1 /*
2  * kmp_runtime.cpp -- KPTS runtime support library
3  */
4 
5 
6 //===----------------------------------------------------------------------===//
7 //
8 //                     The LLVM Compiler Infrastructure
9 //
10 // This file is dual licensed under the MIT and the University of Illinois Open
11 // Source Licenses. See LICENSE.txt for details.
12 //
13 //===----------------------------------------------------------------------===//
14 
15 
16 #include "kmp.h"
17 #include "kmp_affinity.h"
18 #include "kmp_atomic.h"
19 #include "kmp_environment.h"
20 #include "kmp_error.h"
21 #include "kmp_i18n.h"
22 #include "kmp_io.h"
23 #include "kmp_itt.h"
24 #include "kmp_settings.h"
25 #include "kmp_stats.h"
26 #include "kmp_str.h"
27 #include "kmp_wait_release.h"
28 #include "kmp_wrapper_getpid.h"
29 
30 #if OMPT_SUPPORT
31 #include "ompt-specific.h"
32 #endif
33 
34 /* these are temporary issues to be dealt with */
35 #define KMP_USE_PRCTL 0
36 
37 #if KMP_OS_WINDOWS
38 #include <process.h>
39 #endif
40 
41 #include "tsan_annotations.h"
42 
43 #if defined(KMP_GOMP_COMPAT)
44 char const __kmp_version_alt_comp[] =
45     KMP_VERSION_PREFIX "alternative compiler support: yes";
46 #endif /* defined(KMP_GOMP_COMPAT) */
47 
48 char const __kmp_version_omp_api[] = KMP_VERSION_PREFIX "API version: "
49 #if OMP_50_ENABLED
50                                                         "5.0 (201611)";
51 #elif OMP_45_ENABLED
52                                                         "4.5 (201511)";
53 #elif OMP_40_ENABLED
54                                                         "4.0 (201307)";
55 #else
56                                                         "3.1 (201107)";
57 #endif
58 
59 #ifdef KMP_DEBUG
60 char const __kmp_version_lock[] =
61     KMP_VERSION_PREFIX "lock type: run time selectable";
62 #endif /* KMP_DEBUG */
63 
64 #define KMP_MIN(x, y) ((x) < (y) ? (x) : (y))
65 
66 /* ------------------------------------------------------------------------ */
67 
68 kmp_info_t __kmp_monitor;
69 
70 /* Forward declarations */
71 
72 void __kmp_cleanup(void);
73 
74 static void __kmp_initialize_info(kmp_info_t *, kmp_team_t *, int tid,
75                                   int gtid);
76 static void __kmp_initialize_team(kmp_team_t *team, int new_nproc,
77                                   kmp_internal_control_t *new_icvs,
78                                   ident_t *loc);
79 #if OMP_40_ENABLED && KMP_AFFINITY_SUPPORTED
80 static void __kmp_partition_places(kmp_team_t *team,
81                                    int update_master_only = 0);
82 #endif
83 static void __kmp_do_serial_initialize(void);
84 void __kmp_fork_barrier(int gtid, int tid);
85 void __kmp_join_barrier(int gtid);
86 void __kmp_setup_icv_copy(kmp_team_t *team, int new_nproc,
87                           kmp_internal_control_t *new_icvs, ident_t *loc);
88 
89 #ifdef USE_LOAD_BALANCE
90 static int __kmp_load_balance_nproc(kmp_root_t *root, int set_nproc);
91 #endif
92 
93 static int __kmp_expand_threads(int nWish, int nNeed);
94 #if KMP_OS_WINDOWS
95 static int __kmp_unregister_root_other_thread(int gtid);
96 #endif
97 static void __kmp_unregister_library(void); // called by __kmp_internal_end()
98 static void __kmp_reap_thread(kmp_info_t *thread, int is_root);
99 static kmp_info_t *__kmp_thread_pool_insert_pt = NULL;
100 
101 /* Calculate the identifier of the current thread */
102 /* fast (and somewhat portable) way to get unique identifier of executing
103    thread. Returns KMP_GTID_DNE if we haven't been assigned a gtid. */
104 int __kmp_get_global_thread_id() {
105   int i;
106   kmp_info_t **other_threads;
107   size_t stack_data;
108   char *stack_addr;
109   size_t stack_size;
110   char *stack_base;
111 
112   KA_TRACE(
113       1000,
114       ("*** __kmp_get_global_thread_id: entering, nproc=%d  all_nproc=%d\n",
115        __kmp_nth, __kmp_all_nth));
116 
117   /* JPH - to handle the case where __kmpc_end(0) is called immediately prior to
118      a parallel region, made it return KMP_GTID_DNE to force serial_initialize
119      by caller. Had to handle KMP_GTID_DNE at all call-sites, or else guarantee
120      __kmp_init_gtid for this to work. */
121 
122   if (!TCR_4(__kmp_init_gtid))
123     return KMP_GTID_DNE;
124 
125 #ifdef KMP_TDATA_GTID
126   if (TCR_4(__kmp_gtid_mode) >= 3) {
127     KA_TRACE(1000, ("*** __kmp_get_global_thread_id: using TDATA\n"));
128     return __kmp_gtid;
129   }
130 #endif
131   if (TCR_4(__kmp_gtid_mode) >= 2) {
132     KA_TRACE(1000, ("*** __kmp_get_global_thread_id: using keyed TLS\n"));
133     return __kmp_gtid_get_specific();
134   }
135   KA_TRACE(1000, ("*** __kmp_get_global_thread_id: using internal alg.\n"));
136 
137   stack_addr = (char *)&stack_data;
138   other_threads = __kmp_threads;
139 
140   /* ATT: The code below is a source of potential bugs due to unsynchronized
141      access to __kmp_threads array. For example:
142      1. Current thread loads other_threads[i] to thr and checks it, it is
143         non-NULL.
144      2. Current thread is suspended by OS.
145      3. Another thread unregisters and finishes (debug versions of free()
146         may fill memory with something like 0xEF).
147      4. Current thread is resumed.
148      5. Current thread reads junk from *thr.
149      TODO: Fix it.  --ln  */
150 
151   for (i = 0; i < __kmp_threads_capacity; i++) {
152 
153     kmp_info_t *thr = (kmp_info_t *)TCR_SYNC_PTR(other_threads[i]);
154     if (!thr)
155       continue;
156 
157     stack_size = (size_t)TCR_PTR(thr->th.th_info.ds.ds_stacksize);
158     stack_base = (char *)TCR_PTR(thr->th.th_info.ds.ds_stackbase);
159 
160     /* stack grows down -- search through all of the active threads */
161 
162     if (stack_addr <= stack_base) {
163       size_t stack_diff = stack_base - stack_addr;
164 
165       if (stack_diff <= stack_size) {
166         /* The only way we can be closer than the allocated */
167         /* stack size is if we are running on this thread. */
168         KMP_DEBUG_ASSERT(__kmp_gtid_get_specific() == i);
169         return i;
170       }
171     }
172   }
173 
174   /* get specific to try and determine our gtid */
175   KA_TRACE(1000,
176            ("*** __kmp_get_global_thread_id: internal alg. failed to find "
177             "thread, using TLS\n"));
178   i = __kmp_gtid_get_specific();
179 
180   /*fprintf( stderr, "=== %d\n", i );  */ /* GROO */
181 
182   /* if we havn't been assigned a gtid, then return code */
183   if (i < 0)
184     return i;
185 
186   /* dynamically updated stack window for uber threads to avoid get_specific
187      call */
188   if (!TCR_4(other_threads[i]->th.th_info.ds.ds_stackgrow)) {
189     KMP_FATAL(StackOverflow, i);
190   }
191 
192   stack_base = (char *)other_threads[i]->th.th_info.ds.ds_stackbase;
193   if (stack_addr > stack_base) {
194     TCW_PTR(other_threads[i]->th.th_info.ds.ds_stackbase, stack_addr);
195     TCW_PTR(other_threads[i]->th.th_info.ds.ds_stacksize,
196             other_threads[i]->th.th_info.ds.ds_stacksize + stack_addr -
197                 stack_base);
198   } else {
199     TCW_PTR(other_threads[i]->th.th_info.ds.ds_stacksize,
200             stack_base - stack_addr);
201   }
202 
203   /* Reprint stack bounds for ubermaster since they have been refined */
204   if (__kmp_storage_map) {
205     char *stack_end = (char *)other_threads[i]->th.th_info.ds.ds_stackbase;
206     char *stack_beg = stack_end - other_threads[i]->th.th_info.ds.ds_stacksize;
207     __kmp_print_storage_map_gtid(i, stack_beg, stack_end,
208                                  other_threads[i]->th.th_info.ds.ds_stacksize,
209                                  "th_%d stack (refinement)", i);
210   }
211   return i;
212 }
213 
214 int __kmp_get_global_thread_id_reg() {
215   int gtid;
216 
217   if (!__kmp_init_serial) {
218     gtid = KMP_GTID_DNE;
219   } else
220 #ifdef KMP_TDATA_GTID
221       if (TCR_4(__kmp_gtid_mode) >= 3) {
222     KA_TRACE(1000, ("*** __kmp_get_global_thread_id_reg: using TDATA\n"));
223     gtid = __kmp_gtid;
224   } else
225 #endif
226       if (TCR_4(__kmp_gtid_mode) >= 2) {
227     KA_TRACE(1000, ("*** __kmp_get_global_thread_id_reg: using keyed TLS\n"));
228     gtid = __kmp_gtid_get_specific();
229   } else {
230     KA_TRACE(1000,
231              ("*** __kmp_get_global_thread_id_reg: using internal alg.\n"));
232     gtid = __kmp_get_global_thread_id();
233   }
234 
235   /* we must be a new uber master sibling thread */
236   if (gtid == KMP_GTID_DNE) {
237     KA_TRACE(10,
238              ("__kmp_get_global_thread_id_reg: Encountered new root thread. "
239               "Registering a new gtid.\n"));
240     __kmp_acquire_bootstrap_lock(&__kmp_initz_lock);
241     if (!__kmp_init_serial) {
242       __kmp_do_serial_initialize();
243       gtid = __kmp_gtid_get_specific();
244     } else {
245       gtid = __kmp_register_root(FALSE);
246     }
247     __kmp_release_bootstrap_lock(&__kmp_initz_lock);
248     /*__kmp_printf( "+++ %d\n", gtid ); */ /* GROO */
249   }
250 
251   KMP_DEBUG_ASSERT(gtid >= 0);
252 
253   return gtid;
254 }
255 
256 /* caller must hold forkjoin_lock */
257 void __kmp_check_stack_overlap(kmp_info_t *th) {
258   int f;
259   char *stack_beg = NULL;
260   char *stack_end = NULL;
261   int gtid;
262 
263   KA_TRACE(10, ("__kmp_check_stack_overlap: called\n"));
264   if (__kmp_storage_map) {
265     stack_end = (char *)th->th.th_info.ds.ds_stackbase;
266     stack_beg = stack_end - th->th.th_info.ds.ds_stacksize;
267 
268     gtid = __kmp_gtid_from_thread(th);
269 
270     if (gtid == KMP_GTID_MONITOR) {
271       __kmp_print_storage_map_gtid(
272           gtid, stack_beg, stack_end, th->th.th_info.ds.ds_stacksize,
273           "th_%s stack (%s)", "mon",
274           (th->th.th_info.ds.ds_stackgrow) ? "initial" : "actual");
275     } else {
276       __kmp_print_storage_map_gtid(
277           gtid, stack_beg, stack_end, th->th.th_info.ds.ds_stacksize,
278           "th_%d stack (%s)", gtid,
279           (th->th.th_info.ds.ds_stackgrow) ? "initial" : "actual");
280     }
281   }
282 
283   /* No point in checking ubermaster threads since they use refinement and
284    * cannot overlap */
285   gtid = __kmp_gtid_from_thread(th);
286   if (__kmp_env_checks == TRUE && !KMP_UBER_GTID(gtid)) {
287     KA_TRACE(10,
288              ("__kmp_check_stack_overlap: performing extensive checking\n"));
289     if (stack_beg == NULL) {
290       stack_end = (char *)th->th.th_info.ds.ds_stackbase;
291       stack_beg = stack_end - th->th.th_info.ds.ds_stacksize;
292     }
293 
294     for (f = 0; f < __kmp_threads_capacity; f++) {
295       kmp_info_t *f_th = (kmp_info_t *)TCR_SYNC_PTR(__kmp_threads[f]);
296 
297       if (f_th && f_th != th) {
298         char *other_stack_end =
299             (char *)TCR_PTR(f_th->th.th_info.ds.ds_stackbase);
300         char *other_stack_beg =
301             other_stack_end - (size_t)TCR_PTR(f_th->th.th_info.ds.ds_stacksize);
302         if ((stack_beg > other_stack_beg && stack_beg < other_stack_end) ||
303             (stack_end > other_stack_beg && stack_end < other_stack_end)) {
304 
305           /* Print the other stack values before the abort */
306           if (__kmp_storage_map)
307             __kmp_print_storage_map_gtid(
308                 -1, other_stack_beg, other_stack_end,
309                 (size_t)TCR_PTR(f_th->th.th_info.ds.ds_stacksize),
310                 "th_%d stack (overlapped)", __kmp_gtid_from_thread(f_th));
311 
312           __kmp_fatal(KMP_MSG(StackOverlap), KMP_HNT(ChangeStackLimit),
313                       __kmp_msg_null);
314         }
315       }
316     }
317   }
318   KA_TRACE(10, ("__kmp_check_stack_overlap: returning\n"));
319 }
320 
321 /* ------------------------------------------------------------------------ */
322 
323 void __kmp_infinite_loop(void) {
324   static int done = FALSE;
325 
326   while (!done) {
327     KMP_YIELD(1);
328   }
329 }
330 
331 #define MAX_MESSAGE 512
332 
333 void __kmp_print_storage_map_gtid(int gtid, void *p1, void *p2, size_t size,
334                                   char const *format, ...) {
335   char buffer[MAX_MESSAGE];
336   va_list ap;
337 
338   va_start(ap, format);
339   KMP_SNPRINTF(buffer, sizeof(buffer), "OMP storage map: %p %p%8lu %s\n", p1,
340                p2, (unsigned long)size, format);
341   __kmp_acquire_bootstrap_lock(&__kmp_stdio_lock);
342   __kmp_vprintf(kmp_err, buffer, ap);
343 #if KMP_PRINT_DATA_PLACEMENT
344   int node;
345   if (gtid >= 0) {
346     if (p1 <= p2 && (char *)p2 - (char *)p1 == size) {
347       if (__kmp_storage_map_verbose) {
348         node = __kmp_get_host_node(p1);
349         if (node < 0) /* doesn't work, so don't try this next time */
350           __kmp_storage_map_verbose = FALSE;
351         else {
352           char *last;
353           int lastNode;
354           int localProc = __kmp_get_cpu_from_gtid(gtid);
355 
356           const int page_size = KMP_GET_PAGE_SIZE();
357 
358           p1 = (void *)((size_t)p1 & ~((size_t)page_size - 1));
359           p2 = (void *)(((size_t)p2 - 1) & ~((size_t)page_size - 1));
360           if (localProc >= 0)
361             __kmp_printf_no_lock("  GTID %d localNode %d\n", gtid,
362                                  localProc >> 1);
363           else
364             __kmp_printf_no_lock("  GTID %d\n", gtid);
365 #if KMP_USE_PRCTL
366           /* The more elaborate format is disabled for now because of the prctl
367            * hanging bug. */
368           do {
369             last = p1;
370             lastNode = node;
371             /* This loop collates adjacent pages with the same host node. */
372             do {
373               (char *)p1 += page_size;
374             } while (p1 <= p2 && (node = __kmp_get_host_node(p1)) == lastNode);
375             __kmp_printf_no_lock("    %p-%p memNode %d\n", last, (char *)p1 - 1,
376                                  lastNode);
377           } while (p1 <= p2);
378 #else
379           __kmp_printf_no_lock("    %p-%p memNode %d\n", p1,
380                                (char *)p1 + (page_size - 1),
381                                __kmp_get_host_node(p1));
382           if (p1 < p2) {
383             __kmp_printf_no_lock("    %p-%p memNode %d\n", p2,
384                                  (char *)p2 + (page_size - 1),
385                                  __kmp_get_host_node(p2));
386           }
387 #endif
388         }
389       }
390     } else
391       __kmp_printf_no_lock("  %s\n", KMP_I18N_STR(StorageMapWarning));
392   }
393 #endif /* KMP_PRINT_DATA_PLACEMENT */
394   __kmp_release_bootstrap_lock(&__kmp_stdio_lock);
395 }
396 
397 void __kmp_warn(char const *format, ...) {
398   char buffer[MAX_MESSAGE];
399   va_list ap;
400 
401   if (__kmp_generate_warnings == kmp_warnings_off) {
402     return;
403   }
404 
405   va_start(ap, format);
406 
407   KMP_SNPRINTF(buffer, sizeof(buffer), "OMP warning: %s\n", format);
408   __kmp_acquire_bootstrap_lock(&__kmp_stdio_lock);
409   __kmp_vprintf(kmp_err, buffer, ap);
410   __kmp_release_bootstrap_lock(&__kmp_stdio_lock);
411 
412   va_end(ap);
413 }
414 
415 void __kmp_abort_process() {
416   // Later threads may stall here, but that's ok because abort() will kill them.
417   __kmp_acquire_bootstrap_lock(&__kmp_exit_lock);
418 
419   if (__kmp_debug_buf) {
420     __kmp_dump_debug_buffer();
421   }; // if
422 
423   if (KMP_OS_WINDOWS) {
424     // Let other threads know of abnormal termination and prevent deadlock
425     // if abort happened during library initialization or shutdown
426     __kmp_global.g.g_abort = SIGABRT;
427 
428     /* On Windows* OS by default abort() causes pop-up error box, which stalls
429        nightly testing. Unfortunately, we cannot reliably suppress pop-up error
430        boxes. _set_abort_behavior() works well, but this function is not
431        available in VS7 (this is not problem for DLL, but it is a problem for
432        static OpenMP RTL). SetErrorMode (and so, timelimit utility) does not
433        help, at least in some versions of MS C RTL.
434 
435        It seems following sequence is the only way to simulate abort() and
436        avoid pop-up error box. */
437     raise(SIGABRT);
438     _exit(3); // Just in case, if signal ignored, exit anyway.
439   } else {
440     abort();
441   }; // if
442 
443   __kmp_infinite_loop();
444   __kmp_release_bootstrap_lock(&__kmp_exit_lock);
445 
446 } // __kmp_abort_process
447 
448 void __kmp_abort_thread(void) {
449   // TODO: Eliminate g_abort global variable and this function.
450   // In case of abort just call abort(), it will kill all the threads.
451   __kmp_infinite_loop();
452 } // __kmp_abort_thread
453 
454 /* Print out the storage map for the major kmp_info_t thread data structures
455    that are allocated together. */
456 
457 static void __kmp_print_thread_storage_map(kmp_info_t *thr, int gtid) {
458   __kmp_print_storage_map_gtid(gtid, thr, thr + 1, sizeof(kmp_info_t), "th_%d",
459                                gtid);
460 
461   __kmp_print_storage_map_gtid(gtid, &thr->th.th_info, &thr->th.th_team,
462                                sizeof(kmp_desc_t), "th_%d.th_info", gtid);
463 
464   __kmp_print_storage_map_gtid(gtid, &thr->th.th_local, &thr->th.th_pri_head,
465                                sizeof(kmp_local_t), "th_%d.th_local", gtid);
466 
467   __kmp_print_storage_map_gtid(
468       gtid, &thr->th.th_bar[0], &thr->th.th_bar[bs_last_barrier],
469       sizeof(kmp_balign_t) * bs_last_barrier, "th_%d.th_bar", gtid);
470 
471   __kmp_print_storage_map_gtid(gtid, &thr->th.th_bar[bs_plain_barrier],
472                                &thr->th.th_bar[bs_plain_barrier + 1],
473                                sizeof(kmp_balign_t), "th_%d.th_bar[plain]",
474                                gtid);
475 
476   __kmp_print_storage_map_gtid(gtid, &thr->th.th_bar[bs_forkjoin_barrier],
477                                &thr->th.th_bar[bs_forkjoin_barrier + 1],
478                                sizeof(kmp_balign_t), "th_%d.th_bar[forkjoin]",
479                                gtid);
480 
481 #if KMP_FAST_REDUCTION_BARRIER
482   __kmp_print_storage_map_gtid(gtid, &thr->th.th_bar[bs_reduction_barrier],
483                                &thr->th.th_bar[bs_reduction_barrier + 1],
484                                sizeof(kmp_balign_t), "th_%d.th_bar[reduction]",
485                                gtid);
486 #endif // KMP_FAST_REDUCTION_BARRIER
487 }
488 
489 /* Print out the storage map for the major kmp_team_t team data structures
490    that are allocated together. */
491 
492 static void __kmp_print_team_storage_map(const char *header, kmp_team_t *team,
493                                          int team_id, int num_thr) {
494   int num_disp_buff = team->t.t_max_nproc > 1 ? __kmp_dispatch_num_buffers : 2;
495   __kmp_print_storage_map_gtid(-1, team, team + 1, sizeof(kmp_team_t), "%s_%d",
496                                header, team_id);
497 
498   __kmp_print_storage_map_gtid(-1, &team->t.t_bar[0],
499                                &team->t.t_bar[bs_last_barrier],
500                                sizeof(kmp_balign_team_t) * bs_last_barrier,
501                                "%s_%d.t_bar", header, team_id);
502 
503   __kmp_print_storage_map_gtid(-1, &team->t.t_bar[bs_plain_barrier],
504                                &team->t.t_bar[bs_plain_barrier + 1],
505                                sizeof(kmp_balign_team_t), "%s_%d.t_bar[plain]",
506                                header, team_id);
507 
508   __kmp_print_storage_map_gtid(-1, &team->t.t_bar[bs_forkjoin_barrier],
509                                &team->t.t_bar[bs_forkjoin_barrier + 1],
510                                sizeof(kmp_balign_team_t),
511                                "%s_%d.t_bar[forkjoin]", header, team_id);
512 
513 #if KMP_FAST_REDUCTION_BARRIER
514   __kmp_print_storage_map_gtid(-1, &team->t.t_bar[bs_reduction_barrier],
515                                &team->t.t_bar[bs_reduction_barrier + 1],
516                                sizeof(kmp_balign_team_t),
517                                "%s_%d.t_bar[reduction]", header, team_id);
518 #endif // KMP_FAST_REDUCTION_BARRIER
519 
520   __kmp_print_storage_map_gtid(
521       -1, &team->t.t_dispatch[0], &team->t.t_dispatch[num_thr],
522       sizeof(kmp_disp_t) * num_thr, "%s_%d.t_dispatch", header, team_id);
523 
524   __kmp_print_storage_map_gtid(
525       -1, &team->t.t_threads[0], &team->t.t_threads[num_thr],
526       sizeof(kmp_info_t *) * num_thr, "%s_%d.t_threads", header, team_id);
527 
528   __kmp_print_storage_map_gtid(-1, &team->t.t_disp_buffer[0],
529                                &team->t.t_disp_buffer[num_disp_buff],
530                                sizeof(dispatch_shared_info_t) * num_disp_buff,
531                                "%s_%d.t_disp_buffer", header, team_id);
532 
533   __kmp_print_storage_map_gtid(-1, &team->t.t_taskq, &team->t.t_copypriv_data,
534                                sizeof(kmp_taskq_t), "%s_%d.t_taskq", header,
535                                team_id);
536 }
537 
538 static void __kmp_init_allocator() {}
539 static void __kmp_fini_allocator() {}
540 
541 /* ------------------------------------------------------------------------ */
542 
543 #ifdef KMP_DYNAMIC_LIB
544 #if KMP_OS_WINDOWS
545 
546 static void __kmp_reset_lock(kmp_bootstrap_lock_t *lck) {
547   // TODO: Change to __kmp_break_bootstrap_lock().
548   __kmp_init_bootstrap_lock(lck); // make the lock released
549 }
550 
551 static void __kmp_reset_locks_on_process_detach(int gtid_req) {
552   int i;
553   int thread_count;
554 
555   // PROCESS_DETACH is expected to be called by a thread that executes
556   // ProcessExit() or FreeLibrary(). OS terminates other threads (except the one
557   // calling ProcessExit or FreeLibrary). So, it might be safe to access the
558   // __kmp_threads[] without taking the forkjoin_lock. However, in fact, some
559   // threads can be still alive here, although being about to be terminated. The
560   // threads in the array with ds_thread==0 are most suspicious. Actually, it
561   // can be not safe to access the __kmp_threads[].
562 
563   // TODO: does it make sense to check __kmp_roots[] ?
564 
565   // Let's check that there are no other alive threads registered with the OMP
566   // lib.
567   while (1) {
568     thread_count = 0;
569     for (i = 0; i < __kmp_threads_capacity; ++i) {
570       if (!__kmp_threads)
571         continue;
572       kmp_info_t *th = __kmp_threads[i];
573       if (th == NULL)
574         continue;
575       int gtid = th->th.th_info.ds.ds_gtid;
576       if (gtid == gtid_req)
577         continue;
578       if (gtid < 0)
579         continue;
580       DWORD exit_val;
581       int alive = __kmp_is_thread_alive(th, &exit_val);
582       if (alive) {
583         ++thread_count;
584       }
585     }
586     if (thread_count == 0)
587       break; // success
588   }
589 
590   // Assume that I'm alone. Now it might be safe to check and reset locks.
591   // __kmp_forkjoin_lock and __kmp_stdio_lock are expected to be reset.
592   __kmp_reset_lock(&__kmp_forkjoin_lock);
593 #ifdef KMP_DEBUG
594   __kmp_reset_lock(&__kmp_stdio_lock);
595 #endif // KMP_DEBUG
596 }
597 
598 BOOL WINAPI DllMain(HINSTANCE hInstDLL, DWORD fdwReason, LPVOID lpReserved) {
599   //__kmp_acquire_bootstrap_lock( &__kmp_initz_lock );
600 
601   switch (fdwReason) {
602 
603   case DLL_PROCESS_ATTACH:
604     KA_TRACE(10, ("DllMain: PROCESS_ATTACH\n"));
605 
606     return TRUE;
607 
608   case DLL_PROCESS_DETACH:
609     KA_TRACE(10, ("DllMain: PROCESS_DETACH T#%d\n", __kmp_gtid_get_specific()));
610 
611     if (lpReserved != NULL) {
612       // lpReserved is used for telling the difference:
613       //   lpReserved == NULL when FreeLibrary() was called,
614       //   lpReserved != NULL when the process terminates.
615       // When FreeLibrary() is called, worker threads remain alive. So they will
616       // release the forkjoin lock by themselves. When the process terminates,
617       // worker threads disappear triggering the problem of unreleased forkjoin
618       // lock as described below.
619 
620       // A worker thread can take the forkjoin lock. The problem comes up if
621       // that worker thread becomes dead before it releases the forkjoin lock.
622       // The forkjoin lock remains taken, while the thread executing
623       // DllMain()->PROCESS_DETACH->__kmp_internal_end_library() below will try
624       // to take the forkjoin lock and will always fail, so that the application
625       // will never finish [normally]. This scenario is possible if
626       // __kmpc_end() has not been executed. It looks like it's not a corner
627       // case, but common cases:
628       // - the main function was compiled by an alternative compiler;
629       // - the main function was compiled by icl but without /Qopenmp
630       //   (application with plugins);
631       // - application terminates by calling C exit(), Fortran CALL EXIT() or
632       //   Fortran STOP.
633       // - alive foreign thread prevented __kmpc_end from doing cleanup.
634       //
635       // This is a hack to work around the problem.
636       // TODO: !!! figure out something better.
637       __kmp_reset_locks_on_process_detach(__kmp_gtid_get_specific());
638     }
639 
640     __kmp_internal_end_library(__kmp_gtid_get_specific());
641 
642     return TRUE;
643 
644   case DLL_THREAD_ATTACH:
645     KA_TRACE(10, ("DllMain: THREAD_ATTACH\n"));
646 
647     /* if we want to register new siblings all the time here call
648      * __kmp_get_gtid(); */
649     return TRUE;
650 
651   case DLL_THREAD_DETACH:
652     KA_TRACE(10, ("DllMain: THREAD_DETACH T#%d\n", __kmp_gtid_get_specific()));
653 
654     __kmp_internal_end_thread(__kmp_gtid_get_specific());
655     return TRUE;
656   }
657 
658   return TRUE;
659 }
660 
661 #endif /* KMP_OS_WINDOWS */
662 #endif /* KMP_DYNAMIC_LIB */
663 
664 /* Change the library type to "status" and return the old type */
665 /* called from within initialization routines where __kmp_initz_lock is held */
666 int __kmp_change_library(int status) {
667   int old_status;
668 
669   old_status = __kmp_yield_init &
670                1; // check whether KMP_LIBRARY=throughput (even init count)
671 
672   if (status) {
673     __kmp_yield_init |= 1; // throughput => turnaround (odd init count)
674   } else {
675     __kmp_yield_init &= ~1; // turnaround => throughput (even init count)
676   }
677 
678   return old_status; // return previous setting of whether
679   // KMP_LIBRARY=throughput
680 }
681 
682 /* __kmp_parallel_deo -- Wait until it's our turn. */
683 void __kmp_parallel_deo(int *gtid_ref, int *cid_ref, ident_t *loc_ref) {
684   int gtid = *gtid_ref;
685 #ifdef BUILD_PARALLEL_ORDERED
686   kmp_team_t *team = __kmp_team_from_gtid(gtid);
687 #endif /* BUILD_PARALLEL_ORDERED */
688 
689   if (__kmp_env_consistency_check) {
690     if (__kmp_threads[gtid]->th.th_root->r.r_active)
691 #if KMP_USE_DYNAMIC_LOCK
692       __kmp_push_sync(gtid, ct_ordered_in_parallel, loc_ref, NULL, 0);
693 #else
694       __kmp_push_sync(gtid, ct_ordered_in_parallel, loc_ref, NULL);
695 #endif
696   }
697 #ifdef BUILD_PARALLEL_ORDERED
698   if (!team->t.t_serialized) {
699     KMP_MB();
700     KMP_WAIT_YIELD(&team->t.t_ordered.dt.t_value, __kmp_tid_from_gtid(gtid),
701                    KMP_EQ, NULL);
702     KMP_MB();
703   }
704 #endif /* BUILD_PARALLEL_ORDERED */
705 }
706 
707 /* __kmp_parallel_dxo -- Signal the next task. */
708 void __kmp_parallel_dxo(int *gtid_ref, int *cid_ref, ident_t *loc_ref) {
709   int gtid = *gtid_ref;
710 #ifdef BUILD_PARALLEL_ORDERED
711   int tid = __kmp_tid_from_gtid(gtid);
712   kmp_team_t *team = __kmp_team_from_gtid(gtid);
713 #endif /* BUILD_PARALLEL_ORDERED */
714 
715   if (__kmp_env_consistency_check) {
716     if (__kmp_threads[gtid]->th.th_root->r.r_active)
717       __kmp_pop_sync(gtid, ct_ordered_in_parallel, loc_ref);
718   }
719 #ifdef BUILD_PARALLEL_ORDERED
720   if (!team->t.t_serialized) {
721     KMP_MB(); /* Flush all pending memory write invalidates.  */
722 
723     /* use the tid of the next thread in this team */
724     /* TODO replace with general release procedure */
725     team->t.t_ordered.dt.t_value = ((tid + 1) % team->t.t_nproc);
726 
727 #if OMPT_SUPPORT && OMPT_BLAME
728     if (ompt_enabled &&
729         ompt_callbacks.ompt_callback(ompt_event_release_ordered)) {
730       /* accept blame for "ordered" waiting */
731       kmp_info_t *this_thread = __kmp_threads[gtid];
732       ompt_callbacks.ompt_callback(ompt_event_release_ordered)(
733           this_thread->th.ompt_thread_info.wait_id);
734     }
735 #endif
736 
737     KMP_MB(); /* Flush all pending memory write invalidates.  */
738   }
739 #endif /* BUILD_PARALLEL_ORDERED */
740 }
741 
742 /* ------------------------------------------------------------------------ */
743 /* The BARRIER for a SINGLE process section is always explicit   */
744 
745 int __kmp_enter_single(int gtid, ident_t *id_ref, int push_ws) {
746   int status;
747   kmp_info_t *th;
748   kmp_team_t *team;
749 
750   if (!TCR_4(__kmp_init_parallel))
751     __kmp_parallel_initialize();
752 
753   th = __kmp_threads[gtid];
754   team = th->th.th_team;
755   status = 0;
756 
757   th->th.th_ident = id_ref;
758 
759   if (team->t.t_serialized) {
760     status = 1;
761   } else {
762     kmp_int32 old_this = th->th.th_local.this_construct;
763 
764     ++th->th.th_local.this_construct;
765     /* try to set team count to thread count--success means thread got the
766        single block */
767     /* TODO: Should this be acquire or release? */
768     if (team->t.t_construct == old_this) {
769       status = KMP_COMPARE_AND_STORE_ACQ32(&team->t.t_construct, old_this,
770                                            th->th.th_local.this_construct);
771     }
772 #if USE_ITT_BUILD
773     if (__itt_metadata_add_ptr && __kmp_forkjoin_frames_mode == 3 &&
774         KMP_MASTER_GTID(gtid) &&
775 #if OMP_40_ENABLED
776         th->th.th_teams_microtask == NULL &&
777 #endif
778         team->t.t_active_level ==
779             1) { // Only report metadata by master of active team at level 1
780       __kmp_itt_metadata_single(id_ref);
781     }
782 #endif /* USE_ITT_BUILD */
783   }
784 
785   if (__kmp_env_consistency_check) {
786     if (status && push_ws) {
787       __kmp_push_workshare(gtid, ct_psingle, id_ref);
788     } else {
789       __kmp_check_workshare(gtid, ct_psingle, id_ref);
790     }
791   }
792 #if USE_ITT_BUILD
793   if (status) {
794     __kmp_itt_single_start(gtid);
795   }
796 #endif /* USE_ITT_BUILD */
797   return status;
798 }
799 
800 void __kmp_exit_single(int gtid) {
801 #if USE_ITT_BUILD
802   __kmp_itt_single_end(gtid);
803 #endif /* USE_ITT_BUILD */
804   if (__kmp_env_consistency_check)
805     __kmp_pop_workshare(gtid, ct_psingle, NULL);
806 }
807 
808 /* determine if we can go parallel or must use a serialized parallel region and
809  * how many threads we can use
810  * set_nproc is the number of threads requested for the team
811  * returns 0 if we should serialize or only use one thread,
812  * otherwise the number of threads to use
813  * The forkjoin lock is held by the caller. */
814 static int __kmp_reserve_threads(kmp_root_t *root, kmp_team_t *parent_team,
815                                  int master_tid, int set_nthreads
816 #if OMP_40_ENABLED
817                                  ,
818                                  int enter_teams
819 #endif /* OMP_40_ENABLED */
820                                  ) {
821   int capacity;
822   int new_nthreads;
823   KMP_DEBUG_ASSERT(__kmp_init_serial);
824   KMP_DEBUG_ASSERT(root && parent_team);
825 
826   // If dyn-var is set, dynamically adjust the number of desired threads,
827   // according to the method specified by dynamic_mode.
828   new_nthreads = set_nthreads;
829   if (!get__dynamic_2(parent_team, master_tid)) {
830     ;
831   }
832 #ifdef USE_LOAD_BALANCE
833   else if (__kmp_global.g.g_dynamic_mode == dynamic_load_balance) {
834     new_nthreads = __kmp_load_balance_nproc(root, set_nthreads);
835     if (new_nthreads == 1) {
836       KC_TRACE(10, ("__kmp_reserve_threads: T#%d load balance reduced "
837                     "reservation to 1 thread\n",
838                     master_tid));
839       return 1;
840     }
841     if (new_nthreads < set_nthreads) {
842       KC_TRACE(10, ("__kmp_reserve_threads: T#%d load balance reduced "
843                     "reservation to %d threads\n",
844                     master_tid, new_nthreads));
845     }
846   }
847 #endif /* USE_LOAD_BALANCE */
848   else if (__kmp_global.g.g_dynamic_mode == dynamic_thread_limit) {
849     new_nthreads = __kmp_avail_proc - __kmp_nth +
850                    (root->r.r_active ? 1 : root->r.r_hot_team->t.t_nproc);
851     if (new_nthreads <= 1) {
852       KC_TRACE(10, ("__kmp_reserve_threads: T#%d thread limit reduced "
853                     "reservation to 1 thread\n",
854                     master_tid));
855       return 1;
856     }
857     if (new_nthreads < set_nthreads) {
858       KC_TRACE(10, ("__kmp_reserve_threads: T#%d thread limit reduced "
859                     "reservation to %d threads\n",
860                     master_tid, new_nthreads));
861     } else {
862       new_nthreads = set_nthreads;
863     }
864   } else if (__kmp_global.g.g_dynamic_mode == dynamic_random) {
865     if (set_nthreads > 2) {
866       new_nthreads = __kmp_get_random(parent_team->t.t_threads[master_tid]);
867       new_nthreads = (new_nthreads % set_nthreads) + 1;
868       if (new_nthreads == 1) {
869         KC_TRACE(10, ("__kmp_reserve_threads: T#%d dynamic random reduced "
870                       "reservation to 1 thread\n",
871                       master_tid));
872         return 1;
873       }
874       if (new_nthreads < set_nthreads) {
875         KC_TRACE(10, ("__kmp_reserve_threads: T#%d dynamic random reduced "
876                       "reservation to %d threads\n",
877                       master_tid, new_nthreads));
878       }
879     }
880   } else {
881     KMP_ASSERT(0);
882   }
883 
884   // Respect KMP_ALL_THREADS/KMP_DEVICE_THREAD_LIMIT.
885   if (__kmp_nth + new_nthreads -
886           (root->r.r_active ? 1 : root->r.r_hot_team->t.t_nproc) >
887       __kmp_max_nth) {
888     int tl_nthreads = __kmp_max_nth - __kmp_nth +
889                       (root->r.r_active ? 1 : root->r.r_hot_team->t.t_nproc);
890     if (tl_nthreads <= 0) {
891       tl_nthreads = 1;
892     }
893 
894     // If dyn-var is false, emit a 1-time warning.
895     if (!get__dynamic_2(parent_team, master_tid) && (!__kmp_reserve_warn)) {
896       __kmp_reserve_warn = 1;
897       __kmp_msg(kmp_ms_warning,
898                 KMP_MSG(CantFormThrTeam, set_nthreads, tl_nthreads),
899                 KMP_HNT(Unset_ALL_THREADS), __kmp_msg_null);
900     }
901     if (tl_nthreads == 1) {
902       KC_TRACE(10, ("__kmp_reserve_threads: T#%d KMP_DEVICE_THREAD_LIMIT "
903                     "reduced reservation to 1 thread\n",
904                     master_tid));
905       return 1;
906     }
907     KC_TRACE(10, ("__kmp_reserve_threads: T#%d KMP_DEVICE_THREAD_LIMIT reduced "
908                   "reservation to %d threads\n",
909                   master_tid, tl_nthreads));
910     new_nthreads = tl_nthreads;
911   }
912 
913   // Respect OMP_THREAD_LIMIT
914   if (root->r.r_cg_nthreads + new_nthreads -
915           (root->r.r_active ? 1 : root->r.r_hot_team->t.t_nproc) >
916       __kmp_cg_max_nth) {
917     int tl_nthreads = __kmp_cg_max_nth - root->r.r_cg_nthreads +
918                       (root->r.r_active ? 1 : root->r.r_hot_team->t.t_nproc);
919     if (tl_nthreads <= 0) {
920       tl_nthreads = 1;
921     }
922 
923     // If dyn-var is false, emit a 1-time warning.
924     if (!get__dynamic_2(parent_team, master_tid) && (!__kmp_reserve_warn)) {
925       __kmp_reserve_warn = 1;
926       __kmp_msg(kmp_ms_warning,
927                 KMP_MSG(CantFormThrTeam, set_nthreads, tl_nthreads),
928                 KMP_HNT(Unset_ALL_THREADS), __kmp_msg_null);
929     }
930     if (tl_nthreads == 1) {
931       KC_TRACE(10, ("__kmp_reserve_threads: T#%d OMP_THREAD_LIMIT "
932                     "reduced reservation to 1 thread\n",
933                     master_tid));
934       return 1;
935     }
936     KC_TRACE(10, ("__kmp_reserve_threads: T#%d OMP_THREAD_LIMIT reduced "
937                   "reservation to %d threads\n",
938                   master_tid, tl_nthreads));
939     new_nthreads = tl_nthreads;
940   }
941 
942   // Check if the threads array is large enough, or needs expanding.
943   // See comment in __kmp_register_root() about the adjustment if
944   // __kmp_threads[0] == NULL.
945   capacity = __kmp_threads_capacity;
946   if (TCR_PTR(__kmp_threads[0]) == NULL) {
947     --capacity;
948   }
949   if (__kmp_nth + new_nthreads -
950           (root->r.r_active ? 1 : root->r.r_hot_team->t.t_nproc) >
951       capacity) {
952     // Expand the threads array.
953     int slotsRequired = __kmp_nth + new_nthreads -
954                         (root->r.r_active ? 1 : root->r.r_hot_team->t.t_nproc) -
955                         capacity;
956     int slotsAdded = __kmp_expand_threads(slotsRequired, slotsRequired);
957     if (slotsAdded < slotsRequired) {
958       // The threads array was not expanded enough.
959       new_nthreads -= (slotsRequired - slotsAdded);
960       KMP_ASSERT(new_nthreads >= 1);
961 
962       // If dyn-var is false, emit a 1-time warning.
963       if (!get__dynamic_2(parent_team, master_tid) && (!__kmp_reserve_warn)) {
964         __kmp_reserve_warn = 1;
965         if (__kmp_tp_cached) {
966           __kmp_msg(kmp_ms_warning,
967                     KMP_MSG(CantFormThrTeam, set_nthreads, new_nthreads),
968                     KMP_HNT(Set_ALL_THREADPRIVATE, __kmp_tp_capacity),
969                     KMP_HNT(PossibleSystemLimitOnThreads), __kmp_msg_null);
970         } else {
971           __kmp_msg(kmp_ms_warning,
972                     KMP_MSG(CantFormThrTeam, set_nthreads, new_nthreads),
973                     KMP_HNT(SystemLimitOnThreads), __kmp_msg_null);
974         }
975       }
976     }
977   }
978 
979 #ifdef KMP_DEBUG
980   if (new_nthreads == 1) {
981     KC_TRACE(10,
982              ("__kmp_reserve_threads: T#%d serializing team after reclaiming "
983               "dead roots and rechecking; requested %d threads\n",
984               __kmp_get_gtid(), set_nthreads));
985   } else {
986     KC_TRACE(10, ("__kmp_reserve_threads: T#%d allocating %d threads; requested"
987                   " %d threads\n",
988                   __kmp_get_gtid(), new_nthreads, set_nthreads));
989   }
990 #endif // KMP_DEBUG
991   return new_nthreads;
992 }
993 
994 /* Allocate threads from the thread pool and assign them to the new team. We are
995    assured that there are enough threads available, because we checked on that
996    earlier within critical section forkjoin */
997 static void __kmp_fork_team_threads(kmp_root_t *root, kmp_team_t *team,
998                                     kmp_info_t *master_th, int master_gtid) {
999   int i;
1000   int use_hot_team;
1001 
1002   KA_TRACE(10, ("__kmp_fork_team_threads: new_nprocs = %d\n", team->t.t_nproc));
1003   KMP_DEBUG_ASSERT(master_gtid == __kmp_get_gtid());
1004   KMP_MB();
1005 
1006   /* first, let's setup the master thread */
1007   master_th->th.th_info.ds.ds_tid = 0;
1008   master_th->th.th_team = team;
1009   master_th->th.th_team_nproc = team->t.t_nproc;
1010   master_th->th.th_team_master = master_th;
1011   master_th->th.th_team_serialized = FALSE;
1012   master_th->th.th_dispatch = &team->t.t_dispatch[0];
1013 
1014 /* make sure we are not the optimized hot team */
1015 #if KMP_NESTED_HOT_TEAMS
1016   use_hot_team = 0;
1017   kmp_hot_team_ptr_t *hot_teams = master_th->th.th_hot_teams;
1018   if (hot_teams) { // hot teams array is not allocated if
1019     // KMP_HOT_TEAMS_MAX_LEVEL=0
1020     int level = team->t.t_active_level - 1; // index in array of hot teams
1021     if (master_th->th.th_teams_microtask) { // are we inside the teams?
1022       if (master_th->th.th_teams_size.nteams > 1) {
1023         ++level; // level was not increased in teams construct for
1024         // team_of_masters
1025       }
1026       if (team->t.t_pkfn != (microtask_t)__kmp_teams_master &&
1027           master_th->th.th_teams_level == team->t.t_level) {
1028         ++level; // level was not increased in teams construct for
1029         // team_of_workers before the parallel
1030       } // team->t.t_level will be increased inside parallel
1031     }
1032     if (level < __kmp_hot_teams_max_level) {
1033       if (hot_teams[level].hot_team) {
1034         // hot team has already been allocated for given level
1035         KMP_DEBUG_ASSERT(hot_teams[level].hot_team == team);
1036         use_hot_team = 1; // the team is ready to use
1037       } else {
1038         use_hot_team = 0; // AC: threads are not allocated yet
1039         hot_teams[level].hot_team = team; // remember new hot team
1040         hot_teams[level].hot_team_nth = team->t.t_nproc;
1041       }
1042     } else {
1043       use_hot_team = 0;
1044     }
1045   }
1046 #else
1047   use_hot_team = team == root->r.r_hot_team;
1048 #endif
1049   if (!use_hot_team) {
1050 
1051     /* install the master thread */
1052     team->t.t_threads[0] = master_th;
1053     __kmp_initialize_info(master_th, team, 0, master_gtid);
1054 
1055     /* now, install the worker threads */
1056     for (i = 1; i < team->t.t_nproc; i++) {
1057 
1058       /* fork or reallocate a new thread and install it in team */
1059       kmp_info_t *thr = __kmp_allocate_thread(root, team, i);
1060       team->t.t_threads[i] = thr;
1061       KMP_DEBUG_ASSERT(thr);
1062       KMP_DEBUG_ASSERT(thr->th.th_team == team);
1063       /* align team and thread arrived states */
1064       KA_TRACE(20, ("__kmp_fork_team_threads: T#%d(%d:%d) init arrived "
1065                     "T#%d(%d:%d) join =%llu, plain=%llu\n",
1066                     __kmp_gtid_from_tid(0, team), team->t.t_id, 0,
1067                     __kmp_gtid_from_tid(i, team), team->t.t_id, i,
1068                     team->t.t_bar[bs_forkjoin_barrier].b_arrived,
1069                     team->t.t_bar[bs_plain_barrier].b_arrived));
1070 #if OMP_40_ENABLED
1071       thr->th.th_teams_microtask = master_th->th.th_teams_microtask;
1072       thr->th.th_teams_level = master_th->th.th_teams_level;
1073       thr->th.th_teams_size = master_th->th.th_teams_size;
1074 #endif
1075       { // Initialize threads' barrier data.
1076         int b;
1077         kmp_balign_t *balign = team->t.t_threads[i]->th.th_bar;
1078         for (b = 0; b < bs_last_barrier; ++b) {
1079           balign[b].bb.b_arrived = team->t.t_bar[b].b_arrived;
1080           KMP_DEBUG_ASSERT(balign[b].bb.wait_flag != KMP_BARRIER_PARENT_FLAG);
1081 #if USE_DEBUGGER
1082           balign[b].bb.b_worker_arrived = team->t.t_bar[b].b_team_arrived;
1083 #endif
1084         }; // for b
1085       }
1086     }
1087 
1088 #if OMP_40_ENABLED && KMP_AFFINITY_SUPPORTED
1089     __kmp_partition_places(team);
1090 #endif
1091   }
1092 
1093   KMP_MB();
1094 }
1095 
1096 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
1097 // Propagate any changes to the floating point control registers out to the team
1098 // We try to avoid unnecessary writes to the relevant cache line in the team
1099 // structure, so we don't make changes unless they are needed.
1100 inline static void propagateFPControl(kmp_team_t *team) {
1101   if (__kmp_inherit_fp_control) {
1102     kmp_int16 x87_fpu_control_word;
1103     kmp_uint32 mxcsr;
1104 
1105     // Get master values of FPU control flags (both X87 and vector)
1106     __kmp_store_x87_fpu_control_word(&x87_fpu_control_word);
1107     __kmp_store_mxcsr(&mxcsr);
1108     mxcsr &= KMP_X86_MXCSR_MASK;
1109 
1110 // There is no point looking at t_fp_control_saved here.
1111 // If it is TRUE, we still have to update the values if they are different from
1112 // those we now have.
1113 // If it is FALSE we didn't save anything yet, but our objective is the same. We
1114 // have to ensure that the values in the team are the same as those we have.
1115 // So, this code achieves what we need whether or not t_fp_control_saved is
1116 // true. By checking whether the value needs updating we avoid unnecessary
1117 // writes that would put the cache-line into a written state, causing all
1118 // threads in the team to have to read it again.
1119     KMP_CHECK_UPDATE(team->t.t_x87_fpu_control_word, x87_fpu_control_word);
1120     KMP_CHECK_UPDATE(team->t.t_mxcsr, mxcsr);
1121     // Although we don't use this value, other code in the runtime wants to know
1122     // whether it should restore them. So we must ensure it is correct.
1123     KMP_CHECK_UPDATE(team->t.t_fp_control_saved, TRUE);
1124   } else {
1125     // Similarly here. Don't write to this cache-line in the team structure
1126     // unless we have to.
1127     KMP_CHECK_UPDATE(team->t.t_fp_control_saved, FALSE);
1128   }
1129 }
1130 
1131 // Do the opposite, setting the hardware registers to the updated values from
1132 // the team.
1133 inline static void updateHWFPControl(kmp_team_t *team) {
1134   if (__kmp_inherit_fp_control && team->t.t_fp_control_saved) {
1135     // Only reset the fp control regs if they have been changed in the team.
1136     // the parallel region that we are exiting.
1137     kmp_int16 x87_fpu_control_word;
1138     kmp_uint32 mxcsr;
1139     __kmp_store_x87_fpu_control_word(&x87_fpu_control_word);
1140     __kmp_store_mxcsr(&mxcsr);
1141     mxcsr &= KMP_X86_MXCSR_MASK;
1142 
1143     if (team->t.t_x87_fpu_control_word != x87_fpu_control_word) {
1144       __kmp_clear_x87_fpu_status_word();
1145       __kmp_load_x87_fpu_control_word(&team->t.t_x87_fpu_control_word);
1146     }
1147 
1148     if (team->t.t_mxcsr != mxcsr) {
1149       __kmp_load_mxcsr(&team->t.t_mxcsr);
1150     }
1151   }
1152 }
1153 #else
1154 #define propagateFPControl(x) ((void)0)
1155 #define updateHWFPControl(x) ((void)0)
1156 #endif /* KMP_ARCH_X86 || KMP_ARCH_X86_64 */
1157 
1158 static void __kmp_alloc_argv_entries(int argc, kmp_team_t *team,
1159                                      int realloc); // forward declaration
1160 
1161 /* Run a parallel region that has been serialized, so runs only in a team of the
1162    single master thread. */
1163 void __kmp_serialized_parallel(ident_t *loc, kmp_int32 global_tid) {
1164   kmp_info_t *this_thr;
1165   kmp_team_t *serial_team;
1166 
1167   KC_TRACE(10, ("__kmpc_serialized_parallel: called by T#%d\n", global_tid));
1168 
1169   /* Skip all this code for autopar serialized loops since it results in
1170      unacceptable overhead */
1171   if (loc != NULL && (loc->flags & KMP_IDENT_AUTOPAR))
1172     return;
1173 
1174   if (!TCR_4(__kmp_init_parallel))
1175     __kmp_parallel_initialize();
1176 
1177   this_thr = __kmp_threads[global_tid];
1178   serial_team = this_thr->th.th_serial_team;
1179 
1180   /* utilize the serialized team held by this thread */
1181   KMP_DEBUG_ASSERT(serial_team);
1182   KMP_MB();
1183 
1184   if (__kmp_tasking_mode != tskm_immediate_exec) {
1185     KMP_DEBUG_ASSERT(
1186         this_thr->th.th_task_team ==
1187         this_thr->th.th_team->t.t_task_team[this_thr->th.th_task_state]);
1188     KMP_DEBUG_ASSERT(serial_team->t.t_task_team[this_thr->th.th_task_state] ==
1189                      NULL);
1190     KA_TRACE(20, ("__kmpc_serialized_parallel: T#%d pushing task_team %p / "
1191                   "team %p, new task_team = NULL\n",
1192                   global_tid, this_thr->th.th_task_team, this_thr->th.th_team));
1193     this_thr->th.th_task_team = NULL;
1194   }
1195 
1196 #if OMP_40_ENABLED
1197   kmp_proc_bind_t proc_bind = this_thr->th.th_set_proc_bind;
1198   if (this_thr->th.th_current_task->td_icvs.proc_bind == proc_bind_false) {
1199     proc_bind = proc_bind_false;
1200   } else if (proc_bind == proc_bind_default) {
1201     // No proc_bind clause was specified, so use the current value
1202     // of proc-bind-var for this parallel region.
1203     proc_bind = this_thr->th.th_current_task->td_icvs.proc_bind;
1204   }
1205   // Reset for next parallel region
1206   this_thr->th.th_set_proc_bind = proc_bind_default;
1207 #endif /* OMP_40_ENABLED */
1208 
1209   if (this_thr->th.th_team != serial_team) {
1210     // Nested level will be an index in the nested nthreads array
1211     int level = this_thr->th.th_team->t.t_level;
1212 
1213     if (serial_team->t.t_serialized) {
1214       /* this serial team was already used
1215          TODO increase performance by making this locks more specific */
1216       kmp_team_t *new_team;
1217 
1218       __kmp_acquire_bootstrap_lock(&__kmp_forkjoin_lock);
1219 
1220 #if OMPT_SUPPORT
1221       ompt_parallel_id_t ompt_parallel_id = __ompt_parallel_id_new(global_tid);
1222 #endif
1223 
1224       new_team = __kmp_allocate_team(this_thr->th.th_root, 1, 1,
1225 #if OMPT_SUPPORT
1226                                      ompt_parallel_id,
1227 #endif
1228 #if OMP_40_ENABLED
1229                                      proc_bind,
1230 #endif
1231                                      &this_thr->th.th_current_task->td_icvs,
1232                                      0 USE_NESTED_HOT_ARG(NULL));
1233       __kmp_release_bootstrap_lock(&__kmp_forkjoin_lock);
1234       KMP_ASSERT(new_team);
1235 
1236       /* setup new serialized team and install it */
1237       new_team->t.t_threads[0] = this_thr;
1238       new_team->t.t_parent = this_thr->th.th_team;
1239       serial_team = new_team;
1240       this_thr->th.th_serial_team = serial_team;
1241 
1242       KF_TRACE(
1243           10,
1244           ("__kmpc_serialized_parallel: T#%d allocated new serial team %p\n",
1245            global_tid, serial_team));
1246 
1247       /* TODO the above breaks the requirement that if we run out of resources,
1248          then we can still guarantee that serialized teams are ok, since we may
1249          need to allocate a new one */
1250     } else {
1251       KF_TRACE(
1252           10,
1253           ("__kmpc_serialized_parallel: T#%d reusing cached serial team %p\n",
1254            global_tid, serial_team));
1255     }
1256 
1257     /* we have to initialize this serial team */
1258     KMP_DEBUG_ASSERT(serial_team->t.t_threads);
1259     KMP_DEBUG_ASSERT(serial_team->t.t_threads[0] == this_thr);
1260     KMP_DEBUG_ASSERT(this_thr->th.th_team != serial_team);
1261     serial_team->t.t_ident = loc;
1262     serial_team->t.t_serialized = 1;
1263     serial_team->t.t_nproc = 1;
1264     serial_team->t.t_parent = this_thr->th.th_team;
1265     serial_team->t.t_sched = this_thr->th.th_team->t.t_sched;
1266     this_thr->th.th_team = serial_team;
1267     serial_team->t.t_master_tid = this_thr->th.th_info.ds.ds_tid;
1268 
1269     KF_TRACE(10, ("__kmpc_serialized_parallel: T#d curtask=%p\n", global_tid,
1270                   this_thr->th.th_current_task));
1271     KMP_ASSERT(this_thr->th.th_current_task->td_flags.executing == 1);
1272     this_thr->th.th_current_task->td_flags.executing = 0;
1273 
1274     __kmp_push_current_task_to_thread(this_thr, serial_team, 0);
1275 
1276     /* TODO: GEH: do ICVs work for nested serialized teams? Don't we need an
1277        implicit task for each serialized task represented by
1278        team->t.t_serialized? */
1279     copy_icvs(&this_thr->th.th_current_task->td_icvs,
1280               &this_thr->th.th_current_task->td_parent->td_icvs);
1281 
1282     // Thread value exists in the nested nthreads array for the next nested
1283     // level
1284     if (__kmp_nested_nth.used && (level + 1 < __kmp_nested_nth.used)) {
1285       this_thr->th.th_current_task->td_icvs.nproc =
1286           __kmp_nested_nth.nth[level + 1];
1287     }
1288 
1289 #if OMP_40_ENABLED
1290     if (__kmp_nested_proc_bind.used &&
1291         (level + 1 < __kmp_nested_proc_bind.used)) {
1292       this_thr->th.th_current_task->td_icvs.proc_bind =
1293           __kmp_nested_proc_bind.bind_types[level + 1];
1294     }
1295 #endif /* OMP_40_ENABLED */
1296 
1297 #if USE_DEBUGGER
1298     serial_team->t.t_pkfn = (microtask_t)(~0); // For the debugger.
1299 #endif
1300     this_thr->th.th_info.ds.ds_tid = 0;
1301 
1302     /* set thread cache values */
1303     this_thr->th.th_team_nproc = 1;
1304     this_thr->th.th_team_master = this_thr;
1305     this_thr->th.th_team_serialized = 1;
1306 
1307     serial_team->t.t_level = serial_team->t.t_parent->t.t_level + 1;
1308     serial_team->t.t_active_level = serial_team->t.t_parent->t.t_active_level;
1309 
1310     propagateFPControl(serial_team);
1311 
1312     /* check if we need to allocate dispatch buffers stack */
1313     KMP_DEBUG_ASSERT(serial_team->t.t_dispatch);
1314     if (!serial_team->t.t_dispatch->th_disp_buffer) {
1315       serial_team->t.t_dispatch->th_disp_buffer =
1316           (dispatch_private_info_t *)__kmp_allocate(
1317               sizeof(dispatch_private_info_t));
1318     }
1319     this_thr->th.th_dispatch = serial_team->t.t_dispatch;
1320 
1321 #if OMPT_SUPPORT
1322     ompt_parallel_id_t ompt_parallel_id = __ompt_parallel_id_new(global_tid);
1323     __ompt_team_assign_id(serial_team, ompt_parallel_id);
1324 #endif
1325 
1326     KMP_MB();
1327 
1328   } else {
1329     /* this serialized team is already being used,
1330      * that's fine, just add another nested level */
1331     KMP_DEBUG_ASSERT(this_thr->th.th_team == serial_team);
1332     KMP_DEBUG_ASSERT(serial_team->t.t_threads);
1333     KMP_DEBUG_ASSERT(serial_team->t.t_threads[0] == this_thr);
1334     ++serial_team->t.t_serialized;
1335     this_thr->th.th_team_serialized = serial_team->t.t_serialized;
1336 
1337     // Nested level will be an index in the nested nthreads array
1338     int level = this_thr->th.th_team->t.t_level;
1339     // Thread value exists in the nested nthreads array for the next nested
1340     // level
1341     if (__kmp_nested_nth.used && (level + 1 < __kmp_nested_nth.used)) {
1342       this_thr->th.th_current_task->td_icvs.nproc =
1343           __kmp_nested_nth.nth[level + 1];
1344     }
1345     serial_team->t.t_level++;
1346     KF_TRACE(10, ("__kmpc_serialized_parallel: T#%d increasing nesting level "
1347                   "of serial team %p to %d\n",
1348                   global_tid, serial_team, serial_team->t.t_level));
1349 
1350     /* allocate/push dispatch buffers stack */
1351     KMP_DEBUG_ASSERT(serial_team->t.t_dispatch);
1352     {
1353       dispatch_private_info_t *disp_buffer =
1354           (dispatch_private_info_t *)__kmp_allocate(
1355               sizeof(dispatch_private_info_t));
1356       disp_buffer->next = serial_team->t.t_dispatch->th_disp_buffer;
1357       serial_team->t.t_dispatch->th_disp_buffer = disp_buffer;
1358     }
1359     this_thr->th.th_dispatch = serial_team->t.t_dispatch;
1360 
1361     KMP_MB();
1362   }
1363 #if OMP_40_ENABLED
1364   KMP_CHECK_UPDATE(serial_team->t.t_cancel_request, cancel_noreq);
1365 #endif
1366 
1367   if (__kmp_env_consistency_check)
1368     __kmp_push_parallel(global_tid, NULL);
1369 }
1370 
1371 /* most of the work for a fork */
1372 /* return true if we really went parallel, false if serialized */
1373 int __kmp_fork_call(ident_t *loc, int gtid,
1374                     enum fork_context_e call_context, // Intel, GNU, ...
1375                     kmp_int32 argc,
1376 #if OMPT_SUPPORT
1377                     void *unwrapped_task,
1378 #endif
1379                     microtask_t microtask, launch_t invoker,
1380 /* TODO: revert workaround for Intel(R) 64 tracker #96 */
1381 #if (KMP_ARCH_X86_64 || KMP_ARCH_ARM || KMP_ARCH_AARCH64) && KMP_OS_LINUX
1382                     va_list *ap
1383 #else
1384                     va_list ap
1385 #endif
1386                     ) {
1387   void **argv;
1388   int i;
1389   int master_tid;
1390   int master_this_cons;
1391   kmp_team_t *team;
1392   kmp_team_t *parent_team;
1393   kmp_info_t *master_th;
1394   kmp_root_t *root;
1395   int nthreads;
1396   int master_active;
1397   int master_set_numthreads;
1398   int level;
1399 #if OMP_40_ENABLED
1400   int active_level;
1401   int teams_level;
1402 #endif
1403 #if KMP_NESTED_HOT_TEAMS
1404   kmp_hot_team_ptr_t **p_hot_teams;
1405 #endif
1406   { // KMP_TIME_BLOCK
1407     KMP_TIME_DEVELOPER_PARTITIONED_BLOCK(KMP_fork_call);
1408     KMP_COUNT_VALUE(OMP_PARALLEL_args, argc);
1409 
1410     KA_TRACE(20, ("__kmp_fork_call: enter T#%d\n", gtid));
1411     if (__kmp_stkpadding > 0 && __kmp_root[gtid] != NULL) {
1412       /* Some systems prefer the stack for the root thread(s) to start with */
1413       /* some gap from the parent stack to prevent false sharing. */
1414       void *dummy = KMP_ALLOCA(__kmp_stkpadding);
1415       /* These 2 lines below are so this does not get optimized out */
1416       if (__kmp_stkpadding > KMP_MAX_STKPADDING)
1417         __kmp_stkpadding += (short)((kmp_int64)dummy);
1418     }
1419 
1420     /* initialize if needed */
1421     KMP_DEBUG_ASSERT(
1422         __kmp_init_serial); // AC: potentially unsafe, not in sync with shutdown
1423     if (!TCR_4(__kmp_init_parallel))
1424       __kmp_parallel_initialize();
1425 
1426     /* setup current data */
1427     master_th = __kmp_threads[gtid]; // AC: potentially unsafe, not in sync with
1428     // shutdown
1429     parent_team = master_th->th.th_team;
1430     master_tid = master_th->th.th_info.ds.ds_tid;
1431     master_this_cons = master_th->th.th_local.this_construct;
1432     root = master_th->th.th_root;
1433     master_active = root->r.r_active;
1434     master_set_numthreads = master_th->th.th_set_nproc;
1435 
1436 #if OMPT_SUPPORT
1437     ompt_parallel_id_t ompt_parallel_id;
1438     ompt_task_id_t ompt_task_id;
1439     ompt_frame_t *ompt_frame;
1440     ompt_task_id_t my_task_id;
1441     ompt_parallel_id_t my_parallel_id;
1442 
1443     if (ompt_enabled) {
1444       ompt_parallel_id = __ompt_parallel_id_new(gtid);
1445       ompt_task_id = __ompt_get_task_id_internal(0);
1446       ompt_frame = __ompt_get_task_frame_internal(0);
1447     }
1448 #endif
1449 
1450     // Nested level will be an index in the nested nthreads array
1451     level = parent_team->t.t_level;
1452     // used to launch non-serial teams even if nested is not allowed
1453     active_level = parent_team->t.t_active_level;
1454 #if OMP_40_ENABLED
1455     // needed to check nesting inside the teams
1456     teams_level = master_th->th.th_teams_level;
1457 #endif
1458 #if KMP_NESTED_HOT_TEAMS
1459     p_hot_teams = &master_th->th.th_hot_teams;
1460     if (*p_hot_teams == NULL && __kmp_hot_teams_max_level > 0) {
1461       *p_hot_teams = (kmp_hot_team_ptr_t *)__kmp_allocate(
1462           sizeof(kmp_hot_team_ptr_t) * __kmp_hot_teams_max_level);
1463       (*p_hot_teams)[0].hot_team = root->r.r_hot_team;
1464       // it is either actual or not needed (when active_level > 0)
1465       (*p_hot_teams)[0].hot_team_nth = 1;
1466     }
1467 #endif
1468 
1469 #if OMPT_SUPPORT
1470     if (ompt_enabled &&
1471         ompt_callbacks.ompt_callback(ompt_event_parallel_begin)) {
1472       int team_size = master_set_numthreads;
1473 
1474       ompt_callbacks.ompt_callback(ompt_event_parallel_begin)(
1475           ompt_task_id, ompt_frame, ompt_parallel_id, team_size, unwrapped_task,
1476           OMPT_INVOKER(call_context));
1477     }
1478 #endif
1479 
1480     master_th->th.th_ident = loc;
1481 
1482 #if OMP_40_ENABLED
1483     if (master_th->th.th_teams_microtask && ap &&
1484         microtask != (microtask_t)__kmp_teams_master && level == teams_level) {
1485       // AC: This is start of parallel that is nested inside teams construct.
1486       // The team is actual (hot), all workers are ready at the fork barrier.
1487       // No lock needed to initialize the team a bit, then free workers.
1488       parent_team->t.t_ident = loc;
1489       __kmp_alloc_argv_entries(argc, parent_team, TRUE);
1490       parent_team->t.t_argc = argc;
1491       argv = (void **)parent_team->t.t_argv;
1492       for (i = argc - 1; i >= 0; --i)
1493 /* TODO: revert workaround for Intel(R) 64 tracker #96 */
1494 #if (KMP_ARCH_X86_64 || KMP_ARCH_ARM || KMP_ARCH_AARCH64) && KMP_OS_LINUX
1495         *argv++ = va_arg(*ap, void *);
1496 #else
1497         *argv++ = va_arg(ap, void *);
1498 #endif
1499       // Increment our nested depth levels, but not increase the serialization
1500       if (parent_team == master_th->th.th_serial_team) {
1501         // AC: we are in serialized parallel
1502         __kmpc_serialized_parallel(loc, gtid);
1503         KMP_DEBUG_ASSERT(parent_team->t.t_serialized > 1);
1504         // AC: need this in order enquiry functions work
1505         // correctly, will restore at join time
1506         parent_team->t.t_serialized--;
1507 #if OMPT_SUPPORT
1508         void *dummy;
1509         void **exit_runtime_p;
1510 
1511         ompt_lw_taskteam_t lw_taskteam;
1512 
1513         if (ompt_enabled) {
1514           __ompt_lw_taskteam_init(&lw_taskteam, master_th, gtid, unwrapped_task,
1515                                   ompt_parallel_id);
1516           lw_taskteam.ompt_task_info.task_id = __ompt_task_id_new(gtid);
1517           exit_runtime_p =
1518               &(lw_taskteam.ompt_task_info.frame.exit_runtime_frame);
1519 
1520           __ompt_lw_taskteam_link(&lw_taskteam, master_th);
1521 
1522 #if OMPT_TRACE
1523           /* OMPT implicit task begin */
1524           my_task_id = lw_taskteam.ompt_task_info.task_id;
1525           my_parallel_id = parent_team->t.ompt_team_info.parallel_id;
1526           if (ompt_callbacks.ompt_callback(ompt_event_implicit_task_begin)) {
1527             ompt_callbacks.ompt_callback(ompt_event_implicit_task_begin)(
1528                 my_parallel_id, my_task_id);
1529           }
1530 #endif
1531 
1532           /* OMPT state */
1533           master_th->th.ompt_thread_info.state = ompt_state_work_parallel;
1534         } else {
1535           exit_runtime_p = &dummy;
1536         }
1537 #endif
1538 
1539         {
1540           KMP_TIME_PARTITIONED_BLOCK(OMP_parallel);
1541           KMP_SET_THREAD_STATE_BLOCK(IMPLICIT_TASK);
1542           __kmp_invoke_microtask(microtask, gtid, 0, argc, parent_team->t.t_argv
1543 #if OMPT_SUPPORT
1544                                  ,
1545                                  exit_runtime_p
1546 #endif
1547                                  );
1548         }
1549 
1550 #if OMPT_SUPPORT
1551         *exit_runtime_p = NULL;
1552         if (ompt_enabled) {
1553 #if OMPT_TRACE
1554           lw_taskteam.ompt_task_info.frame.exit_runtime_frame = NULL;
1555 
1556           if (ompt_callbacks.ompt_callback(ompt_event_implicit_task_end)) {
1557             ompt_callbacks.ompt_callback(ompt_event_implicit_task_end)(
1558                 ompt_parallel_id, ompt_task_id);
1559           }
1560 
1561           __ompt_lw_taskteam_unlink(master_th);
1562           // reset clear the task id only after unlinking the task
1563           lw_taskteam.ompt_task_info.task_id = ompt_task_id_none;
1564 #endif
1565 
1566           if (ompt_callbacks.ompt_callback(ompt_event_parallel_end)) {
1567             ompt_callbacks.ompt_callback(ompt_event_parallel_end)(
1568                 ompt_parallel_id, ompt_task_id, OMPT_INVOKER(call_context));
1569           }
1570           master_th->th.ompt_thread_info.state = ompt_state_overhead;
1571         }
1572 #endif
1573         return TRUE;
1574       }
1575 
1576       parent_team->t.t_pkfn = microtask;
1577 #if OMPT_SUPPORT
1578       parent_team->t.ompt_team_info.microtask = unwrapped_task;
1579 #endif
1580       parent_team->t.t_invoke = invoker;
1581       KMP_TEST_THEN_INC32((kmp_int32 *)&root->r.r_in_parallel);
1582       parent_team->t.t_active_level++;
1583       parent_team->t.t_level++;
1584 
1585       /* Change number of threads in the team if requested */
1586       if (master_set_numthreads) { // The parallel has num_threads clause
1587         if (master_set_numthreads < master_th->th.th_teams_size.nth) {
1588           // AC: only can reduce number of threads dynamically, can't increase
1589           kmp_info_t **other_threads = parent_team->t.t_threads;
1590           parent_team->t.t_nproc = master_set_numthreads;
1591           for (i = 0; i < master_set_numthreads; ++i) {
1592             other_threads[i]->th.th_team_nproc = master_set_numthreads;
1593           }
1594           // Keep extra threads hot in the team for possible next parallels
1595         }
1596         master_th->th.th_set_nproc = 0;
1597       }
1598 
1599 #if USE_DEBUGGER
1600       if (__kmp_debugging) { // Let debugger override number of threads.
1601         int nth = __kmp_omp_num_threads(loc);
1602         if (nth > 0) { // 0 means debugger doesn't want to change num threads
1603           master_set_numthreads = nth;
1604         }; // if
1605       }; // if
1606 #endif
1607 
1608       KF_TRACE(10, ("__kmp_fork_call: before internal fork: root=%p, team=%p, "
1609                     "master_th=%p, gtid=%d\n",
1610                     root, parent_team, master_th, gtid));
1611       __kmp_internal_fork(loc, gtid, parent_team);
1612       KF_TRACE(10, ("__kmp_fork_call: after internal fork: root=%p, team=%p, "
1613                     "master_th=%p, gtid=%d\n",
1614                     root, parent_team, master_th, gtid));
1615 
1616       /* Invoke microtask for MASTER thread */
1617       KA_TRACE(20, ("__kmp_fork_call: T#%d(%d:0) invoke microtask = %p\n", gtid,
1618                     parent_team->t.t_id, parent_team->t.t_pkfn));
1619 
1620       {
1621         KMP_TIME_PARTITIONED_BLOCK(OMP_parallel);
1622         KMP_SET_THREAD_STATE_BLOCK(IMPLICIT_TASK);
1623         if (!parent_team->t.t_invoke(gtid)) {
1624           KMP_ASSERT2(0, "cannot invoke microtask for MASTER thread");
1625         }
1626       }
1627       KA_TRACE(20, ("__kmp_fork_call: T#%d(%d:0) done microtask = %p\n", gtid,
1628                     parent_team->t.t_id, parent_team->t.t_pkfn));
1629       KMP_MB(); /* Flush all pending memory write invalidates.  */
1630 
1631       KA_TRACE(20, ("__kmp_fork_call: parallel exit T#%d\n", gtid));
1632 
1633       return TRUE;
1634     } // Parallel closely nested in teams construct
1635 #endif /* OMP_40_ENABLED */
1636 
1637 #if KMP_DEBUG
1638     if (__kmp_tasking_mode != tskm_immediate_exec) {
1639       KMP_DEBUG_ASSERT(master_th->th.th_task_team ==
1640                        parent_team->t.t_task_team[master_th->th.th_task_state]);
1641     }
1642 #endif
1643 
1644     if (parent_team->t.t_active_level >=
1645         master_th->th.th_current_task->td_icvs.max_active_levels) {
1646       nthreads = 1;
1647     } else {
1648 #if OMP_40_ENABLED
1649       int enter_teams = ((ap == NULL && active_level == 0) ||
1650                          (ap && teams_level > 0 && teams_level == level));
1651 #endif
1652       nthreads =
1653           master_set_numthreads
1654               ? master_set_numthreads
1655               : get__nproc_2(
1656                     parent_team,
1657                     master_tid); // TODO: get nproc directly from current task
1658 
1659       // Check if we need to take forkjoin lock? (no need for serialized
1660       // parallel out of teams construct). This code moved here from
1661       // __kmp_reserve_threads() to speedup nested serialized parallels.
1662       if (nthreads > 1) {
1663         if ((!get__nested(master_th) && (root->r.r_in_parallel
1664 #if OMP_40_ENABLED
1665                                          && !enter_teams
1666 #endif /* OMP_40_ENABLED */
1667                                          )) ||
1668             (__kmp_library == library_serial)) {
1669           KC_TRACE(10, ("__kmp_fork_call: T#%d serializing team; requested %d"
1670                         " threads\n",
1671                         gtid, nthreads));
1672           nthreads = 1;
1673         }
1674       }
1675       if (nthreads > 1) {
1676         /* determine how many new threads we can use */
1677         __kmp_acquire_bootstrap_lock(&__kmp_forkjoin_lock);
1678         nthreads = __kmp_reserve_threads(
1679             root, parent_team, master_tid, nthreads
1680 #if OMP_40_ENABLED
1681             /* AC: If we execute teams from parallel region (on host), then
1682                teams should be created but each can only have 1 thread if
1683                nesting is disabled. If teams called from serial region, then
1684                teams and their threads should be created regardless of the
1685                nesting setting. */
1686             ,
1687             enter_teams
1688 #endif /* OMP_40_ENABLED */
1689             );
1690         if (nthreads == 1) {
1691           // Free lock for single thread execution here; for multi-thread
1692           // execution it will be freed later after team of threads created
1693           // and initialized
1694           __kmp_release_bootstrap_lock(&__kmp_forkjoin_lock);
1695         }
1696       }
1697     }
1698     KMP_DEBUG_ASSERT(nthreads > 0);
1699 
1700     // If we temporarily changed the set number of threads then restore it now
1701     master_th->th.th_set_nproc = 0;
1702 
1703     /* create a serialized parallel region? */
1704     if (nthreads == 1) {
1705 /* josh todo: hypothetical question: what do we do for OS X*? */
1706 #if KMP_OS_LINUX &&                                                            \
1707     (KMP_ARCH_X86 || KMP_ARCH_X86_64 || KMP_ARCH_ARM || KMP_ARCH_AARCH64)
1708       void *args[argc];
1709 #else
1710       void **args = (void **)KMP_ALLOCA(argc * sizeof(void *));
1711 #endif /* KMP_OS_LINUX && ( KMP_ARCH_X86 || KMP_ARCH_X86_64 || KMP_ARCH_ARM || \
1712           KMP_ARCH_AARCH64) */
1713 
1714       KA_TRACE(20,
1715                ("__kmp_fork_call: T#%d serializing parallel region\n", gtid));
1716 
1717       __kmpc_serialized_parallel(loc, gtid);
1718 
1719       if (call_context == fork_context_intel) {
1720         /* TODO this sucks, use the compiler itself to pass args! :) */
1721         master_th->th.th_serial_team->t.t_ident = loc;
1722 #if OMP_40_ENABLED
1723         if (!ap) {
1724           // revert change made in __kmpc_serialized_parallel()
1725           master_th->th.th_serial_team->t.t_level--;
1726 // Get args from parent team for teams construct
1727 
1728 #if OMPT_SUPPORT
1729           void *dummy;
1730           void **exit_runtime_p;
1731 
1732           ompt_lw_taskteam_t lw_taskteam;
1733 
1734           if (ompt_enabled) {
1735             __ompt_lw_taskteam_init(&lw_taskteam, master_th, gtid,
1736                                     unwrapped_task, ompt_parallel_id);
1737             lw_taskteam.ompt_task_info.task_id = __ompt_task_id_new(gtid);
1738             exit_runtime_p =
1739                 &(lw_taskteam.ompt_task_info.frame.exit_runtime_frame);
1740 
1741             __ompt_lw_taskteam_link(&lw_taskteam, master_th);
1742 
1743 #if OMPT_TRACE
1744             my_task_id = lw_taskteam.ompt_task_info.task_id;
1745             if (ompt_callbacks.ompt_callback(ompt_event_implicit_task_begin)) {
1746               ompt_callbacks.ompt_callback(ompt_event_implicit_task_begin)(
1747                   ompt_parallel_id, my_task_id);
1748             }
1749 #endif
1750 
1751             /* OMPT state */
1752             master_th->th.ompt_thread_info.state = ompt_state_work_parallel;
1753           } else {
1754             exit_runtime_p = &dummy;
1755           }
1756 #endif
1757 
1758           {
1759             KMP_TIME_PARTITIONED_BLOCK(OMP_parallel);
1760             KMP_SET_THREAD_STATE_BLOCK(IMPLICIT_TASK);
1761             __kmp_invoke_microtask(microtask, gtid, 0, argc,
1762                                    parent_team->t.t_argv
1763 #if OMPT_SUPPORT
1764                                    ,
1765                                    exit_runtime_p
1766 #endif
1767                                    );
1768           }
1769 
1770 #if OMPT_SUPPORT
1771           *exit_runtime_p = NULL;
1772           if (ompt_enabled) {
1773             lw_taskteam.ompt_task_info.frame.exit_runtime_frame = NULL;
1774 
1775 #if OMPT_TRACE
1776             if (ompt_callbacks.ompt_callback(ompt_event_implicit_task_end)) {
1777               ompt_callbacks.ompt_callback(ompt_event_implicit_task_end)(
1778                   ompt_parallel_id, ompt_task_id);
1779             }
1780 #endif
1781 
1782             __ompt_lw_taskteam_unlink(master_th);
1783             // reset clear the task id only after unlinking the task
1784             lw_taskteam.ompt_task_info.task_id = ompt_task_id_none;
1785 
1786             if (ompt_callbacks.ompt_callback(ompt_event_parallel_end)) {
1787               ompt_callbacks.ompt_callback(ompt_event_parallel_end)(
1788                   ompt_parallel_id, ompt_task_id, OMPT_INVOKER(call_context));
1789             }
1790             master_th->th.ompt_thread_info.state = ompt_state_overhead;
1791           }
1792 #endif
1793         } else if (microtask == (microtask_t)__kmp_teams_master) {
1794           KMP_DEBUG_ASSERT(master_th->th.th_team ==
1795                            master_th->th.th_serial_team);
1796           team = master_th->th.th_team;
1797           // team->t.t_pkfn = microtask;
1798           team->t.t_invoke = invoker;
1799           __kmp_alloc_argv_entries(argc, team, TRUE);
1800           team->t.t_argc = argc;
1801           argv = (void **)team->t.t_argv;
1802           if (ap) {
1803             for (i = argc - 1; i >= 0; --i)
1804 // TODO: revert workaround for Intel(R) 64 tracker #96
1805 #if (KMP_ARCH_X86_64 || KMP_ARCH_ARM || KMP_ARCH_AARCH64) && KMP_OS_LINUX
1806               *argv++ = va_arg(*ap, void *);
1807 #else
1808               *argv++ = va_arg(ap, void *);
1809 #endif
1810           } else {
1811             for (i = 0; i < argc; ++i)
1812               // Get args from parent team for teams construct
1813               argv[i] = parent_team->t.t_argv[i];
1814           }
1815           // AC: revert change made in __kmpc_serialized_parallel()
1816           //     because initial code in teams should have level=0
1817           team->t.t_level--;
1818           // AC: call special invoker for outer "parallel" of teams construct
1819           {
1820             KMP_TIME_PARTITIONED_BLOCK(OMP_parallel);
1821             KMP_SET_THREAD_STATE_BLOCK(IMPLICIT_TASK);
1822             invoker(gtid);
1823           }
1824         } else {
1825 #endif /* OMP_40_ENABLED */
1826           argv = args;
1827           for (i = argc - 1; i >= 0; --i)
1828 // TODO: revert workaround for Intel(R) 64 tracker #96
1829 #if (KMP_ARCH_X86_64 || KMP_ARCH_ARM || KMP_ARCH_AARCH64) && KMP_OS_LINUX
1830             *argv++ = va_arg(*ap, void *);
1831 #else
1832           *argv++ = va_arg(ap, void *);
1833 #endif
1834           KMP_MB();
1835 
1836 #if OMPT_SUPPORT
1837           void *dummy;
1838           void **exit_runtime_p;
1839 
1840           ompt_lw_taskteam_t lw_taskteam;
1841 
1842           if (ompt_enabled) {
1843             __ompt_lw_taskteam_init(&lw_taskteam, master_th, gtid,
1844                                     unwrapped_task, ompt_parallel_id);
1845             lw_taskteam.ompt_task_info.task_id = __ompt_task_id_new(gtid);
1846             exit_runtime_p =
1847                 &(lw_taskteam.ompt_task_info.frame.exit_runtime_frame);
1848 
1849             __ompt_lw_taskteam_link(&lw_taskteam, master_th);
1850 
1851 #if OMPT_TRACE
1852             /* OMPT implicit task begin */
1853             my_task_id = lw_taskteam.ompt_task_info.task_id;
1854             my_parallel_id = ompt_parallel_id;
1855             if (ompt_callbacks.ompt_callback(ompt_event_implicit_task_begin)) {
1856               ompt_callbacks.ompt_callback(ompt_event_implicit_task_begin)(
1857                   my_parallel_id, my_task_id);
1858             }
1859 #endif
1860 
1861             /* OMPT state */
1862             master_th->th.ompt_thread_info.state = ompt_state_work_parallel;
1863           } else {
1864             exit_runtime_p = &dummy;
1865           }
1866 #endif
1867 
1868           {
1869             KMP_TIME_PARTITIONED_BLOCK(OMP_parallel);
1870             KMP_SET_THREAD_STATE_BLOCK(IMPLICIT_TASK);
1871             __kmp_invoke_microtask(microtask, gtid, 0, argc, args
1872 #if OMPT_SUPPORT
1873                                    ,
1874                                    exit_runtime_p
1875 #endif
1876                                    );
1877           }
1878 
1879 #if OMPT_SUPPORT
1880           *exit_runtime_p = NULL;
1881           if (ompt_enabled) {
1882 #if OMPT_TRACE
1883             lw_taskteam.ompt_task_info.frame.exit_runtime_frame = NULL;
1884 
1885             if (ompt_callbacks.ompt_callback(ompt_event_implicit_task_end)) {
1886               ompt_callbacks.ompt_callback(ompt_event_implicit_task_end)(
1887                   my_parallel_id, my_task_id);
1888             }
1889 #endif
1890 
1891             __ompt_lw_taskteam_unlink(master_th);
1892             // reset clear the task id only after unlinking the task
1893             lw_taskteam.ompt_task_info.task_id = ompt_task_id_none;
1894 
1895             if (ompt_callbacks.ompt_callback(ompt_event_parallel_end)) {
1896               ompt_callbacks.ompt_callback(ompt_event_parallel_end)(
1897                   ompt_parallel_id, ompt_task_id, OMPT_INVOKER(call_context));
1898             }
1899             master_th->th.ompt_thread_info.state = ompt_state_overhead;
1900           }
1901 #endif
1902 #if OMP_40_ENABLED
1903         }
1904 #endif /* OMP_40_ENABLED */
1905       } else if (call_context == fork_context_gnu) {
1906 #if OMPT_SUPPORT
1907         ompt_lw_taskteam_t *lwt =
1908             (ompt_lw_taskteam_t *)__kmp_allocate(sizeof(ompt_lw_taskteam_t));
1909         __ompt_lw_taskteam_init(lwt, master_th, gtid, unwrapped_task,
1910                                 ompt_parallel_id);
1911 
1912         lwt->ompt_task_info.task_id = __ompt_task_id_new(gtid);
1913         lwt->ompt_task_info.frame.exit_runtime_frame = NULL;
1914         __ompt_lw_taskteam_link(lwt, master_th);
1915 #endif
1916 
1917         // we were called from GNU native code
1918         KA_TRACE(20, ("__kmp_fork_call: T#%d serial exit\n", gtid));
1919         return FALSE;
1920       } else {
1921         KMP_ASSERT2(call_context < fork_context_last,
1922                     "__kmp_fork_call: unknown fork_context parameter");
1923       }
1924 
1925       KA_TRACE(20, ("__kmp_fork_call: T#%d serial exit\n", gtid));
1926       KMP_MB();
1927       return FALSE;
1928     }
1929 
1930     // GEH: only modify the executing flag in the case when not serialized
1931     //      serialized case is handled in kmpc_serialized_parallel
1932     KF_TRACE(10, ("__kmp_fork_call: parent_team_aclevel=%d, master_th=%p, "
1933                   "curtask=%p, curtask_max_aclevel=%d\n",
1934                   parent_team->t.t_active_level, master_th,
1935                   master_th->th.th_current_task,
1936                   master_th->th.th_current_task->td_icvs.max_active_levels));
1937     // TODO: GEH - cannot do this assertion because root thread not set up as
1938     // executing
1939     // KMP_ASSERT( master_th->th.th_current_task->td_flags.executing == 1 );
1940     master_th->th.th_current_task->td_flags.executing = 0;
1941 
1942 #if OMP_40_ENABLED
1943     if (!master_th->th.th_teams_microtask || level > teams_level)
1944 #endif /* OMP_40_ENABLED */
1945     {
1946       /* Increment our nested depth level */
1947       KMP_TEST_THEN_INC32((kmp_int32 *)&root->r.r_in_parallel);
1948     }
1949 
1950     // See if we need to make a copy of the ICVs.
1951     int nthreads_icv = master_th->th.th_current_task->td_icvs.nproc;
1952     if ((level + 1 < __kmp_nested_nth.used) &&
1953         (__kmp_nested_nth.nth[level + 1] != nthreads_icv)) {
1954       nthreads_icv = __kmp_nested_nth.nth[level + 1];
1955     } else {
1956       nthreads_icv = 0; // don't update
1957     }
1958 
1959 #if OMP_40_ENABLED
1960     // Figure out the proc_bind_policy for the new team.
1961     kmp_proc_bind_t proc_bind = master_th->th.th_set_proc_bind;
1962     kmp_proc_bind_t proc_bind_icv =
1963         proc_bind_default; // proc_bind_default means don't update
1964     if (master_th->th.th_current_task->td_icvs.proc_bind == proc_bind_false) {
1965       proc_bind = proc_bind_false;
1966     } else {
1967       if (proc_bind == proc_bind_default) {
1968         // No proc_bind clause specified; use current proc-bind-var for this
1969         // parallel region
1970         proc_bind = master_th->th.th_current_task->td_icvs.proc_bind;
1971       }
1972       /* else: The proc_bind policy was specified explicitly on parallel clause.
1973          This overrides proc-bind-var for this parallel region, but does not
1974          change proc-bind-var. */
1975       // Figure the value of proc-bind-var for the child threads.
1976       if ((level + 1 < __kmp_nested_proc_bind.used) &&
1977           (__kmp_nested_proc_bind.bind_types[level + 1] !=
1978            master_th->th.th_current_task->td_icvs.proc_bind)) {
1979         proc_bind_icv = __kmp_nested_proc_bind.bind_types[level + 1];
1980       }
1981     }
1982 
1983     // Reset for next parallel region
1984     master_th->th.th_set_proc_bind = proc_bind_default;
1985 #endif /* OMP_40_ENABLED */
1986 
1987     if ((nthreads_icv > 0)
1988 #if OMP_40_ENABLED
1989         || (proc_bind_icv != proc_bind_default)
1990 #endif /* OMP_40_ENABLED */
1991             ) {
1992       kmp_internal_control_t new_icvs;
1993       copy_icvs(&new_icvs, &master_th->th.th_current_task->td_icvs);
1994       new_icvs.next = NULL;
1995       if (nthreads_icv > 0) {
1996         new_icvs.nproc = nthreads_icv;
1997       }
1998 
1999 #if OMP_40_ENABLED
2000       if (proc_bind_icv != proc_bind_default) {
2001         new_icvs.proc_bind = proc_bind_icv;
2002       }
2003 #endif /* OMP_40_ENABLED */
2004 
2005       /* allocate a new parallel team */
2006       KF_TRACE(10, ("__kmp_fork_call: before __kmp_allocate_team\n"));
2007       team = __kmp_allocate_team(root, nthreads, nthreads,
2008 #if OMPT_SUPPORT
2009                                  ompt_parallel_id,
2010 #endif
2011 #if OMP_40_ENABLED
2012                                  proc_bind,
2013 #endif
2014                                  &new_icvs, argc USE_NESTED_HOT_ARG(master_th));
2015     } else {
2016       /* allocate a new parallel team */
2017       KF_TRACE(10, ("__kmp_fork_call: before __kmp_allocate_team\n"));
2018       team = __kmp_allocate_team(root, nthreads, nthreads,
2019 #if OMPT_SUPPORT
2020                                  ompt_parallel_id,
2021 #endif
2022 #if OMP_40_ENABLED
2023                                  proc_bind,
2024 #endif
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.microtask, unwrapped_task);
2039 #endif
2040     KMP_CHECK_UPDATE(team->t.t_invoke, invoker); // TODO move to root, maybe
2041 // TODO: parent_team->t.t_level == INT_MAX ???
2042 #if OMP_40_ENABLED
2043     if (!master_th->th.th_teams_microtask || level > teams_level) {
2044 #endif /* OMP_40_ENABLED */
2045       int new_level = parent_team->t.t_level + 1;
2046       KMP_CHECK_UPDATE(team->t.t_level, new_level);
2047       new_level = parent_team->t.t_active_level + 1;
2048       KMP_CHECK_UPDATE(team->t.t_active_level, new_level);
2049 #if OMP_40_ENABLED
2050     } else {
2051       // AC: Do not increase parallel level at start of the teams construct
2052       int new_level = parent_team->t.t_level;
2053       KMP_CHECK_UPDATE(team->t.t_level, new_level);
2054       new_level = parent_team->t.t_active_level;
2055       KMP_CHECK_UPDATE(team->t.t_active_level, new_level);
2056     }
2057 #endif /* OMP_40_ENABLED */
2058     kmp_r_sched_t new_sched = get__sched_2(parent_team, master_tid);
2059     if (team->t.t_sched.r_sched_type != new_sched.r_sched_type ||
2060         team->t.t_sched.chunk != new_sched.chunk)
2061       team->t.t_sched =
2062           new_sched; // set master's schedule as new run-time schedule
2063 
2064 #if OMP_40_ENABLED
2065     KMP_CHECK_UPDATE(team->t.t_cancel_request, cancel_noreq);
2066 #endif
2067 
2068     // Update the floating point rounding in the team if required.
2069     propagateFPControl(team);
2070 
2071     if (__kmp_tasking_mode != tskm_immediate_exec) {
2072       // Set master's task team to team's task team. Unless this is hot team, it
2073       // should be NULL.
2074       KMP_DEBUG_ASSERT(master_th->th.th_task_team ==
2075                        parent_team->t.t_task_team[master_th->th.th_task_state]);
2076       KA_TRACE(20, ("__kmp_fork_call: Master T#%d pushing task_team %p / team "
2077                     "%p, new task_team %p / team %p\n",
2078                     __kmp_gtid_from_thread(master_th),
2079                     master_th->th.th_task_team, parent_team,
2080                     team->t.t_task_team[master_th->th.th_task_state], team));
2081 
2082       if (active_level || master_th->th.th_task_team) {
2083         // Take a memo of master's task_state
2084         KMP_DEBUG_ASSERT(master_th->th.th_task_state_memo_stack);
2085         if (master_th->th.th_task_state_top >=
2086             master_th->th.th_task_state_stack_sz) { // increase size
2087           kmp_uint32 new_size = 2 * master_th->th.th_task_state_stack_sz;
2088           kmp_uint8 *old_stack, *new_stack;
2089           kmp_uint32 i;
2090           new_stack = (kmp_uint8 *)__kmp_allocate(new_size);
2091           for (i = 0; i < master_th->th.th_task_state_stack_sz; ++i) {
2092             new_stack[i] = master_th->th.th_task_state_memo_stack[i];
2093           }
2094           for (i = master_th->th.th_task_state_stack_sz; i < new_size;
2095                ++i) { // zero-init rest of stack
2096             new_stack[i] = 0;
2097           }
2098           old_stack = master_th->th.th_task_state_memo_stack;
2099           master_th->th.th_task_state_memo_stack = new_stack;
2100           master_th->th.th_task_state_stack_sz = new_size;
2101           __kmp_free(old_stack);
2102         }
2103         // Store master's task_state on stack
2104         master_th->th
2105             .th_task_state_memo_stack[master_th->th.th_task_state_top] =
2106             master_th->th.th_task_state;
2107         master_th->th.th_task_state_top++;
2108 #if KMP_NESTED_HOT_TEAMS
2109         if (team == master_th->th.th_hot_teams[active_level].hot_team) {
2110           // Restore master's nested state if nested hot team
2111           master_th->th.th_task_state =
2112               master_th->th
2113                   .th_task_state_memo_stack[master_th->th.th_task_state_top];
2114         } else {
2115 #endif
2116           master_th->th.th_task_state = 0;
2117 #if KMP_NESTED_HOT_TEAMS
2118         }
2119 #endif
2120       }
2121 #if !KMP_NESTED_HOT_TEAMS
2122       KMP_DEBUG_ASSERT((master_th->th.th_task_team == NULL) ||
2123                        (team == root->r.r_hot_team));
2124 #endif
2125     }
2126 
2127     KA_TRACE(
2128         20,
2129         ("__kmp_fork_call: T#%d(%d:%d)->(%d:0) created a team of %d threads\n",
2130          gtid, parent_team->t.t_id, team->t.t_master_tid, team->t.t_id,
2131          team->t.t_nproc));
2132     KMP_DEBUG_ASSERT(team != root->r.r_hot_team ||
2133                      (team->t.t_master_tid == 0 &&
2134                       (team->t.t_parent == root->r.r_root_team ||
2135                        team->t.t_parent->t.t_serialized)));
2136     KMP_MB();
2137 
2138     /* now, setup the arguments */
2139     argv = (void **)team->t.t_argv;
2140 #if OMP_40_ENABLED
2141     if (ap) {
2142 #endif /* OMP_40_ENABLED */
2143       for (i = argc - 1; i >= 0; --i) {
2144 // TODO: revert workaround for Intel(R) 64 tracker #96
2145 #if (KMP_ARCH_X86_64 || KMP_ARCH_ARM || KMP_ARCH_AARCH64) && KMP_OS_LINUX
2146         void *new_argv = va_arg(*ap, void *);
2147 #else
2148       void *new_argv = va_arg(ap, void *);
2149 #endif
2150         KMP_CHECK_UPDATE(*argv, new_argv);
2151         argv++;
2152       }
2153 #if OMP_40_ENABLED
2154     } else {
2155       for (i = 0; i < argc; ++i) {
2156         // Get args from parent team for teams construct
2157         KMP_CHECK_UPDATE(argv[i], team->t.t_parent->t.t_argv[i]);
2158       }
2159     }
2160 #endif /* OMP_40_ENABLED */
2161 
2162     /* now actually fork the threads */
2163     KMP_CHECK_UPDATE(team->t.t_master_active, master_active);
2164     if (!root->r.r_active) // Only do assignment if it prevents cache ping-pong
2165       root->r.r_active = TRUE;
2166 
2167     __kmp_fork_team_threads(root, team, master_th, gtid);
2168     __kmp_setup_icv_copy(team, nthreads,
2169                          &master_th->th.th_current_task->td_icvs, loc);
2170 
2171 #if OMPT_SUPPORT
2172     master_th->th.ompt_thread_info.state = ompt_state_work_parallel;
2173 #endif
2174 
2175     __kmp_release_bootstrap_lock(&__kmp_forkjoin_lock);
2176 
2177 #if USE_ITT_BUILD
2178     if (team->t.t_active_level == 1 // only report frames at level 1
2179 #if OMP_40_ENABLED
2180         && !master_th->th.th_teams_microtask // not in teams construct
2181 #endif /* OMP_40_ENABLED */
2182         ) {
2183 #if USE_ITT_NOTIFY
2184       if ((__itt_frame_submit_v3_ptr || KMP_ITT_DEBUG) &&
2185           (__kmp_forkjoin_frames_mode == 3 ||
2186            __kmp_forkjoin_frames_mode == 1)) {
2187         kmp_uint64 tmp_time = 0;
2188         if (__itt_get_timestamp_ptr)
2189           tmp_time = __itt_get_timestamp();
2190         // Internal fork - report frame begin
2191         master_th->th.th_frame_time = tmp_time;
2192         if (__kmp_forkjoin_frames_mode == 3)
2193           team->t.t_region_time = tmp_time;
2194       } else
2195 // only one notification scheme (either "submit" or "forking/joined", not both)
2196 #endif /* USE_ITT_NOTIFY */
2197           if ((__itt_frame_begin_v3_ptr || KMP_ITT_DEBUG) &&
2198               __kmp_forkjoin_frames && !__kmp_forkjoin_frames_mode) {
2199         // Mark start of "parallel" region for VTune.
2200         __kmp_itt_region_forking(gtid, team->t.t_nproc, 0);
2201       }
2202     }
2203 #endif /* USE_ITT_BUILD */
2204 
2205     /* now go on and do the work */
2206     KMP_DEBUG_ASSERT(team == __kmp_threads[gtid]->th.th_team);
2207     KMP_MB();
2208     KF_TRACE(10,
2209              ("__kmp_internal_fork : root=%p, team=%p, master_th=%p, gtid=%d\n",
2210               root, team, master_th, gtid));
2211 
2212 #if USE_ITT_BUILD
2213     if (__itt_stack_caller_create_ptr) {
2214       team->t.t_stack_id =
2215           __kmp_itt_stack_caller_create(); // create new stack stitching id
2216       // before entering fork barrier
2217     }
2218 #endif /* USE_ITT_BUILD */
2219 
2220 #if OMP_40_ENABLED
2221     // AC: skip __kmp_internal_fork at teams construct, let only master
2222     // threads execute
2223     if (ap)
2224 #endif /* OMP_40_ENABLED */
2225     {
2226       __kmp_internal_fork(loc, gtid, team);
2227       KF_TRACE(10, ("__kmp_internal_fork : after : root=%p, team=%p, "
2228                     "master_th=%p, gtid=%d\n",
2229                     root, team, master_th, gtid));
2230     }
2231 
2232     if (call_context == fork_context_gnu) {
2233       KA_TRACE(20, ("__kmp_fork_call: parallel exit T#%d\n", gtid));
2234       return TRUE;
2235     }
2236 
2237     /* Invoke microtask for MASTER thread */
2238     KA_TRACE(20, ("__kmp_fork_call: T#%d(%d:0) invoke microtask = %p\n", gtid,
2239                   team->t.t_id, team->t.t_pkfn));
2240   } // END of timer KMP_fork_call block
2241 
2242   {
2243     KMP_TIME_PARTITIONED_BLOCK(OMP_parallel);
2244     KMP_SET_THREAD_STATE_BLOCK(IMPLICIT_TASK);
2245     if (!team->t.t_invoke(gtid)) {
2246       KMP_ASSERT2(0, "cannot invoke microtask for MASTER thread");
2247     }
2248   }
2249   KA_TRACE(20, ("__kmp_fork_call: T#%d(%d:0) done microtask = %p\n", gtid,
2250                 team->t.t_id, team->t.t_pkfn));
2251   KMP_MB(); /* Flush all pending memory write invalidates.  */
2252 
2253   KA_TRACE(20, ("__kmp_fork_call: parallel exit T#%d\n", gtid));
2254 
2255 #if OMPT_SUPPORT
2256   if (ompt_enabled) {
2257     master_th->th.ompt_thread_info.state = ompt_state_overhead;
2258   }
2259 #endif
2260 
2261   return TRUE;
2262 }
2263 
2264 #if OMPT_SUPPORT
2265 static inline void __kmp_join_restore_state(kmp_info_t *thread,
2266                                             kmp_team_t *team) {
2267   // restore state outside the region
2268   thread->th.ompt_thread_info.state =
2269       ((team->t.t_serialized) ? ompt_state_work_serial
2270                               : ompt_state_work_parallel);
2271 }
2272 
2273 static inline void __kmp_join_ompt(kmp_info_t *thread, kmp_team_t *team,
2274                                    ompt_parallel_id_t parallel_id,
2275                                    fork_context_e fork_context) {
2276   ompt_task_info_t *task_info = __ompt_get_taskinfo(0);
2277   if (ompt_callbacks.ompt_callback(ompt_event_parallel_end)) {
2278     ompt_callbacks.ompt_callback(ompt_event_parallel_end)(
2279         parallel_id, task_info->task_id, OMPT_INVOKER(fork_context));
2280   }
2281 
2282   task_info->frame.reenter_runtime_frame = NULL;
2283   __kmp_join_restore_state(thread, team);
2284 }
2285 #endif
2286 
2287 void __kmp_join_call(ident_t *loc, int gtid
2288 #if OMPT_SUPPORT
2289                      ,
2290                      enum fork_context_e fork_context
2291 #endif
2292 #if OMP_40_ENABLED
2293                      ,
2294                      int exit_teams
2295 #endif /* OMP_40_ENABLED */
2296                      ) {
2297   KMP_TIME_DEVELOPER_PARTITIONED_BLOCK(KMP_join_call);
2298   kmp_team_t *team;
2299   kmp_team_t *parent_team;
2300   kmp_info_t *master_th;
2301   kmp_root_t *root;
2302   int master_active;
2303   int i;
2304 
2305   KA_TRACE(20, ("__kmp_join_call: enter T#%d\n", gtid));
2306 
2307   /* setup current data */
2308   master_th = __kmp_threads[gtid];
2309   root = master_th->th.th_root;
2310   team = master_th->th.th_team;
2311   parent_team = team->t.t_parent;
2312 
2313   master_th->th.th_ident = loc;
2314 
2315 #if OMPT_SUPPORT
2316   if (ompt_enabled) {
2317     master_th->th.ompt_thread_info.state = ompt_state_overhead;
2318   }
2319 #endif
2320 
2321 #if KMP_DEBUG
2322   if (__kmp_tasking_mode != tskm_immediate_exec && !exit_teams) {
2323     KA_TRACE(20, ("__kmp_join_call: T#%d, old team = %p old task_team = %p, "
2324                   "th_task_team = %p\n",
2325                   __kmp_gtid_from_thread(master_th), team,
2326                   team->t.t_task_team[master_th->th.th_task_state],
2327                   master_th->th.th_task_team));
2328     KMP_DEBUG_ASSERT(master_th->th.th_task_team ==
2329                      team->t.t_task_team[master_th->th.th_task_state]);
2330   }
2331 #endif
2332 
2333   if (team->t.t_serialized) {
2334 #if OMP_40_ENABLED
2335     if (master_th->th.th_teams_microtask) {
2336       // We are in teams construct
2337       int level = team->t.t_level;
2338       int tlevel = master_th->th.th_teams_level;
2339       if (level == tlevel) {
2340         // AC: we haven't incremented it earlier at start of teams construct,
2341         //     so do it here - at the end of teams construct
2342         team->t.t_level++;
2343       } else if (level == tlevel + 1) {
2344         // AC: we are exiting parallel inside teams, need to increment
2345         // serialization in order to restore it in the next call to
2346         // __kmpc_end_serialized_parallel
2347         team->t.t_serialized++;
2348       }
2349     }
2350 #endif /* OMP_40_ENABLED */
2351     __kmpc_end_serialized_parallel(loc, gtid);
2352 
2353 #if OMPT_SUPPORT
2354     if (ompt_enabled) {
2355       __kmp_join_restore_state(master_th, parent_team);
2356     }
2357 #endif
2358 
2359     return;
2360   }
2361 
2362   master_active = team->t.t_master_active;
2363 
2364 #if OMP_40_ENABLED
2365   if (!exit_teams)
2366 #endif /* OMP_40_ENABLED */
2367   {
2368     // AC: No barrier for internal teams at exit from teams construct.
2369     //     But there is barrier for external team (league).
2370     __kmp_internal_join(loc, gtid, team);
2371   }
2372 #if OMP_40_ENABLED
2373   else {
2374     master_th->th.th_task_state =
2375         0; // AC: no tasking in teams (out of any parallel)
2376   }
2377 #endif /* OMP_40_ENABLED */
2378 
2379   KMP_MB();
2380 
2381 #if OMPT_SUPPORT
2382   ompt_parallel_id_t parallel_id = team->t.ompt_team_info.parallel_id;
2383 #endif
2384 
2385 #if USE_ITT_BUILD
2386   if (__itt_stack_caller_create_ptr) {
2387     __kmp_itt_stack_caller_destroy(
2388         (__itt_caller)team->t
2389             .t_stack_id); // destroy the stack stitching id after join barrier
2390   }
2391 
2392   // Mark end of "parallel" region for VTune.
2393   if (team->t.t_active_level == 1
2394 #if OMP_40_ENABLED
2395       && !master_th->th.th_teams_microtask /* not in teams construct */
2396 #endif /* OMP_40_ENABLED */
2397       ) {
2398     master_th->th.th_ident = loc;
2399     // only one notification scheme (either "submit" or "forking/joined", not
2400     // both)
2401     if ((__itt_frame_submit_v3_ptr || KMP_ITT_DEBUG) &&
2402         __kmp_forkjoin_frames_mode == 3)
2403       __kmp_itt_frame_submit(gtid, team->t.t_region_time,
2404                              master_th->th.th_frame_time, 0, loc,
2405                              master_th->th.th_team_nproc, 1);
2406     else if ((__itt_frame_end_v3_ptr || KMP_ITT_DEBUG) &&
2407              !__kmp_forkjoin_frames_mode && __kmp_forkjoin_frames)
2408       __kmp_itt_region_joined(gtid);
2409   } // active_level == 1
2410 #endif /* USE_ITT_BUILD */
2411 
2412 #if OMP_40_ENABLED
2413   if (master_th->th.th_teams_microtask && !exit_teams &&
2414       team->t.t_pkfn != (microtask_t)__kmp_teams_master &&
2415       team->t.t_level == master_th->th.th_teams_level + 1) {
2416     // AC: We need to leave the team structure intact at the end of parallel
2417     // inside the teams construct, so that at the next parallel same (hot) team
2418     // works, only adjust nesting levels
2419 
2420     /* Decrement our nested depth level */
2421     team->t.t_level--;
2422     team->t.t_active_level--;
2423     KMP_TEST_THEN_DEC32((kmp_int32 *)&root->r.r_in_parallel);
2424 
2425     /* Restore number of threads in the team if needed */
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 (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 (i = old_num; i < new_num; ++i) {
2436         // Re-initialize thread's barrier data.
2437         int b;
2438         kmp_balign_t *balign = other_threads[i]->th.th_bar;
2439         for (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) {
2455       __kmp_join_ompt(master_th, parent_team, parallel_id, fork_context);
2456     }
2457 #endif
2458 
2459     return;
2460   }
2461 #endif /* OMP_40_ENABLED */
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 OMP_40_ENABLED
2475   if (!master_th->th.th_teams_microtask ||
2476       team->t.t_level > master_th->th.th_teams_level)
2477 #endif /* OMP_40_ENABLED */
2478   {
2479     /* Decrement our nested depth level */
2480     KMP_TEST_THEN_DEC32((kmp_int32 *)&root->r.r_in_parallel);
2481   }
2482   KMP_DEBUG_ASSERT(root->r.r_in_parallel >= 0);
2483 
2484 #if OMPT_SUPPORT && OMPT_TRACE
2485   if (ompt_enabled) {
2486     ompt_task_info_t *task_info = __ompt_get_taskinfo(0);
2487     if (ompt_callbacks.ompt_callback(ompt_event_implicit_task_end)) {
2488       ompt_callbacks.ompt_callback(ompt_event_implicit_task_end)(
2489           parallel_id, task_info->task_id);
2490     }
2491     task_info->frame.exit_runtime_frame = NULL;
2492     task_info->task_id = 0;
2493   }
2494 #endif
2495 
2496   KF_TRACE(10, ("__kmp_join_call1: T#%d, this_thread=%p team=%p\n", 0,
2497                 master_th, team));
2498   __kmp_pop_current_task_from_thread(master_th);
2499 
2500 #if OMP_40_ENABLED && KMP_AFFINITY_SUPPORTED
2501   // Restore master thread's partition.
2502   master_th->th.th_first_place = team->t.t_first_place;
2503   master_th->th.th_last_place = team->t.t_last_place;
2504 #endif /* OMP_40_ENABLED */
2505 
2506   updateHWFPControl(team);
2507 
2508   if (root->r.r_active != master_active)
2509     root->r.r_active = master_active;
2510 
2511   __kmp_free_team(root, team USE_NESTED_HOT_ARG(
2512                             master_th)); // this will free worker threads
2513 
2514   /* this race was fun to find. make sure the following is in the critical
2515      region otherwise assertions may fail occasionally since the old team may be
2516      reallocated and the hierarchy appears inconsistent. it is actually safe to
2517      run and won't cause any bugs, but will cause those assertion failures. it's
2518      only one deref&assign so might as well put this in the critical region */
2519   master_th->th.th_team = parent_team;
2520   master_th->th.th_team_nproc = parent_team->t.t_nproc;
2521   master_th->th.th_team_master = parent_team->t.t_threads[0];
2522   master_th->th.th_team_serialized = parent_team->t.t_serialized;
2523 
2524   /* restore serialized team, if need be */
2525   if (parent_team->t.t_serialized &&
2526       parent_team != master_th->th.th_serial_team &&
2527       parent_team != root->r.r_root_team) {
2528     __kmp_free_team(root,
2529                     master_th->th.th_serial_team USE_NESTED_HOT_ARG(NULL));
2530     master_th->th.th_serial_team = parent_team;
2531   }
2532 
2533   if (__kmp_tasking_mode != tskm_immediate_exec) {
2534     if (master_th->th.th_task_state_top >
2535         0) { // Restore task state from memo stack
2536       KMP_DEBUG_ASSERT(master_th->th.th_task_state_memo_stack);
2537       // Remember master's state if we re-use this nested hot team
2538       master_th->th.th_task_state_memo_stack[master_th->th.th_task_state_top] =
2539           master_th->th.th_task_state;
2540       --master_th->th.th_task_state_top; // pop
2541       // Now restore state at this level
2542       master_th->th.th_task_state =
2543           master_th->th
2544               .th_task_state_memo_stack[master_th->th.th_task_state_top];
2545     }
2546     // Copy the task team from the parent team to the master thread
2547     master_th->th.th_task_team =
2548         parent_team->t.t_task_team[master_th->th.th_task_state];
2549     KA_TRACE(20,
2550              ("__kmp_join_call: Master T#%d restoring task_team %p / team %p\n",
2551               __kmp_gtid_from_thread(master_th), master_th->th.th_task_team,
2552               parent_team));
2553   }
2554 
2555   // TODO: GEH - cannot do this assertion because root thread not set up as
2556   // executing
2557   // KMP_ASSERT( master_th->th.th_current_task->td_flags.executing == 0 );
2558   master_th->th.th_current_task->td_flags.executing = 1;
2559 
2560   __kmp_release_bootstrap_lock(&__kmp_forkjoin_lock);
2561 
2562 #if OMPT_SUPPORT
2563   if (ompt_enabled) {
2564     __kmp_join_ompt(master_th, parent_team, parallel_id, fork_context);
2565   }
2566 #endif
2567 
2568   KMP_MB();
2569   KA_TRACE(20, ("__kmp_join_call: exit T#%d\n", gtid));
2570 }
2571 
2572 /* Check whether we should push an internal control record onto the
2573    serial team stack.  If so, do it.  */
2574 void __kmp_save_internal_controls(kmp_info_t *thread) {
2575 
2576   if (thread->th.th_team != thread->th.th_serial_team) {
2577     return;
2578   }
2579   if (thread->th.th_team->t.t_serialized > 1) {
2580     int push = 0;
2581 
2582     if (thread->th.th_team->t.t_control_stack_top == NULL) {
2583       push = 1;
2584     } else {
2585       if (thread->th.th_team->t.t_control_stack_top->serial_nesting_level !=
2586           thread->th.th_team->t.t_serialized) {
2587         push = 1;
2588       }
2589     }
2590     if (push) { /* push a record on the serial team's stack */
2591       kmp_internal_control_t *control =
2592           (kmp_internal_control_t *)__kmp_allocate(
2593               sizeof(kmp_internal_control_t));
2594 
2595       copy_icvs(control, &thread->th.th_current_task->td_icvs);
2596 
2597       control->serial_nesting_level = thread->th.th_team->t.t_serialized;
2598 
2599       control->next = thread->th.th_team->t.t_control_stack_top;
2600       thread->th.th_team->t.t_control_stack_top = control;
2601     }
2602   }
2603 }
2604 
2605 /* Changes set_nproc */
2606 void __kmp_set_num_threads(int new_nth, int gtid) {
2607   kmp_info_t *thread;
2608   kmp_root_t *root;
2609 
2610   KF_TRACE(10, ("__kmp_set_num_threads: new __kmp_nth = %d\n", new_nth));
2611   KMP_DEBUG_ASSERT(__kmp_init_serial);
2612 
2613   if (new_nth < 1)
2614     new_nth = 1;
2615   else if (new_nth > __kmp_max_nth)
2616     new_nth = __kmp_max_nth;
2617 
2618   KMP_COUNT_VALUE(OMP_set_numthreads, new_nth);
2619   thread = __kmp_threads[gtid];
2620 
2621   __kmp_save_internal_controls(thread);
2622 
2623   set__nproc(thread, new_nth);
2624 
2625   // If this omp_set_num_threads() call will cause the hot team size to be
2626   // reduced (in the absence of a num_threads clause), then reduce it now,
2627   // rather than waiting for the next parallel region.
2628   root = thread->th.th_root;
2629   if (__kmp_init_parallel && (!root->r.r_active) &&
2630       (root->r.r_hot_team->t.t_nproc > new_nth)
2631 #if KMP_NESTED_HOT_TEAMS
2632       && __kmp_hot_teams_max_level && !__kmp_hot_teams_mode
2633 #endif
2634       ) {
2635     kmp_team_t *hot_team = root->r.r_hot_team;
2636     int f;
2637 
2638     __kmp_acquire_bootstrap_lock(&__kmp_forkjoin_lock);
2639 
2640     // Release the extra threads we don't need any more.
2641     for (f = new_nth; f < hot_team->t.t_nproc; f++) {
2642       KMP_DEBUG_ASSERT(hot_team->t.t_threads[f] != NULL);
2643       if (__kmp_tasking_mode != tskm_immediate_exec) {
2644         // When decreasing team size, threads no longer in the team should unref
2645         // task team.
2646         hot_team->t.t_threads[f]->th.th_task_team = NULL;
2647       }
2648       __kmp_free_thread(hot_team->t.t_threads[f]);
2649       hot_team->t.t_threads[f] = NULL;
2650     }
2651     hot_team->t.t_nproc = new_nth;
2652 #if KMP_NESTED_HOT_TEAMS
2653     if (thread->th.th_hot_teams) {
2654       KMP_DEBUG_ASSERT(hot_team == thread->th.th_hot_teams[0].hot_team);
2655       thread->th.th_hot_teams[0].hot_team_nth = new_nth;
2656     }
2657 #endif
2658 
2659     __kmp_release_bootstrap_lock(&__kmp_forkjoin_lock);
2660 
2661     // Update the t_nproc field in the threads that are still active.
2662     for (f = 0; f < new_nth; f++) {
2663       KMP_DEBUG_ASSERT(hot_team->t.t_threads[f] != NULL);
2664       hot_team->t.t_threads[f]->th.th_team_nproc = new_nth;
2665     }
2666     // Special flag in case omp_set_num_threads() call
2667     hot_team->t.t_size_changed = -1;
2668   }
2669 }
2670 
2671 /* Changes max_active_levels */
2672 void __kmp_set_max_active_levels(int gtid, int max_active_levels) {
2673   kmp_info_t *thread;
2674 
2675   KF_TRACE(10, ("__kmp_set_max_active_levels: new max_active_levels for thread "
2676                 "%d = (%d)\n",
2677                 gtid, max_active_levels));
2678   KMP_DEBUG_ASSERT(__kmp_init_serial);
2679 
2680   // validate max_active_levels
2681   if (max_active_levels < 0) {
2682     KMP_WARNING(ActiveLevelsNegative, max_active_levels);
2683     // We ignore this call if the user has specified a negative value.
2684     // The current setting won't be changed. The last valid setting will be
2685     // used. A warning will be issued (if warnings are allowed as controlled by
2686     // the KMP_WARNINGS env var).
2687     KF_TRACE(10, ("__kmp_set_max_active_levels: the call is ignored: new "
2688                   "max_active_levels for thread %d = (%d)\n",
2689                   gtid, max_active_levels));
2690     return;
2691   }
2692   if (max_active_levels <= KMP_MAX_ACTIVE_LEVELS_LIMIT) {
2693     // it's OK, the max_active_levels is within the valid range: [ 0;
2694     // KMP_MAX_ACTIVE_LEVELS_LIMIT ]
2695     // We allow a zero value. (implementation defined behavior)
2696   } else {
2697     KMP_WARNING(ActiveLevelsExceedLimit, max_active_levels,
2698                 KMP_MAX_ACTIVE_LEVELS_LIMIT);
2699     max_active_levels = KMP_MAX_ACTIVE_LEVELS_LIMIT;
2700     // Current upper limit is MAX_INT. (implementation defined behavior)
2701     // If the input exceeds the upper limit, we correct the input to be the
2702     // upper limit. (implementation defined behavior)
2703     // Actually, the flow should never get here until we use MAX_INT limit.
2704   }
2705   KF_TRACE(10, ("__kmp_set_max_active_levels: after validation: new "
2706                 "max_active_levels for thread %d = (%d)\n",
2707                 gtid, max_active_levels));
2708 
2709   thread = __kmp_threads[gtid];
2710 
2711   __kmp_save_internal_controls(thread);
2712 
2713   set__max_active_levels(thread, max_active_levels);
2714 }
2715 
2716 /* Gets max_active_levels */
2717 int __kmp_get_max_active_levels(int gtid) {
2718   kmp_info_t *thread;
2719 
2720   KF_TRACE(10, ("__kmp_get_max_active_levels: thread %d\n", gtid));
2721   KMP_DEBUG_ASSERT(__kmp_init_serial);
2722 
2723   thread = __kmp_threads[gtid];
2724   KMP_DEBUG_ASSERT(thread->th.th_current_task);
2725   KF_TRACE(10, ("__kmp_get_max_active_levels: thread %d, curtask=%p, "
2726                 "curtask_maxaclevel=%d\n",
2727                 gtid, thread->th.th_current_task,
2728                 thread->th.th_current_task->td_icvs.max_active_levels));
2729   return thread->th.th_current_task->td_icvs.max_active_levels;
2730 }
2731 
2732 /* Changes def_sched_var ICV values (run-time schedule kind and chunk) */
2733 void __kmp_set_schedule(int gtid, kmp_sched_t kind, int chunk) {
2734   kmp_info_t *thread;
2735   //    kmp_team_t *team;
2736 
2737   KF_TRACE(10, ("__kmp_set_schedule: new schedule for thread %d = (%d, %d)\n",
2738                 gtid, (int)kind, chunk));
2739   KMP_DEBUG_ASSERT(__kmp_init_serial);
2740 
2741   // Check if the kind parameter is valid, correct if needed.
2742   // Valid parameters should fit in one of two intervals - standard or extended:
2743   //       <lower>, <valid>, <upper_std>, <lower_ext>, <valid>, <upper>
2744   // 2008-01-25: 0,  1 - 4,       5,         100,     101 - 102, 103
2745   if (kind <= kmp_sched_lower || kind >= kmp_sched_upper ||
2746       (kind <= kmp_sched_lower_ext && kind >= kmp_sched_upper_std)) {
2747     // TODO: Hint needs attention in case we change the default schedule.
2748     __kmp_msg(kmp_ms_warning, KMP_MSG(ScheduleKindOutOfRange, kind),
2749               KMP_HNT(DefaultScheduleKindUsed, "static, no chunk"),
2750               __kmp_msg_null);
2751     kind = kmp_sched_default;
2752     chunk = 0; // ignore chunk value in case of bad kind
2753   }
2754 
2755   thread = __kmp_threads[gtid];
2756 
2757   __kmp_save_internal_controls(thread);
2758 
2759   if (kind < kmp_sched_upper_std) {
2760     if (kind == kmp_sched_static && chunk < KMP_DEFAULT_CHUNK) {
2761       // differ static chunked vs. unchunked:  chunk should be invalid to
2762       // indicate unchunked schedule (which is the default)
2763       thread->th.th_current_task->td_icvs.sched.r_sched_type = kmp_sch_static;
2764     } else {
2765       thread->th.th_current_task->td_icvs.sched.r_sched_type =
2766           __kmp_sch_map[kind - kmp_sched_lower - 1];
2767     }
2768   } else {
2769     //    __kmp_sch_map[ kind - kmp_sched_lower_ext + kmp_sched_upper_std -
2770     //    kmp_sched_lower - 2 ];
2771     thread->th.th_current_task->td_icvs.sched.r_sched_type =
2772         __kmp_sch_map[kind - kmp_sched_lower_ext + kmp_sched_upper_std -
2773                       kmp_sched_lower - 2];
2774   }
2775   if (kind == kmp_sched_auto || chunk < 1) {
2776     // ignore parameter chunk for schedule auto
2777     thread->th.th_current_task->td_icvs.sched.chunk = KMP_DEFAULT_CHUNK;
2778   } else {
2779     thread->th.th_current_task->td_icvs.sched.chunk = chunk;
2780   }
2781 }
2782 
2783 /* Gets def_sched_var ICV values */
2784 void __kmp_get_schedule(int gtid, kmp_sched_t *kind, int *chunk) {
2785   kmp_info_t *thread;
2786   enum sched_type th_type;
2787 
2788   KF_TRACE(10, ("__kmp_get_schedule: thread %d\n", gtid));
2789   KMP_DEBUG_ASSERT(__kmp_init_serial);
2790 
2791   thread = __kmp_threads[gtid];
2792 
2793   th_type = thread->th.th_current_task->td_icvs.sched.r_sched_type;
2794 
2795   switch (th_type) {
2796   case kmp_sch_static:
2797   case kmp_sch_static_greedy:
2798   case kmp_sch_static_balanced:
2799     *kind = kmp_sched_static;
2800     *chunk = 0; // chunk was not set, try to show this fact via zero value
2801     return;
2802   case kmp_sch_static_chunked:
2803     *kind = kmp_sched_static;
2804     break;
2805   case kmp_sch_dynamic_chunked:
2806     *kind = kmp_sched_dynamic;
2807     break;
2808   case kmp_sch_guided_chunked:
2809   case kmp_sch_guided_iterative_chunked:
2810   case kmp_sch_guided_analytical_chunked:
2811     *kind = kmp_sched_guided;
2812     break;
2813   case kmp_sch_auto:
2814     *kind = kmp_sched_auto;
2815     break;
2816   case kmp_sch_trapezoidal:
2817     *kind = kmp_sched_trapezoidal;
2818     break;
2819 #if KMP_STATIC_STEAL_ENABLED
2820   case kmp_sch_static_steal:
2821     *kind = kmp_sched_static_steal;
2822     break;
2823 #endif
2824   default:
2825     KMP_FATAL(UnknownSchedulingType, th_type);
2826   }
2827 
2828   *chunk = thread->th.th_current_task->td_icvs.sched.chunk;
2829 }
2830 
2831 int __kmp_get_ancestor_thread_num(int gtid, int level) {
2832 
2833   int ii, dd;
2834   kmp_team_t *team;
2835   kmp_info_t *thr;
2836 
2837   KF_TRACE(10, ("__kmp_get_ancestor_thread_num: thread %d %d\n", gtid, level));
2838   KMP_DEBUG_ASSERT(__kmp_init_serial);
2839 
2840   // validate level
2841   if (level == 0)
2842     return 0;
2843   if (level < 0)
2844     return -1;
2845   thr = __kmp_threads[gtid];
2846   team = thr->th.th_team;
2847   ii = team->t.t_level;
2848   if (level > ii)
2849     return -1;
2850 
2851 #if OMP_40_ENABLED
2852   if (thr->th.th_teams_microtask) {
2853     // AC: we are in teams region where multiple nested teams have same level
2854     int tlevel = thr->th.th_teams_level; // the level of the teams construct
2855     if (level <=
2856         tlevel) { // otherwise usual algorithm works (will not touch the teams)
2857       KMP_DEBUG_ASSERT(ii >= tlevel);
2858       // AC: As we need to pass by the teams league, we need to artificially
2859       // increase ii
2860       if (ii == tlevel) {
2861         ii += 2; // three teams have same level
2862       } else {
2863         ii++; // two teams have same level
2864       }
2865     }
2866   }
2867 #endif
2868 
2869   if (ii == level)
2870     return __kmp_tid_from_gtid(gtid);
2871 
2872   dd = team->t.t_serialized;
2873   level++;
2874   while (ii > level) {
2875     for (dd = team->t.t_serialized; (dd > 0) && (ii > level); dd--, ii--) {
2876     }
2877     if ((team->t.t_serialized) && (!dd)) {
2878       team = team->t.t_parent;
2879       continue;
2880     }
2881     if (ii > level) {
2882       team = team->t.t_parent;
2883       dd = team->t.t_serialized;
2884       ii--;
2885     }
2886   }
2887 
2888   return (dd > 1) ? (0) : (team->t.t_master_tid);
2889 }
2890 
2891 int __kmp_get_team_size(int gtid, int level) {
2892 
2893   int ii, dd;
2894   kmp_team_t *team;
2895   kmp_info_t *thr;
2896 
2897   KF_TRACE(10, ("__kmp_get_team_size: thread %d %d\n", gtid, level));
2898   KMP_DEBUG_ASSERT(__kmp_init_serial);
2899 
2900   // validate level
2901   if (level == 0)
2902     return 1;
2903   if (level < 0)
2904     return -1;
2905   thr = __kmp_threads[gtid];
2906   team = thr->th.th_team;
2907   ii = team->t.t_level;
2908   if (level > ii)
2909     return -1;
2910 
2911 #if OMP_40_ENABLED
2912   if (thr->th.th_teams_microtask) {
2913     // AC: we are in teams region where multiple nested teams have same level
2914     int tlevel = thr->th.th_teams_level; // the level of the teams construct
2915     if (level <=
2916         tlevel) { // otherwise usual algorithm works (will not touch the teams)
2917       KMP_DEBUG_ASSERT(ii >= tlevel);
2918       // AC: As we need to pass by the teams league, we need to artificially
2919       // increase ii
2920       if (ii == tlevel) {
2921         ii += 2; // three teams have same level
2922       } else {
2923         ii++; // two teams have same level
2924       }
2925     }
2926   }
2927 #endif
2928 
2929   while (ii > level) {
2930     for (dd = team->t.t_serialized; (dd > 0) && (ii > level); dd--, ii--) {
2931     }
2932     if (team->t.t_serialized && (!dd)) {
2933       team = team->t.t_parent;
2934       continue;
2935     }
2936     if (ii > level) {
2937       team = team->t.t_parent;
2938       ii--;
2939     }
2940   }
2941 
2942   return team->t.t_nproc;
2943 }
2944 
2945 kmp_r_sched_t __kmp_get_schedule_global() {
2946   // This routine created because pairs (__kmp_sched, __kmp_chunk) and
2947   // (__kmp_static, __kmp_guided) may be changed by kmp_set_defaults
2948   // independently. So one can get the updated schedule here.
2949 
2950   kmp_r_sched_t r_sched;
2951 
2952   // create schedule from 4 globals: __kmp_sched, __kmp_chunk, __kmp_static,
2953   // __kmp_guided. __kmp_sched should keep original value, so that user can set
2954   // KMP_SCHEDULE multiple times, and thus have different run-time schedules in
2955   // different roots (even in OMP 2.5)
2956   if (__kmp_sched == kmp_sch_static) {
2957     r_sched.r_sched_type = __kmp_static; // replace STATIC with more detailed
2958     // schedule (balanced or greedy)
2959   } else if (__kmp_sched == kmp_sch_guided_chunked) {
2960     r_sched.r_sched_type = __kmp_guided; // replace GUIDED with more detailed
2961     // schedule (iterative or analytical)
2962   } else {
2963     r_sched.r_sched_type =
2964         __kmp_sched; // (STATIC_CHUNKED), or (DYNAMIC_CHUNKED), or other
2965   }
2966 
2967   if (__kmp_chunk < KMP_DEFAULT_CHUNK) { // __kmp_chunk may be wrong here (if it
2968     // was not ever set)
2969     r_sched.chunk = KMP_DEFAULT_CHUNK;
2970   } else {
2971     r_sched.chunk = __kmp_chunk;
2972   }
2973 
2974   return r_sched;
2975 }
2976 
2977 /* Allocate (realloc == FALSE) * or reallocate (realloc == TRUE)
2978    at least argc number of *t_argv entries for the requested team. */
2979 static void __kmp_alloc_argv_entries(int argc, kmp_team_t *team, int realloc) {
2980 
2981   KMP_DEBUG_ASSERT(team);
2982   if (!realloc || argc > team->t.t_max_argc) {
2983 
2984     KA_TRACE(100, ("__kmp_alloc_argv_entries: team %d: needed entries=%d, "
2985                    "current entries=%d\n",
2986                    team->t.t_id, argc, (realloc) ? team->t.t_max_argc : 0));
2987     /* if previously allocated heap space for args, free them */
2988     if (realloc && team->t.t_argv != &team->t.t_inline_argv[0])
2989       __kmp_free((void *)team->t.t_argv);
2990 
2991     if (argc <= KMP_INLINE_ARGV_ENTRIES) {
2992       /* use unused space in the cache line for arguments */
2993       team->t.t_max_argc = KMP_INLINE_ARGV_ENTRIES;
2994       KA_TRACE(100, ("__kmp_alloc_argv_entries: team %d: inline allocate %d "
2995                      "argv entries\n",
2996                      team->t.t_id, team->t.t_max_argc));
2997       team->t.t_argv = &team->t.t_inline_argv[0];
2998       if (__kmp_storage_map) {
2999         __kmp_print_storage_map_gtid(
3000             -1, &team->t.t_inline_argv[0],
3001             &team->t.t_inline_argv[KMP_INLINE_ARGV_ENTRIES],
3002             (sizeof(void *) * KMP_INLINE_ARGV_ENTRIES), "team_%d.t_inline_argv",
3003             team->t.t_id);
3004       }
3005     } else {
3006       /* allocate space for arguments in the heap */
3007       team->t.t_max_argc = (argc <= (KMP_MIN_MALLOC_ARGV_ENTRIES >> 1))
3008                                ? KMP_MIN_MALLOC_ARGV_ENTRIES
3009                                : 2 * argc;
3010       KA_TRACE(100, ("__kmp_alloc_argv_entries: team %d: dynamic allocate %d "
3011                      "argv entries\n",
3012                      team->t.t_id, team->t.t_max_argc));
3013       team->t.t_argv =
3014           (void **)__kmp_page_allocate(sizeof(void *) * team->t.t_max_argc);
3015       if (__kmp_storage_map) {
3016         __kmp_print_storage_map_gtid(-1, &team->t.t_argv[0],
3017                                      &team->t.t_argv[team->t.t_max_argc],
3018                                      sizeof(void *) * team->t.t_max_argc,
3019                                      "team_%d.t_argv", team->t.t_id);
3020       }
3021     }
3022   }
3023 }
3024 
3025 static void __kmp_allocate_team_arrays(kmp_team_t *team, int max_nth) {
3026   int i;
3027   int num_disp_buff = max_nth > 1 ? __kmp_dispatch_num_buffers : 2;
3028   team->t.t_threads =
3029       (kmp_info_t **)__kmp_allocate(sizeof(kmp_info_t *) * max_nth);
3030   team->t.t_disp_buffer = (dispatch_shared_info_t *)__kmp_allocate(
3031       sizeof(dispatch_shared_info_t) * num_disp_buff);
3032   team->t.t_dispatch =
3033       (kmp_disp_t *)__kmp_allocate(sizeof(kmp_disp_t) * max_nth);
3034   team->t.t_implicit_task_taskdata =
3035       (kmp_taskdata_t *)__kmp_allocate(sizeof(kmp_taskdata_t) * max_nth);
3036   team->t.t_max_nproc = max_nth;
3037 
3038   /* setup dispatch buffers */
3039   for (i = 0; i < num_disp_buff; ++i) {
3040     team->t.t_disp_buffer[i].buffer_index = i;
3041 #if OMP_45_ENABLED
3042     team->t.t_disp_buffer[i].doacross_buf_idx = i;
3043 #endif
3044   }
3045 }
3046 
3047 static void __kmp_free_team_arrays(kmp_team_t *team) {
3048   /* Note: this does not free the threads in t_threads (__kmp_free_threads) */
3049   int i;
3050   for (i = 0; i < team->t.t_max_nproc; ++i) {
3051     if (team->t.t_dispatch[i].th_disp_buffer != NULL) {
3052       __kmp_free(team->t.t_dispatch[i].th_disp_buffer);
3053       team->t.t_dispatch[i].th_disp_buffer = NULL;
3054     }; // if
3055   }; // for
3056   __kmp_free(team->t.t_threads);
3057   __kmp_free(team->t.t_disp_buffer);
3058   __kmp_free(team->t.t_dispatch);
3059   __kmp_free(team->t.t_implicit_task_taskdata);
3060   team->t.t_threads = NULL;
3061   team->t.t_disp_buffer = NULL;
3062   team->t.t_dispatch = NULL;
3063   team->t.t_implicit_task_taskdata = 0;
3064 }
3065 
3066 static void __kmp_reallocate_team_arrays(kmp_team_t *team, int max_nth) {
3067   kmp_info_t **oldThreads = team->t.t_threads;
3068 
3069   __kmp_free(team->t.t_disp_buffer);
3070   __kmp_free(team->t.t_dispatch);
3071   __kmp_free(team->t.t_implicit_task_taskdata);
3072   __kmp_allocate_team_arrays(team, max_nth);
3073 
3074   KMP_MEMCPY(team->t.t_threads, oldThreads,
3075              team->t.t_nproc * sizeof(kmp_info_t *));
3076 
3077   __kmp_free(oldThreads);
3078 }
3079 
3080 static kmp_internal_control_t __kmp_get_global_icvs(void) {
3081 
3082   kmp_r_sched_t r_sched =
3083       __kmp_get_schedule_global(); // get current state of scheduling globals
3084 
3085 #if OMP_40_ENABLED
3086   KMP_DEBUG_ASSERT(__kmp_nested_proc_bind.used > 0);
3087 #endif /* OMP_40_ENABLED */
3088 
3089   kmp_internal_control_t g_icvs = {
3090     0, // int serial_nesting_level; //corresponds to value of th_team_serialized
3091     (kmp_int8)__kmp_dflt_nested, // int nested; //internal control
3092     // for nested parallelism (per thread)
3093     (kmp_int8)__kmp_global.g.g_dynamic, // internal control for dynamic
3094     // adjustment of threads (per thread)
3095     (kmp_int8)__kmp_env_blocktime, // int bt_set; //internal control for
3096     // whether blocktime is explicitly set
3097     __kmp_dflt_blocktime, // int blocktime; //internal control for blocktime
3098 #if KMP_USE_MONITOR
3099     __kmp_bt_intervals, // int bt_intervals; //internal control for blocktime
3100 // intervals
3101 #endif
3102     __kmp_dflt_team_nth, // int nproc; //internal control for # of threads for
3103     // next parallel region (per thread)
3104     // (use a max ub on value if __kmp_parallel_initialize not called yet)
3105     __kmp_dflt_max_active_levels, // int max_active_levels; //internal control
3106     // for max_active_levels
3107     r_sched, // kmp_r_sched_t sched; //internal control for runtime schedule
3108 // {sched,chunk} pair
3109 #if OMP_40_ENABLED
3110     __kmp_nested_proc_bind.bind_types[0],
3111     __kmp_default_device,
3112 #endif /* OMP_40_ENABLED */
3113     NULL // struct kmp_internal_control *next;
3114   };
3115 
3116   return g_icvs;
3117 }
3118 
3119 static kmp_internal_control_t __kmp_get_x_global_icvs(const kmp_team_t *team) {
3120 
3121   kmp_internal_control_t gx_icvs;
3122   gx_icvs.serial_nesting_level =
3123       0; // probably =team->t.t_serial like in save_inter_controls
3124   copy_icvs(&gx_icvs, &team->t.t_threads[0]->th.th_current_task->td_icvs);
3125   gx_icvs.next = NULL;
3126 
3127   return gx_icvs;
3128 }
3129 
3130 static void __kmp_initialize_root(kmp_root_t *root) {
3131   int f;
3132   kmp_team_t *root_team;
3133   kmp_team_t *hot_team;
3134   int hot_team_max_nth;
3135   kmp_r_sched_t r_sched =
3136       __kmp_get_schedule_global(); // get current state of scheduling globals
3137   kmp_internal_control_t r_icvs = __kmp_get_global_icvs();
3138   KMP_DEBUG_ASSERT(root);
3139   KMP_ASSERT(!root->r.r_begin);
3140 
3141   /* setup the root state structure */
3142   __kmp_init_lock(&root->r.r_begin_lock);
3143   root->r.r_begin = FALSE;
3144   root->r.r_active = FALSE;
3145   root->r.r_in_parallel = 0;
3146   root->r.r_blocktime = __kmp_dflt_blocktime;
3147   root->r.r_nested = __kmp_dflt_nested;
3148   root->r.r_cg_nthreads = 1;
3149 
3150   /* setup the root team for this task */
3151   /* allocate the root team structure */
3152   KF_TRACE(10, ("__kmp_initialize_root: before root_team\n"));
3153 
3154   root_team =
3155       __kmp_allocate_team(root,
3156                           1, // new_nproc
3157                           1, // max_nproc
3158 #if OMPT_SUPPORT
3159                           0, // root parallel id
3160 #endif
3161 #if OMP_40_ENABLED
3162                           __kmp_nested_proc_bind.bind_types[0],
3163 #endif
3164                           &r_icvs,
3165                           0 // argc
3166                           USE_NESTED_HOT_ARG(NULL) // master thread is unknown
3167                           );
3168 #if USE_DEBUGGER
3169   // Non-NULL value should be assigned to make the debugger display the root
3170   // team.
3171   TCW_SYNC_PTR(root_team->t.t_pkfn, (microtask_t)(~0));
3172 #endif
3173 
3174   KF_TRACE(10, ("__kmp_initialize_root: after root_team = %p\n", root_team));
3175 
3176   root->r.r_root_team = root_team;
3177   root_team->t.t_control_stack_top = NULL;
3178 
3179   /* initialize root team */
3180   root_team->t.t_threads[0] = NULL;
3181   root_team->t.t_nproc = 1;
3182   root_team->t.t_serialized = 1;
3183   // TODO???: root_team->t.t_max_active_levels = __kmp_dflt_max_active_levels;
3184   root_team->t.t_sched.r_sched_type = r_sched.r_sched_type;
3185   root_team->t.t_sched.chunk = r_sched.chunk;
3186   KA_TRACE(
3187       20,
3188       ("__kmp_initialize_root: init root team %d arrived: join=%u, plain=%u\n",
3189        root_team->t.t_id, KMP_INIT_BARRIER_STATE, KMP_INIT_BARRIER_STATE));
3190 
3191   /* setup the  hot team for this task */
3192   /* allocate the hot team structure */
3193   KF_TRACE(10, ("__kmp_initialize_root: before hot_team\n"));
3194 
3195   hot_team =
3196       __kmp_allocate_team(root,
3197                           1, // new_nproc
3198                           __kmp_dflt_team_nth_ub * 2, // max_nproc
3199 #if OMPT_SUPPORT
3200                           0, // root parallel id
3201 #endif
3202 #if OMP_40_ENABLED
3203                           __kmp_nested_proc_bind.bind_types[0],
3204 #endif
3205                           &r_icvs,
3206                           0 // argc
3207                           USE_NESTED_HOT_ARG(NULL) // master thread is unknown
3208                           );
3209   KF_TRACE(10, ("__kmp_initialize_root: after hot_team = %p\n", hot_team));
3210 
3211   root->r.r_hot_team = hot_team;
3212   root_team->t.t_control_stack_top = NULL;
3213 
3214   /* first-time initialization */
3215   hot_team->t.t_parent = root_team;
3216 
3217   /* initialize hot team */
3218   hot_team_max_nth = hot_team->t.t_max_nproc;
3219   for (f = 0; f < hot_team_max_nth; ++f) {
3220     hot_team->t.t_threads[f] = NULL;
3221   }; // for
3222   hot_team->t.t_nproc = 1;
3223   // TODO???: hot_team->t.t_max_active_levels = __kmp_dflt_max_active_levels;
3224   hot_team->t.t_sched.r_sched_type = r_sched.r_sched_type;
3225   hot_team->t.t_sched.chunk = r_sched.chunk;
3226   hot_team->t.t_size_changed = 0;
3227 }
3228 
3229 #ifdef KMP_DEBUG
3230 
3231 typedef struct kmp_team_list_item {
3232   kmp_team_p const *entry;
3233   struct kmp_team_list_item *next;
3234 } kmp_team_list_item_t;
3235 typedef kmp_team_list_item_t *kmp_team_list_t;
3236 
3237 static void __kmp_print_structure_team_accum( // Add team to list of teams.
3238     kmp_team_list_t list, // List of teams.
3239     kmp_team_p const *team // Team to add.
3240     ) {
3241 
3242   // List must terminate with item where both entry and next are NULL.
3243   // Team is added to the list only once.
3244   // List is sorted in ascending order by team id.
3245   // Team id is *not* a key.
3246 
3247   kmp_team_list_t l;
3248 
3249   KMP_DEBUG_ASSERT(list != NULL);
3250   if (team == NULL) {
3251     return;
3252   }; // if
3253 
3254   __kmp_print_structure_team_accum(list, team->t.t_parent);
3255   __kmp_print_structure_team_accum(list, team->t.t_next_pool);
3256 
3257   // Search list for the team.
3258   l = list;
3259   while (l->next != NULL && l->entry != team) {
3260     l = l->next;
3261   }; // while
3262   if (l->next != NULL) {
3263     return; // Team has been added before, exit.
3264   }; // if
3265 
3266   // Team is not found. Search list again for insertion point.
3267   l = list;
3268   while (l->next != NULL && l->entry->t.t_id <= team->t.t_id) {
3269     l = l->next;
3270   }; // while
3271 
3272   // Insert team.
3273   {
3274     kmp_team_list_item_t *item = (kmp_team_list_item_t *)KMP_INTERNAL_MALLOC(
3275         sizeof(kmp_team_list_item_t));
3276     *item = *l;
3277     l->entry = team;
3278     l->next = item;
3279   }
3280 }
3281 
3282 static void __kmp_print_structure_team(char const *title, kmp_team_p const *team
3283 
3284                                        ) {
3285   __kmp_printf("%s", title);
3286   if (team != NULL) {
3287     __kmp_printf("%2x %p\n", team->t.t_id, team);
3288   } else {
3289     __kmp_printf(" - (nil)\n");
3290   }; // if
3291 }
3292 
3293 static void __kmp_print_structure_thread(char const *title,
3294                                          kmp_info_p const *thread) {
3295   __kmp_printf("%s", title);
3296   if (thread != NULL) {
3297     __kmp_printf("%2d %p\n", thread->th.th_info.ds.ds_gtid, thread);
3298   } else {
3299     __kmp_printf(" - (nil)\n");
3300   }; // if
3301 }
3302 
3303 void __kmp_print_structure(void) {
3304 
3305   kmp_team_list_t list;
3306 
3307   // Initialize list of teams.
3308   list =
3309       (kmp_team_list_item_t *)KMP_INTERNAL_MALLOC(sizeof(kmp_team_list_item_t));
3310   list->entry = NULL;
3311   list->next = NULL;
3312 
3313   __kmp_printf("\n------------------------------\nGlobal Thread "
3314                "Table\n------------------------------\n");
3315   {
3316     int gtid;
3317     for (gtid = 0; gtid < __kmp_threads_capacity; ++gtid) {
3318       __kmp_printf("%2d", gtid);
3319       if (__kmp_threads != NULL) {
3320         __kmp_printf(" %p", __kmp_threads[gtid]);
3321       }; // if
3322       if (__kmp_root != NULL) {
3323         __kmp_printf(" %p", __kmp_root[gtid]);
3324       }; // if
3325       __kmp_printf("\n");
3326     }; // for gtid
3327   }
3328 
3329   // Print out __kmp_threads array.
3330   __kmp_printf("\n------------------------------\nThreads\n--------------------"
3331                "----------\n");
3332   if (__kmp_threads != NULL) {
3333     int gtid;
3334     for (gtid = 0; gtid < __kmp_threads_capacity; ++gtid) {
3335       kmp_info_t const *thread = __kmp_threads[gtid];
3336       if (thread != NULL) {
3337         __kmp_printf("GTID %2d %p:\n", gtid, thread);
3338         __kmp_printf("    Our Root:        %p\n", thread->th.th_root);
3339         __kmp_print_structure_team("    Our Team:     ", thread->th.th_team);
3340         __kmp_print_structure_team("    Serial Team:  ",
3341                                    thread->th.th_serial_team);
3342         __kmp_printf("    Threads:      %2d\n", thread->th.th_team_nproc);
3343         __kmp_print_structure_thread("    Master:       ",
3344                                      thread->th.th_team_master);
3345         __kmp_printf("    Serialized?:  %2d\n", thread->th.th_team_serialized);
3346         __kmp_printf("    Set NProc:    %2d\n", thread->th.th_set_nproc);
3347 #if OMP_40_ENABLED
3348         __kmp_printf("    Set Proc Bind: %2d\n", thread->th.th_set_proc_bind);
3349 #endif
3350         __kmp_print_structure_thread("    Next in pool: ",
3351                                      thread->th.th_next_pool);
3352         __kmp_printf("\n");
3353         __kmp_print_structure_team_accum(list, thread->th.th_team);
3354         __kmp_print_structure_team_accum(list, thread->th.th_serial_team);
3355       }; // if
3356     }; // for gtid
3357   } else {
3358     __kmp_printf("Threads array is not allocated.\n");
3359   }; // if
3360 
3361   // Print out __kmp_root array.
3362   __kmp_printf("\n------------------------------\nUbers\n----------------------"
3363                "--------\n");
3364   if (__kmp_root != NULL) {
3365     int gtid;
3366     for (gtid = 0; gtid < __kmp_threads_capacity; ++gtid) {
3367       kmp_root_t const *root = __kmp_root[gtid];
3368       if (root != NULL) {
3369         __kmp_printf("GTID %2d %p:\n", gtid, root);
3370         __kmp_print_structure_team("    Root Team:    ", root->r.r_root_team);
3371         __kmp_print_structure_team("    Hot Team:     ", root->r.r_hot_team);
3372         __kmp_print_structure_thread("    Uber Thread:  ",
3373                                      root->r.r_uber_thread);
3374         __kmp_printf("    Active?:      %2d\n", root->r.r_active);
3375         __kmp_printf("    Nested?:      %2d\n", root->r.r_nested);
3376         __kmp_printf("    In Parallel:  %2d\n", root->r.r_in_parallel);
3377         __kmp_printf("\n");
3378         __kmp_print_structure_team_accum(list, root->r.r_root_team);
3379         __kmp_print_structure_team_accum(list, root->r.r_hot_team);
3380       }; // if
3381     }; // for gtid
3382   } else {
3383     __kmp_printf("Ubers array is not allocated.\n");
3384   }; // if
3385 
3386   __kmp_printf("\n------------------------------\nTeams\n----------------------"
3387                "--------\n");
3388   while (list->next != NULL) {
3389     kmp_team_p const *team = list->entry;
3390     int i;
3391     __kmp_printf("Team %2x %p:\n", team->t.t_id, team);
3392     __kmp_print_structure_team("    Parent Team:      ", team->t.t_parent);
3393     __kmp_printf("    Master TID:       %2d\n", team->t.t_master_tid);
3394     __kmp_printf("    Max threads:      %2d\n", team->t.t_max_nproc);
3395     __kmp_printf("    Levels of serial: %2d\n", team->t.t_serialized);
3396     __kmp_printf("    Number threads:   %2d\n", team->t.t_nproc);
3397     for (i = 0; i < team->t.t_nproc; ++i) {
3398       __kmp_printf("    Thread %2d:      ", i);
3399       __kmp_print_structure_thread("", team->t.t_threads[i]);
3400     }; // for i
3401     __kmp_print_structure_team("    Next in pool:     ", team->t.t_next_pool);
3402     __kmp_printf("\n");
3403     list = list->next;
3404   }; // while
3405 
3406   // Print out __kmp_thread_pool and __kmp_team_pool.
3407   __kmp_printf("\n------------------------------\nPools\n----------------------"
3408                "--------\n");
3409   __kmp_print_structure_thread("Thread pool:          ",
3410                                CCAST(kmp_info_t *, __kmp_thread_pool));
3411   __kmp_print_structure_team("Team pool:            ",
3412                              CCAST(kmp_team_t *, __kmp_team_pool));
3413   __kmp_printf("\n");
3414 
3415   // Free team list.
3416   while (list != NULL) {
3417     kmp_team_list_item_t *item = list;
3418     list = list->next;
3419     KMP_INTERNAL_FREE(item);
3420   }; // while
3421 }
3422 
3423 #endif
3424 
3425 //---------------------------------------------------------------------------
3426 //  Stuff for per-thread fast random number generator
3427 //  Table of primes
3428 static const unsigned __kmp_primes[] = {
3429     0x9e3779b1, 0xffe6cc59, 0x2109f6dd, 0x43977ab5, 0xba5703f5, 0xb495a877,
3430     0xe1626741, 0x79695e6b, 0xbc98c09f, 0xd5bee2b3, 0x287488f9, 0x3af18231,
3431     0x9677cd4d, 0xbe3a6929, 0xadc6a877, 0xdcf0674b, 0xbe4d6fe9, 0x5f15e201,
3432     0x99afc3fd, 0xf3f16801, 0xe222cfff, 0x24ba5fdb, 0x0620452d, 0x79f149e3,
3433     0xc8b93f49, 0x972702cd, 0xb07dd827, 0x6c97d5ed, 0x085a3d61, 0x46eb5ea7,
3434     0x3d9910ed, 0x2e687b5b, 0x29609227, 0x6eb081f1, 0x0954c4e1, 0x9d114db9,
3435     0x542acfa9, 0xb3e6bd7b, 0x0742d917, 0xe9f3ffa7, 0x54581edb, 0xf2480f45,
3436     0x0bb9288f, 0xef1affc7, 0x85fa0ca7, 0x3ccc14db, 0xe6baf34b, 0x343377f7,
3437     0x5ca19031, 0xe6d9293b, 0xf0a9f391, 0x5d2e980b, 0xfc411073, 0xc3749363,
3438     0xb892d829, 0x3549366b, 0x629750ad, 0xb98294e5, 0x892d9483, 0xc235baf3,
3439     0x3d2402a3, 0x6bdef3c9, 0xbec333cd, 0x40c9520f};
3440 
3441 //---------------------------------------------------------------------------
3442 //  __kmp_get_random: Get a random number using a linear congruential method.
3443 unsigned short __kmp_get_random(kmp_info_t *thread) {
3444   unsigned x = thread->th.th_x;
3445   unsigned short r = x >> 16;
3446 
3447   thread->th.th_x = x * thread->th.th_a + 1;
3448 
3449   KA_TRACE(30, ("__kmp_get_random: THREAD: %d, RETURN: %u\n",
3450                 thread->th.th_info.ds.ds_tid, r));
3451 
3452   return r;
3453 }
3454 //--------------------------------------------------------
3455 // __kmp_init_random: Initialize a random number generator
3456 void __kmp_init_random(kmp_info_t *thread) {
3457   unsigned seed = thread->th.th_info.ds.ds_tid;
3458 
3459   thread->th.th_a =
3460       __kmp_primes[seed % (sizeof(__kmp_primes) / sizeof(__kmp_primes[0]))];
3461   thread->th.th_x = (seed + 1) * thread->th.th_a + 1;
3462   KA_TRACE(30,
3463            ("__kmp_init_random: THREAD: %u; A: %u\n", seed, thread->th.th_a));
3464 }
3465 
3466 #if KMP_OS_WINDOWS
3467 /* reclaim array entries for root threads that are already dead, returns number
3468  * reclaimed */
3469 static int __kmp_reclaim_dead_roots(void) {
3470   int i, r = 0;
3471 
3472   for (i = 0; i < __kmp_threads_capacity; ++i) {
3473     if (KMP_UBER_GTID(i) &&
3474         !__kmp_still_running((kmp_info_t *)TCR_SYNC_PTR(__kmp_threads[i])) &&
3475         !__kmp_root[i]
3476              ->r.r_active) { // AC: reclaim only roots died in non-active state
3477       r += __kmp_unregister_root_other_thread(i);
3478     }
3479   }
3480   return r;
3481 }
3482 #endif
3483 
3484 /* This function attempts to create free entries in __kmp_threads and
3485    __kmp_root, and returns the number of free entries generated.
3486 
3487    For Windows* OS static library, the first mechanism used is to reclaim array
3488    entries for root threads that are already dead.
3489 
3490    On all platforms, expansion is attempted on the arrays __kmp_threads_ and
3491    __kmp_root, with appropriate update to __kmp_threads_capacity. Array
3492    capacity is increased by doubling with clipping to __kmp_tp_capacity, if
3493    threadprivate cache array has been created. Synchronization with
3494    __kmpc_threadprivate_cached is done using __kmp_tp_cached_lock.
3495 
3496    After any dead root reclamation, if the clipping value allows array expansion
3497    to result in the generation of a total of nWish free slots, the function does
3498    that expansion. If not, but the clipping value allows array expansion to
3499    result in the generation of a total of nNeed free slots, the function does
3500    that expansion. Otherwise, nothing is done beyond the possible initial root
3501    thread reclamation. However, if nNeed is zero, a best-effort attempt is made
3502    to fulfil nWish as far as possible, i.e. the function will attempt to create
3503    as many free slots as possible up to nWish.
3504 
3505    If any argument is negative, the behavior is undefined. */
3506 static int __kmp_expand_threads(int nWish, int nNeed) {
3507   int added = 0;
3508   int old_tp_cached;
3509   int __kmp_actual_max_nth;
3510 
3511   if (nNeed > nWish) /* normalize the arguments */
3512     nWish = nNeed;
3513 #if KMP_OS_WINDOWS && !defined KMP_DYNAMIC_LIB
3514   /* only for Windows static library */
3515   /* reclaim array entries for root threads that are already dead */
3516   added = __kmp_reclaim_dead_roots();
3517 
3518   if (nNeed) {
3519     nNeed -= added;
3520     if (nNeed < 0)
3521       nNeed = 0;
3522   }
3523   if (nWish) {
3524     nWish -= added;
3525     if (nWish < 0)
3526       nWish = 0;
3527   }
3528 #endif
3529   if (nWish <= 0)
3530     return added;
3531 
3532   while (1) {
3533     int nTarget;
3534     int minimumRequiredCapacity;
3535     int newCapacity;
3536     kmp_info_t **newThreads;
3537     kmp_root_t **newRoot;
3538 
3539     // Note that __kmp_threads_capacity is not bounded by __kmp_max_nth. If
3540     // __kmp_max_nth is set to some value less than __kmp_sys_max_nth by the
3541     // user via KMP_DEVICE_THREAD_LIMIT, then __kmp_threads_capacity may become
3542     // > __kmp_max_nth in one of two ways:
3543     //
3544     // 1) The initialization thread (gtid = 0) exits.  __kmp_threads[0]
3545     //    may not be resused by another thread, so we may need to increase
3546     //    __kmp_threads_capacity to __kmp_max_nth + 1.
3547     //
3548     // 2) New foreign root(s) are encountered.  We always register new foreign
3549     //    roots. This may cause a smaller # of threads to be allocated at
3550     //    subsequent parallel regions, but the worker threads hang around (and
3551     //    eventually go to sleep) and need slots in the __kmp_threads[] array.
3552     //
3553     // Anyway, that is the reason for moving the check to see if
3554     // __kmp_max_nth was exceeded into __kmp_reserve_threads()
3555     // instead of having it performed here. -BB
3556     old_tp_cached = __kmp_tp_cached;
3557     __kmp_actual_max_nth =
3558         old_tp_cached ? __kmp_tp_capacity : __kmp_sys_max_nth;
3559     KMP_DEBUG_ASSERT(__kmp_actual_max_nth >= __kmp_threads_capacity);
3560 
3561     /* compute expansion headroom to check if we can expand and whether to aim
3562        for nWish or nNeed */
3563     nTarget = nWish;
3564     if (__kmp_actual_max_nth - __kmp_threads_capacity < nTarget) {
3565       /* can't fulfil nWish, so try nNeed */
3566       if (nNeed) {
3567         nTarget = nNeed;
3568         if (__kmp_actual_max_nth - __kmp_threads_capacity < nTarget) {
3569           /* possible expansion too small -- give up */
3570           break;
3571         }
3572       } else {
3573         /* best-effort */
3574         nTarget = __kmp_actual_max_nth - __kmp_threads_capacity;
3575         if (!nTarget) {
3576           /* can expand at all -- give up */
3577           break;
3578         }
3579       }
3580     }
3581     minimumRequiredCapacity = __kmp_threads_capacity + nTarget;
3582 
3583     newCapacity = __kmp_threads_capacity;
3584     do {
3585       newCapacity = newCapacity <= (__kmp_actual_max_nth >> 1)
3586                         ? (newCapacity << 1)
3587                         : __kmp_actual_max_nth;
3588     } while (newCapacity < minimumRequiredCapacity);
3589     newThreads = (kmp_info_t **)__kmp_allocate(
3590         (sizeof(kmp_info_t *) + sizeof(kmp_root_t *)) * newCapacity +
3591         CACHE_LINE);
3592     newRoot = (kmp_root_t **)((char *)newThreads +
3593                               sizeof(kmp_info_t *) * newCapacity);
3594     KMP_MEMCPY(newThreads, __kmp_threads,
3595                __kmp_threads_capacity * sizeof(kmp_info_t *));
3596     KMP_MEMCPY(newRoot, __kmp_root,
3597                __kmp_threads_capacity * sizeof(kmp_root_t *));
3598     memset(newThreads + __kmp_threads_capacity, 0,
3599            (newCapacity - __kmp_threads_capacity) * sizeof(kmp_info_t *));
3600     memset(newRoot + __kmp_threads_capacity, 0,
3601            (newCapacity - __kmp_threads_capacity) * sizeof(kmp_root_t *));
3602 
3603     if (!old_tp_cached && __kmp_tp_cached && newCapacity > __kmp_tp_capacity) {
3604       /* __kmp_tp_cached has changed, i.e. __kmpc_threadprivate_cached has
3605          allocated a threadprivate cache while we were allocating the expanded
3606          array, and our new capacity is larger than the threadprivate cache
3607          capacity, so we should deallocate the expanded arrays and try again.
3608          This is the first check of a double-check pair. */
3609       __kmp_free(newThreads);
3610       continue; /* start over and try again */
3611     }
3612     __kmp_acquire_bootstrap_lock(&__kmp_tp_cached_lock);
3613     if (!old_tp_cached && __kmp_tp_cached && newCapacity > __kmp_tp_capacity) {
3614       /* Same check as above, but this time with the lock so we can be sure if
3615          we can succeed. */
3616       __kmp_release_bootstrap_lock(&__kmp_tp_cached_lock);
3617       __kmp_free(newThreads);
3618       continue; /* start over and try again */
3619     } else {
3620       /* success */
3621       // __kmp_free( __kmp_threads ); // ATT: It leads to crash. Need to be
3622       // investigated.
3623       *(kmp_info_t * *volatile *)&__kmp_threads = newThreads;
3624       *(kmp_root_t * *volatile *)&__kmp_root = newRoot;
3625       added += newCapacity - __kmp_threads_capacity;
3626       *(volatile int *)&__kmp_threads_capacity = newCapacity;
3627       __kmp_release_bootstrap_lock(&__kmp_tp_cached_lock);
3628       break; /* succeeded, so we can exit the loop */
3629     }
3630   }
3631   return added;
3632 }
3633 
3634 /* Register the current thread as a root thread and obtain our gtid. We must
3635    have the __kmp_initz_lock held at this point. Argument TRUE only if are the
3636    thread that calls from __kmp_do_serial_initialize() */
3637 int __kmp_register_root(int initial_thread) {
3638   kmp_info_t *root_thread;
3639   kmp_root_t *root;
3640   int gtid;
3641   int capacity;
3642   __kmp_acquire_bootstrap_lock(&__kmp_forkjoin_lock);
3643   KA_TRACE(20, ("__kmp_register_root: entered\n"));
3644   KMP_MB();
3645 
3646   /* 2007-03-02:
3647      If initial thread did not invoke OpenMP RTL yet, and this thread is not an
3648      initial one, "__kmp_all_nth >= __kmp_threads_capacity" condition does not
3649      work as expected -- it may return false (that means there is at least one
3650      empty slot in __kmp_threads array), but it is possible the only free slot
3651      is #0, which is reserved for initial thread and so cannot be used for this
3652      one. Following code workarounds this bug.
3653 
3654      However, right solution seems to be not reserving slot #0 for initial
3655      thread because:
3656      (1) there is no magic in slot #0,
3657      (2) we cannot detect initial thread reliably (the first thread which does
3658         serial initialization may be not a real initial thread).
3659   */
3660   capacity = __kmp_threads_capacity;
3661   if (!initial_thread && TCR_PTR(__kmp_threads[0]) == NULL) {
3662     --capacity;
3663   }; // if
3664 
3665   /* see if there are too many threads */
3666   if (__kmp_all_nth >= capacity && !__kmp_expand_threads(1, 1)) {
3667     if (__kmp_tp_cached) {
3668       __kmp_fatal(KMP_MSG(CantRegisterNewThread),
3669                   KMP_HNT(Set_ALL_THREADPRIVATE, __kmp_tp_capacity),
3670                   KMP_HNT(PossibleSystemLimitOnThreads), __kmp_msg_null);
3671     } else {
3672       __kmp_fatal(KMP_MSG(CantRegisterNewThread), KMP_HNT(SystemLimitOnThreads),
3673                   __kmp_msg_null);
3674     }
3675   }; // if
3676 
3677   /* find an available thread slot */
3678   /* Don't reassign the zero slot since we need that to only be used by initial
3679      thread */
3680   for (gtid = (initial_thread ? 0 : 1); TCR_PTR(__kmp_threads[gtid]) != NULL;
3681        gtid++)
3682     ;
3683   KA_TRACE(1,
3684            ("__kmp_register_root: found slot in threads array: T#%d\n", gtid));
3685   KMP_ASSERT(gtid < __kmp_threads_capacity);
3686 
3687   /* update global accounting */
3688   __kmp_all_nth++;
3689   TCW_4(__kmp_nth, __kmp_nth + 1);
3690 
3691   // if __kmp_adjust_gtid_mode is set, then we use method #1 (sp search) for low
3692   // numbers of procs, and method #2 (keyed API call) for higher numbers.
3693   if (__kmp_adjust_gtid_mode) {
3694     if (__kmp_all_nth >= __kmp_tls_gtid_min) {
3695       if (TCR_4(__kmp_gtid_mode) != 2) {
3696         TCW_4(__kmp_gtid_mode, 2);
3697       }
3698     } else {
3699       if (TCR_4(__kmp_gtid_mode) != 1) {
3700         TCW_4(__kmp_gtid_mode, 1);
3701       }
3702     }
3703   }
3704 
3705 #ifdef KMP_ADJUST_BLOCKTIME
3706   /* Adjust blocktime to zero if necessary            */
3707   /* Middle initialization might not have occurred yet */
3708   if (!__kmp_env_blocktime && (__kmp_avail_proc > 0)) {
3709     if (__kmp_nth > __kmp_avail_proc) {
3710       __kmp_zero_bt = TRUE;
3711     }
3712   }
3713 #endif /* KMP_ADJUST_BLOCKTIME */
3714 
3715   /* setup this new hierarchy */
3716   if (!(root = __kmp_root[gtid])) {
3717     root = __kmp_root[gtid] = (kmp_root_t *)__kmp_allocate(sizeof(kmp_root_t));
3718     KMP_DEBUG_ASSERT(!root->r.r_root_team);
3719   }
3720 
3721 #if KMP_STATS_ENABLED
3722   // Initialize stats as soon as possible (right after gtid assignment).
3723   __kmp_stats_thread_ptr = __kmp_stats_list->push_back(gtid);
3724   KMP_START_EXPLICIT_TIMER(OMP_worker_thread_life);
3725   KMP_SET_THREAD_STATE(SERIAL_REGION);
3726   KMP_INIT_PARTITIONED_TIMERS(OMP_serial);
3727 #endif
3728   __kmp_initialize_root(root);
3729 
3730   /* setup new root thread structure */
3731   if (root->r.r_uber_thread) {
3732     root_thread = root->r.r_uber_thread;
3733   } else {
3734     root_thread = (kmp_info_t *)__kmp_allocate(sizeof(kmp_info_t));
3735     if (__kmp_storage_map) {
3736       __kmp_print_thread_storage_map(root_thread, gtid);
3737     }
3738     root_thread->th.th_info.ds.ds_gtid = gtid;
3739     root_thread->th.th_root = root;
3740     if (__kmp_env_consistency_check) {
3741       root_thread->th.th_cons = __kmp_allocate_cons_stack(gtid);
3742     }
3743 #if USE_FAST_MEMORY
3744     __kmp_initialize_fast_memory(root_thread);
3745 #endif /* USE_FAST_MEMORY */
3746 
3747 #if KMP_USE_BGET
3748     KMP_DEBUG_ASSERT(root_thread->th.th_local.bget_data == NULL);
3749     __kmp_initialize_bget(root_thread);
3750 #endif
3751     __kmp_init_random(root_thread); // Initialize random number generator
3752   }
3753 
3754   /* setup the serial team held in reserve by the root thread */
3755   if (!root_thread->th.th_serial_team) {
3756     kmp_internal_control_t r_icvs = __kmp_get_global_icvs();
3757     KF_TRACE(10, ("__kmp_register_root: before serial_team\n"));
3758     root_thread->th.th_serial_team =
3759         __kmp_allocate_team(root, 1, 1,
3760 #if OMPT_SUPPORT
3761                             0, // root parallel id
3762 #endif
3763 #if OMP_40_ENABLED
3764                             proc_bind_default,
3765 #endif
3766                             &r_icvs, 0 USE_NESTED_HOT_ARG(NULL));
3767   }
3768   KMP_ASSERT(root_thread->th.th_serial_team);
3769   KF_TRACE(10, ("__kmp_register_root: after serial_team = %p\n",
3770                 root_thread->th.th_serial_team));
3771 
3772   /* drop root_thread into place */
3773   TCW_SYNC_PTR(__kmp_threads[gtid], root_thread);
3774 
3775   root->r.r_root_team->t.t_threads[0] = root_thread;
3776   root->r.r_hot_team->t.t_threads[0] = root_thread;
3777   root_thread->th.th_serial_team->t.t_threads[0] = root_thread;
3778   // AC: the team created in reserve, not for execution (it is unused for now).
3779   root_thread->th.th_serial_team->t.t_serialized = 0;
3780   root->r.r_uber_thread = root_thread;
3781 
3782   /* initialize the thread, get it ready to go */
3783   __kmp_initialize_info(root_thread, root->r.r_root_team, 0, gtid);
3784   TCW_4(__kmp_init_gtid, TRUE);
3785 
3786   /* prepare the master thread for get_gtid() */
3787   __kmp_gtid_set_specific(gtid);
3788 
3789 #if USE_ITT_BUILD
3790   __kmp_itt_thread_name(gtid);
3791 #endif /* USE_ITT_BUILD */
3792 
3793 #ifdef KMP_TDATA_GTID
3794   __kmp_gtid = gtid;
3795 #endif
3796   __kmp_create_worker(gtid, root_thread, __kmp_stksize);
3797   KMP_DEBUG_ASSERT(__kmp_gtid_get_specific() == gtid);
3798 
3799   KA_TRACE(20, ("__kmp_register_root: T#%d init T#%d(%d:%d) arrived: join=%u, "
3800                 "plain=%u\n",
3801                 gtid, __kmp_gtid_from_tid(0, root->r.r_hot_team),
3802                 root->r.r_hot_team->t.t_id, 0, KMP_INIT_BARRIER_STATE,
3803                 KMP_INIT_BARRIER_STATE));
3804   { // Initialize barrier data.
3805     int b;
3806     for (b = 0; b < bs_last_barrier; ++b) {
3807       root_thread->th.th_bar[b].bb.b_arrived = KMP_INIT_BARRIER_STATE;
3808 #if USE_DEBUGGER
3809       root_thread->th.th_bar[b].bb.b_worker_arrived = 0;
3810 #endif
3811     }; // for
3812   }
3813   KMP_DEBUG_ASSERT(root->r.r_hot_team->t.t_bar[bs_forkjoin_barrier].b_arrived ==
3814                    KMP_INIT_BARRIER_STATE);
3815 
3816 #if KMP_AFFINITY_SUPPORTED
3817 #if OMP_40_ENABLED
3818   root_thread->th.th_current_place = KMP_PLACE_UNDEFINED;
3819   root_thread->th.th_new_place = KMP_PLACE_UNDEFINED;
3820   root_thread->th.th_first_place = KMP_PLACE_UNDEFINED;
3821   root_thread->th.th_last_place = KMP_PLACE_UNDEFINED;
3822 #endif
3823 
3824   if (TCR_4(__kmp_init_middle)) {
3825     __kmp_affinity_set_init_mask(gtid, TRUE);
3826   }
3827 #endif /* KMP_AFFINITY_SUPPORTED */
3828 
3829   __kmp_root_counter++;
3830 
3831   KMP_MB();
3832   __kmp_release_bootstrap_lock(&__kmp_forkjoin_lock);
3833 
3834   return gtid;
3835 }
3836 
3837 #if KMP_NESTED_HOT_TEAMS
3838 static int __kmp_free_hot_teams(kmp_root_t *root, kmp_info_t *thr, int level,
3839                                 const int max_level) {
3840   int i, n, nth;
3841   kmp_hot_team_ptr_t *hot_teams = thr->th.th_hot_teams;
3842   if (!hot_teams || !hot_teams[level].hot_team) {
3843     return 0;
3844   }
3845   KMP_DEBUG_ASSERT(level < max_level);
3846   kmp_team_t *team = hot_teams[level].hot_team;
3847   nth = hot_teams[level].hot_team_nth;
3848   n = nth - 1; // master is not freed
3849   if (level < max_level - 1) {
3850     for (i = 0; i < nth; ++i) {
3851       kmp_info_t *th = team->t.t_threads[i];
3852       n += __kmp_free_hot_teams(root, th, level + 1, max_level);
3853       if (i > 0 && th->th.th_hot_teams) {
3854         __kmp_free(th->th.th_hot_teams);
3855         th->th.th_hot_teams = NULL;
3856       }
3857     }
3858   }
3859   __kmp_free_team(root, team, NULL);
3860   return n;
3861 }
3862 #endif
3863 
3864 // Resets a root thread and clear its root and hot teams.
3865 // Returns the number of __kmp_threads entries directly and indirectly freed.
3866 static int __kmp_reset_root(int gtid, kmp_root_t *root) {
3867   kmp_team_t *root_team = root->r.r_root_team;
3868   kmp_team_t *hot_team = root->r.r_hot_team;
3869   int n = hot_team->t.t_nproc;
3870   int i;
3871 
3872   KMP_DEBUG_ASSERT(!root->r.r_active);
3873 
3874   root->r.r_root_team = NULL;
3875   root->r.r_hot_team = NULL;
3876   // __kmp_free_team() does not free hot teams, so we have to clear r_hot_team
3877   // before call to __kmp_free_team().
3878   __kmp_free_team(root, root_team USE_NESTED_HOT_ARG(NULL));
3879 #if KMP_NESTED_HOT_TEAMS
3880   if (__kmp_hot_teams_max_level >
3881       0) { // need to free nested hot teams and their threads if any
3882     for (i = 0; i < hot_team->t.t_nproc; ++i) {
3883       kmp_info_t *th = hot_team->t.t_threads[i];
3884       if (__kmp_hot_teams_max_level > 1) {
3885         n += __kmp_free_hot_teams(root, th, 1, __kmp_hot_teams_max_level);
3886       }
3887       if (th->th.th_hot_teams) {
3888         __kmp_free(th->th.th_hot_teams);
3889         th->th.th_hot_teams = NULL;
3890       }
3891     }
3892   }
3893 #endif
3894   __kmp_free_team(root, hot_team USE_NESTED_HOT_ARG(NULL));
3895 
3896   // Before we can reap the thread, we need to make certain that all other
3897   // threads in the teams that had this root as ancestor have stopped trying to
3898   // steal tasks.
3899   if (__kmp_tasking_mode != tskm_immediate_exec) {
3900     __kmp_wait_to_unref_task_teams();
3901   }
3902 
3903 #if KMP_OS_WINDOWS
3904   /* Close Handle of root duplicated in __kmp_create_worker (tr #62919) */
3905   KA_TRACE(
3906       10, ("__kmp_reset_root: free handle, th = %p, handle = %" KMP_UINTPTR_SPEC
3907            "\n",
3908            (LPVOID) & (root->r.r_uber_thread->th),
3909            root->r.r_uber_thread->th.th_info.ds.ds_thread));
3910   __kmp_free_handle(root->r.r_uber_thread->th.th_info.ds.ds_thread);
3911 #endif /* KMP_OS_WINDOWS */
3912 
3913 #if OMPT_SUPPORT
3914   if (ompt_enabled && ompt_callbacks.ompt_callback(ompt_event_thread_end)) {
3915     int gtid = __kmp_get_gtid();
3916     __ompt_thread_end(ompt_thread_initial, gtid);
3917   }
3918 #endif
3919 
3920   TCW_4(__kmp_nth,
3921         __kmp_nth - 1); // __kmp_reap_thread will decrement __kmp_all_nth.
3922   root->r.r_cg_nthreads--;
3923 
3924   __kmp_reap_thread(root->r.r_uber_thread, 1);
3925 
3926   // We canot put root thread to __kmp_thread_pool, so we have to reap it istead
3927   // of freeing.
3928   root->r.r_uber_thread = NULL;
3929   /* mark root as no longer in use */
3930   root->r.r_begin = FALSE;
3931 
3932   return n;
3933 }
3934 
3935 void __kmp_unregister_root_current_thread(int gtid) {
3936   KA_TRACE(1, ("__kmp_unregister_root_current_thread: enter T#%d\n", gtid));
3937   /* this lock should be ok, since unregister_root_current_thread is never
3938      called during an abort, only during a normal close. furthermore, if you
3939      have the forkjoin lock, you should never try to get the initz lock */
3940   __kmp_acquire_bootstrap_lock(&__kmp_forkjoin_lock);
3941   if (TCR_4(__kmp_global.g.g_done) || !__kmp_init_serial) {
3942     KC_TRACE(10, ("__kmp_unregister_root_current_thread: already finished, "
3943                   "exiting T#%d\n",
3944                   gtid));
3945     __kmp_release_bootstrap_lock(&__kmp_forkjoin_lock);
3946     return;
3947   }
3948   kmp_root_t *root = __kmp_root[gtid];
3949 
3950   KMP_DEBUG_ASSERT(__kmp_threads && __kmp_threads[gtid]);
3951   KMP_ASSERT(KMP_UBER_GTID(gtid));
3952   KMP_ASSERT(root == __kmp_threads[gtid]->th.th_root);
3953   KMP_ASSERT(root->r.r_active == FALSE);
3954 
3955   KMP_MB();
3956 
3957 #if OMP_45_ENABLED
3958   kmp_info_t *thread = __kmp_threads[gtid];
3959   kmp_team_t *team = thread->th.th_team;
3960   kmp_task_team_t *task_team = thread->th.th_task_team;
3961 
3962   // we need to wait for the proxy tasks before finishing the thread
3963   if (task_team != NULL && task_team->tt.tt_found_proxy_tasks) {
3964 #if OMPT_SUPPORT
3965     // the runtime is shutting down so we won't report any events
3966     thread->th.ompt_thread_info.state = ompt_state_undefined;
3967 #endif
3968     __kmp_task_team_wait(thread, team USE_ITT_BUILD_ARG(NULL));
3969   }
3970 #endif
3971 
3972   __kmp_reset_root(gtid, root);
3973 
3974   /* free up this thread slot */
3975   __kmp_gtid_set_specific(KMP_GTID_DNE);
3976 #ifdef KMP_TDATA_GTID
3977   __kmp_gtid = KMP_GTID_DNE;
3978 #endif
3979 
3980   KMP_MB();
3981   KC_TRACE(10,
3982            ("__kmp_unregister_root_current_thread: T#%d unregistered\n", gtid));
3983 
3984   __kmp_release_bootstrap_lock(&__kmp_forkjoin_lock);
3985 }
3986 
3987 #if KMP_OS_WINDOWS
3988 /* __kmp_forkjoin_lock must be already held
3989    Unregisters a root thread that is not the current thread.  Returns the number
3990    of __kmp_threads entries freed as a result. */
3991 static int __kmp_unregister_root_other_thread(int gtid) {
3992   kmp_root_t *root = __kmp_root[gtid];
3993   int r;
3994 
3995   KA_TRACE(1, ("__kmp_unregister_root_other_thread: enter T#%d\n", gtid));
3996   KMP_DEBUG_ASSERT(__kmp_threads && __kmp_threads[gtid]);
3997   KMP_ASSERT(KMP_UBER_GTID(gtid));
3998   KMP_ASSERT(root == __kmp_threads[gtid]->th.th_root);
3999   KMP_ASSERT(root->r.r_active == FALSE);
4000 
4001   r = __kmp_reset_root(gtid, root);
4002   KC_TRACE(10,
4003            ("__kmp_unregister_root_other_thread: T#%d unregistered\n", gtid));
4004   return r;
4005 }
4006 #endif
4007 
4008 #if KMP_DEBUG
4009 void __kmp_task_info() {
4010 
4011   kmp_int32 gtid = __kmp_entry_gtid();
4012   kmp_int32 tid = __kmp_tid_from_gtid(gtid);
4013   kmp_info_t *this_thr = __kmp_threads[gtid];
4014   kmp_team_t *steam = this_thr->th.th_serial_team;
4015   kmp_team_t *team = this_thr->th.th_team;
4016 
4017   __kmp_printf("__kmp_task_info: gtid=%d tid=%d t_thread=%p team=%p curtask=%p "
4018                "ptask=%p\n",
4019                gtid, tid, this_thr, team, this_thr->th.th_current_task,
4020                team->t.t_implicit_task_taskdata[tid].td_parent);
4021 }
4022 #endif // KMP_DEBUG
4023 
4024 /* TODO optimize with one big memclr, take out what isn't needed, split
4025    responsibility to workers as much as possible, and delay initialization of
4026    features as much as possible  */
4027 static void __kmp_initialize_info(kmp_info_t *this_thr, kmp_team_t *team,
4028                                   int tid, int gtid) {
4029   /* this_thr->th.th_info.ds.ds_gtid is setup in
4030      kmp_allocate_thread/create_worker.
4031      this_thr->th.th_serial_team is setup in __kmp_allocate_thread */
4032   kmp_info_t *master = team->t.t_threads[0];
4033   KMP_DEBUG_ASSERT(this_thr != NULL);
4034   KMP_DEBUG_ASSERT(this_thr->th.th_serial_team);
4035   KMP_DEBUG_ASSERT(team);
4036   KMP_DEBUG_ASSERT(team->t.t_threads);
4037   KMP_DEBUG_ASSERT(team->t.t_dispatch);
4038   KMP_DEBUG_ASSERT(master);
4039   KMP_DEBUG_ASSERT(master->th.th_root);
4040 
4041   KMP_MB();
4042 
4043   TCW_SYNC_PTR(this_thr->th.th_team, team);
4044 
4045   this_thr->th.th_info.ds.ds_tid = tid;
4046   this_thr->th.th_set_nproc = 0;
4047   if (__kmp_tasking_mode != tskm_immediate_exec)
4048     // When tasking is possible, threads are not safe to reap until they are
4049     // done tasking; this will be set when tasking code is exited in wait
4050     this_thr->th.th_reap_state = KMP_NOT_SAFE_TO_REAP;
4051   else // no tasking --> always safe to reap
4052     this_thr->th.th_reap_state = KMP_SAFE_TO_REAP;
4053 #if OMP_40_ENABLED
4054   this_thr->th.th_set_proc_bind = proc_bind_default;
4055 #if KMP_AFFINITY_SUPPORTED
4056   this_thr->th.th_new_place = this_thr->th.th_current_place;
4057 #endif
4058 #endif
4059   this_thr->th.th_root = master->th.th_root;
4060 
4061   /* setup the thread's cache of the team structure */
4062   this_thr->th.th_team_nproc = team->t.t_nproc;
4063   this_thr->th.th_team_master = master;
4064   this_thr->th.th_team_serialized = team->t.t_serialized;
4065   TCW_PTR(this_thr->th.th_sleep_loc, NULL);
4066 
4067   KMP_DEBUG_ASSERT(team->t.t_implicit_task_taskdata);
4068 
4069   KF_TRACE(10, ("__kmp_initialize_info1: T#%d:%d this_thread=%p curtask=%p\n",
4070                 tid, gtid, this_thr, this_thr->th.th_current_task));
4071 
4072   __kmp_init_implicit_task(this_thr->th.th_team_master->th.th_ident, this_thr,
4073                            team, tid, TRUE);
4074 
4075   KF_TRACE(10, ("__kmp_initialize_info2: T#%d:%d this_thread=%p curtask=%p\n",
4076                 tid, gtid, this_thr, this_thr->th.th_current_task));
4077   // TODO: Initialize ICVs from parent; GEH - isn't that already done in
4078   // __kmp_initialize_team()?
4079 
4080   /* TODO no worksharing in speculative threads */
4081   this_thr->th.th_dispatch = &team->t.t_dispatch[tid];
4082 
4083   this_thr->th.th_local.this_construct = 0;
4084 
4085   if (!this_thr->th.th_pri_common) {
4086     this_thr->th.th_pri_common =
4087         (struct common_table *)__kmp_allocate(sizeof(struct common_table));
4088     if (__kmp_storage_map) {
4089       __kmp_print_storage_map_gtid(
4090           gtid, this_thr->th.th_pri_common, this_thr->th.th_pri_common + 1,
4091           sizeof(struct common_table), "th_%d.th_pri_common\n", gtid);
4092     }; // if
4093     this_thr->th.th_pri_head = NULL;
4094   }; // if
4095 
4096   /* Initialize dynamic dispatch */
4097   {
4098     volatile kmp_disp_t *dispatch = this_thr->th.th_dispatch;
4099     // Use team max_nproc since this will never change for the team.
4100     size_t disp_size =
4101         sizeof(dispatch_private_info_t) *
4102         (team->t.t_max_nproc == 1 ? 1 : __kmp_dispatch_num_buffers);
4103     KD_TRACE(10, ("__kmp_initialize_info: T#%d max_nproc: %d\n", gtid,
4104                   team->t.t_max_nproc));
4105     KMP_ASSERT(dispatch);
4106     KMP_DEBUG_ASSERT(team->t.t_dispatch);
4107     KMP_DEBUG_ASSERT(dispatch == &team->t.t_dispatch[tid]);
4108 
4109     dispatch->th_disp_index = 0;
4110 #if OMP_45_ENABLED
4111     dispatch->th_doacross_buf_idx = 0;
4112 #endif
4113     if (!dispatch->th_disp_buffer) {
4114       dispatch->th_disp_buffer =
4115           (dispatch_private_info_t *)__kmp_allocate(disp_size);
4116 
4117       if (__kmp_storage_map) {
4118         __kmp_print_storage_map_gtid(
4119             gtid, &dispatch->th_disp_buffer[0],
4120             &dispatch->th_disp_buffer[team->t.t_max_nproc == 1
4121                                           ? 1
4122                                           : __kmp_dispatch_num_buffers],
4123             disp_size, "th_%d.th_dispatch.th_disp_buffer "
4124                        "(team_%d.t_dispatch[%d].th_disp_buffer)",
4125             gtid, team->t.t_id, gtid);
4126       }
4127     } else {
4128       memset(&dispatch->th_disp_buffer[0], '\0', disp_size);
4129     }
4130 
4131     dispatch->th_dispatch_pr_current = 0;
4132     dispatch->th_dispatch_sh_current = 0;
4133 
4134     dispatch->th_deo_fcn = 0; /* ORDERED     */
4135     dispatch->th_dxo_fcn = 0; /* END ORDERED */
4136   }
4137 
4138   this_thr->th.th_next_pool = NULL;
4139 
4140   if (!this_thr->th.th_task_state_memo_stack) {
4141     size_t i;
4142     this_thr->th.th_task_state_memo_stack =
4143         (kmp_uint8 *)__kmp_allocate(4 * sizeof(kmp_uint8));
4144     this_thr->th.th_task_state_top = 0;
4145     this_thr->th.th_task_state_stack_sz = 4;
4146     for (i = 0; i < this_thr->th.th_task_state_stack_sz;
4147          ++i) // zero init the stack
4148       this_thr->th.th_task_state_memo_stack[i] = 0;
4149   }
4150 
4151   KMP_DEBUG_ASSERT(!this_thr->th.th_spin_here);
4152   KMP_DEBUG_ASSERT(this_thr->th.th_next_waiting == 0);
4153 
4154   KMP_MB();
4155 }
4156 
4157 /* allocate a new thread for the requesting team. this is only called from
4158    within a forkjoin critical section. we will first try to get an available
4159    thread from the thread pool. if none is available, we will fork a new one
4160    assuming we are able to create a new one. this should be assured, as the
4161    caller should check on this first. */
4162 kmp_info_t *__kmp_allocate_thread(kmp_root_t *root, kmp_team_t *team,
4163                                   int new_tid) {
4164   kmp_team_t *serial_team;
4165   kmp_info_t *new_thr;
4166   int new_gtid;
4167 
4168   KA_TRACE(20, ("__kmp_allocate_thread: T#%d\n", __kmp_get_gtid()));
4169   KMP_DEBUG_ASSERT(root && team);
4170 #if !KMP_NESTED_HOT_TEAMS
4171   KMP_DEBUG_ASSERT(KMP_MASTER_GTID(__kmp_get_gtid()));
4172 #endif
4173   KMP_MB();
4174 
4175   /* first, try to get one from the thread pool */
4176   if (__kmp_thread_pool) {
4177 
4178     new_thr = CCAST(kmp_info_t *, __kmp_thread_pool);
4179     __kmp_thread_pool = (volatile kmp_info_t *)new_thr->th.th_next_pool;
4180     if (new_thr == __kmp_thread_pool_insert_pt) {
4181       __kmp_thread_pool_insert_pt = NULL;
4182     }
4183     TCW_4(new_thr->th.th_in_pool, FALSE);
4184     // Don't touch th_active_in_pool or th_active.
4185     // The worker thread adjusts those flags as it sleeps/awakens.
4186     __kmp_thread_pool_nth--;
4187 
4188     KA_TRACE(20, ("__kmp_allocate_thread: T#%d using thread T#%d\n",
4189                   __kmp_get_gtid(), new_thr->th.th_info.ds.ds_gtid));
4190     KMP_ASSERT(!new_thr->th.th_team);
4191     KMP_DEBUG_ASSERT(__kmp_nth < __kmp_threads_capacity);
4192     KMP_DEBUG_ASSERT(__kmp_thread_pool_nth >= 0);
4193 
4194     /* setup the thread structure */
4195     __kmp_initialize_info(new_thr, team, new_tid,
4196                           new_thr->th.th_info.ds.ds_gtid);
4197     KMP_DEBUG_ASSERT(new_thr->th.th_serial_team);
4198 
4199     TCW_4(__kmp_nth, __kmp_nth + 1);
4200     root->r.r_cg_nthreads++;
4201 
4202     new_thr->th.th_task_state = 0;
4203     new_thr->th.th_task_state_top = 0;
4204     new_thr->th.th_task_state_stack_sz = 4;
4205 
4206 #ifdef KMP_ADJUST_BLOCKTIME
4207     /* Adjust blocktime back to zero if necessary */
4208     /* Middle initialization might not have occurred yet */
4209     if (!__kmp_env_blocktime && (__kmp_avail_proc > 0)) {
4210       if (__kmp_nth > __kmp_avail_proc) {
4211         __kmp_zero_bt = TRUE;
4212       }
4213     }
4214 #endif /* KMP_ADJUST_BLOCKTIME */
4215 
4216 #if KMP_DEBUG
4217     // If thread entered pool via __kmp_free_thread, wait_flag should !=
4218     // KMP_BARRIER_PARENT_FLAG.
4219     int b;
4220     kmp_balign_t *balign = new_thr->th.th_bar;
4221     for (b = 0; b < bs_last_barrier; ++b)
4222       KMP_DEBUG_ASSERT(balign[b].bb.wait_flag != KMP_BARRIER_PARENT_FLAG);
4223 #endif
4224 
4225     KF_TRACE(10, ("__kmp_allocate_thread: T#%d using thread %p T#%d\n",
4226                   __kmp_get_gtid(), new_thr, new_thr->th.th_info.ds.ds_gtid));
4227 
4228     KMP_MB();
4229     return new_thr;
4230   }
4231 
4232   /* no, well fork a new one */
4233   KMP_ASSERT(__kmp_nth == __kmp_all_nth);
4234   KMP_ASSERT(__kmp_all_nth < __kmp_threads_capacity);
4235 
4236 #if KMP_USE_MONITOR
4237   // If this is the first worker thread the RTL is creating, then also
4238   // launch the monitor thread.  We try to do this as early as possible.
4239   if (!TCR_4(__kmp_init_monitor)) {
4240     __kmp_acquire_bootstrap_lock(&__kmp_monitor_lock);
4241     if (!TCR_4(__kmp_init_monitor)) {
4242       KF_TRACE(10, ("before __kmp_create_monitor\n"));
4243       TCW_4(__kmp_init_monitor, 1);
4244       __kmp_create_monitor(&__kmp_monitor);
4245       KF_TRACE(10, ("after __kmp_create_monitor\n"));
4246 #if KMP_OS_WINDOWS
4247       // AC: wait until monitor has started. This is a fix for CQ232808.
4248       // The reason is that if the library is loaded/unloaded in a loop with
4249       // small (parallel) work in between, then there is high probability that
4250       // monitor thread started after the library shutdown. At shutdown it is
4251       // too late to cope with the problem, because when the master is in
4252       // DllMain (process detach) the monitor has no chances to start (it is
4253       // blocked), and master has no means to inform the monitor that the
4254       // library has gone, because all the memory which the monitor can access
4255       // is going to be released/reset.
4256       while (TCR_4(__kmp_init_monitor) < 2) {
4257         KMP_YIELD(TRUE);
4258       }
4259       KF_TRACE(10, ("after monitor thread has started\n"));
4260 #endif
4261     }
4262     __kmp_release_bootstrap_lock(&__kmp_monitor_lock);
4263   }
4264 #endif
4265 
4266   KMP_MB();
4267   for (new_gtid = 1; TCR_PTR(__kmp_threads[new_gtid]) != NULL; ++new_gtid) {
4268     KMP_DEBUG_ASSERT(new_gtid < __kmp_threads_capacity);
4269   }
4270 
4271   /* allocate space for it. */
4272   new_thr = (kmp_info_t *)__kmp_allocate(sizeof(kmp_info_t));
4273 
4274   TCW_SYNC_PTR(__kmp_threads[new_gtid], new_thr);
4275 
4276   if (__kmp_storage_map) {
4277     __kmp_print_thread_storage_map(new_thr, new_gtid);
4278   }
4279 
4280   // add the reserve serialized team, initialized from the team's master thread
4281   {
4282     kmp_internal_control_t r_icvs = __kmp_get_x_global_icvs(team);
4283     KF_TRACE(10, ("__kmp_allocate_thread: before th_serial/serial_team\n"));
4284     new_thr->th.th_serial_team = serial_team =
4285         (kmp_team_t *)__kmp_allocate_team(root, 1, 1,
4286 #if OMPT_SUPPORT
4287                                           0, // root parallel id
4288 #endif
4289 #if OMP_40_ENABLED
4290                                           proc_bind_default,
4291 #endif
4292                                           &r_icvs, 0 USE_NESTED_HOT_ARG(NULL));
4293   }
4294   KMP_ASSERT(serial_team);
4295   serial_team->t.t_serialized = 0; // AC: the team created in reserve, not for
4296   // execution (it is unused for now).
4297   serial_team->t.t_threads[0] = new_thr;
4298   KF_TRACE(10,
4299            ("__kmp_allocate_thread: after th_serial/serial_team : new_thr=%p\n",
4300             new_thr));
4301 
4302   /* setup the thread structures */
4303   __kmp_initialize_info(new_thr, team, new_tid, new_gtid);
4304 
4305 #if USE_FAST_MEMORY
4306   __kmp_initialize_fast_memory(new_thr);
4307 #endif /* USE_FAST_MEMORY */
4308 
4309 #if KMP_USE_BGET
4310   KMP_DEBUG_ASSERT(new_thr->th.th_local.bget_data == NULL);
4311   __kmp_initialize_bget(new_thr);
4312 #endif
4313 
4314   __kmp_init_random(new_thr); // Initialize random number generator
4315 
4316   /* Initialize these only once when thread is grabbed for a team allocation */
4317   KA_TRACE(20,
4318            ("__kmp_allocate_thread: T#%d init go fork=%u, plain=%u\n",
4319             __kmp_get_gtid(), KMP_INIT_BARRIER_STATE, KMP_INIT_BARRIER_STATE));
4320 
4321   int b;
4322   kmp_balign_t *balign = new_thr->th.th_bar;
4323   for (b = 0; b < bs_last_barrier; ++b) {
4324     balign[b].bb.b_go = KMP_INIT_BARRIER_STATE;
4325     balign[b].bb.team = NULL;
4326     balign[b].bb.wait_flag = KMP_BARRIER_NOT_WAITING;
4327     balign[b].bb.use_oncore_barrier = 0;
4328   }
4329 
4330   new_thr->th.th_spin_here = FALSE;
4331   new_thr->th.th_next_waiting = 0;
4332 
4333 #if OMP_40_ENABLED && KMP_AFFINITY_SUPPORTED
4334   new_thr->th.th_current_place = KMP_PLACE_UNDEFINED;
4335   new_thr->th.th_new_place = KMP_PLACE_UNDEFINED;
4336   new_thr->th.th_first_place = KMP_PLACE_UNDEFINED;
4337   new_thr->th.th_last_place = KMP_PLACE_UNDEFINED;
4338 #endif
4339 
4340   TCW_4(new_thr->th.th_in_pool, FALSE);
4341   new_thr->th.th_active_in_pool = FALSE;
4342   TCW_4(new_thr->th.th_active, TRUE);
4343 
4344   /* adjust the global counters */
4345   __kmp_all_nth++;
4346   __kmp_nth++;
4347 
4348   root->r.r_cg_nthreads++;
4349 
4350   // if __kmp_adjust_gtid_mode is set, then we use method #1 (sp search) for low
4351   // numbers of procs, and method #2 (keyed API call) for higher numbers.
4352   if (__kmp_adjust_gtid_mode) {
4353     if (__kmp_all_nth >= __kmp_tls_gtid_min) {
4354       if (TCR_4(__kmp_gtid_mode) != 2) {
4355         TCW_4(__kmp_gtid_mode, 2);
4356       }
4357     } else {
4358       if (TCR_4(__kmp_gtid_mode) != 1) {
4359         TCW_4(__kmp_gtid_mode, 1);
4360       }
4361     }
4362   }
4363 
4364 #ifdef KMP_ADJUST_BLOCKTIME
4365   /* Adjust blocktime back to zero if necessary       */
4366   /* Middle initialization might not have occurred yet */
4367   if (!__kmp_env_blocktime && (__kmp_avail_proc > 0)) {
4368     if (__kmp_nth > __kmp_avail_proc) {
4369       __kmp_zero_bt = TRUE;
4370     }
4371   }
4372 #endif /* KMP_ADJUST_BLOCKTIME */
4373 
4374   /* actually fork it and create the new worker thread */
4375   KF_TRACE(
4376       10, ("__kmp_allocate_thread: before __kmp_create_worker: %p\n", new_thr));
4377   __kmp_create_worker(new_gtid, new_thr, __kmp_stksize);
4378   KF_TRACE(10,
4379            ("__kmp_allocate_thread: after __kmp_create_worker: %p\n", new_thr));
4380 
4381   KA_TRACE(20, ("__kmp_allocate_thread: T#%d forked T#%d\n", __kmp_get_gtid(),
4382                 new_gtid));
4383   KMP_MB();
4384   return new_thr;
4385 }
4386 
4387 /* Reinitialize team for reuse.
4388    The hot team code calls this case at every fork barrier, so EPCC barrier
4389    test are extremely sensitive to changes in it, esp. writes to the team
4390    struct, which cause a cache invalidation in all threads.
4391    IF YOU TOUCH THIS ROUTINE, RUN EPCC C SYNCBENCH ON A BIG-IRON MACHINE!!! */
4392 static void __kmp_reinitialize_team(kmp_team_t *team,
4393                                     kmp_internal_control_t *new_icvs,
4394                                     ident_t *loc) {
4395   KF_TRACE(10, ("__kmp_reinitialize_team: enter this_thread=%p team=%p\n",
4396                 team->t.t_threads[0], team));
4397   KMP_DEBUG_ASSERT(team && new_icvs);
4398   KMP_DEBUG_ASSERT((!TCR_4(__kmp_init_parallel)) || new_icvs->nproc);
4399   KMP_CHECK_UPDATE(team->t.t_ident, loc);
4400 
4401   KMP_CHECK_UPDATE(team->t.t_id, KMP_GEN_TEAM_ID());
4402   // Copy ICVs to the master thread's implicit taskdata
4403   __kmp_init_implicit_task(loc, team->t.t_threads[0], team, 0, FALSE);
4404   copy_icvs(&team->t.t_implicit_task_taskdata[0].td_icvs, new_icvs);
4405 
4406   KF_TRACE(10, ("__kmp_reinitialize_team: exit this_thread=%p team=%p\n",
4407                 team->t.t_threads[0], team));
4408 }
4409 
4410 /* Initialize the team data structure.
4411    This assumes the t_threads and t_max_nproc are already set.
4412    Also, we don't touch the arguments */
4413 static void __kmp_initialize_team(kmp_team_t *team, int new_nproc,
4414                                   kmp_internal_control_t *new_icvs,
4415                                   ident_t *loc) {
4416   KF_TRACE(10, ("__kmp_initialize_team: enter: team=%p\n", team));
4417 
4418   /* verify */
4419   KMP_DEBUG_ASSERT(team);
4420   KMP_DEBUG_ASSERT(new_nproc <= team->t.t_max_nproc);
4421   KMP_DEBUG_ASSERT(team->t.t_threads);
4422   KMP_MB();
4423 
4424   team->t.t_master_tid = 0; /* not needed */
4425   /* team->t.t_master_bar;        not needed */
4426   team->t.t_serialized = new_nproc > 1 ? 0 : 1;
4427   team->t.t_nproc = new_nproc;
4428 
4429   /* team->t.t_parent     = NULL; TODO not needed & would mess up hot team */
4430   team->t.t_next_pool = NULL;
4431   /* memset( team->t.t_threads, 0, sizeof(kmp_info_t*)*new_nproc ); would mess
4432    * up hot team */
4433 
4434   TCW_SYNC_PTR(team->t.t_pkfn, NULL); /* not needed */
4435   team->t.t_invoke = NULL; /* not needed */
4436 
4437   // TODO???: team->t.t_max_active_levels       = new_max_active_levels;
4438   team->t.t_sched = new_icvs->sched;
4439 
4440 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
4441   team->t.t_fp_control_saved = FALSE; /* not needed */
4442   team->t.t_x87_fpu_control_word = 0; /* not needed */
4443   team->t.t_mxcsr = 0; /* not needed */
4444 #endif /* KMP_ARCH_X86 || KMP_ARCH_X86_64 */
4445 
4446   team->t.t_construct = 0;
4447   __kmp_init_lock(&team->t.t_single_lock);
4448 
4449   team->t.t_ordered.dt.t_value = 0;
4450   team->t.t_master_active = FALSE;
4451 
4452   memset(&team->t.t_taskq, '\0', sizeof(kmp_taskq_t));
4453 
4454 #ifdef KMP_DEBUG
4455   team->t.t_copypriv_data = NULL; /* not necessary, but nice for debugging */
4456 #endif
4457   team->t.t_copyin_counter = 0; /* for barrier-free copyin implementation */
4458 
4459   team->t.t_control_stack_top = NULL;
4460 
4461   __kmp_reinitialize_team(team, new_icvs, loc);
4462 
4463   KMP_MB();
4464   KF_TRACE(10, ("__kmp_initialize_team: exit: team=%p\n", team));
4465 }
4466 
4467 #if KMP_OS_LINUX && KMP_AFFINITY_SUPPORTED
4468 /* Sets full mask for thread and returns old mask, no changes to structures. */
4469 static void
4470 __kmp_set_thread_affinity_mask_full_tmp(kmp_affin_mask_t *old_mask) {
4471   if (KMP_AFFINITY_CAPABLE()) {
4472     int status;
4473     if (old_mask != NULL) {
4474       status = __kmp_get_system_affinity(old_mask, TRUE);
4475       int error = errno;
4476       if (status != 0) {
4477         __kmp_fatal(KMP_MSG(ChangeThreadAffMaskError), KMP_ERR(error),
4478                     __kmp_msg_null);
4479       }
4480     }
4481     __kmp_set_system_affinity(__kmp_affin_fullMask, TRUE);
4482   }
4483 }
4484 #endif
4485 
4486 #if OMP_40_ENABLED && KMP_AFFINITY_SUPPORTED
4487 
4488 // __kmp_partition_places() is the heart of the OpenMP 4.0 affinity mechanism.
4489 // It calculats the worker + master thread's partition based upon the parent
4490 // thread's partition, and binds each worker to a thread in their partition.
4491 // The master thread's partition should already include its current binding.
4492 static void __kmp_partition_places(kmp_team_t *team, int update_master_only) {
4493   // Copy the master thread's place partion to the team struct
4494   kmp_info_t *master_th = team->t.t_threads[0];
4495   KMP_DEBUG_ASSERT(master_th != NULL);
4496   kmp_proc_bind_t proc_bind = team->t.t_proc_bind;
4497   int first_place = master_th->th.th_first_place;
4498   int last_place = master_th->th.th_last_place;
4499   int masters_place = master_th->th.th_current_place;
4500   team->t.t_first_place = first_place;
4501   team->t.t_last_place = last_place;
4502 
4503   KA_TRACE(20, ("__kmp_partition_places: enter: proc_bind = %d T#%d(%d:0) "
4504                 "bound to place %d partition = [%d,%d]\n",
4505                 proc_bind, __kmp_gtid_from_thread(team->t.t_threads[0]),
4506                 team->t.t_id, masters_place, first_place, last_place));
4507 
4508   switch (proc_bind) {
4509 
4510   case proc_bind_default:
4511     // serial teams might have the proc_bind policy set to proc_bind_default. It
4512     // doesn't matter, as we don't rebind master thread for any proc_bind policy
4513     KMP_DEBUG_ASSERT(team->t.t_nproc == 1);
4514     break;
4515 
4516   case proc_bind_master: {
4517     int f;
4518     int n_th = team->t.t_nproc;
4519     for (f = 1; f < n_th; f++) {
4520       kmp_info_t *th = team->t.t_threads[f];
4521       KMP_DEBUG_ASSERT(th != NULL);
4522       th->th.th_first_place = first_place;
4523       th->th.th_last_place = last_place;
4524       th->th.th_new_place = masters_place;
4525 
4526       KA_TRACE(100, ("__kmp_partition_places: master: T#%d(%d:%d) place %d "
4527                      "partition = [%d,%d]\n",
4528                      __kmp_gtid_from_thread(team->t.t_threads[f]), team->t.t_id,
4529                      f, masters_place, first_place, last_place));
4530     }
4531   } break;
4532 
4533   case proc_bind_close: {
4534     int f;
4535     int n_th = team->t.t_nproc;
4536     int n_places;
4537     if (first_place <= last_place) {
4538       n_places = last_place - first_place + 1;
4539     } else {
4540       n_places = __kmp_affinity_num_masks - first_place + last_place + 1;
4541     }
4542     if (n_th <= n_places) {
4543       int place = masters_place;
4544       for (f = 1; f < n_th; f++) {
4545         kmp_info_t *th = team->t.t_threads[f];
4546         KMP_DEBUG_ASSERT(th != NULL);
4547 
4548         if (place == last_place) {
4549           place = first_place;
4550         } else if (place == (int)(__kmp_affinity_num_masks - 1)) {
4551           place = 0;
4552         } else {
4553           place++;
4554         }
4555         th->th.th_first_place = first_place;
4556         th->th.th_last_place = last_place;
4557         th->th.th_new_place = place;
4558 
4559         KA_TRACE(100, ("__kmp_partition_places: close: T#%d(%d:%d) place %d "
4560                        "partition = [%d,%d]\n",
4561                        __kmp_gtid_from_thread(team->t.t_threads[f]),
4562                        team->t.t_id, f, place, first_place, last_place));
4563       }
4564     } else {
4565       int S, rem, gap, s_count;
4566       S = n_th / n_places;
4567       s_count = 0;
4568       rem = n_th - (S * n_places);
4569       gap = rem > 0 ? n_places / rem : n_places;
4570       int place = masters_place;
4571       int gap_ct = gap;
4572       for (f = 0; f < n_th; f++) {
4573         kmp_info_t *th = team->t.t_threads[f];
4574         KMP_DEBUG_ASSERT(th != NULL);
4575 
4576         th->th.th_first_place = first_place;
4577         th->th.th_last_place = last_place;
4578         th->th.th_new_place = place;
4579         s_count++;
4580 
4581         if ((s_count == S) && rem && (gap_ct == gap)) {
4582           // do nothing, add an extra thread to place on next iteration
4583         } else if ((s_count == S + 1) && rem && (gap_ct == gap)) {
4584           // we added an extra thread to this place; move to next place
4585           if (place == last_place) {
4586             place = first_place;
4587           } else if (place == (int)(__kmp_affinity_num_masks - 1)) {
4588             place = 0;
4589           } else {
4590             place++;
4591           }
4592           s_count = 0;
4593           gap_ct = 1;
4594           rem--;
4595         } else if (s_count == S) { // place full; don't add extra
4596           if (place == last_place) {
4597             place = first_place;
4598           } else if (place == (int)(__kmp_affinity_num_masks - 1)) {
4599             place = 0;
4600           } else {
4601             place++;
4602           }
4603           gap_ct++;
4604           s_count = 0;
4605         }
4606 
4607         KA_TRACE(100,
4608                  ("__kmp_partition_places: close: T#%d(%d:%d) place %d "
4609                   "partition = [%d,%d]\n",
4610                   __kmp_gtid_from_thread(team->t.t_threads[f]), team->t.t_id, f,
4611                   th->th.th_new_place, first_place, last_place));
4612       }
4613       KMP_DEBUG_ASSERT(place == masters_place);
4614     }
4615   } break;
4616 
4617   case proc_bind_spread: {
4618     int f;
4619     int n_th = team->t.t_nproc;
4620     int n_places;
4621     int thidx;
4622     if (first_place <= last_place) {
4623       n_places = last_place - first_place + 1;
4624     } else {
4625       n_places = __kmp_affinity_num_masks - first_place + last_place + 1;
4626     }
4627     if (n_th <= n_places) {
4628       int place = -1;
4629 
4630       if (n_places != static_cast<int>(__kmp_affinity_num_masks)) {
4631         int S = n_places / n_th;
4632         int s_count, rem, gap, gap_ct;
4633 
4634         place = masters_place;
4635         rem = n_places - n_th * S;
4636         gap = rem ? n_th / rem : 1;
4637         gap_ct = gap;
4638         thidx = n_th;
4639         if (update_master_only == 1)
4640           thidx = 1;
4641         for (f = 0; f < thidx; f++) {
4642           kmp_info_t *th = team->t.t_threads[f];
4643           KMP_DEBUG_ASSERT(th != NULL);
4644 
4645           th->th.th_first_place = place;
4646           th->th.th_new_place = place;
4647           s_count = 1;
4648           while (s_count < S) {
4649             if (place == last_place) {
4650               place = first_place;
4651             } else if (place == (int)(__kmp_affinity_num_masks - 1)) {
4652               place = 0;
4653             } else {
4654               place++;
4655             }
4656             s_count++;
4657           }
4658           if (rem && (gap_ct == gap)) {
4659             if (place == last_place) {
4660               place = first_place;
4661             } else if (place == (int)(__kmp_affinity_num_masks - 1)) {
4662               place = 0;
4663             } else {
4664               place++;
4665             }
4666             rem--;
4667             gap_ct = 0;
4668           }
4669           th->th.th_last_place = place;
4670           gap_ct++;
4671 
4672           if (place == last_place) {
4673             place = first_place;
4674           } else if (place == (int)(__kmp_affinity_num_masks - 1)) {
4675             place = 0;
4676           } else {
4677             place++;
4678           }
4679 
4680           KA_TRACE(100, ("__kmp_partition_places: spread: T#%d(%d:%d) place %d "
4681                          "partition = [%d,%d], __kmp_affinity_num_masks: %u\n",
4682                          __kmp_gtid_from_thread(team->t.t_threads[f]),
4683                          team->t.t_id, f, th->th.th_new_place,
4684                          th->th.th_first_place, th->th.th_last_place,
4685                          __kmp_affinity_num_masks));
4686         }
4687       } else {
4688         /* Having uniform space of available computation places I can create
4689            T partitions of round(P/T) size and put threads into the first
4690            place of each partition. */
4691         double current = static_cast<double>(masters_place);
4692         double spacing =
4693                 (static_cast<double>(n_places + 1) / static_cast<double>(n_th));
4694         int first, last;
4695         kmp_info_t *th;
4696 
4697         thidx = n_th + 1;
4698         if (update_master_only == 1)
4699           thidx = 1;
4700         for (f = 0; f < thidx; f++) {
4701           first = static_cast<int>(current);
4702           last = static_cast<int>(current + spacing) - 1;
4703           KMP_DEBUG_ASSERT(last >= first);
4704           if (first >= n_places) {
4705             if (masters_place) {
4706               first -= n_places;
4707               last -= n_places;
4708               if (first == (masters_place + 1)) {
4709                 KMP_DEBUG_ASSERT(f == n_th);
4710                 first--;
4711               }
4712               if (last == masters_place) {
4713                 KMP_DEBUG_ASSERT(f == (n_th - 1));
4714                 last--;
4715               }
4716             } else {
4717               KMP_DEBUG_ASSERT(f == n_th);
4718               first = 0;
4719               last = 0;
4720             }
4721           }
4722           if (last >= n_places) {
4723             last = (n_places - 1);
4724           }
4725           place = first;
4726           current += spacing;
4727           if (f < n_th) {
4728             KMP_DEBUG_ASSERT(0 <= first);
4729             KMP_DEBUG_ASSERT(n_places > first);
4730             KMP_DEBUG_ASSERT(0 <= last);
4731             KMP_DEBUG_ASSERT(n_places > last);
4732             KMP_DEBUG_ASSERT(last_place >= first_place);
4733             th = team->t.t_threads[f];
4734             KMP_DEBUG_ASSERT(th);
4735             th->th.th_first_place = first;
4736             th->th.th_new_place = place;
4737             th->th.th_last_place = last;
4738 
4739             KA_TRACE(100, ("__kmp_partition_places: spread: T#%d(%d:%d) place %d "
4740                            "partition = [%d,%d], spacing = %.4f\n",
4741                            __kmp_gtid_from_thread(team->t.t_threads[f]),
4742                            team->t.t_id, f, th->th.th_new_place,
4743                            th->th.th_first_place, th->th.th_last_place,
4744                            spacing));
4745           }
4746         }
4747       }
4748       KMP_DEBUG_ASSERT(update_master_only || place == masters_place);
4749     } else {
4750       int S, rem, gap, s_count;
4751       S = n_th / n_places;
4752       s_count = 0;
4753       rem = n_th - (S * n_places);
4754       gap = rem > 0 ? n_places / rem : n_places;
4755       int place = masters_place;
4756       int gap_ct = gap;
4757       thidx = n_th;
4758       if (update_master_only == 1)
4759         thidx = 1;
4760       for (f = 0; f < thidx; f++) {
4761         kmp_info_t *th = team->t.t_threads[f];
4762         KMP_DEBUG_ASSERT(th != NULL);
4763 
4764         th->th.th_first_place = place;
4765         th->th.th_last_place = place;
4766         th->th.th_new_place = place;
4767         s_count++;
4768 
4769         if ((s_count == S) && rem && (gap_ct == gap)) {
4770           // do nothing, add an extra thread to place on next iteration
4771         } else if ((s_count == S + 1) && rem && (gap_ct == gap)) {
4772           // we added an extra thread to this place; move on to next place
4773           if (place == last_place) {
4774             place = first_place;
4775           } else if (place == (int)(__kmp_affinity_num_masks - 1)) {
4776             place = 0;
4777           } else {
4778             place++;
4779           }
4780           s_count = 0;
4781           gap_ct = 1;
4782           rem--;
4783         } else if (s_count == S) { // place is full; don't add extra thread
4784           if (place == last_place) {
4785             place = first_place;
4786           } else if (place == (int)(__kmp_affinity_num_masks - 1)) {
4787             place = 0;
4788           } else {
4789             place++;
4790           }
4791           gap_ct++;
4792           s_count = 0;
4793         }
4794 
4795         KA_TRACE(100, ("__kmp_partition_places: spread: T#%d(%d:%d) place %d "
4796                        "partition = [%d,%d]\n",
4797                        __kmp_gtid_from_thread(team->t.t_threads[f]),
4798                        team->t.t_id, f, th->th.th_new_place,
4799                        th->th.th_first_place, th->th.th_last_place));
4800       }
4801       KMP_DEBUG_ASSERT(update_master_only || place == masters_place);
4802     }
4803   } break;
4804 
4805   default:
4806     break;
4807   }
4808 
4809   KA_TRACE(20, ("__kmp_partition_places: exit T#%d\n", team->t.t_id));
4810 }
4811 
4812 #endif /* OMP_40_ENABLED && KMP_AFFINITY_SUPPORTED */
4813 
4814 /* allocate a new team data structure to use.  take one off of the free pool if
4815    available */
4816 kmp_team_t *
4817 __kmp_allocate_team(kmp_root_t *root, int new_nproc, int max_nproc,
4818 #if OMPT_SUPPORT
4819                     ompt_parallel_id_t ompt_parallel_id,
4820 #endif
4821 #if OMP_40_ENABLED
4822                     kmp_proc_bind_t new_proc_bind,
4823 #endif
4824                     kmp_internal_control_t *new_icvs,
4825                     int argc USE_NESTED_HOT_ARG(kmp_info_t *master)) {
4826   KMP_TIME_DEVELOPER_PARTITIONED_BLOCK(KMP_allocate_team);
4827   int f;
4828   kmp_team_t *team;
4829   int use_hot_team = !root->r.r_active;
4830   int level = 0;
4831 
4832   KA_TRACE(20, ("__kmp_allocate_team: called\n"));
4833   KMP_DEBUG_ASSERT(new_nproc >= 1 && argc >= 0);
4834   KMP_DEBUG_ASSERT(max_nproc >= new_nproc);
4835   KMP_MB();
4836 
4837 #if KMP_NESTED_HOT_TEAMS
4838   kmp_hot_team_ptr_t *hot_teams;
4839   if (master) {
4840     team = master->th.th_team;
4841     level = team->t.t_active_level;
4842     if (master->th.th_teams_microtask) { // in teams construct?
4843       if (master->th.th_teams_size.nteams > 1 &&
4844           ( // #teams > 1
4845               team->t.t_pkfn ==
4846                   (microtask_t)__kmp_teams_master || // inner fork of the teams
4847               master->th.th_teams_level <
4848                   team->t.t_level)) { // or nested parallel inside the teams
4849         ++level; // not increment if #teams==1, or for outer fork of the teams;
4850         // increment otherwise
4851       }
4852     }
4853     hot_teams = master->th.th_hot_teams;
4854     if (level < __kmp_hot_teams_max_level && hot_teams &&
4855         hot_teams[level]
4856             .hot_team) { // hot team has already been allocated for given level
4857       use_hot_team = 1;
4858     } else {
4859       use_hot_team = 0;
4860     }
4861   }
4862 #endif
4863   // Optimization to use a "hot" team
4864   if (use_hot_team && new_nproc > 1) {
4865     KMP_DEBUG_ASSERT(new_nproc == max_nproc);
4866 #if KMP_NESTED_HOT_TEAMS
4867     team = hot_teams[level].hot_team;
4868 #else
4869     team = root->r.r_hot_team;
4870 #endif
4871 #if KMP_DEBUG
4872     if (__kmp_tasking_mode != tskm_immediate_exec) {
4873       KA_TRACE(20, ("__kmp_allocate_team: hot team task_team[0] = %p "
4874                     "task_team[1] = %p before reinit\n",
4875                     team->t.t_task_team[0], team->t.t_task_team[1]));
4876     }
4877 #endif
4878 
4879     // Has the number of threads changed?
4880     /* Let's assume the most common case is that the number of threads is
4881        unchanged, and put that case first. */
4882     if (team->t.t_nproc == new_nproc) { // Check changes in number of threads
4883       KA_TRACE(20, ("__kmp_allocate_team: reusing hot team\n"));
4884       // This case can mean that omp_set_num_threads() was called and the hot
4885       // team size was already reduced, so we check the special flag
4886       if (team->t.t_size_changed == -1) {
4887         team->t.t_size_changed = 1;
4888       } else {
4889         KMP_CHECK_UPDATE(team->t.t_size_changed, 0);
4890       }
4891 
4892       // TODO???: team->t.t_max_active_levels = new_max_active_levels;
4893       kmp_r_sched_t new_sched = new_icvs->sched;
4894       if (team->t.t_sched.r_sched_type != new_sched.r_sched_type ||
4895           team->t.t_sched.chunk != new_sched.chunk)
4896         team->t.t_sched =
4897             new_sched; // set master's schedule as new run-time schedule
4898 
4899       __kmp_reinitialize_team(team, new_icvs,
4900                               root->r.r_uber_thread->th.th_ident);
4901 
4902       KF_TRACE(10, ("__kmp_allocate_team2: T#%d, this_thread=%p team=%p\n", 0,
4903                     team->t.t_threads[0], team));
4904       __kmp_push_current_task_to_thread(team->t.t_threads[0], team, 0);
4905 
4906 #if OMP_40_ENABLED
4907 #if KMP_AFFINITY_SUPPORTED
4908       if ((team->t.t_size_changed == 0) &&
4909           (team->t.t_proc_bind == new_proc_bind)) {
4910         if (new_proc_bind == proc_bind_spread) {
4911           __kmp_partition_places(
4912               team, 1); // add flag to update only master for spread
4913         }
4914         KA_TRACE(200, ("__kmp_allocate_team: reusing hot team #%d bindings: "
4915                        "proc_bind = %d, partition = [%d,%d]\n",
4916                        team->t.t_id, new_proc_bind, team->t.t_first_place,
4917                        team->t.t_last_place));
4918       } else {
4919         KMP_CHECK_UPDATE(team->t.t_proc_bind, new_proc_bind);
4920         __kmp_partition_places(team);
4921       }
4922 #else
4923       KMP_CHECK_UPDATE(team->t.t_proc_bind, new_proc_bind);
4924 #endif /* KMP_AFFINITY_SUPPORTED */
4925 #endif /* OMP_40_ENABLED */
4926     } else if (team->t.t_nproc > new_nproc) {
4927       KA_TRACE(20,
4928                ("__kmp_allocate_team: decreasing hot team thread count to %d\n",
4929                 new_nproc));
4930 
4931       team->t.t_size_changed = 1;
4932 #if KMP_NESTED_HOT_TEAMS
4933       if (__kmp_hot_teams_mode == 0) {
4934         // AC: saved number of threads should correspond to team's value in this
4935         // mode, can be bigger in mode 1, when hot team has threads in reserve
4936         KMP_DEBUG_ASSERT(hot_teams[level].hot_team_nth == team->t.t_nproc);
4937         hot_teams[level].hot_team_nth = new_nproc;
4938 #endif // KMP_NESTED_HOT_TEAMS
4939         /* release the extra threads we don't need any more */
4940         for (f = new_nproc; f < team->t.t_nproc; f++) {
4941           KMP_DEBUG_ASSERT(team->t.t_threads[f]);
4942           if (__kmp_tasking_mode != tskm_immediate_exec) {
4943             // When decreasing team size, threads no longer in the team should
4944             // unref task team.
4945             team->t.t_threads[f]->th.th_task_team = NULL;
4946           }
4947           __kmp_free_thread(team->t.t_threads[f]);
4948           team->t.t_threads[f] = NULL;
4949         }
4950 #if KMP_NESTED_HOT_TEAMS
4951       } // (__kmp_hot_teams_mode == 0)
4952       else {
4953         // When keeping extra threads in team, switch threads to wait on own
4954         // b_go flag
4955         for (f = new_nproc; f < team->t.t_nproc; ++f) {
4956           KMP_DEBUG_ASSERT(team->t.t_threads[f]);
4957           kmp_balign_t *balign = team->t.t_threads[f]->th.th_bar;
4958           for (int b = 0; b < bs_last_barrier; ++b) {
4959             if (balign[b].bb.wait_flag == KMP_BARRIER_PARENT_FLAG) {
4960               balign[b].bb.wait_flag = KMP_BARRIER_SWITCH_TO_OWN_FLAG;
4961             }
4962             KMP_CHECK_UPDATE(balign[b].bb.leaf_kids, 0);
4963           }
4964         }
4965       }
4966 #endif // KMP_NESTED_HOT_TEAMS
4967       team->t.t_nproc = new_nproc;
4968       // TODO???: team->t.t_max_active_levels = new_max_active_levels;
4969       if (team->t.t_sched.r_sched_type != new_icvs->sched.r_sched_type ||
4970           team->t.t_sched.chunk != new_icvs->sched.chunk)
4971         team->t.t_sched = new_icvs->sched;
4972       __kmp_reinitialize_team(team, new_icvs,
4973                               root->r.r_uber_thread->th.th_ident);
4974 
4975       /* update the remaining threads */
4976       for (f = 0; f < new_nproc; ++f) {
4977         team->t.t_threads[f]->th.th_team_nproc = new_nproc;
4978       }
4979       // restore the current task state of the master thread: should be the
4980       // implicit task
4981       KF_TRACE(10, ("__kmp_allocate_team: T#%d, this_thread=%p team=%p\n", 0,
4982                     team->t.t_threads[0], team));
4983 
4984       __kmp_push_current_task_to_thread(team->t.t_threads[0], team, 0);
4985 
4986 #ifdef KMP_DEBUG
4987       for (f = 0; f < team->t.t_nproc; f++) {
4988         KMP_DEBUG_ASSERT(team->t.t_threads[f] &&
4989                          team->t.t_threads[f]->th.th_team_nproc ==
4990                              team->t.t_nproc);
4991       }
4992 #endif
4993 
4994 #if OMP_40_ENABLED
4995       KMP_CHECK_UPDATE(team->t.t_proc_bind, new_proc_bind);
4996 #if KMP_AFFINITY_SUPPORTED
4997       __kmp_partition_places(team);
4998 #endif
4999 #endif
5000     } else { // team->t.t_nproc < new_nproc
5001 #if KMP_OS_LINUX && KMP_AFFINITY_SUPPORTED
5002       kmp_affin_mask_t *old_mask;
5003       if (KMP_AFFINITY_CAPABLE()) {
5004         KMP_CPU_ALLOC(old_mask);
5005       }
5006 #endif
5007 
5008       KA_TRACE(20,
5009                ("__kmp_allocate_team: increasing hot team thread count to %d\n",
5010                 new_nproc));
5011 
5012       team->t.t_size_changed = 1;
5013 
5014 #if KMP_NESTED_HOT_TEAMS
5015       int avail_threads = hot_teams[level].hot_team_nth;
5016       if (new_nproc < avail_threads)
5017         avail_threads = new_nproc;
5018       kmp_info_t **other_threads = team->t.t_threads;
5019       for (f = team->t.t_nproc; f < avail_threads; ++f) {
5020         // Adjust barrier data of reserved threads (if any) of the team
5021         // Other data will be set in __kmp_initialize_info() below.
5022         int b;
5023         kmp_balign_t *balign = other_threads[f]->th.th_bar;
5024         for (b = 0; b < bs_last_barrier; ++b) {
5025           balign[b].bb.b_arrived = team->t.t_bar[b].b_arrived;
5026           KMP_DEBUG_ASSERT(balign[b].bb.wait_flag != KMP_BARRIER_PARENT_FLAG);
5027 #if USE_DEBUGGER
5028           balign[b].bb.b_worker_arrived = team->t.t_bar[b].b_team_arrived;
5029 #endif
5030         }
5031       }
5032       if (hot_teams[level].hot_team_nth >= new_nproc) {
5033         // we have all needed threads in reserve, no need to allocate any
5034         // this only possible in mode 1, cannot have reserved threads in mode 0
5035         KMP_DEBUG_ASSERT(__kmp_hot_teams_mode == 1);
5036         team->t.t_nproc = new_nproc; // just get reserved threads involved
5037       } else {
5038         // we may have some threads in reserve, but not enough
5039         team->t.t_nproc =
5040             hot_teams[level]
5041                 .hot_team_nth; // get reserved threads involved if any
5042         hot_teams[level].hot_team_nth = new_nproc; // adjust hot team max size
5043 #endif // KMP_NESTED_HOT_TEAMS
5044         if (team->t.t_max_nproc < new_nproc) {
5045           /* reallocate larger arrays */
5046           __kmp_reallocate_team_arrays(team, new_nproc);
5047           __kmp_reinitialize_team(team, new_icvs, NULL);
5048         }
5049 
5050 #if KMP_OS_LINUX && KMP_AFFINITY_SUPPORTED
5051         /* Temporarily set full mask for master thread before creation of
5052            workers. The reason is that workers inherit the affinity from master,
5053            so if a lot of workers are created on the single core quickly, they
5054            don't get a chance to set their own affinity for a long time. */
5055         __kmp_set_thread_affinity_mask_full_tmp(old_mask);
5056 #endif
5057 
5058         /* allocate new threads for the hot team */
5059         for (f = team->t.t_nproc; f < new_nproc; f++) {
5060           kmp_info_t *new_worker = __kmp_allocate_thread(root, team, f);
5061           KMP_DEBUG_ASSERT(new_worker);
5062           team->t.t_threads[f] = new_worker;
5063 
5064           KA_TRACE(20,
5065                    ("__kmp_allocate_team: team %d init T#%d arrived: "
5066                     "join=%llu, plain=%llu\n",
5067                     team->t.t_id, __kmp_gtid_from_tid(f, team), team->t.t_id, f,
5068                     team->t.t_bar[bs_forkjoin_barrier].b_arrived,
5069                     team->t.t_bar[bs_plain_barrier].b_arrived));
5070 
5071           { // Initialize barrier data for new threads.
5072             int b;
5073             kmp_balign_t *balign = new_worker->th.th_bar;
5074             for (b = 0; b < bs_last_barrier; ++b) {
5075               balign[b].bb.b_arrived = team->t.t_bar[b].b_arrived;
5076               KMP_DEBUG_ASSERT(balign[b].bb.wait_flag !=
5077                                KMP_BARRIER_PARENT_FLAG);
5078 #if USE_DEBUGGER
5079               balign[b].bb.b_worker_arrived = team->t.t_bar[b].b_team_arrived;
5080 #endif
5081             }
5082           }
5083         }
5084 
5085 #if KMP_OS_LINUX && KMP_AFFINITY_SUPPORTED
5086         if (KMP_AFFINITY_CAPABLE()) {
5087           /* Restore initial master thread's affinity mask */
5088           __kmp_set_system_affinity(old_mask, TRUE);
5089           KMP_CPU_FREE(old_mask);
5090         }
5091 #endif
5092 #if KMP_NESTED_HOT_TEAMS
5093       } // end of check of t_nproc vs. new_nproc vs. hot_team_nth
5094 #endif // KMP_NESTED_HOT_TEAMS
5095       /* make sure everyone is syncronized */
5096       int old_nproc = team->t.t_nproc; // save old value and use to update only
5097       // new threads below
5098       __kmp_initialize_team(team, new_nproc, new_icvs,
5099                             root->r.r_uber_thread->th.th_ident);
5100 
5101       /* reinitialize the threads */
5102       KMP_DEBUG_ASSERT(team->t.t_nproc == new_nproc);
5103       for (f = 0; f < team->t.t_nproc; ++f)
5104         __kmp_initialize_info(team->t.t_threads[f], team, f,
5105                               __kmp_gtid_from_tid(f, team));
5106       if (level) { // set th_task_state for new threads in nested hot team
5107         // __kmp_initialize_info() no longer zeroes th_task_state, so we should
5108         // only need to set the th_task_state for the new threads. th_task_state
5109         // for master thread will not be accurate until after this in
5110         // __kmp_fork_call(), so we look to the master's memo_stack to get the
5111         // correct value.
5112         for (f = old_nproc; f < team->t.t_nproc; ++f)
5113           team->t.t_threads[f]->th.th_task_state =
5114               team->t.t_threads[0]->th.th_task_state_memo_stack[level];
5115       } else { // set th_task_state for new threads in non-nested hot team
5116         int old_state =
5117             team->t.t_threads[0]->th.th_task_state; // copy master's state
5118         for (f = old_nproc; f < team->t.t_nproc; ++f)
5119           team->t.t_threads[f]->th.th_task_state = old_state;
5120       }
5121 
5122 #ifdef KMP_DEBUG
5123       for (f = 0; f < team->t.t_nproc; ++f) {
5124         KMP_DEBUG_ASSERT(team->t.t_threads[f] &&
5125                          team->t.t_threads[f]->th.th_team_nproc ==
5126                              team->t.t_nproc);
5127       }
5128 #endif
5129 
5130 #if OMP_40_ENABLED
5131       KMP_CHECK_UPDATE(team->t.t_proc_bind, new_proc_bind);
5132 #if KMP_AFFINITY_SUPPORTED
5133       __kmp_partition_places(team);
5134 #endif
5135 #endif
5136     } // Check changes in number of threads
5137 
5138 #if OMP_40_ENABLED
5139     kmp_info_t *master = team->t.t_threads[0];
5140     if (master->th.th_teams_microtask) {
5141       for (f = 1; f < new_nproc; ++f) {
5142         // propagate teams construct specific info to workers
5143         kmp_info_t *thr = team->t.t_threads[f];
5144         thr->th.th_teams_microtask = master->th.th_teams_microtask;
5145         thr->th.th_teams_level = master->th.th_teams_level;
5146         thr->th.th_teams_size = master->th.th_teams_size;
5147       }
5148     }
5149 #endif /* OMP_40_ENABLED */
5150 #if KMP_NESTED_HOT_TEAMS
5151     if (level) {
5152       // Sync barrier state for nested hot teams, not needed for outermost hot
5153       // team.
5154       for (f = 1; f < new_nproc; ++f) {
5155         kmp_info_t *thr = team->t.t_threads[f];
5156         int b;
5157         kmp_balign_t *balign = thr->th.th_bar;
5158         for (b = 0; b < bs_last_barrier; ++b) {
5159           balign[b].bb.b_arrived = team->t.t_bar[b].b_arrived;
5160           KMP_DEBUG_ASSERT(balign[b].bb.wait_flag != KMP_BARRIER_PARENT_FLAG);
5161 #if USE_DEBUGGER
5162           balign[b].bb.b_worker_arrived = team->t.t_bar[b].b_team_arrived;
5163 #endif
5164         }
5165       }
5166     }
5167 #endif // KMP_NESTED_HOT_TEAMS
5168 
5169     /* reallocate space for arguments if necessary */
5170     __kmp_alloc_argv_entries(argc, team, TRUE);
5171     KMP_CHECK_UPDATE(team->t.t_argc, argc);
5172     // The hot team re-uses the previous task team,
5173     // if untouched during the previous release->gather phase.
5174 
5175     KF_TRACE(10, (" hot_team = %p\n", team));
5176 
5177 #if KMP_DEBUG
5178     if (__kmp_tasking_mode != tskm_immediate_exec) {
5179       KA_TRACE(20, ("__kmp_allocate_team: hot team task_team[0] = %p "
5180                     "task_team[1] = %p after reinit\n",
5181                     team->t.t_task_team[0], team->t.t_task_team[1]));
5182     }
5183 #endif
5184 
5185 #if OMPT_SUPPORT
5186     __ompt_team_assign_id(team, ompt_parallel_id);
5187 #endif
5188 
5189     KMP_MB();
5190 
5191     return team;
5192   }
5193 
5194   /* next, let's try to take one from the team pool */
5195   KMP_MB();
5196   for (team = CCAST(kmp_team_t *, __kmp_team_pool); (team);) {
5197     /* TODO: consider resizing undersized teams instead of reaping them, now
5198        that we have a resizing mechanism */
5199     if (team->t.t_max_nproc >= max_nproc) {
5200       /* take this team from the team pool */
5201       __kmp_team_pool = team->t.t_next_pool;
5202 
5203       /* setup the team for fresh use */
5204       __kmp_initialize_team(team, new_nproc, new_icvs, NULL);
5205 
5206       KA_TRACE(20, ("__kmp_allocate_team: setting task_team[0] %p and "
5207                     "task_team[1] %p to NULL\n",
5208                     &team->t.t_task_team[0], &team->t.t_task_team[1]));
5209       team->t.t_task_team[0] = NULL;
5210       team->t.t_task_team[1] = NULL;
5211 
5212       /* reallocate space for arguments if necessary */
5213       __kmp_alloc_argv_entries(argc, team, TRUE);
5214       KMP_CHECK_UPDATE(team->t.t_argc, argc);
5215 
5216       KA_TRACE(
5217           20, ("__kmp_allocate_team: team %d init arrived: join=%u, plain=%u\n",
5218                team->t.t_id, KMP_INIT_BARRIER_STATE, KMP_INIT_BARRIER_STATE));
5219       { // Initialize barrier data.
5220         int b;
5221         for (b = 0; b < bs_last_barrier; ++b) {
5222           team->t.t_bar[b].b_arrived = KMP_INIT_BARRIER_STATE;
5223 #if USE_DEBUGGER
5224           team->t.t_bar[b].b_master_arrived = 0;
5225           team->t.t_bar[b].b_team_arrived = 0;
5226 #endif
5227         }
5228       }
5229 
5230 #if OMP_40_ENABLED
5231       team->t.t_proc_bind = new_proc_bind;
5232 #endif
5233 
5234       KA_TRACE(20, ("__kmp_allocate_team: using team from pool %d.\n",
5235                     team->t.t_id));
5236 
5237 #if OMPT_SUPPORT
5238       __ompt_team_assign_id(team, ompt_parallel_id);
5239 #endif
5240 
5241       KMP_MB();
5242 
5243       return team;
5244     }
5245 
5246 /* reap team if it is too small, then loop back and check the next one */
5247 // not sure if this is wise, but, will be redone during the hot-teams rewrite.
5248 /* TODO: Use technique to find the right size hot-team, don't reap them */
5249     team = __kmp_reap_team(team);
5250     __kmp_team_pool = team;
5251   }
5252 
5253   /* nothing available in the pool, no matter, make a new team! */
5254   KMP_MB();
5255   team = (kmp_team_t *)__kmp_allocate(sizeof(kmp_team_t));
5256 
5257   /* and set it up */
5258   team->t.t_max_nproc = max_nproc;
5259   /* NOTE well, for some reason allocating one big buffer and dividing it up
5260      seems to really hurt performance a lot on the P4, so, let's not use this */
5261   __kmp_allocate_team_arrays(team, max_nproc);
5262 
5263   KA_TRACE(20, ("__kmp_allocate_team: making a new team\n"));
5264   __kmp_initialize_team(team, new_nproc, new_icvs, NULL);
5265 
5266   KA_TRACE(20, ("__kmp_allocate_team: setting task_team[0] %p and task_team[1] "
5267                 "%p to NULL\n",
5268                 &team->t.t_task_team[0], &team->t.t_task_team[1]));
5269   team->t.t_task_team[0] = NULL; // to be removed, as __kmp_allocate zeroes
5270   // memory, no need to duplicate
5271   team->t.t_task_team[1] = NULL; // to be removed, as __kmp_allocate zeroes
5272   // memory, no need to duplicate
5273 
5274   if (__kmp_storage_map) {
5275     __kmp_print_team_storage_map("team", team, team->t.t_id, new_nproc);
5276   }
5277 
5278   /* allocate space for arguments */
5279   __kmp_alloc_argv_entries(argc, team, FALSE);
5280   team->t.t_argc = argc;
5281 
5282   KA_TRACE(20,
5283            ("__kmp_allocate_team: team %d init arrived: join=%u, plain=%u\n",
5284             team->t.t_id, KMP_INIT_BARRIER_STATE, KMP_INIT_BARRIER_STATE));
5285   { // Initialize barrier data.
5286     int b;
5287     for (b = 0; b < bs_last_barrier; ++b) {
5288       team->t.t_bar[b].b_arrived = KMP_INIT_BARRIER_STATE;
5289 #if USE_DEBUGGER
5290       team->t.t_bar[b].b_master_arrived = 0;
5291       team->t.t_bar[b].b_team_arrived = 0;
5292 #endif
5293     }
5294   }
5295 
5296 #if OMP_40_ENABLED
5297   team->t.t_proc_bind = new_proc_bind;
5298 #endif
5299 
5300 #if OMPT_SUPPORT
5301   __ompt_team_assign_id(team, ompt_parallel_id);
5302   team->t.ompt_serialized_team_info = NULL;
5303 #endif
5304 
5305   KMP_MB();
5306 
5307   KA_TRACE(20, ("__kmp_allocate_team: done creating a new team %d.\n",
5308                 team->t.t_id));
5309 
5310   return team;
5311 }
5312 
5313 /* TODO implement hot-teams at all levels */
5314 /* TODO implement lazy thread release on demand (disband request) */
5315 
5316 /* free the team.  return it to the team pool.  release all the threads
5317  * associated with it */
5318 void __kmp_free_team(kmp_root_t *root,
5319                      kmp_team_t *team USE_NESTED_HOT_ARG(kmp_info_t *master)) {
5320   int f;
5321   KA_TRACE(20, ("__kmp_free_team: T#%d freeing team %d\n", __kmp_get_gtid(),
5322                 team->t.t_id));
5323 
5324   /* verify state */
5325   KMP_DEBUG_ASSERT(root);
5326   KMP_DEBUG_ASSERT(team);
5327   KMP_DEBUG_ASSERT(team->t.t_nproc <= team->t.t_max_nproc);
5328   KMP_DEBUG_ASSERT(team->t.t_threads);
5329 
5330   int use_hot_team = team == root->r.r_hot_team;
5331 #if KMP_NESTED_HOT_TEAMS
5332   int level;
5333   kmp_hot_team_ptr_t *hot_teams;
5334   if (master) {
5335     level = team->t.t_active_level - 1;
5336     if (master->th.th_teams_microtask) { // in teams construct?
5337       if (master->th.th_teams_size.nteams > 1) {
5338         ++level; // level was not increased in teams construct for
5339         // team_of_masters
5340       }
5341       if (team->t.t_pkfn != (microtask_t)__kmp_teams_master &&
5342           master->th.th_teams_level == team->t.t_level) {
5343         ++level; // level was not increased in teams construct for
5344         // team_of_workers before the parallel
5345       } // team->t.t_level will be increased inside parallel
5346     }
5347     hot_teams = master->th.th_hot_teams;
5348     if (level < __kmp_hot_teams_max_level) {
5349       KMP_DEBUG_ASSERT(team == hot_teams[level].hot_team);
5350       use_hot_team = 1;
5351     }
5352   }
5353 #endif // KMP_NESTED_HOT_TEAMS
5354 
5355   /* team is done working */
5356   TCW_SYNC_PTR(team->t.t_pkfn,
5357                NULL); // Important for Debugging Support Library.
5358   team->t.t_copyin_counter = 0; // init counter for possible reuse
5359   // Do not reset pointer to parent team to NULL for hot teams.
5360 
5361   /* if we are non-hot team, release our threads */
5362   if (!use_hot_team) {
5363     if (__kmp_tasking_mode != tskm_immediate_exec) {
5364       // Wait for threads to reach reapable state
5365       for (f = 1; f < team->t.t_nproc; ++f) {
5366         KMP_DEBUG_ASSERT(team->t.t_threads[f]);
5367         kmp_info_t *th = team->t.t_threads[f];
5368         volatile kmp_uint32 *state = &th->th.th_reap_state;
5369         while (*state != KMP_SAFE_TO_REAP) {
5370 #if KMP_OS_WINDOWS
5371           // On Windows a thread can be killed at any time, check this
5372           DWORD ecode;
5373           if (!__kmp_is_thread_alive(th, &ecode)) {
5374             *state = KMP_SAFE_TO_REAP; // reset the flag for dead thread
5375             break;
5376           }
5377 #endif
5378           // first check if thread is sleeping
5379           kmp_flag_64 fl(&th->th.th_bar[bs_forkjoin_barrier].bb.b_go, th);
5380           if (fl.is_sleeping())
5381             fl.resume(__kmp_gtid_from_thread(th));
5382           KMP_CPU_PAUSE();
5383         }
5384       }
5385 
5386       // Delete task teams
5387       int tt_idx;
5388       for (tt_idx = 0; tt_idx < 2; ++tt_idx) {
5389         kmp_task_team_t *task_team = team->t.t_task_team[tt_idx];
5390         if (task_team != NULL) {
5391           for (f = 0; f < team->t.t_nproc;
5392                ++f) { // Have all threads unref task teams
5393             team->t.t_threads[f]->th.th_task_team = NULL;
5394           }
5395           KA_TRACE(
5396               20,
5397               ("__kmp_free_team: T#%d deactivating task_team %p on team %d\n",
5398                __kmp_get_gtid(), task_team, team->t.t_id));
5399 #if KMP_NESTED_HOT_TEAMS
5400           __kmp_free_task_team(master, task_team);
5401 #endif
5402           team->t.t_task_team[tt_idx] = NULL;
5403         }
5404       }
5405     }
5406 
5407     // Reset pointer to parent team only for non-hot teams.
5408     team->t.t_parent = NULL;
5409     team->t.t_level = 0;
5410     team->t.t_active_level = 0;
5411 
5412     /* free the worker threads */
5413     for (f = 1; f < team->t.t_nproc; ++f) {
5414       KMP_DEBUG_ASSERT(team->t.t_threads[f]);
5415       __kmp_free_thread(team->t.t_threads[f]);
5416       team->t.t_threads[f] = NULL;
5417     }
5418 
5419     /* put the team back in the team pool */
5420     /* TODO limit size of team pool, call reap_team if pool too large */
5421     team->t.t_next_pool = CCAST(kmp_team_t *, __kmp_team_pool);
5422     __kmp_team_pool = (volatile kmp_team_t *)team;
5423   }
5424 
5425   KMP_MB();
5426 }
5427 
5428 /* reap the team.  destroy it, reclaim all its resources and free its memory */
5429 kmp_team_t *__kmp_reap_team(kmp_team_t *team) {
5430   kmp_team_t *next_pool = team->t.t_next_pool;
5431 
5432   KMP_DEBUG_ASSERT(team);
5433   KMP_DEBUG_ASSERT(team->t.t_dispatch);
5434   KMP_DEBUG_ASSERT(team->t.t_disp_buffer);
5435   KMP_DEBUG_ASSERT(team->t.t_threads);
5436   KMP_DEBUG_ASSERT(team->t.t_argv);
5437 
5438   /* TODO clean the threads that are a part of this? */
5439 
5440   /* free stuff */
5441   __kmp_free_team_arrays(team);
5442   if (team->t.t_argv != &team->t.t_inline_argv[0])
5443     __kmp_free((void *)team->t.t_argv);
5444   __kmp_free(team);
5445 
5446   KMP_MB();
5447   return next_pool;
5448 }
5449 
5450 // Free the thread.  Don't reap it, just place it on the pool of available
5451 // threads.
5452 //
5453 // Changes for Quad issue 527845: We need a predictable OMP tid <-> gtid
5454 // binding for the affinity mechanism to be useful.
5455 //
5456 // Now, we always keep the free list (__kmp_thread_pool) sorted by gtid.
5457 // However, we want to avoid a potential performance problem by always
5458 // scanning through the list to find the correct point at which to insert
5459 // the thread (potential N**2 behavior).  To do this we keep track of the
5460 // last place a thread struct was inserted (__kmp_thread_pool_insert_pt).
5461 // With single-level parallelism, threads will always be added to the tail
5462 // of the list, kept track of by __kmp_thread_pool_insert_pt.  With nested
5463 // parallelism, all bets are off and we may need to scan through the entire
5464 // free list.
5465 //
5466 // This change also has a potentially large performance benefit, for some
5467 // applications.  Previously, as threads were freed from the hot team, they
5468 // would be placed back on the free list in inverse order.  If the hot team
5469 // grew back to it's original size, then the freed thread would be placed
5470 // back on the hot team in reverse order.  This could cause bad cache
5471 // locality problems on programs where the size of the hot team regularly
5472 // grew and shrunk.
5473 //
5474 // Now, for single-level parallelism, the OMP tid is alway == gtid.
5475 void __kmp_free_thread(kmp_info_t *this_th) {
5476   int gtid;
5477   kmp_info_t **scan;
5478   kmp_root_t *root = this_th->th.th_root;
5479 
5480   KA_TRACE(20, ("__kmp_free_thread: T#%d putting T#%d back on free pool.\n",
5481                 __kmp_get_gtid(), this_th->th.th_info.ds.ds_gtid));
5482 
5483   KMP_DEBUG_ASSERT(this_th);
5484 
5485   // When moving thread to pool, switch thread to wait on own b_go flag, and
5486   // uninitialized (NULL team).
5487   int b;
5488   kmp_balign_t *balign = this_th->th.th_bar;
5489   for (b = 0; b < bs_last_barrier; ++b) {
5490     if (balign[b].bb.wait_flag == KMP_BARRIER_PARENT_FLAG)
5491       balign[b].bb.wait_flag = KMP_BARRIER_SWITCH_TO_OWN_FLAG;
5492     balign[b].bb.team = NULL;
5493     balign[b].bb.leaf_kids = 0;
5494   }
5495   this_th->th.th_task_state = 0;
5496 
5497   /* put thread back on the free pool */
5498   TCW_PTR(this_th->th.th_team, NULL);
5499   TCW_PTR(this_th->th.th_root, NULL);
5500   TCW_PTR(this_th->th.th_dispatch, NULL); /* NOT NEEDED */
5501 
5502   // If the __kmp_thread_pool_insert_pt is already past the new insert
5503   // point, then we need to re-scan the entire list.
5504   gtid = this_th->th.th_info.ds.ds_gtid;
5505   if (__kmp_thread_pool_insert_pt != NULL) {
5506     KMP_DEBUG_ASSERT(__kmp_thread_pool != NULL);
5507     if (__kmp_thread_pool_insert_pt->th.th_info.ds.ds_gtid > gtid) {
5508       __kmp_thread_pool_insert_pt = NULL;
5509     }
5510   }
5511 
5512   // Scan down the list to find the place to insert the thread.
5513   // scan is the address of a link in the list, possibly the address of
5514   // __kmp_thread_pool itself.
5515   //
5516   // In the absence of nested parallism, the for loop will have 0 iterations.
5517   if (__kmp_thread_pool_insert_pt != NULL) {
5518     scan = &(__kmp_thread_pool_insert_pt->th.th_next_pool);
5519   } else {
5520     scan = CCAST(kmp_info_t **, &__kmp_thread_pool);
5521   }
5522   for (; (*scan != NULL) && ((*scan)->th.th_info.ds.ds_gtid < gtid);
5523        scan = &((*scan)->th.th_next_pool))
5524     ;
5525 
5526   // Insert the new element on the list, and set __kmp_thread_pool_insert_pt
5527   // to its address.
5528   TCW_PTR(this_th->th.th_next_pool, *scan);
5529   __kmp_thread_pool_insert_pt = *scan = this_th;
5530   KMP_DEBUG_ASSERT((this_th->th.th_next_pool == NULL) ||
5531                    (this_th->th.th_info.ds.ds_gtid <
5532                     this_th->th.th_next_pool->th.th_info.ds.ds_gtid));
5533   TCW_4(this_th->th.th_in_pool, TRUE);
5534   __kmp_thread_pool_nth++;
5535 
5536   TCW_4(__kmp_nth, __kmp_nth - 1);
5537   root->r.r_cg_nthreads--;
5538 
5539 #ifdef KMP_ADJUST_BLOCKTIME
5540   /* Adjust blocktime back to user setting or default if necessary */
5541   /* Middle initialization might never have occurred                */
5542   if (!__kmp_env_blocktime && (__kmp_avail_proc > 0)) {
5543     KMP_DEBUG_ASSERT(__kmp_avail_proc > 0);
5544     if (__kmp_nth <= __kmp_avail_proc) {
5545       __kmp_zero_bt = FALSE;
5546     }
5547   }
5548 #endif /* KMP_ADJUST_BLOCKTIME */
5549 
5550   KMP_MB();
5551 }
5552 
5553 /* ------------------------------------------------------------------------ */
5554 
5555 void *__kmp_launch_thread(kmp_info_t *this_thr) {
5556   int gtid = this_thr->th.th_info.ds.ds_gtid;
5557   /*    void                 *stack_data;*/
5558   kmp_team_t *(*volatile pteam);
5559 
5560   KMP_MB();
5561   KA_TRACE(10, ("__kmp_launch_thread: T#%d start\n", gtid));
5562 
5563   if (__kmp_env_consistency_check) {
5564     this_thr->th.th_cons = __kmp_allocate_cons_stack(gtid); // ATT: Memory leak?
5565   }
5566 
5567 #if OMPT_SUPPORT
5568   if (ompt_enabled) {
5569     this_thr->th.ompt_thread_info.state = ompt_state_overhead;
5570     this_thr->th.ompt_thread_info.wait_id = 0;
5571     this_thr->th.ompt_thread_info.idle_frame = __builtin_frame_address(0);
5572     if (ompt_callbacks.ompt_callback(ompt_event_thread_begin)) {
5573       __ompt_thread_begin(ompt_thread_worker, gtid);
5574     }
5575   }
5576 #endif
5577 
5578   /* This is the place where threads wait for work */
5579   while (!TCR_4(__kmp_global.g.g_done)) {
5580     KMP_DEBUG_ASSERT(this_thr == __kmp_threads[gtid]);
5581     KMP_MB();
5582 
5583     /* wait for work to do */
5584     KA_TRACE(20, ("__kmp_launch_thread: T#%d waiting for work\n", gtid));
5585 
5586 #if OMPT_SUPPORT
5587     if (ompt_enabled) {
5588       this_thr->th.ompt_thread_info.state = ompt_state_idle;
5589     }
5590 #endif
5591 
5592     /* No tid yet since not part of a team */
5593     __kmp_fork_barrier(gtid, KMP_GTID_DNE);
5594 
5595 #if OMPT_SUPPORT
5596     if (ompt_enabled) {
5597       this_thr->th.ompt_thread_info.state = ompt_state_overhead;
5598     }
5599 #endif
5600 
5601     pteam = (kmp_team_t * (*))(&this_thr->th.th_team);
5602 
5603     /* have we been allocated? */
5604     if (TCR_SYNC_PTR(*pteam) && !TCR_4(__kmp_global.g.g_done)) {
5605 #if OMPT_SUPPORT
5606       ompt_task_info_t *task_info;
5607       ompt_parallel_id_t my_parallel_id;
5608       if (ompt_enabled) {
5609         task_info = __ompt_get_taskinfo(0);
5610         my_parallel_id = (*pteam)->t.ompt_team_info.parallel_id;
5611       }
5612 #endif
5613       /* we were just woken up, so run our new task */
5614       if (TCR_SYNC_PTR((*pteam)->t.t_pkfn) != NULL) {
5615         int rc;
5616         KA_TRACE(20,
5617                  ("__kmp_launch_thread: T#%d(%d:%d) invoke microtask = %p\n",
5618                   gtid, (*pteam)->t.t_id, __kmp_tid_from_gtid(gtid),
5619                   (*pteam)->t.t_pkfn));
5620 
5621         updateHWFPControl(*pteam);
5622 
5623 #if OMPT_SUPPORT
5624         if (ompt_enabled) {
5625           this_thr->th.ompt_thread_info.state = ompt_state_work_parallel;
5626           // Initialize OMPT task id for implicit task.
5627           int tid = __kmp_tid_from_gtid(gtid);
5628           task_info->task_id = __ompt_task_id_new(tid);
5629         }
5630 #endif
5631 
5632         {
5633           KMP_TIME_PARTITIONED_BLOCK(OMP_parallel);
5634           KMP_SET_THREAD_STATE_BLOCK(IMPLICIT_TASK);
5635           rc = (*pteam)->t.t_invoke(gtid);
5636         }
5637         KMP_ASSERT(rc);
5638 
5639 #if OMPT_SUPPORT
5640         if (ompt_enabled) {
5641           /* no frame set while outside task */
5642           task_info->frame.exit_runtime_frame = NULL;
5643 
5644           this_thr->th.ompt_thread_info.state = ompt_state_overhead;
5645         }
5646 #endif
5647         KMP_MB();
5648         KA_TRACE(20, ("__kmp_launch_thread: T#%d(%d:%d) done microtask = %p\n",
5649                       gtid, (*pteam)->t.t_id, __kmp_tid_from_gtid(gtid),
5650                       (*pteam)->t.t_pkfn));
5651       }
5652       /* join barrier after parallel region */
5653       __kmp_join_barrier(gtid);
5654 #if OMPT_SUPPORT && OMPT_TRACE
5655       if (ompt_enabled) {
5656         if (ompt_callbacks.ompt_callback(ompt_event_implicit_task_end)) {
5657           // don't access *pteam here: it may have already been freed
5658           // by the master thread behind the barrier (possible race)
5659           ompt_callbacks.ompt_callback(ompt_event_implicit_task_end)(
5660               my_parallel_id, task_info->task_id);
5661         }
5662         task_info->frame.exit_runtime_frame = NULL;
5663         task_info->task_id = 0;
5664       }
5665 #endif
5666     }
5667   }
5668   TCR_SYNC_PTR((intptr_t)__kmp_global.g.g_done);
5669 
5670 #if OMPT_SUPPORT
5671   if (ompt_enabled && ompt_callbacks.ompt_callback(ompt_event_thread_end)) {
5672     __ompt_thread_end(ompt_thread_worker, gtid);
5673   }
5674 #endif
5675 
5676   this_thr->th.th_task_team = NULL;
5677   /* run the destructors for the threadprivate data for this thread */
5678   __kmp_common_destroy_gtid(gtid);
5679 
5680   KA_TRACE(10, ("__kmp_launch_thread: T#%d done\n", gtid));
5681   KMP_MB();
5682   return this_thr;
5683 }
5684 
5685 /* ------------------------------------------------------------------------ */
5686 
5687 void __kmp_internal_end_dest(void *specific_gtid) {
5688 #if KMP_COMPILER_ICC
5689 #pragma warning(push)
5690 #pragma warning(disable : 810) // conversion from "void *" to "int" may lose
5691 // significant bits
5692 #endif
5693   // Make sure no significant bits are lost
5694   int gtid = (kmp_intptr_t)specific_gtid - 1;
5695 #if KMP_COMPILER_ICC
5696 #pragma warning(pop)
5697 #endif
5698 
5699   KA_TRACE(30, ("__kmp_internal_end_dest: T#%d\n", gtid));
5700   /* NOTE: the gtid is stored as gitd+1 in the thread-local-storage
5701    * this is because 0 is reserved for the nothing-stored case */
5702 
5703   /* josh: One reason for setting the gtid specific data even when it is being
5704      destroyed by pthread is to allow gtid lookup through thread specific data
5705      (__kmp_gtid_get_specific).  Some of the code, especially stat code,
5706      that gets executed in the call to __kmp_internal_end_thread, actually
5707      gets the gtid through the thread specific data.  Setting it here seems
5708      rather inelegant and perhaps wrong, but allows __kmp_internal_end_thread
5709      to run smoothly.
5710      todo: get rid of this after we remove the dependence on
5711      __kmp_gtid_get_specific  */
5712   if (gtid >= 0 && KMP_UBER_GTID(gtid))
5713     __kmp_gtid_set_specific(gtid);
5714 #ifdef KMP_TDATA_GTID
5715   __kmp_gtid = gtid;
5716 #endif
5717   __kmp_internal_end_thread(gtid);
5718 }
5719 
5720 #if KMP_OS_UNIX && KMP_DYNAMIC_LIB
5721 
5722 // 2009-09-08 (lev): It looks the destructor does not work. In simple test cases
5723 // destructors work perfectly, but in real libomp.so I have no evidence it is
5724 // ever called. However, -fini linker option in makefile.mk works fine.
5725 
5726 __attribute__((destructor)) void __kmp_internal_end_dtor(void) {
5727   __kmp_internal_end_atexit();
5728 }
5729 
5730 void __kmp_internal_end_fini(void) { __kmp_internal_end_atexit(); }
5731 
5732 #endif
5733 
5734 /* [Windows] josh: when the atexit handler is called, there may still be more
5735    than one thread alive */
5736 void __kmp_internal_end_atexit(void) {
5737   KA_TRACE(30, ("__kmp_internal_end_atexit\n"));
5738   /* [Windows]
5739      josh: ideally, we want to completely shutdown the library in this atexit
5740      handler, but stat code that depends on thread specific data for gtid fails
5741      because that data becomes unavailable at some point during the shutdown, so
5742      we call __kmp_internal_end_thread instead. We should eventually remove the
5743      dependency on __kmp_get_specific_gtid in the stat code and use
5744      __kmp_internal_end_library to cleanly shutdown the library.
5745 
5746      // TODO: Can some of this comment about GVS be removed?
5747      I suspect that the offending stat code is executed when the calling thread
5748      tries to clean up a dead root thread's data structures, resulting in GVS
5749      code trying to close the GVS structures for that thread, but since the stat
5750      code uses __kmp_get_specific_gtid to get the gtid with the assumption that
5751      the calling thread is cleaning up itself instead of another thread, it get
5752      confused. This happens because allowing a thread to unregister and cleanup
5753      another thread is a recent modification for addressing an issue.
5754      Based on the current design (20050722), a thread may end up
5755      trying to unregister another thread only if thread death does not trigger
5756      the calling of __kmp_internal_end_thread.  For Linux* OS, there is the
5757      thread specific data destructor function to detect thread death. For
5758      Windows dynamic, there is DllMain(THREAD_DETACH). For Windows static, there
5759      is nothing.  Thus, the workaround is applicable only for Windows static
5760      stat library. */
5761   __kmp_internal_end_library(-1);
5762 #if KMP_OS_WINDOWS
5763   __kmp_close_console();
5764 #endif
5765 }
5766 
5767 static void __kmp_reap_thread(kmp_info_t *thread, int is_root) {
5768   // It is assumed __kmp_forkjoin_lock is acquired.
5769 
5770   int gtid;
5771 
5772   KMP_DEBUG_ASSERT(thread != NULL);
5773 
5774   gtid = thread->th.th_info.ds.ds_gtid;
5775 
5776   if (!is_root) {
5777 
5778     if (__kmp_dflt_blocktime != KMP_MAX_BLOCKTIME) {
5779       /* Assume the threads are at the fork barrier here */
5780       KA_TRACE(
5781           20, ("__kmp_reap_thread: releasing T#%d from fork barrier for reap\n",
5782                gtid));
5783       /* Need release fence here to prevent seg faults for tree forkjoin barrier
5784        * (GEH) */
5785       ANNOTATE_HAPPENS_BEFORE(thread);
5786       kmp_flag_64 flag(&thread->th.th_bar[bs_forkjoin_barrier].bb.b_go, thread);
5787       __kmp_release_64(&flag);
5788     }; // if
5789 
5790     // Terminate OS thread.
5791     __kmp_reap_worker(thread);
5792 
5793     // The thread was killed asynchronously.  If it was actively
5794     // spinning in the thread pool, decrement the global count.
5795     //
5796     // There is a small timing hole here - if the worker thread was just waking
5797     // up after sleeping in the pool, had reset it's th_active_in_pool flag but
5798     // not decremented the global counter __kmp_thread_pool_active_nth yet, then
5799     // the global counter might not get updated.
5800     //
5801     // Currently, this can only happen as the library is unloaded,
5802     // so there are no harmful side effects.
5803     if (thread->th.th_active_in_pool) {
5804       thread->th.th_active_in_pool = FALSE;
5805       KMP_TEST_THEN_DEC32(&__kmp_thread_pool_active_nth);
5806       KMP_DEBUG_ASSERT(TCR_4(__kmp_thread_pool_active_nth) >= 0);
5807     }
5808 
5809     // Decrement # of [worker] threads in the pool.
5810     KMP_DEBUG_ASSERT(__kmp_thread_pool_nth > 0);
5811     --__kmp_thread_pool_nth;
5812   }; // if
5813 
5814   __kmp_free_implicit_task(thread);
5815 
5816 // Free the fast memory for tasking
5817 #if USE_FAST_MEMORY
5818   __kmp_free_fast_memory(thread);
5819 #endif /* USE_FAST_MEMORY */
5820 
5821   __kmp_suspend_uninitialize_thread(thread);
5822 
5823   KMP_DEBUG_ASSERT(__kmp_threads[gtid] == thread);
5824   TCW_SYNC_PTR(__kmp_threads[gtid], NULL);
5825 
5826   --__kmp_all_nth;
5827 // __kmp_nth was decremented when thread is added to the pool.
5828 
5829 #ifdef KMP_ADJUST_BLOCKTIME
5830   /* Adjust blocktime back to user setting or default if necessary */
5831   /* Middle initialization might never have occurred                */
5832   if (!__kmp_env_blocktime && (__kmp_avail_proc > 0)) {
5833     KMP_DEBUG_ASSERT(__kmp_avail_proc > 0);
5834     if (__kmp_nth <= __kmp_avail_proc) {
5835       __kmp_zero_bt = FALSE;
5836     }
5837   }
5838 #endif /* KMP_ADJUST_BLOCKTIME */
5839 
5840   /* free the memory being used */
5841   if (__kmp_env_consistency_check) {
5842     if (thread->th.th_cons) {
5843       __kmp_free_cons_stack(thread->th.th_cons);
5844       thread->th.th_cons = NULL;
5845     }; // if
5846   }
5847 
5848   if (thread->th.th_pri_common != NULL) {
5849     __kmp_free(thread->th.th_pri_common);
5850     thread->th.th_pri_common = NULL;
5851   }; // if
5852 
5853   if (thread->th.th_task_state_memo_stack != NULL) {
5854     __kmp_free(thread->th.th_task_state_memo_stack);
5855     thread->th.th_task_state_memo_stack = NULL;
5856   }
5857 
5858 #if KMP_USE_BGET
5859   if (thread->th.th_local.bget_data != NULL) {
5860     __kmp_finalize_bget(thread);
5861   }; // if
5862 #endif
5863 
5864 #if KMP_AFFINITY_SUPPORTED
5865   if (thread->th.th_affin_mask != NULL) {
5866     KMP_CPU_FREE(thread->th.th_affin_mask);
5867     thread->th.th_affin_mask = NULL;
5868   }; // if
5869 #endif /* KMP_AFFINITY_SUPPORTED */
5870 
5871   __kmp_reap_team(thread->th.th_serial_team);
5872   thread->th.th_serial_team = NULL;
5873   __kmp_free(thread);
5874 
5875   KMP_MB();
5876 
5877 } // __kmp_reap_thread
5878 
5879 static void __kmp_internal_end(void) {
5880   int i;
5881 
5882   /* First, unregister the library */
5883   __kmp_unregister_library();
5884 
5885 #if KMP_OS_WINDOWS
5886   /* In Win static library, we can't tell when a root actually dies, so we
5887      reclaim the data structures for any root threads that have died but not
5888      unregistered themselves, in order to shut down cleanly.
5889      In Win dynamic library we also can't tell when a thread dies.  */
5890   __kmp_reclaim_dead_roots(); // AC: moved here to always clean resources of
5891 // dead roots
5892 #endif
5893 
5894   for (i = 0; i < __kmp_threads_capacity; i++)
5895     if (__kmp_root[i])
5896       if (__kmp_root[i]->r.r_active)
5897         break;
5898   KMP_MB(); /* Flush all pending memory write invalidates.  */
5899   TCW_SYNC_4(__kmp_global.g.g_done, TRUE);
5900 
5901   if (i < __kmp_threads_capacity) {
5902 #if KMP_USE_MONITOR
5903     // 2009-09-08 (lev): Other alive roots found. Why do we kill the monitor??
5904     KMP_MB(); /* Flush all pending memory write invalidates.  */
5905 
5906 // Need to check that monitor was initialized before reaping it. If we are
5907 // called form __kmp_atfork_child (which sets __kmp_init_parallel = 0), then
5908 // __kmp_monitor will appear to contain valid data, but it is only valid in the
5909 // parent process, not the child.
5910     // New behavior (201008): instead of keying off of the flag
5911     // __kmp_init_parallel, the monitor thread creation is keyed off
5912     // of the new flag __kmp_init_monitor.
5913     __kmp_acquire_bootstrap_lock(&__kmp_monitor_lock);
5914     if (TCR_4(__kmp_init_monitor)) {
5915       __kmp_reap_monitor(&__kmp_monitor);
5916       TCW_4(__kmp_init_monitor, 0);
5917     }
5918     __kmp_release_bootstrap_lock(&__kmp_monitor_lock);
5919     KA_TRACE(10, ("__kmp_internal_end: monitor reaped\n"));
5920 #endif // KMP_USE_MONITOR
5921   } else {
5922 /* TODO move this to cleanup code */
5923 #ifdef KMP_DEBUG
5924     /* make sure that everything has properly ended */
5925     for (i = 0; i < __kmp_threads_capacity; i++) {
5926       if (__kmp_root[i]) {
5927         //                    KMP_ASSERT( ! KMP_UBER_GTID( i ) );         // AC:
5928         //                    there can be uber threads alive here
5929         KMP_ASSERT(!__kmp_root[i]->r.r_active); // TODO: can they be active?
5930       }
5931     }
5932 #endif
5933 
5934     KMP_MB();
5935 
5936     // Reap the worker threads.
5937     // This is valid for now, but be careful if threads are reaped sooner.
5938     while (__kmp_thread_pool != NULL) { // Loop thru all the thread in the pool.
5939       // Get the next thread from the pool.
5940       kmp_info_t *thread = CCAST(kmp_info_t *, __kmp_thread_pool);
5941       __kmp_thread_pool = thread->th.th_next_pool;
5942       // Reap it.
5943       KMP_DEBUG_ASSERT(thread->th.th_reap_state == KMP_SAFE_TO_REAP);
5944       thread->th.th_next_pool = NULL;
5945       thread->th.th_in_pool = FALSE;
5946       __kmp_reap_thread(thread, 0);
5947     }; // while
5948     __kmp_thread_pool_insert_pt = NULL;
5949 
5950     // Reap teams.
5951     while (__kmp_team_pool != NULL) { // Loop thru all the teams in the pool.
5952       // Get the next team from the pool.
5953       kmp_team_t *team = CCAST(kmp_team_t *, __kmp_team_pool);
5954       __kmp_team_pool = team->t.t_next_pool;
5955       // Reap it.
5956       team->t.t_next_pool = NULL;
5957       __kmp_reap_team(team);
5958     }; // while
5959 
5960     __kmp_reap_task_teams();
5961 
5962     for (i = 0; i < __kmp_threads_capacity; ++i) {
5963       // TBD: Add some checking...
5964       // Something like KMP_DEBUG_ASSERT( __kmp_thread[ i ] == NULL );
5965     }
5966 
5967     /* Make sure all threadprivate destructors get run by joining with all
5968        worker threads before resetting this flag */
5969     TCW_SYNC_4(__kmp_init_common, FALSE);
5970 
5971     KA_TRACE(10, ("__kmp_internal_end: all workers reaped\n"));
5972     KMP_MB();
5973 
5974 #if KMP_USE_MONITOR
5975     // See note above: One of the possible fixes for CQ138434 / CQ140126
5976     //
5977     // FIXME: push both code fragments down and CSE them?
5978     // push them into __kmp_cleanup() ?
5979     __kmp_acquire_bootstrap_lock(&__kmp_monitor_lock);
5980     if (TCR_4(__kmp_init_monitor)) {
5981       __kmp_reap_monitor(&__kmp_monitor);
5982       TCW_4(__kmp_init_monitor, 0);
5983     }
5984     __kmp_release_bootstrap_lock(&__kmp_monitor_lock);
5985     KA_TRACE(10, ("__kmp_internal_end: monitor reaped\n"));
5986 #endif
5987   } /* else !__kmp_global.t_active */
5988   TCW_4(__kmp_init_gtid, FALSE);
5989   KMP_MB(); /* Flush all pending memory write invalidates.  */
5990 
5991   __kmp_cleanup();
5992 #if OMPT_SUPPORT
5993   ompt_fini();
5994 #endif
5995 }
5996 
5997 void __kmp_internal_end_library(int gtid_req) {
5998   /* if we have already cleaned up, don't try again, it wouldn't be pretty */
5999   /* this shouldn't be a race condition because __kmp_internal_end() is the
6000      only place to clear __kmp_serial_init */
6001   /* we'll check this later too, after we get the lock */
6002   // 2009-09-06: We do not set g_abort without setting g_done. This check looks
6003   // redundaant, because the next check will work in any case.
6004   if (__kmp_global.g.g_abort) {
6005     KA_TRACE(11, ("__kmp_internal_end_library: abort, exiting\n"));
6006     /* TODO abort? */
6007     return;
6008   }
6009   if (TCR_4(__kmp_global.g.g_done) || !__kmp_init_serial) {
6010     KA_TRACE(10, ("__kmp_internal_end_library: already finished\n"));
6011     return;
6012   }
6013 
6014   KMP_MB(); /* Flush all pending memory write invalidates.  */
6015 
6016   /* find out who we are and what we should do */
6017   {
6018     int gtid = (gtid_req >= 0) ? gtid_req : __kmp_gtid_get_specific();
6019     KA_TRACE(
6020         10, ("__kmp_internal_end_library: enter T#%d  (%d)\n", gtid, gtid_req));
6021     if (gtid == KMP_GTID_SHUTDOWN) {
6022       KA_TRACE(10, ("__kmp_internal_end_library: !__kmp_init_runtime, system "
6023                     "already shutdown\n"));
6024       return;
6025     } else if (gtid == KMP_GTID_MONITOR) {
6026       KA_TRACE(10, ("__kmp_internal_end_library: monitor thread, gtid not "
6027                     "registered, or system shutdown\n"));
6028       return;
6029     } else if (gtid == KMP_GTID_DNE) {
6030       KA_TRACE(10, ("__kmp_internal_end_library: gtid not registered or system "
6031                     "shutdown\n"));
6032       /* we don't know who we are, but we may still shutdown the library */
6033     } else if (KMP_UBER_GTID(gtid)) {
6034       /* unregister ourselves as an uber thread.  gtid is no longer valid */
6035       if (__kmp_root[gtid]->r.r_active) {
6036         __kmp_global.g.g_abort = -1;
6037         TCW_SYNC_4(__kmp_global.g.g_done, TRUE);
6038         KA_TRACE(10,
6039                  ("__kmp_internal_end_library: root still active, abort T#%d\n",
6040                   gtid));
6041         return;
6042       } else {
6043         KA_TRACE(
6044             10,
6045             ("__kmp_internal_end_library: unregistering sibling T#%d\n", gtid));
6046         __kmp_unregister_root_current_thread(gtid);
6047       }
6048     } else {
6049 /* worker threads may call this function through the atexit handler, if they
6050  * call exit() */
6051 /* For now, skip the usual subsequent processing and just dump the debug buffer.
6052    TODO: do a thorough shutdown instead */
6053 #ifdef DUMP_DEBUG_ON_EXIT
6054       if (__kmp_debug_buf)
6055         __kmp_dump_debug_buffer();
6056 #endif
6057       return;
6058     }
6059   }
6060   /* synchronize the termination process */
6061   __kmp_acquire_bootstrap_lock(&__kmp_initz_lock);
6062 
6063   /* have we already finished */
6064   if (__kmp_global.g.g_abort) {
6065     KA_TRACE(10, ("__kmp_internal_end_library: abort, exiting\n"));
6066     /* TODO abort? */
6067     __kmp_release_bootstrap_lock(&__kmp_initz_lock);
6068     return;
6069   }
6070   if (TCR_4(__kmp_global.g.g_done) || !__kmp_init_serial) {
6071     __kmp_release_bootstrap_lock(&__kmp_initz_lock);
6072     return;
6073   }
6074 
6075   /* We need this lock to enforce mutex between this reading of
6076      __kmp_threads_capacity and the writing by __kmp_register_root.
6077      Alternatively, we can use a counter of roots that is atomically updated by
6078      __kmp_get_global_thread_id_reg, __kmp_do_serial_initialize and
6079      __kmp_internal_end_*.  */
6080   __kmp_acquire_bootstrap_lock(&__kmp_forkjoin_lock);
6081 
6082   /* now we can safely conduct the actual termination */
6083   __kmp_internal_end();
6084 
6085   __kmp_release_bootstrap_lock(&__kmp_forkjoin_lock);
6086   __kmp_release_bootstrap_lock(&__kmp_initz_lock);
6087 
6088   KA_TRACE(10, ("__kmp_internal_end_library: exit\n"));
6089 
6090 #ifdef DUMP_DEBUG_ON_EXIT
6091   if (__kmp_debug_buf)
6092     __kmp_dump_debug_buffer();
6093 #endif
6094 
6095 #if KMP_OS_WINDOWS
6096   __kmp_close_console();
6097 #endif
6098 
6099   __kmp_fini_allocator();
6100 
6101 } // __kmp_internal_end_library
6102 
6103 void __kmp_internal_end_thread(int gtid_req) {
6104   int i;
6105 
6106   /* if we have already cleaned up, don't try again, it wouldn't be pretty */
6107   /* this shouldn't be a race condition because __kmp_internal_end() is the
6108    * only place to clear __kmp_serial_init */
6109   /* we'll check this later too, after we get the lock */
6110   // 2009-09-06: We do not set g_abort without setting g_done. This check looks
6111   // redundant, because the next check will work in any case.
6112   if (__kmp_global.g.g_abort) {
6113     KA_TRACE(11, ("__kmp_internal_end_thread: abort, exiting\n"));
6114     /* TODO abort? */
6115     return;
6116   }
6117   if (TCR_4(__kmp_global.g.g_done) || !__kmp_init_serial) {
6118     KA_TRACE(10, ("__kmp_internal_end_thread: already finished\n"));
6119     return;
6120   }
6121 
6122   KMP_MB(); /* Flush all pending memory write invalidates.  */
6123 
6124   /* find out who we are and what we should do */
6125   {
6126     int gtid = (gtid_req >= 0) ? gtid_req : __kmp_gtid_get_specific();
6127     KA_TRACE(10,
6128              ("__kmp_internal_end_thread: enter T#%d  (%d)\n", gtid, gtid_req));
6129     if (gtid == KMP_GTID_SHUTDOWN) {
6130       KA_TRACE(10, ("__kmp_internal_end_thread: !__kmp_init_runtime, system "
6131                     "already shutdown\n"));
6132       return;
6133     } else if (gtid == KMP_GTID_MONITOR) {
6134       KA_TRACE(10, ("__kmp_internal_end_thread: monitor thread, gtid not "
6135                     "registered, or system shutdown\n"));
6136       return;
6137     } else if (gtid == KMP_GTID_DNE) {
6138       KA_TRACE(10, ("__kmp_internal_end_thread: gtid not registered or system "
6139                     "shutdown\n"));
6140       return;
6141       /* we don't know who we are */
6142     } else if (KMP_UBER_GTID(gtid)) {
6143       /* unregister ourselves as an uber thread.  gtid is no longer valid */
6144       if (__kmp_root[gtid]->r.r_active) {
6145         __kmp_global.g.g_abort = -1;
6146         TCW_SYNC_4(__kmp_global.g.g_done, TRUE);
6147         KA_TRACE(10,
6148                  ("__kmp_internal_end_thread: root still active, abort T#%d\n",
6149                   gtid));
6150         return;
6151       } else {
6152         KA_TRACE(10, ("__kmp_internal_end_thread: unregistering sibling T#%d\n",
6153                       gtid));
6154         __kmp_unregister_root_current_thread(gtid);
6155       }
6156     } else {
6157       /* just a worker thread, let's leave */
6158       KA_TRACE(10, ("__kmp_internal_end_thread: worker thread T#%d\n", gtid));
6159 
6160       if (gtid >= 0) {
6161         __kmp_threads[gtid]->th.th_task_team = NULL;
6162       }
6163 
6164       KA_TRACE(10,
6165                ("__kmp_internal_end_thread: worker thread done, exiting T#%d\n",
6166                 gtid));
6167       return;
6168     }
6169   }
6170 #if defined KMP_DYNAMIC_LIB
6171   // AC: lets not shutdown the Linux* OS dynamic library at the exit of uber
6172   // thread, because we will better shutdown later in the library destructor.
6173   // The reason of this change is performance problem when non-openmp thread in
6174   // a loop forks and joins many openmp threads. We can save a lot of time
6175   // keeping worker threads alive until the program shutdown.
6176   // OM: Removed Linux* OS restriction to fix the crash on OS X* (DPD200239966)
6177   // and Windows(DPD200287443) that occurs when using critical sections from
6178   // foreign threads.
6179   KA_TRACE(10, ("__kmp_internal_end_thread: exiting T#%d\n", gtid_req));
6180   return;
6181 #endif
6182   /* synchronize the termination process */
6183   __kmp_acquire_bootstrap_lock(&__kmp_initz_lock);
6184 
6185   /* have we already finished */
6186   if (__kmp_global.g.g_abort) {
6187     KA_TRACE(10, ("__kmp_internal_end_thread: abort, exiting\n"));
6188     /* TODO abort? */
6189     __kmp_release_bootstrap_lock(&__kmp_initz_lock);
6190     return;
6191   }
6192   if (TCR_4(__kmp_global.g.g_done) || !__kmp_init_serial) {
6193     __kmp_release_bootstrap_lock(&__kmp_initz_lock);
6194     return;
6195   }
6196 
6197   /* We need this lock to enforce mutex between this reading of
6198      __kmp_threads_capacity and the writing by __kmp_register_root.
6199      Alternatively, we can use a counter of roots that is atomically updated by
6200      __kmp_get_global_thread_id_reg, __kmp_do_serial_initialize and
6201      __kmp_internal_end_*.  */
6202 
6203   /* should we finish the run-time?  are all siblings done? */
6204   __kmp_acquire_bootstrap_lock(&__kmp_forkjoin_lock);
6205 
6206   for (i = 0; i < __kmp_threads_capacity; ++i) {
6207     if (KMP_UBER_GTID(i)) {
6208       KA_TRACE(
6209           10,
6210           ("__kmp_internal_end_thread: remaining sibling task: gtid==%d\n", i));
6211       __kmp_release_bootstrap_lock(&__kmp_forkjoin_lock);
6212       __kmp_release_bootstrap_lock(&__kmp_initz_lock);
6213       return;
6214     };
6215   }
6216 
6217   /* now we can safely conduct the actual termination */
6218 
6219   __kmp_internal_end();
6220 
6221   __kmp_release_bootstrap_lock(&__kmp_forkjoin_lock);
6222   __kmp_release_bootstrap_lock(&__kmp_initz_lock);
6223 
6224   KA_TRACE(10, ("__kmp_internal_end_thread: exit T#%d\n", gtid_req));
6225 
6226 #ifdef DUMP_DEBUG_ON_EXIT
6227   if (__kmp_debug_buf)
6228     __kmp_dump_debug_buffer();
6229 #endif
6230 } // __kmp_internal_end_thread
6231 
6232 // -----------------------------------------------------------------------------
6233 // Library registration stuff.
6234 
6235 static long __kmp_registration_flag = 0;
6236 // Random value used to indicate library initialization.
6237 static char *__kmp_registration_str = NULL;
6238 // Value to be saved in env var __KMP_REGISTERED_LIB_<pid>.
6239 
6240 static inline char *__kmp_reg_status_name() {
6241   /* On RHEL 3u5 if linked statically, getpid() returns different values in
6242      each thread. If registration and unregistration go in different threads
6243      (omp_misc_other_root_exit.cpp test case), the name of registered_lib_env
6244      env var can not be found, because the name will contain different pid. */
6245   return __kmp_str_format("__KMP_REGISTERED_LIB_%d", (int)getpid());
6246 } // __kmp_reg_status_get
6247 
6248 void __kmp_register_library_startup(void) {
6249 
6250   char *name = __kmp_reg_status_name(); // Name of the environment variable.
6251   int done = 0;
6252   union {
6253     double dtime;
6254     long ltime;
6255   } time;
6256 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
6257   __kmp_initialize_system_tick();
6258 #endif
6259   __kmp_read_system_time(&time.dtime);
6260   __kmp_registration_flag = 0xCAFE0000L | (time.ltime & 0x0000FFFFL);
6261   __kmp_registration_str =
6262       __kmp_str_format("%p-%lx-%s", &__kmp_registration_flag,
6263                        __kmp_registration_flag, KMP_LIBRARY_FILE);
6264 
6265   KA_TRACE(50, ("__kmp_register_library_startup: %s=\"%s\"\n", name,
6266                 __kmp_registration_str));
6267 
6268   while (!done) {
6269 
6270     char *value = NULL; // Actual value of the environment variable.
6271 
6272     // Set environment variable, but do not overwrite if it is exist.
6273     __kmp_env_set(name, __kmp_registration_str, 0);
6274     // Check the variable is written.
6275     value = __kmp_env_get(name);
6276     if (value != NULL && strcmp(value, __kmp_registration_str) == 0) {
6277 
6278       done = 1; // Ok, environment variable set successfully, exit the loop.
6279 
6280     } else {
6281 
6282       // Oops. Write failed. Another copy of OpenMP RTL is in memory.
6283       // Check whether it alive or dead.
6284       int neighbor = 0; // 0 -- unknown status, 1 -- alive, 2 -- dead.
6285       char *tail = value;
6286       char *flag_addr_str = NULL;
6287       char *flag_val_str = NULL;
6288       char const *file_name = NULL;
6289       __kmp_str_split(tail, '-', &flag_addr_str, &tail);
6290       __kmp_str_split(tail, '-', &flag_val_str, &tail);
6291       file_name = tail;
6292       if (tail != NULL) {
6293         long *flag_addr = 0;
6294         long flag_val = 0;
6295         KMP_SSCANF(flag_addr_str, "%p", &flag_addr);
6296         KMP_SSCANF(flag_val_str, "%lx", &flag_val);
6297         if (flag_addr != 0 && flag_val != 0 && strcmp(file_name, "") != 0) {
6298           // First, check whether environment-encoded address is mapped into
6299           // addr space.
6300           // If so, dereference it to see if it still has the right value.
6301           if (__kmp_is_address_mapped(flag_addr) && *flag_addr == flag_val) {
6302             neighbor = 1;
6303           } else {
6304             // If not, then we know the other copy of the library is no longer
6305             // running.
6306             neighbor = 2;
6307           }; // if
6308         }; // if
6309       }; // if
6310       switch (neighbor) {
6311       case 0: // Cannot parse environment variable -- neighbor status unknown.
6312         // Assume it is the incompatible format of future version of the
6313         // library. Assume the other library is alive.
6314         // WARN( ... ); // TODO: Issue a warning.
6315         file_name = "unknown library";
6316       // Attention! Falling to the next case. That's intentional.
6317       case 1: { // Neighbor is alive.
6318         // Check it is allowed.
6319         char *duplicate_ok = __kmp_env_get("KMP_DUPLICATE_LIB_OK");
6320         if (!__kmp_str_match_true(duplicate_ok)) {
6321           // That's not allowed. Issue fatal error.
6322           __kmp_fatal(KMP_MSG(DuplicateLibrary, KMP_LIBRARY_FILE, file_name),
6323                       KMP_HNT(DuplicateLibrary), __kmp_msg_null);
6324         }; // if
6325         KMP_INTERNAL_FREE(duplicate_ok);
6326         __kmp_duplicate_library_ok = 1;
6327         done = 1; // Exit the loop.
6328       } break;
6329       case 2: { // Neighbor is dead.
6330         // Clear the variable and try to register library again.
6331         __kmp_env_unset(name);
6332       } break;
6333       default: { KMP_DEBUG_ASSERT(0); } break;
6334       }; // switch
6335 
6336     }; // if
6337     KMP_INTERNAL_FREE((void *)value);
6338 
6339   }; // while
6340   KMP_INTERNAL_FREE((void *)name);
6341 
6342 } // func __kmp_register_library_startup
6343 
6344 void __kmp_unregister_library(void) {
6345 
6346   char *name = __kmp_reg_status_name();
6347   char *value = __kmp_env_get(name);
6348 
6349   KMP_DEBUG_ASSERT(__kmp_registration_flag != 0);
6350   KMP_DEBUG_ASSERT(__kmp_registration_str != NULL);
6351   if (value != NULL && strcmp(value, __kmp_registration_str) == 0) {
6352     // Ok, this is our variable. Delete it.
6353     __kmp_env_unset(name);
6354   }; // if
6355 
6356   KMP_INTERNAL_FREE(__kmp_registration_str);
6357   KMP_INTERNAL_FREE(value);
6358   KMP_INTERNAL_FREE(name);
6359 
6360   __kmp_registration_flag = 0;
6361   __kmp_registration_str = NULL;
6362 
6363 } // __kmp_unregister_library
6364 
6365 // End of Library registration stuff.
6366 // -----------------------------------------------------------------------------
6367 
6368 #if KMP_MIC_SUPPORTED
6369 
6370 static void __kmp_check_mic_type() {
6371   kmp_cpuid_t cpuid_state = {0};
6372   kmp_cpuid_t *cs_p = &cpuid_state;
6373   __kmp_x86_cpuid(1, 0, cs_p);
6374   // We don't support mic1 at the moment
6375   if ((cs_p->eax & 0xff0) == 0xB10) {
6376     __kmp_mic_type = mic2;
6377   } else if ((cs_p->eax & 0xf0ff0) == 0x50670) {
6378     __kmp_mic_type = mic3;
6379   } else {
6380     __kmp_mic_type = non_mic;
6381   }
6382 }
6383 
6384 #endif /* KMP_MIC_SUPPORTED */
6385 
6386 static void __kmp_do_serial_initialize(void) {
6387   int i, gtid;
6388   int size;
6389 
6390   KA_TRACE(10, ("__kmp_do_serial_initialize: enter\n"));
6391 
6392   KMP_DEBUG_ASSERT(sizeof(kmp_int32) == 4);
6393   KMP_DEBUG_ASSERT(sizeof(kmp_uint32) == 4);
6394   KMP_DEBUG_ASSERT(sizeof(kmp_int64) == 8);
6395   KMP_DEBUG_ASSERT(sizeof(kmp_uint64) == 8);
6396   KMP_DEBUG_ASSERT(sizeof(kmp_intptr_t) == sizeof(void *));
6397 
6398 #if OMPT_SUPPORT
6399   ompt_pre_init();
6400 #endif
6401 
6402   __kmp_validate_locks();
6403 
6404   /* Initialize internal memory allocator */
6405   __kmp_init_allocator();
6406 
6407   /* Register the library startup via an environment variable and check to see
6408      whether another copy of the library is already registered. */
6409 
6410   __kmp_register_library_startup();
6411 
6412   /* TODO reinitialization of library */
6413   if (TCR_4(__kmp_global.g.g_done)) {
6414     KA_TRACE(10, ("__kmp_do_serial_initialize: reinitialization of library\n"));
6415   }
6416 
6417   __kmp_global.g.g_abort = 0;
6418   TCW_SYNC_4(__kmp_global.g.g_done, FALSE);
6419 
6420 /* initialize the locks */
6421 #if KMP_USE_ADAPTIVE_LOCKS
6422 #if KMP_DEBUG_ADAPTIVE_LOCKS
6423   __kmp_init_speculative_stats();
6424 #endif
6425 #endif
6426 #if KMP_STATS_ENABLED
6427   __kmp_stats_init();
6428 #endif
6429   __kmp_init_lock(&__kmp_global_lock);
6430   __kmp_init_queuing_lock(&__kmp_dispatch_lock);
6431   __kmp_init_lock(&__kmp_debug_lock);
6432   __kmp_init_atomic_lock(&__kmp_atomic_lock);
6433   __kmp_init_atomic_lock(&__kmp_atomic_lock_1i);
6434   __kmp_init_atomic_lock(&__kmp_atomic_lock_2i);
6435   __kmp_init_atomic_lock(&__kmp_atomic_lock_4i);
6436   __kmp_init_atomic_lock(&__kmp_atomic_lock_4r);
6437   __kmp_init_atomic_lock(&__kmp_atomic_lock_8i);
6438   __kmp_init_atomic_lock(&__kmp_atomic_lock_8r);
6439   __kmp_init_atomic_lock(&__kmp_atomic_lock_8c);
6440   __kmp_init_atomic_lock(&__kmp_atomic_lock_10r);
6441   __kmp_init_atomic_lock(&__kmp_atomic_lock_16r);
6442   __kmp_init_atomic_lock(&__kmp_atomic_lock_16c);
6443   __kmp_init_atomic_lock(&__kmp_atomic_lock_20c);
6444   __kmp_init_atomic_lock(&__kmp_atomic_lock_32c);
6445   __kmp_init_bootstrap_lock(&__kmp_forkjoin_lock);
6446   __kmp_init_bootstrap_lock(&__kmp_exit_lock);
6447 #if KMP_USE_MONITOR
6448   __kmp_init_bootstrap_lock(&__kmp_monitor_lock);
6449 #endif
6450   __kmp_init_bootstrap_lock(&__kmp_tp_cached_lock);
6451 
6452   /* conduct initialization and initial setup of configuration */
6453 
6454   __kmp_runtime_initialize();
6455 
6456 #if KMP_MIC_SUPPORTED
6457   __kmp_check_mic_type();
6458 #endif
6459 
6460 // Some global variable initialization moved here from kmp_env_initialize()
6461 #ifdef KMP_DEBUG
6462   kmp_diag = 0;
6463 #endif
6464   __kmp_abort_delay = 0;
6465 
6466   // From __kmp_init_dflt_team_nth()
6467   /* assume the entire machine will be used */
6468   __kmp_dflt_team_nth_ub = __kmp_xproc;
6469   if (__kmp_dflt_team_nth_ub < KMP_MIN_NTH) {
6470     __kmp_dflt_team_nth_ub = KMP_MIN_NTH;
6471   }
6472   if (__kmp_dflt_team_nth_ub > __kmp_sys_max_nth) {
6473     __kmp_dflt_team_nth_ub = __kmp_sys_max_nth;
6474   }
6475   __kmp_max_nth = __kmp_sys_max_nth;
6476   __kmp_cg_max_nth = __kmp_sys_max_nth;
6477   __kmp_teams_max_nth = __kmp_xproc; // set a "reasonable" default
6478   if (__kmp_teams_max_nth > __kmp_sys_max_nth) {
6479     __kmp_teams_max_nth = __kmp_sys_max_nth;
6480   }
6481 
6482   // Three vars below moved here from __kmp_env_initialize() "KMP_BLOCKTIME"
6483   // part
6484   __kmp_dflt_blocktime = KMP_DEFAULT_BLOCKTIME;
6485 #if KMP_USE_MONITOR
6486   __kmp_monitor_wakeups =
6487       KMP_WAKEUPS_FROM_BLOCKTIME(__kmp_dflt_blocktime, __kmp_monitor_wakeups);
6488   __kmp_bt_intervals =
6489       KMP_INTERVALS_FROM_BLOCKTIME(__kmp_dflt_blocktime, __kmp_monitor_wakeups);
6490 #endif
6491   // From "KMP_LIBRARY" part of __kmp_env_initialize()
6492   __kmp_library = library_throughput;
6493   // From KMP_SCHEDULE initialization
6494   __kmp_static = kmp_sch_static_balanced;
6495 // AC: do not use analytical here, because it is non-monotonous
6496 //__kmp_guided = kmp_sch_guided_iterative_chunked;
6497 //__kmp_auto = kmp_sch_guided_analytical_chunked; // AC: it is the default, no
6498 // need to repeat assignment
6499 // Barrier initialization. Moved here from __kmp_env_initialize() Barrier branch
6500 // bit control and barrier method control parts
6501 #if KMP_FAST_REDUCTION_BARRIER
6502 #define kmp_reduction_barrier_gather_bb ((int)1)
6503 #define kmp_reduction_barrier_release_bb ((int)1)
6504 #define kmp_reduction_barrier_gather_pat bp_hyper_bar
6505 #define kmp_reduction_barrier_release_pat bp_hyper_bar
6506 #endif // KMP_FAST_REDUCTION_BARRIER
6507   for (i = bs_plain_barrier; i < bs_last_barrier; i++) {
6508     __kmp_barrier_gather_branch_bits[i] = __kmp_barrier_gather_bb_dflt;
6509     __kmp_barrier_release_branch_bits[i] = __kmp_barrier_release_bb_dflt;
6510     __kmp_barrier_gather_pattern[i] = __kmp_barrier_gather_pat_dflt;
6511     __kmp_barrier_release_pattern[i] = __kmp_barrier_release_pat_dflt;
6512 #if KMP_FAST_REDUCTION_BARRIER
6513     if (i == bs_reduction_barrier) { // tested and confirmed on ALTIX only (
6514       // lin_64 ): hyper,1
6515       __kmp_barrier_gather_branch_bits[i] = kmp_reduction_barrier_gather_bb;
6516       __kmp_barrier_release_branch_bits[i] = kmp_reduction_barrier_release_bb;
6517       __kmp_barrier_gather_pattern[i] = kmp_reduction_barrier_gather_pat;
6518       __kmp_barrier_release_pattern[i] = kmp_reduction_barrier_release_pat;
6519     }
6520 #endif // KMP_FAST_REDUCTION_BARRIER
6521   }
6522 #if KMP_FAST_REDUCTION_BARRIER
6523 #undef kmp_reduction_barrier_release_pat
6524 #undef kmp_reduction_barrier_gather_pat
6525 #undef kmp_reduction_barrier_release_bb
6526 #undef kmp_reduction_barrier_gather_bb
6527 #endif // KMP_FAST_REDUCTION_BARRIER
6528 #if KMP_MIC_SUPPORTED
6529   if (__kmp_mic_type == mic2) { // KNC
6530     // AC: plane=3,2, forkjoin=2,1 are optimal for 240 threads on KNC
6531     __kmp_barrier_gather_branch_bits[bs_plain_barrier] = 3; // plain gather
6532     __kmp_barrier_release_branch_bits[bs_forkjoin_barrier] =
6533         1; // forkjoin release
6534     __kmp_barrier_gather_pattern[bs_forkjoin_barrier] = bp_hierarchical_bar;
6535     __kmp_barrier_release_pattern[bs_forkjoin_barrier] = bp_hierarchical_bar;
6536   }
6537 #if KMP_FAST_REDUCTION_BARRIER
6538   if (__kmp_mic_type == mic2) { // KNC
6539     __kmp_barrier_gather_pattern[bs_reduction_barrier] = bp_hierarchical_bar;
6540     __kmp_barrier_release_pattern[bs_reduction_barrier] = bp_hierarchical_bar;
6541   }
6542 #endif // KMP_FAST_REDUCTION_BARRIER
6543 #endif // KMP_MIC_SUPPORTED
6544 
6545 // From KMP_CHECKS initialization
6546 #ifdef KMP_DEBUG
6547   __kmp_env_checks = TRUE; /* development versions have the extra checks */
6548 #else
6549   __kmp_env_checks = FALSE; /* port versions do not have the extra checks */
6550 #endif
6551 
6552   // From "KMP_FOREIGN_THREADS_THREADPRIVATE" initialization
6553   __kmp_foreign_tp = TRUE;
6554 
6555   __kmp_global.g.g_dynamic = FALSE;
6556   __kmp_global.g.g_dynamic_mode = dynamic_default;
6557 
6558   __kmp_env_initialize(NULL);
6559 
6560 // Print all messages in message catalog for testing purposes.
6561 #ifdef KMP_DEBUG
6562   char const *val = __kmp_env_get("KMP_DUMP_CATALOG");
6563   if (__kmp_str_match_true(val)) {
6564     kmp_str_buf_t buffer;
6565     __kmp_str_buf_init(&buffer);
6566     __kmp_i18n_dump_catalog(&buffer);
6567     __kmp_printf("%s", buffer.str);
6568     __kmp_str_buf_free(&buffer);
6569   }; // if
6570   __kmp_env_free(&val);
6571 #endif
6572 
6573   __kmp_threads_capacity =
6574       __kmp_initial_threads_capacity(__kmp_dflt_team_nth_ub);
6575   // Moved here from __kmp_env_initialize() "KMP_ALL_THREADPRIVATE" part
6576   __kmp_tp_capacity = __kmp_default_tp_capacity(
6577       __kmp_dflt_team_nth_ub, __kmp_max_nth, __kmp_allThreadsSpecified);
6578 
6579   // If the library is shut down properly, both pools must be NULL. Just in
6580   // case, set them to NULL -- some memory may leak, but subsequent code will
6581   // work even if pools are not freed.
6582   KMP_DEBUG_ASSERT(__kmp_thread_pool == NULL);
6583   KMP_DEBUG_ASSERT(__kmp_thread_pool_insert_pt == NULL);
6584   KMP_DEBUG_ASSERT(__kmp_team_pool == NULL);
6585   __kmp_thread_pool = NULL;
6586   __kmp_thread_pool_insert_pt = NULL;
6587   __kmp_team_pool = NULL;
6588 
6589   /* Allocate all of the variable sized records */
6590   /* NOTE: __kmp_threads_capacity entries are allocated, but the arrays are
6591    * expandable */
6592   /* Since allocation is cache-aligned, just add extra padding at the end */
6593   size =
6594       (sizeof(kmp_info_t *) + sizeof(kmp_root_t *)) * __kmp_threads_capacity +
6595       CACHE_LINE;
6596   __kmp_threads = (kmp_info_t **)__kmp_allocate(size);
6597   __kmp_root = (kmp_root_t **)((char *)__kmp_threads +
6598                                sizeof(kmp_info_t *) * __kmp_threads_capacity);
6599 
6600   /* init thread counts */
6601   KMP_DEBUG_ASSERT(__kmp_all_nth ==
6602                    0); // Asserts fail if the library is reinitializing and
6603   KMP_DEBUG_ASSERT(__kmp_nth == 0); // something was wrong in termination.
6604   __kmp_all_nth = 0;
6605   __kmp_nth = 0;
6606 
6607   /* setup the uber master thread and hierarchy */
6608   gtid = __kmp_register_root(TRUE);
6609   KA_TRACE(10, ("__kmp_do_serial_initialize  T#%d\n", gtid));
6610   KMP_ASSERT(KMP_UBER_GTID(gtid));
6611   KMP_ASSERT(KMP_INITIAL_GTID(gtid));
6612 
6613   KMP_MB(); /* Flush all pending memory write invalidates.  */
6614 
6615   __kmp_common_initialize();
6616 
6617 #if KMP_OS_UNIX
6618   /* invoke the child fork handler */
6619   __kmp_register_atfork();
6620 #endif
6621 
6622 #if !defined KMP_DYNAMIC_LIB
6623   {
6624     /* Invoke the exit handler when the program finishes, only for static
6625        library. For dynamic library, we already have _fini and DllMain. */
6626     int rc = atexit(__kmp_internal_end_atexit);
6627     if (rc != 0) {
6628       __kmp_fatal(KMP_MSG(FunctionError, "atexit()"), KMP_ERR(rc),
6629                   __kmp_msg_null);
6630     }; // if
6631   }
6632 #endif
6633 
6634 #if KMP_HANDLE_SIGNALS
6635 #if KMP_OS_UNIX
6636   /* NOTE: make sure that this is called before the user installs their own
6637      signal handlers so that the user handlers are called first. this way they
6638      can return false, not call our handler, avoid terminating the library, and
6639      continue execution where they left off. */
6640   __kmp_install_signals(FALSE);
6641 #endif /* KMP_OS_UNIX */
6642 #if KMP_OS_WINDOWS
6643   __kmp_install_signals(TRUE);
6644 #endif /* KMP_OS_WINDOWS */
6645 #endif
6646 
6647   /* we have finished the serial initialization */
6648   __kmp_init_counter++;
6649 
6650   __kmp_init_serial = TRUE;
6651 
6652   if (__kmp_settings) {
6653     __kmp_env_print();
6654   }
6655 
6656 #if OMP_40_ENABLED
6657   if (__kmp_display_env || __kmp_display_env_verbose) {
6658     __kmp_env_print_2();
6659   }
6660 #endif // OMP_40_ENABLED
6661 
6662 #if OMPT_SUPPORT
6663   ompt_post_init();
6664 #endif
6665 
6666   KMP_MB();
6667 
6668   KA_TRACE(10, ("__kmp_do_serial_initialize: exit\n"));
6669 }
6670 
6671 void __kmp_serial_initialize(void) {
6672   if (__kmp_init_serial) {
6673     return;
6674   }
6675   __kmp_acquire_bootstrap_lock(&__kmp_initz_lock);
6676   if (__kmp_init_serial) {
6677     __kmp_release_bootstrap_lock(&__kmp_initz_lock);
6678     return;
6679   }
6680   __kmp_do_serial_initialize();
6681   __kmp_release_bootstrap_lock(&__kmp_initz_lock);
6682 }
6683 
6684 static void __kmp_do_middle_initialize(void) {
6685   int i, j;
6686   int prev_dflt_team_nth;
6687 
6688   if (!__kmp_init_serial) {
6689     __kmp_do_serial_initialize();
6690   }
6691 
6692   KA_TRACE(10, ("__kmp_middle_initialize: enter\n"));
6693 
6694   // Save the previous value for the __kmp_dflt_team_nth so that
6695   // we can avoid some reinitialization if it hasn't changed.
6696   prev_dflt_team_nth = __kmp_dflt_team_nth;
6697 
6698 #if KMP_AFFINITY_SUPPORTED
6699   // __kmp_affinity_initialize() will try to set __kmp_ncores to the
6700   // number of cores on the machine.
6701   __kmp_affinity_initialize();
6702 
6703   // Run through the __kmp_threads array and set the affinity mask
6704   // for each root thread that is currently registered with the RTL.
6705   for (i = 0; i < __kmp_threads_capacity; i++) {
6706     if (TCR_PTR(__kmp_threads[i]) != NULL) {
6707       __kmp_affinity_set_init_mask(i, TRUE);
6708     }
6709   }
6710 #endif /* KMP_AFFINITY_SUPPORTED */
6711 
6712   KMP_ASSERT(__kmp_xproc > 0);
6713   if (__kmp_avail_proc == 0) {
6714     __kmp_avail_proc = __kmp_xproc;
6715   }
6716 
6717   // If there were empty places in num_threads list (OMP_NUM_THREADS=,,2,3),
6718   // correct them now
6719   j = 0;
6720   while ((j < __kmp_nested_nth.used) && !__kmp_nested_nth.nth[j]) {
6721     __kmp_nested_nth.nth[j] = __kmp_dflt_team_nth = __kmp_dflt_team_nth_ub =
6722         __kmp_avail_proc;
6723     j++;
6724   }
6725 
6726   if (__kmp_dflt_team_nth == 0) {
6727 #ifdef KMP_DFLT_NTH_CORES
6728     // Default #threads = #cores
6729     __kmp_dflt_team_nth = __kmp_ncores;
6730     KA_TRACE(20, ("__kmp_middle_initialize: setting __kmp_dflt_team_nth = "
6731                   "__kmp_ncores (%d)\n",
6732                   __kmp_dflt_team_nth));
6733 #else
6734     // Default #threads = #available OS procs
6735     __kmp_dflt_team_nth = __kmp_avail_proc;
6736     KA_TRACE(20, ("__kmp_middle_initialize: setting __kmp_dflt_team_nth = "
6737                   "__kmp_avail_proc(%d)\n",
6738                   __kmp_dflt_team_nth));
6739 #endif /* KMP_DFLT_NTH_CORES */
6740   }
6741 
6742   if (__kmp_dflt_team_nth < KMP_MIN_NTH) {
6743     __kmp_dflt_team_nth = KMP_MIN_NTH;
6744   }
6745   if (__kmp_dflt_team_nth > __kmp_sys_max_nth) {
6746     __kmp_dflt_team_nth = __kmp_sys_max_nth;
6747   }
6748 
6749   // There's no harm in continuing if the following check fails,
6750   // but it indicates an error in the previous logic.
6751   KMP_DEBUG_ASSERT(__kmp_dflt_team_nth <= __kmp_dflt_team_nth_ub);
6752 
6753   if (__kmp_dflt_team_nth != prev_dflt_team_nth) {
6754     // Run through the __kmp_threads array and set the num threads icv for each
6755     // root thread that is currently registered with the RTL (which has not
6756     // already explicitly set its nthreads-var with a call to
6757     // omp_set_num_threads()).
6758     for (i = 0; i < __kmp_threads_capacity; i++) {
6759       kmp_info_t *thread = __kmp_threads[i];
6760       if (thread == NULL)
6761         continue;
6762       if (thread->th.th_current_task->td_icvs.nproc != 0)
6763         continue;
6764 
6765       set__nproc(__kmp_threads[i], __kmp_dflt_team_nth);
6766     }
6767   }
6768   KA_TRACE(
6769       20,
6770       ("__kmp_middle_initialize: final value for __kmp_dflt_team_nth = %d\n",
6771        __kmp_dflt_team_nth));
6772 
6773 #ifdef KMP_ADJUST_BLOCKTIME
6774   /* Adjust blocktime to zero if necessary  now that __kmp_avail_proc is set */
6775   if (!__kmp_env_blocktime && (__kmp_avail_proc > 0)) {
6776     KMP_DEBUG_ASSERT(__kmp_avail_proc > 0);
6777     if (__kmp_nth > __kmp_avail_proc) {
6778       __kmp_zero_bt = TRUE;
6779     }
6780   }
6781 #endif /* KMP_ADJUST_BLOCKTIME */
6782 
6783   /* we have finished middle initialization */
6784   TCW_SYNC_4(__kmp_init_middle, TRUE);
6785 
6786   KA_TRACE(10, ("__kmp_do_middle_initialize: exit\n"));
6787 }
6788 
6789 void __kmp_middle_initialize(void) {
6790   if (__kmp_init_middle) {
6791     return;
6792   }
6793   __kmp_acquire_bootstrap_lock(&__kmp_initz_lock);
6794   if (__kmp_init_middle) {
6795     __kmp_release_bootstrap_lock(&__kmp_initz_lock);
6796     return;
6797   }
6798   __kmp_do_middle_initialize();
6799   __kmp_release_bootstrap_lock(&__kmp_initz_lock);
6800 }
6801 
6802 void __kmp_parallel_initialize(void) {
6803   int gtid = __kmp_entry_gtid(); // this might be a new root
6804 
6805   /* synchronize parallel initialization (for sibling) */
6806   if (TCR_4(__kmp_init_parallel))
6807     return;
6808   __kmp_acquire_bootstrap_lock(&__kmp_initz_lock);
6809   if (TCR_4(__kmp_init_parallel)) {
6810     __kmp_release_bootstrap_lock(&__kmp_initz_lock);
6811     return;
6812   }
6813 
6814   /* TODO reinitialization after we have already shut down */
6815   if (TCR_4(__kmp_global.g.g_done)) {
6816     KA_TRACE(
6817         10,
6818         ("__kmp_parallel_initialize: attempt to init while shutting down\n"));
6819     __kmp_infinite_loop();
6820   }
6821 
6822   /* jc: The lock __kmp_initz_lock is already held, so calling
6823      __kmp_serial_initialize would cause a deadlock.  So we call
6824      __kmp_do_serial_initialize directly. */
6825   if (!__kmp_init_middle) {
6826     __kmp_do_middle_initialize();
6827   }
6828 
6829   /* begin initialization */
6830   KA_TRACE(10, ("__kmp_parallel_initialize: enter\n"));
6831   KMP_ASSERT(KMP_UBER_GTID(gtid));
6832 
6833 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
6834   // Save the FP control regs.
6835   // Worker threads will set theirs to these values at thread startup.
6836   __kmp_store_x87_fpu_control_word(&__kmp_init_x87_fpu_control_word);
6837   __kmp_store_mxcsr(&__kmp_init_mxcsr);
6838   __kmp_init_mxcsr &= KMP_X86_MXCSR_MASK;
6839 #endif /* KMP_ARCH_X86 || KMP_ARCH_X86_64 */
6840 
6841 #if KMP_OS_UNIX
6842 #if KMP_HANDLE_SIGNALS
6843   /*  must be after __kmp_serial_initialize  */
6844   __kmp_install_signals(TRUE);
6845 #endif
6846 #endif
6847 
6848   __kmp_suspend_initialize();
6849 
6850 #if defined(USE_LOAD_BALANCE)
6851   if (__kmp_global.g.g_dynamic_mode == dynamic_default) {
6852     __kmp_global.g.g_dynamic_mode = dynamic_load_balance;
6853   }
6854 #else
6855   if (__kmp_global.g.g_dynamic_mode == dynamic_default) {
6856     __kmp_global.g.g_dynamic_mode = dynamic_thread_limit;
6857   }
6858 #endif
6859 
6860   if (__kmp_version) {
6861     __kmp_print_version_2();
6862   }
6863 
6864   /* we have finished parallel initialization */
6865   TCW_SYNC_4(__kmp_init_parallel, TRUE);
6866 
6867   KMP_MB();
6868   KA_TRACE(10, ("__kmp_parallel_initialize: exit\n"));
6869 
6870   __kmp_release_bootstrap_lock(&__kmp_initz_lock);
6871 }
6872 
6873 /* ------------------------------------------------------------------------ */
6874 
6875 void __kmp_run_before_invoked_task(int gtid, int tid, kmp_info_t *this_thr,
6876                                    kmp_team_t *team) {
6877   kmp_disp_t *dispatch;
6878 
6879   KMP_MB();
6880 
6881   /* none of the threads have encountered any constructs, yet. */
6882   this_thr->th.th_local.this_construct = 0;
6883 #if KMP_CACHE_MANAGE
6884   KMP_CACHE_PREFETCH(&this_thr->th.th_bar[bs_forkjoin_barrier].bb.b_arrived);
6885 #endif /* KMP_CACHE_MANAGE */
6886   dispatch = (kmp_disp_t *)TCR_PTR(this_thr->th.th_dispatch);
6887   KMP_DEBUG_ASSERT(dispatch);
6888   KMP_DEBUG_ASSERT(team->t.t_dispatch);
6889   // KMP_DEBUG_ASSERT( this_thr->th.th_dispatch == &team->t.t_dispatch[
6890   // this_thr->th.th_info.ds.ds_tid ] );
6891 
6892   dispatch->th_disp_index = 0; /* reset the dispatch buffer counter */
6893 #if OMP_45_ENABLED
6894   dispatch->th_doacross_buf_idx =
6895       0; /* reset the doacross dispatch buffer counter */
6896 #endif
6897   if (__kmp_env_consistency_check)
6898     __kmp_push_parallel(gtid, team->t.t_ident);
6899 
6900   KMP_MB(); /* Flush all pending memory write invalidates.  */
6901 }
6902 
6903 void __kmp_run_after_invoked_task(int gtid, int tid, kmp_info_t *this_thr,
6904                                   kmp_team_t *team) {
6905   if (__kmp_env_consistency_check)
6906     __kmp_pop_parallel(gtid, team->t.t_ident);
6907 
6908   __kmp_finish_implicit_task(this_thr);
6909 }
6910 
6911 int __kmp_invoke_task_func(int gtid) {
6912   int rc;
6913   int tid = __kmp_tid_from_gtid(gtid);
6914   kmp_info_t *this_thr = __kmp_threads[gtid];
6915   kmp_team_t *team = this_thr->th.th_team;
6916 
6917   __kmp_run_before_invoked_task(gtid, tid, this_thr, team);
6918 #if USE_ITT_BUILD
6919   if (__itt_stack_caller_create_ptr) {
6920     __kmp_itt_stack_callee_enter(
6921         (__itt_caller)
6922             team->t.t_stack_id); // inform ittnotify about entering user's code
6923   }
6924 #endif /* USE_ITT_BUILD */
6925 #if INCLUDE_SSC_MARKS
6926   SSC_MARK_INVOKING();
6927 #endif
6928 
6929 #if OMPT_SUPPORT
6930   void *dummy;
6931   void **exit_runtime_p;
6932   ompt_task_id_t my_task_id;
6933   ompt_parallel_id_t my_parallel_id;
6934 
6935   if (ompt_enabled) {
6936     exit_runtime_p = &(team->t.t_implicit_task_taskdata[tid]
6937                            .ompt_task_info.frame.exit_runtime_frame);
6938   } else {
6939     exit_runtime_p = &dummy;
6940   }
6941 
6942 #if OMPT_TRACE
6943   my_task_id = team->t.t_implicit_task_taskdata[tid].ompt_task_info.task_id;
6944   my_parallel_id = team->t.ompt_team_info.parallel_id;
6945   if (ompt_enabled &&
6946       ompt_callbacks.ompt_callback(ompt_event_implicit_task_begin)) {
6947     ompt_callbacks.ompt_callback(ompt_event_implicit_task_begin)(my_parallel_id,
6948                                                                  my_task_id);
6949   }
6950 #endif
6951 #endif
6952 
6953   {
6954     KMP_TIME_PARTITIONED_BLOCK(OMP_parallel);
6955     KMP_SET_THREAD_STATE_BLOCK(IMPLICIT_TASK);
6956     rc =
6957         __kmp_invoke_microtask((microtask_t)TCR_SYNC_PTR(team->t.t_pkfn), gtid,
6958                                tid, (int)team->t.t_argc, (void **)team->t.t_argv
6959 #if OMPT_SUPPORT
6960                                ,
6961                                exit_runtime_p
6962 #endif
6963                                );
6964 #if OMPT_SUPPORT
6965     *exit_runtime_p = NULL;
6966 #endif
6967   }
6968 
6969 #if USE_ITT_BUILD
6970   if (__itt_stack_caller_create_ptr) {
6971     __kmp_itt_stack_callee_leave(
6972         (__itt_caller)
6973             team->t.t_stack_id); // inform ittnotify about leaving user's code
6974   }
6975 #endif /* USE_ITT_BUILD */
6976   __kmp_run_after_invoked_task(gtid, tid, this_thr, team);
6977 
6978   return rc;
6979 }
6980 
6981 #if OMP_40_ENABLED
6982 void __kmp_teams_master(int gtid) {
6983   // This routine is called by all master threads in teams construct
6984   kmp_info_t *thr = __kmp_threads[gtid];
6985   kmp_team_t *team = thr->th.th_team;
6986   ident_t *loc = team->t.t_ident;
6987   thr->th.th_set_nproc = thr->th.th_teams_size.nth;
6988   KMP_DEBUG_ASSERT(thr->th.th_teams_microtask);
6989   KMP_DEBUG_ASSERT(thr->th.th_set_nproc);
6990   KA_TRACE(20, ("__kmp_teams_master: T#%d, Tid %d, microtask %p\n", gtid,
6991                 __kmp_tid_from_gtid(gtid), thr->th.th_teams_microtask));
6992 // Launch league of teams now, but not let workers execute
6993 // (they hang on fork barrier until next parallel)
6994 #if INCLUDE_SSC_MARKS
6995   SSC_MARK_FORKING();
6996 #endif
6997   __kmp_fork_call(loc, gtid, fork_context_intel, team->t.t_argc,
6998 #if OMPT_SUPPORT
6999                   (void *)thr->th.th_teams_microtask, // "unwrapped" task
7000 #endif
7001                   (microtask_t)thr->th.th_teams_microtask, // "wrapped" task
7002                   VOLATILE_CAST(launch_t) __kmp_invoke_task_func, NULL);
7003 #if INCLUDE_SSC_MARKS
7004   SSC_MARK_JOINING();
7005 #endif
7006 
7007   // AC: last parameter "1" eliminates join barrier which won't work because
7008   // worker threads are in a fork barrier waiting for more parallel regions
7009   __kmp_join_call(loc, gtid
7010 #if OMPT_SUPPORT
7011                   ,
7012                   fork_context_intel
7013 #endif
7014                   ,
7015                   1);
7016 }
7017 
7018 int __kmp_invoke_teams_master(int gtid) {
7019   kmp_info_t *this_thr = __kmp_threads[gtid];
7020   kmp_team_t *team = this_thr->th.th_team;
7021 #if KMP_DEBUG
7022   if (!__kmp_threads[gtid]->th.th_team->t.t_serialized)
7023     KMP_DEBUG_ASSERT((void *)__kmp_threads[gtid]->th.th_team->t.t_pkfn ==
7024                      (void *)__kmp_teams_master);
7025 #endif
7026   __kmp_run_before_invoked_task(gtid, 0, this_thr, team);
7027   __kmp_teams_master(gtid);
7028   __kmp_run_after_invoked_task(gtid, 0, this_thr, team);
7029   return 1;
7030 }
7031 #endif /* OMP_40_ENABLED */
7032 
7033 /* this sets the requested number of threads for the next parallel region
7034    encountered by this team. since this should be enclosed in the forkjoin
7035    critical section it should avoid race conditions with assymmetrical nested
7036    parallelism */
7037 
7038 void __kmp_push_num_threads(ident_t *id, int gtid, int num_threads) {
7039   kmp_info_t *thr = __kmp_threads[gtid];
7040 
7041   if (num_threads > 0)
7042     thr->th.th_set_nproc = num_threads;
7043 }
7044 
7045 #if OMP_40_ENABLED
7046 
7047 /* this sets the requested number of teams for the teams region and/or
7048    the number of threads for the next parallel region encountered  */
7049 void __kmp_push_num_teams(ident_t *id, int gtid, int num_teams,
7050                           int num_threads) {
7051   kmp_info_t *thr = __kmp_threads[gtid];
7052   KMP_DEBUG_ASSERT(num_teams >= 0);
7053   KMP_DEBUG_ASSERT(num_threads >= 0);
7054 
7055   if (num_teams == 0)
7056     num_teams = 1; // default number of teams is 1.
7057   if (num_teams > __kmp_teams_max_nth) { // if too many teams requested?
7058     if (!__kmp_reserve_warn) {
7059       __kmp_reserve_warn = 1;
7060       __kmp_msg(kmp_ms_warning,
7061                 KMP_MSG(CantFormThrTeam, num_teams, __kmp_teams_max_nth),
7062                 KMP_HNT(Unset_ALL_THREADS), __kmp_msg_null);
7063     }
7064     num_teams = __kmp_teams_max_nth;
7065   }
7066   // Set number of teams (number of threads in the outer "parallel" of the
7067   // teams)
7068   thr->th.th_set_nproc = thr->th.th_teams_size.nteams = num_teams;
7069 
7070   // Remember the number of threads for inner parallel regions
7071   if (num_threads == 0) {
7072     if (!TCR_4(__kmp_init_middle))
7073       __kmp_middle_initialize(); // get __kmp_avail_proc calculated
7074     num_threads = __kmp_avail_proc / num_teams;
7075     if (num_teams * num_threads > __kmp_teams_max_nth) {
7076       // adjust num_threads w/o warning as it is not user setting
7077       num_threads = __kmp_teams_max_nth / num_teams;
7078     }
7079   } else {
7080     if (num_teams * num_threads > __kmp_teams_max_nth) {
7081       int new_threads = __kmp_teams_max_nth / num_teams;
7082       if (!__kmp_reserve_warn) { // user asked for too many threads
7083         __kmp_reserve_warn = 1; // that conflicts with KMP_TEAMS_THREAD_LIMIT
7084         __kmp_msg(kmp_ms_warning,
7085                   KMP_MSG(CantFormThrTeam, num_threads, new_threads),
7086                   KMP_HNT(Unset_ALL_THREADS), __kmp_msg_null);
7087       }
7088       num_threads = new_threads;
7089     }
7090   }
7091   thr->th.th_teams_size.nth = num_threads;
7092 }
7093 
7094 // Set the proc_bind var to use in the following parallel region.
7095 void __kmp_push_proc_bind(ident_t *id, int gtid, kmp_proc_bind_t proc_bind) {
7096   kmp_info_t *thr = __kmp_threads[gtid];
7097   thr->th.th_set_proc_bind = proc_bind;
7098 }
7099 
7100 #endif /* OMP_40_ENABLED */
7101 
7102 /* Launch the worker threads into the microtask. */
7103 
7104 void __kmp_internal_fork(ident_t *id, int gtid, kmp_team_t *team) {
7105   kmp_info_t *this_thr = __kmp_threads[gtid];
7106 
7107 #ifdef KMP_DEBUG
7108   int f;
7109 #endif /* KMP_DEBUG */
7110 
7111   KMP_DEBUG_ASSERT(team);
7112   KMP_DEBUG_ASSERT(this_thr->th.th_team == team);
7113   KMP_ASSERT(KMP_MASTER_GTID(gtid));
7114   KMP_MB(); /* Flush all pending memory write invalidates.  */
7115 
7116   team->t.t_construct = 0; /* no single directives seen yet */
7117   team->t.t_ordered.dt.t_value =
7118       0; /* thread 0 enters the ordered section first */
7119 
7120   /* Reset the identifiers on the dispatch buffer */
7121   KMP_DEBUG_ASSERT(team->t.t_disp_buffer);
7122   if (team->t.t_max_nproc > 1) {
7123     int i;
7124     for (i = 0; i < __kmp_dispatch_num_buffers; ++i) {
7125       team->t.t_disp_buffer[i].buffer_index = i;
7126 #if OMP_45_ENABLED
7127       team->t.t_disp_buffer[i].doacross_buf_idx = i;
7128 #endif
7129     }
7130   } else {
7131     team->t.t_disp_buffer[0].buffer_index = 0;
7132 #if OMP_45_ENABLED
7133     team->t.t_disp_buffer[0].doacross_buf_idx = 0;
7134 #endif
7135   }
7136 
7137   KMP_MB(); /* Flush all pending memory write invalidates.  */
7138   KMP_ASSERT(this_thr->th.th_team == team);
7139 
7140 #ifdef KMP_DEBUG
7141   for (f = 0; f < team->t.t_nproc; f++) {
7142     KMP_DEBUG_ASSERT(team->t.t_threads[f] &&
7143                      team->t.t_threads[f]->th.th_team_nproc == team->t.t_nproc);
7144   }
7145 #endif /* KMP_DEBUG */
7146 
7147   /* release the worker threads so they may begin working */
7148   __kmp_fork_barrier(gtid, 0);
7149 }
7150 
7151 void __kmp_internal_join(ident_t *id, int gtid, kmp_team_t *team) {
7152   kmp_info_t *this_thr = __kmp_threads[gtid];
7153 
7154   KMP_DEBUG_ASSERT(team);
7155   KMP_DEBUG_ASSERT(this_thr->th.th_team == team);
7156   KMP_ASSERT(KMP_MASTER_GTID(gtid));
7157   KMP_MB(); /* Flush all pending memory write invalidates.  */
7158 
7159 /* Join barrier after fork */
7160 
7161 #ifdef KMP_DEBUG
7162   if (__kmp_threads[gtid] &&
7163       __kmp_threads[gtid]->th.th_team_nproc != team->t.t_nproc) {
7164     __kmp_printf("GTID: %d, __kmp_threads[%d]=%p\n", gtid, gtid,
7165                  __kmp_threads[gtid]);
7166     __kmp_printf("__kmp_threads[%d]->th.th_team_nproc=%d, TEAM: %p, "
7167                  "team->t.t_nproc=%d\n",
7168                  gtid, __kmp_threads[gtid]->th.th_team_nproc, team,
7169                  team->t.t_nproc);
7170     __kmp_print_structure();
7171   }
7172   KMP_DEBUG_ASSERT(__kmp_threads[gtid] &&
7173                    __kmp_threads[gtid]->th.th_team_nproc == team->t.t_nproc);
7174 #endif /* KMP_DEBUG */
7175 
7176   __kmp_join_barrier(gtid); /* wait for everyone */
7177 
7178   KMP_MB(); /* Flush all pending memory write invalidates.  */
7179   KMP_ASSERT(this_thr->th.th_team == team);
7180 }
7181 
7182 /* ------------------------------------------------------------------------ */
7183 
7184 #ifdef USE_LOAD_BALANCE
7185 
7186 // Return the worker threads actively spinning in the hot team, if we
7187 // are at the outermost level of parallelism.  Otherwise, return 0.
7188 static int __kmp_active_hot_team_nproc(kmp_root_t *root) {
7189   int i;
7190   int retval;
7191   kmp_team_t *hot_team;
7192 
7193   if (root->r.r_active) {
7194     return 0;
7195   }
7196   hot_team = root->r.r_hot_team;
7197   if (__kmp_dflt_blocktime == KMP_MAX_BLOCKTIME) {
7198     return hot_team->t.t_nproc - 1; // Don't count master thread
7199   }
7200 
7201   // Skip the master thread - it is accounted for elsewhere.
7202   retval = 0;
7203   for (i = 1; i < hot_team->t.t_nproc; i++) {
7204     if (hot_team->t.t_threads[i]->th.th_active) {
7205       retval++;
7206     }
7207   }
7208   return retval;
7209 }
7210 
7211 // Perform an automatic adjustment to the number of
7212 // threads used by the next parallel region.
7213 static int __kmp_load_balance_nproc(kmp_root_t *root, int set_nproc) {
7214   int retval;
7215   int pool_active;
7216   int hot_team_active;
7217   int team_curr_active;
7218   int system_active;
7219 
7220   KB_TRACE(20, ("__kmp_load_balance_nproc: called root:%p set_nproc:%d\n", root,
7221                 set_nproc));
7222   KMP_DEBUG_ASSERT(root);
7223   KMP_DEBUG_ASSERT(root->r.r_root_team->t.t_threads[0]
7224                        ->th.th_current_task->td_icvs.dynamic == TRUE);
7225   KMP_DEBUG_ASSERT(set_nproc > 1);
7226 
7227   if (set_nproc == 1) {
7228     KB_TRACE(20, ("__kmp_load_balance_nproc: serial execution.\n"));
7229     return 1;
7230   }
7231 
7232   // Threads that are active in the thread pool, active in the hot team for this
7233   // particular root (if we are at the outer par level), and the currently
7234   // executing thread (to become the master) are available to add to the new
7235   // team, but are currently contributing to the system load, and must be
7236   // accounted for.
7237   pool_active = TCR_4(__kmp_thread_pool_active_nth);
7238   hot_team_active = __kmp_active_hot_team_nproc(root);
7239   team_curr_active = pool_active + hot_team_active + 1;
7240 
7241   // Check the system load.
7242   system_active = __kmp_get_load_balance(__kmp_avail_proc + team_curr_active);
7243   KB_TRACE(30, ("__kmp_load_balance_nproc: system active = %d pool active = %d "
7244                 "hot team active = %d\n",
7245                 system_active, pool_active, hot_team_active));
7246 
7247   if (system_active < 0) {
7248     // There was an error reading the necessary info from /proc, so use the
7249     // thread limit algorithm instead. Once we set __kmp_global.g.g_dynamic_mode
7250     // = dynamic_thread_limit, we shouldn't wind up getting back here.
7251     __kmp_global.g.g_dynamic_mode = dynamic_thread_limit;
7252     KMP_WARNING(CantLoadBalUsing, "KMP_DYNAMIC_MODE=thread limit");
7253 
7254     // Make this call behave like the thread limit algorithm.
7255     retval = __kmp_avail_proc - __kmp_nth +
7256              (root->r.r_active ? 1 : root->r.r_hot_team->t.t_nproc);
7257     if (retval > set_nproc) {
7258       retval = set_nproc;
7259     }
7260     if (retval < KMP_MIN_NTH) {
7261       retval = KMP_MIN_NTH;
7262     }
7263 
7264     KB_TRACE(20, ("__kmp_load_balance_nproc: thread limit exit. retval:%d\n",
7265                   retval));
7266     return retval;
7267   }
7268 
7269   // There is a slight delay in the load balance algorithm in detecting new
7270   // running procs. The real system load at this instant should be at least as
7271   // large as the #active omp thread that are available to add to the team.
7272   if (system_active < team_curr_active) {
7273     system_active = team_curr_active;
7274   }
7275   retval = __kmp_avail_proc - system_active + team_curr_active;
7276   if (retval > set_nproc) {
7277     retval = set_nproc;
7278   }
7279   if (retval < KMP_MIN_NTH) {
7280     retval = KMP_MIN_NTH;
7281   }
7282 
7283   KB_TRACE(20, ("__kmp_load_balance_nproc: exit. retval:%d\n", retval));
7284   return retval;
7285 } // __kmp_load_balance_nproc()
7286 
7287 #endif /* USE_LOAD_BALANCE */
7288 
7289 /* ------------------------------------------------------------------------ */
7290 
7291 /* NOTE: this is called with the __kmp_init_lock held */
7292 void __kmp_cleanup(void) {
7293   int f;
7294 
7295   KA_TRACE(10, ("__kmp_cleanup: enter\n"));
7296 
7297   if (TCR_4(__kmp_init_parallel)) {
7298 #if KMP_HANDLE_SIGNALS
7299     __kmp_remove_signals();
7300 #endif
7301     TCW_4(__kmp_init_parallel, FALSE);
7302   }
7303 
7304   if (TCR_4(__kmp_init_middle)) {
7305 #if KMP_AFFINITY_SUPPORTED
7306     __kmp_affinity_uninitialize();
7307 #endif /* KMP_AFFINITY_SUPPORTED */
7308     __kmp_cleanup_hierarchy();
7309     TCW_4(__kmp_init_middle, FALSE);
7310   }
7311 
7312   KA_TRACE(10, ("__kmp_cleanup: go serial cleanup\n"));
7313 
7314   if (__kmp_init_serial) {
7315     __kmp_runtime_destroy();
7316     __kmp_init_serial = FALSE;
7317   }
7318 
7319   for (f = 0; f < __kmp_threads_capacity; f++) {
7320     if (__kmp_root[f] != NULL) {
7321       __kmp_free(__kmp_root[f]);
7322       __kmp_root[f] = NULL;
7323     }
7324   }
7325   __kmp_free(__kmp_threads);
7326   // __kmp_threads and __kmp_root were allocated at once, as single block, so
7327   // there is no need in freeing __kmp_root.
7328   __kmp_threads = NULL;
7329   __kmp_root = NULL;
7330   __kmp_threads_capacity = 0;
7331 
7332 #if KMP_USE_DYNAMIC_LOCK
7333   __kmp_cleanup_indirect_user_locks();
7334 #else
7335   __kmp_cleanup_user_locks();
7336 #endif
7337 
7338 #if KMP_AFFINITY_SUPPORTED
7339   KMP_INTERNAL_FREE(CCAST(char *, __kmp_cpuinfo_file));
7340   __kmp_cpuinfo_file = NULL;
7341 #endif /* KMP_AFFINITY_SUPPORTED */
7342 
7343 #if KMP_USE_ADAPTIVE_LOCKS
7344 #if KMP_DEBUG_ADAPTIVE_LOCKS
7345   __kmp_print_speculative_stats();
7346 #endif
7347 #endif
7348   KMP_INTERNAL_FREE(__kmp_nested_nth.nth);
7349   __kmp_nested_nth.nth = NULL;
7350   __kmp_nested_nth.size = 0;
7351   __kmp_nested_nth.used = 0;
7352   KMP_INTERNAL_FREE(__kmp_nested_proc_bind.bind_types);
7353   __kmp_nested_proc_bind.bind_types = NULL;
7354   __kmp_nested_proc_bind.size = 0;
7355   __kmp_nested_proc_bind.used = 0;
7356 
7357   __kmp_i18n_catclose();
7358 
7359 #if KMP_STATS_ENABLED
7360   __kmp_stats_fini();
7361 #endif
7362 
7363   KA_TRACE(10, ("__kmp_cleanup: exit\n"));
7364 }
7365 
7366 /* ------------------------------------------------------------------------ */
7367 
7368 int __kmp_ignore_mppbeg(void) {
7369   char *env;
7370 
7371   if ((env = getenv("KMP_IGNORE_MPPBEG")) != NULL) {
7372     if (__kmp_str_match_false(env))
7373       return FALSE;
7374   }
7375   // By default __kmpc_begin() is no-op.
7376   return TRUE;
7377 }
7378 
7379 int __kmp_ignore_mppend(void) {
7380   char *env;
7381 
7382   if ((env = getenv("KMP_IGNORE_MPPEND")) != NULL) {
7383     if (__kmp_str_match_false(env))
7384       return FALSE;
7385   }
7386   // By default __kmpc_end() is no-op.
7387   return TRUE;
7388 }
7389 
7390 void __kmp_internal_begin(void) {
7391   int gtid;
7392   kmp_root_t *root;
7393 
7394   /* this is a very important step as it will register new sibling threads
7395      and assign these new uber threads a new gtid */
7396   gtid = __kmp_entry_gtid();
7397   root = __kmp_threads[gtid]->th.th_root;
7398   KMP_ASSERT(KMP_UBER_GTID(gtid));
7399 
7400   if (root->r.r_begin)
7401     return;
7402   __kmp_acquire_lock(&root->r.r_begin_lock, gtid);
7403   if (root->r.r_begin) {
7404     __kmp_release_lock(&root->r.r_begin_lock, gtid);
7405     return;
7406   }
7407 
7408   root->r.r_begin = TRUE;
7409 
7410   __kmp_release_lock(&root->r.r_begin_lock, gtid);
7411 }
7412 
7413 /* ------------------------------------------------------------------------ */
7414 
7415 void __kmp_user_set_library(enum library_type arg) {
7416   int gtid;
7417   kmp_root_t *root;
7418   kmp_info_t *thread;
7419 
7420   /* first, make sure we are initialized so we can get our gtid */
7421 
7422   gtid = __kmp_entry_gtid();
7423   thread = __kmp_threads[gtid];
7424 
7425   root = thread->th.th_root;
7426 
7427   KA_TRACE(20, ("__kmp_user_set_library: enter T#%d, arg: %d, %d\n", gtid, arg,
7428                 library_serial));
7429   if (root->r.r_in_parallel) { /* Must be called in serial section of top-level
7430                                   thread */
7431     KMP_WARNING(SetLibraryIncorrectCall);
7432     return;
7433   }
7434 
7435   switch (arg) {
7436   case library_serial:
7437     thread->th.th_set_nproc = 0;
7438     set__nproc(thread, 1);
7439     break;
7440   case library_turnaround:
7441     thread->th.th_set_nproc = 0;
7442     set__nproc(thread, __kmp_dflt_team_nth ? __kmp_dflt_team_nth
7443                                            : __kmp_dflt_team_nth_ub);
7444     break;
7445   case library_throughput:
7446     thread->th.th_set_nproc = 0;
7447     set__nproc(thread, __kmp_dflt_team_nth ? __kmp_dflt_team_nth
7448                                            : __kmp_dflt_team_nth_ub);
7449     break;
7450   default:
7451     KMP_FATAL(UnknownLibraryType, arg);
7452   }
7453 
7454   __kmp_aux_set_library(arg);
7455 }
7456 
7457 void __kmp_aux_set_stacksize(size_t arg) {
7458   if (!__kmp_init_serial)
7459     __kmp_serial_initialize();
7460 
7461 #if KMP_OS_DARWIN
7462   if (arg & (0x1000 - 1)) {
7463     arg &= ~(0x1000 - 1);
7464     if (arg + 0x1000) /* check for overflow if we round up */
7465       arg += 0x1000;
7466   }
7467 #endif
7468   __kmp_acquire_bootstrap_lock(&__kmp_initz_lock);
7469 
7470   /* only change the default stacksize before the first parallel region */
7471   if (!TCR_4(__kmp_init_parallel)) {
7472     size_t value = arg; /* argument is in bytes */
7473 
7474     if (value < __kmp_sys_min_stksize)
7475       value = __kmp_sys_min_stksize;
7476     else if (value > KMP_MAX_STKSIZE)
7477       value = KMP_MAX_STKSIZE;
7478 
7479     __kmp_stksize = value;
7480 
7481     __kmp_env_stksize = TRUE; /* was KMP_STACKSIZE specified? */
7482   }
7483 
7484   __kmp_release_bootstrap_lock(&__kmp_initz_lock);
7485 }
7486 
7487 /* set the behaviour of the runtime library */
7488 /* TODO this can cause some odd behaviour with sibling parallelism... */
7489 void __kmp_aux_set_library(enum library_type arg) {
7490   __kmp_library = arg;
7491 
7492   switch (__kmp_library) {
7493   case library_serial: {
7494     KMP_INFORM(LibraryIsSerial);
7495     (void)__kmp_change_library(TRUE);
7496   } break;
7497   case library_turnaround:
7498     (void)__kmp_change_library(TRUE);
7499     break;
7500   case library_throughput:
7501     (void)__kmp_change_library(FALSE);
7502     break;
7503   default:
7504     KMP_FATAL(UnknownLibraryType, arg);
7505   }
7506 }
7507 
7508 /* ------------------------------------------------------------------------ */
7509 
7510 void __kmp_aux_set_blocktime(int arg, kmp_info_t *thread, int tid) {
7511   int blocktime = arg; /* argument is in milliseconds */
7512 #if KMP_USE_MONITOR
7513   int bt_intervals;
7514 #endif
7515   int bt_set;
7516 
7517   __kmp_save_internal_controls(thread);
7518 
7519   /* Normalize and set blocktime for the teams */
7520   if (blocktime < KMP_MIN_BLOCKTIME)
7521     blocktime = KMP_MIN_BLOCKTIME;
7522   else if (blocktime > KMP_MAX_BLOCKTIME)
7523     blocktime = KMP_MAX_BLOCKTIME;
7524 
7525   set__blocktime_team(thread->th.th_team, tid, blocktime);
7526   set__blocktime_team(thread->th.th_serial_team, 0, blocktime);
7527 
7528 #if KMP_USE_MONITOR
7529   /* Calculate and set blocktime intervals for the teams */
7530   bt_intervals = KMP_INTERVALS_FROM_BLOCKTIME(blocktime, __kmp_monitor_wakeups);
7531 
7532   set__bt_intervals_team(thread->th.th_team, tid, bt_intervals);
7533   set__bt_intervals_team(thread->th.th_serial_team, 0, bt_intervals);
7534 #endif
7535 
7536   /* Set whether blocktime has been set to "TRUE" */
7537   bt_set = TRUE;
7538 
7539   set__bt_set_team(thread->th.th_team, tid, bt_set);
7540   set__bt_set_team(thread->th.th_serial_team, 0, bt_set);
7541 #if KMP_USE_MONITOR
7542   KF_TRACE(10, ("kmp_set_blocktime: T#%d(%d:%d), blocktime=%d, "
7543                 "bt_intervals=%d, monitor_updates=%d\n",
7544                 __kmp_gtid_from_tid(tid, thread->th.th_team),
7545                 thread->th.th_team->t.t_id, tid, blocktime, bt_intervals,
7546                 __kmp_monitor_wakeups));
7547 #else
7548   KF_TRACE(10, ("kmp_set_blocktime: T#%d(%d:%d), blocktime=%d\n",
7549                 __kmp_gtid_from_tid(tid, thread->th.th_team),
7550                 thread->th.th_team->t.t_id, tid, blocktime));
7551 #endif
7552 }
7553 
7554 void __kmp_aux_set_defaults(char const *str, int len) {
7555   if (!__kmp_init_serial) {
7556     __kmp_serial_initialize();
7557   };
7558   __kmp_env_initialize(str);
7559 
7560   if (__kmp_settings
7561 #if OMP_40_ENABLED
7562       || __kmp_display_env || __kmp_display_env_verbose
7563 #endif // OMP_40_ENABLED
7564       ) {
7565     __kmp_env_print();
7566   }
7567 } // __kmp_aux_set_defaults
7568 
7569 /* ------------------------------------------------------------------------ */
7570 /* internal fast reduction routines */
7571 
7572 PACKED_REDUCTION_METHOD_T
7573 __kmp_determine_reduction_method(
7574     ident_t *loc, kmp_int32 global_tid, kmp_int32 num_vars, size_t reduce_size,
7575     void *reduce_data, void (*reduce_func)(void *lhs_data, void *rhs_data),
7576     kmp_critical_name *lck) {
7577 
7578   // Default reduction method: critical construct ( lck != NULL, like in current
7579   // PAROPT )
7580   // If ( reduce_data!=NULL && reduce_func!=NULL ): the tree-reduction method
7581   // can be selected by RTL
7582   // If loc->flags contains KMP_IDENT_ATOMIC_REDUCE, the atomic reduce method
7583   // can be selected by RTL
7584   // Finally, it's up to OpenMP RTL to make a decision on which method to select
7585   // among generated by PAROPT.
7586 
7587   PACKED_REDUCTION_METHOD_T retval;
7588 
7589   int team_size;
7590 
7591   KMP_DEBUG_ASSERT(loc); // it would be nice to test ( loc != 0 )
7592   KMP_DEBUG_ASSERT(lck); // it would be nice to test ( lck != 0 )
7593 
7594 #define FAST_REDUCTION_ATOMIC_METHOD_GENERATED                                 \
7595   ((loc->flags & (KMP_IDENT_ATOMIC_REDUCE)) == (KMP_IDENT_ATOMIC_REDUCE))
7596 #define FAST_REDUCTION_TREE_METHOD_GENERATED ((reduce_data) && (reduce_func))
7597 
7598   retval = critical_reduce_block;
7599 
7600   // another choice of getting a team size (with 1 dynamic deference) is slower
7601   team_size = __kmp_get_team_num_threads(global_tid);
7602   if (team_size == 1) {
7603 
7604     retval = empty_reduce_block;
7605 
7606   } else {
7607 
7608     int atomic_available = FAST_REDUCTION_ATOMIC_METHOD_GENERATED;
7609     int tree_available = FAST_REDUCTION_TREE_METHOD_GENERATED;
7610 
7611 #if KMP_ARCH_X86_64 || KMP_ARCH_PPC64 || KMP_ARCH_AARCH64 || KMP_ARCH_MIPS64
7612 
7613 #if KMP_OS_LINUX || KMP_OS_FREEBSD || KMP_OS_NETBSD || KMP_OS_WINDOWS ||       \
7614     KMP_OS_DARWIN
7615 
7616     int teamsize_cutoff = 4;
7617 
7618 #if KMP_MIC_SUPPORTED
7619     if (__kmp_mic_type != non_mic) {
7620       teamsize_cutoff = 8;
7621     }
7622 #endif
7623     if (tree_available) {
7624       if (team_size <= teamsize_cutoff) {
7625         if (atomic_available) {
7626           retval = atomic_reduce_block;
7627         }
7628       } else {
7629         retval = TREE_REDUCE_BLOCK_WITH_REDUCTION_BARRIER;
7630       }
7631     } else if (atomic_available) {
7632       retval = atomic_reduce_block;
7633     }
7634 #else
7635 #error "Unknown or unsupported OS"
7636 #endif // KMP_OS_LINUX || KMP_OS_FREEBSD || KMP_OS_NETBSD || KMP_OS_WINDOWS ||
7637 // KMP_OS_DARWIN
7638 
7639 #elif KMP_ARCH_X86 || KMP_ARCH_ARM || KMP_ARCH_AARCH || KMP_ARCH_MIPS
7640 
7641 #if KMP_OS_LINUX || KMP_OS_WINDOWS
7642 
7643     // basic tuning
7644 
7645     if (atomic_available) {
7646       if (num_vars <= 2) { // && ( team_size <= 8 ) due to false-sharing ???
7647         retval = atomic_reduce_block;
7648       }
7649     } // otherwise: use critical section
7650 
7651 #elif KMP_OS_DARWIN
7652 
7653     if (atomic_available && (num_vars <= 3)) {
7654       retval = atomic_reduce_block;
7655     } else if (tree_available) {
7656       if ((reduce_size > (9 * sizeof(kmp_real64))) &&
7657           (reduce_size < (2000 * sizeof(kmp_real64)))) {
7658         retval = TREE_REDUCE_BLOCK_WITH_PLAIN_BARRIER;
7659       }
7660     } // otherwise: use critical section
7661 
7662 #else
7663 #error "Unknown or unsupported OS"
7664 #endif
7665 
7666 #else
7667 #error "Unknown or unsupported architecture"
7668 #endif
7669   }
7670 
7671   // KMP_FORCE_REDUCTION
7672 
7673   // If the team is serialized (team_size == 1), ignore the forced reduction
7674   // method and stay with the unsynchronized method (empty_reduce_block)
7675   if (__kmp_force_reduction_method != reduction_method_not_defined &&
7676       team_size != 1) {
7677 
7678     PACKED_REDUCTION_METHOD_T forced_retval = critical_reduce_block;
7679 
7680     int atomic_available, tree_available;
7681 
7682     switch ((forced_retval = __kmp_force_reduction_method)) {
7683     case critical_reduce_block:
7684       KMP_ASSERT(lck); // lck should be != 0
7685       break;
7686 
7687     case atomic_reduce_block:
7688       atomic_available = FAST_REDUCTION_ATOMIC_METHOD_GENERATED;
7689       if (!atomic_available) {
7690         KMP_WARNING(RedMethodNotSupported, "atomic");
7691         forced_retval = critical_reduce_block;
7692       }
7693       break;
7694 
7695     case tree_reduce_block:
7696       tree_available = FAST_REDUCTION_TREE_METHOD_GENERATED;
7697       if (!tree_available) {
7698         KMP_WARNING(RedMethodNotSupported, "tree");
7699         forced_retval = critical_reduce_block;
7700       } else {
7701 #if KMP_FAST_REDUCTION_BARRIER
7702         forced_retval = TREE_REDUCE_BLOCK_WITH_REDUCTION_BARRIER;
7703 #endif
7704       }
7705       break;
7706 
7707     default:
7708       KMP_ASSERT(0); // "unsupported method specified"
7709     }
7710 
7711     retval = forced_retval;
7712   }
7713 
7714   KA_TRACE(10, ("reduction method selected=%08x\n", retval));
7715 
7716 #undef FAST_REDUCTION_TREE_METHOD_GENERATED
7717 #undef FAST_REDUCTION_ATOMIC_METHOD_GENERATED
7718 
7719   return (retval);
7720 }
7721 
7722 // this function is for testing set/get/determine reduce method
7723 kmp_int32 __kmp_get_reduce_method(void) {
7724   return ((__kmp_entry_thread()->th.th_local.packed_reduction_method) >> 8);
7725 }
7726