1 /*
2  * kmp_runtime.cpp -- KPTS runtime support library
3  */
4 
5 //===----------------------------------------------------------------------===//
6 //
7 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
8 // See https://llvm.org/LICENSE.txt for license information.
9 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
10 //
11 //===----------------------------------------------------------------------===//
12 
13 #include "kmp.h"
14 #include "kmp_affinity.h"
15 #include "kmp_atomic.h"
16 #include "kmp_environment.h"
17 #include "kmp_error.h"
18 #include "kmp_i18n.h"
19 #include "kmp_io.h"
20 #include "kmp_itt.h"
21 #include "kmp_settings.h"
22 #include "kmp_stats.h"
23 #include "kmp_str.h"
24 #include "kmp_wait_release.h"
25 #include "kmp_wrapper_getpid.h"
26 #include "kmp_dispatch.h"
27 #if KMP_USE_HIER_SCHED
28 #include "kmp_dispatch_hier.h"
29 #endif
30 
31 #if OMPT_SUPPORT
32 #include "ompt-specific.h"
33 #endif
34 
35 /* these are temporary issues to be dealt with */
36 #define KMP_USE_PRCTL 0
37 
38 #if KMP_OS_WINDOWS
39 #include <process.h>
40 #endif
41 
42 #include "tsan_annotations.h"
43 
44 #if defined(KMP_GOMP_COMPAT)
45 char const __kmp_version_alt_comp[] =
46     KMP_VERSION_PREFIX "alternative compiler support: yes";
47 #endif /* defined(KMP_GOMP_COMPAT) */
48 
49 char const __kmp_version_omp_api[] =
50     KMP_VERSION_PREFIX "API version: 5.0 (201611)";
51 
52 #ifdef KMP_DEBUG
53 char const __kmp_version_lock[] =
54     KMP_VERSION_PREFIX "lock type: run time selectable";
55 #endif /* KMP_DEBUG */
56 
57 #define KMP_MIN(x, y) ((x) < (y) ? (x) : (y))
58 
59 /* ------------------------------------------------------------------------ */
60 
61 #if KMP_USE_MONITOR
62 kmp_info_t __kmp_monitor;
63 #endif
64 
65 /* Forward declarations */
66 
67 void __kmp_cleanup(void);
68 
69 static void __kmp_initialize_info(kmp_info_t *, kmp_team_t *, int tid,
70                                   int gtid);
71 static void __kmp_initialize_team(kmp_team_t *team, int new_nproc,
72                                   kmp_internal_control_t *new_icvs,
73                                   ident_t *loc);
74 #if KMP_AFFINITY_SUPPORTED
75 static void __kmp_partition_places(kmp_team_t *team,
76                                    int update_master_only = 0);
77 #endif
78 static void __kmp_do_serial_initialize(void);
79 void __kmp_fork_barrier(int gtid, int tid);
80 void __kmp_join_barrier(int gtid);
81 void __kmp_setup_icv_copy(kmp_team_t *team, int new_nproc,
82                           kmp_internal_control_t *new_icvs, ident_t *loc);
83 
84 #ifdef USE_LOAD_BALANCE
85 static int __kmp_load_balance_nproc(kmp_root_t *root, int set_nproc);
86 #endif
87 
88 static int __kmp_expand_threads(int nNeed);
89 #if KMP_OS_WINDOWS
90 static int __kmp_unregister_root_other_thread(int gtid);
91 #endif
92 static void __kmp_unregister_library(void); // called by __kmp_internal_end()
93 static void __kmp_reap_thread(kmp_info_t *thread, int is_root);
94 kmp_info_t *__kmp_thread_pool_insert_pt = NULL;
95 
96 /* Calculate the identifier of the current thread */
97 /* fast (and somewhat portable) way to get unique identifier of executing
98    thread. Returns KMP_GTID_DNE if we haven't been assigned a gtid. */
99 int __kmp_get_global_thread_id() {
100   int i;
101   kmp_info_t **other_threads;
102   size_t stack_data;
103   char *stack_addr;
104   size_t stack_size;
105   char *stack_base;
106 
107   KA_TRACE(
108       1000,
109       ("*** __kmp_get_global_thread_id: entering, nproc=%d  all_nproc=%d\n",
110        __kmp_nth, __kmp_all_nth));
111 
112   /* JPH - to handle the case where __kmpc_end(0) is called immediately prior to
113      a parallel region, made it return KMP_GTID_DNE to force serial_initialize
114      by caller. Had to handle KMP_GTID_DNE at all call-sites, or else guarantee
115      __kmp_init_gtid for this to work. */
116 
117   if (!TCR_4(__kmp_init_gtid))
118     return KMP_GTID_DNE;
119 
120 #ifdef KMP_TDATA_GTID
121   if (TCR_4(__kmp_gtid_mode) >= 3) {
122     KA_TRACE(1000, ("*** __kmp_get_global_thread_id: using TDATA\n"));
123     return __kmp_gtid;
124   }
125 #endif
126   if (TCR_4(__kmp_gtid_mode) >= 2) {
127     KA_TRACE(1000, ("*** __kmp_get_global_thread_id: using keyed TLS\n"));
128     return __kmp_gtid_get_specific();
129   }
130   KA_TRACE(1000, ("*** __kmp_get_global_thread_id: using internal alg.\n"));
131 
132   stack_addr = (char *)&stack_data;
133   other_threads = __kmp_threads;
134 
135   /* ATT: The code below is a source of potential bugs due to unsynchronized
136      access to __kmp_threads array. For example:
137      1. Current thread loads other_threads[i] to thr and checks it, it is
138         non-NULL.
139      2. Current thread is suspended by OS.
140      3. Another thread unregisters and finishes (debug versions of free()
141         may fill memory with something like 0xEF).
142      4. Current thread is resumed.
143      5. Current thread reads junk from *thr.
144      TODO: Fix it.  --ln  */
145 
146   for (i = 0; i < __kmp_threads_capacity; i++) {
147 
148     kmp_info_t *thr = (kmp_info_t *)TCR_SYNC_PTR(other_threads[i]);
149     if (!thr)
150       continue;
151 
152     stack_size = (size_t)TCR_PTR(thr->th.th_info.ds.ds_stacksize);
153     stack_base = (char *)TCR_PTR(thr->th.th_info.ds.ds_stackbase);
154 
155     /* stack grows down -- search through all of the active threads */
156 
157     if (stack_addr <= stack_base) {
158       size_t stack_diff = stack_base - stack_addr;
159 
160       if (stack_diff <= stack_size) {
161         /* The only way we can be closer than the allocated */
162         /* stack size is if we are running on this thread. */
163         KMP_DEBUG_ASSERT(__kmp_gtid_get_specific() == i);
164         return i;
165       }
166     }
167   }
168 
169   /* get specific to try and determine our gtid */
170   KA_TRACE(1000,
171            ("*** __kmp_get_global_thread_id: internal alg. failed to find "
172             "thread, using TLS\n"));
173   i = __kmp_gtid_get_specific();
174 
175   /*fprintf( stderr, "=== %d\n", i );  */ /* GROO */
176 
177   /* if we havn't been assigned a gtid, then return code */
178   if (i < 0)
179     return i;
180 
181   /* dynamically updated stack window for uber threads to avoid get_specific
182      call */
183   if (!TCR_4(other_threads[i]->th.th_info.ds.ds_stackgrow)) {
184     KMP_FATAL(StackOverflow, i);
185   }
186 
187   stack_base = (char *)other_threads[i]->th.th_info.ds.ds_stackbase;
188   if (stack_addr > stack_base) {
189     TCW_PTR(other_threads[i]->th.th_info.ds.ds_stackbase, stack_addr);
190     TCW_PTR(other_threads[i]->th.th_info.ds.ds_stacksize,
191             other_threads[i]->th.th_info.ds.ds_stacksize + stack_addr -
192                 stack_base);
193   } else {
194     TCW_PTR(other_threads[i]->th.th_info.ds.ds_stacksize,
195             stack_base - stack_addr);
196   }
197 
198   /* Reprint stack bounds for ubermaster since they have been refined */
199   if (__kmp_storage_map) {
200     char *stack_end = (char *)other_threads[i]->th.th_info.ds.ds_stackbase;
201     char *stack_beg = stack_end - other_threads[i]->th.th_info.ds.ds_stacksize;
202     __kmp_print_storage_map_gtid(i, stack_beg, stack_end,
203                                  other_threads[i]->th.th_info.ds.ds_stacksize,
204                                  "th_%d stack (refinement)", i);
205   }
206   return i;
207 }
208 
209 int __kmp_get_global_thread_id_reg() {
210   int gtid;
211 
212   if (!__kmp_init_serial) {
213     gtid = KMP_GTID_DNE;
214   } else
215 #ifdef KMP_TDATA_GTID
216       if (TCR_4(__kmp_gtid_mode) >= 3) {
217     KA_TRACE(1000, ("*** __kmp_get_global_thread_id_reg: using TDATA\n"));
218     gtid = __kmp_gtid;
219   } else
220 #endif
221       if (TCR_4(__kmp_gtid_mode) >= 2) {
222     KA_TRACE(1000, ("*** __kmp_get_global_thread_id_reg: using keyed TLS\n"));
223     gtid = __kmp_gtid_get_specific();
224   } else {
225     KA_TRACE(1000,
226              ("*** __kmp_get_global_thread_id_reg: using internal alg.\n"));
227     gtid = __kmp_get_global_thread_id();
228   }
229 
230   /* we must be a new uber master sibling thread */
231   if (gtid == KMP_GTID_DNE) {
232     KA_TRACE(10,
233              ("__kmp_get_global_thread_id_reg: Encountered new root thread. "
234               "Registering a new gtid.\n"));
235     __kmp_acquire_bootstrap_lock(&__kmp_initz_lock);
236     if (!__kmp_init_serial) {
237       __kmp_do_serial_initialize();
238       gtid = __kmp_gtid_get_specific();
239     } else {
240       gtid = __kmp_register_root(FALSE);
241     }
242     __kmp_release_bootstrap_lock(&__kmp_initz_lock);
243     /*__kmp_printf( "+++ %d\n", gtid ); */ /* GROO */
244   }
245 
246   KMP_DEBUG_ASSERT(gtid >= 0);
247 
248   return gtid;
249 }
250 
251 /* caller must hold forkjoin_lock */
252 void __kmp_check_stack_overlap(kmp_info_t *th) {
253   int f;
254   char *stack_beg = NULL;
255   char *stack_end = NULL;
256   int gtid;
257 
258   KA_TRACE(10, ("__kmp_check_stack_overlap: called\n"));
259   if (__kmp_storage_map) {
260     stack_end = (char *)th->th.th_info.ds.ds_stackbase;
261     stack_beg = stack_end - th->th.th_info.ds.ds_stacksize;
262 
263     gtid = __kmp_gtid_from_thread(th);
264 
265     if (gtid == KMP_GTID_MONITOR) {
266       __kmp_print_storage_map_gtid(
267           gtid, stack_beg, stack_end, th->th.th_info.ds.ds_stacksize,
268           "th_%s stack (%s)", "mon",
269           (th->th.th_info.ds.ds_stackgrow) ? "initial" : "actual");
270     } else {
271       __kmp_print_storage_map_gtid(
272           gtid, stack_beg, stack_end, th->th.th_info.ds.ds_stacksize,
273           "th_%d stack (%s)", gtid,
274           (th->th.th_info.ds.ds_stackgrow) ? "initial" : "actual");
275     }
276   }
277 
278   /* No point in checking ubermaster threads since they use refinement and
279    * cannot overlap */
280   gtid = __kmp_gtid_from_thread(th);
281   if (__kmp_env_checks == TRUE && !KMP_UBER_GTID(gtid)) {
282     KA_TRACE(10,
283              ("__kmp_check_stack_overlap: performing extensive checking\n"));
284     if (stack_beg == NULL) {
285       stack_end = (char *)th->th.th_info.ds.ds_stackbase;
286       stack_beg = stack_end - th->th.th_info.ds.ds_stacksize;
287     }
288 
289     for (f = 0; f < __kmp_threads_capacity; f++) {
290       kmp_info_t *f_th = (kmp_info_t *)TCR_SYNC_PTR(__kmp_threads[f]);
291 
292       if (f_th && f_th != th) {
293         char *other_stack_end =
294             (char *)TCR_PTR(f_th->th.th_info.ds.ds_stackbase);
295         char *other_stack_beg =
296             other_stack_end - (size_t)TCR_PTR(f_th->th.th_info.ds.ds_stacksize);
297         if ((stack_beg > other_stack_beg && stack_beg < other_stack_end) ||
298             (stack_end > other_stack_beg && stack_end < other_stack_end)) {
299 
300           /* Print the other stack values before the abort */
301           if (__kmp_storage_map)
302             __kmp_print_storage_map_gtid(
303                 -1, other_stack_beg, other_stack_end,
304                 (size_t)TCR_PTR(f_th->th.th_info.ds.ds_stacksize),
305                 "th_%d stack (overlapped)", __kmp_gtid_from_thread(f_th));
306 
307           __kmp_fatal(KMP_MSG(StackOverlap), KMP_HNT(ChangeStackLimit),
308                       __kmp_msg_null);
309         }
310       }
311     }
312   }
313   KA_TRACE(10, ("__kmp_check_stack_overlap: returning\n"));
314 }
315 
316 /* ------------------------------------------------------------------------ */
317 
318 void __kmp_infinite_loop(void) {
319   static int done = FALSE;
320 
321   while (!done) {
322     KMP_YIELD(TRUE);
323   }
324 }
325 
326 #define MAX_MESSAGE 512
327 
328 void __kmp_print_storage_map_gtid(int gtid, void *p1, void *p2, size_t size,
329                                   char const *format, ...) {
330   char buffer[MAX_MESSAGE];
331   va_list ap;
332 
333   va_start(ap, format);
334   KMP_SNPRINTF(buffer, sizeof(buffer), "OMP storage map: %p %p%8lu %s\n", p1,
335                p2, (unsigned long)size, format);
336   __kmp_acquire_bootstrap_lock(&__kmp_stdio_lock);
337   __kmp_vprintf(kmp_err, buffer, ap);
338 #if KMP_PRINT_DATA_PLACEMENT
339   int node;
340   if (gtid >= 0) {
341     if (p1 <= p2 && (char *)p2 - (char *)p1 == size) {
342       if (__kmp_storage_map_verbose) {
343         node = __kmp_get_host_node(p1);
344         if (node < 0) /* doesn't work, so don't try this next time */
345           __kmp_storage_map_verbose = FALSE;
346         else {
347           char *last;
348           int lastNode;
349           int localProc = __kmp_get_cpu_from_gtid(gtid);
350 
351           const int page_size = KMP_GET_PAGE_SIZE();
352 
353           p1 = (void *)((size_t)p1 & ~((size_t)page_size - 1));
354           p2 = (void *)(((size_t)p2 - 1) & ~((size_t)page_size - 1));
355           if (localProc >= 0)
356             __kmp_printf_no_lock("  GTID %d localNode %d\n", gtid,
357                                  localProc >> 1);
358           else
359             __kmp_printf_no_lock("  GTID %d\n", gtid);
360 #if KMP_USE_PRCTL
361           /* The more elaborate format is disabled for now because of the prctl
362            * hanging bug. */
363           do {
364             last = p1;
365             lastNode = node;
366             /* This loop collates adjacent pages with the same host node. */
367             do {
368               (char *)p1 += page_size;
369             } while (p1 <= p2 && (node = __kmp_get_host_node(p1)) == lastNode);
370             __kmp_printf_no_lock("    %p-%p memNode %d\n", last, (char *)p1 - 1,
371                                  lastNode);
372           } while (p1 <= p2);
373 #else
374           __kmp_printf_no_lock("    %p-%p memNode %d\n", p1,
375                                (char *)p1 + (page_size - 1),
376                                __kmp_get_host_node(p1));
377           if (p1 < p2) {
378             __kmp_printf_no_lock("    %p-%p memNode %d\n", p2,
379                                  (char *)p2 + (page_size - 1),
380                                  __kmp_get_host_node(p2));
381           }
382 #endif
383         }
384       }
385     } else
386       __kmp_printf_no_lock("  %s\n", KMP_I18N_STR(StorageMapWarning));
387   }
388 #endif /* KMP_PRINT_DATA_PLACEMENT */
389   __kmp_release_bootstrap_lock(&__kmp_stdio_lock);
390 }
391 
392 void __kmp_warn(char const *format, ...) {
393   char buffer[MAX_MESSAGE];
394   va_list ap;
395 
396   if (__kmp_generate_warnings == kmp_warnings_off) {
397     return;
398   }
399 
400   va_start(ap, format);
401 
402   KMP_SNPRINTF(buffer, sizeof(buffer), "OMP warning: %s\n", format);
403   __kmp_acquire_bootstrap_lock(&__kmp_stdio_lock);
404   __kmp_vprintf(kmp_err, buffer, ap);
405   __kmp_release_bootstrap_lock(&__kmp_stdio_lock);
406 
407   va_end(ap);
408 }
409 
410 void __kmp_abort_process() {
411   // Later threads may stall here, but that's ok because abort() will kill them.
412   __kmp_acquire_bootstrap_lock(&__kmp_exit_lock);
413 
414   if (__kmp_debug_buf) {
415     __kmp_dump_debug_buffer();
416   }
417 
418   if (KMP_OS_WINDOWS) {
419     // Let other threads know of abnormal termination and prevent deadlock
420     // if abort happened during library initialization or shutdown
421     __kmp_global.g.g_abort = SIGABRT;
422 
423     /* On Windows* OS by default abort() causes pop-up error box, which stalls
424        nightly testing. Unfortunately, we cannot reliably suppress pop-up error
425        boxes. _set_abort_behavior() works well, but this function is not
426        available in VS7 (this is not problem for DLL, but it is a problem for
427        static OpenMP RTL). SetErrorMode (and so, timelimit utility) does not
428        help, at least in some versions of MS C RTL.
429 
430        It seems following sequence is the only way to simulate abort() and
431        avoid pop-up error box. */
432     raise(SIGABRT);
433     _exit(3); // Just in case, if signal ignored, exit anyway.
434   } else {
435     abort();
436   }
437 
438   __kmp_infinite_loop();
439   __kmp_release_bootstrap_lock(&__kmp_exit_lock);
440 
441 } // __kmp_abort_process
442 
443 void __kmp_abort_thread(void) {
444   // TODO: Eliminate g_abort global variable and this function.
445   // In case of abort just call abort(), it will kill all the threads.
446   __kmp_infinite_loop();
447 } // __kmp_abort_thread
448 
449 /* Print out the storage map for the major kmp_info_t thread data structures
450    that are allocated together. */
451 
452 static void __kmp_print_thread_storage_map(kmp_info_t *thr, int gtid) {
453   __kmp_print_storage_map_gtid(gtid, thr, thr + 1, sizeof(kmp_info_t), "th_%d",
454                                gtid);
455 
456   __kmp_print_storage_map_gtid(gtid, &thr->th.th_info, &thr->th.th_team,
457                                sizeof(kmp_desc_t), "th_%d.th_info", gtid);
458 
459   __kmp_print_storage_map_gtid(gtid, &thr->th.th_local, &thr->th.th_pri_head,
460                                sizeof(kmp_local_t), "th_%d.th_local", gtid);
461 
462   __kmp_print_storage_map_gtid(
463       gtid, &thr->th.th_bar[0], &thr->th.th_bar[bs_last_barrier],
464       sizeof(kmp_balign_t) * bs_last_barrier, "th_%d.th_bar", gtid);
465 
466   __kmp_print_storage_map_gtid(gtid, &thr->th.th_bar[bs_plain_barrier],
467                                &thr->th.th_bar[bs_plain_barrier + 1],
468                                sizeof(kmp_balign_t), "th_%d.th_bar[plain]",
469                                gtid);
470 
471   __kmp_print_storage_map_gtid(gtid, &thr->th.th_bar[bs_forkjoin_barrier],
472                                &thr->th.th_bar[bs_forkjoin_barrier + 1],
473                                sizeof(kmp_balign_t), "th_%d.th_bar[forkjoin]",
474                                gtid);
475 
476 #if KMP_FAST_REDUCTION_BARRIER
477   __kmp_print_storage_map_gtid(gtid, &thr->th.th_bar[bs_reduction_barrier],
478                                &thr->th.th_bar[bs_reduction_barrier + 1],
479                                sizeof(kmp_balign_t), "th_%d.th_bar[reduction]",
480                                gtid);
481 #endif // KMP_FAST_REDUCTION_BARRIER
482 }
483 
484 /* Print out the storage map for the major kmp_team_t team data structures
485    that are allocated together. */
486 
487 static void __kmp_print_team_storage_map(const char *header, kmp_team_t *team,
488                                          int team_id, int num_thr) {
489   int num_disp_buff = team->t.t_max_nproc > 1 ? __kmp_dispatch_num_buffers : 2;
490   __kmp_print_storage_map_gtid(-1, team, team + 1, sizeof(kmp_team_t), "%s_%d",
491                                header, team_id);
492 
493   __kmp_print_storage_map_gtid(-1, &team->t.t_bar[0],
494                                &team->t.t_bar[bs_last_barrier],
495                                sizeof(kmp_balign_team_t) * bs_last_barrier,
496                                "%s_%d.t_bar", header, team_id);
497 
498   __kmp_print_storage_map_gtid(-1, &team->t.t_bar[bs_plain_barrier],
499                                &team->t.t_bar[bs_plain_barrier + 1],
500                                sizeof(kmp_balign_team_t), "%s_%d.t_bar[plain]",
501                                header, team_id);
502 
503   __kmp_print_storage_map_gtid(-1, &team->t.t_bar[bs_forkjoin_barrier],
504                                &team->t.t_bar[bs_forkjoin_barrier + 1],
505                                sizeof(kmp_balign_team_t),
506                                "%s_%d.t_bar[forkjoin]", header, team_id);
507 
508 #if KMP_FAST_REDUCTION_BARRIER
509   __kmp_print_storage_map_gtid(-1, &team->t.t_bar[bs_reduction_barrier],
510                                &team->t.t_bar[bs_reduction_barrier + 1],
511                                sizeof(kmp_balign_team_t),
512                                "%s_%d.t_bar[reduction]", header, team_id);
513 #endif // KMP_FAST_REDUCTION_BARRIER
514 
515   __kmp_print_storage_map_gtid(
516       -1, &team->t.t_dispatch[0], &team->t.t_dispatch[num_thr],
517       sizeof(kmp_disp_t) * num_thr, "%s_%d.t_dispatch", header, team_id);
518 
519   __kmp_print_storage_map_gtid(
520       -1, &team->t.t_threads[0], &team->t.t_threads[num_thr],
521       sizeof(kmp_info_t *) * num_thr, "%s_%d.t_threads", header, team_id);
522 
523   __kmp_print_storage_map_gtid(-1, &team->t.t_disp_buffer[0],
524                                &team->t.t_disp_buffer[num_disp_buff],
525                                sizeof(dispatch_shared_info_t) * num_disp_buff,
526                                "%s_%d.t_disp_buffer", header, team_id);
527 }
528 
529 static void __kmp_init_allocator() { __kmp_init_memkind(); }
530 static void __kmp_fini_allocator() { __kmp_fini_memkind(); }
531 
532 /* ------------------------------------------------------------------------ */
533 
534 #if KMP_DYNAMIC_LIB
535 #if KMP_OS_WINDOWS
536 
537 static void __kmp_reset_lock(kmp_bootstrap_lock_t *lck) {
538   // TODO: Change to __kmp_break_bootstrap_lock().
539   __kmp_init_bootstrap_lock(lck); // make the lock released
540 }
541 
542 static void __kmp_reset_locks_on_process_detach(int gtid_req) {
543   int i;
544   int thread_count;
545 
546   // PROCESS_DETACH is expected to be called by a thread that executes
547   // ProcessExit() or FreeLibrary(). OS terminates other threads (except the one
548   // calling ProcessExit or FreeLibrary). So, it might be safe to access the
549   // __kmp_threads[] without taking the forkjoin_lock. However, in fact, some
550   // threads can be still alive here, although being about to be terminated. The
551   // threads in the array with ds_thread==0 are most suspicious. Actually, it
552   // can be not safe to access the __kmp_threads[].
553 
554   // TODO: does it make sense to check __kmp_roots[] ?
555 
556   // Let's check that there are no other alive threads registered with the OMP
557   // lib.
558   while (1) {
559     thread_count = 0;
560     for (i = 0; i < __kmp_threads_capacity; ++i) {
561       if (!__kmp_threads)
562         continue;
563       kmp_info_t *th = __kmp_threads[i];
564       if (th == NULL)
565         continue;
566       int gtid = th->th.th_info.ds.ds_gtid;
567       if (gtid == gtid_req)
568         continue;
569       if (gtid < 0)
570         continue;
571       DWORD exit_val;
572       int alive = __kmp_is_thread_alive(th, &exit_val);
573       if (alive) {
574         ++thread_count;
575       }
576     }
577     if (thread_count == 0)
578       break; // success
579   }
580 
581   // Assume that I'm alone. Now it might be safe to check and reset locks.
582   // __kmp_forkjoin_lock and __kmp_stdio_lock are expected to be reset.
583   __kmp_reset_lock(&__kmp_forkjoin_lock);
584 #ifdef KMP_DEBUG
585   __kmp_reset_lock(&__kmp_stdio_lock);
586 #endif // KMP_DEBUG
587 }
588 
589 BOOL WINAPI DllMain(HINSTANCE hInstDLL, DWORD fdwReason, LPVOID lpReserved) {
590   //__kmp_acquire_bootstrap_lock( &__kmp_initz_lock );
591 
592   switch (fdwReason) {
593 
594   case DLL_PROCESS_ATTACH:
595     KA_TRACE(10, ("DllMain: PROCESS_ATTACH\n"));
596 
597     return TRUE;
598 
599   case DLL_PROCESS_DETACH:
600     KA_TRACE(10, ("DllMain: PROCESS_DETACH T#%d\n", __kmp_gtid_get_specific()));
601 
602     if (lpReserved != NULL) {
603       // lpReserved is used for telling the difference:
604       //   lpReserved == NULL when FreeLibrary() was called,
605       //   lpReserved != NULL when the process terminates.
606       // When FreeLibrary() is called, worker threads remain alive. So they will
607       // release the forkjoin lock by themselves. When the process terminates,
608       // worker threads disappear triggering the problem of unreleased forkjoin
609       // lock as described below.
610 
611       // A worker thread can take the forkjoin lock. The problem comes up if
612       // that worker thread becomes dead before it releases the forkjoin lock.
613       // The forkjoin lock remains taken, while the thread executing
614       // DllMain()->PROCESS_DETACH->__kmp_internal_end_library() below will try
615       // to take the forkjoin lock and will always fail, so that the application
616       // will never finish [normally]. This scenario is possible if
617       // __kmpc_end() has not been executed. It looks like it's not a corner
618       // case, but common cases:
619       // - the main function was compiled by an alternative compiler;
620       // - the main function was compiled by icl but without /Qopenmp
621       //   (application with plugins);
622       // - application terminates by calling C exit(), Fortran CALL EXIT() or
623       //   Fortran STOP.
624       // - alive foreign thread prevented __kmpc_end from doing cleanup.
625       //
626       // This is a hack to work around the problem.
627       // TODO: !!! figure out something better.
628       __kmp_reset_locks_on_process_detach(__kmp_gtid_get_specific());
629     }
630 
631     __kmp_internal_end_library(__kmp_gtid_get_specific());
632 
633     return TRUE;
634 
635   case DLL_THREAD_ATTACH:
636     KA_TRACE(10, ("DllMain: THREAD_ATTACH\n"));
637 
638     /* if we want to register new siblings all the time here call
639      * __kmp_get_gtid(); */
640     return TRUE;
641 
642   case DLL_THREAD_DETACH:
643     KA_TRACE(10, ("DllMain: THREAD_DETACH T#%d\n", __kmp_gtid_get_specific()));
644 
645     __kmp_internal_end_thread(__kmp_gtid_get_specific());
646     return TRUE;
647   }
648 
649   return TRUE;
650 }
651 
652 #endif /* KMP_OS_WINDOWS */
653 #endif /* KMP_DYNAMIC_LIB */
654 
655 /* __kmp_parallel_deo -- Wait until it's our turn. */
656 void __kmp_parallel_deo(int *gtid_ref, int *cid_ref, ident_t *loc_ref) {
657   int gtid = *gtid_ref;
658 #ifdef BUILD_PARALLEL_ORDERED
659   kmp_team_t *team = __kmp_team_from_gtid(gtid);
660 #endif /* BUILD_PARALLEL_ORDERED */
661 
662   if (__kmp_env_consistency_check) {
663     if (__kmp_threads[gtid]->th.th_root->r.r_active)
664 #if KMP_USE_DYNAMIC_LOCK
665       __kmp_push_sync(gtid, ct_ordered_in_parallel, loc_ref, NULL, 0);
666 #else
667       __kmp_push_sync(gtid, ct_ordered_in_parallel, loc_ref, NULL);
668 #endif
669   }
670 #ifdef BUILD_PARALLEL_ORDERED
671   if (!team->t.t_serialized) {
672     KMP_MB();
673     KMP_WAIT(&team->t.t_ordered.dt.t_value, __kmp_tid_from_gtid(gtid), KMP_EQ,
674              NULL);
675     KMP_MB();
676   }
677 #endif /* BUILD_PARALLEL_ORDERED */
678 }
679 
680 /* __kmp_parallel_dxo -- Signal the next task. */
681 void __kmp_parallel_dxo(int *gtid_ref, int *cid_ref, ident_t *loc_ref) {
682   int gtid = *gtid_ref;
683 #ifdef BUILD_PARALLEL_ORDERED
684   int tid = __kmp_tid_from_gtid(gtid);
685   kmp_team_t *team = __kmp_team_from_gtid(gtid);
686 #endif /* BUILD_PARALLEL_ORDERED */
687 
688   if (__kmp_env_consistency_check) {
689     if (__kmp_threads[gtid]->th.th_root->r.r_active)
690       __kmp_pop_sync(gtid, ct_ordered_in_parallel, loc_ref);
691   }
692 #ifdef BUILD_PARALLEL_ORDERED
693   if (!team->t.t_serialized) {
694     KMP_MB(); /* Flush all pending memory write invalidates.  */
695 
696     /* use the tid of the next thread in this team */
697     /* TODO replace with general release procedure */
698     team->t.t_ordered.dt.t_value = ((tid + 1) % team->t.t_nproc);
699 
700     KMP_MB(); /* Flush all pending memory write invalidates.  */
701   }
702 #endif /* BUILD_PARALLEL_ORDERED */
703 }
704 
705 /* ------------------------------------------------------------------------ */
706 /* The BARRIER for a SINGLE process section is always explicit   */
707 
708 int __kmp_enter_single(int gtid, ident_t *id_ref, int push_ws) {
709   int status;
710   kmp_info_t *th;
711   kmp_team_t *team;
712 
713   if (!TCR_4(__kmp_init_parallel))
714     __kmp_parallel_initialize();
715   __kmp_resume_if_soft_paused();
716 
717   th = __kmp_threads[gtid];
718   team = th->th.th_team;
719   status = 0;
720 
721   th->th.th_ident = id_ref;
722 
723   if (team->t.t_serialized) {
724     status = 1;
725   } else {
726     kmp_int32 old_this = th->th.th_local.this_construct;
727 
728     ++th->th.th_local.this_construct;
729     /* try to set team count to thread count--success means thread got the
730        single block */
731     /* TODO: Should this be acquire or release? */
732     if (team->t.t_construct == old_this) {
733       status = __kmp_atomic_compare_store_acq(&team->t.t_construct, old_this,
734                                               th->th.th_local.this_construct);
735     }
736 #if USE_ITT_BUILD
737     if (__itt_metadata_add_ptr && __kmp_forkjoin_frames_mode == 3 &&
738         KMP_MASTER_GTID(gtid) && th->th.th_teams_microtask == NULL &&
739         team->t.t_active_level ==
740             1) { // Only report metadata by master of active team at level 1
741       __kmp_itt_metadata_single(id_ref);
742     }
743 #endif /* USE_ITT_BUILD */
744   }
745 
746   if (__kmp_env_consistency_check) {
747     if (status && push_ws) {
748       __kmp_push_workshare(gtid, ct_psingle, id_ref);
749     } else {
750       __kmp_check_workshare(gtid, ct_psingle, id_ref);
751     }
752   }
753 #if USE_ITT_BUILD
754   if (status) {
755     __kmp_itt_single_start(gtid);
756   }
757 #endif /* USE_ITT_BUILD */
758   return status;
759 }
760 
761 void __kmp_exit_single(int gtid) {
762 #if USE_ITT_BUILD
763   __kmp_itt_single_end(gtid);
764 #endif /* USE_ITT_BUILD */
765   if (__kmp_env_consistency_check)
766     __kmp_pop_workshare(gtid, ct_psingle, NULL);
767 }
768 
769 /* determine if we can go parallel or must use a serialized parallel region and
770  * how many threads we can use
771  * set_nproc is the number of threads requested for the team
772  * returns 0 if we should serialize or only use one thread,
773  * otherwise the number of threads to use
774  * The forkjoin lock is held by the caller. */
775 static int __kmp_reserve_threads(kmp_root_t *root, kmp_team_t *parent_team,
776                                  int master_tid, int set_nthreads,
777                                  int enter_teams) {
778   int capacity;
779   int new_nthreads;
780   KMP_DEBUG_ASSERT(__kmp_init_serial);
781   KMP_DEBUG_ASSERT(root && parent_team);
782   kmp_info_t *this_thr = parent_team->t.t_threads[master_tid];
783 
784   // If dyn-var is set, dynamically adjust the number of desired threads,
785   // according to the method specified by dynamic_mode.
786   new_nthreads = set_nthreads;
787   if (!get__dynamic_2(parent_team, master_tid)) {
788     ;
789   }
790 #ifdef USE_LOAD_BALANCE
791   else if (__kmp_global.g.g_dynamic_mode == dynamic_load_balance) {
792     new_nthreads = __kmp_load_balance_nproc(root, set_nthreads);
793     if (new_nthreads == 1) {
794       KC_TRACE(10, ("__kmp_reserve_threads: T#%d load balance reduced "
795                     "reservation to 1 thread\n",
796                     master_tid));
797       return 1;
798     }
799     if (new_nthreads < set_nthreads) {
800       KC_TRACE(10, ("__kmp_reserve_threads: T#%d load balance reduced "
801                     "reservation to %d threads\n",
802                     master_tid, new_nthreads));
803     }
804   }
805 #endif /* USE_LOAD_BALANCE */
806   else if (__kmp_global.g.g_dynamic_mode == dynamic_thread_limit) {
807     new_nthreads = __kmp_avail_proc - __kmp_nth +
808                    (root->r.r_active ? 1 : root->r.r_hot_team->t.t_nproc);
809     if (new_nthreads <= 1) {
810       KC_TRACE(10, ("__kmp_reserve_threads: T#%d thread limit reduced "
811                     "reservation to 1 thread\n",
812                     master_tid));
813       return 1;
814     }
815     if (new_nthreads < set_nthreads) {
816       KC_TRACE(10, ("__kmp_reserve_threads: T#%d thread limit reduced "
817                     "reservation to %d threads\n",
818                     master_tid, new_nthreads));
819     } else {
820       new_nthreads = set_nthreads;
821     }
822   } else if (__kmp_global.g.g_dynamic_mode == dynamic_random) {
823     if (set_nthreads > 2) {
824       new_nthreads = __kmp_get_random(parent_team->t.t_threads[master_tid]);
825       new_nthreads = (new_nthreads % set_nthreads) + 1;
826       if (new_nthreads == 1) {
827         KC_TRACE(10, ("__kmp_reserve_threads: T#%d dynamic random reduced "
828                       "reservation to 1 thread\n",
829                       master_tid));
830         return 1;
831       }
832       if (new_nthreads < set_nthreads) {
833         KC_TRACE(10, ("__kmp_reserve_threads: T#%d dynamic random reduced "
834                       "reservation to %d threads\n",
835                       master_tid, new_nthreads));
836       }
837     }
838   } else {
839     KMP_ASSERT(0);
840   }
841 
842   // Respect KMP_ALL_THREADS/KMP_DEVICE_THREAD_LIMIT.
843   if (__kmp_nth + new_nthreads -
844           (root->r.r_active ? 1 : root->r.r_hot_team->t.t_nproc) >
845       __kmp_max_nth) {
846     int tl_nthreads = __kmp_max_nth - __kmp_nth +
847                       (root->r.r_active ? 1 : root->r.r_hot_team->t.t_nproc);
848     if (tl_nthreads <= 0) {
849       tl_nthreads = 1;
850     }
851 
852     // If dyn-var is false, emit a 1-time warning.
853     if (!get__dynamic_2(parent_team, master_tid) && (!__kmp_reserve_warn)) {
854       __kmp_reserve_warn = 1;
855       __kmp_msg(kmp_ms_warning,
856                 KMP_MSG(CantFormThrTeam, set_nthreads, tl_nthreads),
857                 KMP_HNT(Unset_ALL_THREADS), __kmp_msg_null);
858     }
859     if (tl_nthreads == 1) {
860       KC_TRACE(10, ("__kmp_reserve_threads: T#%d KMP_DEVICE_THREAD_LIMIT "
861                     "reduced reservation to 1 thread\n",
862                     master_tid));
863       return 1;
864     }
865     KC_TRACE(10, ("__kmp_reserve_threads: T#%d KMP_DEVICE_THREAD_LIMIT reduced "
866                   "reservation to %d threads\n",
867                   master_tid, tl_nthreads));
868     new_nthreads = tl_nthreads;
869   }
870 
871   // Respect OMP_THREAD_LIMIT
872   int cg_nthreads = this_thr->th.th_cg_roots->cg_nthreads;
873   int max_cg_threads = this_thr->th.th_cg_roots->cg_thread_limit;
874   if (cg_nthreads + new_nthreads -
875           (root->r.r_active ? 1 : root->r.r_hot_team->t.t_nproc) >
876       max_cg_threads) {
877     int tl_nthreads = max_cg_threads - cg_nthreads +
878                       (root->r.r_active ? 1 : root->r.r_hot_team->t.t_nproc);
879     if (tl_nthreads <= 0) {
880       tl_nthreads = 1;
881     }
882 
883     // If dyn-var is false, emit a 1-time warning.
884     if (!get__dynamic_2(parent_team, master_tid) && (!__kmp_reserve_warn)) {
885       __kmp_reserve_warn = 1;
886       __kmp_msg(kmp_ms_warning,
887                 KMP_MSG(CantFormThrTeam, set_nthreads, tl_nthreads),
888                 KMP_HNT(Unset_ALL_THREADS), __kmp_msg_null);
889     }
890     if (tl_nthreads == 1) {
891       KC_TRACE(10, ("__kmp_reserve_threads: T#%d OMP_THREAD_LIMIT "
892                     "reduced reservation to 1 thread\n",
893                     master_tid));
894       return 1;
895     }
896     KC_TRACE(10, ("__kmp_reserve_threads: T#%d OMP_THREAD_LIMIT reduced "
897                   "reservation to %d threads\n",
898                   master_tid, tl_nthreads));
899     new_nthreads = tl_nthreads;
900   }
901 
902   // Check if the threads array is large enough, or needs expanding.
903   // See comment in __kmp_register_root() about the adjustment if
904   // __kmp_threads[0] == NULL.
905   capacity = __kmp_threads_capacity;
906   if (TCR_PTR(__kmp_threads[0]) == NULL) {
907     --capacity;
908   }
909   if (__kmp_nth + new_nthreads -
910           (root->r.r_active ? 1 : root->r.r_hot_team->t.t_nproc) >
911       capacity) {
912     // Expand the threads array.
913     int slotsRequired = __kmp_nth + new_nthreads -
914                         (root->r.r_active ? 1 : root->r.r_hot_team->t.t_nproc) -
915                         capacity;
916     int slotsAdded = __kmp_expand_threads(slotsRequired);
917     if (slotsAdded < slotsRequired) {
918       // The threads array was not expanded enough.
919       new_nthreads -= (slotsRequired - slotsAdded);
920       KMP_ASSERT(new_nthreads >= 1);
921 
922       // If dyn-var is false, emit a 1-time warning.
923       if (!get__dynamic_2(parent_team, master_tid) && (!__kmp_reserve_warn)) {
924         __kmp_reserve_warn = 1;
925         if (__kmp_tp_cached) {
926           __kmp_msg(kmp_ms_warning,
927                     KMP_MSG(CantFormThrTeam, set_nthreads, new_nthreads),
928                     KMP_HNT(Set_ALL_THREADPRIVATE, __kmp_tp_capacity),
929                     KMP_HNT(PossibleSystemLimitOnThreads), __kmp_msg_null);
930         } else {
931           __kmp_msg(kmp_ms_warning,
932                     KMP_MSG(CantFormThrTeam, set_nthreads, new_nthreads),
933                     KMP_HNT(SystemLimitOnThreads), __kmp_msg_null);
934         }
935       }
936     }
937   }
938 
939 #ifdef KMP_DEBUG
940   if (new_nthreads == 1) {
941     KC_TRACE(10,
942              ("__kmp_reserve_threads: T#%d serializing team after reclaiming "
943               "dead roots and rechecking; requested %d threads\n",
944               __kmp_get_gtid(), set_nthreads));
945   } else {
946     KC_TRACE(10, ("__kmp_reserve_threads: T#%d allocating %d threads; requested"
947                   " %d threads\n",
948                   __kmp_get_gtid(), new_nthreads, set_nthreads));
949   }
950 #endif // KMP_DEBUG
951   return new_nthreads;
952 }
953 
954 /* Allocate threads from the thread pool and assign them to the new team. We are
955    assured that there are enough threads available, because we checked on that
956    earlier within critical section forkjoin */
957 static void __kmp_fork_team_threads(kmp_root_t *root, kmp_team_t *team,
958                                     kmp_info_t *master_th, int master_gtid) {
959   int i;
960   int use_hot_team;
961 
962   KA_TRACE(10, ("__kmp_fork_team_threads: new_nprocs = %d\n", team->t.t_nproc));
963   KMP_DEBUG_ASSERT(master_gtid == __kmp_get_gtid());
964   KMP_MB();
965 
966   /* first, let's setup the master thread */
967   master_th->th.th_info.ds.ds_tid = 0;
968   master_th->th.th_team = team;
969   master_th->th.th_team_nproc = team->t.t_nproc;
970   master_th->th.th_team_master = master_th;
971   master_th->th.th_team_serialized = FALSE;
972   master_th->th.th_dispatch = &team->t.t_dispatch[0];
973 
974 /* make sure we are not the optimized hot team */
975 #if KMP_NESTED_HOT_TEAMS
976   use_hot_team = 0;
977   kmp_hot_team_ptr_t *hot_teams = master_th->th.th_hot_teams;
978   if (hot_teams) { // hot teams array is not allocated if
979     // KMP_HOT_TEAMS_MAX_LEVEL=0
980     int level = team->t.t_active_level - 1; // index in array of hot teams
981     if (master_th->th.th_teams_microtask) { // are we inside the teams?
982       if (master_th->th.th_teams_size.nteams > 1) {
983         ++level; // level was not increased in teams construct for
984         // team_of_masters
985       }
986       if (team->t.t_pkfn != (microtask_t)__kmp_teams_master &&
987           master_th->th.th_teams_level == team->t.t_level) {
988         ++level; // level was not increased in teams construct for
989         // team_of_workers before the parallel
990       } // team->t.t_level will be increased inside parallel
991     }
992     if (level < __kmp_hot_teams_max_level) {
993       if (hot_teams[level].hot_team) {
994         // hot team has already been allocated for given level
995         KMP_DEBUG_ASSERT(hot_teams[level].hot_team == team);
996         use_hot_team = 1; // the team is ready to use
997       } else {
998         use_hot_team = 0; // AC: threads are not allocated yet
999         hot_teams[level].hot_team = team; // remember new hot team
1000         hot_teams[level].hot_team_nth = team->t.t_nproc;
1001       }
1002     } else {
1003       use_hot_team = 0;
1004     }
1005   }
1006 #else
1007   use_hot_team = team == root->r.r_hot_team;
1008 #endif
1009   if (!use_hot_team) {
1010 
1011     /* install the master thread */
1012     team->t.t_threads[0] = master_th;
1013     __kmp_initialize_info(master_th, team, 0, master_gtid);
1014 
1015     /* now, install the worker threads */
1016     for (i = 1; i < team->t.t_nproc; i++) {
1017 
1018       /* fork or reallocate a new thread and install it in team */
1019       kmp_info_t *thr = __kmp_allocate_thread(root, team, i);
1020       team->t.t_threads[i] = thr;
1021       KMP_DEBUG_ASSERT(thr);
1022       KMP_DEBUG_ASSERT(thr->th.th_team == team);
1023       /* align team and thread arrived states */
1024       KA_TRACE(20, ("__kmp_fork_team_threads: T#%d(%d:%d) init arrived "
1025                     "T#%d(%d:%d) join =%llu, plain=%llu\n",
1026                     __kmp_gtid_from_tid(0, team), team->t.t_id, 0,
1027                     __kmp_gtid_from_tid(i, team), team->t.t_id, i,
1028                     team->t.t_bar[bs_forkjoin_barrier].b_arrived,
1029                     team->t.t_bar[bs_plain_barrier].b_arrived));
1030       thr->th.th_teams_microtask = master_th->th.th_teams_microtask;
1031       thr->th.th_teams_level = master_th->th.th_teams_level;
1032       thr->th.th_teams_size = master_th->th.th_teams_size;
1033       { // Initialize threads' barrier data.
1034         int b;
1035         kmp_balign_t *balign = team->t.t_threads[i]->th.th_bar;
1036         for (b = 0; b < bs_last_barrier; ++b) {
1037           balign[b].bb.b_arrived = team->t.t_bar[b].b_arrived;
1038           KMP_DEBUG_ASSERT(balign[b].bb.wait_flag != KMP_BARRIER_PARENT_FLAG);
1039 #if USE_DEBUGGER
1040           balign[b].bb.b_worker_arrived = team->t.t_bar[b].b_team_arrived;
1041 #endif
1042         }
1043       }
1044     }
1045 
1046 #if KMP_AFFINITY_SUPPORTED
1047     __kmp_partition_places(team);
1048 #endif
1049   }
1050 
1051   if (__kmp_display_affinity && team->t.t_display_affinity != 1) {
1052     for (i = 0; i < team->t.t_nproc; i++) {
1053       kmp_info_t *thr = team->t.t_threads[i];
1054       if (thr->th.th_prev_num_threads != team->t.t_nproc ||
1055           thr->th.th_prev_level != team->t.t_level) {
1056         team->t.t_display_affinity = 1;
1057         break;
1058       }
1059     }
1060   }
1061 
1062   KMP_MB();
1063 }
1064 
1065 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
1066 // Propagate any changes to the floating point control registers out to the team
1067 // We try to avoid unnecessary writes to the relevant cache line in the team
1068 // structure, so we don't make changes unless they are needed.
1069 inline static void propagateFPControl(kmp_team_t *team) {
1070   if (__kmp_inherit_fp_control) {
1071     kmp_int16 x87_fpu_control_word;
1072     kmp_uint32 mxcsr;
1073 
1074     // Get master values of FPU control flags (both X87 and vector)
1075     __kmp_store_x87_fpu_control_word(&x87_fpu_control_word);
1076     __kmp_store_mxcsr(&mxcsr);
1077     mxcsr &= KMP_X86_MXCSR_MASK;
1078 
1079     // There is no point looking at t_fp_control_saved here.
1080     // If it is TRUE, we still have to update the values if they are different
1081     // from those we now have. If it is FALSE we didn't save anything yet, but
1082     // our objective is the same. We have to ensure that the values in the team
1083     // are the same as those we have.
1084     // So, this code achieves what we need whether or not t_fp_control_saved is
1085     // true. By checking whether the value needs updating we avoid unnecessary
1086     // writes that would put the cache-line into a written state, causing all
1087     // threads in the team to have to read it again.
1088     KMP_CHECK_UPDATE(team->t.t_x87_fpu_control_word, x87_fpu_control_word);
1089     KMP_CHECK_UPDATE(team->t.t_mxcsr, mxcsr);
1090     // Although we don't use this value, other code in the runtime wants to know
1091     // whether it should restore them. So we must ensure it is correct.
1092     KMP_CHECK_UPDATE(team->t.t_fp_control_saved, TRUE);
1093   } else {
1094     // Similarly here. Don't write to this cache-line in the team structure
1095     // unless we have to.
1096     KMP_CHECK_UPDATE(team->t.t_fp_control_saved, FALSE);
1097   }
1098 }
1099 
1100 // Do the opposite, setting the hardware registers to the updated values from
1101 // the team.
1102 inline static void updateHWFPControl(kmp_team_t *team) {
1103   if (__kmp_inherit_fp_control && team->t.t_fp_control_saved) {
1104     // Only reset the fp control regs if they have been changed in the team.
1105     // the parallel region that we are exiting.
1106     kmp_int16 x87_fpu_control_word;
1107     kmp_uint32 mxcsr;
1108     __kmp_store_x87_fpu_control_word(&x87_fpu_control_word);
1109     __kmp_store_mxcsr(&mxcsr);
1110     mxcsr &= KMP_X86_MXCSR_MASK;
1111 
1112     if (team->t.t_x87_fpu_control_word != x87_fpu_control_word) {
1113       __kmp_clear_x87_fpu_status_word();
1114       __kmp_load_x87_fpu_control_word(&team->t.t_x87_fpu_control_word);
1115     }
1116 
1117     if (team->t.t_mxcsr != mxcsr) {
1118       __kmp_load_mxcsr(&team->t.t_mxcsr);
1119     }
1120   }
1121 }
1122 #else
1123 #define propagateFPControl(x) ((void)0)
1124 #define updateHWFPControl(x) ((void)0)
1125 #endif /* KMP_ARCH_X86 || KMP_ARCH_X86_64 */
1126 
1127 static void __kmp_alloc_argv_entries(int argc, kmp_team_t *team,
1128                                      int realloc); // forward declaration
1129 
1130 /* Run a parallel region that has been serialized, so runs only in a team of the
1131    single master thread. */
1132 void __kmp_serialized_parallel(ident_t *loc, kmp_int32 global_tid) {
1133   kmp_info_t *this_thr;
1134   kmp_team_t *serial_team;
1135 
1136   KC_TRACE(10, ("__kmpc_serialized_parallel: called by T#%d\n", global_tid));
1137 
1138   /* Skip all this code for autopar serialized loops since it results in
1139      unacceptable overhead */
1140   if (loc != NULL && (loc->flags & KMP_IDENT_AUTOPAR))
1141     return;
1142 
1143   if (!TCR_4(__kmp_init_parallel))
1144     __kmp_parallel_initialize();
1145   __kmp_resume_if_soft_paused();
1146 
1147   this_thr = __kmp_threads[global_tid];
1148   serial_team = this_thr->th.th_serial_team;
1149 
1150   /* utilize the serialized team held by this thread */
1151   KMP_DEBUG_ASSERT(serial_team);
1152   KMP_MB();
1153 
1154   if (__kmp_tasking_mode != tskm_immediate_exec) {
1155     KMP_DEBUG_ASSERT(
1156         this_thr->th.th_task_team ==
1157         this_thr->th.th_team->t.t_task_team[this_thr->th.th_task_state]);
1158     KMP_DEBUG_ASSERT(serial_team->t.t_task_team[this_thr->th.th_task_state] ==
1159                      NULL);
1160     KA_TRACE(20, ("__kmpc_serialized_parallel: T#%d pushing task_team %p / "
1161                   "team %p, new task_team = NULL\n",
1162                   global_tid, this_thr->th.th_task_team, this_thr->th.th_team));
1163     this_thr->th.th_task_team = NULL;
1164   }
1165 
1166   kmp_proc_bind_t proc_bind = this_thr->th.th_set_proc_bind;
1167   if (this_thr->th.th_current_task->td_icvs.proc_bind == proc_bind_false) {
1168     proc_bind = proc_bind_false;
1169   } else if (proc_bind == proc_bind_default) {
1170     // No proc_bind clause was specified, so use the current value
1171     // of proc-bind-var for this parallel region.
1172     proc_bind = this_thr->th.th_current_task->td_icvs.proc_bind;
1173   }
1174   // Reset for next parallel region
1175   this_thr->th.th_set_proc_bind = proc_bind_default;
1176 
1177 #if OMPT_SUPPORT
1178   ompt_data_t ompt_parallel_data = ompt_data_none;
1179   ompt_data_t *implicit_task_data;
1180   void *codeptr = OMPT_LOAD_RETURN_ADDRESS(global_tid);
1181   if (ompt_enabled.enabled &&
1182       this_thr->th.ompt_thread_info.state != ompt_state_overhead) {
1183 
1184     ompt_task_info_t *parent_task_info;
1185     parent_task_info = OMPT_CUR_TASK_INFO(this_thr);
1186 
1187     parent_task_info->frame.enter_frame.ptr = OMPT_GET_FRAME_ADDRESS(0);
1188     if (ompt_enabled.ompt_callback_parallel_begin) {
1189       int team_size = 1;
1190 
1191       ompt_callbacks.ompt_callback(ompt_callback_parallel_begin)(
1192           &(parent_task_info->task_data), &(parent_task_info->frame),
1193           &ompt_parallel_data, team_size,
1194           ompt_parallel_invoker_program | ompt_parallel_team, codeptr);
1195     }
1196   }
1197 #endif // OMPT_SUPPORT
1198 
1199   if (this_thr->th.th_team != serial_team) {
1200     // Nested level will be an index in the nested nthreads array
1201     int level = this_thr->th.th_team->t.t_level;
1202 
1203     if (serial_team->t.t_serialized) {
1204       /* this serial team was already used
1205          TODO increase performance by making this locks more specific */
1206       kmp_team_t *new_team;
1207 
1208       __kmp_acquire_bootstrap_lock(&__kmp_forkjoin_lock);
1209 
1210       new_team =
1211           __kmp_allocate_team(this_thr->th.th_root, 1, 1,
1212 #if OMPT_SUPPORT
1213                               ompt_parallel_data,
1214 #endif
1215                               proc_bind, &this_thr->th.th_current_task->td_icvs,
1216                               0 USE_NESTED_HOT_ARG(NULL));
1217       __kmp_release_bootstrap_lock(&__kmp_forkjoin_lock);
1218       KMP_ASSERT(new_team);
1219 
1220       /* setup new serialized team and install it */
1221       new_team->t.t_threads[0] = this_thr;
1222       new_team->t.t_parent = this_thr->th.th_team;
1223       serial_team = new_team;
1224       this_thr->th.th_serial_team = serial_team;
1225 
1226       KF_TRACE(
1227           10,
1228           ("__kmpc_serialized_parallel: T#%d allocated new serial team %p\n",
1229            global_tid, serial_team));
1230 
1231       /* TODO the above breaks the requirement that if we run out of resources,
1232          then we can still guarantee that serialized teams are ok, since we may
1233          need to allocate a new one */
1234     } else {
1235       KF_TRACE(
1236           10,
1237           ("__kmpc_serialized_parallel: T#%d reusing cached serial team %p\n",
1238            global_tid, serial_team));
1239     }
1240 
1241     /* we have to initialize this serial team */
1242     KMP_DEBUG_ASSERT(serial_team->t.t_threads);
1243     KMP_DEBUG_ASSERT(serial_team->t.t_threads[0] == this_thr);
1244     KMP_DEBUG_ASSERT(this_thr->th.th_team != serial_team);
1245     serial_team->t.t_ident = loc;
1246     serial_team->t.t_serialized = 1;
1247     serial_team->t.t_nproc = 1;
1248     serial_team->t.t_parent = this_thr->th.th_team;
1249     serial_team->t.t_sched.sched = this_thr->th.th_team->t.t_sched.sched;
1250     this_thr->th.th_team = serial_team;
1251     serial_team->t.t_master_tid = this_thr->th.th_info.ds.ds_tid;
1252 
1253     KF_TRACE(10, ("__kmpc_serialized_parallel: T#d curtask=%p\n", global_tid,
1254                   this_thr->th.th_current_task));
1255     KMP_ASSERT(this_thr->th.th_current_task->td_flags.executing == 1);
1256     this_thr->th.th_current_task->td_flags.executing = 0;
1257 
1258     __kmp_push_current_task_to_thread(this_thr, serial_team, 0);
1259 
1260     /* TODO: GEH: do ICVs work for nested serialized teams? Don't we need an
1261        implicit task for each serialized task represented by
1262        team->t.t_serialized? */
1263     copy_icvs(&this_thr->th.th_current_task->td_icvs,
1264               &this_thr->th.th_current_task->td_parent->td_icvs);
1265 
1266     // Thread value exists in the nested nthreads array for the next nested
1267     // level
1268     if (__kmp_nested_nth.used && (level + 1 < __kmp_nested_nth.used)) {
1269       this_thr->th.th_current_task->td_icvs.nproc =
1270           __kmp_nested_nth.nth[level + 1];
1271     }
1272 
1273     if (__kmp_nested_proc_bind.used &&
1274         (level + 1 < __kmp_nested_proc_bind.used)) {
1275       this_thr->th.th_current_task->td_icvs.proc_bind =
1276           __kmp_nested_proc_bind.bind_types[level + 1];
1277     }
1278 
1279 #if USE_DEBUGGER
1280     serial_team->t.t_pkfn = (microtask_t)(~0); // For the debugger.
1281 #endif
1282     this_thr->th.th_info.ds.ds_tid = 0;
1283 
1284     /* set thread cache values */
1285     this_thr->th.th_team_nproc = 1;
1286     this_thr->th.th_team_master = this_thr;
1287     this_thr->th.th_team_serialized = 1;
1288 
1289     serial_team->t.t_level = serial_team->t.t_parent->t.t_level + 1;
1290     serial_team->t.t_active_level = serial_team->t.t_parent->t.t_active_level;
1291     serial_team->t.t_def_allocator = this_thr->th.th_def_allocator; // save
1292 
1293     propagateFPControl(serial_team);
1294 
1295     /* check if we need to allocate dispatch buffers stack */
1296     KMP_DEBUG_ASSERT(serial_team->t.t_dispatch);
1297     if (!serial_team->t.t_dispatch->th_disp_buffer) {
1298       serial_team->t.t_dispatch->th_disp_buffer =
1299           (dispatch_private_info_t *)__kmp_allocate(
1300               sizeof(dispatch_private_info_t));
1301     }
1302     this_thr->th.th_dispatch = serial_team->t.t_dispatch;
1303 
1304     KMP_MB();
1305 
1306   } else {
1307     /* this serialized team is already being used,
1308      * that's fine, just add another nested level */
1309     KMP_DEBUG_ASSERT(this_thr->th.th_team == serial_team);
1310     KMP_DEBUG_ASSERT(serial_team->t.t_threads);
1311     KMP_DEBUG_ASSERT(serial_team->t.t_threads[0] == this_thr);
1312     ++serial_team->t.t_serialized;
1313     this_thr->th.th_team_serialized = serial_team->t.t_serialized;
1314 
1315     // Nested level will be an index in the nested nthreads array
1316     int level = this_thr->th.th_team->t.t_level;
1317     // Thread value exists in the nested nthreads array for the next nested
1318     // level
1319     if (__kmp_nested_nth.used && (level + 1 < __kmp_nested_nth.used)) {
1320       this_thr->th.th_current_task->td_icvs.nproc =
1321           __kmp_nested_nth.nth[level + 1];
1322     }
1323     serial_team->t.t_level++;
1324     KF_TRACE(10, ("__kmpc_serialized_parallel: T#%d increasing nesting level "
1325                   "of serial team %p to %d\n",
1326                   global_tid, serial_team, serial_team->t.t_level));
1327 
1328     /* allocate/push dispatch buffers stack */
1329     KMP_DEBUG_ASSERT(serial_team->t.t_dispatch);
1330     {
1331       dispatch_private_info_t *disp_buffer =
1332           (dispatch_private_info_t *)__kmp_allocate(
1333               sizeof(dispatch_private_info_t));
1334       disp_buffer->next = serial_team->t.t_dispatch->th_disp_buffer;
1335       serial_team->t.t_dispatch->th_disp_buffer = disp_buffer;
1336     }
1337     this_thr->th.th_dispatch = serial_team->t.t_dispatch;
1338 
1339     KMP_MB();
1340   }
1341   KMP_CHECK_UPDATE(serial_team->t.t_cancel_request, cancel_noreq);
1342 
1343   // Perform the display affinity functionality for
1344   // serialized parallel regions
1345   if (__kmp_display_affinity) {
1346     if (this_thr->th.th_prev_level != serial_team->t.t_level ||
1347         this_thr->th.th_prev_num_threads != 1) {
1348       // NULL means use the affinity-format-var ICV
1349       __kmp_aux_display_affinity(global_tid, NULL);
1350       this_thr->th.th_prev_level = serial_team->t.t_level;
1351       this_thr->th.th_prev_num_threads = 1;
1352     }
1353   }
1354 
1355   if (__kmp_env_consistency_check)
1356     __kmp_push_parallel(global_tid, NULL);
1357 #if OMPT_SUPPORT
1358   serial_team->t.ompt_team_info.master_return_address = codeptr;
1359   if (ompt_enabled.enabled &&
1360       this_thr->th.ompt_thread_info.state != ompt_state_overhead) {
1361     OMPT_CUR_TASK_INFO(this_thr)->frame.exit_frame.ptr = OMPT_GET_FRAME_ADDRESS(0);
1362 
1363     ompt_lw_taskteam_t lw_taskteam;
1364     __ompt_lw_taskteam_init(&lw_taskteam, this_thr, global_tid,
1365                             &ompt_parallel_data, codeptr);
1366 
1367     __ompt_lw_taskteam_link(&lw_taskteam, this_thr, 1);
1368     // don't use lw_taskteam after linking. content was swaped
1369 
1370     /* OMPT implicit task begin */
1371     implicit_task_data = OMPT_CUR_TASK_DATA(this_thr);
1372     if (ompt_enabled.ompt_callback_implicit_task) {
1373       ompt_callbacks.ompt_callback(ompt_callback_implicit_task)(
1374           ompt_scope_begin, OMPT_CUR_TEAM_DATA(this_thr),
1375           OMPT_CUR_TASK_DATA(this_thr), 1, __kmp_tid_from_gtid(global_tid), ompt_task_implicit); // TODO: Can this be ompt_task_initial?
1376       OMPT_CUR_TASK_INFO(this_thr)
1377           ->thread_num = __kmp_tid_from_gtid(global_tid);
1378     }
1379 
1380     /* OMPT state */
1381     this_thr->th.ompt_thread_info.state = ompt_state_work_parallel;
1382     OMPT_CUR_TASK_INFO(this_thr)->frame.exit_frame.ptr = OMPT_GET_FRAME_ADDRESS(0);
1383   }
1384 #endif
1385 }
1386 
1387 /* most of the work for a fork */
1388 /* return true if we really went parallel, false if serialized */
1389 int __kmp_fork_call(ident_t *loc, int gtid,
1390                     enum fork_context_e call_context, // Intel, GNU, ...
1391                     kmp_int32 argc, microtask_t microtask, launch_t invoker,
1392                     kmp_va_list ap) {
1393   void **argv;
1394   int i;
1395   int master_tid;
1396   int master_this_cons;
1397   kmp_team_t *team;
1398   kmp_team_t *parent_team;
1399   kmp_info_t *master_th;
1400   kmp_root_t *root;
1401   int nthreads;
1402   int master_active;
1403   int master_set_numthreads;
1404   int level;
1405   int active_level;
1406   int teams_level;
1407 #if KMP_NESTED_HOT_TEAMS
1408   kmp_hot_team_ptr_t **p_hot_teams;
1409 #endif
1410   { // KMP_TIME_BLOCK
1411     KMP_TIME_DEVELOPER_PARTITIONED_BLOCK(KMP_fork_call);
1412     KMP_COUNT_VALUE(OMP_PARALLEL_args, argc);
1413 
1414     KA_TRACE(20, ("__kmp_fork_call: enter T#%d\n", gtid));
1415     if (__kmp_stkpadding > 0 && __kmp_root[gtid] != NULL) {
1416       /* Some systems prefer the stack for the root thread(s) to start with */
1417       /* some gap from the parent stack to prevent false sharing. */
1418       void *dummy = KMP_ALLOCA(__kmp_stkpadding);
1419       /* These 2 lines below are so this does not get optimized out */
1420       if (__kmp_stkpadding > KMP_MAX_STKPADDING)
1421         __kmp_stkpadding += (short)((kmp_int64)dummy);
1422     }
1423 
1424     /* initialize if needed */
1425     KMP_DEBUG_ASSERT(
1426         __kmp_init_serial); // AC: potentially unsafe, not in sync with shutdown
1427     if (!TCR_4(__kmp_init_parallel))
1428       __kmp_parallel_initialize();
1429     __kmp_resume_if_soft_paused();
1430 
1431     /* setup current data */
1432     master_th = __kmp_threads[gtid]; // AC: potentially unsafe, not in sync with
1433     // shutdown
1434     parent_team = master_th->th.th_team;
1435     master_tid = master_th->th.th_info.ds.ds_tid;
1436     master_this_cons = master_th->th.th_local.this_construct;
1437     root = master_th->th.th_root;
1438     master_active = root->r.r_active;
1439     master_set_numthreads = master_th->th.th_set_nproc;
1440 
1441 #if OMPT_SUPPORT
1442     ompt_data_t ompt_parallel_data = ompt_data_none;
1443     ompt_data_t *parent_task_data;
1444     ompt_frame_t *ompt_frame;
1445     ompt_data_t *implicit_task_data;
1446     void *return_address = NULL;
1447 
1448     if (ompt_enabled.enabled) {
1449       __ompt_get_task_info_internal(0, NULL, &parent_task_data, &ompt_frame,
1450                                     NULL, NULL);
1451       return_address = OMPT_LOAD_RETURN_ADDRESS(gtid);
1452     }
1453 #endif
1454 
1455     // Nested level will be an index in the nested nthreads array
1456     level = parent_team->t.t_level;
1457     // used to launch non-serial teams even if nested is not allowed
1458     active_level = parent_team->t.t_active_level;
1459     // needed to check nesting inside the teams
1460     teams_level = master_th->th.th_teams_level;
1461 #if KMP_NESTED_HOT_TEAMS
1462     p_hot_teams = &master_th->th.th_hot_teams;
1463     if (*p_hot_teams == NULL && __kmp_hot_teams_max_level > 0) {
1464       *p_hot_teams = (kmp_hot_team_ptr_t *)__kmp_allocate(
1465           sizeof(kmp_hot_team_ptr_t) * __kmp_hot_teams_max_level);
1466       (*p_hot_teams)[0].hot_team = root->r.r_hot_team;
1467       // it is either actual or not needed (when active_level > 0)
1468       (*p_hot_teams)[0].hot_team_nth = 1;
1469     }
1470 #endif
1471 
1472 #if OMPT_SUPPORT
1473     if (ompt_enabled.enabled) {
1474       if (ompt_enabled.ompt_callback_parallel_begin) {
1475         int team_size = master_set_numthreads
1476                             ? master_set_numthreads
1477                             : get__nproc_2(parent_team, master_tid);
1478         int flags = OMPT_INVOKER(call_context) |
1479                     ((microtask == (microtask_t)__kmp_teams_master)
1480                          ? ompt_parallel_league
1481                          : ompt_parallel_team);
1482         ompt_callbacks.ompt_callback(ompt_callback_parallel_begin)(
1483             parent_task_data, ompt_frame, &ompt_parallel_data, team_size, flags,
1484             return_address);
1485       }
1486       master_th->th.ompt_thread_info.state = ompt_state_overhead;
1487     }
1488 #endif
1489 
1490     master_th->th.th_ident = loc;
1491 
1492     if (master_th->th.th_teams_microtask && ap &&
1493         microtask != (microtask_t)__kmp_teams_master && level == teams_level) {
1494       // AC: This is start of parallel that is nested inside teams construct.
1495       // The team is actual (hot), all workers are ready at the fork barrier.
1496       // No lock needed to initialize the team a bit, then free workers.
1497       parent_team->t.t_ident = loc;
1498       __kmp_alloc_argv_entries(argc, parent_team, TRUE);
1499       parent_team->t.t_argc = argc;
1500       argv = (void **)parent_team->t.t_argv;
1501       for (i = argc - 1; i >= 0; --i)
1502         *argv++ = va_arg(kmp_va_deref(ap), void *);
1503       // Increment our nested depth levels, but not increase the serialization
1504       if (parent_team == master_th->th.th_serial_team) {
1505         // AC: we are in serialized parallel
1506         __kmpc_serialized_parallel(loc, gtid);
1507         KMP_DEBUG_ASSERT(parent_team->t.t_serialized > 1);
1508 
1509         if (call_context == fork_context_gnu) {
1510           // AC: need to decrement t_serialized for enquiry functions to work
1511           // correctly, will restore at join time
1512           parent_team->t.t_serialized--;
1513           return TRUE;
1514         }
1515 
1516 #if OMPT_SUPPORT
1517         void *dummy;
1518         void **exit_frame_p;
1519 
1520         ompt_lw_taskteam_t lw_taskteam;
1521 
1522         if (ompt_enabled.enabled) {
1523           __ompt_lw_taskteam_init(&lw_taskteam, master_th, gtid,
1524                                   &ompt_parallel_data, return_address);
1525           exit_frame_p = &(lw_taskteam.ompt_task_info.frame.exit_frame.ptr);
1526 
1527           __ompt_lw_taskteam_link(&lw_taskteam, master_th, 0);
1528           // don't use lw_taskteam after linking. content was swaped
1529 
1530           /* OMPT implicit task begin */
1531           implicit_task_data = OMPT_CUR_TASK_DATA(master_th);
1532           if (ompt_enabled.ompt_callback_implicit_task) {
1533             OMPT_CUR_TASK_INFO(master_th)
1534                 ->thread_num = __kmp_tid_from_gtid(gtid);
1535             ompt_callbacks.ompt_callback(ompt_callback_implicit_task)(
1536                 ompt_scope_begin, OMPT_CUR_TEAM_DATA(master_th),
1537                 implicit_task_data, 1,
1538                 OMPT_CUR_TASK_INFO(master_th)->thread_num, ompt_task_implicit);
1539           }
1540 
1541           /* OMPT state */
1542           master_th->th.ompt_thread_info.state = ompt_state_work_parallel;
1543         } else {
1544           exit_frame_p = &dummy;
1545         }
1546 #endif
1547         // AC: need to decrement t_serialized for enquiry functions to work
1548         // correctly, will restore at join time
1549         parent_team->t.t_serialized--;
1550 
1551         {
1552           KMP_TIME_PARTITIONED_BLOCK(OMP_parallel);
1553           KMP_SET_THREAD_STATE_BLOCK(IMPLICIT_TASK);
1554           __kmp_invoke_microtask(microtask, gtid, 0, argc, parent_team->t.t_argv
1555 #if OMPT_SUPPORT
1556                                  ,
1557                                  exit_frame_p
1558 #endif
1559                                  );
1560         }
1561 
1562 #if OMPT_SUPPORT
1563         if (ompt_enabled.enabled) {
1564           *exit_frame_p = NULL;
1565           OMPT_CUR_TASK_INFO(master_th)->frame.exit_frame = ompt_data_none;
1566           if (ompt_enabled.ompt_callback_implicit_task) {
1567             ompt_callbacks.ompt_callback(ompt_callback_implicit_task)(
1568                 ompt_scope_end, NULL, implicit_task_data, 1,
1569                 OMPT_CUR_TASK_INFO(master_th)->thread_num, ompt_task_implicit);
1570           }
1571           ompt_parallel_data = *OMPT_CUR_TEAM_DATA(master_th);
1572           __ompt_lw_taskteam_unlink(master_th);
1573           if (ompt_enabled.ompt_callback_parallel_end) {
1574             ompt_callbacks.ompt_callback(ompt_callback_parallel_end)(
1575                 &ompt_parallel_data, OMPT_CUR_TASK_DATA(master_th),
1576                 OMPT_INVOKER(call_context) | ompt_parallel_team,
1577                 return_address);
1578           }
1579           master_th->th.ompt_thread_info.state = ompt_state_overhead;
1580         }
1581 #endif
1582         return TRUE;
1583       }
1584 
1585       parent_team->t.t_pkfn = microtask;
1586       parent_team->t.t_invoke = invoker;
1587       KMP_ATOMIC_INC(&root->r.r_in_parallel);
1588       parent_team->t.t_active_level++;
1589       parent_team->t.t_level++;
1590       parent_team->t.t_def_allocator = master_th->th.th_def_allocator; // save
1591 
1592 #if OMPT_SUPPORT
1593       if (ompt_enabled.enabled) {
1594         ompt_lw_taskteam_t lw_taskteam;
1595         __ompt_lw_taskteam_init(&lw_taskteam, master_th, gtid,
1596                                 &ompt_parallel_data, return_address);
1597         __ompt_lw_taskteam_link(&lw_taskteam, master_th, 1, true);
1598       }
1599 #endif
1600 
1601       /* Change number of threads in the team if requested */
1602       if (master_set_numthreads) { // The parallel has num_threads clause
1603         if (master_set_numthreads < master_th->th.th_teams_size.nth) {
1604           // AC: only can reduce number of threads dynamically, can't increase
1605           kmp_info_t **other_threads = parent_team->t.t_threads;
1606           parent_team->t.t_nproc = master_set_numthreads;
1607           for (i = 0; i < master_set_numthreads; ++i) {
1608             other_threads[i]->th.th_team_nproc = master_set_numthreads;
1609           }
1610           // Keep extra threads hot in the team for possible next parallels
1611         }
1612         master_th->th.th_set_nproc = 0;
1613       }
1614 
1615 #if USE_DEBUGGER
1616       if (__kmp_debugging) { // Let debugger override number of threads.
1617         int nth = __kmp_omp_num_threads(loc);
1618         if (nth > 0) { // 0 means debugger doesn't want to change num threads
1619           master_set_numthreads = nth;
1620         }
1621       }
1622 #endif
1623 
1624 #if USE_ITT_BUILD
1625       if (((__itt_frame_submit_v3_ptr && __itt_get_timestamp_ptr) ||
1626            KMP_ITT_DEBUG) &&
1627           __kmp_forkjoin_frames_mode == 3 &&
1628           parent_team->t.t_active_level == 1 // only report frames at level 1
1629           && master_th->th.th_teams_size.nteams == 1) {
1630         kmp_uint64 tmp_time = __itt_get_timestamp();
1631         master_th->th.th_frame_time = tmp_time;
1632         parent_team->t.t_region_time = tmp_time;
1633       }
1634       if (__itt_stack_caller_create_ptr) {
1635         // create new stack stitching id before entering fork barrier
1636         parent_team->t.t_stack_id = __kmp_itt_stack_caller_create();
1637       }
1638 #endif /* USE_ITT_BUILD */
1639 
1640       KF_TRACE(10, ("__kmp_fork_call: before internal fork: root=%p, team=%p, "
1641                     "master_th=%p, gtid=%d\n",
1642                     root, parent_team, master_th, gtid));
1643       __kmp_internal_fork(loc, gtid, parent_team);
1644       KF_TRACE(10, ("__kmp_fork_call: after internal fork: root=%p, team=%p, "
1645                     "master_th=%p, gtid=%d\n",
1646                     root, parent_team, master_th, gtid));
1647 
1648       if (call_context == fork_context_gnu)
1649         return TRUE;
1650 
1651       /* Invoke microtask for MASTER thread */
1652       KA_TRACE(20, ("__kmp_fork_call: T#%d(%d:0) invoke microtask = %p\n", gtid,
1653                     parent_team->t.t_id, parent_team->t.t_pkfn));
1654 
1655       if (!parent_team->t.t_invoke(gtid)) {
1656         KMP_ASSERT2(0, "cannot invoke microtask for MASTER thread");
1657       }
1658       KA_TRACE(20, ("__kmp_fork_call: T#%d(%d:0) done microtask = %p\n", gtid,
1659                     parent_team->t.t_id, parent_team->t.t_pkfn));
1660       KMP_MB(); /* Flush all pending memory write invalidates.  */
1661 
1662       KA_TRACE(20, ("__kmp_fork_call: parallel exit T#%d\n", gtid));
1663 
1664       return TRUE;
1665     } // Parallel closely nested in teams construct
1666 
1667 #if KMP_DEBUG
1668     if (__kmp_tasking_mode != tskm_immediate_exec) {
1669       KMP_DEBUG_ASSERT(master_th->th.th_task_team ==
1670                        parent_team->t.t_task_team[master_th->th.th_task_state]);
1671     }
1672 #endif
1673 
1674     if (parent_team->t.t_active_level >=
1675         master_th->th.th_current_task->td_icvs.max_active_levels) {
1676       nthreads = 1;
1677     } else {
1678       int enter_teams = ((ap == NULL && active_level == 0) ||
1679                          (ap && teams_level > 0 && teams_level == level));
1680       nthreads =
1681           master_set_numthreads
1682               ? master_set_numthreads
1683               : get__nproc_2(
1684                     parent_team,
1685                     master_tid); // TODO: get nproc directly from current task
1686 
1687       // Check if we need to take forkjoin lock? (no need for serialized
1688       // parallel out of teams construct). This code moved here from
1689       // __kmp_reserve_threads() to speedup nested serialized parallels.
1690       if (nthreads > 1) {
1691         if ((get__max_active_levels(master_th) == 1 &&
1692              (root->r.r_in_parallel && !enter_teams)) ||
1693             (__kmp_library == library_serial)) {
1694           KC_TRACE(10, ("__kmp_fork_call: T#%d serializing team; requested %d"
1695                         " threads\n",
1696                         gtid, nthreads));
1697           nthreads = 1;
1698         }
1699       }
1700       if (nthreads > 1) {
1701         /* determine how many new threads we can use */
1702         __kmp_acquire_bootstrap_lock(&__kmp_forkjoin_lock);
1703         /* AC: If we execute teams from parallel region (on host), then teams
1704            should be created but each can only have 1 thread if nesting is
1705            disabled. If teams called from serial region, then teams and their
1706            threads should be created regardless of the nesting setting. */
1707         nthreads = __kmp_reserve_threads(root, parent_team, master_tid,
1708                                          nthreads, enter_teams);
1709         if (nthreads == 1) {
1710           // Free lock for single thread execution here; for multi-thread
1711           // execution it will be freed later after team of threads created
1712           // and initialized
1713           __kmp_release_bootstrap_lock(&__kmp_forkjoin_lock);
1714         }
1715       }
1716     }
1717     KMP_DEBUG_ASSERT(nthreads > 0);
1718 
1719     // If we temporarily changed the set number of threads then restore it now
1720     master_th->th.th_set_nproc = 0;
1721 
1722     /* create a serialized parallel region? */
1723     if (nthreads == 1) {
1724 /* josh todo: hypothetical question: what do we do for OS X*? */
1725 #if KMP_OS_LINUX &&                                                            \
1726     (KMP_ARCH_X86 || KMP_ARCH_X86_64 || KMP_ARCH_ARM || KMP_ARCH_AARCH64)
1727       void *args[argc];
1728 #else
1729       void **args = (void **)KMP_ALLOCA(argc * sizeof(void *));
1730 #endif /* KMP_OS_LINUX && ( KMP_ARCH_X86 || KMP_ARCH_X86_64 || KMP_ARCH_ARM || \
1731           KMP_ARCH_AARCH64) */
1732 
1733       KA_TRACE(20,
1734                ("__kmp_fork_call: T#%d serializing parallel region\n", gtid));
1735 
1736       __kmpc_serialized_parallel(loc, gtid);
1737 
1738       if (call_context == fork_context_intel) {
1739         /* TODO this sucks, use the compiler itself to pass args! :) */
1740         master_th->th.th_serial_team->t.t_ident = loc;
1741         if (!ap) {
1742           // revert change made in __kmpc_serialized_parallel()
1743           master_th->th.th_serial_team->t.t_level--;
1744 // Get args from parent team for teams construct
1745 
1746 #if OMPT_SUPPORT
1747           void *dummy;
1748           void **exit_frame_p;
1749           ompt_task_info_t *task_info;
1750 
1751           ompt_lw_taskteam_t lw_taskteam;
1752 
1753           if (ompt_enabled.enabled) {
1754             __ompt_lw_taskteam_init(&lw_taskteam, master_th, gtid,
1755                                     &ompt_parallel_data, return_address);
1756 
1757             __ompt_lw_taskteam_link(&lw_taskteam, master_th, 0);
1758             // don't use lw_taskteam after linking. content was swaped
1759 
1760             task_info = OMPT_CUR_TASK_INFO(master_th);
1761             exit_frame_p = &(task_info->frame.exit_frame.ptr);
1762             if (ompt_enabled.ompt_callback_implicit_task) {
1763               OMPT_CUR_TASK_INFO(master_th)
1764                   ->thread_num = __kmp_tid_from_gtid(gtid);
1765               ompt_callbacks.ompt_callback(ompt_callback_implicit_task)(
1766                   ompt_scope_begin, OMPT_CUR_TEAM_DATA(master_th),
1767                   &(task_info->task_data), 1,
1768                   OMPT_CUR_TASK_INFO(master_th)->thread_num,
1769                   ompt_task_implicit);
1770             }
1771 
1772             /* OMPT state */
1773             master_th->th.ompt_thread_info.state = ompt_state_work_parallel;
1774           } else {
1775             exit_frame_p = &dummy;
1776           }
1777 #endif
1778 
1779           {
1780             KMP_TIME_PARTITIONED_BLOCK(OMP_parallel);
1781             KMP_SET_THREAD_STATE_BLOCK(IMPLICIT_TASK);
1782             __kmp_invoke_microtask(microtask, gtid, 0, argc,
1783                                    parent_team->t.t_argv
1784 #if OMPT_SUPPORT
1785                                    ,
1786                                    exit_frame_p
1787 #endif
1788                                    );
1789           }
1790 
1791 #if OMPT_SUPPORT
1792           if (ompt_enabled.enabled) {
1793             *exit_frame_p = NULL;
1794             if (ompt_enabled.ompt_callback_implicit_task) {
1795               ompt_callbacks.ompt_callback(ompt_callback_implicit_task)(
1796                   ompt_scope_end, NULL, &(task_info->task_data), 1,
1797                   OMPT_CUR_TASK_INFO(master_th)->thread_num,
1798                   ompt_task_implicit);
1799             }
1800             ompt_parallel_data = *OMPT_CUR_TEAM_DATA(master_th);
1801             __ompt_lw_taskteam_unlink(master_th);
1802             if (ompt_enabled.ompt_callback_parallel_end) {
1803               ompt_callbacks.ompt_callback(ompt_callback_parallel_end)(
1804                   &ompt_parallel_data, parent_task_data,
1805                   OMPT_INVOKER(call_context) | ompt_parallel_team,
1806                   return_address);
1807             }
1808             master_th->th.ompt_thread_info.state = ompt_state_overhead;
1809           }
1810 #endif
1811         } else if (microtask == (microtask_t)__kmp_teams_master) {
1812           KMP_DEBUG_ASSERT(master_th->th.th_team ==
1813                            master_th->th.th_serial_team);
1814           team = master_th->th.th_team;
1815           // team->t.t_pkfn = microtask;
1816           team->t.t_invoke = invoker;
1817           __kmp_alloc_argv_entries(argc, team, TRUE);
1818           team->t.t_argc = argc;
1819           argv = (void **)team->t.t_argv;
1820           if (ap) {
1821             for (i = argc - 1; i >= 0; --i)
1822               *argv++ = va_arg(kmp_va_deref(ap), void *);
1823           } else {
1824             for (i = 0; i < argc; ++i)
1825               // Get args from parent team for teams construct
1826               argv[i] = parent_team->t.t_argv[i];
1827           }
1828           // AC: revert change made in __kmpc_serialized_parallel()
1829           //     because initial code in teams should have level=0
1830           team->t.t_level--;
1831           // AC: call special invoker for outer "parallel" of teams construct
1832           invoker(gtid);
1833 #if OMPT_SUPPORT
1834           if (ompt_enabled.enabled) {
1835             ompt_task_info_t *task_info = OMPT_CUR_TASK_INFO(master_th);
1836             if (ompt_enabled.ompt_callback_implicit_task) {
1837               ompt_callbacks.ompt_callback(ompt_callback_implicit_task)(
1838                   ompt_scope_end, NULL, &(task_info->task_data), 0,
1839                   OMPT_CUR_TASK_INFO(master_th)->thread_num, ompt_task_initial);
1840             }
1841             if (ompt_enabled.ompt_callback_parallel_end) {
1842               ompt_callbacks.ompt_callback(ompt_callback_parallel_end)(
1843                   &ompt_parallel_data, parent_task_data,
1844                   OMPT_INVOKER(call_context) | ompt_parallel_league,
1845                   return_address);
1846             }
1847             master_th->th.ompt_thread_info.state = ompt_state_overhead;
1848           }
1849 #endif
1850         } else {
1851           argv = args;
1852           for (i = argc - 1; i >= 0; --i)
1853             *argv++ = va_arg(kmp_va_deref(ap), void *);
1854           KMP_MB();
1855 
1856 #if OMPT_SUPPORT
1857           void *dummy;
1858           void **exit_frame_p;
1859           ompt_task_info_t *task_info;
1860 
1861           ompt_lw_taskteam_t lw_taskteam;
1862 
1863           if (ompt_enabled.enabled) {
1864             __ompt_lw_taskteam_init(&lw_taskteam, master_th, gtid,
1865                                     &ompt_parallel_data, return_address);
1866             __ompt_lw_taskteam_link(&lw_taskteam, master_th, 0);
1867             // don't use lw_taskteam after linking. content was swaped
1868             task_info = OMPT_CUR_TASK_INFO(master_th);
1869             exit_frame_p = &(task_info->frame.exit_frame.ptr);
1870 
1871             /* OMPT implicit task begin */
1872             implicit_task_data = OMPT_CUR_TASK_DATA(master_th);
1873             if (ompt_enabled.ompt_callback_implicit_task) {
1874               ompt_callbacks.ompt_callback(ompt_callback_implicit_task)(
1875                   ompt_scope_begin, OMPT_CUR_TEAM_DATA(master_th),
1876                   implicit_task_data, 1, __kmp_tid_from_gtid(gtid),
1877                   ompt_task_implicit);
1878               OMPT_CUR_TASK_INFO(master_th)
1879                   ->thread_num = __kmp_tid_from_gtid(gtid);
1880             }
1881 
1882             /* OMPT state */
1883             master_th->th.ompt_thread_info.state = ompt_state_work_parallel;
1884           } else {
1885             exit_frame_p = &dummy;
1886           }
1887 #endif
1888 
1889           {
1890             KMP_TIME_PARTITIONED_BLOCK(OMP_parallel);
1891             KMP_SET_THREAD_STATE_BLOCK(IMPLICIT_TASK);
1892             __kmp_invoke_microtask(microtask, gtid, 0, argc, args
1893 #if OMPT_SUPPORT
1894                                    ,
1895                                    exit_frame_p
1896 #endif
1897                                    );
1898           }
1899 
1900 #if OMPT_SUPPORT
1901           if (ompt_enabled.enabled) {
1902             *exit_frame_p = NULL;
1903             if (ompt_enabled.ompt_callback_implicit_task) {
1904               ompt_callbacks.ompt_callback(ompt_callback_implicit_task)(
1905                   ompt_scope_end, NULL, &(task_info->task_data), 1,
1906                   OMPT_CUR_TASK_INFO(master_th)->thread_num,
1907                   ompt_task_implicit);
1908             }
1909 
1910             ompt_parallel_data = *OMPT_CUR_TEAM_DATA(master_th);
1911             __ompt_lw_taskteam_unlink(master_th);
1912             if (ompt_enabled.ompt_callback_parallel_end) {
1913               ompt_callbacks.ompt_callback(ompt_callback_parallel_end)(
1914                   &ompt_parallel_data, parent_task_data,
1915                   OMPT_INVOKER(call_context) | ompt_parallel_team,
1916                   return_address);
1917             }
1918             master_th->th.ompt_thread_info.state = ompt_state_overhead;
1919           }
1920 #endif
1921         }
1922       } else if (call_context == fork_context_gnu) {
1923 #if OMPT_SUPPORT
1924         ompt_lw_taskteam_t lwt;
1925         __ompt_lw_taskteam_init(&lwt, master_th, gtid, &ompt_parallel_data,
1926                                 return_address);
1927 
1928         lwt.ompt_task_info.frame.exit_frame = ompt_data_none;
1929         __ompt_lw_taskteam_link(&lwt, master_th, 1);
1930 // don't use lw_taskteam after linking. content was swaped
1931 #endif
1932 
1933         // we were called from GNU native code
1934         KA_TRACE(20, ("__kmp_fork_call: T#%d serial exit\n", gtid));
1935         return FALSE;
1936       } else {
1937         KMP_ASSERT2(call_context < fork_context_last,
1938                     "__kmp_fork_call: unknown fork_context parameter");
1939       }
1940 
1941       KA_TRACE(20, ("__kmp_fork_call: T#%d serial exit\n", gtid));
1942       KMP_MB();
1943       return FALSE;
1944     } // if (nthreads == 1)
1945 
1946     // GEH: only modify the executing flag in the case when not serialized
1947     //      serialized case is handled in kmpc_serialized_parallel
1948     KF_TRACE(10, ("__kmp_fork_call: parent_team_aclevel=%d, master_th=%p, "
1949                   "curtask=%p, curtask_max_aclevel=%d\n",
1950                   parent_team->t.t_active_level, master_th,
1951                   master_th->th.th_current_task,
1952                   master_th->th.th_current_task->td_icvs.max_active_levels));
1953     // TODO: GEH - cannot do this assertion because root thread not set up as
1954     // executing
1955     // KMP_ASSERT( master_th->th.th_current_task->td_flags.executing == 1 );
1956     master_th->th.th_current_task->td_flags.executing = 0;
1957 
1958     if (!master_th->th.th_teams_microtask || level > teams_level) {
1959       /* Increment our nested depth level */
1960       KMP_ATOMIC_INC(&root->r.r_in_parallel);
1961     }
1962 
1963     // See if we need to make a copy of the ICVs.
1964     int nthreads_icv = master_th->th.th_current_task->td_icvs.nproc;
1965     if ((level + 1 < __kmp_nested_nth.used) &&
1966         (__kmp_nested_nth.nth[level + 1] != nthreads_icv)) {
1967       nthreads_icv = __kmp_nested_nth.nth[level + 1];
1968     } else {
1969       nthreads_icv = 0; // don't update
1970     }
1971 
1972     // Figure out the proc_bind_policy for the new team.
1973     kmp_proc_bind_t proc_bind = master_th->th.th_set_proc_bind;
1974     kmp_proc_bind_t proc_bind_icv =
1975         proc_bind_default; // proc_bind_default means don't update
1976     if (master_th->th.th_current_task->td_icvs.proc_bind == proc_bind_false) {
1977       proc_bind = proc_bind_false;
1978     } else {
1979       if (proc_bind == proc_bind_default) {
1980         // No proc_bind clause specified; use current proc-bind-var for this
1981         // parallel region
1982         proc_bind = master_th->th.th_current_task->td_icvs.proc_bind;
1983       }
1984       /* else: The proc_bind policy was specified explicitly on parallel clause.
1985          This overrides proc-bind-var for this parallel region, but does not
1986          change proc-bind-var. */
1987       // Figure the value of proc-bind-var for the child threads.
1988       if ((level + 1 < __kmp_nested_proc_bind.used) &&
1989           (__kmp_nested_proc_bind.bind_types[level + 1] !=
1990            master_th->th.th_current_task->td_icvs.proc_bind)) {
1991         proc_bind_icv = __kmp_nested_proc_bind.bind_types[level + 1];
1992       }
1993     }
1994 
1995     // Reset for next parallel region
1996     master_th->th.th_set_proc_bind = proc_bind_default;
1997 
1998     if ((nthreads_icv > 0) || (proc_bind_icv != proc_bind_default)) {
1999       kmp_internal_control_t new_icvs;
2000       copy_icvs(&new_icvs, &master_th->th.th_current_task->td_icvs);
2001       new_icvs.next = NULL;
2002       if (nthreads_icv > 0) {
2003         new_icvs.nproc = nthreads_icv;
2004       }
2005       if (proc_bind_icv != proc_bind_default) {
2006         new_icvs.proc_bind = proc_bind_icv;
2007       }
2008 
2009       /* allocate a new parallel team */
2010       KF_TRACE(10, ("__kmp_fork_call: before __kmp_allocate_team\n"));
2011       team = __kmp_allocate_team(root, nthreads, nthreads,
2012 #if OMPT_SUPPORT
2013                                  ompt_parallel_data,
2014 #endif
2015                                  proc_bind, &new_icvs,
2016                                  argc USE_NESTED_HOT_ARG(master_th));
2017     } else {
2018       /* allocate a new parallel team */
2019       KF_TRACE(10, ("__kmp_fork_call: before __kmp_allocate_team\n"));
2020       team = __kmp_allocate_team(root, nthreads, nthreads,
2021 #if OMPT_SUPPORT
2022                                  ompt_parallel_data,
2023 #endif
2024                                  proc_bind,
2025                                  &master_th->th.th_current_task->td_icvs,
2026                                  argc USE_NESTED_HOT_ARG(master_th));
2027     }
2028     KF_TRACE(
2029         10, ("__kmp_fork_call: after __kmp_allocate_team - team = %p\n", team));
2030 
2031     /* setup the new team */
2032     KMP_CHECK_UPDATE(team->t.t_master_tid, master_tid);
2033     KMP_CHECK_UPDATE(team->t.t_master_this_cons, master_this_cons);
2034     KMP_CHECK_UPDATE(team->t.t_ident, loc);
2035     KMP_CHECK_UPDATE(team->t.t_parent, parent_team);
2036     KMP_CHECK_UPDATE_SYNC(team->t.t_pkfn, microtask);
2037 #if OMPT_SUPPORT
2038     KMP_CHECK_UPDATE_SYNC(team->t.ompt_team_info.master_return_address,
2039                           return_address);
2040 #endif
2041     KMP_CHECK_UPDATE(team->t.t_invoke, invoker); // TODO move to root, maybe
2042     // TODO: parent_team->t.t_level == INT_MAX ???
2043     if (!master_th->th.th_teams_microtask || level > teams_level) {
2044       int new_level = parent_team->t.t_level + 1;
2045       KMP_CHECK_UPDATE(team->t.t_level, new_level);
2046       new_level = parent_team->t.t_active_level + 1;
2047       KMP_CHECK_UPDATE(team->t.t_active_level, new_level);
2048     } else {
2049       // AC: Do not increase parallel level at start of the teams construct
2050       int new_level = parent_team->t.t_level;
2051       KMP_CHECK_UPDATE(team->t.t_level, new_level);
2052       new_level = parent_team->t.t_active_level;
2053       KMP_CHECK_UPDATE(team->t.t_active_level, new_level);
2054     }
2055     kmp_r_sched_t new_sched = get__sched_2(parent_team, master_tid);
2056     // set master's schedule as new run-time schedule
2057     KMP_CHECK_UPDATE(team->t.t_sched.sched, new_sched.sched);
2058 
2059     KMP_CHECK_UPDATE(team->t.t_cancel_request, cancel_noreq);
2060     KMP_CHECK_UPDATE(team->t.t_def_allocator, master_th->th.th_def_allocator);
2061 
2062     // Update the floating point rounding in the team if required.
2063     propagateFPControl(team);
2064 
2065     if (__kmp_tasking_mode != tskm_immediate_exec) {
2066       // Set master's task team to team's task team. Unless this is hot team, it
2067       // should be NULL.
2068       KMP_DEBUG_ASSERT(master_th->th.th_task_team ==
2069                        parent_team->t.t_task_team[master_th->th.th_task_state]);
2070       KA_TRACE(20, ("__kmp_fork_call: Master T#%d pushing task_team %p / team "
2071                     "%p, new task_team %p / team %p\n",
2072                     __kmp_gtid_from_thread(master_th),
2073                     master_th->th.th_task_team, parent_team,
2074                     team->t.t_task_team[master_th->th.th_task_state], team));
2075 
2076       if (active_level || master_th->th.th_task_team) {
2077         // Take a memo of master's task_state
2078         KMP_DEBUG_ASSERT(master_th->th.th_task_state_memo_stack);
2079         if (master_th->th.th_task_state_top >=
2080             master_th->th.th_task_state_stack_sz) { // increase size
2081           kmp_uint32 new_size = 2 * master_th->th.th_task_state_stack_sz;
2082           kmp_uint8 *old_stack, *new_stack;
2083           kmp_uint32 i;
2084           new_stack = (kmp_uint8 *)__kmp_allocate(new_size);
2085           for (i = 0; i < master_th->th.th_task_state_stack_sz; ++i) {
2086             new_stack[i] = master_th->th.th_task_state_memo_stack[i];
2087           }
2088           for (i = master_th->th.th_task_state_stack_sz; i < new_size;
2089                ++i) { // zero-init rest of stack
2090             new_stack[i] = 0;
2091           }
2092           old_stack = master_th->th.th_task_state_memo_stack;
2093           master_th->th.th_task_state_memo_stack = new_stack;
2094           master_th->th.th_task_state_stack_sz = new_size;
2095           __kmp_free(old_stack);
2096         }
2097         // Store master's task_state on stack
2098         master_th->th
2099             .th_task_state_memo_stack[master_th->th.th_task_state_top] =
2100             master_th->th.th_task_state;
2101         master_th->th.th_task_state_top++;
2102 #if KMP_NESTED_HOT_TEAMS
2103         if (master_th->th.th_hot_teams &&
2104             active_level < __kmp_hot_teams_max_level &&
2105             team == master_th->th.th_hot_teams[active_level].hot_team) {
2106           // Restore master's nested state if nested hot team
2107           master_th->th.th_task_state =
2108               master_th->th
2109                   .th_task_state_memo_stack[master_th->th.th_task_state_top];
2110         } else {
2111 #endif
2112           master_th->th.th_task_state = 0;
2113 #if KMP_NESTED_HOT_TEAMS
2114         }
2115 #endif
2116       }
2117 #if !KMP_NESTED_HOT_TEAMS
2118       KMP_DEBUG_ASSERT((master_th->th.th_task_team == NULL) ||
2119                        (team == root->r.r_hot_team));
2120 #endif
2121     }
2122 
2123     KA_TRACE(
2124         20,
2125         ("__kmp_fork_call: T#%d(%d:%d)->(%d:0) created a team of %d threads\n",
2126          gtid, parent_team->t.t_id, team->t.t_master_tid, team->t.t_id,
2127          team->t.t_nproc));
2128     KMP_DEBUG_ASSERT(team != root->r.r_hot_team ||
2129                      (team->t.t_master_tid == 0 &&
2130                       (team->t.t_parent == root->r.r_root_team ||
2131                        team->t.t_parent->t.t_serialized)));
2132     KMP_MB();
2133 
2134     /* now, setup the arguments */
2135     argv = (void **)team->t.t_argv;
2136     if (ap) {
2137       for (i = argc - 1; i >= 0; --i) {
2138         void *new_argv = va_arg(kmp_va_deref(ap), void *);
2139         KMP_CHECK_UPDATE(*argv, new_argv);
2140         argv++;
2141       }
2142     } else {
2143       for (i = 0; i < argc; ++i) {
2144         // Get args from parent team for teams construct
2145         KMP_CHECK_UPDATE(argv[i], team->t.t_parent->t.t_argv[i]);
2146       }
2147     }
2148 
2149     /* now actually fork the threads */
2150     KMP_CHECK_UPDATE(team->t.t_master_active, master_active);
2151     if (!root->r.r_active) // Only do assignment if it prevents cache ping-pong
2152       root->r.r_active = TRUE;
2153 
2154     __kmp_fork_team_threads(root, team, master_th, gtid);
2155     __kmp_setup_icv_copy(team, nthreads,
2156                          &master_th->th.th_current_task->td_icvs, loc);
2157 
2158 #if OMPT_SUPPORT
2159     master_th->th.ompt_thread_info.state = ompt_state_work_parallel;
2160 #endif
2161 
2162     __kmp_release_bootstrap_lock(&__kmp_forkjoin_lock);
2163 
2164 #if USE_ITT_BUILD
2165     if (team->t.t_active_level == 1 // only report frames at level 1
2166         && !master_th->th.th_teams_microtask) { // not in teams construct
2167 #if USE_ITT_NOTIFY
2168       if ((__itt_frame_submit_v3_ptr || KMP_ITT_DEBUG) &&
2169           (__kmp_forkjoin_frames_mode == 3 ||
2170            __kmp_forkjoin_frames_mode == 1)) {
2171         kmp_uint64 tmp_time = 0;
2172         if (__itt_get_timestamp_ptr)
2173           tmp_time = __itt_get_timestamp();
2174         // Internal fork - report frame begin
2175         master_th->th.th_frame_time = tmp_time;
2176         if (__kmp_forkjoin_frames_mode == 3)
2177           team->t.t_region_time = tmp_time;
2178       } else
2179 // only one notification scheme (either "submit" or "forking/joined", not both)
2180 #endif /* USE_ITT_NOTIFY */
2181           if ((__itt_frame_begin_v3_ptr || KMP_ITT_DEBUG) &&
2182               __kmp_forkjoin_frames && !__kmp_forkjoin_frames_mode) {
2183         // Mark start of "parallel" region for Intel(R) VTune(TM) analyzer.
2184         __kmp_itt_region_forking(gtid, team->t.t_nproc, 0);
2185       }
2186     }
2187 #endif /* USE_ITT_BUILD */
2188 
2189     /* now go on and do the work */
2190     KMP_DEBUG_ASSERT(team == __kmp_threads[gtid]->th.th_team);
2191     KMP_MB();
2192     KF_TRACE(10,
2193              ("__kmp_internal_fork : root=%p, team=%p, master_th=%p, gtid=%d\n",
2194               root, team, master_th, gtid));
2195 
2196 #if USE_ITT_BUILD
2197     if (__itt_stack_caller_create_ptr) {
2198       team->t.t_stack_id =
2199           __kmp_itt_stack_caller_create(); // create new stack stitching id
2200       // before entering fork barrier
2201     }
2202 #endif /* USE_ITT_BUILD */
2203 
2204     // AC: skip __kmp_internal_fork at teams construct, let only master
2205     // threads execute
2206     if (ap) {
2207       __kmp_internal_fork(loc, gtid, team);
2208       KF_TRACE(10, ("__kmp_internal_fork : after : root=%p, team=%p, "
2209                     "master_th=%p, gtid=%d\n",
2210                     root, team, master_th, gtid));
2211     }
2212 
2213     if (call_context == fork_context_gnu) {
2214       KA_TRACE(20, ("__kmp_fork_call: parallel exit T#%d\n", gtid));
2215       return TRUE;
2216     }
2217 
2218     /* Invoke microtask for MASTER thread */
2219     KA_TRACE(20, ("__kmp_fork_call: T#%d(%d:0) invoke microtask = %p\n", gtid,
2220                   team->t.t_id, team->t.t_pkfn));
2221   } // END of timer KMP_fork_call block
2222 
2223 #if KMP_STATS_ENABLED
2224   // If beginning a teams construct, then change thread state
2225   stats_state_e previous_state = KMP_GET_THREAD_STATE();
2226   if (!ap) {
2227     KMP_SET_THREAD_STATE(stats_state_e::TEAMS_REGION);
2228   }
2229 #endif
2230 
2231   if (!team->t.t_invoke(gtid)) {
2232     KMP_ASSERT2(0, "cannot invoke microtask for MASTER thread");
2233   }
2234 
2235 #if KMP_STATS_ENABLED
2236   // If was beginning of a teams construct, then reset thread state
2237   if (!ap) {
2238     KMP_SET_THREAD_STATE(previous_state);
2239   }
2240 #endif
2241 
2242   KA_TRACE(20, ("__kmp_fork_call: T#%d(%d:0) done microtask = %p\n", gtid,
2243                 team->t.t_id, team->t.t_pkfn));
2244   KMP_MB(); /* Flush all pending memory write invalidates.  */
2245 
2246   KA_TRACE(20, ("__kmp_fork_call: parallel exit T#%d\n", gtid));
2247 
2248 #if OMPT_SUPPORT
2249   if (ompt_enabled.enabled) {
2250     master_th->th.ompt_thread_info.state = ompt_state_overhead;
2251   }
2252 #endif
2253 
2254   return TRUE;
2255 }
2256 
2257 #if OMPT_SUPPORT
2258 static inline void __kmp_join_restore_state(kmp_info_t *thread,
2259                                             kmp_team_t *team) {
2260   // restore state outside the region
2261   thread->th.ompt_thread_info.state =
2262       ((team->t.t_serialized) ? ompt_state_work_serial
2263                               : ompt_state_work_parallel);
2264 }
2265 
2266 static inline void __kmp_join_ompt(int gtid, kmp_info_t *thread,
2267                                    kmp_team_t *team, ompt_data_t *parallel_data,
2268                                    int flags, void *codeptr) {
2269   ompt_task_info_t *task_info = __ompt_get_task_info_object(0);
2270   if (ompt_enabled.ompt_callback_parallel_end) {
2271     ompt_callbacks.ompt_callback(ompt_callback_parallel_end)(
2272         parallel_data, &(task_info->task_data), flags, codeptr);
2273   }
2274 
2275   task_info->frame.enter_frame = ompt_data_none;
2276   __kmp_join_restore_state(thread, team);
2277 }
2278 #endif
2279 
2280 void __kmp_join_call(ident_t *loc, int gtid
2281 #if OMPT_SUPPORT
2282                      ,
2283                      enum fork_context_e fork_context
2284 #endif
2285                      ,
2286                      int exit_teams) {
2287   KMP_TIME_DEVELOPER_PARTITIONED_BLOCK(KMP_join_call);
2288   kmp_team_t *team;
2289   kmp_team_t *parent_team;
2290   kmp_info_t *master_th;
2291   kmp_root_t *root;
2292   int master_active;
2293 
2294   KA_TRACE(20, ("__kmp_join_call: enter T#%d\n", gtid));
2295 
2296   /* setup current data */
2297   master_th = __kmp_threads[gtid];
2298   root = master_th->th.th_root;
2299   team = master_th->th.th_team;
2300   parent_team = team->t.t_parent;
2301 
2302   master_th->th.th_ident = loc;
2303 
2304 #if OMPT_SUPPORT
2305   void *team_microtask = (void *)team->t.t_pkfn;
2306   // For GOMP interface with serialized parallel, need the
2307   // __kmpc_end_serialized_parallel to call hooks for OMPT end-implicit-task
2308   // and end-parallel events.
2309   if (ompt_enabled.enabled &&
2310       !(team->t.t_serialized && fork_context == fork_context_gnu)) {
2311     master_th->th.ompt_thread_info.state = ompt_state_overhead;
2312   }
2313 #endif
2314 
2315 #if KMP_DEBUG
2316   if (__kmp_tasking_mode != tskm_immediate_exec && !exit_teams) {
2317     KA_TRACE(20, ("__kmp_join_call: T#%d, old team = %p old task_team = %p, "
2318                   "th_task_team = %p\n",
2319                   __kmp_gtid_from_thread(master_th), team,
2320                   team->t.t_task_team[master_th->th.th_task_state],
2321                   master_th->th.th_task_team));
2322     KMP_DEBUG_ASSERT(master_th->th.th_task_team ==
2323                      team->t.t_task_team[master_th->th.th_task_state]);
2324   }
2325 #endif
2326 
2327   if (team->t.t_serialized) {
2328     if (master_th->th.th_teams_microtask) {
2329       // We are in teams construct
2330       int level = team->t.t_level;
2331       int tlevel = master_th->th.th_teams_level;
2332       if (level == tlevel) {
2333         // AC: we haven't incremented it earlier at start of teams construct,
2334         //     so do it here - at the end of teams construct
2335         team->t.t_level++;
2336       } else if (level == tlevel + 1) {
2337         // AC: we are exiting parallel inside teams, need to increment
2338         // serialization in order to restore it in the next call to
2339         // __kmpc_end_serialized_parallel
2340         team->t.t_serialized++;
2341       }
2342     }
2343     __kmpc_end_serialized_parallel(loc, gtid);
2344 
2345 #if OMPT_SUPPORT
2346     if (ompt_enabled.enabled) {
2347       __kmp_join_restore_state(master_th, parent_team);
2348     }
2349 #endif
2350 
2351     return;
2352   }
2353 
2354   master_active = team->t.t_master_active;
2355 
2356   if (!exit_teams) {
2357     // AC: No barrier for internal teams at exit from teams construct.
2358     //     But there is barrier for external team (league).
2359     __kmp_internal_join(loc, gtid, team);
2360   } else {
2361     master_th->th.th_task_state =
2362         0; // AC: no tasking in teams (out of any parallel)
2363   }
2364 
2365   KMP_MB();
2366 
2367 #if OMPT_SUPPORT
2368   ompt_data_t *parallel_data = &(team->t.ompt_team_info.parallel_data);
2369   void *codeptr = team->t.ompt_team_info.master_return_address;
2370 #endif
2371 
2372 #if USE_ITT_BUILD
2373   if (__itt_stack_caller_create_ptr) {
2374     // destroy the stack stitching id after join barrier
2375     __kmp_itt_stack_caller_destroy((__itt_caller)team->t.t_stack_id);
2376   }
2377   // Mark end of "parallel" region for Intel(R) VTune(TM) analyzer.
2378   if (team->t.t_active_level == 1 &&
2379       (!master_th->th.th_teams_microtask || /* not in teams construct */
2380        master_th->th.th_teams_size.nteams == 1)) {
2381     master_th->th.th_ident = loc;
2382     // only one notification scheme (either "submit" or "forking/joined", not
2383     // both)
2384     if ((__itt_frame_submit_v3_ptr || KMP_ITT_DEBUG) &&
2385         __kmp_forkjoin_frames_mode == 3)
2386       __kmp_itt_frame_submit(gtid, team->t.t_region_time,
2387                              master_th->th.th_frame_time, 0, loc,
2388                              master_th->th.th_team_nproc, 1);
2389     else if ((__itt_frame_end_v3_ptr || KMP_ITT_DEBUG) &&
2390              !__kmp_forkjoin_frames_mode && __kmp_forkjoin_frames)
2391       __kmp_itt_region_joined(gtid);
2392   } // active_level == 1
2393 #endif /* USE_ITT_BUILD */
2394 
2395   if (master_th->th.th_teams_microtask && !exit_teams &&
2396       team->t.t_pkfn != (microtask_t)__kmp_teams_master &&
2397       team->t.t_level == master_th->th.th_teams_level + 1) {
2398 // AC: We need to leave the team structure intact at the end of parallel
2399 // inside the teams construct, so that at the next parallel same (hot) team
2400 // works, only adjust nesting levels
2401 #if OMPT_SUPPORT
2402     ompt_data_t ompt_parallel_data = ompt_data_none;
2403     if (ompt_enabled.enabled) {
2404       ompt_task_info_t *task_info = __ompt_get_task_info_object(0);
2405       if (ompt_enabled.ompt_callback_implicit_task) {
2406         int ompt_team_size = team->t.t_nproc;
2407         ompt_callbacks.ompt_callback(ompt_callback_implicit_task)(
2408             ompt_scope_end, NULL, &(task_info->task_data), ompt_team_size,
2409             OMPT_CUR_TASK_INFO(master_th)->thread_num, ompt_task_implicit);
2410       }
2411       task_info->frame.exit_frame = ompt_data_none;
2412       task_info->task_data = ompt_data_none;
2413       ompt_parallel_data = *OMPT_CUR_TEAM_DATA(master_th);
2414       __ompt_lw_taskteam_unlink(master_th);
2415     }
2416 #endif
2417     /* Decrement our nested depth level */
2418     team->t.t_level--;
2419     team->t.t_active_level--;
2420     KMP_ATOMIC_DEC(&root->r.r_in_parallel);
2421 
2422     // Restore number of threads in the team if needed. This code relies on
2423     // the proper adjustment of th_teams_size.nth after the fork in
2424     // __kmp_teams_master on each teams master in the case that
2425     // __kmp_reserve_threads reduced it.
2426     if (master_th->th.th_team_nproc < master_th->th.th_teams_size.nth) {
2427       int old_num = master_th->th.th_team_nproc;
2428       int new_num = master_th->th.th_teams_size.nth;
2429       kmp_info_t **other_threads = team->t.t_threads;
2430       team->t.t_nproc = new_num;
2431       for (int i = 0; i < old_num; ++i) {
2432         other_threads[i]->th.th_team_nproc = new_num;
2433       }
2434       // Adjust states of non-used threads of the team
2435       for (int i = old_num; i < new_num; ++i) {
2436         // Re-initialize thread's barrier data.
2437         KMP_DEBUG_ASSERT(other_threads[i]);
2438         kmp_balign_t *balign = other_threads[i]->th.th_bar;
2439         for (int b = 0; b < bs_last_barrier; ++b) {
2440           balign[b].bb.b_arrived = team->t.t_bar[b].b_arrived;
2441           KMP_DEBUG_ASSERT(balign[b].bb.wait_flag != KMP_BARRIER_PARENT_FLAG);
2442 #if USE_DEBUGGER
2443           balign[b].bb.b_worker_arrived = team->t.t_bar[b].b_team_arrived;
2444 #endif
2445         }
2446         if (__kmp_tasking_mode != tskm_immediate_exec) {
2447           // Synchronize thread's task state
2448           other_threads[i]->th.th_task_state = master_th->th.th_task_state;
2449         }
2450       }
2451     }
2452 
2453 #if OMPT_SUPPORT
2454     if (ompt_enabled.enabled) {
2455       __kmp_join_ompt(gtid, master_th, parent_team, &ompt_parallel_data,
2456                       OMPT_INVOKER(fork_context) | ompt_parallel_team, codeptr);
2457     }
2458 #endif
2459 
2460     return;
2461   }
2462 
2463   /* do cleanup and restore the parent team */
2464   master_th->th.th_info.ds.ds_tid = team->t.t_master_tid;
2465   master_th->th.th_local.this_construct = team->t.t_master_this_cons;
2466 
2467   master_th->th.th_dispatch = &parent_team->t.t_dispatch[team->t.t_master_tid];
2468 
2469   /* jc: The following lock has instructions with REL and ACQ semantics,
2470      separating the parallel user code called in this parallel region
2471      from the serial user code called after this function returns. */
2472   __kmp_acquire_bootstrap_lock(&__kmp_forkjoin_lock);
2473 
2474   if (!master_th->th.th_teams_microtask ||
2475       team->t.t_level > master_th->th.th_teams_level) {
2476     /* Decrement our nested depth level */
2477     KMP_ATOMIC_DEC(&root->r.r_in_parallel);
2478   }
2479   KMP_DEBUG_ASSERT(root->r.r_in_parallel >= 0);
2480 
2481 #if OMPT_SUPPORT
2482   if (ompt_enabled.enabled) {
2483     ompt_task_info_t *task_info = __ompt_get_task_info_object(0);
2484     if (ompt_enabled.ompt_callback_implicit_task) {
2485       int flags = (team_microtask == (void *)__kmp_teams_master)
2486                       ? ompt_task_initial
2487                       : ompt_task_implicit;
2488       int ompt_team_size = (flags == ompt_task_initial) ? 0 : team->t.t_nproc;
2489       ompt_callbacks.ompt_callback(ompt_callback_implicit_task)(
2490           ompt_scope_end, NULL, &(task_info->task_data), ompt_team_size,
2491           OMPT_CUR_TASK_INFO(master_th)->thread_num, flags);
2492     }
2493     task_info->frame.exit_frame = ompt_data_none;
2494     task_info->task_data = ompt_data_none;
2495   }
2496 #endif
2497 
2498   KF_TRACE(10, ("__kmp_join_call1: T#%d, this_thread=%p team=%p\n", 0,
2499                 master_th, team));
2500   __kmp_pop_current_task_from_thread(master_th);
2501 
2502 #if KMP_AFFINITY_SUPPORTED
2503   // Restore master thread's partition.
2504   master_th->th.th_first_place = team->t.t_first_place;
2505   master_th->th.th_last_place = team->t.t_last_place;
2506 #endif // KMP_AFFINITY_SUPPORTED
2507   master_th->th.th_def_allocator = team->t.t_def_allocator;
2508 
2509   updateHWFPControl(team);
2510 
2511   if (root->r.r_active != master_active)
2512     root->r.r_active = master_active;
2513 
2514   __kmp_free_team(root, team USE_NESTED_HOT_ARG(
2515                             master_th)); // this will free worker threads
2516 
2517   /* this race was fun to find. make sure the following is in the critical
2518      region otherwise assertions may fail occasionally since the old team may be
2519      reallocated and the hierarchy appears inconsistent. it is actually safe to
2520      run and won't cause any bugs, but will cause those assertion failures. it's
2521      only one deref&assign so might as well put this in the critical region */
2522   master_th->th.th_team = parent_team;
2523   master_th->th.th_team_nproc = parent_team->t.t_nproc;
2524   master_th->th.th_team_master = parent_team->t.t_threads[0];
2525   master_th->th.th_team_serialized = parent_team->t.t_serialized;
2526 
2527   /* restore serialized team, if need be */
2528   if (parent_team->t.t_serialized &&
2529       parent_team != master_th->th.th_serial_team &&
2530       parent_team != root->r.r_root_team) {
2531     __kmp_free_team(root,
2532                     master_th->th.th_serial_team USE_NESTED_HOT_ARG(NULL));
2533     master_th->th.th_serial_team = parent_team;
2534   }
2535 
2536   if (__kmp_tasking_mode != tskm_immediate_exec) {
2537     if (master_th->th.th_task_state_top >
2538         0) { // Restore task state from memo stack
2539       KMP_DEBUG_ASSERT(master_th->th.th_task_state_memo_stack);
2540       // Remember master's state if we re-use this nested hot team
2541       master_th->th.th_task_state_memo_stack[master_th->th.th_task_state_top] =
2542           master_th->th.th_task_state;
2543       --master_th->th.th_task_state_top; // pop
2544       // Now restore state at this level
2545       master_th->th.th_task_state =
2546           master_th->th
2547               .th_task_state_memo_stack[master_th->th.th_task_state_top];
2548     }
2549     // Copy the task team from the parent team to the master thread
2550     master_th->th.th_task_team =
2551         parent_team->t.t_task_team[master_th->th.th_task_state];
2552     KA_TRACE(20,
2553              ("__kmp_join_call: Master T#%d restoring task_team %p / team %p\n",
2554               __kmp_gtid_from_thread(master_th), master_th->th.th_task_team,
2555               parent_team));
2556   }
2557 
2558   // TODO: GEH - cannot do this assertion because root thread not set up as
2559   // executing
2560   // KMP_ASSERT( master_th->th.th_current_task->td_flags.executing == 0 );
2561   master_th->th.th_current_task->td_flags.executing = 1;
2562 
2563   __kmp_release_bootstrap_lock(&__kmp_forkjoin_lock);
2564 
2565 #if OMPT_SUPPORT
2566   int flags =
2567       OMPT_INVOKER(fork_context) |
2568       ((team_microtask == (void *)__kmp_teams_master) ? ompt_parallel_league
2569                                                       : ompt_parallel_team);
2570   if (ompt_enabled.enabled) {
2571     __kmp_join_ompt(gtid, master_th, parent_team, parallel_data, flags,
2572                     codeptr);
2573   }
2574 #endif
2575 
2576   KMP_MB();
2577   KA_TRACE(20, ("__kmp_join_call: exit T#%d\n", gtid));
2578 }
2579 
2580 /* Check whether we should push an internal control record onto the
2581    serial team stack.  If so, do it.  */
2582 void __kmp_save_internal_controls(kmp_info_t *thread) {
2583 
2584   if (thread->th.th_team != thread->th.th_serial_team) {
2585     return;
2586   }
2587   if (thread->th.th_team->t.t_serialized > 1) {
2588     int push = 0;
2589 
2590     if (thread->th.th_team->t.t_control_stack_top == NULL) {
2591       push = 1;
2592     } else {
2593       if (thread->th.th_team->t.t_control_stack_top->serial_nesting_level !=
2594           thread->th.th_team->t.t_serialized) {
2595         push = 1;
2596       }
2597     }
2598     if (push) { /* push a record on the serial team's stack */
2599       kmp_internal_control_t *control =
2600           (kmp_internal_control_t *)__kmp_allocate(
2601               sizeof(kmp_internal_control_t));
2602 
2603       copy_icvs(control, &thread->th.th_current_task->td_icvs);
2604 
2605       control->serial_nesting_level = thread->th.th_team->t.t_serialized;
2606 
2607       control->next = thread->th.th_team->t.t_control_stack_top;
2608       thread->th.th_team->t.t_control_stack_top = control;
2609     }
2610   }
2611 }
2612 
2613 /* Changes set_nproc */
2614 void __kmp_set_num_threads(int new_nth, int gtid) {
2615   kmp_info_t *thread;
2616   kmp_root_t *root;
2617 
2618   KF_TRACE(10, ("__kmp_set_num_threads: new __kmp_nth = %d\n", new_nth));
2619   KMP_DEBUG_ASSERT(__kmp_init_serial);
2620 
2621   if (new_nth < 1)
2622     new_nth = 1;
2623   else if (new_nth > __kmp_max_nth)
2624     new_nth = __kmp_max_nth;
2625 
2626   KMP_COUNT_VALUE(OMP_set_numthreads, new_nth);
2627   thread = __kmp_threads[gtid];
2628   if (thread->th.th_current_task->td_icvs.nproc == new_nth)
2629     return; // nothing to do
2630 
2631   __kmp_save_internal_controls(thread);
2632 
2633   set__nproc(thread, new_nth);
2634 
2635   // If this omp_set_num_threads() call will cause the hot team size to be
2636   // reduced (in the absence of a num_threads clause), then reduce it now,
2637   // rather than waiting for the next parallel region.
2638   root = thread->th.th_root;
2639   if (__kmp_init_parallel && (!root->r.r_active) &&
2640       (root->r.r_hot_team->t.t_nproc > new_nth)
2641 #if KMP_NESTED_HOT_TEAMS
2642       && __kmp_hot_teams_max_level && !__kmp_hot_teams_mode
2643 #endif
2644       ) {
2645     kmp_team_t *hot_team = root->r.r_hot_team;
2646     int f;
2647 
2648     __kmp_acquire_bootstrap_lock(&__kmp_forkjoin_lock);
2649 
2650     // Release the extra threads we don't need any more.
2651     for (f = new_nth; f < hot_team->t.t_nproc; f++) {
2652       KMP_DEBUG_ASSERT(hot_team->t.t_threads[f] != NULL);
2653       if (__kmp_tasking_mode != tskm_immediate_exec) {
2654         // When decreasing team size, threads no longer in the team should unref
2655         // task team.
2656         hot_team->t.t_threads[f]->th.th_task_team = NULL;
2657       }
2658       __kmp_free_thread(hot_team->t.t_threads[f]);
2659       hot_team->t.t_threads[f] = NULL;
2660     }
2661     hot_team->t.t_nproc = new_nth;
2662 #if KMP_NESTED_HOT_TEAMS
2663     if (thread->th.th_hot_teams) {
2664       KMP_DEBUG_ASSERT(hot_team == thread->th.th_hot_teams[0].hot_team);
2665       thread->th.th_hot_teams[0].hot_team_nth = new_nth;
2666     }
2667 #endif
2668 
2669     __kmp_release_bootstrap_lock(&__kmp_forkjoin_lock);
2670 
2671     // Update the t_nproc field in the threads that are still active.
2672     for (f = 0; f < new_nth; f++) {
2673       KMP_DEBUG_ASSERT(hot_team->t.t_threads[f] != NULL);
2674       hot_team->t.t_threads[f]->th.th_team_nproc = new_nth;
2675     }
2676     // Special flag in case omp_set_num_threads() call
2677     hot_team->t.t_size_changed = -1;
2678   }
2679 }
2680 
2681 /* Changes max_active_levels */
2682 void __kmp_set_max_active_levels(int gtid, int max_active_levels) {
2683   kmp_info_t *thread;
2684 
2685   KF_TRACE(10, ("__kmp_set_max_active_levels: new max_active_levels for thread "
2686                 "%d = (%d)\n",
2687                 gtid, max_active_levels));
2688   KMP_DEBUG_ASSERT(__kmp_init_serial);
2689 
2690   // validate max_active_levels
2691   if (max_active_levels < 0) {
2692     KMP_WARNING(ActiveLevelsNegative, max_active_levels);
2693     // We ignore this call if the user has specified a negative value.
2694     // The current setting won't be changed. The last valid setting will be
2695     // used. A warning will be issued (if warnings are allowed as controlled by
2696     // the KMP_WARNINGS env var).
2697     KF_TRACE(10, ("__kmp_set_max_active_levels: the call is ignored: new "
2698                   "max_active_levels for thread %d = (%d)\n",
2699                   gtid, max_active_levels));
2700     return;
2701   }
2702   if (max_active_levels <= KMP_MAX_ACTIVE_LEVELS_LIMIT) {
2703     // it's OK, the max_active_levels is within the valid range: [ 0;
2704     // KMP_MAX_ACTIVE_LEVELS_LIMIT ]
2705     // We allow a zero value. (implementation defined behavior)
2706   } else {
2707     KMP_WARNING(ActiveLevelsExceedLimit, max_active_levels,
2708                 KMP_MAX_ACTIVE_LEVELS_LIMIT);
2709     max_active_levels = KMP_MAX_ACTIVE_LEVELS_LIMIT;
2710     // Current upper limit is MAX_INT. (implementation defined behavior)
2711     // If the input exceeds the upper limit, we correct the input to be the
2712     // upper limit. (implementation defined behavior)
2713     // Actually, the flow should never get here until we use MAX_INT limit.
2714   }
2715   KF_TRACE(10, ("__kmp_set_max_active_levels: after validation: new "
2716                 "max_active_levels for thread %d = (%d)\n",
2717                 gtid, max_active_levels));
2718 
2719   thread = __kmp_threads[gtid];
2720 
2721   __kmp_save_internal_controls(thread);
2722 
2723   set__max_active_levels(thread, max_active_levels);
2724 }
2725 
2726 /* Gets max_active_levels */
2727 int __kmp_get_max_active_levels(int gtid) {
2728   kmp_info_t *thread;
2729 
2730   KF_TRACE(10, ("__kmp_get_max_active_levels: thread %d\n", gtid));
2731   KMP_DEBUG_ASSERT(__kmp_init_serial);
2732 
2733   thread = __kmp_threads[gtid];
2734   KMP_DEBUG_ASSERT(thread->th.th_current_task);
2735   KF_TRACE(10, ("__kmp_get_max_active_levels: thread %d, curtask=%p, "
2736                 "curtask_maxaclevel=%d\n",
2737                 gtid, thread->th.th_current_task,
2738                 thread->th.th_current_task->td_icvs.max_active_levels));
2739   return thread->th.th_current_task->td_icvs.max_active_levels;
2740 }
2741 
2742 KMP_BUILD_ASSERT(sizeof(kmp_sched_t) == sizeof(int));
2743 KMP_BUILD_ASSERT(sizeof(enum sched_type) == sizeof(int));
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_sched_t orig_kind;
2749   //    kmp_team_t *team;
2750 
2751   KF_TRACE(10, ("__kmp_set_schedule: new schedule for thread %d = (%d, %d)\n",
2752                 gtid, (int)kind, chunk));
2753   KMP_DEBUG_ASSERT(__kmp_init_serial);
2754 
2755   // Check if the kind parameter is valid, correct if needed.
2756   // Valid parameters should fit in one of two intervals - standard or extended:
2757   //       <lower>, <valid>, <upper_std>, <lower_ext>, <valid>, <upper>
2758   // 2008-01-25: 0,  1 - 4,       5,         100,     101 - 102, 103
2759   orig_kind = kind;
2760   kind = __kmp_sched_without_mods(kind);
2761 
2762   if (kind <= kmp_sched_lower || kind >= kmp_sched_upper ||
2763       (kind <= kmp_sched_lower_ext && kind >= kmp_sched_upper_std)) {
2764     // TODO: Hint needs attention in case we change the default schedule.
2765     __kmp_msg(kmp_ms_warning, KMP_MSG(ScheduleKindOutOfRange, kind),
2766               KMP_HNT(DefaultScheduleKindUsed, "static, no chunk"),
2767               __kmp_msg_null);
2768     kind = kmp_sched_default;
2769     chunk = 0; // ignore chunk value in case of bad kind
2770   }
2771 
2772   thread = __kmp_threads[gtid];
2773 
2774   __kmp_save_internal_controls(thread);
2775 
2776   if (kind < kmp_sched_upper_std) {
2777     if (kind == kmp_sched_static && chunk < KMP_DEFAULT_CHUNK) {
2778       // differ static chunked vs. unchunked:  chunk should be invalid to
2779       // indicate unchunked schedule (which is the default)
2780       thread->th.th_current_task->td_icvs.sched.r_sched_type = kmp_sch_static;
2781     } else {
2782       thread->th.th_current_task->td_icvs.sched.r_sched_type =
2783           __kmp_sch_map[kind - kmp_sched_lower - 1];
2784     }
2785   } else {
2786     //    __kmp_sch_map[ kind - kmp_sched_lower_ext + kmp_sched_upper_std -
2787     //    kmp_sched_lower - 2 ];
2788     thread->th.th_current_task->td_icvs.sched.r_sched_type =
2789         __kmp_sch_map[kind - kmp_sched_lower_ext + kmp_sched_upper_std -
2790                       kmp_sched_lower - 2];
2791   }
2792   __kmp_sched_apply_mods_intkind(
2793       orig_kind, &(thread->th.th_current_task->td_icvs.sched.r_sched_type));
2794   if (kind == kmp_sched_auto || chunk < 1) {
2795     // ignore parameter chunk for schedule auto
2796     thread->th.th_current_task->td_icvs.sched.chunk = KMP_DEFAULT_CHUNK;
2797   } else {
2798     thread->th.th_current_task->td_icvs.sched.chunk = chunk;
2799   }
2800 }
2801 
2802 /* Gets def_sched_var ICV values */
2803 void __kmp_get_schedule(int gtid, kmp_sched_t *kind, int *chunk) {
2804   kmp_info_t *thread;
2805   enum sched_type th_type;
2806 
2807   KF_TRACE(10, ("__kmp_get_schedule: thread %d\n", gtid));
2808   KMP_DEBUG_ASSERT(__kmp_init_serial);
2809 
2810   thread = __kmp_threads[gtid];
2811 
2812   th_type = thread->th.th_current_task->td_icvs.sched.r_sched_type;
2813   switch (SCHEDULE_WITHOUT_MODIFIERS(th_type)) {
2814   case kmp_sch_static:
2815   case kmp_sch_static_greedy:
2816   case kmp_sch_static_balanced:
2817     *kind = kmp_sched_static;
2818     __kmp_sched_apply_mods_stdkind(kind, th_type);
2819     *chunk = 0; // chunk was not set, try to show this fact via zero value
2820     return;
2821   case kmp_sch_static_chunked:
2822     *kind = kmp_sched_static;
2823     break;
2824   case kmp_sch_dynamic_chunked:
2825     *kind = kmp_sched_dynamic;
2826     break;
2827   case kmp_sch_guided_chunked:
2828   case kmp_sch_guided_iterative_chunked:
2829   case kmp_sch_guided_analytical_chunked:
2830     *kind = kmp_sched_guided;
2831     break;
2832   case kmp_sch_auto:
2833     *kind = kmp_sched_auto;
2834     break;
2835   case kmp_sch_trapezoidal:
2836     *kind = kmp_sched_trapezoidal;
2837     break;
2838 #if KMP_STATIC_STEAL_ENABLED
2839   case kmp_sch_static_steal:
2840     *kind = kmp_sched_static_steal;
2841     break;
2842 #endif
2843   default:
2844     KMP_FATAL(UnknownSchedulingType, th_type);
2845   }
2846 
2847   __kmp_sched_apply_mods_stdkind(kind, th_type);
2848   *chunk = thread->th.th_current_task->td_icvs.sched.chunk;
2849 }
2850 
2851 int __kmp_get_ancestor_thread_num(int gtid, int level) {
2852 
2853   int ii, dd;
2854   kmp_team_t *team;
2855   kmp_info_t *thr;
2856 
2857   KF_TRACE(10, ("__kmp_get_ancestor_thread_num: thread %d %d\n", gtid, level));
2858   KMP_DEBUG_ASSERT(__kmp_init_serial);
2859 
2860   // validate level
2861   if (level == 0)
2862     return 0;
2863   if (level < 0)
2864     return -1;
2865   thr = __kmp_threads[gtid];
2866   team = thr->th.th_team;
2867   ii = team->t.t_level;
2868   if (level > ii)
2869     return -1;
2870 
2871   if (thr->th.th_teams_microtask) {
2872     // AC: we are in teams region where multiple nested teams have same level
2873     int tlevel = thr->th.th_teams_level; // the level of the teams construct
2874     if (level <=
2875         tlevel) { // otherwise usual algorithm works (will not touch the teams)
2876       KMP_DEBUG_ASSERT(ii >= tlevel);
2877       // AC: As we need to pass by the teams league, we need to artificially
2878       // increase ii
2879       if (ii == tlevel) {
2880         ii += 2; // three teams have same level
2881       } else {
2882         ii++; // two teams have same level
2883       }
2884     }
2885   }
2886 
2887   if (ii == level)
2888     return __kmp_tid_from_gtid(gtid);
2889 
2890   dd = team->t.t_serialized;
2891   level++;
2892   while (ii > level) {
2893     for (dd = team->t.t_serialized; (dd > 0) && (ii > level); dd--, ii--) {
2894     }
2895     if ((team->t.t_serialized) && (!dd)) {
2896       team = team->t.t_parent;
2897       continue;
2898     }
2899     if (ii > level) {
2900       team = team->t.t_parent;
2901       dd = team->t.t_serialized;
2902       ii--;
2903     }
2904   }
2905 
2906   return (dd > 1) ? (0) : (team->t.t_master_tid);
2907 }
2908 
2909 int __kmp_get_team_size(int gtid, int level) {
2910 
2911   int ii, dd;
2912   kmp_team_t *team;
2913   kmp_info_t *thr;
2914 
2915   KF_TRACE(10, ("__kmp_get_team_size: thread %d %d\n", gtid, level));
2916   KMP_DEBUG_ASSERT(__kmp_init_serial);
2917 
2918   // validate level
2919   if (level == 0)
2920     return 1;
2921   if (level < 0)
2922     return -1;
2923   thr = __kmp_threads[gtid];
2924   team = thr->th.th_team;
2925   ii = team->t.t_level;
2926   if (level > ii)
2927     return -1;
2928 
2929   if (thr->th.th_teams_microtask) {
2930     // AC: we are in teams region where multiple nested teams have same level
2931     int tlevel = thr->th.th_teams_level; // the level of the teams construct
2932     if (level <=
2933         tlevel) { // otherwise usual algorithm works (will not touch the teams)
2934       KMP_DEBUG_ASSERT(ii >= tlevel);
2935       // AC: As we need to pass by the teams league, we need to artificially
2936       // increase ii
2937       if (ii == tlevel) {
2938         ii += 2; // three teams have same level
2939       } else {
2940         ii++; // two teams have same level
2941       }
2942     }
2943   }
2944 
2945   while (ii > level) {
2946     for (dd = team->t.t_serialized; (dd > 0) && (ii > level); dd--, ii--) {
2947     }
2948     if (team->t.t_serialized && (!dd)) {
2949       team = team->t.t_parent;
2950       continue;
2951     }
2952     if (ii > level) {
2953       team = team->t.t_parent;
2954       ii--;
2955     }
2956   }
2957 
2958   return team->t.t_nproc;
2959 }
2960 
2961 kmp_r_sched_t __kmp_get_schedule_global() {
2962   // This routine created because pairs (__kmp_sched, __kmp_chunk) and
2963   // (__kmp_static, __kmp_guided) may be changed by kmp_set_defaults
2964   // independently. So one can get the updated schedule here.
2965 
2966   kmp_r_sched_t r_sched;
2967 
2968   // create schedule from 4 globals: __kmp_sched, __kmp_chunk, __kmp_static,
2969   // __kmp_guided. __kmp_sched should keep original value, so that user can set
2970   // KMP_SCHEDULE multiple times, and thus have different run-time schedules in
2971   // different roots (even in OMP 2.5)
2972   enum sched_type s = SCHEDULE_WITHOUT_MODIFIERS(__kmp_sched);
2973   enum sched_type sched_modifiers = SCHEDULE_GET_MODIFIERS(__kmp_sched);
2974   if (s == kmp_sch_static) {
2975     // replace STATIC with more detailed schedule (balanced or greedy)
2976     r_sched.r_sched_type = __kmp_static;
2977   } else if (s == kmp_sch_guided_chunked) {
2978     // replace GUIDED with more detailed schedule (iterative or analytical)
2979     r_sched.r_sched_type = __kmp_guided;
2980   } else { // (STATIC_CHUNKED), or (DYNAMIC_CHUNKED), or other
2981     r_sched.r_sched_type = __kmp_sched;
2982   }
2983   SCHEDULE_SET_MODIFIERS(r_sched.r_sched_type, sched_modifiers);
2984 
2985   if (__kmp_chunk < KMP_DEFAULT_CHUNK) {
2986     // __kmp_chunk may be wrong here (if it was not ever set)
2987     r_sched.chunk = KMP_DEFAULT_CHUNK;
2988   } else {
2989     r_sched.chunk = __kmp_chunk;
2990   }
2991 
2992   return r_sched;
2993 }
2994 
2995 /* Allocate (realloc == FALSE) * or reallocate (realloc == TRUE)
2996    at least argc number of *t_argv entries for the requested team. */
2997 static void __kmp_alloc_argv_entries(int argc, kmp_team_t *team, int realloc) {
2998 
2999   KMP_DEBUG_ASSERT(team);
3000   if (!realloc || argc > team->t.t_max_argc) {
3001 
3002     KA_TRACE(100, ("__kmp_alloc_argv_entries: team %d: needed entries=%d, "
3003                    "current entries=%d\n",
3004                    team->t.t_id, argc, (realloc) ? team->t.t_max_argc : 0));
3005     /* if previously allocated heap space for args, free them */
3006     if (realloc && team->t.t_argv != &team->t.t_inline_argv[0])
3007       __kmp_free((void *)team->t.t_argv);
3008 
3009     if (argc <= KMP_INLINE_ARGV_ENTRIES) {
3010       /* use unused space in the cache line for arguments */
3011       team->t.t_max_argc = KMP_INLINE_ARGV_ENTRIES;
3012       KA_TRACE(100, ("__kmp_alloc_argv_entries: team %d: inline allocate %d "
3013                      "argv entries\n",
3014                      team->t.t_id, team->t.t_max_argc));
3015       team->t.t_argv = &team->t.t_inline_argv[0];
3016       if (__kmp_storage_map) {
3017         __kmp_print_storage_map_gtid(
3018             -1, &team->t.t_inline_argv[0],
3019             &team->t.t_inline_argv[KMP_INLINE_ARGV_ENTRIES],
3020             (sizeof(void *) * KMP_INLINE_ARGV_ENTRIES), "team_%d.t_inline_argv",
3021             team->t.t_id);
3022       }
3023     } else {
3024       /* allocate space for arguments in the heap */
3025       team->t.t_max_argc = (argc <= (KMP_MIN_MALLOC_ARGV_ENTRIES >> 1))
3026                                ? KMP_MIN_MALLOC_ARGV_ENTRIES
3027                                : 2 * argc;
3028       KA_TRACE(100, ("__kmp_alloc_argv_entries: team %d: dynamic allocate %d "
3029                      "argv entries\n",
3030                      team->t.t_id, team->t.t_max_argc));
3031       team->t.t_argv =
3032           (void **)__kmp_page_allocate(sizeof(void *) * team->t.t_max_argc);
3033       if (__kmp_storage_map) {
3034         __kmp_print_storage_map_gtid(-1, &team->t.t_argv[0],
3035                                      &team->t.t_argv[team->t.t_max_argc],
3036                                      sizeof(void *) * team->t.t_max_argc,
3037                                      "team_%d.t_argv", team->t.t_id);
3038       }
3039     }
3040   }
3041 }
3042 
3043 static void __kmp_allocate_team_arrays(kmp_team_t *team, int max_nth) {
3044   int i;
3045   int num_disp_buff = max_nth > 1 ? __kmp_dispatch_num_buffers : 2;
3046   team->t.t_threads =
3047       (kmp_info_t **)__kmp_allocate(sizeof(kmp_info_t *) * max_nth);
3048   team->t.t_disp_buffer = (dispatch_shared_info_t *)__kmp_allocate(
3049       sizeof(dispatch_shared_info_t) * num_disp_buff);
3050   team->t.t_dispatch =
3051       (kmp_disp_t *)__kmp_allocate(sizeof(kmp_disp_t) * max_nth);
3052   team->t.t_implicit_task_taskdata =
3053       (kmp_taskdata_t *)__kmp_allocate(sizeof(kmp_taskdata_t) * max_nth);
3054   team->t.t_max_nproc = max_nth;
3055 
3056   /* setup dispatch buffers */
3057   for (i = 0; i < num_disp_buff; ++i) {
3058     team->t.t_disp_buffer[i].buffer_index = i;
3059     team->t.t_disp_buffer[i].doacross_buf_idx = i;
3060   }
3061 }
3062 
3063 static void __kmp_free_team_arrays(kmp_team_t *team) {
3064   /* Note: this does not free the threads in t_threads (__kmp_free_threads) */
3065   int i;
3066   for (i = 0; i < team->t.t_max_nproc; ++i) {
3067     if (team->t.t_dispatch[i].th_disp_buffer != NULL) {
3068       __kmp_free(team->t.t_dispatch[i].th_disp_buffer);
3069       team->t.t_dispatch[i].th_disp_buffer = NULL;
3070     }
3071   }
3072 #if KMP_USE_HIER_SCHED
3073   __kmp_dispatch_free_hierarchies(team);
3074 #endif
3075   __kmp_free(team->t.t_threads);
3076   __kmp_free(team->t.t_disp_buffer);
3077   __kmp_free(team->t.t_dispatch);
3078   __kmp_free(team->t.t_implicit_task_taskdata);
3079   team->t.t_threads = NULL;
3080   team->t.t_disp_buffer = NULL;
3081   team->t.t_dispatch = NULL;
3082   team->t.t_implicit_task_taskdata = 0;
3083 }
3084 
3085 static void __kmp_reallocate_team_arrays(kmp_team_t *team, int max_nth) {
3086   kmp_info_t **oldThreads = team->t.t_threads;
3087 
3088   __kmp_free(team->t.t_disp_buffer);
3089   __kmp_free(team->t.t_dispatch);
3090   __kmp_free(team->t.t_implicit_task_taskdata);
3091   __kmp_allocate_team_arrays(team, max_nth);
3092 
3093   KMP_MEMCPY(team->t.t_threads, oldThreads,
3094              team->t.t_nproc * sizeof(kmp_info_t *));
3095 
3096   __kmp_free(oldThreads);
3097 }
3098 
3099 static kmp_internal_control_t __kmp_get_global_icvs(void) {
3100 
3101   kmp_r_sched_t r_sched =
3102       __kmp_get_schedule_global(); // get current state of scheduling globals
3103 
3104   KMP_DEBUG_ASSERT(__kmp_nested_proc_bind.used > 0);
3105 
3106   kmp_internal_control_t g_icvs = {
3107     0, // int serial_nesting_level; //corresponds to value of th_team_serialized
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_cg_max_nth, // int thread_limit;
3121     __kmp_dflt_max_active_levels, // int max_active_levels; //internal control
3122     // for max_active_levels
3123     r_sched, // kmp_r_sched_t sched; //internal control for runtime schedule
3124     // {sched,chunk} pair
3125     __kmp_nested_proc_bind.bind_types[0],
3126     __kmp_default_device,
3127     NULL // struct kmp_internal_control *next;
3128   };
3129 
3130   return g_icvs;
3131 }
3132 
3133 static kmp_internal_control_t __kmp_get_x_global_icvs(const kmp_team_t *team) {
3134 
3135   kmp_internal_control_t gx_icvs;
3136   gx_icvs.serial_nesting_level =
3137       0; // probably =team->t.t_serial like in save_inter_controls
3138   copy_icvs(&gx_icvs, &team->t.t_threads[0]->th.th_current_task->td_icvs);
3139   gx_icvs.next = NULL;
3140 
3141   return gx_icvs;
3142 }
3143 
3144 static void __kmp_initialize_root(kmp_root_t *root) {
3145   int f;
3146   kmp_team_t *root_team;
3147   kmp_team_t *hot_team;
3148   int hot_team_max_nth;
3149   kmp_r_sched_t r_sched =
3150       __kmp_get_schedule_global(); // get current state of scheduling globals
3151   kmp_internal_control_t r_icvs = __kmp_get_global_icvs();
3152   KMP_DEBUG_ASSERT(root);
3153   KMP_ASSERT(!root->r.r_begin);
3154 
3155   /* setup the root state structure */
3156   __kmp_init_lock(&root->r.r_begin_lock);
3157   root->r.r_begin = FALSE;
3158   root->r.r_active = FALSE;
3159   root->r.r_in_parallel = 0;
3160   root->r.r_blocktime = __kmp_dflt_blocktime;
3161 
3162   /* setup the root team for this task */
3163   /* allocate the root team structure */
3164   KF_TRACE(10, ("__kmp_initialize_root: before root_team\n"));
3165 
3166   root_team =
3167       __kmp_allocate_team(root,
3168                           1, // new_nproc
3169                           1, // max_nproc
3170 #if OMPT_SUPPORT
3171                           ompt_data_none, // root parallel id
3172 #endif
3173                           __kmp_nested_proc_bind.bind_types[0], &r_icvs,
3174                           0 // argc
3175                           USE_NESTED_HOT_ARG(NULL) // master thread is unknown
3176                           );
3177 #if USE_DEBUGGER
3178   // Non-NULL value should be assigned to make the debugger display the root
3179   // team.
3180   TCW_SYNC_PTR(root_team->t.t_pkfn, (microtask_t)(~0));
3181 #endif
3182 
3183   KF_TRACE(10, ("__kmp_initialize_root: after root_team = %p\n", root_team));
3184 
3185   root->r.r_root_team = root_team;
3186   root_team->t.t_control_stack_top = NULL;
3187 
3188   /* initialize root team */
3189   root_team->t.t_threads[0] = NULL;
3190   root_team->t.t_nproc = 1;
3191   root_team->t.t_serialized = 1;
3192   // TODO???: root_team->t.t_max_active_levels = __kmp_dflt_max_active_levels;
3193   root_team->t.t_sched.sched = r_sched.sched;
3194   KA_TRACE(
3195       20,
3196       ("__kmp_initialize_root: init root team %d arrived: join=%u, plain=%u\n",
3197        root_team->t.t_id, KMP_INIT_BARRIER_STATE, KMP_INIT_BARRIER_STATE));
3198 
3199   /* setup the  hot team for this task */
3200   /* allocate the hot team structure */
3201   KF_TRACE(10, ("__kmp_initialize_root: before hot_team\n"));
3202 
3203   hot_team =
3204       __kmp_allocate_team(root,
3205                           1, // new_nproc
3206                           __kmp_dflt_team_nth_ub * 2, // max_nproc
3207 #if OMPT_SUPPORT
3208                           ompt_data_none, // root parallel id
3209 #endif
3210                           __kmp_nested_proc_bind.bind_types[0], &r_icvs,
3211                           0 // argc
3212                           USE_NESTED_HOT_ARG(NULL) // master thread is unknown
3213                           );
3214   KF_TRACE(10, ("__kmp_initialize_root: after hot_team = %p\n", hot_team));
3215 
3216   root->r.r_hot_team = hot_team;
3217   root_team->t.t_control_stack_top = NULL;
3218 
3219   /* first-time initialization */
3220   hot_team->t.t_parent = root_team;
3221 
3222   /* initialize hot team */
3223   hot_team_max_nth = hot_team->t.t_max_nproc;
3224   for (f = 0; f < hot_team_max_nth; ++f) {
3225     hot_team->t.t_threads[f] = NULL;
3226   }
3227   hot_team->t.t_nproc = 1;
3228   // TODO???: hot_team->t.t_max_active_levels = __kmp_dflt_max_active_levels;
3229   hot_team->t.t_sched.sched = r_sched.sched;
3230   hot_team->t.t_size_changed = 0;
3231 }
3232 
3233 #ifdef KMP_DEBUG
3234 
3235 typedef struct kmp_team_list_item {
3236   kmp_team_p const *entry;
3237   struct kmp_team_list_item *next;
3238 } kmp_team_list_item_t;
3239 typedef kmp_team_list_item_t *kmp_team_list_t;
3240 
3241 static void __kmp_print_structure_team_accum( // Add team to list of teams.
3242     kmp_team_list_t list, // List of teams.
3243     kmp_team_p const *team // Team to add.
3244     ) {
3245 
3246   // List must terminate with item where both entry and next are NULL.
3247   // Team is added to the list only once.
3248   // List is sorted in ascending order by team id.
3249   // Team id is *not* a key.
3250 
3251   kmp_team_list_t l;
3252 
3253   KMP_DEBUG_ASSERT(list != NULL);
3254   if (team == NULL) {
3255     return;
3256   }
3257 
3258   __kmp_print_structure_team_accum(list, team->t.t_parent);
3259   __kmp_print_structure_team_accum(list, team->t.t_next_pool);
3260 
3261   // Search list for the team.
3262   l = list;
3263   while (l->next != NULL && l->entry != team) {
3264     l = l->next;
3265   }
3266   if (l->next != NULL) {
3267     return; // Team has been added before, exit.
3268   }
3269 
3270   // Team is not found. Search list again for insertion point.
3271   l = list;
3272   while (l->next != NULL && l->entry->t.t_id <= team->t.t_id) {
3273     l = l->next;
3274   }
3275 
3276   // Insert team.
3277   {
3278     kmp_team_list_item_t *item = (kmp_team_list_item_t *)KMP_INTERNAL_MALLOC(
3279         sizeof(kmp_team_list_item_t));
3280     *item = *l;
3281     l->entry = team;
3282     l->next = item;
3283   }
3284 }
3285 
3286 static void __kmp_print_structure_team(char const *title, kmp_team_p const *team
3287 
3288                                        ) {
3289   __kmp_printf("%s", title);
3290   if (team != NULL) {
3291     __kmp_printf("%2x %p\n", team->t.t_id, team);
3292   } else {
3293     __kmp_printf(" - (nil)\n");
3294   }
3295 }
3296 
3297 static void __kmp_print_structure_thread(char const *title,
3298                                          kmp_info_p const *thread) {
3299   __kmp_printf("%s", title);
3300   if (thread != NULL) {
3301     __kmp_printf("%2d %p\n", thread->th.th_info.ds.ds_gtid, thread);
3302   } else {
3303     __kmp_printf(" - (nil)\n");
3304   }
3305 }
3306 
3307 void __kmp_print_structure(void) {
3308 
3309   kmp_team_list_t list;
3310 
3311   // Initialize list of teams.
3312   list =
3313       (kmp_team_list_item_t *)KMP_INTERNAL_MALLOC(sizeof(kmp_team_list_item_t));
3314   list->entry = NULL;
3315   list->next = NULL;
3316 
3317   __kmp_printf("\n------------------------------\nGlobal Thread "
3318                "Table\n------------------------------\n");
3319   {
3320     int gtid;
3321     for (gtid = 0; gtid < __kmp_threads_capacity; ++gtid) {
3322       __kmp_printf("%2d", gtid);
3323       if (__kmp_threads != NULL) {
3324         __kmp_printf(" %p", __kmp_threads[gtid]);
3325       }
3326       if (__kmp_root != NULL) {
3327         __kmp_printf(" %p", __kmp_root[gtid]);
3328       }
3329       __kmp_printf("\n");
3330     }
3331   }
3332 
3333   // Print out __kmp_threads array.
3334   __kmp_printf("\n------------------------------\nThreads\n--------------------"
3335                "----------\n");
3336   if (__kmp_threads != NULL) {
3337     int gtid;
3338     for (gtid = 0; gtid < __kmp_threads_capacity; ++gtid) {
3339       kmp_info_t const *thread = __kmp_threads[gtid];
3340       if (thread != NULL) {
3341         __kmp_printf("GTID %2d %p:\n", gtid, thread);
3342         __kmp_printf("    Our Root:        %p\n", thread->th.th_root);
3343         __kmp_print_structure_team("    Our Team:     ", thread->th.th_team);
3344         __kmp_print_structure_team("    Serial Team:  ",
3345                                    thread->th.th_serial_team);
3346         __kmp_printf("    Threads:      %2d\n", thread->th.th_team_nproc);
3347         __kmp_print_structure_thread("    Master:       ",
3348                                      thread->th.th_team_master);
3349         __kmp_printf("    Serialized?:  %2d\n", thread->th.th_team_serialized);
3350         __kmp_printf("    Set NProc:    %2d\n", thread->th.th_set_nproc);
3351         __kmp_printf("    Set Proc Bind: %2d\n", thread->th.th_set_proc_bind);
3352         __kmp_print_structure_thread("    Next in pool: ",
3353                                      thread->th.th_next_pool);
3354         __kmp_printf("\n");
3355         __kmp_print_structure_team_accum(list, thread->th.th_team);
3356         __kmp_print_structure_team_accum(list, thread->th.th_serial_team);
3357       }
3358     }
3359   } else {
3360     __kmp_printf("Threads array is not allocated.\n");
3361   }
3362 
3363   // Print out __kmp_root array.
3364   __kmp_printf("\n------------------------------\nUbers\n----------------------"
3365                "--------\n");
3366   if (__kmp_root != NULL) {
3367     int gtid;
3368     for (gtid = 0; gtid < __kmp_threads_capacity; ++gtid) {
3369       kmp_root_t const *root = __kmp_root[gtid];
3370       if (root != NULL) {
3371         __kmp_printf("GTID %2d %p:\n", gtid, root);
3372         __kmp_print_structure_team("    Root Team:    ", root->r.r_root_team);
3373         __kmp_print_structure_team("    Hot Team:     ", root->r.r_hot_team);
3374         __kmp_print_structure_thread("    Uber Thread:  ",
3375                                      root->r.r_uber_thread);
3376         __kmp_printf("    Active?:      %2d\n", root->r.r_active);
3377         __kmp_printf("    In Parallel:  %2d\n",
3378                      KMP_ATOMIC_LD_RLX(&root->r.r_in_parallel));
3379         __kmp_printf("\n");
3380         __kmp_print_structure_team_accum(list, root->r.r_root_team);
3381         __kmp_print_structure_team_accum(list, root->r.r_hot_team);
3382       }
3383     }
3384   } else {
3385     __kmp_printf("Ubers array is not allocated.\n");
3386   }
3387 
3388   __kmp_printf("\n------------------------------\nTeams\n----------------------"
3389                "--------\n");
3390   while (list->next != NULL) {
3391     kmp_team_p const *team = list->entry;
3392     int i;
3393     __kmp_printf("Team %2x %p:\n", team->t.t_id, team);
3394     __kmp_print_structure_team("    Parent Team:      ", team->t.t_parent);
3395     __kmp_printf("    Master TID:       %2d\n", team->t.t_master_tid);
3396     __kmp_printf("    Max threads:      %2d\n", team->t.t_max_nproc);
3397     __kmp_printf("    Levels of serial: %2d\n", team->t.t_serialized);
3398     __kmp_printf("    Number threads:   %2d\n", team->t.t_nproc);
3399     for (i = 0; i < team->t.t_nproc; ++i) {
3400       __kmp_printf("    Thread %2d:      ", i);
3401       __kmp_print_structure_thread("", team->t.t_threads[i]);
3402     }
3403     __kmp_print_structure_team("    Next in pool:     ", team->t.t_next_pool);
3404     __kmp_printf("\n");
3405     list = list->next;
3406   }
3407 
3408   // Print out __kmp_thread_pool and __kmp_team_pool.
3409   __kmp_printf("\n------------------------------\nPools\n----------------------"
3410                "--------\n");
3411   __kmp_print_structure_thread("Thread pool:          ",
3412                                CCAST(kmp_info_t *, __kmp_thread_pool));
3413   __kmp_print_structure_team("Team pool:            ",
3414                              CCAST(kmp_team_t *, __kmp_team_pool));
3415   __kmp_printf("\n");
3416 
3417   // Free team list.
3418   while (list != NULL) {
3419     kmp_team_list_item_t *item = list;
3420     list = list->next;
3421     KMP_INTERNAL_FREE(item);
3422   }
3423 }
3424 
3425 #endif
3426 
3427 //---------------------------------------------------------------------------
3428 //  Stuff for per-thread fast random number generator
3429 //  Table of primes
3430 static const unsigned __kmp_primes[] = {
3431     0x9e3779b1, 0xffe6cc59, 0x2109f6dd, 0x43977ab5, 0xba5703f5, 0xb495a877,
3432     0xe1626741, 0x79695e6b, 0xbc98c09f, 0xd5bee2b3, 0x287488f9, 0x3af18231,
3433     0x9677cd4d, 0xbe3a6929, 0xadc6a877, 0xdcf0674b, 0xbe4d6fe9, 0x5f15e201,
3434     0x99afc3fd, 0xf3f16801, 0xe222cfff, 0x24ba5fdb, 0x0620452d, 0x79f149e3,
3435     0xc8b93f49, 0x972702cd, 0xb07dd827, 0x6c97d5ed, 0x085a3d61, 0x46eb5ea7,
3436     0x3d9910ed, 0x2e687b5b, 0x29609227, 0x6eb081f1, 0x0954c4e1, 0x9d114db9,
3437     0x542acfa9, 0xb3e6bd7b, 0x0742d917, 0xe9f3ffa7, 0x54581edb, 0xf2480f45,
3438     0x0bb9288f, 0xef1affc7, 0x85fa0ca7, 0x3ccc14db, 0xe6baf34b, 0x343377f7,
3439     0x5ca19031, 0xe6d9293b, 0xf0a9f391, 0x5d2e980b, 0xfc411073, 0xc3749363,
3440     0xb892d829, 0x3549366b, 0x629750ad, 0xb98294e5, 0x892d9483, 0xc235baf3,
3441     0x3d2402a3, 0x6bdef3c9, 0xbec333cd, 0x40c9520f};
3442 
3443 //---------------------------------------------------------------------------
3444 //  __kmp_get_random: Get a random number using a linear congruential method.
3445 unsigned short __kmp_get_random(kmp_info_t *thread) {
3446   unsigned x = thread->th.th_x;
3447   unsigned short r = x >> 16;
3448 
3449   thread->th.th_x = x * thread->th.th_a + 1;
3450 
3451   KA_TRACE(30, ("__kmp_get_random: THREAD: %d, RETURN: %u\n",
3452                 thread->th.th_info.ds.ds_tid, r));
3453 
3454   return r;
3455 }
3456 //--------------------------------------------------------
3457 // __kmp_init_random: Initialize a random number generator
3458 void __kmp_init_random(kmp_info_t *thread) {
3459   unsigned seed = thread->th.th_info.ds.ds_tid;
3460 
3461   thread->th.th_a =
3462       __kmp_primes[seed % (sizeof(__kmp_primes) / sizeof(__kmp_primes[0]))];
3463   thread->th.th_x = (seed + 1) * thread->th.th_a + 1;
3464   KA_TRACE(30,
3465            ("__kmp_init_random: THREAD: %u; A: %u\n", seed, thread->th.th_a));
3466 }
3467 
3468 #if KMP_OS_WINDOWS
3469 /* reclaim array entries for root threads that are already dead, returns number
3470  * reclaimed */
3471 static int __kmp_reclaim_dead_roots(void) {
3472   int i, r = 0;
3473 
3474   for (i = 0; i < __kmp_threads_capacity; ++i) {
3475     if (KMP_UBER_GTID(i) &&
3476         !__kmp_still_running((kmp_info_t *)TCR_SYNC_PTR(__kmp_threads[i])) &&
3477         !__kmp_root[i]
3478              ->r.r_active) { // AC: reclaim only roots died in non-active state
3479       r += __kmp_unregister_root_other_thread(i);
3480     }
3481   }
3482   return r;
3483 }
3484 #endif
3485 
3486 /* This function attempts to create free entries in __kmp_threads and
3487    __kmp_root, and returns the number of free entries generated.
3488 
3489    For Windows* OS static library, the first mechanism used is to reclaim array
3490    entries for root threads that are already dead.
3491 
3492    On all platforms, expansion is attempted on the arrays __kmp_threads_ and
3493    __kmp_root, with appropriate update to __kmp_threads_capacity. Array
3494    capacity is increased by doubling with clipping to __kmp_tp_capacity, if
3495    threadprivate cache array has been created. Synchronization with
3496    __kmpc_threadprivate_cached is done using __kmp_tp_cached_lock.
3497 
3498    After any dead root reclamation, if the clipping value allows array expansion
3499    to result in the generation of a total of nNeed free slots, the function does
3500    that expansion. If not, nothing is done beyond the possible initial root
3501    thread reclamation.
3502 
3503    If any argument is negative, the behavior is undefined. */
3504 static int __kmp_expand_threads(int nNeed) {
3505   int added = 0;
3506   int minimumRequiredCapacity;
3507   int newCapacity;
3508   kmp_info_t **newThreads;
3509   kmp_root_t **newRoot;
3510 
3511 // All calls to __kmp_expand_threads should be under __kmp_forkjoin_lock, so
3512 // resizing __kmp_threads does not need additional protection if foreign
3513 // threads are present
3514 
3515 #if KMP_OS_WINDOWS && !KMP_DYNAMIC_LIB
3516   /* only for Windows static library */
3517   /* reclaim array entries for root threads that are already dead */
3518   added = __kmp_reclaim_dead_roots();
3519 
3520   if (nNeed) {
3521     nNeed -= added;
3522     if (nNeed < 0)
3523       nNeed = 0;
3524   }
3525 #endif
3526   if (nNeed <= 0)
3527     return added;
3528 
3529   // Note that __kmp_threads_capacity is not bounded by __kmp_max_nth. If
3530   // __kmp_max_nth is set to some value less than __kmp_sys_max_nth by the
3531   // user via KMP_DEVICE_THREAD_LIMIT, then __kmp_threads_capacity may become
3532   // > __kmp_max_nth in one of two ways:
3533   //
3534   // 1) The initialization thread (gtid = 0) exits.  __kmp_threads[0]
3535   //    may not be reused by another thread, so we may need to increase
3536   //    __kmp_threads_capacity to __kmp_max_nth + 1.
3537   //
3538   // 2) New foreign root(s) are encountered.  We always register new foreign
3539   //    roots. This may cause a smaller # of threads to be allocated at
3540   //    subsequent parallel regions, but the worker threads hang around (and
3541   //    eventually go to sleep) and need slots in the __kmp_threads[] array.
3542   //
3543   // Anyway, that is the reason for moving the check to see if
3544   // __kmp_max_nth was exceeded into __kmp_reserve_threads()
3545   // instead of having it performed here. -BB
3546 
3547   KMP_DEBUG_ASSERT(__kmp_sys_max_nth >= __kmp_threads_capacity);
3548 
3549   /* compute expansion headroom to check if we can expand */
3550   if (__kmp_sys_max_nth - __kmp_threads_capacity < nNeed) {
3551     /* possible expansion too small -- give up */
3552     return added;
3553   }
3554   minimumRequiredCapacity = __kmp_threads_capacity + nNeed;
3555 
3556   newCapacity = __kmp_threads_capacity;
3557   do {
3558     newCapacity = newCapacity <= (__kmp_sys_max_nth >> 1) ? (newCapacity << 1)
3559                                                           : __kmp_sys_max_nth;
3560   } while (newCapacity < minimumRequiredCapacity);
3561   newThreads = (kmp_info_t **)__kmp_allocate(
3562       (sizeof(kmp_info_t *) + sizeof(kmp_root_t *)) * newCapacity + CACHE_LINE);
3563   newRoot =
3564       (kmp_root_t **)((char *)newThreads + sizeof(kmp_info_t *) * newCapacity);
3565   KMP_MEMCPY(newThreads, __kmp_threads,
3566              __kmp_threads_capacity * sizeof(kmp_info_t *));
3567   KMP_MEMCPY(newRoot, __kmp_root,
3568              __kmp_threads_capacity * sizeof(kmp_root_t *));
3569 
3570   kmp_info_t **temp_threads = __kmp_threads;
3571   *(kmp_info_t * *volatile *)&__kmp_threads = newThreads;
3572   *(kmp_root_t * *volatile *)&__kmp_root = newRoot;
3573   __kmp_free(temp_threads);
3574   added += newCapacity - __kmp_threads_capacity;
3575   *(volatile int *)&__kmp_threads_capacity = newCapacity;
3576 
3577   if (newCapacity > __kmp_tp_capacity) {
3578     __kmp_acquire_bootstrap_lock(&__kmp_tp_cached_lock);
3579     if (__kmp_tp_cached && newCapacity > __kmp_tp_capacity) {
3580       __kmp_threadprivate_resize_cache(newCapacity);
3581     } else { // increase __kmp_tp_capacity to correspond with kmp_threads size
3582       *(volatile int *)&__kmp_tp_capacity = newCapacity;
3583     }
3584     __kmp_release_bootstrap_lock(&__kmp_tp_cached_lock);
3585   }
3586 
3587   return added;
3588 }
3589 
3590 /* Register the current thread as a root thread and obtain our gtid. We must
3591    have the __kmp_initz_lock held at this point. Argument TRUE only if are the
3592    thread that calls from __kmp_do_serial_initialize() */
3593 int __kmp_register_root(int initial_thread) {
3594   kmp_info_t *root_thread;
3595   kmp_root_t *root;
3596   int gtid;
3597   int capacity;
3598   __kmp_acquire_bootstrap_lock(&__kmp_forkjoin_lock);
3599   KA_TRACE(20, ("__kmp_register_root: entered\n"));
3600   KMP_MB();
3601 
3602   /* 2007-03-02:
3603      If initial thread did not invoke OpenMP RTL yet, and this thread is not an
3604      initial one, "__kmp_all_nth >= __kmp_threads_capacity" condition does not
3605      work as expected -- it may return false (that means there is at least one
3606      empty slot in __kmp_threads array), but it is possible the only free slot
3607      is #0, which is reserved for initial thread and so cannot be used for this
3608      one. Following code workarounds this bug.
3609 
3610      However, right solution seems to be not reserving slot #0 for initial
3611      thread because:
3612      (1) there is no magic in slot #0,
3613      (2) we cannot detect initial thread reliably (the first thread which does
3614         serial initialization may be not a real initial thread).
3615   */
3616   capacity = __kmp_threads_capacity;
3617   if (!initial_thread && TCR_PTR(__kmp_threads[0]) == NULL) {
3618     --capacity;
3619   }
3620 
3621   /* see if there are too many threads */
3622   if (__kmp_all_nth >= capacity && !__kmp_expand_threads(1)) {
3623     if (__kmp_tp_cached) {
3624       __kmp_fatal(KMP_MSG(CantRegisterNewThread),
3625                   KMP_HNT(Set_ALL_THREADPRIVATE, __kmp_tp_capacity),
3626                   KMP_HNT(PossibleSystemLimitOnThreads), __kmp_msg_null);
3627     } else {
3628       __kmp_fatal(KMP_MSG(CantRegisterNewThread), KMP_HNT(SystemLimitOnThreads),
3629                   __kmp_msg_null);
3630     }
3631   }
3632 
3633   /* find an available thread slot */
3634   /* Don't reassign the zero slot since we need that to only be used by initial
3635      thread */
3636   for (gtid = (initial_thread ? 0 : 1); TCR_PTR(__kmp_threads[gtid]) != NULL;
3637        gtid++)
3638     ;
3639   KA_TRACE(1,
3640            ("__kmp_register_root: found slot in threads array: T#%d\n", gtid));
3641   KMP_ASSERT(gtid < __kmp_threads_capacity);
3642 
3643   /* update global accounting */
3644   __kmp_all_nth++;
3645   TCW_4(__kmp_nth, __kmp_nth + 1);
3646 
3647   // if __kmp_adjust_gtid_mode is set, then we use method #1 (sp search) for low
3648   // numbers of procs, and method #2 (keyed API call) for higher numbers.
3649   if (__kmp_adjust_gtid_mode) {
3650     if (__kmp_all_nth >= __kmp_tls_gtid_min) {
3651       if (TCR_4(__kmp_gtid_mode) != 2) {
3652         TCW_4(__kmp_gtid_mode, 2);
3653       }
3654     } else {
3655       if (TCR_4(__kmp_gtid_mode) != 1) {
3656         TCW_4(__kmp_gtid_mode, 1);
3657       }
3658     }
3659   }
3660 
3661 #ifdef KMP_ADJUST_BLOCKTIME
3662   /* Adjust blocktime to zero if necessary            */
3663   /* Middle initialization might not have occurred yet */
3664   if (!__kmp_env_blocktime && (__kmp_avail_proc > 0)) {
3665     if (__kmp_nth > __kmp_avail_proc) {
3666       __kmp_zero_bt = TRUE;
3667     }
3668   }
3669 #endif /* KMP_ADJUST_BLOCKTIME */
3670 
3671   /* setup this new hierarchy */
3672   if (!(root = __kmp_root[gtid])) {
3673     root = __kmp_root[gtid] = (kmp_root_t *)__kmp_allocate(sizeof(kmp_root_t));
3674     KMP_DEBUG_ASSERT(!root->r.r_root_team);
3675   }
3676 
3677 #if KMP_STATS_ENABLED
3678   // Initialize stats as soon as possible (right after gtid assignment).
3679   __kmp_stats_thread_ptr = __kmp_stats_list->push_back(gtid);
3680   __kmp_stats_thread_ptr->startLife();
3681   KMP_SET_THREAD_STATE(SERIAL_REGION);
3682   KMP_INIT_PARTITIONED_TIMERS(OMP_serial);
3683 #endif
3684   __kmp_initialize_root(root);
3685 
3686   /* setup new root thread structure */
3687   if (root->r.r_uber_thread) {
3688     root_thread = root->r.r_uber_thread;
3689   } else {
3690     root_thread = (kmp_info_t *)__kmp_allocate(sizeof(kmp_info_t));
3691     if (__kmp_storage_map) {
3692       __kmp_print_thread_storage_map(root_thread, gtid);
3693     }
3694     root_thread->th.th_info.ds.ds_gtid = gtid;
3695 #if OMPT_SUPPORT
3696     root_thread->th.ompt_thread_info.thread_data = ompt_data_none;
3697 #endif
3698     root_thread->th.th_root = root;
3699     if (__kmp_env_consistency_check) {
3700       root_thread->th.th_cons = __kmp_allocate_cons_stack(gtid);
3701     }
3702 #if USE_FAST_MEMORY
3703     __kmp_initialize_fast_memory(root_thread);
3704 #endif /* USE_FAST_MEMORY */
3705 
3706 #if KMP_USE_BGET
3707     KMP_DEBUG_ASSERT(root_thread->th.th_local.bget_data == NULL);
3708     __kmp_initialize_bget(root_thread);
3709 #endif
3710     __kmp_init_random(root_thread); // Initialize random number generator
3711   }
3712 
3713   /* setup the serial team held in reserve by the root thread */
3714   if (!root_thread->th.th_serial_team) {
3715     kmp_internal_control_t r_icvs = __kmp_get_global_icvs();
3716     KF_TRACE(10, ("__kmp_register_root: before serial_team\n"));
3717     root_thread->th.th_serial_team = __kmp_allocate_team(
3718         root, 1, 1,
3719 #if OMPT_SUPPORT
3720         ompt_data_none, // root parallel id
3721 #endif
3722         proc_bind_default, &r_icvs, 0 USE_NESTED_HOT_ARG(NULL));
3723   }
3724   KMP_ASSERT(root_thread->th.th_serial_team);
3725   KF_TRACE(10, ("__kmp_register_root: after serial_team = %p\n",
3726                 root_thread->th.th_serial_team));
3727 
3728   /* drop root_thread into place */
3729   TCW_SYNC_PTR(__kmp_threads[gtid], root_thread);
3730 
3731   root->r.r_root_team->t.t_threads[0] = root_thread;
3732   root->r.r_hot_team->t.t_threads[0] = root_thread;
3733   root_thread->th.th_serial_team->t.t_threads[0] = root_thread;
3734   // AC: the team created in reserve, not for execution (it is unused for now).
3735   root_thread->th.th_serial_team->t.t_serialized = 0;
3736   root->r.r_uber_thread = root_thread;
3737 
3738   /* initialize the thread, get it ready to go */
3739   __kmp_initialize_info(root_thread, root->r.r_root_team, 0, gtid);
3740   TCW_4(__kmp_init_gtid, TRUE);
3741 
3742   /* prepare the master thread for get_gtid() */
3743   __kmp_gtid_set_specific(gtid);
3744 
3745 #if USE_ITT_BUILD
3746   __kmp_itt_thread_name(gtid);
3747 #endif /* USE_ITT_BUILD */
3748 
3749 #ifdef KMP_TDATA_GTID
3750   __kmp_gtid = gtid;
3751 #endif
3752   __kmp_create_worker(gtid, root_thread, __kmp_stksize);
3753   KMP_DEBUG_ASSERT(__kmp_gtid_get_specific() == gtid);
3754 
3755   KA_TRACE(20, ("__kmp_register_root: T#%d init T#%d(%d:%d) arrived: join=%u, "
3756                 "plain=%u\n",
3757                 gtid, __kmp_gtid_from_tid(0, root->r.r_hot_team),
3758                 root->r.r_hot_team->t.t_id, 0, KMP_INIT_BARRIER_STATE,
3759                 KMP_INIT_BARRIER_STATE));
3760   { // Initialize barrier data.
3761     int b;
3762     for (b = 0; b < bs_last_barrier; ++b) {
3763       root_thread->th.th_bar[b].bb.b_arrived = KMP_INIT_BARRIER_STATE;
3764 #if USE_DEBUGGER
3765       root_thread->th.th_bar[b].bb.b_worker_arrived = 0;
3766 #endif
3767     }
3768   }
3769   KMP_DEBUG_ASSERT(root->r.r_hot_team->t.t_bar[bs_forkjoin_barrier].b_arrived ==
3770                    KMP_INIT_BARRIER_STATE);
3771 
3772 #if KMP_AFFINITY_SUPPORTED
3773   root_thread->th.th_current_place = KMP_PLACE_UNDEFINED;
3774   root_thread->th.th_new_place = KMP_PLACE_UNDEFINED;
3775   root_thread->th.th_first_place = KMP_PLACE_UNDEFINED;
3776   root_thread->th.th_last_place = KMP_PLACE_UNDEFINED;
3777   if (TCR_4(__kmp_init_middle)) {
3778     __kmp_affinity_set_init_mask(gtid, TRUE);
3779   }
3780 #endif /* KMP_AFFINITY_SUPPORTED */
3781   root_thread->th.th_def_allocator = __kmp_def_allocator;
3782   root_thread->th.th_prev_level = 0;
3783   root_thread->th.th_prev_num_threads = 1;
3784 
3785   kmp_cg_root_t *tmp = (kmp_cg_root_t *)__kmp_allocate(sizeof(kmp_cg_root_t));
3786   tmp->cg_root = root_thread;
3787   tmp->cg_thread_limit = __kmp_cg_max_nth;
3788   tmp->cg_nthreads = 1;
3789   KA_TRACE(100, ("__kmp_register_root: Thread %p created node %p with"
3790                  " cg_nthreads init to 1\n",
3791                  root_thread, tmp));
3792   tmp->up = NULL;
3793   root_thread->th.th_cg_roots = tmp;
3794 
3795   __kmp_root_counter++;
3796 
3797 #if OMPT_SUPPORT
3798   if (!initial_thread && ompt_enabled.enabled) {
3799 
3800     kmp_info_t *root_thread = ompt_get_thread();
3801 
3802     ompt_set_thread_state(root_thread, ompt_state_overhead);
3803 
3804     if (ompt_enabled.ompt_callback_thread_begin) {
3805       ompt_callbacks.ompt_callback(ompt_callback_thread_begin)(
3806           ompt_thread_initial, __ompt_get_thread_data_internal());
3807     }
3808     ompt_data_t *task_data;
3809     ompt_data_t *parallel_data;
3810     __ompt_get_task_info_internal(0, NULL, &task_data, NULL, &parallel_data, NULL);
3811     if (ompt_enabled.ompt_callback_implicit_task) {
3812       ompt_callbacks.ompt_callback(ompt_callback_implicit_task)(
3813           ompt_scope_begin, parallel_data, task_data, 1, 1, ompt_task_initial);
3814     }
3815 
3816     ompt_set_thread_state(root_thread, ompt_state_work_serial);
3817   }
3818 #endif
3819 
3820   KMP_MB();
3821   __kmp_release_bootstrap_lock(&__kmp_forkjoin_lock);
3822 
3823   return gtid;
3824 }
3825 
3826 #if KMP_NESTED_HOT_TEAMS
3827 static int __kmp_free_hot_teams(kmp_root_t *root, kmp_info_t *thr, int level,
3828                                 const int max_level) {
3829   int i, n, nth;
3830   kmp_hot_team_ptr_t *hot_teams = thr->th.th_hot_teams;
3831   if (!hot_teams || !hot_teams[level].hot_team) {
3832     return 0;
3833   }
3834   KMP_DEBUG_ASSERT(level < max_level);
3835   kmp_team_t *team = hot_teams[level].hot_team;
3836   nth = hot_teams[level].hot_team_nth;
3837   n = nth - 1; // master is not freed
3838   if (level < max_level - 1) {
3839     for (i = 0; i < nth; ++i) {
3840       kmp_info_t *th = team->t.t_threads[i];
3841       n += __kmp_free_hot_teams(root, th, level + 1, max_level);
3842       if (i > 0 && th->th.th_hot_teams) {
3843         __kmp_free(th->th.th_hot_teams);
3844         th->th.th_hot_teams = NULL;
3845       }
3846     }
3847   }
3848   __kmp_free_team(root, team, NULL);
3849   return n;
3850 }
3851 #endif
3852 
3853 // Resets a root thread and clear its root and hot teams.
3854 // Returns the number of __kmp_threads entries directly and indirectly freed.
3855 static int __kmp_reset_root(int gtid, kmp_root_t *root) {
3856   kmp_team_t *root_team = root->r.r_root_team;
3857   kmp_team_t *hot_team = root->r.r_hot_team;
3858   int n = hot_team->t.t_nproc;
3859   int i;
3860 
3861   KMP_DEBUG_ASSERT(!root->r.r_active);
3862 
3863   root->r.r_root_team = NULL;
3864   root->r.r_hot_team = NULL;
3865   // __kmp_free_team() does not free hot teams, so we have to clear r_hot_team
3866   // before call to __kmp_free_team().
3867   __kmp_free_team(root, root_team USE_NESTED_HOT_ARG(NULL));
3868 #if KMP_NESTED_HOT_TEAMS
3869   if (__kmp_hot_teams_max_level >
3870       0) { // need to free nested hot teams and their threads if any
3871     for (i = 0; i < hot_team->t.t_nproc; ++i) {
3872       kmp_info_t *th = hot_team->t.t_threads[i];
3873       if (__kmp_hot_teams_max_level > 1) {
3874         n += __kmp_free_hot_teams(root, th, 1, __kmp_hot_teams_max_level);
3875       }
3876       if (th->th.th_hot_teams) {
3877         __kmp_free(th->th.th_hot_teams);
3878         th->th.th_hot_teams = NULL;
3879       }
3880     }
3881   }
3882 #endif
3883   __kmp_free_team(root, hot_team USE_NESTED_HOT_ARG(NULL));
3884 
3885   // Before we can reap the thread, we need to make certain that all other
3886   // threads in the teams that had this root as ancestor have stopped trying to
3887   // steal tasks.
3888   if (__kmp_tasking_mode != tskm_immediate_exec) {
3889     __kmp_wait_to_unref_task_teams();
3890   }
3891 
3892 #if KMP_OS_WINDOWS
3893   /* Close Handle of root duplicated in __kmp_create_worker (tr #62919) */
3894   KA_TRACE(
3895       10, ("__kmp_reset_root: free handle, th = %p, handle = %" KMP_UINTPTR_SPEC
3896            "\n",
3897            (LPVOID) & (root->r.r_uber_thread->th),
3898            root->r.r_uber_thread->th.th_info.ds.ds_thread));
3899   __kmp_free_handle(root->r.r_uber_thread->th.th_info.ds.ds_thread);
3900 #endif /* KMP_OS_WINDOWS */
3901 
3902 #if OMPT_SUPPORT
3903   ompt_data_t *task_data;
3904   ompt_data_t *parallel_data;
3905   __ompt_get_task_info_internal(0, NULL, &task_data, NULL, &parallel_data, NULL);
3906   if (ompt_enabled.ompt_callback_implicit_task) {
3907     ompt_callbacks.ompt_callback(ompt_callback_implicit_task)(
3908         ompt_scope_end, parallel_data, task_data, 0, 1, ompt_task_initial);
3909   }
3910   if (ompt_enabled.ompt_callback_thread_end) {
3911     ompt_callbacks.ompt_callback(ompt_callback_thread_end)(
3912         &(root->r.r_uber_thread->th.ompt_thread_info.thread_data));
3913   }
3914 #endif
3915 
3916   TCW_4(__kmp_nth,
3917         __kmp_nth - 1); // __kmp_reap_thread will decrement __kmp_all_nth.
3918   i = root->r.r_uber_thread->th.th_cg_roots->cg_nthreads--;
3919   KA_TRACE(100, ("__kmp_reset_root: Thread %p decrement cg_nthreads on node %p"
3920                  " to %d\n",
3921                  root->r.r_uber_thread, root->r.r_uber_thread->th.th_cg_roots,
3922                  root->r.r_uber_thread->th.th_cg_roots->cg_nthreads));
3923   if (i == 1) {
3924     // need to free contention group structure
3925     KMP_DEBUG_ASSERT(root->r.r_uber_thread ==
3926                      root->r.r_uber_thread->th.th_cg_roots->cg_root);
3927     KMP_DEBUG_ASSERT(root->r.r_uber_thread->th.th_cg_roots->up == NULL);
3928     __kmp_free(root->r.r_uber_thread->th.th_cg_roots);
3929     root->r.r_uber_thread->th.th_cg_roots = NULL;
3930   }
3931   __kmp_reap_thread(root->r.r_uber_thread, 1);
3932 
3933   // We canot put root thread to __kmp_thread_pool, so we have to reap it
3934   // instead of freeing.
3935   root->r.r_uber_thread = NULL;
3936   /* mark root as no longer in use */
3937   root->r.r_begin = FALSE;
3938 
3939   return n;
3940 }
3941 
3942 void __kmp_unregister_root_current_thread(int gtid) {
3943   KA_TRACE(1, ("__kmp_unregister_root_current_thread: enter T#%d\n", gtid));
3944   /* this lock should be ok, since unregister_root_current_thread is never
3945      called during an abort, only during a normal close. furthermore, if you
3946      have the forkjoin lock, you should never try to get the initz lock */
3947   __kmp_acquire_bootstrap_lock(&__kmp_forkjoin_lock);
3948   if (TCR_4(__kmp_global.g.g_done) || !__kmp_init_serial) {
3949     KC_TRACE(10, ("__kmp_unregister_root_current_thread: already finished, "
3950                   "exiting T#%d\n",
3951                   gtid));
3952     __kmp_release_bootstrap_lock(&__kmp_forkjoin_lock);
3953     return;
3954   }
3955   kmp_root_t *root = __kmp_root[gtid];
3956 
3957   KMP_DEBUG_ASSERT(__kmp_threads && __kmp_threads[gtid]);
3958   KMP_ASSERT(KMP_UBER_GTID(gtid));
3959   KMP_ASSERT(root == __kmp_threads[gtid]->th.th_root);
3960   KMP_ASSERT(root->r.r_active == FALSE);
3961 
3962   KMP_MB();
3963 
3964   kmp_info_t *thread = __kmp_threads[gtid];
3965   kmp_team_t *team = thread->th.th_team;
3966   kmp_task_team_t *task_team = thread->th.th_task_team;
3967 
3968   // we need to wait for the proxy tasks before finishing the thread
3969   if (task_team != NULL && task_team->tt.tt_found_proxy_tasks) {
3970 #if OMPT_SUPPORT
3971     // the runtime is shutting down so we won't report any events
3972     thread->th.ompt_thread_info.state = ompt_state_undefined;
3973 #endif
3974     __kmp_task_team_wait(thread, team USE_ITT_BUILD_ARG(NULL));
3975   }
3976 
3977   __kmp_reset_root(gtid, root);
3978 
3979   /* free up this thread slot */
3980   __kmp_gtid_set_specific(KMP_GTID_DNE);
3981 #ifdef KMP_TDATA_GTID
3982   __kmp_gtid = KMP_GTID_DNE;
3983 #endif
3984 
3985   KMP_MB();
3986   KC_TRACE(10,
3987            ("__kmp_unregister_root_current_thread: T#%d unregistered\n", gtid));
3988 
3989   __kmp_release_bootstrap_lock(&__kmp_forkjoin_lock);
3990 }
3991 
3992 #if KMP_OS_WINDOWS
3993 /* __kmp_forkjoin_lock must be already held
3994    Unregisters a root thread that is not the current thread.  Returns the number
3995    of __kmp_threads entries freed as a result. */
3996 static int __kmp_unregister_root_other_thread(int gtid) {
3997   kmp_root_t *root = __kmp_root[gtid];
3998   int r;
3999 
4000   KA_TRACE(1, ("__kmp_unregister_root_other_thread: enter T#%d\n", gtid));
4001   KMP_DEBUG_ASSERT(__kmp_threads && __kmp_threads[gtid]);
4002   KMP_ASSERT(KMP_UBER_GTID(gtid));
4003   KMP_ASSERT(root == __kmp_threads[gtid]->th.th_root);
4004   KMP_ASSERT(root->r.r_active == FALSE);
4005 
4006   r = __kmp_reset_root(gtid, root);
4007   KC_TRACE(10,
4008            ("__kmp_unregister_root_other_thread: T#%d unregistered\n", gtid));
4009   return r;
4010 }
4011 #endif
4012 
4013 #if KMP_DEBUG
4014 void __kmp_task_info() {
4015 
4016   kmp_int32 gtid = __kmp_entry_gtid();
4017   kmp_int32 tid = __kmp_tid_from_gtid(gtid);
4018   kmp_info_t *this_thr = __kmp_threads[gtid];
4019   kmp_team_t *steam = this_thr->th.th_serial_team;
4020   kmp_team_t *team = this_thr->th.th_team;
4021 
4022   __kmp_printf(
4023       "__kmp_task_info: gtid=%d tid=%d t_thread=%p team=%p steam=%p curtask=%p "
4024       "ptask=%p\n",
4025       gtid, tid, this_thr, team, steam, this_thr->th.th_current_task,
4026       team->t.t_implicit_task_taskdata[tid].td_parent);
4027 }
4028 #endif // KMP_DEBUG
4029 
4030 /* TODO optimize with one big memclr, take out what isn't needed, split
4031    responsibility to workers as much as possible, and delay initialization of
4032    features as much as possible  */
4033 static void __kmp_initialize_info(kmp_info_t *this_thr, kmp_team_t *team,
4034                                   int tid, int gtid) {
4035   /* this_thr->th.th_info.ds.ds_gtid is setup in
4036      kmp_allocate_thread/create_worker.
4037      this_thr->th.th_serial_team is setup in __kmp_allocate_thread */
4038   kmp_info_t *master = team->t.t_threads[0];
4039   KMP_DEBUG_ASSERT(this_thr != NULL);
4040   KMP_DEBUG_ASSERT(this_thr->th.th_serial_team);
4041   KMP_DEBUG_ASSERT(team);
4042   KMP_DEBUG_ASSERT(team->t.t_threads);
4043   KMP_DEBUG_ASSERT(team->t.t_dispatch);
4044   KMP_DEBUG_ASSERT(master);
4045   KMP_DEBUG_ASSERT(master->th.th_root);
4046 
4047   KMP_MB();
4048 
4049   TCW_SYNC_PTR(this_thr->th.th_team, team);
4050 
4051   this_thr->th.th_info.ds.ds_tid = tid;
4052   this_thr->th.th_set_nproc = 0;
4053   if (__kmp_tasking_mode != tskm_immediate_exec)
4054     // When tasking is possible, threads are not safe to reap until they are
4055     // done tasking; this will be set when tasking code is exited in wait
4056     this_thr->th.th_reap_state = KMP_NOT_SAFE_TO_REAP;
4057   else // no tasking --> always safe to reap
4058     this_thr->th.th_reap_state = KMP_SAFE_TO_REAP;
4059   this_thr->th.th_set_proc_bind = proc_bind_default;
4060 #if KMP_AFFINITY_SUPPORTED
4061   this_thr->th.th_new_place = this_thr->th.th_current_place;
4062 #endif
4063   this_thr->th.th_root = master->th.th_root;
4064 
4065   /* setup the thread's cache of the team structure */
4066   this_thr->th.th_team_nproc = team->t.t_nproc;
4067   this_thr->th.th_team_master = master;
4068   this_thr->th.th_team_serialized = team->t.t_serialized;
4069   TCW_PTR(this_thr->th.th_sleep_loc, NULL);
4070 
4071   KMP_DEBUG_ASSERT(team->t.t_implicit_task_taskdata);
4072 
4073   KF_TRACE(10, ("__kmp_initialize_info1: T#%d:%d this_thread=%p curtask=%p\n",
4074                 tid, gtid, this_thr, this_thr->th.th_current_task));
4075 
4076   __kmp_init_implicit_task(this_thr->th.th_team_master->th.th_ident, this_thr,
4077                            team, tid, TRUE);
4078 
4079   KF_TRACE(10, ("__kmp_initialize_info2: T#%d:%d this_thread=%p curtask=%p\n",
4080                 tid, gtid, this_thr, this_thr->th.th_current_task));
4081   // TODO: Initialize ICVs from parent; GEH - isn't that already done in
4082   // __kmp_initialize_team()?
4083 
4084   /* TODO no worksharing in speculative threads */
4085   this_thr->th.th_dispatch = &team->t.t_dispatch[tid];
4086 
4087   this_thr->th.th_local.this_construct = 0;
4088 
4089   if (!this_thr->th.th_pri_common) {
4090     this_thr->th.th_pri_common =
4091         (struct common_table *)__kmp_allocate(sizeof(struct common_table));
4092     if (__kmp_storage_map) {
4093       __kmp_print_storage_map_gtid(
4094           gtid, this_thr->th.th_pri_common, this_thr->th.th_pri_common + 1,
4095           sizeof(struct common_table), "th_%d.th_pri_common\n", gtid);
4096     }
4097     this_thr->th.th_pri_head = NULL;
4098   }
4099 
4100   if (this_thr != master && // Master's CG root is initialized elsewhere
4101       this_thr->th.th_cg_roots != master->th.th_cg_roots) { // CG root not set
4102     // Make new thread's CG root same as master's
4103     KMP_DEBUG_ASSERT(master->th.th_cg_roots);
4104     kmp_cg_root_t *tmp = this_thr->th.th_cg_roots;
4105     if (tmp) {
4106       // worker changes CG, need to check if old CG should be freed
4107       int i = tmp->cg_nthreads--;
4108       KA_TRACE(100, ("__kmp_initialize_info: Thread %p decrement cg_nthreads"
4109                      " on node %p of thread %p to %d\n",
4110                      this_thr, tmp, tmp->cg_root, tmp->cg_nthreads));
4111       if (i == 1) {
4112         __kmp_free(tmp); // last thread left CG --> free it
4113       }
4114     }
4115     this_thr->th.th_cg_roots = master->th.th_cg_roots;
4116     // Increment new thread's CG root's counter to add the new thread
4117     this_thr->th.th_cg_roots->cg_nthreads++;
4118     KA_TRACE(100, ("__kmp_initialize_info: Thread %p increment cg_nthreads on"
4119                    " node %p of thread %p to %d\n",
4120                    this_thr, this_thr->th.th_cg_roots,
4121                    this_thr->th.th_cg_roots->cg_root,
4122                    this_thr->th.th_cg_roots->cg_nthreads));
4123     this_thr->th.th_current_task->td_icvs.thread_limit =
4124         this_thr->th.th_cg_roots->cg_thread_limit;
4125   }
4126 
4127   /* Initialize dynamic dispatch */
4128   {
4129     volatile kmp_disp_t *dispatch = this_thr->th.th_dispatch;
4130     // Use team max_nproc since this will never change for the team.
4131     size_t disp_size =
4132         sizeof(dispatch_private_info_t) *
4133         (team->t.t_max_nproc == 1 ? 1 : __kmp_dispatch_num_buffers);
4134     KD_TRACE(10, ("__kmp_initialize_info: T#%d max_nproc: %d\n", gtid,
4135                   team->t.t_max_nproc));
4136     KMP_ASSERT(dispatch);
4137     KMP_DEBUG_ASSERT(team->t.t_dispatch);
4138     KMP_DEBUG_ASSERT(dispatch == &team->t.t_dispatch[tid]);
4139 
4140     dispatch->th_disp_index = 0;
4141     dispatch->th_doacross_buf_idx = 0;
4142     if (!dispatch->th_disp_buffer) {
4143       dispatch->th_disp_buffer =
4144           (dispatch_private_info_t *)__kmp_allocate(disp_size);
4145 
4146       if (__kmp_storage_map) {
4147         __kmp_print_storage_map_gtid(
4148             gtid, &dispatch->th_disp_buffer[0],
4149             &dispatch->th_disp_buffer[team->t.t_max_nproc == 1
4150                                           ? 1
4151                                           : __kmp_dispatch_num_buffers],
4152             disp_size, "th_%d.th_dispatch.th_disp_buffer "
4153                        "(team_%d.t_dispatch[%d].th_disp_buffer)",
4154             gtid, team->t.t_id, gtid);
4155       }
4156     } else {
4157       memset(&dispatch->th_disp_buffer[0], '\0', disp_size);
4158     }
4159 
4160     dispatch->th_dispatch_pr_current = 0;
4161     dispatch->th_dispatch_sh_current = 0;
4162 
4163     dispatch->th_deo_fcn = 0; /* ORDERED     */
4164     dispatch->th_dxo_fcn = 0; /* END ORDERED */
4165   }
4166 
4167   this_thr->th.th_next_pool = NULL;
4168 
4169   if (!this_thr->th.th_task_state_memo_stack) {
4170     size_t i;
4171     this_thr->th.th_task_state_memo_stack =
4172         (kmp_uint8 *)__kmp_allocate(4 * sizeof(kmp_uint8));
4173     this_thr->th.th_task_state_top = 0;
4174     this_thr->th.th_task_state_stack_sz = 4;
4175     for (i = 0; i < this_thr->th.th_task_state_stack_sz;
4176          ++i) // zero init the stack
4177       this_thr->th.th_task_state_memo_stack[i] = 0;
4178   }
4179 
4180   KMP_DEBUG_ASSERT(!this_thr->th.th_spin_here);
4181   KMP_DEBUG_ASSERT(this_thr->th.th_next_waiting == 0);
4182 
4183   KMP_MB();
4184 }
4185 
4186 /* allocate a new thread for the requesting team. this is only called from
4187    within a forkjoin critical section. we will first try to get an available
4188    thread from the thread pool. if none is available, we will fork a new one
4189    assuming we are able to create a new one. this should be assured, as the
4190    caller should check on this first. */
4191 kmp_info_t *__kmp_allocate_thread(kmp_root_t *root, kmp_team_t *team,
4192                                   int new_tid) {
4193   kmp_team_t *serial_team;
4194   kmp_info_t *new_thr;
4195   int new_gtid;
4196 
4197   KA_TRACE(20, ("__kmp_allocate_thread: T#%d\n", __kmp_get_gtid()));
4198   KMP_DEBUG_ASSERT(root && team);
4199 #if !KMP_NESTED_HOT_TEAMS
4200   KMP_DEBUG_ASSERT(KMP_MASTER_GTID(__kmp_get_gtid()));
4201 #endif
4202   KMP_MB();
4203 
4204   /* first, try to get one from the thread pool */
4205   if (__kmp_thread_pool) {
4206     new_thr = CCAST(kmp_info_t *, __kmp_thread_pool);
4207     __kmp_thread_pool = (volatile kmp_info_t *)new_thr->th.th_next_pool;
4208     if (new_thr == __kmp_thread_pool_insert_pt) {
4209       __kmp_thread_pool_insert_pt = NULL;
4210     }
4211     TCW_4(new_thr->th.th_in_pool, FALSE);
4212     __kmp_suspend_initialize_thread(new_thr);
4213     __kmp_lock_suspend_mx(new_thr);
4214     if (new_thr->th.th_active_in_pool == TRUE) {
4215       KMP_DEBUG_ASSERT(new_thr->th.th_active == TRUE);
4216       KMP_ATOMIC_DEC(&__kmp_thread_pool_active_nth);
4217       new_thr->th.th_active_in_pool = FALSE;
4218     }
4219     __kmp_unlock_suspend_mx(new_thr);
4220 
4221     KA_TRACE(20, ("__kmp_allocate_thread: T#%d using thread T#%d\n",
4222                   __kmp_get_gtid(), new_thr->th.th_info.ds.ds_gtid));
4223     KMP_ASSERT(!new_thr->th.th_team);
4224     KMP_DEBUG_ASSERT(__kmp_nth < __kmp_threads_capacity);
4225 
4226     /* setup the thread structure */
4227     __kmp_initialize_info(new_thr, team, new_tid,
4228                           new_thr->th.th_info.ds.ds_gtid);
4229     KMP_DEBUG_ASSERT(new_thr->th.th_serial_team);
4230 
4231     TCW_4(__kmp_nth, __kmp_nth + 1);
4232 
4233     new_thr->th.th_task_state = 0;
4234     new_thr->th.th_task_state_top = 0;
4235     new_thr->th.th_task_state_stack_sz = 4;
4236 
4237 #ifdef KMP_ADJUST_BLOCKTIME
4238     /* Adjust blocktime back to zero if necessary */
4239     /* Middle initialization might not have occurred yet */
4240     if (!__kmp_env_blocktime && (__kmp_avail_proc > 0)) {
4241       if (__kmp_nth > __kmp_avail_proc) {
4242         __kmp_zero_bt = TRUE;
4243       }
4244     }
4245 #endif /* KMP_ADJUST_BLOCKTIME */
4246 
4247 #if KMP_DEBUG
4248     // If thread entered pool via __kmp_free_thread, wait_flag should !=
4249     // KMP_BARRIER_PARENT_FLAG.
4250     int b;
4251     kmp_balign_t *balign = new_thr->th.th_bar;
4252     for (b = 0; b < bs_last_barrier; ++b)
4253       KMP_DEBUG_ASSERT(balign[b].bb.wait_flag != KMP_BARRIER_PARENT_FLAG);
4254 #endif
4255 
4256     KF_TRACE(10, ("__kmp_allocate_thread: T#%d using thread %p T#%d\n",
4257                   __kmp_get_gtid(), new_thr, new_thr->th.th_info.ds.ds_gtid));
4258 
4259     KMP_MB();
4260     return new_thr;
4261   }
4262 
4263   /* no, well fork a new one */
4264   KMP_ASSERT(__kmp_nth == __kmp_all_nth);
4265   KMP_ASSERT(__kmp_all_nth < __kmp_threads_capacity);
4266 
4267 #if KMP_USE_MONITOR
4268   // If this is the first worker thread the RTL is creating, then also
4269   // launch the monitor thread.  We try to do this as early as possible.
4270   if (!TCR_4(__kmp_init_monitor)) {
4271     __kmp_acquire_bootstrap_lock(&__kmp_monitor_lock);
4272     if (!TCR_4(__kmp_init_monitor)) {
4273       KF_TRACE(10, ("before __kmp_create_monitor\n"));
4274       TCW_4(__kmp_init_monitor, 1);
4275       __kmp_create_monitor(&__kmp_monitor);
4276       KF_TRACE(10, ("after __kmp_create_monitor\n"));
4277 #if KMP_OS_WINDOWS
4278       // AC: wait until monitor has started. This is a fix for CQ232808.
4279       // The reason is that if the library is loaded/unloaded in a loop with
4280       // small (parallel) work in between, then there is high probability that
4281       // monitor thread started after the library shutdown. At shutdown it is
4282       // too late to cope with the problem, because when the master is in
4283       // DllMain (process detach) the monitor has no chances to start (it is
4284       // blocked), and master has no means to inform the monitor that the
4285       // library has gone, because all the memory which the monitor can access
4286       // is going to be released/reset.
4287       while (TCR_4(__kmp_init_monitor) < 2) {
4288         KMP_YIELD(TRUE);
4289       }
4290       KF_TRACE(10, ("after monitor thread has started\n"));
4291 #endif
4292     }
4293     __kmp_release_bootstrap_lock(&__kmp_monitor_lock);
4294   }
4295 #endif
4296 
4297   KMP_MB();
4298   for (new_gtid = 1; TCR_PTR(__kmp_threads[new_gtid]) != NULL; ++new_gtid) {
4299     KMP_DEBUG_ASSERT(new_gtid < __kmp_threads_capacity);
4300   }
4301 
4302   /* allocate space for it. */
4303   new_thr = (kmp_info_t *)__kmp_allocate(sizeof(kmp_info_t));
4304 
4305   TCW_SYNC_PTR(__kmp_threads[new_gtid], new_thr);
4306 
4307 #if USE_ITT_BUILD && USE_ITT_NOTIFY && KMP_DEBUG
4308   // suppress race conditions detection on synchronization flags in debug mode
4309   // this helps to analyze library internals eliminating false positives
4310   __itt_suppress_mark_range(
4311       __itt_suppress_range, __itt_suppress_threading_errors,
4312       &new_thr->th.th_sleep_loc, sizeof(new_thr->th.th_sleep_loc));
4313   __itt_suppress_mark_range(
4314       __itt_suppress_range, __itt_suppress_threading_errors,
4315       &new_thr->th.th_reap_state, sizeof(new_thr->th.th_reap_state));
4316 #if KMP_OS_WINDOWS
4317   __itt_suppress_mark_range(
4318       __itt_suppress_range, __itt_suppress_threading_errors,
4319       &new_thr->th.th_suspend_init, sizeof(new_thr->th.th_suspend_init));
4320 #else
4321   __itt_suppress_mark_range(__itt_suppress_range,
4322                             __itt_suppress_threading_errors,
4323                             &new_thr->th.th_suspend_init_count,
4324                             sizeof(new_thr->th.th_suspend_init_count));
4325 #endif
4326   // TODO: check if we need to also suppress b_arrived flags
4327   __itt_suppress_mark_range(__itt_suppress_range,
4328                             __itt_suppress_threading_errors,
4329                             CCAST(kmp_uint64 *, &new_thr->th.th_bar[0].bb.b_go),
4330                             sizeof(new_thr->th.th_bar[0].bb.b_go));
4331   __itt_suppress_mark_range(__itt_suppress_range,
4332                             __itt_suppress_threading_errors,
4333                             CCAST(kmp_uint64 *, &new_thr->th.th_bar[1].bb.b_go),
4334                             sizeof(new_thr->th.th_bar[1].bb.b_go));
4335   __itt_suppress_mark_range(__itt_suppress_range,
4336                             __itt_suppress_threading_errors,
4337                             CCAST(kmp_uint64 *, &new_thr->th.th_bar[2].bb.b_go),
4338                             sizeof(new_thr->th.th_bar[2].bb.b_go));
4339 #endif /* USE_ITT_BUILD && USE_ITT_NOTIFY && KMP_DEBUG */
4340   if (__kmp_storage_map) {
4341     __kmp_print_thread_storage_map(new_thr, new_gtid);
4342   }
4343 
4344   // add the reserve serialized team, initialized from the team's master thread
4345   {
4346     kmp_internal_control_t r_icvs = __kmp_get_x_global_icvs(team);
4347     KF_TRACE(10, ("__kmp_allocate_thread: before th_serial/serial_team\n"));
4348     new_thr->th.th_serial_team = serial_team =
4349         (kmp_team_t *)__kmp_allocate_team(root, 1, 1,
4350 #if OMPT_SUPPORT
4351                                           ompt_data_none, // root parallel id
4352 #endif
4353                                           proc_bind_default, &r_icvs,
4354                                           0 USE_NESTED_HOT_ARG(NULL));
4355   }
4356   KMP_ASSERT(serial_team);
4357   serial_team->t.t_serialized = 0; // AC: the team created in reserve, not for
4358   // execution (it is unused for now).
4359   serial_team->t.t_threads[0] = new_thr;
4360   KF_TRACE(10,
4361            ("__kmp_allocate_thread: after th_serial/serial_team : new_thr=%p\n",
4362             new_thr));
4363 
4364   /* setup the thread structures */
4365   __kmp_initialize_info(new_thr, team, new_tid, new_gtid);
4366 
4367 #if USE_FAST_MEMORY
4368   __kmp_initialize_fast_memory(new_thr);
4369 #endif /* USE_FAST_MEMORY */
4370 
4371 #if KMP_USE_BGET
4372   KMP_DEBUG_ASSERT(new_thr->th.th_local.bget_data == NULL);
4373   __kmp_initialize_bget(new_thr);
4374 #endif
4375 
4376   __kmp_init_random(new_thr); // Initialize random number generator
4377 
4378   /* Initialize these only once when thread is grabbed for a team allocation */
4379   KA_TRACE(20,
4380            ("__kmp_allocate_thread: T#%d init go fork=%u, plain=%u\n",
4381             __kmp_get_gtid(), KMP_INIT_BARRIER_STATE, KMP_INIT_BARRIER_STATE));
4382 
4383   int b;
4384   kmp_balign_t *balign = new_thr->th.th_bar;
4385   for (b = 0; b < bs_last_barrier; ++b) {
4386     balign[b].bb.b_go = KMP_INIT_BARRIER_STATE;
4387     balign[b].bb.team = NULL;
4388     balign[b].bb.wait_flag = KMP_BARRIER_NOT_WAITING;
4389     balign[b].bb.use_oncore_barrier = 0;
4390   }
4391 
4392   new_thr->th.th_spin_here = FALSE;
4393   new_thr->th.th_next_waiting = 0;
4394 #if KMP_OS_UNIX
4395   new_thr->th.th_blocking = false;
4396 #endif
4397 
4398 #if KMP_AFFINITY_SUPPORTED
4399   new_thr->th.th_current_place = KMP_PLACE_UNDEFINED;
4400   new_thr->th.th_new_place = KMP_PLACE_UNDEFINED;
4401   new_thr->th.th_first_place = KMP_PLACE_UNDEFINED;
4402   new_thr->th.th_last_place = KMP_PLACE_UNDEFINED;
4403 #endif
4404   new_thr->th.th_def_allocator = __kmp_def_allocator;
4405   new_thr->th.th_prev_level = 0;
4406   new_thr->th.th_prev_num_threads = 1;
4407 
4408   TCW_4(new_thr->th.th_in_pool, FALSE);
4409   new_thr->th.th_active_in_pool = FALSE;
4410   TCW_4(new_thr->th.th_active, TRUE);
4411 
4412   /* adjust the global counters */
4413   __kmp_all_nth++;
4414   __kmp_nth++;
4415 
4416   // if __kmp_adjust_gtid_mode is set, then we use method #1 (sp search) for low
4417   // numbers of procs, and method #2 (keyed API call) for higher numbers.
4418   if (__kmp_adjust_gtid_mode) {
4419     if (__kmp_all_nth >= __kmp_tls_gtid_min) {
4420       if (TCR_4(__kmp_gtid_mode) != 2) {
4421         TCW_4(__kmp_gtid_mode, 2);
4422       }
4423     } else {
4424       if (TCR_4(__kmp_gtid_mode) != 1) {
4425         TCW_4(__kmp_gtid_mode, 1);
4426       }
4427     }
4428   }
4429 
4430 #ifdef KMP_ADJUST_BLOCKTIME
4431   /* Adjust blocktime back to zero if necessary       */
4432   /* Middle initialization might not have occurred yet */
4433   if (!__kmp_env_blocktime && (__kmp_avail_proc > 0)) {
4434     if (__kmp_nth > __kmp_avail_proc) {
4435       __kmp_zero_bt = TRUE;
4436     }
4437   }
4438 #endif /* KMP_ADJUST_BLOCKTIME */
4439 
4440   /* actually fork it and create the new worker thread */
4441   KF_TRACE(
4442       10, ("__kmp_allocate_thread: before __kmp_create_worker: %p\n", new_thr));
4443   __kmp_create_worker(new_gtid, new_thr, __kmp_stksize);
4444   KF_TRACE(10,
4445            ("__kmp_allocate_thread: after __kmp_create_worker: %p\n", new_thr));
4446 
4447   KA_TRACE(20, ("__kmp_allocate_thread: T#%d forked T#%d\n", __kmp_get_gtid(),
4448                 new_gtid));
4449   KMP_MB();
4450   return new_thr;
4451 }
4452 
4453 /* Reinitialize team for reuse.
4454    The hot team code calls this case at every fork barrier, so EPCC barrier
4455    test are extremely sensitive to changes in it, esp. writes to the team
4456    struct, which cause a cache invalidation in all threads.
4457    IF YOU TOUCH THIS ROUTINE, RUN EPCC C SYNCBENCH ON A BIG-IRON MACHINE!!! */
4458 static void __kmp_reinitialize_team(kmp_team_t *team,
4459                                     kmp_internal_control_t *new_icvs,
4460                                     ident_t *loc) {
4461   KF_TRACE(10, ("__kmp_reinitialize_team: enter this_thread=%p team=%p\n",
4462                 team->t.t_threads[0], team));
4463   KMP_DEBUG_ASSERT(team && new_icvs);
4464   KMP_DEBUG_ASSERT((!TCR_4(__kmp_init_parallel)) || new_icvs->nproc);
4465   KMP_CHECK_UPDATE(team->t.t_ident, loc);
4466 
4467   KMP_CHECK_UPDATE(team->t.t_id, KMP_GEN_TEAM_ID());
4468   // Copy ICVs to the master thread's implicit taskdata
4469   __kmp_init_implicit_task(loc, team->t.t_threads[0], team, 0, FALSE);
4470   copy_icvs(&team->t.t_implicit_task_taskdata[0].td_icvs, new_icvs);
4471 
4472   KF_TRACE(10, ("__kmp_reinitialize_team: exit this_thread=%p team=%p\n",
4473                 team->t.t_threads[0], team));
4474 }
4475 
4476 /* Initialize the team data structure.
4477    This assumes the t_threads and t_max_nproc are already set.
4478    Also, we don't touch the arguments */
4479 static void __kmp_initialize_team(kmp_team_t *team, int new_nproc,
4480                                   kmp_internal_control_t *new_icvs,
4481                                   ident_t *loc) {
4482   KF_TRACE(10, ("__kmp_initialize_team: enter: team=%p\n", team));
4483 
4484   /* verify */
4485   KMP_DEBUG_ASSERT(team);
4486   KMP_DEBUG_ASSERT(new_nproc <= team->t.t_max_nproc);
4487   KMP_DEBUG_ASSERT(team->t.t_threads);
4488   KMP_MB();
4489 
4490   team->t.t_master_tid = 0; /* not needed */
4491   /* team->t.t_master_bar;        not needed */
4492   team->t.t_serialized = new_nproc > 1 ? 0 : 1;
4493   team->t.t_nproc = new_nproc;
4494 
4495   /* team->t.t_parent     = NULL; TODO not needed & would mess up hot team */
4496   team->t.t_next_pool = NULL;
4497   /* memset( team->t.t_threads, 0, sizeof(kmp_info_t*)*new_nproc ); would mess
4498    * up hot team */
4499 
4500   TCW_SYNC_PTR(team->t.t_pkfn, NULL); /* not needed */
4501   team->t.t_invoke = NULL; /* not needed */
4502 
4503   // TODO???: team->t.t_max_active_levels       = new_max_active_levels;
4504   team->t.t_sched.sched = new_icvs->sched.sched;
4505 
4506 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
4507   team->t.t_fp_control_saved = FALSE; /* not needed */
4508   team->t.t_x87_fpu_control_word = 0; /* not needed */
4509   team->t.t_mxcsr = 0; /* not needed */
4510 #endif /* KMP_ARCH_X86 || KMP_ARCH_X86_64 */
4511 
4512   team->t.t_construct = 0;
4513 
4514   team->t.t_ordered.dt.t_value = 0;
4515   team->t.t_master_active = FALSE;
4516 
4517 #ifdef KMP_DEBUG
4518   team->t.t_copypriv_data = NULL; /* not necessary, but nice for debugging */
4519 #endif
4520 #if KMP_OS_WINDOWS
4521   team->t.t_copyin_counter = 0; /* for barrier-free copyin implementation */
4522 #endif
4523 
4524   team->t.t_control_stack_top = NULL;
4525 
4526   __kmp_reinitialize_team(team, new_icvs, loc);
4527 
4528   KMP_MB();
4529   KF_TRACE(10, ("__kmp_initialize_team: exit: team=%p\n", team));
4530 }
4531 
4532 #if (KMP_OS_LINUX || KMP_OS_FREEBSD) && KMP_AFFINITY_SUPPORTED
4533 /* Sets full mask for thread and returns old mask, no changes to structures. */
4534 static void
4535 __kmp_set_thread_affinity_mask_full_tmp(kmp_affin_mask_t *old_mask) {
4536   if (KMP_AFFINITY_CAPABLE()) {
4537     int status;
4538     if (old_mask != NULL) {
4539       status = __kmp_get_system_affinity(old_mask, TRUE);
4540       int error = errno;
4541       if (status != 0) {
4542         __kmp_fatal(KMP_MSG(ChangeThreadAffMaskError), KMP_ERR(error),
4543                     __kmp_msg_null);
4544       }
4545     }
4546     __kmp_set_system_affinity(__kmp_affin_fullMask, TRUE);
4547   }
4548 }
4549 #endif
4550 
4551 #if KMP_AFFINITY_SUPPORTED
4552 
4553 // __kmp_partition_places() is the heart of the OpenMP 4.0 affinity mechanism.
4554 // It calculates the worker + master thread's partition based upon the parent
4555 // thread's partition, and binds each worker to a thread in their partition.
4556 // The master thread's partition should already include its current binding.
4557 static void __kmp_partition_places(kmp_team_t *team, int update_master_only) {
4558   // Copy the master thread's place partition to the team struct
4559   kmp_info_t *master_th = team->t.t_threads[0];
4560   KMP_DEBUG_ASSERT(master_th != NULL);
4561   kmp_proc_bind_t proc_bind = team->t.t_proc_bind;
4562   int first_place = master_th->th.th_first_place;
4563   int last_place = master_th->th.th_last_place;
4564   int masters_place = master_th->th.th_current_place;
4565   team->t.t_first_place = first_place;
4566   team->t.t_last_place = last_place;
4567 
4568   KA_TRACE(20, ("__kmp_partition_places: enter: proc_bind = %d T#%d(%d:0) "
4569                 "bound to place %d partition = [%d,%d]\n",
4570                 proc_bind, __kmp_gtid_from_thread(team->t.t_threads[0]),
4571                 team->t.t_id, masters_place, first_place, last_place));
4572 
4573   switch (proc_bind) {
4574 
4575   case proc_bind_default:
4576     // serial teams might have the proc_bind policy set to proc_bind_default. It
4577     // doesn't matter, as we don't rebind master thread for any proc_bind policy
4578     KMP_DEBUG_ASSERT(team->t.t_nproc == 1);
4579     break;
4580 
4581   case proc_bind_master: {
4582     int f;
4583     int n_th = team->t.t_nproc;
4584     for (f = 1; f < n_th; f++) {
4585       kmp_info_t *th = team->t.t_threads[f];
4586       KMP_DEBUG_ASSERT(th != NULL);
4587       th->th.th_first_place = first_place;
4588       th->th.th_last_place = last_place;
4589       th->th.th_new_place = masters_place;
4590       if (__kmp_display_affinity && masters_place != th->th.th_current_place &&
4591           team->t.t_display_affinity != 1) {
4592         team->t.t_display_affinity = 1;
4593       }
4594 
4595       KA_TRACE(100, ("__kmp_partition_places: master: T#%d(%d:%d) place %d "
4596                      "partition = [%d,%d]\n",
4597                      __kmp_gtid_from_thread(team->t.t_threads[f]), team->t.t_id,
4598                      f, masters_place, first_place, last_place));
4599     }
4600   } break;
4601 
4602   case proc_bind_close: {
4603     int f;
4604     int n_th = team->t.t_nproc;
4605     int n_places;
4606     if (first_place <= last_place) {
4607       n_places = last_place - first_place + 1;
4608     } else {
4609       n_places = __kmp_affinity_num_masks - first_place + last_place + 1;
4610     }
4611     if (n_th <= n_places) {
4612       int place = masters_place;
4613       for (f = 1; f < n_th; f++) {
4614         kmp_info_t *th = team->t.t_threads[f];
4615         KMP_DEBUG_ASSERT(th != NULL);
4616 
4617         if (place == last_place) {
4618           place = first_place;
4619         } else if (place == (int)(__kmp_affinity_num_masks - 1)) {
4620           place = 0;
4621         } else {
4622           place++;
4623         }
4624         th->th.th_first_place = first_place;
4625         th->th.th_last_place = last_place;
4626         th->th.th_new_place = place;
4627         if (__kmp_display_affinity && place != th->th.th_current_place &&
4628             team->t.t_display_affinity != 1) {
4629           team->t.t_display_affinity = 1;
4630         }
4631 
4632         KA_TRACE(100, ("__kmp_partition_places: close: T#%d(%d:%d) place %d "
4633                        "partition = [%d,%d]\n",
4634                        __kmp_gtid_from_thread(team->t.t_threads[f]),
4635                        team->t.t_id, f, place, first_place, last_place));
4636       }
4637     } else {
4638       int S, rem, gap, s_count;
4639       S = n_th / n_places;
4640       s_count = 0;
4641       rem = n_th - (S * n_places);
4642       gap = rem > 0 ? n_places / rem : n_places;
4643       int place = masters_place;
4644       int gap_ct = gap;
4645       for (f = 0; f < n_th; f++) {
4646         kmp_info_t *th = team->t.t_threads[f];
4647         KMP_DEBUG_ASSERT(th != NULL);
4648 
4649         th->th.th_first_place = first_place;
4650         th->th.th_last_place = last_place;
4651         th->th.th_new_place = place;
4652         if (__kmp_display_affinity && place != th->th.th_current_place &&
4653             team->t.t_display_affinity != 1) {
4654           team->t.t_display_affinity = 1;
4655         }
4656         s_count++;
4657 
4658         if ((s_count == S) && rem && (gap_ct == gap)) {
4659           // do nothing, add an extra thread to place on next iteration
4660         } else if ((s_count == S + 1) && rem && (gap_ct == gap)) {
4661           // we added an extra thread to this place; move to next place
4662           if (place == last_place) {
4663             place = first_place;
4664           } else if (place == (int)(__kmp_affinity_num_masks - 1)) {
4665             place = 0;
4666           } else {
4667             place++;
4668           }
4669           s_count = 0;
4670           gap_ct = 1;
4671           rem--;
4672         } else if (s_count == S) { // place full; don't add extra
4673           if (place == last_place) {
4674             place = first_place;
4675           } else if (place == (int)(__kmp_affinity_num_masks - 1)) {
4676             place = 0;
4677           } else {
4678             place++;
4679           }
4680           gap_ct++;
4681           s_count = 0;
4682         }
4683 
4684         KA_TRACE(100,
4685                  ("__kmp_partition_places: close: T#%d(%d:%d) place %d "
4686                   "partition = [%d,%d]\n",
4687                   __kmp_gtid_from_thread(team->t.t_threads[f]), team->t.t_id, f,
4688                   th->th.th_new_place, first_place, last_place));
4689       }
4690       KMP_DEBUG_ASSERT(place == masters_place);
4691     }
4692   } break;
4693 
4694   case proc_bind_spread: {
4695     int f;
4696     int n_th = team->t.t_nproc;
4697     int n_places;
4698     int thidx;
4699     if (first_place <= last_place) {
4700       n_places = last_place - first_place + 1;
4701     } else {
4702       n_places = __kmp_affinity_num_masks - first_place + last_place + 1;
4703     }
4704     if (n_th <= n_places) {
4705       int place = -1;
4706 
4707       if (n_places != static_cast<int>(__kmp_affinity_num_masks)) {
4708         int S = n_places / n_th;
4709         int s_count, rem, gap, gap_ct;
4710 
4711         place = masters_place;
4712         rem = n_places - n_th * S;
4713         gap = rem ? n_th / rem : 1;
4714         gap_ct = gap;
4715         thidx = n_th;
4716         if (update_master_only == 1)
4717           thidx = 1;
4718         for (f = 0; f < thidx; f++) {
4719           kmp_info_t *th = team->t.t_threads[f];
4720           KMP_DEBUG_ASSERT(th != NULL);
4721 
4722           th->th.th_first_place = place;
4723           th->th.th_new_place = place;
4724           if (__kmp_display_affinity && place != th->th.th_current_place &&
4725               team->t.t_display_affinity != 1) {
4726             team->t.t_display_affinity = 1;
4727           }
4728           s_count = 1;
4729           while (s_count < S) {
4730             if (place == last_place) {
4731               place = first_place;
4732             } else if (place == (int)(__kmp_affinity_num_masks - 1)) {
4733               place = 0;
4734             } else {
4735               place++;
4736             }
4737             s_count++;
4738           }
4739           if (rem && (gap_ct == gap)) {
4740             if (place == last_place) {
4741               place = first_place;
4742             } else if (place == (int)(__kmp_affinity_num_masks - 1)) {
4743               place = 0;
4744             } else {
4745               place++;
4746             }
4747             rem--;
4748             gap_ct = 0;
4749           }
4750           th->th.th_last_place = place;
4751           gap_ct++;
4752 
4753           if (place == last_place) {
4754             place = first_place;
4755           } else if (place == (int)(__kmp_affinity_num_masks - 1)) {
4756             place = 0;
4757           } else {
4758             place++;
4759           }
4760 
4761           KA_TRACE(100,
4762                    ("__kmp_partition_places: spread: T#%d(%d:%d) place %d "
4763                     "partition = [%d,%d], __kmp_affinity_num_masks: %u\n",
4764                     __kmp_gtid_from_thread(team->t.t_threads[f]), team->t.t_id,
4765                     f, th->th.th_new_place, th->th.th_first_place,
4766                     th->th.th_last_place, __kmp_affinity_num_masks));
4767         }
4768       } else {
4769         /* Having uniform space of available computation places I can create
4770            T partitions of round(P/T) size and put threads into the first
4771            place of each partition. */
4772         double current = static_cast<double>(masters_place);
4773         double spacing =
4774             (static_cast<double>(n_places + 1) / static_cast<double>(n_th));
4775         int first, last;
4776         kmp_info_t *th;
4777 
4778         thidx = n_th + 1;
4779         if (update_master_only == 1)
4780           thidx = 1;
4781         for (f = 0; f < thidx; f++) {
4782           first = static_cast<int>(current);
4783           last = static_cast<int>(current + spacing) - 1;
4784           KMP_DEBUG_ASSERT(last >= first);
4785           if (first >= n_places) {
4786             if (masters_place) {
4787               first -= n_places;
4788               last -= n_places;
4789               if (first == (masters_place + 1)) {
4790                 KMP_DEBUG_ASSERT(f == n_th);
4791                 first--;
4792               }
4793               if (last == masters_place) {
4794                 KMP_DEBUG_ASSERT(f == (n_th - 1));
4795                 last--;
4796               }
4797             } else {
4798               KMP_DEBUG_ASSERT(f == n_th);
4799               first = 0;
4800               last = 0;
4801             }
4802           }
4803           if (last >= n_places) {
4804             last = (n_places - 1);
4805           }
4806           place = first;
4807           current += spacing;
4808           if (f < n_th) {
4809             KMP_DEBUG_ASSERT(0 <= first);
4810             KMP_DEBUG_ASSERT(n_places > first);
4811             KMP_DEBUG_ASSERT(0 <= last);
4812             KMP_DEBUG_ASSERT(n_places > last);
4813             KMP_DEBUG_ASSERT(last_place >= first_place);
4814             th = team->t.t_threads[f];
4815             KMP_DEBUG_ASSERT(th);
4816             th->th.th_first_place = first;
4817             th->th.th_new_place = place;
4818             th->th.th_last_place = last;
4819             if (__kmp_display_affinity && place != th->th.th_current_place &&
4820                 team->t.t_display_affinity != 1) {
4821               team->t.t_display_affinity = 1;
4822             }
4823             KA_TRACE(100,
4824                      ("__kmp_partition_places: spread: T#%d(%d:%d) place %d "
4825                       "partition = [%d,%d], spacing = %.4f\n",
4826                       __kmp_gtid_from_thread(team->t.t_threads[f]),
4827                       team->t.t_id, f, th->th.th_new_place,
4828                       th->th.th_first_place, th->th.th_last_place, spacing));
4829           }
4830         }
4831       }
4832       KMP_DEBUG_ASSERT(update_master_only || place == masters_place);
4833     } else {
4834       int S, rem, gap, s_count;
4835       S = n_th / n_places;
4836       s_count = 0;
4837       rem = n_th - (S * n_places);
4838       gap = rem > 0 ? n_places / rem : n_places;
4839       int place = masters_place;
4840       int gap_ct = gap;
4841       thidx = n_th;
4842       if (update_master_only == 1)
4843         thidx = 1;
4844       for (f = 0; f < thidx; f++) {
4845         kmp_info_t *th = team->t.t_threads[f];
4846         KMP_DEBUG_ASSERT(th != NULL);
4847 
4848         th->th.th_first_place = place;
4849         th->th.th_last_place = place;
4850         th->th.th_new_place = place;
4851         if (__kmp_display_affinity && place != th->th.th_current_place &&
4852             team->t.t_display_affinity != 1) {
4853           team->t.t_display_affinity = 1;
4854         }
4855         s_count++;
4856 
4857         if ((s_count == S) && rem && (gap_ct == gap)) {
4858           // do nothing, add an extra thread to place on next iteration
4859         } else if ((s_count == S + 1) && rem && (gap_ct == gap)) {
4860           // we added an extra thread to this place; move on to next place
4861           if (place == last_place) {
4862             place = first_place;
4863           } else if (place == (int)(__kmp_affinity_num_masks - 1)) {
4864             place = 0;
4865           } else {
4866             place++;
4867           }
4868           s_count = 0;
4869           gap_ct = 1;
4870           rem--;
4871         } else if (s_count == S) { // place is full; don't add extra thread
4872           if (place == last_place) {
4873             place = first_place;
4874           } else if (place == (int)(__kmp_affinity_num_masks - 1)) {
4875             place = 0;
4876           } else {
4877             place++;
4878           }
4879           gap_ct++;
4880           s_count = 0;
4881         }
4882 
4883         KA_TRACE(100, ("__kmp_partition_places: spread: T#%d(%d:%d) place %d "
4884                        "partition = [%d,%d]\n",
4885                        __kmp_gtid_from_thread(team->t.t_threads[f]),
4886                        team->t.t_id, f, th->th.th_new_place,
4887                        th->th.th_first_place, th->th.th_last_place));
4888       }
4889       KMP_DEBUG_ASSERT(update_master_only || place == masters_place);
4890     }
4891   } break;
4892 
4893   default:
4894     break;
4895   }
4896 
4897   KA_TRACE(20, ("__kmp_partition_places: exit T#%d\n", team->t.t_id));
4898 }
4899 
4900 #endif // KMP_AFFINITY_SUPPORTED
4901 
4902 /* allocate a new team data structure to use.  take one off of the free pool if
4903    available */
4904 kmp_team_t *
4905 __kmp_allocate_team(kmp_root_t *root, int new_nproc, int max_nproc,
4906 #if OMPT_SUPPORT
4907                     ompt_data_t ompt_parallel_data,
4908 #endif
4909                     kmp_proc_bind_t new_proc_bind,
4910                     kmp_internal_control_t *new_icvs,
4911                     int argc USE_NESTED_HOT_ARG(kmp_info_t *master)) {
4912   KMP_TIME_DEVELOPER_PARTITIONED_BLOCK(KMP_allocate_team);
4913   int f;
4914   kmp_team_t *team;
4915   int use_hot_team = !root->r.r_active;
4916   int level = 0;
4917 
4918   KA_TRACE(20, ("__kmp_allocate_team: called\n"));
4919   KMP_DEBUG_ASSERT(new_nproc >= 1 && argc >= 0);
4920   KMP_DEBUG_ASSERT(max_nproc >= new_nproc);
4921   KMP_MB();
4922 
4923 #if KMP_NESTED_HOT_TEAMS
4924   kmp_hot_team_ptr_t *hot_teams;
4925   if (master) {
4926     team = master->th.th_team;
4927     level = team->t.t_active_level;
4928     if (master->th.th_teams_microtask) { // in teams construct?
4929       if (master->th.th_teams_size.nteams > 1 &&
4930           ( // #teams > 1
4931               team->t.t_pkfn ==
4932                   (microtask_t)__kmp_teams_master || // inner fork of the teams
4933               master->th.th_teams_level <
4934                   team->t.t_level)) { // or nested parallel inside the teams
4935         ++level; // not increment if #teams==1, or for outer fork of the teams;
4936         // increment otherwise
4937       }
4938     }
4939     hot_teams = master->th.th_hot_teams;
4940     if (level < __kmp_hot_teams_max_level && hot_teams &&
4941         hot_teams[level].hot_team) {
4942       // hot team has already been allocated for given level
4943       use_hot_team = 1;
4944     } else {
4945       use_hot_team = 0;
4946     }
4947   } else {
4948     // check we won't access uninitialized hot_teams, just in case
4949     KMP_DEBUG_ASSERT(new_nproc == 1);
4950   }
4951 #endif
4952   // Optimization to use a "hot" team
4953   if (use_hot_team && new_nproc > 1) {
4954     KMP_DEBUG_ASSERT(new_nproc <= max_nproc);
4955 #if KMP_NESTED_HOT_TEAMS
4956     team = hot_teams[level].hot_team;
4957 #else
4958     team = root->r.r_hot_team;
4959 #endif
4960 #if KMP_DEBUG
4961     if (__kmp_tasking_mode != tskm_immediate_exec) {
4962       KA_TRACE(20, ("__kmp_allocate_team: hot team task_team[0] = %p "
4963                     "task_team[1] = %p before reinit\n",
4964                     team->t.t_task_team[0], team->t.t_task_team[1]));
4965     }
4966 #endif
4967 
4968     // Has the number of threads changed?
4969     /* Let's assume the most common case is that the number of threads is
4970        unchanged, and put that case first. */
4971     if (team->t.t_nproc == new_nproc) { // Check changes in number of threads
4972       KA_TRACE(20, ("__kmp_allocate_team: reusing hot team\n"));
4973       // This case can mean that omp_set_num_threads() was called and the hot
4974       // team size was already reduced, so we check the special flag
4975       if (team->t.t_size_changed == -1) {
4976         team->t.t_size_changed = 1;
4977       } else {
4978         KMP_CHECK_UPDATE(team->t.t_size_changed, 0);
4979       }
4980 
4981       // TODO???: team->t.t_max_active_levels = new_max_active_levels;
4982       kmp_r_sched_t new_sched = new_icvs->sched;
4983       // set master's schedule as new run-time schedule
4984       KMP_CHECK_UPDATE(team->t.t_sched.sched, new_sched.sched);
4985 
4986       __kmp_reinitialize_team(team, new_icvs,
4987                               root->r.r_uber_thread->th.th_ident);
4988 
4989       KF_TRACE(10, ("__kmp_allocate_team2: T#%d, this_thread=%p team=%p\n", 0,
4990                     team->t.t_threads[0], team));
4991       __kmp_push_current_task_to_thread(team->t.t_threads[0], team, 0);
4992 
4993 #if KMP_AFFINITY_SUPPORTED
4994       if ((team->t.t_size_changed == 0) &&
4995           (team->t.t_proc_bind == new_proc_bind)) {
4996         if (new_proc_bind == proc_bind_spread) {
4997           __kmp_partition_places(
4998               team, 1); // add flag to update only master for spread
4999         }
5000         KA_TRACE(200, ("__kmp_allocate_team: reusing hot team #%d bindings: "
5001                        "proc_bind = %d, partition = [%d,%d]\n",
5002                        team->t.t_id, new_proc_bind, team->t.t_first_place,
5003                        team->t.t_last_place));
5004       } else {
5005         KMP_CHECK_UPDATE(team->t.t_proc_bind, new_proc_bind);
5006         __kmp_partition_places(team);
5007       }
5008 #else
5009       KMP_CHECK_UPDATE(team->t.t_proc_bind, new_proc_bind);
5010 #endif /* KMP_AFFINITY_SUPPORTED */
5011     } else if (team->t.t_nproc > new_nproc) {
5012       KA_TRACE(20,
5013                ("__kmp_allocate_team: decreasing hot team thread count to %d\n",
5014                 new_nproc));
5015 
5016       team->t.t_size_changed = 1;
5017 #if KMP_NESTED_HOT_TEAMS
5018       if (__kmp_hot_teams_mode == 0) {
5019         // AC: saved number of threads should correspond to team's value in this
5020         // mode, can be bigger in mode 1, when hot team has threads in reserve
5021         KMP_DEBUG_ASSERT(hot_teams[level].hot_team_nth == team->t.t_nproc);
5022         hot_teams[level].hot_team_nth = new_nproc;
5023 #endif // KMP_NESTED_HOT_TEAMS
5024         /* release the extra threads we don't need any more */
5025         for (f = new_nproc; f < team->t.t_nproc; f++) {
5026           KMP_DEBUG_ASSERT(team->t.t_threads[f]);
5027           if (__kmp_tasking_mode != tskm_immediate_exec) {
5028             // When decreasing team size, threads no longer in the team should
5029             // unref task team.
5030             team->t.t_threads[f]->th.th_task_team = NULL;
5031           }
5032           __kmp_free_thread(team->t.t_threads[f]);
5033           team->t.t_threads[f] = NULL;
5034         }
5035 #if KMP_NESTED_HOT_TEAMS
5036       } // (__kmp_hot_teams_mode == 0)
5037       else {
5038         // When keeping extra threads in team, switch threads to wait on own
5039         // b_go flag
5040         for (f = new_nproc; f < team->t.t_nproc; ++f) {
5041           KMP_DEBUG_ASSERT(team->t.t_threads[f]);
5042           kmp_balign_t *balign = team->t.t_threads[f]->th.th_bar;
5043           for (int b = 0; b < bs_last_barrier; ++b) {
5044             if (balign[b].bb.wait_flag == KMP_BARRIER_PARENT_FLAG) {
5045               balign[b].bb.wait_flag = KMP_BARRIER_SWITCH_TO_OWN_FLAG;
5046             }
5047             KMP_CHECK_UPDATE(balign[b].bb.leaf_kids, 0);
5048           }
5049         }
5050       }
5051 #endif // KMP_NESTED_HOT_TEAMS
5052       team->t.t_nproc = new_nproc;
5053       // TODO???: team->t.t_max_active_levels = new_max_active_levels;
5054       KMP_CHECK_UPDATE(team->t.t_sched.sched, new_icvs->sched.sched);
5055       __kmp_reinitialize_team(team, new_icvs,
5056                               root->r.r_uber_thread->th.th_ident);
5057 
5058       // Update remaining threads
5059       for (f = 0; f < new_nproc; ++f) {
5060         team->t.t_threads[f]->th.th_team_nproc = new_nproc;
5061       }
5062 
5063       // restore the current task state of the master thread: should be the
5064       // implicit task
5065       KF_TRACE(10, ("__kmp_allocate_team: T#%d, this_thread=%p team=%p\n", 0,
5066                     team->t.t_threads[0], team));
5067 
5068       __kmp_push_current_task_to_thread(team->t.t_threads[0], team, 0);
5069 
5070 #ifdef KMP_DEBUG
5071       for (f = 0; f < team->t.t_nproc; f++) {
5072         KMP_DEBUG_ASSERT(team->t.t_threads[f] &&
5073                          team->t.t_threads[f]->th.th_team_nproc ==
5074                              team->t.t_nproc);
5075       }
5076 #endif
5077 
5078       KMP_CHECK_UPDATE(team->t.t_proc_bind, new_proc_bind);
5079 #if KMP_AFFINITY_SUPPORTED
5080       __kmp_partition_places(team);
5081 #endif
5082     } else { // team->t.t_nproc < new_nproc
5083 #if (KMP_OS_LINUX || KMP_OS_FREEBSD) && KMP_AFFINITY_SUPPORTED
5084       kmp_affin_mask_t *old_mask;
5085       if (KMP_AFFINITY_CAPABLE()) {
5086         KMP_CPU_ALLOC(old_mask);
5087       }
5088 #endif
5089 
5090       KA_TRACE(20,
5091                ("__kmp_allocate_team: increasing hot team thread count to %d\n",
5092                 new_nproc));
5093 
5094       team->t.t_size_changed = 1;
5095 
5096 #if KMP_NESTED_HOT_TEAMS
5097       int avail_threads = hot_teams[level].hot_team_nth;
5098       if (new_nproc < avail_threads)
5099         avail_threads = new_nproc;
5100       kmp_info_t **other_threads = team->t.t_threads;
5101       for (f = team->t.t_nproc; f < avail_threads; ++f) {
5102         // Adjust barrier data of reserved threads (if any) of the team
5103         // Other data will be set in __kmp_initialize_info() below.
5104         int b;
5105         kmp_balign_t *balign = other_threads[f]->th.th_bar;
5106         for (b = 0; b < bs_last_barrier; ++b) {
5107           balign[b].bb.b_arrived = team->t.t_bar[b].b_arrived;
5108           KMP_DEBUG_ASSERT(balign[b].bb.wait_flag != KMP_BARRIER_PARENT_FLAG);
5109 #if USE_DEBUGGER
5110           balign[b].bb.b_worker_arrived = team->t.t_bar[b].b_team_arrived;
5111 #endif
5112         }
5113       }
5114       if (hot_teams[level].hot_team_nth >= new_nproc) {
5115         // we have all needed threads in reserve, no need to allocate any
5116         // this only possible in mode 1, cannot have reserved threads in mode 0
5117         KMP_DEBUG_ASSERT(__kmp_hot_teams_mode == 1);
5118         team->t.t_nproc = new_nproc; // just get reserved threads involved
5119       } else {
5120         // we may have some threads in reserve, but not enough
5121         team->t.t_nproc =
5122             hot_teams[level]
5123                 .hot_team_nth; // get reserved threads involved if any
5124         hot_teams[level].hot_team_nth = new_nproc; // adjust hot team max size
5125 #endif // KMP_NESTED_HOT_TEAMS
5126         if (team->t.t_max_nproc < new_nproc) {
5127           /* reallocate larger arrays */
5128           __kmp_reallocate_team_arrays(team, new_nproc);
5129           __kmp_reinitialize_team(team, new_icvs, NULL);
5130         }
5131 
5132 #if (KMP_OS_LINUX || KMP_OS_FREEBSD) && KMP_AFFINITY_SUPPORTED
5133         /* Temporarily set full mask for master thread before creation of
5134            workers. The reason is that workers inherit the affinity from master,
5135            so if a lot of workers are created on the single core quickly, they
5136            don't get a chance to set their own affinity for a long time. */
5137         __kmp_set_thread_affinity_mask_full_tmp(old_mask);
5138 #endif
5139 
5140         /* allocate new threads for the hot team */
5141         for (f = team->t.t_nproc; f < new_nproc; f++) {
5142           kmp_info_t *new_worker = __kmp_allocate_thread(root, team, f);
5143           KMP_DEBUG_ASSERT(new_worker);
5144           team->t.t_threads[f] = new_worker;
5145 
5146           KA_TRACE(20,
5147                    ("__kmp_allocate_team: team %d init T#%d arrived: "
5148                     "join=%llu, plain=%llu\n",
5149                     team->t.t_id, __kmp_gtid_from_tid(f, team), team->t.t_id, f,
5150                     team->t.t_bar[bs_forkjoin_barrier].b_arrived,
5151                     team->t.t_bar[bs_plain_barrier].b_arrived));
5152 
5153           { // Initialize barrier data for new threads.
5154             int b;
5155             kmp_balign_t *balign = new_worker->th.th_bar;
5156             for (b = 0; b < bs_last_barrier; ++b) {
5157               balign[b].bb.b_arrived = team->t.t_bar[b].b_arrived;
5158               KMP_DEBUG_ASSERT(balign[b].bb.wait_flag !=
5159                                KMP_BARRIER_PARENT_FLAG);
5160 #if USE_DEBUGGER
5161               balign[b].bb.b_worker_arrived = team->t.t_bar[b].b_team_arrived;
5162 #endif
5163             }
5164           }
5165         }
5166 
5167 #if (KMP_OS_LINUX || KMP_OS_FREEBSD) && KMP_AFFINITY_SUPPORTED
5168         if (KMP_AFFINITY_CAPABLE()) {
5169           /* Restore initial master thread's affinity mask */
5170           __kmp_set_system_affinity(old_mask, TRUE);
5171           KMP_CPU_FREE(old_mask);
5172         }
5173 #endif
5174 #if KMP_NESTED_HOT_TEAMS
5175       } // end of check of t_nproc vs. new_nproc vs. hot_team_nth
5176 #endif // KMP_NESTED_HOT_TEAMS
5177       /* make sure everyone is syncronized */
5178       int old_nproc = team->t.t_nproc; // save old value and use to update only
5179       // new threads below
5180       __kmp_initialize_team(team, new_nproc, new_icvs,
5181                             root->r.r_uber_thread->th.th_ident);
5182 
5183       /* reinitialize the threads */
5184       KMP_DEBUG_ASSERT(team->t.t_nproc == new_nproc);
5185       for (f = 0; f < team->t.t_nproc; ++f)
5186         __kmp_initialize_info(team->t.t_threads[f], team, f,
5187                               __kmp_gtid_from_tid(f, team));
5188 
5189       if (level) { // set th_task_state for new threads in nested hot team
5190         // __kmp_initialize_info() no longer zeroes th_task_state, so we should
5191         // only need to set the th_task_state for the new threads. th_task_state
5192         // for master thread will not be accurate until after this in
5193         // __kmp_fork_call(), so we look to the master's memo_stack to get the
5194         // correct value.
5195         for (f = old_nproc; f < team->t.t_nproc; ++f)
5196           team->t.t_threads[f]->th.th_task_state =
5197               team->t.t_threads[0]->th.th_task_state_memo_stack[level];
5198       } else { // set th_task_state for new threads in non-nested hot team
5199         int old_state =
5200             team->t.t_threads[0]->th.th_task_state; // copy master's state
5201         for (f = old_nproc; f < team->t.t_nproc; ++f)
5202           team->t.t_threads[f]->th.th_task_state = old_state;
5203       }
5204 
5205 #ifdef KMP_DEBUG
5206       for (f = 0; f < team->t.t_nproc; ++f) {
5207         KMP_DEBUG_ASSERT(team->t.t_threads[f] &&
5208                          team->t.t_threads[f]->th.th_team_nproc ==
5209                              team->t.t_nproc);
5210       }
5211 #endif
5212 
5213       KMP_CHECK_UPDATE(team->t.t_proc_bind, new_proc_bind);
5214 #if KMP_AFFINITY_SUPPORTED
5215       __kmp_partition_places(team);
5216 #endif
5217     } // Check changes in number of threads
5218 
5219     kmp_info_t *master = team->t.t_threads[0];
5220     if (master->th.th_teams_microtask) {
5221       for (f = 1; f < new_nproc; ++f) {
5222         // propagate teams construct specific info to workers
5223         kmp_info_t *thr = team->t.t_threads[f];
5224         thr->th.th_teams_microtask = master->th.th_teams_microtask;
5225         thr->th.th_teams_level = master->th.th_teams_level;
5226         thr->th.th_teams_size = master->th.th_teams_size;
5227       }
5228     }
5229 #if KMP_NESTED_HOT_TEAMS
5230     if (level) {
5231       // Sync barrier state for nested hot teams, not needed for outermost hot
5232       // team.
5233       for (f = 1; f < new_nproc; ++f) {
5234         kmp_info_t *thr = team->t.t_threads[f];
5235         int b;
5236         kmp_balign_t *balign = thr->th.th_bar;
5237         for (b = 0; b < bs_last_barrier; ++b) {
5238           balign[b].bb.b_arrived = team->t.t_bar[b].b_arrived;
5239           KMP_DEBUG_ASSERT(balign[b].bb.wait_flag != KMP_BARRIER_PARENT_FLAG);
5240 #if USE_DEBUGGER
5241           balign[b].bb.b_worker_arrived = team->t.t_bar[b].b_team_arrived;
5242 #endif
5243         }
5244       }
5245     }
5246 #endif // KMP_NESTED_HOT_TEAMS
5247 
5248     /* reallocate space for arguments if necessary */
5249     __kmp_alloc_argv_entries(argc, team, TRUE);
5250     KMP_CHECK_UPDATE(team->t.t_argc, argc);
5251     // The hot team re-uses the previous task team,
5252     // if untouched during the previous release->gather phase.
5253 
5254     KF_TRACE(10, (" hot_team = %p\n", team));
5255 
5256 #if KMP_DEBUG
5257     if (__kmp_tasking_mode != tskm_immediate_exec) {
5258       KA_TRACE(20, ("__kmp_allocate_team: hot team task_team[0] = %p "
5259                     "task_team[1] = %p after reinit\n",
5260                     team->t.t_task_team[0], team->t.t_task_team[1]));
5261     }
5262 #endif
5263 
5264 #if OMPT_SUPPORT
5265     __ompt_team_assign_id(team, ompt_parallel_data);
5266 #endif
5267 
5268     KMP_MB();
5269 
5270     return team;
5271   }
5272 
5273   /* next, let's try to take one from the team pool */
5274   KMP_MB();
5275   for (team = CCAST(kmp_team_t *, __kmp_team_pool); (team);) {
5276     /* TODO: consider resizing undersized teams instead of reaping them, now
5277        that we have a resizing mechanism */
5278     if (team->t.t_max_nproc >= max_nproc) {
5279       /* take this team from the team pool */
5280       __kmp_team_pool = team->t.t_next_pool;
5281 
5282       /* setup the team for fresh use */
5283       __kmp_initialize_team(team, new_nproc, new_icvs, NULL);
5284 
5285       KA_TRACE(20, ("__kmp_allocate_team: setting task_team[0] %p and "
5286                     "task_team[1] %p to NULL\n",
5287                     &team->t.t_task_team[0], &team->t.t_task_team[1]));
5288       team->t.t_task_team[0] = NULL;
5289       team->t.t_task_team[1] = NULL;
5290 
5291       /* reallocate space for arguments if necessary */
5292       __kmp_alloc_argv_entries(argc, team, TRUE);
5293       KMP_CHECK_UPDATE(team->t.t_argc, argc);
5294 
5295       KA_TRACE(
5296           20, ("__kmp_allocate_team: team %d init arrived: join=%u, plain=%u\n",
5297                team->t.t_id, KMP_INIT_BARRIER_STATE, KMP_INIT_BARRIER_STATE));
5298       { // Initialize barrier data.
5299         int b;
5300         for (b = 0; b < bs_last_barrier; ++b) {
5301           team->t.t_bar[b].b_arrived = KMP_INIT_BARRIER_STATE;
5302 #if USE_DEBUGGER
5303           team->t.t_bar[b].b_master_arrived = 0;
5304           team->t.t_bar[b].b_team_arrived = 0;
5305 #endif
5306         }
5307       }
5308 
5309       team->t.t_proc_bind = new_proc_bind;
5310 
5311       KA_TRACE(20, ("__kmp_allocate_team: using team from pool %d.\n",
5312                     team->t.t_id));
5313 
5314 #if OMPT_SUPPORT
5315       __ompt_team_assign_id(team, ompt_parallel_data);
5316 #endif
5317 
5318       KMP_MB();
5319 
5320       return team;
5321     }
5322 
5323     /* reap team if it is too small, then loop back and check the next one */
5324     // not sure if this is wise, but, will be redone during the hot-teams
5325     // rewrite.
5326     /* TODO: Use technique to find the right size hot-team, don't reap them */
5327     team = __kmp_reap_team(team);
5328     __kmp_team_pool = team;
5329   }
5330 
5331   /* nothing available in the pool, no matter, make a new team! */
5332   KMP_MB();
5333   team = (kmp_team_t *)__kmp_allocate(sizeof(kmp_team_t));
5334 
5335   /* and set it up */
5336   team->t.t_max_nproc = max_nproc;
5337   /* NOTE well, for some reason allocating one big buffer and dividing it up
5338      seems to really hurt performance a lot on the P4, so, let's not use this */
5339   __kmp_allocate_team_arrays(team, max_nproc);
5340 
5341   KA_TRACE(20, ("__kmp_allocate_team: making a new team\n"));
5342   __kmp_initialize_team(team, new_nproc, new_icvs, NULL);
5343 
5344   KA_TRACE(20, ("__kmp_allocate_team: setting task_team[0] %p and task_team[1] "
5345                 "%p to NULL\n",
5346                 &team->t.t_task_team[0], &team->t.t_task_team[1]));
5347   team->t.t_task_team[0] = NULL; // to be removed, as __kmp_allocate zeroes
5348   // memory, no need to duplicate
5349   team->t.t_task_team[1] = NULL; // to be removed, as __kmp_allocate zeroes
5350   // memory, no need to duplicate
5351 
5352   if (__kmp_storage_map) {
5353     __kmp_print_team_storage_map("team", team, team->t.t_id, new_nproc);
5354   }
5355 
5356   /* allocate space for arguments */
5357   __kmp_alloc_argv_entries(argc, team, FALSE);
5358   team->t.t_argc = argc;
5359 
5360   KA_TRACE(20,
5361            ("__kmp_allocate_team: team %d init arrived: join=%u, plain=%u\n",
5362             team->t.t_id, KMP_INIT_BARRIER_STATE, KMP_INIT_BARRIER_STATE));
5363   { // Initialize barrier data.
5364     int b;
5365     for (b = 0; b < bs_last_barrier; ++b) {
5366       team->t.t_bar[b].b_arrived = KMP_INIT_BARRIER_STATE;
5367 #if USE_DEBUGGER
5368       team->t.t_bar[b].b_master_arrived = 0;
5369       team->t.t_bar[b].b_team_arrived = 0;
5370 #endif
5371     }
5372   }
5373 
5374   team->t.t_proc_bind = new_proc_bind;
5375 
5376 #if OMPT_SUPPORT
5377   __ompt_team_assign_id(team, ompt_parallel_data);
5378   team->t.ompt_serialized_team_info = NULL;
5379 #endif
5380 
5381   KMP_MB();
5382 
5383   KA_TRACE(20, ("__kmp_allocate_team: done creating a new team %d.\n",
5384                 team->t.t_id));
5385 
5386   return team;
5387 }
5388 
5389 /* TODO implement hot-teams at all levels */
5390 /* TODO implement lazy thread release on demand (disband request) */
5391 
5392 /* free the team.  return it to the team pool.  release all the threads
5393  * associated with it */
5394 void __kmp_free_team(kmp_root_t *root,
5395                      kmp_team_t *team USE_NESTED_HOT_ARG(kmp_info_t *master)) {
5396   int f;
5397   KA_TRACE(20, ("__kmp_free_team: T#%d freeing team %d\n", __kmp_get_gtid(),
5398                 team->t.t_id));
5399 
5400   /* verify state */
5401   KMP_DEBUG_ASSERT(root);
5402   KMP_DEBUG_ASSERT(team);
5403   KMP_DEBUG_ASSERT(team->t.t_nproc <= team->t.t_max_nproc);
5404   KMP_DEBUG_ASSERT(team->t.t_threads);
5405 
5406   int use_hot_team = team == root->r.r_hot_team;
5407 #if KMP_NESTED_HOT_TEAMS
5408   int level;
5409   kmp_hot_team_ptr_t *hot_teams;
5410   if (master) {
5411     level = team->t.t_active_level - 1;
5412     if (master->th.th_teams_microtask) { // in teams construct?
5413       if (master->th.th_teams_size.nteams > 1) {
5414         ++level; // level was not increased in teams construct for
5415         // team_of_masters
5416       }
5417       if (team->t.t_pkfn != (microtask_t)__kmp_teams_master &&
5418           master->th.th_teams_level == team->t.t_level) {
5419         ++level; // level was not increased in teams construct for
5420         // team_of_workers before the parallel
5421       } // team->t.t_level will be increased inside parallel
5422     }
5423     hot_teams = master->th.th_hot_teams;
5424     if (level < __kmp_hot_teams_max_level) {
5425       KMP_DEBUG_ASSERT(team == hot_teams[level].hot_team);
5426       use_hot_team = 1;
5427     }
5428   }
5429 #endif // KMP_NESTED_HOT_TEAMS
5430 
5431   /* team is done working */
5432   TCW_SYNC_PTR(team->t.t_pkfn,
5433                NULL); // Important for Debugging Support Library.
5434 #if KMP_OS_WINDOWS
5435   team->t.t_copyin_counter = 0; // init counter for possible reuse
5436 #endif
5437   // Do not reset pointer to parent team to NULL for hot teams.
5438 
5439   /* if we are non-hot team, release our threads */
5440   if (!use_hot_team) {
5441     if (__kmp_tasking_mode != tskm_immediate_exec) {
5442       // Wait for threads to reach reapable state
5443       for (f = 1; f < team->t.t_nproc; ++f) {
5444         KMP_DEBUG_ASSERT(team->t.t_threads[f]);
5445         kmp_info_t *th = team->t.t_threads[f];
5446         volatile kmp_uint32 *state = &th->th.th_reap_state;
5447         while (*state != KMP_SAFE_TO_REAP) {
5448 #if KMP_OS_WINDOWS
5449           // On Windows a thread can be killed at any time, check this
5450           DWORD ecode;
5451           if (!__kmp_is_thread_alive(th, &ecode)) {
5452             *state = KMP_SAFE_TO_REAP; // reset the flag for dead thread
5453             break;
5454           }
5455 #endif
5456           // first check if thread is sleeping
5457           kmp_flag_64 fl(&th->th.th_bar[bs_forkjoin_barrier].bb.b_go, th);
5458           if (fl.is_sleeping())
5459             fl.resume(__kmp_gtid_from_thread(th));
5460           KMP_CPU_PAUSE();
5461         }
5462       }
5463 
5464       // Delete task teams
5465       int tt_idx;
5466       for (tt_idx = 0; tt_idx < 2; ++tt_idx) {
5467         kmp_task_team_t *task_team = team->t.t_task_team[tt_idx];
5468         if (task_team != NULL) {
5469           for (f = 0; f < team->t.t_nproc; ++f) { // threads unref task teams
5470             KMP_DEBUG_ASSERT(team->t.t_threads[f]);
5471             team->t.t_threads[f]->th.th_task_team = NULL;
5472           }
5473           KA_TRACE(
5474               20,
5475               ("__kmp_free_team: T#%d deactivating task_team %p on team %d\n",
5476                __kmp_get_gtid(), task_team, team->t.t_id));
5477 #if KMP_NESTED_HOT_TEAMS
5478           __kmp_free_task_team(master, task_team);
5479 #endif
5480           team->t.t_task_team[tt_idx] = NULL;
5481         }
5482       }
5483     }
5484 
5485     // Reset pointer to parent team only for non-hot teams.
5486     team->t.t_parent = NULL;
5487     team->t.t_level = 0;
5488     team->t.t_active_level = 0;
5489 
5490     /* free the worker threads */
5491     for (f = 1; f < team->t.t_nproc; ++f) {
5492       KMP_DEBUG_ASSERT(team->t.t_threads[f]);
5493       __kmp_free_thread(team->t.t_threads[f]);
5494       team->t.t_threads[f] = NULL;
5495     }
5496 
5497     /* put the team back in the team pool */
5498     /* TODO limit size of team pool, call reap_team if pool too large */
5499     team->t.t_next_pool = CCAST(kmp_team_t *, __kmp_team_pool);
5500     __kmp_team_pool = (volatile kmp_team_t *)team;
5501   } else { // Check if team was created for the masters in a teams construct
5502     // See if first worker is a CG root
5503     KMP_DEBUG_ASSERT(team->t.t_threads[1] &&
5504                      team->t.t_threads[1]->th.th_cg_roots);
5505     if (team->t.t_threads[1]->th.th_cg_roots->cg_root == team->t.t_threads[1]) {
5506       // Clean up the CG root nodes on workers so that this team can be re-used
5507       for (f = 1; f < team->t.t_nproc; ++f) {
5508         kmp_info_t *thr = team->t.t_threads[f];
5509         KMP_DEBUG_ASSERT(thr && thr->th.th_cg_roots &&
5510                          thr->th.th_cg_roots->cg_root == thr);
5511         // Pop current CG root off list
5512         kmp_cg_root_t *tmp = thr->th.th_cg_roots;
5513         thr->th.th_cg_roots = tmp->up;
5514         KA_TRACE(100, ("__kmp_free_team: Thread %p popping node %p and moving"
5515                        " up to node %p. cg_nthreads was %d\n",
5516                        thr, tmp, thr->th.th_cg_roots, tmp->cg_nthreads));
5517         int i = tmp->cg_nthreads--;
5518         if (i == 1) {
5519           __kmp_free(tmp); // free CG if we are the last thread in it
5520         }
5521         // Restore current task's thread_limit from CG root
5522         if (thr->th.th_cg_roots)
5523           thr->th.th_current_task->td_icvs.thread_limit =
5524               thr->th.th_cg_roots->cg_thread_limit;
5525       }
5526     }
5527   }
5528 
5529   KMP_MB();
5530 }
5531 
5532 /* reap the team.  destroy it, reclaim all its resources and free its memory */
5533 kmp_team_t *__kmp_reap_team(kmp_team_t *team) {
5534   kmp_team_t *next_pool = team->t.t_next_pool;
5535 
5536   KMP_DEBUG_ASSERT(team);
5537   KMP_DEBUG_ASSERT(team->t.t_dispatch);
5538   KMP_DEBUG_ASSERT(team->t.t_disp_buffer);
5539   KMP_DEBUG_ASSERT(team->t.t_threads);
5540   KMP_DEBUG_ASSERT(team->t.t_argv);
5541 
5542   /* TODO clean the threads that are a part of this? */
5543 
5544   /* free stuff */
5545   __kmp_free_team_arrays(team);
5546   if (team->t.t_argv != &team->t.t_inline_argv[0])
5547     __kmp_free((void *)team->t.t_argv);
5548   __kmp_free(team);
5549 
5550   KMP_MB();
5551   return next_pool;
5552 }
5553 
5554 // Free the thread.  Don't reap it, just place it on the pool of available
5555 // threads.
5556 //
5557 // Changes for Quad issue 527845: We need a predictable OMP tid <-> gtid
5558 // binding for the affinity mechanism to be useful.
5559 //
5560 // Now, we always keep the free list (__kmp_thread_pool) sorted by gtid.
5561 // However, we want to avoid a potential performance problem by always
5562 // scanning through the list to find the correct point at which to insert
5563 // the thread (potential N**2 behavior).  To do this we keep track of the
5564 // last place a thread struct was inserted (__kmp_thread_pool_insert_pt).
5565 // With single-level parallelism, threads will always be added to the tail
5566 // of the list, kept track of by __kmp_thread_pool_insert_pt.  With nested
5567 // parallelism, all bets are off and we may need to scan through the entire
5568 // free list.
5569 //
5570 // This change also has a potentially large performance benefit, for some
5571 // applications.  Previously, as threads were freed from the hot team, they
5572 // would be placed back on the free list in inverse order.  If the hot team
5573 // grew back to it's original size, then the freed thread would be placed
5574 // back on the hot team in reverse order.  This could cause bad cache
5575 // locality problems on programs where the size of the hot team regularly
5576 // grew and shrunk.
5577 //
5578 // Now, for single-level parallelism, the OMP tid is always == gtid.
5579 void __kmp_free_thread(kmp_info_t *this_th) {
5580   int gtid;
5581   kmp_info_t **scan;
5582 
5583   KA_TRACE(20, ("__kmp_free_thread: T#%d putting T#%d back on free pool.\n",
5584                 __kmp_get_gtid(), this_th->th.th_info.ds.ds_gtid));
5585 
5586   KMP_DEBUG_ASSERT(this_th);
5587 
5588   // When moving thread to pool, switch thread to wait on own b_go flag, and
5589   // uninitialized (NULL team).
5590   int b;
5591   kmp_balign_t *balign = this_th->th.th_bar;
5592   for (b = 0; b < bs_last_barrier; ++b) {
5593     if (balign[b].bb.wait_flag == KMP_BARRIER_PARENT_FLAG)
5594       balign[b].bb.wait_flag = KMP_BARRIER_SWITCH_TO_OWN_FLAG;
5595     balign[b].bb.team = NULL;
5596     balign[b].bb.leaf_kids = 0;
5597   }
5598   this_th->th.th_task_state = 0;
5599   this_th->th.th_reap_state = KMP_SAFE_TO_REAP;
5600 
5601   /* put thread back on the free pool */
5602   TCW_PTR(this_th->th.th_team, NULL);
5603   TCW_PTR(this_th->th.th_root, NULL);
5604   TCW_PTR(this_th->th.th_dispatch, NULL); /* NOT NEEDED */
5605 
5606   while (this_th->th.th_cg_roots) {
5607     this_th->th.th_cg_roots->cg_nthreads--;
5608     KA_TRACE(100, ("__kmp_free_thread: Thread %p decrement cg_nthreads on node"
5609                    " %p of thread  %p to %d\n",
5610                    this_th, this_th->th.th_cg_roots,
5611                    this_th->th.th_cg_roots->cg_root,
5612                    this_th->th.th_cg_roots->cg_nthreads));
5613     kmp_cg_root_t *tmp = this_th->th.th_cg_roots;
5614     if (tmp->cg_root == this_th) { // Thread is a cg_root
5615       KMP_DEBUG_ASSERT(tmp->cg_nthreads == 0);
5616       KA_TRACE(
5617           5, ("__kmp_free_thread: Thread %p freeing node %p\n", this_th, tmp));
5618       this_th->th.th_cg_roots = tmp->up;
5619       __kmp_free(tmp);
5620     } else { // Worker thread
5621       if (tmp->cg_nthreads == 0) { // last thread leaves contention group
5622         __kmp_free(tmp);
5623       }
5624       this_th->th.th_cg_roots = NULL;
5625       break;
5626     }
5627   }
5628 
5629   /* If the implicit task assigned to this thread can be used by other threads
5630    * -> multiple threads can share the data and try to free the task at
5631    * __kmp_reap_thread at exit. This duplicate use of the task data can happen
5632    * with higher probability when hot team is disabled but can occurs even when
5633    * the hot team is enabled */
5634   __kmp_free_implicit_task(this_th);
5635   this_th->th.th_current_task = NULL;
5636 
5637   // If the __kmp_thread_pool_insert_pt is already past the new insert
5638   // point, then we need to re-scan the entire list.
5639   gtid = this_th->th.th_info.ds.ds_gtid;
5640   if (__kmp_thread_pool_insert_pt != NULL) {
5641     KMP_DEBUG_ASSERT(__kmp_thread_pool != NULL);
5642     if (__kmp_thread_pool_insert_pt->th.th_info.ds.ds_gtid > gtid) {
5643       __kmp_thread_pool_insert_pt = NULL;
5644     }
5645   }
5646 
5647   // Scan down the list to find the place to insert the thread.
5648   // scan is the address of a link in the list, possibly the address of
5649   // __kmp_thread_pool itself.
5650   //
5651   // In the absence of nested parallelism, the for loop will have 0 iterations.
5652   if (__kmp_thread_pool_insert_pt != NULL) {
5653     scan = &(__kmp_thread_pool_insert_pt->th.th_next_pool);
5654   } else {
5655     scan = CCAST(kmp_info_t **, &__kmp_thread_pool);
5656   }
5657   for (; (*scan != NULL) && ((*scan)->th.th_info.ds.ds_gtid < gtid);
5658        scan = &((*scan)->th.th_next_pool))
5659     ;
5660 
5661   // Insert the new element on the list, and set __kmp_thread_pool_insert_pt
5662   // to its address.
5663   TCW_PTR(this_th->th.th_next_pool, *scan);
5664   __kmp_thread_pool_insert_pt = *scan = this_th;
5665   KMP_DEBUG_ASSERT((this_th->th.th_next_pool == NULL) ||
5666                    (this_th->th.th_info.ds.ds_gtid <
5667                     this_th->th.th_next_pool->th.th_info.ds.ds_gtid));
5668   TCW_4(this_th->th.th_in_pool, TRUE);
5669   __kmp_suspend_initialize_thread(this_th);
5670   __kmp_lock_suspend_mx(this_th);
5671   if (this_th->th.th_active == TRUE) {
5672     KMP_ATOMIC_INC(&__kmp_thread_pool_active_nth);
5673     this_th->th.th_active_in_pool = TRUE;
5674   }
5675 #if KMP_DEBUG
5676   else {
5677     KMP_DEBUG_ASSERT(this_th->th.th_active_in_pool == FALSE);
5678   }
5679 #endif
5680   __kmp_unlock_suspend_mx(this_th);
5681 
5682   TCW_4(__kmp_nth, __kmp_nth - 1);
5683 
5684 #ifdef KMP_ADJUST_BLOCKTIME
5685   /* Adjust blocktime back to user setting or default if necessary */
5686   /* Middle initialization might never have occurred                */
5687   if (!__kmp_env_blocktime && (__kmp_avail_proc > 0)) {
5688     KMP_DEBUG_ASSERT(__kmp_avail_proc > 0);
5689     if (__kmp_nth <= __kmp_avail_proc) {
5690       __kmp_zero_bt = FALSE;
5691     }
5692   }
5693 #endif /* KMP_ADJUST_BLOCKTIME */
5694 
5695   KMP_MB();
5696 }
5697 
5698 /* ------------------------------------------------------------------------ */
5699 
5700 void *__kmp_launch_thread(kmp_info_t *this_thr) {
5701   int gtid = this_thr->th.th_info.ds.ds_gtid;
5702   /*    void                 *stack_data;*/
5703   kmp_team_t **volatile pteam;
5704 
5705   KMP_MB();
5706   KA_TRACE(10, ("__kmp_launch_thread: T#%d start\n", gtid));
5707 
5708   if (__kmp_env_consistency_check) {
5709     this_thr->th.th_cons = __kmp_allocate_cons_stack(gtid); // ATT: Memory leak?
5710   }
5711 
5712 #if OMPT_SUPPORT
5713   ompt_data_t *thread_data;
5714   if (ompt_enabled.enabled) {
5715     thread_data = &(this_thr->th.ompt_thread_info.thread_data);
5716     *thread_data = ompt_data_none;
5717 
5718     this_thr->th.ompt_thread_info.state = ompt_state_overhead;
5719     this_thr->th.ompt_thread_info.wait_id = 0;
5720     this_thr->th.ompt_thread_info.idle_frame = OMPT_GET_FRAME_ADDRESS(0);
5721     this_thr->th.ompt_thread_info.parallel_flags = 0;
5722     if (ompt_enabled.ompt_callback_thread_begin) {
5723       ompt_callbacks.ompt_callback(ompt_callback_thread_begin)(
5724           ompt_thread_worker, thread_data);
5725     }
5726     this_thr->th.ompt_thread_info.state = ompt_state_idle;
5727   }
5728 #endif
5729 
5730   /* This is the place where threads wait for work */
5731   while (!TCR_4(__kmp_global.g.g_done)) {
5732     KMP_DEBUG_ASSERT(this_thr == __kmp_threads[gtid]);
5733     KMP_MB();
5734 
5735     /* wait for work to do */
5736     KA_TRACE(20, ("__kmp_launch_thread: T#%d waiting for work\n", gtid));
5737 
5738     /* No tid yet since not part of a team */
5739     __kmp_fork_barrier(gtid, KMP_GTID_DNE);
5740 
5741 #if OMPT_SUPPORT
5742     if (ompt_enabled.enabled) {
5743       this_thr->th.ompt_thread_info.state = ompt_state_overhead;
5744     }
5745 #endif
5746 
5747     pteam = &this_thr->th.th_team;
5748 
5749     /* have we been allocated? */
5750     if (TCR_SYNC_PTR(*pteam) && !TCR_4(__kmp_global.g.g_done)) {
5751       /* we were just woken up, so run our new task */
5752       if (TCR_SYNC_PTR((*pteam)->t.t_pkfn) != NULL) {
5753         int rc;
5754         KA_TRACE(20,
5755                  ("__kmp_launch_thread: T#%d(%d:%d) invoke microtask = %p\n",
5756                   gtid, (*pteam)->t.t_id, __kmp_tid_from_gtid(gtid),
5757                   (*pteam)->t.t_pkfn));
5758 
5759         updateHWFPControl(*pteam);
5760 
5761 #if OMPT_SUPPORT
5762         if (ompt_enabled.enabled) {
5763           this_thr->th.ompt_thread_info.state = ompt_state_work_parallel;
5764         }
5765 #endif
5766 
5767         rc = (*pteam)->t.t_invoke(gtid);
5768         KMP_ASSERT(rc);
5769 
5770         KMP_MB();
5771         KA_TRACE(20, ("__kmp_launch_thread: T#%d(%d:%d) done microtask = %p\n",
5772                       gtid, (*pteam)->t.t_id, __kmp_tid_from_gtid(gtid),
5773                       (*pteam)->t.t_pkfn));
5774       }
5775 #if OMPT_SUPPORT
5776       if (ompt_enabled.enabled) {
5777         /* no frame set while outside task */
5778         __ompt_get_task_info_object(0)->frame.exit_frame = ompt_data_none;
5779 
5780         this_thr->th.ompt_thread_info.state = ompt_state_overhead;
5781       }
5782 #endif
5783       /* join barrier after parallel region */
5784       __kmp_join_barrier(gtid);
5785     }
5786   }
5787   TCR_SYNC_PTR((intptr_t)__kmp_global.g.g_done);
5788 
5789 #if OMPT_SUPPORT
5790   if (ompt_enabled.ompt_callback_thread_end) {
5791     ompt_callbacks.ompt_callback(ompt_callback_thread_end)(thread_data);
5792   }
5793 #endif
5794 
5795   this_thr->th.th_task_team = NULL;
5796   /* run the destructors for the threadprivate data for this thread */
5797   __kmp_common_destroy_gtid(gtid);
5798 
5799   KA_TRACE(10, ("__kmp_launch_thread: T#%d done\n", gtid));
5800   KMP_MB();
5801   return this_thr;
5802 }
5803 
5804 /* ------------------------------------------------------------------------ */
5805 
5806 void __kmp_internal_end_dest(void *specific_gtid) {
5807 #if KMP_COMPILER_ICC
5808 #pragma warning(push)
5809 #pragma warning(disable : 810) // conversion from "void *" to "int" may lose
5810 // significant bits
5811 #endif
5812   // Make sure no significant bits are lost
5813   int gtid = (kmp_intptr_t)specific_gtid - 1;
5814 #if KMP_COMPILER_ICC
5815 #pragma warning(pop)
5816 #endif
5817 
5818   KA_TRACE(30, ("__kmp_internal_end_dest: T#%d\n", gtid));
5819   /* NOTE: the gtid is stored as gitd+1 in the thread-local-storage
5820    * this is because 0 is reserved for the nothing-stored case */
5821 
5822   /* josh: One reason for setting the gtid specific data even when it is being
5823      destroyed by pthread is to allow gtid lookup through thread specific data
5824      (__kmp_gtid_get_specific).  Some of the code, especially stat code,
5825      that gets executed in the call to __kmp_internal_end_thread, actually
5826      gets the gtid through the thread specific data.  Setting it here seems
5827      rather inelegant and perhaps wrong, but allows __kmp_internal_end_thread
5828      to run smoothly.
5829      todo: get rid of this after we remove the dependence on
5830      __kmp_gtid_get_specific  */
5831   if (gtid >= 0 && KMP_UBER_GTID(gtid))
5832     __kmp_gtid_set_specific(gtid);
5833 #ifdef KMP_TDATA_GTID
5834   __kmp_gtid = gtid;
5835 #endif
5836   __kmp_internal_end_thread(gtid);
5837 }
5838 
5839 #if KMP_OS_UNIX && KMP_DYNAMIC_LIB
5840 
5841 __attribute__((destructor)) void __kmp_internal_end_dtor(void) {
5842   __kmp_internal_end_atexit();
5843 }
5844 
5845 #endif
5846 
5847 /* [Windows] josh: when the atexit handler is called, there may still be more
5848    than one thread alive */
5849 void __kmp_internal_end_atexit(void) {
5850   KA_TRACE(30, ("__kmp_internal_end_atexit\n"));
5851   /* [Windows]
5852      josh: ideally, we want to completely shutdown the library in this atexit
5853      handler, but stat code that depends on thread specific data for gtid fails
5854      because that data becomes unavailable at some point during the shutdown, so
5855      we call __kmp_internal_end_thread instead. We should eventually remove the
5856      dependency on __kmp_get_specific_gtid in the stat code and use
5857      __kmp_internal_end_library to cleanly shutdown the library.
5858 
5859      // TODO: Can some of this comment about GVS be removed?
5860      I suspect that the offending stat code is executed when the calling thread
5861      tries to clean up a dead root thread's data structures, resulting in GVS
5862      code trying to close the GVS structures for that thread, but since the stat
5863      code uses __kmp_get_specific_gtid to get the gtid with the assumption that
5864      the calling thread is cleaning up itself instead of another thread, it get
5865      confused. This happens because allowing a thread to unregister and cleanup
5866      another thread is a recent modification for addressing an issue.
5867      Based on the current design (20050722), a thread may end up
5868      trying to unregister another thread only if thread death does not trigger
5869      the calling of __kmp_internal_end_thread.  For Linux* OS, there is the
5870      thread specific data destructor function to detect thread death. For
5871      Windows dynamic, there is DllMain(THREAD_DETACH). For Windows static, there
5872      is nothing.  Thus, the workaround is applicable only for Windows static
5873      stat library. */
5874   __kmp_internal_end_library(-1);
5875 #if KMP_OS_WINDOWS
5876   __kmp_close_console();
5877 #endif
5878 }
5879 
5880 static void __kmp_reap_thread(kmp_info_t *thread, int is_root) {
5881   // It is assumed __kmp_forkjoin_lock is acquired.
5882 
5883   int gtid;
5884 
5885   KMP_DEBUG_ASSERT(thread != NULL);
5886 
5887   gtid = thread->th.th_info.ds.ds_gtid;
5888 
5889   if (!is_root) {
5890     if (__kmp_dflt_blocktime != KMP_MAX_BLOCKTIME) {
5891       /* Assume the threads are at the fork barrier here */
5892       KA_TRACE(
5893           20, ("__kmp_reap_thread: releasing T#%d from fork barrier for reap\n",
5894                gtid));
5895       /* Need release fence here to prevent seg faults for tree forkjoin barrier
5896        * (GEH) */
5897       ANNOTATE_HAPPENS_BEFORE(thread);
5898       kmp_flag_64 flag(&thread->th.th_bar[bs_forkjoin_barrier].bb.b_go, thread);
5899       __kmp_release_64(&flag);
5900     }
5901 
5902     // Terminate OS thread.
5903     __kmp_reap_worker(thread);
5904 
5905     // The thread was killed asynchronously.  If it was actively
5906     // spinning in the thread pool, decrement the global count.
5907     //
5908     // There is a small timing hole here - if the worker thread was just waking
5909     // up after sleeping in the pool, had reset it's th_active_in_pool flag but
5910     // not decremented the global counter __kmp_thread_pool_active_nth yet, then
5911     // the global counter might not get updated.
5912     //
5913     // Currently, this can only happen as the library is unloaded,
5914     // so there are no harmful side effects.
5915     if (thread->th.th_active_in_pool) {
5916       thread->th.th_active_in_pool = FALSE;
5917       KMP_ATOMIC_DEC(&__kmp_thread_pool_active_nth);
5918       KMP_DEBUG_ASSERT(__kmp_thread_pool_active_nth >= 0);
5919     }
5920   }
5921 
5922   __kmp_free_implicit_task(thread);
5923 
5924 // Free the fast memory for tasking
5925 #if USE_FAST_MEMORY
5926   __kmp_free_fast_memory(thread);
5927 #endif /* USE_FAST_MEMORY */
5928 
5929   __kmp_suspend_uninitialize_thread(thread);
5930 
5931   KMP_DEBUG_ASSERT(__kmp_threads[gtid] == thread);
5932   TCW_SYNC_PTR(__kmp_threads[gtid], NULL);
5933 
5934   --__kmp_all_nth;
5935 // __kmp_nth was decremented when thread is added to the pool.
5936 
5937 #ifdef KMP_ADJUST_BLOCKTIME
5938   /* Adjust blocktime back to user setting or default if necessary */
5939   /* Middle initialization might never have occurred                */
5940   if (!__kmp_env_blocktime && (__kmp_avail_proc > 0)) {
5941     KMP_DEBUG_ASSERT(__kmp_avail_proc > 0);
5942     if (__kmp_nth <= __kmp_avail_proc) {
5943       __kmp_zero_bt = FALSE;
5944     }
5945   }
5946 #endif /* KMP_ADJUST_BLOCKTIME */
5947 
5948   /* free the memory being used */
5949   if (__kmp_env_consistency_check) {
5950     if (thread->th.th_cons) {
5951       __kmp_free_cons_stack(thread->th.th_cons);
5952       thread->th.th_cons = NULL;
5953     }
5954   }
5955 
5956   if (thread->th.th_pri_common != NULL) {
5957     __kmp_free(thread->th.th_pri_common);
5958     thread->th.th_pri_common = NULL;
5959   }
5960 
5961   if (thread->th.th_task_state_memo_stack != NULL) {
5962     __kmp_free(thread->th.th_task_state_memo_stack);
5963     thread->th.th_task_state_memo_stack = NULL;
5964   }
5965 
5966 #if KMP_USE_BGET
5967   if (thread->th.th_local.bget_data != NULL) {
5968     __kmp_finalize_bget(thread);
5969   }
5970 #endif
5971 
5972 #if KMP_AFFINITY_SUPPORTED
5973   if (thread->th.th_affin_mask != NULL) {
5974     KMP_CPU_FREE(thread->th.th_affin_mask);
5975     thread->th.th_affin_mask = NULL;
5976   }
5977 #endif /* KMP_AFFINITY_SUPPORTED */
5978 
5979 #if KMP_USE_HIER_SCHED
5980   if (thread->th.th_hier_bar_data != NULL) {
5981     __kmp_free(thread->th.th_hier_bar_data);
5982     thread->th.th_hier_bar_data = NULL;
5983   }
5984 #endif
5985 
5986   __kmp_reap_team(thread->th.th_serial_team);
5987   thread->th.th_serial_team = NULL;
5988   __kmp_free(thread);
5989 
5990   KMP_MB();
5991 
5992 } // __kmp_reap_thread
5993 
5994 static void __kmp_internal_end(void) {
5995   int i;
5996 
5997   /* First, unregister the library */
5998   __kmp_unregister_library();
5999 
6000 #if KMP_OS_WINDOWS
6001   /* In Win static library, we can't tell when a root actually dies, so we
6002      reclaim the data structures for any root threads that have died but not
6003      unregistered themselves, in order to shut down cleanly.
6004      In Win dynamic library we also can't tell when a thread dies.  */
6005   __kmp_reclaim_dead_roots(); // AC: moved here to always clean resources of
6006 // dead roots
6007 #endif
6008 
6009   for (i = 0; i < __kmp_threads_capacity; i++)
6010     if (__kmp_root[i])
6011       if (__kmp_root[i]->r.r_active)
6012         break;
6013   KMP_MB(); /* Flush all pending memory write invalidates.  */
6014   TCW_SYNC_4(__kmp_global.g.g_done, TRUE);
6015 
6016   if (i < __kmp_threads_capacity) {
6017 #if KMP_USE_MONITOR
6018     // 2009-09-08 (lev): Other alive roots found. Why do we kill the monitor??
6019     KMP_MB(); /* Flush all pending memory write invalidates.  */
6020 
6021     // Need to check that monitor was initialized before reaping it. If we are
6022     // called form __kmp_atfork_child (which sets __kmp_init_parallel = 0), then
6023     // __kmp_monitor will appear to contain valid data, but it is only valid in
6024     // the parent process, not the child.
6025     // New behavior (201008): instead of keying off of the flag
6026     // __kmp_init_parallel, the monitor thread creation is keyed off
6027     // of the new flag __kmp_init_monitor.
6028     __kmp_acquire_bootstrap_lock(&__kmp_monitor_lock);
6029     if (TCR_4(__kmp_init_monitor)) {
6030       __kmp_reap_monitor(&__kmp_monitor);
6031       TCW_4(__kmp_init_monitor, 0);
6032     }
6033     __kmp_release_bootstrap_lock(&__kmp_monitor_lock);
6034     KA_TRACE(10, ("__kmp_internal_end: monitor reaped\n"));
6035 #endif // KMP_USE_MONITOR
6036   } else {
6037 /* TODO move this to cleanup code */
6038 #ifdef KMP_DEBUG
6039     /* make sure that everything has properly ended */
6040     for (i = 0; i < __kmp_threads_capacity; i++) {
6041       if (__kmp_root[i]) {
6042         //                    KMP_ASSERT( ! KMP_UBER_GTID( i ) );         // AC:
6043         //                    there can be uber threads alive here
6044         KMP_ASSERT(!__kmp_root[i]->r.r_active); // TODO: can they be active?
6045       }
6046     }
6047 #endif
6048 
6049     KMP_MB();
6050 
6051     // Reap the worker threads.
6052     // This is valid for now, but be careful if threads are reaped sooner.
6053     while (__kmp_thread_pool != NULL) { // Loop thru all the thread in the pool.
6054       // Get the next thread from the pool.
6055       kmp_info_t *thread = CCAST(kmp_info_t *, __kmp_thread_pool);
6056       __kmp_thread_pool = thread->th.th_next_pool;
6057       // Reap it.
6058       KMP_DEBUG_ASSERT(thread->th.th_reap_state == KMP_SAFE_TO_REAP);
6059       thread->th.th_next_pool = NULL;
6060       thread->th.th_in_pool = FALSE;
6061       __kmp_reap_thread(thread, 0);
6062     }
6063     __kmp_thread_pool_insert_pt = NULL;
6064 
6065     // Reap teams.
6066     while (__kmp_team_pool != NULL) { // Loop thru all the teams in the pool.
6067       // Get the next team from the pool.
6068       kmp_team_t *team = CCAST(kmp_team_t *, __kmp_team_pool);
6069       __kmp_team_pool = team->t.t_next_pool;
6070       // Reap it.
6071       team->t.t_next_pool = NULL;
6072       __kmp_reap_team(team);
6073     }
6074 
6075     __kmp_reap_task_teams();
6076 
6077 #if KMP_OS_UNIX
6078     // Threads that are not reaped should not access any resources since they
6079     // are going to be deallocated soon, so the shutdown sequence should wait
6080     // until all threads either exit the final spin-waiting loop or begin
6081     // sleeping after the given blocktime.
6082     for (i = 0; i < __kmp_threads_capacity; i++) {
6083       kmp_info_t *thr = __kmp_threads[i];
6084       while (thr && KMP_ATOMIC_LD_ACQ(&thr->th.th_blocking))
6085         KMP_CPU_PAUSE();
6086     }
6087 #endif
6088 
6089     for (i = 0; i < __kmp_threads_capacity; ++i) {
6090       // TBD: Add some checking...
6091       // Something like KMP_DEBUG_ASSERT( __kmp_thread[ i ] == NULL );
6092     }
6093 
6094     /* Make sure all threadprivate destructors get run by joining with all
6095        worker threads before resetting this flag */
6096     TCW_SYNC_4(__kmp_init_common, FALSE);
6097 
6098     KA_TRACE(10, ("__kmp_internal_end: all workers reaped\n"));
6099     KMP_MB();
6100 
6101 #if KMP_USE_MONITOR
6102     // See note above: One of the possible fixes for CQ138434 / CQ140126
6103     //
6104     // FIXME: push both code fragments down and CSE them?
6105     // push them into __kmp_cleanup() ?
6106     __kmp_acquire_bootstrap_lock(&__kmp_monitor_lock);
6107     if (TCR_4(__kmp_init_monitor)) {
6108       __kmp_reap_monitor(&__kmp_monitor);
6109       TCW_4(__kmp_init_monitor, 0);
6110     }
6111     __kmp_release_bootstrap_lock(&__kmp_monitor_lock);
6112     KA_TRACE(10, ("__kmp_internal_end: monitor reaped\n"));
6113 #endif
6114   } /* else !__kmp_global.t_active */
6115   TCW_4(__kmp_init_gtid, FALSE);
6116   KMP_MB(); /* Flush all pending memory write invalidates.  */
6117 
6118   __kmp_cleanup();
6119 #if OMPT_SUPPORT
6120   ompt_fini();
6121 #endif
6122 }
6123 
6124 void __kmp_internal_end_library(int gtid_req) {
6125   /* if we have already cleaned up, don't try again, it wouldn't be pretty */
6126   /* this shouldn't be a race condition because __kmp_internal_end() is the
6127      only place to clear __kmp_serial_init */
6128   /* we'll check this later too, after we get the lock */
6129   // 2009-09-06: We do not set g_abort without setting g_done. This check looks
6130   // redundant, because the next check will work in any case.
6131   if (__kmp_global.g.g_abort) {
6132     KA_TRACE(11, ("__kmp_internal_end_library: abort, exiting\n"));
6133     /* TODO abort? */
6134     return;
6135   }
6136   if (TCR_4(__kmp_global.g.g_done) || !__kmp_init_serial) {
6137     KA_TRACE(10, ("__kmp_internal_end_library: already finished\n"));
6138     return;
6139   }
6140 
6141   KMP_MB(); /* Flush all pending memory write invalidates.  */
6142 
6143   /* find out who we are and what we should do */
6144   {
6145     int gtid = (gtid_req >= 0) ? gtid_req : __kmp_gtid_get_specific();
6146     KA_TRACE(
6147         10, ("__kmp_internal_end_library: enter T#%d  (%d)\n", gtid, gtid_req));
6148     if (gtid == KMP_GTID_SHUTDOWN) {
6149       KA_TRACE(10, ("__kmp_internal_end_library: !__kmp_init_runtime, system "
6150                     "already shutdown\n"));
6151       return;
6152     } else if (gtid == KMP_GTID_MONITOR) {
6153       KA_TRACE(10, ("__kmp_internal_end_library: monitor thread, gtid not "
6154                     "registered, or system shutdown\n"));
6155       return;
6156     } else if (gtid == KMP_GTID_DNE) {
6157       KA_TRACE(10, ("__kmp_internal_end_library: gtid not registered or system "
6158                     "shutdown\n"));
6159       /* we don't know who we are, but we may still shutdown the library */
6160     } else if (KMP_UBER_GTID(gtid)) {
6161       /* unregister ourselves as an uber thread.  gtid is no longer valid */
6162       if (__kmp_root[gtid]->r.r_active) {
6163         __kmp_global.g.g_abort = -1;
6164         TCW_SYNC_4(__kmp_global.g.g_done, TRUE);
6165         KA_TRACE(10,
6166                  ("__kmp_internal_end_library: root still active, abort T#%d\n",
6167                   gtid));
6168         return;
6169       } else {
6170         KA_TRACE(
6171             10,
6172             ("__kmp_internal_end_library: unregistering sibling T#%d\n", gtid));
6173         __kmp_unregister_root_current_thread(gtid);
6174       }
6175     } else {
6176 /* worker threads may call this function through the atexit handler, if they
6177  * call exit() */
6178 /* For now, skip the usual subsequent processing and just dump the debug buffer.
6179    TODO: do a thorough shutdown instead */
6180 #ifdef DUMP_DEBUG_ON_EXIT
6181       if (__kmp_debug_buf)
6182         __kmp_dump_debug_buffer();
6183 #endif
6184       return;
6185     }
6186   }
6187   /* synchronize the termination process */
6188   __kmp_acquire_bootstrap_lock(&__kmp_initz_lock);
6189 
6190   /* have we already finished */
6191   if (__kmp_global.g.g_abort) {
6192     KA_TRACE(10, ("__kmp_internal_end_library: abort, exiting\n"));
6193     /* TODO abort? */
6194     __kmp_release_bootstrap_lock(&__kmp_initz_lock);
6195     return;
6196   }
6197   if (TCR_4(__kmp_global.g.g_done) || !__kmp_init_serial) {
6198     __kmp_release_bootstrap_lock(&__kmp_initz_lock);
6199     return;
6200   }
6201 
6202   /* We need this lock to enforce mutex between this reading of
6203      __kmp_threads_capacity and the writing by __kmp_register_root.
6204      Alternatively, we can use a counter of roots that is atomically updated by
6205      __kmp_get_global_thread_id_reg, __kmp_do_serial_initialize and
6206      __kmp_internal_end_*.  */
6207   __kmp_acquire_bootstrap_lock(&__kmp_forkjoin_lock);
6208 
6209   /* now we can safely conduct the actual termination */
6210   __kmp_internal_end();
6211 
6212   __kmp_release_bootstrap_lock(&__kmp_forkjoin_lock);
6213   __kmp_release_bootstrap_lock(&__kmp_initz_lock);
6214 
6215   KA_TRACE(10, ("__kmp_internal_end_library: exit\n"));
6216 
6217 #ifdef DUMP_DEBUG_ON_EXIT
6218   if (__kmp_debug_buf)
6219     __kmp_dump_debug_buffer();
6220 #endif
6221 
6222 #if KMP_OS_WINDOWS
6223   __kmp_close_console();
6224 #endif
6225 
6226   __kmp_fini_allocator();
6227 
6228 } // __kmp_internal_end_library
6229 
6230 void __kmp_internal_end_thread(int gtid_req) {
6231   int i;
6232 
6233   /* if we have already cleaned up, don't try again, it wouldn't be pretty */
6234   /* this shouldn't be a race condition because __kmp_internal_end() is the
6235    * only place to clear __kmp_serial_init */
6236   /* we'll check this later too, after we get the lock */
6237   // 2009-09-06: We do not set g_abort without setting g_done. This check looks
6238   // redundant, because the next check will work in any case.
6239   if (__kmp_global.g.g_abort) {
6240     KA_TRACE(11, ("__kmp_internal_end_thread: abort, exiting\n"));
6241     /* TODO abort? */
6242     return;
6243   }
6244   if (TCR_4(__kmp_global.g.g_done) || !__kmp_init_serial) {
6245     KA_TRACE(10, ("__kmp_internal_end_thread: already finished\n"));
6246     return;
6247   }
6248 
6249   KMP_MB(); /* Flush all pending memory write invalidates.  */
6250 
6251   /* find out who we are and what we should do */
6252   {
6253     int gtid = (gtid_req >= 0) ? gtid_req : __kmp_gtid_get_specific();
6254     KA_TRACE(10,
6255              ("__kmp_internal_end_thread: enter T#%d  (%d)\n", gtid, gtid_req));
6256     if (gtid == KMP_GTID_SHUTDOWN) {
6257       KA_TRACE(10, ("__kmp_internal_end_thread: !__kmp_init_runtime, system "
6258                     "already shutdown\n"));
6259       return;
6260     } else if (gtid == KMP_GTID_MONITOR) {
6261       KA_TRACE(10, ("__kmp_internal_end_thread: monitor thread, gtid not "
6262                     "registered, or system shutdown\n"));
6263       return;
6264     } else if (gtid == KMP_GTID_DNE) {
6265       KA_TRACE(10, ("__kmp_internal_end_thread: gtid not registered or system "
6266                     "shutdown\n"));
6267       return;
6268       /* we don't know who we are */
6269     } else if (KMP_UBER_GTID(gtid)) {
6270       /* unregister ourselves as an uber thread.  gtid is no longer valid */
6271       if (__kmp_root[gtid]->r.r_active) {
6272         __kmp_global.g.g_abort = -1;
6273         TCW_SYNC_4(__kmp_global.g.g_done, TRUE);
6274         KA_TRACE(10,
6275                  ("__kmp_internal_end_thread: root still active, abort T#%d\n",
6276                   gtid));
6277         return;
6278       } else {
6279         KA_TRACE(10, ("__kmp_internal_end_thread: unregistering sibling T#%d\n",
6280                       gtid));
6281         __kmp_unregister_root_current_thread(gtid);
6282       }
6283     } else {
6284       /* just a worker thread, let's leave */
6285       KA_TRACE(10, ("__kmp_internal_end_thread: worker thread T#%d\n", gtid));
6286 
6287       if (gtid >= 0) {
6288         __kmp_threads[gtid]->th.th_task_team = NULL;
6289       }
6290 
6291       KA_TRACE(10,
6292                ("__kmp_internal_end_thread: worker thread done, exiting T#%d\n",
6293                 gtid));
6294       return;
6295     }
6296   }
6297 #if KMP_DYNAMIC_LIB
6298   if (__kmp_pause_status != kmp_hard_paused)
6299   // AC: lets not shutdown the dynamic library at the exit of uber thread,
6300   // because we will better shutdown later in the library destructor.
6301   {
6302     KA_TRACE(10, ("__kmp_internal_end_thread: exiting T#%d\n", gtid_req));
6303     return;
6304   }
6305 #endif
6306   /* synchronize the termination process */
6307   __kmp_acquire_bootstrap_lock(&__kmp_initz_lock);
6308 
6309   /* have we already finished */
6310   if (__kmp_global.g.g_abort) {
6311     KA_TRACE(10, ("__kmp_internal_end_thread: abort, exiting\n"));
6312     /* TODO abort? */
6313     __kmp_release_bootstrap_lock(&__kmp_initz_lock);
6314     return;
6315   }
6316   if (TCR_4(__kmp_global.g.g_done) || !__kmp_init_serial) {
6317     __kmp_release_bootstrap_lock(&__kmp_initz_lock);
6318     return;
6319   }
6320 
6321   /* We need this lock to enforce mutex between this reading of
6322      __kmp_threads_capacity and the writing by __kmp_register_root.
6323      Alternatively, we can use a counter of roots that is atomically updated by
6324      __kmp_get_global_thread_id_reg, __kmp_do_serial_initialize and
6325      __kmp_internal_end_*.  */
6326 
6327   /* should we finish the run-time?  are all siblings done? */
6328   __kmp_acquire_bootstrap_lock(&__kmp_forkjoin_lock);
6329 
6330   for (i = 0; i < __kmp_threads_capacity; ++i) {
6331     if (KMP_UBER_GTID(i)) {
6332       KA_TRACE(
6333           10,
6334           ("__kmp_internal_end_thread: remaining sibling task: gtid==%d\n", i));
6335       __kmp_release_bootstrap_lock(&__kmp_forkjoin_lock);
6336       __kmp_release_bootstrap_lock(&__kmp_initz_lock);
6337       return;
6338     }
6339   }
6340 
6341   /* now we can safely conduct the actual termination */
6342 
6343   __kmp_internal_end();
6344 
6345   __kmp_release_bootstrap_lock(&__kmp_forkjoin_lock);
6346   __kmp_release_bootstrap_lock(&__kmp_initz_lock);
6347 
6348   KA_TRACE(10, ("__kmp_internal_end_thread: exit T#%d\n", gtid_req));
6349 
6350 #ifdef DUMP_DEBUG_ON_EXIT
6351   if (__kmp_debug_buf)
6352     __kmp_dump_debug_buffer();
6353 #endif
6354 } // __kmp_internal_end_thread
6355 
6356 // -----------------------------------------------------------------------------
6357 // Library registration stuff.
6358 
6359 static long __kmp_registration_flag = 0;
6360 // Random value used to indicate library initialization.
6361 static char *__kmp_registration_str = NULL;
6362 // Value to be saved in env var __KMP_REGISTERED_LIB_<pid>.
6363 
6364 static inline char *__kmp_reg_status_name() {
6365   /* On RHEL 3u5 if linked statically, getpid() returns different values in
6366      each thread. If registration and unregistration go in different threads
6367      (omp_misc_other_root_exit.cpp test case), the name of registered_lib_env
6368      env var can not be found, because the name will contain different pid. */
6369   return __kmp_str_format("__KMP_REGISTERED_LIB_%d", (int)getpid());
6370 } // __kmp_reg_status_get
6371 
6372 void __kmp_register_library_startup(void) {
6373 
6374   char *name = __kmp_reg_status_name(); // Name of the environment variable.
6375   int done = 0;
6376   union {
6377     double dtime;
6378     long ltime;
6379   } time;
6380 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
6381   __kmp_initialize_system_tick();
6382 #endif
6383   __kmp_read_system_time(&time.dtime);
6384   __kmp_registration_flag = 0xCAFE0000L | (time.ltime & 0x0000FFFFL);
6385   __kmp_registration_str =
6386       __kmp_str_format("%p-%lx-%s", &__kmp_registration_flag,
6387                        __kmp_registration_flag, KMP_LIBRARY_FILE);
6388 
6389   KA_TRACE(50, ("__kmp_register_library_startup: %s=\"%s\"\n", name,
6390                 __kmp_registration_str));
6391 
6392   while (!done) {
6393 
6394     char *value = NULL; // Actual value of the environment variable.
6395 
6396     // Set environment variable, but do not overwrite if it is exist.
6397     __kmp_env_set(name, __kmp_registration_str, 0);
6398     // Check the variable is written.
6399     value = __kmp_env_get(name);
6400     if (value != NULL && strcmp(value, __kmp_registration_str) == 0) {
6401 
6402       done = 1; // Ok, environment variable set successfully, exit the loop.
6403 
6404     } else {
6405 
6406       // Oops. Write failed. Another copy of OpenMP RTL is in memory.
6407       // Check whether it alive or dead.
6408       int neighbor = 0; // 0 -- unknown status, 1 -- alive, 2 -- dead.
6409       char *tail = value;
6410       char *flag_addr_str = NULL;
6411       char *flag_val_str = NULL;
6412       char const *file_name = NULL;
6413       __kmp_str_split(tail, '-', &flag_addr_str, &tail);
6414       __kmp_str_split(tail, '-', &flag_val_str, &tail);
6415       file_name = tail;
6416       if (tail != NULL) {
6417         long *flag_addr = 0;
6418         long flag_val = 0;
6419         KMP_SSCANF(flag_addr_str, "%p", RCAST(void**, &flag_addr));
6420         KMP_SSCANF(flag_val_str, "%lx", &flag_val);
6421         if (flag_addr != 0 && flag_val != 0 && strcmp(file_name, "") != 0) {
6422           // First, check whether environment-encoded address is mapped into
6423           // addr space.
6424           // If so, dereference it to see if it still has the right value.
6425           if (__kmp_is_address_mapped(flag_addr) && *flag_addr == flag_val) {
6426             neighbor = 1;
6427           } else {
6428             // If not, then we know the other copy of the library is no longer
6429             // running.
6430             neighbor = 2;
6431           }
6432         }
6433       }
6434       switch (neighbor) {
6435       case 0: // Cannot parse environment variable -- neighbor status unknown.
6436         // Assume it is the incompatible format of future version of the
6437         // library. Assume the other library is alive.
6438         // WARN( ... ); // TODO: Issue a warning.
6439         file_name = "unknown library";
6440         KMP_FALLTHROUGH();
6441       // Attention! Falling to the next case. That's intentional.
6442       case 1: { // Neighbor is alive.
6443         // Check it is allowed.
6444         char *duplicate_ok = __kmp_env_get("KMP_DUPLICATE_LIB_OK");
6445         if (!__kmp_str_match_true(duplicate_ok)) {
6446           // That's not allowed. Issue fatal error.
6447           __kmp_fatal(KMP_MSG(DuplicateLibrary, KMP_LIBRARY_FILE, file_name),
6448                       KMP_HNT(DuplicateLibrary), __kmp_msg_null);
6449         }
6450         KMP_INTERNAL_FREE(duplicate_ok);
6451         __kmp_duplicate_library_ok = 1;
6452         done = 1; // Exit the loop.
6453       } break;
6454       case 2: { // Neighbor is dead.
6455         // Clear the variable and try to register library again.
6456         __kmp_env_unset(name);
6457       } break;
6458       default: { KMP_DEBUG_ASSERT(0); } break;
6459       }
6460     }
6461     KMP_INTERNAL_FREE((void *)value);
6462   }
6463   KMP_INTERNAL_FREE((void *)name);
6464 
6465 } // func __kmp_register_library_startup
6466 
6467 void __kmp_unregister_library(void) {
6468 
6469   char *name = __kmp_reg_status_name();
6470   char *value = __kmp_env_get(name);
6471 
6472   KMP_DEBUG_ASSERT(__kmp_registration_flag != 0);
6473   KMP_DEBUG_ASSERT(__kmp_registration_str != NULL);
6474   if (value != NULL && strcmp(value, __kmp_registration_str) == 0) {
6475     // Ok, this is our variable. Delete it.
6476     __kmp_env_unset(name);
6477   }
6478 
6479   KMP_INTERNAL_FREE(__kmp_registration_str);
6480   KMP_INTERNAL_FREE(value);
6481   KMP_INTERNAL_FREE(name);
6482 
6483   __kmp_registration_flag = 0;
6484   __kmp_registration_str = NULL;
6485 
6486 } // __kmp_unregister_library
6487 
6488 // End of Library registration stuff.
6489 // -----------------------------------------------------------------------------
6490 
6491 #if KMP_MIC_SUPPORTED
6492 
6493 static void __kmp_check_mic_type() {
6494   kmp_cpuid_t cpuid_state = {0};
6495   kmp_cpuid_t *cs_p = &cpuid_state;
6496   __kmp_x86_cpuid(1, 0, cs_p);
6497   // We don't support mic1 at the moment
6498   if ((cs_p->eax & 0xff0) == 0xB10) {
6499     __kmp_mic_type = mic2;
6500   } else if ((cs_p->eax & 0xf0ff0) == 0x50670) {
6501     __kmp_mic_type = mic3;
6502   } else {
6503     __kmp_mic_type = non_mic;
6504   }
6505 }
6506 
6507 #endif /* KMP_MIC_SUPPORTED */
6508 
6509 static void __kmp_do_serial_initialize(void) {
6510   int i, gtid;
6511   int size;
6512 
6513   KA_TRACE(10, ("__kmp_do_serial_initialize: enter\n"));
6514 
6515   KMP_DEBUG_ASSERT(sizeof(kmp_int32) == 4);
6516   KMP_DEBUG_ASSERT(sizeof(kmp_uint32) == 4);
6517   KMP_DEBUG_ASSERT(sizeof(kmp_int64) == 8);
6518   KMP_DEBUG_ASSERT(sizeof(kmp_uint64) == 8);
6519   KMP_DEBUG_ASSERT(sizeof(kmp_intptr_t) == sizeof(void *));
6520 
6521 #if OMPT_SUPPORT
6522   ompt_pre_init();
6523 #endif
6524 
6525   __kmp_validate_locks();
6526 
6527   /* Initialize internal memory allocator */
6528   __kmp_init_allocator();
6529 
6530   /* Register the library startup via an environment variable and check to see
6531      whether another copy of the library is already registered. */
6532 
6533   __kmp_register_library_startup();
6534 
6535   /* TODO reinitialization of library */
6536   if (TCR_4(__kmp_global.g.g_done)) {
6537     KA_TRACE(10, ("__kmp_do_serial_initialize: reinitialization of library\n"));
6538   }
6539 
6540   __kmp_global.g.g_abort = 0;
6541   TCW_SYNC_4(__kmp_global.g.g_done, FALSE);
6542 
6543 /* initialize the locks */
6544 #if KMP_USE_ADAPTIVE_LOCKS
6545 #if KMP_DEBUG_ADAPTIVE_LOCKS
6546   __kmp_init_speculative_stats();
6547 #endif
6548 #endif
6549 #if KMP_STATS_ENABLED
6550   __kmp_stats_init();
6551 #endif
6552   __kmp_init_lock(&__kmp_global_lock);
6553   __kmp_init_queuing_lock(&__kmp_dispatch_lock);
6554   __kmp_init_lock(&__kmp_debug_lock);
6555   __kmp_init_atomic_lock(&__kmp_atomic_lock);
6556   __kmp_init_atomic_lock(&__kmp_atomic_lock_1i);
6557   __kmp_init_atomic_lock(&__kmp_atomic_lock_2i);
6558   __kmp_init_atomic_lock(&__kmp_atomic_lock_4i);
6559   __kmp_init_atomic_lock(&__kmp_atomic_lock_4r);
6560   __kmp_init_atomic_lock(&__kmp_atomic_lock_8i);
6561   __kmp_init_atomic_lock(&__kmp_atomic_lock_8r);
6562   __kmp_init_atomic_lock(&__kmp_atomic_lock_8c);
6563   __kmp_init_atomic_lock(&__kmp_atomic_lock_10r);
6564   __kmp_init_atomic_lock(&__kmp_atomic_lock_16r);
6565   __kmp_init_atomic_lock(&__kmp_atomic_lock_16c);
6566   __kmp_init_atomic_lock(&__kmp_atomic_lock_20c);
6567   __kmp_init_atomic_lock(&__kmp_atomic_lock_32c);
6568   __kmp_init_bootstrap_lock(&__kmp_forkjoin_lock);
6569   __kmp_init_bootstrap_lock(&__kmp_exit_lock);
6570 #if KMP_USE_MONITOR
6571   __kmp_init_bootstrap_lock(&__kmp_monitor_lock);
6572 #endif
6573   __kmp_init_bootstrap_lock(&__kmp_tp_cached_lock);
6574 
6575   /* conduct initialization and initial setup of configuration */
6576 
6577   __kmp_runtime_initialize();
6578 
6579 #if KMP_MIC_SUPPORTED
6580   __kmp_check_mic_type();
6581 #endif
6582 
6583 // Some global variable initialization moved here from kmp_env_initialize()
6584 #ifdef KMP_DEBUG
6585   kmp_diag = 0;
6586 #endif
6587   __kmp_abort_delay = 0;
6588 
6589   // From __kmp_init_dflt_team_nth()
6590   /* assume the entire machine will be used */
6591   __kmp_dflt_team_nth_ub = __kmp_xproc;
6592   if (__kmp_dflt_team_nth_ub < KMP_MIN_NTH) {
6593     __kmp_dflt_team_nth_ub = KMP_MIN_NTH;
6594   }
6595   if (__kmp_dflt_team_nth_ub > __kmp_sys_max_nth) {
6596     __kmp_dflt_team_nth_ub = __kmp_sys_max_nth;
6597   }
6598   __kmp_max_nth = __kmp_sys_max_nth;
6599   __kmp_cg_max_nth = __kmp_sys_max_nth;
6600   __kmp_teams_max_nth = __kmp_xproc; // set a "reasonable" default
6601   if (__kmp_teams_max_nth > __kmp_sys_max_nth) {
6602     __kmp_teams_max_nth = __kmp_sys_max_nth;
6603   }
6604 
6605   // Three vars below moved here from __kmp_env_initialize() "KMP_BLOCKTIME"
6606   // part
6607   __kmp_dflt_blocktime = KMP_DEFAULT_BLOCKTIME;
6608 #if KMP_USE_MONITOR
6609   __kmp_monitor_wakeups =
6610       KMP_WAKEUPS_FROM_BLOCKTIME(__kmp_dflt_blocktime, __kmp_monitor_wakeups);
6611   __kmp_bt_intervals =
6612       KMP_INTERVALS_FROM_BLOCKTIME(__kmp_dflt_blocktime, __kmp_monitor_wakeups);
6613 #endif
6614   // From "KMP_LIBRARY" part of __kmp_env_initialize()
6615   __kmp_library = library_throughput;
6616   // From KMP_SCHEDULE initialization
6617   __kmp_static = kmp_sch_static_balanced;
6618 // AC: do not use analytical here, because it is non-monotonous
6619 //__kmp_guided = kmp_sch_guided_iterative_chunked;
6620 //__kmp_auto = kmp_sch_guided_analytical_chunked; // AC: it is the default, no
6621 // need to repeat assignment
6622 // Barrier initialization. Moved here from __kmp_env_initialize() Barrier branch
6623 // bit control and barrier method control parts
6624 #if KMP_FAST_REDUCTION_BARRIER
6625 #define kmp_reduction_barrier_gather_bb ((int)1)
6626 #define kmp_reduction_barrier_release_bb ((int)1)
6627 #define kmp_reduction_barrier_gather_pat bp_hyper_bar
6628 #define kmp_reduction_barrier_release_pat bp_hyper_bar
6629 #endif // KMP_FAST_REDUCTION_BARRIER
6630   for (i = bs_plain_barrier; i < bs_last_barrier; i++) {
6631     __kmp_barrier_gather_branch_bits[i] = __kmp_barrier_gather_bb_dflt;
6632     __kmp_barrier_release_branch_bits[i] = __kmp_barrier_release_bb_dflt;
6633     __kmp_barrier_gather_pattern[i] = __kmp_barrier_gather_pat_dflt;
6634     __kmp_barrier_release_pattern[i] = __kmp_barrier_release_pat_dflt;
6635 #if KMP_FAST_REDUCTION_BARRIER
6636     if (i == bs_reduction_barrier) { // tested and confirmed on ALTIX only (
6637       // lin_64 ): hyper,1
6638       __kmp_barrier_gather_branch_bits[i] = kmp_reduction_barrier_gather_bb;
6639       __kmp_barrier_release_branch_bits[i] = kmp_reduction_barrier_release_bb;
6640       __kmp_barrier_gather_pattern[i] = kmp_reduction_barrier_gather_pat;
6641       __kmp_barrier_release_pattern[i] = kmp_reduction_barrier_release_pat;
6642     }
6643 #endif // KMP_FAST_REDUCTION_BARRIER
6644   }
6645 #if KMP_FAST_REDUCTION_BARRIER
6646 #undef kmp_reduction_barrier_release_pat
6647 #undef kmp_reduction_barrier_gather_pat
6648 #undef kmp_reduction_barrier_release_bb
6649 #undef kmp_reduction_barrier_gather_bb
6650 #endif // KMP_FAST_REDUCTION_BARRIER
6651 #if KMP_MIC_SUPPORTED
6652   if (__kmp_mic_type == mic2) { // KNC
6653     // AC: plane=3,2, forkjoin=2,1 are optimal for 240 threads on KNC
6654     __kmp_barrier_gather_branch_bits[bs_plain_barrier] = 3; // plain gather
6655     __kmp_barrier_release_branch_bits[bs_forkjoin_barrier] =
6656         1; // forkjoin release
6657     __kmp_barrier_gather_pattern[bs_forkjoin_barrier] = bp_hierarchical_bar;
6658     __kmp_barrier_release_pattern[bs_forkjoin_barrier] = bp_hierarchical_bar;
6659   }
6660 #if KMP_FAST_REDUCTION_BARRIER
6661   if (__kmp_mic_type == mic2) { // KNC
6662     __kmp_barrier_gather_pattern[bs_reduction_barrier] = bp_hierarchical_bar;
6663     __kmp_barrier_release_pattern[bs_reduction_barrier] = bp_hierarchical_bar;
6664   }
6665 #endif // KMP_FAST_REDUCTION_BARRIER
6666 #endif // KMP_MIC_SUPPORTED
6667 
6668 // From KMP_CHECKS initialization
6669 #ifdef KMP_DEBUG
6670   __kmp_env_checks = TRUE; /* development versions have the extra checks */
6671 #else
6672   __kmp_env_checks = FALSE; /* port versions do not have the extra checks */
6673 #endif
6674 
6675   // From "KMP_FOREIGN_THREADS_THREADPRIVATE" initialization
6676   __kmp_foreign_tp = TRUE;
6677 
6678   __kmp_global.g.g_dynamic = FALSE;
6679   __kmp_global.g.g_dynamic_mode = dynamic_default;
6680 
6681   __kmp_env_initialize(NULL);
6682 
6683 // Print all messages in message catalog for testing purposes.
6684 #ifdef KMP_DEBUG
6685   char const *val = __kmp_env_get("KMP_DUMP_CATALOG");
6686   if (__kmp_str_match_true(val)) {
6687     kmp_str_buf_t buffer;
6688     __kmp_str_buf_init(&buffer);
6689     __kmp_i18n_dump_catalog(&buffer);
6690     __kmp_printf("%s", buffer.str);
6691     __kmp_str_buf_free(&buffer);
6692   }
6693   __kmp_env_free(&val);
6694 #endif
6695 
6696   __kmp_threads_capacity =
6697       __kmp_initial_threads_capacity(__kmp_dflt_team_nth_ub);
6698   // Moved here from __kmp_env_initialize() "KMP_ALL_THREADPRIVATE" part
6699   __kmp_tp_capacity = __kmp_default_tp_capacity(
6700       __kmp_dflt_team_nth_ub, __kmp_max_nth, __kmp_allThreadsSpecified);
6701 
6702   // If the library is shut down properly, both pools must be NULL. Just in
6703   // case, set them to NULL -- some memory may leak, but subsequent code will
6704   // work even if pools are not freed.
6705   KMP_DEBUG_ASSERT(__kmp_thread_pool == NULL);
6706   KMP_DEBUG_ASSERT(__kmp_thread_pool_insert_pt == NULL);
6707   KMP_DEBUG_ASSERT(__kmp_team_pool == NULL);
6708   __kmp_thread_pool = NULL;
6709   __kmp_thread_pool_insert_pt = NULL;
6710   __kmp_team_pool = NULL;
6711 
6712   /* Allocate all of the variable sized records */
6713   /* NOTE: __kmp_threads_capacity entries are allocated, but the arrays are
6714    * expandable */
6715   /* Since allocation is cache-aligned, just add extra padding at the end */
6716   size =
6717       (sizeof(kmp_info_t *) + sizeof(kmp_root_t *)) * __kmp_threads_capacity +
6718       CACHE_LINE;
6719   __kmp_threads = (kmp_info_t **)__kmp_allocate(size);
6720   __kmp_root = (kmp_root_t **)((char *)__kmp_threads +
6721                                sizeof(kmp_info_t *) * __kmp_threads_capacity);
6722 
6723   /* init thread counts */
6724   KMP_DEBUG_ASSERT(__kmp_all_nth ==
6725                    0); // Asserts fail if the library is reinitializing and
6726   KMP_DEBUG_ASSERT(__kmp_nth == 0); // something was wrong in termination.
6727   __kmp_all_nth = 0;
6728   __kmp_nth = 0;
6729 
6730   /* setup the uber master thread and hierarchy */
6731   gtid = __kmp_register_root(TRUE);
6732   KA_TRACE(10, ("__kmp_do_serial_initialize  T#%d\n", gtid));
6733   KMP_ASSERT(KMP_UBER_GTID(gtid));
6734   KMP_ASSERT(KMP_INITIAL_GTID(gtid));
6735 
6736   KMP_MB(); /* Flush all pending memory write invalidates.  */
6737 
6738   __kmp_common_initialize();
6739 
6740 #if KMP_OS_UNIX
6741   /* invoke the child fork handler */
6742   __kmp_register_atfork();
6743 #endif
6744 
6745 #if !KMP_DYNAMIC_LIB
6746   {
6747     /* Invoke the exit handler when the program finishes, only for static
6748        library. For dynamic library, we already have _fini and DllMain. */
6749     int rc = atexit(__kmp_internal_end_atexit);
6750     if (rc != 0) {
6751       __kmp_fatal(KMP_MSG(FunctionError, "atexit()"), KMP_ERR(rc),
6752                   __kmp_msg_null);
6753     }
6754   }
6755 #endif
6756 
6757 #if KMP_HANDLE_SIGNALS
6758 #if KMP_OS_UNIX
6759   /* NOTE: make sure that this is called before the user installs their own
6760      signal handlers so that the user handlers are called first. this way they
6761      can return false, not call our handler, avoid terminating the library, and
6762      continue execution where they left off. */
6763   __kmp_install_signals(FALSE);
6764 #endif /* KMP_OS_UNIX */
6765 #if KMP_OS_WINDOWS
6766   __kmp_install_signals(TRUE);
6767 #endif /* KMP_OS_WINDOWS */
6768 #endif
6769 
6770   /* we have finished the serial initialization */
6771   __kmp_init_counter++;
6772 
6773   __kmp_init_serial = TRUE;
6774 
6775   if (__kmp_settings) {
6776     __kmp_env_print();
6777   }
6778 
6779   if (__kmp_display_env || __kmp_display_env_verbose) {
6780     __kmp_env_print_2();
6781   }
6782 
6783 #if OMPT_SUPPORT
6784   ompt_post_init();
6785 #endif
6786 
6787   KMP_MB();
6788 
6789   KA_TRACE(10, ("__kmp_do_serial_initialize: exit\n"));
6790 }
6791 
6792 void __kmp_serial_initialize(void) {
6793   if (__kmp_init_serial) {
6794     return;
6795   }
6796   __kmp_acquire_bootstrap_lock(&__kmp_initz_lock);
6797   if (__kmp_init_serial) {
6798     __kmp_release_bootstrap_lock(&__kmp_initz_lock);
6799     return;
6800   }
6801   __kmp_do_serial_initialize();
6802   __kmp_release_bootstrap_lock(&__kmp_initz_lock);
6803 }
6804 
6805 static void __kmp_do_middle_initialize(void) {
6806   int i, j;
6807   int prev_dflt_team_nth;
6808 
6809   if (!__kmp_init_serial) {
6810     __kmp_do_serial_initialize();
6811   }
6812 
6813   KA_TRACE(10, ("__kmp_middle_initialize: enter\n"));
6814 
6815   // Save the previous value for the __kmp_dflt_team_nth so that
6816   // we can avoid some reinitialization if it hasn't changed.
6817   prev_dflt_team_nth = __kmp_dflt_team_nth;
6818 
6819 #if KMP_AFFINITY_SUPPORTED
6820   // __kmp_affinity_initialize() will try to set __kmp_ncores to the
6821   // number of cores on the machine.
6822   __kmp_affinity_initialize();
6823 
6824   // Run through the __kmp_threads array and set the affinity mask
6825   // for each root thread that is currently registered with the RTL.
6826   for (i = 0; i < __kmp_threads_capacity; i++) {
6827     if (TCR_PTR(__kmp_threads[i]) != NULL) {
6828       __kmp_affinity_set_init_mask(i, TRUE);
6829     }
6830   }
6831 #endif /* KMP_AFFINITY_SUPPORTED */
6832 
6833   KMP_ASSERT(__kmp_xproc > 0);
6834   if (__kmp_avail_proc == 0) {
6835     __kmp_avail_proc = __kmp_xproc;
6836   }
6837 
6838   // If there were empty places in num_threads list (OMP_NUM_THREADS=,,2,3),
6839   // correct them now
6840   j = 0;
6841   while ((j < __kmp_nested_nth.used) && !__kmp_nested_nth.nth[j]) {
6842     __kmp_nested_nth.nth[j] = __kmp_dflt_team_nth = __kmp_dflt_team_nth_ub =
6843         __kmp_avail_proc;
6844     j++;
6845   }
6846 
6847   if (__kmp_dflt_team_nth == 0) {
6848 #ifdef KMP_DFLT_NTH_CORES
6849     // Default #threads = #cores
6850     __kmp_dflt_team_nth = __kmp_ncores;
6851     KA_TRACE(20, ("__kmp_middle_initialize: setting __kmp_dflt_team_nth = "
6852                   "__kmp_ncores (%d)\n",
6853                   __kmp_dflt_team_nth));
6854 #else
6855     // Default #threads = #available OS procs
6856     __kmp_dflt_team_nth = __kmp_avail_proc;
6857     KA_TRACE(20, ("__kmp_middle_initialize: setting __kmp_dflt_team_nth = "
6858                   "__kmp_avail_proc(%d)\n",
6859                   __kmp_dflt_team_nth));
6860 #endif /* KMP_DFLT_NTH_CORES */
6861   }
6862 
6863   if (__kmp_dflt_team_nth < KMP_MIN_NTH) {
6864     __kmp_dflt_team_nth = KMP_MIN_NTH;
6865   }
6866   if (__kmp_dflt_team_nth > __kmp_sys_max_nth) {
6867     __kmp_dflt_team_nth = __kmp_sys_max_nth;
6868   }
6869 
6870   // There's no harm in continuing if the following check fails,
6871   // but it indicates an error in the previous logic.
6872   KMP_DEBUG_ASSERT(__kmp_dflt_team_nth <= __kmp_dflt_team_nth_ub);
6873 
6874   if (__kmp_dflt_team_nth != prev_dflt_team_nth) {
6875     // Run through the __kmp_threads array and set the num threads icv for each
6876     // root thread that is currently registered with the RTL (which has not
6877     // already explicitly set its nthreads-var with a call to
6878     // omp_set_num_threads()).
6879     for (i = 0; i < __kmp_threads_capacity; i++) {
6880       kmp_info_t *thread = __kmp_threads[i];
6881       if (thread == NULL)
6882         continue;
6883       if (thread->th.th_current_task->td_icvs.nproc != 0)
6884         continue;
6885 
6886       set__nproc(__kmp_threads[i], __kmp_dflt_team_nth);
6887     }
6888   }
6889   KA_TRACE(
6890       20,
6891       ("__kmp_middle_initialize: final value for __kmp_dflt_team_nth = %d\n",
6892        __kmp_dflt_team_nth));
6893 
6894 #ifdef KMP_ADJUST_BLOCKTIME
6895   /* Adjust blocktime to zero if necessary  now that __kmp_avail_proc is set */
6896   if (!__kmp_env_blocktime && (__kmp_avail_proc > 0)) {
6897     KMP_DEBUG_ASSERT(__kmp_avail_proc > 0);
6898     if (__kmp_nth > __kmp_avail_proc) {
6899       __kmp_zero_bt = TRUE;
6900     }
6901   }
6902 #endif /* KMP_ADJUST_BLOCKTIME */
6903 
6904   /* we have finished middle initialization */
6905   TCW_SYNC_4(__kmp_init_middle, TRUE);
6906 
6907   KA_TRACE(10, ("__kmp_do_middle_initialize: exit\n"));
6908 }
6909 
6910 void __kmp_middle_initialize(void) {
6911   if (__kmp_init_middle) {
6912     return;
6913   }
6914   __kmp_acquire_bootstrap_lock(&__kmp_initz_lock);
6915   if (__kmp_init_middle) {
6916     __kmp_release_bootstrap_lock(&__kmp_initz_lock);
6917     return;
6918   }
6919   __kmp_do_middle_initialize();
6920   __kmp_release_bootstrap_lock(&__kmp_initz_lock);
6921 }
6922 
6923 void __kmp_parallel_initialize(void) {
6924   int gtid = __kmp_entry_gtid(); // this might be a new root
6925 
6926   /* synchronize parallel initialization (for sibling) */
6927   if (TCR_4(__kmp_init_parallel))
6928     return;
6929   __kmp_acquire_bootstrap_lock(&__kmp_initz_lock);
6930   if (TCR_4(__kmp_init_parallel)) {
6931     __kmp_release_bootstrap_lock(&__kmp_initz_lock);
6932     return;
6933   }
6934 
6935   /* TODO reinitialization after we have already shut down */
6936   if (TCR_4(__kmp_global.g.g_done)) {
6937     KA_TRACE(
6938         10,
6939         ("__kmp_parallel_initialize: attempt to init while shutting down\n"));
6940     __kmp_infinite_loop();
6941   }
6942 
6943   /* jc: The lock __kmp_initz_lock is already held, so calling
6944      __kmp_serial_initialize would cause a deadlock.  So we call
6945      __kmp_do_serial_initialize directly. */
6946   if (!__kmp_init_middle) {
6947     __kmp_do_middle_initialize();
6948   }
6949   __kmp_resume_if_hard_paused();
6950 
6951   /* begin initialization */
6952   KA_TRACE(10, ("__kmp_parallel_initialize: enter\n"));
6953   KMP_ASSERT(KMP_UBER_GTID(gtid));
6954 
6955 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
6956   // Save the FP control regs.
6957   // Worker threads will set theirs to these values at thread startup.
6958   __kmp_store_x87_fpu_control_word(&__kmp_init_x87_fpu_control_word);
6959   __kmp_store_mxcsr(&__kmp_init_mxcsr);
6960   __kmp_init_mxcsr &= KMP_X86_MXCSR_MASK;
6961 #endif /* KMP_ARCH_X86 || KMP_ARCH_X86_64 */
6962 
6963 #if KMP_OS_UNIX
6964 #if KMP_HANDLE_SIGNALS
6965   /*  must be after __kmp_serial_initialize  */
6966   __kmp_install_signals(TRUE);
6967 #endif
6968 #endif
6969 
6970   __kmp_suspend_initialize();
6971 
6972 #if defined(USE_LOAD_BALANCE)
6973   if (__kmp_global.g.g_dynamic_mode == dynamic_default) {
6974     __kmp_global.g.g_dynamic_mode = dynamic_load_balance;
6975   }
6976 #else
6977   if (__kmp_global.g.g_dynamic_mode == dynamic_default) {
6978     __kmp_global.g.g_dynamic_mode = dynamic_thread_limit;
6979   }
6980 #endif
6981 
6982   if (__kmp_version) {
6983     __kmp_print_version_2();
6984   }
6985 
6986   /* we have finished parallel initialization */
6987   TCW_SYNC_4(__kmp_init_parallel, TRUE);
6988 
6989   KMP_MB();
6990   KA_TRACE(10, ("__kmp_parallel_initialize: exit\n"));
6991 
6992   __kmp_release_bootstrap_lock(&__kmp_initz_lock);
6993 }
6994 
6995 /* ------------------------------------------------------------------------ */
6996 
6997 void __kmp_run_before_invoked_task(int gtid, int tid, kmp_info_t *this_thr,
6998                                    kmp_team_t *team) {
6999   kmp_disp_t *dispatch;
7000 
7001   KMP_MB();
7002 
7003   /* none of the threads have encountered any constructs, yet. */
7004   this_thr->th.th_local.this_construct = 0;
7005 #if KMP_CACHE_MANAGE
7006   KMP_CACHE_PREFETCH(&this_thr->th.th_bar[bs_forkjoin_barrier].bb.b_arrived);
7007 #endif /* KMP_CACHE_MANAGE */
7008   dispatch = (kmp_disp_t *)TCR_PTR(this_thr->th.th_dispatch);
7009   KMP_DEBUG_ASSERT(dispatch);
7010   KMP_DEBUG_ASSERT(team->t.t_dispatch);
7011   // KMP_DEBUG_ASSERT( this_thr->th.th_dispatch == &team->t.t_dispatch[
7012   // this_thr->th.th_info.ds.ds_tid ] );
7013 
7014   dispatch->th_disp_index = 0; /* reset the dispatch buffer counter */
7015   dispatch->th_doacross_buf_idx = 0; // reset doacross dispatch buffer counter
7016   if (__kmp_env_consistency_check)
7017     __kmp_push_parallel(gtid, team->t.t_ident);
7018 
7019   KMP_MB(); /* Flush all pending memory write invalidates.  */
7020 }
7021 
7022 void __kmp_run_after_invoked_task(int gtid, int tid, kmp_info_t *this_thr,
7023                                   kmp_team_t *team) {
7024   if (__kmp_env_consistency_check)
7025     __kmp_pop_parallel(gtid, team->t.t_ident);
7026 
7027   __kmp_finish_implicit_task(this_thr);
7028 }
7029 
7030 int __kmp_invoke_task_func(int gtid) {
7031   int rc;
7032   int tid = __kmp_tid_from_gtid(gtid);
7033   kmp_info_t *this_thr = __kmp_threads[gtid];
7034   kmp_team_t *team = this_thr->th.th_team;
7035 
7036   __kmp_run_before_invoked_task(gtid, tid, this_thr, team);
7037 #if USE_ITT_BUILD
7038   if (__itt_stack_caller_create_ptr) {
7039     __kmp_itt_stack_callee_enter(
7040         (__itt_caller)
7041             team->t.t_stack_id); // inform ittnotify about entering user's code
7042   }
7043 #endif /* USE_ITT_BUILD */
7044 #if INCLUDE_SSC_MARKS
7045   SSC_MARK_INVOKING();
7046 #endif
7047 
7048 #if OMPT_SUPPORT
7049   void *dummy;
7050   void **exit_frame_p;
7051   ompt_data_t *my_task_data;
7052   ompt_data_t *my_parallel_data;
7053   int ompt_team_size;
7054 
7055   if (ompt_enabled.enabled) {
7056     exit_frame_p = &(
7057         team->t.t_implicit_task_taskdata[tid].ompt_task_info.frame.exit_frame.ptr);
7058   } else {
7059     exit_frame_p = &dummy;
7060   }
7061 
7062   my_task_data =
7063       &(team->t.t_implicit_task_taskdata[tid].ompt_task_info.task_data);
7064   my_parallel_data = &(team->t.ompt_team_info.parallel_data);
7065   if (ompt_enabled.ompt_callback_implicit_task) {
7066     ompt_team_size = team->t.t_nproc;
7067     ompt_callbacks.ompt_callback(ompt_callback_implicit_task)(
7068         ompt_scope_begin, my_parallel_data, my_task_data, ompt_team_size,
7069         __kmp_tid_from_gtid(gtid), ompt_task_implicit);
7070     OMPT_CUR_TASK_INFO(this_thr)->thread_num = __kmp_tid_from_gtid(gtid);
7071   }
7072 #endif
7073 
7074 #if KMP_STATS_ENABLED
7075   stats_state_e previous_state = KMP_GET_THREAD_STATE();
7076   if (previous_state == stats_state_e::TEAMS_REGION) {
7077     KMP_PUSH_PARTITIONED_TIMER(OMP_teams);
7078   } else {
7079     KMP_PUSH_PARTITIONED_TIMER(OMP_parallel);
7080   }
7081   KMP_SET_THREAD_STATE(IMPLICIT_TASK);
7082 #endif
7083 
7084   rc = __kmp_invoke_microtask((microtask_t)TCR_SYNC_PTR(team->t.t_pkfn), gtid,
7085                               tid, (int)team->t.t_argc, (void **)team->t.t_argv
7086 #if OMPT_SUPPORT
7087                               ,
7088                               exit_frame_p
7089 #endif
7090                               );
7091 #if OMPT_SUPPORT
7092   *exit_frame_p = NULL;
7093    this_thr->th.ompt_thread_info.parallel_flags |= ompt_parallel_team;
7094 #endif
7095 
7096 #if KMP_STATS_ENABLED
7097   if (previous_state == stats_state_e::TEAMS_REGION) {
7098     KMP_SET_THREAD_STATE(previous_state);
7099   }
7100   KMP_POP_PARTITIONED_TIMER();
7101 #endif
7102 
7103 #if USE_ITT_BUILD
7104   if (__itt_stack_caller_create_ptr) {
7105     __kmp_itt_stack_callee_leave(
7106         (__itt_caller)
7107             team->t.t_stack_id); // inform ittnotify about leaving user's code
7108   }
7109 #endif /* USE_ITT_BUILD */
7110   __kmp_run_after_invoked_task(gtid, tid, this_thr, team);
7111 
7112   return rc;
7113 }
7114 
7115 void __kmp_teams_master(int gtid) {
7116   // This routine is called by all master threads in teams construct
7117   kmp_info_t *thr = __kmp_threads[gtid];
7118   kmp_team_t *team = thr->th.th_team;
7119   ident_t *loc = team->t.t_ident;
7120   thr->th.th_set_nproc = thr->th.th_teams_size.nth;
7121   KMP_DEBUG_ASSERT(thr->th.th_teams_microtask);
7122   KMP_DEBUG_ASSERT(thr->th.th_set_nproc);
7123   KA_TRACE(20, ("__kmp_teams_master: T#%d, Tid %d, microtask %p\n", gtid,
7124                 __kmp_tid_from_gtid(gtid), thr->th.th_teams_microtask));
7125 
7126   // This thread is a new CG root.  Set up the proper variables.
7127   kmp_cg_root_t *tmp = (kmp_cg_root_t *)__kmp_allocate(sizeof(kmp_cg_root_t));
7128   tmp->cg_root = thr; // Make thr the CG root
7129   // Init to thread limit that was stored when league masters were forked
7130   tmp->cg_thread_limit = thr->th.th_current_task->td_icvs.thread_limit;
7131   tmp->cg_nthreads = 1; // Init counter to one active thread, this one
7132   KA_TRACE(100, ("__kmp_teams_master: Thread %p created node %p and init"
7133                  " cg_nthreads to 1\n",
7134                  thr, tmp));
7135   tmp->up = thr->th.th_cg_roots;
7136   thr->th.th_cg_roots = tmp;
7137 
7138 // Launch league of teams now, but not let workers execute
7139 // (they hang on fork barrier until next parallel)
7140 #if INCLUDE_SSC_MARKS
7141   SSC_MARK_FORKING();
7142 #endif
7143   __kmp_fork_call(loc, gtid, fork_context_intel, team->t.t_argc,
7144                   (microtask_t)thr->th.th_teams_microtask, // "wrapped" task
7145                   VOLATILE_CAST(launch_t) __kmp_invoke_task_func, NULL);
7146 #if INCLUDE_SSC_MARKS
7147   SSC_MARK_JOINING();
7148 #endif
7149   // If the team size was reduced from the limit, set it to the new size
7150   if (thr->th.th_team_nproc < thr->th.th_teams_size.nth)
7151     thr->th.th_teams_size.nth = thr->th.th_team_nproc;
7152   // AC: last parameter "1" eliminates join barrier which won't work because
7153   // worker threads are in a fork barrier waiting for more parallel regions
7154   __kmp_join_call(loc, gtid
7155 #if OMPT_SUPPORT
7156                   ,
7157                   fork_context_intel
7158 #endif
7159                   ,
7160                   1);
7161 }
7162 
7163 int __kmp_invoke_teams_master(int gtid) {
7164   kmp_info_t *this_thr = __kmp_threads[gtid];
7165   kmp_team_t *team = this_thr->th.th_team;
7166 #if KMP_DEBUG
7167   if (!__kmp_threads[gtid]->th.th_team->t.t_serialized)
7168     KMP_DEBUG_ASSERT((void *)__kmp_threads[gtid]->th.th_team->t.t_pkfn ==
7169                      (void *)__kmp_teams_master);
7170 #endif
7171   __kmp_run_before_invoked_task(gtid, 0, this_thr, team);
7172 #if OMPT_SUPPORT
7173   int tid = __kmp_tid_from_gtid(gtid);
7174   ompt_data_t *task_data =
7175       &team->t.t_implicit_task_taskdata[tid].ompt_task_info.task_data;
7176   ompt_data_t *parallel_data = &team->t.ompt_team_info.parallel_data;
7177   if (ompt_enabled.ompt_callback_implicit_task) {
7178     ompt_callbacks.ompt_callback(ompt_callback_implicit_task)(
7179         ompt_scope_begin, parallel_data, task_data, team->t.t_nproc, tid,
7180         ompt_task_initial);
7181     OMPT_CUR_TASK_INFO(this_thr)->thread_num = tid;
7182   }
7183 #endif
7184   __kmp_teams_master(gtid);
7185 #if OMPT_SUPPORT
7186   this_thr->th.ompt_thread_info.parallel_flags |= ompt_parallel_league;
7187 #endif
7188   __kmp_run_after_invoked_task(gtid, 0, this_thr, team);
7189   return 1;
7190 }
7191 
7192 /* this sets the requested number of threads for the next parallel region
7193    encountered by this team. since this should be enclosed in the forkjoin
7194    critical section it should avoid race conditions with asymmetrical nested
7195    parallelism */
7196 
7197 void __kmp_push_num_threads(ident_t *id, int gtid, int num_threads) {
7198   kmp_info_t *thr = __kmp_threads[gtid];
7199 
7200   if (num_threads > 0)
7201     thr->th.th_set_nproc = num_threads;
7202 }
7203 
7204 /* this sets the requested number of teams for the teams region and/or
7205    the number of threads for the next parallel region encountered  */
7206 void __kmp_push_num_teams(ident_t *id, int gtid, int num_teams,
7207                           int num_threads) {
7208   kmp_info_t *thr = __kmp_threads[gtid];
7209   KMP_DEBUG_ASSERT(num_teams >= 0);
7210   KMP_DEBUG_ASSERT(num_threads >= 0);
7211 
7212   if (num_teams == 0)
7213     num_teams = 1; // default number of teams is 1.
7214   if (num_teams > __kmp_teams_max_nth) { // if too many teams requested?
7215     if (!__kmp_reserve_warn) {
7216       __kmp_reserve_warn = 1;
7217       __kmp_msg(kmp_ms_warning,
7218                 KMP_MSG(CantFormThrTeam, num_teams, __kmp_teams_max_nth),
7219                 KMP_HNT(Unset_ALL_THREADS), __kmp_msg_null);
7220     }
7221     num_teams = __kmp_teams_max_nth;
7222   }
7223   // Set number of teams (number of threads in the outer "parallel" of the
7224   // teams)
7225   thr->th.th_set_nproc = thr->th.th_teams_size.nteams = num_teams;
7226 
7227   // Remember the number of threads for inner parallel regions
7228   if (!TCR_4(__kmp_init_middle))
7229     __kmp_middle_initialize(); // get internal globals calculated
7230   KMP_DEBUG_ASSERT(__kmp_avail_proc);
7231   KMP_DEBUG_ASSERT(__kmp_dflt_team_nth);
7232   if (num_threads == 0) {
7233     num_threads = __kmp_avail_proc / num_teams;
7234     // adjust num_threads w/o warning as it is not user setting
7235     // num_threads = min(num_threads, nthreads-var, thread-limit-var)
7236     // no thread_limit clause specified -  do not change thread-limit-var ICV
7237     if (num_threads > __kmp_dflt_team_nth) {
7238       num_threads = __kmp_dflt_team_nth; // honor nthreads-var ICV
7239     }
7240     if (num_threads > thr->th.th_current_task->td_icvs.thread_limit) {
7241       num_threads = thr->th.th_current_task->td_icvs.thread_limit;
7242     } // prevent team size to exceed thread-limit-var
7243     if (num_teams * num_threads > __kmp_teams_max_nth) {
7244       num_threads = __kmp_teams_max_nth / num_teams;
7245     }
7246   } else {
7247     // This thread will be the master of the league masters
7248     // Store new thread limit; old limit is saved in th_cg_roots list
7249     thr->th.th_current_task->td_icvs.thread_limit = num_threads;
7250     // num_threads = min(num_threads, nthreads-var)
7251     if (num_threads > __kmp_dflt_team_nth) {
7252       num_threads = __kmp_dflt_team_nth; // honor nthreads-var ICV
7253     }
7254     if (num_teams * num_threads > __kmp_teams_max_nth) {
7255       int new_threads = __kmp_teams_max_nth / num_teams;
7256       if (!__kmp_reserve_warn) { // user asked for too many threads
7257         __kmp_reserve_warn = 1; // conflicts with KMP_TEAMS_THREAD_LIMIT
7258         __kmp_msg(kmp_ms_warning,
7259                   KMP_MSG(CantFormThrTeam, num_threads, new_threads),
7260                   KMP_HNT(Unset_ALL_THREADS), __kmp_msg_null);
7261       }
7262       num_threads = new_threads;
7263     }
7264   }
7265   thr->th.th_teams_size.nth = num_threads;
7266 }
7267 
7268 // Set the proc_bind var to use in the following parallel region.
7269 void __kmp_push_proc_bind(ident_t *id, int gtid, kmp_proc_bind_t proc_bind) {
7270   kmp_info_t *thr = __kmp_threads[gtid];
7271   thr->th.th_set_proc_bind = proc_bind;
7272 }
7273 
7274 /* Launch the worker threads into the microtask. */
7275 
7276 void __kmp_internal_fork(ident_t *id, int gtid, kmp_team_t *team) {
7277   kmp_info_t *this_thr = __kmp_threads[gtid];
7278 
7279 #ifdef KMP_DEBUG
7280   int f;
7281 #endif /* KMP_DEBUG */
7282 
7283   KMP_DEBUG_ASSERT(team);
7284   KMP_DEBUG_ASSERT(this_thr->th.th_team == team);
7285   KMP_ASSERT(KMP_MASTER_GTID(gtid));
7286   KMP_MB(); /* Flush all pending memory write invalidates.  */
7287 
7288   team->t.t_construct = 0; /* no single directives seen yet */
7289   team->t.t_ordered.dt.t_value =
7290       0; /* thread 0 enters the ordered section first */
7291 
7292   /* Reset the identifiers on the dispatch buffer */
7293   KMP_DEBUG_ASSERT(team->t.t_disp_buffer);
7294   if (team->t.t_max_nproc > 1) {
7295     int i;
7296     for (i = 0; i < __kmp_dispatch_num_buffers; ++i) {
7297       team->t.t_disp_buffer[i].buffer_index = i;
7298       team->t.t_disp_buffer[i].doacross_buf_idx = i;
7299     }
7300   } else {
7301     team->t.t_disp_buffer[0].buffer_index = 0;
7302     team->t.t_disp_buffer[0].doacross_buf_idx = 0;
7303   }
7304 
7305   KMP_MB(); /* Flush all pending memory write invalidates.  */
7306   KMP_ASSERT(this_thr->th.th_team == team);
7307 
7308 #ifdef KMP_DEBUG
7309   for (f = 0; f < team->t.t_nproc; f++) {
7310     KMP_DEBUG_ASSERT(team->t.t_threads[f] &&
7311                      team->t.t_threads[f]->th.th_team_nproc == team->t.t_nproc);
7312   }
7313 #endif /* KMP_DEBUG */
7314 
7315   /* release the worker threads so they may begin working */
7316   __kmp_fork_barrier(gtid, 0);
7317 }
7318 
7319 void __kmp_internal_join(ident_t *id, int gtid, kmp_team_t *team) {
7320   kmp_info_t *this_thr = __kmp_threads[gtid];
7321 
7322   KMP_DEBUG_ASSERT(team);
7323   KMP_DEBUG_ASSERT(this_thr->th.th_team == team);
7324   KMP_ASSERT(KMP_MASTER_GTID(gtid));
7325   KMP_MB(); /* Flush all pending memory write invalidates.  */
7326 
7327 /* Join barrier after fork */
7328 
7329 #ifdef KMP_DEBUG
7330   if (__kmp_threads[gtid] &&
7331       __kmp_threads[gtid]->th.th_team_nproc != team->t.t_nproc) {
7332     __kmp_printf("GTID: %d, __kmp_threads[%d]=%p\n", gtid, gtid,
7333                  __kmp_threads[gtid]);
7334     __kmp_printf("__kmp_threads[%d]->th.th_team_nproc=%d, TEAM: %p, "
7335                  "team->t.t_nproc=%d\n",
7336                  gtid, __kmp_threads[gtid]->th.th_team_nproc, team,
7337                  team->t.t_nproc);
7338     __kmp_print_structure();
7339   }
7340   KMP_DEBUG_ASSERT(__kmp_threads[gtid] &&
7341                    __kmp_threads[gtid]->th.th_team_nproc == team->t.t_nproc);
7342 #endif /* KMP_DEBUG */
7343 
7344   __kmp_join_barrier(gtid); /* wait for everyone */
7345 #if OMPT_SUPPORT
7346   if (ompt_enabled.enabled &&
7347       this_thr->th.ompt_thread_info.state == ompt_state_wait_barrier_implicit) {
7348     int ds_tid = this_thr->th.th_info.ds.ds_tid;
7349     ompt_data_t *task_data = OMPT_CUR_TASK_DATA(this_thr);
7350     this_thr->th.ompt_thread_info.state = ompt_state_overhead;
7351 #if OMPT_OPTIONAL
7352     void *codeptr = NULL;
7353     if (KMP_MASTER_TID(ds_tid) &&
7354         (ompt_callbacks.ompt_callback(ompt_callback_sync_region_wait) ||
7355          ompt_callbacks.ompt_callback(ompt_callback_sync_region)))
7356       codeptr = OMPT_CUR_TEAM_INFO(this_thr)->master_return_address;
7357 
7358     if (ompt_enabled.ompt_callback_sync_region_wait) {
7359       ompt_callbacks.ompt_callback(ompt_callback_sync_region_wait)(
7360           ompt_sync_region_barrier_implicit, ompt_scope_end, NULL, task_data,
7361           codeptr);
7362     }
7363     if (ompt_enabled.ompt_callback_sync_region) {
7364       ompt_callbacks.ompt_callback(ompt_callback_sync_region)(
7365           ompt_sync_region_barrier_implicit, ompt_scope_end, NULL, task_data,
7366           codeptr);
7367     }
7368 #endif
7369     if (!KMP_MASTER_TID(ds_tid) && ompt_enabled.ompt_callback_implicit_task) {
7370       ompt_callbacks.ompt_callback(ompt_callback_implicit_task)(
7371           ompt_scope_end, NULL, task_data, 0, ds_tid, ompt_task_implicit); // TODO: Can this be ompt_task_initial?
7372     }
7373   }
7374 #endif
7375 
7376   KMP_MB(); /* Flush all pending memory write invalidates.  */
7377   KMP_ASSERT(this_thr->th.th_team == team);
7378 }
7379 
7380 /* ------------------------------------------------------------------------ */
7381 
7382 #ifdef USE_LOAD_BALANCE
7383 
7384 // Return the worker threads actively spinning in the hot team, if we
7385 // are at the outermost level of parallelism.  Otherwise, return 0.
7386 static int __kmp_active_hot_team_nproc(kmp_root_t *root) {
7387   int i;
7388   int retval;
7389   kmp_team_t *hot_team;
7390 
7391   if (root->r.r_active) {
7392     return 0;
7393   }
7394   hot_team = root->r.r_hot_team;
7395   if (__kmp_dflt_blocktime == KMP_MAX_BLOCKTIME) {
7396     return hot_team->t.t_nproc - 1; // Don't count master thread
7397   }
7398 
7399   // Skip the master thread - it is accounted for elsewhere.
7400   retval = 0;
7401   for (i = 1; i < hot_team->t.t_nproc; i++) {
7402     if (hot_team->t.t_threads[i]->th.th_active) {
7403       retval++;
7404     }
7405   }
7406   return retval;
7407 }
7408 
7409 // Perform an automatic adjustment to the number of
7410 // threads used by the next parallel region.
7411 static int __kmp_load_balance_nproc(kmp_root_t *root, int set_nproc) {
7412   int retval;
7413   int pool_active;
7414   int hot_team_active;
7415   int team_curr_active;
7416   int system_active;
7417 
7418   KB_TRACE(20, ("__kmp_load_balance_nproc: called root:%p set_nproc:%d\n", root,
7419                 set_nproc));
7420   KMP_DEBUG_ASSERT(root);
7421   KMP_DEBUG_ASSERT(root->r.r_root_team->t.t_threads[0]
7422                        ->th.th_current_task->td_icvs.dynamic == TRUE);
7423   KMP_DEBUG_ASSERT(set_nproc > 1);
7424 
7425   if (set_nproc == 1) {
7426     KB_TRACE(20, ("__kmp_load_balance_nproc: serial execution.\n"));
7427     return 1;
7428   }
7429 
7430   // Threads that are active in the thread pool, active in the hot team for this
7431   // particular root (if we are at the outer par level), and the currently
7432   // executing thread (to become the master) are available to add to the new
7433   // team, but are currently contributing to the system load, and must be
7434   // accounted for.
7435   pool_active = __kmp_thread_pool_active_nth;
7436   hot_team_active = __kmp_active_hot_team_nproc(root);
7437   team_curr_active = pool_active + hot_team_active + 1;
7438 
7439   // Check the system load.
7440   system_active = __kmp_get_load_balance(__kmp_avail_proc + team_curr_active);
7441   KB_TRACE(30, ("__kmp_load_balance_nproc: system active = %d pool active = %d "
7442                 "hot team active = %d\n",
7443                 system_active, pool_active, hot_team_active));
7444 
7445   if (system_active < 0) {
7446     // There was an error reading the necessary info from /proc, so use the
7447     // thread limit algorithm instead. Once we set __kmp_global.g.g_dynamic_mode
7448     // = dynamic_thread_limit, we shouldn't wind up getting back here.
7449     __kmp_global.g.g_dynamic_mode = dynamic_thread_limit;
7450     KMP_WARNING(CantLoadBalUsing, "KMP_DYNAMIC_MODE=thread limit");
7451 
7452     // Make this call behave like the thread limit algorithm.
7453     retval = __kmp_avail_proc - __kmp_nth +
7454              (root->r.r_active ? 1 : root->r.r_hot_team->t.t_nproc);
7455     if (retval > set_nproc) {
7456       retval = set_nproc;
7457     }
7458     if (retval < KMP_MIN_NTH) {
7459       retval = KMP_MIN_NTH;
7460     }
7461 
7462     KB_TRACE(20, ("__kmp_load_balance_nproc: thread limit exit. retval:%d\n",
7463                   retval));
7464     return retval;
7465   }
7466 
7467   // There is a slight delay in the load balance algorithm in detecting new
7468   // running procs. The real system load at this instant should be at least as
7469   // large as the #active omp thread that are available to add to the team.
7470   if (system_active < team_curr_active) {
7471     system_active = team_curr_active;
7472   }
7473   retval = __kmp_avail_proc - system_active + team_curr_active;
7474   if (retval > set_nproc) {
7475     retval = set_nproc;
7476   }
7477   if (retval < KMP_MIN_NTH) {
7478     retval = KMP_MIN_NTH;
7479   }
7480 
7481   KB_TRACE(20, ("__kmp_load_balance_nproc: exit. retval:%d\n", retval));
7482   return retval;
7483 } // __kmp_load_balance_nproc()
7484 
7485 #endif /* USE_LOAD_BALANCE */
7486 
7487 /* ------------------------------------------------------------------------ */
7488 
7489 /* NOTE: this is called with the __kmp_init_lock held */
7490 void __kmp_cleanup(void) {
7491   int f;
7492 
7493   KA_TRACE(10, ("__kmp_cleanup: enter\n"));
7494 
7495   if (TCR_4(__kmp_init_parallel)) {
7496 #if KMP_HANDLE_SIGNALS
7497     __kmp_remove_signals();
7498 #endif
7499     TCW_4(__kmp_init_parallel, FALSE);
7500   }
7501 
7502   if (TCR_4(__kmp_init_middle)) {
7503 #if KMP_AFFINITY_SUPPORTED
7504     __kmp_affinity_uninitialize();
7505 #endif /* KMP_AFFINITY_SUPPORTED */
7506     __kmp_cleanup_hierarchy();
7507     TCW_4(__kmp_init_middle, FALSE);
7508   }
7509 
7510   KA_TRACE(10, ("__kmp_cleanup: go serial cleanup\n"));
7511 
7512   if (__kmp_init_serial) {
7513     __kmp_runtime_destroy();
7514     __kmp_init_serial = FALSE;
7515   }
7516 
7517   __kmp_cleanup_threadprivate_caches();
7518 
7519   for (f = 0; f < __kmp_threads_capacity; f++) {
7520     if (__kmp_root[f] != NULL) {
7521       __kmp_free(__kmp_root[f]);
7522       __kmp_root[f] = NULL;
7523     }
7524   }
7525   __kmp_free(__kmp_threads);
7526   // __kmp_threads and __kmp_root were allocated at once, as single block, so
7527   // there is no need in freeing __kmp_root.
7528   __kmp_threads = NULL;
7529   __kmp_root = NULL;
7530   __kmp_threads_capacity = 0;
7531 
7532 #if KMP_USE_DYNAMIC_LOCK
7533   __kmp_cleanup_indirect_user_locks();
7534 #else
7535   __kmp_cleanup_user_locks();
7536 #endif
7537 
7538 #if KMP_AFFINITY_SUPPORTED
7539   KMP_INTERNAL_FREE(CCAST(char *, __kmp_cpuinfo_file));
7540   __kmp_cpuinfo_file = NULL;
7541 #endif /* KMP_AFFINITY_SUPPORTED */
7542 
7543 #if KMP_USE_ADAPTIVE_LOCKS
7544 #if KMP_DEBUG_ADAPTIVE_LOCKS
7545   __kmp_print_speculative_stats();
7546 #endif
7547 #endif
7548   KMP_INTERNAL_FREE(__kmp_nested_nth.nth);
7549   __kmp_nested_nth.nth = NULL;
7550   __kmp_nested_nth.size = 0;
7551   __kmp_nested_nth.used = 0;
7552   KMP_INTERNAL_FREE(__kmp_nested_proc_bind.bind_types);
7553   __kmp_nested_proc_bind.bind_types = NULL;
7554   __kmp_nested_proc_bind.size = 0;
7555   __kmp_nested_proc_bind.used = 0;
7556   if (__kmp_affinity_format) {
7557     KMP_INTERNAL_FREE(__kmp_affinity_format);
7558     __kmp_affinity_format = NULL;
7559   }
7560 
7561   __kmp_i18n_catclose();
7562 
7563 #if KMP_USE_HIER_SCHED
7564   __kmp_hier_scheds.deallocate();
7565 #endif
7566 
7567 #if KMP_STATS_ENABLED
7568   __kmp_stats_fini();
7569 #endif
7570 
7571   KA_TRACE(10, ("__kmp_cleanup: exit\n"));
7572 }
7573 
7574 /* ------------------------------------------------------------------------ */
7575 
7576 int __kmp_ignore_mppbeg(void) {
7577   char *env;
7578 
7579   if ((env = getenv("KMP_IGNORE_MPPBEG")) != NULL) {
7580     if (__kmp_str_match_false(env))
7581       return FALSE;
7582   }
7583   // By default __kmpc_begin() is no-op.
7584   return TRUE;
7585 }
7586 
7587 int __kmp_ignore_mppend(void) {
7588   char *env;
7589 
7590   if ((env = getenv("KMP_IGNORE_MPPEND")) != NULL) {
7591     if (__kmp_str_match_false(env))
7592       return FALSE;
7593   }
7594   // By default __kmpc_end() is no-op.
7595   return TRUE;
7596 }
7597 
7598 void __kmp_internal_begin(void) {
7599   int gtid;
7600   kmp_root_t *root;
7601 
7602   /* this is a very important step as it will register new sibling threads
7603      and assign these new uber threads a new gtid */
7604   gtid = __kmp_entry_gtid();
7605   root = __kmp_threads[gtid]->th.th_root;
7606   KMP_ASSERT(KMP_UBER_GTID(gtid));
7607 
7608   if (root->r.r_begin)
7609     return;
7610   __kmp_acquire_lock(&root->r.r_begin_lock, gtid);
7611   if (root->r.r_begin) {
7612     __kmp_release_lock(&root->r.r_begin_lock, gtid);
7613     return;
7614   }
7615 
7616   root->r.r_begin = TRUE;
7617 
7618   __kmp_release_lock(&root->r.r_begin_lock, gtid);
7619 }
7620 
7621 /* ------------------------------------------------------------------------ */
7622 
7623 void __kmp_user_set_library(enum library_type arg) {
7624   int gtid;
7625   kmp_root_t *root;
7626   kmp_info_t *thread;
7627 
7628   /* first, make sure we are initialized so we can get our gtid */
7629 
7630   gtid = __kmp_entry_gtid();
7631   thread = __kmp_threads[gtid];
7632 
7633   root = thread->th.th_root;
7634 
7635   KA_TRACE(20, ("__kmp_user_set_library: enter T#%d, arg: %d, %d\n", gtid, arg,
7636                 library_serial));
7637   if (root->r.r_in_parallel) { /* Must be called in serial section of top-level
7638                                   thread */
7639     KMP_WARNING(SetLibraryIncorrectCall);
7640     return;
7641   }
7642 
7643   switch (arg) {
7644   case library_serial:
7645     thread->th.th_set_nproc = 0;
7646     set__nproc(thread, 1);
7647     break;
7648   case library_turnaround:
7649     thread->th.th_set_nproc = 0;
7650     set__nproc(thread, __kmp_dflt_team_nth ? __kmp_dflt_team_nth
7651                                            : __kmp_dflt_team_nth_ub);
7652     break;
7653   case library_throughput:
7654     thread->th.th_set_nproc = 0;
7655     set__nproc(thread, __kmp_dflt_team_nth ? __kmp_dflt_team_nth
7656                                            : __kmp_dflt_team_nth_ub);
7657     break;
7658   default:
7659     KMP_FATAL(UnknownLibraryType, arg);
7660   }
7661 
7662   __kmp_aux_set_library(arg);
7663 }
7664 
7665 void __kmp_aux_set_stacksize(size_t arg) {
7666   if (!__kmp_init_serial)
7667     __kmp_serial_initialize();
7668 
7669 #if KMP_OS_DARWIN
7670   if (arg & (0x1000 - 1)) {
7671     arg &= ~(0x1000 - 1);
7672     if (arg + 0x1000) /* check for overflow if we round up */
7673       arg += 0x1000;
7674   }
7675 #endif
7676   __kmp_acquire_bootstrap_lock(&__kmp_initz_lock);
7677 
7678   /* only change the default stacksize before the first parallel region */
7679   if (!TCR_4(__kmp_init_parallel)) {
7680     size_t value = arg; /* argument is in bytes */
7681 
7682     if (value < __kmp_sys_min_stksize)
7683       value = __kmp_sys_min_stksize;
7684     else if (value > KMP_MAX_STKSIZE)
7685       value = KMP_MAX_STKSIZE;
7686 
7687     __kmp_stksize = value;
7688 
7689     __kmp_env_stksize = TRUE; /* was KMP_STACKSIZE specified? */
7690   }
7691 
7692   __kmp_release_bootstrap_lock(&__kmp_initz_lock);
7693 }
7694 
7695 /* set the behaviour of the runtime library */
7696 /* TODO this can cause some odd behaviour with sibling parallelism... */
7697 void __kmp_aux_set_library(enum library_type arg) {
7698   __kmp_library = arg;
7699 
7700   switch (__kmp_library) {
7701   case library_serial: {
7702     KMP_INFORM(LibraryIsSerial);
7703   } break;
7704   case library_turnaround:
7705     if (__kmp_use_yield == 1 && !__kmp_use_yield_exp_set)
7706       __kmp_use_yield = 2; // only yield when oversubscribed
7707     break;
7708   case library_throughput:
7709     if (__kmp_dflt_blocktime == KMP_MAX_BLOCKTIME)
7710       __kmp_dflt_blocktime = 200;
7711     break;
7712   default:
7713     KMP_FATAL(UnknownLibraryType, arg);
7714   }
7715 }
7716 
7717 /* Getting team information common for all team API */
7718 // Returns NULL if not in teams construct
7719 static kmp_team_t *__kmp_aux_get_team_info(int &teams_serialized) {
7720   kmp_info_t *thr = __kmp_entry_thread();
7721   teams_serialized = 0;
7722   if (thr->th.th_teams_microtask) {
7723     kmp_team_t *team = thr->th.th_team;
7724     int tlevel = thr->th.th_teams_level; // the level of the teams construct
7725     int ii = team->t.t_level;
7726     teams_serialized = team->t.t_serialized;
7727     int level = tlevel + 1;
7728     KMP_DEBUG_ASSERT(ii >= tlevel);
7729     while (ii > level) {
7730       for (teams_serialized = team->t.t_serialized;
7731            (teams_serialized > 0) && (ii > level); teams_serialized--, ii--) {
7732       }
7733       if (team->t.t_serialized && (!teams_serialized)) {
7734         team = team->t.t_parent;
7735         continue;
7736       }
7737       if (ii > level) {
7738         team = team->t.t_parent;
7739         ii--;
7740       }
7741     }
7742     return team;
7743   }
7744   return NULL;
7745 }
7746 
7747 int __kmp_aux_get_team_num() {
7748   int serialized;
7749   kmp_team_t *team = __kmp_aux_get_team_info(serialized);
7750   if (team) {
7751     if (serialized > 1) {
7752       return 0; // teams region is serialized ( 1 team of 1 thread ).
7753     } else {
7754       return team->t.t_master_tid;
7755     }
7756   }
7757   return 0;
7758 }
7759 
7760 int __kmp_aux_get_num_teams() {
7761   int serialized;
7762   kmp_team_t *team = __kmp_aux_get_team_info(serialized);
7763   if (team) {
7764     if (serialized > 1) {
7765       return 1;
7766     } else {
7767       return team->t.t_parent->t.t_nproc;
7768     }
7769   }
7770   return 1;
7771 }
7772 
7773 /* ------------------------------------------------------------------------ */
7774 
7775 /*
7776  * Affinity Format Parser
7777  *
7778  * Field is in form of: %[[[0].]size]type
7779  * % and type are required (%% means print a literal '%')
7780  * type is either single char or long name surrounded by {},
7781  * e.g., N or {num_threads}
7782  * 0 => leading zeros
7783  * . => right justified when size is specified
7784  * by default output is left justified
7785  * size is the *minimum* field length
7786  * All other characters are printed as is
7787  *
7788  * Available field types:
7789  * L {thread_level}      - omp_get_level()
7790  * n {thread_num}        - omp_get_thread_num()
7791  * h {host}              - name of host machine
7792  * P {process_id}        - process id (integer)
7793  * T {thread_identifier} - native thread identifier (integer)
7794  * N {num_threads}       - omp_get_num_threads()
7795  * A {ancestor_tnum}     - omp_get_ancestor_thread_num(omp_get_level()-1)
7796  * a {thread_affinity}   - comma separated list of integers or integer ranges
7797  *                         (values of affinity mask)
7798  *
7799  * Implementation-specific field types can be added
7800  * If a type is unknown, print "undefined"
7801 */
7802 
7803 // Structure holding the short name, long name, and corresponding data type
7804 // for snprintf.  A table of these will represent the entire valid keyword
7805 // field types.
7806 typedef struct kmp_affinity_format_field_t {
7807   char short_name; // from spec e.g., L -> thread level
7808   const char *long_name; // from spec thread_level -> thread level
7809   char field_format; // data type for snprintf (typically 'd' or 's'
7810   // for integer or string)
7811 } kmp_affinity_format_field_t;
7812 
7813 static const kmp_affinity_format_field_t __kmp_affinity_format_table[] = {
7814 #if KMP_AFFINITY_SUPPORTED
7815     {'A', "thread_affinity", 's'},
7816 #endif
7817     {'t', "team_num", 'd'},
7818     {'T', "num_teams", 'd'},
7819     {'L', "nesting_level", 'd'},
7820     {'n', "thread_num", 'd'},
7821     {'N', "num_threads", 'd'},
7822     {'a', "ancestor_tnum", 'd'},
7823     {'H', "host", 's'},
7824     {'P', "process_id", 'd'},
7825     {'i', "native_thread_id", 'd'}};
7826 
7827 // Return the number of characters it takes to hold field
7828 static int __kmp_aux_capture_affinity_field(int gtid, const kmp_info_t *th,
7829                                             const char **ptr,
7830                                             kmp_str_buf_t *field_buffer) {
7831   int rc, format_index, field_value;
7832   const char *width_left, *width_right;
7833   bool pad_zeros, right_justify, parse_long_name, found_valid_name;
7834   static const int FORMAT_SIZE = 20;
7835   char format[FORMAT_SIZE] = {0};
7836   char absolute_short_name = 0;
7837 
7838   KMP_DEBUG_ASSERT(gtid >= 0);
7839   KMP_DEBUG_ASSERT(th);
7840   KMP_DEBUG_ASSERT(**ptr == '%');
7841   KMP_DEBUG_ASSERT(field_buffer);
7842 
7843   __kmp_str_buf_clear(field_buffer);
7844 
7845   // Skip the initial %
7846   (*ptr)++;
7847 
7848   // Check for %% first
7849   if (**ptr == '%') {
7850     __kmp_str_buf_cat(field_buffer, "%", 1);
7851     (*ptr)++; // skip over the second %
7852     return 1;
7853   }
7854 
7855   // Parse field modifiers if they are present
7856   pad_zeros = false;
7857   if (**ptr == '0') {
7858     pad_zeros = true;
7859     (*ptr)++; // skip over 0
7860   }
7861   right_justify = false;
7862   if (**ptr == '.') {
7863     right_justify = true;
7864     (*ptr)++; // skip over .
7865   }
7866   // Parse width of field: [width_left, width_right)
7867   width_left = width_right = NULL;
7868   if (**ptr >= '0' && **ptr <= '9') {
7869     width_left = *ptr;
7870     SKIP_DIGITS(*ptr);
7871     width_right = *ptr;
7872   }
7873 
7874   // Create the format for KMP_SNPRINTF based on flags parsed above
7875   format_index = 0;
7876   format[format_index++] = '%';
7877   if (!right_justify)
7878     format[format_index++] = '-';
7879   if (pad_zeros)
7880     format[format_index++] = '0';
7881   if (width_left && width_right) {
7882     int i = 0;
7883     // Only allow 8 digit number widths.
7884     // This also prevents overflowing format variable
7885     while (i < 8 && width_left < width_right) {
7886       format[format_index++] = *width_left;
7887       width_left++;
7888       i++;
7889     }
7890   }
7891 
7892   // Parse a name (long or short)
7893   // Canonicalize the name into absolute_short_name
7894   found_valid_name = false;
7895   parse_long_name = (**ptr == '{');
7896   if (parse_long_name)
7897     (*ptr)++; // skip initial left brace
7898   for (size_t i = 0; i < sizeof(__kmp_affinity_format_table) /
7899                              sizeof(__kmp_affinity_format_table[0]);
7900        ++i) {
7901     char short_name = __kmp_affinity_format_table[i].short_name;
7902     const char *long_name = __kmp_affinity_format_table[i].long_name;
7903     char field_format = __kmp_affinity_format_table[i].field_format;
7904     if (parse_long_name) {
7905       int length = KMP_STRLEN(long_name);
7906       if (strncmp(*ptr, long_name, length) == 0) {
7907         found_valid_name = true;
7908         (*ptr) += length; // skip the long name
7909       }
7910     } else if (**ptr == short_name) {
7911       found_valid_name = true;
7912       (*ptr)++; // skip the short name
7913     }
7914     if (found_valid_name) {
7915       format[format_index++] = field_format;
7916       format[format_index++] = '\0';
7917       absolute_short_name = short_name;
7918       break;
7919     }
7920   }
7921   if (parse_long_name) {
7922     if (**ptr != '}') {
7923       absolute_short_name = 0;
7924     } else {
7925       (*ptr)++; // skip over the right brace
7926     }
7927   }
7928 
7929   // Attempt to fill the buffer with the requested
7930   // value using snprintf within __kmp_str_buf_print()
7931   switch (absolute_short_name) {
7932   case 't':
7933     rc = __kmp_str_buf_print(field_buffer, format, __kmp_aux_get_team_num());
7934     break;
7935   case 'T':
7936     rc = __kmp_str_buf_print(field_buffer, format, __kmp_aux_get_num_teams());
7937     break;
7938   case 'L':
7939     rc = __kmp_str_buf_print(field_buffer, format, th->th.th_team->t.t_level);
7940     break;
7941   case 'n':
7942     rc = __kmp_str_buf_print(field_buffer, format, __kmp_tid_from_gtid(gtid));
7943     break;
7944   case 'H': {
7945     static const int BUFFER_SIZE = 256;
7946     char buf[BUFFER_SIZE];
7947     __kmp_expand_host_name(buf, BUFFER_SIZE);
7948     rc = __kmp_str_buf_print(field_buffer, format, buf);
7949   } break;
7950   case 'P':
7951     rc = __kmp_str_buf_print(field_buffer, format, getpid());
7952     break;
7953   case 'i':
7954     rc = __kmp_str_buf_print(field_buffer, format, __kmp_gettid());
7955     break;
7956   case 'N':
7957     rc = __kmp_str_buf_print(field_buffer, format, th->th.th_team->t.t_nproc);
7958     break;
7959   case 'a':
7960     field_value =
7961         __kmp_get_ancestor_thread_num(gtid, th->th.th_team->t.t_level - 1);
7962     rc = __kmp_str_buf_print(field_buffer, format, field_value);
7963     break;
7964 #if KMP_AFFINITY_SUPPORTED
7965   case 'A': {
7966     kmp_str_buf_t buf;
7967     __kmp_str_buf_init(&buf);
7968     __kmp_affinity_str_buf_mask(&buf, th->th.th_affin_mask);
7969     rc = __kmp_str_buf_print(field_buffer, format, buf.str);
7970     __kmp_str_buf_free(&buf);
7971   } break;
7972 #endif
7973   default:
7974     // According to spec, If an implementation does not have info for field
7975     // type, then "undefined" is printed
7976     rc = __kmp_str_buf_print(field_buffer, "%s", "undefined");
7977     // Skip the field
7978     if (parse_long_name) {
7979       SKIP_TOKEN(*ptr);
7980       if (**ptr == '}')
7981         (*ptr)++;
7982     } else {
7983       (*ptr)++;
7984     }
7985   }
7986 
7987   KMP_ASSERT(format_index <= FORMAT_SIZE);
7988   return rc;
7989 }
7990 
7991 /*
7992  * Return number of characters needed to hold the affinity string
7993  * (not including null byte character)
7994  * The resultant string is printed to buffer, which the caller can then
7995  * handle afterwards
7996 */
7997 size_t __kmp_aux_capture_affinity(int gtid, const char *format,
7998                                   kmp_str_buf_t *buffer) {
7999   const char *parse_ptr;
8000   size_t retval;
8001   const kmp_info_t *th;
8002   kmp_str_buf_t field;
8003 
8004   KMP_DEBUG_ASSERT(buffer);
8005   KMP_DEBUG_ASSERT(gtid >= 0);
8006 
8007   __kmp_str_buf_init(&field);
8008   __kmp_str_buf_clear(buffer);
8009 
8010   th = __kmp_threads[gtid];
8011   retval = 0;
8012 
8013   // If format is NULL or zero-length string, then we use
8014   // affinity-format-var ICV
8015   parse_ptr = format;
8016   if (parse_ptr == NULL || *parse_ptr == '\0') {
8017     parse_ptr = __kmp_affinity_format;
8018   }
8019   KMP_DEBUG_ASSERT(parse_ptr);
8020 
8021   while (*parse_ptr != '\0') {
8022     // Parse a field
8023     if (*parse_ptr == '%') {
8024       // Put field in the buffer
8025       int rc = __kmp_aux_capture_affinity_field(gtid, th, &parse_ptr, &field);
8026       __kmp_str_buf_catbuf(buffer, &field);
8027       retval += rc;
8028     } else {
8029       // Put literal character in buffer
8030       __kmp_str_buf_cat(buffer, parse_ptr, 1);
8031       retval++;
8032       parse_ptr++;
8033     }
8034   }
8035   __kmp_str_buf_free(&field);
8036   return retval;
8037 }
8038 
8039 // Displays the affinity string to stdout
8040 void __kmp_aux_display_affinity(int gtid, const char *format) {
8041   kmp_str_buf_t buf;
8042   __kmp_str_buf_init(&buf);
8043   __kmp_aux_capture_affinity(gtid, format, &buf);
8044   __kmp_fprintf(kmp_out, "%s" KMP_END_OF_LINE, buf.str);
8045   __kmp_str_buf_free(&buf);
8046 }
8047 
8048 /* ------------------------------------------------------------------------ */
8049 
8050 void __kmp_aux_set_blocktime(int arg, kmp_info_t *thread, int tid) {
8051   int blocktime = arg; /* argument is in milliseconds */
8052 #if KMP_USE_MONITOR
8053   int bt_intervals;
8054 #endif
8055   int bt_set;
8056 
8057   __kmp_save_internal_controls(thread);
8058 
8059   /* Normalize and set blocktime for the teams */
8060   if (blocktime < KMP_MIN_BLOCKTIME)
8061     blocktime = KMP_MIN_BLOCKTIME;
8062   else if (blocktime > KMP_MAX_BLOCKTIME)
8063     blocktime = KMP_MAX_BLOCKTIME;
8064 
8065   set__blocktime_team(thread->th.th_team, tid, blocktime);
8066   set__blocktime_team(thread->th.th_serial_team, 0, blocktime);
8067 
8068 #if KMP_USE_MONITOR
8069   /* Calculate and set blocktime intervals for the teams */
8070   bt_intervals = KMP_INTERVALS_FROM_BLOCKTIME(blocktime, __kmp_monitor_wakeups);
8071 
8072   set__bt_intervals_team(thread->th.th_team, tid, bt_intervals);
8073   set__bt_intervals_team(thread->th.th_serial_team, 0, bt_intervals);
8074 #endif
8075 
8076   /* Set whether blocktime has been set to "TRUE" */
8077   bt_set = TRUE;
8078 
8079   set__bt_set_team(thread->th.th_team, tid, bt_set);
8080   set__bt_set_team(thread->th.th_serial_team, 0, bt_set);
8081 #if KMP_USE_MONITOR
8082   KF_TRACE(10, ("kmp_set_blocktime: T#%d(%d:%d), blocktime=%d, "
8083                 "bt_intervals=%d, monitor_updates=%d\n",
8084                 __kmp_gtid_from_tid(tid, thread->th.th_team),
8085                 thread->th.th_team->t.t_id, tid, blocktime, bt_intervals,
8086                 __kmp_monitor_wakeups));
8087 #else
8088   KF_TRACE(10, ("kmp_set_blocktime: T#%d(%d:%d), blocktime=%d\n",
8089                 __kmp_gtid_from_tid(tid, thread->th.th_team),
8090                 thread->th.th_team->t.t_id, tid, blocktime));
8091 #endif
8092 }
8093 
8094 void __kmp_aux_set_defaults(char const *str, int len) {
8095   if (!__kmp_init_serial) {
8096     __kmp_serial_initialize();
8097   }
8098   __kmp_env_initialize(str);
8099 
8100   if (__kmp_settings || __kmp_display_env || __kmp_display_env_verbose) {
8101     __kmp_env_print();
8102   }
8103 } // __kmp_aux_set_defaults
8104 
8105 /* ------------------------------------------------------------------------ */
8106 /* internal fast reduction routines */
8107 
8108 PACKED_REDUCTION_METHOD_T
8109 __kmp_determine_reduction_method(
8110     ident_t *loc, kmp_int32 global_tid, kmp_int32 num_vars, size_t reduce_size,
8111     void *reduce_data, void (*reduce_func)(void *lhs_data, void *rhs_data),
8112     kmp_critical_name *lck) {
8113 
8114   // Default reduction method: critical construct ( lck != NULL, like in current
8115   // PAROPT )
8116   // If ( reduce_data!=NULL && reduce_func!=NULL ): the tree-reduction method
8117   // can be selected by RTL
8118   // If loc->flags contains KMP_IDENT_ATOMIC_REDUCE, the atomic reduce method
8119   // can be selected by RTL
8120   // Finally, it's up to OpenMP RTL to make a decision on which method to select
8121   // among generated by PAROPT.
8122 
8123   PACKED_REDUCTION_METHOD_T retval;
8124 
8125   int team_size;
8126 
8127   KMP_DEBUG_ASSERT(loc); // it would be nice to test ( loc != 0 )
8128   KMP_DEBUG_ASSERT(lck); // it would be nice to test ( lck != 0 )
8129 
8130 #define FAST_REDUCTION_ATOMIC_METHOD_GENERATED                                 \
8131   ((loc->flags & (KMP_IDENT_ATOMIC_REDUCE)) == (KMP_IDENT_ATOMIC_REDUCE))
8132 #define FAST_REDUCTION_TREE_METHOD_GENERATED ((reduce_data) && (reduce_func))
8133 
8134   retval = critical_reduce_block;
8135 
8136   // another choice of getting a team size (with 1 dynamic deference) is slower
8137   team_size = __kmp_get_team_num_threads(global_tid);
8138   if (team_size == 1) {
8139 
8140     retval = empty_reduce_block;
8141 
8142   } else {
8143 
8144     int atomic_available = FAST_REDUCTION_ATOMIC_METHOD_GENERATED;
8145 
8146 #if KMP_ARCH_X86_64 || KMP_ARCH_PPC64 || KMP_ARCH_AARCH64 ||                   \
8147     KMP_ARCH_MIPS64 || KMP_ARCH_RISCV64
8148 
8149 #if KMP_OS_LINUX || KMP_OS_DRAGONFLY || KMP_OS_FREEBSD || KMP_OS_NETBSD ||     \
8150     KMP_OS_OPENBSD || KMP_OS_WINDOWS || KMP_OS_DARWIN || KMP_OS_HURD
8151 
8152     int teamsize_cutoff = 4;
8153 
8154 #if KMP_MIC_SUPPORTED
8155     if (__kmp_mic_type != non_mic) {
8156       teamsize_cutoff = 8;
8157     }
8158 #endif
8159     int tree_available = FAST_REDUCTION_TREE_METHOD_GENERATED;
8160     if (tree_available) {
8161       if (team_size <= teamsize_cutoff) {
8162         if (atomic_available) {
8163           retval = atomic_reduce_block;
8164         }
8165       } else {
8166         retval = TREE_REDUCE_BLOCK_WITH_REDUCTION_BARRIER;
8167       }
8168     } else if (atomic_available) {
8169       retval = atomic_reduce_block;
8170     }
8171 #else
8172 #error "Unknown or unsupported OS"
8173 #endif // KMP_OS_LINUX || KMP_OS_DRAGONFLY || KMP_OS_FREEBSD || KMP_OS_NETBSD ||
8174        // KMP_OS_OPENBSD || KMP_OS_WINDOWS || KMP_OS_DARWIN || KMP_OS_HURD
8175 
8176 #elif KMP_ARCH_X86 || KMP_ARCH_ARM || KMP_ARCH_AARCH || KMP_ARCH_MIPS
8177 
8178 #if KMP_OS_LINUX || KMP_OS_FREEBSD || KMP_OS_WINDOWS || KMP_OS_HURD
8179 
8180     // basic tuning
8181 
8182     if (atomic_available) {
8183       if (num_vars <= 2) { // && ( team_size <= 8 ) due to false-sharing ???
8184         retval = atomic_reduce_block;
8185       }
8186     } // otherwise: use critical section
8187 
8188 #elif KMP_OS_DARWIN
8189 
8190     int tree_available = FAST_REDUCTION_TREE_METHOD_GENERATED;
8191     if (atomic_available && (num_vars <= 3)) {
8192       retval = atomic_reduce_block;
8193     } else if (tree_available) {
8194       if ((reduce_size > (9 * sizeof(kmp_real64))) &&
8195           (reduce_size < (2000 * sizeof(kmp_real64)))) {
8196         retval = TREE_REDUCE_BLOCK_WITH_PLAIN_BARRIER;
8197       }
8198     } // otherwise: use critical section
8199 
8200 #else
8201 #error "Unknown or unsupported OS"
8202 #endif
8203 
8204 #else
8205 #error "Unknown or unsupported architecture"
8206 #endif
8207   }
8208 
8209   // KMP_FORCE_REDUCTION
8210 
8211   // If the team is serialized (team_size == 1), ignore the forced reduction
8212   // method and stay with the unsynchronized method (empty_reduce_block)
8213   if (__kmp_force_reduction_method != reduction_method_not_defined &&
8214       team_size != 1) {
8215 
8216     PACKED_REDUCTION_METHOD_T forced_retval = critical_reduce_block;
8217 
8218     int atomic_available, tree_available;
8219 
8220     switch ((forced_retval = __kmp_force_reduction_method)) {
8221     case critical_reduce_block:
8222       KMP_ASSERT(lck); // lck should be != 0
8223       break;
8224 
8225     case atomic_reduce_block:
8226       atomic_available = FAST_REDUCTION_ATOMIC_METHOD_GENERATED;
8227       if (!atomic_available) {
8228         KMP_WARNING(RedMethodNotSupported, "atomic");
8229         forced_retval = critical_reduce_block;
8230       }
8231       break;
8232 
8233     case tree_reduce_block:
8234       tree_available = FAST_REDUCTION_TREE_METHOD_GENERATED;
8235       if (!tree_available) {
8236         KMP_WARNING(RedMethodNotSupported, "tree");
8237         forced_retval = critical_reduce_block;
8238       } else {
8239 #if KMP_FAST_REDUCTION_BARRIER
8240         forced_retval = TREE_REDUCE_BLOCK_WITH_REDUCTION_BARRIER;
8241 #endif
8242       }
8243       break;
8244 
8245     default:
8246       KMP_ASSERT(0); // "unsupported method specified"
8247     }
8248 
8249     retval = forced_retval;
8250   }
8251 
8252   KA_TRACE(10, ("reduction method selected=%08x\n", retval));
8253 
8254 #undef FAST_REDUCTION_TREE_METHOD_GENERATED
8255 #undef FAST_REDUCTION_ATOMIC_METHOD_GENERATED
8256 
8257   return (retval);
8258 }
8259 // this function is for testing set/get/determine reduce method
8260 kmp_int32 __kmp_get_reduce_method(void) {
8261   return ((__kmp_entry_thread()->th.th_local.packed_reduction_method) >> 8);
8262 }
8263 
8264 // Soft pause sets up threads to ignore blocktime and just go to sleep.
8265 // Spin-wait code checks __kmp_pause_status and reacts accordingly.
8266 void __kmp_soft_pause() { __kmp_pause_status = kmp_soft_paused; }
8267 
8268 // Hard pause shuts down the runtime completely.  Resume happens naturally when
8269 // OpenMP is used subsequently.
8270 void __kmp_hard_pause() {
8271   __kmp_pause_status = kmp_hard_paused;
8272   __kmp_internal_end_thread(-1);
8273 }
8274 
8275 // Soft resume sets __kmp_pause_status, and wakes up all threads.
8276 void __kmp_resume_if_soft_paused() {
8277   if (__kmp_pause_status == kmp_soft_paused) {
8278     __kmp_pause_status = kmp_not_paused;
8279 
8280     for (int gtid = 1; gtid < __kmp_threads_capacity; ++gtid) {
8281       kmp_info_t *thread = __kmp_threads[gtid];
8282       if (thread) { // Wake it if sleeping
8283         kmp_flag_64 fl(&thread->th.th_bar[bs_forkjoin_barrier].bb.b_go, thread);
8284         if (fl.is_sleeping())
8285           fl.resume(gtid);
8286         else if (__kmp_try_suspend_mx(thread)) { // got suspend lock
8287           __kmp_unlock_suspend_mx(thread); // unlock it; it won't sleep
8288         } else { // thread holds the lock and may sleep soon
8289           do { // until either the thread sleeps, or we can get the lock
8290             if (fl.is_sleeping()) {
8291               fl.resume(gtid);
8292               break;
8293             } else if (__kmp_try_suspend_mx(thread)) {
8294               __kmp_unlock_suspend_mx(thread);
8295               break;
8296             }
8297           } while (1);
8298         }
8299       }
8300     }
8301   }
8302 }
8303 
8304 // This function is called via __kmpc_pause_resource. Returns 0 if successful.
8305 // TODO: add warning messages
8306 int __kmp_pause_resource(kmp_pause_status_t level) {
8307   if (level == kmp_not_paused) { // requesting resume
8308     if (__kmp_pause_status == kmp_not_paused) {
8309       // error message about runtime not being paused, so can't resume
8310       return 1;
8311     } else {
8312       KMP_DEBUG_ASSERT(__kmp_pause_status == kmp_soft_paused ||
8313                        __kmp_pause_status == kmp_hard_paused);
8314       __kmp_pause_status = kmp_not_paused;
8315       return 0;
8316     }
8317   } else if (level == kmp_soft_paused) { // requesting soft pause
8318     if (__kmp_pause_status != kmp_not_paused) {
8319       // error message about already being paused
8320       return 1;
8321     } else {
8322       __kmp_soft_pause();
8323       return 0;
8324     }
8325   } else if (level == kmp_hard_paused) { // requesting hard pause
8326     if (__kmp_pause_status != kmp_not_paused) {
8327       // error message about already being paused
8328       return 1;
8329     } else {
8330       __kmp_hard_pause();
8331       return 0;
8332     }
8333   } else {
8334     // error message about invalid level
8335     return 1;
8336   }
8337 }
8338 
8339 
8340 void __kmp_omp_display_env(int verbose) {
8341   __kmp_acquire_bootstrap_lock(&__kmp_initz_lock);
8342   if (__kmp_init_serial == 0)
8343     __kmp_do_serial_initialize();
8344   __kmp_display_env_impl(!verbose, verbose);
8345   __kmp_release_bootstrap_lock(&__kmp_initz_lock);
8346 }
8347