1 /*
2  * kmp_runtime.cpp -- KPTS runtime support library
3  */
4 
5 //===----------------------------------------------------------------------===//
6 //
7 //                     The LLVM Compiler Infrastructure
8 //
9 // This file is dual licensed under the MIT and the University of Illinois Open
10 // Source Licenses. See LICENSE.txt for details.
11 //
12 //===----------------------------------------------------------------------===//
13 
14 #include "kmp.h"
15 #include "kmp_affinity.h"
16 #include "kmp_atomic.h"
17 #include "kmp_environment.h"
18 #include "kmp_error.h"
19 #include "kmp_i18n.h"
20 #include "kmp_io.h"
21 #include "kmp_itt.h"
22 #include "kmp_settings.h"
23 #include "kmp_stats.h"
24 #include "kmp_str.h"
25 #include "kmp_wait_release.h"
26 #include "kmp_wrapper_getpid.h"
27 #include "kmp_dispatch.h"
28 #if KMP_USE_HIER_SCHED
29 #include "kmp_dispatch_hier.h"
30 #endif
31 
32 #if OMPT_SUPPORT
33 #include "ompt-specific.h"
34 #endif
35 
36 /* these are temporary issues to be dealt with */
37 #define KMP_USE_PRCTL 0
38 
39 #if KMP_OS_WINDOWS
40 #include <process.h>
41 #endif
42 
43 #include "tsan_annotations.h"
44 
45 #if defined(KMP_GOMP_COMPAT)
46 char const __kmp_version_alt_comp[] =
47     KMP_VERSION_PREFIX "alternative compiler support: yes";
48 #endif /* defined(KMP_GOMP_COMPAT) */
49 
50 char const __kmp_version_omp_api[] = KMP_VERSION_PREFIX "API version: "
51 #if OMP_50_ENABLED
52                                                         "5.0 (201611)";
53 #elif OMP_45_ENABLED
54                                                         "4.5 (201511)";
55 #elif OMP_40_ENABLED
56                                                         "4.0 (201307)";
57 #else
58                                                         "3.1 (201107)";
59 #endif
60 
61 #ifdef KMP_DEBUG
62 char const __kmp_version_lock[] =
63     KMP_VERSION_PREFIX "lock type: run time selectable";
64 #endif /* KMP_DEBUG */
65 
66 #define KMP_MIN(x, y) ((x) < (y) ? (x) : (y))
67 
68 /* ------------------------------------------------------------------------ */
69 
70 #if KMP_USE_MONITOR
71 kmp_info_t __kmp_monitor;
72 #endif
73 
74 /* Forward declarations */
75 
76 void __kmp_cleanup(void);
77 
78 static void __kmp_initialize_info(kmp_info_t *, kmp_team_t *, int tid,
79                                   int gtid);
80 static void __kmp_initialize_team(kmp_team_t *team, int new_nproc,
81                                   kmp_internal_control_t *new_icvs,
82                                   ident_t *loc);
83 #if OMP_40_ENABLED && KMP_AFFINITY_SUPPORTED
84 static void __kmp_partition_places(kmp_team_t *team,
85                                    int update_master_only = 0);
86 #endif
87 static void __kmp_do_serial_initialize(void);
88 void __kmp_fork_barrier(int gtid, int tid);
89 void __kmp_join_barrier(int gtid);
90 void __kmp_setup_icv_copy(kmp_team_t *team, int new_nproc,
91                           kmp_internal_control_t *new_icvs, ident_t *loc);
92 
93 #ifdef USE_LOAD_BALANCE
94 static int __kmp_load_balance_nproc(kmp_root_t *root, int set_nproc);
95 #endif
96 
97 static int __kmp_expand_threads(int nNeed);
98 #if KMP_OS_WINDOWS
99 static int __kmp_unregister_root_other_thread(int gtid);
100 #endif
101 static void __kmp_unregister_library(void); // called by __kmp_internal_end()
102 static void __kmp_reap_thread(kmp_info_t *thread, int is_root);
103 kmp_info_t *__kmp_thread_pool_insert_pt = NULL;
104 
105 /* Calculate the identifier of the current thread */
106 /* fast (and somewhat portable) way to get unique identifier of executing
107    thread. Returns KMP_GTID_DNE if we haven't been assigned a gtid. */
108 int __kmp_get_global_thread_id() {
109   int i;
110   kmp_info_t **other_threads;
111   size_t stack_data;
112   char *stack_addr;
113   size_t stack_size;
114   char *stack_base;
115 
116   KA_TRACE(
117       1000,
118       ("*** __kmp_get_global_thread_id: entering, nproc=%d  all_nproc=%d\n",
119        __kmp_nth, __kmp_all_nth));
120 
121   /* JPH - to handle the case where __kmpc_end(0) is called immediately prior to
122      a parallel region, made it return KMP_GTID_DNE to force serial_initialize
123      by caller. Had to handle KMP_GTID_DNE at all call-sites, or else guarantee
124      __kmp_init_gtid for this to work. */
125 
126   if (!TCR_4(__kmp_init_gtid))
127     return KMP_GTID_DNE;
128 
129 #ifdef KMP_TDATA_GTID
130   if (TCR_4(__kmp_gtid_mode) >= 3) {
131     KA_TRACE(1000, ("*** __kmp_get_global_thread_id: using TDATA\n"));
132     return __kmp_gtid;
133   }
134 #endif
135   if (TCR_4(__kmp_gtid_mode) >= 2) {
136     KA_TRACE(1000, ("*** __kmp_get_global_thread_id: using keyed TLS\n"));
137     return __kmp_gtid_get_specific();
138   }
139   KA_TRACE(1000, ("*** __kmp_get_global_thread_id: using internal alg.\n"));
140 
141   stack_addr = (char *)&stack_data;
142   other_threads = __kmp_threads;
143 
144   /* ATT: The code below is a source of potential bugs due to unsynchronized
145      access to __kmp_threads array. For example:
146      1. Current thread loads other_threads[i] to thr and checks it, it is
147         non-NULL.
148      2. Current thread is suspended by OS.
149      3. Another thread unregisters and finishes (debug versions of free()
150         may fill memory with something like 0xEF).
151      4. Current thread is resumed.
152      5. Current thread reads junk from *thr.
153      TODO: Fix it.  --ln  */
154 
155   for (i = 0; i < __kmp_threads_capacity; i++) {
156 
157     kmp_info_t *thr = (kmp_info_t *)TCR_SYNC_PTR(other_threads[i]);
158     if (!thr)
159       continue;
160 
161     stack_size = (size_t)TCR_PTR(thr->th.th_info.ds.ds_stacksize);
162     stack_base = (char *)TCR_PTR(thr->th.th_info.ds.ds_stackbase);
163 
164     /* stack grows down -- search through all of the active threads */
165 
166     if (stack_addr <= stack_base) {
167       size_t stack_diff = stack_base - stack_addr;
168 
169       if (stack_diff <= stack_size) {
170         /* The only way we can be closer than the allocated */
171         /* stack size is if we are running on this thread. */
172         KMP_DEBUG_ASSERT(__kmp_gtid_get_specific() == i);
173         return i;
174       }
175     }
176   }
177 
178   /* get specific to try and determine our gtid */
179   KA_TRACE(1000,
180            ("*** __kmp_get_global_thread_id: internal alg. failed to find "
181             "thread, using TLS\n"));
182   i = __kmp_gtid_get_specific();
183 
184   /*fprintf( stderr, "=== %d\n", i );  */ /* GROO */
185 
186   /* if we havn't been assigned a gtid, then return code */
187   if (i < 0)
188     return i;
189 
190   /* dynamically updated stack window for uber threads to avoid get_specific
191      call */
192   if (!TCR_4(other_threads[i]->th.th_info.ds.ds_stackgrow)) {
193     KMP_FATAL(StackOverflow, i);
194   }
195 
196   stack_base = (char *)other_threads[i]->th.th_info.ds.ds_stackbase;
197   if (stack_addr > stack_base) {
198     TCW_PTR(other_threads[i]->th.th_info.ds.ds_stackbase, stack_addr);
199     TCW_PTR(other_threads[i]->th.th_info.ds.ds_stacksize,
200             other_threads[i]->th.th_info.ds.ds_stacksize + stack_addr -
201                 stack_base);
202   } else {
203     TCW_PTR(other_threads[i]->th.th_info.ds.ds_stacksize,
204             stack_base - stack_addr);
205   }
206 
207   /* Reprint stack bounds for ubermaster since they have been refined */
208   if (__kmp_storage_map) {
209     char *stack_end = (char *)other_threads[i]->th.th_info.ds.ds_stackbase;
210     char *stack_beg = stack_end - other_threads[i]->th.th_info.ds.ds_stacksize;
211     __kmp_print_storage_map_gtid(i, stack_beg, stack_end,
212                                  other_threads[i]->th.th_info.ds.ds_stacksize,
213                                  "th_%d stack (refinement)", i);
214   }
215   return i;
216 }
217 
218 int __kmp_get_global_thread_id_reg() {
219   int gtid;
220 
221   if (!__kmp_init_serial) {
222     gtid = KMP_GTID_DNE;
223   } else
224 #ifdef KMP_TDATA_GTID
225       if (TCR_4(__kmp_gtid_mode) >= 3) {
226     KA_TRACE(1000, ("*** __kmp_get_global_thread_id_reg: using TDATA\n"));
227     gtid = __kmp_gtid;
228   } else
229 #endif
230       if (TCR_4(__kmp_gtid_mode) >= 2) {
231     KA_TRACE(1000, ("*** __kmp_get_global_thread_id_reg: using keyed TLS\n"));
232     gtid = __kmp_gtid_get_specific();
233   } else {
234     KA_TRACE(1000,
235              ("*** __kmp_get_global_thread_id_reg: using internal alg.\n"));
236     gtid = __kmp_get_global_thread_id();
237   }
238 
239   /* we must be a new uber master sibling thread */
240   if (gtid == KMP_GTID_DNE) {
241     KA_TRACE(10,
242              ("__kmp_get_global_thread_id_reg: Encountered new root thread. "
243               "Registering a new gtid.\n"));
244     __kmp_acquire_bootstrap_lock(&__kmp_initz_lock);
245     if (!__kmp_init_serial) {
246       __kmp_do_serial_initialize();
247       gtid = __kmp_gtid_get_specific();
248     } else {
249       gtid = __kmp_register_root(FALSE);
250     }
251     __kmp_release_bootstrap_lock(&__kmp_initz_lock);
252     /*__kmp_printf( "+++ %d\n", gtid ); */ /* GROO */
253   }
254 
255   KMP_DEBUG_ASSERT(gtid >= 0);
256 
257   return gtid;
258 }
259 
260 /* caller must hold forkjoin_lock */
261 void __kmp_check_stack_overlap(kmp_info_t *th) {
262   int f;
263   char *stack_beg = NULL;
264   char *stack_end = NULL;
265   int gtid;
266 
267   KA_TRACE(10, ("__kmp_check_stack_overlap: called\n"));
268   if (__kmp_storage_map) {
269     stack_end = (char *)th->th.th_info.ds.ds_stackbase;
270     stack_beg = stack_end - th->th.th_info.ds.ds_stacksize;
271 
272     gtid = __kmp_gtid_from_thread(th);
273 
274     if (gtid == KMP_GTID_MONITOR) {
275       __kmp_print_storage_map_gtid(
276           gtid, stack_beg, stack_end, th->th.th_info.ds.ds_stacksize,
277           "th_%s stack (%s)", "mon",
278           (th->th.th_info.ds.ds_stackgrow) ? "initial" : "actual");
279     } else {
280       __kmp_print_storage_map_gtid(
281           gtid, stack_beg, stack_end, th->th.th_info.ds.ds_stacksize,
282           "th_%d stack (%s)", gtid,
283           (th->th.th_info.ds.ds_stackgrow) ? "initial" : "actual");
284     }
285   }
286 
287   /* No point in checking ubermaster threads since they use refinement and
288    * cannot overlap */
289   gtid = __kmp_gtid_from_thread(th);
290   if (__kmp_env_checks == TRUE && !KMP_UBER_GTID(gtid)) {
291     KA_TRACE(10,
292              ("__kmp_check_stack_overlap: performing extensive checking\n"));
293     if (stack_beg == NULL) {
294       stack_end = (char *)th->th.th_info.ds.ds_stackbase;
295       stack_beg = stack_end - th->th.th_info.ds.ds_stacksize;
296     }
297 
298     for (f = 0; f < __kmp_threads_capacity; f++) {
299       kmp_info_t *f_th = (kmp_info_t *)TCR_SYNC_PTR(__kmp_threads[f]);
300 
301       if (f_th && f_th != th) {
302         char *other_stack_end =
303             (char *)TCR_PTR(f_th->th.th_info.ds.ds_stackbase);
304         char *other_stack_beg =
305             other_stack_end - (size_t)TCR_PTR(f_th->th.th_info.ds.ds_stacksize);
306         if ((stack_beg > other_stack_beg && stack_beg < other_stack_end) ||
307             (stack_end > other_stack_beg && stack_end < other_stack_end)) {
308 
309           /* Print the other stack values before the abort */
310           if (__kmp_storage_map)
311             __kmp_print_storage_map_gtid(
312                 -1, other_stack_beg, other_stack_end,
313                 (size_t)TCR_PTR(f_th->th.th_info.ds.ds_stacksize),
314                 "th_%d stack (overlapped)", __kmp_gtid_from_thread(f_th));
315 
316           __kmp_fatal(KMP_MSG(StackOverlap), KMP_HNT(ChangeStackLimit),
317                       __kmp_msg_null);
318         }
319       }
320     }
321   }
322   KA_TRACE(10, ("__kmp_check_stack_overlap: returning\n"));
323 }
324 
325 /* ------------------------------------------------------------------------ */
326 
327 void __kmp_infinite_loop(void) {
328   static int done = FALSE;
329 
330   while (!done) {
331     KMP_YIELD(1);
332   }
333 }
334 
335 #define MAX_MESSAGE 512
336 
337 void __kmp_print_storage_map_gtid(int gtid, void *p1, void *p2, size_t size,
338                                   char const *format, ...) {
339   char buffer[MAX_MESSAGE];
340   va_list ap;
341 
342   va_start(ap, format);
343   KMP_SNPRINTF(buffer, sizeof(buffer), "OMP storage map: %p %p%8lu %s\n", p1,
344                p2, (unsigned long)size, format);
345   __kmp_acquire_bootstrap_lock(&__kmp_stdio_lock);
346   __kmp_vprintf(kmp_err, buffer, ap);
347 #if KMP_PRINT_DATA_PLACEMENT
348   int node;
349   if (gtid >= 0) {
350     if (p1 <= p2 && (char *)p2 - (char *)p1 == size) {
351       if (__kmp_storage_map_verbose) {
352         node = __kmp_get_host_node(p1);
353         if (node < 0) /* doesn't work, so don't try this next time */
354           __kmp_storage_map_verbose = FALSE;
355         else {
356           char *last;
357           int lastNode;
358           int localProc = __kmp_get_cpu_from_gtid(gtid);
359 
360           const int page_size = KMP_GET_PAGE_SIZE();
361 
362           p1 = (void *)((size_t)p1 & ~((size_t)page_size - 1));
363           p2 = (void *)(((size_t)p2 - 1) & ~((size_t)page_size - 1));
364           if (localProc >= 0)
365             __kmp_printf_no_lock("  GTID %d localNode %d\n", gtid,
366                                  localProc >> 1);
367           else
368             __kmp_printf_no_lock("  GTID %d\n", gtid);
369 #if KMP_USE_PRCTL
370           /* The more elaborate format is disabled for now because of the prctl
371            * hanging bug. */
372           do {
373             last = p1;
374             lastNode = node;
375             /* This loop collates adjacent pages with the same host node. */
376             do {
377               (char *)p1 += page_size;
378             } while (p1 <= p2 && (node = __kmp_get_host_node(p1)) == lastNode);
379             __kmp_printf_no_lock("    %p-%p memNode %d\n", last, (char *)p1 - 1,
380                                  lastNode);
381           } while (p1 <= p2);
382 #else
383           __kmp_printf_no_lock("    %p-%p memNode %d\n", p1,
384                                (char *)p1 + (page_size - 1),
385                                __kmp_get_host_node(p1));
386           if (p1 < p2) {
387             __kmp_printf_no_lock("    %p-%p memNode %d\n", p2,
388                                  (char *)p2 + (page_size - 1),
389                                  __kmp_get_host_node(p2));
390           }
391 #endif
392         }
393       }
394     } else
395       __kmp_printf_no_lock("  %s\n", KMP_I18N_STR(StorageMapWarning));
396   }
397 #endif /* KMP_PRINT_DATA_PLACEMENT */
398   __kmp_release_bootstrap_lock(&__kmp_stdio_lock);
399 }
400 
401 void __kmp_warn(char const *format, ...) {
402   char buffer[MAX_MESSAGE];
403   va_list ap;
404 
405   if (__kmp_generate_warnings == kmp_warnings_off) {
406     return;
407   }
408 
409   va_start(ap, format);
410 
411   KMP_SNPRINTF(buffer, sizeof(buffer), "OMP warning: %s\n", format);
412   __kmp_acquire_bootstrap_lock(&__kmp_stdio_lock);
413   __kmp_vprintf(kmp_err, buffer, ap);
414   __kmp_release_bootstrap_lock(&__kmp_stdio_lock);
415 
416   va_end(ap);
417 }
418 
419 void __kmp_abort_process() {
420   // Later threads may stall here, but that's ok because abort() will kill them.
421   __kmp_acquire_bootstrap_lock(&__kmp_exit_lock);
422 
423   if (__kmp_debug_buf) {
424     __kmp_dump_debug_buffer();
425   }
426 
427   if (KMP_OS_WINDOWS) {
428     // Let other threads know of abnormal termination and prevent deadlock
429     // if abort happened during library initialization or shutdown
430     __kmp_global.g.g_abort = SIGABRT;
431 
432     /* On Windows* OS by default abort() causes pop-up error box, which stalls
433        nightly testing. Unfortunately, we cannot reliably suppress pop-up error
434        boxes. _set_abort_behavior() works well, but this function is not
435        available in VS7 (this is not problem for DLL, but it is a problem for
436        static OpenMP RTL). SetErrorMode (and so, timelimit utility) does not
437        help, at least in some versions of MS C RTL.
438 
439        It seems following sequence is the only way to simulate abort() and
440        avoid pop-up error box. */
441     raise(SIGABRT);
442     _exit(3); // Just in case, if signal ignored, exit anyway.
443   } else {
444     abort();
445   }
446 
447   __kmp_infinite_loop();
448   __kmp_release_bootstrap_lock(&__kmp_exit_lock);
449 
450 } // __kmp_abort_process
451 
452 void __kmp_abort_thread(void) {
453   // TODO: Eliminate g_abort global variable and this function.
454   // In case of abort just call abort(), it will kill all the threads.
455   __kmp_infinite_loop();
456 } // __kmp_abort_thread
457 
458 /* Print out the storage map for the major kmp_info_t thread data structures
459    that are allocated together. */
460 
461 static void __kmp_print_thread_storage_map(kmp_info_t *thr, int gtid) {
462   __kmp_print_storage_map_gtid(gtid, thr, thr + 1, sizeof(kmp_info_t), "th_%d",
463                                gtid);
464 
465   __kmp_print_storage_map_gtid(gtid, &thr->th.th_info, &thr->th.th_team,
466                                sizeof(kmp_desc_t), "th_%d.th_info", gtid);
467 
468   __kmp_print_storage_map_gtid(gtid, &thr->th.th_local, &thr->th.th_pri_head,
469                                sizeof(kmp_local_t), "th_%d.th_local", gtid);
470 
471   __kmp_print_storage_map_gtid(
472       gtid, &thr->th.th_bar[0], &thr->th.th_bar[bs_last_barrier],
473       sizeof(kmp_balign_t) * bs_last_barrier, "th_%d.th_bar", gtid);
474 
475   __kmp_print_storage_map_gtid(gtid, &thr->th.th_bar[bs_plain_barrier],
476                                &thr->th.th_bar[bs_plain_barrier + 1],
477                                sizeof(kmp_balign_t), "th_%d.th_bar[plain]",
478                                gtid);
479 
480   __kmp_print_storage_map_gtid(gtid, &thr->th.th_bar[bs_forkjoin_barrier],
481                                &thr->th.th_bar[bs_forkjoin_barrier + 1],
482                                sizeof(kmp_balign_t), "th_%d.th_bar[forkjoin]",
483                                gtid);
484 
485 #if KMP_FAST_REDUCTION_BARRIER
486   __kmp_print_storage_map_gtid(gtid, &thr->th.th_bar[bs_reduction_barrier],
487                                &thr->th.th_bar[bs_reduction_barrier + 1],
488                                sizeof(kmp_balign_t), "th_%d.th_bar[reduction]",
489                                gtid);
490 #endif // KMP_FAST_REDUCTION_BARRIER
491 }
492 
493 /* Print out the storage map for the major kmp_team_t team data structures
494    that are allocated together. */
495 
496 static void __kmp_print_team_storage_map(const char *header, kmp_team_t *team,
497                                          int team_id, int num_thr) {
498   int num_disp_buff = team->t.t_max_nproc > 1 ? __kmp_dispatch_num_buffers : 2;
499   __kmp_print_storage_map_gtid(-1, team, team + 1, sizeof(kmp_team_t), "%s_%d",
500                                header, team_id);
501 
502   __kmp_print_storage_map_gtid(-1, &team->t.t_bar[0],
503                                &team->t.t_bar[bs_last_barrier],
504                                sizeof(kmp_balign_team_t) * bs_last_barrier,
505                                "%s_%d.t_bar", header, team_id);
506 
507   __kmp_print_storage_map_gtid(-1, &team->t.t_bar[bs_plain_barrier],
508                                &team->t.t_bar[bs_plain_barrier + 1],
509                                sizeof(kmp_balign_team_t), "%s_%d.t_bar[plain]",
510                                header, team_id);
511 
512   __kmp_print_storage_map_gtid(-1, &team->t.t_bar[bs_forkjoin_barrier],
513                                &team->t.t_bar[bs_forkjoin_barrier + 1],
514                                sizeof(kmp_balign_team_t),
515                                "%s_%d.t_bar[forkjoin]", header, team_id);
516 
517 #if KMP_FAST_REDUCTION_BARRIER
518   __kmp_print_storage_map_gtid(-1, &team->t.t_bar[bs_reduction_barrier],
519                                &team->t.t_bar[bs_reduction_barrier + 1],
520                                sizeof(kmp_balign_team_t),
521                                "%s_%d.t_bar[reduction]", header, team_id);
522 #endif // KMP_FAST_REDUCTION_BARRIER
523 
524   __kmp_print_storage_map_gtid(
525       -1, &team->t.t_dispatch[0], &team->t.t_dispatch[num_thr],
526       sizeof(kmp_disp_t) * num_thr, "%s_%d.t_dispatch", header, team_id);
527 
528   __kmp_print_storage_map_gtid(
529       -1, &team->t.t_threads[0], &team->t.t_threads[num_thr],
530       sizeof(kmp_info_t *) * num_thr, "%s_%d.t_threads", header, team_id);
531 
532   __kmp_print_storage_map_gtid(-1, &team->t.t_disp_buffer[0],
533                                &team->t.t_disp_buffer[num_disp_buff],
534                                sizeof(dispatch_shared_info_t) * num_disp_buff,
535                                "%s_%d.t_disp_buffer", header, team_id);
536 
537   __kmp_print_storage_map_gtid(-1, &team->t.t_taskq, &team->t.t_copypriv_data,
538                                sizeof(kmp_taskq_t), "%s_%d.t_taskq", header,
539                                team_id);
540 }
541 
542 static void __kmp_init_allocator() {}
543 static void __kmp_fini_allocator() {}
544 
545 /* ------------------------------------------------------------------------ */
546 
547 #ifdef KMP_DYNAMIC_LIB
548 #if KMP_OS_WINDOWS
549 
550 static void __kmp_reset_lock(kmp_bootstrap_lock_t *lck) {
551   // TODO: Change to __kmp_break_bootstrap_lock().
552   __kmp_init_bootstrap_lock(lck); // make the lock released
553 }
554 
555 static void __kmp_reset_locks_on_process_detach(int gtid_req) {
556   int i;
557   int thread_count;
558 
559   // PROCESS_DETACH is expected to be called by a thread that executes
560   // ProcessExit() or FreeLibrary(). OS terminates other threads (except the one
561   // calling ProcessExit or FreeLibrary). So, it might be safe to access the
562   // __kmp_threads[] without taking the forkjoin_lock. However, in fact, some
563   // threads can be still alive here, although being about to be terminated. The
564   // threads in the array with ds_thread==0 are most suspicious. Actually, it
565   // can be not safe to access the __kmp_threads[].
566 
567   // TODO: does it make sense to check __kmp_roots[] ?
568 
569   // Let's check that there are no other alive threads registered with the OMP
570   // lib.
571   while (1) {
572     thread_count = 0;
573     for (i = 0; i < __kmp_threads_capacity; ++i) {
574       if (!__kmp_threads)
575         continue;
576       kmp_info_t *th = __kmp_threads[i];
577       if (th == NULL)
578         continue;
579       int gtid = th->th.th_info.ds.ds_gtid;
580       if (gtid == gtid_req)
581         continue;
582       if (gtid < 0)
583         continue;
584       DWORD exit_val;
585       int alive = __kmp_is_thread_alive(th, &exit_val);
586       if (alive) {
587         ++thread_count;
588       }
589     }
590     if (thread_count == 0)
591       break; // success
592   }
593 
594   // Assume that I'm alone. Now it might be safe to check and reset locks.
595   // __kmp_forkjoin_lock and __kmp_stdio_lock are expected to be reset.
596   __kmp_reset_lock(&__kmp_forkjoin_lock);
597 #ifdef KMP_DEBUG
598   __kmp_reset_lock(&__kmp_stdio_lock);
599 #endif // KMP_DEBUG
600 }
601 
602 BOOL WINAPI DllMain(HINSTANCE hInstDLL, DWORD fdwReason, LPVOID lpReserved) {
603   //__kmp_acquire_bootstrap_lock( &__kmp_initz_lock );
604 
605   switch (fdwReason) {
606 
607   case DLL_PROCESS_ATTACH:
608     KA_TRACE(10, ("DllMain: PROCESS_ATTACH\n"));
609 
610     return TRUE;
611 
612   case DLL_PROCESS_DETACH:
613     KA_TRACE(10, ("DllMain: PROCESS_DETACH T#%d\n", __kmp_gtid_get_specific()));
614 
615     if (lpReserved != NULL) {
616       // lpReserved is used for telling the difference:
617       //   lpReserved == NULL when FreeLibrary() was called,
618       //   lpReserved != NULL when the process terminates.
619       // When FreeLibrary() is called, worker threads remain alive. So they will
620       // release the forkjoin lock by themselves. When the process terminates,
621       // worker threads disappear triggering the problem of unreleased forkjoin
622       // lock as described below.
623 
624       // A worker thread can take the forkjoin lock. The problem comes up if
625       // that worker thread becomes dead before it releases the forkjoin lock.
626       // The forkjoin lock remains taken, while the thread executing
627       // DllMain()->PROCESS_DETACH->__kmp_internal_end_library() below will try
628       // to take the forkjoin lock and will always fail, so that the application
629       // will never finish [normally]. This scenario is possible if
630       // __kmpc_end() has not been executed. It looks like it's not a corner
631       // case, but common cases:
632       // - the main function was compiled by an alternative compiler;
633       // - the main function was compiled by icl but without /Qopenmp
634       //   (application with plugins);
635       // - application terminates by calling C exit(), Fortran CALL EXIT() or
636       //   Fortran STOP.
637       // - alive foreign thread prevented __kmpc_end from doing cleanup.
638       //
639       // This is a hack to work around the problem.
640       // TODO: !!! figure out something better.
641       __kmp_reset_locks_on_process_detach(__kmp_gtid_get_specific());
642     }
643 
644     __kmp_internal_end_library(__kmp_gtid_get_specific());
645 
646     return TRUE;
647 
648   case DLL_THREAD_ATTACH:
649     KA_TRACE(10, ("DllMain: THREAD_ATTACH\n"));
650 
651     /* if we want to register new siblings all the time here call
652      * __kmp_get_gtid(); */
653     return TRUE;
654 
655   case DLL_THREAD_DETACH:
656     KA_TRACE(10, ("DllMain: THREAD_DETACH T#%d\n", __kmp_gtid_get_specific()));
657 
658     __kmp_internal_end_thread(__kmp_gtid_get_specific());
659     return TRUE;
660   }
661 
662   return TRUE;
663 }
664 
665 #endif /* KMP_OS_WINDOWS */
666 #endif /* KMP_DYNAMIC_LIB */
667 
668 /* Change the library type to "status" and return the old type */
669 /* called from within initialization routines where __kmp_initz_lock is held */
670 int __kmp_change_library(int status) {
671   int old_status;
672 
673   old_status = __kmp_yield_init &
674                1; // check whether KMP_LIBRARY=throughput (even init count)
675 
676   if (status) {
677     __kmp_yield_init |= 1; // throughput => turnaround (odd init count)
678   } else {
679     __kmp_yield_init &= ~1; // turnaround => throughput (even init count)
680   }
681 
682   return old_status; // return previous setting of whether
683   // KMP_LIBRARY=throughput
684 }
685 
686 /* __kmp_parallel_deo -- Wait until it's our turn. */
687 void __kmp_parallel_deo(int *gtid_ref, int *cid_ref, ident_t *loc_ref) {
688   int gtid = *gtid_ref;
689 #ifdef BUILD_PARALLEL_ORDERED
690   kmp_team_t *team = __kmp_team_from_gtid(gtid);
691 #endif /* BUILD_PARALLEL_ORDERED */
692 
693   if (__kmp_env_consistency_check) {
694     if (__kmp_threads[gtid]->th.th_root->r.r_active)
695 #if KMP_USE_DYNAMIC_LOCK
696       __kmp_push_sync(gtid, ct_ordered_in_parallel, loc_ref, NULL, 0);
697 #else
698       __kmp_push_sync(gtid, ct_ordered_in_parallel, loc_ref, NULL);
699 #endif
700   }
701 #ifdef BUILD_PARALLEL_ORDERED
702   if (!team->t.t_serialized) {
703     KMP_MB();
704     KMP_WAIT_YIELD(&team->t.t_ordered.dt.t_value, __kmp_tid_from_gtid(gtid),
705                    KMP_EQ, NULL);
706     KMP_MB();
707   }
708 #endif /* BUILD_PARALLEL_ORDERED */
709 }
710 
711 /* __kmp_parallel_dxo -- Signal the next task. */
712 void __kmp_parallel_dxo(int *gtid_ref, int *cid_ref, ident_t *loc_ref) {
713   int gtid = *gtid_ref;
714 #ifdef BUILD_PARALLEL_ORDERED
715   int tid = __kmp_tid_from_gtid(gtid);
716   kmp_team_t *team = __kmp_team_from_gtid(gtid);
717 #endif /* BUILD_PARALLEL_ORDERED */
718 
719   if (__kmp_env_consistency_check) {
720     if (__kmp_threads[gtid]->th.th_root->r.r_active)
721       __kmp_pop_sync(gtid, ct_ordered_in_parallel, loc_ref);
722   }
723 #ifdef BUILD_PARALLEL_ORDERED
724   if (!team->t.t_serialized) {
725     KMP_MB(); /* Flush all pending memory write invalidates.  */
726 
727     /* use the tid of the next thread in this team */
728     /* TODO replace with general release procedure */
729     team->t.t_ordered.dt.t_value = ((tid + 1) % team->t.t_nproc);
730 
731     KMP_MB(); /* Flush all pending memory write invalidates.  */
732   }
733 #endif /* BUILD_PARALLEL_ORDERED */
734 }
735 
736 /* ------------------------------------------------------------------------ */
737 /* The BARRIER for a SINGLE process section is always explicit   */
738 
739 int __kmp_enter_single(int gtid, ident_t *id_ref, int push_ws) {
740   int status;
741   kmp_info_t *th;
742   kmp_team_t *team;
743 
744   if (!TCR_4(__kmp_init_parallel))
745     __kmp_parallel_initialize();
746 
747   th = __kmp_threads[gtid];
748   team = th->th.th_team;
749   status = 0;
750 
751   th->th.th_ident = id_ref;
752 
753   if (team->t.t_serialized) {
754     status = 1;
755   } else {
756     kmp_int32 old_this = th->th.th_local.this_construct;
757 
758     ++th->th.th_local.this_construct;
759     /* try to set team count to thread count--success means thread got the
760        single block */
761     /* TODO: Should this be acquire or release? */
762     if (team->t.t_construct == old_this) {
763       status = __kmp_atomic_compare_store_acq(&team->t.t_construct, old_this,
764                                               th->th.th_local.this_construct);
765     }
766 #if USE_ITT_BUILD
767     if (__itt_metadata_add_ptr && __kmp_forkjoin_frames_mode == 3 &&
768         KMP_MASTER_GTID(gtid) &&
769 #if OMP_40_ENABLED
770         th->th.th_teams_microtask == NULL &&
771 #endif
772         team->t.t_active_level ==
773             1) { // Only report metadata by master of active team at level 1
774       __kmp_itt_metadata_single(id_ref);
775     }
776 #endif /* USE_ITT_BUILD */
777   }
778 
779   if (__kmp_env_consistency_check) {
780     if (status && push_ws) {
781       __kmp_push_workshare(gtid, ct_psingle, id_ref);
782     } else {
783       __kmp_check_workshare(gtid, ct_psingle, id_ref);
784     }
785   }
786 #if USE_ITT_BUILD
787   if (status) {
788     __kmp_itt_single_start(gtid);
789   }
790 #endif /* USE_ITT_BUILD */
791   return status;
792 }
793 
794 void __kmp_exit_single(int gtid) {
795 #if USE_ITT_BUILD
796   __kmp_itt_single_end(gtid);
797 #endif /* USE_ITT_BUILD */
798   if (__kmp_env_consistency_check)
799     __kmp_pop_workshare(gtid, ct_psingle, NULL);
800 }
801 
802 /* determine if we can go parallel or must use a serialized parallel region and
803  * how many threads we can use
804  * set_nproc is the number of threads requested for the team
805  * returns 0 if we should serialize or only use one thread,
806  * otherwise the number of threads to use
807  * The forkjoin lock is held by the caller. */
808 static int __kmp_reserve_threads(kmp_root_t *root, kmp_team_t *parent_team,
809                                  int master_tid, int set_nthreads
810 #if OMP_40_ENABLED
811                                  ,
812                                  int enter_teams
813 #endif /* OMP_40_ENABLED */
814                                  ) {
815   int capacity;
816   int new_nthreads;
817   KMP_DEBUG_ASSERT(__kmp_init_serial);
818   KMP_DEBUG_ASSERT(root && parent_team);
819 
820   // If dyn-var is set, dynamically adjust the number of desired threads,
821   // according to the method specified by dynamic_mode.
822   new_nthreads = set_nthreads;
823   if (!get__dynamic_2(parent_team, master_tid)) {
824     ;
825   }
826 #ifdef USE_LOAD_BALANCE
827   else if (__kmp_global.g.g_dynamic_mode == dynamic_load_balance) {
828     new_nthreads = __kmp_load_balance_nproc(root, set_nthreads);
829     if (new_nthreads == 1) {
830       KC_TRACE(10, ("__kmp_reserve_threads: T#%d load balance reduced "
831                     "reservation to 1 thread\n",
832                     master_tid));
833       return 1;
834     }
835     if (new_nthreads < set_nthreads) {
836       KC_TRACE(10, ("__kmp_reserve_threads: T#%d load balance reduced "
837                     "reservation to %d threads\n",
838                     master_tid, new_nthreads));
839     }
840   }
841 #endif /* USE_LOAD_BALANCE */
842   else if (__kmp_global.g.g_dynamic_mode == dynamic_thread_limit) {
843     new_nthreads = __kmp_avail_proc - __kmp_nth +
844                    (root->r.r_active ? 1 : root->r.r_hot_team->t.t_nproc);
845     if (new_nthreads <= 1) {
846       KC_TRACE(10, ("__kmp_reserve_threads: T#%d thread limit reduced "
847                     "reservation to 1 thread\n",
848                     master_tid));
849       return 1;
850     }
851     if (new_nthreads < set_nthreads) {
852       KC_TRACE(10, ("__kmp_reserve_threads: T#%d thread limit reduced "
853                     "reservation to %d threads\n",
854                     master_tid, new_nthreads));
855     } else {
856       new_nthreads = set_nthreads;
857     }
858   } else if (__kmp_global.g.g_dynamic_mode == dynamic_random) {
859     if (set_nthreads > 2) {
860       new_nthreads = __kmp_get_random(parent_team->t.t_threads[master_tid]);
861       new_nthreads = (new_nthreads % set_nthreads) + 1;
862       if (new_nthreads == 1) {
863         KC_TRACE(10, ("__kmp_reserve_threads: T#%d dynamic random reduced "
864                       "reservation to 1 thread\n",
865                       master_tid));
866         return 1;
867       }
868       if (new_nthreads < set_nthreads) {
869         KC_TRACE(10, ("__kmp_reserve_threads: T#%d dynamic random reduced "
870                       "reservation to %d threads\n",
871                       master_tid, new_nthreads));
872       }
873     }
874   } else {
875     KMP_ASSERT(0);
876   }
877 
878   // Respect KMP_ALL_THREADS/KMP_DEVICE_THREAD_LIMIT.
879   if (__kmp_nth + new_nthreads -
880           (root->r.r_active ? 1 : root->r.r_hot_team->t.t_nproc) >
881       __kmp_max_nth) {
882     int tl_nthreads = __kmp_max_nth - __kmp_nth +
883                       (root->r.r_active ? 1 : root->r.r_hot_team->t.t_nproc);
884     if (tl_nthreads <= 0) {
885       tl_nthreads = 1;
886     }
887 
888     // If dyn-var is false, emit a 1-time warning.
889     if (!get__dynamic_2(parent_team, master_tid) && (!__kmp_reserve_warn)) {
890       __kmp_reserve_warn = 1;
891       __kmp_msg(kmp_ms_warning,
892                 KMP_MSG(CantFormThrTeam, set_nthreads, tl_nthreads),
893                 KMP_HNT(Unset_ALL_THREADS), __kmp_msg_null);
894     }
895     if (tl_nthreads == 1) {
896       KC_TRACE(10, ("__kmp_reserve_threads: T#%d KMP_DEVICE_THREAD_LIMIT "
897                     "reduced reservation to 1 thread\n",
898                     master_tid));
899       return 1;
900     }
901     KC_TRACE(10, ("__kmp_reserve_threads: T#%d KMP_DEVICE_THREAD_LIMIT reduced "
902                   "reservation to %d threads\n",
903                   master_tid, tl_nthreads));
904     new_nthreads = tl_nthreads;
905   }
906 
907   // Respect OMP_THREAD_LIMIT
908   if (root->r.r_cg_nthreads + new_nthreads -
909           (root->r.r_active ? 1 : root->r.r_hot_team->t.t_nproc) >
910       __kmp_cg_max_nth) {
911     int tl_nthreads = __kmp_cg_max_nth - root->r.r_cg_nthreads +
912                       (root->r.r_active ? 1 : root->r.r_hot_team->t.t_nproc);
913     if (tl_nthreads <= 0) {
914       tl_nthreads = 1;
915     }
916 
917     // If dyn-var is false, emit a 1-time warning.
918     if (!get__dynamic_2(parent_team, master_tid) && (!__kmp_reserve_warn)) {
919       __kmp_reserve_warn = 1;
920       __kmp_msg(kmp_ms_warning,
921                 KMP_MSG(CantFormThrTeam, set_nthreads, tl_nthreads),
922                 KMP_HNT(Unset_ALL_THREADS), __kmp_msg_null);
923     }
924     if (tl_nthreads == 1) {
925       KC_TRACE(10, ("__kmp_reserve_threads: T#%d OMP_THREAD_LIMIT "
926                     "reduced reservation to 1 thread\n",
927                     master_tid));
928       return 1;
929     }
930     KC_TRACE(10, ("__kmp_reserve_threads: T#%d OMP_THREAD_LIMIT reduced "
931                   "reservation to %d threads\n",
932                   master_tid, tl_nthreads));
933     new_nthreads = tl_nthreads;
934   }
935 
936   // Check if the threads array is large enough, or needs expanding.
937   // See comment in __kmp_register_root() about the adjustment if
938   // __kmp_threads[0] == NULL.
939   capacity = __kmp_threads_capacity;
940   if (TCR_PTR(__kmp_threads[0]) == NULL) {
941     --capacity;
942   }
943   if (__kmp_nth + new_nthreads -
944           (root->r.r_active ? 1 : root->r.r_hot_team->t.t_nproc) >
945       capacity) {
946     // Expand the threads array.
947     int slotsRequired = __kmp_nth + new_nthreads -
948                         (root->r.r_active ? 1 : root->r.r_hot_team->t.t_nproc) -
949                         capacity;
950     int slotsAdded = __kmp_expand_threads(slotsRequired);
951     if (slotsAdded < slotsRequired) {
952       // The threads array was not expanded enough.
953       new_nthreads -= (slotsRequired - slotsAdded);
954       KMP_ASSERT(new_nthreads >= 1);
955 
956       // If dyn-var is false, emit a 1-time warning.
957       if (!get__dynamic_2(parent_team, master_tid) && (!__kmp_reserve_warn)) {
958         __kmp_reserve_warn = 1;
959         if (__kmp_tp_cached) {
960           __kmp_msg(kmp_ms_warning,
961                     KMP_MSG(CantFormThrTeam, set_nthreads, new_nthreads),
962                     KMP_HNT(Set_ALL_THREADPRIVATE, __kmp_tp_capacity),
963                     KMP_HNT(PossibleSystemLimitOnThreads), __kmp_msg_null);
964         } else {
965           __kmp_msg(kmp_ms_warning,
966                     KMP_MSG(CantFormThrTeam, set_nthreads, new_nthreads),
967                     KMP_HNT(SystemLimitOnThreads), __kmp_msg_null);
968         }
969       }
970     }
971   }
972 
973 #ifdef KMP_DEBUG
974   if (new_nthreads == 1) {
975     KC_TRACE(10,
976              ("__kmp_reserve_threads: T#%d serializing team after reclaiming "
977               "dead roots and rechecking; requested %d threads\n",
978               __kmp_get_gtid(), set_nthreads));
979   } else {
980     KC_TRACE(10, ("__kmp_reserve_threads: T#%d allocating %d threads; requested"
981                   " %d threads\n",
982                   __kmp_get_gtid(), new_nthreads, set_nthreads));
983   }
984 #endif // KMP_DEBUG
985   return new_nthreads;
986 }
987 
988 /* Allocate threads from the thread pool and assign them to the new team. We are
989    assured that there are enough threads available, because we checked on that
990    earlier within critical section forkjoin */
991 static void __kmp_fork_team_threads(kmp_root_t *root, kmp_team_t *team,
992                                     kmp_info_t *master_th, int master_gtid) {
993   int i;
994   int use_hot_team;
995 
996   KA_TRACE(10, ("__kmp_fork_team_threads: new_nprocs = %d\n", team->t.t_nproc));
997   KMP_DEBUG_ASSERT(master_gtid == __kmp_get_gtid());
998   KMP_MB();
999 
1000   /* first, let's setup the master thread */
1001   master_th->th.th_info.ds.ds_tid = 0;
1002   master_th->th.th_team = team;
1003   master_th->th.th_team_nproc = team->t.t_nproc;
1004   master_th->th.th_team_master = master_th;
1005   master_th->th.th_team_serialized = FALSE;
1006   master_th->th.th_dispatch = &team->t.t_dispatch[0];
1007 
1008 /* make sure we are not the optimized hot team */
1009 #if KMP_NESTED_HOT_TEAMS
1010   use_hot_team = 0;
1011   kmp_hot_team_ptr_t *hot_teams = master_th->th.th_hot_teams;
1012   if (hot_teams) { // hot teams array is not allocated if
1013     // KMP_HOT_TEAMS_MAX_LEVEL=0
1014     int level = team->t.t_active_level - 1; // index in array of hot teams
1015     if (master_th->th.th_teams_microtask) { // are we inside the teams?
1016       if (master_th->th.th_teams_size.nteams > 1) {
1017         ++level; // level was not increased in teams construct for
1018         // team_of_masters
1019       }
1020       if (team->t.t_pkfn != (microtask_t)__kmp_teams_master &&
1021           master_th->th.th_teams_level == team->t.t_level) {
1022         ++level; // level was not increased in teams construct for
1023         // team_of_workers before the parallel
1024       } // team->t.t_level will be increased inside parallel
1025     }
1026     if (level < __kmp_hot_teams_max_level) {
1027       if (hot_teams[level].hot_team) {
1028         // hot team has already been allocated for given level
1029         KMP_DEBUG_ASSERT(hot_teams[level].hot_team == team);
1030         use_hot_team = 1; // the team is ready to use
1031       } else {
1032         use_hot_team = 0; // AC: threads are not allocated yet
1033         hot_teams[level].hot_team = team; // remember new hot team
1034         hot_teams[level].hot_team_nth = team->t.t_nproc;
1035       }
1036     } else {
1037       use_hot_team = 0;
1038     }
1039   }
1040 #else
1041   use_hot_team = team == root->r.r_hot_team;
1042 #endif
1043   if (!use_hot_team) {
1044 
1045     /* install the master thread */
1046     team->t.t_threads[0] = master_th;
1047     __kmp_initialize_info(master_th, team, 0, master_gtid);
1048 
1049     /* now, install the worker threads */
1050     for (i = 1; i < team->t.t_nproc; i++) {
1051 
1052       /* fork or reallocate a new thread and install it in team */
1053       kmp_info_t *thr = __kmp_allocate_thread(root, team, i);
1054       team->t.t_threads[i] = thr;
1055       KMP_DEBUG_ASSERT(thr);
1056       KMP_DEBUG_ASSERT(thr->th.th_team == team);
1057       /* align team and thread arrived states */
1058       KA_TRACE(20, ("__kmp_fork_team_threads: T#%d(%d:%d) init arrived "
1059                     "T#%d(%d:%d) join =%llu, plain=%llu\n",
1060                     __kmp_gtid_from_tid(0, team), team->t.t_id, 0,
1061                     __kmp_gtid_from_tid(i, team), team->t.t_id, i,
1062                     team->t.t_bar[bs_forkjoin_barrier].b_arrived,
1063                     team->t.t_bar[bs_plain_barrier].b_arrived));
1064 #if OMP_40_ENABLED
1065       thr->th.th_teams_microtask = master_th->th.th_teams_microtask;
1066       thr->th.th_teams_level = master_th->th.th_teams_level;
1067       thr->th.th_teams_size = master_th->th.th_teams_size;
1068 #endif
1069       { // Initialize threads' barrier data.
1070         int b;
1071         kmp_balign_t *balign = team->t.t_threads[i]->th.th_bar;
1072         for (b = 0; b < bs_last_barrier; ++b) {
1073           balign[b].bb.b_arrived = team->t.t_bar[b].b_arrived;
1074           KMP_DEBUG_ASSERT(balign[b].bb.wait_flag != KMP_BARRIER_PARENT_FLAG);
1075 #if USE_DEBUGGER
1076           balign[b].bb.b_worker_arrived = team->t.t_bar[b].b_team_arrived;
1077 #endif
1078         }
1079       }
1080     }
1081 
1082 #if OMP_40_ENABLED && KMP_AFFINITY_SUPPORTED
1083     __kmp_partition_places(team);
1084 #endif
1085   }
1086 
1087   KMP_MB();
1088 }
1089 
1090 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
1091 // Propagate any changes to the floating point control registers out to the team
1092 // We try to avoid unnecessary writes to the relevant cache line in the team
1093 // structure, so we don't make changes unless they are needed.
1094 inline static void propagateFPControl(kmp_team_t *team) {
1095   if (__kmp_inherit_fp_control) {
1096     kmp_int16 x87_fpu_control_word;
1097     kmp_uint32 mxcsr;
1098 
1099     // Get master values of FPU control flags (both X87 and vector)
1100     __kmp_store_x87_fpu_control_word(&x87_fpu_control_word);
1101     __kmp_store_mxcsr(&mxcsr);
1102     mxcsr &= KMP_X86_MXCSR_MASK;
1103 
1104     // There is no point looking at t_fp_control_saved here.
1105     // If it is TRUE, we still have to update the values if they are different
1106     // from those we now have. If it is FALSE we didn't save anything yet, but
1107     // our objective is the same. We have to ensure that the values in the team
1108     // are the same as those we have.
1109     // So, this code achieves what we need whether or not t_fp_control_saved is
1110     // true. By checking whether the value needs updating we avoid unnecessary
1111     // writes that would put the cache-line into a written state, causing all
1112     // threads in the team to have to read it again.
1113     KMP_CHECK_UPDATE(team->t.t_x87_fpu_control_word, x87_fpu_control_word);
1114     KMP_CHECK_UPDATE(team->t.t_mxcsr, mxcsr);
1115     // Although we don't use this value, other code in the runtime wants to know
1116     // whether it should restore them. So we must ensure it is correct.
1117     KMP_CHECK_UPDATE(team->t.t_fp_control_saved, TRUE);
1118   } else {
1119     // Similarly here. Don't write to this cache-line in the team structure
1120     // unless we have to.
1121     KMP_CHECK_UPDATE(team->t.t_fp_control_saved, FALSE);
1122   }
1123 }
1124 
1125 // Do the opposite, setting the hardware registers to the updated values from
1126 // the team.
1127 inline static void updateHWFPControl(kmp_team_t *team) {
1128   if (__kmp_inherit_fp_control && team->t.t_fp_control_saved) {
1129     // Only reset the fp control regs if they have been changed in the team.
1130     // the parallel region that we are exiting.
1131     kmp_int16 x87_fpu_control_word;
1132     kmp_uint32 mxcsr;
1133     __kmp_store_x87_fpu_control_word(&x87_fpu_control_word);
1134     __kmp_store_mxcsr(&mxcsr);
1135     mxcsr &= KMP_X86_MXCSR_MASK;
1136 
1137     if (team->t.t_x87_fpu_control_word != x87_fpu_control_word) {
1138       __kmp_clear_x87_fpu_status_word();
1139       __kmp_load_x87_fpu_control_word(&team->t.t_x87_fpu_control_word);
1140     }
1141 
1142     if (team->t.t_mxcsr != mxcsr) {
1143       __kmp_load_mxcsr(&team->t.t_mxcsr);
1144     }
1145   }
1146 }
1147 #else
1148 #define propagateFPControl(x) ((void)0)
1149 #define updateHWFPControl(x) ((void)0)
1150 #endif /* KMP_ARCH_X86 || KMP_ARCH_X86_64 */
1151 
1152 static void __kmp_alloc_argv_entries(int argc, kmp_team_t *team,
1153                                      int realloc); // forward declaration
1154 
1155 /* Run a parallel region that has been serialized, so runs only in a team of the
1156    single master thread. */
1157 void __kmp_serialized_parallel(ident_t *loc, kmp_int32 global_tid) {
1158   kmp_info_t *this_thr;
1159   kmp_team_t *serial_team;
1160 
1161   KC_TRACE(10, ("__kmpc_serialized_parallel: called by T#%d\n", global_tid));
1162 
1163   /* Skip all this code for autopar serialized loops since it results in
1164      unacceptable overhead */
1165   if (loc != NULL && (loc->flags & KMP_IDENT_AUTOPAR))
1166     return;
1167 
1168   if (!TCR_4(__kmp_init_parallel))
1169     __kmp_parallel_initialize();
1170 
1171   this_thr = __kmp_threads[global_tid];
1172   serial_team = this_thr->th.th_serial_team;
1173 
1174   /* utilize the serialized team held by this thread */
1175   KMP_DEBUG_ASSERT(serial_team);
1176   KMP_MB();
1177 
1178   if (__kmp_tasking_mode != tskm_immediate_exec) {
1179     KMP_DEBUG_ASSERT(
1180         this_thr->th.th_task_team ==
1181         this_thr->th.th_team->t.t_task_team[this_thr->th.th_task_state]);
1182     KMP_DEBUG_ASSERT(serial_team->t.t_task_team[this_thr->th.th_task_state] ==
1183                      NULL);
1184     KA_TRACE(20, ("__kmpc_serialized_parallel: T#%d pushing task_team %p / "
1185                   "team %p, new task_team = NULL\n",
1186                   global_tid, this_thr->th.th_task_team, this_thr->th.th_team));
1187     this_thr->th.th_task_team = NULL;
1188   }
1189 
1190 #if OMP_40_ENABLED
1191   kmp_proc_bind_t proc_bind = this_thr->th.th_set_proc_bind;
1192   if (this_thr->th.th_current_task->td_icvs.proc_bind == proc_bind_false) {
1193     proc_bind = proc_bind_false;
1194   } else if (proc_bind == proc_bind_default) {
1195     // No proc_bind clause was specified, so use the current value
1196     // of proc-bind-var for this parallel region.
1197     proc_bind = this_thr->th.th_current_task->td_icvs.proc_bind;
1198   }
1199   // Reset for next parallel region
1200   this_thr->th.th_set_proc_bind = proc_bind_default;
1201 #endif /* OMP_40_ENABLED */
1202 
1203 #if OMPT_SUPPORT
1204   ompt_data_t ompt_parallel_data;
1205   ompt_parallel_data.ptr = NULL;
1206   ompt_data_t *implicit_task_data;
1207   void *codeptr = OMPT_LOAD_RETURN_ADDRESS(global_tid);
1208   if (ompt_enabled.enabled &&
1209       this_thr->th.ompt_thread_info.state != omp_state_overhead) {
1210 
1211     ompt_task_info_t *parent_task_info;
1212     parent_task_info = OMPT_CUR_TASK_INFO(this_thr);
1213 
1214     parent_task_info->frame.enter_frame = OMPT_GET_FRAME_ADDRESS(1);
1215     if (ompt_enabled.ompt_callback_parallel_begin) {
1216       int team_size = 1;
1217 
1218       ompt_callbacks.ompt_callback(ompt_callback_parallel_begin)(
1219           &(parent_task_info->task_data), &(parent_task_info->frame),
1220           &ompt_parallel_data, team_size, ompt_invoker_program, codeptr);
1221     }
1222   }
1223 #endif // OMPT_SUPPORT
1224 
1225   if (this_thr->th.th_team != serial_team) {
1226     // Nested level will be an index in the nested nthreads array
1227     int level = this_thr->th.th_team->t.t_level;
1228 
1229     if (serial_team->t.t_serialized) {
1230       /* this serial team was already used
1231          TODO increase performance by making this locks more specific */
1232       kmp_team_t *new_team;
1233 
1234       __kmp_acquire_bootstrap_lock(&__kmp_forkjoin_lock);
1235 
1236       new_team = __kmp_allocate_team(this_thr->th.th_root, 1, 1,
1237 #if OMPT_SUPPORT
1238                                      ompt_parallel_data,
1239 #endif
1240 #if OMP_40_ENABLED
1241                                      proc_bind,
1242 #endif
1243                                      &this_thr->th.th_current_task->td_icvs,
1244                                      0 USE_NESTED_HOT_ARG(NULL));
1245       __kmp_release_bootstrap_lock(&__kmp_forkjoin_lock);
1246       KMP_ASSERT(new_team);
1247 
1248       /* setup new serialized team and install it */
1249       new_team->t.t_threads[0] = this_thr;
1250       new_team->t.t_parent = this_thr->th.th_team;
1251       serial_team = new_team;
1252       this_thr->th.th_serial_team = serial_team;
1253 
1254       KF_TRACE(
1255           10,
1256           ("__kmpc_serialized_parallel: T#%d allocated new serial team %p\n",
1257            global_tid, serial_team));
1258 
1259       /* TODO the above breaks the requirement that if we run out of resources,
1260          then we can still guarantee that serialized teams are ok, since we may
1261          need to allocate a new one */
1262     } else {
1263       KF_TRACE(
1264           10,
1265           ("__kmpc_serialized_parallel: T#%d reusing cached serial team %p\n",
1266            global_tid, serial_team));
1267     }
1268 
1269     /* we have to initialize this serial team */
1270     KMP_DEBUG_ASSERT(serial_team->t.t_threads);
1271     KMP_DEBUG_ASSERT(serial_team->t.t_threads[0] == this_thr);
1272     KMP_DEBUG_ASSERT(this_thr->th.th_team != serial_team);
1273     serial_team->t.t_ident = loc;
1274     serial_team->t.t_serialized = 1;
1275     serial_team->t.t_nproc = 1;
1276     serial_team->t.t_parent = this_thr->th.th_team;
1277     serial_team->t.t_sched.sched = this_thr->th.th_team->t.t_sched.sched;
1278     this_thr->th.th_team = serial_team;
1279     serial_team->t.t_master_tid = this_thr->th.th_info.ds.ds_tid;
1280 
1281     KF_TRACE(10, ("__kmpc_serialized_parallel: T#d curtask=%p\n", global_tid,
1282                   this_thr->th.th_current_task));
1283     KMP_ASSERT(this_thr->th.th_current_task->td_flags.executing == 1);
1284     this_thr->th.th_current_task->td_flags.executing = 0;
1285 
1286     __kmp_push_current_task_to_thread(this_thr, serial_team, 0);
1287 
1288     /* TODO: GEH: do ICVs work for nested serialized teams? Don't we need an
1289        implicit task for each serialized task represented by
1290        team->t.t_serialized? */
1291     copy_icvs(&this_thr->th.th_current_task->td_icvs,
1292               &this_thr->th.th_current_task->td_parent->td_icvs);
1293 
1294     // Thread value exists in the nested nthreads array for the next nested
1295     // level
1296     if (__kmp_nested_nth.used && (level + 1 < __kmp_nested_nth.used)) {
1297       this_thr->th.th_current_task->td_icvs.nproc =
1298           __kmp_nested_nth.nth[level + 1];
1299     }
1300 
1301 #if OMP_40_ENABLED
1302     if (__kmp_nested_proc_bind.used &&
1303         (level + 1 < __kmp_nested_proc_bind.used)) {
1304       this_thr->th.th_current_task->td_icvs.proc_bind =
1305           __kmp_nested_proc_bind.bind_types[level + 1];
1306     }
1307 #endif /* OMP_40_ENABLED */
1308 
1309 #if USE_DEBUGGER
1310     serial_team->t.t_pkfn = (microtask_t)(~0); // For the debugger.
1311 #endif
1312     this_thr->th.th_info.ds.ds_tid = 0;
1313 
1314     /* set thread cache values */
1315     this_thr->th.th_team_nproc = 1;
1316     this_thr->th.th_team_master = this_thr;
1317     this_thr->th.th_team_serialized = 1;
1318 
1319     serial_team->t.t_level = serial_team->t.t_parent->t.t_level + 1;
1320     serial_team->t.t_active_level = serial_team->t.t_parent->t.t_active_level;
1321 
1322     propagateFPControl(serial_team);
1323 
1324     /* check if we need to allocate dispatch buffers stack */
1325     KMP_DEBUG_ASSERT(serial_team->t.t_dispatch);
1326     if (!serial_team->t.t_dispatch->th_disp_buffer) {
1327       serial_team->t.t_dispatch->th_disp_buffer =
1328           (dispatch_private_info_t *)__kmp_allocate(
1329               sizeof(dispatch_private_info_t));
1330     }
1331     this_thr->th.th_dispatch = serial_team->t.t_dispatch;
1332 
1333     KMP_MB();
1334 
1335   } else {
1336     /* this serialized team is already being used,
1337      * that's fine, just add another nested level */
1338     KMP_DEBUG_ASSERT(this_thr->th.th_team == serial_team);
1339     KMP_DEBUG_ASSERT(serial_team->t.t_threads);
1340     KMP_DEBUG_ASSERT(serial_team->t.t_threads[0] == this_thr);
1341     ++serial_team->t.t_serialized;
1342     this_thr->th.th_team_serialized = serial_team->t.t_serialized;
1343 
1344     // Nested level will be an index in the nested nthreads array
1345     int level = this_thr->th.th_team->t.t_level;
1346     // Thread value exists in the nested nthreads array for the next nested
1347     // level
1348     if (__kmp_nested_nth.used && (level + 1 < __kmp_nested_nth.used)) {
1349       this_thr->th.th_current_task->td_icvs.nproc =
1350           __kmp_nested_nth.nth[level + 1];
1351     }
1352     serial_team->t.t_level++;
1353     KF_TRACE(10, ("__kmpc_serialized_parallel: T#%d increasing nesting level "
1354                   "of serial team %p to %d\n",
1355                   global_tid, serial_team, serial_team->t.t_level));
1356 
1357     /* allocate/push dispatch buffers stack */
1358     KMP_DEBUG_ASSERT(serial_team->t.t_dispatch);
1359     {
1360       dispatch_private_info_t *disp_buffer =
1361           (dispatch_private_info_t *)__kmp_allocate(
1362               sizeof(dispatch_private_info_t));
1363       disp_buffer->next = serial_team->t.t_dispatch->th_disp_buffer;
1364       serial_team->t.t_dispatch->th_disp_buffer = disp_buffer;
1365     }
1366     this_thr->th.th_dispatch = serial_team->t.t_dispatch;
1367 
1368     KMP_MB();
1369   }
1370 #if OMP_40_ENABLED
1371   KMP_CHECK_UPDATE(serial_team->t.t_cancel_request, cancel_noreq);
1372 #endif
1373 
1374   if (__kmp_env_consistency_check)
1375     __kmp_push_parallel(global_tid, NULL);
1376 #if OMPT_SUPPORT
1377   serial_team->t.ompt_team_info.master_return_address = codeptr;
1378   if (ompt_enabled.enabled &&
1379       this_thr->th.ompt_thread_info.state != omp_state_overhead) {
1380     OMPT_CUR_TASK_INFO(this_thr)->frame.exit_frame = OMPT_GET_FRAME_ADDRESS(1);
1381 
1382     ompt_lw_taskteam_t lw_taskteam;
1383     __ompt_lw_taskteam_init(&lw_taskteam, this_thr, global_tid,
1384                             &ompt_parallel_data, codeptr);
1385 
1386     __ompt_lw_taskteam_link(&lw_taskteam, this_thr, 1);
1387     // don't use lw_taskteam after linking. content was swaped
1388 
1389     /* OMPT implicit task begin */
1390     implicit_task_data = OMPT_CUR_TASK_DATA(this_thr);
1391     if (ompt_enabled.ompt_callback_implicit_task) {
1392       ompt_callbacks.ompt_callback(ompt_callback_implicit_task)(
1393           ompt_scope_begin, OMPT_CUR_TEAM_DATA(this_thr),
1394           OMPT_CUR_TASK_DATA(this_thr), 1, __kmp_tid_from_gtid(global_tid));
1395       OMPT_CUR_TASK_INFO(this_thr)
1396           ->thread_num = __kmp_tid_from_gtid(global_tid);
1397     }
1398 
1399     /* OMPT state */
1400     this_thr->th.ompt_thread_info.state = omp_state_work_parallel;
1401     OMPT_CUR_TASK_INFO(this_thr)->frame.exit_frame = OMPT_GET_FRAME_ADDRESS(1);
1402   }
1403 #endif
1404 }
1405 
1406 /* most of the work for a fork */
1407 /* return true if we really went parallel, false if serialized */
1408 int __kmp_fork_call(ident_t *loc, int gtid,
1409                     enum fork_context_e call_context, // Intel, GNU, ...
1410                     kmp_int32 argc, microtask_t microtask, launch_t invoker,
1411 /* TODO: revert workaround for Intel(R) 64 tracker #96 */
1412 #if (KMP_ARCH_X86_64 || KMP_ARCH_ARM || KMP_ARCH_AARCH64) && KMP_OS_LINUX
1413                     va_list *ap
1414 #else
1415                     va_list ap
1416 #endif
1417                     ) {
1418   void **argv;
1419   int i;
1420   int master_tid;
1421   int master_this_cons;
1422   kmp_team_t *team;
1423   kmp_team_t *parent_team;
1424   kmp_info_t *master_th;
1425   kmp_root_t *root;
1426   int nthreads;
1427   int master_active;
1428   int master_set_numthreads;
1429   int level;
1430 #if OMP_40_ENABLED
1431   int active_level;
1432   int teams_level;
1433 #endif
1434 #if KMP_NESTED_HOT_TEAMS
1435   kmp_hot_team_ptr_t **p_hot_teams;
1436 #endif
1437   { // KMP_TIME_BLOCK
1438     KMP_TIME_DEVELOPER_PARTITIONED_BLOCK(KMP_fork_call);
1439     KMP_COUNT_VALUE(OMP_PARALLEL_args, argc);
1440 
1441     KA_TRACE(20, ("__kmp_fork_call: enter T#%d\n", gtid));
1442     if (__kmp_stkpadding > 0 && __kmp_root[gtid] != NULL) {
1443       /* Some systems prefer the stack for the root thread(s) to start with */
1444       /* some gap from the parent stack to prevent false sharing. */
1445       void *dummy = KMP_ALLOCA(__kmp_stkpadding);
1446       /* These 2 lines below are so this does not get optimized out */
1447       if (__kmp_stkpadding > KMP_MAX_STKPADDING)
1448         __kmp_stkpadding += (short)((kmp_int64)dummy);
1449     }
1450 
1451     /* initialize if needed */
1452     KMP_DEBUG_ASSERT(
1453         __kmp_init_serial); // AC: potentially unsafe, not in sync with shutdown
1454     if (!TCR_4(__kmp_init_parallel))
1455       __kmp_parallel_initialize();
1456 
1457     /* setup current data */
1458     master_th = __kmp_threads[gtid]; // AC: potentially unsafe, not in sync with
1459     // shutdown
1460     parent_team = master_th->th.th_team;
1461     master_tid = master_th->th.th_info.ds.ds_tid;
1462     master_this_cons = master_th->th.th_local.this_construct;
1463     root = master_th->th.th_root;
1464     master_active = root->r.r_active;
1465     master_set_numthreads = master_th->th.th_set_nproc;
1466 
1467 #if OMPT_SUPPORT
1468     ompt_data_t ompt_parallel_data;
1469     ompt_parallel_data.ptr = NULL;
1470     ompt_data_t *parent_task_data;
1471     omp_frame_t *ompt_frame;
1472     ompt_data_t *implicit_task_data;
1473     void *return_address = NULL;
1474 
1475     if (ompt_enabled.enabled) {
1476       __ompt_get_task_info_internal(0, NULL, &parent_task_data, &ompt_frame,
1477                                     NULL, NULL);
1478       return_address = OMPT_LOAD_RETURN_ADDRESS(gtid);
1479     }
1480 #endif
1481 
1482     // Nested level will be an index in the nested nthreads array
1483     level = parent_team->t.t_level;
1484     // used to launch non-serial teams even if nested is not allowed
1485     active_level = parent_team->t.t_active_level;
1486 #if OMP_40_ENABLED
1487     // needed to check nesting inside the teams
1488     teams_level = master_th->th.th_teams_level;
1489 #endif
1490 #if KMP_NESTED_HOT_TEAMS
1491     p_hot_teams = &master_th->th.th_hot_teams;
1492     if (*p_hot_teams == NULL && __kmp_hot_teams_max_level > 0) {
1493       *p_hot_teams = (kmp_hot_team_ptr_t *)__kmp_allocate(
1494           sizeof(kmp_hot_team_ptr_t) * __kmp_hot_teams_max_level);
1495       (*p_hot_teams)[0].hot_team = root->r.r_hot_team;
1496       // it is either actual or not needed (when active_level > 0)
1497       (*p_hot_teams)[0].hot_team_nth = 1;
1498     }
1499 #endif
1500 
1501 #if OMPT_SUPPORT
1502     if (ompt_enabled.enabled) {
1503       if (ompt_enabled.ompt_callback_parallel_begin) {
1504         int team_size = master_set_numthreads
1505                             ? master_set_numthreads
1506                             : get__nproc_2(parent_team, master_tid);
1507         ompt_callbacks.ompt_callback(ompt_callback_parallel_begin)(
1508             parent_task_data, ompt_frame, &ompt_parallel_data, team_size,
1509             OMPT_INVOKER(call_context), return_address);
1510       }
1511       master_th->th.ompt_thread_info.state = omp_state_overhead;
1512     }
1513 #endif
1514 
1515     master_th->th.th_ident = loc;
1516 
1517 #if OMP_40_ENABLED
1518     if (master_th->th.th_teams_microtask && ap &&
1519         microtask != (microtask_t)__kmp_teams_master && level == teams_level) {
1520       // AC: This is start of parallel that is nested inside teams construct.
1521       // The team is actual (hot), all workers are ready at the fork barrier.
1522       // No lock needed to initialize the team a bit, then free workers.
1523       parent_team->t.t_ident = loc;
1524       __kmp_alloc_argv_entries(argc, parent_team, TRUE);
1525       parent_team->t.t_argc = argc;
1526       argv = (void **)parent_team->t.t_argv;
1527       for (i = argc - 1; i >= 0; --i)
1528 /* TODO: revert workaround for Intel(R) 64 tracker #96 */
1529 #if (KMP_ARCH_X86_64 || KMP_ARCH_ARM || KMP_ARCH_AARCH64) && KMP_OS_LINUX
1530         *argv++ = va_arg(*ap, void *);
1531 #else
1532         *argv++ = va_arg(ap, void *);
1533 #endif
1534       // Increment our nested depth levels, but not increase the serialization
1535       if (parent_team == master_th->th.th_serial_team) {
1536         // AC: we are in serialized parallel
1537         __kmpc_serialized_parallel(loc, gtid);
1538         KMP_DEBUG_ASSERT(parent_team->t.t_serialized > 1);
1539         // AC: need this in order enquiry functions work
1540         // correctly, will restore at join time
1541         parent_team->t.t_serialized--;
1542 #if OMPT_SUPPORT
1543         void *dummy;
1544         void **exit_runtime_p;
1545 
1546         ompt_lw_taskteam_t lw_taskteam;
1547 
1548         if (ompt_enabled.enabled) {
1549           __ompt_lw_taskteam_init(&lw_taskteam, master_th, gtid,
1550                                   &ompt_parallel_data, return_address);
1551           exit_runtime_p = &(lw_taskteam.ompt_task_info.frame.exit_frame);
1552 
1553           __ompt_lw_taskteam_link(&lw_taskteam, master_th, 0);
1554           // don't use lw_taskteam after linking. content was swaped
1555 
1556           /* OMPT implicit task begin */
1557           implicit_task_data = OMPT_CUR_TASK_DATA(master_th);
1558           if (ompt_enabled.ompt_callback_implicit_task) {
1559             ompt_callbacks.ompt_callback(ompt_callback_implicit_task)(
1560                 ompt_scope_begin, OMPT_CUR_TEAM_DATA(master_th),
1561                 implicit_task_data, 1, __kmp_tid_from_gtid(gtid));
1562             OMPT_CUR_TASK_INFO(master_th)
1563                 ->thread_num = __kmp_tid_from_gtid(gtid);
1564           }
1565 
1566           /* OMPT state */
1567           master_th->th.ompt_thread_info.state = omp_state_work_parallel;
1568         } else {
1569           exit_runtime_p = &dummy;
1570         }
1571 #endif
1572 
1573         {
1574           KMP_TIME_PARTITIONED_BLOCK(OMP_parallel);
1575           KMP_SET_THREAD_STATE_BLOCK(IMPLICIT_TASK);
1576           __kmp_invoke_microtask(microtask, gtid, 0, argc, parent_team->t.t_argv
1577 #if OMPT_SUPPORT
1578                                  ,
1579                                  exit_runtime_p
1580 #endif
1581                                  );
1582         }
1583 
1584 #if OMPT_SUPPORT
1585         *exit_runtime_p = NULL;
1586         if (ompt_enabled.enabled) {
1587           OMPT_CUR_TASK_INFO(master_th)->frame.exit_frame = NULL;
1588           if (ompt_enabled.ompt_callback_implicit_task) {
1589             ompt_callbacks.ompt_callback(ompt_callback_implicit_task)(
1590                 ompt_scope_end, NULL, implicit_task_data, 1,
1591                 OMPT_CUR_TASK_INFO(master_th)->thread_num);
1592           }
1593           __ompt_lw_taskteam_unlink(master_th);
1594 
1595           if (ompt_enabled.ompt_callback_parallel_end) {
1596             ompt_callbacks.ompt_callback(ompt_callback_parallel_end)(
1597                 OMPT_CUR_TEAM_DATA(master_th), OMPT_CUR_TASK_DATA(master_th),
1598                 OMPT_INVOKER(call_context), return_address);
1599           }
1600           master_th->th.ompt_thread_info.state = omp_state_overhead;
1601         }
1602 #endif
1603         return TRUE;
1604       }
1605 
1606       parent_team->t.t_pkfn = microtask;
1607       parent_team->t.t_invoke = invoker;
1608       KMP_ATOMIC_INC(&root->r.r_in_parallel);
1609       parent_team->t.t_active_level++;
1610       parent_team->t.t_level++;
1611 
1612       /* Change number of threads in the team if requested */
1613       if (master_set_numthreads) { // The parallel has num_threads clause
1614         if (master_set_numthreads < master_th->th.th_teams_size.nth) {
1615           // AC: only can reduce number of threads dynamically, can't increase
1616           kmp_info_t **other_threads = parent_team->t.t_threads;
1617           parent_team->t.t_nproc = master_set_numthreads;
1618           for (i = 0; i < master_set_numthreads; ++i) {
1619             other_threads[i]->th.th_team_nproc = master_set_numthreads;
1620           }
1621           // Keep extra threads hot in the team for possible next parallels
1622         }
1623         master_th->th.th_set_nproc = 0;
1624       }
1625 
1626 #if USE_DEBUGGER
1627       if (__kmp_debugging) { // Let debugger override number of threads.
1628         int nth = __kmp_omp_num_threads(loc);
1629         if (nth > 0) { // 0 means debugger doesn't want to change num threads
1630           master_set_numthreads = nth;
1631         }
1632       }
1633 #endif
1634 
1635       KF_TRACE(10, ("__kmp_fork_call: before internal fork: root=%p, team=%p, "
1636                     "master_th=%p, gtid=%d\n",
1637                     root, parent_team, master_th, gtid));
1638       __kmp_internal_fork(loc, gtid, parent_team);
1639       KF_TRACE(10, ("__kmp_fork_call: after internal fork: root=%p, team=%p, "
1640                     "master_th=%p, gtid=%d\n",
1641                     root, parent_team, master_th, gtid));
1642 
1643       /* Invoke microtask for MASTER thread */
1644       KA_TRACE(20, ("__kmp_fork_call: T#%d(%d:0) invoke microtask = %p\n", gtid,
1645                     parent_team->t.t_id, parent_team->t.t_pkfn));
1646 
1647       if (!parent_team->t.t_invoke(gtid)) {
1648         KMP_ASSERT2(0, "cannot invoke microtask for MASTER thread");
1649       }
1650       KA_TRACE(20, ("__kmp_fork_call: T#%d(%d:0) done microtask = %p\n", gtid,
1651                     parent_team->t.t_id, parent_team->t.t_pkfn));
1652       KMP_MB(); /* Flush all pending memory write invalidates.  */
1653 
1654       KA_TRACE(20, ("__kmp_fork_call: parallel exit T#%d\n", gtid));
1655 
1656       return TRUE;
1657     } // Parallel closely nested in teams construct
1658 #endif /* OMP_40_ENABLED */
1659 
1660 #if KMP_DEBUG
1661     if (__kmp_tasking_mode != tskm_immediate_exec) {
1662       KMP_DEBUG_ASSERT(master_th->th.th_task_team ==
1663                        parent_team->t.t_task_team[master_th->th.th_task_state]);
1664     }
1665 #endif
1666 
1667     if (parent_team->t.t_active_level >=
1668         master_th->th.th_current_task->td_icvs.max_active_levels) {
1669       nthreads = 1;
1670     } else {
1671 #if OMP_40_ENABLED
1672       int enter_teams = ((ap == NULL && active_level == 0) ||
1673                          (ap && teams_level > 0 && teams_level == level));
1674 #endif
1675       nthreads =
1676           master_set_numthreads
1677               ? master_set_numthreads
1678               : get__nproc_2(
1679                     parent_team,
1680                     master_tid); // TODO: get nproc directly from current task
1681 
1682       // Check if we need to take forkjoin lock? (no need for serialized
1683       // parallel out of teams construct). This code moved here from
1684       // __kmp_reserve_threads() to speedup nested serialized parallels.
1685       if (nthreads > 1) {
1686         if ((!get__nested(master_th) && (root->r.r_in_parallel
1687 #if OMP_40_ENABLED
1688                                          && !enter_teams
1689 #endif /* OMP_40_ENABLED */
1690                                          )) ||
1691             (__kmp_library == library_serial)) {
1692           KC_TRACE(10, ("__kmp_fork_call: T#%d serializing team; requested %d"
1693                         " threads\n",
1694                         gtid, nthreads));
1695           nthreads = 1;
1696         }
1697       }
1698       if (nthreads > 1) {
1699         /* determine how many new threads we can use */
1700         __kmp_acquire_bootstrap_lock(&__kmp_forkjoin_lock);
1701         nthreads = __kmp_reserve_threads(
1702             root, parent_team, master_tid, nthreads
1703 #if OMP_40_ENABLED
1704             /* AC: If we execute teams from parallel region (on host), then
1705                teams should be created but each can only have 1 thread if
1706                nesting is disabled. If teams called from serial region, then
1707                teams and their threads should be created regardless of the
1708                nesting setting. */
1709             ,
1710             enter_teams
1711 #endif /* OMP_40_ENABLED */
1712             );
1713         if (nthreads == 1) {
1714           // Free lock for single thread execution here; for multi-thread
1715           // execution it will be freed later after team of threads created
1716           // and initialized
1717           __kmp_release_bootstrap_lock(&__kmp_forkjoin_lock);
1718         }
1719       }
1720     }
1721     KMP_DEBUG_ASSERT(nthreads > 0);
1722 
1723     // If we temporarily changed the set number of threads then restore it now
1724     master_th->th.th_set_nproc = 0;
1725 
1726     /* create a serialized parallel region? */
1727     if (nthreads == 1) {
1728 /* josh todo: hypothetical question: what do we do for OS X*? */
1729 #if KMP_OS_LINUX &&                                                            \
1730     (KMP_ARCH_X86 || KMP_ARCH_X86_64 || KMP_ARCH_ARM || KMP_ARCH_AARCH64)
1731       void *args[argc];
1732 #else
1733       void **args = (void **)KMP_ALLOCA(argc * sizeof(void *));
1734 #endif /* KMP_OS_LINUX && ( KMP_ARCH_X86 || KMP_ARCH_X86_64 || KMP_ARCH_ARM || \
1735           KMP_ARCH_AARCH64) */
1736 
1737       KA_TRACE(20,
1738                ("__kmp_fork_call: T#%d serializing parallel region\n", gtid));
1739 
1740       __kmpc_serialized_parallel(loc, gtid);
1741 
1742       if (call_context == fork_context_intel) {
1743         /* TODO this sucks, use the compiler itself to pass args! :) */
1744         master_th->th.th_serial_team->t.t_ident = loc;
1745 #if OMP_40_ENABLED
1746         if (!ap) {
1747           // revert change made in __kmpc_serialized_parallel()
1748           master_th->th.th_serial_team->t.t_level--;
1749 // Get args from parent team for teams construct
1750 
1751 #if OMPT_SUPPORT
1752           void *dummy;
1753           void **exit_runtime_p;
1754           ompt_task_info_t *task_info;
1755 
1756           ompt_lw_taskteam_t lw_taskteam;
1757 
1758           if (ompt_enabled.enabled) {
1759             __ompt_lw_taskteam_init(&lw_taskteam, master_th, gtid,
1760                                     &ompt_parallel_data, return_address);
1761 
1762             __ompt_lw_taskteam_link(&lw_taskteam, master_th, 0);
1763             // don't use lw_taskteam after linking. content was swaped
1764 
1765             task_info = OMPT_CUR_TASK_INFO(master_th);
1766             exit_runtime_p = &(task_info->frame.exit_frame);
1767             if (ompt_enabled.ompt_callback_implicit_task) {
1768               ompt_callbacks.ompt_callback(ompt_callback_implicit_task)(
1769                   ompt_scope_begin, OMPT_CUR_TEAM_DATA(master_th),
1770                   &(task_info->task_data), 1, __kmp_tid_from_gtid(gtid));
1771               OMPT_CUR_TASK_INFO(master_th)
1772                   ->thread_num = __kmp_tid_from_gtid(gtid);
1773             }
1774 
1775             /* OMPT state */
1776             master_th->th.ompt_thread_info.state = omp_state_work_parallel;
1777           } else {
1778             exit_runtime_p = &dummy;
1779           }
1780 #endif
1781 
1782           {
1783             KMP_TIME_PARTITIONED_BLOCK(OMP_parallel);
1784             KMP_SET_THREAD_STATE_BLOCK(IMPLICIT_TASK);
1785             __kmp_invoke_microtask(microtask, gtid, 0, argc,
1786                                    parent_team->t.t_argv
1787 #if OMPT_SUPPORT
1788                                    ,
1789                                    exit_runtime_p
1790 #endif
1791                                    );
1792           }
1793 
1794 #if OMPT_SUPPORT
1795           if (ompt_enabled.enabled) {
1796             exit_runtime_p = NULL;
1797             if (ompt_enabled.ompt_callback_implicit_task) {
1798               ompt_callbacks.ompt_callback(ompt_callback_implicit_task)(
1799                   ompt_scope_end, NULL, &(task_info->task_data), 1,
1800                   OMPT_CUR_TASK_INFO(master_th)->thread_num);
1801             }
1802 
1803             __ompt_lw_taskteam_unlink(master_th);
1804             if (ompt_enabled.ompt_callback_parallel_end) {
1805               ompt_callbacks.ompt_callback(ompt_callback_parallel_end)(
1806                   OMPT_CUR_TEAM_DATA(master_th), parent_task_data,
1807                   OMPT_INVOKER(call_context), return_address);
1808             }
1809             master_th->th.ompt_thread_info.state = omp_state_overhead;
1810           }
1811 #endif
1812         } else if (microtask == (microtask_t)__kmp_teams_master) {
1813           KMP_DEBUG_ASSERT(master_th->th.th_team ==
1814                            master_th->th.th_serial_team);
1815           team = master_th->th.th_team;
1816           // team->t.t_pkfn = microtask;
1817           team->t.t_invoke = invoker;
1818           __kmp_alloc_argv_entries(argc, team, TRUE);
1819           team->t.t_argc = argc;
1820           argv = (void **)team->t.t_argv;
1821           if (ap) {
1822             for (i = argc - 1; i >= 0; --i)
1823 // TODO: revert workaround for Intel(R) 64 tracker #96
1824 #if (KMP_ARCH_X86_64 || KMP_ARCH_ARM || KMP_ARCH_AARCH64) && KMP_OS_LINUX
1825               *argv++ = va_arg(*ap, void *);
1826 #else
1827               *argv++ = va_arg(ap, void *);
1828 #endif
1829           } else {
1830             for (i = 0; i < argc; ++i)
1831               // Get args from parent team for teams construct
1832               argv[i] = parent_team->t.t_argv[i];
1833           }
1834           // AC: revert change made in __kmpc_serialized_parallel()
1835           //     because initial code in teams should have level=0
1836           team->t.t_level--;
1837           // AC: call special invoker for outer "parallel" of teams construct
1838           invoker(gtid);
1839         } else {
1840 #endif /* OMP_40_ENABLED */
1841           argv = args;
1842           for (i = argc - 1; i >= 0; --i)
1843 // TODO: revert workaround for Intel(R) 64 tracker #96
1844 #if (KMP_ARCH_X86_64 || KMP_ARCH_ARM || KMP_ARCH_AARCH64) && KMP_OS_LINUX
1845             *argv++ = va_arg(*ap, void *);
1846 #else
1847           *argv++ = va_arg(ap, void *);
1848 #endif
1849           KMP_MB();
1850 
1851 #if OMPT_SUPPORT
1852           void *dummy;
1853           void **exit_runtime_p;
1854           ompt_task_info_t *task_info;
1855 
1856           ompt_lw_taskteam_t lw_taskteam;
1857 
1858           if (ompt_enabled.enabled) {
1859             __ompt_lw_taskteam_init(&lw_taskteam, master_th, gtid,
1860                                     &ompt_parallel_data, return_address);
1861             __ompt_lw_taskteam_link(&lw_taskteam, master_th, 0);
1862             // don't use lw_taskteam after linking. content was swaped
1863             task_info = OMPT_CUR_TASK_INFO(master_th);
1864             exit_runtime_p = &(task_info->frame.exit_frame);
1865 
1866             /* OMPT implicit task begin */
1867             implicit_task_data = OMPT_CUR_TASK_DATA(master_th);
1868             if (ompt_enabled.ompt_callback_implicit_task) {
1869               ompt_callbacks.ompt_callback(ompt_callback_implicit_task)(
1870                   ompt_scope_begin, OMPT_CUR_TEAM_DATA(master_th),
1871                   implicit_task_data, 1, __kmp_tid_from_gtid(gtid));
1872               OMPT_CUR_TASK_INFO(master_th)
1873                   ->thread_num = __kmp_tid_from_gtid(gtid);
1874             }
1875 
1876             /* OMPT state */
1877             master_th->th.ompt_thread_info.state = omp_state_work_parallel;
1878           } else {
1879             exit_runtime_p = &dummy;
1880           }
1881 #endif
1882 
1883           {
1884             KMP_TIME_PARTITIONED_BLOCK(OMP_parallel);
1885             KMP_SET_THREAD_STATE_BLOCK(IMPLICIT_TASK);
1886             __kmp_invoke_microtask(microtask, gtid, 0, argc, args
1887 #if OMPT_SUPPORT
1888                                    ,
1889                                    exit_runtime_p
1890 #endif
1891                                    );
1892           }
1893 
1894 #if OMPT_SUPPORT
1895           if (ompt_enabled.enabled) {
1896             *exit_runtime_p = NULL;
1897             if (ompt_enabled.ompt_callback_implicit_task) {
1898               ompt_callbacks.ompt_callback(ompt_callback_implicit_task)(
1899                   ompt_scope_end, NULL, &(task_info->task_data), 1,
1900                   OMPT_CUR_TASK_INFO(master_th)->thread_num);
1901             }
1902 
1903             ompt_parallel_data = *OMPT_CUR_TEAM_DATA(master_th);
1904             __ompt_lw_taskteam_unlink(master_th);
1905             if (ompt_enabled.ompt_callback_parallel_end) {
1906               ompt_callbacks.ompt_callback(ompt_callback_parallel_end)(
1907                   &ompt_parallel_data, parent_task_data,
1908                   OMPT_INVOKER(call_context), return_address);
1909             }
1910             master_th->th.ompt_thread_info.state = omp_state_overhead;
1911           }
1912 #endif
1913 #if OMP_40_ENABLED
1914         }
1915 #endif /* OMP_40_ENABLED */
1916       } else if (call_context == fork_context_gnu) {
1917 #if OMPT_SUPPORT
1918         ompt_lw_taskteam_t lwt;
1919         __ompt_lw_taskteam_init(&lwt, master_th, gtid, &ompt_parallel_data,
1920                                 return_address);
1921 
1922         lwt.ompt_task_info.frame.exit_frame = NULL;
1923         __ompt_lw_taskteam_link(&lwt, master_th, 1);
1924 // don't use lw_taskteam after linking. content was swaped
1925 #endif
1926 
1927         // we were called from GNU native code
1928         KA_TRACE(20, ("__kmp_fork_call: T#%d serial exit\n", gtid));
1929         return FALSE;
1930       } else {
1931         KMP_ASSERT2(call_context < fork_context_last,
1932                     "__kmp_fork_call: unknown fork_context parameter");
1933       }
1934 
1935       KA_TRACE(20, ("__kmp_fork_call: T#%d serial exit\n", gtid));
1936       KMP_MB();
1937       return FALSE;
1938     }
1939 
1940     // GEH: only modify the executing flag in the case when not serialized
1941     //      serialized case is handled in kmpc_serialized_parallel
1942     KF_TRACE(10, ("__kmp_fork_call: parent_team_aclevel=%d, master_th=%p, "
1943                   "curtask=%p, curtask_max_aclevel=%d\n",
1944                   parent_team->t.t_active_level, master_th,
1945                   master_th->th.th_current_task,
1946                   master_th->th.th_current_task->td_icvs.max_active_levels));
1947     // TODO: GEH - cannot do this assertion because root thread not set up as
1948     // executing
1949     // KMP_ASSERT( master_th->th.th_current_task->td_flags.executing == 1 );
1950     master_th->th.th_current_task->td_flags.executing = 0;
1951 
1952 #if OMP_40_ENABLED
1953     if (!master_th->th.th_teams_microtask || level > teams_level)
1954 #endif /* OMP_40_ENABLED */
1955     {
1956       /* Increment our nested depth level */
1957       KMP_ATOMIC_INC(&root->r.r_in_parallel);
1958     }
1959 
1960     // See if we need to make a copy of the ICVs.
1961     int nthreads_icv = master_th->th.th_current_task->td_icvs.nproc;
1962     if ((level + 1 < __kmp_nested_nth.used) &&
1963         (__kmp_nested_nth.nth[level + 1] != nthreads_icv)) {
1964       nthreads_icv = __kmp_nested_nth.nth[level + 1];
1965     } else {
1966       nthreads_icv = 0; // don't update
1967     }
1968 
1969 #if OMP_40_ENABLED
1970     // Figure out the proc_bind_policy for the new team.
1971     kmp_proc_bind_t proc_bind = master_th->th.th_set_proc_bind;
1972     kmp_proc_bind_t proc_bind_icv =
1973         proc_bind_default; // proc_bind_default means don't update
1974     if (master_th->th.th_current_task->td_icvs.proc_bind == proc_bind_false) {
1975       proc_bind = proc_bind_false;
1976     } else {
1977       if (proc_bind == proc_bind_default) {
1978         // No proc_bind clause specified; use current proc-bind-var for this
1979         // parallel region
1980         proc_bind = master_th->th.th_current_task->td_icvs.proc_bind;
1981       }
1982       /* else: The proc_bind policy was specified explicitly on parallel clause.
1983          This overrides proc-bind-var for this parallel region, but does not
1984          change proc-bind-var. */
1985       // Figure the value of proc-bind-var for the child threads.
1986       if ((level + 1 < __kmp_nested_proc_bind.used) &&
1987           (__kmp_nested_proc_bind.bind_types[level + 1] !=
1988            master_th->th.th_current_task->td_icvs.proc_bind)) {
1989         proc_bind_icv = __kmp_nested_proc_bind.bind_types[level + 1];
1990       }
1991     }
1992 
1993     // Reset for next parallel region
1994     master_th->th.th_set_proc_bind = proc_bind_default;
1995 #endif /* OMP_40_ENABLED */
1996 
1997     if ((nthreads_icv > 0)
1998 #if OMP_40_ENABLED
1999         || (proc_bind_icv != proc_bind_default)
2000 #endif /* OMP_40_ENABLED */
2001             ) {
2002       kmp_internal_control_t new_icvs;
2003       copy_icvs(&new_icvs, &master_th->th.th_current_task->td_icvs);
2004       new_icvs.next = NULL;
2005       if (nthreads_icv > 0) {
2006         new_icvs.nproc = nthreads_icv;
2007       }
2008 
2009 #if OMP_40_ENABLED
2010       if (proc_bind_icv != proc_bind_default) {
2011         new_icvs.proc_bind = proc_bind_icv;
2012       }
2013 #endif /* OMP_40_ENABLED */
2014 
2015       /* allocate a new parallel team */
2016       KF_TRACE(10, ("__kmp_fork_call: before __kmp_allocate_team\n"));
2017       team = __kmp_allocate_team(root, nthreads, nthreads,
2018 #if OMPT_SUPPORT
2019                                  ompt_parallel_data,
2020 #endif
2021 #if OMP_40_ENABLED
2022                                  proc_bind,
2023 #endif
2024                                  &new_icvs, argc USE_NESTED_HOT_ARG(master_th));
2025     } else {
2026       /* allocate a new parallel team */
2027       KF_TRACE(10, ("__kmp_fork_call: before __kmp_allocate_team\n"));
2028       team = __kmp_allocate_team(root, nthreads, nthreads,
2029 #if OMPT_SUPPORT
2030                                  ompt_parallel_data,
2031 #endif
2032 #if OMP_40_ENABLED
2033                                  proc_bind,
2034 #endif
2035                                  &master_th->th.th_current_task->td_icvs,
2036                                  argc USE_NESTED_HOT_ARG(master_th));
2037     }
2038     KF_TRACE(
2039         10, ("__kmp_fork_call: after __kmp_allocate_team - team = %p\n", team));
2040 
2041     /* setup the new team */
2042     KMP_CHECK_UPDATE(team->t.t_master_tid, master_tid);
2043     KMP_CHECK_UPDATE(team->t.t_master_this_cons, master_this_cons);
2044     KMP_CHECK_UPDATE(team->t.t_ident, loc);
2045     KMP_CHECK_UPDATE(team->t.t_parent, parent_team);
2046     KMP_CHECK_UPDATE_SYNC(team->t.t_pkfn, microtask);
2047 #if OMPT_SUPPORT
2048     KMP_CHECK_UPDATE_SYNC(team->t.ompt_team_info.master_return_address,
2049                           return_address);
2050 #endif
2051     KMP_CHECK_UPDATE(team->t.t_invoke, invoker); // TODO move to root, maybe
2052 // TODO: parent_team->t.t_level == INT_MAX ???
2053 #if OMP_40_ENABLED
2054     if (!master_th->th.th_teams_microtask || level > teams_level) {
2055 #endif /* OMP_40_ENABLED */
2056       int new_level = parent_team->t.t_level + 1;
2057       KMP_CHECK_UPDATE(team->t.t_level, new_level);
2058       new_level = parent_team->t.t_active_level + 1;
2059       KMP_CHECK_UPDATE(team->t.t_active_level, new_level);
2060 #if OMP_40_ENABLED
2061     } else {
2062       // AC: Do not increase parallel level at start of the teams construct
2063       int new_level = parent_team->t.t_level;
2064       KMP_CHECK_UPDATE(team->t.t_level, new_level);
2065       new_level = parent_team->t.t_active_level;
2066       KMP_CHECK_UPDATE(team->t.t_active_level, new_level);
2067     }
2068 #endif /* OMP_40_ENABLED */
2069     kmp_r_sched_t new_sched = get__sched_2(parent_team, master_tid);
2070     // set master's schedule as new run-time schedule
2071     KMP_CHECK_UPDATE(team->t.t_sched.sched, new_sched.sched);
2072 
2073 #if OMP_40_ENABLED
2074     KMP_CHECK_UPDATE(team->t.t_cancel_request, cancel_noreq);
2075 #endif
2076 
2077     // Update the floating point rounding in the team if required.
2078     propagateFPControl(team);
2079 
2080     if (__kmp_tasking_mode != tskm_immediate_exec) {
2081       // Set master's task team to team's task team. Unless this is hot team, it
2082       // should be NULL.
2083       KMP_DEBUG_ASSERT(master_th->th.th_task_team ==
2084                        parent_team->t.t_task_team[master_th->th.th_task_state]);
2085       KA_TRACE(20, ("__kmp_fork_call: Master T#%d pushing task_team %p / team "
2086                     "%p, new task_team %p / team %p\n",
2087                     __kmp_gtid_from_thread(master_th),
2088                     master_th->th.th_task_team, parent_team,
2089                     team->t.t_task_team[master_th->th.th_task_state], team));
2090 
2091       if (active_level || master_th->th.th_task_team) {
2092         // Take a memo of master's task_state
2093         KMP_DEBUG_ASSERT(master_th->th.th_task_state_memo_stack);
2094         if (master_th->th.th_task_state_top >=
2095             master_th->th.th_task_state_stack_sz) { // increase size
2096           kmp_uint32 new_size = 2 * master_th->th.th_task_state_stack_sz;
2097           kmp_uint8 *old_stack, *new_stack;
2098           kmp_uint32 i;
2099           new_stack = (kmp_uint8 *)__kmp_allocate(new_size);
2100           for (i = 0; i < master_th->th.th_task_state_stack_sz; ++i) {
2101             new_stack[i] = master_th->th.th_task_state_memo_stack[i];
2102           }
2103           for (i = master_th->th.th_task_state_stack_sz; i < new_size;
2104                ++i) { // zero-init rest of stack
2105             new_stack[i] = 0;
2106           }
2107           old_stack = master_th->th.th_task_state_memo_stack;
2108           master_th->th.th_task_state_memo_stack = new_stack;
2109           master_th->th.th_task_state_stack_sz = new_size;
2110           __kmp_free(old_stack);
2111         }
2112         // Store master's task_state on stack
2113         master_th->th
2114             .th_task_state_memo_stack[master_th->th.th_task_state_top] =
2115             master_th->th.th_task_state;
2116         master_th->th.th_task_state_top++;
2117 #if KMP_NESTED_HOT_TEAMS
2118         if (master_th->th.th_hot_teams &&
2119             team == master_th->th.th_hot_teams[active_level].hot_team) {
2120           // Restore master's nested state if nested hot team
2121           master_th->th.th_task_state =
2122               master_th->th
2123                   .th_task_state_memo_stack[master_th->th.th_task_state_top];
2124         } else {
2125 #endif
2126           master_th->th.th_task_state = 0;
2127 #if KMP_NESTED_HOT_TEAMS
2128         }
2129 #endif
2130       }
2131 #if !KMP_NESTED_HOT_TEAMS
2132       KMP_DEBUG_ASSERT((master_th->th.th_task_team == NULL) ||
2133                        (team == root->r.r_hot_team));
2134 #endif
2135     }
2136 
2137     KA_TRACE(
2138         20,
2139         ("__kmp_fork_call: T#%d(%d:%d)->(%d:0) created a team of %d threads\n",
2140          gtid, parent_team->t.t_id, team->t.t_master_tid, team->t.t_id,
2141          team->t.t_nproc));
2142     KMP_DEBUG_ASSERT(team != root->r.r_hot_team ||
2143                      (team->t.t_master_tid == 0 &&
2144                       (team->t.t_parent == root->r.r_root_team ||
2145                        team->t.t_parent->t.t_serialized)));
2146     KMP_MB();
2147 
2148     /* now, setup the arguments */
2149     argv = (void **)team->t.t_argv;
2150 #if OMP_40_ENABLED
2151     if (ap) {
2152 #endif /* OMP_40_ENABLED */
2153       for (i = argc - 1; i >= 0; --i) {
2154 // TODO: revert workaround for Intel(R) 64 tracker #96
2155 #if (KMP_ARCH_X86_64 || KMP_ARCH_ARM || KMP_ARCH_AARCH64) && KMP_OS_LINUX
2156         void *new_argv = va_arg(*ap, void *);
2157 #else
2158       void *new_argv = va_arg(ap, void *);
2159 #endif
2160         KMP_CHECK_UPDATE(*argv, new_argv);
2161         argv++;
2162       }
2163 #if OMP_40_ENABLED
2164     } else {
2165       for (i = 0; i < argc; ++i) {
2166         // Get args from parent team for teams construct
2167         KMP_CHECK_UPDATE(argv[i], team->t.t_parent->t.t_argv[i]);
2168       }
2169     }
2170 #endif /* OMP_40_ENABLED */
2171 
2172     /* now actually fork the threads */
2173     KMP_CHECK_UPDATE(team->t.t_master_active, master_active);
2174     if (!root->r.r_active) // Only do assignment if it prevents cache ping-pong
2175       root->r.r_active = TRUE;
2176 
2177     __kmp_fork_team_threads(root, team, master_th, gtid);
2178     __kmp_setup_icv_copy(team, nthreads,
2179                          &master_th->th.th_current_task->td_icvs, loc);
2180 
2181 #if OMPT_SUPPORT
2182     master_th->th.ompt_thread_info.state = omp_state_work_parallel;
2183 #endif
2184 
2185     __kmp_release_bootstrap_lock(&__kmp_forkjoin_lock);
2186 
2187 #if USE_ITT_BUILD
2188     if (team->t.t_active_level == 1 // only report frames at level 1
2189 #if OMP_40_ENABLED
2190         && !master_th->th.th_teams_microtask // not in teams construct
2191 #endif /* OMP_40_ENABLED */
2192         ) {
2193 #if USE_ITT_NOTIFY
2194       if ((__itt_frame_submit_v3_ptr || KMP_ITT_DEBUG) &&
2195           (__kmp_forkjoin_frames_mode == 3 ||
2196            __kmp_forkjoin_frames_mode == 1)) {
2197         kmp_uint64 tmp_time = 0;
2198         if (__itt_get_timestamp_ptr)
2199           tmp_time = __itt_get_timestamp();
2200         // Internal fork - report frame begin
2201         master_th->th.th_frame_time = tmp_time;
2202         if (__kmp_forkjoin_frames_mode == 3)
2203           team->t.t_region_time = tmp_time;
2204       } else
2205 // only one notification scheme (either "submit" or "forking/joined", not both)
2206 #endif /* USE_ITT_NOTIFY */
2207           if ((__itt_frame_begin_v3_ptr || KMP_ITT_DEBUG) &&
2208               __kmp_forkjoin_frames && !__kmp_forkjoin_frames_mode) {
2209         // Mark start of "parallel" region for Intel(R) VTune(TM) analyzer.
2210         __kmp_itt_region_forking(gtid, team->t.t_nproc, 0);
2211       }
2212     }
2213 #endif /* USE_ITT_BUILD */
2214 
2215     /* now go on and do the work */
2216     KMP_DEBUG_ASSERT(team == __kmp_threads[gtid]->th.th_team);
2217     KMP_MB();
2218     KF_TRACE(10,
2219              ("__kmp_internal_fork : root=%p, team=%p, master_th=%p, gtid=%d\n",
2220               root, team, master_th, gtid));
2221 
2222 #if USE_ITT_BUILD
2223     if (__itt_stack_caller_create_ptr) {
2224       team->t.t_stack_id =
2225           __kmp_itt_stack_caller_create(); // create new stack stitching id
2226       // before entering fork barrier
2227     }
2228 #endif /* USE_ITT_BUILD */
2229 
2230 #if OMP_40_ENABLED
2231     // AC: skip __kmp_internal_fork at teams construct, let only master
2232     // threads execute
2233     if (ap)
2234 #endif /* OMP_40_ENABLED */
2235     {
2236       __kmp_internal_fork(loc, gtid, team);
2237       KF_TRACE(10, ("__kmp_internal_fork : after : root=%p, team=%p, "
2238                     "master_th=%p, gtid=%d\n",
2239                     root, team, master_th, gtid));
2240     }
2241 
2242     if (call_context == fork_context_gnu) {
2243       KA_TRACE(20, ("__kmp_fork_call: parallel exit T#%d\n", gtid));
2244       return TRUE;
2245     }
2246 
2247     /* Invoke microtask for MASTER thread */
2248     KA_TRACE(20, ("__kmp_fork_call: T#%d(%d:0) invoke microtask = %p\n", gtid,
2249                   team->t.t_id, team->t.t_pkfn));
2250   } // END of timer KMP_fork_call block
2251 
2252   if (!team->t.t_invoke(gtid)) {
2253     KMP_ASSERT2(0, "cannot invoke microtask for MASTER thread");
2254   }
2255   KA_TRACE(20, ("__kmp_fork_call: T#%d(%d:0) done microtask = %p\n", gtid,
2256                 team->t.t_id, team->t.t_pkfn));
2257   KMP_MB(); /* Flush all pending memory write invalidates.  */
2258 
2259   KA_TRACE(20, ("__kmp_fork_call: parallel exit T#%d\n", gtid));
2260 
2261 #if OMPT_SUPPORT
2262   if (ompt_enabled.enabled) {
2263     master_th->th.ompt_thread_info.state = omp_state_overhead;
2264   }
2265 #endif
2266 
2267   return TRUE;
2268 }
2269 
2270 #if OMPT_SUPPORT
2271 static inline void __kmp_join_restore_state(kmp_info_t *thread,
2272                                             kmp_team_t *team) {
2273   // restore state outside the region
2274   thread->th.ompt_thread_info.state =
2275       ((team->t.t_serialized) ? omp_state_work_serial
2276                               : omp_state_work_parallel);
2277 }
2278 
2279 static inline void __kmp_join_ompt(int gtid, kmp_info_t *thread,
2280                                    kmp_team_t *team, ompt_data_t *parallel_data,
2281                                    fork_context_e fork_context, void *codeptr) {
2282   ompt_task_info_t *task_info = __ompt_get_task_info_object(0);
2283   if (ompt_enabled.ompt_callback_parallel_end) {
2284     ompt_callbacks.ompt_callback(ompt_callback_parallel_end)(
2285         parallel_data, &(task_info->task_data), OMPT_INVOKER(fork_context),
2286         codeptr);
2287   }
2288 
2289   task_info->frame.enter_frame = NULL;
2290   __kmp_join_restore_state(thread, team);
2291 }
2292 #endif
2293 
2294 void __kmp_join_call(ident_t *loc, int gtid
2295 #if OMPT_SUPPORT
2296                      ,
2297                      enum fork_context_e fork_context
2298 #endif
2299 #if OMP_40_ENABLED
2300                      ,
2301                      int exit_teams
2302 #endif /* OMP_40_ENABLED */
2303                      ) {
2304   KMP_TIME_DEVELOPER_PARTITIONED_BLOCK(KMP_join_call);
2305   kmp_team_t *team;
2306   kmp_team_t *parent_team;
2307   kmp_info_t *master_th;
2308   kmp_root_t *root;
2309   int master_active;
2310   int i;
2311 
2312   KA_TRACE(20, ("__kmp_join_call: enter T#%d\n", gtid));
2313 
2314   /* setup current data */
2315   master_th = __kmp_threads[gtid];
2316   root = master_th->th.th_root;
2317   team = master_th->th.th_team;
2318   parent_team = team->t.t_parent;
2319 
2320   master_th->th.th_ident = loc;
2321 
2322 #if OMPT_SUPPORT
2323   if (ompt_enabled.enabled) {
2324     master_th->th.ompt_thread_info.state = omp_state_overhead;
2325   }
2326 #endif
2327 
2328 #if KMP_DEBUG
2329   if (__kmp_tasking_mode != tskm_immediate_exec && !exit_teams) {
2330     KA_TRACE(20, ("__kmp_join_call: T#%d, old team = %p old task_team = %p, "
2331                   "th_task_team = %p\n",
2332                   __kmp_gtid_from_thread(master_th), team,
2333                   team->t.t_task_team[master_th->th.th_task_state],
2334                   master_th->th.th_task_team));
2335     KMP_DEBUG_ASSERT(master_th->th.th_task_team ==
2336                      team->t.t_task_team[master_th->th.th_task_state]);
2337   }
2338 #endif
2339 
2340   if (team->t.t_serialized) {
2341 #if OMP_40_ENABLED
2342     if (master_th->th.th_teams_microtask) {
2343       // We are in teams construct
2344       int level = team->t.t_level;
2345       int tlevel = master_th->th.th_teams_level;
2346       if (level == tlevel) {
2347         // AC: we haven't incremented it earlier at start of teams construct,
2348         //     so do it here - at the end of teams construct
2349         team->t.t_level++;
2350       } else if (level == tlevel + 1) {
2351         // AC: we are exiting parallel inside teams, need to increment
2352         // serialization in order to restore it in the next call to
2353         // __kmpc_end_serialized_parallel
2354         team->t.t_serialized++;
2355       }
2356     }
2357 #endif /* OMP_40_ENABLED */
2358     __kmpc_end_serialized_parallel(loc, gtid);
2359 
2360 #if OMPT_SUPPORT
2361     if (ompt_enabled.enabled) {
2362       __kmp_join_restore_state(master_th, parent_team);
2363     }
2364 #endif
2365 
2366     return;
2367   }
2368 
2369   master_active = team->t.t_master_active;
2370 
2371 #if OMP_40_ENABLED
2372   if (!exit_teams)
2373 #endif /* OMP_40_ENABLED */
2374   {
2375     // AC: No barrier for internal teams at exit from teams construct.
2376     //     But there is barrier for external team (league).
2377     __kmp_internal_join(loc, gtid, team);
2378   }
2379 #if OMP_40_ENABLED
2380   else {
2381     master_th->th.th_task_state =
2382         0; // AC: no tasking in teams (out of any parallel)
2383   }
2384 #endif /* OMP_40_ENABLED */
2385 
2386   KMP_MB();
2387 
2388 #if OMPT_SUPPORT
2389   ompt_data_t *parallel_data = &(team->t.ompt_team_info.parallel_data);
2390   void *codeptr = team->t.ompt_team_info.master_return_address;
2391 #endif
2392 
2393 #if USE_ITT_BUILD
2394   if (__itt_stack_caller_create_ptr) {
2395     __kmp_itt_stack_caller_destroy(
2396         (__itt_caller)team->t
2397             .t_stack_id); // destroy the stack stitching id after join barrier
2398   }
2399 
2400   // Mark end of "parallel" region for Intel(R) VTune(TM) analyzer.
2401   if (team->t.t_active_level == 1
2402 #if OMP_40_ENABLED
2403       && !master_th->th.th_teams_microtask /* not in teams construct */
2404 #endif /* OMP_40_ENABLED */
2405       ) {
2406     master_th->th.th_ident = loc;
2407     // only one notification scheme (either "submit" or "forking/joined", not
2408     // both)
2409     if ((__itt_frame_submit_v3_ptr || KMP_ITT_DEBUG) &&
2410         __kmp_forkjoin_frames_mode == 3)
2411       __kmp_itt_frame_submit(gtid, team->t.t_region_time,
2412                              master_th->th.th_frame_time, 0, loc,
2413                              master_th->th.th_team_nproc, 1);
2414     else if ((__itt_frame_end_v3_ptr || KMP_ITT_DEBUG) &&
2415              !__kmp_forkjoin_frames_mode && __kmp_forkjoin_frames)
2416       __kmp_itt_region_joined(gtid);
2417   } // active_level == 1
2418 #endif /* USE_ITT_BUILD */
2419 
2420 #if OMP_40_ENABLED
2421   if (master_th->th.th_teams_microtask && !exit_teams &&
2422       team->t.t_pkfn != (microtask_t)__kmp_teams_master &&
2423       team->t.t_level == master_th->th.th_teams_level + 1) {
2424     // AC: We need to leave the team structure intact at the end of parallel
2425     // inside the teams construct, so that at the next parallel same (hot) team
2426     // works, only adjust nesting levels
2427 
2428     /* Decrement our nested depth level */
2429     team->t.t_level--;
2430     team->t.t_active_level--;
2431     KMP_ATOMIC_DEC(&root->r.r_in_parallel);
2432 
2433     /* Restore number of threads in the team if needed */
2434     if (master_th->th.th_team_nproc < master_th->th.th_teams_size.nth) {
2435       int old_num = master_th->th.th_team_nproc;
2436       int new_num = master_th->th.th_teams_size.nth;
2437       kmp_info_t **other_threads = team->t.t_threads;
2438       team->t.t_nproc = new_num;
2439       for (i = 0; i < old_num; ++i) {
2440         other_threads[i]->th.th_team_nproc = new_num;
2441       }
2442       // Adjust states of non-used threads of the team
2443       for (i = old_num; i < new_num; ++i) {
2444         // Re-initialize thread's barrier data.
2445         int b;
2446         kmp_balign_t *balign = other_threads[i]->th.th_bar;
2447         for (b = 0; b < bs_last_barrier; ++b) {
2448           balign[b].bb.b_arrived = team->t.t_bar[b].b_arrived;
2449           KMP_DEBUG_ASSERT(balign[b].bb.wait_flag != KMP_BARRIER_PARENT_FLAG);
2450 #if USE_DEBUGGER
2451           balign[b].bb.b_worker_arrived = team->t.t_bar[b].b_team_arrived;
2452 #endif
2453         }
2454         if (__kmp_tasking_mode != tskm_immediate_exec) {
2455           // Synchronize thread's task state
2456           other_threads[i]->th.th_task_state = master_th->th.th_task_state;
2457         }
2458       }
2459     }
2460 
2461 #if OMPT_SUPPORT
2462     if (ompt_enabled.enabled) {
2463       __kmp_join_ompt(gtid, master_th, parent_team, parallel_data, fork_context,
2464                       codeptr);
2465     }
2466 #endif
2467 
2468     return;
2469   }
2470 #endif /* OMP_40_ENABLED */
2471 
2472   /* do cleanup and restore the parent team */
2473   master_th->th.th_info.ds.ds_tid = team->t.t_master_tid;
2474   master_th->th.th_local.this_construct = team->t.t_master_this_cons;
2475 
2476   master_th->th.th_dispatch = &parent_team->t.t_dispatch[team->t.t_master_tid];
2477 
2478   /* jc: The following lock has instructions with REL and ACQ semantics,
2479      separating the parallel user code called in this parallel region
2480      from the serial user code called after this function returns. */
2481   __kmp_acquire_bootstrap_lock(&__kmp_forkjoin_lock);
2482 
2483 #if OMP_40_ENABLED
2484   if (!master_th->th.th_teams_microtask ||
2485       team->t.t_level > master_th->th.th_teams_level)
2486 #endif /* OMP_40_ENABLED */
2487   {
2488     /* Decrement our nested depth level */
2489     KMP_ATOMIC_DEC(&root->r.r_in_parallel);
2490   }
2491   KMP_DEBUG_ASSERT(root->r.r_in_parallel >= 0);
2492 
2493 #if OMPT_SUPPORT
2494   if (ompt_enabled.enabled) {
2495     ompt_task_info_t *task_info = __ompt_get_task_info_object(0);
2496     if (ompt_enabled.ompt_callback_implicit_task) {
2497       int ompt_team_size = team->t.t_nproc;
2498       ompt_callbacks.ompt_callback(ompt_callback_implicit_task)(
2499           ompt_scope_end, NULL, &(task_info->task_data), ompt_team_size,
2500           OMPT_CUR_TASK_INFO(master_th)->thread_num);
2501     }
2502 
2503     task_info->frame.exit_frame = NULL;
2504     task_info->task_data = ompt_data_none;
2505   }
2506 #endif
2507 
2508   KF_TRACE(10, ("__kmp_join_call1: T#%d, this_thread=%p team=%p\n", 0,
2509                 master_th, team));
2510   __kmp_pop_current_task_from_thread(master_th);
2511 
2512 #if OMP_40_ENABLED && KMP_AFFINITY_SUPPORTED
2513   // Restore master thread's partition.
2514   master_th->th.th_first_place = team->t.t_first_place;
2515   master_th->th.th_last_place = team->t.t_last_place;
2516 #endif /* OMP_40_ENABLED */
2517 
2518   updateHWFPControl(team);
2519 
2520   if (root->r.r_active != master_active)
2521     root->r.r_active = master_active;
2522 
2523   __kmp_free_team(root, team USE_NESTED_HOT_ARG(
2524                             master_th)); // this will free worker threads
2525 
2526   /* this race was fun to find. make sure the following is in the critical
2527      region otherwise assertions may fail occasionally since the old team may be
2528      reallocated and the hierarchy appears inconsistent. it is actually safe to
2529      run and won't cause any bugs, but will cause those assertion failures. it's
2530      only one deref&assign so might as well put this in the critical region */
2531   master_th->th.th_team = parent_team;
2532   master_th->th.th_team_nproc = parent_team->t.t_nproc;
2533   master_th->th.th_team_master = parent_team->t.t_threads[0];
2534   master_th->th.th_team_serialized = parent_team->t.t_serialized;
2535 
2536   /* restore serialized team, if need be */
2537   if (parent_team->t.t_serialized &&
2538       parent_team != master_th->th.th_serial_team &&
2539       parent_team != root->r.r_root_team) {
2540     __kmp_free_team(root,
2541                     master_th->th.th_serial_team USE_NESTED_HOT_ARG(NULL));
2542     master_th->th.th_serial_team = parent_team;
2543   }
2544 
2545   if (__kmp_tasking_mode != tskm_immediate_exec) {
2546     if (master_th->th.th_task_state_top >
2547         0) { // Restore task state from memo stack
2548       KMP_DEBUG_ASSERT(master_th->th.th_task_state_memo_stack);
2549       // Remember master's state if we re-use this nested hot team
2550       master_th->th.th_task_state_memo_stack[master_th->th.th_task_state_top] =
2551           master_th->th.th_task_state;
2552       --master_th->th.th_task_state_top; // pop
2553       // Now restore state at this level
2554       master_th->th.th_task_state =
2555           master_th->th
2556               .th_task_state_memo_stack[master_th->th.th_task_state_top];
2557     }
2558     // Copy the task team from the parent team to the master thread
2559     master_th->th.th_task_team =
2560         parent_team->t.t_task_team[master_th->th.th_task_state];
2561     KA_TRACE(20,
2562              ("__kmp_join_call: Master T#%d restoring task_team %p / team %p\n",
2563               __kmp_gtid_from_thread(master_th), master_th->th.th_task_team,
2564               parent_team));
2565   }
2566 
2567   // TODO: GEH - cannot do this assertion because root thread not set up as
2568   // executing
2569   // KMP_ASSERT( master_th->th.th_current_task->td_flags.executing == 0 );
2570   master_th->th.th_current_task->td_flags.executing = 1;
2571 
2572   __kmp_release_bootstrap_lock(&__kmp_forkjoin_lock);
2573 
2574 #if OMPT_SUPPORT
2575   if (ompt_enabled.enabled) {
2576     __kmp_join_ompt(gtid, master_th, parent_team, parallel_data, fork_context,
2577                     codeptr);
2578   }
2579 #endif
2580 
2581   KMP_MB();
2582   KA_TRACE(20, ("__kmp_join_call: exit T#%d\n", gtid));
2583 }
2584 
2585 /* Check whether we should push an internal control record onto the
2586    serial team stack.  If so, do it.  */
2587 void __kmp_save_internal_controls(kmp_info_t *thread) {
2588 
2589   if (thread->th.th_team != thread->th.th_serial_team) {
2590     return;
2591   }
2592   if (thread->th.th_team->t.t_serialized > 1) {
2593     int push = 0;
2594 
2595     if (thread->th.th_team->t.t_control_stack_top == NULL) {
2596       push = 1;
2597     } else {
2598       if (thread->th.th_team->t.t_control_stack_top->serial_nesting_level !=
2599           thread->th.th_team->t.t_serialized) {
2600         push = 1;
2601       }
2602     }
2603     if (push) { /* push a record on the serial team's stack */
2604       kmp_internal_control_t *control =
2605           (kmp_internal_control_t *)__kmp_allocate(
2606               sizeof(kmp_internal_control_t));
2607 
2608       copy_icvs(control, &thread->th.th_current_task->td_icvs);
2609 
2610       control->serial_nesting_level = thread->th.th_team->t.t_serialized;
2611 
2612       control->next = thread->th.th_team->t.t_control_stack_top;
2613       thread->th.th_team->t.t_control_stack_top = control;
2614     }
2615   }
2616 }
2617 
2618 /* Changes set_nproc */
2619 void __kmp_set_num_threads(int new_nth, int gtid) {
2620   kmp_info_t *thread;
2621   kmp_root_t *root;
2622 
2623   KF_TRACE(10, ("__kmp_set_num_threads: new __kmp_nth = %d\n", new_nth));
2624   KMP_DEBUG_ASSERT(__kmp_init_serial);
2625 
2626   if (new_nth < 1)
2627     new_nth = 1;
2628   else if (new_nth > __kmp_max_nth)
2629     new_nth = __kmp_max_nth;
2630 
2631   KMP_COUNT_VALUE(OMP_set_numthreads, new_nth);
2632   thread = __kmp_threads[gtid];
2633 
2634   __kmp_save_internal_controls(thread);
2635 
2636   set__nproc(thread, new_nth);
2637 
2638   // If this omp_set_num_threads() call will cause the hot team size to be
2639   // reduced (in the absence of a num_threads clause), then reduce it now,
2640   // rather than waiting for the next parallel region.
2641   root = thread->th.th_root;
2642   if (__kmp_init_parallel && (!root->r.r_active) &&
2643       (root->r.r_hot_team->t.t_nproc > new_nth)
2644 #if KMP_NESTED_HOT_TEAMS
2645       && __kmp_hot_teams_max_level && !__kmp_hot_teams_mode
2646 #endif
2647       ) {
2648     kmp_team_t *hot_team = root->r.r_hot_team;
2649     int f;
2650 
2651     __kmp_acquire_bootstrap_lock(&__kmp_forkjoin_lock);
2652 
2653     // Release the extra threads we don't need any more.
2654     for (f = new_nth; f < hot_team->t.t_nproc; f++) {
2655       KMP_DEBUG_ASSERT(hot_team->t.t_threads[f] != NULL);
2656       if (__kmp_tasking_mode != tskm_immediate_exec) {
2657         // When decreasing team size, threads no longer in the team should unref
2658         // task team.
2659         hot_team->t.t_threads[f]->th.th_task_team = NULL;
2660       }
2661       __kmp_free_thread(hot_team->t.t_threads[f]);
2662       hot_team->t.t_threads[f] = NULL;
2663     }
2664     hot_team->t.t_nproc = new_nth;
2665 #if KMP_NESTED_HOT_TEAMS
2666     if (thread->th.th_hot_teams) {
2667       KMP_DEBUG_ASSERT(hot_team == thread->th.th_hot_teams[0].hot_team);
2668       thread->th.th_hot_teams[0].hot_team_nth = new_nth;
2669     }
2670 #endif
2671 
2672     __kmp_release_bootstrap_lock(&__kmp_forkjoin_lock);
2673 
2674     // Update the t_nproc field in the threads that are still active.
2675     for (f = 0; f < new_nth; f++) {
2676       KMP_DEBUG_ASSERT(hot_team->t.t_threads[f] != NULL);
2677       hot_team->t.t_threads[f]->th.th_team_nproc = new_nth;
2678     }
2679     // Special flag in case omp_set_num_threads() call
2680     hot_team->t.t_size_changed = -1;
2681   }
2682 }
2683 
2684 /* Changes max_active_levels */
2685 void __kmp_set_max_active_levels(int gtid, int max_active_levels) {
2686   kmp_info_t *thread;
2687 
2688   KF_TRACE(10, ("__kmp_set_max_active_levels: new max_active_levels for thread "
2689                 "%d = (%d)\n",
2690                 gtid, max_active_levels));
2691   KMP_DEBUG_ASSERT(__kmp_init_serial);
2692 
2693   // validate max_active_levels
2694   if (max_active_levels < 0) {
2695     KMP_WARNING(ActiveLevelsNegative, max_active_levels);
2696     // We ignore this call if the user has specified a negative value.
2697     // The current setting won't be changed. The last valid setting will be
2698     // used. A warning will be issued (if warnings are allowed as controlled by
2699     // the KMP_WARNINGS env var).
2700     KF_TRACE(10, ("__kmp_set_max_active_levels: the call is ignored: new "
2701                   "max_active_levels for thread %d = (%d)\n",
2702                   gtid, max_active_levels));
2703     return;
2704   }
2705   if (max_active_levels <= KMP_MAX_ACTIVE_LEVELS_LIMIT) {
2706     // it's OK, the max_active_levels is within the valid range: [ 0;
2707     // KMP_MAX_ACTIVE_LEVELS_LIMIT ]
2708     // We allow a zero value. (implementation defined behavior)
2709   } else {
2710     KMP_WARNING(ActiveLevelsExceedLimit, max_active_levels,
2711                 KMP_MAX_ACTIVE_LEVELS_LIMIT);
2712     max_active_levels = KMP_MAX_ACTIVE_LEVELS_LIMIT;
2713     // Current upper limit is MAX_INT. (implementation defined behavior)
2714     // If the input exceeds the upper limit, we correct the input to be the
2715     // upper limit. (implementation defined behavior)
2716     // Actually, the flow should never get here until we use MAX_INT limit.
2717   }
2718   KF_TRACE(10, ("__kmp_set_max_active_levels: after validation: new "
2719                 "max_active_levels for thread %d = (%d)\n",
2720                 gtid, max_active_levels));
2721 
2722   thread = __kmp_threads[gtid];
2723 
2724   __kmp_save_internal_controls(thread);
2725 
2726   set__max_active_levels(thread, max_active_levels);
2727 }
2728 
2729 /* Gets max_active_levels */
2730 int __kmp_get_max_active_levels(int gtid) {
2731   kmp_info_t *thread;
2732 
2733   KF_TRACE(10, ("__kmp_get_max_active_levels: thread %d\n", gtid));
2734   KMP_DEBUG_ASSERT(__kmp_init_serial);
2735 
2736   thread = __kmp_threads[gtid];
2737   KMP_DEBUG_ASSERT(thread->th.th_current_task);
2738   KF_TRACE(10, ("__kmp_get_max_active_levels: thread %d, curtask=%p, "
2739                 "curtask_maxaclevel=%d\n",
2740                 gtid, thread->th.th_current_task,
2741                 thread->th.th_current_task->td_icvs.max_active_levels));
2742   return thread->th.th_current_task->td_icvs.max_active_levels;
2743 }
2744 
2745 /* Changes def_sched_var ICV values (run-time schedule kind and chunk) */
2746 void __kmp_set_schedule(int gtid, kmp_sched_t kind, int chunk) {
2747   kmp_info_t *thread;
2748   //    kmp_team_t *team;
2749 
2750   KF_TRACE(10, ("__kmp_set_schedule: new schedule for thread %d = (%d, %d)\n",
2751                 gtid, (int)kind, chunk));
2752   KMP_DEBUG_ASSERT(__kmp_init_serial);
2753 
2754   // Check if the kind parameter is valid, correct if needed.
2755   // Valid parameters should fit in one of two intervals - standard or extended:
2756   //       <lower>, <valid>, <upper_std>, <lower_ext>, <valid>, <upper>
2757   // 2008-01-25: 0,  1 - 4,       5,         100,     101 - 102, 103
2758   if (kind <= kmp_sched_lower || kind >= kmp_sched_upper ||
2759       (kind <= kmp_sched_lower_ext && kind >= kmp_sched_upper_std)) {
2760     // TODO: Hint needs attention in case we change the default schedule.
2761     __kmp_msg(kmp_ms_warning, KMP_MSG(ScheduleKindOutOfRange, kind),
2762               KMP_HNT(DefaultScheduleKindUsed, "static, no chunk"),
2763               __kmp_msg_null);
2764     kind = kmp_sched_default;
2765     chunk = 0; // ignore chunk value in case of bad kind
2766   }
2767 
2768   thread = __kmp_threads[gtid];
2769 
2770   __kmp_save_internal_controls(thread);
2771 
2772   if (kind < kmp_sched_upper_std) {
2773     if (kind == kmp_sched_static && chunk < KMP_DEFAULT_CHUNK) {
2774       // differ static chunked vs. unchunked:  chunk should be invalid to
2775       // indicate unchunked schedule (which is the default)
2776       thread->th.th_current_task->td_icvs.sched.r_sched_type = kmp_sch_static;
2777     } else {
2778       thread->th.th_current_task->td_icvs.sched.r_sched_type =
2779           __kmp_sch_map[kind - kmp_sched_lower - 1];
2780     }
2781   } else {
2782     //    __kmp_sch_map[ kind - kmp_sched_lower_ext + kmp_sched_upper_std -
2783     //    kmp_sched_lower - 2 ];
2784     thread->th.th_current_task->td_icvs.sched.r_sched_type =
2785         __kmp_sch_map[kind - kmp_sched_lower_ext + kmp_sched_upper_std -
2786                       kmp_sched_lower - 2];
2787   }
2788   if (kind == kmp_sched_auto || chunk < 1) {
2789     // ignore parameter chunk for schedule auto
2790     thread->th.th_current_task->td_icvs.sched.chunk = KMP_DEFAULT_CHUNK;
2791   } else {
2792     thread->th.th_current_task->td_icvs.sched.chunk = chunk;
2793   }
2794 }
2795 
2796 /* Gets def_sched_var ICV values */
2797 void __kmp_get_schedule(int gtid, kmp_sched_t *kind, int *chunk) {
2798   kmp_info_t *thread;
2799   enum sched_type th_type;
2800 
2801   KF_TRACE(10, ("__kmp_get_schedule: thread %d\n", gtid));
2802   KMP_DEBUG_ASSERT(__kmp_init_serial);
2803 
2804   thread = __kmp_threads[gtid];
2805 
2806   th_type = thread->th.th_current_task->td_icvs.sched.r_sched_type;
2807 
2808   switch (th_type) {
2809   case kmp_sch_static:
2810   case kmp_sch_static_greedy:
2811   case kmp_sch_static_balanced:
2812     *kind = kmp_sched_static;
2813     *chunk = 0; // chunk was not set, try to show this fact via zero value
2814     return;
2815   case kmp_sch_static_chunked:
2816     *kind = kmp_sched_static;
2817     break;
2818   case kmp_sch_dynamic_chunked:
2819     *kind = kmp_sched_dynamic;
2820     break;
2821   case kmp_sch_guided_chunked:
2822   case kmp_sch_guided_iterative_chunked:
2823   case kmp_sch_guided_analytical_chunked:
2824     *kind = kmp_sched_guided;
2825     break;
2826   case kmp_sch_auto:
2827     *kind = kmp_sched_auto;
2828     break;
2829   case kmp_sch_trapezoidal:
2830     *kind = kmp_sched_trapezoidal;
2831     break;
2832 #if KMP_STATIC_STEAL_ENABLED
2833   case kmp_sch_static_steal:
2834     *kind = kmp_sched_static_steal;
2835     break;
2836 #endif
2837   default:
2838     KMP_FATAL(UnknownSchedulingType, th_type);
2839   }
2840 
2841   *chunk = thread->th.th_current_task->td_icvs.sched.chunk;
2842 }
2843 
2844 int __kmp_get_ancestor_thread_num(int gtid, int level) {
2845 
2846   int ii, dd;
2847   kmp_team_t *team;
2848   kmp_info_t *thr;
2849 
2850   KF_TRACE(10, ("__kmp_get_ancestor_thread_num: thread %d %d\n", gtid, level));
2851   KMP_DEBUG_ASSERT(__kmp_init_serial);
2852 
2853   // validate level
2854   if (level == 0)
2855     return 0;
2856   if (level < 0)
2857     return -1;
2858   thr = __kmp_threads[gtid];
2859   team = thr->th.th_team;
2860   ii = team->t.t_level;
2861   if (level > ii)
2862     return -1;
2863 
2864 #if OMP_40_ENABLED
2865   if (thr->th.th_teams_microtask) {
2866     // AC: we are in teams region where multiple nested teams have same level
2867     int tlevel = thr->th.th_teams_level; // the level of the teams construct
2868     if (level <=
2869         tlevel) { // otherwise usual algorithm works (will not touch the teams)
2870       KMP_DEBUG_ASSERT(ii >= tlevel);
2871       // AC: As we need to pass by the teams league, we need to artificially
2872       // increase ii
2873       if (ii == tlevel) {
2874         ii += 2; // three teams have same level
2875       } else {
2876         ii++; // two teams have same level
2877       }
2878     }
2879   }
2880 #endif
2881 
2882   if (ii == level)
2883     return __kmp_tid_from_gtid(gtid);
2884 
2885   dd = team->t.t_serialized;
2886   level++;
2887   while (ii > level) {
2888     for (dd = team->t.t_serialized; (dd > 0) && (ii > level); dd--, ii--) {
2889     }
2890     if ((team->t.t_serialized) && (!dd)) {
2891       team = team->t.t_parent;
2892       continue;
2893     }
2894     if (ii > level) {
2895       team = team->t.t_parent;
2896       dd = team->t.t_serialized;
2897       ii--;
2898     }
2899   }
2900 
2901   return (dd > 1) ? (0) : (team->t.t_master_tid);
2902 }
2903 
2904 int __kmp_get_team_size(int gtid, int level) {
2905 
2906   int ii, dd;
2907   kmp_team_t *team;
2908   kmp_info_t *thr;
2909 
2910   KF_TRACE(10, ("__kmp_get_team_size: thread %d %d\n", gtid, level));
2911   KMP_DEBUG_ASSERT(__kmp_init_serial);
2912 
2913   // validate level
2914   if (level == 0)
2915     return 1;
2916   if (level < 0)
2917     return -1;
2918   thr = __kmp_threads[gtid];
2919   team = thr->th.th_team;
2920   ii = team->t.t_level;
2921   if (level > ii)
2922     return -1;
2923 
2924 #if OMP_40_ENABLED
2925   if (thr->th.th_teams_microtask) {
2926     // AC: we are in teams region where multiple nested teams have same level
2927     int tlevel = thr->th.th_teams_level; // the level of the teams construct
2928     if (level <=
2929         tlevel) { // otherwise usual algorithm works (will not touch the teams)
2930       KMP_DEBUG_ASSERT(ii >= tlevel);
2931       // AC: As we need to pass by the teams league, we need to artificially
2932       // increase ii
2933       if (ii == tlevel) {
2934         ii += 2; // three teams have same level
2935       } else {
2936         ii++; // two teams have same level
2937       }
2938     }
2939   }
2940 #endif
2941 
2942   while (ii > level) {
2943     for (dd = team->t.t_serialized; (dd > 0) && (ii > level); dd--, ii--) {
2944     }
2945     if (team->t.t_serialized && (!dd)) {
2946       team = team->t.t_parent;
2947       continue;
2948     }
2949     if (ii > level) {
2950       team = team->t.t_parent;
2951       ii--;
2952     }
2953   }
2954 
2955   return team->t.t_nproc;
2956 }
2957 
2958 kmp_r_sched_t __kmp_get_schedule_global() {
2959   // This routine created because pairs (__kmp_sched, __kmp_chunk) and
2960   // (__kmp_static, __kmp_guided) may be changed by kmp_set_defaults
2961   // independently. So one can get the updated schedule here.
2962 
2963   kmp_r_sched_t r_sched;
2964 
2965   // create schedule from 4 globals: __kmp_sched, __kmp_chunk, __kmp_static,
2966   // __kmp_guided. __kmp_sched should keep original value, so that user can set
2967   // KMP_SCHEDULE multiple times, and thus have different run-time schedules in
2968   // different roots (even in OMP 2.5)
2969   if (__kmp_sched == kmp_sch_static) {
2970     // replace STATIC with more detailed schedule (balanced or greedy)
2971     r_sched.r_sched_type = __kmp_static;
2972   } else if (__kmp_sched == kmp_sch_guided_chunked) {
2973     // replace GUIDED with more detailed schedule (iterative or analytical)
2974     r_sched.r_sched_type = __kmp_guided;
2975   } else { // (STATIC_CHUNKED), or (DYNAMIC_CHUNKED), or other
2976     r_sched.r_sched_type = __kmp_sched;
2977   }
2978 
2979   if (__kmp_chunk < KMP_DEFAULT_CHUNK) {
2980     // __kmp_chunk may be wrong here (if it was not ever set)
2981     r_sched.chunk = KMP_DEFAULT_CHUNK;
2982   } else {
2983     r_sched.chunk = __kmp_chunk;
2984   }
2985 
2986   return r_sched;
2987 }
2988 
2989 /* Allocate (realloc == FALSE) * or reallocate (realloc == TRUE)
2990    at least argc number of *t_argv entries for the requested team. */
2991 static void __kmp_alloc_argv_entries(int argc, kmp_team_t *team, int realloc) {
2992 
2993   KMP_DEBUG_ASSERT(team);
2994   if (!realloc || argc > team->t.t_max_argc) {
2995 
2996     KA_TRACE(100, ("__kmp_alloc_argv_entries: team %d: needed entries=%d, "
2997                    "current entries=%d\n",
2998                    team->t.t_id, argc, (realloc) ? team->t.t_max_argc : 0));
2999     /* if previously allocated heap space for args, free them */
3000     if (realloc && team->t.t_argv != &team->t.t_inline_argv[0])
3001       __kmp_free((void *)team->t.t_argv);
3002 
3003     if (argc <= KMP_INLINE_ARGV_ENTRIES) {
3004       /* use unused space in the cache line for arguments */
3005       team->t.t_max_argc = KMP_INLINE_ARGV_ENTRIES;
3006       KA_TRACE(100, ("__kmp_alloc_argv_entries: team %d: inline allocate %d "
3007                      "argv entries\n",
3008                      team->t.t_id, team->t.t_max_argc));
3009       team->t.t_argv = &team->t.t_inline_argv[0];
3010       if (__kmp_storage_map) {
3011         __kmp_print_storage_map_gtid(
3012             -1, &team->t.t_inline_argv[0],
3013             &team->t.t_inline_argv[KMP_INLINE_ARGV_ENTRIES],
3014             (sizeof(void *) * KMP_INLINE_ARGV_ENTRIES), "team_%d.t_inline_argv",
3015             team->t.t_id);
3016       }
3017     } else {
3018       /* allocate space for arguments in the heap */
3019       team->t.t_max_argc = (argc <= (KMP_MIN_MALLOC_ARGV_ENTRIES >> 1))
3020                                ? KMP_MIN_MALLOC_ARGV_ENTRIES
3021                                : 2 * argc;
3022       KA_TRACE(100, ("__kmp_alloc_argv_entries: team %d: dynamic allocate %d "
3023                      "argv entries\n",
3024                      team->t.t_id, team->t.t_max_argc));
3025       team->t.t_argv =
3026           (void **)__kmp_page_allocate(sizeof(void *) * team->t.t_max_argc);
3027       if (__kmp_storage_map) {
3028         __kmp_print_storage_map_gtid(-1, &team->t.t_argv[0],
3029                                      &team->t.t_argv[team->t.t_max_argc],
3030                                      sizeof(void *) * team->t.t_max_argc,
3031                                      "team_%d.t_argv", team->t.t_id);
3032       }
3033     }
3034   }
3035 }
3036 
3037 static void __kmp_allocate_team_arrays(kmp_team_t *team, int max_nth) {
3038   int i;
3039   int num_disp_buff = max_nth > 1 ? __kmp_dispatch_num_buffers : 2;
3040   team->t.t_threads =
3041       (kmp_info_t **)__kmp_allocate(sizeof(kmp_info_t *) * max_nth);
3042   team->t.t_disp_buffer = (dispatch_shared_info_t *)__kmp_allocate(
3043       sizeof(dispatch_shared_info_t) * num_disp_buff);
3044   team->t.t_dispatch =
3045       (kmp_disp_t *)__kmp_allocate(sizeof(kmp_disp_t) * max_nth);
3046   team->t.t_implicit_task_taskdata =
3047       (kmp_taskdata_t *)__kmp_allocate(sizeof(kmp_taskdata_t) * max_nth);
3048   team->t.t_max_nproc = max_nth;
3049 
3050   /* setup dispatch buffers */
3051   for (i = 0; i < num_disp_buff; ++i) {
3052     team->t.t_disp_buffer[i].buffer_index = i;
3053 #if OMP_45_ENABLED
3054     team->t.t_disp_buffer[i].doacross_buf_idx = i;
3055 #endif
3056   }
3057 }
3058 
3059 static void __kmp_free_team_arrays(kmp_team_t *team) {
3060   /* Note: this does not free the threads in t_threads (__kmp_free_threads) */
3061   int i;
3062   for (i = 0; i < team->t.t_max_nproc; ++i) {
3063     if (team->t.t_dispatch[i].th_disp_buffer != NULL) {
3064       __kmp_free(team->t.t_dispatch[i].th_disp_buffer);
3065       team->t.t_dispatch[i].th_disp_buffer = NULL;
3066     }
3067   }
3068 #if KMP_USE_HIER_SCHED
3069   __kmp_dispatch_free_hierarchies(team);
3070 #endif
3071   __kmp_free(team->t.t_threads);
3072   __kmp_free(team->t.t_disp_buffer);
3073   __kmp_free(team->t.t_dispatch);
3074   __kmp_free(team->t.t_implicit_task_taskdata);
3075   team->t.t_threads = NULL;
3076   team->t.t_disp_buffer = NULL;
3077   team->t.t_dispatch = NULL;
3078   team->t.t_implicit_task_taskdata = 0;
3079 }
3080 
3081 static void __kmp_reallocate_team_arrays(kmp_team_t *team, int max_nth) {
3082   kmp_info_t **oldThreads = team->t.t_threads;
3083 
3084   __kmp_free(team->t.t_disp_buffer);
3085   __kmp_free(team->t.t_dispatch);
3086   __kmp_free(team->t.t_implicit_task_taskdata);
3087   __kmp_allocate_team_arrays(team, max_nth);
3088 
3089   KMP_MEMCPY(team->t.t_threads, oldThreads,
3090              team->t.t_nproc * sizeof(kmp_info_t *));
3091 
3092   __kmp_free(oldThreads);
3093 }
3094 
3095 static kmp_internal_control_t __kmp_get_global_icvs(void) {
3096 
3097   kmp_r_sched_t r_sched =
3098       __kmp_get_schedule_global(); // get current state of scheduling globals
3099 
3100 #if OMP_40_ENABLED
3101   KMP_DEBUG_ASSERT(__kmp_nested_proc_bind.used > 0);
3102 #endif /* OMP_40_ENABLED */
3103 
3104   kmp_internal_control_t g_icvs = {
3105     0, // int serial_nesting_level; //corresponds to value of th_team_serialized
3106     (kmp_int8)__kmp_dflt_nested, // int nested; //internal control
3107     // for nested parallelism (per thread)
3108     (kmp_int8)__kmp_global.g.g_dynamic, // internal control for dynamic
3109     // adjustment of threads (per thread)
3110     (kmp_int8)__kmp_env_blocktime, // int bt_set; //internal control for
3111     // whether blocktime is explicitly set
3112     __kmp_dflt_blocktime, // int blocktime; //internal control for blocktime
3113 #if KMP_USE_MONITOR
3114     __kmp_bt_intervals, // int bt_intervals; //internal control for blocktime
3115 // intervals
3116 #endif
3117     __kmp_dflt_team_nth, // int nproc; //internal control for # of threads for
3118     // next parallel region (per thread)
3119     // (use a max ub on value if __kmp_parallel_initialize not called yet)
3120     __kmp_dflt_max_active_levels, // int max_active_levels; //internal control
3121     // for max_active_levels
3122     r_sched, // kmp_r_sched_t sched; //internal control for runtime schedule
3123 // {sched,chunk} pair
3124 #if OMP_40_ENABLED
3125     __kmp_nested_proc_bind.bind_types[0],
3126     __kmp_default_device,
3127 #endif /* OMP_40_ENABLED */
3128     NULL // struct kmp_internal_control *next;
3129   };
3130 
3131   return g_icvs;
3132 }
3133 
3134 static kmp_internal_control_t __kmp_get_x_global_icvs(const kmp_team_t *team) {
3135 
3136   kmp_internal_control_t gx_icvs;
3137   gx_icvs.serial_nesting_level =
3138       0; // probably =team->t.t_serial like in save_inter_controls
3139   copy_icvs(&gx_icvs, &team->t.t_threads[0]->th.th_current_task->td_icvs);
3140   gx_icvs.next = NULL;
3141 
3142   return gx_icvs;
3143 }
3144 
3145 static void __kmp_initialize_root(kmp_root_t *root) {
3146   int f;
3147   kmp_team_t *root_team;
3148   kmp_team_t *hot_team;
3149   int hot_team_max_nth;
3150   kmp_r_sched_t r_sched =
3151       __kmp_get_schedule_global(); // get current state of scheduling globals
3152   kmp_internal_control_t r_icvs = __kmp_get_global_icvs();
3153   KMP_DEBUG_ASSERT(root);
3154   KMP_ASSERT(!root->r.r_begin);
3155 
3156   /* setup the root state structure */
3157   __kmp_init_lock(&root->r.r_begin_lock);
3158   root->r.r_begin = FALSE;
3159   root->r.r_active = FALSE;
3160   root->r.r_in_parallel = 0;
3161   root->r.r_blocktime = __kmp_dflt_blocktime;
3162   root->r.r_nested = __kmp_dflt_nested;
3163   root->r.r_cg_nthreads = 1;
3164 
3165   /* setup the root team for this task */
3166   /* allocate the root team structure */
3167   KF_TRACE(10, ("__kmp_initialize_root: before root_team\n"));
3168 
3169   root_team =
3170       __kmp_allocate_team(root,
3171                           1, // new_nproc
3172                           1, // max_nproc
3173 #if OMPT_SUPPORT
3174                           ompt_data_none, // root parallel id
3175 #endif
3176 #if OMP_40_ENABLED
3177                           __kmp_nested_proc_bind.bind_types[0],
3178 #endif
3179                           &r_icvs,
3180                           0 // argc
3181                           USE_NESTED_HOT_ARG(NULL) // master thread is unknown
3182                           );
3183 #if USE_DEBUGGER
3184   // Non-NULL value should be assigned to make the debugger display the root
3185   // team.
3186   TCW_SYNC_PTR(root_team->t.t_pkfn, (microtask_t)(~0));
3187 #endif
3188 
3189   KF_TRACE(10, ("__kmp_initialize_root: after root_team = %p\n", root_team));
3190 
3191   root->r.r_root_team = root_team;
3192   root_team->t.t_control_stack_top = NULL;
3193 
3194   /* initialize root team */
3195   root_team->t.t_threads[0] = NULL;
3196   root_team->t.t_nproc = 1;
3197   root_team->t.t_serialized = 1;
3198   // TODO???: root_team->t.t_max_active_levels = __kmp_dflt_max_active_levels;
3199   root_team->t.t_sched.sched = r_sched.sched;
3200   KA_TRACE(
3201       20,
3202       ("__kmp_initialize_root: init root team %d arrived: join=%u, plain=%u\n",
3203        root_team->t.t_id, KMP_INIT_BARRIER_STATE, KMP_INIT_BARRIER_STATE));
3204 
3205   /* setup the  hot team for this task */
3206   /* allocate the hot team structure */
3207   KF_TRACE(10, ("__kmp_initialize_root: before hot_team\n"));
3208 
3209   hot_team =
3210       __kmp_allocate_team(root,
3211                           1, // new_nproc
3212                           __kmp_dflt_team_nth_ub * 2, // max_nproc
3213 #if OMPT_SUPPORT
3214                           ompt_data_none, // root parallel id
3215 #endif
3216 #if OMP_40_ENABLED
3217                           __kmp_nested_proc_bind.bind_types[0],
3218 #endif
3219                           &r_icvs,
3220                           0 // argc
3221                           USE_NESTED_HOT_ARG(NULL) // master thread is unknown
3222                           );
3223   KF_TRACE(10, ("__kmp_initialize_root: after hot_team = %p\n", hot_team));
3224 
3225   root->r.r_hot_team = hot_team;
3226   root_team->t.t_control_stack_top = NULL;
3227 
3228   /* first-time initialization */
3229   hot_team->t.t_parent = root_team;
3230 
3231   /* initialize hot team */
3232   hot_team_max_nth = hot_team->t.t_max_nproc;
3233   for (f = 0; f < hot_team_max_nth; ++f) {
3234     hot_team->t.t_threads[f] = NULL;
3235   }
3236   hot_team->t.t_nproc = 1;
3237   // TODO???: hot_team->t.t_max_active_levels = __kmp_dflt_max_active_levels;
3238   hot_team->t.t_sched.sched = r_sched.sched;
3239   hot_team->t.t_size_changed = 0;
3240 }
3241 
3242 #ifdef KMP_DEBUG
3243 
3244 typedef struct kmp_team_list_item {
3245   kmp_team_p const *entry;
3246   struct kmp_team_list_item *next;
3247 } kmp_team_list_item_t;
3248 typedef kmp_team_list_item_t *kmp_team_list_t;
3249 
3250 static void __kmp_print_structure_team_accum( // Add team to list of teams.
3251     kmp_team_list_t list, // List of teams.
3252     kmp_team_p const *team // Team to add.
3253     ) {
3254 
3255   // List must terminate with item where both entry and next are NULL.
3256   // Team is added to the list only once.
3257   // List is sorted in ascending order by team id.
3258   // Team id is *not* a key.
3259 
3260   kmp_team_list_t l;
3261 
3262   KMP_DEBUG_ASSERT(list != NULL);
3263   if (team == NULL) {
3264     return;
3265   }
3266 
3267   __kmp_print_structure_team_accum(list, team->t.t_parent);
3268   __kmp_print_structure_team_accum(list, team->t.t_next_pool);
3269 
3270   // Search list for the team.
3271   l = list;
3272   while (l->next != NULL && l->entry != team) {
3273     l = l->next;
3274   }
3275   if (l->next != NULL) {
3276     return; // Team has been added before, exit.
3277   }
3278 
3279   // Team is not found. Search list again for insertion point.
3280   l = list;
3281   while (l->next != NULL && l->entry->t.t_id <= team->t.t_id) {
3282     l = l->next;
3283   }
3284 
3285   // Insert team.
3286   {
3287     kmp_team_list_item_t *item = (kmp_team_list_item_t *)KMP_INTERNAL_MALLOC(
3288         sizeof(kmp_team_list_item_t));
3289     *item = *l;
3290     l->entry = team;
3291     l->next = item;
3292   }
3293 }
3294 
3295 static void __kmp_print_structure_team(char const *title, kmp_team_p const *team
3296 
3297                                        ) {
3298   __kmp_printf("%s", title);
3299   if (team != NULL) {
3300     __kmp_printf("%2x %p\n", team->t.t_id, team);
3301   } else {
3302     __kmp_printf(" - (nil)\n");
3303   }
3304 }
3305 
3306 static void __kmp_print_structure_thread(char const *title,
3307                                          kmp_info_p const *thread) {
3308   __kmp_printf("%s", title);
3309   if (thread != NULL) {
3310     __kmp_printf("%2d %p\n", thread->th.th_info.ds.ds_gtid, thread);
3311   } else {
3312     __kmp_printf(" - (nil)\n");
3313   }
3314 }
3315 
3316 void __kmp_print_structure(void) {
3317 
3318   kmp_team_list_t list;
3319 
3320   // Initialize list of teams.
3321   list =
3322       (kmp_team_list_item_t *)KMP_INTERNAL_MALLOC(sizeof(kmp_team_list_item_t));
3323   list->entry = NULL;
3324   list->next = NULL;
3325 
3326   __kmp_printf("\n------------------------------\nGlobal Thread "
3327                "Table\n------------------------------\n");
3328   {
3329     int gtid;
3330     for (gtid = 0; gtid < __kmp_threads_capacity; ++gtid) {
3331       __kmp_printf("%2d", gtid);
3332       if (__kmp_threads != NULL) {
3333         __kmp_printf(" %p", __kmp_threads[gtid]);
3334       }
3335       if (__kmp_root != NULL) {
3336         __kmp_printf(" %p", __kmp_root[gtid]);
3337       }
3338       __kmp_printf("\n");
3339     }
3340   }
3341 
3342   // Print out __kmp_threads array.
3343   __kmp_printf("\n------------------------------\nThreads\n--------------------"
3344                "----------\n");
3345   if (__kmp_threads != NULL) {
3346     int gtid;
3347     for (gtid = 0; gtid < __kmp_threads_capacity; ++gtid) {
3348       kmp_info_t const *thread = __kmp_threads[gtid];
3349       if (thread != NULL) {
3350         __kmp_printf("GTID %2d %p:\n", gtid, thread);
3351         __kmp_printf("    Our Root:        %p\n", thread->th.th_root);
3352         __kmp_print_structure_team("    Our Team:     ", thread->th.th_team);
3353         __kmp_print_structure_team("    Serial Team:  ",
3354                                    thread->th.th_serial_team);
3355         __kmp_printf("    Threads:      %2d\n", thread->th.th_team_nproc);
3356         __kmp_print_structure_thread("    Master:       ",
3357                                      thread->th.th_team_master);
3358         __kmp_printf("    Serialized?:  %2d\n", thread->th.th_team_serialized);
3359         __kmp_printf("    Set NProc:    %2d\n", thread->th.th_set_nproc);
3360 #if OMP_40_ENABLED
3361         __kmp_printf("    Set Proc Bind: %2d\n", thread->th.th_set_proc_bind);
3362 #endif
3363         __kmp_print_structure_thread("    Next in pool: ",
3364                                      thread->th.th_next_pool);
3365         __kmp_printf("\n");
3366         __kmp_print_structure_team_accum(list, thread->th.th_team);
3367         __kmp_print_structure_team_accum(list, thread->th.th_serial_team);
3368       }
3369     }
3370   } else {
3371     __kmp_printf("Threads array is not allocated.\n");
3372   }
3373 
3374   // Print out __kmp_root array.
3375   __kmp_printf("\n------------------------------\nUbers\n----------------------"
3376                "--------\n");
3377   if (__kmp_root != NULL) {
3378     int gtid;
3379     for (gtid = 0; gtid < __kmp_threads_capacity; ++gtid) {
3380       kmp_root_t const *root = __kmp_root[gtid];
3381       if (root != NULL) {
3382         __kmp_printf("GTID %2d %p:\n", gtid, root);
3383         __kmp_print_structure_team("    Root Team:    ", root->r.r_root_team);
3384         __kmp_print_structure_team("    Hot Team:     ", root->r.r_hot_team);
3385         __kmp_print_structure_thread("    Uber Thread:  ",
3386                                      root->r.r_uber_thread);
3387         __kmp_printf("    Active?:      %2d\n", root->r.r_active);
3388         __kmp_printf("    Nested?:      %2d\n", root->r.r_nested);
3389         __kmp_printf("    In Parallel:  %2d\n",
3390                      KMP_ATOMIC_LD_RLX(&root->r.r_in_parallel));
3391         __kmp_printf("\n");
3392         __kmp_print_structure_team_accum(list, root->r.r_root_team);
3393         __kmp_print_structure_team_accum(list, root->r.r_hot_team);
3394       }
3395     }
3396   } else {
3397     __kmp_printf("Ubers array is not allocated.\n");
3398   }
3399 
3400   __kmp_printf("\n------------------------------\nTeams\n----------------------"
3401                "--------\n");
3402   while (list->next != NULL) {
3403     kmp_team_p const *team = list->entry;
3404     int i;
3405     __kmp_printf("Team %2x %p:\n", team->t.t_id, team);
3406     __kmp_print_structure_team("    Parent Team:      ", team->t.t_parent);
3407     __kmp_printf("    Master TID:       %2d\n", team->t.t_master_tid);
3408     __kmp_printf("    Max threads:      %2d\n", team->t.t_max_nproc);
3409     __kmp_printf("    Levels of serial: %2d\n", team->t.t_serialized);
3410     __kmp_printf("    Number threads:   %2d\n", team->t.t_nproc);
3411     for (i = 0; i < team->t.t_nproc; ++i) {
3412       __kmp_printf("    Thread %2d:      ", i);
3413       __kmp_print_structure_thread("", team->t.t_threads[i]);
3414     }
3415     __kmp_print_structure_team("    Next in pool:     ", team->t.t_next_pool);
3416     __kmp_printf("\n");
3417     list = list->next;
3418   }
3419 
3420   // Print out __kmp_thread_pool and __kmp_team_pool.
3421   __kmp_printf("\n------------------------------\nPools\n----------------------"
3422                "--------\n");
3423   __kmp_print_structure_thread("Thread pool:          ",
3424                                CCAST(kmp_info_t *, __kmp_thread_pool));
3425   __kmp_print_structure_team("Team pool:            ",
3426                              CCAST(kmp_team_t *, __kmp_team_pool));
3427   __kmp_printf("\n");
3428 
3429   // Free team list.
3430   while (list != NULL) {
3431     kmp_team_list_item_t *item = list;
3432     list = list->next;
3433     KMP_INTERNAL_FREE(item);
3434   }
3435 }
3436 
3437 #endif
3438 
3439 //---------------------------------------------------------------------------
3440 //  Stuff for per-thread fast random number generator
3441 //  Table of primes
3442 static const unsigned __kmp_primes[] = {
3443     0x9e3779b1, 0xffe6cc59, 0x2109f6dd, 0x43977ab5, 0xba5703f5, 0xb495a877,
3444     0xe1626741, 0x79695e6b, 0xbc98c09f, 0xd5bee2b3, 0x287488f9, 0x3af18231,
3445     0x9677cd4d, 0xbe3a6929, 0xadc6a877, 0xdcf0674b, 0xbe4d6fe9, 0x5f15e201,
3446     0x99afc3fd, 0xf3f16801, 0xe222cfff, 0x24ba5fdb, 0x0620452d, 0x79f149e3,
3447     0xc8b93f49, 0x972702cd, 0xb07dd827, 0x6c97d5ed, 0x085a3d61, 0x46eb5ea7,
3448     0x3d9910ed, 0x2e687b5b, 0x29609227, 0x6eb081f1, 0x0954c4e1, 0x9d114db9,
3449     0x542acfa9, 0xb3e6bd7b, 0x0742d917, 0xe9f3ffa7, 0x54581edb, 0xf2480f45,
3450     0x0bb9288f, 0xef1affc7, 0x85fa0ca7, 0x3ccc14db, 0xe6baf34b, 0x343377f7,
3451     0x5ca19031, 0xe6d9293b, 0xf0a9f391, 0x5d2e980b, 0xfc411073, 0xc3749363,
3452     0xb892d829, 0x3549366b, 0x629750ad, 0xb98294e5, 0x892d9483, 0xc235baf3,
3453     0x3d2402a3, 0x6bdef3c9, 0xbec333cd, 0x40c9520f};
3454 
3455 //---------------------------------------------------------------------------
3456 //  __kmp_get_random: Get a random number using a linear congruential method.
3457 unsigned short __kmp_get_random(kmp_info_t *thread) {
3458   unsigned x = thread->th.th_x;
3459   unsigned short r = x >> 16;
3460 
3461   thread->th.th_x = x * thread->th.th_a + 1;
3462 
3463   KA_TRACE(30, ("__kmp_get_random: THREAD: %d, RETURN: %u\n",
3464                 thread->th.th_info.ds.ds_tid, r));
3465 
3466   return r;
3467 }
3468 //--------------------------------------------------------
3469 // __kmp_init_random: Initialize a random number generator
3470 void __kmp_init_random(kmp_info_t *thread) {
3471   unsigned seed = thread->th.th_info.ds.ds_tid;
3472 
3473   thread->th.th_a =
3474       __kmp_primes[seed % (sizeof(__kmp_primes) / sizeof(__kmp_primes[0]))];
3475   thread->th.th_x = (seed + 1) * thread->th.th_a + 1;
3476   KA_TRACE(30,
3477            ("__kmp_init_random: THREAD: %u; A: %u\n", seed, thread->th.th_a));
3478 }
3479 
3480 #if KMP_OS_WINDOWS
3481 /* reclaim array entries for root threads that are already dead, returns number
3482  * reclaimed */
3483 static int __kmp_reclaim_dead_roots(void) {
3484   int i, r = 0;
3485 
3486   for (i = 0; i < __kmp_threads_capacity; ++i) {
3487     if (KMP_UBER_GTID(i) &&
3488         !__kmp_still_running((kmp_info_t *)TCR_SYNC_PTR(__kmp_threads[i])) &&
3489         !__kmp_root[i]
3490              ->r.r_active) { // AC: reclaim only roots died in non-active state
3491       r += __kmp_unregister_root_other_thread(i);
3492     }
3493   }
3494   return r;
3495 }
3496 #endif
3497 
3498 /* This function attempts to create free entries in __kmp_threads and
3499    __kmp_root, and returns the number of free entries generated.
3500 
3501    For Windows* OS static library, the first mechanism used is to reclaim array
3502    entries for root threads that are already dead.
3503 
3504    On all platforms, expansion is attempted on the arrays __kmp_threads_ and
3505    __kmp_root, with appropriate update to __kmp_threads_capacity. Array
3506    capacity is increased by doubling with clipping to __kmp_tp_capacity, if
3507    threadprivate cache array has been created. Synchronization with
3508    __kmpc_threadprivate_cached is done using __kmp_tp_cached_lock.
3509 
3510    After any dead root reclamation, if the clipping value allows array expansion
3511    to result in the generation of a total of nNeed free slots, the function does
3512    that expansion. If not, nothing is done beyond the possible initial root
3513    thread reclamation.
3514 
3515    If any argument is negative, the behavior is undefined. */
3516 static int __kmp_expand_threads(int nNeed) {
3517   int added = 0;
3518   int minimumRequiredCapacity;
3519   int newCapacity;
3520   kmp_info_t **newThreads;
3521   kmp_root_t **newRoot;
3522 
3523 // All calls to __kmp_expand_threads should be under __kmp_forkjoin_lock, so
3524 // resizing __kmp_threads does not need additional protection if foreign
3525 // threads are present
3526 
3527 #if KMP_OS_WINDOWS && !defined KMP_DYNAMIC_LIB
3528   /* only for Windows static library */
3529   /* reclaim array entries for root threads that are already dead */
3530   added = __kmp_reclaim_dead_roots();
3531 
3532   if (nNeed) {
3533     nNeed -= added;
3534     if (nNeed < 0)
3535       nNeed = 0;
3536   }
3537 #endif
3538   if (nNeed <= 0)
3539     return added;
3540 
3541   // Note that __kmp_threads_capacity is not bounded by __kmp_max_nth. If
3542   // __kmp_max_nth is set to some value less than __kmp_sys_max_nth by the
3543   // user via KMP_DEVICE_THREAD_LIMIT, then __kmp_threads_capacity may become
3544   // > __kmp_max_nth in one of two ways:
3545   //
3546   // 1) The initialization thread (gtid = 0) exits.  __kmp_threads[0]
3547   //    may not be resused by another thread, so we may need to increase
3548   //    __kmp_threads_capacity to __kmp_max_nth + 1.
3549   //
3550   // 2) New foreign root(s) are encountered.  We always register new foreign
3551   //    roots. This may cause a smaller # of threads to be allocated at
3552   //    subsequent parallel regions, but the worker threads hang around (and
3553   //    eventually go to sleep) and need slots in the __kmp_threads[] array.
3554   //
3555   // Anyway, that is the reason for moving the check to see if
3556   // __kmp_max_nth was exceeded into __kmp_reserve_threads()
3557   // instead of having it performed here. -BB
3558 
3559   KMP_DEBUG_ASSERT(__kmp_sys_max_nth >= __kmp_threads_capacity);
3560 
3561   /* compute expansion headroom to check if we can expand */
3562   if (__kmp_sys_max_nth - __kmp_threads_capacity < nNeed) {
3563     /* possible expansion too small -- give up */
3564     return added;
3565   }
3566   minimumRequiredCapacity = __kmp_threads_capacity + nNeed;
3567 
3568   newCapacity = __kmp_threads_capacity;
3569   do {
3570     newCapacity = newCapacity <= (__kmp_sys_max_nth >> 1) ? (newCapacity << 1)
3571                                                           : __kmp_sys_max_nth;
3572   } while (newCapacity < minimumRequiredCapacity);
3573   newThreads = (kmp_info_t **)__kmp_allocate(
3574       (sizeof(kmp_info_t *) + sizeof(kmp_root_t *)) * newCapacity + CACHE_LINE);
3575   newRoot =
3576       (kmp_root_t **)((char *)newThreads + sizeof(kmp_info_t *) * newCapacity);
3577   KMP_MEMCPY(newThreads, __kmp_threads,
3578              __kmp_threads_capacity * sizeof(kmp_info_t *));
3579   KMP_MEMCPY(newRoot, __kmp_root,
3580              __kmp_threads_capacity * sizeof(kmp_root_t *));
3581 
3582   kmp_info_t **temp_threads = __kmp_threads;
3583   *(kmp_info_t * *volatile *)&__kmp_threads = newThreads;
3584   *(kmp_root_t * *volatile *)&__kmp_root = newRoot;
3585   __kmp_free(temp_threads);
3586   added += newCapacity - __kmp_threads_capacity;
3587   *(volatile int *)&__kmp_threads_capacity = newCapacity;
3588 
3589   if (newCapacity > __kmp_tp_capacity) {
3590     __kmp_acquire_bootstrap_lock(&__kmp_tp_cached_lock);
3591     if (__kmp_tp_cached && newCapacity > __kmp_tp_capacity) {
3592       __kmp_threadprivate_resize_cache(newCapacity);
3593     } else { // increase __kmp_tp_capacity to correspond with kmp_threads size
3594       *(volatile int *)&__kmp_tp_capacity = newCapacity;
3595     }
3596     __kmp_release_bootstrap_lock(&__kmp_tp_cached_lock);
3597   }
3598 
3599   return added;
3600 }
3601 
3602 /* Register the current thread as a root thread and obtain our gtid. We must
3603    have the __kmp_initz_lock held at this point. Argument TRUE only if are the
3604    thread that calls from __kmp_do_serial_initialize() */
3605 int __kmp_register_root(int initial_thread) {
3606   kmp_info_t *root_thread;
3607   kmp_root_t *root;
3608   int gtid;
3609   int capacity;
3610   __kmp_acquire_bootstrap_lock(&__kmp_forkjoin_lock);
3611   KA_TRACE(20, ("__kmp_register_root: entered\n"));
3612   KMP_MB();
3613 
3614   /* 2007-03-02:
3615      If initial thread did not invoke OpenMP RTL yet, and this thread is not an
3616      initial one, "__kmp_all_nth >= __kmp_threads_capacity" condition does not
3617      work as expected -- it may return false (that means there is at least one
3618      empty slot in __kmp_threads array), but it is possible the only free slot
3619      is #0, which is reserved for initial thread and so cannot be used for this
3620      one. Following code workarounds this bug.
3621 
3622      However, right solution seems to be not reserving slot #0 for initial
3623      thread because:
3624      (1) there is no magic in slot #0,
3625      (2) we cannot detect initial thread reliably (the first thread which does
3626         serial initialization may be not a real initial thread).
3627   */
3628   capacity = __kmp_threads_capacity;
3629   if (!initial_thread && TCR_PTR(__kmp_threads[0]) == NULL) {
3630     --capacity;
3631   }
3632 
3633   /* see if there are too many threads */
3634   if (__kmp_all_nth >= capacity && !__kmp_expand_threads(1)) {
3635     if (__kmp_tp_cached) {
3636       __kmp_fatal(KMP_MSG(CantRegisterNewThread),
3637                   KMP_HNT(Set_ALL_THREADPRIVATE, __kmp_tp_capacity),
3638                   KMP_HNT(PossibleSystemLimitOnThreads), __kmp_msg_null);
3639     } else {
3640       __kmp_fatal(KMP_MSG(CantRegisterNewThread), KMP_HNT(SystemLimitOnThreads),
3641                   __kmp_msg_null);
3642     }
3643   }
3644 
3645   /* find an available thread slot */
3646   /* Don't reassign the zero slot since we need that to only be used by initial
3647      thread */
3648   for (gtid = (initial_thread ? 0 : 1); TCR_PTR(__kmp_threads[gtid]) != NULL;
3649        gtid++)
3650     ;
3651   KA_TRACE(1,
3652            ("__kmp_register_root: found slot in threads array: T#%d\n", gtid));
3653   KMP_ASSERT(gtid < __kmp_threads_capacity);
3654 
3655   /* update global accounting */
3656   __kmp_all_nth++;
3657   TCW_4(__kmp_nth, __kmp_nth + 1);
3658 
3659   // if __kmp_adjust_gtid_mode is set, then we use method #1 (sp search) for low
3660   // numbers of procs, and method #2 (keyed API call) for higher numbers.
3661   if (__kmp_adjust_gtid_mode) {
3662     if (__kmp_all_nth >= __kmp_tls_gtid_min) {
3663       if (TCR_4(__kmp_gtid_mode) != 2) {
3664         TCW_4(__kmp_gtid_mode, 2);
3665       }
3666     } else {
3667       if (TCR_4(__kmp_gtid_mode) != 1) {
3668         TCW_4(__kmp_gtid_mode, 1);
3669       }
3670     }
3671   }
3672 
3673 #ifdef KMP_ADJUST_BLOCKTIME
3674   /* Adjust blocktime to zero if necessary            */
3675   /* Middle initialization might not have occurred yet */
3676   if (!__kmp_env_blocktime && (__kmp_avail_proc > 0)) {
3677     if (__kmp_nth > __kmp_avail_proc) {
3678       __kmp_zero_bt = TRUE;
3679     }
3680   }
3681 #endif /* KMP_ADJUST_BLOCKTIME */
3682 
3683   /* setup this new hierarchy */
3684   if (!(root = __kmp_root[gtid])) {
3685     root = __kmp_root[gtid] = (kmp_root_t *)__kmp_allocate(sizeof(kmp_root_t));
3686     KMP_DEBUG_ASSERT(!root->r.r_root_team);
3687   }
3688 
3689 #if KMP_STATS_ENABLED
3690   // Initialize stats as soon as possible (right after gtid assignment).
3691   __kmp_stats_thread_ptr = __kmp_stats_list->push_back(gtid);
3692   __kmp_stats_thread_ptr->startLife();
3693   KMP_SET_THREAD_STATE(SERIAL_REGION);
3694   KMP_INIT_PARTITIONED_TIMERS(OMP_serial);
3695 #endif
3696   __kmp_initialize_root(root);
3697 
3698   /* setup new root thread structure */
3699   if (root->r.r_uber_thread) {
3700     root_thread = root->r.r_uber_thread;
3701   } else {
3702     root_thread = (kmp_info_t *)__kmp_allocate(sizeof(kmp_info_t));
3703     if (__kmp_storage_map) {
3704       __kmp_print_thread_storage_map(root_thread, gtid);
3705     }
3706     root_thread->th.th_info.ds.ds_gtid = gtid;
3707 #if OMPT_SUPPORT
3708     root_thread->th.ompt_thread_info.thread_data.ptr = NULL;
3709 #endif
3710     root_thread->th.th_root = root;
3711     if (__kmp_env_consistency_check) {
3712       root_thread->th.th_cons = __kmp_allocate_cons_stack(gtid);
3713     }
3714 #if USE_FAST_MEMORY
3715     __kmp_initialize_fast_memory(root_thread);
3716 #endif /* USE_FAST_MEMORY */
3717 
3718 #if KMP_USE_BGET
3719     KMP_DEBUG_ASSERT(root_thread->th.th_local.bget_data == NULL);
3720     __kmp_initialize_bget(root_thread);
3721 #endif
3722     __kmp_init_random(root_thread); // Initialize random number generator
3723   }
3724 
3725   /* setup the serial team held in reserve by the root thread */
3726   if (!root_thread->th.th_serial_team) {
3727     kmp_internal_control_t r_icvs = __kmp_get_global_icvs();
3728     KF_TRACE(10, ("__kmp_register_root: before serial_team\n"));
3729     root_thread->th.th_serial_team =
3730         __kmp_allocate_team(root, 1, 1,
3731 #if OMPT_SUPPORT
3732                             ompt_data_none, // root parallel id
3733 #endif
3734 #if OMP_40_ENABLED
3735                             proc_bind_default,
3736 #endif
3737                             &r_icvs, 0 USE_NESTED_HOT_ARG(NULL));
3738   }
3739   KMP_ASSERT(root_thread->th.th_serial_team);
3740   KF_TRACE(10, ("__kmp_register_root: after serial_team = %p\n",
3741                 root_thread->th.th_serial_team));
3742 
3743   /* drop root_thread into place */
3744   TCW_SYNC_PTR(__kmp_threads[gtid], root_thread);
3745 
3746   root->r.r_root_team->t.t_threads[0] = root_thread;
3747   root->r.r_hot_team->t.t_threads[0] = root_thread;
3748   root_thread->th.th_serial_team->t.t_threads[0] = root_thread;
3749   // AC: the team created in reserve, not for execution (it is unused for now).
3750   root_thread->th.th_serial_team->t.t_serialized = 0;
3751   root->r.r_uber_thread = root_thread;
3752 
3753   /* initialize the thread, get it ready to go */
3754   __kmp_initialize_info(root_thread, root->r.r_root_team, 0, gtid);
3755   TCW_4(__kmp_init_gtid, TRUE);
3756 
3757   /* prepare the master thread for get_gtid() */
3758   __kmp_gtid_set_specific(gtid);
3759 
3760 #if USE_ITT_BUILD
3761   __kmp_itt_thread_name(gtid);
3762 #endif /* USE_ITT_BUILD */
3763 
3764 #ifdef KMP_TDATA_GTID
3765   __kmp_gtid = gtid;
3766 #endif
3767   __kmp_create_worker(gtid, root_thread, __kmp_stksize);
3768   KMP_DEBUG_ASSERT(__kmp_gtid_get_specific() == gtid);
3769 
3770   KA_TRACE(20, ("__kmp_register_root: T#%d init T#%d(%d:%d) arrived: join=%u, "
3771                 "plain=%u\n",
3772                 gtid, __kmp_gtid_from_tid(0, root->r.r_hot_team),
3773                 root->r.r_hot_team->t.t_id, 0, KMP_INIT_BARRIER_STATE,
3774                 KMP_INIT_BARRIER_STATE));
3775   { // Initialize barrier data.
3776     int b;
3777     for (b = 0; b < bs_last_barrier; ++b) {
3778       root_thread->th.th_bar[b].bb.b_arrived = KMP_INIT_BARRIER_STATE;
3779 #if USE_DEBUGGER
3780       root_thread->th.th_bar[b].bb.b_worker_arrived = 0;
3781 #endif
3782     }
3783   }
3784   KMP_DEBUG_ASSERT(root->r.r_hot_team->t.t_bar[bs_forkjoin_barrier].b_arrived ==
3785                    KMP_INIT_BARRIER_STATE);
3786 
3787 #if KMP_AFFINITY_SUPPORTED
3788 #if OMP_40_ENABLED
3789   root_thread->th.th_current_place = KMP_PLACE_UNDEFINED;
3790   root_thread->th.th_new_place = KMP_PLACE_UNDEFINED;
3791   root_thread->th.th_first_place = KMP_PLACE_UNDEFINED;
3792   root_thread->th.th_last_place = KMP_PLACE_UNDEFINED;
3793 #endif
3794 
3795   if (TCR_4(__kmp_init_middle)) {
3796     __kmp_affinity_set_init_mask(gtid, TRUE);
3797   }
3798 #endif /* KMP_AFFINITY_SUPPORTED */
3799 
3800   __kmp_root_counter++;
3801 
3802 #if OMPT_SUPPORT
3803   if (!initial_thread && ompt_enabled.enabled) {
3804 
3805     ompt_thread_t *root_thread = ompt_get_thread();
3806 
3807     ompt_set_thread_state(root_thread, omp_state_overhead);
3808 
3809     if (ompt_enabled.ompt_callback_thread_begin) {
3810       ompt_callbacks.ompt_callback(ompt_callback_thread_begin)(
3811           ompt_thread_initial, __ompt_get_thread_data_internal());
3812     }
3813     ompt_data_t *task_data;
3814     __ompt_get_task_info_internal(0, NULL, &task_data, NULL, NULL, NULL);
3815     if (ompt_enabled.ompt_callback_task_create) {
3816       ompt_callbacks.ompt_callback(ompt_callback_task_create)(
3817           NULL, NULL, task_data, ompt_task_initial, 0, NULL);
3818       // initial task has nothing to return to
3819     }
3820 
3821     ompt_set_thread_state(root_thread, omp_state_work_serial);
3822   }
3823 #endif
3824 
3825   KMP_MB();
3826   __kmp_release_bootstrap_lock(&__kmp_forkjoin_lock);
3827 
3828   return gtid;
3829 }
3830 
3831 #if KMP_NESTED_HOT_TEAMS
3832 static int __kmp_free_hot_teams(kmp_root_t *root, kmp_info_t *thr, int level,
3833                                 const int max_level) {
3834   int i, n, nth;
3835   kmp_hot_team_ptr_t *hot_teams = thr->th.th_hot_teams;
3836   if (!hot_teams || !hot_teams[level].hot_team) {
3837     return 0;
3838   }
3839   KMP_DEBUG_ASSERT(level < max_level);
3840   kmp_team_t *team = hot_teams[level].hot_team;
3841   nth = hot_teams[level].hot_team_nth;
3842   n = nth - 1; // master is not freed
3843   if (level < max_level - 1) {
3844     for (i = 0; i < nth; ++i) {
3845       kmp_info_t *th = team->t.t_threads[i];
3846       n += __kmp_free_hot_teams(root, th, level + 1, max_level);
3847       if (i > 0 && th->th.th_hot_teams) {
3848         __kmp_free(th->th.th_hot_teams);
3849         th->th.th_hot_teams = NULL;
3850       }
3851     }
3852   }
3853   __kmp_free_team(root, team, NULL);
3854   return n;
3855 }
3856 #endif
3857 
3858 // Resets a root thread and clear its root and hot teams.
3859 // Returns the number of __kmp_threads entries directly and indirectly freed.
3860 static int __kmp_reset_root(int gtid, kmp_root_t *root) {
3861   kmp_team_t *root_team = root->r.r_root_team;
3862   kmp_team_t *hot_team = root->r.r_hot_team;
3863   int n = hot_team->t.t_nproc;
3864   int i;
3865 
3866   KMP_DEBUG_ASSERT(!root->r.r_active);
3867 
3868   root->r.r_root_team = NULL;
3869   root->r.r_hot_team = NULL;
3870   // __kmp_free_team() does not free hot teams, so we have to clear r_hot_team
3871   // before call to __kmp_free_team().
3872   __kmp_free_team(root, root_team USE_NESTED_HOT_ARG(NULL));
3873 #if KMP_NESTED_HOT_TEAMS
3874   if (__kmp_hot_teams_max_level >
3875       0) { // need to free nested hot teams and their threads if any
3876     for (i = 0; i < hot_team->t.t_nproc; ++i) {
3877       kmp_info_t *th = hot_team->t.t_threads[i];
3878       if (__kmp_hot_teams_max_level > 1) {
3879         n += __kmp_free_hot_teams(root, th, 1, __kmp_hot_teams_max_level);
3880       }
3881       if (th->th.th_hot_teams) {
3882         __kmp_free(th->th.th_hot_teams);
3883         th->th.th_hot_teams = NULL;
3884       }
3885     }
3886   }
3887 #endif
3888   __kmp_free_team(root, hot_team USE_NESTED_HOT_ARG(NULL));
3889 
3890   // Before we can reap the thread, we need to make certain that all other
3891   // threads in the teams that had this root as ancestor have stopped trying to
3892   // steal tasks.
3893   if (__kmp_tasking_mode != tskm_immediate_exec) {
3894     __kmp_wait_to_unref_task_teams();
3895   }
3896 
3897 #if KMP_OS_WINDOWS
3898   /* Close Handle of root duplicated in __kmp_create_worker (tr #62919) */
3899   KA_TRACE(
3900       10, ("__kmp_reset_root: free handle, th = %p, handle = %" KMP_UINTPTR_SPEC
3901            "\n",
3902            (LPVOID) & (root->r.r_uber_thread->th),
3903            root->r.r_uber_thread->th.th_info.ds.ds_thread));
3904   __kmp_free_handle(root->r.r_uber_thread->th.th_info.ds.ds_thread);
3905 #endif /* KMP_OS_WINDOWS */
3906 
3907 #if OMPT_SUPPORT
3908   if (ompt_enabled.ompt_callback_thread_end) {
3909     ompt_callbacks.ompt_callback(ompt_callback_thread_end)(
3910         &(root->r.r_uber_thread->th.ompt_thread_info.thread_data));
3911   }
3912 #endif
3913 
3914   TCW_4(__kmp_nth,
3915         __kmp_nth - 1); // __kmp_reap_thread will decrement __kmp_all_nth.
3916   root->r.r_cg_nthreads--;
3917 
3918   __kmp_reap_thread(root->r.r_uber_thread, 1);
3919 
3920   // We canot put root thread to __kmp_thread_pool, so we have to reap it istead
3921   // of freeing.
3922   root->r.r_uber_thread = NULL;
3923   /* mark root as no longer in use */
3924   root->r.r_begin = FALSE;
3925 
3926   return n;
3927 }
3928 
3929 void __kmp_unregister_root_current_thread(int gtid) {
3930   KA_TRACE(1, ("__kmp_unregister_root_current_thread: enter T#%d\n", gtid));
3931   /* this lock should be ok, since unregister_root_current_thread is never
3932      called during an abort, only during a normal close. furthermore, if you
3933      have the forkjoin lock, you should never try to get the initz lock */
3934   __kmp_acquire_bootstrap_lock(&__kmp_forkjoin_lock);
3935   if (TCR_4(__kmp_global.g.g_done) || !__kmp_init_serial) {
3936     KC_TRACE(10, ("__kmp_unregister_root_current_thread: already finished, "
3937                   "exiting T#%d\n",
3938                   gtid));
3939     __kmp_release_bootstrap_lock(&__kmp_forkjoin_lock);
3940     return;
3941   }
3942   kmp_root_t *root = __kmp_root[gtid];
3943 
3944   KMP_DEBUG_ASSERT(__kmp_threads && __kmp_threads[gtid]);
3945   KMP_ASSERT(KMP_UBER_GTID(gtid));
3946   KMP_ASSERT(root == __kmp_threads[gtid]->th.th_root);
3947   KMP_ASSERT(root->r.r_active == FALSE);
3948 
3949   KMP_MB();
3950 
3951 #if OMP_45_ENABLED
3952   kmp_info_t *thread = __kmp_threads[gtid];
3953   kmp_team_t *team = thread->th.th_team;
3954   kmp_task_team_t *task_team = thread->th.th_task_team;
3955 
3956   // we need to wait for the proxy tasks before finishing the thread
3957   if (task_team != NULL && task_team->tt.tt_found_proxy_tasks) {
3958 #if OMPT_SUPPORT
3959     // the runtime is shutting down so we won't report any events
3960     thread->th.ompt_thread_info.state = omp_state_undefined;
3961 #endif
3962     __kmp_task_team_wait(thread, team USE_ITT_BUILD_ARG(NULL));
3963   }
3964 #endif
3965 
3966   __kmp_reset_root(gtid, root);
3967 
3968   /* free up this thread slot */
3969   __kmp_gtid_set_specific(KMP_GTID_DNE);
3970 #ifdef KMP_TDATA_GTID
3971   __kmp_gtid = KMP_GTID_DNE;
3972 #endif
3973 
3974   KMP_MB();
3975   KC_TRACE(10,
3976            ("__kmp_unregister_root_current_thread: T#%d unregistered\n", gtid));
3977 
3978   __kmp_release_bootstrap_lock(&__kmp_forkjoin_lock);
3979 }
3980 
3981 #if KMP_OS_WINDOWS
3982 /* __kmp_forkjoin_lock must be already held
3983    Unregisters a root thread that is not the current thread.  Returns the number
3984    of __kmp_threads entries freed as a result. */
3985 static int __kmp_unregister_root_other_thread(int gtid) {
3986   kmp_root_t *root = __kmp_root[gtid];
3987   int r;
3988 
3989   KA_TRACE(1, ("__kmp_unregister_root_other_thread: enter T#%d\n", gtid));
3990   KMP_DEBUG_ASSERT(__kmp_threads && __kmp_threads[gtid]);
3991   KMP_ASSERT(KMP_UBER_GTID(gtid));
3992   KMP_ASSERT(root == __kmp_threads[gtid]->th.th_root);
3993   KMP_ASSERT(root->r.r_active == FALSE);
3994 
3995   r = __kmp_reset_root(gtid, root);
3996   KC_TRACE(10,
3997            ("__kmp_unregister_root_other_thread: T#%d unregistered\n", gtid));
3998   return r;
3999 }
4000 #endif
4001 
4002 #if KMP_DEBUG
4003 void __kmp_task_info() {
4004 
4005   kmp_int32 gtid = __kmp_entry_gtid();
4006   kmp_int32 tid = __kmp_tid_from_gtid(gtid);
4007   kmp_info_t *this_thr = __kmp_threads[gtid];
4008   kmp_team_t *steam = this_thr->th.th_serial_team;
4009   kmp_team_t *team = this_thr->th.th_team;
4010 
4011   __kmp_printf(
4012       "__kmp_task_info: gtid=%d tid=%d t_thread=%p team=%p steam=%p curtask=%p "
4013       "ptask=%p\n",
4014       gtid, tid, this_thr, team, steam, this_thr->th.th_current_task,
4015       team->t.t_implicit_task_taskdata[tid].td_parent);
4016 }
4017 #endif // KMP_DEBUG
4018 
4019 /* TODO optimize with one big memclr, take out what isn't needed, split
4020    responsibility to workers as much as possible, and delay initialization of
4021    features as much as possible  */
4022 static void __kmp_initialize_info(kmp_info_t *this_thr, kmp_team_t *team,
4023                                   int tid, int gtid) {
4024   /* this_thr->th.th_info.ds.ds_gtid is setup in
4025      kmp_allocate_thread/create_worker.
4026      this_thr->th.th_serial_team is setup in __kmp_allocate_thread */
4027   kmp_info_t *master = team->t.t_threads[0];
4028   KMP_DEBUG_ASSERT(this_thr != NULL);
4029   KMP_DEBUG_ASSERT(this_thr->th.th_serial_team);
4030   KMP_DEBUG_ASSERT(team);
4031   KMP_DEBUG_ASSERT(team->t.t_threads);
4032   KMP_DEBUG_ASSERT(team->t.t_dispatch);
4033   KMP_DEBUG_ASSERT(master);
4034   KMP_DEBUG_ASSERT(master->th.th_root);
4035 
4036   KMP_MB();
4037 
4038   TCW_SYNC_PTR(this_thr->th.th_team, team);
4039 
4040   this_thr->th.th_info.ds.ds_tid = tid;
4041   this_thr->th.th_set_nproc = 0;
4042   if (__kmp_tasking_mode != tskm_immediate_exec)
4043     // When tasking is possible, threads are not safe to reap until they are
4044     // done tasking; this will be set when tasking code is exited in wait
4045     this_thr->th.th_reap_state = KMP_NOT_SAFE_TO_REAP;
4046   else // no tasking --> always safe to reap
4047     this_thr->th.th_reap_state = KMP_SAFE_TO_REAP;
4048 #if OMP_40_ENABLED
4049   this_thr->th.th_set_proc_bind = proc_bind_default;
4050 #if KMP_AFFINITY_SUPPORTED
4051   this_thr->th.th_new_place = this_thr->th.th_current_place;
4052 #endif
4053 #endif
4054   this_thr->th.th_root = master->th.th_root;
4055 
4056   /* setup the thread's cache of the team structure */
4057   this_thr->th.th_team_nproc = team->t.t_nproc;
4058   this_thr->th.th_team_master = master;
4059   this_thr->th.th_team_serialized = team->t.t_serialized;
4060   TCW_PTR(this_thr->th.th_sleep_loc, NULL);
4061 
4062   KMP_DEBUG_ASSERT(team->t.t_implicit_task_taskdata);
4063 
4064   KF_TRACE(10, ("__kmp_initialize_info1: T#%d:%d this_thread=%p curtask=%p\n",
4065                 tid, gtid, this_thr, this_thr->th.th_current_task));
4066 
4067   __kmp_init_implicit_task(this_thr->th.th_team_master->th.th_ident, this_thr,
4068                            team, tid, TRUE);
4069 
4070   KF_TRACE(10, ("__kmp_initialize_info2: T#%d:%d this_thread=%p curtask=%p\n",
4071                 tid, gtid, this_thr, this_thr->th.th_current_task));
4072   // TODO: Initialize ICVs from parent; GEH - isn't that already done in
4073   // __kmp_initialize_team()?
4074 
4075   /* TODO no worksharing in speculative threads */
4076   this_thr->th.th_dispatch = &team->t.t_dispatch[tid];
4077 
4078   this_thr->th.th_local.this_construct = 0;
4079 
4080   if (!this_thr->th.th_pri_common) {
4081     this_thr->th.th_pri_common =
4082         (struct common_table *)__kmp_allocate(sizeof(struct common_table));
4083     if (__kmp_storage_map) {
4084       __kmp_print_storage_map_gtid(
4085           gtid, this_thr->th.th_pri_common, this_thr->th.th_pri_common + 1,
4086           sizeof(struct common_table), "th_%d.th_pri_common\n", gtid);
4087     }
4088     this_thr->th.th_pri_head = NULL;
4089   }
4090 
4091   /* Initialize dynamic dispatch */
4092   {
4093     volatile kmp_disp_t *dispatch = this_thr->th.th_dispatch;
4094     // Use team max_nproc since this will never change for the team.
4095     size_t disp_size =
4096         sizeof(dispatch_private_info_t) *
4097         (team->t.t_max_nproc == 1 ? 1 : __kmp_dispatch_num_buffers);
4098     KD_TRACE(10, ("__kmp_initialize_info: T#%d max_nproc: %d\n", gtid,
4099                   team->t.t_max_nproc));
4100     KMP_ASSERT(dispatch);
4101     KMP_DEBUG_ASSERT(team->t.t_dispatch);
4102     KMP_DEBUG_ASSERT(dispatch == &team->t.t_dispatch[tid]);
4103 
4104     dispatch->th_disp_index = 0;
4105 #if OMP_45_ENABLED
4106     dispatch->th_doacross_buf_idx = 0;
4107 #endif
4108     if (!dispatch->th_disp_buffer) {
4109       dispatch->th_disp_buffer =
4110           (dispatch_private_info_t *)__kmp_allocate(disp_size);
4111 
4112       if (__kmp_storage_map) {
4113         __kmp_print_storage_map_gtid(
4114             gtid, &dispatch->th_disp_buffer[0],
4115             &dispatch->th_disp_buffer[team->t.t_max_nproc == 1
4116                                           ? 1
4117                                           : __kmp_dispatch_num_buffers],
4118             disp_size, "th_%d.th_dispatch.th_disp_buffer "
4119                        "(team_%d.t_dispatch[%d].th_disp_buffer)",
4120             gtid, team->t.t_id, gtid);
4121       }
4122     } else {
4123       memset(&dispatch->th_disp_buffer[0], '\0', disp_size);
4124     }
4125 
4126     dispatch->th_dispatch_pr_current = 0;
4127     dispatch->th_dispatch_sh_current = 0;
4128 
4129     dispatch->th_deo_fcn = 0; /* ORDERED     */
4130     dispatch->th_dxo_fcn = 0; /* END ORDERED */
4131   }
4132 
4133   this_thr->th.th_next_pool = NULL;
4134 
4135   if (!this_thr->th.th_task_state_memo_stack) {
4136     size_t i;
4137     this_thr->th.th_task_state_memo_stack =
4138         (kmp_uint8 *)__kmp_allocate(4 * sizeof(kmp_uint8));
4139     this_thr->th.th_task_state_top = 0;
4140     this_thr->th.th_task_state_stack_sz = 4;
4141     for (i = 0; i < this_thr->th.th_task_state_stack_sz;
4142          ++i) // zero init the stack
4143       this_thr->th.th_task_state_memo_stack[i] = 0;
4144   }
4145 
4146   KMP_DEBUG_ASSERT(!this_thr->th.th_spin_here);
4147   KMP_DEBUG_ASSERT(this_thr->th.th_next_waiting == 0);
4148 
4149   KMP_MB();
4150 }
4151 
4152 /* allocate a new thread for the requesting team. this is only called from
4153    within a forkjoin critical section. we will first try to get an available
4154    thread from the thread pool. if none is available, we will fork a new one
4155    assuming we are able to create a new one. this should be assured, as the
4156    caller should check on this first. */
4157 kmp_info_t *__kmp_allocate_thread(kmp_root_t *root, kmp_team_t *team,
4158                                   int new_tid) {
4159   kmp_team_t *serial_team;
4160   kmp_info_t *new_thr;
4161   int new_gtid;
4162 
4163   KA_TRACE(20, ("__kmp_allocate_thread: T#%d\n", __kmp_get_gtid()));
4164   KMP_DEBUG_ASSERT(root && team);
4165 #if !KMP_NESTED_HOT_TEAMS
4166   KMP_DEBUG_ASSERT(KMP_MASTER_GTID(__kmp_get_gtid()));
4167 #endif
4168   KMP_MB();
4169 
4170   /* first, try to get one from the thread pool */
4171   if (__kmp_thread_pool) {
4172 
4173     new_thr = CCAST(kmp_info_t *, __kmp_thread_pool);
4174     __kmp_thread_pool = (volatile kmp_info_t *)new_thr->th.th_next_pool;
4175     if (new_thr == __kmp_thread_pool_insert_pt) {
4176       __kmp_thread_pool_insert_pt = NULL;
4177     }
4178     TCW_4(new_thr->th.th_in_pool, FALSE);
4179     // Don't touch th_active_in_pool or th_active.
4180     // The worker thread adjusts those flags as it sleeps/awakens.
4181     __kmp_thread_pool_nth--;
4182 
4183     KA_TRACE(20, ("__kmp_allocate_thread: T#%d using thread T#%d\n",
4184                   __kmp_get_gtid(), new_thr->th.th_info.ds.ds_gtid));
4185     KMP_ASSERT(!new_thr->th.th_team);
4186     KMP_DEBUG_ASSERT(__kmp_nth < __kmp_threads_capacity);
4187     KMP_DEBUG_ASSERT(__kmp_thread_pool_nth >= 0);
4188 
4189     /* setup the thread structure */
4190     __kmp_initialize_info(new_thr, team, new_tid,
4191                           new_thr->th.th_info.ds.ds_gtid);
4192     KMP_DEBUG_ASSERT(new_thr->th.th_serial_team);
4193 
4194     TCW_4(__kmp_nth, __kmp_nth + 1);
4195     root->r.r_cg_nthreads++;
4196 
4197     new_thr->th.th_task_state = 0;
4198     new_thr->th.th_task_state_top = 0;
4199     new_thr->th.th_task_state_stack_sz = 4;
4200 
4201 #ifdef KMP_ADJUST_BLOCKTIME
4202     /* Adjust blocktime back to zero if necessary */
4203     /* Middle initialization might not have occurred yet */
4204     if (!__kmp_env_blocktime && (__kmp_avail_proc > 0)) {
4205       if (__kmp_nth > __kmp_avail_proc) {
4206         __kmp_zero_bt = TRUE;
4207       }
4208     }
4209 #endif /* KMP_ADJUST_BLOCKTIME */
4210 
4211 #if KMP_DEBUG
4212     // If thread entered pool via __kmp_free_thread, wait_flag should !=
4213     // KMP_BARRIER_PARENT_FLAG.
4214     int b;
4215     kmp_balign_t *balign = new_thr->th.th_bar;
4216     for (b = 0; b < bs_last_barrier; ++b)
4217       KMP_DEBUG_ASSERT(balign[b].bb.wait_flag != KMP_BARRIER_PARENT_FLAG);
4218 #endif
4219 
4220     KF_TRACE(10, ("__kmp_allocate_thread: T#%d using thread %p T#%d\n",
4221                   __kmp_get_gtid(), new_thr, new_thr->th.th_info.ds.ds_gtid));
4222 
4223     KMP_MB();
4224     return new_thr;
4225   }
4226 
4227   /* no, well fork a new one */
4228   KMP_ASSERT(__kmp_nth == __kmp_all_nth);
4229   KMP_ASSERT(__kmp_all_nth < __kmp_threads_capacity);
4230 
4231 #if KMP_USE_MONITOR
4232   // If this is the first worker thread the RTL is creating, then also
4233   // launch the monitor thread.  We try to do this as early as possible.
4234   if (!TCR_4(__kmp_init_monitor)) {
4235     __kmp_acquire_bootstrap_lock(&__kmp_monitor_lock);
4236     if (!TCR_4(__kmp_init_monitor)) {
4237       KF_TRACE(10, ("before __kmp_create_monitor\n"));
4238       TCW_4(__kmp_init_monitor, 1);
4239       __kmp_create_monitor(&__kmp_monitor);
4240       KF_TRACE(10, ("after __kmp_create_monitor\n"));
4241 #if KMP_OS_WINDOWS
4242       // AC: wait until monitor has started. This is a fix for CQ232808.
4243       // The reason is that if the library is loaded/unloaded in a loop with
4244       // small (parallel) work in between, then there is high probability that
4245       // monitor thread started after the library shutdown. At shutdown it is
4246       // too late to cope with the problem, because when the master is in
4247       // DllMain (process detach) the monitor has no chances to start (it is
4248       // blocked), and master has no means to inform the monitor that the
4249       // library has gone, because all the memory which the monitor can access
4250       // is going to be released/reset.
4251       while (TCR_4(__kmp_init_monitor) < 2) {
4252         KMP_YIELD(TRUE);
4253       }
4254       KF_TRACE(10, ("after monitor thread has started\n"));
4255 #endif
4256     }
4257     __kmp_release_bootstrap_lock(&__kmp_monitor_lock);
4258   }
4259 #endif
4260 
4261   KMP_MB();
4262   for (new_gtid = 1; TCR_PTR(__kmp_threads[new_gtid]) != NULL; ++new_gtid) {
4263     KMP_DEBUG_ASSERT(new_gtid < __kmp_threads_capacity);
4264   }
4265 
4266   /* allocate space for it. */
4267   new_thr = (kmp_info_t *)__kmp_allocate(sizeof(kmp_info_t));
4268 
4269   TCW_SYNC_PTR(__kmp_threads[new_gtid], new_thr);
4270 
4271   if (__kmp_storage_map) {
4272     __kmp_print_thread_storage_map(new_thr, new_gtid);
4273   }
4274 
4275   // add the reserve serialized team, initialized from the team's master thread
4276   {
4277     kmp_internal_control_t r_icvs = __kmp_get_x_global_icvs(team);
4278     KF_TRACE(10, ("__kmp_allocate_thread: before th_serial/serial_team\n"));
4279     new_thr->th.th_serial_team = serial_team =
4280         (kmp_team_t *)__kmp_allocate_team(root, 1, 1,
4281 #if OMPT_SUPPORT
4282                                           ompt_data_none, // root parallel id
4283 #endif
4284 #if OMP_40_ENABLED
4285                                           proc_bind_default,
4286 #endif
4287                                           &r_icvs, 0 USE_NESTED_HOT_ARG(NULL));
4288   }
4289   KMP_ASSERT(serial_team);
4290   serial_team->t.t_serialized = 0; // AC: the team created in reserve, not for
4291   // execution (it is unused for now).
4292   serial_team->t.t_threads[0] = new_thr;
4293   KF_TRACE(10,
4294            ("__kmp_allocate_thread: after th_serial/serial_team : new_thr=%p\n",
4295             new_thr));
4296 
4297   /* setup the thread structures */
4298   __kmp_initialize_info(new_thr, team, new_tid, new_gtid);
4299 
4300 #if USE_FAST_MEMORY
4301   __kmp_initialize_fast_memory(new_thr);
4302 #endif /* USE_FAST_MEMORY */
4303 
4304 #if KMP_USE_BGET
4305   KMP_DEBUG_ASSERT(new_thr->th.th_local.bget_data == NULL);
4306   __kmp_initialize_bget(new_thr);
4307 #endif
4308 
4309   __kmp_init_random(new_thr); // Initialize random number generator
4310 
4311   /* Initialize these only once when thread is grabbed for a team allocation */
4312   KA_TRACE(20,
4313            ("__kmp_allocate_thread: T#%d init go fork=%u, plain=%u\n",
4314             __kmp_get_gtid(), KMP_INIT_BARRIER_STATE, KMP_INIT_BARRIER_STATE));
4315 
4316   int b;
4317   kmp_balign_t *balign = new_thr->th.th_bar;
4318   for (b = 0; b < bs_last_barrier; ++b) {
4319     balign[b].bb.b_go = KMP_INIT_BARRIER_STATE;
4320     balign[b].bb.team = NULL;
4321     balign[b].bb.wait_flag = KMP_BARRIER_NOT_WAITING;
4322     balign[b].bb.use_oncore_barrier = 0;
4323   }
4324 
4325   new_thr->th.th_spin_here = FALSE;
4326   new_thr->th.th_next_waiting = 0;
4327 #if KMP_OS_UNIX
4328   new_thr->th.th_blocking = false;
4329 #endif
4330 
4331 #if OMP_40_ENABLED && KMP_AFFINITY_SUPPORTED
4332   new_thr->th.th_current_place = KMP_PLACE_UNDEFINED;
4333   new_thr->th.th_new_place = KMP_PLACE_UNDEFINED;
4334   new_thr->th.th_first_place = KMP_PLACE_UNDEFINED;
4335   new_thr->th.th_last_place = KMP_PLACE_UNDEFINED;
4336 #endif
4337 
4338   TCW_4(new_thr->th.th_in_pool, FALSE);
4339   new_thr->th.th_active_in_pool = FALSE;
4340   TCW_4(new_thr->th.th_active, TRUE);
4341 
4342   /* adjust the global counters */
4343   __kmp_all_nth++;
4344   __kmp_nth++;
4345 
4346   root->r.r_cg_nthreads++;
4347 
4348   // if __kmp_adjust_gtid_mode is set, then we use method #1 (sp search) for low
4349   // numbers of procs, and method #2 (keyed API call) for higher numbers.
4350   if (__kmp_adjust_gtid_mode) {
4351     if (__kmp_all_nth >= __kmp_tls_gtid_min) {
4352       if (TCR_4(__kmp_gtid_mode) != 2) {
4353         TCW_4(__kmp_gtid_mode, 2);
4354       }
4355     } else {
4356       if (TCR_4(__kmp_gtid_mode) != 1) {
4357         TCW_4(__kmp_gtid_mode, 1);
4358       }
4359     }
4360   }
4361 
4362 #ifdef KMP_ADJUST_BLOCKTIME
4363   /* Adjust blocktime back to zero if necessary       */
4364   /* Middle initialization might not have occurred yet */
4365   if (!__kmp_env_blocktime && (__kmp_avail_proc > 0)) {
4366     if (__kmp_nth > __kmp_avail_proc) {
4367       __kmp_zero_bt = TRUE;
4368     }
4369   }
4370 #endif /* KMP_ADJUST_BLOCKTIME */
4371 
4372   /* actually fork it and create the new worker thread */
4373   KF_TRACE(
4374       10, ("__kmp_allocate_thread: before __kmp_create_worker: %p\n", new_thr));
4375   __kmp_create_worker(new_gtid, new_thr, __kmp_stksize);
4376   KF_TRACE(10,
4377            ("__kmp_allocate_thread: after __kmp_create_worker: %p\n", new_thr));
4378 
4379   KA_TRACE(20, ("__kmp_allocate_thread: T#%d forked T#%d\n", __kmp_get_gtid(),
4380                 new_gtid));
4381   KMP_MB();
4382   return new_thr;
4383 }
4384 
4385 /* Reinitialize team for reuse.
4386    The hot team code calls this case at every fork barrier, so EPCC barrier
4387    test are extremely sensitive to changes in it, esp. writes to the team
4388    struct, which cause a cache invalidation in all threads.
4389    IF YOU TOUCH THIS ROUTINE, RUN EPCC C SYNCBENCH ON A BIG-IRON MACHINE!!! */
4390 static void __kmp_reinitialize_team(kmp_team_t *team,
4391                                     kmp_internal_control_t *new_icvs,
4392                                     ident_t *loc) {
4393   KF_TRACE(10, ("__kmp_reinitialize_team: enter this_thread=%p team=%p\n",
4394                 team->t.t_threads[0], team));
4395   KMP_DEBUG_ASSERT(team && new_icvs);
4396   KMP_DEBUG_ASSERT((!TCR_4(__kmp_init_parallel)) || new_icvs->nproc);
4397   KMP_CHECK_UPDATE(team->t.t_ident, loc);
4398 
4399   KMP_CHECK_UPDATE(team->t.t_id, KMP_GEN_TEAM_ID());
4400   // Copy ICVs to the master thread's implicit taskdata
4401   __kmp_init_implicit_task(loc, team->t.t_threads[0], team, 0, FALSE);
4402   copy_icvs(&team->t.t_implicit_task_taskdata[0].td_icvs, new_icvs);
4403 
4404   KF_TRACE(10, ("__kmp_reinitialize_team: exit this_thread=%p team=%p\n",
4405                 team->t.t_threads[0], team));
4406 }
4407 
4408 /* Initialize the team data structure.
4409    This assumes the t_threads and t_max_nproc are already set.
4410    Also, we don't touch the arguments */
4411 static void __kmp_initialize_team(kmp_team_t *team, int new_nproc,
4412                                   kmp_internal_control_t *new_icvs,
4413                                   ident_t *loc) {
4414   KF_TRACE(10, ("__kmp_initialize_team: enter: team=%p\n", team));
4415 
4416   /* verify */
4417   KMP_DEBUG_ASSERT(team);
4418   KMP_DEBUG_ASSERT(new_nproc <= team->t.t_max_nproc);
4419   KMP_DEBUG_ASSERT(team->t.t_threads);
4420   KMP_MB();
4421 
4422   team->t.t_master_tid = 0; /* not needed */
4423   /* team->t.t_master_bar;        not needed */
4424   team->t.t_serialized = new_nproc > 1 ? 0 : 1;
4425   team->t.t_nproc = new_nproc;
4426 
4427   /* team->t.t_parent     = NULL; TODO not needed & would mess up hot team */
4428   team->t.t_next_pool = NULL;
4429   /* memset( team->t.t_threads, 0, sizeof(kmp_info_t*)*new_nproc ); would mess
4430    * up hot team */
4431 
4432   TCW_SYNC_PTR(team->t.t_pkfn, NULL); /* not needed */
4433   team->t.t_invoke = NULL; /* not needed */
4434 
4435   // TODO???: team->t.t_max_active_levels       = new_max_active_levels;
4436   team->t.t_sched.sched = new_icvs->sched.sched;
4437 
4438 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
4439   team->t.t_fp_control_saved = FALSE; /* not needed */
4440   team->t.t_x87_fpu_control_word = 0; /* not needed */
4441   team->t.t_mxcsr = 0; /* not needed */
4442 #endif /* KMP_ARCH_X86 || KMP_ARCH_X86_64 */
4443 
4444   team->t.t_construct = 0;
4445 
4446   team->t.t_ordered.dt.t_value = 0;
4447   team->t.t_master_active = FALSE;
4448 
4449   memset(&team->t.t_taskq, '\0', sizeof(kmp_taskq_t));
4450 
4451 #ifdef KMP_DEBUG
4452   team->t.t_copypriv_data = NULL; /* not necessary, but nice for debugging */
4453 #endif
4454 #if KMP_OS_WINDOWS
4455   team->t.t_copyin_counter = 0; /* for barrier-free copyin implementation */
4456 #endif
4457 
4458   team->t.t_control_stack_top = NULL;
4459 
4460   __kmp_reinitialize_team(team, new_icvs, loc);
4461 
4462   KMP_MB();
4463   KF_TRACE(10, ("__kmp_initialize_team: exit: team=%p\n", team));
4464 }
4465 
4466 #if KMP_OS_LINUX && KMP_AFFINITY_SUPPORTED
4467 /* Sets full mask for thread and returns old mask, no changes to structures. */
4468 static void
4469 __kmp_set_thread_affinity_mask_full_tmp(kmp_affin_mask_t *old_mask) {
4470   if (KMP_AFFINITY_CAPABLE()) {
4471     int status;
4472     if (old_mask != NULL) {
4473       status = __kmp_get_system_affinity(old_mask, TRUE);
4474       int error = errno;
4475       if (status != 0) {
4476         __kmp_fatal(KMP_MSG(ChangeThreadAffMaskError), KMP_ERR(error),
4477                     __kmp_msg_null);
4478       }
4479     }
4480     __kmp_set_system_affinity(__kmp_affin_fullMask, TRUE);
4481   }
4482 }
4483 #endif
4484 
4485 #if OMP_40_ENABLED && KMP_AFFINITY_SUPPORTED
4486 
4487 // __kmp_partition_places() is the heart of the OpenMP 4.0 affinity mechanism.
4488 // It calculats the worker + master thread's partition based upon the parent
4489 // thread's partition, and binds each worker to a thread in their partition.
4490 // The master thread's partition should already include its current binding.
4491 static void __kmp_partition_places(kmp_team_t *team, int update_master_only) {
4492   // Copy the master thread's place partion to the team struct
4493   kmp_info_t *master_th = team->t.t_threads[0];
4494   KMP_DEBUG_ASSERT(master_th != NULL);
4495   kmp_proc_bind_t proc_bind = team->t.t_proc_bind;
4496   int first_place = master_th->th.th_first_place;
4497   int last_place = master_th->th.th_last_place;
4498   int masters_place = master_th->th.th_current_place;
4499   team->t.t_first_place = first_place;
4500   team->t.t_last_place = last_place;
4501 
4502   KA_TRACE(20, ("__kmp_partition_places: enter: proc_bind = %d T#%d(%d:0) "
4503                 "bound to place %d partition = [%d,%d]\n",
4504                 proc_bind, __kmp_gtid_from_thread(team->t.t_threads[0]),
4505                 team->t.t_id, masters_place, first_place, last_place));
4506 
4507   switch (proc_bind) {
4508 
4509   case proc_bind_default:
4510     // serial teams might have the proc_bind policy set to proc_bind_default. It
4511     // doesn't matter, as we don't rebind master thread for any proc_bind policy
4512     KMP_DEBUG_ASSERT(team->t.t_nproc == 1);
4513     break;
4514 
4515   case proc_bind_master: {
4516     int f;
4517     int n_th = team->t.t_nproc;
4518     for (f = 1; f < n_th; f++) {
4519       kmp_info_t *th = team->t.t_threads[f];
4520       KMP_DEBUG_ASSERT(th != NULL);
4521       th->th.th_first_place = first_place;
4522       th->th.th_last_place = last_place;
4523       th->th.th_new_place = masters_place;
4524 
4525       KA_TRACE(100, ("__kmp_partition_places: master: T#%d(%d:%d) place %d "
4526                      "partition = [%d,%d]\n",
4527                      __kmp_gtid_from_thread(team->t.t_threads[f]), team->t.t_id,
4528                      f, masters_place, first_place, last_place));
4529     }
4530   } break;
4531 
4532   case proc_bind_close: {
4533     int f;
4534     int n_th = team->t.t_nproc;
4535     int n_places;
4536     if (first_place <= last_place) {
4537       n_places = last_place - first_place + 1;
4538     } else {
4539       n_places = __kmp_affinity_num_masks - first_place + last_place + 1;
4540     }
4541     if (n_th <= n_places) {
4542       int place = masters_place;
4543       for (f = 1; f < n_th; f++) {
4544         kmp_info_t *th = team->t.t_threads[f];
4545         KMP_DEBUG_ASSERT(th != NULL);
4546 
4547         if (place == last_place) {
4548           place = first_place;
4549         } else if (place == (int)(__kmp_affinity_num_masks - 1)) {
4550           place = 0;
4551         } else {
4552           place++;
4553         }
4554         th->th.th_first_place = first_place;
4555         th->th.th_last_place = last_place;
4556         th->th.th_new_place = place;
4557 
4558         KA_TRACE(100, ("__kmp_partition_places: close: T#%d(%d:%d) place %d "
4559                        "partition = [%d,%d]\n",
4560                        __kmp_gtid_from_thread(team->t.t_threads[f]),
4561                        team->t.t_id, f, place, first_place, last_place));
4562       }
4563     } else {
4564       int S, rem, gap, s_count;
4565       S = n_th / n_places;
4566       s_count = 0;
4567       rem = n_th - (S * n_places);
4568       gap = rem > 0 ? n_places / rem : n_places;
4569       int place = masters_place;
4570       int gap_ct = gap;
4571       for (f = 0; f < n_th; f++) {
4572         kmp_info_t *th = team->t.t_threads[f];
4573         KMP_DEBUG_ASSERT(th != NULL);
4574 
4575         th->th.th_first_place = first_place;
4576         th->th.th_last_place = last_place;
4577         th->th.th_new_place = place;
4578         s_count++;
4579 
4580         if ((s_count == S) && rem && (gap_ct == gap)) {
4581           // do nothing, add an extra thread to place on next iteration
4582         } else if ((s_count == S + 1) && rem && (gap_ct == gap)) {
4583           // we added an extra thread to this place; move to next place
4584           if (place == last_place) {
4585             place = first_place;
4586           } else if (place == (int)(__kmp_affinity_num_masks - 1)) {
4587             place = 0;
4588           } else {
4589             place++;
4590           }
4591           s_count = 0;
4592           gap_ct = 1;
4593           rem--;
4594         } else if (s_count == S) { // place full; don't add extra
4595           if (place == last_place) {
4596             place = first_place;
4597           } else if (place == (int)(__kmp_affinity_num_masks - 1)) {
4598             place = 0;
4599           } else {
4600             place++;
4601           }
4602           gap_ct++;
4603           s_count = 0;
4604         }
4605 
4606         KA_TRACE(100,
4607                  ("__kmp_partition_places: close: T#%d(%d:%d) place %d "
4608                   "partition = [%d,%d]\n",
4609                   __kmp_gtid_from_thread(team->t.t_threads[f]), team->t.t_id, f,
4610                   th->th.th_new_place, first_place, last_place));
4611       }
4612       KMP_DEBUG_ASSERT(place == masters_place);
4613     }
4614   } break;
4615 
4616   case proc_bind_spread: {
4617     int f;
4618     int n_th = team->t.t_nproc;
4619     int n_places;
4620     int thidx;
4621     if (first_place <= last_place) {
4622       n_places = last_place - first_place + 1;
4623     } else {
4624       n_places = __kmp_affinity_num_masks - first_place + last_place + 1;
4625     }
4626     if (n_th <= n_places) {
4627       int place = -1;
4628 
4629       if (n_places != static_cast<int>(__kmp_affinity_num_masks)) {
4630         int S = n_places / n_th;
4631         int s_count, rem, gap, gap_ct;
4632 
4633         place = masters_place;
4634         rem = n_places - n_th * S;
4635         gap = rem ? n_th / rem : 1;
4636         gap_ct = gap;
4637         thidx = n_th;
4638         if (update_master_only == 1)
4639           thidx = 1;
4640         for (f = 0; f < thidx; f++) {
4641           kmp_info_t *th = team->t.t_threads[f];
4642           KMP_DEBUG_ASSERT(th != NULL);
4643 
4644           th->th.th_first_place = place;
4645           th->th.th_new_place = place;
4646           s_count = 1;
4647           while (s_count < S) {
4648             if (place == last_place) {
4649               place = first_place;
4650             } else if (place == (int)(__kmp_affinity_num_masks - 1)) {
4651               place = 0;
4652             } else {
4653               place++;
4654             }
4655             s_count++;
4656           }
4657           if (rem && (gap_ct == gap)) {
4658             if (place == last_place) {
4659               place = first_place;
4660             } else if (place == (int)(__kmp_affinity_num_masks - 1)) {
4661               place = 0;
4662             } else {
4663               place++;
4664             }
4665             rem--;
4666             gap_ct = 0;
4667           }
4668           th->th.th_last_place = place;
4669           gap_ct++;
4670 
4671           if (place == last_place) {
4672             place = first_place;
4673           } else if (place == (int)(__kmp_affinity_num_masks - 1)) {
4674             place = 0;
4675           } else {
4676             place++;
4677           }
4678 
4679           KA_TRACE(100,
4680                    ("__kmp_partition_places: spread: T#%d(%d:%d) place %d "
4681                     "partition = [%d,%d], __kmp_affinity_num_masks: %u\n",
4682                     __kmp_gtid_from_thread(team->t.t_threads[f]), team->t.t_id,
4683                     f, th->th.th_new_place, th->th.th_first_place,
4684                     th->th.th_last_place, __kmp_affinity_num_masks));
4685         }
4686       } else {
4687         /* Having uniform space of available computation places I can create
4688            T partitions of round(P/T) size and put threads into the first
4689            place of each partition. */
4690         double current = static_cast<double>(masters_place);
4691         double spacing =
4692             (static_cast<double>(n_places + 1) / static_cast<double>(n_th));
4693         int first, last;
4694         kmp_info_t *th;
4695 
4696         thidx = n_th + 1;
4697         if (update_master_only == 1)
4698           thidx = 1;
4699         for (f = 0; f < thidx; f++) {
4700           first = static_cast<int>(current);
4701           last = static_cast<int>(current + spacing) - 1;
4702           KMP_DEBUG_ASSERT(last >= first);
4703           if (first >= n_places) {
4704             if (masters_place) {
4705               first -= n_places;
4706               last -= n_places;
4707               if (first == (masters_place + 1)) {
4708                 KMP_DEBUG_ASSERT(f == n_th);
4709                 first--;
4710               }
4711               if (last == masters_place) {
4712                 KMP_DEBUG_ASSERT(f == (n_th - 1));
4713                 last--;
4714               }
4715             } else {
4716               KMP_DEBUG_ASSERT(f == n_th);
4717               first = 0;
4718               last = 0;
4719             }
4720           }
4721           if (last >= n_places) {
4722             last = (n_places - 1);
4723           }
4724           place = first;
4725           current += spacing;
4726           if (f < n_th) {
4727             KMP_DEBUG_ASSERT(0 <= first);
4728             KMP_DEBUG_ASSERT(n_places > first);
4729             KMP_DEBUG_ASSERT(0 <= last);
4730             KMP_DEBUG_ASSERT(n_places > last);
4731             KMP_DEBUG_ASSERT(last_place >= first_place);
4732             th = team->t.t_threads[f];
4733             KMP_DEBUG_ASSERT(th);
4734             th->th.th_first_place = first;
4735             th->th.th_new_place = place;
4736             th->th.th_last_place = last;
4737 
4738             KA_TRACE(100,
4739                      ("__kmp_partition_places: spread: T#%d(%d:%d) place %d "
4740                       "partition = [%d,%d], spacing = %.4f\n",
4741                       __kmp_gtid_from_thread(team->t.t_threads[f]),
4742                       team->t.t_id, f, th->th.th_new_place,
4743                       th->th.th_first_place, th->th.th_last_place, spacing));
4744           }
4745         }
4746       }
4747       KMP_DEBUG_ASSERT(update_master_only || place == masters_place);
4748     } else {
4749       int S, rem, gap, s_count;
4750       S = n_th / n_places;
4751       s_count = 0;
4752       rem = n_th - (S * n_places);
4753       gap = rem > 0 ? n_places / rem : n_places;
4754       int place = masters_place;
4755       int gap_ct = gap;
4756       thidx = n_th;
4757       if (update_master_only == 1)
4758         thidx = 1;
4759       for (f = 0; f < thidx; f++) {
4760         kmp_info_t *th = team->t.t_threads[f];
4761         KMP_DEBUG_ASSERT(th != NULL);
4762 
4763         th->th.th_first_place = place;
4764         th->th.th_last_place = place;
4765         th->th.th_new_place = place;
4766         s_count++;
4767 
4768         if ((s_count == S) && rem && (gap_ct == gap)) {
4769           // do nothing, add an extra thread to place on next iteration
4770         } else if ((s_count == S + 1) && rem && (gap_ct == gap)) {
4771           // we added an extra thread to this place; move on to next place
4772           if (place == last_place) {
4773             place = first_place;
4774           } else if (place == (int)(__kmp_affinity_num_masks - 1)) {
4775             place = 0;
4776           } else {
4777             place++;
4778           }
4779           s_count = 0;
4780           gap_ct = 1;
4781           rem--;
4782         } else if (s_count == S) { // place is full; don't add extra thread
4783           if (place == last_place) {
4784             place = first_place;
4785           } else if (place == (int)(__kmp_affinity_num_masks - 1)) {
4786             place = 0;
4787           } else {
4788             place++;
4789           }
4790           gap_ct++;
4791           s_count = 0;
4792         }
4793 
4794         KA_TRACE(100, ("__kmp_partition_places: spread: T#%d(%d:%d) place %d "
4795                        "partition = [%d,%d]\n",
4796                        __kmp_gtid_from_thread(team->t.t_threads[f]),
4797                        team->t.t_id, f, th->th.th_new_place,
4798                        th->th.th_first_place, th->th.th_last_place));
4799       }
4800       KMP_DEBUG_ASSERT(update_master_only || place == masters_place);
4801     }
4802   } break;
4803 
4804   default:
4805     break;
4806   }
4807 
4808   KA_TRACE(20, ("__kmp_partition_places: exit T#%d\n", team->t.t_id));
4809 }
4810 
4811 #endif /* OMP_40_ENABLED && KMP_AFFINITY_SUPPORTED */
4812 
4813 /* allocate a new team data structure to use.  take one off of the free pool if
4814    available */
4815 kmp_team_t *
4816 __kmp_allocate_team(kmp_root_t *root, int new_nproc, int max_nproc,
4817 #if OMPT_SUPPORT
4818                     ompt_data_t ompt_parallel_data,
4819 #endif
4820 #if OMP_40_ENABLED
4821                     kmp_proc_bind_t new_proc_bind,
4822 #endif
4823                     kmp_internal_control_t *new_icvs,
4824                     int argc USE_NESTED_HOT_ARG(kmp_info_t *master)) {
4825   KMP_TIME_DEVELOPER_PARTITIONED_BLOCK(KMP_allocate_team);
4826   int f;
4827   kmp_team_t *team;
4828   int use_hot_team = !root->r.r_active;
4829   int level = 0;
4830 
4831   KA_TRACE(20, ("__kmp_allocate_team: called\n"));
4832   KMP_DEBUG_ASSERT(new_nproc >= 1 && argc >= 0);
4833   KMP_DEBUG_ASSERT(max_nproc >= new_nproc);
4834   KMP_MB();
4835 
4836 #if KMP_NESTED_HOT_TEAMS
4837   kmp_hot_team_ptr_t *hot_teams;
4838   if (master) {
4839     team = master->th.th_team;
4840     level = team->t.t_active_level;
4841     if (master->th.th_teams_microtask) { // in teams construct?
4842       if (master->th.th_teams_size.nteams > 1 &&
4843           ( // #teams > 1
4844               team->t.t_pkfn ==
4845                   (microtask_t)__kmp_teams_master || // inner fork of the teams
4846               master->th.th_teams_level <
4847                   team->t.t_level)) { // or nested parallel inside the teams
4848         ++level; // not increment if #teams==1, or for outer fork of the teams;
4849         // increment otherwise
4850       }
4851     }
4852     hot_teams = master->th.th_hot_teams;
4853     if (level < __kmp_hot_teams_max_level && hot_teams &&
4854         hot_teams[level]
4855             .hot_team) { // hot team has already been allocated for given level
4856       use_hot_team = 1;
4857     } else {
4858       use_hot_team = 0;
4859     }
4860   }
4861 #endif
4862   // Optimization to use a "hot" team
4863   if (use_hot_team && new_nproc > 1) {
4864     KMP_DEBUG_ASSERT(new_nproc == max_nproc);
4865 #if KMP_NESTED_HOT_TEAMS
4866     team = hot_teams[level].hot_team;
4867 #else
4868     team = root->r.r_hot_team;
4869 #endif
4870 #if KMP_DEBUG
4871     if (__kmp_tasking_mode != tskm_immediate_exec) {
4872       KA_TRACE(20, ("__kmp_allocate_team: hot team task_team[0] = %p "
4873                     "task_team[1] = %p before reinit\n",
4874                     team->t.t_task_team[0], team->t.t_task_team[1]));
4875     }
4876 #endif
4877 
4878     // Has the number of threads changed?
4879     /* Let's assume the most common case is that the number of threads is
4880        unchanged, and put that case first. */
4881     if (team->t.t_nproc == new_nproc) { // Check changes in number of threads
4882       KA_TRACE(20, ("__kmp_allocate_team: reusing hot team\n"));
4883       // This case can mean that omp_set_num_threads() was called and the hot
4884       // team size was already reduced, so we check the special flag
4885       if (team->t.t_size_changed == -1) {
4886         team->t.t_size_changed = 1;
4887       } else {
4888         KMP_CHECK_UPDATE(team->t.t_size_changed, 0);
4889       }
4890 
4891       // TODO???: team->t.t_max_active_levels = new_max_active_levels;
4892       kmp_r_sched_t new_sched = new_icvs->sched;
4893       // set master's schedule as new run-time schedule
4894       KMP_CHECK_UPDATE(team->t.t_sched.sched, new_sched.sched);
4895 
4896       __kmp_reinitialize_team(team, new_icvs,
4897                               root->r.r_uber_thread->th.th_ident);
4898 
4899       KF_TRACE(10, ("__kmp_allocate_team2: T#%d, this_thread=%p team=%p\n", 0,
4900                     team->t.t_threads[0], team));
4901       __kmp_push_current_task_to_thread(team->t.t_threads[0], team, 0);
4902 
4903 #if OMP_40_ENABLED
4904 #if KMP_AFFINITY_SUPPORTED
4905       if ((team->t.t_size_changed == 0) &&
4906           (team->t.t_proc_bind == new_proc_bind)) {
4907         if (new_proc_bind == proc_bind_spread) {
4908           __kmp_partition_places(
4909               team, 1); // add flag to update only master for spread
4910         }
4911         KA_TRACE(200, ("__kmp_allocate_team: reusing hot team #%d bindings: "
4912                        "proc_bind = %d, partition = [%d,%d]\n",
4913                        team->t.t_id, new_proc_bind, team->t.t_first_place,
4914                        team->t.t_last_place));
4915       } else {
4916         KMP_CHECK_UPDATE(team->t.t_proc_bind, new_proc_bind);
4917         __kmp_partition_places(team);
4918       }
4919 #else
4920       KMP_CHECK_UPDATE(team->t.t_proc_bind, new_proc_bind);
4921 #endif /* KMP_AFFINITY_SUPPORTED */
4922 #endif /* OMP_40_ENABLED */
4923     } else if (team->t.t_nproc > new_nproc) {
4924       KA_TRACE(20,
4925                ("__kmp_allocate_team: decreasing hot team thread count to %d\n",
4926                 new_nproc));
4927 
4928       team->t.t_size_changed = 1;
4929 #if KMP_NESTED_HOT_TEAMS
4930       if (__kmp_hot_teams_mode == 0) {
4931         // AC: saved number of threads should correspond to team's value in this
4932         // mode, can be bigger in mode 1, when hot team has threads in reserve
4933         KMP_DEBUG_ASSERT(hot_teams[level].hot_team_nth == team->t.t_nproc);
4934         hot_teams[level].hot_team_nth = new_nproc;
4935 #endif // KMP_NESTED_HOT_TEAMS
4936         /* release the extra threads we don't need any more */
4937         for (f = new_nproc; f < team->t.t_nproc; f++) {
4938           KMP_DEBUG_ASSERT(team->t.t_threads[f]);
4939           if (__kmp_tasking_mode != tskm_immediate_exec) {
4940             // When decreasing team size, threads no longer in the team should
4941             // unref task team.
4942             team->t.t_threads[f]->th.th_task_team = NULL;
4943           }
4944           __kmp_free_thread(team->t.t_threads[f]);
4945           team->t.t_threads[f] = NULL;
4946         }
4947 #if KMP_NESTED_HOT_TEAMS
4948       } // (__kmp_hot_teams_mode == 0)
4949       else {
4950         // When keeping extra threads in team, switch threads to wait on own
4951         // b_go flag
4952         for (f = new_nproc; f < team->t.t_nproc; ++f) {
4953           KMP_DEBUG_ASSERT(team->t.t_threads[f]);
4954           kmp_balign_t *balign = team->t.t_threads[f]->th.th_bar;
4955           for (int b = 0; b < bs_last_barrier; ++b) {
4956             if (balign[b].bb.wait_flag == KMP_BARRIER_PARENT_FLAG) {
4957               balign[b].bb.wait_flag = KMP_BARRIER_SWITCH_TO_OWN_FLAG;
4958             }
4959             KMP_CHECK_UPDATE(balign[b].bb.leaf_kids, 0);
4960           }
4961         }
4962       }
4963 #endif // KMP_NESTED_HOT_TEAMS
4964       team->t.t_nproc = new_nproc;
4965       // TODO???: team->t.t_max_active_levels = new_max_active_levels;
4966       KMP_CHECK_UPDATE(team->t.t_sched.sched, new_icvs->sched.sched);
4967       __kmp_reinitialize_team(team, new_icvs,
4968                               root->r.r_uber_thread->th.th_ident);
4969 
4970       /* update the remaining threads */
4971       for (f = 0; f < new_nproc; ++f) {
4972         team->t.t_threads[f]->th.th_team_nproc = new_nproc;
4973       }
4974       // restore the current task state of the master thread: should be the
4975       // implicit task
4976       KF_TRACE(10, ("__kmp_allocate_team: T#%d, this_thread=%p team=%p\n", 0,
4977                     team->t.t_threads[0], team));
4978 
4979       __kmp_push_current_task_to_thread(team->t.t_threads[0], team, 0);
4980 
4981 #ifdef KMP_DEBUG
4982       for (f = 0; f < team->t.t_nproc; f++) {
4983         KMP_DEBUG_ASSERT(team->t.t_threads[f] &&
4984                          team->t.t_threads[f]->th.th_team_nproc ==
4985                              team->t.t_nproc);
4986       }
4987 #endif
4988 
4989 #if OMP_40_ENABLED
4990       KMP_CHECK_UPDATE(team->t.t_proc_bind, new_proc_bind);
4991 #if KMP_AFFINITY_SUPPORTED
4992       __kmp_partition_places(team);
4993 #endif
4994 #endif
4995     } else { // team->t.t_nproc < new_nproc
4996 #if KMP_OS_LINUX && KMP_AFFINITY_SUPPORTED
4997       kmp_affin_mask_t *old_mask;
4998       if (KMP_AFFINITY_CAPABLE()) {
4999         KMP_CPU_ALLOC(old_mask);
5000       }
5001 #endif
5002 
5003       KA_TRACE(20,
5004                ("__kmp_allocate_team: increasing hot team thread count to %d\n",
5005                 new_nproc));
5006 
5007       team->t.t_size_changed = 1;
5008 
5009 #if KMP_NESTED_HOT_TEAMS
5010       int avail_threads = hot_teams[level].hot_team_nth;
5011       if (new_nproc < avail_threads)
5012         avail_threads = new_nproc;
5013       kmp_info_t **other_threads = team->t.t_threads;
5014       for (f = team->t.t_nproc; f < avail_threads; ++f) {
5015         // Adjust barrier data of reserved threads (if any) of the team
5016         // Other data will be set in __kmp_initialize_info() below.
5017         int b;
5018         kmp_balign_t *balign = other_threads[f]->th.th_bar;
5019         for (b = 0; b < bs_last_barrier; ++b) {
5020           balign[b].bb.b_arrived = team->t.t_bar[b].b_arrived;
5021           KMP_DEBUG_ASSERT(balign[b].bb.wait_flag != KMP_BARRIER_PARENT_FLAG);
5022 #if USE_DEBUGGER
5023           balign[b].bb.b_worker_arrived = team->t.t_bar[b].b_team_arrived;
5024 #endif
5025         }
5026       }
5027       if (hot_teams[level].hot_team_nth >= new_nproc) {
5028         // we have all needed threads in reserve, no need to allocate any
5029         // this only possible in mode 1, cannot have reserved threads in mode 0
5030         KMP_DEBUG_ASSERT(__kmp_hot_teams_mode == 1);
5031         team->t.t_nproc = new_nproc; // just get reserved threads involved
5032       } else {
5033         // we may have some threads in reserve, but not enough
5034         team->t.t_nproc =
5035             hot_teams[level]
5036                 .hot_team_nth; // get reserved threads involved if any
5037         hot_teams[level].hot_team_nth = new_nproc; // adjust hot team max size
5038 #endif // KMP_NESTED_HOT_TEAMS
5039         if (team->t.t_max_nproc < new_nproc) {
5040           /* reallocate larger arrays */
5041           __kmp_reallocate_team_arrays(team, new_nproc);
5042           __kmp_reinitialize_team(team, new_icvs, NULL);
5043         }
5044 
5045 #if KMP_OS_LINUX && KMP_AFFINITY_SUPPORTED
5046         /* Temporarily set full mask for master thread before creation of
5047            workers. The reason is that workers inherit the affinity from master,
5048            so if a lot of workers are created on the single core quickly, they
5049            don't get a chance to set their own affinity for a long time. */
5050         __kmp_set_thread_affinity_mask_full_tmp(old_mask);
5051 #endif
5052 
5053         /* allocate new threads for the hot team */
5054         for (f = team->t.t_nproc; f < new_nproc; f++) {
5055           kmp_info_t *new_worker = __kmp_allocate_thread(root, team, f);
5056           KMP_DEBUG_ASSERT(new_worker);
5057           team->t.t_threads[f] = new_worker;
5058 
5059           KA_TRACE(20,
5060                    ("__kmp_allocate_team: team %d init T#%d arrived: "
5061                     "join=%llu, plain=%llu\n",
5062                     team->t.t_id, __kmp_gtid_from_tid(f, team), team->t.t_id, f,
5063                     team->t.t_bar[bs_forkjoin_barrier].b_arrived,
5064                     team->t.t_bar[bs_plain_barrier].b_arrived));
5065 
5066           { // Initialize barrier data for new threads.
5067             int b;
5068             kmp_balign_t *balign = new_worker->th.th_bar;
5069             for (b = 0; b < bs_last_barrier; ++b) {
5070               balign[b].bb.b_arrived = team->t.t_bar[b].b_arrived;
5071               KMP_DEBUG_ASSERT(balign[b].bb.wait_flag !=
5072                                KMP_BARRIER_PARENT_FLAG);
5073 #if USE_DEBUGGER
5074               balign[b].bb.b_worker_arrived = team->t.t_bar[b].b_team_arrived;
5075 #endif
5076             }
5077           }
5078         }
5079 
5080 #if KMP_OS_LINUX && KMP_AFFINITY_SUPPORTED
5081         if (KMP_AFFINITY_CAPABLE()) {
5082           /* Restore initial master thread's affinity mask */
5083           __kmp_set_system_affinity(old_mask, TRUE);
5084           KMP_CPU_FREE(old_mask);
5085         }
5086 #endif
5087 #if KMP_NESTED_HOT_TEAMS
5088       } // end of check of t_nproc vs. new_nproc vs. hot_team_nth
5089 #endif // KMP_NESTED_HOT_TEAMS
5090       /* make sure everyone is syncronized */
5091       int old_nproc = team->t.t_nproc; // save old value and use to update only
5092       // new threads below
5093       __kmp_initialize_team(team, new_nproc, new_icvs,
5094                             root->r.r_uber_thread->th.th_ident);
5095 
5096       /* reinitialize the threads */
5097       KMP_DEBUG_ASSERT(team->t.t_nproc == new_nproc);
5098       for (f = 0; f < team->t.t_nproc; ++f)
5099         __kmp_initialize_info(team->t.t_threads[f], team, f,
5100                               __kmp_gtid_from_tid(f, team));
5101       if (level) { // set th_task_state for new threads in nested hot team
5102         // __kmp_initialize_info() no longer zeroes th_task_state, so we should
5103         // only need to set the th_task_state for the new threads. th_task_state
5104         // for master thread will not be accurate until after this in
5105         // __kmp_fork_call(), so we look to the master's memo_stack to get the
5106         // correct value.
5107         for (f = old_nproc; f < team->t.t_nproc; ++f)
5108           team->t.t_threads[f]->th.th_task_state =
5109               team->t.t_threads[0]->th.th_task_state_memo_stack[level];
5110       } else { // set th_task_state for new threads in non-nested hot team
5111         int old_state =
5112             team->t.t_threads[0]->th.th_task_state; // copy master's state
5113         for (f = old_nproc; f < team->t.t_nproc; ++f)
5114           team->t.t_threads[f]->th.th_task_state = old_state;
5115       }
5116 
5117 #ifdef KMP_DEBUG
5118       for (f = 0; f < team->t.t_nproc; ++f) {
5119         KMP_DEBUG_ASSERT(team->t.t_threads[f] &&
5120                          team->t.t_threads[f]->th.th_team_nproc ==
5121                              team->t.t_nproc);
5122       }
5123 #endif
5124 
5125 #if OMP_40_ENABLED
5126       KMP_CHECK_UPDATE(team->t.t_proc_bind, new_proc_bind);
5127 #if KMP_AFFINITY_SUPPORTED
5128       __kmp_partition_places(team);
5129 #endif
5130 #endif
5131     } // Check changes in number of threads
5132 
5133 #if OMP_40_ENABLED
5134     kmp_info_t *master = team->t.t_threads[0];
5135     if (master->th.th_teams_microtask) {
5136       for (f = 1; f < new_nproc; ++f) {
5137         // propagate teams construct specific info to workers
5138         kmp_info_t *thr = team->t.t_threads[f];
5139         thr->th.th_teams_microtask = master->th.th_teams_microtask;
5140         thr->th.th_teams_level = master->th.th_teams_level;
5141         thr->th.th_teams_size = master->th.th_teams_size;
5142       }
5143     }
5144 #endif /* OMP_40_ENABLED */
5145 #if KMP_NESTED_HOT_TEAMS
5146     if (level) {
5147       // Sync barrier state for nested hot teams, not needed for outermost hot
5148       // team.
5149       for (f = 1; f < new_nproc; ++f) {
5150         kmp_info_t *thr = team->t.t_threads[f];
5151         int b;
5152         kmp_balign_t *balign = thr->th.th_bar;
5153         for (b = 0; b < bs_last_barrier; ++b) {
5154           balign[b].bb.b_arrived = team->t.t_bar[b].b_arrived;
5155           KMP_DEBUG_ASSERT(balign[b].bb.wait_flag != KMP_BARRIER_PARENT_FLAG);
5156 #if USE_DEBUGGER
5157           balign[b].bb.b_worker_arrived = team->t.t_bar[b].b_team_arrived;
5158 #endif
5159         }
5160       }
5161     }
5162 #endif // KMP_NESTED_HOT_TEAMS
5163 
5164     /* reallocate space for arguments if necessary */
5165     __kmp_alloc_argv_entries(argc, team, TRUE);
5166     KMP_CHECK_UPDATE(team->t.t_argc, argc);
5167     // The hot team re-uses the previous task team,
5168     // if untouched during the previous release->gather phase.
5169 
5170     KF_TRACE(10, (" hot_team = %p\n", team));
5171 
5172 #if KMP_DEBUG
5173     if (__kmp_tasking_mode != tskm_immediate_exec) {
5174       KA_TRACE(20, ("__kmp_allocate_team: hot team task_team[0] = %p "
5175                     "task_team[1] = %p after reinit\n",
5176                     team->t.t_task_team[0], team->t.t_task_team[1]));
5177     }
5178 #endif
5179 
5180 #if OMPT_SUPPORT
5181     __ompt_team_assign_id(team, ompt_parallel_data);
5182 #endif
5183 
5184     KMP_MB();
5185 
5186     return team;
5187   }
5188 
5189   /* next, let's try to take one from the team pool */
5190   KMP_MB();
5191   for (team = CCAST(kmp_team_t *, __kmp_team_pool); (team);) {
5192     /* TODO: consider resizing undersized teams instead of reaping them, now
5193        that we have a resizing mechanism */
5194     if (team->t.t_max_nproc >= max_nproc) {
5195       /* take this team from the team pool */
5196       __kmp_team_pool = team->t.t_next_pool;
5197 
5198       /* setup the team for fresh use */
5199       __kmp_initialize_team(team, new_nproc, new_icvs, NULL);
5200 
5201       KA_TRACE(20, ("__kmp_allocate_team: setting task_team[0] %p and "
5202                     "task_team[1] %p to NULL\n",
5203                     &team->t.t_task_team[0], &team->t.t_task_team[1]));
5204       team->t.t_task_team[0] = NULL;
5205       team->t.t_task_team[1] = NULL;
5206 
5207       /* reallocate space for arguments if necessary */
5208       __kmp_alloc_argv_entries(argc, team, TRUE);
5209       KMP_CHECK_UPDATE(team->t.t_argc, argc);
5210 
5211       KA_TRACE(
5212           20, ("__kmp_allocate_team: team %d init arrived: join=%u, plain=%u\n",
5213                team->t.t_id, KMP_INIT_BARRIER_STATE, KMP_INIT_BARRIER_STATE));
5214       { // Initialize barrier data.
5215         int b;
5216         for (b = 0; b < bs_last_barrier; ++b) {
5217           team->t.t_bar[b].b_arrived = KMP_INIT_BARRIER_STATE;
5218 #if USE_DEBUGGER
5219           team->t.t_bar[b].b_master_arrived = 0;
5220           team->t.t_bar[b].b_team_arrived = 0;
5221 #endif
5222         }
5223       }
5224 
5225 #if OMP_40_ENABLED
5226       team->t.t_proc_bind = new_proc_bind;
5227 #endif
5228 
5229       KA_TRACE(20, ("__kmp_allocate_team: using team from pool %d.\n",
5230                     team->t.t_id));
5231 
5232 #if OMPT_SUPPORT
5233       __ompt_team_assign_id(team, ompt_parallel_data);
5234 #endif
5235 
5236       KMP_MB();
5237 
5238       return team;
5239     }
5240 
5241     /* reap team if it is too small, then loop back and check the next one */
5242     // not sure if this is wise, but, will be redone during the hot-teams
5243     // rewrite.
5244     /* TODO: Use technique to find the right size hot-team, don't reap them */
5245     team = __kmp_reap_team(team);
5246     __kmp_team_pool = team;
5247   }
5248 
5249   /* nothing available in the pool, no matter, make a new team! */
5250   KMP_MB();
5251   team = (kmp_team_t *)__kmp_allocate(sizeof(kmp_team_t));
5252 
5253   /* and set it up */
5254   team->t.t_max_nproc = max_nproc;
5255   /* NOTE well, for some reason allocating one big buffer and dividing it up
5256      seems to really hurt performance a lot on the P4, so, let's not use this */
5257   __kmp_allocate_team_arrays(team, max_nproc);
5258 
5259   KA_TRACE(20, ("__kmp_allocate_team: making a new team\n"));
5260   __kmp_initialize_team(team, new_nproc, new_icvs, NULL);
5261 
5262   KA_TRACE(20, ("__kmp_allocate_team: setting task_team[0] %p and task_team[1] "
5263                 "%p to NULL\n",
5264                 &team->t.t_task_team[0], &team->t.t_task_team[1]));
5265   team->t.t_task_team[0] = NULL; // to be removed, as __kmp_allocate zeroes
5266   // memory, no need to duplicate
5267   team->t.t_task_team[1] = NULL; // to be removed, as __kmp_allocate zeroes
5268   // memory, no need to duplicate
5269 
5270   if (__kmp_storage_map) {
5271     __kmp_print_team_storage_map("team", team, team->t.t_id, new_nproc);
5272   }
5273 
5274   /* allocate space for arguments */
5275   __kmp_alloc_argv_entries(argc, team, FALSE);
5276   team->t.t_argc = argc;
5277 
5278   KA_TRACE(20,
5279            ("__kmp_allocate_team: team %d init arrived: join=%u, plain=%u\n",
5280             team->t.t_id, KMP_INIT_BARRIER_STATE, KMP_INIT_BARRIER_STATE));
5281   { // Initialize barrier data.
5282     int b;
5283     for (b = 0; b < bs_last_barrier; ++b) {
5284       team->t.t_bar[b].b_arrived = KMP_INIT_BARRIER_STATE;
5285 #if USE_DEBUGGER
5286       team->t.t_bar[b].b_master_arrived = 0;
5287       team->t.t_bar[b].b_team_arrived = 0;
5288 #endif
5289     }
5290   }
5291 
5292 #if OMP_40_ENABLED
5293   team->t.t_proc_bind = new_proc_bind;
5294 #endif
5295 
5296 #if OMPT_SUPPORT
5297   __ompt_team_assign_id(team, ompt_parallel_data);
5298   team->t.ompt_serialized_team_info = NULL;
5299 #endif
5300 
5301   KMP_MB();
5302 
5303   KA_TRACE(20, ("__kmp_allocate_team: done creating a new team %d.\n",
5304                 team->t.t_id));
5305 
5306   return team;
5307 }
5308 
5309 /* TODO implement hot-teams at all levels */
5310 /* TODO implement lazy thread release on demand (disband request) */
5311 
5312 /* free the team.  return it to the team pool.  release all the threads
5313  * associated with it */
5314 void __kmp_free_team(kmp_root_t *root,
5315                      kmp_team_t *team USE_NESTED_HOT_ARG(kmp_info_t *master)) {
5316   int f;
5317   KA_TRACE(20, ("__kmp_free_team: T#%d freeing team %d\n", __kmp_get_gtid(),
5318                 team->t.t_id));
5319 
5320   /* verify state */
5321   KMP_DEBUG_ASSERT(root);
5322   KMP_DEBUG_ASSERT(team);
5323   KMP_DEBUG_ASSERT(team->t.t_nproc <= team->t.t_max_nproc);
5324   KMP_DEBUG_ASSERT(team->t.t_threads);
5325 
5326   int use_hot_team = team == root->r.r_hot_team;
5327 #if KMP_NESTED_HOT_TEAMS
5328   int level;
5329   kmp_hot_team_ptr_t *hot_teams;
5330   if (master) {
5331     level = team->t.t_active_level - 1;
5332     if (master->th.th_teams_microtask) { // in teams construct?
5333       if (master->th.th_teams_size.nteams > 1) {
5334         ++level; // level was not increased in teams construct for
5335         // team_of_masters
5336       }
5337       if (team->t.t_pkfn != (microtask_t)__kmp_teams_master &&
5338           master->th.th_teams_level == team->t.t_level) {
5339         ++level; // level was not increased in teams construct for
5340         // team_of_workers before the parallel
5341       } // team->t.t_level will be increased inside parallel
5342     }
5343     hot_teams = master->th.th_hot_teams;
5344     if (level < __kmp_hot_teams_max_level) {
5345       KMP_DEBUG_ASSERT(team == hot_teams[level].hot_team);
5346       use_hot_team = 1;
5347     }
5348   }
5349 #endif // KMP_NESTED_HOT_TEAMS
5350 
5351   /* team is done working */
5352   TCW_SYNC_PTR(team->t.t_pkfn,
5353                NULL); // Important for Debugging Support Library.
5354 #if KMP_OS_WINDOWS
5355   team->t.t_copyin_counter = 0; // init counter for possible reuse
5356 #endif
5357   // Do not reset pointer to parent team to NULL for hot teams.
5358 
5359   /* if we are non-hot team, release our threads */
5360   if (!use_hot_team) {
5361     if (__kmp_tasking_mode != tskm_immediate_exec) {
5362       // Wait for threads to reach reapable state
5363       for (f = 1; f < team->t.t_nproc; ++f) {
5364         KMP_DEBUG_ASSERT(team->t.t_threads[f]);
5365         kmp_info_t *th = team->t.t_threads[f];
5366         volatile kmp_uint32 *state = &th->th.th_reap_state;
5367         while (*state != KMP_SAFE_TO_REAP) {
5368 #if KMP_OS_WINDOWS
5369           // On Windows a thread can be killed at any time, check this
5370           DWORD ecode;
5371           if (!__kmp_is_thread_alive(th, &ecode)) {
5372             *state = KMP_SAFE_TO_REAP; // reset the flag for dead thread
5373             break;
5374           }
5375 #endif
5376           // first check if thread is sleeping
5377           kmp_flag_64 fl(&th->th.th_bar[bs_forkjoin_barrier].bb.b_go, th);
5378           if (fl.is_sleeping())
5379             fl.resume(__kmp_gtid_from_thread(th));
5380           KMP_CPU_PAUSE();
5381         }
5382       }
5383 
5384       // Delete task teams
5385       int tt_idx;
5386       for (tt_idx = 0; tt_idx < 2; ++tt_idx) {
5387         kmp_task_team_t *task_team = team->t.t_task_team[tt_idx];
5388         if (task_team != NULL) {
5389           for (f = 0; f < team->t.t_nproc;
5390                ++f) { // Have all threads unref task teams
5391             team->t.t_threads[f]->th.th_task_team = NULL;
5392           }
5393           KA_TRACE(
5394               20,
5395               ("__kmp_free_team: T#%d deactivating task_team %p on team %d\n",
5396                __kmp_get_gtid(), task_team, team->t.t_id));
5397 #if KMP_NESTED_HOT_TEAMS
5398           __kmp_free_task_team(master, task_team);
5399 #endif
5400           team->t.t_task_team[tt_idx] = NULL;
5401         }
5402       }
5403     }
5404 
5405     // Reset pointer to parent team only for non-hot teams.
5406     team->t.t_parent = NULL;
5407     team->t.t_level = 0;
5408     team->t.t_active_level = 0;
5409 
5410     /* free the worker threads */
5411     for (f = 1; f < team->t.t_nproc; ++f) {
5412       KMP_DEBUG_ASSERT(team->t.t_threads[f]);
5413       __kmp_free_thread(team->t.t_threads[f]);
5414       team->t.t_threads[f] = NULL;
5415     }
5416 
5417     /* put the team back in the team pool */
5418     /* TODO limit size of team pool, call reap_team if pool too large */
5419     team->t.t_next_pool = CCAST(kmp_team_t *, __kmp_team_pool);
5420     __kmp_team_pool = (volatile kmp_team_t *)team;
5421   }
5422 
5423   KMP_MB();
5424 }
5425 
5426 /* reap the team.  destroy it, reclaim all its resources and free its memory */
5427 kmp_team_t *__kmp_reap_team(kmp_team_t *team) {
5428   kmp_team_t *next_pool = team->t.t_next_pool;
5429 
5430   KMP_DEBUG_ASSERT(team);
5431   KMP_DEBUG_ASSERT(team->t.t_dispatch);
5432   KMP_DEBUG_ASSERT(team->t.t_disp_buffer);
5433   KMP_DEBUG_ASSERT(team->t.t_threads);
5434   KMP_DEBUG_ASSERT(team->t.t_argv);
5435 
5436   /* TODO clean the threads that are a part of this? */
5437 
5438   /* free stuff */
5439   __kmp_free_team_arrays(team);
5440   if (team->t.t_argv != &team->t.t_inline_argv[0])
5441     __kmp_free((void *)team->t.t_argv);
5442   __kmp_free(team);
5443 
5444   KMP_MB();
5445   return next_pool;
5446 }
5447 
5448 // Free the thread.  Don't reap it, just place it on the pool of available
5449 // threads.
5450 //
5451 // Changes for Quad issue 527845: We need a predictable OMP tid <-> gtid
5452 // binding for the affinity mechanism to be useful.
5453 //
5454 // Now, we always keep the free list (__kmp_thread_pool) sorted by gtid.
5455 // However, we want to avoid a potential performance problem by always
5456 // scanning through the list to find the correct point at which to insert
5457 // the thread (potential N**2 behavior).  To do this we keep track of the
5458 // last place a thread struct was inserted (__kmp_thread_pool_insert_pt).
5459 // With single-level parallelism, threads will always be added to the tail
5460 // of the list, kept track of by __kmp_thread_pool_insert_pt.  With nested
5461 // parallelism, all bets are off and we may need to scan through the entire
5462 // free list.
5463 //
5464 // This change also has a potentially large performance benefit, for some
5465 // applications.  Previously, as threads were freed from the hot team, they
5466 // would be placed back on the free list in inverse order.  If the hot team
5467 // grew back to it's original size, then the freed thread would be placed
5468 // back on the hot team in reverse order.  This could cause bad cache
5469 // locality problems on programs where the size of the hot team regularly
5470 // grew and shrunk.
5471 //
5472 // Now, for single-level parallelism, the OMP tid is alway == gtid.
5473 void __kmp_free_thread(kmp_info_t *this_th) {
5474   int gtid;
5475   kmp_info_t **scan;
5476   kmp_root_t *root = this_th->th.th_root;
5477 
5478   KA_TRACE(20, ("__kmp_free_thread: T#%d putting T#%d back on free pool.\n",
5479                 __kmp_get_gtid(), this_th->th.th_info.ds.ds_gtid));
5480 
5481   KMP_DEBUG_ASSERT(this_th);
5482 
5483   // When moving thread to pool, switch thread to wait on own b_go flag, and
5484   // uninitialized (NULL team).
5485   int b;
5486   kmp_balign_t *balign = this_th->th.th_bar;
5487   for (b = 0; b < bs_last_barrier; ++b) {
5488     if (balign[b].bb.wait_flag == KMP_BARRIER_PARENT_FLAG)
5489       balign[b].bb.wait_flag = KMP_BARRIER_SWITCH_TO_OWN_FLAG;
5490     balign[b].bb.team = NULL;
5491     balign[b].bb.leaf_kids = 0;
5492   }
5493   this_th->th.th_task_state = 0;
5494   this_th->th.th_reap_state = KMP_SAFE_TO_REAP;
5495 
5496   /* put thread back on the free pool */
5497   TCW_PTR(this_th->th.th_team, NULL);
5498   TCW_PTR(this_th->th.th_root, NULL);
5499   TCW_PTR(this_th->th.th_dispatch, NULL); /* NOT NEEDED */
5500 
5501   /* If the implicit task assigned to this thread can be used by other threads
5502    * -> multiple threads can share the data and try to free the task at
5503    * __kmp_reap_thread at exit. This duplicate use of the task data can happen
5504    * with higher probability when hot team is disabled but can occurs even when
5505    * the hot team is enabled */
5506   __kmp_free_implicit_task(this_th);
5507   this_th->th.th_current_task = NULL;
5508 
5509   // If the __kmp_thread_pool_insert_pt is already past the new insert
5510   // point, then we need to re-scan the entire list.
5511   gtid = this_th->th.th_info.ds.ds_gtid;
5512   if (__kmp_thread_pool_insert_pt != NULL) {
5513     KMP_DEBUG_ASSERT(__kmp_thread_pool != NULL);
5514     if (__kmp_thread_pool_insert_pt->th.th_info.ds.ds_gtid > gtid) {
5515       __kmp_thread_pool_insert_pt = NULL;
5516     }
5517   }
5518 
5519   // Scan down the list to find the place to insert the thread.
5520   // scan is the address of a link in the list, possibly the address of
5521   // __kmp_thread_pool itself.
5522   //
5523   // In the absence of nested parallism, the for loop will have 0 iterations.
5524   if (__kmp_thread_pool_insert_pt != NULL) {
5525     scan = &(__kmp_thread_pool_insert_pt->th.th_next_pool);
5526   } else {
5527     scan = CCAST(kmp_info_t **, &__kmp_thread_pool);
5528   }
5529   for (; (*scan != NULL) && ((*scan)->th.th_info.ds.ds_gtid < gtid);
5530        scan = &((*scan)->th.th_next_pool))
5531     ;
5532 
5533   // Insert the new element on the list, and set __kmp_thread_pool_insert_pt
5534   // to its address.
5535   TCW_PTR(this_th->th.th_next_pool, *scan);
5536   __kmp_thread_pool_insert_pt = *scan = this_th;
5537   KMP_DEBUG_ASSERT((this_th->th.th_next_pool == NULL) ||
5538                    (this_th->th.th_info.ds.ds_gtid <
5539                     this_th->th.th_next_pool->th.th_info.ds.ds_gtid));
5540   TCW_4(this_th->th.th_in_pool, TRUE);
5541   __kmp_thread_pool_nth++;
5542 
5543   TCW_4(__kmp_nth, __kmp_nth - 1);
5544   root->r.r_cg_nthreads--;
5545 
5546 #ifdef KMP_ADJUST_BLOCKTIME
5547   /* Adjust blocktime back to user setting or default if necessary */
5548   /* Middle initialization might never have occurred                */
5549   if (!__kmp_env_blocktime && (__kmp_avail_proc > 0)) {
5550     KMP_DEBUG_ASSERT(__kmp_avail_proc > 0);
5551     if (__kmp_nth <= __kmp_avail_proc) {
5552       __kmp_zero_bt = FALSE;
5553     }
5554   }
5555 #endif /* KMP_ADJUST_BLOCKTIME */
5556 
5557   KMP_MB();
5558 }
5559 
5560 /* ------------------------------------------------------------------------ */
5561 
5562 void *__kmp_launch_thread(kmp_info_t *this_thr) {
5563   int gtid = this_thr->th.th_info.ds.ds_gtid;
5564   /*    void                 *stack_data;*/
5565   kmp_team_t *(*volatile pteam);
5566 
5567   KMP_MB();
5568   KA_TRACE(10, ("__kmp_launch_thread: T#%d start\n", gtid));
5569 
5570   if (__kmp_env_consistency_check) {
5571     this_thr->th.th_cons = __kmp_allocate_cons_stack(gtid); // ATT: Memory leak?
5572   }
5573 
5574 #if OMPT_SUPPORT
5575   ompt_data_t *thread_data;
5576   if (ompt_enabled.enabled) {
5577     thread_data = &(this_thr->th.ompt_thread_info.thread_data);
5578     thread_data->ptr = NULL;
5579 
5580     this_thr->th.ompt_thread_info.state = omp_state_overhead;
5581     this_thr->th.ompt_thread_info.wait_id = 0;
5582     this_thr->th.ompt_thread_info.idle_frame = OMPT_GET_FRAME_ADDRESS(0);
5583     if (ompt_enabled.ompt_callback_thread_begin) {
5584       ompt_callbacks.ompt_callback(ompt_callback_thread_begin)(
5585           ompt_thread_worker, thread_data);
5586     }
5587   }
5588 #endif
5589 
5590 #if OMPT_SUPPORT
5591   if (ompt_enabled.enabled) {
5592     this_thr->th.ompt_thread_info.state = omp_state_idle;
5593   }
5594 #endif
5595   /* This is the place where threads wait for work */
5596   while (!TCR_4(__kmp_global.g.g_done)) {
5597     KMP_DEBUG_ASSERT(this_thr == __kmp_threads[gtid]);
5598     KMP_MB();
5599 
5600     /* wait for work to do */
5601     KA_TRACE(20, ("__kmp_launch_thread: T#%d waiting for work\n", gtid));
5602 
5603     /* No tid yet since not part of a team */
5604     __kmp_fork_barrier(gtid, KMP_GTID_DNE);
5605 
5606 #if OMPT_SUPPORT
5607     if (ompt_enabled.enabled) {
5608       this_thr->th.ompt_thread_info.state = omp_state_overhead;
5609     }
5610 #endif
5611 
5612     pteam = (kmp_team_t * (*))(&this_thr->th.th_team);
5613 
5614     /* have we been allocated? */
5615     if (TCR_SYNC_PTR(*pteam) && !TCR_4(__kmp_global.g.g_done)) {
5616       /* we were just woken up, so run our new task */
5617       if (TCR_SYNC_PTR((*pteam)->t.t_pkfn) != NULL) {
5618         int rc;
5619         KA_TRACE(20,
5620                  ("__kmp_launch_thread: T#%d(%d:%d) invoke microtask = %p\n",
5621                   gtid, (*pteam)->t.t_id, __kmp_tid_from_gtid(gtid),
5622                   (*pteam)->t.t_pkfn));
5623 
5624         updateHWFPControl(*pteam);
5625 
5626 #if OMPT_SUPPORT
5627         if (ompt_enabled.enabled) {
5628           this_thr->th.ompt_thread_info.state = omp_state_work_parallel;
5629         }
5630 #endif
5631 
5632         rc = (*pteam)->t.t_invoke(gtid);
5633         KMP_ASSERT(rc);
5634 
5635         KMP_MB();
5636         KA_TRACE(20, ("__kmp_launch_thread: T#%d(%d:%d) done microtask = %p\n",
5637                       gtid, (*pteam)->t.t_id, __kmp_tid_from_gtid(gtid),
5638                       (*pteam)->t.t_pkfn));
5639       }
5640 #if OMPT_SUPPORT
5641       if (ompt_enabled.enabled) {
5642         /* no frame set while outside task */
5643         __ompt_get_task_info_object(0)->frame.exit_frame = NULL;
5644 
5645         this_thr->th.ompt_thread_info.state = omp_state_overhead;
5646       }
5647 #endif
5648       /* join barrier after parallel region */
5649       __kmp_join_barrier(gtid);
5650     }
5651   }
5652   TCR_SYNC_PTR((intptr_t)__kmp_global.g.g_done);
5653 
5654 #if OMPT_SUPPORT
5655   if (ompt_enabled.ompt_callback_thread_end) {
5656     ompt_callbacks.ompt_callback(ompt_callback_thread_end)(thread_data);
5657   }
5658 #endif
5659 
5660   this_thr->th.th_task_team = NULL;
5661   /* run the destructors for the threadprivate data for this thread */
5662   __kmp_common_destroy_gtid(gtid);
5663 
5664   KA_TRACE(10, ("__kmp_launch_thread: T#%d done\n", gtid));
5665   KMP_MB();
5666   return this_thr;
5667 }
5668 
5669 /* ------------------------------------------------------------------------ */
5670 
5671 void __kmp_internal_end_dest(void *specific_gtid) {
5672 #if KMP_COMPILER_ICC
5673 #pragma warning(push)
5674 #pragma warning(disable : 810) // conversion from "void *" to "int" may lose
5675 // significant bits
5676 #endif
5677   // Make sure no significant bits are lost
5678   int gtid = (kmp_intptr_t)specific_gtid - 1;
5679 #if KMP_COMPILER_ICC
5680 #pragma warning(pop)
5681 #endif
5682 
5683   KA_TRACE(30, ("__kmp_internal_end_dest: T#%d\n", gtid));
5684   /* NOTE: the gtid is stored as gitd+1 in the thread-local-storage
5685    * this is because 0 is reserved for the nothing-stored case */
5686 
5687   /* josh: One reason for setting the gtid specific data even when it is being
5688      destroyed by pthread is to allow gtid lookup through thread specific data
5689      (__kmp_gtid_get_specific).  Some of the code, especially stat code,
5690      that gets executed in the call to __kmp_internal_end_thread, actually
5691      gets the gtid through the thread specific data.  Setting it here seems
5692      rather inelegant and perhaps wrong, but allows __kmp_internal_end_thread
5693      to run smoothly.
5694      todo: get rid of this after we remove the dependence on
5695      __kmp_gtid_get_specific  */
5696   if (gtid >= 0 && KMP_UBER_GTID(gtid))
5697     __kmp_gtid_set_specific(gtid);
5698 #ifdef KMP_TDATA_GTID
5699   __kmp_gtid = gtid;
5700 #endif
5701   __kmp_internal_end_thread(gtid);
5702 }
5703 
5704 #if KMP_OS_UNIX && KMP_DYNAMIC_LIB
5705 
5706 // 2009-09-08 (lev): It looks the destructor does not work. In simple test cases
5707 // destructors work perfectly, but in real libomp.so I have no evidence it is
5708 // ever called. However, -fini linker option in makefile.mk works fine.
5709 
5710 __attribute__((destructor)) void __kmp_internal_end_dtor(void) {
5711   __kmp_internal_end_atexit();
5712 }
5713 
5714 void __kmp_internal_end_fini(void) { __kmp_internal_end_atexit(); }
5715 
5716 #endif
5717 
5718 /* [Windows] josh: when the atexit handler is called, there may still be more
5719    than one thread alive */
5720 void __kmp_internal_end_atexit(void) {
5721   KA_TRACE(30, ("__kmp_internal_end_atexit\n"));
5722   /* [Windows]
5723      josh: ideally, we want to completely shutdown the library in this atexit
5724      handler, but stat code that depends on thread specific data for gtid fails
5725      because that data becomes unavailable at some point during the shutdown, so
5726      we call __kmp_internal_end_thread instead. We should eventually remove the
5727      dependency on __kmp_get_specific_gtid in the stat code and use
5728      __kmp_internal_end_library to cleanly shutdown the library.
5729 
5730      // TODO: Can some of this comment about GVS be removed?
5731      I suspect that the offending stat code is executed when the calling thread
5732      tries to clean up a dead root thread's data structures, resulting in GVS
5733      code trying to close the GVS structures for that thread, but since the stat
5734      code uses __kmp_get_specific_gtid to get the gtid with the assumption that
5735      the calling thread is cleaning up itself instead of another thread, it get
5736      confused. This happens because allowing a thread to unregister and cleanup
5737      another thread is a recent modification for addressing an issue.
5738      Based on the current design (20050722), a thread may end up
5739      trying to unregister another thread only if thread death does not trigger
5740      the calling of __kmp_internal_end_thread.  For Linux* OS, there is the
5741      thread specific data destructor function to detect thread death. For
5742      Windows dynamic, there is DllMain(THREAD_DETACH). For Windows static, there
5743      is nothing.  Thus, the workaround is applicable only for Windows static
5744      stat library. */
5745   __kmp_internal_end_library(-1);
5746 #if KMP_OS_WINDOWS
5747   __kmp_close_console();
5748 #endif
5749 }
5750 
5751 static void __kmp_reap_thread(kmp_info_t *thread, int is_root) {
5752   // It is assumed __kmp_forkjoin_lock is acquired.
5753 
5754   int gtid;
5755 
5756   KMP_DEBUG_ASSERT(thread != NULL);
5757 
5758   gtid = thread->th.th_info.ds.ds_gtid;
5759 
5760   if (!is_root) {
5761 
5762     if (__kmp_dflt_blocktime != KMP_MAX_BLOCKTIME) {
5763       /* Assume the threads are at the fork barrier here */
5764       KA_TRACE(
5765           20, ("__kmp_reap_thread: releasing T#%d from fork barrier for reap\n",
5766                gtid));
5767       /* Need release fence here to prevent seg faults for tree forkjoin barrier
5768        * (GEH) */
5769       ANNOTATE_HAPPENS_BEFORE(thread);
5770       kmp_flag_64 flag(&thread->th.th_bar[bs_forkjoin_barrier].bb.b_go, thread);
5771       __kmp_release_64(&flag);
5772     }
5773 
5774     // Terminate OS thread.
5775     __kmp_reap_worker(thread);
5776 
5777     // The thread was killed asynchronously.  If it was actively
5778     // spinning in the thread pool, decrement the global count.
5779     //
5780     // There is a small timing hole here - if the worker thread was just waking
5781     // up after sleeping in the pool, had reset it's th_active_in_pool flag but
5782     // not decremented the global counter __kmp_thread_pool_active_nth yet, then
5783     // the global counter might not get updated.
5784     //
5785     // Currently, this can only happen as the library is unloaded,
5786     // so there are no harmful side effects.
5787     if (thread->th.th_active_in_pool) {
5788       thread->th.th_active_in_pool = FALSE;
5789       KMP_ATOMIC_DEC(&__kmp_thread_pool_active_nth);
5790       KMP_DEBUG_ASSERT(__kmp_thread_pool_active_nth >= 0);
5791     }
5792 
5793     // Decrement # of [worker] threads in the pool.
5794     KMP_DEBUG_ASSERT(__kmp_thread_pool_nth > 0);
5795     --__kmp_thread_pool_nth;
5796   }
5797 
5798   __kmp_free_implicit_task(thread);
5799 
5800 // Free the fast memory for tasking
5801 #if USE_FAST_MEMORY
5802   __kmp_free_fast_memory(thread);
5803 #endif /* USE_FAST_MEMORY */
5804 
5805   __kmp_suspend_uninitialize_thread(thread);
5806 
5807   KMP_DEBUG_ASSERT(__kmp_threads[gtid] == thread);
5808   TCW_SYNC_PTR(__kmp_threads[gtid], NULL);
5809 
5810   --__kmp_all_nth;
5811 // __kmp_nth was decremented when thread is added to the pool.
5812 
5813 #ifdef KMP_ADJUST_BLOCKTIME
5814   /* Adjust blocktime back to user setting or default if necessary */
5815   /* Middle initialization might never have occurred                */
5816   if (!__kmp_env_blocktime && (__kmp_avail_proc > 0)) {
5817     KMP_DEBUG_ASSERT(__kmp_avail_proc > 0);
5818     if (__kmp_nth <= __kmp_avail_proc) {
5819       __kmp_zero_bt = FALSE;
5820     }
5821   }
5822 #endif /* KMP_ADJUST_BLOCKTIME */
5823 
5824   /* free the memory being used */
5825   if (__kmp_env_consistency_check) {
5826     if (thread->th.th_cons) {
5827       __kmp_free_cons_stack(thread->th.th_cons);
5828       thread->th.th_cons = NULL;
5829     }
5830   }
5831 
5832   if (thread->th.th_pri_common != NULL) {
5833     __kmp_free(thread->th.th_pri_common);
5834     thread->th.th_pri_common = NULL;
5835   }
5836 
5837   if (thread->th.th_task_state_memo_stack != NULL) {
5838     __kmp_free(thread->th.th_task_state_memo_stack);
5839     thread->th.th_task_state_memo_stack = NULL;
5840   }
5841 
5842 #if KMP_USE_BGET
5843   if (thread->th.th_local.bget_data != NULL) {
5844     __kmp_finalize_bget(thread);
5845   }
5846 #endif
5847 
5848 #if KMP_AFFINITY_SUPPORTED
5849   if (thread->th.th_affin_mask != NULL) {
5850     KMP_CPU_FREE(thread->th.th_affin_mask);
5851     thread->th.th_affin_mask = NULL;
5852   }
5853 #endif /* KMP_AFFINITY_SUPPORTED */
5854 
5855 #if KMP_USE_HIER_SCHED
5856   if (thread->th.th_hier_bar_data != NULL) {
5857     __kmp_free(thread->th.th_hier_bar_data);
5858     thread->th.th_hier_bar_data = NULL;
5859   }
5860 #endif
5861 
5862   __kmp_reap_team(thread->th.th_serial_team);
5863   thread->th.th_serial_team = NULL;
5864   __kmp_free(thread);
5865 
5866   KMP_MB();
5867 
5868 } // __kmp_reap_thread
5869 
5870 static void __kmp_internal_end(void) {
5871   int i;
5872 
5873   /* First, unregister the library */
5874   __kmp_unregister_library();
5875 
5876 #if KMP_OS_WINDOWS
5877   /* In Win static library, we can't tell when a root actually dies, so we
5878      reclaim the data structures for any root threads that have died but not
5879      unregistered themselves, in order to shut down cleanly.
5880      In Win dynamic library we also can't tell when a thread dies.  */
5881   __kmp_reclaim_dead_roots(); // AC: moved here to always clean resources of
5882 // dead roots
5883 #endif
5884 
5885   for (i = 0; i < __kmp_threads_capacity; i++)
5886     if (__kmp_root[i])
5887       if (__kmp_root[i]->r.r_active)
5888         break;
5889   KMP_MB(); /* Flush all pending memory write invalidates.  */
5890   TCW_SYNC_4(__kmp_global.g.g_done, TRUE);
5891 
5892   if (i < __kmp_threads_capacity) {
5893 #if KMP_USE_MONITOR
5894     // 2009-09-08 (lev): Other alive roots found. Why do we kill the monitor??
5895     KMP_MB(); /* Flush all pending memory write invalidates.  */
5896 
5897     // Need to check that monitor was initialized before reaping it. If we are
5898     // called form __kmp_atfork_child (which sets __kmp_init_parallel = 0), then
5899     // __kmp_monitor will appear to contain valid data, but it is only valid in
5900     // the parent process, not the child.
5901     // New behavior (201008): instead of keying off of the flag
5902     // __kmp_init_parallel, the monitor thread creation is keyed off
5903     // of the new flag __kmp_init_monitor.
5904     __kmp_acquire_bootstrap_lock(&__kmp_monitor_lock);
5905     if (TCR_4(__kmp_init_monitor)) {
5906       __kmp_reap_monitor(&__kmp_monitor);
5907       TCW_4(__kmp_init_monitor, 0);
5908     }
5909     __kmp_release_bootstrap_lock(&__kmp_monitor_lock);
5910     KA_TRACE(10, ("__kmp_internal_end: monitor reaped\n"));
5911 #endif // KMP_USE_MONITOR
5912   } else {
5913 /* TODO move this to cleanup code */
5914 #ifdef KMP_DEBUG
5915     /* make sure that everything has properly ended */
5916     for (i = 0; i < __kmp_threads_capacity; i++) {
5917       if (__kmp_root[i]) {
5918         //                    KMP_ASSERT( ! KMP_UBER_GTID( i ) );         // AC:
5919         //                    there can be uber threads alive here
5920         KMP_ASSERT(!__kmp_root[i]->r.r_active); // TODO: can they be active?
5921       }
5922     }
5923 #endif
5924 
5925     KMP_MB();
5926 
5927     // Reap the worker threads.
5928     // This is valid for now, but be careful if threads are reaped sooner.
5929     while (__kmp_thread_pool != NULL) { // Loop thru all the thread in the pool.
5930       // Get the next thread from the pool.
5931       kmp_info_t *thread = CCAST(kmp_info_t *, __kmp_thread_pool);
5932       __kmp_thread_pool = thread->th.th_next_pool;
5933       // Reap it.
5934       KMP_DEBUG_ASSERT(thread->th.th_reap_state == KMP_SAFE_TO_REAP);
5935       thread->th.th_next_pool = NULL;
5936       thread->th.th_in_pool = FALSE;
5937       __kmp_reap_thread(thread, 0);
5938     }
5939     __kmp_thread_pool_insert_pt = NULL;
5940 
5941     // Reap teams.
5942     while (__kmp_team_pool != NULL) { // Loop thru all the teams in the pool.
5943       // Get the next team from the pool.
5944       kmp_team_t *team = CCAST(kmp_team_t *, __kmp_team_pool);
5945       __kmp_team_pool = team->t.t_next_pool;
5946       // Reap it.
5947       team->t.t_next_pool = NULL;
5948       __kmp_reap_team(team);
5949     }
5950 
5951     __kmp_reap_task_teams();
5952 
5953 #if KMP_OS_UNIX
5954     // Threads that are not reaped should not access any resources since they
5955     // are going to be deallocated soon, so the shutdown sequence should wait
5956     // until all threads either exit the final spin-waiting loop or begin
5957     // sleeping after the given blocktime.
5958     for (i = 0; i < __kmp_threads_capacity; i++) {
5959       kmp_info_t *thr = __kmp_threads[i];
5960       while (thr && KMP_ATOMIC_LD_ACQ(&thr->th.th_blocking))
5961         KMP_CPU_PAUSE();
5962     }
5963 #endif
5964 
5965     for (i = 0; i < __kmp_threads_capacity; ++i) {
5966       // TBD: Add some checking...
5967       // Something like KMP_DEBUG_ASSERT( __kmp_thread[ i ] == NULL );
5968     }
5969 
5970     /* Make sure all threadprivate destructors get run by joining with all
5971        worker threads before resetting this flag */
5972     TCW_SYNC_4(__kmp_init_common, FALSE);
5973 
5974     KA_TRACE(10, ("__kmp_internal_end: all workers reaped\n"));
5975     KMP_MB();
5976 
5977 #if KMP_USE_MONITOR
5978     // See note above: One of the possible fixes for CQ138434 / CQ140126
5979     //
5980     // FIXME: push both code fragments down and CSE them?
5981     // push them into __kmp_cleanup() ?
5982     __kmp_acquire_bootstrap_lock(&__kmp_monitor_lock);
5983     if (TCR_4(__kmp_init_monitor)) {
5984       __kmp_reap_monitor(&__kmp_monitor);
5985       TCW_4(__kmp_init_monitor, 0);
5986     }
5987     __kmp_release_bootstrap_lock(&__kmp_monitor_lock);
5988     KA_TRACE(10, ("__kmp_internal_end: monitor reaped\n"));
5989 #endif
5990   } /* else !__kmp_global.t_active */
5991   TCW_4(__kmp_init_gtid, FALSE);
5992   KMP_MB(); /* Flush all pending memory write invalidates.  */
5993 
5994   __kmp_cleanup();
5995 #if OMPT_SUPPORT
5996   ompt_fini();
5997 #endif
5998 }
5999 
6000 void __kmp_internal_end_library(int gtid_req) {
6001   /* if we have already cleaned up, don't try again, it wouldn't be pretty */
6002   /* this shouldn't be a race condition because __kmp_internal_end() is the
6003      only place to clear __kmp_serial_init */
6004   /* we'll check this later too, after we get the lock */
6005   // 2009-09-06: We do not set g_abort without setting g_done. This check looks
6006   // redundaant, because the next check will work in any case.
6007   if (__kmp_global.g.g_abort) {
6008     KA_TRACE(11, ("__kmp_internal_end_library: abort, exiting\n"));
6009     /* TODO abort? */
6010     return;
6011   }
6012   if (TCR_4(__kmp_global.g.g_done) || !__kmp_init_serial) {
6013     KA_TRACE(10, ("__kmp_internal_end_library: already finished\n"));
6014     return;
6015   }
6016 
6017   KMP_MB(); /* Flush all pending memory write invalidates.  */
6018 
6019   /* find out who we are and what we should do */
6020   {
6021     int gtid = (gtid_req >= 0) ? gtid_req : __kmp_gtid_get_specific();
6022     KA_TRACE(
6023         10, ("__kmp_internal_end_library: enter T#%d  (%d)\n", gtid, gtid_req));
6024     if (gtid == KMP_GTID_SHUTDOWN) {
6025       KA_TRACE(10, ("__kmp_internal_end_library: !__kmp_init_runtime, system "
6026                     "already shutdown\n"));
6027       return;
6028     } else if (gtid == KMP_GTID_MONITOR) {
6029       KA_TRACE(10, ("__kmp_internal_end_library: monitor thread, gtid not "
6030                     "registered, or system shutdown\n"));
6031       return;
6032     } else if (gtid == KMP_GTID_DNE) {
6033       KA_TRACE(10, ("__kmp_internal_end_library: gtid not registered or system "
6034                     "shutdown\n"));
6035       /* we don't know who we are, but we may still shutdown the library */
6036     } else if (KMP_UBER_GTID(gtid)) {
6037       /* unregister ourselves as an uber thread.  gtid is no longer valid */
6038       if (__kmp_root[gtid]->r.r_active) {
6039         __kmp_global.g.g_abort = -1;
6040         TCW_SYNC_4(__kmp_global.g.g_done, TRUE);
6041         KA_TRACE(10,
6042                  ("__kmp_internal_end_library: root still active, abort T#%d\n",
6043                   gtid));
6044         return;
6045       } else {
6046         KA_TRACE(
6047             10,
6048             ("__kmp_internal_end_library: unregistering sibling T#%d\n", gtid));
6049         __kmp_unregister_root_current_thread(gtid);
6050       }
6051     } else {
6052 /* worker threads may call this function through the atexit handler, if they
6053  * call exit() */
6054 /* For now, skip the usual subsequent processing and just dump the debug buffer.
6055    TODO: do a thorough shutdown instead */
6056 #ifdef DUMP_DEBUG_ON_EXIT
6057       if (__kmp_debug_buf)
6058         __kmp_dump_debug_buffer();
6059 #endif
6060       return;
6061     }
6062   }
6063   /* synchronize the termination process */
6064   __kmp_acquire_bootstrap_lock(&__kmp_initz_lock);
6065 
6066   /* have we already finished */
6067   if (__kmp_global.g.g_abort) {
6068     KA_TRACE(10, ("__kmp_internal_end_library: abort, exiting\n"));
6069     /* TODO abort? */
6070     __kmp_release_bootstrap_lock(&__kmp_initz_lock);
6071     return;
6072   }
6073   if (TCR_4(__kmp_global.g.g_done) || !__kmp_init_serial) {
6074     __kmp_release_bootstrap_lock(&__kmp_initz_lock);
6075     return;
6076   }
6077 
6078   /* We need this lock to enforce mutex between this reading of
6079      __kmp_threads_capacity and the writing by __kmp_register_root.
6080      Alternatively, we can use a counter of roots that is atomically updated by
6081      __kmp_get_global_thread_id_reg, __kmp_do_serial_initialize and
6082      __kmp_internal_end_*.  */
6083   __kmp_acquire_bootstrap_lock(&__kmp_forkjoin_lock);
6084 
6085   /* now we can safely conduct the actual termination */
6086   __kmp_internal_end();
6087 
6088   __kmp_release_bootstrap_lock(&__kmp_forkjoin_lock);
6089   __kmp_release_bootstrap_lock(&__kmp_initz_lock);
6090 
6091   KA_TRACE(10, ("__kmp_internal_end_library: exit\n"));
6092 
6093 #ifdef DUMP_DEBUG_ON_EXIT
6094   if (__kmp_debug_buf)
6095     __kmp_dump_debug_buffer();
6096 #endif
6097 
6098 #if KMP_OS_WINDOWS
6099   __kmp_close_console();
6100 #endif
6101 
6102   __kmp_fini_allocator();
6103 
6104 } // __kmp_internal_end_library
6105 
6106 void __kmp_internal_end_thread(int gtid_req) {
6107   int i;
6108 
6109   /* if we have already cleaned up, don't try again, it wouldn't be pretty */
6110   /* this shouldn't be a race condition because __kmp_internal_end() is the
6111    * only place to clear __kmp_serial_init */
6112   /* we'll check this later too, after we get the lock */
6113   // 2009-09-06: We do not set g_abort without setting g_done. This check looks
6114   // redundant, because the next check will work in any case.
6115   if (__kmp_global.g.g_abort) {
6116     KA_TRACE(11, ("__kmp_internal_end_thread: abort, exiting\n"));
6117     /* TODO abort? */
6118     return;
6119   }
6120   if (TCR_4(__kmp_global.g.g_done) || !__kmp_init_serial) {
6121     KA_TRACE(10, ("__kmp_internal_end_thread: already finished\n"));
6122     return;
6123   }
6124 
6125   KMP_MB(); /* Flush all pending memory write invalidates.  */
6126 
6127   /* find out who we are and what we should do */
6128   {
6129     int gtid = (gtid_req >= 0) ? gtid_req : __kmp_gtid_get_specific();
6130     KA_TRACE(10,
6131              ("__kmp_internal_end_thread: enter T#%d  (%d)\n", gtid, gtid_req));
6132     if (gtid == KMP_GTID_SHUTDOWN) {
6133       KA_TRACE(10, ("__kmp_internal_end_thread: !__kmp_init_runtime, system "
6134                     "already shutdown\n"));
6135       return;
6136     } else if (gtid == KMP_GTID_MONITOR) {
6137       KA_TRACE(10, ("__kmp_internal_end_thread: monitor thread, gtid not "
6138                     "registered, or system shutdown\n"));
6139       return;
6140     } else if (gtid == KMP_GTID_DNE) {
6141       KA_TRACE(10, ("__kmp_internal_end_thread: gtid not registered or system "
6142                     "shutdown\n"));
6143       return;
6144       /* we don't know who we are */
6145     } else if (KMP_UBER_GTID(gtid)) {
6146       /* unregister ourselves as an uber thread.  gtid is no longer valid */
6147       if (__kmp_root[gtid]->r.r_active) {
6148         __kmp_global.g.g_abort = -1;
6149         TCW_SYNC_4(__kmp_global.g.g_done, TRUE);
6150         KA_TRACE(10,
6151                  ("__kmp_internal_end_thread: root still active, abort T#%d\n",
6152                   gtid));
6153         return;
6154       } else {
6155         KA_TRACE(10, ("__kmp_internal_end_thread: unregistering sibling T#%d\n",
6156                       gtid));
6157         __kmp_unregister_root_current_thread(gtid);
6158       }
6159     } else {
6160       /* just a worker thread, let's leave */
6161       KA_TRACE(10, ("__kmp_internal_end_thread: worker thread T#%d\n", gtid));
6162 
6163       if (gtid >= 0) {
6164         __kmp_threads[gtid]->th.th_task_team = NULL;
6165       }
6166 
6167       KA_TRACE(10,
6168                ("__kmp_internal_end_thread: worker thread done, exiting T#%d\n",
6169                 gtid));
6170       return;
6171     }
6172   }
6173 #if defined KMP_DYNAMIC_LIB
6174   // AC: lets not shutdown the Linux* OS dynamic library at the exit of uber
6175   // thread, because we will better shutdown later in the library destructor.
6176   // The reason of this change is performance problem when non-openmp thread in
6177   // a loop forks and joins many openmp threads. We can save a lot of time
6178   // keeping worker threads alive until the program shutdown.
6179   // OM: Removed Linux* OS restriction to fix the crash on OS X* (DPD200239966)
6180   // and Windows(DPD200287443) that occurs when using critical sections from
6181   // foreign threads.
6182   KA_TRACE(10, ("__kmp_internal_end_thread: exiting T#%d\n", gtid_req));
6183   return;
6184 #endif
6185   /* synchronize the termination process */
6186   __kmp_acquire_bootstrap_lock(&__kmp_initz_lock);
6187 
6188   /* have we already finished */
6189   if (__kmp_global.g.g_abort) {
6190     KA_TRACE(10, ("__kmp_internal_end_thread: abort, exiting\n"));
6191     /* TODO abort? */
6192     __kmp_release_bootstrap_lock(&__kmp_initz_lock);
6193     return;
6194   }
6195   if (TCR_4(__kmp_global.g.g_done) || !__kmp_init_serial) {
6196     __kmp_release_bootstrap_lock(&__kmp_initz_lock);
6197     return;
6198   }
6199 
6200   /* We need this lock to enforce mutex between this reading of
6201      __kmp_threads_capacity and the writing by __kmp_register_root.
6202      Alternatively, we can use a counter of roots that is atomically updated by
6203      __kmp_get_global_thread_id_reg, __kmp_do_serial_initialize and
6204      __kmp_internal_end_*.  */
6205 
6206   /* should we finish the run-time?  are all siblings done? */
6207   __kmp_acquire_bootstrap_lock(&__kmp_forkjoin_lock);
6208 
6209   for (i = 0; i < __kmp_threads_capacity; ++i) {
6210     if (KMP_UBER_GTID(i)) {
6211       KA_TRACE(
6212           10,
6213           ("__kmp_internal_end_thread: remaining sibling task: gtid==%d\n", i));
6214       __kmp_release_bootstrap_lock(&__kmp_forkjoin_lock);
6215       __kmp_release_bootstrap_lock(&__kmp_initz_lock);
6216       return;
6217     }
6218   }
6219 
6220   /* now we can safely conduct the actual termination */
6221 
6222   __kmp_internal_end();
6223 
6224   __kmp_release_bootstrap_lock(&__kmp_forkjoin_lock);
6225   __kmp_release_bootstrap_lock(&__kmp_initz_lock);
6226 
6227   KA_TRACE(10, ("__kmp_internal_end_thread: exit T#%d\n", gtid_req));
6228 
6229 #ifdef DUMP_DEBUG_ON_EXIT
6230   if (__kmp_debug_buf)
6231     __kmp_dump_debug_buffer();
6232 #endif
6233 } // __kmp_internal_end_thread
6234 
6235 // -----------------------------------------------------------------------------
6236 // Library registration stuff.
6237 
6238 static long __kmp_registration_flag = 0;
6239 // Random value used to indicate library initialization.
6240 static char *__kmp_registration_str = NULL;
6241 // Value to be saved in env var __KMP_REGISTERED_LIB_<pid>.
6242 
6243 static inline char *__kmp_reg_status_name() {
6244   /* On RHEL 3u5 if linked statically, getpid() returns different values in
6245      each thread. If registration and unregistration go in different threads
6246      (omp_misc_other_root_exit.cpp test case), the name of registered_lib_env
6247      env var can not be found, because the name will contain different pid. */
6248   return __kmp_str_format("__KMP_REGISTERED_LIB_%d", (int)getpid());
6249 } // __kmp_reg_status_get
6250 
6251 void __kmp_register_library_startup(void) {
6252 
6253   char *name = __kmp_reg_status_name(); // Name of the environment variable.
6254   int done = 0;
6255   union {
6256     double dtime;
6257     long ltime;
6258   } time;
6259 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
6260   __kmp_initialize_system_tick();
6261 #endif
6262   __kmp_read_system_time(&time.dtime);
6263   __kmp_registration_flag = 0xCAFE0000L | (time.ltime & 0x0000FFFFL);
6264   __kmp_registration_str =
6265       __kmp_str_format("%p-%lx-%s", &__kmp_registration_flag,
6266                        __kmp_registration_flag, KMP_LIBRARY_FILE);
6267 
6268   KA_TRACE(50, ("__kmp_register_library_startup: %s=\"%s\"\n", name,
6269                 __kmp_registration_str));
6270 
6271   while (!done) {
6272 
6273     char *value = NULL; // Actual value of the environment variable.
6274 
6275     // Set environment variable, but do not overwrite if it is exist.
6276     __kmp_env_set(name, __kmp_registration_str, 0);
6277     // Check the variable is written.
6278     value = __kmp_env_get(name);
6279     if (value != NULL && strcmp(value, __kmp_registration_str) == 0) {
6280 
6281       done = 1; // Ok, environment variable set successfully, exit the loop.
6282 
6283     } else {
6284 
6285       // Oops. Write failed. Another copy of OpenMP RTL is in memory.
6286       // Check whether it alive or dead.
6287       int neighbor = 0; // 0 -- unknown status, 1 -- alive, 2 -- dead.
6288       char *tail = value;
6289       char *flag_addr_str = NULL;
6290       char *flag_val_str = NULL;
6291       char const *file_name = NULL;
6292       __kmp_str_split(tail, '-', &flag_addr_str, &tail);
6293       __kmp_str_split(tail, '-', &flag_val_str, &tail);
6294       file_name = tail;
6295       if (tail != NULL) {
6296         long *flag_addr = 0;
6297         long flag_val = 0;
6298         KMP_SSCANF(flag_addr_str, "%p", RCAST(void**, &flag_addr));
6299         KMP_SSCANF(flag_val_str, "%lx", &flag_val);
6300         if (flag_addr != 0 && flag_val != 0 && strcmp(file_name, "") != 0) {
6301           // First, check whether environment-encoded address is mapped into
6302           // addr space.
6303           // If so, dereference it to see if it still has the right value.
6304           if (__kmp_is_address_mapped(flag_addr) && *flag_addr == flag_val) {
6305             neighbor = 1;
6306           } else {
6307             // If not, then we know the other copy of the library is no longer
6308             // running.
6309             neighbor = 2;
6310           }
6311         }
6312       }
6313       switch (neighbor) {
6314       case 0: // Cannot parse environment variable -- neighbor status unknown.
6315         // Assume it is the incompatible format of future version of the
6316         // library. Assume the other library is alive.
6317         // WARN( ... ); // TODO: Issue a warning.
6318         file_name = "unknown library";
6319       // Attention! Falling to the next case. That's intentional.
6320       case 1: { // Neighbor is alive.
6321         // Check it is allowed.
6322         char *duplicate_ok = __kmp_env_get("KMP_DUPLICATE_LIB_OK");
6323         if (!__kmp_str_match_true(duplicate_ok)) {
6324           // That's not allowed. Issue fatal error.
6325           __kmp_fatal(KMP_MSG(DuplicateLibrary, KMP_LIBRARY_FILE, file_name),
6326                       KMP_HNT(DuplicateLibrary), __kmp_msg_null);
6327         }
6328         KMP_INTERNAL_FREE(duplicate_ok);
6329         __kmp_duplicate_library_ok = 1;
6330         done = 1; // Exit the loop.
6331       } break;
6332       case 2: { // Neighbor is dead.
6333         // Clear the variable and try to register library again.
6334         __kmp_env_unset(name);
6335       } break;
6336       default: { KMP_DEBUG_ASSERT(0); } break;
6337       }
6338     }
6339     KMP_INTERNAL_FREE((void *)value);
6340   }
6341   KMP_INTERNAL_FREE((void *)name);
6342 
6343 } // func __kmp_register_library_startup
6344 
6345 void __kmp_unregister_library(void) {
6346 
6347   char *name = __kmp_reg_status_name();
6348   char *value = __kmp_env_get(name);
6349 
6350   KMP_DEBUG_ASSERT(__kmp_registration_flag != 0);
6351   KMP_DEBUG_ASSERT(__kmp_registration_str != NULL);
6352   if (value != NULL && strcmp(value, __kmp_registration_str) == 0) {
6353     // Ok, this is our variable. Delete it.
6354     __kmp_env_unset(name);
6355   }
6356 
6357   KMP_INTERNAL_FREE(__kmp_registration_str);
6358   KMP_INTERNAL_FREE(value);
6359   KMP_INTERNAL_FREE(name);
6360 
6361   __kmp_registration_flag = 0;
6362   __kmp_registration_str = NULL;
6363 
6364 } // __kmp_unregister_library
6365 
6366 // End of Library registration stuff.
6367 // -----------------------------------------------------------------------------
6368 
6369 #if KMP_MIC_SUPPORTED
6370 
6371 static void __kmp_check_mic_type() {
6372   kmp_cpuid_t cpuid_state = {0};
6373   kmp_cpuid_t *cs_p = &cpuid_state;
6374   __kmp_x86_cpuid(1, 0, cs_p);
6375   // We don't support mic1 at the moment
6376   if ((cs_p->eax & 0xff0) == 0xB10) {
6377     __kmp_mic_type = mic2;
6378   } else if ((cs_p->eax & 0xf0ff0) == 0x50670) {
6379     __kmp_mic_type = mic3;
6380   } else {
6381     __kmp_mic_type = non_mic;
6382   }
6383 }
6384 
6385 #endif /* KMP_MIC_SUPPORTED */
6386 
6387 static void __kmp_do_serial_initialize(void) {
6388   int i, gtid;
6389   int size;
6390 
6391   KA_TRACE(10, ("__kmp_do_serial_initialize: enter\n"));
6392 
6393   KMP_DEBUG_ASSERT(sizeof(kmp_int32) == 4);
6394   KMP_DEBUG_ASSERT(sizeof(kmp_uint32) == 4);
6395   KMP_DEBUG_ASSERT(sizeof(kmp_int64) == 8);
6396   KMP_DEBUG_ASSERT(sizeof(kmp_uint64) == 8);
6397   KMP_DEBUG_ASSERT(sizeof(kmp_intptr_t) == sizeof(void *));
6398 
6399 #if OMPT_SUPPORT
6400   ompt_pre_init();
6401 #endif
6402 
6403   __kmp_validate_locks();
6404 
6405   /* Initialize internal memory allocator */
6406   __kmp_init_allocator();
6407 
6408   /* Register the library startup via an environment variable and check to see
6409      whether another copy of the library is already registered. */
6410 
6411   __kmp_register_library_startup();
6412 
6413   /* TODO reinitialization of library */
6414   if (TCR_4(__kmp_global.g.g_done)) {
6415     KA_TRACE(10, ("__kmp_do_serial_initialize: reinitialization of library\n"));
6416   }
6417 
6418   __kmp_global.g.g_abort = 0;
6419   TCW_SYNC_4(__kmp_global.g.g_done, FALSE);
6420 
6421 /* initialize the locks */
6422 #if KMP_USE_ADAPTIVE_LOCKS
6423 #if KMP_DEBUG_ADAPTIVE_LOCKS
6424   __kmp_init_speculative_stats();
6425 #endif
6426 #endif
6427 #if KMP_STATS_ENABLED
6428   __kmp_stats_init();
6429 #endif
6430   __kmp_init_lock(&__kmp_global_lock);
6431   __kmp_init_queuing_lock(&__kmp_dispatch_lock);
6432   __kmp_init_lock(&__kmp_debug_lock);
6433   __kmp_init_atomic_lock(&__kmp_atomic_lock);
6434   __kmp_init_atomic_lock(&__kmp_atomic_lock_1i);
6435   __kmp_init_atomic_lock(&__kmp_atomic_lock_2i);
6436   __kmp_init_atomic_lock(&__kmp_atomic_lock_4i);
6437   __kmp_init_atomic_lock(&__kmp_atomic_lock_4r);
6438   __kmp_init_atomic_lock(&__kmp_atomic_lock_8i);
6439   __kmp_init_atomic_lock(&__kmp_atomic_lock_8r);
6440   __kmp_init_atomic_lock(&__kmp_atomic_lock_8c);
6441   __kmp_init_atomic_lock(&__kmp_atomic_lock_10r);
6442   __kmp_init_atomic_lock(&__kmp_atomic_lock_16r);
6443   __kmp_init_atomic_lock(&__kmp_atomic_lock_16c);
6444   __kmp_init_atomic_lock(&__kmp_atomic_lock_20c);
6445   __kmp_init_atomic_lock(&__kmp_atomic_lock_32c);
6446   __kmp_init_bootstrap_lock(&__kmp_forkjoin_lock);
6447   __kmp_init_bootstrap_lock(&__kmp_exit_lock);
6448 #if KMP_USE_MONITOR
6449   __kmp_init_bootstrap_lock(&__kmp_monitor_lock);
6450 #endif
6451   __kmp_init_bootstrap_lock(&__kmp_tp_cached_lock);
6452 
6453   /* conduct initialization and initial setup of configuration */
6454 
6455   __kmp_runtime_initialize();
6456 
6457 #if KMP_MIC_SUPPORTED
6458   __kmp_check_mic_type();
6459 #endif
6460 
6461 // Some global variable initialization moved here from kmp_env_initialize()
6462 #ifdef KMP_DEBUG
6463   kmp_diag = 0;
6464 #endif
6465   __kmp_abort_delay = 0;
6466 
6467   // From __kmp_init_dflt_team_nth()
6468   /* assume the entire machine will be used */
6469   __kmp_dflt_team_nth_ub = __kmp_xproc;
6470   if (__kmp_dflt_team_nth_ub < KMP_MIN_NTH) {
6471     __kmp_dflt_team_nth_ub = KMP_MIN_NTH;
6472   }
6473   if (__kmp_dflt_team_nth_ub > __kmp_sys_max_nth) {
6474     __kmp_dflt_team_nth_ub = __kmp_sys_max_nth;
6475   }
6476   __kmp_max_nth = __kmp_sys_max_nth;
6477   __kmp_cg_max_nth = __kmp_sys_max_nth;
6478   __kmp_teams_max_nth = __kmp_xproc; // set a "reasonable" default
6479   if (__kmp_teams_max_nth > __kmp_sys_max_nth) {
6480     __kmp_teams_max_nth = __kmp_sys_max_nth;
6481   }
6482 
6483   // Three vars below moved here from __kmp_env_initialize() "KMP_BLOCKTIME"
6484   // part
6485   __kmp_dflt_blocktime = KMP_DEFAULT_BLOCKTIME;
6486 #if KMP_USE_MONITOR
6487   __kmp_monitor_wakeups =
6488       KMP_WAKEUPS_FROM_BLOCKTIME(__kmp_dflt_blocktime, __kmp_monitor_wakeups);
6489   __kmp_bt_intervals =
6490       KMP_INTERVALS_FROM_BLOCKTIME(__kmp_dflt_blocktime, __kmp_monitor_wakeups);
6491 #endif
6492   // From "KMP_LIBRARY" part of __kmp_env_initialize()
6493   __kmp_library = library_throughput;
6494   // From KMP_SCHEDULE initialization
6495   __kmp_static = kmp_sch_static_balanced;
6496 // AC: do not use analytical here, because it is non-monotonous
6497 //__kmp_guided = kmp_sch_guided_iterative_chunked;
6498 //__kmp_auto = kmp_sch_guided_analytical_chunked; // AC: it is the default, no
6499 // need to repeat assignment
6500 // Barrier initialization. Moved here from __kmp_env_initialize() Barrier branch
6501 // bit control and barrier method control parts
6502 #if KMP_FAST_REDUCTION_BARRIER
6503 #define kmp_reduction_barrier_gather_bb ((int)1)
6504 #define kmp_reduction_barrier_release_bb ((int)1)
6505 #define kmp_reduction_barrier_gather_pat bp_hyper_bar
6506 #define kmp_reduction_barrier_release_pat bp_hyper_bar
6507 #endif // KMP_FAST_REDUCTION_BARRIER
6508   for (i = bs_plain_barrier; i < bs_last_barrier; i++) {
6509     __kmp_barrier_gather_branch_bits[i] = __kmp_barrier_gather_bb_dflt;
6510     __kmp_barrier_release_branch_bits[i] = __kmp_barrier_release_bb_dflt;
6511     __kmp_barrier_gather_pattern[i] = __kmp_barrier_gather_pat_dflt;
6512     __kmp_barrier_release_pattern[i] = __kmp_barrier_release_pat_dflt;
6513 #if KMP_FAST_REDUCTION_BARRIER
6514     if (i == bs_reduction_barrier) { // tested and confirmed on ALTIX only (
6515       // lin_64 ): hyper,1
6516       __kmp_barrier_gather_branch_bits[i] = kmp_reduction_barrier_gather_bb;
6517       __kmp_barrier_release_branch_bits[i] = kmp_reduction_barrier_release_bb;
6518       __kmp_barrier_gather_pattern[i] = kmp_reduction_barrier_gather_pat;
6519       __kmp_barrier_release_pattern[i] = kmp_reduction_barrier_release_pat;
6520     }
6521 #endif // KMP_FAST_REDUCTION_BARRIER
6522   }
6523 #if KMP_FAST_REDUCTION_BARRIER
6524 #undef kmp_reduction_barrier_release_pat
6525 #undef kmp_reduction_barrier_gather_pat
6526 #undef kmp_reduction_barrier_release_bb
6527 #undef kmp_reduction_barrier_gather_bb
6528 #endif // KMP_FAST_REDUCTION_BARRIER
6529 #if KMP_MIC_SUPPORTED
6530   if (__kmp_mic_type == mic2) { // KNC
6531     // AC: plane=3,2, forkjoin=2,1 are optimal for 240 threads on KNC
6532     __kmp_barrier_gather_branch_bits[bs_plain_barrier] = 3; // plain gather
6533     __kmp_barrier_release_branch_bits[bs_forkjoin_barrier] =
6534         1; // forkjoin release
6535     __kmp_barrier_gather_pattern[bs_forkjoin_barrier] = bp_hierarchical_bar;
6536     __kmp_barrier_release_pattern[bs_forkjoin_barrier] = bp_hierarchical_bar;
6537   }
6538 #if KMP_FAST_REDUCTION_BARRIER
6539   if (__kmp_mic_type == mic2) { // KNC
6540     __kmp_barrier_gather_pattern[bs_reduction_barrier] = bp_hierarchical_bar;
6541     __kmp_barrier_release_pattern[bs_reduction_barrier] = bp_hierarchical_bar;
6542   }
6543 #endif // KMP_FAST_REDUCTION_BARRIER
6544 #endif // KMP_MIC_SUPPORTED
6545 
6546 // From KMP_CHECKS initialization
6547 #ifdef KMP_DEBUG
6548   __kmp_env_checks = TRUE; /* development versions have the extra checks */
6549 #else
6550   __kmp_env_checks = FALSE; /* port versions do not have the extra checks */
6551 #endif
6552 
6553   // From "KMP_FOREIGN_THREADS_THREADPRIVATE" initialization
6554   __kmp_foreign_tp = TRUE;
6555 
6556   __kmp_global.g.g_dynamic = FALSE;
6557   __kmp_global.g.g_dynamic_mode = dynamic_default;
6558 
6559   __kmp_env_initialize(NULL);
6560 
6561 // Print all messages in message catalog for testing purposes.
6562 #ifdef KMP_DEBUG
6563   char const *val = __kmp_env_get("KMP_DUMP_CATALOG");
6564   if (__kmp_str_match_true(val)) {
6565     kmp_str_buf_t buffer;
6566     __kmp_str_buf_init(&buffer);
6567     __kmp_i18n_dump_catalog(&buffer);
6568     __kmp_printf("%s", buffer.str);
6569     __kmp_str_buf_free(&buffer);
6570   }
6571   __kmp_env_free(&val);
6572 #endif
6573 
6574   __kmp_threads_capacity =
6575       __kmp_initial_threads_capacity(__kmp_dflt_team_nth_ub);
6576   // Moved here from __kmp_env_initialize() "KMP_ALL_THREADPRIVATE" part
6577   __kmp_tp_capacity = __kmp_default_tp_capacity(
6578       __kmp_dflt_team_nth_ub, __kmp_max_nth, __kmp_allThreadsSpecified);
6579 
6580   // If the library is shut down properly, both pools must be NULL. Just in
6581   // case, set them to NULL -- some memory may leak, but subsequent code will
6582   // work even if pools are not freed.
6583   KMP_DEBUG_ASSERT(__kmp_thread_pool == NULL);
6584   KMP_DEBUG_ASSERT(__kmp_thread_pool_insert_pt == NULL);
6585   KMP_DEBUG_ASSERT(__kmp_team_pool == NULL);
6586   __kmp_thread_pool = NULL;
6587   __kmp_thread_pool_insert_pt = NULL;
6588   __kmp_team_pool = NULL;
6589 
6590   /* Allocate all of the variable sized records */
6591   /* NOTE: __kmp_threads_capacity entries are allocated, but the arrays are
6592    * expandable */
6593   /* Since allocation is cache-aligned, just add extra padding at the end */
6594   size =
6595       (sizeof(kmp_info_t *) + sizeof(kmp_root_t *)) * __kmp_threads_capacity +
6596       CACHE_LINE;
6597   __kmp_threads = (kmp_info_t **)__kmp_allocate(size);
6598   __kmp_root = (kmp_root_t **)((char *)__kmp_threads +
6599                                sizeof(kmp_info_t *) * __kmp_threads_capacity);
6600 
6601   /* init thread counts */
6602   KMP_DEBUG_ASSERT(__kmp_all_nth ==
6603                    0); // Asserts fail if the library is reinitializing and
6604   KMP_DEBUG_ASSERT(__kmp_nth == 0); // something was wrong in termination.
6605   __kmp_all_nth = 0;
6606   __kmp_nth = 0;
6607 
6608   /* setup the uber master thread and hierarchy */
6609   gtid = __kmp_register_root(TRUE);
6610   KA_TRACE(10, ("__kmp_do_serial_initialize  T#%d\n", gtid));
6611   KMP_ASSERT(KMP_UBER_GTID(gtid));
6612   KMP_ASSERT(KMP_INITIAL_GTID(gtid));
6613 
6614   KMP_MB(); /* Flush all pending memory write invalidates.  */
6615 
6616   __kmp_common_initialize();
6617 
6618 #if KMP_OS_UNIX
6619   /* invoke the child fork handler */
6620   __kmp_register_atfork();
6621 #endif
6622 
6623 #if !defined KMP_DYNAMIC_LIB
6624   {
6625     /* Invoke the exit handler when the program finishes, only for static
6626        library. For dynamic library, we already have _fini and DllMain. */
6627     int rc = atexit(__kmp_internal_end_atexit);
6628     if (rc != 0) {
6629       __kmp_fatal(KMP_MSG(FunctionError, "atexit()"), KMP_ERR(rc),
6630                   __kmp_msg_null);
6631     }
6632   }
6633 #endif
6634 
6635 #if KMP_HANDLE_SIGNALS
6636 #if KMP_OS_UNIX
6637   /* NOTE: make sure that this is called before the user installs their own
6638      signal handlers so that the user handlers are called first. this way they
6639      can return false, not call our handler, avoid terminating the library, and
6640      continue execution where they left off. */
6641   __kmp_install_signals(FALSE);
6642 #endif /* KMP_OS_UNIX */
6643 #if KMP_OS_WINDOWS
6644   __kmp_install_signals(TRUE);
6645 #endif /* KMP_OS_WINDOWS */
6646 #endif
6647 
6648   /* we have finished the serial initialization */
6649   __kmp_init_counter++;
6650 
6651   __kmp_init_serial = TRUE;
6652 
6653   if (__kmp_settings) {
6654     __kmp_env_print();
6655   }
6656 
6657 #if OMP_40_ENABLED
6658   if (__kmp_display_env || __kmp_display_env_verbose) {
6659     __kmp_env_print_2();
6660   }
6661 #endif // OMP_40_ENABLED
6662 
6663 #if OMPT_SUPPORT
6664   ompt_post_init();
6665 #endif
6666 
6667   KMP_MB();
6668 
6669   KA_TRACE(10, ("__kmp_do_serial_initialize: exit\n"));
6670 }
6671 
6672 void __kmp_serial_initialize(void) {
6673   if (__kmp_init_serial) {
6674     return;
6675   }
6676   __kmp_acquire_bootstrap_lock(&__kmp_initz_lock);
6677   if (__kmp_init_serial) {
6678     __kmp_release_bootstrap_lock(&__kmp_initz_lock);
6679     return;
6680   }
6681   __kmp_do_serial_initialize();
6682   __kmp_release_bootstrap_lock(&__kmp_initz_lock);
6683 }
6684 
6685 static void __kmp_do_middle_initialize(void) {
6686   int i, j;
6687   int prev_dflt_team_nth;
6688 
6689   if (!__kmp_init_serial) {
6690     __kmp_do_serial_initialize();
6691   }
6692 
6693   KA_TRACE(10, ("__kmp_middle_initialize: enter\n"));
6694 
6695   // Save the previous value for the __kmp_dflt_team_nth so that
6696   // we can avoid some reinitialization if it hasn't changed.
6697   prev_dflt_team_nth = __kmp_dflt_team_nth;
6698 
6699 #if KMP_AFFINITY_SUPPORTED
6700   // __kmp_affinity_initialize() will try to set __kmp_ncores to the
6701   // number of cores on the machine.
6702   __kmp_affinity_initialize();
6703 
6704   // Run through the __kmp_threads array and set the affinity mask
6705   // for each root thread that is currently registered with the RTL.
6706   for (i = 0; i < __kmp_threads_capacity; i++) {
6707     if (TCR_PTR(__kmp_threads[i]) != NULL) {
6708       __kmp_affinity_set_init_mask(i, TRUE);
6709     }
6710   }
6711 #endif /* KMP_AFFINITY_SUPPORTED */
6712 
6713   KMP_ASSERT(__kmp_xproc > 0);
6714   if (__kmp_avail_proc == 0) {
6715     __kmp_avail_proc = __kmp_xproc;
6716   }
6717 
6718   // If there were empty places in num_threads list (OMP_NUM_THREADS=,,2,3),
6719   // correct them now
6720   j = 0;
6721   while ((j < __kmp_nested_nth.used) && !__kmp_nested_nth.nth[j]) {
6722     __kmp_nested_nth.nth[j] = __kmp_dflt_team_nth = __kmp_dflt_team_nth_ub =
6723         __kmp_avail_proc;
6724     j++;
6725   }
6726 
6727   if (__kmp_dflt_team_nth == 0) {
6728 #ifdef KMP_DFLT_NTH_CORES
6729     // Default #threads = #cores
6730     __kmp_dflt_team_nth = __kmp_ncores;
6731     KA_TRACE(20, ("__kmp_middle_initialize: setting __kmp_dflt_team_nth = "
6732                   "__kmp_ncores (%d)\n",
6733                   __kmp_dflt_team_nth));
6734 #else
6735     // Default #threads = #available OS procs
6736     __kmp_dflt_team_nth = __kmp_avail_proc;
6737     KA_TRACE(20, ("__kmp_middle_initialize: setting __kmp_dflt_team_nth = "
6738                   "__kmp_avail_proc(%d)\n",
6739                   __kmp_dflt_team_nth));
6740 #endif /* KMP_DFLT_NTH_CORES */
6741   }
6742 
6743   if (__kmp_dflt_team_nth < KMP_MIN_NTH) {
6744     __kmp_dflt_team_nth = KMP_MIN_NTH;
6745   }
6746   if (__kmp_dflt_team_nth > __kmp_sys_max_nth) {
6747     __kmp_dflt_team_nth = __kmp_sys_max_nth;
6748   }
6749 
6750   // There's no harm in continuing if the following check fails,
6751   // but it indicates an error in the previous logic.
6752   KMP_DEBUG_ASSERT(__kmp_dflt_team_nth <= __kmp_dflt_team_nth_ub);
6753 
6754   if (__kmp_dflt_team_nth != prev_dflt_team_nth) {
6755     // Run through the __kmp_threads array and set the num threads icv for each
6756     // root thread that is currently registered with the RTL (which has not
6757     // already explicitly set its nthreads-var with a call to
6758     // omp_set_num_threads()).
6759     for (i = 0; i < __kmp_threads_capacity; i++) {
6760       kmp_info_t *thread = __kmp_threads[i];
6761       if (thread == NULL)
6762         continue;
6763       if (thread->th.th_current_task->td_icvs.nproc != 0)
6764         continue;
6765 
6766       set__nproc(__kmp_threads[i], __kmp_dflt_team_nth);
6767     }
6768   }
6769   KA_TRACE(
6770       20,
6771       ("__kmp_middle_initialize: final value for __kmp_dflt_team_nth = %d\n",
6772        __kmp_dflt_team_nth));
6773 
6774 #ifdef KMP_ADJUST_BLOCKTIME
6775   /* Adjust blocktime to zero if necessary  now that __kmp_avail_proc is set */
6776   if (!__kmp_env_blocktime && (__kmp_avail_proc > 0)) {
6777     KMP_DEBUG_ASSERT(__kmp_avail_proc > 0);
6778     if (__kmp_nth > __kmp_avail_proc) {
6779       __kmp_zero_bt = TRUE;
6780     }
6781   }
6782 #endif /* KMP_ADJUST_BLOCKTIME */
6783 
6784   /* we have finished middle initialization */
6785   TCW_SYNC_4(__kmp_init_middle, TRUE);
6786 
6787   KA_TRACE(10, ("__kmp_do_middle_initialize: exit\n"));
6788 }
6789 
6790 void __kmp_middle_initialize(void) {
6791   if (__kmp_init_middle) {
6792     return;
6793   }
6794   __kmp_acquire_bootstrap_lock(&__kmp_initz_lock);
6795   if (__kmp_init_middle) {
6796     __kmp_release_bootstrap_lock(&__kmp_initz_lock);
6797     return;
6798   }
6799   __kmp_do_middle_initialize();
6800   __kmp_release_bootstrap_lock(&__kmp_initz_lock);
6801 }
6802 
6803 void __kmp_parallel_initialize(void) {
6804   int gtid = __kmp_entry_gtid(); // this might be a new root
6805 
6806   /* synchronize parallel initialization (for sibling) */
6807   if (TCR_4(__kmp_init_parallel))
6808     return;
6809   __kmp_acquire_bootstrap_lock(&__kmp_initz_lock);
6810   if (TCR_4(__kmp_init_parallel)) {
6811     __kmp_release_bootstrap_lock(&__kmp_initz_lock);
6812     return;
6813   }
6814 
6815   /* TODO reinitialization after we have already shut down */
6816   if (TCR_4(__kmp_global.g.g_done)) {
6817     KA_TRACE(
6818         10,
6819         ("__kmp_parallel_initialize: attempt to init while shutting down\n"));
6820     __kmp_infinite_loop();
6821   }
6822 
6823   /* jc: The lock __kmp_initz_lock is already held, so calling
6824      __kmp_serial_initialize would cause a deadlock.  So we call
6825      __kmp_do_serial_initialize directly. */
6826   if (!__kmp_init_middle) {
6827     __kmp_do_middle_initialize();
6828   }
6829 
6830   /* begin initialization */
6831   KA_TRACE(10, ("__kmp_parallel_initialize: enter\n"));
6832   KMP_ASSERT(KMP_UBER_GTID(gtid));
6833 
6834 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
6835   // Save the FP control regs.
6836   // Worker threads will set theirs to these values at thread startup.
6837   __kmp_store_x87_fpu_control_word(&__kmp_init_x87_fpu_control_word);
6838   __kmp_store_mxcsr(&__kmp_init_mxcsr);
6839   __kmp_init_mxcsr &= KMP_X86_MXCSR_MASK;
6840 #endif /* KMP_ARCH_X86 || KMP_ARCH_X86_64 */
6841 
6842 #if KMP_OS_UNIX
6843 #if KMP_HANDLE_SIGNALS
6844   /*  must be after __kmp_serial_initialize  */
6845   __kmp_install_signals(TRUE);
6846 #endif
6847 #endif
6848 
6849   __kmp_suspend_initialize();
6850 
6851 #if defined(USE_LOAD_BALANCE)
6852   if (__kmp_global.g.g_dynamic_mode == dynamic_default) {
6853     __kmp_global.g.g_dynamic_mode = dynamic_load_balance;
6854   }
6855 #else
6856   if (__kmp_global.g.g_dynamic_mode == dynamic_default) {
6857     __kmp_global.g.g_dynamic_mode = dynamic_thread_limit;
6858   }
6859 #endif
6860 
6861   if (__kmp_version) {
6862     __kmp_print_version_2();
6863   }
6864 
6865   /* we have finished parallel initialization */
6866   TCW_SYNC_4(__kmp_init_parallel, TRUE);
6867 
6868   KMP_MB();
6869   KA_TRACE(10, ("__kmp_parallel_initialize: exit\n"));
6870 
6871   __kmp_release_bootstrap_lock(&__kmp_initz_lock);
6872 }
6873 
6874 /* ------------------------------------------------------------------------ */
6875 
6876 void __kmp_run_before_invoked_task(int gtid, int tid, kmp_info_t *this_thr,
6877                                    kmp_team_t *team) {
6878   kmp_disp_t *dispatch;
6879 
6880   KMP_MB();
6881 
6882   /* none of the threads have encountered any constructs, yet. */
6883   this_thr->th.th_local.this_construct = 0;
6884 #if KMP_CACHE_MANAGE
6885   KMP_CACHE_PREFETCH(&this_thr->th.th_bar[bs_forkjoin_barrier].bb.b_arrived);
6886 #endif /* KMP_CACHE_MANAGE */
6887   dispatch = (kmp_disp_t *)TCR_PTR(this_thr->th.th_dispatch);
6888   KMP_DEBUG_ASSERT(dispatch);
6889   KMP_DEBUG_ASSERT(team->t.t_dispatch);
6890   // KMP_DEBUG_ASSERT( this_thr->th.th_dispatch == &team->t.t_dispatch[
6891   // this_thr->th.th_info.ds.ds_tid ] );
6892 
6893   dispatch->th_disp_index = 0; /* reset the dispatch buffer counter */
6894 #if OMP_45_ENABLED
6895   dispatch->th_doacross_buf_idx =
6896       0; /* reset the doacross dispatch buffer counter */
6897 #endif
6898   if (__kmp_env_consistency_check)
6899     __kmp_push_parallel(gtid, team->t.t_ident);
6900 
6901   KMP_MB(); /* Flush all pending memory write invalidates.  */
6902 }
6903 
6904 void __kmp_run_after_invoked_task(int gtid, int tid, kmp_info_t *this_thr,
6905                                   kmp_team_t *team) {
6906   if (__kmp_env_consistency_check)
6907     __kmp_pop_parallel(gtid, team->t.t_ident);
6908 
6909   __kmp_finish_implicit_task(this_thr);
6910 }
6911 
6912 int __kmp_invoke_task_func(int gtid) {
6913   int rc;
6914   int tid = __kmp_tid_from_gtid(gtid);
6915   kmp_info_t *this_thr = __kmp_threads[gtid];
6916   kmp_team_t *team = this_thr->th.th_team;
6917 
6918   __kmp_run_before_invoked_task(gtid, tid, this_thr, team);
6919 #if USE_ITT_BUILD
6920   if (__itt_stack_caller_create_ptr) {
6921     __kmp_itt_stack_callee_enter(
6922         (__itt_caller)
6923             team->t.t_stack_id); // inform ittnotify about entering user's code
6924   }
6925 #endif /* USE_ITT_BUILD */
6926 #if INCLUDE_SSC_MARKS
6927   SSC_MARK_INVOKING();
6928 #endif
6929 
6930 #if OMPT_SUPPORT
6931   void *dummy;
6932   void **exit_runtime_p;
6933   ompt_data_t *my_task_data;
6934   ompt_data_t *my_parallel_data;
6935   int ompt_team_size;
6936 
6937   if (ompt_enabled.enabled) {
6938     exit_runtime_p = &(
6939         team->t.t_implicit_task_taskdata[tid].ompt_task_info.frame.exit_frame);
6940   } else {
6941     exit_runtime_p = &dummy;
6942   }
6943 
6944   my_task_data =
6945       &(team->t.t_implicit_task_taskdata[tid].ompt_task_info.task_data);
6946   my_parallel_data = &(team->t.ompt_team_info.parallel_data);
6947   if (ompt_enabled.ompt_callback_implicit_task) {
6948     ompt_team_size = team->t.t_nproc;
6949     ompt_callbacks.ompt_callback(ompt_callback_implicit_task)(
6950         ompt_scope_begin, my_parallel_data, my_task_data, ompt_team_size,
6951         __kmp_tid_from_gtid(gtid));
6952     OMPT_CUR_TASK_INFO(this_thr)->thread_num = __kmp_tid_from_gtid(gtid);
6953   }
6954 #endif
6955 
6956   {
6957     KMP_TIME_PARTITIONED_BLOCK(OMP_parallel);
6958     KMP_SET_THREAD_STATE_BLOCK(IMPLICIT_TASK);
6959     rc =
6960         __kmp_invoke_microtask((microtask_t)TCR_SYNC_PTR(team->t.t_pkfn), gtid,
6961                                tid, (int)team->t.t_argc, (void **)team->t.t_argv
6962 #if OMPT_SUPPORT
6963                                ,
6964                                exit_runtime_p
6965 #endif
6966                                );
6967 #if OMPT_SUPPORT
6968     *exit_runtime_p = NULL;
6969 #endif
6970   }
6971 
6972 #if USE_ITT_BUILD
6973   if (__itt_stack_caller_create_ptr) {
6974     __kmp_itt_stack_callee_leave(
6975         (__itt_caller)
6976             team->t.t_stack_id); // inform ittnotify about leaving user's code
6977   }
6978 #endif /* USE_ITT_BUILD */
6979   __kmp_run_after_invoked_task(gtid, tid, this_thr, team);
6980 
6981   return rc;
6982 }
6983 
6984 #if OMP_40_ENABLED
6985 void __kmp_teams_master(int gtid) {
6986   // This routine is called by all master threads in teams construct
6987   kmp_info_t *thr = __kmp_threads[gtid];
6988   kmp_team_t *team = thr->th.th_team;
6989   ident_t *loc = team->t.t_ident;
6990   thr->th.th_set_nproc = thr->th.th_teams_size.nth;
6991   KMP_DEBUG_ASSERT(thr->th.th_teams_microtask);
6992   KMP_DEBUG_ASSERT(thr->th.th_set_nproc);
6993   KA_TRACE(20, ("__kmp_teams_master: T#%d, Tid %d, microtask %p\n", gtid,
6994                 __kmp_tid_from_gtid(gtid), thr->th.th_teams_microtask));
6995 // Launch league of teams now, but not let workers execute
6996 // (they hang on fork barrier until next parallel)
6997 #if INCLUDE_SSC_MARKS
6998   SSC_MARK_FORKING();
6999 #endif
7000   __kmp_fork_call(loc, gtid, fork_context_intel, team->t.t_argc,
7001                   (microtask_t)thr->th.th_teams_microtask, // "wrapped" task
7002                   VOLATILE_CAST(launch_t) __kmp_invoke_task_func, NULL);
7003 #if INCLUDE_SSC_MARKS
7004   SSC_MARK_JOINING();
7005 #endif
7006 
7007   // AC: last parameter "1" eliminates join barrier which won't work because
7008   // worker threads are in a fork barrier waiting for more parallel regions
7009   __kmp_join_call(loc, gtid
7010 #if OMPT_SUPPORT
7011                   ,
7012                   fork_context_intel
7013 #endif
7014                   ,
7015                   1);
7016 }
7017 
7018 int __kmp_invoke_teams_master(int gtid) {
7019   kmp_info_t *this_thr = __kmp_threads[gtid];
7020   kmp_team_t *team = this_thr->th.th_team;
7021 #if KMP_DEBUG
7022   if (!__kmp_threads[gtid]->th.th_team->t.t_serialized)
7023     KMP_DEBUG_ASSERT((void *)__kmp_threads[gtid]->th.th_team->t.t_pkfn ==
7024                      (void *)__kmp_teams_master);
7025 #endif
7026   __kmp_run_before_invoked_task(gtid, 0, this_thr, team);
7027   __kmp_teams_master(gtid);
7028   __kmp_run_after_invoked_task(gtid, 0, this_thr, team);
7029   return 1;
7030 }
7031 #endif /* OMP_40_ENABLED */
7032 
7033 /* this sets the requested number of threads for the next parallel region
7034    encountered by this team. since this should be enclosed in the forkjoin
7035    critical section it should avoid race conditions with assymmetrical nested
7036    parallelism */
7037 
7038 void __kmp_push_num_threads(ident_t *id, int gtid, int num_threads) {
7039   kmp_info_t *thr = __kmp_threads[gtid];
7040 
7041   if (num_threads > 0)
7042     thr->th.th_set_nproc = num_threads;
7043 }
7044 
7045 #if OMP_40_ENABLED
7046 
7047 /* this sets the requested number of teams for the teams region and/or
7048    the number of threads for the next parallel region encountered  */
7049 void __kmp_push_num_teams(ident_t *id, int gtid, int num_teams,
7050                           int num_threads) {
7051   kmp_info_t *thr = __kmp_threads[gtid];
7052   KMP_DEBUG_ASSERT(num_teams >= 0);
7053   KMP_DEBUG_ASSERT(num_threads >= 0);
7054 
7055   if (num_teams == 0)
7056     num_teams = 1; // default number of teams is 1.
7057   if (num_teams > __kmp_teams_max_nth) { // if too many teams requested?
7058     if (!__kmp_reserve_warn) {
7059       __kmp_reserve_warn = 1;
7060       __kmp_msg(kmp_ms_warning,
7061                 KMP_MSG(CantFormThrTeam, num_teams, __kmp_teams_max_nth),
7062                 KMP_HNT(Unset_ALL_THREADS), __kmp_msg_null);
7063     }
7064     num_teams = __kmp_teams_max_nth;
7065   }
7066   // Set number of teams (number of threads in the outer "parallel" of the
7067   // teams)
7068   thr->th.th_set_nproc = thr->th.th_teams_size.nteams = num_teams;
7069 
7070   // Remember the number of threads for inner parallel regions
7071   if (num_threads == 0) {
7072     if (!TCR_4(__kmp_init_middle))
7073       __kmp_middle_initialize(); // get __kmp_avail_proc calculated
7074     num_threads = __kmp_avail_proc / num_teams;
7075     if (num_teams * num_threads > __kmp_teams_max_nth) {
7076       // adjust num_threads w/o warning as it is not user setting
7077       num_threads = __kmp_teams_max_nth / num_teams;
7078     }
7079   } else {
7080     if (num_teams * num_threads > __kmp_teams_max_nth) {
7081       int new_threads = __kmp_teams_max_nth / num_teams;
7082       if (!__kmp_reserve_warn) { // user asked for too many threads
7083         __kmp_reserve_warn = 1; // that conflicts with KMP_TEAMS_THREAD_LIMIT
7084         __kmp_msg(kmp_ms_warning,
7085                   KMP_MSG(CantFormThrTeam, num_threads, new_threads),
7086                   KMP_HNT(Unset_ALL_THREADS), __kmp_msg_null);
7087       }
7088       num_threads = new_threads;
7089     }
7090   }
7091   thr->th.th_teams_size.nth = num_threads;
7092 }
7093 
7094 // Set the proc_bind var to use in the following parallel region.
7095 void __kmp_push_proc_bind(ident_t *id, int gtid, kmp_proc_bind_t proc_bind) {
7096   kmp_info_t *thr = __kmp_threads[gtid];
7097   thr->th.th_set_proc_bind = proc_bind;
7098 }
7099 
7100 #endif /* OMP_40_ENABLED */
7101 
7102 /* Launch the worker threads into the microtask. */
7103 
7104 void __kmp_internal_fork(ident_t *id, int gtid, kmp_team_t *team) {
7105   kmp_info_t *this_thr = __kmp_threads[gtid];
7106 
7107 #ifdef KMP_DEBUG
7108   int f;
7109 #endif /* KMP_DEBUG */
7110 
7111   KMP_DEBUG_ASSERT(team);
7112   KMP_DEBUG_ASSERT(this_thr->th.th_team == team);
7113   KMP_ASSERT(KMP_MASTER_GTID(gtid));
7114   KMP_MB(); /* Flush all pending memory write invalidates.  */
7115 
7116   team->t.t_construct = 0; /* no single directives seen yet */
7117   team->t.t_ordered.dt.t_value =
7118       0; /* thread 0 enters the ordered section first */
7119 
7120   /* Reset the identifiers on the dispatch buffer */
7121   KMP_DEBUG_ASSERT(team->t.t_disp_buffer);
7122   if (team->t.t_max_nproc > 1) {
7123     int i;
7124     for (i = 0; i < __kmp_dispatch_num_buffers; ++i) {
7125       team->t.t_disp_buffer[i].buffer_index = i;
7126 #if OMP_45_ENABLED
7127       team->t.t_disp_buffer[i].doacross_buf_idx = i;
7128 #endif
7129     }
7130   } else {
7131     team->t.t_disp_buffer[0].buffer_index = 0;
7132 #if OMP_45_ENABLED
7133     team->t.t_disp_buffer[0].doacross_buf_idx = 0;
7134 #endif
7135   }
7136 
7137   KMP_MB(); /* Flush all pending memory write invalidates.  */
7138   KMP_ASSERT(this_thr->th.th_team == team);
7139 
7140 #ifdef KMP_DEBUG
7141   for (f = 0; f < team->t.t_nproc; f++) {
7142     KMP_DEBUG_ASSERT(team->t.t_threads[f] &&
7143                      team->t.t_threads[f]->th.th_team_nproc == team->t.t_nproc);
7144   }
7145 #endif /* KMP_DEBUG */
7146 
7147   /* release the worker threads so they may begin working */
7148   __kmp_fork_barrier(gtid, 0);
7149 }
7150 
7151 void __kmp_internal_join(ident_t *id, int gtid, kmp_team_t *team) {
7152   kmp_info_t *this_thr = __kmp_threads[gtid];
7153 
7154   KMP_DEBUG_ASSERT(team);
7155   KMP_DEBUG_ASSERT(this_thr->th.th_team == team);
7156   KMP_ASSERT(KMP_MASTER_GTID(gtid));
7157   KMP_MB(); /* Flush all pending memory write invalidates.  */
7158 
7159 /* Join barrier after fork */
7160 
7161 #ifdef KMP_DEBUG
7162   if (__kmp_threads[gtid] &&
7163       __kmp_threads[gtid]->th.th_team_nproc != team->t.t_nproc) {
7164     __kmp_printf("GTID: %d, __kmp_threads[%d]=%p\n", gtid, gtid,
7165                  __kmp_threads[gtid]);
7166     __kmp_printf("__kmp_threads[%d]->th.th_team_nproc=%d, TEAM: %p, "
7167                  "team->t.t_nproc=%d\n",
7168                  gtid, __kmp_threads[gtid]->th.th_team_nproc, team,
7169                  team->t.t_nproc);
7170     __kmp_print_structure();
7171   }
7172   KMP_DEBUG_ASSERT(__kmp_threads[gtid] &&
7173                    __kmp_threads[gtid]->th.th_team_nproc == team->t.t_nproc);
7174 #endif /* KMP_DEBUG */
7175 
7176   __kmp_join_barrier(gtid); /* wait for everyone */
7177 #if OMPT_SUPPORT
7178   if (ompt_enabled.enabled &&
7179       this_thr->th.ompt_thread_info.state == omp_state_wait_barrier_implicit) {
7180     int ds_tid = this_thr->th.th_info.ds.ds_tid;
7181     ompt_data_t *task_data = OMPT_CUR_TASK_DATA(this_thr);
7182     this_thr->th.ompt_thread_info.state = omp_state_overhead;
7183 #if OMPT_OPTIONAL
7184     void *codeptr = NULL;
7185     if (KMP_MASTER_TID(ds_tid) &&
7186         (ompt_callbacks.ompt_callback(ompt_callback_sync_region_wait) ||
7187          ompt_callbacks.ompt_callback(ompt_callback_sync_region)))
7188       codeptr = OMPT_CUR_TEAM_INFO(this_thr)->master_return_address;
7189 
7190     if (ompt_enabled.ompt_callback_sync_region_wait) {
7191       ompt_callbacks.ompt_callback(ompt_callback_sync_region_wait)(
7192           ompt_sync_region_barrier, ompt_scope_end, NULL, task_data, codeptr);
7193     }
7194     if (ompt_enabled.ompt_callback_sync_region) {
7195       ompt_callbacks.ompt_callback(ompt_callback_sync_region)(
7196           ompt_sync_region_barrier, ompt_scope_end, NULL, task_data, codeptr);
7197     }
7198 #endif
7199     if (!KMP_MASTER_TID(ds_tid) && ompt_enabled.ompt_callback_implicit_task) {
7200       ompt_callbacks.ompt_callback(ompt_callback_implicit_task)(
7201           ompt_scope_end, NULL, task_data, 0, ds_tid);
7202     }
7203   }
7204 #endif
7205 
7206   KMP_MB(); /* Flush all pending memory write invalidates.  */
7207   KMP_ASSERT(this_thr->th.th_team == team);
7208 }
7209 
7210 /* ------------------------------------------------------------------------ */
7211 
7212 #ifdef USE_LOAD_BALANCE
7213 
7214 // Return the worker threads actively spinning in the hot team, if we
7215 // are at the outermost level of parallelism.  Otherwise, return 0.
7216 static int __kmp_active_hot_team_nproc(kmp_root_t *root) {
7217   int i;
7218   int retval;
7219   kmp_team_t *hot_team;
7220 
7221   if (root->r.r_active) {
7222     return 0;
7223   }
7224   hot_team = root->r.r_hot_team;
7225   if (__kmp_dflt_blocktime == KMP_MAX_BLOCKTIME) {
7226     return hot_team->t.t_nproc - 1; // Don't count master thread
7227   }
7228 
7229   // Skip the master thread - it is accounted for elsewhere.
7230   retval = 0;
7231   for (i = 1; i < hot_team->t.t_nproc; i++) {
7232     if (hot_team->t.t_threads[i]->th.th_active) {
7233       retval++;
7234     }
7235   }
7236   return retval;
7237 }
7238 
7239 // Perform an automatic adjustment to the number of
7240 // threads used by the next parallel region.
7241 static int __kmp_load_balance_nproc(kmp_root_t *root, int set_nproc) {
7242   int retval;
7243   int pool_active;
7244   int hot_team_active;
7245   int team_curr_active;
7246   int system_active;
7247 
7248   KB_TRACE(20, ("__kmp_load_balance_nproc: called root:%p set_nproc:%d\n", root,
7249                 set_nproc));
7250   KMP_DEBUG_ASSERT(root);
7251   KMP_DEBUG_ASSERT(root->r.r_root_team->t.t_threads[0]
7252                        ->th.th_current_task->td_icvs.dynamic == TRUE);
7253   KMP_DEBUG_ASSERT(set_nproc > 1);
7254 
7255   if (set_nproc == 1) {
7256     KB_TRACE(20, ("__kmp_load_balance_nproc: serial execution.\n"));
7257     return 1;
7258   }
7259 
7260   // Threads that are active in the thread pool, active in the hot team for this
7261   // particular root (if we are at the outer par level), and the currently
7262   // executing thread (to become the master) are available to add to the new
7263   // team, but are currently contributing to the system load, and must be
7264   // accounted for.
7265   pool_active = __kmp_thread_pool_active_nth;
7266   hot_team_active = __kmp_active_hot_team_nproc(root);
7267   team_curr_active = pool_active + hot_team_active + 1;
7268 
7269   // Check the system load.
7270   system_active = __kmp_get_load_balance(__kmp_avail_proc + team_curr_active);
7271   KB_TRACE(30, ("__kmp_load_balance_nproc: system active = %d pool active = %d "
7272                 "hot team active = %d\n",
7273                 system_active, pool_active, hot_team_active));
7274 
7275   if (system_active < 0) {
7276     // There was an error reading the necessary info from /proc, so use the
7277     // thread limit algorithm instead. Once we set __kmp_global.g.g_dynamic_mode
7278     // = dynamic_thread_limit, we shouldn't wind up getting back here.
7279     __kmp_global.g.g_dynamic_mode = dynamic_thread_limit;
7280     KMP_WARNING(CantLoadBalUsing, "KMP_DYNAMIC_MODE=thread limit");
7281 
7282     // Make this call behave like the thread limit algorithm.
7283     retval = __kmp_avail_proc - __kmp_nth +
7284              (root->r.r_active ? 1 : root->r.r_hot_team->t.t_nproc);
7285     if (retval > set_nproc) {
7286       retval = set_nproc;
7287     }
7288     if (retval < KMP_MIN_NTH) {
7289       retval = KMP_MIN_NTH;
7290     }
7291 
7292     KB_TRACE(20, ("__kmp_load_balance_nproc: thread limit exit. retval:%d\n",
7293                   retval));
7294     return retval;
7295   }
7296 
7297   // There is a slight delay in the load balance algorithm in detecting new
7298   // running procs. The real system load at this instant should be at least as
7299   // large as the #active omp thread that are available to add to the team.
7300   if (system_active < team_curr_active) {
7301     system_active = team_curr_active;
7302   }
7303   retval = __kmp_avail_proc - system_active + team_curr_active;
7304   if (retval > set_nproc) {
7305     retval = set_nproc;
7306   }
7307   if (retval < KMP_MIN_NTH) {
7308     retval = KMP_MIN_NTH;
7309   }
7310 
7311   KB_TRACE(20, ("__kmp_load_balance_nproc: exit. retval:%d\n", retval));
7312   return retval;
7313 } // __kmp_load_balance_nproc()
7314 
7315 #endif /* USE_LOAD_BALANCE */
7316 
7317 /* ------------------------------------------------------------------------ */
7318 
7319 /* NOTE: this is called with the __kmp_init_lock held */
7320 void __kmp_cleanup(void) {
7321   int f;
7322 
7323   KA_TRACE(10, ("__kmp_cleanup: enter\n"));
7324 
7325   if (TCR_4(__kmp_init_parallel)) {
7326 #if KMP_HANDLE_SIGNALS
7327     __kmp_remove_signals();
7328 #endif
7329     TCW_4(__kmp_init_parallel, FALSE);
7330   }
7331 
7332   if (TCR_4(__kmp_init_middle)) {
7333 #if KMP_AFFINITY_SUPPORTED
7334     __kmp_affinity_uninitialize();
7335 #endif /* KMP_AFFINITY_SUPPORTED */
7336     __kmp_cleanup_hierarchy();
7337     TCW_4(__kmp_init_middle, FALSE);
7338   }
7339 
7340   KA_TRACE(10, ("__kmp_cleanup: go serial cleanup\n"));
7341 
7342   if (__kmp_init_serial) {
7343     __kmp_runtime_destroy();
7344     __kmp_init_serial = FALSE;
7345   }
7346 
7347   __kmp_cleanup_threadprivate_caches();
7348 
7349   for (f = 0; f < __kmp_threads_capacity; f++) {
7350     if (__kmp_root[f] != NULL) {
7351       __kmp_free(__kmp_root[f]);
7352       __kmp_root[f] = NULL;
7353     }
7354   }
7355   __kmp_free(__kmp_threads);
7356   // __kmp_threads and __kmp_root were allocated at once, as single block, so
7357   // there is no need in freeing __kmp_root.
7358   __kmp_threads = NULL;
7359   __kmp_root = NULL;
7360   __kmp_threads_capacity = 0;
7361 
7362 #if KMP_USE_DYNAMIC_LOCK
7363   __kmp_cleanup_indirect_user_locks();
7364 #else
7365   __kmp_cleanup_user_locks();
7366 #endif
7367 
7368 #if KMP_AFFINITY_SUPPORTED
7369   KMP_INTERNAL_FREE(CCAST(char *, __kmp_cpuinfo_file));
7370   __kmp_cpuinfo_file = NULL;
7371 #endif /* KMP_AFFINITY_SUPPORTED */
7372 
7373 #if KMP_USE_ADAPTIVE_LOCKS
7374 #if KMP_DEBUG_ADAPTIVE_LOCKS
7375   __kmp_print_speculative_stats();
7376 #endif
7377 #endif
7378   KMP_INTERNAL_FREE(__kmp_nested_nth.nth);
7379   __kmp_nested_nth.nth = NULL;
7380   __kmp_nested_nth.size = 0;
7381   __kmp_nested_nth.used = 0;
7382   KMP_INTERNAL_FREE(__kmp_nested_proc_bind.bind_types);
7383   __kmp_nested_proc_bind.bind_types = NULL;
7384   __kmp_nested_proc_bind.size = 0;
7385   __kmp_nested_proc_bind.used = 0;
7386 
7387   __kmp_i18n_catclose();
7388 
7389 #if KMP_USE_HIER_SCHED
7390   __kmp_hier_scheds.deallocate();
7391 #endif
7392 
7393 #if KMP_STATS_ENABLED
7394   __kmp_stats_fini();
7395 #endif
7396 
7397   KA_TRACE(10, ("__kmp_cleanup: exit\n"));
7398 }
7399 
7400 /* ------------------------------------------------------------------------ */
7401 
7402 int __kmp_ignore_mppbeg(void) {
7403   char *env;
7404 
7405   if ((env = getenv("KMP_IGNORE_MPPBEG")) != NULL) {
7406     if (__kmp_str_match_false(env))
7407       return FALSE;
7408   }
7409   // By default __kmpc_begin() is no-op.
7410   return TRUE;
7411 }
7412 
7413 int __kmp_ignore_mppend(void) {
7414   char *env;
7415 
7416   if ((env = getenv("KMP_IGNORE_MPPEND")) != NULL) {
7417     if (__kmp_str_match_false(env))
7418       return FALSE;
7419   }
7420   // By default __kmpc_end() is no-op.
7421   return TRUE;
7422 }
7423 
7424 void __kmp_internal_begin(void) {
7425   int gtid;
7426   kmp_root_t *root;
7427 
7428   /* this is a very important step as it will register new sibling threads
7429      and assign these new uber threads a new gtid */
7430   gtid = __kmp_entry_gtid();
7431   root = __kmp_threads[gtid]->th.th_root;
7432   KMP_ASSERT(KMP_UBER_GTID(gtid));
7433 
7434   if (root->r.r_begin)
7435     return;
7436   __kmp_acquire_lock(&root->r.r_begin_lock, gtid);
7437   if (root->r.r_begin) {
7438     __kmp_release_lock(&root->r.r_begin_lock, gtid);
7439     return;
7440   }
7441 
7442   root->r.r_begin = TRUE;
7443 
7444   __kmp_release_lock(&root->r.r_begin_lock, gtid);
7445 }
7446 
7447 /* ------------------------------------------------------------------------ */
7448 
7449 void __kmp_user_set_library(enum library_type arg) {
7450   int gtid;
7451   kmp_root_t *root;
7452   kmp_info_t *thread;
7453 
7454   /* first, make sure we are initialized so we can get our gtid */
7455 
7456   gtid = __kmp_entry_gtid();
7457   thread = __kmp_threads[gtid];
7458 
7459   root = thread->th.th_root;
7460 
7461   KA_TRACE(20, ("__kmp_user_set_library: enter T#%d, arg: %d, %d\n", gtid, arg,
7462                 library_serial));
7463   if (root->r.r_in_parallel) { /* Must be called in serial section of top-level
7464                                   thread */
7465     KMP_WARNING(SetLibraryIncorrectCall);
7466     return;
7467   }
7468 
7469   switch (arg) {
7470   case library_serial:
7471     thread->th.th_set_nproc = 0;
7472     set__nproc(thread, 1);
7473     break;
7474   case library_turnaround:
7475     thread->th.th_set_nproc = 0;
7476     set__nproc(thread, __kmp_dflt_team_nth ? __kmp_dflt_team_nth
7477                                            : __kmp_dflt_team_nth_ub);
7478     break;
7479   case library_throughput:
7480     thread->th.th_set_nproc = 0;
7481     set__nproc(thread, __kmp_dflt_team_nth ? __kmp_dflt_team_nth
7482                                            : __kmp_dflt_team_nth_ub);
7483     break;
7484   default:
7485     KMP_FATAL(UnknownLibraryType, arg);
7486   }
7487 
7488   __kmp_aux_set_library(arg);
7489 }
7490 
7491 void __kmp_aux_set_stacksize(size_t arg) {
7492   if (!__kmp_init_serial)
7493     __kmp_serial_initialize();
7494 
7495 #if KMP_OS_DARWIN
7496   if (arg & (0x1000 - 1)) {
7497     arg &= ~(0x1000 - 1);
7498     if (arg + 0x1000) /* check for overflow if we round up */
7499       arg += 0x1000;
7500   }
7501 #endif
7502   __kmp_acquire_bootstrap_lock(&__kmp_initz_lock);
7503 
7504   /* only change the default stacksize before the first parallel region */
7505   if (!TCR_4(__kmp_init_parallel)) {
7506     size_t value = arg; /* argument is in bytes */
7507 
7508     if (value < __kmp_sys_min_stksize)
7509       value = __kmp_sys_min_stksize;
7510     else if (value > KMP_MAX_STKSIZE)
7511       value = KMP_MAX_STKSIZE;
7512 
7513     __kmp_stksize = value;
7514 
7515     __kmp_env_stksize = TRUE; /* was KMP_STACKSIZE specified? */
7516   }
7517 
7518   __kmp_release_bootstrap_lock(&__kmp_initz_lock);
7519 }
7520 
7521 /* set the behaviour of the runtime library */
7522 /* TODO this can cause some odd behaviour with sibling parallelism... */
7523 void __kmp_aux_set_library(enum library_type arg) {
7524   __kmp_library = arg;
7525 
7526   switch (__kmp_library) {
7527   case library_serial: {
7528     KMP_INFORM(LibraryIsSerial);
7529     (void)__kmp_change_library(TRUE);
7530   } break;
7531   case library_turnaround:
7532     (void)__kmp_change_library(TRUE);
7533     break;
7534   case library_throughput:
7535     (void)__kmp_change_library(FALSE);
7536     break;
7537   default:
7538     KMP_FATAL(UnknownLibraryType, arg);
7539   }
7540 }
7541 
7542 /* ------------------------------------------------------------------------ */
7543 
7544 void __kmp_aux_set_blocktime(int arg, kmp_info_t *thread, int tid) {
7545   int blocktime = arg; /* argument is in milliseconds */
7546 #if KMP_USE_MONITOR
7547   int bt_intervals;
7548 #endif
7549   int bt_set;
7550 
7551   __kmp_save_internal_controls(thread);
7552 
7553   /* Normalize and set blocktime for the teams */
7554   if (blocktime < KMP_MIN_BLOCKTIME)
7555     blocktime = KMP_MIN_BLOCKTIME;
7556   else if (blocktime > KMP_MAX_BLOCKTIME)
7557     blocktime = KMP_MAX_BLOCKTIME;
7558 
7559   set__blocktime_team(thread->th.th_team, tid, blocktime);
7560   set__blocktime_team(thread->th.th_serial_team, 0, blocktime);
7561 
7562 #if KMP_USE_MONITOR
7563   /* Calculate and set blocktime intervals for the teams */
7564   bt_intervals = KMP_INTERVALS_FROM_BLOCKTIME(blocktime, __kmp_monitor_wakeups);
7565 
7566   set__bt_intervals_team(thread->th.th_team, tid, bt_intervals);
7567   set__bt_intervals_team(thread->th.th_serial_team, 0, bt_intervals);
7568 #endif
7569 
7570   /* Set whether blocktime has been set to "TRUE" */
7571   bt_set = TRUE;
7572 
7573   set__bt_set_team(thread->th.th_team, tid, bt_set);
7574   set__bt_set_team(thread->th.th_serial_team, 0, bt_set);
7575 #if KMP_USE_MONITOR
7576   KF_TRACE(10, ("kmp_set_blocktime: T#%d(%d:%d), blocktime=%d, "
7577                 "bt_intervals=%d, monitor_updates=%d\n",
7578                 __kmp_gtid_from_tid(tid, thread->th.th_team),
7579                 thread->th.th_team->t.t_id, tid, blocktime, bt_intervals,
7580                 __kmp_monitor_wakeups));
7581 #else
7582   KF_TRACE(10, ("kmp_set_blocktime: T#%d(%d:%d), blocktime=%d\n",
7583                 __kmp_gtid_from_tid(tid, thread->th.th_team),
7584                 thread->th.th_team->t.t_id, tid, blocktime));
7585 #endif
7586 }
7587 
7588 void __kmp_aux_set_defaults(char const *str, int len) {
7589   if (!__kmp_init_serial) {
7590     __kmp_serial_initialize();
7591   }
7592   __kmp_env_initialize(str);
7593 
7594   if (__kmp_settings
7595 #if OMP_40_ENABLED
7596       || __kmp_display_env || __kmp_display_env_verbose
7597 #endif // OMP_40_ENABLED
7598       ) {
7599     __kmp_env_print();
7600   }
7601 } // __kmp_aux_set_defaults
7602 
7603 /* ------------------------------------------------------------------------ */
7604 /* internal fast reduction routines */
7605 
7606 PACKED_REDUCTION_METHOD_T
7607 __kmp_determine_reduction_method(
7608     ident_t *loc, kmp_int32 global_tid, kmp_int32 num_vars, size_t reduce_size,
7609     void *reduce_data, void (*reduce_func)(void *lhs_data, void *rhs_data),
7610     kmp_critical_name *lck) {
7611 
7612   // Default reduction method: critical construct ( lck != NULL, like in current
7613   // PAROPT )
7614   // If ( reduce_data!=NULL && reduce_func!=NULL ): the tree-reduction method
7615   // can be selected by RTL
7616   // If loc->flags contains KMP_IDENT_ATOMIC_REDUCE, the atomic reduce method
7617   // can be selected by RTL
7618   // Finally, it's up to OpenMP RTL to make a decision on which method to select
7619   // among generated by PAROPT.
7620 
7621   PACKED_REDUCTION_METHOD_T retval;
7622 
7623   int team_size;
7624 
7625   KMP_DEBUG_ASSERT(loc); // it would be nice to test ( loc != 0 )
7626   KMP_DEBUG_ASSERT(lck); // it would be nice to test ( lck != 0 )
7627 
7628 #define FAST_REDUCTION_ATOMIC_METHOD_GENERATED                                 \
7629   ((loc->flags & (KMP_IDENT_ATOMIC_REDUCE)) == (KMP_IDENT_ATOMIC_REDUCE))
7630 #define FAST_REDUCTION_TREE_METHOD_GENERATED ((reduce_data) && (reduce_func))
7631 
7632   retval = critical_reduce_block;
7633 
7634   // another choice of getting a team size (with 1 dynamic deference) is slower
7635   team_size = __kmp_get_team_num_threads(global_tid);
7636   if (team_size == 1) {
7637 
7638     retval = empty_reduce_block;
7639 
7640   } else {
7641 
7642     int atomic_available = FAST_REDUCTION_ATOMIC_METHOD_GENERATED;
7643 
7644 #if KMP_ARCH_X86_64 || KMP_ARCH_PPC64 || KMP_ARCH_AARCH64 || KMP_ARCH_MIPS64
7645 
7646 #if KMP_OS_LINUX || KMP_OS_FREEBSD || KMP_OS_NETBSD || KMP_OS_WINDOWS ||       \
7647     KMP_OS_DARWIN
7648 
7649     int teamsize_cutoff = 4;
7650 
7651 #if KMP_MIC_SUPPORTED
7652     if (__kmp_mic_type != non_mic) {
7653       teamsize_cutoff = 8;
7654     }
7655 #endif
7656     int tree_available = FAST_REDUCTION_TREE_METHOD_GENERATED;
7657     if (tree_available) {
7658       if (team_size <= teamsize_cutoff) {
7659         if (atomic_available) {
7660           retval = atomic_reduce_block;
7661         }
7662       } else {
7663         retval = TREE_REDUCE_BLOCK_WITH_REDUCTION_BARRIER;
7664       }
7665     } else if (atomic_available) {
7666       retval = atomic_reduce_block;
7667     }
7668 #else
7669 #error "Unknown or unsupported OS"
7670 #endif // KMP_OS_LINUX || KMP_OS_FREEBSD || KMP_OS_NETBSD || KMP_OS_WINDOWS ||
7671 // KMP_OS_DARWIN
7672 
7673 #elif KMP_ARCH_X86 || KMP_ARCH_ARM || KMP_ARCH_AARCH || KMP_ARCH_MIPS
7674 
7675 #if KMP_OS_LINUX || KMP_OS_WINDOWS
7676 
7677     // basic tuning
7678 
7679     if (atomic_available) {
7680       if (num_vars <= 2) { // && ( team_size <= 8 ) due to false-sharing ???
7681         retval = atomic_reduce_block;
7682       }
7683     } // otherwise: use critical section
7684 
7685 #elif KMP_OS_DARWIN
7686 
7687     int tree_available = FAST_REDUCTION_TREE_METHOD_GENERATED;
7688     if (atomic_available && (num_vars <= 3)) {
7689       retval = atomic_reduce_block;
7690     } else if (tree_available) {
7691       if ((reduce_size > (9 * sizeof(kmp_real64))) &&
7692           (reduce_size < (2000 * sizeof(kmp_real64)))) {
7693         retval = TREE_REDUCE_BLOCK_WITH_PLAIN_BARRIER;
7694       }
7695     } // otherwise: use critical section
7696 
7697 #else
7698 #error "Unknown or unsupported OS"
7699 #endif
7700 
7701 #else
7702 #error "Unknown or unsupported architecture"
7703 #endif
7704   }
7705 
7706   // KMP_FORCE_REDUCTION
7707 
7708   // If the team is serialized (team_size == 1), ignore the forced reduction
7709   // method and stay with the unsynchronized method (empty_reduce_block)
7710   if (__kmp_force_reduction_method != reduction_method_not_defined &&
7711       team_size != 1) {
7712 
7713     PACKED_REDUCTION_METHOD_T forced_retval = critical_reduce_block;
7714 
7715     int atomic_available, tree_available;
7716 
7717     switch ((forced_retval = __kmp_force_reduction_method)) {
7718     case critical_reduce_block:
7719       KMP_ASSERT(lck); // lck should be != 0
7720       break;
7721 
7722     case atomic_reduce_block:
7723       atomic_available = FAST_REDUCTION_ATOMIC_METHOD_GENERATED;
7724       if (!atomic_available) {
7725         KMP_WARNING(RedMethodNotSupported, "atomic");
7726         forced_retval = critical_reduce_block;
7727       }
7728       break;
7729 
7730     case tree_reduce_block:
7731       tree_available = FAST_REDUCTION_TREE_METHOD_GENERATED;
7732       if (!tree_available) {
7733         KMP_WARNING(RedMethodNotSupported, "tree");
7734         forced_retval = critical_reduce_block;
7735       } else {
7736 #if KMP_FAST_REDUCTION_BARRIER
7737         forced_retval = TREE_REDUCE_BLOCK_WITH_REDUCTION_BARRIER;
7738 #endif
7739       }
7740       break;
7741 
7742     default:
7743       KMP_ASSERT(0); // "unsupported method specified"
7744     }
7745 
7746     retval = forced_retval;
7747   }
7748 
7749   KA_TRACE(10, ("reduction method selected=%08x\n", retval));
7750 
7751 #undef FAST_REDUCTION_TREE_METHOD_GENERATED
7752 #undef FAST_REDUCTION_ATOMIC_METHOD_GENERATED
7753 
7754   return (retval);
7755 }
7756 
7757 // this function is for testing set/get/determine reduce method
7758 kmp_int32 __kmp_get_reduce_method(void) {
7759   return ((__kmp_entry_thread()->th.th_local.packed_reduction_method) >> 8);
7760 }
7761