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_data_t *parallel_data;
3901     __ompt_get_task_info_internal(0, NULL, &task_data, NULL, &parallel_data, NULL);
3902     if (ompt_enabled.ompt_callback_implicit_task) {
3903       ompt_callbacks.ompt_callback(ompt_callback_implicit_task)(
3904           ompt_scope_begin, parallel_data, task_data, 1, 1, ompt_task_initial);
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   ompt_data_t *task_data;
3995   ompt_data_t *parallel_data;
3996   __ompt_get_task_info_internal(0, NULL, &task_data, NULL, &parallel_data, NULL);
3997   if (ompt_enabled.ompt_callback_implicit_task) {
3998     ompt_callbacks.ompt_callback(ompt_callback_implicit_task)(
3999         ompt_scope_end, parallel_data, task_data, 0, 1, ompt_task_initial);
4000   }
4001   if (ompt_enabled.ompt_callback_thread_end) {
4002     ompt_callbacks.ompt_callback(ompt_callback_thread_end)(
4003         &(root->r.r_uber_thread->th.ompt_thread_info.thread_data));
4004   }
4005 #endif
4006 
4007   TCW_4(__kmp_nth,
4008         __kmp_nth - 1); // __kmp_reap_thread will decrement __kmp_all_nth.
4009   i = root->r.r_uber_thread->th.th_cg_roots->cg_nthreads--;
4010   KA_TRACE(100, ("__kmp_reset_root: Thread %p decrement cg_nthreads on node %p"
4011                  " to %d\n",
4012                  root->r.r_uber_thread, root->r.r_uber_thread->th.th_cg_roots,
4013                  root->r.r_uber_thread->th.th_cg_roots->cg_nthreads));
4014   if (i == 1) {
4015     // need to free contention group structure
4016     KMP_DEBUG_ASSERT(root->r.r_uber_thread ==
4017                      root->r.r_uber_thread->th.th_cg_roots->cg_root);
4018     KMP_DEBUG_ASSERT(root->r.r_uber_thread->th.th_cg_roots->up == NULL);
4019     __kmp_free(root->r.r_uber_thread->th.th_cg_roots);
4020     root->r.r_uber_thread->th.th_cg_roots = NULL;
4021   }
4022   __kmp_reap_thread(root->r.r_uber_thread, 1);
4023 
4024   // We canot put root thread to __kmp_thread_pool, so we have to reap it istead
4025   // of freeing.
4026   root->r.r_uber_thread = NULL;
4027   /* mark root as no longer in use */
4028   root->r.r_begin = FALSE;
4029 
4030   return n;
4031 }
4032 
4033 void __kmp_unregister_root_current_thread(int gtid) {
4034   KA_TRACE(1, ("__kmp_unregister_root_current_thread: enter T#%d\n", gtid));
4035   /* this lock should be ok, since unregister_root_current_thread is never
4036      called during an abort, only during a normal close. furthermore, if you
4037      have the forkjoin lock, you should never try to get the initz lock */
4038   __kmp_acquire_bootstrap_lock(&__kmp_forkjoin_lock);
4039   if (TCR_4(__kmp_global.g.g_done) || !__kmp_init_serial) {
4040     KC_TRACE(10, ("__kmp_unregister_root_current_thread: already finished, "
4041                   "exiting T#%d\n",
4042                   gtid));
4043     __kmp_release_bootstrap_lock(&__kmp_forkjoin_lock);
4044     return;
4045   }
4046   kmp_root_t *root = __kmp_root[gtid];
4047 
4048   KMP_DEBUG_ASSERT(__kmp_threads && __kmp_threads[gtid]);
4049   KMP_ASSERT(KMP_UBER_GTID(gtid));
4050   KMP_ASSERT(root == __kmp_threads[gtid]->th.th_root);
4051   KMP_ASSERT(root->r.r_active == FALSE);
4052 
4053   KMP_MB();
4054 
4055 #if OMP_45_ENABLED
4056   kmp_info_t *thread = __kmp_threads[gtid];
4057   kmp_team_t *team = thread->th.th_team;
4058   kmp_task_team_t *task_team = thread->th.th_task_team;
4059 
4060   // we need to wait for the proxy tasks before finishing the thread
4061   if (task_team != NULL && task_team->tt.tt_found_proxy_tasks) {
4062 #if OMPT_SUPPORT
4063     // the runtime is shutting down so we won't report any events
4064     thread->th.ompt_thread_info.state = ompt_state_undefined;
4065 #endif
4066     __kmp_task_team_wait(thread, team USE_ITT_BUILD_ARG(NULL));
4067   }
4068 #endif
4069 
4070   __kmp_reset_root(gtid, root);
4071 
4072   /* free up this thread slot */
4073   __kmp_gtid_set_specific(KMP_GTID_DNE);
4074 #ifdef KMP_TDATA_GTID
4075   __kmp_gtid = KMP_GTID_DNE;
4076 #endif
4077 
4078   KMP_MB();
4079   KC_TRACE(10,
4080            ("__kmp_unregister_root_current_thread: T#%d unregistered\n", gtid));
4081 
4082   __kmp_release_bootstrap_lock(&__kmp_forkjoin_lock);
4083 }
4084 
4085 #if KMP_OS_WINDOWS
4086 /* __kmp_forkjoin_lock must be already held
4087    Unregisters a root thread that is not the current thread.  Returns the number
4088    of __kmp_threads entries freed as a result. */
4089 static int __kmp_unregister_root_other_thread(int gtid) {
4090   kmp_root_t *root = __kmp_root[gtid];
4091   int r;
4092 
4093   KA_TRACE(1, ("__kmp_unregister_root_other_thread: enter T#%d\n", gtid));
4094   KMP_DEBUG_ASSERT(__kmp_threads && __kmp_threads[gtid]);
4095   KMP_ASSERT(KMP_UBER_GTID(gtid));
4096   KMP_ASSERT(root == __kmp_threads[gtid]->th.th_root);
4097   KMP_ASSERT(root->r.r_active == FALSE);
4098 
4099   r = __kmp_reset_root(gtid, root);
4100   KC_TRACE(10,
4101            ("__kmp_unregister_root_other_thread: T#%d unregistered\n", gtid));
4102   return r;
4103 }
4104 #endif
4105 
4106 #if KMP_DEBUG
4107 void __kmp_task_info() {
4108 
4109   kmp_int32 gtid = __kmp_entry_gtid();
4110   kmp_int32 tid = __kmp_tid_from_gtid(gtid);
4111   kmp_info_t *this_thr = __kmp_threads[gtid];
4112   kmp_team_t *steam = this_thr->th.th_serial_team;
4113   kmp_team_t *team = this_thr->th.th_team;
4114 
4115   __kmp_printf(
4116       "__kmp_task_info: gtid=%d tid=%d t_thread=%p team=%p steam=%p curtask=%p "
4117       "ptask=%p\n",
4118       gtid, tid, this_thr, team, steam, this_thr->th.th_current_task,
4119       team->t.t_implicit_task_taskdata[tid].td_parent);
4120 }
4121 #endif // KMP_DEBUG
4122 
4123 /* TODO optimize with one big memclr, take out what isn't needed, split
4124    responsibility to workers as much as possible, and delay initialization of
4125    features as much as possible  */
4126 static void __kmp_initialize_info(kmp_info_t *this_thr, kmp_team_t *team,
4127                                   int tid, int gtid) {
4128   /* this_thr->th.th_info.ds.ds_gtid is setup in
4129      kmp_allocate_thread/create_worker.
4130      this_thr->th.th_serial_team is setup in __kmp_allocate_thread */
4131   kmp_info_t *master = team->t.t_threads[0];
4132   KMP_DEBUG_ASSERT(this_thr != NULL);
4133   KMP_DEBUG_ASSERT(this_thr->th.th_serial_team);
4134   KMP_DEBUG_ASSERT(team);
4135   KMP_DEBUG_ASSERT(team->t.t_threads);
4136   KMP_DEBUG_ASSERT(team->t.t_dispatch);
4137   KMP_DEBUG_ASSERT(master);
4138   KMP_DEBUG_ASSERT(master->th.th_root);
4139 
4140   KMP_MB();
4141 
4142   TCW_SYNC_PTR(this_thr->th.th_team, team);
4143 
4144   this_thr->th.th_info.ds.ds_tid = tid;
4145   this_thr->th.th_set_nproc = 0;
4146   if (__kmp_tasking_mode != tskm_immediate_exec)
4147     // When tasking is possible, threads are not safe to reap until they are
4148     // done tasking; this will be set when tasking code is exited in wait
4149     this_thr->th.th_reap_state = KMP_NOT_SAFE_TO_REAP;
4150   else // no tasking --> always safe to reap
4151     this_thr->th.th_reap_state = KMP_SAFE_TO_REAP;
4152 #if OMP_40_ENABLED
4153   this_thr->th.th_set_proc_bind = proc_bind_default;
4154 #if KMP_AFFINITY_SUPPORTED
4155   this_thr->th.th_new_place = this_thr->th.th_current_place;
4156 #endif
4157 #endif
4158   this_thr->th.th_root = master->th.th_root;
4159 
4160   /* setup the thread's cache of the team structure */
4161   this_thr->th.th_team_nproc = team->t.t_nproc;
4162   this_thr->th.th_team_master = master;
4163   this_thr->th.th_team_serialized = team->t.t_serialized;
4164   TCW_PTR(this_thr->th.th_sleep_loc, NULL);
4165 
4166   KMP_DEBUG_ASSERT(team->t.t_implicit_task_taskdata);
4167 
4168   KF_TRACE(10, ("__kmp_initialize_info1: T#%d:%d this_thread=%p curtask=%p\n",
4169                 tid, gtid, this_thr, this_thr->th.th_current_task));
4170 
4171   __kmp_init_implicit_task(this_thr->th.th_team_master->th.th_ident, this_thr,
4172                            team, tid, TRUE);
4173 
4174   KF_TRACE(10, ("__kmp_initialize_info2: T#%d:%d this_thread=%p curtask=%p\n",
4175                 tid, gtid, this_thr, this_thr->th.th_current_task));
4176   // TODO: Initialize ICVs from parent; GEH - isn't that already done in
4177   // __kmp_initialize_team()?
4178 
4179   /* TODO no worksharing in speculative threads */
4180   this_thr->th.th_dispatch = &team->t.t_dispatch[tid];
4181 
4182   this_thr->th.th_local.this_construct = 0;
4183 
4184   if (!this_thr->th.th_pri_common) {
4185     this_thr->th.th_pri_common =
4186         (struct common_table *)__kmp_allocate(sizeof(struct common_table));
4187     if (__kmp_storage_map) {
4188       __kmp_print_storage_map_gtid(
4189           gtid, this_thr->th.th_pri_common, this_thr->th.th_pri_common + 1,
4190           sizeof(struct common_table), "th_%d.th_pri_common\n", gtid);
4191     }
4192     this_thr->th.th_pri_head = NULL;
4193   }
4194 
4195   if (this_thr != master && // Master's CG root is initialized elsewhere
4196       this_thr->th.th_cg_roots != master->th.th_cg_roots) { // CG root not set
4197     // Make new thread's CG root same as master's
4198     KMP_DEBUG_ASSERT(master->th.th_cg_roots);
4199     this_thr->th.th_cg_roots = master->th.th_cg_roots;
4200     // Increment new thread's CG root's counter to add the new thread
4201     this_thr->th.th_cg_roots->cg_nthreads++;
4202     KA_TRACE(100, ("__kmp_initialize_info: Thread %p increment cg_nthreads on"
4203                    " node %p of thread %p to %d\n",
4204                    this_thr, this_thr->th.th_cg_roots,
4205                    this_thr->th.th_cg_roots->cg_root,
4206                    this_thr->th.th_cg_roots->cg_nthreads));
4207     this_thr->th.th_current_task->td_icvs.thread_limit =
4208         this_thr->th.th_cg_roots->cg_thread_limit;
4209   }
4210 
4211   /* Initialize dynamic dispatch */
4212   {
4213     volatile kmp_disp_t *dispatch = this_thr->th.th_dispatch;
4214     // Use team max_nproc since this will never change for the team.
4215     size_t disp_size =
4216         sizeof(dispatch_private_info_t) *
4217         (team->t.t_max_nproc == 1 ? 1 : __kmp_dispatch_num_buffers);
4218     KD_TRACE(10, ("__kmp_initialize_info: T#%d max_nproc: %d\n", gtid,
4219                   team->t.t_max_nproc));
4220     KMP_ASSERT(dispatch);
4221     KMP_DEBUG_ASSERT(team->t.t_dispatch);
4222     KMP_DEBUG_ASSERT(dispatch == &team->t.t_dispatch[tid]);
4223 
4224     dispatch->th_disp_index = 0;
4225 #if OMP_45_ENABLED
4226     dispatch->th_doacross_buf_idx = 0;
4227 #endif
4228     if (!dispatch->th_disp_buffer) {
4229       dispatch->th_disp_buffer =
4230           (dispatch_private_info_t *)__kmp_allocate(disp_size);
4231 
4232       if (__kmp_storage_map) {
4233         __kmp_print_storage_map_gtid(
4234             gtid, &dispatch->th_disp_buffer[0],
4235             &dispatch->th_disp_buffer[team->t.t_max_nproc == 1
4236                                           ? 1
4237                                           : __kmp_dispatch_num_buffers],
4238             disp_size, "th_%d.th_dispatch.th_disp_buffer "
4239                        "(team_%d.t_dispatch[%d].th_disp_buffer)",
4240             gtid, team->t.t_id, gtid);
4241       }
4242     } else {
4243       memset(&dispatch->th_disp_buffer[0], '\0', disp_size);
4244     }
4245 
4246     dispatch->th_dispatch_pr_current = 0;
4247     dispatch->th_dispatch_sh_current = 0;
4248 
4249     dispatch->th_deo_fcn = 0; /* ORDERED     */
4250     dispatch->th_dxo_fcn = 0; /* END ORDERED */
4251   }
4252 
4253   this_thr->th.th_next_pool = NULL;
4254 
4255   if (!this_thr->th.th_task_state_memo_stack) {
4256     size_t i;
4257     this_thr->th.th_task_state_memo_stack =
4258         (kmp_uint8 *)__kmp_allocate(4 * sizeof(kmp_uint8));
4259     this_thr->th.th_task_state_top = 0;
4260     this_thr->th.th_task_state_stack_sz = 4;
4261     for (i = 0; i < this_thr->th.th_task_state_stack_sz;
4262          ++i) // zero init the stack
4263       this_thr->th.th_task_state_memo_stack[i] = 0;
4264   }
4265 
4266   KMP_DEBUG_ASSERT(!this_thr->th.th_spin_here);
4267   KMP_DEBUG_ASSERT(this_thr->th.th_next_waiting == 0);
4268 
4269   KMP_MB();
4270 }
4271 
4272 /* allocate a new thread for the requesting team. this is only called from
4273    within a forkjoin critical section. we will first try to get an available
4274    thread from the thread pool. if none is available, we will fork a new one
4275    assuming we are able to create a new one. this should be assured, as the
4276    caller should check on this first. */
4277 kmp_info_t *__kmp_allocate_thread(kmp_root_t *root, kmp_team_t *team,
4278                                   int new_tid) {
4279   kmp_team_t *serial_team;
4280   kmp_info_t *new_thr;
4281   int new_gtid;
4282 
4283   KA_TRACE(20, ("__kmp_allocate_thread: T#%d\n", __kmp_get_gtid()));
4284   KMP_DEBUG_ASSERT(root && team);
4285 #if !KMP_NESTED_HOT_TEAMS
4286   KMP_DEBUG_ASSERT(KMP_MASTER_GTID(__kmp_get_gtid()));
4287 #endif
4288   KMP_MB();
4289 
4290   /* first, try to get one from the thread pool */
4291   if (__kmp_thread_pool) {
4292     new_thr = CCAST(kmp_info_t *, __kmp_thread_pool);
4293     __kmp_thread_pool = (volatile kmp_info_t *)new_thr->th.th_next_pool;
4294     if (new_thr == __kmp_thread_pool_insert_pt) {
4295       __kmp_thread_pool_insert_pt = NULL;
4296     }
4297     TCW_4(new_thr->th.th_in_pool, FALSE);
4298     __kmp_suspend_initialize_thread(new_thr);
4299     __kmp_lock_suspend_mx(new_thr);
4300     if (new_thr->th.th_active_in_pool == TRUE) {
4301       KMP_DEBUG_ASSERT(new_thr->th.th_active == TRUE);
4302       KMP_ATOMIC_DEC(&__kmp_thread_pool_active_nth);
4303       new_thr->th.th_active_in_pool = FALSE;
4304     }
4305 #if KMP_DEBUG
4306     else {
4307       KMP_DEBUG_ASSERT(new_thr->th.th_active == FALSE);
4308     }
4309 #endif
4310     __kmp_unlock_suspend_mx(new_thr);
4311 
4312     KA_TRACE(20, ("__kmp_allocate_thread: T#%d using thread T#%d\n",
4313                   __kmp_get_gtid(), new_thr->th.th_info.ds.ds_gtid));
4314     KMP_ASSERT(!new_thr->th.th_team);
4315     KMP_DEBUG_ASSERT(__kmp_nth < __kmp_threads_capacity);
4316 
4317     /* setup the thread structure */
4318     __kmp_initialize_info(new_thr, team, new_tid,
4319                           new_thr->th.th_info.ds.ds_gtid);
4320     KMP_DEBUG_ASSERT(new_thr->th.th_serial_team);
4321 
4322     TCW_4(__kmp_nth, __kmp_nth + 1);
4323 
4324     new_thr->th.th_task_state = 0;
4325     new_thr->th.th_task_state_top = 0;
4326     new_thr->th.th_task_state_stack_sz = 4;
4327 
4328 #ifdef KMP_ADJUST_BLOCKTIME
4329     /* Adjust blocktime back to zero if necessary */
4330     /* Middle initialization might not have occurred yet */
4331     if (!__kmp_env_blocktime && (__kmp_avail_proc > 0)) {
4332       if (__kmp_nth > __kmp_avail_proc) {
4333         __kmp_zero_bt = TRUE;
4334       }
4335     }
4336 #endif /* KMP_ADJUST_BLOCKTIME */
4337 
4338 #if KMP_DEBUG
4339     // If thread entered pool via __kmp_free_thread, wait_flag should !=
4340     // KMP_BARRIER_PARENT_FLAG.
4341     int b;
4342     kmp_balign_t *balign = new_thr->th.th_bar;
4343     for (b = 0; b < bs_last_barrier; ++b)
4344       KMP_DEBUG_ASSERT(balign[b].bb.wait_flag != KMP_BARRIER_PARENT_FLAG);
4345 #endif
4346 
4347     KF_TRACE(10, ("__kmp_allocate_thread: T#%d using thread %p T#%d\n",
4348                   __kmp_get_gtid(), new_thr, new_thr->th.th_info.ds.ds_gtid));
4349 
4350     KMP_MB();
4351     return new_thr;
4352   }
4353 
4354   /* no, well fork a new one */
4355   KMP_ASSERT(__kmp_nth == __kmp_all_nth);
4356   KMP_ASSERT(__kmp_all_nth < __kmp_threads_capacity);
4357 
4358 #if KMP_USE_MONITOR
4359   // If this is the first worker thread the RTL is creating, then also
4360   // launch the monitor thread.  We try to do this as early as possible.
4361   if (!TCR_4(__kmp_init_monitor)) {
4362     __kmp_acquire_bootstrap_lock(&__kmp_monitor_lock);
4363     if (!TCR_4(__kmp_init_monitor)) {
4364       KF_TRACE(10, ("before __kmp_create_monitor\n"));
4365       TCW_4(__kmp_init_monitor, 1);
4366       __kmp_create_monitor(&__kmp_monitor);
4367       KF_TRACE(10, ("after __kmp_create_monitor\n"));
4368 #if KMP_OS_WINDOWS
4369       // AC: wait until monitor has started. This is a fix for CQ232808.
4370       // The reason is that if the library is loaded/unloaded in a loop with
4371       // small (parallel) work in between, then there is high probability that
4372       // monitor thread started after the library shutdown. At shutdown it is
4373       // too late to cope with the problem, because when the master is in
4374       // DllMain (process detach) the monitor has no chances to start (it is
4375       // blocked), and master has no means to inform the monitor that the
4376       // library has gone, because all the memory which the monitor can access
4377       // is going to be released/reset.
4378       while (TCR_4(__kmp_init_monitor) < 2) {
4379         KMP_YIELD(TRUE);
4380       }
4381       KF_TRACE(10, ("after monitor thread has started\n"));
4382 #endif
4383     }
4384     __kmp_release_bootstrap_lock(&__kmp_monitor_lock);
4385   }
4386 #endif
4387 
4388   KMP_MB();
4389   for (new_gtid = 1; TCR_PTR(__kmp_threads[new_gtid]) != NULL; ++new_gtid) {
4390     KMP_DEBUG_ASSERT(new_gtid < __kmp_threads_capacity);
4391   }
4392 
4393   /* allocate space for it. */
4394   new_thr = (kmp_info_t *)__kmp_allocate(sizeof(kmp_info_t));
4395 
4396   TCW_SYNC_PTR(__kmp_threads[new_gtid], new_thr);
4397 
4398   if (__kmp_storage_map) {
4399     __kmp_print_thread_storage_map(new_thr, new_gtid);
4400   }
4401 
4402   // add the reserve serialized team, initialized from the team's master thread
4403   {
4404     kmp_internal_control_t r_icvs = __kmp_get_x_global_icvs(team);
4405     KF_TRACE(10, ("__kmp_allocate_thread: before th_serial/serial_team\n"));
4406     new_thr->th.th_serial_team = serial_team =
4407         (kmp_team_t *)__kmp_allocate_team(root, 1, 1,
4408 #if OMPT_SUPPORT
4409                                           ompt_data_none, // root parallel id
4410 #endif
4411 #if OMP_40_ENABLED
4412                                           proc_bind_default,
4413 #endif
4414                                           &r_icvs, 0 USE_NESTED_HOT_ARG(NULL));
4415   }
4416   KMP_ASSERT(serial_team);
4417   serial_team->t.t_serialized = 0; // AC: the team created in reserve, not for
4418   // execution (it is unused for now).
4419   serial_team->t.t_threads[0] = new_thr;
4420   KF_TRACE(10,
4421            ("__kmp_allocate_thread: after th_serial/serial_team : new_thr=%p\n",
4422             new_thr));
4423 
4424   /* setup the thread structures */
4425   __kmp_initialize_info(new_thr, team, new_tid, new_gtid);
4426 
4427 #if USE_FAST_MEMORY
4428   __kmp_initialize_fast_memory(new_thr);
4429 #endif /* USE_FAST_MEMORY */
4430 
4431 #if KMP_USE_BGET
4432   KMP_DEBUG_ASSERT(new_thr->th.th_local.bget_data == NULL);
4433   __kmp_initialize_bget(new_thr);
4434 #endif
4435 
4436   __kmp_init_random(new_thr); // Initialize random number generator
4437 
4438   /* Initialize these only once when thread is grabbed for a team allocation */
4439   KA_TRACE(20,
4440            ("__kmp_allocate_thread: T#%d init go fork=%u, plain=%u\n",
4441             __kmp_get_gtid(), KMP_INIT_BARRIER_STATE, KMP_INIT_BARRIER_STATE));
4442 
4443   int b;
4444   kmp_balign_t *balign = new_thr->th.th_bar;
4445   for (b = 0; b < bs_last_barrier; ++b) {
4446     balign[b].bb.b_go = KMP_INIT_BARRIER_STATE;
4447     balign[b].bb.team = NULL;
4448     balign[b].bb.wait_flag = KMP_BARRIER_NOT_WAITING;
4449     balign[b].bb.use_oncore_barrier = 0;
4450   }
4451 
4452   new_thr->th.th_spin_here = FALSE;
4453   new_thr->th.th_next_waiting = 0;
4454 #if KMP_OS_UNIX
4455   new_thr->th.th_blocking = false;
4456 #endif
4457 
4458 #if OMP_40_ENABLED && KMP_AFFINITY_SUPPORTED
4459   new_thr->th.th_current_place = KMP_PLACE_UNDEFINED;
4460   new_thr->th.th_new_place = KMP_PLACE_UNDEFINED;
4461   new_thr->th.th_first_place = KMP_PLACE_UNDEFINED;
4462   new_thr->th.th_last_place = KMP_PLACE_UNDEFINED;
4463 #endif
4464 #if OMP_50_ENABLED
4465   new_thr->th.th_def_allocator = __kmp_def_allocator;
4466   new_thr->th.th_prev_level = 0;
4467   new_thr->th.th_prev_num_threads = 1;
4468 #endif
4469 
4470   TCW_4(new_thr->th.th_in_pool, FALSE);
4471   new_thr->th.th_active_in_pool = FALSE;
4472   TCW_4(new_thr->th.th_active, TRUE);
4473 
4474   /* adjust the global counters */
4475   __kmp_all_nth++;
4476   __kmp_nth++;
4477 
4478   // if __kmp_adjust_gtid_mode is set, then we use method #1 (sp search) for low
4479   // numbers of procs, and method #2 (keyed API call) for higher numbers.
4480   if (__kmp_adjust_gtid_mode) {
4481     if (__kmp_all_nth >= __kmp_tls_gtid_min) {
4482       if (TCR_4(__kmp_gtid_mode) != 2) {
4483         TCW_4(__kmp_gtid_mode, 2);
4484       }
4485     } else {
4486       if (TCR_4(__kmp_gtid_mode) != 1) {
4487         TCW_4(__kmp_gtid_mode, 1);
4488       }
4489     }
4490   }
4491 
4492 #ifdef KMP_ADJUST_BLOCKTIME
4493   /* Adjust blocktime back to zero if necessary       */
4494   /* Middle initialization might not have occurred yet */
4495   if (!__kmp_env_blocktime && (__kmp_avail_proc > 0)) {
4496     if (__kmp_nth > __kmp_avail_proc) {
4497       __kmp_zero_bt = TRUE;
4498     }
4499   }
4500 #endif /* KMP_ADJUST_BLOCKTIME */
4501 
4502   /* actually fork it and create the new worker thread */
4503   KF_TRACE(
4504       10, ("__kmp_allocate_thread: before __kmp_create_worker: %p\n", new_thr));
4505   __kmp_create_worker(new_gtid, new_thr, __kmp_stksize);
4506   KF_TRACE(10,
4507            ("__kmp_allocate_thread: after __kmp_create_worker: %p\n", new_thr));
4508 
4509   KA_TRACE(20, ("__kmp_allocate_thread: T#%d forked T#%d\n", __kmp_get_gtid(),
4510                 new_gtid));
4511   KMP_MB();
4512   return new_thr;
4513 }
4514 
4515 /* Reinitialize team for reuse.
4516    The hot team code calls this case at every fork barrier, so EPCC barrier
4517    test are extremely sensitive to changes in it, esp. writes to the team
4518    struct, which cause a cache invalidation in all threads.
4519    IF YOU TOUCH THIS ROUTINE, RUN EPCC C SYNCBENCH ON A BIG-IRON MACHINE!!! */
4520 static void __kmp_reinitialize_team(kmp_team_t *team,
4521                                     kmp_internal_control_t *new_icvs,
4522                                     ident_t *loc) {
4523   KF_TRACE(10, ("__kmp_reinitialize_team: enter this_thread=%p team=%p\n",
4524                 team->t.t_threads[0], team));
4525   KMP_DEBUG_ASSERT(team && new_icvs);
4526   KMP_DEBUG_ASSERT((!TCR_4(__kmp_init_parallel)) || new_icvs->nproc);
4527   KMP_CHECK_UPDATE(team->t.t_ident, loc);
4528 
4529   KMP_CHECK_UPDATE(team->t.t_id, KMP_GEN_TEAM_ID());
4530   // Copy ICVs to the master thread's implicit taskdata
4531   __kmp_init_implicit_task(loc, team->t.t_threads[0], team, 0, FALSE);
4532   copy_icvs(&team->t.t_implicit_task_taskdata[0].td_icvs, new_icvs);
4533 
4534   KF_TRACE(10, ("__kmp_reinitialize_team: exit this_thread=%p team=%p\n",
4535                 team->t.t_threads[0], team));
4536 }
4537 
4538 /* Initialize the team data structure.
4539    This assumes the t_threads and t_max_nproc are already set.
4540    Also, we don't touch the arguments */
4541 static void __kmp_initialize_team(kmp_team_t *team, int new_nproc,
4542                                   kmp_internal_control_t *new_icvs,
4543                                   ident_t *loc) {
4544   KF_TRACE(10, ("__kmp_initialize_team: enter: team=%p\n", team));
4545 
4546   /* verify */
4547   KMP_DEBUG_ASSERT(team);
4548   KMP_DEBUG_ASSERT(new_nproc <= team->t.t_max_nproc);
4549   KMP_DEBUG_ASSERT(team->t.t_threads);
4550   KMP_MB();
4551 
4552   team->t.t_master_tid = 0; /* not needed */
4553   /* team->t.t_master_bar;        not needed */
4554   team->t.t_serialized = new_nproc > 1 ? 0 : 1;
4555   team->t.t_nproc = new_nproc;
4556 
4557   /* team->t.t_parent     = NULL; TODO not needed & would mess up hot team */
4558   team->t.t_next_pool = NULL;
4559   /* memset( team->t.t_threads, 0, sizeof(kmp_info_t*)*new_nproc ); would mess
4560    * up hot team */
4561 
4562   TCW_SYNC_PTR(team->t.t_pkfn, NULL); /* not needed */
4563   team->t.t_invoke = NULL; /* not needed */
4564 
4565   // TODO???: team->t.t_max_active_levels       = new_max_active_levels;
4566   team->t.t_sched.sched = new_icvs->sched.sched;
4567 
4568 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
4569   team->t.t_fp_control_saved = FALSE; /* not needed */
4570   team->t.t_x87_fpu_control_word = 0; /* not needed */
4571   team->t.t_mxcsr = 0; /* not needed */
4572 #endif /* KMP_ARCH_X86 || KMP_ARCH_X86_64 */
4573 
4574   team->t.t_construct = 0;
4575 
4576   team->t.t_ordered.dt.t_value = 0;
4577   team->t.t_master_active = FALSE;
4578 
4579 #ifdef KMP_DEBUG
4580   team->t.t_copypriv_data = NULL; /* not necessary, but nice for debugging */
4581 #endif
4582 #if KMP_OS_WINDOWS
4583   team->t.t_copyin_counter = 0; /* for barrier-free copyin implementation */
4584 #endif
4585 
4586   team->t.t_control_stack_top = NULL;
4587 
4588   __kmp_reinitialize_team(team, new_icvs, loc);
4589 
4590   KMP_MB();
4591   KF_TRACE(10, ("__kmp_initialize_team: exit: team=%p\n", team));
4592 }
4593 
4594 #if KMP_OS_LINUX && KMP_AFFINITY_SUPPORTED
4595 /* Sets full mask for thread and returns old mask, no changes to structures. */
4596 static void
4597 __kmp_set_thread_affinity_mask_full_tmp(kmp_affin_mask_t *old_mask) {
4598   if (KMP_AFFINITY_CAPABLE()) {
4599     int status;
4600     if (old_mask != NULL) {
4601       status = __kmp_get_system_affinity(old_mask, TRUE);
4602       int error = errno;
4603       if (status != 0) {
4604         __kmp_fatal(KMP_MSG(ChangeThreadAffMaskError), KMP_ERR(error),
4605                     __kmp_msg_null);
4606       }
4607     }
4608     __kmp_set_system_affinity(__kmp_affin_fullMask, TRUE);
4609   }
4610 }
4611 #endif
4612 
4613 #if OMP_40_ENABLED && KMP_AFFINITY_SUPPORTED
4614 
4615 // __kmp_partition_places() is the heart of the OpenMP 4.0 affinity mechanism.
4616 // It calculats the worker + master thread's partition based upon the parent
4617 // thread's partition, and binds each worker to a thread in their partition.
4618 // The master thread's partition should already include its current binding.
4619 static void __kmp_partition_places(kmp_team_t *team, int update_master_only) {
4620   // Copy the master thread's place partion to the team struct
4621   kmp_info_t *master_th = team->t.t_threads[0];
4622   KMP_DEBUG_ASSERT(master_th != NULL);
4623   kmp_proc_bind_t proc_bind = team->t.t_proc_bind;
4624   int first_place = master_th->th.th_first_place;
4625   int last_place = master_th->th.th_last_place;
4626   int masters_place = master_th->th.th_current_place;
4627   team->t.t_first_place = first_place;
4628   team->t.t_last_place = last_place;
4629 
4630   KA_TRACE(20, ("__kmp_partition_places: enter: proc_bind = %d T#%d(%d:0) "
4631                 "bound to place %d partition = [%d,%d]\n",
4632                 proc_bind, __kmp_gtid_from_thread(team->t.t_threads[0]),
4633                 team->t.t_id, masters_place, first_place, last_place));
4634 
4635   switch (proc_bind) {
4636 
4637   case proc_bind_default:
4638     // serial teams might have the proc_bind policy set to proc_bind_default. It
4639     // doesn't matter, as we don't rebind master thread for any proc_bind policy
4640     KMP_DEBUG_ASSERT(team->t.t_nproc == 1);
4641     break;
4642 
4643   case proc_bind_master: {
4644     int f;
4645     int n_th = team->t.t_nproc;
4646     for (f = 1; f < n_th; f++) {
4647       kmp_info_t *th = team->t.t_threads[f];
4648       KMP_DEBUG_ASSERT(th != NULL);
4649       th->th.th_first_place = first_place;
4650       th->th.th_last_place = last_place;
4651       th->th.th_new_place = masters_place;
4652 #if OMP_50_ENABLED
4653       if (__kmp_display_affinity && masters_place != th->th.th_current_place &&
4654           team->t.t_display_affinity != 1) {
4655         team->t.t_display_affinity = 1;
4656       }
4657 #endif
4658 
4659       KA_TRACE(100, ("__kmp_partition_places: master: T#%d(%d:%d) place %d "
4660                      "partition = [%d,%d]\n",
4661                      __kmp_gtid_from_thread(team->t.t_threads[f]), team->t.t_id,
4662                      f, masters_place, first_place, last_place));
4663     }
4664   } break;
4665 
4666   case proc_bind_close: {
4667     int f;
4668     int n_th = team->t.t_nproc;
4669     int n_places;
4670     if (first_place <= last_place) {
4671       n_places = last_place - first_place + 1;
4672     } else {
4673       n_places = __kmp_affinity_num_masks - first_place + last_place + 1;
4674     }
4675     if (n_th <= n_places) {
4676       int place = masters_place;
4677       for (f = 1; f < n_th; f++) {
4678         kmp_info_t *th = team->t.t_threads[f];
4679         KMP_DEBUG_ASSERT(th != NULL);
4680 
4681         if (place == last_place) {
4682           place = first_place;
4683         } else if (place == (int)(__kmp_affinity_num_masks - 1)) {
4684           place = 0;
4685         } else {
4686           place++;
4687         }
4688         th->th.th_first_place = first_place;
4689         th->th.th_last_place = last_place;
4690         th->th.th_new_place = place;
4691 #if OMP_50_ENABLED
4692         if (__kmp_display_affinity && place != th->th.th_current_place &&
4693             team->t.t_display_affinity != 1) {
4694           team->t.t_display_affinity = 1;
4695         }
4696 #endif
4697 
4698         KA_TRACE(100, ("__kmp_partition_places: close: T#%d(%d:%d) place %d "
4699                        "partition = [%d,%d]\n",
4700                        __kmp_gtid_from_thread(team->t.t_threads[f]),
4701                        team->t.t_id, f, place, first_place, last_place));
4702       }
4703     } else {
4704       int S, rem, gap, s_count;
4705       S = n_th / n_places;
4706       s_count = 0;
4707       rem = n_th - (S * n_places);
4708       gap = rem > 0 ? n_places / rem : n_places;
4709       int place = masters_place;
4710       int gap_ct = gap;
4711       for (f = 0; f < n_th; f++) {
4712         kmp_info_t *th = team->t.t_threads[f];
4713         KMP_DEBUG_ASSERT(th != NULL);
4714 
4715         th->th.th_first_place = first_place;
4716         th->th.th_last_place = last_place;
4717         th->th.th_new_place = place;
4718 #if OMP_50_ENABLED
4719         if (__kmp_display_affinity && place != th->th.th_current_place &&
4720             team->t.t_display_affinity != 1) {
4721           team->t.t_display_affinity = 1;
4722         }
4723 #endif
4724         s_count++;
4725 
4726         if ((s_count == S) && rem && (gap_ct == gap)) {
4727           // do nothing, add an extra thread to place on next iteration
4728         } else if ((s_count == S + 1) && rem && (gap_ct == gap)) {
4729           // we added an extra thread to this place; move to next place
4730           if (place == last_place) {
4731             place = first_place;
4732           } else if (place == (int)(__kmp_affinity_num_masks - 1)) {
4733             place = 0;
4734           } else {
4735             place++;
4736           }
4737           s_count = 0;
4738           gap_ct = 1;
4739           rem--;
4740         } else if (s_count == S) { // place full; don't add extra
4741           if (place == last_place) {
4742             place = first_place;
4743           } else if (place == (int)(__kmp_affinity_num_masks - 1)) {
4744             place = 0;
4745           } else {
4746             place++;
4747           }
4748           gap_ct++;
4749           s_count = 0;
4750         }
4751 
4752         KA_TRACE(100,
4753                  ("__kmp_partition_places: close: T#%d(%d:%d) place %d "
4754                   "partition = [%d,%d]\n",
4755                   __kmp_gtid_from_thread(team->t.t_threads[f]), team->t.t_id, f,
4756                   th->th.th_new_place, first_place, last_place));
4757       }
4758       KMP_DEBUG_ASSERT(place == masters_place);
4759     }
4760   } break;
4761 
4762   case proc_bind_spread: {
4763     int f;
4764     int n_th = team->t.t_nproc;
4765     int n_places;
4766     int thidx;
4767     if (first_place <= last_place) {
4768       n_places = last_place - first_place + 1;
4769     } else {
4770       n_places = __kmp_affinity_num_masks - first_place + last_place + 1;
4771     }
4772     if (n_th <= n_places) {
4773       int place = -1;
4774 
4775       if (n_places != static_cast<int>(__kmp_affinity_num_masks)) {
4776         int S = n_places / n_th;
4777         int s_count, rem, gap, gap_ct;
4778 
4779         place = masters_place;
4780         rem = n_places - n_th * S;
4781         gap = rem ? n_th / rem : 1;
4782         gap_ct = gap;
4783         thidx = n_th;
4784         if (update_master_only == 1)
4785           thidx = 1;
4786         for (f = 0; f < thidx; f++) {
4787           kmp_info_t *th = team->t.t_threads[f];
4788           KMP_DEBUG_ASSERT(th != NULL);
4789 
4790           th->th.th_first_place = place;
4791           th->th.th_new_place = place;
4792 #if OMP_50_ENABLED
4793           if (__kmp_display_affinity && place != th->th.th_current_place &&
4794               team->t.t_display_affinity != 1) {
4795             team->t.t_display_affinity = 1;
4796           }
4797 #endif
4798           s_count = 1;
4799           while (s_count < S) {
4800             if (place == last_place) {
4801               place = first_place;
4802             } else if (place == (int)(__kmp_affinity_num_masks - 1)) {
4803               place = 0;
4804             } else {
4805               place++;
4806             }
4807             s_count++;
4808           }
4809           if (rem && (gap_ct == gap)) {
4810             if (place == last_place) {
4811               place = first_place;
4812             } else if (place == (int)(__kmp_affinity_num_masks - 1)) {
4813               place = 0;
4814             } else {
4815               place++;
4816             }
4817             rem--;
4818             gap_ct = 0;
4819           }
4820           th->th.th_last_place = place;
4821           gap_ct++;
4822 
4823           if (place == last_place) {
4824             place = first_place;
4825           } else if (place == (int)(__kmp_affinity_num_masks - 1)) {
4826             place = 0;
4827           } else {
4828             place++;
4829           }
4830 
4831           KA_TRACE(100,
4832                    ("__kmp_partition_places: spread: T#%d(%d:%d) place %d "
4833                     "partition = [%d,%d], __kmp_affinity_num_masks: %u\n",
4834                     __kmp_gtid_from_thread(team->t.t_threads[f]), team->t.t_id,
4835                     f, th->th.th_new_place, th->th.th_first_place,
4836                     th->th.th_last_place, __kmp_affinity_num_masks));
4837         }
4838       } else {
4839         /* Having uniform space of available computation places I can create
4840            T partitions of round(P/T) size and put threads into the first
4841            place of each partition. */
4842         double current = static_cast<double>(masters_place);
4843         double spacing =
4844             (static_cast<double>(n_places + 1) / static_cast<double>(n_th));
4845         int first, last;
4846         kmp_info_t *th;
4847 
4848         thidx = n_th + 1;
4849         if (update_master_only == 1)
4850           thidx = 1;
4851         for (f = 0; f < thidx; f++) {
4852           first = static_cast<int>(current);
4853           last = static_cast<int>(current + spacing) - 1;
4854           KMP_DEBUG_ASSERT(last >= first);
4855           if (first >= n_places) {
4856             if (masters_place) {
4857               first -= n_places;
4858               last -= n_places;
4859               if (first == (masters_place + 1)) {
4860                 KMP_DEBUG_ASSERT(f == n_th);
4861                 first--;
4862               }
4863               if (last == masters_place) {
4864                 KMP_DEBUG_ASSERT(f == (n_th - 1));
4865                 last--;
4866               }
4867             } else {
4868               KMP_DEBUG_ASSERT(f == n_th);
4869               first = 0;
4870               last = 0;
4871             }
4872           }
4873           if (last >= n_places) {
4874             last = (n_places - 1);
4875           }
4876           place = first;
4877           current += spacing;
4878           if (f < n_th) {
4879             KMP_DEBUG_ASSERT(0 <= first);
4880             KMP_DEBUG_ASSERT(n_places > first);
4881             KMP_DEBUG_ASSERT(0 <= last);
4882             KMP_DEBUG_ASSERT(n_places > last);
4883             KMP_DEBUG_ASSERT(last_place >= first_place);
4884             th = team->t.t_threads[f];
4885             KMP_DEBUG_ASSERT(th);
4886             th->th.th_first_place = first;
4887             th->th.th_new_place = place;
4888             th->th.th_last_place = last;
4889 #if OMP_50_ENABLED
4890             if (__kmp_display_affinity && place != th->th.th_current_place &&
4891                 team->t.t_display_affinity != 1) {
4892               team->t.t_display_affinity = 1;
4893             }
4894 #endif
4895             KA_TRACE(100,
4896                      ("__kmp_partition_places: spread: T#%d(%d:%d) place %d "
4897                       "partition = [%d,%d], spacing = %.4f\n",
4898                       __kmp_gtid_from_thread(team->t.t_threads[f]),
4899                       team->t.t_id, f, th->th.th_new_place,
4900                       th->th.th_first_place, th->th.th_last_place, spacing));
4901           }
4902         }
4903       }
4904       KMP_DEBUG_ASSERT(update_master_only || place == masters_place);
4905     } else {
4906       int S, rem, gap, s_count;
4907       S = n_th / n_places;
4908       s_count = 0;
4909       rem = n_th - (S * n_places);
4910       gap = rem > 0 ? n_places / rem : n_places;
4911       int place = masters_place;
4912       int gap_ct = gap;
4913       thidx = n_th;
4914       if (update_master_only == 1)
4915         thidx = 1;
4916       for (f = 0; f < thidx; f++) {
4917         kmp_info_t *th = team->t.t_threads[f];
4918         KMP_DEBUG_ASSERT(th != NULL);
4919 
4920         th->th.th_first_place = place;
4921         th->th.th_last_place = place;
4922         th->th.th_new_place = place;
4923 #if OMP_50_ENABLED
4924         if (__kmp_display_affinity && place != th->th.th_current_place &&
4925             team->t.t_display_affinity != 1) {
4926           team->t.t_display_affinity = 1;
4927         }
4928 #endif
4929         s_count++;
4930 
4931         if ((s_count == S) && rem && (gap_ct == gap)) {
4932           // do nothing, add an extra thread to place on next iteration
4933         } else if ((s_count == S + 1) && rem && (gap_ct == gap)) {
4934           // we added an extra thread to this place; move on to next place
4935           if (place == last_place) {
4936             place = first_place;
4937           } else if (place == (int)(__kmp_affinity_num_masks - 1)) {
4938             place = 0;
4939           } else {
4940             place++;
4941           }
4942           s_count = 0;
4943           gap_ct = 1;
4944           rem--;
4945         } else if (s_count == S) { // place is full; don't add extra thread
4946           if (place == last_place) {
4947             place = first_place;
4948           } else if (place == (int)(__kmp_affinity_num_masks - 1)) {
4949             place = 0;
4950           } else {
4951             place++;
4952           }
4953           gap_ct++;
4954           s_count = 0;
4955         }
4956 
4957         KA_TRACE(100, ("__kmp_partition_places: spread: T#%d(%d:%d) place %d "
4958                        "partition = [%d,%d]\n",
4959                        __kmp_gtid_from_thread(team->t.t_threads[f]),
4960                        team->t.t_id, f, th->th.th_new_place,
4961                        th->th.th_first_place, th->th.th_last_place));
4962       }
4963       KMP_DEBUG_ASSERT(update_master_only || place == masters_place);
4964     }
4965   } break;
4966 
4967   default:
4968     break;
4969   }
4970 
4971   KA_TRACE(20, ("__kmp_partition_places: exit T#%d\n", team->t.t_id));
4972 }
4973 
4974 #endif /* OMP_40_ENABLED && KMP_AFFINITY_SUPPORTED */
4975 
4976 /* allocate a new team data structure to use.  take one off of the free pool if
4977    available */
4978 kmp_team_t *
4979 __kmp_allocate_team(kmp_root_t *root, int new_nproc, int max_nproc,
4980 #if OMPT_SUPPORT
4981                     ompt_data_t ompt_parallel_data,
4982 #endif
4983 #if OMP_40_ENABLED
4984                     kmp_proc_bind_t new_proc_bind,
4985 #endif
4986                     kmp_internal_control_t *new_icvs,
4987                     int argc USE_NESTED_HOT_ARG(kmp_info_t *master)) {
4988   KMP_TIME_DEVELOPER_PARTITIONED_BLOCK(KMP_allocate_team);
4989   int f;
4990   kmp_team_t *team;
4991   int use_hot_team = !root->r.r_active;
4992   int level = 0;
4993 
4994   KA_TRACE(20, ("__kmp_allocate_team: called\n"));
4995   KMP_DEBUG_ASSERT(new_nproc >= 1 && argc >= 0);
4996   KMP_DEBUG_ASSERT(max_nproc >= new_nproc);
4997   KMP_MB();
4998 
4999 #if KMP_NESTED_HOT_TEAMS
5000   kmp_hot_team_ptr_t *hot_teams;
5001   if (master) {
5002     team = master->th.th_team;
5003     level = team->t.t_active_level;
5004     if (master->th.th_teams_microtask) { // in teams construct?
5005       if (master->th.th_teams_size.nteams > 1 &&
5006           ( // #teams > 1
5007               team->t.t_pkfn ==
5008                   (microtask_t)__kmp_teams_master || // inner fork of the teams
5009               master->th.th_teams_level <
5010                   team->t.t_level)) { // or nested parallel inside the teams
5011         ++level; // not increment if #teams==1, or for outer fork of the teams;
5012         // increment otherwise
5013       }
5014     }
5015     hot_teams = master->th.th_hot_teams;
5016     if (level < __kmp_hot_teams_max_level && hot_teams &&
5017         hot_teams[level]
5018             .hot_team) { // hot team has already been allocated for given level
5019       use_hot_team = 1;
5020     } else {
5021       use_hot_team = 0;
5022     }
5023   }
5024 #endif
5025   // Optimization to use a "hot" team
5026   if (use_hot_team && new_nproc > 1) {
5027     KMP_DEBUG_ASSERT(new_nproc <= max_nproc);
5028 #if KMP_NESTED_HOT_TEAMS
5029     team = hot_teams[level].hot_team;
5030 #else
5031     team = root->r.r_hot_team;
5032 #endif
5033 #if KMP_DEBUG
5034     if (__kmp_tasking_mode != tskm_immediate_exec) {
5035       KA_TRACE(20, ("__kmp_allocate_team: hot team task_team[0] = %p "
5036                     "task_team[1] = %p before reinit\n",
5037                     team->t.t_task_team[0], team->t.t_task_team[1]));
5038     }
5039 #endif
5040 
5041     // Has the number of threads changed?
5042     /* Let's assume the most common case is that the number of threads is
5043        unchanged, and put that case first. */
5044     if (team->t.t_nproc == new_nproc) { // Check changes in number of threads
5045       KA_TRACE(20, ("__kmp_allocate_team: reusing hot team\n"));
5046       // This case can mean that omp_set_num_threads() was called and the hot
5047       // team size was already reduced, so we check the special flag
5048       if (team->t.t_size_changed == -1) {
5049         team->t.t_size_changed = 1;
5050       } else {
5051         KMP_CHECK_UPDATE(team->t.t_size_changed, 0);
5052       }
5053 
5054       // TODO???: team->t.t_max_active_levels = new_max_active_levels;
5055       kmp_r_sched_t new_sched = new_icvs->sched;
5056       // set master's schedule as new run-time schedule
5057       KMP_CHECK_UPDATE(team->t.t_sched.sched, new_sched.sched);
5058 
5059       __kmp_reinitialize_team(team, new_icvs,
5060                               root->r.r_uber_thread->th.th_ident);
5061 
5062       KF_TRACE(10, ("__kmp_allocate_team2: T#%d, this_thread=%p team=%p\n", 0,
5063                     team->t.t_threads[0], team));
5064       __kmp_push_current_task_to_thread(team->t.t_threads[0], team, 0);
5065 
5066 #if OMP_40_ENABLED
5067 #if KMP_AFFINITY_SUPPORTED
5068       if ((team->t.t_size_changed == 0) &&
5069           (team->t.t_proc_bind == new_proc_bind)) {
5070         if (new_proc_bind == proc_bind_spread) {
5071           __kmp_partition_places(
5072               team, 1); // add flag to update only master for spread
5073         }
5074         KA_TRACE(200, ("__kmp_allocate_team: reusing hot team #%d bindings: "
5075                        "proc_bind = %d, partition = [%d,%d]\n",
5076                        team->t.t_id, new_proc_bind, team->t.t_first_place,
5077                        team->t.t_last_place));
5078       } else {
5079         KMP_CHECK_UPDATE(team->t.t_proc_bind, new_proc_bind);
5080         __kmp_partition_places(team);
5081       }
5082 #else
5083       KMP_CHECK_UPDATE(team->t.t_proc_bind, new_proc_bind);
5084 #endif /* KMP_AFFINITY_SUPPORTED */
5085 #endif /* OMP_40_ENABLED */
5086     } else if (team->t.t_nproc > new_nproc) {
5087       KA_TRACE(20,
5088                ("__kmp_allocate_team: decreasing hot team thread count to %d\n",
5089                 new_nproc));
5090 
5091       team->t.t_size_changed = 1;
5092 #if KMP_NESTED_HOT_TEAMS
5093       if (__kmp_hot_teams_mode == 0) {
5094         // AC: saved number of threads should correspond to team's value in this
5095         // mode, can be bigger in mode 1, when hot team has threads in reserve
5096         KMP_DEBUG_ASSERT(hot_teams[level].hot_team_nth == team->t.t_nproc);
5097         hot_teams[level].hot_team_nth = new_nproc;
5098 #endif // KMP_NESTED_HOT_TEAMS
5099         /* release the extra threads we don't need any more */
5100         for (f = new_nproc; f < team->t.t_nproc; f++) {
5101           KMP_DEBUG_ASSERT(team->t.t_threads[f]);
5102           if (__kmp_tasking_mode != tskm_immediate_exec) {
5103             // When decreasing team size, threads no longer in the team should
5104             // unref task team.
5105             team->t.t_threads[f]->th.th_task_team = NULL;
5106           }
5107           __kmp_free_thread(team->t.t_threads[f]);
5108           team->t.t_threads[f] = NULL;
5109         }
5110 #if KMP_NESTED_HOT_TEAMS
5111       } // (__kmp_hot_teams_mode == 0)
5112       else {
5113         // When keeping extra threads in team, switch threads to wait on own
5114         // b_go flag
5115         for (f = new_nproc; f < team->t.t_nproc; ++f) {
5116           KMP_DEBUG_ASSERT(team->t.t_threads[f]);
5117           kmp_balign_t *balign = team->t.t_threads[f]->th.th_bar;
5118           for (int b = 0; b < bs_last_barrier; ++b) {
5119             if (balign[b].bb.wait_flag == KMP_BARRIER_PARENT_FLAG) {
5120               balign[b].bb.wait_flag = KMP_BARRIER_SWITCH_TO_OWN_FLAG;
5121             }
5122             KMP_CHECK_UPDATE(balign[b].bb.leaf_kids, 0);
5123           }
5124         }
5125       }
5126 #endif // KMP_NESTED_HOT_TEAMS
5127       team->t.t_nproc = new_nproc;
5128       // TODO???: team->t.t_max_active_levels = new_max_active_levels;
5129       KMP_CHECK_UPDATE(team->t.t_sched.sched, new_icvs->sched.sched);
5130       __kmp_reinitialize_team(team, new_icvs,
5131                               root->r.r_uber_thread->th.th_ident);
5132 
5133       // Update remaining threads
5134       for (f = 0; f < new_nproc; ++f) {
5135         team->t.t_threads[f]->th.th_team_nproc = new_nproc;
5136       }
5137 
5138       // restore the current task state of the master thread: should be the
5139       // implicit task
5140       KF_TRACE(10, ("__kmp_allocate_team: T#%d, this_thread=%p team=%p\n", 0,
5141                     team->t.t_threads[0], team));
5142 
5143       __kmp_push_current_task_to_thread(team->t.t_threads[0], team, 0);
5144 
5145 #ifdef KMP_DEBUG
5146       for (f = 0; f < team->t.t_nproc; f++) {
5147         KMP_DEBUG_ASSERT(team->t.t_threads[f] &&
5148                          team->t.t_threads[f]->th.th_team_nproc ==
5149                              team->t.t_nproc);
5150       }
5151 #endif
5152 
5153 #if OMP_40_ENABLED
5154       KMP_CHECK_UPDATE(team->t.t_proc_bind, new_proc_bind);
5155 #if KMP_AFFINITY_SUPPORTED
5156       __kmp_partition_places(team);
5157 #endif
5158 #endif
5159     } else { // team->t.t_nproc < new_nproc
5160 #if KMP_OS_LINUX && KMP_AFFINITY_SUPPORTED
5161       kmp_affin_mask_t *old_mask;
5162       if (KMP_AFFINITY_CAPABLE()) {
5163         KMP_CPU_ALLOC(old_mask);
5164       }
5165 #endif
5166 
5167       KA_TRACE(20,
5168                ("__kmp_allocate_team: increasing hot team thread count to %d\n",
5169                 new_nproc));
5170 
5171       team->t.t_size_changed = 1;
5172 
5173 #if KMP_NESTED_HOT_TEAMS
5174       int avail_threads = hot_teams[level].hot_team_nth;
5175       if (new_nproc < avail_threads)
5176         avail_threads = new_nproc;
5177       kmp_info_t **other_threads = team->t.t_threads;
5178       for (f = team->t.t_nproc; f < avail_threads; ++f) {
5179         // Adjust barrier data of reserved threads (if any) of the team
5180         // Other data will be set in __kmp_initialize_info() below.
5181         int b;
5182         kmp_balign_t *balign = other_threads[f]->th.th_bar;
5183         for (b = 0; b < bs_last_barrier; ++b) {
5184           balign[b].bb.b_arrived = team->t.t_bar[b].b_arrived;
5185           KMP_DEBUG_ASSERT(balign[b].bb.wait_flag != KMP_BARRIER_PARENT_FLAG);
5186 #if USE_DEBUGGER
5187           balign[b].bb.b_worker_arrived = team->t.t_bar[b].b_team_arrived;
5188 #endif
5189         }
5190       }
5191       if (hot_teams[level].hot_team_nth >= new_nproc) {
5192         // we have all needed threads in reserve, no need to allocate any
5193         // this only possible in mode 1, cannot have reserved threads in mode 0
5194         KMP_DEBUG_ASSERT(__kmp_hot_teams_mode == 1);
5195         team->t.t_nproc = new_nproc; // just get reserved threads involved
5196       } else {
5197         // we may have some threads in reserve, but not enough
5198         team->t.t_nproc =
5199             hot_teams[level]
5200                 .hot_team_nth; // get reserved threads involved if any
5201         hot_teams[level].hot_team_nth = new_nproc; // adjust hot team max size
5202 #endif // KMP_NESTED_HOT_TEAMS
5203         if (team->t.t_max_nproc < new_nproc) {
5204           /* reallocate larger arrays */
5205           __kmp_reallocate_team_arrays(team, new_nproc);
5206           __kmp_reinitialize_team(team, new_icvs, NULL);
5207         }
5208 
5209 #if KMP_OS_LINUX && KMP_AFFINITY_SUPPORTED
5210         /* Temporarily set full mask for master thread before creation of
5211            workers. The reason is that workers inherit the affinity from master,
5212            so if a lot of workers are created on the single core quickly, they
5213            don't get a chance to set their own affinity for a long time. */
5214         __kmp_set_thread_affinity_mask_full_tmp(old_mask);
5215 #endif
5216 
5217         /* allocate new threads for the hot team */
5218         for (f = team->t.t_nproc; f < new_nproc; f++) {
5219           kmp_info_t *new_worker = __kmp_allocate_thread(root, team, f);
5220           KMP_DEBUG_ASSERT(new_worker);
5221           team->t.t_threads[f] = new_worker;
5222 
5223           KA_TRACE(20,
5224                    ("__kmp_allocate_team: team %d init T#%d arrived: "
5225                     "join=%llu, plain=%llu\n",
5226                     team->t.t_id, __kmp_gtid_from_tid(f, team), team->t.t_id, f,
5227                     team->t.t_bar[bs_forkjoin_barrier].b_arrived,
5228                     team->t.t_bar[bs_plain_barrier].b_arrived));
5229 
5230           { // Initialize barrier data for new threads.
5231             int b;
5232             kmp_balign_t *balign = new_worker->th.th_bar;
5233             for (b = 0; b < bs_last_barrier; ++b) {
5234               balign[b].bb.b_arrived = team->t.t_bar[b].b_arrived;
5235               KMP_DEBUG_ASSERT(balign[b].bb.wait_flag !=
5236                                KMP_BARRIER_PARENT_FLAG);
5237 #if USE_DEBUGGER
5238               balign[b].bb.b_worker_arrived = team->t.t_bar[b].b_team_arrived;
5239 #endif
5240             }
5241           }
5242         }
5243 
5244 #if KMP_OS_LINUX && KMP_AFFINITY_SUPPORTED
5245         if (KMP_AFFINITY_CAPABLE()) {
5246           /* Restore initial master thread's affinity mask */
5247           __kmp_set_system_affinity(old_mask, TRUE);
5248           KMP_CPU_FREE(old_mask);
5249         }
5250 #endif
5251 #if KMP_NESTED_HOT_TEAMS
5252       } // end of check of t_nproc vs. new_nproc vs. hot_team_nth
5253 #endif // KMP_NESTED_HOT_TEAMS
5254       /* make sure everyone is syncronized */
5255       int old_nproc = team->t.t_nproc; // save old value and use to update only
5256       // new threads below
5257       __kmp_initialize_team(team, new_nproc, new_icvs,
5258                             root->r.r_uber_thread->th.th_ident);
5259 
5260       /* reinitialize the threads */
5261       KMP_DEBUG_ASSERT(team->t.t_nproc == new_nproc);
5262       for (f = 0; f < team->t.t_nproc; ++f)
5263         __kmp_initialize_info(team->t.t_threads[f], team, f,
5264                               __kmp_gtid_from_tid(f, team));
5265 
5266       if (level) { // set th_task_state for new threads in nested hot team
5267         // __kmp_initialize_info() no longer zeroes th_task_state, so we should
5268         // only need to set the th_task_state for the new threads. th_task_state
5269         // for master thread will not be accurate until after this in
5270         // __kmp_fork_call(), so we look to the master's memo_stack to get the
5271         // correct value.
5272         for (f = old_nproc; f < team->t.t_nproc; ++f)
5273           team->t.t_threads[f]->th.th_task_state =
5274               team->t.t_threads[0]->th.th_task_state_memo_stack[level];
5275       } else { // set th_task_state for new threads in non-nested hot team
5276         int old_state =
5277             team->t.t_threads[0]->th.th_task_state; // copy master's state
5278         for (f = old_nproc; f < team->t.t_nproc; ++f)
5279           team->t.t_threads[f]->th.th_task_state = old_state;
5280       }
5281 
5282 #ifdef KMP_DEBUG
5283       for (f = 0; f < team->t.t_nproc; ++f) {
5284         KMP_DEBUG_ASSERT(team->t.t_threads[f] &&
5285                          team->t.t_threads[f]->th.th_team_nproc ==
5286                              team->t.t_nproc);
5287       }
5288 #endif
5289 
5290 #if OMP_40_ENABLED
5291       KMP_CHECK_UPDATE(team->t.t_proc_bind, new_proc_bind);
5292 #if KMP_AFFINITY_SUPPORTED
5293       __kmp_partition_places(team);
5294 #endif
5295 #endif
5296     } // Check changes in number of threads
5297 
5298 #if OMP_40_ENABLED
5299     kmp_info_t *master = team->t.t_threads[0];
5300     if (master->th.th_teams_microtask) {
5301       for (f = 1; f < new_nproc; ++f) {
5302         // propagate teams construct specific info to workers
5303         kmp_info_t *thr = team->t.t_threads[f];
5304         thr->th.th_teams_microtask = master->th.th_teams_microtask;
5305         thr->th.th_teams_level = master->th.th_teams_level;
5306         thr->th.th_teams_size = master->th.th_teams_size;
5307       }
5308     }
5309 #endif /* OMP_40_ENABLED */
5310 #if KMP_NESTED_HOT_TEAMS
5311     if (level) {
5312       // Sync barrier state for nested hot teams, not needed for outermost hot
5313       // team.
5314       for (f = 1; f < new_nproc; ++f) {
5315         kmp_info_t *thr = team->t.t_threads[f];
5316         int b;
5317         kmp_balign_t *balign = thr->th.th_bar;
5318         for (b = 0; b < bs_last_barrier; ++b) {
5319           balign[b].bb.b_arrived = team->t.t_bar[b].b_arrived;
5320           KMP_DEBUG_ASSERT(balign[b].bb.wait_flag != KMP_BARRIER_PARENT_FLAG);
5321 #if USE_DEBUGGER
5322           balign[b].bb.b_worker_arrived = team->t.t_bar[b].b_team_arrived;
5323 #endif
5324         }
5325       }
5326     }
5327 #endif // KMP_NESTED_HOT_TEAMS
5328 
5329     /* reallocate space for arguments if necessary */
5330     __kmp_alloc_argv_entries(argc, team, TRUE);
5331     KMP_CHECK_UPDATE(team->t.t_argc, argc);
5332     // The hot team re-uses the previous task team,
5333     // if untouched during the previous release->gather phase.
5334 
5335     KF_TRACE(10, (" hot_team = %p\n", team));
5336 
5337 #if KMP_DEBUG
5338     if (__kmp_tasking_mode != tskm_immediate_exec) {
5339       KA_TRACE(20, ("__kmp_allocate_team: hot team task_team[0] = %p "
5340                     "task_team[1] = %p after reinit\n",
5341                     team->t.t_task_team[0], team->t.t_task_team[1]));
5342     }
5343 #endif
5344 
5345 #if OMPT_SUPPORT
5346     __ompt_team_assign_id(team, ompt_parallel_data);
5347 #endif
5348 
5349     KMP_MB();
5350 
5351     return team;
5352   }
5353 
5354   /* next, let's try to take one from the team pool */
5355   KMP_MB();
5356   for (team = CCAST(kmp_team_t *, __kmp_team_pool); (team);) {
5357     /* TODO: consider resizing undersized teams instead of reaping them, now
5358        that we have a resizing mechanism */
5359     if (team->t.t_max_nproc >= max_nproc) {
5360       /* take this team from the team pool */
5361       __kmp_team_pool = team->t.t_next_pool;
5362 
5363       /* setup the team for fresh use */
5364       __kmp_initialize_team(team, new_nproc, new_icvs, NULL);
5365 
5366       KA_TRACE(20, ("__kmp_allocate_team: setting task_team[0] %p and "
5367                     "task_team[1] %p to NULL\n",
5368                     &team->t.t_task_team[0], &team->t.t_task_team[1]));
5369       team->t.t_task_team[0] = NULL;
5370       team->t.t_task_team[1] = NULL;
5371 
5372       /* reallocate space for arguments if necessary */
5373       __kmp_alloc_argv_entries(argc, team, TRUE);
5374       KMP_CHECK_UPDATE(team->t.t_argc, argc);
5375 
5376       KA_TRACE(
5377           20, ("__kmp_allocate_team: team %d init arrived: join=%u, plain=%u\n",
5378                team->t.t_id, KMP_INIT_BARRIER_STATE, KMP_INIT_BARRIER_STATE));
5379       { // Initialize barrier data.
5380         int b;
5381         for (b = 0; b < bs_last_barrier; ++b) {
5382           team->t.t_bar[b].b_arrived = KMP_INIT_BARRIER_STATE;
5383 #if USE_DEBUGGER
5384           team->t.t_bar[b].b_master_arrived = 0;
5385           team->t.t_bar[b].b_team_arrived = 0;
5386 #endif
5387         }
5388       }
5389 
5390 #if OMP_40_ENABLED
5391       team->t.t_proc_bind = new_proc_bind;
5392 #endif
5393 
5394       KA_TRACE(20, ("__kmp_allocate_team: using team from pool %d.\n",
5395                     team->t.t_id));
5396 
5397 #if OMPT_SUPPORT
5398       __ompt_team_assign_id(team, ompt_parallel_data);
5399 #endif
5400 
5401       KMP_MB();
5402 
5403       return team;
5404     }
5405 
5406     /* reap team if it is too small, then loop back and check the next one */
5407     // not sure if this is wise, but, will be redone during the hot-teams
5408     // rewrite.
5409     /* TODO: Use technique to find the right size hot-team, don't reap them */
5410     team = __kmp_reap_team(team);
5411     __kmp_team_pool = team;
5412   }
5413 
5414   /* nothing available in the pool, no matter, make a new team! */
5415   KMP_MB();
5416   team = (kmp_team_t *)__kmp_allocate(sizeof(kmp_team_t));
5417 
5418   /* and set it up */
5419   team->t.t_max_nproc = max_nproc;
5420   /* NOTE well, for some reason allocating one big buffer and dividing it up
5421      seems to really hurt performance a lot on the P4, so, let's not use this */
5422   __kmp_allocate_team_arrays(team, max_nproc);
5423 
5424   KA_TRACE(20, ("__kmp_allocate_team: making a new team\n"));
5425   __kmp_initialize_team(team, new_nproc, new_icvs, NULL);
5426 
5427   KA_TRACE(20, ("__kmp_allocate_team: setting task_team[0] %p and task_team[1] "
5428                 "%p to NULL\n",
5429                 &team->t.t_task_team[0], &team->t.t_task_team[1]));
5430   team->t.t_task_team[0] = NULL; // to be removed, as __kmp_allocate zeroes
5431   // memory, no need to duplicate
5432   team->t.t_task_team[1] = NULL; // to be removed, as __kmp_allocate zeroes
5433   // memory, no need to duplicate
5434 
5435   if (__kmp_storage_map) {
5436     __kmp_print_team_storage_map("team", team, team->t.t_id, new_nproc);
5437   }
5438 
5439   /* allocate space for arguments */
5440   __kmp_alloc_argv_entries(argc, team, FALSE);
5441   team->t.t_argc = argc;
5442 
5443   KA_TRACE(20,
5444            ("__kmp_allocate_team: team %d init arrived: join=%u, plain=%u\n",
5445             team->t.t_id, KMP_INIT_BARRIER_STATE, KMP_INIT_BARRIER_STATE));
5446   { // Initialize barrier data.
5447     int b;
5448     for (b = 0; b < bs_last_barrier; ++b) {
5449       team->t.t_bar[b].b_arrived = KMP_INIT_BARRIER_STATE;
5450 #if USE_DEBUGGER
5451       team->t.t_bar[b].b_master_arrived = 0;
5452       team->t.t_bar[b].b_team_arrived = 0;
5453 #endif
5454     }
5455   }
5456 
5457 #if OMP_40_ENABLED
5458   team->t.t_proc_bind = new_proc_bind;
5459 #endif
5460 
5461 #if OMPT_SUPPORT
5462   __ompt_team_assign_id(team, ompt_parallel_data);
5463   team->t.ompt_serialized_team_info = NULL;
5464 #endif
5465 
5466   KMP_MB();
5467 
5468   KA_TRACE(20, ("__kmp_allocate_team: done creating a new team %d.\n",
5469                 team->t.t_id));
5470 
5471   return team;
5472 }
5473 
5474 /* TODO implement hot-teams at all levels */
5475 /* TODO implement lazy thread release on demand (disband request) */
5476 
5477 /* free the team.  return it to the team pool.  release all the threads
5478  * associated with it */
5479 void __kmp_free_team(kmp_root_t *root,
5480                      kmp_team_t *team USE_NESTED_HOT_ARG(kmp_info_t *master)) {
5481   int f;
5482   KA_TRACE(20, ("__kmp_free_team: T#%d freeing team %d\n", __kmp_get_gtid(),
5483                 team->t.t_id));
5484 
5485   /* verify state */
5486   KMP_DEBUG_ASSERT(root);
5487   KMP_DEBUG_ASSERT(team);
5488   KMP_DEBUG_ASSERT(team->t.t_nproc <= team->t.t_max_nproc);
5489   KMP_DEBUG_ASSERT(team->t.t_threads);
5490 
5491   int use_hot_team = team == root->r.r_hot_team;
5492 #if KMP_NESTED_HOT_TEAMS
5493   int level;
5494   kmp_hot_team_ptr_t *hot_teams;
5495   if (master) {
5496     level = team->t.t_active_level - 1;
5497     if (master->th.th_teams_microtask) { // in teams construct?
5498       if (master->th.th_teams_size.nteams > 1) {
5499         ++level; // level was not increased in teams construct for
5500         // team_of_masters
5501       }
5502       if (team->t.t_pkfn != (microtask_t)__kmp_teams_master &&
5503           master->th.th_teams_level == team->t.t_level) {
5504         ++level; // level was not increased in teams construct for
5505         // team_of_workers before the parallel
5506       } // team->t.t_level will be increased inside parallel
5507     }
5508     hot_teams = master->th.th_hot_teams;
5509     if (level < __kmp_hot_teams_max_level) {
5510       KMP_DEBUG_ASSERT(team == hot_teams[level].hot_team);
5511       use_hot_team = 1;
5512     }
5513   }
5514 #endif // KMP_NESTED_HOT_TEAMS
5515 
5516   /* team is done working */
5517   TCW_SYNC_PTR(team->t.t_pkfn,
5518                NULL); // Important for Debugging Support Library.
5519 #if KMP_OS_WINDOWS
5520   team->t.t_copyin_counter = 0; // init counter for possible reuse
5521 #endif
5522   // Do not reset pointer to parent team to NULL for hot teams.
5523 
5524   /* if we are non-hot team, release our threads */
5525   if (!use_hot_team) {
5526     if (__kmp_tasking_mode != tskm_immediate_exec) {
5527       // Wait for threads to reach reapable state
5528       for (f = 1; f < team->t.t_nproc; ++f) {
5529         KMP_DEBUG_ASSERT(team->t.t_threads[f]);
5530         kmp_info_t *th = team->t.t_threads[f];
5531         volatile kmp_uint32 *state = &th->th.th_reap_state;
5532         while (*state != KMP_SAFE_TO_REAP) {
5533 #if KMP_OS_WINDOWS
5534           // On Windows a thread can be killed at any time, check this
5535           DWORD ecode;
5536           if (!__kmp_is_thread_alive(th, &ecode)) {
5537             *state = KMP_SAFE_TO_REAP; // reset the flag for dead thread
5538             break;
5539           }
5540 #endif
5541           // first check if thread is sleeping
5542           kmp_flag_64 fl(&th->th.th_bar[bs_forkjoin_barrier].bb.b_go, th);
5543           if (fl.is_sleeping())
5544             fl.resume(__kmp_gtid_from_thread(th));
5545           KMP_CPU_PAUSE();
5546         }
5547       }
5548 
5549       // Delete task teams
5550       int tt_idx;
5551       for (tt_idx = 0; tt_idx < 2; ++tt_idx) {
5552         kmp_task_team_t *task_team = team->t.t_task_team[tt_idx];
5553         if (task_team != NULL) {
5554           for (f = 0; f < team->t.t_nproc; ++f) { // threads unref task teams
5555             KMP_DEBUG_ASSERT(team->t.t_threads[f]);
5556             team->t.t_threads[f]->th.th_task_team = NULL;
5557           }
5558           KA_TRACE(
5559               20,
5560               ("__kmp_free_team: T#%d deactivating task_team %p on team %d\n",
5561                __kmp_get_gtid(), task_team, team->t.t_id));
5562 #if KMP_NESTED_HOT_TEAMS
5563           __kmp_free_task_team(master, task_team);
5564 #endif
5565           team->t.t_task_team[tt_idx] = NULL;
5566         }
5567       }
5568     }
5569 
5570     // Reset pointer to parent team only for non-hot teams.
5571     team->t.t_parent = NULL;
5572     team->t.t_level = 0;
5573     team->t.t_active_level = 0;
5574 
5575     /* free the worker threads */
5576     for (f = 1; f < team->t.t_nproc; ++f) {
5577       KMP_DEBUG_ASSERT(team->t.t_threads[f]);
5578       __kmp_free_thread(team->t.t_threads[f]);
5579       team->t.t_threads[f] = NULL;
5580     }
5581 
5582     /* put the team back in the team pool */
5583     /* TODO limit size of team pool, call reap_team if pool too large */
5584     team->t.t_next_pool = CCAST(kmp_team_t *, __kmp_team_pool);
5585     __kmp_team_pool = (volatile kmp_team_t *)team;
5586   } else { // Check if team was created for the masters in a teams construct
5587     // See if first worker is a CG root
5588     KMP_DEBUG_ASSERT(team->t.t_threads[1] &&
5589                      team->t.t_threads[1]->th.th_cg_roots);
5590     if (team->t.t_threads[1]->th.th_cg_roots->cg_root == team->t.t_threads[1]) {
5591       // Clean up the CG root nodes on workers so that this team can be re-used
5592       for (f = 1; f < team->t.t_nproc; ++f) {
5593         kmp_info_t *thr = team->t.t_threads[f];
5594         KMP_DEBUG_ASSERT(thr && thr->th.th_cg_roots &&
5595                          thr->th.th_cg_roots->cg_root == thr);
5596         // Pop current CG root off list
5597         kmp_cg_root_t *tmp = thr->th.th_cg_roots;
5598         thr->th.th_cg_roots = tmp->up;
5599         KA_TRACE(100, ("__kmp_free_team: Thread %p popping node %p and moving"
5600                        " up to node %p. cg_nthreads was %d\n",
5601                        thr, tmp, thr->th.th_cg_roots, tmp->cg_nthreads));
5602         __kmp_free(tmp);
5603         // Restore current task's thread_limit from CG root
5604         if (thr->th.th_cg_roots)
5605           thr->th.th_current_task->td_icvs.thread_limit =
5606               thr->th.th_cg_roots->cg_thread_limit;
5607       }
5608     }
5609   }
5610 
5611   KMP_MB();
5612 }
5613 
5614 /* reap the team.  destroy it, reclaim all its resources and free its memory */
5615 kmp_team_t *__kmp_reap_team(kmp_team_t *team) {
5616   kmp_team_t *next_pool = team->t.t_next_pool;
5617 
5618   KMP_DEBUG_ASSERT(team);
5619   KMP_DEBUG_ASSERT(team->t.t_dispatch);
5620   KMP_DEBUG_ASSERT(team->t.t_disp_buffer);
5621   KMP_DEBUG_ASSERT(team->t.t_threads);
5622   KMP_DEBUG_ASSERT(team->t.t_argv);
5623 
5624   /* TODO clean the threads that are a part of this? */
5625 
5626   /* free stuff */
5627   __kmp_free_team_arrays(team);
5628   if (team->t.t_argv != &team->t.t_inline_argv[0])
5629     __kmp_free((void *)team->t.t_argv);
5630   __kmp_free(team);
5631 
5632   KMP_MB();
5633   return next_pool;
5634 }
5635 
5636 // Free the thread.  Don't reap it, just place it on the pool of available
5637 // threads.
5638 //
5639 // Changes for Quad issue 527845: We need a predictable OMP tid <-> gtid
5640 // binding for the affinity mechanism to be useful.
5641 //
5642 // Now, we always keep the free list (__kmp_thread_pool) sorted by gtid.
5643 // However, we want to avoid a potential performance problem by always
5644 // scanning through the list to find the correct point at which to insert
5645 // the thread (potential N**2 behavior).  To do this we keep track of the
5646 // last place a thread struct was inserted (__kmp_thread_pool_insert_pt).
5647 // With single-level parallelism, threads will always be added to the tail
5648 // of the list, kept track of by __kmp_thread_pool_insert_pt.  With nested
5649 // parallelism, all bets are off and we may need to scan through the entire
5650 // free list.
5651 //
5652 // This change also has a potentially large performance benefit, for some
5653 // applications.  Previously, as threads were freed from the hot team, they
5654 // would be placed back on the free list in inverse order.  If the hot team
5655 // grew back to it's original size, then the freed thread would be placed
5656 // back on the hot team in reverse order.  This could cause bad cache
5657 // locality problems on programs where the size of the hot team regularly
5658 // grew and shrunk.
5659 //
5660 // Now, for single-level parallelism, the OMP tid is alway == gtid.
5661 void __kmp_free_thread(kmp_info_t *this_th) {
5662   int gtid;
5663   kmp_info_t **scan;
5664 
5665   KA_TRACE(20, ("__kmp_free_thread: T#%d putting T#%d back on free pool.\n",
5666                 __kmp_get_gtid(), this_th->th.th_info.ds.ds_gtid));
5667 
5668   KMP_DEBUG_ASSERT(this_th);
5669 
5670   // When moving thread to pool, switch thread to wait on own b_go flag, and
5671   // uninitialized (NULL team).
5672   int b;
5673   kmp_balign_t *balign = this_th->th.th_bar;
5674   for (b = 0; b < bs_last_barrier; ++b) {
5675     if (balign[b].bb.wait_flag == KMP_BARRIER_PARENT_FLAG)
5676       balign[b].bb.wait_flag = KMP_BARRIER_SWITCH_TO_OWN_FLAG;
5677     balign[b].bb.team = NULL;
5678     balign[b].bb.leaf_kids = 0;
5679   }
5680   this_th->th.th_task_state = 0;
5681   this_th->th.th_reap_state = KMP_SAFE_TO_REAP;
5682 
5683   /* put thread back on the free pool */
5684   TCW_PTR(this_th->th.th_team, NULL);
5685   TCW_PTR(this_th->th.th_root, NULL);
5686   TCW_PTR(this_th->th.th_dispatch, NULL); /* NOT NEEDED */
5687 
5688   while (this_th->th.th_cg_roots) {
5689     this_th->th.th_cg_roots->cg_nthreads--;
5690     KA_TRACE(100, ("__kmp_free_thread: Thread %p decrement cg_nthreads on node"
5691                    " %p of thread  %p to %d\n",
5692                    this_th, this_th->th.th_cg_roots,
5693                    this_th->th.th_cg_roots->cg_root,
5694                    this_th->th.th_cg_roots->cg_nthreads));
5695     kmp_cg_root_t *tmp = this_th->th.th_cg_roots;
5696     if (tmp->cg_root == this_th) { // Thread is a cg_root
5697       KMP_DEBUG_ASSERT(tmp->cg_nthreads == 0);
5698       KA_TRACE(
5699           5, ("__kmp_free_thread: Thread %p freeing node %p\n", this_th, tmp));
5700       this_th->th.th_cg_roots = tmp->up;
5701       __kmp_free(tmp);
5702     } else { // Worker thread
5703       this_th->th.th_cg_roots = NULL;
5704       break;
5705     }
5706   }
5707 
5708   /* If the implicit task assigned to this thread can be used by other threads
5709    * -> multiple threads can share the data and try to free the task at
5710    * __kmp_reap_thread at exit. This duplicate use of the task data can happen
5711    * with higher probability when hot team is disabled but can occurs even when
5712    * the hot team is enabled */
5713   __kmp_free_implicit_task(this_th);
5714   this_th->th.th_current_task = NULL;
5715 
5716   // If the __kmp_thread_pool_insert_pt is already past the new insert
5717   // point, then we need to re-scan the entire list.
5718   gtid = this_th->th.th_info.ds.ds_gtid;
5719   if (__kmp_thread_pool_insert_pt != NULL) {
5720     KMP_DEBUG_ASSERT(__kmp_thread_pool != NULL);
5721     if (__kmp_thread_pool_insert_pt->th.th_info.ds.ds_gtid > gtid) {
5722       __kmp_thread_pool_insert_pt = NULL;
5723     }
5724   }
5725 
5726   // Scan down the list to find the place to insert the thread.
5727   // scan is the address of a link in the list, possibly the address of
5728   // __kmp_thread_pool itself.
5729   //
5730   // In the absence of nested parallism, the for loop will have 0 iterations.
5731   if (__kmp_thread_pool_insert_pt != NULL) {
5732     scan = &(__kmp_thread_pool_insert_pt->th.th_next_pool);
5733   } else {
5734     scan = CCAST(kmp_info_t **, &__kmp_thread_pool);
5735   }
5736   for (; (*scan != NULL) && ((*scan)->th.th_info.ds.ds_gtid < gtid);
5737        scan = &((*scan)->th.th_next_pool))
5738     ;
5739 
5740   // Insert the new element on the list, and set __kmp_thread_pool_insert_pt
5741   // to its address.
5742   TCW_PTR(this_th->th.th_next_pool, *scan);
5743   __kmp_thread_pool_insert_pt = *scan = this_th;
5744   KMP_DEBUG_ASSERT((this_th->th.th_next_pool == NULL) ||
5745                    (this_th->th.th_info.ds.ds_gtid <
5746                     this_th->th.th_next_pool->th.th_info.ds.ds_gtid));
5747   TCW_4(this_th->th.th_in_pool, TRUE);
5748   __kmp_suspend_initialize_thread(this_th);
5749   __kmp_lock_suspend_mx(this_th);
5750   if (this_th->th.th_active == TRUE) {
5751     KMP_ATOMIC_INC(&__kmp_thread_pool_active_nth);
5752     this_th->th.th_active_in_pool = TRUE;
5753   }
5754 #if KMP_DEBUG
5755   else {
5756     KMP_DEBUG_ASSERT(this_th->th.th_active_in_pool == FALSE);
5757   }
5758 #endif
5759   __kmp_unlock_suspend_mx(this_th);
5760 
5761   TCW_4(__kmp_nth, __kmp_nth - 1);
5762 
5763 #ifdef KMP_ADJUST_BLOCKTIME
5764   /* Adjust blocktime back to user setting or default if necessary */
5765   /* Middle initialization might never have occurred                */
5766   if (!__kmp_env_blocktime && (__kmp_avail_proc > 0)) {
5767     KMP_DEBUG_ASSERT(__kmp_avail_proc > 0);
5768     if (__kmp_nth <= __kmp_avail_proc) {
5769       __kmp_zero_bt = FALSE;
5770     }
5771   }
5772 #endif /* KMP_ADJUST_BLOCKTIME */
5773 
5774   KMP_MB();
5775 }
5776 
5777 /* ------------------------------------------------------------------------ */
5778 
5779 void *__kmp_launch_thread(kmp_info_t *this_thr) {
5780   int gtid = this_thr->th.th_info.ds.ds_gtid;
5781   /*    void                 *stack_data;*/
5782   kmp_team_t *(*volatile pteam);
5783 
5784   KMP_MB();
5785   KA_TRACE(10, ("__kmp_launch_thread: T#%d start\n", gtid));
5786 
5787   if (__kmp_env_consistency_check) {
5788     this_thr->th.th_cons = __kmp_allocate_cons_stack(gtid); // ATT: Memory leak?
5789   }
5790 
5791 #if OMPT_SUPPORT
5792   ompt_data_t *thread_data;
5793   if (ompt_enabled.enabled) {
5794     thread_data = &(this_thr->th.ompt_thread_info.thread_data);
5795     *thread_data = ompt_data_none;
5796 
5797     this_thr->th.ompt_thread_info.state = ompt_state_overhead;
5798     this_thr->th.ompt_thread_info.wait_id = 0;
5799     this_thr->th.ompt_thread_info.idle_frame = OMPT_GET_FRAME_ADDRESS(0);
5800     if (ompt_enabled.ompt_callback_thread_begin) {
5801       ompt_callbacks.ompt_callback(ompt_callback_thread_begin)(
5802           ompt_thread_worker, thread_data);
5803     }
5804   }
5805 #endif
5806 
5807 #if OMPT_SUPPORT
5808   if (ompt_enabled.enabled) {
5809     this_thr->th.ompt_thread_info.state = ompt_state_idle;
5810   }
5811 #endif
5812   /* This is the place where threads wait for work */
5813   while (!TCR_4(__kmp_global.g.g_done)) {
5814     KMP_DEBUG_ASSERT(this_thr == __kmp_threads[gtid]);
5815     KMP_MB();
5816 
5817     /* wait for work to do */
5818     KA_TRACE(20, ("__kmp_launch_thread: T#%d waiting for work\n", gtid));
5819 
5820     /* No tid yet since not part of a team */
5821     __kmp_fork_barrier(gtid, KMP_GTID_DNE);
5822 
5823 #if OMPT_SUPPORT
5824     if (ompt_enabled.enabled) {
5825       this_thr->th.ompt_thread_info.state = ompt_state_overhead;
5826     }
5827 #endif
5828 
5829     pteam = (kmp_team_t * (*))(&this_thr->th.th_team);
5830 
5831     /* have we been allocated? */
5832     if (TCR_SYNC_PTR(*pteam) && !TCR_4(__kmp_global.g.g_done)) {
5833       /* we were just woken up, so run our new task */
5834       if (TCR_SYNC_PTR((*pteam)->t.t_pkfn) != NULL) {
5835         int rc;
5836         KA_TRACE(20,
5837                  ("__kmp_launch_thread: T#%d(%d:%d) invoke microtask = %p\n",
5838                   gtid, (*pteam)->t.t_id, __kmp_tid_from_gtid(gtid),
5839                   (*pteam)->t.t_pkfn));
5840 
5841         updateHWFPControl(*pteam);
5842 
5843 #if OMPT_SUPPORT
5844         if (ompt_enabled.enabled) {
5845           this_thr->th.ompt_thread_info.state = ompt_state_work_parallel;
5846         }
5847 #endif
5848 
5849         rc = (*pteam)->t.t_invoke(gtid);
5850         KMP_ASSERT(rc);
5851 
5852         KMP_MB();
5853         KA_TRACE(20, ("__kmp_launch_thread: T#%d(%d:%d) done microtask = %p\n",
5854                       gtid, (*pteam)->t.t_id, __kmp_tid_from_gtid(gtid),
5855                       (*pteam)->t.t_pkfn));
5856       }
5857 #if OMPT_SUPPORT
5858       if (ompt_enabled.enabled) {
5859         /* no frame set while outside task */
5860         __ompt_get_task_info_object(0)->frame.exit_frame = ompt_data_none;
5861 
5862         this_thr->th.ompt_thread_info.state = ompt_state_overhead;
5863       }
5864 #endif
5865       /* join barrier after parallel region */
5866       __kmp_join_barrier(gtid);
5867     }
5868   }
5869   TCR_SYNC_PTR((intptr_t)__kmp_global.g.g_done);
5870 
5871 #if OMPT_SUPPORT
5872   if (ompt_enabled.ompt_callback_thread_end) {
5873     ompt_callbacks.ompt_callback(ompt_callback_thread_end)(thread_data);
5874   }
5875 #endif
5876 
5877   this_thr->th.th_task_team = NULL;
5878   /* run the destructors for the threadprivate data for this thread */
5879   __kmp_common_destroy_gtid(gtid);
5880 
5881   KA_TRACE(10, ("__kmp_launch_thread: T#%d done\n", gtid));
5882   KMP_MB();
5883   return this_thr;
5884 }
5885 
5886 /* ------------------------------------------------------------------------ */
5887 
5888 void __kmp_internal_end_dest(void *specific_gtid) {
5889 #if KMP_COMPILER_ICC
5890 #pragma warning(push)
5891 #pragma warning(disable : 810) // conversion from "void *" to "int" may lose
5892 // significant bits
5893 #endif
5894   // Make sure no significant bits are lost
5895   int gtid = (kmp_intptr_t)specific_gtid - 1;
5896 #if KMP_COMPILER_ICC
5897 #pragma warning(pop)
5898 #endif
5899 
5900   KA_TRACE(30, ("__kmp_internal_end_dest: T#%d\n", gtid));
5901   /* NOTE: the gtid is stored as gitd+1 in the thread-local-storage
5902    * this is because 0 is reserved for the nothing-stored case */
5903 
5904   /* josh: One reason for setting the gtid specific data even when it is being
5905      destroyed by pthread is to allow gtid lookup through thread specific data
5906      (__kmp_gtid_get_specific).  Some of the code, especially stat code,
5907      that gets executed in the call to __kmp_internal_end_thread, actually
5908      gets the gtid through the thread specific data.  Setting it here seems
5909      rather inelegant and perhaps wrong, but allows __kmp_internal_end_thread
5910      to run smoothly.
5911      todo: get rid of this after we remove the dependence on
5912      __kmp_gtid_get_specific  */
5913   if (gtid >= 0 && KMP_UBER_GTID(gtid))
5914     __kmp_gtid_set_specific(gtid);
5915 #ifdef KMP_TDATA_GTID
5916   __kmp_gtid = gtid;
5917 #endif
5918   __kmp_internal_end_thread(gtid);
5919 }
5920 
5921 #if KMP_OS_UNIX && KMP_DYNAMIC_LIB
5922 
5923 // 2009-09-08 (lev): It looks the destructor does not work. In simple test cases
5924 // destructors work perfectly, but in real libomp.so I have no evidence it is
5925 // ever called. However, -fini linker option in makefile.mk works fine.
5926 
5927 __attribute__((destructor)) void __kmp_internal_end_dtor(void) {
5928   __kmp_internal_end_atexit();
5929 }
5930 
5931 void __kmp_internal_end_fini(void) { __kmp_internal_end_atexit(); }
5932 
5933 #endif
5934 
5935 /* [Windows] josh: when the atexit handler is called, there may still be more
5936    than one thread alive */
5937 void __kmp_internal_end_atexit(void) {
5938   KA_TRACE(30, ("__kmp_internal_end_atexit\n"));
5939   /* [Windows]
5940      josh: ideally, we want to completely shutdown the library in this atexit
5941      handler, but stat code that depends on thread specific data for gtid fails
5942      because that data becomes unavailable at some point during the shutdown, so
5943      we call __kmp_internal_end_thread instead. We should eventually remove the
5944      dependency on __kmp_get_specific_gtid in the stat code and use
5945      __kmp_internal_end_library to cleanly shutdown the library.
5946 
5947      // TODO: Can some of this comment about GVS be removed?
5948      I suspect that the offending stat code is executed when the calling thread
5949      tries to clean up a dead root thread's data structures, resulting in GVS
5950      code trying to close the GVS structures for that thread, but since the stat
5951      code uses __kmp_get_specific_gtid to get the gtid with the assumption that
5952      the calling thread is cleaning up itself instead of another thread, it get
5953      confused. This happens because allowing a thread to unregister and cleanup
5954      another thread is a recent modification for addressing an issue.
5955      Based on the current design (20050722), a thread may end up
5956      trying to unregister another thread only if thread death does not trigger
5957      the calling of __kmp_internal_end_thread.  For Linux* OS, there is the
5958      thread specific data destructor function to detect thread death. For
5959      Windows dynamic, there is DllMain(THREAD_DETACH). For Windows static, there
5960      is nothing.  Thus, the workaround is applicable only for Windows static
5961      stat library. */
5962   __kmp_internal_end_library(-1);
5963 #if KMP_OS_WINDOWS
5964   __kmp_close_console();
5965 #endif
5966 }
5967 
5968 static void __kmp_reap_thread(kmp_info_t *thread, int is_root) {
5969   // It is assumed __kmp_forkjoin_lock is acquired.
5970 
5971   int gtid;
5972 
5973   KMP_DEBUG_ASSERT(thread != NULL);
5974 
5975   gtid = thread->th.th_info.ds.ds_gtid;
5976 
5977   if (!is_root) {
5978     if (__kmp_dflt_blocktime != KMP_MAX_BLOCKTIME) {
5979       /* Assume the threads are at the fork barrier here */
5980       KA_TRACE(
5981           20, ("__kmp_reap_thread: releasing T#%d from fork barrier for reap\n",
5982                gtid));
5983       /* Need release fence here to prevent seg faults for tree forkjoin barrier
5984        * (GEH) */
5985       ANNOTATE_HAPPENS_BEFORE(thread);
5986       kmp_flag_64 flag(&thread->th.th_bar[bs_forkjoin_barrier].bb.b_go, thread);
5987       __kmp_release_64(&flag);
5988     }
5989 
5990     // Terminate OS thread.
5991     __kmp_reap_worker(thread);
5992 
5993     // The thread was killed asynchronously.  If it was actively
5994     // spinning in the thread pool, decrement the global count.
5995     //
5996     // There is a small timing hole here - if the worker thread was just waking
5997     // up after sleeping in the pool, had reset it's th_active_in_pool flag but
5998     // not decremented the global counter __kmp_thread_pool_active_nth yet, then
5999     // the global counter might not get updated.
6000     //
6001     // Currently, this can only happen as the library is unloaded,
6002     // so there are no harmful side effects.
6003     if (thread->th.th_active_in_pool) {
6004       thread->th.th_active_in_pool = FALSE;
6005       KMP_ATOMIC_DEC(&__kmp_thread_pool_active_nth);
6006       KMP_DEBUG_ASSERT(__kmp_thread_pool_active_nth >= 0);
6007     }
6008   }
6009 
6010   __kmp_free_implicit_task(thread);
6011 
6012 // Free the fast memory for tasking
6013 #if USE_FAST_MEMORY
6014   __kmp_free_fast_memory(thread);
6015 #endif /* USE_FAST_MEMORY */
6016 
6017   __kmp_suspend_uninitialize_thread(thread);
6018 
6019   KMP_DEBUG_ASSERT(__kmp_threads[gtid] == thread);
6020   TCW_SYNC_PTR(__kmp_threads[gtid], NULL);
6021 
6022   --__kmp_all_nth;
6023 // __kmp_nth was decremented when thread is added to the pool.
6024 
6025 #ifdef KMP_ADJUST_BLOCKTIME
6026   /* Adjust blocktime back to user setting or default if necessary */
6027   /* Middle initialization might never have occurred                */
6028   if (!__kmp_env_blocktime && (__kmp_avail_proc > 0)) {
6029     KMP_DEBUG_ASSERT(__kmp_avail_proc > 0);
6030     if (__kmp_nth <= __kmp_avail_proc) {
6031       __kmp_zero_bt = FALSE;
6032     }
6033   }
6034 #endif /* KMP_ADJUST_BLOCKTIME */
6035 
6036   /* free the memory being used */
6037   if (__kmp_env_consistency_check) {
6038     if (thread->th.th_cons) {
6039       __kmp_free_cons_stack(thread->th.th_cons);
6040       thread->th.th_cons = NULL;
6041     }
6042   }
6043 
6044   if (thread->th.th_pri_common != NULL) {
6045     __kmp_free(thread->th.th_pri_common);
6046     thread->th.th_pri_common = NULL;
6047   }
6048 
6049   if (thread->th.th_task_state_memo_stack != NULL) {
6050     __kmp_free(thread->th.th_task_state_memo_stack);
6051     thread->th.th_task_state_memo_stack = NULL;
6052   }
6053 
6054 #if KMP_USE_BGET
6055   if (thread->th.th_local.bget_data != NULL) {
6056     __kmp_finalize_bget(thread);
6057   }
6058 #endif
6059 
6060 #if KMP_AFFINITY_SUPPORTED
6061   if (thread->th.th_affin_mask != NULL) {
6062     KMP_CPU_FREE(thread->th.th_affin_mask);
6063     thread->th.th_affin_mask = NULL;
6064   }
6065 #endif /* KMP_AFFINITY_SUPPORTED */
6066 
6067 #if KMP_USE_HIER_SCHED
6068   if (thread->th.th_hier_bar_data != NULL) {
6069     __kmp_free(thread->th.th_hier_bar_data);
6070     thread->th.th_hier_bar_data = NULL;
6071   }
6072 #endif
6073 
6074   __kmp_reap_team(thread->th.th_serial_team);
6075   thread->th.th_serial_team = NULL;
6076   __kmp_free(thread);
6077 
6078   KMP_MB();
6079 
6080 } // __kmp_reap_thread
6081 
6082 static void __kmp_internal_end(void) {
6083   int i;
6084 
6085   /* First, unregister the library */
6086   __kmp_unregister_library();
6087 
6088 #if KMP_OS_WINDOWS
6089   /* In Win static library, we can't tell when a root actually dies, so we
6090      reclaim the data structures for any root threads that have died but not
6091      unregistered themselves, in order to shut down cleanly.
6092      In Win dynamic library we also can't tell when a thread dies.  */
6093   __kmp_reclaim_dead_roots(); // AC: moved here to always clean resources of
6094 // dead roots
6095 #endif
6096 
6097   for (i = 0; i < __kmp_threads_capacity; i++)
6098     if (__kmp_root[i])
6099       if (__kmp_root[i]->r.r_active)
6100         break;
6101   KMP_MB(); /* Flush all pending memory write invalidates.  */
6102   TCW_SYNC_4(__kmp_global.g.g_done, TRUE);
6103 
6104   if (i < __kmp_threads_capacity) {
6105 #if KMP_USE_MONITOR
6106     // 2009-09-08 (lev): Other alive roots found. Why do we kill the monitor??
6107     KMP_MB(); /* Flush all pending memory write invalidates.  */
6108 
6109     // Need to check that monitor was initialized before reaping it. If we are
6110     // called form __kmp_atfork_child (which sets __kmp_init_parallel = 0), then
6111     // __kmp_monitor will appear to contain valid data, but it is only valid in
6112     // the parent process, not the child.
6113     // New behavior (201008): instead of keying off of the flag
6114     // __kmp_init_parallel, the monitor thread creation is keyed off
6115     // of the new flag __kmp_init_monitor.
6116     __kmp_acquire_bootstrap_lock(&__kmp_monitor_lock);
6117     if (TCR_4(__kmp_init_monitor)) {
6118       __kmp_reap_monitor(&__kmp_monitor);
6119       TCW_4(__kmp_init_monitor, 0);
6120     }
6121     __kmp_release_bootstrap_lock(&__kmp_monitor_lock);
6122     KA_TRACE(10, ("__kmp_internal_end: monitor reaped\n"));
6123 #endif // KMP_USE_MONITOR
6124   } else {
6125 /* TODO move this to cleanup code */
6126 #ifdef KMP_DEBUG
6127     /* make sure that everything has properly ended */
6128     for (i = 0; i < __kmp_threads_capacity; i++) {
6129       if (__kmp_root[i]) {
6130         //                    KMP_ASSERT( ! KMP_UBER_GTID( i ) );         // AC:
6131         //                    there can be uber threads alive here
6132         KMP_ASSERT(!__kmp_root[i]->r.r_active); // TODO: can they be active?
6133       }
6134     }
6135 #endif
6136 
6137     KMP_MB();
6138 
6139     // Reap the worker threads.
6140     // This is valid for now, but be careful if threads are reaped sooner.
6141     while (__kmp_thread_pool != NULL) { // Loop thru all the thread in the pool.
6142       // Get the next thread from the pool.
6143       kmp_info_t *thread = CCAST(kmp_info_t *, __kmp_thread_pool);
6144       __kmp_thread_pool = thread->th.th_next_pool;
6145       // Reap it.
6146       KMP_DEBUG_ASSERT(thread->th.th_reap_state == KMP_SAFE_TO_REAP);
6147       thread->th.th_next_pool = NULL;
6148       thread->th.th_in_pool = FALSE;
6149       __kmp_reap_thread(thread, 0);
6150     }
6151     __kmp_thread_pool_insert_pt = NULL;
6152 
6153     // Reap teams.
6154     while (__kmp_team_pool != NULL) { // Loop thru all the teams in the pool.
6155       // Get the next team from the pool.
6156       kmp_team_t *team = CCAST(kmp_team_t *, __kmp_team_pool);
6157       __kmp_team_pool = team->t.t_next_pool;
6158       // Reap it.
6159       team->t.t_next_pool = NULL;
6160       __kmp_reap_team(team);
6161     }
6162 
6163     __kmp_reap_task_teams();
6164 
6165 #if KMP_OS_UNIX
6166     // Threads that are not reaped should not access any resources since they
6167     // are going to be deallocated soon, so the shutdown sequence should wait
6168     // until all threads either exit the final spin-waiting loop or begin
6169     // sleeping after the given blocktime.
6170     for (i = 0; i < __kmp_threads_capacity; i++) {
6171       kmp_info_t *thr = __kmp_threads[i];
6172       while (thr && KMP_ATOMIC_LD_ACQ(&thr->th.th_blocking))
6173         KMP_CPU_PAUSE();
6174     }
6175 #endif
6176 
6177     for (i = 0; i < __kmp_threads_capacity; ++i) {
6178       // TBD: Add some checking...
6179       // Something like KMP_DEBUG_ASSERT( __kmp_thread[ i ] == NULL );
6180     }
6181 
6182     /* Make sure all threadprivate destructors get run by joining with all
6183        worker threads before resetting this flag */
6184     TCW_SYNC_4(__kmp_init_common, FALSE);
6185 
6186     KA_TRACE(10, ("__kmp_internal_end: all workers reaped\n"));
6187     KMP_MB();
6188 
6189 #if KMP_USE_MONITOR
6190     // See note above: One of the possible fixes for CQ138434 / CQ140126
6191     //
6192     // FIXME: push both code fragments down and CSE them?
6193     // push them into __kmp_cleanup() ?
6194     __kmp_acquire_bootstrap_lock(&__kmp_monitor_lock);
6195     if (TCR_4(__kmp_init_monitor)) {
6196       __kmp_reap_monitor(&__kmp_monitor);
6197       TCW_4(__kmp_init_monitor, 0);
6198     }
6199     __kmp_release_bootstrap_lock(&__kmp_monitor_lock);
6200     KA_TRACE(10, ("__kmp_internal_end: monitor reaped\n"));
6201 #endif
6202   } /* else !__kmp_global.t_active */
6203   TCW_4(__kmp_init_gtid, FALSE);
6204   KMP_MB(); /* Flush all pending memory write invalidates.  */
6205 
6206   __kmp_cleanup();
6207 #if OMPT_SUPPORT
6208   ompt_fini();
6209 #endif
6210 }
6211 
6212 void __kmp_internal_end_library(int gtid_req) {
6213   /* if we have already cleaned up, don't try again, it wouldn't be pretty */
6214   /* this shouldn't be a race condition because __kmp_internal_end() is the
6215      only place to clear __kmp_serial_init */
6216   /* we'll check this later too, after we get the lock */
6217   // 2009-09-06: We do not set g_abort without setting g_done. This check looks
6218   // redundaant, because the next check will work in any case.
6219   if (__kmp_global.g.g_abort) {
6220     KA_TRACE(11, ("__kmp_internal_end_library: abort, exiting\n"));
6221     /* TODO abort? */
6222     return;
6223   }
6224   if (TCR_4(__kmp_global.g.g_done) || !__kmp_init_serial) {
6225     KA_TRACE(10, ("__kmp_internal_end_library: already finished\n"));
6226     return;
6227   }
6228 
6229   KMP_MB(); /* Flush all pending memory write invalidates.  */
6230 
6231   /* find out who we are and what we should do */
6232   {
6233     int gtid = (gtid_req >= 0) ? gtid_req : __kmp_gtid_get_specific();
6234     KA_TRACE(
6235         10, ("__kmp_internal_end_library: enter T#%d  (%d)\n", gtid, gtid_req));
6236     if (gtid == KMP_GTID_SHUTDOWN) {
6237       KA_TRACE(10, ("__kmp_internal_end_library: !__kmp_init_runtime, system "
6238                     "already shutdown\n"));
6239       return;
6240     } else if (gtid == KMP_GTID_MONITOR) {
6241       KA_TRACE(10, ("__kmp_internal_end_library: monitor thread, gtid not "
6242                     "registered, or system shutdown\n"));
6243       return;
6244     } else if (gtid == KMP_GTID_DNE) {
6245       KA_TRACE(10, ("__kmp_internal_end_library: gtid not registered or system "
6246                     "shutdown\n"));
6247       /* we don't know who we are, but we may still shutdown the library */
6248     } else if (KMP_UBER_GTID(gtid)) {
6249       /* unregister ourselves as an uber thread.  gtid is no longer valid */
6250       if (__kmp_root[gtid]->r.r_active) {
6251         __kmp_global.g.g_abort = -1;
6252         TCW_SYNC_4(__kmp_global.g.g_done, TRUE);
6253         KA_TRACE(10,
6254                  ("__kmp_internal_end_library: root still active, abort T#%d\n",
6255                   gtid));
6256         return;
6257       } else {
6258         KA_TRACE(
6259             10,
6260             ("__kmp_internal_end_library: unregistering sibling T#%d\n", gtid));
6261         __kmp_unregister_root_current_thread(gtid);
6262       }
6263     } else {
6264 /* worker threads may call this function through the atexit handler, if they
6265  * call exit() */
6266 /* For now, skip the usual subsequent processing and just dump the debug buffer.
6267    TODO: do a thorough shutdown instead */
6268 #ifdef DUMP_DEBUG_ON_EXIT
6269       if (__kmp_debug_buf)
6270         __kmp_dump_debug_buffer();
6271 #endif
6272       return;
6273     }
6274   }
6275   /* synchronize the termination process */
6276   __kmp_acquire_bootstrap_lock(&__kmp_initz_lock);
6277 
6278   /* have we already finished */
6279   if (__kmp_global.g.g_abort) {
6280     KA_TRACE(10, ("__kmp_internal_end_library: abort, exiting\n"));
6281     /* TODO abort? */
6282     __kmp_release_bootstrap_lock(&__kmp_initz_lock);
6283     return;
6284   }
6285   if (TCR_4(__kmp_global.g.g_done) || !__kmp_init_serial) {
6286     __kmp_release_bootstrap_lock(&__kmp_initz_lock);
6287     return;
6288   }
6289 
6290   /* We need this lock to enforce mutex between this reading of
6291      __kmp_threads_capacity and the writing by __kmp_register_root.
6292      Alternatively, we can use a counter of roots that is atomically updated by
6293      __kmp_get_global_thread_id_reg, __kmp_do_serial_initialize and
6294      __kmp_internal_end_*.  */
6295   __kmp_acquire_bootstrap_lock(&__kmp_forkjoin_lock);
6296 
6297   /* now we can safely conduct the actual termination */
6298   __kmp_internal_end();
6299 
6300   __kmp_release_bootstrap_lock(&__kmp_forkjoin_lock);
6301   __kmp_release_bootstrap_lock(&__kmp_initz_lock);
6302 
6303   KA_TRACE(10, ("__kmp_internal_end_library: exit\n"));
6304 
6305 #ifdef DUMP_DEBUG_ON_EXIT
6306   if (__kmp_debug_buf)
6307     __kmp_dump_debug_buffer();
6308 #endif
6309 
6310 #if KMP_OS_WINDOWS
6311   __kmp_close_console();
6312 #endif
6313 
6314   __kmp_fini_allocator();
6315 
6316 } // __kmp_internal_end_library
6317 
6318 void __kmp_internal_end_thread(int gtid_req) {
6319   int i;
6320 
6321   /* if we have already cleaned up, don't try again, it wouldn't be pretty */
6322   /* this shouldn't be a race condition because __kmp_internal_end() is the
6323    * only place to clear __kmp_serial_init */
6324   /* we'll check this later too, after we get the lock */
6325   // 2009-09-06: We do not set g_abort without setting g_done. This check looks
6326   // redundant, because the next check will work in any case.
6327   if (__kmp_global.g.g_abort) {
6328     KA_TRACE(11, ("__kmp_internal_end_thread: abort, exiting\n"));
6329     /* TODO abort? */
6330     return;
6331   }
6332   if (TCR_4(__kmp_global.g.g_done) || !__kmp_init_serial) {
6333     KA_TRACE(10, ("__kmp_internal_end_thread: already finished\n"));
6334     return;
6335   }
6336 
6337   KMP_MB(); /* Flush all pending memory write invalidates.  */
6338 
6339   /* find out who we are and what we should do */
6340   {
6341     int gtid = (gtid_req >= 0) ? gtid_req : __kmp_gtid_get_specific();
6342     KA_TRACE(10,
6343              ("__kmp_internal_end_thread: enter T#%d  (%d)\n", gtid, gtid_req));
6344     if (gtid == KMP_GTID_SHUTDOWN) {
6345       KA_TRACE(10, ("__kmp_internal_end_thread: !__kmp_init_runtime, system "
6346                     "already shutdown\n"));
6347       return;
6348     } else if (gtid == KMP_GTID_MONITOR) {
6349       KA_TRACE(10, ("__kmp_internal_end_thread: monitor thread, gtid not "
6350                     "registered, or system shutdown\n"));
6351       return;
6352     } else if (gtid == KMP_GTID_DNE) {
6353       KA_TRACE(10, ("__kmp_internal_end_thread: gtid not registered or system "
6354                     "shutdown\n"));
6355       return;
6356       /* we don't know who we are */
6357     } else if (KMP_UBER_GTID(gtid)) {
6358       /* unregister ourselves as an uber thread.  gtid is no longer valid */
6359       if (__kmp_root[gtid]->r.r_active) {
6360         __kmp_global.g.g_abort = -1;
6361         TCW_SYNC_4(__kmp_global.g.g_done, TRUE);
6362         KA_TRACE(10,
6363                  ("__kmp_internal_end_thread: root still active, abort T#%d\n",
6364                   gtid));
6365         return;
6366       } else {
6367         KA_TRACE(10, ("__kmp_internal_end_thread: unregistering sibling T#%d\n",
6368                       gtid));
6369         __kmp_unregister_root_current_thread(gtid);
6370       }
6371     } else {
6372       /* just a worker thread, let's leave */
6373       KA_TRACE(10, ("__kmp_internal_end_thread: worker thread T#%d\n", gtid));
6374 
6375       if (gtid >= 0) {
6376         __kmp_threads[gtid]->th.th_task_team = NULL;
6377       }
6378 
6379       KA_TRACE(10,
6380                ("__kmp_internal_end_thread: worker thread done, exiting T#%d\n",
6381                 gtid));
6382       return;
6383     }
6384   }
6385 #if KMP_DYNAMIC_LIB
6386 #if OMP_50_ENABLED
6387   if (__kmp_pause_status != kmp_hard_paused)
6388 #endif
6389   // AC: lets not shutdown the dynamic library at the exit of uber thread,
6390   // because we will better shutdown later in the library destructor.
6391   {
6392     KA_TRACE(10, ("__kmp_internal_end_thread: exiting T#%d\n", gtid_req));
6393     return;
6394   }
6395 #endif
6396   /* synchronize the termination process */
6397   __kmp_acquire_bootstrap_lock(&__kmp_initz_lock);
6398 
6399   /* have we already finished */
6400   if (__kmp_global.g.g_abort) {
6401     KA_TRACE(10, ("__kmp_internal_end_thread: abort, exiting\n"));
6402     /* TODO abort? */
6403     __kmp_release_bootstrap_lock(&__kmp_initz_lock);
6404     return;
6405   }
6406   if (TCR_4(__kmp_global.g.g_done) || !__kmp_init_serial) {
6407     __kmp_release_bootstrap_lock(&__kmp_initz_lock);
6408     return;
6409   }
6410 
6411   /* We need this lock to enforce mutex between this reading of
6412      __kmp_threads_capacity and the writing by __kmp_register_root.
6413      Alternatively, we can use a counter of roots that is atomically updated by
6414      __kmp_get_global_thread_id_reg, __kmp_do_serial_initialize and
6415      __kmp_internal_end_*.  */
6416 
6417   /* should we finish the run-time?  are all siblings done? */
6418   __kmp_acquire_bootstrap_lock(&__kmp_forkjoin_lock);
6419 
6420   for (i = 0; i < __kmp_threads_capacity; ++i) {
6421     if (KMP_UBER_GTID(i)) {
6422       KA_TRACE(
6423           10,
6424           ("__kmp_internal_end_thread: remaining sibling task: gtid==%d\n", i));
6425       __kmp_release_bootstrap_lock(&__kmp_forkjoin_lock);
6426       __kmp_release_bootstrap_lock(&__kmp_initz_lock);
6427       return;
6428     }
6429   }
6430 
6431   /* now we can safely conduct the actual termination */
6432 
6433   __kmp_internal_end();
6434 
6435   __kmp_release_bootstrap_lock(&__kmp_forkjoin_lock);
6436   __kmp_release_bootstrap_lock(&__kmp_initz_lock);
6437 
6438   KA_TRACE(10, ("__kmp_internal_end_thread: exit T#%d\n", gtid_req));
6439 
6440 #ifdef DUMP_DEBUG_ON_EXIT
6441   if (__kmp_debug_buf)
6442     __kmp_dump_debug_buffer();
6443 #endif
6444 } // __kmp_internal_end_thread
6445 
6446 // -----------------------------------------------------------------------------
6447 // Library registration stuff.
6448 
6449 static long __kmp_registration_flag = 0;
6450 // Random value used to indicate library initialization.
6451 static char *__kmp_registration_str = NULL;
6452 // Value to be saved in env var __KMP_REGISTERED_LIB_<pid>.
6453 
6454 static inline char *__kmp_reg_status_name() {
6455   /* On RHEL 3u5 if linked statically, getpid() returns different values in
6456      each thread. If registration and unregistration go in different threads
6457      (omp_misc_other_root_exit.cpp test case), the name of registered_lib_env
6458      env var can not be found, because the name will contain different pid. */
6459   return __kmp_str_format("__KMP_REGISTERED_LIB_%d", (int)getpid());
6460 } // __kmp_reg_status_get
6461 
6462 void __kmp_register_library_startup(void) {
6463 
6464   char *name = __kmp_reg_status_name(); // Name of the environment variable.
6465   int done = 0;
6466   union {
6467     double dtime;
6468     long ltime;
6469   } time;
6470 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
6471   __kmp_initialize_system_tick();
6472 #endif
6473   __kmp_read_system_time(&time.dtime);
6474   __kmp_registration_flag = 0xCAFE0000L | (time.ltime & 0x0000FFFFL);
6475   __kmp_registration_str =
6476       __kmp_str_format("%p-%lx-%s", &__kmp_registration_flag,
6477                        __kmp_registration_flag, KMP_LIBRARY_FILE);
6478 
6479   KA_TRACE(50, ("__kmp_register_library_startup: %s=\"%s\"\n", name,
6480                 __kmp_registration_str));
6481 
6482   while (!done) {
6483 
6484     char *value = NULL; // Actual value of the environment variable.
6485 
6486     // Set environment variable, but do not overwrite if it is exist.
6487     __kmp_env_set(name, __kmp_registration_str, 0);
6488     // Check the variable is written.
6489     value = __kmp_env_get(name);
6490     if (value != NULL && strcmp(value, __kmp_registration_str) == 0) {
6491 
6492       done = 1; // Ok, environment variable set successfully, exit the loop.
6493 
6494     } else {
6495 
6496       // Oops. Write failed. Another copy of OpenMP RTL is in memory.
6497       // Check whether it alive or dead.
6498       int neighbor = 0; // 0 -- unknown status, 1 -- alive, 2 -- dead.
6499       char *tail = value;
6500       char *flag_addr_str = NULL;
6501       char *flag_val_str = NULL;
6502       char const *file_name = NULL;
6503       __kmp_str_split(tail, '-', &flag_addr_str, &tail);
6504       __kmp_str_split(tail, '-', &flag_val_str, &tail);
6505       file_name = tail;
6506       if (tail != NULL) {
6507         long *flag_addr = 0;
6508         long flag_val = 0;
6509         KMP_SSCANF(flag_addr_str, "%p", RCAST(void**, &flag_addr));
6510         KMP_SSCANF(flag_val_str, "%lx", &flag_val);
6511         if (flag_addr != 0 && flag_val != 0 && strcmp(file_name, "") != 0) {
6512           // First, check whether environment-encoded address is mapped into
6513           // addr space.
6514           // If so, dereference it to see if it still has the right value.
6515           if (__kmp_is_address_mapped(flag_addr) && *flag_addr == flag_val) {
6516             neighbor = 1;
6517           } else {
6518             // If not, then we know the other copy of the library is no longer
6519             // running.
6520             neighbor = 2;
6521           }
6522         }
6523       }
6524       switch (neighbor) {
6525       case 0: // Cannot parse environment variable -- neighbor status unknown.
6526         // Assume it is the incompatible format of future version of the
6527         // library. Assume the other library is alive.
6528         // WARN( ... ); // TODO: Issue a warning.
6529         file_name = "unknown library";
6530         KMP_FALLTHROUGH();
6531       // Attention! Falling to the next case. That's intentional.
6532       case 1: { // Neighbor is alive.
6533         // Check it is allowed.
6534         char *duplicate_ok = __kmp_env_get("KMP_DUPLICATE_LIB_OK");
6535         if (!__kmp_str_match_true(duplicate_ok)) {
6536           // That's not allowed. Issue fatal error.
6537           __kmp_fatal(KMP_MSG(DuplicateLibrary, KMP_LIBRARY_FILE, file_name),
6538                       KMP_HNT(DuplicateLibrary), __kmp_msg_null);
6539         }
6540         KMP_INTERNAL_FREE(duplicate_ok);
6541         __kmp_duplicate_library_ok = 1;
6542         done = 1; // Exit the loop.
6543       } break;
6544       case 2: { // Neighbor is dead.
6545         // Clear the variable and try to register library again.
6546         __kmp_env_unset(name);
6547       } break;
6548       default: { KMP_DEBUG_ASSERT(0); } break;
6549       }
6550     }
6551     KMP_INTERNAL_FREE((void *)value);
6552   }
6553   KMP_INTERNAL_FREE((void *)name);
6554 
6555 } // func __kmp_register_library_startup
6556 
6557 void __kmp_unregister_library(void) {
6558 
6559   char *name = __kmp_reg_status_name();
6560   char *value = __kmp_env_get(name);
6561 
6562   KMP_DEBUG_ASSERT(__kmp_registration_flag != 0);
6563   KMP_DEBUG_ASSERT(__kmp_registration_str != NULL);
6564   if (value != NULL && strcmp(value, __kmp_registration_str) == 0) {
6565     // Ok, this is our variable. Delete it.
6566     __kmp_env_unset(name);
6567   }
6568 
6569   KMP_INTERNAL_FREE(__kmp_registration_str);
6570   KMP_INTERNAL_FREE(value);
6571   KMP_INTERNAL_FREE(name);
6572 
6573   __kmp_registration_flag = 0;
6574   __kmp_registration_str = NULL;
6575 
6576 } // __kmp_unregister_library
6577 
6578 // End of Library registration stuff.
6579 // -----------------------------------------------------------------------------
6580 
6581 #if KMP_MIC_SUPPORTED
6582 
6583 static void __kmp_check_mic_type() {
6584   kmp_cpuid_t cpuid_state = {0};
6585   kmp_cpuid_t *cs_p = &cpuid_state;
6586   __kmp_x86_cpuid(1, 0, cs_p);
6587   // We don't support mic1 at the moment
6588   if ((cs_p->eax & 0xff0) == 0xB10) {
6589     __kmp_mic_type = mic2;
6590   } else if ((cs_p->eax & 0xf0ff0) == 0x50670) {
6591     __kmp_mic_type = mic3;
6592   } else {
6593     __kmp_mic_type = non_mic;
6594   }
6595 }
6596 
6597 #endif /* KMP_MIC_SUPPORTED */
6598 
6599 static void __kmp_do_serial_initialize(void) {
6600   int i, gtid;
6601   int size;
6602 
6603   KA_TRACE(10, ("__kmp_do_serial_initialize: enter\n"));
6604 
6605   KMP_DEBUG_ASSERT(sizeof(kmp_int32) == 4);
6606   KMP_DEBUG_ASSERT(sizeof(kmp_uint32) == 4);
6607   KMP_DEBUG_ASSERT(sizeof(kmp_int64) == 8);
6608   KMP_DEBUG_ASSERT(sizeof(kmp_uint64) == 8);
6609   KMP_DEBUG_ASSERT(sizeof(kmp_intptr_t) == sizeof(void *));
6610 
6611 #if OMPT_SUPPORT
6612   ompt_pre_init();
6613 #endif
6614 
6615   __kmp_validate_locks();
6616 
6617   /* Initialize internal memory allocator */
6618   __kmp_init_allocator();
6619 
6620   /* Register the library startup via an environment variable and check to see
6621      whether another copy of the library is already registered. */
6622 
6623   __kmp_register_library_startup();
6624 
6625   /* TODO reinitialization of library */
6626   if (TCR_4(__kmp_global.g.g_done)) {
6627     KA_TRACE(10, ("__kmp_do_serial_initialize: reinitialization of library\n"));
6628   }
6629 
6630   __kmp_global.g.g_abort = 0;
6631   TCW_SYNC_4(__kmp_global.g.g_done, FALSE);
6632 
6633 /* initialize the locks */
6634 #if KMP_USE_ADAPTIVE_LOCKS
6635 #if KMP_DEBUG_ADAPTIVE_LOCKS
6636   __kmp_init_speculative_stats();
6637 #endif
6638 #endif
6639 #if KMP_STATS_ENABLED
6640   __kmp_stats_init();
6641 #endif
6642   __kmp_init_lock(&__kmp_global_lock);
6643   __kmp_init_queuing_lock(&__kmp_dispatch_lock);
6644   __kmp_init_lock(&__kmp_debug_lock);
6645   __kmp_init_atomic_lock(&__kmp_atomic_lock);
6646   __kmp_init_atomic_lock(&__kmp_atomic_lock_1i);
6647   __kmp_init_atomic_lock(&__kmp_atomic_lock_2i);
6648   __kmp_init_atomic_lock(&__kmp_atomic_lock_4i);
6649   __kmp_init_atomic_lock(&__kmp_atomic_lock_4r);
6650   __kmp_init_atomic_lock(&__kmp_atomic_lock_8i);
6651   __kmp_init_atomic_lock(&__kmp_atomic_lock_8r);
6652   __kmp_init_atomic_lock(&__kmp_atomic_lock_8c);
6653   __kmp_init_atomic_lock(&__kmp_atomic_lock_10r);
6654   __kmp_init_atomic_lock(&__kmp_atomic_lock_16r);
6655   __kmp_init_atomic_lock(&__kmp_atomic_lock_16c);
6656   __kmp_init_atomic_lock(&__kmp_atomic_lock_20c);
6657   __kmp_init_atomic_lock(&__kmp_atomic_lock_32c);
6658   __kmp_init_bootstrap_lock(&__kmp_forkjoin_lock);
6659   __kmp_init_bootstrap_lock(&__kmp_exit_lock);
6660 #if KMP_USE_MONITOR
6661   __kmp_init_bootstrap_lock(&__kmp_monitor_lock);
6662 #endif
6663   __kmp_init_bootstrap_lock(&__kmp_tp_cached_lock);
6664 
6665   /* conduct initialization and initial setup of configuration */
6666 
6667   __kmp_runtime_initialize();
6668 
6669 #if KMP_MIC_SUPPORTED
6670   __kmp_check_mic_type();
6671 #endif
6672 
6673 // Some global variable initialization moved here from kmp_env_initialize()
6674 #ifdef KMP_DEBUG
6675   kmp_diag = 0;
6676 #endif
6677   __kmp_abort_delay = 0;
6678 
6679   // From __kmp_init_dflt_team_nth()
6680   /* assume the entire machine will be used */
6681   __kmp_dflt_team_nth_ub = __kmp_xproc;
6682   if (__kmp_dflt_team_nth_ub < KMP_MIN_NTH) {
6683     __kmp_dflt_team_nth_ub = KMP_MIN_NTH;
6684   }
6685   if (__kmp_dflt_team_nth_ub > __kmp_sys_max_nth) {
6686     __kmp_dflt_team_nth_ub = __kmp_sys_max_nth;
6687   }
6688   __kmp_max_nth = __kmp_sys_max_nth;
6689   __kmp_cg_max_nth = __kmp_sys_max_nth;
6690   __kmp_teams_max_nth = __kmp_xproc; // set a "reasonable" default
6691   if (__kmp_teams_max_nth > __kmp_sys_max_nth) {
6692     __kmp_teams_max_nth = __kmp_sys_max_nth;
6693   }
6694 
6695   // Three vars below moved here from __kmp_env_initialize() "KMP_BLOCKTIME"
6696   // part
6697   __kmp_dflt_blocktime = KMP_DEFAULT_BLOCKTIME;
6698 #if KMP_USE_MONITOR
6699   __kmp_monitor_wakeups =
6700       KMP_WAKEUPS_FROM_BLOCKTIME(__kmp_dflt_blocktime, __kmp_monitor_wakeups);
6701   __kmp_bt_intervals =
6702       KMP_INTERVALS_FROM_BLOCKTIME(__kmp_dflt_blocktime, __kmp_monitor_wakeups);
6703 #endif
6704   // From "KMP_LIBRARY" part of __kmp_env_initialize()
6705   __kmp_library = library_throughput;
6706   // From KMP_SCHEDULE initialization
6707   __kmp_static = kmp_sch_static_balanced;
6708 // AC: do not use analytical here, because it is non-monotonous
6709 //__kmp_guided = kmp_sch_guided_iterative_chunked;
6710 //__kmp_auto = kmp_sch_guided_analytical_chunked; // AC: it is the default, no
6711 // need to repeat assignment
6712 // Barrier initialization. Moved here from __kmp_env_initialize() Barrier branch
6713 // bit control and barrier method control parts
6714 #if KMP_FAST_REDUCTION_BARRIER
6715 #define kmp_reduction_barrier_gather_bb ((int)1)
6716 #define kmp_reduction_barrier_release_bb ((int)1)
6717 #define kmp_reduction_barrier_gather_pat bp_hyper_bar
6718 #define kmp_reduction_barrier_release_pat bp_hyper_bar
6719 #endif // KMP_FAST_REDUCTION_BARRIER
6720   for (i = bs_plain_barrier; i < bs_last_barrier; i++) {
6721     __kmp_barrier_gather_branch_bits[i] = __kmp_barrier_gather_bb_dflt;
6722     __kmp_barrier_release_branch_bits[i] = __kmp_barrier_release_bb_dflt;
6723     __kmp_barrier_gather_pattern[i] = __kmp_barrier_gather_pat_dflt;
6724     __kmp_barrier_release_pattern[i] = __kmp_barrier_release_pat_dflt;
6725 #if KMP_FAST_REDUCTION_BARRIER
6726     if (i == bs_reduction_barrier) { // tested and confirmed on ALTIX only (
6727       // lin_64 ): hyper,1
6728       __kmp_barrier_gather_branch_bits[i] = kmp_reduction_barrier_gather_bb;
6729       __kmp_barrier_release_branch_bits[i] = kmp_reduction_barrier_release_bb;
6730       __kmp_barrier_gather_pattern[i] = kmp_reduction_barrier_gather_pat;
6731       __kmp_barrier_release_pattern[i] = kmp_reduction_barrier_release_pat;
6732     }
6733 #endif // KMP_FAST_REDUCTION_BARRIER
6734   }
6735 #if KMP_FAST_REDUCTION_BARRIER
6736 #undef kmp_reduction_barrier_release_pat
6737 #undef kmp_reduction_barrier_gather_pat
6738 #undef kmp_reduction_barrier_release_bb
6739 #undef kmp_reduction_barrier_gather_bb
6740 #endif // KMP_FAST_REDUCTION_BARRIER
6741 #if KMP_MIC_SUPPORTED
6742   if (__kmp_mic_type == mic2) { // KNC
6743     // AC: plane=3,2, forkjoin=2,1 are optimal for 240 threads on KNC
6744     __kmp_barrier_gather_branch_bits[bs_plain_barrier] = 3; // plain gather
6745     __kmp_barrier_release_branch_bits[bs_forkjoin_barrier] =
6746         1; // forkjoin release
6747     __kmp_barrier_gather_pattern[bs_forkjoin_barrier] = bp_hierarchical_bar;
6748     __kmp_barrier_release_pattern[bs_forkjoin_barrier] = bp_hierarchical_bar;
6749   }
6750 #if KMP_FAST_REDUCTION_BARRIER
6751   if (__kmp_mic_type == mic2) { // KNC
6752     __kmp_barrier_gather_pattern[bs_reduction_barrier] = bp_hierarchical_bar;
6753     __kmp_barrier_release_pattern[bs_reduction_barrier] = bp_hierarchical_bar;
6754   }
6755 #endif // KMP_FAST_REDUCTION_BARRIER
6756 #endif // KMP_MIC_SUPPORTED
6757 
6758 // From KMP_CHECKS initialization
6759 #ifdef KMP_DEBUG
6760   __kmp_env_checks = TRUE; /* development versions have the extra checks */
6761 #else
6762   __kmp_env_checks = FALSE; /* port versions do not have the extra checks */
6763 #endif
6764 
6765   // From "KMP_FOREIGN_THREADS_THREADPRIVATE" initialization
6766   __kmp_foreign_tp = TRUE;
6767 
6768   __kmp_global.g.g_dynamic = FALSE;
6769   __kmp_global.g.g_dynamic_mode = dynamic_default;
6770 
6771   __kmp_env_initialize(NULL);
6772 
6773 // Print all messages in message catalog for testing purposes.
6774 #ifdef KMP_DEBUG
6775   char const *val = __kmp_env_get("KMP_DUMP_CATALOG");
6776   if (__kmp_str_match_true(val)) {
6777     kmp_str_buf_t buffer;
6778     __kmp_str_buf_init(&buffer);
6779     __kmp_i18n_dump_catalog(&buffer);
6780     __kmp_printf("%s", buffer.str);
6781     __kmp_str_buf_free(&buffer);
6782   }
6783   __kmp_env_free(&val);
6784 #endif
6785 
6786   __kmp_threads_capacity =
6787       __kmp_initial_threads_capacity(__kmp_dflt_team_nth_ub);
6788   // Moved here from __kmp_env_initialize() "KMP_ALL_THREADPRIVATE" part
6789   __kmp_tp_capacity = __kmp_default_tp_capacity(
6790       __kmp_dflt_team_nth_ub, __kmp_max_nth, __kmp_allThreadsSpecified);
6791 
6792   // If the library is shut down properly, both pools must be NULL. Just in
6793   // case, set them to NULL -- some memory may leak, but subsequent code will
6794   // work even if pools are not freed.
6795   KMP_DEBUG_ASSERT(__kmp_thread_pool == NULL);
6796   KMP_DEBUG_ASSERT(__kmp_thread_pool_insert_pt == NULL);
6797   KMP_DEBUG_ASSERT(__kmp_team_pool == NULL);
6798   __kmp_thread_pool = NULL;
6799   __kmp_thread_pool_insert_pt = NULL;
6800   __kmp_team_pool = NULL;
6801 
6802   /* Allocate all of the variable sized records */
6803   /* NOTE: __kmp_threads_capacity entries are allocated, but the arrays are
6804    * expandable */
6805   /* Since allocation is cache-aligned, just add extra padding at the end */
6806   size =
6807       (sizeof(kmp_info_t *) + sizeof(kmp_root_t *)) * __kmp_threads_capacity +
6808       CACHE_LINE;
6809   __kmp_threads = (kmp_info_t **)__kmp_allocate(size);
6810   __kmp_root = (kmp_root_t **)((char *)__kmp_threads +
6811                                sizeof(kmp_info_t *) * __kmp_threads_capacity);
6812 
6813   /* init thread counts */
6814   KMP_DEBUG_ASSERT(__kmp_all_nth ==
6815                    0); // Asserts fail if the library is reinitializing and
6816   KMP_DEBUG_ASSERT(__kmp_nth == 0); // something was wrong in termination.
6817   __kmp_all_nth = 0;
6818   __kmp_nth = 0;
6819 
6820   /* setup the uber master thread and hierarchy */
6821   gtid = __kmp_register_root(TRUE);
6822   KA_TRACE(10, ("__kmp_do_serial_initialize  T#%d\n", gtid));
6823   KMP_ASSERT(KMP_UBER_GTID(gtid));
6824   KMP_ASSERT(KMP_INITIAL_GTID(gtid));
6825 
6826   KMP_MB(); /* Flush all pending memory write invalidates.  */
6827 
6828   __kmp_common_initialize();
6829 
6830 #if KMP_OS_UNIX
6831   /* invoke the child fork handler */
6832   __kmp_register_atfork();
6833 #endif
6834 
6835 #if !KMP_DYNAMIC_LIB
6836   {
6837     /* Invoke the exit handler when the program finishes, only for static
6838        library. For dynamic library, we already have _fini and DllMain. */
6839     int rc = atexit(__kmp_internal_end_atexit);
6840     if (rc != 0) {
6841       __kmp_fatal(KMP_MSG(FunctionError, "atexit()"), KMP_ERR(rc),
6842                   __kmp_msg_null);
6843     }
6844   }
6845 #endif
6846 
6847 #if KMP_HANDLE_SIGNALS
6848 #if KMP_OS_UNIX
6849   /* NOTE: make sure that this is called before the user installs their own
6850      signal handlers so that the user handlers are called first. this way they
6851      can return false, not call our handler, avoid terminating the library, and
6852      continue execution where they left off. */
6853   __kmp_install_signals(FALSE);
6854 #endif /* KMP_OS_UNIX */
6855 #if KMP_OS_WINDOWS
6856   __kmp_install_signals(TRUE);
6857 #endif /* KMP_OS_WINDOWS */
6858 #endif
6859 
6860   /* we have finished the serial initialization */
6861   __kmp_init_counter++;
6862 
6863   __kmp_init_serial = TRUE;
6864 
6865   if (__kmp_settings) {
6866     __kmp_env_print();
6867   }
6868 
6869 #if OMP_40_ENABLED
6870   if (__kmp_display_env || __kmp_display_env_verbose) {
6871     __kmp_env_print_2();
6872   }
6873 #endif // OMP_40_ENABLED
6874 
6875 #if OMPT_SUPPORT
6876   ompt_post_init();
6877 #endif
6878 
6879   KMP_MB();
6880 
6881   KA_TRACE(10, ("__kmp_do_serial_initialize: exit\n"));
6882 }
6883 
6884 void __kmp_serial_initialize(void) {
6885   if (__kmp_init_serial) {
6886     return;
6887   }
6888   __kmp_acquire_bootstrap_lock(&__kmp_initz_lock);
6889   if (__kmp_init_serial) {
6890     __kmp_release_bootstrap_lock(&__kmp_initz_lock);
6891     return;
6892   }
6893   __kmp_do_serial_initialize();
6894   __kmp_release_bootstrap_lock(&__kmp_initz_lock);
6895 }
6896 
6897 static void __kmp_do_middle_initialize(void) {
6898   int i, j;
6899   int prev_dflt_team_nth;
6900 
6901   if (!__kmp_init_serial) {
6902     __kmp_do_serial_initialize();
6903   }
6904 
6905   KA_TRACE(10, ("__kmp_middle_initialize: enter\n"));
6906 
6907   // Save the previous value for the __kmp_dflt_team_nth so that
6908   // we can avoid some reinitialization if it hasn't changed.
6909   prev_dflt_team_nth = __kmp_dflt_team_nth;
6910 
6911 #if KMP_AFFINITY_SUPPORTED
6912   // __kmp_affinity_initialize() will try to set __kmp_ncores to the
6913   // number of cores on the machine.
6914   __kmp_affinity_initialize();
6915 
6916   // Run through the __kmp_threads array and set the affinity mask
6917   // for each root thread that is currently registered with the RTL.
6918   for (i = 0; i < __kmp_threads_capacity; i++) {
6919     if (TCR_PTR(__kmp_threads[i]) != NULL) {
6920       __kmp_affinity_set_init_mask(i, TRUE);
6921     }
6922   }
6923 #endif /* KMP_AFFINITY_SUPPORTED */
6924 
6925   KMP_ASSERT(__kmp_xproc > 0);
6926   if (__kmp_avail_proc == 0) {
6927     __kmp_avail_proc = __kmp_xproc;
6928   }
6929 
6930   // If there were empty places in num_threads list (OMP_NUM_THREADS=,,2,3),
6931   // correct them now
6932   j = 0;
6933   while ((j < __kmp_nested_nth.used) && !__kmp_nested_nth.nth[j]) {
6934     __kmp_nested_nth.nth[j] = __kmp_dflt_team_nth = __kmp_dflt_team_nth_ub =
6935         __kmp_avail_proc;
6936     j++;
6937   }
6938 
6939   if (__kmp_dflt_team_nth == 0) {
6940 #ifdef KMP_DFLT_NTH_CORES
6941     // Default #threads = #cores
6942     __kmp_dflt_team_nth = __kmp_ncores;
6943     KA_TRACE(20, ("__kmp_middle_initialize: setting __kmp_dflt_team_nth = "
6944                   "__kmp_ncores (%d)\n",
6945                   __kmp_dflt_team_nth));
6946 #else
6947     // Default #threads = #available OS procs
6948     __kmp_dflt_team_nth = __kmp_avail_proc;
6949     KA_TRACE(20, ("__kmp_middle_initialize: setting __kmp_dflt_team_nth = "
6950                   "__kmp_avail_proc(%d)\n",
6951                   __kmp_dflt_team_nth));
6952 #endif /* KMP_DFLT_NTH_CORES */
6953   }
6954 
6955   if (__kmp_dflt_team_nth < KMP_MIN_NTH) {
6956     __kmp_dflt_team_nth = KMP_MIN_NTH;
6957   }
6958   if (__kmp_dflt_team_nth > __kmp_sys_max_nth) {
6959     __kmp_dflt_team_nth = __kmp_sys_max_nth;
6960   }
6961 
6962   // There's no harm in continuing if the following check fails,
6963   // but it indicates an error in the previous logic.
6964   KMP_DEBUG_ASSERT(__kmp_dflt_team_nth <= __kmp_dflt_team_nth_ub);
6965 
6966   if (__kmp_dflt_team_nth != prev_dflt_team_nth) {
6967     // Run through the __kmp_threads array and set the num threads icv for each
6968     // root thread that is currently registered with the RTL (which has not
6969     // already explicitly set its nthreads-var with a call to
6970     // omp_set_num_threads()).
6971     for (i = 0; i < __kmp_threads_capacity; i++) {
6972       kmp_info_t *thread = __kmp_threads[i];
6973       if (thread == NULL)
6974         continue;
6975       if (thread->th.th_current_task->td_icvs.nproc != 0)
6976         continue;
6977 
6978       set__nproc(__kmp_threads[i], __kmp_dflt_team_nth);
6979     }
6980   }
6981   KA_TRACE(
6982       20,
6983       ("__kmp_middle_initialize: final value for __kmp_dflt_team_nth = %d\n",
6984        __kmp_dflt_team_nth));
6985 
6986 #ifdef KMP_ADJUST_BLOCKTIME
6987   /* Adjust blocktime to zero if necessary  now that __kmp_avail_proc is set */
6988   if (!__kmp_env_blocktime && (__kmp_avail_proc > 0)) {
6989     KMP_DEBUG_ASSERT(__kmp_avail_proc > 0);
6990     if (__kmp_nth > __kmp_avail_proc) {
6991       __kmp_zero_bt = TRUE;
6992     }
6993   }
6994 #endif /* KMP_ADJUST_BLOCKTIME */
6995 
6996   /* we have finished middle initialization */
6997   TCW_SYNC_4(__kmp_init_middle, TRUE);
6998 
6999   KA_TRACE(10, ("__kmp_do_middle_initialize: exit\n"));
7000 }
7001 
7002 void __kmp_middle_initialize(void) {
7003   if (__kmp_init_middle) {
7004     return;
7005   }
7006   __kmp_acquire_bootstrap_lock(&__kmp_initz_lock);
7007   if (__kmp_init_middle) {
7008     __kmp_release_bootstrap_lock(&__kmp_initz_lock);
7009     return;
7010   }
7011   __kmp_do_middle_initialize();
7012   __kmp_release_bootstrap_lock(&__kmp_initz_lock);
7013 }
7014 
7015 void __kmp_parallel_initialize(void) {
7016   int gtid = __kmp_entry_gtid(); // this might be a new root
7017 
7018   /* synchronize parallel initialization (for sibling) */
7019   if (TCR_4(__kmp_init_parallel))
7020     return;
7021   __kmp_acquire_bootstrap_lock(&__kmp_initz_lock);
7022   if (TCR_4(__kmp_init_parallel)) {
7023     __kmp_release_bootstrap_lock(&__kmp_initz_lock);
7024     return;
7025   }
7026 
7027   /* TODO reinitialization after we have already shut down */
7028   if (TCR_4(__kmp_global.g.g_done)) {
7029     KA_TRACE(
7030         10,
7031         ("__kmp_parallel_initialize: attempt to init while shutting down\n"));
7032     __kmp_infinite_loop();
7033   }
7034 
7035   /* jc: The lock __kmp_initz_lock is already held, so calling
7036      __kmp_serial_initialize would cause a deadlock.  So we call
7037      __kmp_do_serial_initialize directly. */
7038   if (!__kmp_init_middle) {
7039     __kmp_do_middle_initialize();
7040   }
7041 
7042 #if OMP_50_ENABLED
7043   __kmp_resume_if_hard_paused();
7044 #endif
7045 
7046   /* begin initialization */
7047   KA_TRACE(10, ("__kmp_parallel_initialize: enter\n"));
7048   KMP_ASSERT(KMP_UBER_GTID(gtid));
7049 
7050 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
7051   // Save the FP control regs.
7052   // Worker threads will set theirs to these values at thread startup.
7053   __kmp_store_x87_fpu_control_word(&__kmp_init_x87_fpu_control_word);
7054   __kmp_store_mxcsr(&__kmp_init_mxcsr);
7055   __kmp_init_mxcsr &= KMP_X86_MXCSR_MASK;
7056 #endif /* KMP_ARCH_X86 || KMP_ARCH_X86_64 */
7057 
7058 #if KMP_OS_UNIX
7059 #if KMP_HANDLE_SIGNALS
7060   /*  must be after __kmp_serial_initialize  */
7061   __kmp_install_signals(TRUE);
7062 #endif
7063 #endif
7064 
7065   __kmp_suspend_initialize();
7066 
7067 #if defined(USE_LOAD_BALANCE)
7068   if (__kmp_global.g.g_dynamic_mode == dynamic_default) {
7069     __kmp_global.g.g_dynamic_mode = dynamic_load_balance;
7070   }
7071 #else
7072   if (__kmp_global.g.g_dynamic_mode == dynamic_default) {
7073     __kmp_global.g.g_dynamic_mode = dynamic_thread_limit;
7074   }
7075 #endif
7076 
7077   if (__kmp_version) {
7078     __kmp_print_version_2();
7079   }
7080 
7081   /* we have finished parallel initialization */
7082   TCW_SYNC_4(__kmp_init_parallel, TRUE);
7083 
7084   KMP_MB();
7085   KA_TRACE(10, ("__kmp_parallel_initialize: exit\n"));
7086 
7087   __kmp_release_bootstrap_lock(&__kmp_initz_lock);
7088 }
7089 
7090 /* ------------------------------------------------------------------------ */
7091 
7092 void __kmp_run_before_invoked_task(int gtid, int tid, kmp_info_t *this_thr,
7093                                    kmp_team_t *team) {
7094   kmp_disp_t *dispatch;
7095 
7096   KMP_MB();
7097 
7098   /* none of the threads have encountered any constructs, yet. */
7099   this_thr->th.th_local.this_construct = 0;
7100 #if KMP_CACHE_MANAGE
7101   KMP_CACHE_PREFETCH(&this_thr->th.th_bar[bs_forkjoin_barrier].bb.b_arrived);
7102 #endif /* KMP_CACHE_MANAGE */
7103   dispatch = (kmp_disp_t *)TCR_PTR(this_thr->th.th_dispatch);
7104   KMP_DEBUG_ASSERT(dispatch);
7105   KMP_DEBUG_ASSERT(team->t.t_dispatch);
7106   // KMP_DEBUG_ASSERT( this_thr->th.th_dispatch == &team->t.t_dispatch[
7107   // this_thr->th.th_info.ds.ds_tid ] );
7108 
7109   dispatch->th_disp_index = 0; /* reset the dispatch buffer counter */
7110 #if OMP_45_ENABLED
7111   dispatch->th_doacross_buf_idx =
7112       0; /* reset the doacross dispatch buffer counter */
7113 #endif
7114   if (__kmp_env_consistency_check)
7115     __kmp_push_parallel(gtid, team->t.t_ident);
7116 
7117   KMP_MB(); /* Flush all pending memory write invalidates.  */
7118 }
7119 
7120 void __kmp_run_after_invoked_task(int gtid, int tid, kmp_info_t *this_thr,
7121                                   kmp_team_t *team) {
7122   if (__kmp_env_consistency_check)
7123     __kmp_pop_parallel(gtid, team->t.t_ident);
7124 
7125   __kmp_finish_implicit_task(this_thr);
7126 }
7127 
7128 int __kmp_invoke_task_func(int gtid) {
7129   int rc;
7130   int tid = __kmp_tid_from_gtid(gtid);
7131   kmp_info_t *this_thr = __kmp_threads[gtid];
7132   kmp_team_t *team = this_thr->th.th_team;
7133 
7134   __kmp_run_before_invoked_task(gtid, tid, this_thr, team);
7135 #if USE_ITT_BUILD
7136   if (__itt_stack_caller_create_ptr) {
7137     __kmp_itt_stack_callee_enter(
7138         (__itt_caller)
7139             team->t.t_stack_id); // inform ittnotify about entering user's code
7140   }
7141 #endif /* USE_ITT_BUILD */
7142 #if INCLUDE_SSC_MARKS
7143   SSC_MARK_INVOKING();
7144 #endif
7145 
7146 #if OMPT_SUPPORT
7147   void *dummy;
7148   void **exit_runtime_p;
7149   ompt_data_t *my_task_data;
7150   ompt_data_t *my_parallel_data;
7151   int ompt_team_size;
7152 
7153   if (ompt_enabled.enabled) {
7154     exit_runtime_p = &(
7155         team->t.t_implicit_task_taskdata[tid].ompt_task_info.frame.exit_frame.ptr);
7156   } else {
7157     exit_runtime_p = &dummy;
7158   }
7159 
7160   my_task_data =
7161       &(team->t.t_implicit_task_taskdata[tid].ompt_task_info.task_data);
7162   my_parallel_data = &(team->t.ompt_team_info.parallel_data);
7163   if (ompt_enabled.ompt_callback_implicit_task) {
7164     ompt_team_size = team->t.t_nproc;
7165     ompt_callbacks.ompt_callback(ompt_callback_implicit_task)(
7166         ompt_scope_begin, my_parallel_data, my_task_data, ompt_team_size,
7167         __kmp_tid_from_gtid(gtid), ompt_task_implicit); // TODO: Can this be ompt_task_initial?
7168     OMPT_CUR_TASK_INFO(this_thr)->thread_num = __kmp_tid_from_gtid(gtid);
7169   }
7170 #endif
7171 
7172 #if KMP_STATS_ENABLED
7173   stats_state_e previous_state = KMP_GET_THREAD_STATE();
7174   if (previous_state == stats_state_e::TEAMS_REGION) {
7175     KMP_PUSH_PARTITIONED_TIMER(OMP_teams);
7176   } else {
7177     KMP_PUSH_PARTITIONED_TIMER(OMP_parallel);
7178   }
7179   KMP_SET_THREAD_STATE(IMPLICIT_TASK);
7180 #endif
7181 
7182   rc = __kmp_invoke_microtask((microtask_t)TCR_SYNC_PTR(team->t.t_pkfn), gtid,
7183                               tid, (int)team->t.t_argc, (void **)team->t.t_argv
7184 #if OMPT_SUPPORT
7185                               ,
7186                               exit_runtime_p
7187 #endif
7188                               );
7189 #if OMPT_SUPPORT
7190   *exit_runtime_p = NULL;
7191 #endif
7192 
7193 #if KMP_STATS_ENABLED
7194   if (previous_state == stats_state_e::TEAMS_REGION) {
7195     KMP_SET_THREAD_STATE(previous_state);
7196   }
7197   KMP_POP_PARTITIONED_TIMER();
7198 #endif
7199 
7200 #if USE_ITT_BUILD
7201   if (__itt_stack_caller_create_ptr) {
7202     __kmp_itt_stack_callee_leave(
7203         (__itt_caller)
7204             team->t.t_stack_id); // inform ittnotify about leaving user's code
7205   }
7206 #endif /* USE_ITT_BUILD */
7207   __kmp_run_after_invoked_task(gtid, tid, this_thr, team);
7208 
7209   return rc;
7210 }
7211 
7212 #if OMP_40_ENABLED
7213 void __kmp_teams_master(int gtid) {
7214   // This routine is called by all master threads in teams construct
7215   kmp_info_t *thr = __kmp_threads[gtid];
7216   kmp_team_t *team = thr->th.th_team;
7217   ident_t *loc = team->t.t_ident;
7218   thr->th.th_set_nproc = thr->th.th_teams_size.nth;
7219   KMP_DEBUG_ASSERT(thr->th.th_teams_microtask);
7220   KMP_DEBUG_ASSERT(thr->th.th_set_nproc);
7221   KA_TRACE(20, ("__kmp_teams_master: T#%d, Tid %d, microtask %p\n", gtid,
7222                 __kmp_tid_from_gtid(gtid), thr->th.th_teams_microtask));
7223 
7224   // This thread is a new CG root.  Set up the proper variables.
7225   kmp_cg_root_t *tmp = (kmp_cg_root_t *)__kmp_allocate(sizeof(kmp_cg_root_t));
7226   tmp->cg_root = thr; // Make thr the CG root
7227   // Init to thread limit that was stored when league masters were forked
7228   tmp->cg_thread_limit = thr->th.th_current_task->td_icvs.thread_limit;
7229   tmp->cg_nthreads = 1; // Init counter to one active thread, this one
7230   KA_TRACE(100, ("__kmp_teams_master: Thread %p created node %p and init"
7231                  " cg_threads to 1\n",
7232                  thr, tmp));
7233   tmp->up = thr->th.th_cg_roots;
7234   thr->th.th_cg_roots = tmp;
7235 
7236 // Launch league of teams now, but not let workers execute
7237 // (they hang on fork barrier until next parallel)
7238 #if INCLUDE_SSC_MARKS
7239   SSC_MARK_FORKING();
7240 #endif
7241   __kmp_fork_call(loc, gtid, fork_context_intel, team->t.t_argc,
7242                   (microtask_t)thr->th.th_teams_microtask, // "wrapped" task
7243                   VOLATILE_CAST(launch_t) __kmp_invoke_task_func, NULL);
7244 #if INCLUDE_SSC_MARKS
7245   SSC_MARK_JOINING();
7246 #endif
7247   // If the team size was reduced from the limit, set it to the new size
7248   if (thr->th.th_team_nproc < thr->th.th_teams_size.nth)
7249     thr->th.th_teams_size.nth = thr->th.th_team_nproc;
7250   // AC: last parameter "1" eliminates join barrier which won't work because
7251   // worker threads are in a fork barrier waiting for more parallel regions
7252   __kmp_join_call(loc, gtid
7253 #if OMPT_SUPPORT
7254                   ,
7255                   fork_context_intel
7256 #endif
7257                   ,
7258                   1);
7259 }
7260 
7261 int __kmp_invoke_teams_master(int gtid) {
7262   kmp_info_t *this_thr = __kmp_threads[gtid];
7263   kmp_team_t *team = this_thr->th.th_team;
7264 #if KMP_DEBUG
7265   if (!__kmp_threads[gtid]->th.th_team->t.t_serialized)
7266     KMP_DEBUG_ASSERT((void *)__kmp_threads[gtid]->th.th_team->t.t_pkfn ==
7267                      (void *)__kmp_teams_master);
7268 #endif
7269   __kmp_run_before_invoked_task(gtid, 0, this_thr, team);
7270   __kmp_teams_master(gtid);
7271   __kmp_run_after_invoked_task(gtid, 0, this_thr, team);
7272   return 1;
7273 }
7274 #endif /* OMP_40_ENABLED */
7275 
7276 /* this sets the requested number of threads for the next parallel region
7277    encountered by this team. since this should be enclosed in the forkjoin
7278    critical section it should avoid race conditions with assymmetrical nested
7279    parallelism */
7280 
7281 void __kmp_push_num_threads(ident_t *id, int gtid, int num_threads) {
7282   kmp_info_t *thr = __kmp_threads[gtid];
7283 
7284   if (num_threads > 0)
7285     thr->th.th_set_nproc = num_threads;
7286 }
7287 
7288 #if OMP_40_ENABLED
7289 
7290 /* this sets the requested number of teams for the teams region and/or
7291    the number of threads for the next parallel region encountered  */
7292 void __kmp_push_num_teams(ident_t *id, int gtid, int num_teams,
7293                           int num_threads) {
7294   kmp_info_t *thr = __kmp_threads[gtid];
7295   KMP_DEBUG_ASSERT(num_teams >= 0);
7296   KMP_DEBUG_ASSERT(num_threads >= 0);
7297 
7298   if (num_teams == 0)
7299     num_teams = 1; // default number of teams is 1.
7300   if (num_teams > __kmp_teams_max_nth) { // if too many teams requested?
7301     if (!__kmp_reserve_warn) {
7302       __kmp_reserve_warn = 1;
7303       __kmp_msg(kmp_ms_warning,
7304                 KMP_MSG(CantFormThrTeam, num_teams, __kmp_teams_max_nth),
7305                 KMP_HNT(Unset_ALL_THREADS), __kmp_msg_null);
7306     }
7307     num_teams = __kmp_teams_max_nth;
7308   }
7309   // Set number of teams (number of threads in the outer "parallel" of the
7310   // teams)
7311   thr->th.th_set_nproc = thr->th.th_teams_size.nteams = num_teams;
7312 
7313   // Remember the number of threads for inner parallel regions
7314   if (num_threads == 0) {
7315     if (!TCR_4(__kmp_init_middle))
7316       __kmp_middle_initialize(); // get __kmp_avail_proc calculated
7317     num_threads = __kmp_avail_proc / num_teams;
7318     if (num_teams * num_threads > __kmp_teams_max_nth) {
7319       // adjust num_threads w/o warning as it is not user setting
7320       num_threads = __kmp_teams_max_nth / num_teams;
7321     }
7322   } else {
7323     // This thread will be the master of the league masters
7324     // Store new thread limit; old limit is saved in th_cg_roots list
7325     thr->th.th_current_task->td_icvs.thread_limit = num_threads;
7326 
7327     if (num_teams * num_threads > __kmp_teams_max_nth) {
7328       int new_threads = __kmp_teams_max_nth / num_teams;
7329       if (!__kmp_reserve_warn) { // user asked for too many threads
7330         __kmp_reserve_warn = 1; // conflicts with KMP_TEAMS_THREAD_LIMIT
7331         __kmp_msg(kmp_ms_warning,
7332                   KMP_MSG(CantFormThrTeam, num_threads, new_threads),
7333                   KMP_HNT(Unset_ALL_THREADS), __kmp_msg_null);
7334       }
7335       num_threads = new_threads;
7336     }
7337   }
7338   thr->th.th_teams_size.nth = num_threads;
7339 }
7340 
7341 // Set the proc_bind var to use in the following parallel region.
7342 void __kmp_push_proc_bind(ident_t *id, int gtid, kmp_proc_bind_t proc_bind) {
7343   kmp_info_t *thr = __kmp_threads[gtid];
7344   thr->th.th_set_proc_bind = proc_bind;
7345 }
7346 
7347 #endif /* OMP_40_ENABLED */
7348 
7349 /* Launch the worker threads into the microtask. */
7350 
7351 void __kmp_internal_fork(ident_t *id, int gtid, kmp_team_t *team) {
7352   kmp_info_t *this_thr = __kmp_threads[gtid];
7353 
7354 #ifdef KMP_DEBUG
7355   int f;
7356 #endif /* KMP_DEBUG */
7357 
7358   KMP_DEBUG_ASSERT(team);
7359   KMP_DEBUG_ASSERT(this_thr->th.th_team == team);
7360   KMP_ASSERT(KMP_MASTER_GTID(gtid));
7361   KMP_MB(); /* Flush all pending memory write invalidates.  */
7362 
7363   team->t.t_construct = 0; /* no single directives seen yet */
7364   team->t.t_ordered.dt.t_value =
7365       0; /* thread 0 enters the ordered section first */
7366 
7367   /* Reset the identifiers on the dispatch buffer */
7368   KMP_DEBUG_ASSERT(team->t.t_disp_buffer);
7369   if (team->t.t_max_nproc > 1) {
7370     int i;
7371     for (i = 0; i < __kmp_dispatch_num_buffers; ++i) {
7372       team->t.t_disp_buffer[i].buffer_index = i;
7373 #if OMP_45_ENABLED
7374       team->t.t_disp_buffer[i].doacross_buf_idx = i;
7375 #endif
7376     }
7377   } else {
7378     team->t.t_disp_buffer[0].buffer_index = 0;
7379 #if OMP_45_ENABLED
7380     team->t.t_disp_buffer[0].doacross_buf_idx = 0;
7381 #endif
7382   }
7383 
7384   KMP_MB(); /* Flush all pending memory write invalidates.  */
7385   KMP_ASSERT(this_thr->th.th_team == team);
7386 
7387 #ifdef KMP_DEBUG
7388   for (f = 0; f < team->t.t_nproc; f++) {
7389     KMP_DEBUG_ASSERT(team->t.t_threads[f] &&
7390                      team->t.t_threads[f]->th.th_team_nproc == team->t.t_nproc);
7391   }
7392 #endif /* KMP_DEBUG */
7393 
7394   /* release the worker threads so they may begin working */
7395   __kmp_fork_barrier(gtid, 0);
7396 }
7397 
7398 void __kmp_internal_join(ident_t *id, int gtid, kmp_team_t *team) {
7399   kmp_info_t *this_thr = __kmp_threads[gtid];
7400 
7401   KMP_DEBUG_ASSERT(team);
7402   KMP_DEBUG_ASSERT(this_thr->th.th_team == team);
7403   KMP_ASSERT(KMP_MASTER_GTID(gtid));
7404   KMP_MB(); /* Flush all pending memory write invalidates.  */
7405 
7406 /* Join barrier after fork */
7407 
7408 #ifdef KMP_DEBUG
7409   if (__kmp_threads[gtid] &&
7410       __kmp_threads[gtid]->th.th_team_nproc != team->t.t_nproc) {
7411     __kmp_printf("GTID: %d, __kmp_threads[%d]=%p\n", gtid, gtid,
7412                  __kmp_threads[gtid]);
7413     __kmp_printf("__kmp_threads[%d]->th.th_team_nproc=%d, TEAM: %p, "
7414                  "team->t.t_nproc=%d\n",
7415                  gtid, __kmp_threads[gtid]->th.th_team_nproc, team,
7416                  team->t.t_nproc);
7417     __kmp_print_structure();
7418   }
7419   KMP_DEBUG_ASSERT(__kmp_threads[gtid] &&
7420                    __kmp_threads[gtid]->th.th_team_nproc == team->t.t_nproc);
7421 #endif /* KMP_DEBUG */
7422 
7423   __kmp_join_barrier(gtid); /* wait for everyone */
7424 #if OMPT_SUPPORT
7425   if (ompt_enabled.enabled &&
7426       this_thr->th.ompt_thread_info.state == ompt_state_wait_barrier_implicit) {
7427     int ds_tid = this_thr->th.th_info.ds.ds_tid;
7428     ompt_data_t *task_data = OMPT_CUR_TASK_DATA(this_thr);
7429     this_thr->th.ompt_thread_info.state = ompt_state_overhead;
7430 #if OMPT_OPTIONAL
7431     void *codeptr = NULL;
7432     if (KMP_MASTER_TID(ds_tid) &&
7433         (ompt_callbacks.ompt_callback(ompt_callback_sync_region_wait) ||
7434          ompt_callbacks.ompt_callback(ompt_callback_sync_region)))
7435       codeptr = OMPT_CUR_TEAM_INFO(this_thr)->master_return_address;
7436 
7437     if (ompt_enabled.ompt_callback_sync_region_wait) {
7438       ompt_callbacks.ompt_callback(ompt_callback_sync_region_wait)(
7439           ompt_sync_region_barrier_implicit, ompt_scope_end, NULL, task_data,
7440           codeptr);
7441     }
7442     if (ompt_enabled.ompt_callback_sync_region) {
7443       ompt_callbacks.ompt_callback(ompt_callback_sync_region)(
7444           ompt_sync_region_barrier_implicit, ompt_scope_end, NULL, task_data,
7445           codeptr);
7446     }
7447 #endif
7448     if (!KMP_MASTER_TID(ds_tid) && ompt_enabled.ompt_callback_implicit_task) {
7449       ompt_callbacks.ompt_callback(ompt_callback_implicit_task)(
7450           ompt_scope_end, NULL, task_data, 0, ds_tid, ompt_task_implicit); // TODO: Can this be ompt_task_initial?
7451     }
7452   }
7453 #endif
7454 
7455   KMP_MB(); /* Flush all pending memory write invalidates.  */
7456   KMP_ASSERT(this_thr->th.th_team == team);
7457 }
7458 
7459 /* ------------------------------------------------------------------------ */
7460 
7461 #ifdef USE_LOAD_BALANCE
7462 
7463 // Return the worker threads actively spinning in the hot team, if we
7464 // are at the outermost level of parallelism.  Otherwise, return 0.
7465 static int __kmp_active_hot_team_nproc(kmp_root_t *root) {
7466   int i;
7467   int retval;
7468   kmp_team_t *hot_team;
7469 
7470   if (root->r.r_active) {
7471     return 0;
7472   }
7473   hot_team = root->r.r_hot_team;
7474   if (__kmp_dflt_blocktime == KMP_MAX_BLOCKTIME) {
7475     return hot_team->t.t_nproc - 1; // Don't count master thread
7476   }
7477 
7478   // Skip the master thread - it is accounted for elsewhere.
7479   retval = 0;
7480   for (i = 1; i < hot_team->t.t_nproc; i++) {
7481     if (hot_team->t.t_threads[i]->th.th_active) {
7482       retval++;
7483     }
7484   }
7485   return retval;
7486 }
7487 
7488 // Perform an automatic adjustment to the number of
7489 // threads used by the next parallel region.
7490 static int __kmp_load_balance_nproc(kmp_root_t *root, int set_nproc) {
7491   int retval;
7492   int pool_active;
7493   int hot_team_active;
7494   int team_curr_active;
7495   int system_active;
7496 
7497   KB_TRACE(20, ("__kmp_load_balance_nproc: called root:%p set_nproc:%d\n", root,
7498                 set_nproc));
7499   KMP_DEBUG_ASSERT(root);
7500   KMP_DEBUG_ASSERT(root->r.r_root_team->t.t_threads[0]
7501                        ->th.th_current_task->td_icvs.dynamic == TRUE);
7502   KMP_DEBUG_ASSERT(set_nproc > 1);
7503 
7504   if (set_nproc == 1) {
7505     KB_TRACE(20, ("__kmp_load_balance_nproc: serial execution.\n"));
7506     return 1;
7507   }
7508 
7509   // Threads that are active in the thread pool, active in the hot team for this
7510   // particular root (if we are at the outer par level), and the currently
7511   // executing thread (to become the master) are available to add to the new
7512   // team, but are currently contributing to the system load, and must be
7513   // accounted for.
7514   pool_active = __kmp_thread_pool_active_nth;
7515   hot_team_active = __kmp_active_hot_team_nproc(root);
7516   team_curr_active = pool_active + hot_team_active + 1;
7517 
7518   // Check the system load.
7519   system_active = __kmp_get_load_balance(__kmp_avail_proc + team_curr_active);
7520   KB_TRACE(30, ("__kmp_load_balance_nproc: system active = %d pool active = %d "
7521                 "hot team active = %d\n",
7522                 system_active, pool_active, hot_team_active));
7523 
7524   if (system_active < 0) {
7525     // There was an error reading the necessary info from /proc, so use the
7526     // thread limit algorithm instead. Once we set __kmp_global.g.g_dynamic_mode
7527     // = dynamic_thread_limit, we shouldn't wind up getting back here.
7528     __kmp_global.g.g_dynamic_mode = dynamic_thread_limit;
7529     KMP_WARNING(CantLoadBalUsing, "KMP_DYNAMIC_MODE=thread limit");
7530 
7531     // Make this call behave like the thread limit algorithm.
7532     retval = __kmp_avail_proc - __kmp_nth +
7533              (root->r.r_active ? 1 : root->r.r_hot_team->t.t_nproc);
7534     if (retval > set_nproc) {
7535       retval = set_nproc;
7536     }
7537     if (retval < KMP_MIN_NTH) {
7538       retval = KMP_MIN_NTH;
7539     }
7540 
7541     KB_TRACE(20, ("__kmp_load_balance_nproc: thread limit exit. retval:%d\n",
7542                   retval));
7543     return retval;
7544   }
7545 
7546   // There is a slight delay in the load balance algorithm in detecting new
7547   // running procs. The real system load at this instant should be at least as
7548   // large as the #active omp thread that are available to add to the team.
7549   if (system_active < team_curr_active) {
7550     system_active = team_curr_active;
7551   }
7552   retval = __kmp_avail_proc - system_active + team_curr_active;
7553   if (retval > set_nproc) {
7554     retval = set_nproc;
7555   }
7556   if (retval < KMP_MIN_NTH) {
7557     retval = KMP_MIN_NTH;
7558   }
7559 
7560   KB_TRACE(20, ("__kmp_load_balance_nproc: exit. retval:%d\n", retval));
7561   return retval;
7562 } // __kmp_load_balance_nproc()
7563 
7564 #endif /* USE_LOAD_BALANCE */
7565 
7566 /* ------------------------------------------------------------------------ */
7567 
7568 /* NOTE: this is called with the __kmp_init_lock held */
7569 void __kmp_cleanup(void) {
7570   int f;
7571 
7572   KA_TRACE(10, ("__kmp_cleanup: enter\n"));
7573 
7574   if (TCR_4(__kmp_init_parallel)) {
7575 #if KMP_HANDLE_SIGNALS
7576     __kmp_remove_signals();
7577 #endif
7578     TCW_4(__kmp_init_parallel, FALSE);
7579   }
7580 
7581   if (TCR_4(__kmp_init_middle)) {
7582 #if KMP_AFFINITY_SUPPORTED
7583     __kmp_affinity_uninitialize();
7584 #endif /* KMP_AFFINITY_SUPPORTED */
7585     __kmp_cleanup_hierarchy();
7586     TCW_4(__kmp_init_middle, FALSE);
7587   }
7588 
7589   KA_TRACE(10, ("__kmp_cleanup: go serial cleanup\n"));
7590 
7591   if (__kmp_init_serial) {
7592     __kmp_runtime_destroy();
7593     __kmp_init_serial = FALSE;
7594   }
7595 
7596   __kmp_cleanup_threadprivate_caches();
7597 
7598   for (f = 0; f < __kmp_threads_capacity; f++) {
7599     if (__kmp_root[f] != NULL) {
7600       __kmp_free(__kmp_root[f]);
7601       __kmp_root[f] = NULL;
7602     }
7603   }
7604   __kmp_free(__kmp_threads);
7605   // __kmp_threads and __kmp_root were allocated at once, as single block, so
7606   // there is no need in freeing __kmp_root.
7607   __kmp_threads = NULL;
7608   __kmp_root = NULL;
7609   __kmp_threads_capacity = 0;
7610 
7611 #if KMP_USE_DYNAMIC_LOCK
7612   __kmp_cleanup_indirect_user_locks();
7613 #else
7614   __kmp_cleanup_user_locks();
7615 #endif
7616 
7617 #if KMP_AFFINITY_SUPPORTED
7618   KMP_INTERNAL_FREE(CCAST(char *, __kmp_cpuinfo_file));
7619   __kmp_cpuinfo_file = NULL;
7620 #endif /* KMP_AFFINITY_SUPPORTED */
7621 
7622 #if KMP_USE_ADAPTIVE_LOCKS
7623 #if KMP_DEBUG_ADAPTIVE_LOCKS
7624   __kmp_print_speculative_stats();
7625 #endif
7626 #endif
7627   KMP_INTERNAL_FREE(__kmp_nested_nth.nth);
7628   __kmp_nested_nth.nth = NULL;
7629   __kmp_nested_nth.size = 0;
7630   __kmp_nested_nth.used = 0;
7631   KMP_INTERNAL_FREE(__kmp_nested_proc_bind.bind_types);
7632   __kmp_nested_proc_bind.bind_types = NULL;
7633   __kmp_nested_proc_bind.size = 0;
7634   __kmp_nested_proc_bind.used = 0;
7635 #if OMP_50_ENABLED
7636   if (__kmp_affinity_format) {
7637     KMP_INTERNAL_FREE(__kmp_affinity_format);
7638     __kmp_affinity_format = NULL;
7639   }
7640 #endif
7641 
7642   __kmp_i18n_catclose();
7643 
7644 #if KMP_USE_HIER_SCHED
7645   __kmp_hier_scheds.deallocate();
7646 #endif
7647 
7648 #if KMP_STATS_ENABLED
7649   __kmp_stats_fini();
7650 #endif
7651 
7652   KA_TRACE(10, ("__kmp_cleanup: exit\n"));
7653 }
7654 
7655 /* ------------------------------------------------------------------------ */
7656 
7657 int __kmp_ignore_mppbeg(void) {
7658   char *env;
7659 
7660   if ((env = getenv("KMP_IGNORE_MPPBEG")) != NULL) {
7661     if (__kmp_str_match_false(env))
7662       return FALSE;
7663   }
7664   // By default __kmpc_begin() is no-op.
7665   return TRUE;
7666 }
7667 
7668 int __kmp_ignore_mppend(void) {
7669   char *env;
7670 
7671   if ((env = getenv("KMP_IGNORE_MPPEND")) != NULL) {
7672     if (__kmp_str_match_false(env))
7673       return FALSE;
7674   }
7675   // By default __kmpc_end() is no-op.
7676   return TRUE;
7677 }
7678 
7679 void __kmp_internal_begin(void) {
7680   int gtid;
7681   kmp_root_t *root;
7682 
7683   /* this is a very important step as it will register new sibling threads
7684      and assign these new uber threads a new gtid */
7685   gtid = __kmp_entry_gtid();
7686   root = __kmp_threads[gtid]->th.th_root;
7687   KMP_ASSERT(KMP_UBER_GTID(gtid));
7688 
7689   if (root->r.r_begin)
7690     return;
7691   __kmp_acquire_lock(&root->r.r_begin_lock, gtid);
7692   if (root->r.r_begin) {
7693     __kmp_release_lock(&root->r.r_begin_lock, gtid);
7694     return;
7695   }
7696 
7697   root->r.r_begin = TRUE;
7698 
7699   __kmp_release_lock(&root->r.r_begin_lock, gtid);
7700 }
7701 
7702 /* ------------------------------------------------------------------------ */
7703 
7704 void __kmp_user_set_library(enum library_type arg) {
7705   int gtid;
7706   kmp_root_t *root;
7707   kmp_info_t *thread;
7708 
7709   /* first, make sure we are initialized so we can get our gtid */
7710 
7711   gtid = __kmp_entry_gtid();
7712   thread = __kmp_threads[gtid];
7713 
7714   root = thread->th.th_root;
7715 
7716   KA_TRACE(20, ("__kmp_user_set_library: enter T#%d, arg: %d, %d\n", gtid, arg,
7717                 library_serial));
7718   if (root->r.r_in_parallel) { /* Must be called in serial section of top-level
7719                                   thread */
7720     KMP_WARNING(SetLibraryIncorrectCall);
7721     return;
7722   }
7723 
7724   switch (arg) {
7725   case library_serial:
7726     thread->th.th_set_nproc = 0;
7727     set__nproc(thread, 1);
7728     break;
7729   case library_turnaround:
7730     thread->th.th_set_nproc = 0;
7731     set__nproc(thread, __kmp_dflt_team_nth ? __kmp_dflt_team_nth
7732                                            : __kmp_dflt_team_nth_ub);
7733     break;
7734   case library_throughput:
7735     thread->th.th_set_nproc = 0;
7736     set__nproc(thread, __kmp_dflt_team_nth ? __kmp_dflt_team_nth
7737                                            : __kmp_dflt_team_nth_ub);
7738     break;
7739   default:
7740     KMP_FATAL(UnknownLibraryType, arg);
7741   }
7742 
7743   __kmp_aux_set_library(arg);
7744 }
7745 
7746 void __kmp_aux_set_stacksize(size_t arg) {
7747   if (!__kmp_init_serial)
7748     __kmp_serial_initialize();
7749 
7750 #if KMP_OS_DARWIN
7751   if (arg & (0x1000 - 1)) {
7752     arg &= ~(0x1000 - 1);
7753     if (arg + 0x1000) /* check for overflow if we round up */
7754       arg += 0x1000;
7755   }
7756 #endif
7757   __kmp_acquire_bootstrap_lock(&__kmp_initz_lock);
7758 
7759   /* only change the default stacksize before the first parallel region */
7760   if (!TCR_4(__kmp_init_parallel)) {
7761     size_t value = arg; /* argument is in bytes */
7762 
7763     if (value < __kmp_sys_min_stksize)
7764       value = __kmp_sys_min_stksize;
7765     else if (value > KMP_MAX_STKSIZE)
7766       value = KMP_MAX_STKSIZE;
7767 
7768     __kmp_stksize = value;
7769 
7770     __kmp_env_stksize = TRUE; /* was KMP_STACKSIZE specified? */
7771   }
7772 
7773   __kmp_release_bootstrap_lock(&__kmp_initz_lock);
7774 }
7775 
7776 /* set the behaviour of the runtime library */
7777 /* TODO this can cause some odd behaviour with sibling parallelism... */
7778 void __kmp_aux_set_library(enum library_type arg) {
7779   __kmp_library = arg;
7780 
7781   switch (__kmp_library) {
7782   case library_serial: {
7783     KMP_INFORM(LibraryIsSerial);
7784   } break;
7785   case library_turnaround:
7786     if (__kmp_use_yield == 1 && !__kmp_use_yield_exp_set)
7787       __kmp_use_yield = 2; // only yield when oversubscribed
7788     break;
7789   case library_throughput:
7790     if (__kmp_dflt_blocktime == KMP_MAX_BLOCKTIME)
7791       __kmp_dflt_blocktime = 200;
7792     break;
7793   default:
7794     KMP_FATAL(UnknownLibraryType, arg);
7795   }
7796 }
7797 
7798 /* Getting team information common for all team API */
7799 // Returns NULL if not in teams construct
7800 static kmp_team_t *__kmp_aux_get_team_info(int &teams_serialized) {
7801   kmp_info_t *thr = __kmp_entry_thread();
7802   teams_serialized = 0;
7803   if (thr->th.th_teams_microtask) {
7804     kmp_team_t *team = thr->th.th_team;
7805     int tlevel = thr->th.th_teams_level; // the level of the teams construct
7806     int ii = team->t.t_level;
7807     teams_serialized = team->t.t_serialized;
7808     int level = tlevel + 1;
7809     KMP_DEBUG_ASSERT(ii >= tlevel);
7810     while (ii > level) {
7811       for (teams_serialized = team->t.t_serialized;
7812            (teams_serialized > 0) && (ii > level); teams_serialized--, ii--) {
7813       }
7814       if (team->t.t_serialized && (!teams_serialized)) {
7815         team = team->t.t_parent;
7816         continue;
7817       }
7818       if (ii > level) {
7819         team = team->t.t_parent;
7820         ii--;
7821       }
7822     }
7823     return team;
7824   }
7825   return NULL;
7826 }
7827 
7828 int __kmp_aux_get_team_num() {
7829   int serialized;
7830   kmp_team_t *team = __kmp_aux_get_team_info(serialized);
7831   if (team) {
7832     if (serialized > 1) {
7833       return 0; // teams region is serialized ( 1 team of 1 thread ).
7834     } else {
7835       return team->t.t_master_tid;
7836     }
7837   }
7838   return 0;
7839 }
7840 
7841 int __kmp_aux_get_num_teams() {
7842   int serialized;
7843   kmp_team_t *team = __kmp_aux_get_team_info(serialized);
7844   if (team) {
7845     if (serialized > 1) {
7846       return 1;
7847     } else {
7848       return team->t.t_parent->t.t_nproc;
7849     }
7850   }
7851   return 1;
7852 }
7853 
7854 /* ------------------------------------------------------------------------ */
7855 
7856 #if OMP_50_ENABLED
7857 /*
7858  * Affinity Format Parser
7859  *
7860  * Field is in form of: %[[[0].]size]type
7861  * % and type are required (%% means print a literal '%')
7862  * type is either single char or long name surrounded by {},
7863  * e.g., N or {num_threads}
7864  * 0 => leading zeros
7865  * . => right justified when size is specified
7866  * by default output is left justified
7867  * size is the *minimum* field length
7868  * All other characters are printed as is
7869  *
7870  * Available field types:
7871  * L {thread_level}      - omp_get_level()
7872  * n {thread_num}        - omp_get_thread_num()
7873  * h {host}              - name of host machine
7874  * P {process_id}        - process id (integer)
7875  * T {thread_identifier} - native thread identifier (integer)
7876  * N {num_threads}       - omp_get_num_threads()
7877  * A {ancestor_tnum}     - omp_get_ancestor_thread_num(omp_get_level()-1)
7878  * a {thread_affinity}   - comma separated list of integers or integer ranges
7879  *                         (values of affinity mask)
7880  *
7881  * Implementation-specific field types can be added
7882  * If a type is unknown, print "undefined"
7883 */
7884 
7885 // Structure holding the short name, long name, and corresponding data type
7886 // for snprintf.  A table of these will represent the entire valid keyword
7887 // field types.
7888 typedef struct kmp_affinity_format_field_t {
7889   char short_name; // from spec e.g., L -> thread level
7890   const char *long_name; // from spec thread_level -> thread level
7891   char field_format; // data type for snprintf (typically 'd' or 's'
7892   // for integer or string)
7893 } kmp_affinity_format_field_t;
7894 
7895 static const kmp_affinity_format_field_t __kmp_affinity_format_table[] = {
7896 #if KMP_AFFINITY_SUPPORTED
7897     {'A', "thread_affinity", 's'},
7898 #endif
7899     {'t', "team_num", 'd'},
7900     {'T', "num_teams", 'd'},
7901     {'L', "nesting_level", 'd'},
7902     {'n', "thread_num", 'd'},
7903     {'N', "num_threads", 'd'},
7904     {'a', "ancestor_tnum", 'd'},
7905     {'H', "host", 's'},
7906     {'P', "process_id", 'd'},
7907     {'i', "native_thread_id", 'd'}};
7908 
7909 // Return the number of characters it takes to hold field
7910 static int __kmp_aux_capture_affinity_field(int gtid, const kmp_info_t *th,
7911                                             const char **ptr,
7912                                             kmp_str_buf_t *field_buffer) {
7913   int rc, format_index, field_value;
7914   const char *width_left, *width_right;
7915   bool pad_zeros, right_justify, parse_long_name, found_valid_name;
7916   static const int FORMAT_SIZE = 20;
7917   char format[FORMAT_SIZE] = {0};
7918   char absolute_short_name = 0;
7919 
7920   KMP_DEBUG_ASSERT(gtid >= 0);
7921   KMP_DEBUG_ASSERT(th);
7922   KMP_DEBUG_ASSERT(**ptr == '%');
7923   KMP_DEBUG_ASSERT(field_buffer);
7924 
7925   __kmp_str_buf_clear(field_buffer);
7926 
7927   // Skip the initial %
7928   (*ptr)++;
7929 
7930   // Check for %% first
7931   if (**ptr == '%') {
7932     __kmp_str_buf_cat(field_buffer, "%", 1);
7933     (*ptr)++; // skip over the second %
7934     return 1;
7935   }
7936 
7937   // Parse field modifiers if they are present
7938   pad_zeros = false;
7939   if (**ptr == '0') {
7940     pad_zeros = true;
7941     (*ptr)++; // skip over 0
7942   }
7943   right_justify = false;
7944   if (**ptr == '.') {
7945     right_justify = true;
7946     (*ptr)++; // skip over .
7947   }
7948   // Parse width of field: [width_left, width_right)
7949   width_left = width_right = NULL;
7950   if (**ptr >= '0' && **ptr <= '9') {
7951     width_left = *ptr;
7952     SKIP_DIGITS(*ptr);
7953     width_right = *ptr;
7954   }
7955 
7956   // Create the format for KMP_SNPRINTF based on flags parsed above
7957   format_index = 0;
7958   format[format_index++] = '%';
7959   if (!right_justify)
7960     format[format_index++] = '-';
7961   if (pad_zeros)
7962     format[format_index++] = '0';
7963   if (width_left && width_right) {
7964     int i = 0;
7965     // Only allow 8 digit number widths.
7966     // This also prevents overflowing format variable
7967     while (i < 8 && width_left < width_right) {
7968       format[format_index++] = *width_left;
7969       width_left++;
7970       i++;
7971     }
7972   }
7973 
7974   // Parse a name (long or short)
7975   // Canonicalize the name into absolute_short_name
7976   found_valid_name = false;
7977   parse_long_name = (**ptr == '{');
7978   if (parse_long_name)
7979     (*ptr)++; // skip initial left brace
7980   for (size_t i = 0; i < sizeof(__kmp_affinity_format_table) /
7981                              sizeof(__kmp_affinity_format_table[0]);
7982        ++i) {
7983     char short_name = __kmp_affinity_format_table[i].short_name;
7984     const char *long_name = __kmp_affinity_format_table[i].long_name;
7985     char field_format = __kmp_affinity_format_table[i].field_format;
7986     if (parse_long_name) {
7987       int length = KMP_STRLEN(long_name);
7988       if (strncmp(*ptr, long_name, length) == 0) {
7989         found_valid_name = true;
7990         (*ptr) += length; // skip the long name
7991       }
7992     } else if (**ptr == short_name) {
7993       found_valid_name = true;
7994       (*ptr)++; // skip the short name
7995     }
7996     if (found_valid_name) {
7997       format[format_index++] = field_format;
7998       format[format_index++] = '\0';
7999       absolute_short_name = short_name;
8000       break;
8001     }
8002   }
8003   if (parse_long_name) {
8004     if (**ptr != '}') {
8005       absolute_short_name = 0;
8006     } else {
8007       (*ptr)++; // skip over the right brace
8008     }
8009   }
8010 
8011   // Attempt to fill the buffer with the requested
8012   // value using snprintf within __kmp_str_buf_print()
8013   switch (absolute_short_name) {
8014   case 't':
8015     rc = __kmp_str_buf_print(field_buffer, format, __kmp_aux_get_team_num());
8016     break;
8017   case 'T':
8018     rc = __kmp_str_buf_print(field_buffer, format, __kmp_aux_get_num_teams());
8019     break;
8020   case 'L':
8021     rc = __kmp_str_buf_print(field_buffer, format, th->th.th_team->t.t_level);
8022     break;
8023   case 'n':
8024     rc = __kmp_str_buf_print(field_buffer, format, __kmp_tid_from_gtid(gtid));
8025     break;
8026   case 'H': {
8027     static const int BUFFER_SIZE = 256;
8028     char buf[BUFFER_SIZE];
8029     __kmp_expand_host_name(buf, BUFFER_SIZE);
8030     rc = __kmp_str_buf_print(field_buffer, format, buf);
8031   } break;
8032   case 'P':
8033     rc = __kmp_str_buf_print(field_buffer, format, getpid());
8034     break;
8035   case 'i':
8036     rc = __kmp_str_buf_print(field_buffer, format, __kmp_gettid());
8037     break;
8038   case 'N':
8039     rc = __kmp_str_buf_print(field_buffer, format, th->th.th_team->t.t_nproc);
8040     break;
8041   case 'a':
8042     field_value =
8043         __kmp_get_ancestor_thread_num(gtid, th->th.th_team->t.t_level - 1);
8044     rc = __kmp_str_buf_print(field_buffer, format, field_value);
8045     break;
8046 #if KMP_AFFINITY_SUPPORTED
8047   case 'A': {
8048     kmp_str_buf_t buf;
8049     __kmp_str_buf_init(&buf);
8050     __kmp_affinity_str_buf_mask(&buf, th->th.th_affin_mask);
8051     rc = __kmp_str_buf_print(field_buffer, format, buf.str);
8052     __kmp_str_buf_free(&buf);
8053   } break;
8054 #endif
8055   default:
8056     // According to spec, If an implementation does not have info for field
8057     // type, then "undefined" is printed
8058     rc = __kmp_str_buf_print(field_buffer, "%s", "undefined");
8059     // Skip the field
8060     if (parse_long_name) {
8061       SKIP_TOKEN(*ptr);
8062       if (**ptr == '}')
8063         (*ptr)++;
8064     } else {
8065       (*ptr)++;
8066     }
8067   }
8068 
8069   KMP_ASSERT(format_index <= FORMAT_SIZE);
8070   return rc;
8071 }
8072 
8073 /*
8074  * Return number of characters needed to hold the affinity string
8075  * (not including null byte character)
8076  * The resultant string is printed to buffer, which the caller can then
8077  * handle afterwards
8078 */
8079 size_t __kmp_aux_capture_affinity(int gtid, const char *format,
8080                                   kmp_str_buf_t *buffer) {
8081   const char *parse_ptr;
8082   size_t retval;
8083   const kmp_info_t *th;
8084   kmp_str_buf_t field;
8085 
8086   KMP_DEBUG_ASSERT(buffer);
8087   KMP_DEBUG_ASSERT(gtid >= 0);
8088 
8089   __kmp_str_buf_init(&field);
8090   __kmp_str_buf_clear(buffer);
8091 
8092   th = __kmp_threads[gtid];
8093   retval = 0;
8094 
8095   // If format is NULL or zero-length string, then we use
8096   // affinity-format-var ICV
8097   parse_ptr = format;
8098   if (parse_ptr == NULL || *parse_ptr == '\0') {
8099     parse_ptr = __kmp_affinity_format;
8100   }
8101   KMP_DEBUG_ASSERT(parse_ptr);
8102 
8103   while (*parse_ptr != '\0') {
8104     // Parse a field
8105     if (*parse_ptr == '%') {
8106       // Put field in the buffer
8107       int rc = __kmp_aux_capture_affinity_field(gtid, th, &parse_ptr, &field);
8108       __kmp_str_buf_catbuf(buffer, &field);
8109       retval += rc;
8110     } else {
8111       // Put literal character in buffer
8112       __kmp_str_buf_cat(buffer, parse_ptr, 1);
8113       retval++;
8114       parse_ptr++;
8115     }
8116   }
8117   __kmp_str_buf_free(&field);
8118   return retval;
8119 }
8120 
8121 // Displays the affinity string to stdout
8122 void __kmp_aux_display_affinity(int gtid, const char *format) {
8123   kmp_str_buf_t buf;
8124   __kmp_str_buf_init(&buf);
8125   __kmp_aux_capture_affinity(gtid, format, &buf);
8126   __kmp_fprintf(kmp_out, "%s" KMP_END_OF_LINE, buf.str);
8127   __kmp_str_buf_free(&buf);
8128 }
8129 #endif // OMP_50_ENABLED
8130 
8131 /* ------------------------------------------------------------------------ */
8132 
8133 void __kmp_aux_set_blocktime(int arg, kmp_info_t *thread, int tid) {
8134   int blocktime = arg; /* argument is in milliseconds */
8135 #if KMP_USE_MONITOR
8136   int bt_intervals;
8137 #endif
8138   int bt_set;
8139 
8140   __kmp_save_internal_controls(thread);
8141 
8142   /* Normalize and set blocktime for the teams */
8143   if (blocktime < KMP_MIN_BLOCKTIME)
8144     blocktime = KMP_MIN_BLOCKTIME;
8145   else if (blocktime > KMP_MAX_BLOCKTIME)
8146     blocktime = KMP_MAX_BLOCKTIME;
8147 
8148   set__blocktime_team(thread->th.th_team, tid, blocktime);
8149   set__blocktime_team(thread->th.th_serial_team, 0, blocktime);
8150 
8151 #if KMP_USE_MONITOR
8152   /* Calculate and set blocktime intervals for the teams */
8153   bt_intervals = KMP_INTERVALS_FROM_BLOCKTIME(blocktime, __kmp_monitor_wakeups);
8154 
8155   set__bt_intervals_team(thread->th.th_team, tid, bt_intervals);
8156   set__bt_intervals_team(thread->th.th_serial_team, 0, bt_intervals);
8157 #endif
8158 
8159   /* Set whether blocktime has been set to "TRUE" */
8160   bt_set = TRUE;
8161 
8162   set__bt_set_team(thread->th.th_team, tid, bt_set);
8163   set__bt_set_team(thread->th.th_serial_team, 0, bt_set);
8164 #if KMP_USE_MONITOR
8165   KF_TRACE(10, ("kmp_set_blocktime: T#%d(%d:%d), blocktime=%d, "
8166                 "bt_intervals=%d, monitor_updates=%d\n",
8167                 __kmp_gtid_from_tid(tid, thread->th.th_team),
8168                 thread->th.th_team->t.t_id, tid, blocktime, bt_intervals,
8169                 __kmp_monitor_wakeups));
8170 #else
8171   KF_TRACE(10, ("kmp_set_blocktime: T#%d(%d:%d), blocktime=%d\n",
8172                 __kmp_gtid_from_tid(tid, thread->th.th_team),
8173                 thread->th.th_team->t.t_id, tid, blocktime));
8174 #endif
8175 }
8176 
8177 void __kmp_aux_set_defaults(char const *str, int len) {
8178   if (!__kmp_init_serial) {
8179     __kmp_serial_initialize();
8180   }
8181   __kmp_env_initialize(str);
8182 
8183   if (__kmp_settings
8184 #if OMP_40_ENABLED
8185       || __kmp_display_env || __kmp_display_env_verbose
8186 #endif // OMP_40_ENABLED
8187       ) {
8188     __kmp_env_print();
8189   }
8190 } // __kmp_aux_set_defaults
8191 
8192 /* ------------------------------------------------------------------------ */
8193 /* internal fast reduction routines */
8194 
8195 PACKED_REDUCTION_METHOD_T
8196 __kmp_determine_reduction_method(
8197     ident_t *loc, kmp_int32 global_tid, kmp_int32 num_vars, size_t reduce_size,
8198     void *reduce_data, void (*reduce_func)(void *lhs_data, void *rhs_data),
8199     kmp_critical_name *lck) {
8200 
8201   // Default reduction method: critical construct ( lck != NULL, like in current
8202   // PAROPT )
8203   // If ( reduce_data!=NULL && reduce_func!=NULL ): the tree-reduction method
8204   // can be selected by RTL
8205   // If loc->flags contains KMP_IDENT_ATOMIC_REDUCE, the atomic reduce method
8206   // can be selected by RTL
8207   // Finally, it's up to OpenMP RTL to make a decision on which method to select
8208   // among generated by PAROPT.
8209 
8210   PACKED_REDUCTION_METHOD_T retval;
8211 
8212   int team_size;
8213 
8214   KMP_DEBUG_ASSERT(loc); // it would be nice to test ( loc != 0 )
8215   KMP_DEBUG_ASSERT(lck); // it would be nice to test ( lck != 0 )
8216 
8217 #define FAST_REDUCTION_ATOMIC_METHOD_GENERATED                                 \
8218   ((loc->flags & (KMP_IDENT_ATOMIC_REDUCE)) == (KMP_IDENT_ATOMIC_REDUCE))
8219 #define FAST_REDUCTION_TREE_METHOD_GENERATED ((reduce_data) && (reduce_func))
8220 
8221   retval = critical_reduce_block;
8222 
8223   // another choice of getting a team size (with 1 dynamic deference) is slower
8224   team_size = __kmp_get_team_num_threads(global_tid);
8225   if (team_size == 1) {
8226 
8227     retval = empty_reduce_block;
8228 
8229   } else {
8230 
8231     int atomic_available = FAST_REDUCTION_ATOMIC_METHOD_GENERATED;
8232 
8233 #if KMP_ARCH_X86_64 || KMP_ARCH_PPC64 || KMP_ARCH_AARCH64 || KMP_ARCH_MIPS64
8234 
8235 #if KMP_OS_LINUX || KMP_OS_DRAGONFLY || KMP_OS_FREEBSD || KMP_OS_NETBSD ||     \
8236     KMP_OS_OPENBSD || KMP_OS_WINDOWS || KMP_OS_DARWIN || KMP_OS_HURD
8237 
8238     int teamsize_cutoff = 4;
8239 
8240 #if KMP_MIC_SUPPORTED
8241     if (__kmp_mic_type != non_mic) {
8242       teamsize_cutoff = 8;
8243     }
8244 #endif
8245     int tree_available = FAST_REDUCTION_TREE_METHOD_GENERATED;
8246     if (tree_available) {
8247       if (team_size <= teamsize_cutoff) {
8248         if (atomic_available) {
8249           retval = atomic_reduce_block;
8250         }
8251       } else {
8252         retval = TREE_REDUCE_BLOCK_WITH_REDUCTION_BARRIER;
8253       }
8254     } else if (atomic_available) {
8255       retval = atomic_reduce_block;
8256     }
8257 #else
8258 #error "Unknown or unsupported OS"
8259 #endif // KMP_OS_LINUX || KMP_OS_DRAGONFLY || KMP_OS_FREEBSD || KMP_OS_NETBSD ||
8260        // KMP_OS_OPENBSD || KMP_OS_WINDOWS || KMP_OS_DARWIN || KMP_OS_HURD
8261 
8262 #elif KMP_ARCH_X86 || KMP_ARCH_ARM || KMP_ARCH_AARCH || KMP_ARCH_MIPS
8263 
8264 #if KMP_OS_LINUX || KMP_OS_FREEBSD || KMP_OS_WINDOWS || KMP_OS_HURD
8265 
8266     // basic tuning
8267 
8268     if (atomic_available) {
8269       if (num_vars <= 2) { // && ( team_size <= 8 ) due to false-sharing ???
8270         retval = atomic_reduce_block;
8271       }
8272     } // otherwise: use critical section
8273 
8274 #elif KMP_OS_DARWIN
8275 
8276     int tree_available = FAST_REDUCTION_TREE_METHOD_GENERATED;
8277     if (atomic_available && (num_vars <= 3)) {
8278       retval = atomic_reduce_block;
8279     } else if (tree_available) {
8280       if ((reduce_size > (9 * sizeof(kmp_real64))) &&
8281           (reduce_size < (2000 * sizeof(kmp_real64)))) {
8282         retval = TREE_REDUCE_BLOCK_WITH_PLAIN_BARRIER;
8283       }
8284     } // otherwise: use critical section
8285 
8286 #else
8287 #error "Unknown or unsupported OS"
8288 #endif
8289 
8290 #else
8291 #error "Unknown or unsupported architecture"
8292 #endif
8293   }
8294 
8295   // KMP_FORCE_REDUCTION
8296 
8297   // If the team is serialized (team_size == 1), ignore the forced reduction
8298   // method and stay with the unsynchronized method (empty_reduce_block)
8299   if (__kmp_force_reduction_method != reduction_method_not_defined &&
8300       team_size != 1) {
8301 
8302     PACKED_REDUCTION_METHOD_T forced_retval = critical_reduce_block;
8303 
8304     int atomic_available, tree_available;
8305 
8306     switch ((forced_retval = __kmp_force_reduction_method)) {
8307     case critical_reduce_block:
8308       KMP_ASSERT(lck); // lck should be != 0
8309       break;
8310 
8311     case atomic_reduce_block:
8312       atomic_available = FAST_REDUCTION_ATOMIC_METHOD_GENERATED;
8313       if (!atomic_available) {
8314         KMP_WARNING(RedMethodNotSupported, "atomic");
8315         forced_retval = critical_reduce_block;
8316       }
8317       break;
8318 
8319     case tree_reduce_block:
8320       tree_available = FAST_REDUCTION_TREE_METHOD_GENERATED;
8321       if (!tree_available) {
8322         KMP_WARNING(RedMethodNotSupported, "tree");
8323         forced_retval = critical_reduce_block;
8324       } else {
8325 #if KMP_FAST_REDUCTION_BARRIER
8326         forced_retval = TREE_REDUCE_BLOCK_WITH_REDUCTION_BARRIER;
8327 #endif
8328       }
8329       break;
8330 
8331     default:
8332       KMP_ASSERT(0); // "unsupported method specified"
8333     }
8334 
8335     retval = forced_retval;
8336   }
8337 
8338   KA_TRACE(10, ("reduction method selected=%08x\n", retval));
8339 
8340 #undef FAST_REDUCTION_TREE_METHOD_GENERATED
8341 #undef FAST_REDUCTION_ATOMIC_METHOD_GENERATED
8342 
8343   return (retval);
8344 }
8345 
8346 // this function is for testing set/get/determine reduce method
8347 kmp_int32 __kmp_get_reduce_method(void) {
8348   return ((__kmp_entry_thread()->th.th_local.packed_reduction_method) >> 8);
8349 }
8350 
8351 #if OMP_50_ENABLED
8352 
8353 // Soft pause sets up threads to ignore blocktime and just go to sleep.
8354 // Spin-wait code checks __kmp_pause_status and reacts accordingly.
8355 void __kmp_soft_pause() { __kmp_pause_status = kmp_soft_paused; }
8356 
8357 // Hard pause shuts down the runtime completely.  Resume happens naturally when
8358 // OpenMP is used subsequently.
8359 void __kmp_hard_pause() {
8360   __kmp_pause_status = kmp_hard_paused;
8361   __kmp_internal_end_thread(-1);
8362 }
8363 
8364 // Soft resume sets __kmp_pause_status, and wakes up all threads.
8365 void __kmp_resume_if_soft_paused() {
8366   if (__kmp_pause_status == kmp_soft_paused) {
8367     __kmp_pause_status = kmp_not_paused;
8368 
8369     for (int gtid = 1; gtid < __kmp_threads_capacity; ++gtid) {
8370       kmp_info_t *thread = __kmp_threads[gtid];
8371       if (thread) { // Wake it if sleeping
8372         kmp_flag_64 fl(&thread->th.th_bar[bs_forkjoin_barrier].bb.b_go, thread);
8373         if (fl.is_sleeping())
8374           fl.resume(gtid);
8375         else if (__kmp_try_suspend_mx(thread)) { // got suspend lock
8376           __kmp_unlock_suspend_mx(thread); // unlock it; it won't sleep
8377         } else { // thread holds the lock and may sleep soon
8378           do { // until either the thread sleeps, or we can get the lock
8379             if (fl.is_sleeping()) {
8380               fl.resume(gtid);
8381               break;
8382             } else if (__kmp_try_suspend_mx(thread)) {
8383               __kmp_unlock_suspend_mx(thread);
8384               break;
8385             }
8386           } while (1);
8387         }
8388       }
8389     }
8390   }
8391 }
8392 
8393 // This function is called via __kmpc_pause_resource. Returns 0 if successful.
8394 // TODO: add warning messages
8395 int __kmp_pause_resource(kmp_pause_status_t level) {
8396   if (level == kmp_not_paused) { // requesting resume
8397     if (__kmp_pause_status == kmp_not_paused) {
8398       // error message about runtime not being paused, so can't resume
8399       return 1;
8400     } else {
8401       KMP_DEBUG_ASSERT(__kmp_pause_status == kmp_soft_paused ||
8402                        __kmp_pause_status == kmp_hard_paused);
8403       __kmp_pause_status = kmp_not_paused;
8404       return 0;
8405     }
8406   } else if (level == kmp_soft_paused) { // requesting soft pause
8407     if (__kmp_pause_status != kmp_not_paused) {
8408       // error message about already being paused
8409       return 1;
8410     } else {
8411       __kmp_soft_pause();
8412       return 0;
8413     }
8414   } else if (level == kmp_hard_paused) { // requesting hard pause
8415     if (__kmp_pause_status != kmp_not_paused) {
8416       // error message about already being paused
8417       return 1;
8418     } else {
8419       __kmp_hard_pause();
8420       return 0;
8421     }
8422   } else {
8423     // error message about invalid level
8424     return 1;
8425   }
8426 }
8427 
8428 #endif // OMP_50_ENABLED
8429