT(team->t.t_threads[f]);
5110           kmp_balign_t *balign = team->t.t_threads[f]->th.th_bar;
5111           for (int b = 0; b < bs_last_barrier; ++b) {
5112             if (balign[b].bb.wait_flag == KMP_BARRIER_PARENT_FLAG) {
5113               balign[b].bb.wait_flag = KMP_BARRIER_SWITCH_TO_OWN_FLAG;
5114             }
5115             KMP_CHECK_UPDATE(balign[b].bb.leaf_kids, 0);
5116           }
5117         }
5118       }
5119 #endif // KMP_NESTED_HOT_TEAMS
5120       team->t.t_nproc = new_nproc;
5121       // TODO???: team->t.t_max_active_levels = new_max_active_levels;
5122       KMP_CHECK_UPDATE(team->t.t_sched.sched, new_icvs->sched.sched);
5123       __kmp_reinitialize_team(team, new_icvs,
5124                               root->r.r_uber_thread->th.th_ident);
5125 
5126       // Update remaining threads
5127       for (f = 0; f < new_nproc; ++f) {
5128         team->t.t_threads[f]->th.th_team_nproc = new_nproc;
5129       }
5130 
5131       // restore the current task state of the master thread: should be the
5132       // implicit task
5133       KF_TRACE(10, ("__kmp_allocate_team: T#%d, this_thread=%p team=%p\n", 0,
5134                     team->t.t_threads[0], team));
5135 
5136       __kmp_push_current_task_to_thread(team->t.t_threads[0], team, 0);
5137 
5138 #ifdef KMP_DEBUG
5139       for (f = 0; f < team->t.t_nproc; f++) {
5140         KMP_DEBUG_ASSERT(team->t.t_threads[f] &&
5141                          team->t.t_threads[f]->th.th_team_nproc ==
5142                              team->t.t_nproc);
5143       }
5144 #endif
5145 
5146 #if OMP_40_ENABLED
5147       KMP_CHECK_UPDATE(team->t.t_proc_bind, new_proc_bind);
5148 #if KMP_AFFINITY_SUPPORTED
5149       __kmp_partition_places(team);
5150 #endif
5151 #endif
5152     } else { // team->t.t_nproc < new_nproc
5153 #if KMP_OS_LINUX && KMP_AFFINITY_SUPPORTED
5154       kmp_affin_mask_t *old_mask;
5155       if (KMP_AFFINITY_CAPABLE()) {
5156         KMP_CPU_ALLOC(old_mask);
5157       }
5158 #endif
5159 
5160       KA_TRACE(20,
5161                ("__kmp_allocate_team: increasing hot team thread count to %d\n",
5162                 new_nproc));
5163 
5164       team->t.t_size_changed = 1;
5165 
5166 #if KMP_NESTED_HOT_TEAMS
5167       int avail_threads = hot_teams[level].hot_team_nth;
5168       if (new_nproc < avail_threads)
5169         avail_threads = new_nproc;
5170       kmp_info_t **other_threads = team->t.t_threads;
5171       for (f = team->t.t_nproc; f < avail_threads; ++f) {
5172         // Adjust barrier data of reserved threads (if any) of the team
5173         // Other data will be set in __kmp_initialize_info() below.
5174         int b;
5175         kmp_balign_t *balign = other_threads[f]->th.th_bar;
5176         for (b = 0; b < bs_last_barrier; ++b) {
5177           balign[b].bb.b_arrived = team->t.t_bar[b].b_arrived;
5178           KMP_DEBUG_ASSERT(balign[b].bb.wait_flag != KMP_BARRIER_PARENT_FLAG);
5179 #if USE_DEBUGGER
5180           balign[b].bb.b_worker_arrived = team->t.t_bar[b].b_team_arrived;
5181 #endif
5182         }
5183       }
5184       if (hot_teams[level].hot_team_nth >= new_nproc) {
5185         // we have all needed threads in reserve, no need to allocate any
5186         // this only possible in mode 1, cannot have reserved threads in mode 0
5187         KMP_DEBUG_ASSERT(__kmp_hot_teams_mode == 1);
5188         team->t.t_nproc = new_nproc; // just get reserved threads involved
5189       } else {
5190         // we may have some threads in reserve, but not enough
5191         team->t.t_nproc =
5192             hot_teams[level]
5193                 .hot_team_nth; // get reserved threads involved if any
5194         hot_teams[level].hot_team_nth = new_nproc; // adjust hot team max size
5195 #endif // KMP_NESTED_HOT_TEAMS
5196         if (team->t.t_max_nproc < new_nproc) {
5197           /* reallocate larger arrays */
5198           __kmp_reallocate_team_arrays(team, new_nproc);
5199           __kmp_reinitialize_team(team, new_icvs, NULL);
5200         }
5201 
5202 #if KMP_OS_LINUX && KMP_AFFINITY_SUPPORTED
5203         /* Temporarily set full mask for master thread before creation of
5204            workers. The reason is that workers inherit the affinity from master,
5205            so if a lot of workers are created on the single core quickly, they
5206            don't get a chance to set their own affinity for a long time. */
5207         __kmp_set_thread_affinity_mask_full_tmp(old_mask);
5208 #endif
5209 
5210         /* allocate new threads for the hot team */
5211         for (f = team->t.t_nproc; f < new_nproc; f++) {
5212           kmp_info_t *new_worker = __kmp_allocate_thread(root, team, f);
5213           KMP_DEBUG_ASSERT(new_worker);
5214           team->t.t_threads[f] = new_worker;
5215 
5216           KA_TRACE(20,
5217                    ("__kmp_allocate_team: team %d init T#%d arrived: "
5218                     "join=%llu, plain=%llu\n",
5219                     team->t.t_id, __kmp_gtid_from_tid(f, team), team->t.t_id, f,
5220                     team->t.t_bar[bs_forkjoin_barrier].b_arrived,
5221                     team->t.t_bar[bs_plain_barrier].b_arrived));
5222 
5223           { // Initialize barrier data for new threads.
5224             int b;
5225             kmp_balign_t *balign = new_worker->th.th_bar;
5226             for (b = 0; b < bs_last_barrier; ++b) {
5227               balign[b].bb.b_arrived = team->t.t_bar[b].b_arrived;
5228               KMP_DEBUG_ASSERT(balign[b].bb.wait_flag !=
5229                                KMP_BARRIER_PARENT_FLAG);
5230 #if USE_DEBUGGER
5231               balign[b].bb.b_worker_arrived = team->t.t_bar[b].b_team_arrived;
5232 #endif
5233             }
5234           }
5235         }
5236 
5237 #if KMP_OS_LINUX && KMP_AFFINITY_SUPPORTED
5238         if (KMP_AFFINITY_CAPABLE()) {
5239           /* Restore initial master thread's affinity mask */
5240           __kmp_set_system_affinity(old_mask, TRUE);
5241           KMP_CPU_FREE(old_mask);
5242         }
5243 #endif
5244 #if KMP_NESTED_HOT_TEAMS
5245       } // end of check of t_nproc vs. new_nproc vs. hot_team_nth
5246 #endif // KMP_NESTED_HOT_TEAMS
5247       /* make sure everyone is syncronized */
5248       int old_nproc = team->t.t_nproc; // save old value and use to update only
5249       // new threads below
5250       __kmp_initialize_team(team, new_nproc, new_icvs,
5251                             root->r.r_uber_thread->th.th_ident);
5252 
5253       /* reinitialize the threads */
5254       KMP_DEBUG_ASSERT(team->t.t_nproc == new_nproc);
5255       for (f = 0; f < team->t.t_nproc; ++f)
5256         __kmp_initialize_info(team->t.t_threads[f], team, f,
5257                               __kmp_gtid_from_tid(f, team));
5258 
5259       if (level) { // set th_task_state for new threads in nested hot team
5260         // __kmp_initialize_info() no longer zeroes th_task_state, so we should
5261         // only need to set the th_task_state for the new threads. th_task_state
5262         // for master thread will not be accurate until after this in
5263         // __kmp_fork_call(), so we look to the master's memo_stack to get the
5264         // correct value.
5265         for (f = old_nproc; f < team->t.t_nproc; ++f)
5266           team->t.t_threads[f]->th.th_task_state =
5267               team->t.t_threads[0]->th.th_task_state_memo_stack[level];
5268       } else { // set th_task_state for new threads in non-nested hot team
5269         int old_state =
5270             team->t.t_threads[0]->th.th_task_state; // copy master's state
5271         for (f = old_nproc; f < team->t.t_nproc; ++f)
5272           team->t.t_threads[f]->th.th_task_state = old_state;
5273       }
5274 
5275 #ifdef KMP_DEBUG
5276       for (f = 0; f < team->t.t_nproc; ++f) {
5277         KMP_DEBUG_ASSERT(team->t.t_threads[f] &&
5278                          team->t.t_threads[f]->th.th_team_nproc ==
5279                              team->t.t_nproc);
5280       }
5281 #endif
5282 
5283 #if OMP_40_ENABLED
5284       KMP_CHECK_UPDATE(team->t.t_proc_bind, new_proc_bind);
5285 #if KMP_AFFINITY_SUPPORTED
5286       __kmp_partition_places(team);
5287 #endif
5288 #endif
5289     } // Check changes in number of threads
5290 
5291 #if OMP_40_ENABLED
5292     kmp_info_t *master = team->t.t_threads[0];
5293     if (master->th.th_teams_microtask) {
5294       for (f = 1; f < new_nproc; ++f) {
5295         // propagate teams construct specific info to workers
5296         kmp_info_t *thr = team->t.t_threads[f];
5297         thr->th.th_teams_microtask = master->th.th_teams_microtask;
5298         thr->th.th_teams_level = master->th.th_teams_level;
5299         thr->th.th_teams_size = master->th.th_teams_size;
5300       }
5301     }
5302 #endif /* OMP_40_ENABLED */
5303 #if KMP_NESTED_HOT_TEAMS
5304     if (level) {
5305       // Sync barrier state for nested hot teams, not needed for outermost hot
5306       // team.
5307       for (f = 1; f < new_nproc; ++f) {
5308         kmp_info_t *thr = team->t.t_threads[f];
5309         int b;
5310         kmp_balign_t *balign = thr->th.th_bar;
5311         for (b = 0; b < bs_last_barrier; ++b) {
5312           balign[b].bb.b_arrived = team->t.t_bar[b].b_arrived;
5313           KMP_DEBUG_ASSERT(balign[b].bb.wait_flag != KMP_BARRIER_PARENT_FLAG);
5314 #if USE_DEBUGGER
5315           balign[b].bb.b_worker_arrived = team->t.t_bar[b].b_team_arrived;
5316 #endif
5317         }
5318       }
5319     }
5320 #endif // KMP_NESTED_HOT_TEAMS
5321 
5322     /* reallocate space for arguments if necessary */
5323     __kmp_alloc_argv_entries(argc, team, TRUE);
5324     KMP_CHECK_UPDATE(team->t.t_argc, argc);
5325     // The hot team re-uses the previous task team,
5326     // if untouched during the previous release->gather phase.
5327 
5328     KF_TRACE(10, (" hot_team = %p\n", team));
5329 
5330 #if KMP_DEBUG
5331     if (__kmp_tasking_mode != tskm_immediate_exec) {
5332       KA_TRACE(20, ("__kmp_allocate_team: hot team task_team[0] = %p "
5333                     "task_team[1] = %p after reinit\n",
5334                     team->t.t_task_team[0], team->t.t_task_team[1]));
5335     }
5336 #endif
5337 
5338 #if OMPT_SUPPORT
5339     __ompt_team_assign_id(team, ompt_parallel_data);
5340 #endif
5341 
5342     KMP_MB();
5343 
5344     return team;
5345   }
5346 
5347   /* next, let's try to take one from the team pool */
5348   KMP_MB();
5349   for (team = CCAST(kmp_team_t *, __kmp_team_pool); (team);) {
5350     /* TODO: consider resizing undersized teams instead of reaping them, now
5351        that we have a resizing mechanism */
5352     if (team->t.t_max_nproc >= max_nproc) {
5353       /* take this team from the team pool */
5354       __kmp_team_pool = team->t.t_next_pool;
5355 
5356       /* setup the team for fresh use */
5357       __kmp_initialize_team(team, new_nproc, new_icvs, NULL);
5358 
5359       KA_TRACE(20, ("__kmp_allocate_team: setting task_team[0] %p and "
5360                     "task_team[1] %p to NULL\n",
5361                     &team->t.t_task_team[0], &team->t.t_task_team[1]));
5362       team->t.t_task_team[0] = NULL;
5363       team->t.t_task_team[1] = NULL;
5364 
5365       /* reallocate space for arguments if necessary */
5366       __kmp_alloc_argv_entries(argc, team, TRUE);
5367       KMP_CHECK_UPDATE(team->t.t_argc, argc);
5368 
5369       KA_TRACE(
5370           20, ("__kmp_allocate_team: team %d init arrived: join=%u, plain=%u\n",
5371                team->t.t_id, KMP_INIT_BARRIER_STATE, KMP_INIT_BARRIER_STATE));
5372       { // Initialize barrier data.
5373         int b;
5374         for (b = 0; b < bs_last_barrier; ++b) {
5375           team->t.t_bar[b].b_arrived = KMP_INIT_BARRIER_STATE;
5376 #if USE_DEBUGGER
5377           team->t.t_bar[b].b_master_arrived = 0;
5378           team->t.t_bar[b].b_team_arrived = 0;
5379 #endif
5380         }
5381       }
5382 
5383 #if OMP_40_ENABLED
5384       team->t.t_proc_bind = new_proc_bind;
5385 #endif
5386 
5387       KA_TRACE(20, ("__kmp_allocate_team: using team from pool %d.\n",
5388                     team->t.t_id));
5389 
5390 #if OMPT_SUPPORT
5391       __ompt_team_assign_id(team, ompt_parallel_data);
5392 #endif
5393 
5394       KMP_MB();
5395 
5396       return team;
5397     }
5398 
5399     /* reap team if it is too small, then loop back and check the next one */
5400     // not sure if this is wise, but, will be redone during the hot-teams
5401     // rewrite.
5402     /* TODO: Use technique to find the right size hot-team, don't reap them */
5403     team = __kmp_reap_team(team);
5404     __kmp_team_pool = team;
5405   }
5406 
5407   /* nothing available in the pool, no matter, make a new team! */
5408   KMP_MB();
5409   team = (kmp_team_t *)__kmp_allocate(sizeof(kmp_team_t));
5410 
5411   /* and set it up */
5412   team->t.t_max_nproc = max_nproc;
5413   /* NOTE well, for some reason allocating one big buffer and dividing it up
5414      seems to really hurt performance a lot on the P4, so, let's not use this */
5415   __kmp_allocate_team_arrays(team, max_nproc);
5416 
5417   KA_TRACE(20, ("__kmp_allocate_team: making a new team\n"));
5418   __kmp_initialize_team(team, new_nproc, new_icvs, NULL);
5419 
5420   KA_TRACE(20, ("__kmp_allocate_team: setting task_team[0] %p and task_team[1] "
5421                 "%p to NULL\n",
5422                 &team->t.t_task_team[0], &team->t.t_task_team[1]));
5423   team->t.t_task_team[0] = NULL; // to be removed, as __kmp_allocate zeroes
5424   // memory, no need to duplicate
5425   team->t.t_task_team[1] = NULL; // to be removed, as __kmp_allocate zeroes
5426   // memory, no need to duplicate
5427 
5428   if (__kmp_storage_map) {
5429     __kmp_print_team_storage_map("team", team, team->t.t_id, new_nproc);
5430   }
5431 
5432   /* allocate space for arguments */
5433   __kmp_alloc_argv_entries(argc, team, FALSE);
5434   team->t.t_argc = argc;
5435 
5436   KA_TRACE(20,
5437            ("__kmp_allocate_team: team %d init arrived: join=%u, plain=%u\n",
5438             team->t.t_id, KMP_INIT_BARRIER_STATE, KMP_INIT_BARRIER_STATE));
5439   { // Initialize barrier data.
5440     int b;
5441     for (b = 0; b < bs_last_barrier; ++b) {
5442       team->t.t_bar[b].b_arrived = KMP_INIT_BARRIER_STATE;
5443 #if USE_DEBUGGER
5444       team->t.t_bar[b].b_master_arrived = 0;
5445       team->t.t_bar[b].b_team_arrived = 0;
5446 #endif
5447     }
5448   }
5449 
5450 #if OMP_40_ENABLED
5451   team->t.t_proc_bind = new_proc_bind;
5452 #endif
5453 
5454 #if OMPT_SUPPORT
5455   __ompt_team_assign_id(team, ompt_parallel_data);
5456   team->t.ompt_serialized_team_info = NULL;
5457 #endif
5458 
5459   KMP_MB();
5460 
5461   KA_TRACE(20, ("__kmp_allocate_team: done creating a new team %d.\n",
5462                 team->t.t_id));
5463 
5464   return team;
5465 }
5466 
5467 /* TODO implement hot-teams at all levels */
5468 /* TODO implement lazy thread release on demand (disband request) */
5469 
5470 /* free the team.  return it to the team pool.  release all the threads
5471  * associated with it */
5472 void __kmp_free_team(kmp_root_t *root,
5473                      kmp_team_t *team USE_NESTED_HOT_ARG(kmp_info_t *master)) {
5474   int f;
5475   KA_TRACE(20, ("__kmp_free_team: T#%d freeing team %d\n", __kmp_get_gtid(),
5476                 team->t.t_id));
5477 
5478   /* verify state */
5479   KMP_DEBUG_ASSERT(root);
5480   KMP_DEBUG_ASSERT(team);
5481   KMP_DEBUG_ASSERT(team->t.t_nproc <= team->t.t_max_nproc);
5482   KMP_DEBUG_ASSERT(team->t.t_threads);
5483 
5484   int use_hot_team = team == root->r.r_hot_team;
5485 #if KMP_NESTED_HOT_TEAMS
5486   int level;
5487   kmp_hot_team_ptr_t *hot_teams;
5488   if (master) {
5489     level = team->t.t_active_level - 1;
5490     if (master->th.th_teams_microtask) { // in teams construct?
5491       if (master->th.th_teams_size.nteams > 1) {
5492         ++level; // level was not increased in teams construct for
5493         // team_of_masters
5494       }
5495       if (team->t.t_pkfn != (microtask_t)__kmp_teams_master &&
5496           master->th.th_teams_level == team->t.t_level) {
5497         ++level; // level was not increased in teams construct for
5498         // team_of_workers before the parallel
5499       } // team->t.t_level will be increased inside parallel
5500     }
5501     hot_teams = master->th.th_hot_teams;
5502     if (level < __kmp_hot_teams_max_level) {
5503       KMP_DEBUG_ASSERT(team == hot_teams[level].hot_team);
5504       use_hot_team = 1;
5505     }
5506   }
5507 #endif // KMP_NESTED_HOT_TEAMS
5508 
5509   /* team is done working */
5510   TCW_SYNC_PTR(team->t.t_pkfn,
5511                NULL); // Important for Debugging Support Library.
5512 #if KMP_OS_WINDOWS
5513   team->t.t_copyin_counter = 0; // init counter for possible reuse
5514 #endif
5515   // Do not reset pointer to parent team to NULL for hot teams.
5516 
5517   /* if we are non-hot team, release our threads */
5518   if (!use_hot_team) {
5519     if (__kmp_tasking_mode != tskm_immediate_exec) {
5520       // Wait for threads to reach reapable state
5521       for (f = 1; f < team->t.t_nproc; ++f) {
5522         KMP_DEBUG_ASSERT(team->t.t_threads[f]);
5523         kmp_info_t *th = team->t.t_threads[f];
5524         volatile kmp_uint32 *state = &th->th.th_reap_state;
5525         while (*state != KMP_SAFE_TO_REAP) {
5526 #if KMP_OS_WINDOWS
5527           // On Windows a thread can be killed at any time, check this
5528           DWORD ecode;
5529           if (!__kmp_is_thread_alive(th, &ecode)) {
5530             *state = KMP_SAFE_TO_REAP; // reset the flag for dead thread
5531             break;
5532           }
5533 #endif
5534           // first check if thread is sleeping
5535           kmp_flag_64 fl(&th->th.th_bar[bs_forkjoin_barrier].bb.b_go, th);
5536           if (fl.is_sleeping())
5537             fl.resume(__kmp_gtid_from_thread(th));
5538           KMP_CPU_PAUSE();
5539         }
5540       }
5541 
5542       // Delete task teams
5543       int tt_idx;
5544       for (tt_idx = 0; tt_idx < 2; ++tt_idx) {
5545         kmp_task_team_t *task_team = team->t.t_task_team[tt_idx];
5546         if (task_team != NULL) {
5547           for (f = 0; f < team->t.t_nproc; ++f) { // threads unref task teams
5548             KMP_DEBUG_ASSERT(team->t.t_threads[f]);
5549             team->t.t_threads[f]->th.th_task_team = NULL;
5550           }
5551           KA_TRACE(
5552               20,
5553               ("__kmp_free_team: T#%d deactivating task_team %p on team %d\n",
5554                __kmp_get_gtid(), task_team, team->t.t_id));
5555 #if KMP_NESTED_HOT_TEAMS
5556           __kmp_free_task_team(master, task_team);
5557 #endif
5558           team->t.t_task_team[tt_idx] = NULL;
5559         }
5560       }
5561     }
5562 
5563     // Reset pointer to parent team only for non-hot teams.
5564     team->t.t_parent = NULL;
5565     team->t.t_level = 0;
5566     team->t.t_active_level = 0;
5567 
5568     /* free the worker threads */
5569     for (f = 1; f < team->t.t_nproc; ++f) {
5570       KMP_DEBUG_ASSERT(team->t.t_threads[f]);
5571       __kmp_free_thread(team->t.t_threads[f]);
5572       team->t.t_threads[f] = NULL;
5573     }
5574 
5575     /* put the team back in the team pool */
5576     /* TODO limit size of team pool, call reap_team if pool too large */
5577     team->t.t_next_pool = CCAST(kmp_team_t *, __kmp_team_pool);
5578     __kmp_team_pool = (volatile kmp_team_t *)team;
5579   } else { // Check if team was created for the masters in a teams construct
5580     // See if first worker is a CG root
5581     KMP_DEBUG_ASSERT(team->t.t_threads[1] &&
5582                      team->t.t_threads[1]->th.th_cg_roots);
5583     if (team->t.t_threads[1]->th.th_cg_roots->cg_root == team->t.t_threads[1]) {
5584       // Clean up the CG root nodes on workers so that this team can be re-used
5585       for (f = 1; f < team->t.t_nproc; ++f) {
5586         kmp_info_t *thr = team->t.t_threads[f];
5587         KMP_DEBUG_ASSERT(thr && thr->th.th_cg_roots &&
5588                          thr->th.th_cg_roots->cg_root == thr);
5589         // Pop current CG root off list
5590         kmp_cg_root_t *tmp = thr->th.th_cg_roots;
5591         thr->th.th_cg_roots = tmp->up;
5592         KA_TRACE(100, ("__kmp_free_team: Thread %p popping node %p and moving"
5593                        " up to node %p. cg_nthreads was %d\n",
5594                        thr, tmp, thr->th.th_cg_roots, tmp->cg_nthreads));
5595         __kmp_free(tmp);
5596         // Restore current task's thread_limit from CG root
5597         if (thr->th.th_cg_roots)
5598           thr->th.th_current_task->td_icvs.thread_limit =
5599               thr->th.th_cg_roots->cg_thread_limit;
5600       }
5601     }
5602   }
5603 
5604   KMP_MB();
5605 }
5606 
5607 /* reap the team.  destroy it, reclaim all its resources and free its memory */
5608 kmp_team_t *__kmp_reap_team(kmp_team_t *team) {
5609   kmp_team_t *next_pool = team->t.t_next_pool;
5610 
5611   KMP_DEBUG_ASSERT(team);
5612   KMP_DEBUG_ASSERT(team->t.t_dispatch);
5613   KMP_DEBUG_ASSERT(team->t.t_disp_buffer);
5614   KMP_DEBUG_ASSERT(team->t.t_threads);
5615   KMP_DEBUG_ASSERT(team->t.t_argv);
5616 
5617   /* TODO clean the threads that are a part of this? */
5618 
5619   /* free stuff */
5620   __kmp_free_team_arrays(team);
5621   if (team->t.t_argv != &team->t.t_inline_argv[0])
5622     __kmp_free((void *)team->t.t_argv);
5623   __kmp_free(team);
5624 
5625   KMP_MB();
5626   return next_pool;
5627 }
5628 
5629 // Free the thread.  Don't reap it, just place it on the pool of available
5630 // threads.
5631 //
5632 // Changes for Quad issue 527845: We need a predictable OMP tid <-> gtid
5633 // binding for the affinity mechanism to be useful.
5634 //
5635 // Now, we always keep the free list (__kmp_thread_pool) sorted by gtid.
5636 // However, we want to avoid a potential performance problem by always
5637 // scanning through the list to find the correct point at which to insert
5638 // the thread (potential N**2 behavior).  To do this we keep track of the
5639 // last place a thread struct was inserted (__kmp_thread_pool_insert_pt).
5640 // With single-level parallelism, threads will always be added to the tail
5641 // of the list, kept track of by __kmp_thread_pool_insert_pt.  With nested
5642 // parallelism, all bets are off and we may need to scan through the entire
5643 // free list.
5644 //
5645 // This change also has a potentially large performance benefit, for some
5646 // applications.  Previously, as threads were freed from the hot team, they
5647 // would be placed back on the free list in inverse order.  If the hot team
5648 // grew back to it's original size, then the freed thread would be placed
5649 // back on the hot team in reverse order.  This could cause bad cache
5650 // locality problems on programs where the size of the hot team regularly
5651 // grew and shrunk.
5652 //
5653 // Now, for single-level parallelism, the OMP tid is alway == gtid.
5654 void __kmp_free_thread(kmp_info_t *this_th) {
5655   int gtid;
5656   kmp_info_t **scan;
5657 
5658   KA_TRACE(20, ("__kmp_free_thread: T#%d putting T#%d back on free pool.\n",
5659                 __kmp_get_gtid(), this_th->th.th_info.ds.ds_gtid));
5660 
5661   KMP_DEBUG_ASSERT(this_th);
5662 
5663   // When moving thread to pool, switch thread to wait on own b_go flag, and
5664   // uninitialized (NULL team).
5665   int b;
5666   kmp_balign_t *balign = this_th->th.th_bar;
5667   for (b = 0; b < bs_last_barrier; ++b) {
5668     if (balign[b].bb.wait_flag == KMP_BARRIER_PARENT_FLAG)
5669       balign[b].bb.wait_flag = KMP_BARRIER_SWITCH_TO_OWN_FLAG;
5670     balign[b].bb.team = NULL;
5671     balign[b].bb.leaf_kids = 0;
5672   }
5673   this_th->th.th_task_state = 0;
5674   this_th->th.th_reap_state = KMP_SAFE_TO_REAP;
5675 
5676   /* put thread back on the free pool */
5677   TCW_PTR(this_th->th.th_team, NULL);
5678   TCW_PTR(this_th->th.th_root, NULL);
5679   TCW_PTR(this_th->th.th_dispatch, NULL); /* NOT NEEDED */
5680 
5681   while (this_th->th.th_cg_roots) {
5682     this_th->th.th_cg_roots->cg_nthreads--;
5683     KA_TRACE(100, ("__kmp_free_thread: Thread %p decrement cg_nthreads on node"
5684                    " %p of thread  %p to %d\n",
5685                    this_th, this_th->th.th_cg_roots,
5686                    this_th->th.th_cg_roots->cg_root,
5687                    this_th->th.th_cg_roots->cg_nthreads));
5688     kmp_cg_root_t *tmp = this_th->th.th_cg_roots;
5689     if (tmp->cg_root == this_th) { // Thread is a cg_root
5690       KMP_DEBUG_ASSERT(tmp->cg_nthreads == 0);
5691       KA_TRACE(
5692           5, ("__kmp_free_thread: Thread %p freeing node %p\n", this_th, tmp));
5693       this_th->th.th_cg_roots = tmp->up;
5694       __kmp_free(tmp);
5695     } else { // Worker thread
5696       this_th->th.th_cg_roots = NULL;
5697       break;
5698     }
5699   }
5700 
5701   /* If the implicit task assigned to this thread can be used by other threads
5702    * -> multiple threads can share the data and try to free the task at
5703    * __kmp_reap_thread at exit. This duplicate use of the task data can happen
5704    * with higher probability when hot team is disabled but can occurs even when
5705    * the hot team is enabled */
5706   __kmp_free_implicit_task(this_th);
5707   this_th->th.th_current_task = NULL;
5708 
5709   // If the __kmp_thread_pool_insert_pt is already past the new insert
5710   // point, then we need to re-scan the entire list.
5711   gtid = this_th->th.th_info.ds.ds_gtid;
5712   if (__kmp_thread_pool_insert_pt != NULL) {
5713     KMP_DEBUG_ASSERT(__kmp_thread_pool != NULL);
5714     if (__kmp_thread_pool_insert_pt->th.th_info.ds.ds_gtid > gtid) {
5715       __kmp_thread_pool_insert_pt = NULL;
5716     }
5717   }
5718 
5719   // Scan down the list to find the place to insert the thread.
5720   // scan is the address of a link in the list, possibly the address of
5721   // __kmp_thread_pool itself.
5722   //
5723   // In the absence of nested parallism, the for loop will have 0 iterations.
5724   if (__kmp_thread_pool_insert_pt != NULL) {
5725     scan = &(__kmp_thread_pool_insert_pt->th.th_next_pool);
5726   } else {
5727     scan = CCAST(kmp_info_t **, &__kmp_thread_pool);
5728   }
5729   for (; (*scan != NULL) && ((*scan)->th.th_info.ds.ds_gtid < gtid);
5730        scan = &((*scan)->th.th_next_pool))
5731     ;
5732 
5733   // Insert the new element on the list, and set __kmp_thread_pool_insert_pt
5734   // to its address.
5735   TCW_PTR(this_th->th.th_next_pool, *scan);
5736   __kmp_thread_pool_insert_pt = *scan = this_th;
5737   KMP_DEBUG_ASSERT((this_th->th.th_next_pool == NULL) ||
5738                    (this_th->th.th_info.ds.ds_gtid <
5739                     this_th->th.th_next_pool->th.th_info.ds.ds_gtid));
5740   TCW_4(this_th->th.th_in_pool, TRUE);
5741   __kmp_suspend_initialize_thread(this_th);
5742   __kmp_lock_suspend_mx(this_th);
5743   if (this_th->th.th_active == TRUE) {
5744     KMP_ATOMIC_INC(&__kmp_thread_pool_active_nth);
5745     this_th->th.th_active_in_pool = TRUE;
5746   }
5747 #if KMP_DEBUG
5748   else {
5749     KMP_DEBUG_ASSERT(this_th->th.th_active_in_pool == FALSE);
5750   }
5751 #endif
5752   __kmp_unlock_suspend_mx(this_th);
5753 
5754   TCW_4(__kmp_nth, __kmp_nth - 1);
5755 
5756 #ifdef KMP_ADJUST_BLOCKTIME
5757   /* Adjust blocktime back to user setting or default if necessary */
5758   /* Middle initialization might never have occurred                */
5759   if (!__kmp_env_blocktime && (__kmp_avail_proc > 0)) {
5760     KMP_DEBUG_ASSERT(__kmp_avail_proc > 0);
5761     if (__kmp_nth <= __kmp_avail_proc) {
5762       __kmp_zero_bt = FALSE;
5763     }
5764   }
5765 #endif /* KMP_ADJUST_BLOCKTIME */
5766 
5767   KMP_MB();
5768 }
5769 
5770 /* ------------------------------------------------------------------------ */
5771 
5772 void *__kmp_launch_thread(kmp_info_t *this_thr) {
5773   int gtid = this_thr->th.th_info.ds.ds_gtid;
5774   /*    void                 *stack_data;*/
5775   kmp_team_t *(*volatile pteam);
5776 
5777   KMP_MB();
5778   KA_TRACE(10, ("__kmp_launch_thread: T#%d start\n", gtid));
5779 
5780   if (__kmp_env_consistency_check) {
5781     this_thr->th.th_cons = __kmp_allocate_cons_stack(gtid); // ATT: Memory leak?
5782   }
5783 
5784 #if OMPT_SUPPORT
5785   ompt_data_t *thread_data;
5786   if (ompt_enabled.enabled) {
5787     thread_data = &(this_thr->th.ompt_thread_info.thread_data);
5788     *thread_data = ompt_data_none;
5789 
5790     this_thr->th.ompt_thread_info.state = ompt_state_overhead;
5791     this_thr->th.ompt_thread_info.wait_id = 0;
5792     this_thr->th.ompt_thread_info.idle_frame = OMPT_GET_FRAME_ADDRESS(0);
5793     if (ompt_enabled.ompt_callback_thread_begin) {
5794       ompt_callbacks.ompt_callback(ompt_callback_thread_begin)(
5795           ompt_thread_worker, thread_data);
5796     }
5797   }
5798 #endif
5799 
5800 #if OMPT_SUPPORT
5801   if (ompt_enabled.enabled) {
5802     this_thr->th.ompt_thread_info.state = ompt_state_idle;
5803   }
5804 #endif
5805   /* This is the place where threads wait for work */
5806   while (!TCR_4(__kmp_global.g.g_done)) {
5807     KMP_DEBUG_ASSERT(this_thr == __kmp_threads[gtid]);
5808     KMP_MB();
5809 
5810     /* wait for work to do */
5811     KA_TRACE(20, ("__kmp_launch_thread: T#%d waiting for work\n", gtid));
5812 
5813     /* No tid yet since not part of a team */
5814     __kmp_fork_barrier(gtid, KMP_GTID_DNE);
5815 
5816 #if OMPT_SUPPORT
5817     if (ompt_enabled.enabled) {
5818       this_thr->th.ompt_thread_info.state = ompt_state_overhead;
5819     }
5820 #endif
5821 
5822     pteam = (kmp_team_t * (*))(&this_thr->th.th_team);
5823 
5824     /* have we been allocated? */
5825     if (TCR_SYNC_PTR(*pteam) && !TCR_4(__kmp_global.g.g_done)) {
5826       /* we were just woken up, so run our new task */
5827       if (TCR_SYNC_PTR((*pteam)->t.t_pkfn) != NULL) {
5828         int rc;
5829         KA_TRACE(20,
5830                  ("__kmp_launch_thread: T#%d(%d:%d) invoke microtask = %p\n",
5831                   gtid, (*pteam)->t.t_id, __kmp_tid_from_gtid(gtid),
5832                   (*pteam)->t.t_pkfn));
5833 
5834         updateHWFPControl(*pteam);
5835 
5836 #if OMPT_SUPPORT
5837         if (ompt_enabled.enabled) {
5838           this_thr->th.ompt_thread_info.state = ompt_state_work_parallel;
5839         }
5840 #endif
5841 
5842         rc = (*pteam)->t.t_invoke(gtid);
5843         KMP_ASSERT(rc);
5844 
5845         KMP_MB();
5846         KA_TRACE(20, ("__kmp_launch_thread: T#%d(%d:%d) done microtask = %p\n",
5847                       gtid, (*pteam)->t.t_id, __kmp_tid_from_gtid(gtid),
5848                       (*pteam)->t.t_pkfn));
5849       }
5850 #if OMPT_SUPPORT
5851       if (ompt_enabled.enabled) {
5852         /* no frame set while outside task */
5853         __ompt_get_task_info_object(0)->frame.exit_frame = ompt_data_none;
5854 
5855         this_thr->th.ompt_thread_info.state = ompt_state_overhead;
5856       }
5857 #endif
5858       /* join barrier after parallel region */
5859       __kmp_join_barrier(gtid);
5860     }
5861   }
5862   TCR_SYNC_PTR((intptr_t)__kmp_global.g.g_done);
5863 
5864 #if OMPT_SUPPORT
5865   if (ompt_enabled.ompt_callback_thread_end) {
5866     ompt_callbacks.ompt_callback(ompt_callback_thread_end)(thread_data);
5867   }
5868 #endif
5869 
5870   this_thr->th.th_task_team = NULL;
5871   /* run the destructors for the threadprivate data for this thread */
5872   __kmp_common_destroy_gtid(gtid);
5873 
5874   KA_TRACE(10, ("__kmp_launch_thread: T#%d done\n", gtid));
5875   KMP_MB();
5876   return this_thr;
5877 }
5878 
5879 /* ------------------------------------------------------------------------ */
5880 
5881 void __kmp_internal_end_dest(void *specific_gtid) {
5882 #if KMP_COMPILER_ICC
5883 #pragma warning(push)
5884 #pragma warning(disable : 810) // conversion from "void *" to "int" may lose
5885 // significant bits
5886 #endif
5887   // Make sure no significant bits are lost
5888   int gtid = (kmp_intptr_t)specific_gtid - 1;
5889 #if KMP_COMPILER_ICC
5890 #pragma warning(pop)
5891 #endif
5892 
5893   KA_TRACE(30, ("__kmp_internal_end_dest: T#%d\n", gtid));
5894   /* NOTE: the gtid is stored as gitd+1 in the thread-local-storage
5895    * this is because 0 is reserved for the nothing-stored case */
5896 
5897   /* josh: One reason for setting the gtid specific data even when it is being
5898      destroyed by pthread is to allow gtid lookup through thread specific data
5899      (__kmp_gtid_get_specific).  Some of the code, especially stat code,
5900      that gets executed in the call to __kmp_internal_end_thread, actually
5901      gets the gtid through the thread specific data.  Setting it here seems
5902      rather inelegant and perhaps wrong, but allows __kmp_internal_end_thread
5903      to run smoothly.
5904      todo: get rid of this after we remove the dependence on
5905      __kmp_gtid_get_specific  */
5906   if (gtid >= 0 && KMP_UBER_GTID(gtid))
5907     __kmp_gtid_set_specific(gtid);
5908 #ifdef KMP_TDATA_GTID
5909   __kmp_gtid = gtid;
5910 #endif
5911   __kmp_internal_end_thread(gtid);
5912 }
5913 
5914 #if KMP_OS_UNIX && KMP_DYNAMIC_LIB
5915 
5916 // 2009-09-08 (lev): It looks the destructor does not work. In simple test cases
5917 // destructors work perfectly, but in real libomp.so I have no evidence it is
5918 // ever called. However, -fini linker option in makefile.mk works fine.
5919 
5920 __attribute__((destructor)) void __kmp_internal_end_dtor(void) {
5921   __kmp_internal_end_atexit();
5922 }
5923 
5924 void __kmp_internal_end_fini(void) { __kmp_internal_end_atexit(); }
5925 
5926 #endif
5927 
5928 /* [Windows] josh: when the atexit handler is called, there may still be more
5929    than one thread alive */
5930 void __kmp_internal_end_atexit(void) {
5931   KA_TRACE(30, ("__kmp_internal_end_atexit\n"));
5932   /* [Windows]
5933      josh: ideally, we want to completely shutdown the library in this atexit
5934      handler, but stat code that depends on thread specific data for gtid fails
5935      because that data becomes unavailable at some point during the shutdown, so
5936      we call __kmp_internal_end_thread instead. We should eventually remove the
5937      dependency on __kmp_get_specific_gtid in the stat code and use
5938      __kmp_internal_end_library to cleanly shutdown the library.
5939 
5940      // TODO: Can some of this comment about GVS be removed?
5941      I suspect that the offending stat code is executed when the calling thread
5942      tries to clean up a dead root thread's data structures, resulting in GVS
5943      code trying to close the GVS structures for that thread, but since the stat
5944      code uses __kmp_get_specific_gtid to get the gtid with the assumption that
5945      the calling thread is cleaning up itself instead of another thread, it get
5946      confused. This happens because allowing a thread to unregister and cleanup
5947      another thread is a recent modification for addressing an issue.
5948      Based on the current design (20050722), a thread may end up
5949      trying to unregister another thread only if thread death does not trigger
5950      the calling of __kmp_internal_end_thread.  For Linux* OS, there is the
5951      thread specific data destructor function to detect thread death. For
5952      Windows dynamic, there is DllMain(THREAD_DETACH). For Windows static, there
5953      is nothing.  Thus, the workaround is applicable only for Windows static
5954      stat library. */
5955   __kmp_internal_end_library(-1);
5956 #if KMP_OS_WINDOWS
5957   __kmp_close_console();
5958 #endif
5959 }
5960 
5961 static void __kmp_reap_thread(kmp_info_t *thread, int is_root) {
5962   // It is assumed __kmp_forkjoin_lock is acquired.
5963 
5964   int gtid;
5965 
5966   KMP_DEBUG_ASSERT(thread != NULL);
5967 
5968   gtid = thread->th.th_info.ds.ds_gtid;
5969 
5970   if (!is_root) {
5971     if (__kmp_dflt_blocktime != KMP_MAX_BLOCKTIME) {
5972       /* Assume the threads are at the fork barrier here */
5973       KA_TRACE(
5974           20, ("__kmp_reap_thread: releasing T#%d from fork barrier for reap\n",
5975                gtid));
5976       /* Need release fence here to prevent seg faults for tree forkjoin barrier
5977        * (GEH) */
5978       ANNOTATE_HAPPENS_BEFORE(thread);
5979       kmp_flag_64 flag(&thread->th.th_bar[bs_forkjoin_barrier].bb.b_go, thread);
5980       __kmp_release_64(&flag);
5981     }
5982 
5983     // Terminate OS thread.
5984     __kmp_reap_worker(thread);
5985 
5986     // The thread was killed asynchronously.  If it was actively
5987     // spinning in the thread pool, decrement the global count.
5988     //
5989     // There is a small timing hole here - if the worker thread was just waking
5990     // up after sleeping in the pool, had reset it's th_active_in_pool flag but
5991     // not decremented the global counter __kmp_thread_pool_active_nth yet, then
5992     // the global counter might not get updated.
5993     //
5994     // Currently, this can only happen as the library is unloaded,
5995     // so there are no harmful side effects.
5996     if (thread->th.th_active_in_pool) {
5997       thread->th.th_active_in_pool = FALSE;
5998       KMP_ATOMIC_DEC(&__kmp_thread_pool_active_nth);
5999       KMP_DEBUG_ASSERT(__kmp_thread_pool_active_nth >= 0);
6000     }
6001   }
6002 
6003   __kmp_free_implicit_task(thread);
6004 
6005 // Free the fast memory for tasking
6006 #if USE_FAST_MEMORY
6007   __kmp_free_fast_memory(thread);
6008 #endif /* USE_FAST_MEMORY */
6009 
6010   __kmp_suspend_uninitialize_thread(thread);
6011 
6012   KMP_DEBUG_ASSERT(__kmp_threads[gtid] == thread);
6013   TCW_SYNC_PTR(__kmp_threads[gtid], NULL);
6014 
6015   --__kmp_all_nth;
6016 // __kmp_nth was decremented when thread is added to the pool.
6017 
6018 #ifdef KMP_ADJUST_BLOCKTIME
6019   /* Adjust blocktime back to user setting or default if necessary */
6020   /* Middle initialization might never have occurred                */
6021   if (!__kmp_env_blocktime && (__kmp_avail_proc > 0)) {
6022     KMP_DEBUG_ASSERT(__kmp_avail_proc > 0);
6023     if (__kmp_nth <= __kmp_avail_proc) {
6024       __kmp_zero_bt = FALSE;
6025     }
6026   }
6027 #endif /* KMP_ADJUST_BLOCKTIME */
6028 
6029   /* free the memory being used */
6030   if (__kmp_env_consistency_check) {
6031     if (thread->th.th_cons) {
6032       __kmp_free_cons_stack(thread->th.th_cons);
6033       thread->th.th_cons = NULL;
6034     }
6035   }
6036 
6037   if (thread->th.th_pri_common != NULL) {
6038     __kmp_free(thread->th.th_pri_common);
6039     thread->th.th_pri_common = NULL;
6040   }
6041 
6042   if (thread->th.th_task_state_memo_stack != NULL) {
6043     __kmp_free(thread->th.th_task_state_memo_stack);
6044     thread->th.th_task_state_memo_stack = NULL;
6045   }
6046 
6047 #if KMP_USE_BGET
6048   if (thread->th.th_local.bget_data != NULL) {
6049     __kmp_finalize_bget(thread);
6050   }
6051 #endif
6052 
6053 #if KMP_AFFINITY_SUPPORTED
6054   if (thread->th.th_affin_mask != NULL) {
6055     KMP_CPU_FREE(thread->th.th_affin_mask);
6056     thread->th.th_affin_mask = NULL;
6057   }
6058 #endif /* KMP_AFFINITY_SUPPORTED */
6059 
6060 #if KMP_USE_HIER_SCHED
6061   if (thread->th.th_hier_bar_data != NULL) {
6062     __kmp_free(thread->th.th_hier_bar_data);
6063     thread->th.th_hier_bar_data = NULL;
6064   }
6065 #endif
6066 
6067   __kmp_reap_team(thread->th.th_serial_team);
6068   thread->th.th_serial_team = NULL;
6069   __kmp_free(thread);
6070 
6071   KMP_MB();
6072 
6073 } // __kmp_reap_thread
6074 
6075 static void __kmp_internal_end(void) {
6076   int i;
6077 
6078   /* First, unregister the library */
6079   __kmp_unregister_library();
6080 
6081 #if KMP_OS_WINDOWS
6082   /* In Win static library, we can't tell when a root actually dies, so we
6083      reclaim the data structures for any root threads that have died but not
6084      unregistered themselves, in order to shut down cleanly.
6085      In Win dynamic library we also can't tell when a thread dies.  */
6086   __kmp_reclaim_dead_roots(); // AC: moved here to always clean resources of
6087 // dead roots
6088 #endif
6089 
6090   for (i = 0; i < __kmp_threads_capacity; i++)
6091     if (__kmp_root[i])
6092       if (__kmp_root[i]->r.r_active)
6093         break;
6094   KMP_MB(); /* Flush all pending memory write invalidates.  */
6095   TCW_SYNC_4(__kmp_global.g.g_done, TRUE);
6096 
6097   if (i < __kmp_threads_capacity) {
6098 #if KMP_USE_MONITOR
6099     // 2009-09-08 (lev): Other alive roots found. Why do we kill the monitor??
6100     KMP_MB(); /* Flush all pending memory write invalidates.  */
6101 
6102     // Need to check that monitor was initialized before reaping it. If we are
6103     // called form __kmp_atfork_child (which sets __kmp_init_parallel = 0), then
6104     // __kmp_monitor will appear to contain valid data, but it is only valid in
6105     // the parent process, not the child.
6106     // New behavior (201008): instead of keying off of the flag
6107     // __kmp_init_parallel, the monitor thread creation is keyed off
6108     // of the new flag __kmp_init_monitor.
6109     __kmp_acquire_bootstrap_lock(&__kmp_monitor_lock);
6110     if (TCR_4(__kmp_init_monitor)) {
6111       __kmp_reap_monitor(&__kmp_monitor);
6112       TCW_4(__kmp_init_monitor, 0);
6113     }
6114     __kmp_release_bootstrap_lock(&__kmp_monitor_lock);
6115     KA_TRACE(10, ("__kmp_internal_end: monitor reaped\n"));
6116 #endif // KMP_USE_MONITOR
6117   } else {
6118 /* TODO move this to cleanup code */
6119 #ifdef KMP_DEBUG
6120     /* make sure that everything has properly ended */
6121     for (i = 0; i < __kmp_threads_capacity; i++) {
6122       if (__kmp_root[i]) {
6123         //                    KMP_ASSERT( ! KMP_UBER_GTID( i ) );         // AC:
6124         //                    there can be uber threads alive here
6125         KMP_ASSERT(!__kmp_root[i]->r.r_active); // TODO: can they be active?
6126       }
6127     }
6128 #endif
6129 
6130     KMP_MB();
6131 
6132     // Reap the worker threads.
6133     // This is valid for now, but be careful if threads are reaped sooner.
6134     while (__kmp_thread_pool != NULL) { // Loop thru all the thread in the pool.
6135       // Get the next thread from the pool.
6136       kmp_info_t *thread = CCAST(kmp_info_t *, __kmp_thread_pool);
6137       __kmp_thread_pool = thread->th.th_next_pool;
6138       // Reap it.
6139       KMP_DEBUG_ASSERT(thread->th.th_reap_state == KMP_SAFE_TO_REAP);
6140       thread->th.th_next_pool = NULL;
6141       thread->th.th_in_pool = FALSE;
6142       __kmp_reap_thread(thread, 0);
6143     }
6144     __kmp_thread_pool_insert_pt = NULL;
6145 
6146     // Reap teams.
6147     while (__kmp_team_pool != NULL) { // Loop thru all the teams in the pool.
6148       // Get the next team from the pool.
6149       kmp_team_t *team = CCAST(kmp_team_t *, __kmp_team_pool);
6150       __kmp_team_pool = team->t.t_next_pool;
6151       // Reap it.
6152       team->t.t_next_pool = NULL;
6153       __kmp_reap_team(team);
6154     }
6155 
6156     __kmp_reap_task_teams();
6157 
6158 #if KMP_OS_UNIX
6159     // Threads that are not reaped should not access any resources since they
6160     // are going to be deallocated soon, so the shutdown sequence should wait
6161     // until all threads either exit the final spin-waiting loop or begin
6162     // sleeping after the given blocktime.
6163     for (i = 0; i < __kmp_threads_capacity; i++) {
6164       kmp_info_t *thr = __kmp_threads[i];
6165       while (thr && KMP_ATOMIC_LD_ACQ(&thr->th.th_blocking))
6166         KMP_CPU_PAUSE();
6167     }
6168 #endif
6169 
6170     for (i = 0; i < __kmp_threads_capacity; ++i) {
6171       // TBD: Add some checking...
6172       // Something like KMP_DEBUG_ASSERT( __kmp_thread[ i ] == NULL );
6173     }
6174 
6175     /* Make sure all threadprivate destructors get run by joining with all
6176        worker threads before resetting this flag */
6177     TCW_SYNC_4(__kmp_init_common, FALSE);
6178 
6179     KA_TRACE(10, ("__kmp_internal_end: all workers reaped\n"));
6180     KMP_MB();
6181 
6182 #if KMP_USE_MONITOR
6183     // See note above: One of the possible fixes for CQ138434 / CQ140126
6184     //
6185     // FIXME: push both code fragments down and CSE them?
6186     // push them into __kmp_cleanup() ?
6187     __kmp_acquire_bootstrap_lock(&__kmp_monitor_lock);
6188     if (TCR_4(__kmp_init_monitor)) {
6189       __kmp_reap_monitor(&__kmp_monitor);
6190       TCW_4(__kmp_init_monitor, 0);
6191     }
6192     __kmp_release_bootstrap_lock(&__kmp_monitor_lock);
6193     KA_TRACE(10, ("__kmp_internal_end: monitor reaped\n"));
6194 #endif
6195   } /* else !__kmp_global.t_active */
6196   TCW_4(__kmp_init_gtid, FALSE);
6197   KMP_MB(); /* Flush all pending memory write invalidates.  */
6198 
6199   __kmp_cleanup();
6200 #if OMPT_SUPPORT
6201   ompt_fini();
6202 #endif
6203 }
6204 
6205 void __kmp_internal_end_library(int gtid_req) {
6206   /* if we have already cleaned up, don't try again, it wouldn't be pretty */
6207   /* this shouldn't be a race condition because __kmp_internal_end() is the
6208      only place to clear __kmp_serial_init */
6209   /* we'll check this later too, after we get the lock */
6210   // 2009-09-06: We do not set g_abort without setting g_done. This check looks
6211   // redundaant, because the next check will work in any case.
6212   if (__kmp_global.g.g_abort) {
6213     KA_TRACE(11, ("__kmp_internal_end_library: abort, exiting\n"));
6214     /* TODO abort? */
6215     return;
6216   }
6217   if (TCR_4(__kmp_global.g.g_done) || !__kmp_init_serial) {
6218     KA_TRACE(10, ("__kmp_internal_end_library: already finished\n"));
6219     return;
6220   }
6221 
6222   KMP_MB(); /* Flush all pending memory write invalidates.  */
6223 
6224   /* find out who we are and what we should do */
6225   {
6226     int gtid = (gtid_req >= 0) ? gtid_req : __kmp_gtid_get_specific();
6227     KA_TRACE(
6228         10, ("__kmp_internal_end_library: enter T#%d  (%d)\n", gtid, gtid_req));
6229     if (gtid == KMP_GTID_SHUTDOWN) {
6230       KA_TRACE(10, ("__kmp_internal_end_library: !__kmp_init_runtime, system "
6231                     "already shutdown\n"));
6232       return;
6233     } else if (gtid == KMP_GTID_MONITOR) {
6234       KA_TRACE(10, ("__kmp_internal_end_library: monitor thread, gtid not "
6235                     "registered, or system shutdown\n"));
6236       return;
6237     } else if (gtid == KMP_GTID_DNE) {
6238       KA_TRACE(10, ("__kmp_internal_end_library: gtid not registered or system "
6239                     "shutdown\n"));
6240       /* we don't know who we are, but we may still shutdown the library */
6241     } else if (KMP_UBER_GTID(gtid)) {
6242       /* unregister ourselves as an uber thread.  gtid is no longer valid */
6243       if (__kmp_root[gtid]->r.r_active) {
6244         __kmp_global.g.g_abort = -1;
6245         TCW_SYNC_4(__kmp_global.g.g_done, TRUE);
6246         KA_TRACE(10,
6247                  ("__kmp_internal_end_library: root still active, abort T#%d\n",
6248                   gtid));
6249         return;
6250       } else {
6251         KA_TRACE(
6252             10,
6253             ("__kmp_internal_end_library: unregistering sibling T#%d\n", gtid));
6254         __kmp_unregister_root_current_thread(gtid);
6255       }
6256     } else {
6257 /* worker threads may call this function through the atexit handler, if they
6258  * call exit() */
6259 /* For now, skip the usual subsequent processing and just dump the debug buffer.
6260    TODO: do a thorough shutdown instead */
6261 #ifdef DUMP_DEBUG_ON_EXIT
6262       if (__kmp_debug_buf)
6263         __kmp_dump_debug_buffer();
6264 #endif
6265       return;
6266     }
6267   }
6268   /* synchronize the termination process */
6269   __kmp_acquire_bootstrap_lock(&__kmp_initz_lock);
6270 
6271   /* have we already finished */
6272   if (__kmp_global.g.g_abort) {
6273     KA_TRACE(10, ("__kmp_internal_end_library: abort, exiting\n"));
6274     /* TODO abort? */
6275     __kmp_release_bootstrap_lock(&__kmp_initz_lock);
6276     return;
6277   }
6278   if (TCR_4(__kmp_global.g.g_done) || !__kmp_init_serial) {
6279     __kmp_release_bootstrap_lock(&__kmp_initz_lock);
6280     return;
6281   }
6282 
6283   /* We need this lock to enforce mutex between this reading of
6284      __kmp_threads_capacity and the writing by __kmp_register_root.
6285      Alternatively, we can use a counter of roots that is atomically updated by
6286      __kmp_get_global_thread_id_reg, __kmp_do_serial_initialize and
6287      __kmp_internal_end_*.  */
6288   __kmp_acquire_bootstrap_lock(&__kmp_forkjoin_lock);
6289 
6290   /* now we can safely conduct the actual termination */
6291   __kmp_internal_end();
6292 
6293   __kmp_release_bootstrap_lock(&__kmp_forkjoin_lock);
6294   __kmp_release_bootstrap_lock(&__kmp_initz_lock);
6295 
6296   KA_TRACE(10, ("__kmp_internal_end_library: exit\n"));
6297 
6298 #ifdef DUMP_DEBUG_ON_EXIT
6299   if (__kmp_debug_buf)
6300     __kmp_dump_debug_buffer();
6301 #endif
6302 
6303 #if KMP_OS_WINDOWS
6304   __kmp_close_console();
6305 #endif
6306 
6307   __kmp_fini_allocator();
6308 
6309 } // __kmp_internal_end_library
6310 
6311 void __kmp_internal_end_thread(int gtid_req) {
6312   int i;
6313 
6314   /* if we have already cleaned up, don't try again, it wouldn't be pretty */
6315   /* this shouldn't be a race condition because __kmp_internal_end() is the
6316    * only place to clear __kmp_serial_init */
6317   /* we'll check this later too, after we get the lock */
6318   // 2009-09-06: We do not set g_abort without setting g_done. This check looks
6319   // redundant, because the next check will work in any case.
6320   if (__kmp_global.g.g_abort) {
6321     KA_TRACE(11, ("__kmp_internal_end_thread: abort, exiting\n"));
6322     /* TODO abort? */
6323     return;
6324   }
6325   if (TCR_4(__kmp_global.g.g_done) || !__kmp_init_serial) {
6326     KA_TRACE(10, ("__kmp_internal_end_thread: already finished\n"));
6327     return;
6328   }
6329 
6330   KMP_MB(); /* Flush all pending memory write invalidates.  */
6331 
6332   /* find out who we are and what we should do */
6333   {
6334     int gtid = (gtid_req >= 0) ? gtid_req : __kmp_gtid_get_specific();
6335     KA_TRACE(10,
6336              ("__kmp_internal_end_thread: enter T#%d  (%d)\n", gtid, gtid_req));
6337     if (gtid == KMP_GTID_SHUTDOWN) {
6338       KA_TRACE(10, ("__kmp_internal_end_thread: !__kmp_init_runtime, system "
6339                     "already shutdown\n"));
6340       return;
6341     } else if (gtid == KMP_GTID_MONITOR) {
6342       KA_TRACE(10, ("__kmp_internal_end_thread: monitor thread, gtid not "
6343                     "registered, or system shutdown\n"));
6344       return;
6345     } else if (gtid == KMP_GTID_DNE) {
6346       KA_TRACE(10, ("__kmp_internal_end_thread: gtid not registered or system "
6347                     "shutdown\n"));
6348       return;
6349       /* we don't know who we are */
6350     } else if (KMP_UBER_GTID(gtid)) {
6351       /* unregister ourselves as an uber thread.  gtid is no longer valid */
6352       if (__kmp_root[gtid]->r.r_active) {
6353         __kmp_global.g.g_abort = -1;
6354         TCW_SYNC_4(__kmp_global.g.g_done, TRUE);
6355         KA_TRACE(10,
6356                  ("__kmp_internal_end_thread: root still active, abort T#%d\n",
6357                   gtid));
6358         return;
6359       } else {
6360         KA_TRACE(10, ("__kmp_internal_end_thread: unregistering sibling T#%d\n",
6361                       gtid));
6362         __kmp_unregister_root_current_thread(gtid);
6363       }
6364     } else {
6365       /* just a worker thread, let's leave */
6366       KA_TRACE(10, ("__kmp_internal_end_thread: worker thread T#%d\n", gtid));
6367 
6368       if (gtid >= 0) {
6369         __kmp_threads[gtid]->th.th_task_team = NULL;
6370       }
6371 
6372       KA_TRACE(10,
6373                ("__kmp_internal_end_thread: worker thread done, exiting T#%d\n",
6374                 gtid));
6375       return;
6376     }
6377   }
6378 #if KMP_DYNAMIC_LIB
6379 #if OMP_50_ENABLED
6380   if (__kmp_pause_status != kmp_hard_paused)
6381 #endif
6382   // AC: lets not shutdown the dynamic library at the exit of uber thread,
6383   // because we will better shutdown later in the library destructor.
6384   {
6385     KA_TRACE(10, ("__kmp_internal_end_thread: exiting T#%d\n", gtid_req));
6386     return;
6387   }
6388 #endif
6389   /* synchronize the termination process */
6390   __kmp_acquire_bootstrap_lock(&__kmp_initz_lock);
6391 
6392   /* have we already finished */
6393   if (__kmp_global.g.g_abort) {
6394     KA_TRACE(10, ("__kmp_internal_end_thread: abort, exiting\n"));
6395     /* TODO abort? */
6396     __kmp_release_bootstrap_lock(&__kmp_initz_lock);
6397     return;
6398   }
6399   if (TCR_4(__kmp_global.g.g_done) || !__kmp_init_serial) {
6400     __kmp_release_bootstrap_lock(&__kmp_initz_lock);
6401     return;
6402   }
6403 
6404   /* We need this lock to enforce mutex between this reading of
6405      __kmp_threads_capacity and the writing by __kmp_register_root.
6406      Alternatively, we can use a counter of roots that is atomically updated by
6407      __kmp_get_global_thread_id_reg, __kmp_do_serial_initialize and
6408      __kmp_internal_end_*.  */
6409 
6410   /* should we finish the run-time?  are all siblings done? */
6411   __kmp_acquire_bootstrap_lock(&__kmp_forkjoin_lock);
6412 
6413   for (i = 0; i < __kmp_threads_capacity; ++i) {
6414     if (KMP_UBER_GTID(i)) {
6415       KA_TRACE(
6416           10,
6417           ("__kmp_internal_end_thread: remaining sibling task: gtid==%d\n", i));
6418       __kmp_release_bootstrap_lock(&__kmp_forkjoin_lock);
6419       __kmp_release_bootstrap_lock(&__kmp_initz_lock);
6420       return;
6421     }
6422   }
6423 
6424   /* now we can safely conduct the actual termination */
6425 
6426   __kmp_internal_end();
6427 
6428   __kmp_release_bootstrap_lock(&__kmp_forkjoin_lock);
6429   __kmp_release_bootstrap_lock(&__kmp_initz_lock);
6430 
6431   KA_TRACE(10, ("__kmp_internal_end_thread: exit T#%d\n", gtid_req));
6432 
6433 #ifdef DUMP_DEBUG_ON_EXIT
6434   if (__kmp_debug_buf)
6435     __kmp_dump_debug_buffer();
6436 #endif
6437 } // __kmp_internal_end_thread
6438 
6439 // -----------------------------------------------------------------------------
6440 // Library registration stuff.
6441 
6442 static long __kmp_registration_flag = 0;
6443 // Random value used to indicate library initialization.
6444 static char *__kmp_registration_str = NULL;
6445 // Value to be saved in env var __KMP_REGISTERED_LIB_<pid>.
6446 
6447 static inline char *__kmp_reg_status_name() {
6448   /* On RHEL 3u5 if linked statically, getpid() returns different values in
6449      each thread. If registration and unregistration go in different threads
6450      (omp_misc_other_root_exit.cpp test case), the name of registered_lib_env
6451      env var can not be found, because the name will contain different pid. */
6452   return __kmp_str_format("__KMP_REGISTERED_LIB_%d", (int)getpid());
6453 } // __kmp_reg_status_get
6454 
6455 void __kmp_register_library_startup(void) {
6456 
6457   char *name = __kmp_reg_status_name(); // Name of the environment variable.
6458   int done = 0;
6459   union {
6460     double dtime;
6461     long ltime;
6462   } time;
6463 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
6464   __kmp_initialize_system_tick();
6465 #endif
6466   __kmp_read_system_time(&time.dtime);
6467   __kmp_registration_flag = 0xCAFE0000L | (time.ltime & 0x0000FFFFL);
6468   __kmp_registration_str =
6469       __kmp_str_format("%p-%lx-%s", &__kmp_registration_flag,
6470                        __kmp_registration_flag, KMP_LIBRARY_FILE);
6471 
6472   KA_TRACE(50, ("__kmp_register_library_startup: %s=\"%s\"\n", name,
6473                 __kmp_registration_str));
6474 
6475   while (!done) {
6476 
6477     char *value = NULL; // Actual value of the environment variable.
6478 
6479     // Set environment variable, but do not overwrite if it is exist.
6480     __kmp_env_set(name, __kmp_registration_str, 0);
6481     // Check the variable is written.
6482     value = __kmp_env_get(name);
6483     if (value != NULL && strcmp(value, __kmp_registration_str) == 0) {
6484 
6485       done = 1; // Ok, environment variable set successfully, exit the loop.
6486 
6487     } else {
6488 
6489       // Oops. Write failed. Another copy of OpenMP RTL is in memory.
6490       // Check whether it alive or dead.
6491       int neighbor = 0; // 0 -- unknown status, 1 -- alive, 2 -- dead.
6492       char *tail = value;
6493       char *flag_addr_str = NULL;
6494       char *flag_val_str = NULL;
6495       char const *file_name = NULL;
6496       __kmp_str_split(tail, '-', &flag_addr_str, &tail);
6497       __kmp_str_split(tail, '-', &flag_val_str, &tail);
6498       file_name = tail;
6499       if (tail != NULL) {
6500         long *flag_addr = 0;
6501         long flag_val = 0;
6502         KMP_SSCANF(flag_addr_str, "%p", RCAST(void**, &flag_addr));
6503         KMP_SSCANF(flag_val_str, "%lx", &flag_val);
6504         if (flag_addr != 0 && flag_val != 0 && strcmp(file_name, "") != 0) {
6505           // First, check whether environment-encoded address is mapped into
6506           // addr space.
6507           // If so, dereference it to see if it still has the right value.
6508           if (__kmp_is_address_mapped(flag_addr) && *flag_addr == flag_val) {
6509             neighbor = 1;
6510           } else {
6511             // If not, then we know the other copy of the library is no longer
6512             // running.
6513             neighbor = 2;
6514           }
6515         }
6516       }
6517       switch (neighbor) {
6518       case 0: // Cannot parse environment variable -- neighbor status unknown.
6519         // Assume it is the incompatible format of future version of the
6520         // library. Assume the other library is alive.
6521         // WARN( ... ); // TODO: Issue a warning.
6522         file_name = "unknown library";
6523         KMP_FALLTHROUGH();
6524       // Attention! Falling to the next case. That's intentional.
6525       case 1: { // Neighbor is alive.
6526         // Check it is allowed.
6527         char *duplicate_ok = __kmp_env_get("KMP_DUPLICATE_LIB_OK");
6528         if (!__kmp_str_match_true(duplicate_ok)) {
6529           // That's not allowed. Issue fatal error.
6530           __kmp_fatal(KMP_MSG(DuplicateLibrary, KMP_LIBRARY_FILE, file_name),
6531                       KMP_HNT(DuplicateLibrary), __kmp_msg_null);
6532         }
6533         KMP_INTERNAL_FREE(duplicate_ok);
6534         __kmp_duplicate_library_ok = 1;
6535         done = 1; // Exit the loop.
6536       } break;
6537       case 2: { // Neighbor is dead.
6538         // Clear the variable and try to register library again.
6539         __kmp_env_unset(name);
6540       } break;
6541       default: { KMP_DEBUG_ASSERT(0); } break;
6542       }
6543     }
6544     KMP_INTERNAL_FREE((void *)value);
6545   }
6546   KMP_INTERNAL_FREE((void *)name);
6547 
6548 } // func __kmp_register_library_startup
6549 
6550 void __kmp_unregister_library(void) {
6551 
6552   char *name = __kmp_reg_status_name();
6553   char *value = __kmp_env_get(name);
6554 
6555   KMP_DEBUG_ASSERT(__kmp_registration_flag != 0);
6556   KMP_DEBUG_ASSERT(__kmp_registration_str != NULL);
6557   if (value != NULL && strcmp(value, __kmp_registration_str) == 0) {
6558     // Ok, this is our variable. Delete it.
6559     __kmp_env_unset(name);
6560   }
6561 
6562   KMP_INTERNAL_FREE(__kmp_registration_str);
6563   KMP_INTERNAL_FREE(value);
6564   KMP_INTERNAL_FREE(name);
6565 
6566   __kmp_registration_flag = 0;
6567   __kmp_registration_str = NULL;
6568 
6569 } // __kmp_unregister_library
6570 
6571 // End of Library registration stuff.
6572 // -----------------------------------------------------------------------------
6573 
6574 #if KMP_MIC_SUPPORTED
6575 
6576 static void __kmp_check_mic_type() {
6577   kmp_cpuid_t cpuid_state = {0};
6578   kmp_cpuid_t *cs_p = &cpuid_state;
6579   __kmp_x86_cpuid(1, 0, cs_p);
6580   // We don't support mic1 at the moment
6581   if ((cs_p->eax & 0xff0) == 0xB10) {
6582     __kmp_mic_type = mic2;
6583   } else if ((cs_p->eax & 0xf0ff0) == 0x50670) {
6584     __kmp_mic_type = mic3;
6585   } else {
6586     __kmp_mic_type = non_mic;
6587   }
6588 }
6589 
6590 #endif /* KMP_MIC_SUPPORTED */
6591 
6592 static void __kmp_do_serial_initialize(void) {
6593   int i, gtid;
6594   int size;
6595 
6596   KA_TRACE(10, ("__kmp_do_serial_initialize: enter\n"));
6597 
6598   KMP_DEBUG_ASSERT(sizeof(kmp_int32) == 4);
6599   KMP_DEBUG_ASSERT(sizeof(kmp_uint32) == 4);
6600   KMP_DEBUG_ASSERT(sizeof(kmp_int64) == 8);
6601   KMP_DEBUG_ASSERT(sizeof(kmp_uint64) == 8);
6602   KMP_DEBUG_ASSERT(sizeof(kmp_intptr_t) == sizeof(void *));
6603 
6604 #if OMPT_SUPPORT
6605   ompt_pre_init();
6606 #endif
6607 
6608   __kmp_validate_locks();
6609 
6610   /* Initialize internal memory allocator */
6611   __kmp_init_allocator();
6612 
6613   /* Register the library startup via an environment variable and check to see
6614      whether another copy of the library is already registered. */
6615 
6616   __kmp_register_library_startup();
6617 
6618   /* TODO reinitialization of library */
6619   if (TCR_4(__kmp_global.g.g_done)) {
6620     KA_TRACE(10, ("__kmp_do_serial_initialize: reinitialization of library\n"));
6621   }
6622 
6623   __kmp_global.g.g_abort = 0;
6624   TCW_SYNC_4(__kmp_global.g.g_done, FALSE);
6625 
6626 /* initialize the locks */
6627 #if KMP_USE_ADAPTIVE_LOCKS
6628 #if KMP_DEBUG_ADAPTIVE_LOCKS
6629   __kmp_init_speculative_stats();
6630 #endif
6631 #endif
6632 #if KMP_STATS_ENABLED
6633   __kmp_stats_init();
6634 #endif
6635   __kmp_init_lock(&__kmp_global_lock);
6636   __kmp_init_queuing_lock(&__kmp_dispatch_lock);
6637   __kmp_init_lock(&__kmp_debug_lock);
6638   __kmp_init_atomic_lock(&__kmp_atomic_lock);
6639   __kmp_init_atomic_lock(&__kmp_atomic_lock_1i);
6640   __kmp_init_atomic_lock(&__kmp_atomic_lock_2i);
6641   __kmp_init_atomic_lock(&__kmp_atomic_lock_4i);
6642   __kmp_init_atomic_lock(&__kmp_atomic_lock_4r);
6643   __kmp_init_atomic_lock(&__kmp_atomic_lock_8i);
6644   __kmp_init_atomic_lock(&__kmp_atomic_lock_8r);
6645   __kmp_init_atomic_lock(&__kmp_atomic_lock_8c);
6646   __kmp_init_atomic_lock(&__kmp_atomic_lock_10r);
6647   __kmp_init_atomic_lock(&__kmp_atomic_lock_16r);
6648   __kmp_init_atomic_lock(&__kmp_atomic_lock_16c);
6649   __kmp_init_atomic_lock(&__kmp_atomic_lock_20c);
6650   __kmp_init_atomic_lock(&__kmp_atomic_lock_32c);
6651   __kmp_init_bootstrap_lock(&__kmp_forkjoin_lock);
6652   __kmp_init_bootstrap_lock(&__kmp_exit_lock);
6653 #if KMP_USE_MONITOR
6654   __kmp_init_bootstrap_lock(&__kmp_monitor_lock);
6655 #endif
6656   __kmp_init_bootstrap_lock(&__kmp_tp_cached_lock);
6657 
6658   /* conduct initialization and initial setup of configuration */
6659 
6660   __kmp_runtime_initialize();
6661 
6662 #if KMP_MIC_SUPPORTED
6663   __kmp_check_mic_type();
6664 #endif
6665 
6666 // Some global variable initialization moved here from kmp_env_initialize()
6667 #ifdef KMP_DEBUG
6668   kmp_diag = 0;
6669 #endif
6670   __kmp_abort_delay = 0;
6671 
6672   // From __kmp_init_dflt_team_nth()
6673   /* assume the entire machine will be used */
6674   __kmp_dflt_team_nth_ub = __kmp_xproc;
6675   if (__kmp_dflt_team_nth_ub < KMP_MIN_NTH) {
6676     __kmp_dflt_team_nth_ub = KMP_MIN_NTH;
6677   }
6678   if (__kmp_dflt_team_nth_ub > __kmp_sys_max_nth) {
6679     __kmp_dflt_team_nth_ub = __kmp_sys_max_nth;
6680   }
6681   __kmp_max_nth = __kmp_sys_max_nth;
6682   __kmp_cg_max_nth = __kmp_sys_max_nth;
6683   __kmp_teams_max_nth = __kmp_xproc; // set a "reasonable" default
6684   if (__kmp_teams_max_nth > __kmp_sys_max_nth) {
6685     __kmp_teams_max_nth = __kmp_sys_max_nth;
6686   }
6687 
6688   // Three vars below moved here from __kmp_env_initialize() "KMP_BLOCKTIME"
6689   // part
6690   __kmp_dflt_blocktime = KMP_DEFAULT_BLOCKTIME;
6691 #if KMP_USE_MONITOR
6692   __kmp_monitor_wakeups =
6693       KMP_WAKEUPS_FROM_BLOCKTIME(__kmp_dflt_blocktime, __kmp_monitor_wakeups);
6694   __kmp_bt_intervals =
6695       KMP_INTERVALS_FROM_BLOCKTIME(__kmp_dflt_blocktime, __kmp_monitor_wakeups);
6696 #endif
6697   // From "KMP_LIBRARY" part of __kmp_env_initialize()
6698   __kmp_library = library_throughput;
6699   // From KMP_SCHEDULE initialization
6700   __kmp_static = kmp_sch_static_balanced;
6701 // AC: do not use analytical here, because it is non-monotonous
6702 //__kmp_guided = kmp_sch_guided_iterative_chunked;
6703 //__kmp_auto = kmp_sch_guided_analytical_chunked; // AC: it is the default, no
6704 // need to repeat assignment
6705 // Barrier initialization. Moved here from __kmp_env_initialize() Barrier branch
6706 // bit control and barrier method control parts
6707 #if KMP_FAST_REDUCTION_BARRIER
6708 #define kmp_reduction_barrier_gather_bb ((int)1)
6709 #define kmp_reduction_barrier_release_bb ((int)1)
6710 #define kmp_reduction_barrier_gather_pat bp_hyper_bar
6711 #define kmp_reduction_barrier_release_pat bp_hyper_bar
6712 #endif // KMP_FAST_REDUCTION_BARRIER
6713   for (i = bs_plain_barrier; i < bs_last_barrier; i++) {
6714     __kmp_barrier_gather_branch_bits[i] = __kmp_barrier_gather_bb_dflt;
6715     __kmp_barrier_release_branch_bits[i] = __kmp_barrier_release_bb_dflt;
6716     __kmp_barrier_gather_pattern[i] = __kmp_barrier_gather_pat_dflt;
6717     __kmp_barrier_release_pattern[i] = __kmp_barrier_release_pat_dflt;
6718 #if KMP_FAST_REDUCTION_BARRIER
6719     if (i == bs_reduction_barrier) { // tested and confirmed on ALTIX only (
6720       // lin_64 ): hyper,1
6721       __kmp_barrier_gather_branch_bits[i] = kmp_reduction_barrier_gather_bb;
6722       __kmp_barrier_release_branch_bits[i] = kmp_reduction_barrier_release_bb;
6723       __kmp_barrier_gather_pattern[i] = kmp_reduction_barrier_gather_pat;
6724       __kmp_barrier_release_pattern[i] = kmp_reduction_barrier_release_pat;
6725     }
6726 #endif // KMP_FAST_REDUCTION_BARRIER
6727   }
6728 #if KMP_FAST_REDUCTION_BARRIER
6729 #undef kmp_reduction_barrier_release_pat
6730 #undef kmp_reduction_barrier_gather_pat
6731 #undef kmp_reduction_barrier_release_bb
6732 #undef kmp_reduction_barrier_gather_bb
6733 #endif // KMP_FAST_REDUCTION_BARRIER
6734 #if KMP_MIC_SUPPORTED
6735   if (__kmp_mic_type == mic2) { // KNC
6736     // AC: plane=3,2, forkjoin=2,1 are optimal for 240 threads on KNC
6737     __kmp_barrier_gather_branch_bits[bs_plain_barrier] = 3; // plain gather
6738     __kmp_barrier_release_branch_bits[bs_forkjoin_barrier] =
6739         1; // forkjoin release
6740     __kmp_barrier_gather_pattern[bs_forkjoin_barrier] = bp_hierarchical_bar;
6741     __kmp_barrier_release_pattern[bs_forkjoin_barrier] = bp_hierarchical_bar;
6742   }
6743 #if KMP_FAST_REDUCTION_BARRIER
6744   if (__kmp_mic_type == mic2) { // KNC
6745     __kmp_barrier_gather_pattern[bs_reduction_barrier] = bp_hierarchical_bar;
6746     __kmp_barrier_release_pattern[bs_reduction_barrier] = bp_hierarchical_bar;
6747   }
6748 #endif // KMP_FAST_REDUCTION_BARRIER
6749 #endif // KMP_MIC_SUPPORTED
6750 
6751 // From KMP_CHECKS initialization
6752 #ifdef KMP_DEBUG
6753   __kmp_env_checks = TRUE; /* development versions have the extra checks */
6754 #else
6755   __kmp_env_checks = FALSE; /* port versions do not have the extra checks */
6756 #endif
6757 
6758   // From "KMP_FOREIGN_THREADS_THREADPRIVATE" initialization
6759   __kmp_foreign_tp = TRUE;
6760 
6761   __kmp_global.g.g_dynamic = FALSE;
6762   __kmp_global.g.g_dynamic_mode = dynamic_default;
6763 
6764   __kmp_env_initialize(NULL);
6765 
6766 // Print all messages in message catalog for testing purposes.
6767 #ifdef KMP_DEBUG
6768   char const *val = __kmp_env_get("KMP_DUMP_CATALOG");
6769   if (__kmp_str_match_true(val)) {
6770     kmp_str_buf_t buffer;
6771     __kmp_str_buf_init(&buffer);
6772     __kmp_i18n_dump_catalog(&buffer);
6773     __kmp_printf("%s", buffer.str);
6774     __kmp_str_buf_free(&buffer);
6775   }
6776   __kmp_env_free(&val);
6777 #endif
6778 
6779   __kmp_threads_capacity =
6780       __kmp_initial_threads_capacity(__kmp_dflt_team_nth_ub);
6781   // Moved here from __kmp_env_initialize() "KMP_ALL_THREADPRIVATE" part
6782   __kmp_tp_capacity = __kmp_default_tp_capacity(
6783       __kmp_dflt_team_nth_ub, __kmp_max_nth, __kmp_allThreadsSpecified);
6784 
6785   // If the library is shut down properly, both pools must be NULL. Just in
6786   // case, set them to NULL -- some memory may leak, but subsequent code will
6787   // work even if pools are not freed.
6788   KMP_DEBUG_ASSERT(__kmp_thread_pool == NULL);
6789   KMP_DEBUG_ASSERT(__kmp_thread_pool_insert_pt == NULL);
6790   KMP_DEBUG_ASSERT(__kmp_team_pool == NULL);
6791   __kmp_thread_pool = NULL;
6792   __kmp_thread_pool_insert_pt = NULL;
6793   __kmp_team_pool = NULL;
6794 
6795   /* Allocate all of the variable sized records */
6796   /* NOTE: __kmp_threads_capacity entries are allocated, but the arrays are
6797    * expandable */
6798   /* Since allocation is cache-aligned, just add extra padding at the end */
6799   size =
6800       (sizeof(kmp_info_t *) + sizeof(kmp_root_t *)) * __kmp_threads_capacity +
6801       CACHE_LINE;
6802   __kmp_threads = (kmp_info_t **)__kmp_allocate(size);
6803   __kmp_root = (kmp_root_t **)((char *)__kmp_threads +
6804                                sizeof(kmp_info_t *) * __kmp_threads_capacity);
6805 
6806   /* init thread counts */
6807   KMP_DEBUG_ASSERT(__kmp_all_nth ==
6808                    0); // Asserts fail if the library is reinitializing and
6809   KMP_DEBUG_ASSERT(__kmp_nth == 0); // something was wrong in termination.
6810   __kmp_all_nth = 0;
6811   __kmp_nth = 0;
6812 
6813   /* setup the uber master thread and hierarchy */
6814   gtid = __kmp_register_root(TRUE);
6815   KA_TRACE(10, ("__kmp_do_serial_initialize  T#%d\n", gtid));
6816   KMP_ASSERT(KMP_UBER_GTID(gtid));
6817   KMP_ASSERT(KMP_INITIAL_GTID(gtid));
6818 
6819   KMP_MB(); /* Flush all pending memory write invalidates.  */
6820 
6821   __kmp_common_initialize();
6822 
6823 #if KMP_OS_UNIX
6824   /* invoke the child fork handler */
6825   __kmp_register_atfork();
6826 #endif
6827 
6828 #if !KMP_DYNAMIC_LIB
6829   {
6830     /* Invoke the exit handler when the program finishes, only for static
6831        library. For dynamic library, we already have _fini and DllMain. */
6832     int rc = atexit(__kmp_internal_end_atexit);
6833     if (rc != 0) {
6834       __kmp_fatal(KMP_MSG(FunctionError, "atexit()"), KMP_ERR(rc),
6835                   __kmp_msg_null);
6836     }
6837   }
6838 #endif
6839 
6840 #if KMP_HANDLE_SIGNALS
6841 #if KMP_OS_UNIX
6842   /* NOTE: make sure that this is called before the user installs their own
6843      signal handlers so that the user handlers are called first. this way they
6844      can return false, not call our handler, avoid terminating the library, and
6845      continue execution where they left off. */
6846   __kmp_install_signals(FALSE);
6847 #endif /* KMP_OS_UNIX */
6848 #if KMP_OS_WINDOWS
6849   __kmp_install_signals(TRUE);
6850 #endif /* KMP_OS_WINDOWS */
6851 #endif
6852 
6853   /* we have finished the serial initialization */
6854   __kmp_init_counter++;
6855 
6856   __kmp_init_serial = TRUE;
6857 
6858   if (__kmp_settings) {
6859     __kmp_env_print();
6860   }
6861 
6862 #if OMP_40_ENABLED
6863   if (__kmp_display_env || __kmp_display_env_verbose) {
6864     __kmp_env_print_2();
6865   }
6866 #endif // OMP_40_ENABLED
6867 
6868 #if OMPT_SUPPORT
6869   ompt_post_init();
6870 #endif
6871 
6872   KMP_MB();
6873 
6874   KA_TRACE(10, ("__kmp_do_serial_initialize: exit\n"));
6875 }
6876 
6877 void __kmp_serial_initialize(void) {
6878   if (__kmp_init_serial) {
6879     return;
6880   }
6881   __kmp_acquire_bootstrap_lock(&__kmp_initz_lock);
6882   if (__kmp_init_serial) {
6883     __kmp_release_bootstrap_lock(&__kmp_initz_lock);
6884     return;
6885   }
6886   __kmp_do_serial_initialize();
6887   __kmp_release_bootstrap_lock(&__kmp_initz_lock);
6888 }
6889 
6890 static void __kmp_do_middle_initialize(void) {
6891   int i, j;
6892   int prev_dflt_team_nth;
6893 
6894   if (!__kmp_init_serial) {
6895     __kmp_do_serial_initialize();
6896   }
6897 
6898   KA_TRACE(10, ("__kmp_middle_initialize: enter\n"));
6899 
6900   // Save the previous value for the __kmp_dflt_team_nth so that
6901   // we can avoid some reinitialization if it hasn't changed.
6902   prev_dflt_team_nth = __kmp_dflt_team_nth;
6903 
6904 #if KMP_AFFINITY_SUPPORTED
6905   // __kmp_affinity_initialize() will try to set __kmp_ncores to the
6906   // number of cores on the machine.
6907   __kmp_affinity_initialize();
6908 
6909   // Run through the __kmp_threads array and set the affinity mask
6910   // for each root thread that is currently registered with the RTL.
6911   for (i = 0; i < __kmp_threads_capacity; i++) {
6912     if (TCR_PTR(__kmp_threads[i]) != NULL) {
6913       __kmp_affinity_set_init_mask(i, TRUE);
6914     }
6915   }
6916 #endif /* KMP_AFFINITY_SUPPORTED */
6917 
6918   KMP_ASSERT(__kmp_xproc > 0);
6919   if (__kmp_avail_proc == 0) {
6920     __kmp_avail_proc = __kmp_xproc;
6921   }
6922 
6923   // If there were empty places in num_threads list (OMP_NUM_THREADS=,,2,3),
6924   // correct them now
6925   j = 0;
6926   while ((j < __kmp_nested_nth.used) && !__kmp_nested_nth.nth[j]) {
6927     __kmp_nested_nth.nth[j] = __kmp_dflt_team_nth = __kmp_dflt_team_nth_ub =
6928         __kmp_avail_proc;
6929     j++;
6930   }
6931 
6932   if (__kmp_dflt_team_nth == 0) {
6933 #ifdef KMP_DFLT_NTH_CORES
6934     // Default #threads = #cores
6935     __kmp_dflt_team_nth = __kmp_ncores;
6936     KA_TRACE(20, ("__kmp_middle_initialize: setting __kmp_dflt_team_nth = "
6937                   "__kmp_ncores (%d)\n",
6938                   __kmp_dflt_team_nth));
6939 #else
6940     // Default #threads = #available OS procs
6941     __kmp_dflt_team_nth = __kmp_avail_proc;
6942     KA_TRACE(20, ("__kmp_middle_initialize: setting __kmp_dflt_team_nth = "
6943                   "__kmp_avail_proc(%d)\n",
6944                   __kmp_dflt_team_nth));
6945 #endif /* KMP_DFLT_NTH_CORES */
6946   }
6947 
6948   if (__kmp_dflt_team_nth < KMP_MIN_NTH) {
6949     __kmp_dflt_team_nth = KMP_MIN_NTH;
6950   }
6951   if (__kmp_dflt_team_nth > __kmp_sys_max_nth) {
6952     __kmp_dflt_team_nth = __kmp_sys_max_nth;
6953   }
6954 
6955   // There's no harm in continuing if the following check fails,
6956   // but it indicates an error in the previous logic.
6957   KMP_DEBUG_ASSERT(__kmp_dflt_team_nth <= __kmp_dflt_team_nth_ub);
6958 
6959   if (__kmp_dflt_team_nth != prev_dflt_team_nth) {
6960     // Run through the __kmp_threads array and set the num threads icv for each
6961     // root thread that is currently registered with the RTL (which has not
6962     // already explicitly set its nthreads-var with a call to
6963     // omp_set_num_threads()).
6964     for (i = 0; i < __kmp_threads_capacity; i++) {
6965       kmp_info_t *thread = __kmp_threads[i];
6966       if (thread == NULL)
6967         continue;
6968       if (thread->th.th_current_task->td_icvs.nproc != 0)
6969         continue;
6970 
6971       set__nproc(__kmp_threads[i], __kmp_dflt_team_nth);
6972     }
6973   }
6974   KA_TRACE(
6975       20,
6976       ("__kmp_middle_initialize: final value for __kmp_dflt_team_nth = %d\n",
6977        __kmp_dflt_team_nth));
6978 
6979 #ifdef KMP_ADJUST_BLOCKTIME
6980   /* Adjust blocktime to zero if necessary  now that __kmp_avail_proc is set */
6981   if (!__kmp_env_blocktime && (__kmp_avail_proc > 0)) {
6982     KMP_DEBUG_ASSERT(__kmp_avail_proc > 0);
6983     if (__kmp_nth > __kmp_avail_proc) {
6984       __kmp_zero_bt = TRUE;
6985     }
6986   }
6987 #endif /* KMP_ADJUST_BLOCKTIME */
6988 
6989   /* we have finished middle initialization */
6990   TCW_SYNC_4(__kmp_init_middle, TRUE);
6991 
6992   KA_TRACE(10, ("__kmp_do_middle_initialize: exit\n"));
6993 }
6994 
6995 void __kmp_middle_initialize(void) {
6996   if (__kmp_init_middle) {
6997     return;
6998   }
6999   __kmp_acquire_bootstrap_lock(&__kmp_initz_lock);
7000   if (__kmp_init_middle) {
7001     __kmp_release_bootstrap_lock(&__kmp_initz_lock);
7002     return;
7003   }
7004   __kmp_do_middle_initialize();
7005   __kmp_release_bootstrap_lock(&__kmp_initz_lock);
7006 }
7007 
7008 void __kmp_parallel_initialize(void) {
7009   int gtid = __kmp_entry_gtid(); // this might be a new root
7010 
7011   /* synchronize parallel initialization (for sibling) */
7012   if (TCR_4(__kmp_init_parallel))
7013     return;
7014   __kmp_acquire_bootstrap_lock(&__kmp_initz_lock);
7015   if (TCR_4(__kmp_init_parallel)) {
7016     __kmp_release_bootstrap_lock(&__kmp_initz_lock);
7017     return;
7018   }
7019 
7020   /* TODO reinitialization after we have already shut down */
7021   if (TCR_4(__kmp_global.g.g_done)) {
7022     KA_TRACE(
7023         10,
7024         ("__kmp_parallel_initialize: attempt to init while shutting down\n"));
7025     __kmp_infinite_loop();
7026   }
7027 
7028   /* jc: The lock __kmp_initz_lock is already held, so calling
7029      __kmp_serial_initialize would cause a deadlock.  So we call
7030      __kmp_do_serial_initialize directly. */
7031   if (!__kmp_init_middle) {
7032     __kmp_do_middle_initialize();
7033   }
7034 
7035 #if OMP_50_ENABLED
7036   __kmp_resume_if_hard_paused();
7037 #endif
7038 
7039   /* begin initialization */
7040   KA_TRACE(10, ("__kmp_parallel_initialize: enter\n"));
7041   KMP_ASSERT(KMP_UBER_GTID(gtid));
7042 
7043 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
7044   // Save the FP control regs.
7045   // Worker threads will set theirs to these values at thread startup.
7046   __kmp_store_x87_fpu_control_word(&__kmp_init_x87_fpu_control_word);
7047   __kmp_store_mxcsr(&__kmp_init_mxcsr);
7048   __kmp_init_mxcsr &= KMP_X86_MXCSR_MASK;
7049 #endif /* KMP_ARCH_X86 || KMP_ARCH_X86_64 */
7050 
7051 #if KMP_OS_UNIX
7052 #if KMP_HANDLE_SIGNALS
7053   /*  must be after __kmp_serial_initialize  */
7054   __kmp_install_signals(TRUE);
7055 #endif
7056 #endif
7057 
7058   __kmp_suspend_initialize();
7059 
7060 #if defined(USE_LOAD_BALANCE)
7061   if (__kmp_global.g.g_dynamic_mode == dynamic_default) {
7062     __kmp_global.g.g_dynamic_mode = dynamic_load_balance;
7063   }
7064 #else
7065   if (__kmp_global.g.g_dynamic_mode == dynamic_default) {
7066     __kmp_global.g.g_dynamic_mode = dynamic_thread_limit;
7067   }
7068 #endif
7069 
7070   if (__kmp_version) {
7071     __kmp_print_version_2();
7072   }
7073 
7074   /* we have finished parallel initialization */
7075   TCW_SYNC_4(__kmp_init_parallel, TRUE);
7076 
7077   KMP_MB();
7078   KA_TRACE(10, ("__kmp_parallel_initialize: exit\n"));
7079 
7080   __kmp_release_bootstrap_lock(&__kmp_initz_lock);
7081 }
7082 
7083 /* ------------------------------------------------------------------------ */
7084 
7085 void __kmp_run_before_invoked_task(int gtid, int tid, kmp_info_t *this_thr,
7086                                    kmp_team_t *team) {
7087   kmp_disp_t *dispatch;
7088 
7089   KMP_MB();
7090 
7091   /* none of the threads have encountered any constructs, yet. */
7092   this_thr->th.th_local.this_construct = 0;
7093 #if KMP_CACHE_MANAGE
7094   KMP_CACHE_PREFETCH(&this_thr->th.th_bar[bs_forkjoin_barrier].bb.b_arrived);
7095 #endif /* KMP_CACHE_MANAGE */
7096   dispatch = (kmp_disp_t *)TCR_PTR(this_thr->th.th_dispatch);
7097   KMP_DEBUG_ASSERT(dispatch);
7098   KMP_DEBUG_ASSERT(team->t.t_dispatch);
7099   // KMP_DEBUG_ASSERT( this_thr->th.th_dispatch == &team->t.t_dispatch[
7100   // this_thr->th.th_info.ds.ds_tid ] );
7101 
7102   dispatch->th_disp_index = 0; /* reset the dispatch buffer counter */
7103 #if OMP_45_ENABLED
7104   dispatch->th_doacross_buf_idx =
7105       0; /* reset the doacross dispatch buffer counter */
7106 #endif
7107   if (__kmp_env_consistency_check)
7108     __kmp_push_parallel(gtid, team->t.t_ident);
7109 
7110   KMP_MB(); /* Flush all pending memory write invalidates.  */
7111 }
7112 
7113 void __kmp_run_after_invoked_task(int gtid, int tid, kmp_info_t *this_thr,
7114                                   kmp_team_t *team) {
7115   if (__kmp_env_consistency_check)
7116     __kmp_pop_parallel(gtid, team->t.t_ident);
7117 
7118   __kmp_finish_implicit_task(this_thr);
7119 }
7120 
7121 int __kmp_invoke_task_func(int gtid) {
7122   int rc;
7123   int tid = __kmp_tid_from_gtid(gtid);
7124   kmp_info_t *this_thr = __kmp_threads[gtid];
7125   kmp_team_t *team = this_thr->th.th_team;
7126 
7127   __kmp_run_before_invoked_task(gtid, tid, this_thr, team);
7128 #if USE_ITT_BUILD
7129   if (__itt_stack_caller_create_ptr) {
7130     __kmp_itt_stack_callee_enter(
7131         (__itt_caller)
7132             team->t.t_stack_id); // inform ittnotify about entering user's code
7133   }
7134 #endif /* USE_ITT_BUILD */
7135 #if INCLUDE_SSC_MARKS
7136   SSC_MARK_INVOKING();
7137 #endif
7138 
7139 #if OMPT_SUPPORT
7140   void *dummy;
7141   void **exit_runtime_p;
7142   ompt_data_t *my_task_data;
7143   ompt_data_t *my_parallel_data;
7144   int ompt_team_size;
7145 
7146   if (ompt_enabled.enabled) {
7147     exit_runtime_p = &(
7148         team->t.t_implicit_task_taskdata[tid].ompt_task_info.frame.exit_frame.ptr);
7149   } else {
7150     exit_runtime_p = &dummy;
7151   }
7152 
7153   my_task_data =
7154       &(team->t.t_implicit_task_taskdata[tid].ompt_task_info.task_data);
7155   my_parallel_data = &(team->t.ompt_team_info.parallel_data);
7156   if (ompt_enabled.ompt_callback_implicit_task) {
7157     ompt_team_size = team->t.t_nproc;
7158     ompt_callbacks.ompt_callback(ompt_callback_implicit_task)(
7159         ompt_scope_begin, my_parallel_data, my_task_data, ompt_team_size,
7160         __kmp_tid_from_gtid(gtid), ompt_task_implicit); // TODO: Can this be ompt_task_initial?
7161     OMPT_CUR_TASK_INFO(this_thr)->thread_num = __kmp_tid_from_gtid(gtid);
7162   }
7163 #endif
7164 
7165 #if KMP_STATS_ENABLED
7166   stats_state_e previous_state = KMP_GET_THREAD_STATE();
7167   if (previous_state == stats_state_e::TEAMS_REGION) {
7168     KMP_PUSH_PARTITIONED_TIMER(OMP_teams);
7169   } else {
7170     KMP_PUSH_PARTITIONED_TIMER(OMP_parallel);
7171   }
7172   KMP_SET_THREAD_STATE(IMPLICIT_TASK);
7173 #endif
7174 
7175   rc = __kmp_invoke_microtask((microtask_t)TCR_SYNC_PTR(team->t.t_pkfn), gtid,
7176                               tid, (int)team->t.t_argc, (void **)team->t.t_argv
7177 #if OMPT_SUPPORT
7178                               ,
7179                               exit_runtime_p
7180 #endif
7181                               );
7182 #if OMPT_SUPPORT
7183   *exit_runtime_p = NULL;
7184 #endif
7185 
7186 #if KMP_STATS_ENABLED
7187   if (previous_state == stats_state_e::TEAMS_REGION) {
7188     KMP_SET_THREAD_STATE(previous_state);
7189   }
7190   KMP_POP_PARTITIONED_TIMER();
7191 #endif
7192 
7193 #if USE_ITT_BUILD
7194   if (__itt_stack_caller_create_ptr) {
7195     __kmp_itt_stack_callee_leave(
7196         (__itt_caller)
7197             team->t.t_stack_id); // inform ittnotify about leaving user's code
7198   }
7199 #endif /* USE_ITT_BUILD */
7200   __kmp_run_after_invoked_task(gtid, tid, this_thr, team);
7201 
7202   return rc;
7203 }
7204 
7205 #if OMP_40_ENABLED
7206 void __kmp_teams_master(int gtid) {
7207   // This routine is called by all master threads in teams construct
7208   kmp_info_t *thr = __kmp_threads[gtid];
7209   kmp_team_t *team = thr->th.th_team;
7210   ident_t *loc = team->t.t_ident;
7211   thr->th.th_set_nproc = thr->th.th_teams_size.nth;
7212   KMP_DEBUG_ASSERT(thr->th.th_teams_microtask);
7213   KMP_DEBUG_ASSERT(thr->th.th_set_nproc);
7214   KA_TRACE(20, ("__kmp_teams_master: T#%d, Tid %d, microtask %p\n", gtid,
7215                 __kmp_tid_from_gtid(gtid), thr->th.th_teams_microtask));
7216 
7217   // This thread is a new CG root.  Set up the proper variables.
7218   kmp_cg_root_t *tmp = (kmp_cg_root_t *)__kmp_allocate(sizeof(kmp_cg_root_t));
7219   tmp->cg_root = thr; // Make thr the CG root
7220   // Init to thread limit that was stored when league masters were forked
7221   tmp->cg_thread_limit = thr->th.th_current_task->td_icvs.thread_limit;
7222   tmp->cg_nthreads = 1; // Init counter to one active thread, this one
7223   KA_TRACE(100, ("__kmp_teams_master: Thread %p created node %p and init"
7224                  " cg_threads to 1\n",
7225                  thr, tmp));
7226   tmp->up = thr->th.th_cg_roots;
7227   thr->th.th_cg_roots = tmp;
7228 
7229 // Launch league of teams now, but not let workers execute
7230 // (they hang on fork barrier until next parallel)
7231 #if INCLUDE_SSC_MARKS
7232   SSC_MARK_FORKING();
7233 #endif
7234   __kmp_fork_call(loc, gtid, fork_context_intel, team->t.t_argc,
7235                   (microtask_t)thr->th.th_teams_microtask, // "wrapped" task
7236                   VOLATILE_CAST(launch_t) __kmp_invoke_task_func, NULL);
7237 #if INCLUDE_SSC_MARKS
7238   SSC_MARK_JOINING();
7239 #endif
7240   // If the team size was reduced from the limit, set it to the new size
7241   if (thr->th.th_team_nproc < thr->th.th_teams_size.nth)
7242     thr->th.th_teams_size.nth = thr->th.th_team_nproc;
7243   // AC: last parameter "1" eliminates join barrier which won't work because
7244   // worker threads are in a fork barrier waiting for more parallel regions
7245   __kmp_join_call(loc, gtid
7246 #if OMPT_SUPPORT
7247                   ,
7248                   fork_context_intel
7249 #endif
7250                   ,
7251                   1);
7252 }
7253 
7254 int __kmp_invoke_teams_master(int gtid) {
7255   kmp_info_t *this_thr = __kmp_threads[gtid];
7256   kmp_team_t *team = this_thr->th.th_team;
7257 #if KMP_DEBUG
7258   if (!__kmp_threads[gtid]->th.th_team->t.t_serialized)
7259     KMP_DEBUG_ASSERT((void *)__kmp_threads[gtid]->th.th_team->t.t_pkfn ==
7260                      (void *)__kmp_teams_master);
7261 #endif
7262   __kmp_run_before_invoked_task(gtid, 0, this_thr, team);
7263   __kmp_teams_master(gtid);
7264   __kmp_run_after_invoked_task(gtid, 0, this_thr, team);
7265   return 1;
7266 }
7267 #endif /* OMP_40_ENABLED */
7268 
7269 /* this sets the requested number of threads for the next parallel region
7270    encountered by this team. since this should be enclosed in the forkjoin
7271    critical section it should avoid race conditions with assymmetrical nested
7272    parallelism */
7273 
7274 void __kmp_push_num_threads(ident_t *id, int gtid, int num_threads) {
7275   kmp_info_t *thr = __kmp_threads[gtid];
7276 
7277   if (num_threads > 0)
7278     thr->th.th_set_nproc = num_threads;
7279 }
7280 
7281 #if OMP_40_ENABLED
7282 
7283 /* this sets the requested number of teams for the teams region and/or
7284    the number of threads for the next parallel region encountered  */
7285 void __kmp_push_num_teams(ident_t *id, int gtid, int num_teams,
7286                           int num_threads) {
7287   kmp_info_t *thr = __kmp_threads[gtid];
7288   KMP_DEBUG_ASSERT(num_teams >= 0);
7289   KMP_DEBUG_ASSERT(num_threads >= 0);
7290 
7291   if (num_teams == 0)
7292     num_teams = 1; // default number of teams is 1.
7293   if (num_teams > __kmp_teams_max_nth) { // if too many teams requested?
7294     if (!__kmp_reserve_warn) {
7295       __kmp_reserve_warn = 1;
7296       __kmp_msg(kmp_ms_warning,
7297                 KMP_MSG(CantFormThrTeam, num_teams, __kmp_teams_max_nth),
7298                 KMP_HNT(Unset_ALL_THREADS), __kmp_msg_null);
7299     }
7300     num_teams = __kmp_teams_max_nth;
7301   }
7302   // Set number of teams (number of threads in the outer "parallel" of the
7303   // teams)
7304   thr->th.th_set_nproc = thr->th.th_teams_size.nteams = num_teams;
7305 
7306   // Remember the number of threads for inner parallel regions
7307   if (num_threads == 0) {
7308     if (!TCR_4(__kmp_init_middle))
7309       __kmp_middle_initialize(); // get __kmp_avail_proc calculated
7310     num_threads = __kmp_avail_proc / num_teams;
7311     if (num_teams * num_threads > __kmp_teams_max_nth) {
7312       // adjust num_threads w/o warning as it is not user setting
7313       num_threads = __kmp_teams_max_nth / num_teams;
7314     }
7315   } else {
7316     // This thread will be the master of the league masters
7317     // Store new thread limit; old limit is saved in th_cg_roots list
7318     thr->th.th_current_task->td_icvs.thread_limit = num_threads;
7319 
7320     if (num_teams * num_threads > __kmp_teams_max_nth) {
7321       int new_threads = __kmp_teams_max_nth / num_teams;
7322       if (!__kmp_reserve_warn) { // user asked for too many threads
7323         __kmp_reserve_warn = 1; // conflicts with KMP_TEAMS_THREAD_LIMIT
7324         __kmp_msg(kmp_ms_warning,
7325                   KMP_MSG(CantFormThrTeam, num_threads, new_threads),
7326                   KMP_HNT(Unset_ALL_THREADS), __kmp_msg_null);
7327       }
7328       num_threads = new_threads;
7329     }
7330   }
7331   thr->th.th_teams_size.nth = num_threads;
7332 }
7333 
7334 // Set the proc_bind var to use in the following parallel region.
7335 void __kmp_push_proc_bind(ident_t *id, int gtid, kmp_proc_bind_t proc_bind) {
7336   kmp_info_t *thr = __kmp_threads[gtid];
7337   thr->th.th_set_proc_bind = proc_bind;
7338 }
7339 
7340 #endif /* OMP_40_ENABLED */
7341 
7342 /* Launch the worker threads into the microtask. */
7343 
7344 void __kmp_internal_fork(ident_t *id, int gtid, kmp_team_t *team) {
7345   kmp_info_t *this_thr = __kmp_threads[gtid];
7346 
7347 #ifdef KMP_DEBUG
7348   int f;
7349 #endif /* KMP_DEBUG */
7350 
7351   KMP_DEBUG_ASSERT(team);
7352   KMP_DEBUG_ASSERT(this_thr->th.th_team == team);
7353   KMP_ASSERT(KMP_MASTER_GTID(gtid));
7354   KMP_MB(); /* Flush all pending memory write invalidates.  */
7355 
7356   team->t.t_construct = 0; /* no single directives seen yet */
7357   team->t.t_ordered.dt.t_value =
7358       0; /* thread 0 enters the ordered section first */
7359 
7360   /* Reset the identifiers on the dispatch buffer */
7361   KMP_DEBUG_ASSERT(team->t.t_disp_buffer);
7362   if (team->t.t_max_nproc > 1) {
7363     int i;
7364     for (i = 0; i < __kmp_dispatch_num_buffers; ++i) {
7365       team->t.t_disp_buffer[i].buffer_index = i;
7366 #if OMP_45_ENABLED
7367       team->t.t_disp_buffer[i].doacross_buf_idx = i;
7368 #endif
7369     }
7370   } else {
7371     team->t.t_disp_buffer[0].buffer_index = 0;
7372 #if OMP_45_ENABLED
7373     team->t.t_disp_buffer[0].doacross_buf_idx = 0;
7374 #endif
7375   }
7376 
7377   KMP_MB(); /* Flush all pending memory write invalidates.  */
7378   KMP_ASSERT(this_thr->th.th_team == team);
7379 
7380 #ifdef KMP_DEBUG
7381   for (f = 0; f < team->t.t_nproc; f++) {
7382     KMP_DEBUG_ASSERT(team->t.t_threads[f] &&
7383                      team->t.t_threads[f]->th.th_team_nproc == team->t.t_nproc);
7384   }
7385 #endif /* KMP_DEBUG */
7386 
7387   /* release the worker threads so they may begin working */
7388   __kmp_fork_barrier(gtid, 0);
7389 }
7390 
7391 void __kmp_internal_join(ident_t *id, int gtid, kmp_team_t *team) {
7392   kmp_info_t *this_thr = __kmp_threads[gtid];
7393 
7394   KMP_DEBUG_ASSERT(team);
7395   KMP_DEBUG_ASSERT(this_thr->th.th_team == team);
7396   KMP_ASSERT(KMP_MASTER_GTID(gtid));
7397   KMP_MB(); /* Flush all pending memory write invalidates.  */
7398 
7399 /* Join barrier after fork */
7400 
7401 #ifdef KMP_DEBUG
7402   if (__kmp_threads[gtid] &&
7403       __kmp_threads[gtid]->th.th_team_nproc != team->t.t_nproc) {
7404     __kmp_printf("GTID: %d, __kmp_threads[%d]=%p\n", gtid, gtid,
7405                  __kmp_threads[gtid]);
7406     __kmp_printf("__kmp_threads[%d]->th.th_team_nproc=%d, TEAM: %p, "
7407                  "team->t.t_nproc=%d\n",
7408                  gtid, __kmp_threads[gtid]->th.th_team_nproc, team,
7409                  team->t.t_nproc);
7410     __kmp_print_structure();
7411   }
7412   KMP_DEBUG_ASSERT(__kmp_threads[gtid] &&
7413                    __kmp_threads[gtid]->th.th_team_nproc == team->t.t_nproc);
7414 #endif /* KMP_DEBUG */
7415 
7416   __kmp_join_barrier(gtid); /* wait for everyone */
7417 #if OMPT_SUPPORT
7418   if (ompt_enabled.enabled &&
7419       this_thr->th.ompt_thread_info.state == ompt_state_wait_barrier_implicit) {
7420     int ds_tid = this_thr->th.th_info.ds.ds_tid;
7421     ompt_data_t *task_data = OMPT_CUR_TASK_DATA(this_thr);
7422     this_thr->th.ompt_thread_info.state = ompt_state_overhead;
7423 #if OMPT_OPTIONAL
7424     void *codeptr = NULL;
7425     if (KMP_MASTER_TID(ds_tid) &&
7426         (ompt_callbacks.ompt_callback(ompt_callback_sync_region_wait) ||
7427          ompt_callbacks.ompt_callback(ompt_callback_sync_region)))
7428       codeptr = OMPT_CUR_TEAM_INFO(this_thr)->master_return_address;
7429 
7430     if (ompt_enabled.ompt_callback_sync_region_wait) {
7431       ompt_callbacks.ompt_callback(ompt_callback_sync_region_wait)(
7432           ompt_sync_region_barrier_implicit, ompt_scope_end, NULL, task_data,
7433           codeptr);
7434     }
7435     if (ompt_enabled.ompt_callback_sync_region) {
7436       ompt_callbacks.ompt_callback(ompt_callback_sync_region)(
7437           ompt_sync_region_barrier_implicit, ompt_scope_end, NULL, task_data,
7438           codeptr);
7439     }
7440 #endif
7441     if (!KMP_MASTER_TID(ds_tid) && ompt_enabled.ompt_callback_implicit_task) {
7442       ompt_callbacks.ompt_callback(ompt_callback_implicit_task)(
7443           ompt_scope_end, NULL, task_data, 0, ds_tid, ompt_task_implicit); // TODO: Can this be ompt_task_initial?
7444     }
7445   }
7446 #endif
7447 
7448   KMP_MB(); /* Flush all pending memory write invalidates.  */
7449   KMP_ASSERT(this_thr->th.th_team == team);
7450 }
7451 
7452 /* ------------------------------------------------------------------------ */
7453 
7454 #ifdef USE_LOAD_BALANCE
7455 
7456 // Return the worker threads actively spinning in the hot team, if we
7457 // are at the outermost level of parallelism.  Otherwise, return 0.
7458 static int __kmp_active_hot_team_nproc(kmp_root_t *root) {
7459   int i;
7460   int retval;
7461   kmp_team_t *hot_team;
7462 
7463   if (root->r.r_active) {
7464     return 0;
7465   }
7466   hot_team = root->r.r_hot_team;
7467   if (__kmp_dflt_blocktime == KMP_MAX_BLOCKTIME) {
7468     return hot_team->t.t_nproc - 1; // Don't count master thread
7469   }
7470 
7471   // Skip the master thread - it is accounted for elsewhere.
7472   retval = 0;
7473   for (i = 1; i < hot_team->t.t_nproc; i++) {
7474     if (hot_team->t.t_threads[i]->th.th_active) {
7475       retval++;
7476     }
7477   }
7478   return retval;
7479 }
7480 
7481 // Perform an automatic adjustment to the number of
7482 // threads used by the next parallel region.
7483 static int __kmp_load_balance_nproc(kmp_root_t *root, int set_nproc) {
7484   int retval;
7485   int pool_active;
7486   int hot_team_active;
7487   int team_curr_active;
7488   int system_active;
7489 
7490   KB_TRACE(20, ("__kmp_load_balance_nproc: called root:%p set_nproc:%d\n", root,
7491                 set_nproc));
7492   KMP_DEBUG_ASSERT(root);
7493   KMP_DEBUG_ASSERT(root->r.r_root_team->t.t_threads[0]
7494                        ->th.th_current_task->td_icvs.dynamic == TRUE);
7495   KMP_DEBUG_ASSERT(set_nproc > 1);
7496 
7497   if (set_nproc == 1) {
7498     KB_TRACE(20, ("__kmp_load_balance_nproc: serial execution.\n"));
7499     return 1;
7500   }
7501 
7502   // Threads that are active in the thread pool, active in the hot team for this
7503   // particular root (if we are at the outer par level), and the currently
7504   // executing thread (to become the master) are available to add to the new
7505   // team, but are currently contributing to the system load, and must be
7506   // accounted for.
7507   pool_active = __kmp_thread_pool_active_nth;
7508   hot_team_active = __kmp_active_hot_team_nproc(root);
7509   team_curr_active = pool_active + hot_team_active + 1;
7510 
7511   // Check the system load.
7512   system_active = __kmp_get_load_balance(__kmp_avail_proc + team_curr_active);
7513   KB_TRACE(30, ("__kmp_load_balance_nproc: system active = %d pool active = %d "
7514                 "hot team active = %d\n",
7515                 system_active, pool_active, hot_team_active));
7516 
7517   if (system_active < 0) {
7518     // There was an error reading the necessary info from /proc, so use the
7519     // thread limit algorithm instead. Once we set __kmp_global.g.g_dynamic_mode
7520     // = dynamic_thread_limit, we shouldn't wind up getting back here.
7521     __kmp_global.g.g_dynamic_mode = dynamic_thread_limit;
7522     KMP_WARNING(CantLoadBalUsing, "KMP_DYNAMIC_MODE=thread limit");
7523 
7524     // Make this call behave like the thread limit algorithm.
7525     retval = __kmp_avail_proc - __kmp_nth +
7526              (root->r.r_active ? 1 : root->r.r_hot_team->t.t_nproc);
7527     if (retval > set_nproc) {
7528       retval = set_nproc;
7529     }
7530     if (retval < KMP_MIN_NTH) {
7531       retval = KMP_MIN_NTH;
7532     }
7533 
7534     KB_TRACE(20, ("__kmp_load_balance_nproc: thread limit exit. retval:%d\n",
7535                   retval));
7536     return retval;
7537   }
7538 
7539   // There is a slight delay in the load balance algorithm in detecting new
7540   // running procs. The real system load at this instant should be at least as
7541   // large as the #active omp thread that are available to add to the team.
7542   if (system_active < team_curr_active) {
7543     system_active = team_curr_active;
7544   }
7545   retval = __kmp_avail_proc - system_active + team_curr_active;
7546   if (retval > set_nproc) {
7547     retval = set_nproc;
7548   }
7549   if (retval < KMP_MIN_NTH) {
7550     retval = KMP_MIN_NTH;
7551   }
7552 
7553   KB_TRACE(20, ("__kmp_load_balance_nproc: exit. retval:%d\n", retval));
7554   return retval;
7555 } // __kmp_load_balance_nproc()
7556 
7557 #endif /* USE_LOAD_BALANCE */
7558 
7559 /* ------------------------------------------------------------------------ */
7560 
7561 /* NOTE: this is called with the __kmp_init_lock held */
7562 void __kmp_cleanup(void) {
7563   int f;
7564 
7565   KA_TRACE(10, ("__kmp_cleanup: enter\n"));
7566 
7567   if (TCR_4(__kmp_init_parallel)) {
7568 #if KMP_HANDLE_SIGNALS
7569     __kmp_remove_signals();
7570 #endif
7571     TCW_4(__kmp_init_parallel, FALSE);
7572   }
7573 
7574   if (TCR_4(__kmp_init_middle)) {
7575 #if KMP_AFFINITY_SUPPORTED
7576     __kmp_affinity_uninitialize();
7577 #endif /* KMP_AFFINITY_SUPPORTED */
7578     __kmp_cleanup_hierarchy();
7579     TCW_4(__kmp_init_middle, FALSE);
7580   }
7581 
7582   KA_TRACE(10, ("__kmp_cleanup: go serial cleanup\n"));
7583 
7584   if (__kmp_init_serial) {
7585     __kmp_runtime_destroy();
7586     __kmp_init_serial = FALSE;
7587   }
7588 
7589   __kmp_cleanup_threadprivate_caches();
7590 
7591   for (f = 0; f < __kmp_threads_capacity; f++) {
7592     if (__kmp_root[f] != NULL) {
7593       __kmp_free(__kmp_root[f]);
7594       __kmp_root[f] = NULL;
7595     }
7596   }
7597   __kmp_free(__kmp_threads);
7598   // __kmp_threads and __kmp_root were allocated at once, as single block, so
7599   // there is no need in freeing __kmp_root.
7600   __kmp_threads = NULL;
7601   __kmp_root = NULL;
7602   __kmp_threads_capacity = 0;
7603 
7604 #if KMP_USE_DYNAMIC_LOCK
7605   __kmp_cleanup_indirect_user_locks();
7606 #else
7607   __kmp_cleanup_user_locks();
7608 #endif
7609 
7610 #if KMP_AFFINITY_SUPPORTED
7611   KMP_INTERNAL_FREE(CCAST(char *, __kmp_cpuinfo_file));
7612   __kmp_cpuinfo_file = NULL;
7613 #endif /* KMP_AFFINITY_SUPPORTED */
7614 
7615 #if KMP_USE_ADAPTIVE_LOCKS
7616 #if KMP_DEBUG_ADAPTIVE_LOCKS
7617   __kmp_print_speculative_stats();
7618 #endif
7619 #endif
7620   KMP_INTERNAL_FREE(__kmp_nested_nth.nth);
7621   __kmp_nested_nth.nth = NULL;
7622   __kmp_nested_nth.size = 0;
7623   __kmp_nested_nth.used = 0;
7624   KMP_INTERNAL_FREE(__kmp_nested_proc_bind.bind_types);
7625   __kmp_nested_proc_bind.bind_types = NULL;
7626   __kmp_nested_proc_bind.size = 0;
7627   __kmp_nested_proc_bind.used = 0;
7628 #if OMP_50_ENABLED
7629   if (__kmp_affinity_format) {
7630     KMP_INTERNAL_FREE(__kmp_affinity_format);
7631     __kmp_affinity_format = NULL;
7632   }
7633 #endif
7634 
7635   __kmp_i18n_catclose();
7636 
7637 #if KMP_USE_HIER_SCHED
7638   __kmp_hier_scheds.deallocate();
7639 #endif
7640 
7641 #if KMP_STATS_ENABLED
7642   __kmp_stats_fini();
7643 #endif
7644 
7645   KA_TRACE(10, ("__kmp_cleanup: exit\n"));
7646 }
7647 
7648 /* ------------------------------------------------------------------------ */
7649 
7650 int __kmp_ignore_mppbeg(void) {
7651   char *env;
7652 
7653   if ((env = getenv("KMP_IGNORE_MPPBEG")) != NULL) {
7654     if (__kmp_str_match_false(env))
7655       return FALSE;
7656   }
7657   // By default __kmpc_begin() is no-op.
7658   return TRUE;
7659 }
7660 
7661 int __kmp_ignore_mppend(void) {
7662   char *env;
7663 
7664   if ((env = getenv("KMP_IGNORE_MPPEND")) != NULL) {
7665     if (__kmp_str_match_false(env))
7666       return FALSE;
7667   }
7668   // By default __kmpc_end() is no-op.
7669   return TRUE;
7670 }
7671 
7672 void __kmp_internal_begin(void) {
7673   int gtid;
7674   kmp_root_t *root;
7675 
7676   /* this is a very important step as it will register new sibling threads
7677      and assign these new uber threads a new gtid */
7678   gtid = __kmp_entry_gtid();
7679   root = __kmp_threads[gtid]->th.th_root;
7680   KMP_ASSERT(KMP_UBER_GTID(gtid));
7681 
7682   if (root->r.r_begin)
7683     return;
7684   __kmp_acquire_lock(&root->r.r_begin_lock, gtid);
7685   if (root->r.r_begin) {
7686     __kmp_release_lock(&root->r.r_begin_lock, gtid);
7687     return;
7688   }
7689 
7690   root->r.r_begin = TRUE;
7691 
7692   __kmp_release_lock(&root->r.r_begin_lock, gtid);
7693 }
7694 
7695 /* ------------------------------------------------------------------------ */
7696 
7697 void __kmp_user_set_library(enum library_type arg) {
7698   int gtid;
7699   kmp_root_t *root;
7700   kmp_info_t *thread;
7701 
7702   /* first, make sure we are initialized so we can get our gtid */
7703 
7704   gtid = __kmp_entry_gtid();
7705   thread = __kmp_threads[gtid];
7706 
7707   root = thread->th.th_root;
7708 
7709   KA_TRACE(20, ("__kmp_user_set_library: enter T#%d, arg: %d, %d\n", gtid, arg,
7710                 library_serial));
7711   if (root->r.r_in_parallel) { /* Must be called in serial section of top-level
7712                                   thread */
7713     KMP_WARNING(SetLibraryIncorrectCall);
7714     return;
7715   }
7716 
7717   switch (arg) {
7718   case library_serial:
7719     thread->th.th_set_nproc = 0;
7720     set__nproc(thread, 1);
7721     break;
7722   case library_turnaround:
7723     thread->th.th_set_nproc = 0;
7724     set__nproc(thread, __kmp_dflt_team_nth ? __kmp_dflt_team_nth
7725                                            : __kmp_dflt_team_nth_ub);
7726     break;
7727   case library_throughput:
7728     thread->th.th_set_nproc = 0;
7729     set__nproc(thread, __kmp_dflt_team_nth ? __kmp_dflt_team_nth
7730                                            : __kmp_dflt_team_nth_ub);
7731     break;
7732   default:
7733     KMP_FATAL(UnknownLibraryType, arg);
7734   }
7735 
7736   __kmp_aux_set_library(arg);
7737 }
7738 
7739 void __kmp_aux_set_stacksize(size_t arg) {
7740   if (!__kmp_init_serial)
7741     __kmp_serial_initialize();
7742 
7743 #if KMP_OS_DARWIN
7744   if (arg & (0x1000 - 1)) {
7745     arg &= ~(0x1000 - 1);
7746     if (arg + 0x1000) /* check for overflow if we round up */
7747       arg += 0x1000;
7748   }
7749 #endif
7750   __kmp_acquire_bootstrap_lock(&__kmp_initz_lock);
7751 
7752   /* only change the default stacksize before the first parallel region */
7753   if (!TCR_4(__kmp_init_parallel)) {
7754     size_t value = arg; /* argument is in bytes */
7755 
7756     if (value < __kmp_sys_min_stksize)
7757       value = __kmp_sys_min_stksize;
7758     else if (value > KMP_MAX_STKSIZE)
7759       value = KMP_MAX_STKSIZE;
7760 
7761     __kmp_stksize = value;
7762 
7763     __kmp_env_stksize = TRUE; /* was KMP_STACKSIZE specified? */
7764   }
7765 
7766   __kmp_release_bootstrap_lock(&__kmp_initz_lock);
7767 }
7768 
7769 /* set the behaviour of the runtime library */
7770 /* TODO this can cause some odd behaviour with sibling parallelism... */
7771 void __kmp_aux_set_library(enum library_type arg) {
7772   __kmp_library = arg;
7773 
7774   switch (__kmp_library) {
7775   case library_serial: {
7776     KMP_INFORM(LibraryIsSerial);
7777   } break;
7778   case library_turnaround:
7779     if (__kmp_use_yield == 1 && !__kmp_use_yield_exp_set)
7780       __kmp_use_yield = 2; // only yield when oversubscribed
7781     break;
7782   case library_throughput:
7783     if (__kmp_dflt_blocktime == KMP_MAX_BLOCKTIME)
7784       __kmp_dflt_blocktime = 200;
7785     break;
7786   default:
7787     KMP_FATAL(UnknownLibraryType, arg);
7788   }
7789 }
7790 
7791 /* Getting team information common for all team API */
7792 // Returns NULL if not in teams construct
7793 static kmp_team_t *__kmp_aux_get_team_info(int &teams_serialized) {
7794   kmp_info_t *thr = __kmp_entry_thread();
7795   teams_serialized = 0;
7796   if (thr->th.th_teams_microtask) {
7797     kmp_team_t *team = thr->th.th_team;
7798     int tlevel = thr->th.th_teams_level; // the level of the teams construct
7799     int ii = team->t.t_level;
7800     teams_serialized = team->t.t_serialized;
7801     int level = tlevel + 1;
7802     KMP_DEBUG_ASSERT(ii >= tlevel);
7803     while (ii > level) {
7804       for (teams_serialized = team->t.t_serialized;
7805            (teams_serialized > 0) && (ii > level); teams_serialized--, ii--) {
7806       }
7807       if (team->t.t_serialized && (!teams_serialized)) {
7808         team = team->t.t_parent;
7809         continue;
7810       }
7811       if (ii > level) {
7812         team = team->t.t_parent;
7813         ii--;
7814       }
7815     }
7816     return team;
7817   }
7818   return NULL;
7819 }
7820 
7821 int __kmp_aux_get_team_num() {
7822   int serialized;
7823   kmp_team_t *team = __kmp_aux_get_team_info(serialized);
7824   if (team) {
7825     if (serialized > 1) {
7826       return 0; // teams region is serialized ( 1 team of 1 thread ).
7827     } else {
7828       return team->t.t_master_tid;
7829     }
7830   }
7831   return 0;
7832 }
7833 
7834 int __kmp_aux_get_num_teams() {
7835   int serialized;
7836   kmp_team_t *team = __kmp_aux_get_team_info(serialized);
7837   if (team) {
7838     if (serialized > 1) {
7839       return 1;
7840     } else {
7841       return team->t.t_parent->t.t_nproc;
7842     }
7843   }
7844   return 1;
7845 }
7846 
7847 /* ------------------------------------------------------------------------ */
7848 
7849 #if OMP_50_ENABLED
7850 /*
7851  * Affinity Format Parser
7852  *
7853  * Field is in form of: %[[[0].]size]type
7854  * % and type are required (%% means print a literal '%')
7855  * type is either single char or long name surrounded by {},
7856  * e.g., N or {num_threads}
7857  * 0 => leading zeros
7858  * . => right justified when size is specified
7859  * by default output is left justified
7860  * size is the *minimum* field length
7861  * All other characters are printed as is
7862  *
7863  * Available field types:
7864  * L {thread_level}      - omp_get_level()
7865  * n {thread_num}        - omp_get_thread_num()
7866  * h {host}              - name of host machine
7867  * P {process_id}        - process id (integer)
7868  * T {thread_identifier} - native thread identifier (integer)
7869  * N {num_threads}       - omp_get_num_threads()
7870  * A {ancestor_tnum}     - omp_get_ancestor_thread_num(omp_get_level()-1)
7871  * a {thread_affinity}   - comma separated list of integers or integer ranges
7872  *                         (values of affinity mask)
7873  *
7874  * Implementation-specific field types can be added
7875  * If a type is unknown, print "undefined"
7876 */
7877 
7878 // Structure holding the short name, long name, and corresponding data type
7879 // for snprintf.  A table of these will represent the entire valid keyword
7880 // field types.
7881 typedef struct kmp_affinity_format_field_t {
7882   char short_name; // from spec e.g., L -> thread level
7883   const char *long_name; // from spec thread_level -> thread level
7884   char field_format; // data type for snprintf (typically 'd' or 's'
7885   // for integer or string)
7886 } kmp_affinity_format_field_t;
7887 
7888 static const kmp_affinity_format_field_t __kmp_affinity_format_table[] = {
7889 #if KMP_AFFINITY_SUPPORTED
7890     {'A', "thread_affinity", 's'},
7891 #endif
7892     {'t', "team_num", 'd'},
7893     {'T', "num_teams", 'd'},
7894     {'L', "nesting_level", 'd'},
7895     {'n', "thread_num", 'd'},
7896     {'N', "num_threads", 'd'},
7897     {'a', "ancestor_tnum", 'd'},
7898     {'H', "host", 's'},
7899     {'P', "process_id", 'd'},
7900     {'i', "native_thread_id", 'd'}};
7901 
7902 // Return the number of characters it takes to hold field
7903 static int __kmp_aux_capture_affinity_field(int gtid, const kmp_info_t *th,
7904                                             const char **ptr,
7905                                             kmp_str_buf_t *field_buffer) {
7906   int rc, format_index, field_value;
7907   const char *width_left, *width_right;
7908   bool pad_zeros, right_justify, parse_long_name, found_valid_name;
7909   static const int FORMAT_SIZE = 20;
7910   char format[FORMAT_SIZE] = {0};
7911   char absolute_short_name = 0;
7912 
7913   KMP_DEBUG_ASSERT(gtid >= 0);
7914   KMP_DEBUG_ASSERT(th);
7915   KMP_DEBUG_ASSERT(**ptr == '%');
7916   KMP_DEBUG_ASSERT(field_buffer);
7917 
7918   __kmp_str_buf_clear(field_buffer);
7919 
7920   // Skip the initial %
7921   (*ptr)++;
7922 
7923   // Check for %% first
7924   if (**ptr == '%') {
7925     __kmp_str_buf_cat(field_buffer, "%", 1);
7926     (*ptr)++; // skip over the second %
7927     return 1;
7928   }
7929 
7930   // Parse field modifiers if they are present
7931   pad_zeros = false;
7932   if (**ptr == '0') {
7933     pad_zeros = true;
7934     (*ptr)++; // skip over 0
7935   }
7936   right_justify = false;
7937   if (**ptr == '.') {
7938     right_justify = true;
7939     (*ptr)++; // skip over .
7940   }
7941   // Parse width of field: [width_left, width_right)
7942   width_left = width_right = NULL;
7943   if (**ptr >= '0' && **ptr <= '9') {
7944     width_left = *ptr;
7945     SKIP_DIGITS(*ptr);
7946     width_right = *ptr;
7947   }
7948 
7949   // Create the format for KMP_SNPRINTF based on flags parsed above
7950   format_index = 0;
7951   format[format_index++] = '%';
7952   if (!right_justify)
7953     format[format_index++] = '-';
7954   if (pad_zeros)
7955     format[format_index++] = '0';
7956   if (width_left && width_right) {
7957     int i = 0;
7958     // Only allow 8 digit number widths.
7959     // This also prevents overflowing format variable
7960     while (i < 8 && width_left < width_right) {
7961       format[format_index++] = *width_left;
7962       width_left++;
7963       i++;
7964     }
7965   }
7966 
7967   // Parse a name (long or short)
7968   // Canonicalize the name into absolute_short_name
7969   found_valid_name = false;
7970   parse_long_name = (**ptr == '{');
7971   if (parse_long_name)
7972     (*ptr)++; // skip initial left brace
7973   for (size_t i = 0; i < sizeof(__kmp_affinity_format_table) /
7974                              sizeof(__kmp_affinity_format_table[0]);
7975        ++i) {
7976     char short_name = __kmp_affinity_format_table[i].short_name;
7977     const char *long_name = __kmp_affinity_format_table[i].long_name;
7978     char field_format = __kmp_affinity_format_table[i].field_format;
7979     if (parse_long_name) {
7980       int length = KMP_STRLEN(long_name);
7981       if (strncmp(*ptr, long_name, length) == 0) {
7982         found_valid_name = true;
7983         (*ptr) += length; // skip the long name
7984       }
7985     } else if (**ptr == short_name) {
7986       found_valid_name = true;
7987       (*ptr)++; // skip the short name
7988     }
7989     if (found_valid_name) {
7990       format[format_index++] = field_format;
7991       format[format_index++] = '\0';
7992       absolute_short_name = short_name;
7993       break;
7994     }
7995   }
7996   if (parse_long_name) {
7997     if (**ptr != '}') {
7998       absolute_short_name = 0;
7999     } else {
8000       (*ptr)++; // skip over the right brace
8001     }
8002   }
8003 
8004   // Attempt to fill the buffer with the requested
8005   // value using snprintf within __kmp_str_buf_print()
8006   switch (absolute_short_name) {
8007   case 't':
8008     rc = __kmp_str_buf_print(field_buffer, format, __kmp_aux_get_team_num());
8009     break;
8010   case 'T':
8011     rc = __kmp_str_buf_print(field_buffer, format, __kmp_aux_get_num_teams());
8012     break;
8013   case 'L':
8014     rc = __kmp_str_buf_print(field_buffer, format, th->th.th_team->t.t_level);
8015     break;
8016   case 'n':
8017     rc = __kmp_str_buf_print(field_buffer, format, __kmp_tid_from_gtid(gtid));
8018     break;
8019   case 'H': {
8020     static const int BUFFER_SIZE = 256;
8021     char buf[BUFFER_SIZE];
8022     __kmp_expand_host_name(buf, BUFFER_SIZE);
8023     rc = __kmp_str_buf_print(field_buffer, format, buf);
8024   } break;
8025   case 'P':
8026     rc = __kmp_str_buf_print(field_buffer, format, getpid());
8027     break;
8028   case 'i':
8029     rc = __kmp_str_buf_print(field_buffer, format, __kmp_gettid());
8030     break;
8031   case 'N':
8032     rc = __kmp_str_buf_print(field_buffer, format, th->th.th_team->t.t_nproc);
8033     break;
8034   case 'a':
8035     field_value =
8036         __kmp_get_ancestor_thread_num(gtid, th->th.th_team->t.t_level - 1);
8037     rc = __kmp_str_buf_print(field_buffer, format, field_value);
8038     break;
8039 #if KMP_AFFINITY_SUPPORTED
8040   case 'A': {
8041     kmp_str_buf_t buf;
8042     __kmp_str_buf_init(&buf);
8043     __kmp_affinity_str_buf_mask(&buf, th->th.th_affin_mask);
8044     rc = __kmp_str_buf_print(field_buffer, format, buf.str);
8045     __kmp_str_buf_free(&buf);
8046   } break;
8047 #endif
8048   default:
8049     // According to spec, If an implementation does not have info for field
8050     // type, then "undefined" is printed
8051     rc = __kmp_str_buf_print(field_buffer, "%s", "undefined");
8052     // Skip the field
8053     if (parse_long_name) {
8054       SKIP_TOKEN(*ptr);
8055       if (**ptr == '}')
8056         (*ptr)++;
8057     } else {
8058       (*ptr)++;
8059     }
8060   }
8061 
8062   KMP_ASSERT(format_index <= FORMAT_SIZE);
8063   return rc;
8064 }
8065 
8066 /*
8067  * Return number of characters needed to hold the affinity string
8068  * (not including null byte character)
8069  * The resultant string is printed to buffer, which the caller can then
8070  * handle afterwards
8071 */
8072 size_t __kmp_aux_capture_affinity(int gtid, const char *format,
8073                                   kmp_str_buf_t *buffer) {
8074   const char *parse_ptr;
8075   size_t retval;
8076   const kmp_info_t *th;
8077   kmp_str_buf_t field;
8078 
8079   KMP_DEBUG_ASSERT(buffer);
8080   KMP_DEBUG_ASSERT(gtid >= 0);
8081 
8082   __kmp_str_buf_init(&field);
8083   __kmp_str_buf_clear(buffer);
8084 
8085   th = __kmp_threads[gtid];
8086   retval = 0;
8087 
8088   // If format is NULL or zero-length string, then we use
8089   // affinity-format-var ICV
8090   parse_ptr = format;
8091   if (parse_ptr == NULL || *parse_ptr == '\0') {
8092     parse_ptr = __kmp_affinity_format;
8093   }
8094   KMP_DEBUG_ASSERT(parse_ptr);
8095 
8096   while (*parse_ptr != '\0') {
8097     // Parse a field
8098     if (*parse_ptr == '%') {
8099       // Put field in the buffer
8100       int rc = __kmp_aux_capture_affinity_field(gtid, th, &parse_ptr, &field);
8101       __kmp_str_buf_catbuf(buffer, &field);
8102       retval += rc;
8103     } else {
8104       // Put literal character in buffer
8105       __kmp_str_buf_cat(buffer, parse_ptr, 1);
8106       retval++;
8107       parse_ptr++;
8108     }
8109   }
8110   __kmp_str_buf_free(&field);
8111   return retval;
8112 }
8113 
8114 // Displays the affinity string to stdout
8115 void __kmp_aux_display_affinity(int gtid, const char *format) {
8116   kmp_str_buf_t buf;
8117   __kmp_str_buf_init(&buf);
8118   __kmp_aux_capture_affinity(gtid, format, &buf);
8119   __kmp_fprintf(kmp_out, "%s" KMP_END_OF_LINE, buf.str);
8120   __kmp_str_buf_free(&buf);
8121 }
8122 #endif // OMP_50_ENABLED
8123 
8124 /* ------------------------------------------------------------------------ */
8125 
8126 void __kmp_aux_set_blocktime(int arg, kmp_info_t *thread, int tid) {
8127   int blocktime = arg; /* argument is in milliseconds */
8128 #if KMP_USE_MONITOR
8129   int bt_intervals;
8130 #endif
8131   int bt_set;
8132 
8133   __kmp_save_internal_controls(thread);
8134 
8135   /* Normalize and set blocktime for the teams */
8136   if (blocktime < KMP_MIN_BLOCKTIME)
8137     blocktime = KMP_MIN_BLOCKTIME;
8138   else if (blocktime > KMP_MAX_BLOCKTIME)
8139     blocktime = KMP_MAX_BLOCKTIME;
8140 
8141   set__blocktime_team(thread->th.th_team, tid, blocktime);
8142   set__blocktime_team(thread->th.th_serial_team, 0, blocktime);
8143 
8144 #if KMP_USE_MONITOR
8145   /* Calculate and set blocktime intervals for the teams */
8146   bt_intervals = KMP_INTERVALS_FROM_BLOCKTIME(blocktime, __kmp_monitor_wakeups);
8147 
8148   set__bt_intervals_team(thread->th.th_team, tid, bt_intervals);
8149   set__bt_intervals_team(thread->th.th_serial_team, 0, bt_intervals);
8150 #endif
8151 
8152   /* Set whether blocktime has been set to "TRUE" */
8153   bt_set = TRUE;
8154 
8155   set__bt_set_team(thread->th.th_team, tid, bt_set);
8156   set__bt_set_team(thread->th.th_serial_team, 0, bt_set);
8157 #if KMP_USE_MONITOR
8158   KF_TRACE(10, ("kmp_set_blocktime: T#%d(%d:%d), blocktime=%d, "
8159                 "bt_intervals=%d, monitor_updates=%d\n",
8160                 __kmp_gtid_from_tid(tid, thread->th.th_team),
8161                 thread->th.th_team->t.t_id, tid, blocktime, bt_intervals,
8162                 __kmp_monitor_wakeups));
8163 #else
8164   KF_TRACE(10, ("kmp_set_blocktime: T#%d(%d:%d), blocktime=%d\n",
8165                 __kmp_gtid_from_tid(tid, thread->th.th_team),
8166                 thread->th.th_team->t.t_id, tid, blocktime));
8167 #endif
8168 }
8169 
8170 void __kmp_aux_set_defaults(char const *str, int len) {
8171   if (!__kmp_init_serial) {
8172     __kmp_serial_initialize();
8173   }
8174   __kmp_env_initialize(str);
8175 
8176   if (__kmp_settings
8177 #if OMP_40_ENABLED
8178       || __kmp_display_env || __kmp_display_env_verbose
8179 #endif // OMP_40_ENABLED
8180       ) {
8181     __kmp_env_print();
8182   }
8183 } // __kmp_aux_set_defaults
8184 
8185 /* ------------------------------------------------------------------------ */
8186 /* internal fast reduction routines */
8187 
8188 PACKED_REDUCTION_METHOD_T
8189 __kmp_determine_reduction_method(
8190     ident_t *loc, kmp_int32 global_tid, kmp_int32 num_vars, size_t reduce_size,
8191     void *reduce_data, void (*reduce_func)(void *lhs_data, void *rhs_data),
8192     kmp_critical_name *lck) {
8193 
8194   // Default reduction method: critical construct ( lck != NULL, like in current
8195   // PAROPT )
8196   // If ( reduce_data!=NULL && reduce_func!=NULL ): the tree-reduction method
8197   // can be selected by RTL
8198   // If loc->flags contains KMP_IDENT_ATOMIC_REDUCE, the atomic reduce method
8199   // can be selected by RTL
8200   // Finally, it's up to OpenMP RTL to make a decision on which method to select
8201   // among generated by PAROPT.
8202 
8203   PACKED_REDUCTION_METHOD_T retval;
8204 
8205   int team_size;
8206 
8207   KMP_DEBUG_ASSERT(loc); // it would be nice to test ( loc != 0 )
8208   KMP_DEBUG_ASSERT(lck); // it would be nice to test ( lck != 0 )
8209 
8210 #define FAST_REDUCTION_ATOMIC_METHOD_GENERATED                                 \
8211   ((loc->flags & (KMP_IDENT_ATOMIC_REDUCE)) == (KMP_IDENT_ATOMIC_REDUCE))
8212 #define FAST_REDUCTION_TREE_METHOD_GENERATED ((reduce_data) && (reduce_func))
8213 
8214   retval = critical_reduce_block;
8215 
8216   // another choice of getting a team size (with 1 dynamic deference) is slower
8217   team_size = __kmp_get_team_num_threads(global_tid);
8218   if (team_size == 1) {
8219 
8220     retval = empty_reduce_block;
8221 
8222   } else {
8223 
8224     int atomic_available = FAST_REDUCTION_ATOMIC_METHOD_GENERATED;
8225 
8226 #if KMP_ARCH_X86_64 || KMP_ARCH_PPC64 || KMP_ARCH_AARCH64 || KMP_ARCH_MIPS64
8227 
8228 #if KMP_OS_LINUX || KMP_OS_DRAGONFLY || KMP_OS_FREEBSD || KMP_OS_NETBSD ||     \
8229     KMP_OS_OPENBSD || KMP_OS_WINDOWS || KMP_OS_DARWIN || KMP_OS_HURD
8230 
8231     int teamsize_cutoff = 4;
8232 
8233 #if KMP_MIC_SUPPORTED
8234     if (__kmp_mic_type != non_mic) {
8235       teamsize_cutoff = 8;
8236     }
8237 #endif
8238     int tree_available = FAST_REDUCTION_TREE_METHOD_GENERATED;
8239     if (tree_available) {
8240       if (team_size <= teamsize_cutoff) {
8241         if (atomic_available) {
8242           retval = atomic_reduce_block;
8243         }
8244       } else {
8245         retval = TREE_REDUCE_BLOCK_WITH_REDUCTION_BARRIER;
8246       }
8247     } else if (atomic_available) {
8248       retval = atomic_reduce_block;
8249     }
8250 #else
8251 #error "Unknown or unsupported OS"
8252 #endif // KMP_OS_LINUX || KMP_OS_DRAGONFLY || KMP_OS_FREEBSD || KMP_OS_NETBSD ||
8253        // KMP_OS_OPENBSD || KMP_OS_WINDOWS || KMP_OS_DARWIN || KMP_OS_HURD
8254 
8255 #elif KMP_ARCH_X86 || KMP_ARCH_ARM || KMP_ARCH_AARCH || KMP_ARCH_MIPS
8256 
8257 #if KMP_OS_LINUX || KMP_OS_FREEBSD || KMP_OS_WINDOWS || KMP_OS_HURD
8258 
8259     // basic tuning
8260 
8261     if (atomic_available) {
8262       if (num_vars <= 2) { // && ( team_size <= 8 ) due to false-sharing ???
8263         retval = atomic_reduce_block;
8264       }
8265     } // otherwise: use critical section
8266 
8267 #elif KMP_OS_DARWIN
8268 
8269     int tree_available = FAST_REDUCTION_TREE_METHOD_GENERATED;
8270     if (atomic_available && (num_vars <= 3)) {
8271       retval = atomic_reduce_block;
8272     } else if (tree_available) {
8273       if ((reduce_size > (9 * sizeof(kmp_real64))) &&
8274           (reduce_size < (2000 * sizeof(kmp_real64)))) {
8275         retval = TREE_REDUCE_BLOCK_WITH_PLAIN_BARRIER;
8276       }
8277     } // otherwise: use critical section
8278 
8279 #else
8280 #error "Unknown or unsupported OS"
8281 #endif
8282 
8283 #else
8284 #error "Unknown or unsupported architecture"
8285 #endif
8286   }
8287 
8288   // KMP_FORCE_REDUCTION
8289 
8290   // If the team is serialized (team_size == 1), ignore the forced reduction
8291   // method and stay with the unsynchronized method (empty_reduce_block)
8292   if (__kmp_force_reduction_method != reduction_method_not_defined &&
8293       team_size != 1) {
8294 
8295     PACKED_REDUCTION_METHOD_T forced_retval = critical_reduce_block;
8296 
8297     int atomic_available, tree_available;
8298 
8299     switch ((forced_retval = __kmp_force_reduction_method)) {
8300     case critical_reduce_block:
8301       KMP_ASSERT(lck); // lck should be != 0
8302       break;
8303 
8304     case atomic_reduce_block:
8305       atomic_available = FAST_REDUCTION_ATOMIC_METHOD_GENERATED;
8306       if (!atomic_available) {
8307         KMP_WARNING(RedMethodNotSupported, "atomic");
8308         forced_retval = critical_reduce_block;
8309       }
8310       break;
8311 
8312     case tree_reduce_block:
8313       tree_available = FAST_REDUCTION_TREE_METHOD_GENERATED;
8314       if (!tree_available) {
8315         KMP_WARNING(RedMethodNotSupported, "tree");
8316         forced_retval = critical_reduce_block;
8317       } else {
8318 #if KMP_FAST_REDUCTION_BARRIER
8319         forced_retval = TREE_REDUCE_BLOCK_WITH_REDUCTION_BARRIER;
8320 #endif
8321       }
8322       break;
8323 
8324     default:
8325       KMP_ASSERT(0); // "unsupported method specified"
8326     }
8327 
8328     retval = forced_retval;
8329   }
8330 
8331   KA_TRACE(10, ("reduction method selected=%08x\n", retval));
8332 
8333 #undef FAST_REDUCTION_TREE_METHOD_GENERATED
8334 #undef FAST_REDUCTION_ATOMIC_METHOD_GENERATED
8335 
8336   return (retval);
8337 }
8338 
8339 // this function is for testing set/get/determine reduce method
8340 kmp_int32 __kmp_get_reduce_method(void) {
8341   return ((__kmp_entry_thread()->th.th_local.packed_reduction_method) >> 8);
8342 }
8343 
8344 #if OMP_50_ENABLED
8345 
8346 // Soft pause sets up threads to ignore blocktime and just go to sleep.
8347 // Spin-wait code checks __kmp_pause_status and reacts accordingly.
8348 void __kmp_soft_pause() { __kmp_pause_status = kmp_soft_paused; }
8349 
8350 // Hard pause shuts down the runtime completely.  Resume happens naturally when
8351 // OpenMP is used subsequently.
8352 void __kmp_hard_pause() {
8353   __kmp_pause_status = kmp_hard_paused;
8354   __kmp_internal_end_thread(-1);
8355 }
8356 
8357 // Soft resume sets __kmp_pause_status, and wakes up all threads.
8358 void __kmp_resume_if_soft_paused() {
8359   if (__kmp_pause_status == kmp_soft_paused) {
8360     __kmp_pause_status = kmp_not_paused;
8361 
8362     for (int gtid = 1; gtid < __kmp_threads_capacity; ++gtid) {
8363       kmp_info_t *thread = __kmp_threads[gtid];
8364       if (thread) { // Wake it if sleeping
8365         kmp_flag_64 fl(&thread->th.th_bar[bs_forkjoin_barrier].bb.b_go, thread);
8366         if (fl.is_sleeping())
8367           fl.resume(gtid);
8368         else if (__kmp_try_suspend_mx(thread)) { // got suspend lock
8369           __kmp_unlock_suspend_mx(thread); // unlock it; it won't sleep
8370         } else { // thread holds the lock and may sleep soon
8371           do { // until either the thread sleeps, or we can get the lock
8372             if (fl.is_sleeping()) {
8373               fl.resume(gtid);
8374               break;
8375             } else if (__kmp_try_suspend_mx(thread)) {
8376               __kmp_unlock_suspend_mx(thread);
8377               break;
8378             }
8379           } while (1);
8380         }
8381       }
8382     }
8383   }
8384 }
8385 
8386 // This function is called via __kmpc_pause_resource. Returns 0 if successful.
8387 // TODO: add warning messages
8388 int __kmp_pause_resource(kmp_pause_status_t level) {
8389   if (level == kmp_not_paused) { // requesting resume
8390     if (__kmp_pause_status == kmp_not_paused) {
8391       // error message about runtime not being paused, so can't resume
8392       return 1;
8393     } else {
8394       KMP_DEBUG_ASSERT(__kmp_pause_status == kmp_soft_paused ||
8395                        __kmp_pause_status == kmp_hard_paused);
8396       __kmp_pause_status = kmp_not_paused;
8397       return 0;
8398     }
8399   } else if (level == kmp_soft_paused) { // requesting soft pause
8400     if (__kmp_pause_status != kmp_not_paused) {
8401       // error message about already being paused
8402       return 1;
8403     } else {
8404       __kmp_soft_pause();
8405       return 0;
8406     }
8407   } else if (level == kmp_hard_paused) { // requesting hard pause
8408     if (__kmp_pause_status != kmp_not_paused) {
8409       // error message about already being paused
8410       return 1;
8411     } else {
8412       __kmp_hard_pause();
8413       return 0;
8414     }
8415   } else {
8416     // error message about invalid level
8417     return 1;
8418   }
8419 }
8420 
8421 #endif // OMP_50_ENABLED
8422