libgomp/team.c

   1 /* Copyright (C) 2005-2017 Free Software Foundation, Inc.
   2    Contributed by Richard Henderson <rth@redhat.com>.
   3
   4    This file is part of the GNU Offloading and Multi Processing Library
   5    (libgomp).
   6
   7    Libgomp is free software; you can redistribute it and/or modify it
   8    under the terms of the GNU General Public License as published by
   9    the Free Software Foundation; either version 3, or (at your option)
  10    any later version.
  11
  12    Libgomp is distributed in the hope that it will be useful, but WITHOUT ANY
  13    WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS
  14    FOR A PARTICULAR PURPOSE.  See the GNU General Public License for
  15    more details.
  16
  17    Under Section 7 of GPL version 3, you are granted additional
  18    permissions described in the GCC Runtime Library Exception, version
  19    3.1, as published by the Free Software Foundation.
  20
  21    You should have received a copy of the GNU General Public License and
  22    a copy of the GCC Runtime Library Exception along with this program;
  23    see the files COPYING3 and COPYING.RUNTIME respectively.  If not, see
  24    <http://www.gnu.org/licenses/>.  */
  25
  26 /* This file handles the maintainence of threads in response to team
  27    creation and termination.  */
  28
  29 #include "libgomp.h"
  30 #include "pool.h"
  31 #include <stdlib.h>
  32 #include <string.h>
  33
  34 #ifdef LIBGOMP_USE_PTHREADS
  35 /* This attribute contains PTHREAD_CREATE_DETACHED.  */
  36 pthread_attr_t gomp_thread_attr;
  37
  38 /* This key is for the thread destructor.  */
  39 pthread_key_t gomp_thread_destructor;
  40
  41
  42 /* This is the libgomp per-thread data structure.  */
  43 #if defined HAVE_TLS || defined USE_EMUTLS
  44 __thread struct gomp_thread gomp_tls_data;
  45 #else
  46 pthread_key_t gomp_tls_key;
  47 #endif
  48
  49
  50 /* This structure is used to communicate across pthread_create.  */
  51
  52 struct gomp_thread_start_data
  53 {
  54   void (*fn) (void *);
  55   void *fn_data;
  56   struct gomp_team_state ts;
  57   struct gomp_task *task;
  58   struct gomp_thread_pool *thread_pool;
  59   unsigned int place;
  60   bool nested;
  61 };
  62
  63
  64 /* This function is a pthread_create entry point.  This contains the idle
  65    loop in which a thread waits to be called up to become part of a team.  */
  66
  67 static void *
  68 gomp_thread_start (void *xdata)
  69 {
  70   struct gomp_thread_start_data *data = xdata;
  71   struct gomp_thread *thr;
  72   struct gomp_thread_pool *pool;
  73   void (*local_fn) (void *);
  74   void *local_data;
  75
  76 #if defined HAVE_TLS || defined USE_EMUTLS
  77   thr = &gomp_tls_data;
  78 #else
  79   struct gomp_thread local_thr;
  80   thr = &local_thr;
  81   pthread_setspecific (gomp_tls_key, thr);
  82 #endif
  83   gomp_sem_init (&thr->release, 0);
  84
  85   /* Extract what we need from data.  */
  86   local_fn = data->fn;
  87   local_data = data->fn_data;
  88   thr->thread_pool = data->thread_pool;
  89   thr->ts = data->ts;
  90   thr->task = data->task;
  91   thr->place = data->place;
  92
  93   thr->ts.team->ordered_release[thr->ts.team_id] = &thr->release;
  94
  95   /* Make thread pool local. */
  96   pool = thr->thread_pool;
  97
  98   if (data->nested)
  99     {
 100       struct gomp_team *team = thr->ts.team;
 101       struct gomp_task *task = thr->task;
 102
 103       gomp_barrier_wait (&team->barrier);
 104
 105       local_fn (local_data);
 106       gomp_team_barrier_wait_final (&team->barrier);
 107       gomp_finish_task (task);
 108       gomp_barrier_wait_last (&team->barrier);
 109     }
 110   else
 111     {
 112       pool->threads[thr->ts.team_id] = thr;
 113
 114       gomp_simple_barrier_wait (&pool->threads_dock);
 115       do
 116         {
 117           struct gomp_team *team = thr->ts.team;
 118           struct gomp_task *task = thr->task;
 119
 120           local_fn (local_data);
 121           gomp_team_barrier_wait_final (&team->barrier);
 122           gomp_finish_task (task);
 123
 124           gomp_simple_barrier_wait (&pool->threads_dock);
 125
 126           local_fn = thr->fn;
 127           local_data = thr->data;
 128           thr->fn = NULL;
 129         }
 130       while (local_fn);
 131     }
 132
 133   gomp_sem_destroy (&thr->release);
 134   thr->thread_pool = NULL;
 135   thr->task = NULL;
 136   return NULL;
 137 }
 138 #endif
 139
 140 static inline struct gomp_team *
 141 get_last_team (unsigned nthreads)
 142 {
 143   struct gomp_thread *thr = gomp_thread ();
 144   if (thr->ts.team == NULL)
 145     {
 146       struct gomp_thread_pool *pool = gomp_get_thread_pool (thr, nthreads);
 147       struct gomp_team *last_team = pool->last_team;
 148       if (last_team != NULL && last_team->nthreads == nthreads)
 149         {
 150           pool->last_team = NULL;
 151           return last_team;
 152         }
 153     }
 154   return NULL;
 155 }
 156
 157 /* Create a new team data structure.  */
 158
 159 struct gomp_team *
 160 gomp_new_team (unsigned nthreads)
 161 {
 162   struct gomp_team *team;
 163   int i;
 164
 165   team = get_last_team (nthreads);
 166   if (team == NULL)
 167     {
 168       size_t extra = sizeof (team->ordered_release[0])
 169                      + sizeof (team->implicit_task[0]);
 170       team = gomp_malloc (sizeof (*team) + nthreads * extra);
 171
 172 #ifndef HAVE_SYNC_BUILTINS
 173       gomp_mutex_init (&team->work_share_list_free_lock);
 174 #endif
 175       gomp_barrier_init (&team->barrier, nthreads);
 176       gomp_mutex_init (&team->task_lock);
 177
 178       team->nthreads = nthreads;
 179     }
 180
 181   team->work_share_chunk = 8;
 182 #ifdef HAVE_SYNC_BUILTINS
 183   team->single_count = 0;
 184 #endif
 185   team->work_shares_to_free = &team->work_shares[0];
 186   gomp_init_work_share (&team->work_shares[0], false, nthreads);
 187   team->work_shares[0].next_alloc = NULL;
 188   team->work_share_list_free = NULL;
 189   team->work_share_list_alloc = &team->work_shares[1];
 190   for (i = 1; i < 7; i++)
 191     team->work_shares[i].next_free = &team->work_shares[i + 1];
 192   team->work_shares[i].next_free = NULL;
 193
 194   gomp_sem_init (&team->master_release, 0);
 195   team->ordered_release = (void *) &team->implicit_task[nthreads];
 196   team->ordered_release[0] = &team->master_release;
 197
 198   priority_queue_init (&team->task_queue);
 199   team->task_count = 0;
 200   team->task_queued_count = 0;
 201   team->task_running_count = 0;
 202   team->work_share_cancelled = 0;
 203   team->team_cancelled = 0;
 204
 205   return team;
 206 }
 207
 208
 209 /* Free a team data structure.  */
 210
 211 static void
 212 free_team (struct gomp_team *team)
 213 {
 214 #ifndef HAVE_SYNC_BUILTINS
 215   gomp_mutex_destroy (&team->work_share_list_free_lock);
 216 #endif
 217   gomp_barrier_destroy (&team->barrier);
 218   gomp_mutex_destroy (&team->task_lock);
 219   priority_queue_free (&team->task_queue);
 220   free (team);
 221 }
 222
 223 static void
 224 gomp_free_pool_helper (void *thread_pool)
 225 {
 226   struct gomp_thread *thr = gomp_thread ();
 227   struct gomp_thread_pool *pool
 228     = (struct gomp_thread_pool *) thread_pool;
 229   gomp_simple_barrier_wait_last (&pool->threads_dock);
 230   gomp_sem_destroy (&thr->release);
 231   thr->thread_pool = NULL;
 232   thr->task = NULL;
 233 #ifdef LIBGOMP_USE_PTHREADS
 234   pthread_exit (NULL);
 235 #elif defined(__nvptx__)
 236   asm ("exit;");
 237 #else
 238 #error gomp_free_pool_helper must terminate the thread
 239 #endif
 240 }
 241
 242 /* Free a thread pool and release its threads. */
 243
 244 void
 245 gomp_free_thread (void *arg __attribute__((unused)))
 246 {
 247   struct gomp_thread *thr = gomp_thread ();
 248   struct gomp_thread_pool *pool = thr->thread_pool;
 249   if (pool)
 250     {
 251       if (pool->threads_used > 0)
 252         {
 253           int i;
 254           for (i = 1; i < pool->threads_used; i++)
 255             {
 256               struct gomp_thread *nthr = pool->threads[i];
 257               nthr->fn = gomp_free_pool_helper;
 258               nthr->data = pool;
 259             }
 260           /* This barrier undocks threads docked on pool->threads_dock.  */
 261           gomp_simple_barrier_wait (&pool->threads_dock);
 262           /* And this waits till all threads have called gomp_barrier_wait_last
 263              in gomp_free_pool_helper.  */
 264           gomp_simple_barrier_wait (&pool->threads_dock);
 265           /* Now it is safe to destroy the barrier and free the pool.  */
 266           gomp_simple_barrier_destroy (&pool->threads_dock);
 267
 268 #ifdef HAVE_SYNC_BUILTINS
 269           __sync_fetch_and_add (&gomp_managed_threads,
 270                                 1L - pool->threads_used);
 271 #else
 272           gomp_mutex_lock (&gomp_managed_threads_lock);
 273           gomp_managed_threads -= pool->threads_used - 1L;
 274           gomp_mutex_unlock (&gomp_managed_threads_lock);
 275 #endif
 276         }
 277       if (pool->last_team)
 278         free_team (pool->last_team);
 279 #ifndef __nvptx__
 280       free (pool->threads);
 281       free (pool);
 282 #endif
 283       thr->thread_pool = NULL;
 284     }
 285   if (thr->ts.level == 0 && __builtin_expect (thr->ts.team != NULL, 0))
 286     gomp_team_end ();
 287   if (thr->task != NULL)
 288     {
 289       struct gomp_task *task = thr->task;
 290       gomp_end_task ();
 291       free (task);
 292     }
 293 }
 294
 295 /* Launch a team.  */
 296
 297 #ifdef LIBGOMP_USE_PTHREADS
 298 void
 299 gomp_team_start (void (*fn) (void *), void *data, unsigned nthreads,
 300                  unsigned flags, struct gomp_team *team)
 301 {
 302   struct gomp_thread_start_data *start_data;
 303   struct gomp_thread *thr, *nthr;
 304   struct gomp_task *task;
 305   struct gomp_task_icv *icv;
 306   bool nested;
 307   struct gomp_thread_pool *pool;
 308   unsigned i, n, old_threads_used = 0;
 309   pthread_attr_t thread_attr, *attr;
 310   unsigned long nthreads_var;
 311   char bind, bind_var;
 312   unsigned int s = 0, rest = 0, p = 0, k = 0;
 313   unsigned int affinity_count = 0;
 314   struct gomp_thread **affinity_thr = NULL;
 315
 316   thr = gomp_thread ();
 317   nested = thr->ts.level;
 318   pool = thr->thread_pool;
 319   task = thr->task;
 320   icv = task ? &task->icv : &gomp_global_icv;
 321   if (__builtin_expect (gomp_places_list != NULL, 0) && thr->place == 0)
 322     gomp_init_affinity ();
 323
 324   /* Always save the previous state, even if this isn't a nested team.
 325      In particular, we should save any work share state from an outer
 326      orphaned work share construct.  */
 327   team->prev_ts = thr->ts;
 328
 329   thr->ts.team = team;
 330   thr->ts.team_id = 0;
 331   ++thr->ts.level;
 332   if (nthreads > 1)
 333     ++thr->ts.active_level;
 334   thr->ts.work_share = &team->work_shares[0];
 335   thr->ts.last_work_share = NULL;
 336 #ifdef HAVE_SYNC_BUILTINS
 337   thr->ts.single_count = 0;
 338 #endif
 339   thr->ts.static_trip = 0;
 340   thr->task = &team->implicit_task[0];
 341   nthreads_var = icv->nthreads_var;
 342   if (__builtin_expect (gomp_nthreads_var_list != NULL, 0)
 343       && thr->ts.level < gomp_nthreads_var_list_len)
 344     nthreads_var = gomp_nthreads_var_list[thr->ts.level];
 345   bind_var = icv->bind_var;
 346   if (bind_var != omp_proc_bind_false && (flags & 7) != omp_proc_bind_false)
 347     bind_var = flags & 7;
 348   bind = bind_var;
 349   if (__builtin_expect (gomp_bind_var_list != NULL, 0)
 350       && thr->ts.level < gomp_bind_var_list_len)
 351     bind_var = gomp_bind_var_list[thr->ts.level];
 352   gomp_init_task (thr->task, task, icv);
 353   team->implicit_task[0].icv.nthreads_var = nthreads_var;
 354   team->implicit_task[0].icv.bind_var = bind_var;
 355
 356   if (nthreads == 1)
 357     return;
 358
 359   i = 1;
 360
 361   if (__builtin_expect (gomp_places_list != NULL, 0))
 362     {
 363       /* Depending on chosen proc_bind model, set subpartition
 364          for the master thread and initialize helper variables
 365          P and optionally S, K and/or REST used by later place
 366          computation for each additional thread.  */
 367       p = thr->place - 1;
 368       switch (bind)
 369         {
 370         case omp_proc_bind_true:
 371         case omp_proc_bind_close:
 372           if (nthreads > thr->ts.place_partition_len)
 373             {
 374               /* T > P.  S threads will be placed in each place,
 375                  and the final REM threads placed one by one
 376                  into the already occupied places.  */
 377               s = nthreads / thr->ts.place_partition_len;
 378               rest = nthreads % thr->ts.place_partition_len;
 379             }
 380           else
 381             s = 1;
 382           k = 1;
 383           break;
 384         case omp_proc_bind_master:
 385           /* Each thread will be bound to master's place.  */
 386           break;
 387         case omp_proc_bind_spread:
 388           if (nthreads <= thr->ts.place_partition_len)
 389             {
 390               /* T <= P.  Each subpartition will have in between s
 391                  and s+1 places (subpartitions starting at or
 392                  after rest will have s places, earlier s+1 places),
 393                  each thread will be bound to the first place in
 394                  its subpartition (except for the master thread
 395                  that can be bound to another place in its
 396                  subpartition).  */
 397               s = thr->ts.place_partition_len / nthreads;
 398               rest = thr->ts.place_partition_len % nthreads;
 399               rest = (s + 1) * rest + thr->ts.place_partition_off;
 400               if (p < rest)
 401                 {
 402                   p -= (p - thr->ts.place_partition_off) % (s + 1);
 403                   thr->ts.place_partition_len = s + 1;
 404                 }
 405               else
 406                 {
 407                   p -= (p - rest) % s;
 408                   thr->ts.place_partition_len = s;
 409                 }
 410               thr->ts.place_partition_off = p;
 411             }
 412           else
 413             {
 414               /* T > P.  Each subpartition will have just a single
 415                  place and we'll place between s and s+1
 416                  threads into each subpartition.  */
 417               s = nthreads / thr->ts.place_partition_len;
 418               rest = nthreads % thr->ts.place_partition_len;
 419               thr->ts.place_partition_off = p;
 420               thr->ts.place_partition_len = 1;
 421               k = 1;
 422             }
 423           break;
 424         }
 425     }
 426   else
 427     bind = omp_proc_bind_false;
 428
 429   /* We only allow the reuse of idle threads for non-nested PARALLEL
 430      regions.  This appears to be implied by the semantics of
 431      threadprivate variables, but perhaps that's reading too much into
 432      things.  Certainly it does prevent any locking problems, since
 433      only the initial program thread will modify gomp_threads.  */
 434   if (!nested)
 435     {
 436       old_threads_used = pool->threads_used;
 437
 438       if (nthreads <= old_threads_used)
 439         n = nthreads;
 440       else if (old_threads_used == 0)
 441         {
 442           n = 0;
 443           gomp_simple_barrier_init (&pool->threads_dock, nthreads);
 444         }
 445       else
 446         {
 447           n = old_threads_used;
 448
 449           /* Increase the barrier threshold to make sure all new
 450              threads arrive before the team is released.  */
 451           gomp_simple_barrier_reinit (&pool->threads_dock, nthreads);
 452         }
 453
 454       /* Not true yet, but soon will be.  We're going to release all
 455          threads from the dock, and those that aren't part of the
 456          team will exit.  */
 457       pool->threads_used = nthreads;
 458
 459       /* If necessary, expand the size of the gomp_threads array.  It is
 460          expected that changes in the number of threads are rare, thus we
 461          make no effort to expand gomp_threads_size geometrically.  */
 462       if (nthreads >= pool->threads_size)
 463         {
 464           pool->threads_size = nthreads + 1;
 465           pool->threads
 466             = gomp_realloc (pool->threads,
 467                             pool->threads_size
 468                             * sizeof (struct gomp_thread_data *));
 469         }
 470
 471       /* Release existing idle threads.  */
 472       for (; i < n; ++i)
 473         {
 474           unsigned int place_partition_off = thr->ts.place_partition_off;
 475           unsigned int place_partition_len = thr->ts.place_partition_len;
 476           unsigned int place = 0;
 477           if (__builtin_expect (gomp_places_list != NULL, 0))
 478             {
 479               switch (bind)
 480                 {
 481                 case omp_proc_bind_true:
 482                 case omp_proc_bind_close:
 483                   if (k == s)
 484                     {
 485                       ++p;
 486                       if (p == (team->prev_ts.place_partition_off
 487                                 + team->prev_ts.place_partition_len))
 488                         p = team->prev_ts.place_partition_off;
 489                       k = 1;
 490                       if (i == nthreads - rest)
 491                         s = 1;
 492                     }
 493                   else
 494                     ++k;
 495                   break;
 496                 case omp_proc_bind_master:
 497                   break;
 498                 case omp_proc_bind_spread:
 499                   if (k == 0)
 500                     {
 501                       /* T <= P.  */
 502                       if (p < rest)
 503                         p += s + 1;
 504                       else
 505                         p += s;
 506                       if (p == (team->prev_ts.place_partition_off
 507                                 + team->prev_ts.place_partition_len))
 508                         p = team->prev_ts.place_partition_off;
 509                       place_partition_off = p;
 510                       if (p < rest)
 511                         place_partition_len = s + 1;
 512                       else
 513                         place_partition_len = s;
 514                     }
 515                   else
 516                     {
 517                       /* T > P.  */
 518                       if (k == s)
 519                         {
 520                           ++p;
 521                           if (p == (team->prev_ts.place_partition_off
 522                                     + team->prev_ts.place_partition_len))
 523                             p = team->prev_ts.place_partition_off;
 524                           k = 1;
 525                           if (i == nthreads - rest)
 526                             s = 1;
 527                         }
 528                       else
 529                         ++k;
 530                       place_partition_off = p;
 531                       place_partition_len = 1;
 532                     }
 533                   break;
 534                 }
 535               if (affinity_thr != NULL
 536                   || (bind != omp_proc_bind_true
 537                       && pool->threads[i]->place != p + 1)
 538                   || pool->threads[i]->place <= place_partition_off
 539                   || pool->threads[i]->place > (place_partition_off
 540                                                 + place_partition_len))
 541                 {
 542                   unsigned int l;
 543                   if (affinity_thr == NULL)
 544                     {
 545                       unsigned int j;
 546
 547                       if (team->prev_ts.place_partition_len > 64)
 548                         affinity_thr
 549                           = gomp_malloc (team->prev_ts.place_partition_len
 550                                          * sizeof (struct gomp_thread *));
 551                       else
 552                         affinity_thr
 553                           = gomp_alloca (team->prev_ts.place_partition_len
 554                                          * sizeof (struct gomp_thread *));
 555                       memset (affinity_thr, '\0',
 556                               team->prev_ts.place_partition_len
 557                               * sizeof (struct gomp_thread *));
 558                       for (j = i; j < old_threads_used; j++)
 559                         {
 560                           if (pool->threads[j]->place
 561                               > team->prev_ts.place_partition_off
 562                               && (pool->threads[j]->place
 563                                   <= (team->prev_ts.place_partition_off
 564                                       + team->prev_ts.place_partition_len)))
 565                             {
 566                               l = pool->threads[j]->place - 1
 567                                   - team->prev_ts.place_partition_off;
 568                               pool->threads[j]->data = affinity_thr[l];
 569                               affinity_thr[l] = pool->threads[j];
 570                             }
 571                           pool->threads[j] = NULL;
 572                         }
 573                       if (nthreads > old_threads_used)
 574                         memset (&pool->threads[old_threads_used],
 575                                 '\0', ((nthreads - old_threads_used)
 576                                        * sizeof (struct gomp_thread *)));
 577                       n = nthreads;
 578                       affinity_count = old_threads_used - i;
 579                     }
 580                   if (affinity_count == 0)
 581                     break;
 582                   l = p;
 583                   if (affinity_thr[l - team->prev_ts.place_partition_off]
 584                       == NULL)
 585                     {
 586                       if (bind != omp_proc_bind_true)
 587                         continue;
 588                       for (l = place_partition_off;
 589                            l < place_partition_off + place_partition_len;
 590                            l++)
 591                         if (affinity_thr[l - team->prev_ts.place_partition_off]
 592                             != NULL)
 593                           break;
 594                       if (l == place_partition_off + place_partition_len)
 595                         continue;
 596                     }
 597                   nthr = affinity_thr[l - team->prev_ts.place_partition_off];
 598                   affinity_thr[l - team->prev_ts.place_partition_off]
 599                     = (struct gomp_thread *) nthr->data;
 600                   affinity_count--;
 601                   pool->threads[i] = nthr;
 602                 }
 603               else
 604                 nthr = pool->threads[i];
 605               place = p + 1;
 606             }
 607           else
 608             nthr = pool->threads[i];
 609           nthr->ts.team = team;
 610           nthr->ts.work_share = &team->work_shares[0];
 611           nthr->ts.last_work_share = NULL;
 612           nthr->ts.team_id = i;
 613           nthr->ts.level = team->prev_ts.level + 1;
 614           nthr->ts.active_level = thr->ts.active_level;
 615           nthr->ts.place_partition_off = place_partition_off;
 616           nthr->ts.place_partition_len = place_partition_len;
 617 #ifdef HAVE_SYNC_BUILTINS
 618           nthr->ts.single_count = 0;
 619 #endif
 620           nthr->ts.static_trip = 0;
 621           nthr->task = &team->implicit_task[i];
 622           nthr->place = place;
 623           gomp_init_task (nthr->task, task, icv);
 624           team->implicit_task[i].icv.nthreads_var = nthreads_var;
 625           team->implicit_task[i].icv.bind_var = bind_var;
 626           nthr->fn = fn;
 627           nthr->data = data;
 628           team->ordered_release[i] = &nthr->release;
 629         }
 630
 631       if (__builtin_expect (affinity_thr != NULL, 0))
 632         {
 633           /* If AFFINITY_THR is non-NULL just because we had to
 634              permute some threads in the pool, but we've managed
 635              to find exactly as many old threads as we'd find
 636              without affinity, we don't need to handle this
 637              specially anymore.  */
 638           if (nthreads <= old_threads_used
 639               ? (affinity_count == old_threads_used - nthreads)
 640               : (i == old_threads_used))
 641             {
 642               if (team->prev_ts.place_partition_len > 64)
 643                 free (affinity_thr);
 644               affinity_thr = NULL;
 645               affinity_count = 0;
 646             }
 647           else
 648             {
 649               i = 1;
 650               /* We are going to compute the places/subpartitions
 651                  again from the beginning.  So, we need to reinitialize
 652                  vars modified by the switch (bind) above inside
 653                  of the loop, to the state they had after the initial
 654                  switch (bind).  */
 655               switch (bind)
 656                 {
 657                 case omp_proc_bind_true:
 658                 case omp_proc_bind_close:
 659                   if (nthreads > thr->ts.place_partition_len)
 660                     /* T > P.  S has been changed, so needs
 661                        to be recomputed.  */
 662                     s = nthreads / thr->ts.place_partition_len;
 663                   k = 1;
 664                   p = thr->place - 1;
 665                   break;
 666                 case omp_proc_bind_master:
 667                   /* No vars have been changed.  */
 668                   break;
 669                 case omp_proc_bind_spread:
 670                   p = thr->ts.place_partition_off;
 671                   if (k != 0)
 672                     {
 673                       /* T > P.  */
 674                       s = nthreads / team->prev_ts.place_partition_len;
 675                       k = 1;
 676                     }
 677                   break;
 678                 }
 679
 680               /* Increase the barrier threshold to make sure all new
 681                  threads and all the threads we're going to let die
 682                  arrive before the team is released.  */
 683               if (affinity_count)
 684                 gomp_simple_barrier_reinit (&pool->threads_dock,
 685                                             nthreads + affinity_count);
 686             }
 687         }
 688
 689       if (i == nthreads)
 690         goto do_release;
 691
 692     }
 693
 694   if (__builtin_expect (nthreads + affinity_count > old_threads_used, 0))
 695     {
 696       long diff = (long) (nthreads + affinity_count) - (long) old_threads_used;
 697
 698       if (old_threads_used == 0)
 699         --diff;
 700
 701 #ifdef HAVE_SYNC_BUILTINS
 702       __sync_fetch_and_add (&gomp_managed_threads, diff);
 703 #else
 704       gomp_mutex_lock (&gomp_managed_threads_lock);
 705       gomp_managed_threads += diff;
 706       gomp_mutex_unlock (&gomp_managed_threads_lock);
 707 #endif
 708     }
 709
 710   attr = &gomp_thread_attr;
 711   if (__builtin_expect (gomp_places_list != NULL, 0))
 712     {
 713       size_t stacksize;
 714       pthread_attr_init (&thread_attr);
 715       pthread_attr_setdetachstate (&thread_attr, PTHREAD_CREATE_DETACHED);
 716       if (! pthread_attr_getstacksize (&gomp_thread_attr, &stacksize))
 717         pthread_attr_setstacksize (&thread_attr, stacksize);
 718       attr = &thread_attr;
 719     }
 720
 721   start_data = gomp_alloca (sizeof (struct gomp_thread_start_data)
 722                             * (nthreads-i));
 723
 724   /* Launch new threads.  */
 725   for (; i < nthreads; ++i)
 726     {
 727       pthread_t pt;
 728       int err;
 729
 730       start_data->ts.place_partition_off = thr->ts.place_partition_off;
 731       start_data->ts.place_partition_len = thr->ts.place_partition_len;
 732       start_data->place = 0;
 733       if (__builtin_expect (gomp_places_list != NULL, 0))
 734         {
 735           switch (bind)
 736             {
 737             case omp_proc_bind_true:
 738             case omp_proc_bind_close:
 739               if (k == s)
 740                 {
 741                   ++p;
 742                   if (p == (team->prev_ts.place_partition_off
 743                             + team->prev_ts.place_partition_len))
 744                     p = team->prev_ts.place_partition_off;
 745                   k = 1;
 746                   if (i == nthreads - rest)
 747                     s = 1;
 748                 }
 749               else
 750                 ++k;
 751               break;
 752             case omp_proc_bind_master:
 753               break;
 754             case omp_proc_bind_spread:
 755               if (k == 0)
 756                 {
 757                   /* T <= P.  */
 758                   if (p < rest)
 759                     p += s + 1;
 760                   else
 761                     p += s;
 762                   if (p == (team->prev_ts.place_partition_off
 763                             + team->prev_ts.place_partition_len))
 764                     p = team->prev_ts.place_partition_off;
 765                   start_data->ts.place_partition_off = p;
 766                   if (p < rest)
 767                     start_data->ts.place_partition_len = s + 1;
 768                   else
 769                     start_data->ts.place_partition_len = s;
 770                 }
 771               else
 772                 {
 773                   /* T > P.  */
 774                   if (k == s)
 775                     {
 776                       ++p;
 777                       if (p == (team->prev_ts.place_partition_off
 778                                 + team->prev_ts.place_partition_len))
 779                         p = team->prev_ts.place_partition_off;
 780                       k = 1;
 781                       if (i == nthreads - rest)
 782                         s = 1;
 783                     }
 784                   else
 785                     ++k;
 786                   start_data->ts.place_partition_off = p;
 787                   start_data->ts.place_partition_len = 1;
 788                 }
 789               break;
 790             }
 791           start_data->place = p + 1;
 792           if (affinity_thr != NULL && pool->threads[i] != NULL)
 793             continue;
 794           gomp_init_thread_affinity (attr, p);
 795         }
 796
 797       start_data->fn = fn;
 798       start_data->fn_data = data;
 799       start_data->ts.team = team;
 800       start_data->ts.work_share = &team->work_shares[0];
 801       start_data->ts.last_work_share = NULL;
 802       start_data->ts.team_id = i;
 803       start_data->ts.level = team->prev_ts.level + 1;
 804       start_data->ts.active_level = thr->ts.active_level;
 805 #ifdef HAVE_SYNC_BUILTINS
 806       start_data->ts.single_count = 0;
 807 #endif
 808       start_data->ts.static_trip = 0;
 809       start_data->task = &team->implicit_task[i];
 810       gomp_init_task (start_data->task, task, icv);
 811       team->implicit_task[i].icv.nthreads_var = nthreads_var;
 812       team->implicit_task[i].icv.bind_var = bind_var;
 813       start_data->thread_pool = pool;
 814       start_data->nested = nested;
 815
 816       attr = gomp_adjust_thread_attr (attr, &thread_attr);
 817       err = pthread_create (&pt, attr, gomp_thread_start, start_data++);
 818       if (err != 0)
 819         gomp_fatal ("Thread creation failed: %s", strerror (err));
 820     }
 821
 822   if (__builtin_expect (attr == &thread_attr, 0))
 823     pthread_attr_destroy (&thread_attr);
 824
 825  do_release:
 826   if (nested)
 827     gomp_barrier_wait (&team->barrier);
 828   else
 829     gomp_simple_barrier_wait (&pool->threads_dock);
 830
 831   /* Decrease the barrier threshold to match the number of threads
 832      that should arrive back at the end of this team.  The extra
 833      threads should be exiting.  Note that we arrange for this test
 834      to never be true for nested teams.  If AFFINITY_COUNT is non-zero,
 835      the barrier as well as gomp_managed_threads was temporarily
 836      set to NTHREADS + AFFINITY_COUNT.  For NTHREADS < OLD_THREADS_COUNT,
 837      AFFINITY_COUNT if non-zero will be always at least
 838      OLD_THREADS_COUNT - NTHREADS.  */
 839   if (__builtin_expect (nthreads < old_threads_used, 0)
 840       || __builtin_expect (affinity_count, 0))
 841     {
 842       long diff = (long) nthreads - (long) old_threads_used;
 843
 844       if (affinity_count)
 845         diff = -affinity_count;
 846
 847       gomp_simple_barrier_reinit (&pool->threads_dock, nthreads);
 848
 849 #ifdef HAVE_SYNC_BUILTINS
 850       __sync_fetch_and_add (&gomp_managed_threads, diff);
 851 #else
 852       gomp_mutex_lock (&gomp_managed_threads_lock);
 853       gomp_managed_threads += diff;
 854       gomp_mutex_unlock (&gomp_managed_threads_lock);
 855 #endif
 856     }
 857   if (__builtin_expect (affinity_thr != NULL, 0)
 858       && team->prev_ts.place_partition_len > 64)
 859     free (affinity_thr);
 860 }
 861 #endif
 862
 863
 864 /* Terminate the current team.  This is only to be called by the master
 865    thread.  We assume that we must wait for the other threads.  */
 866
 867 void
 868 gomp_team_end (void)
 869 {
 870   struct gomp_thread *thr = gomp_thread ();
 871   struct gomp_team *team = thr->ts.team;
 872
 873   /* This barrier handles all pending explicit threads.
 874      As #pragma omp cancel parallel might get awaited count in
 875      team->barrier in a inconsistent state, we need to use a different
 876      counter here.  */
 877   gomp_team_barrier_wait_final (&team->barrier);
 878   if (__builtin_expect (team->team_cancelled, 0))
 879     {
 880       struct gomp_work_share *ws = team->work_shares_to_free;
 881       do
 882         {
 883           struct gomp_work_share *next_ws = gomp_ptrlock_get (&ws->next_ws);
 884           if (next_ws == NULL)
 885             gomp_ptrlock_set (&ws->next_ws, ws);
 886           gomp_fini_work_share (ws);
 887           ws = next_ws;
 888         }
 889       while (ws != NULL);
 890     }
 891   else
 892     gomp_fini_work_share (thr->ts.work_share);
 893
 894   gomp_end_task ();
 895   thr->ts = team->prev_ts;
 896
 897   if (__builtin_expect (thr->ts.team != NULL, 0))
 898     {
 899 #ifdef HAVE_SYNC_BUILTINS
 900       __sync_fetch_and_add (&gomp_managed_threads, 1L - team->nthreads);
 901 #else
 902       gomp_mutex_lock (&gomp_managed_threads_lock);
 903       gomp_managed_threads -= team->nthreads - 1L;
 904       gomp_mutex_unlock (&gomp_managed_threads_lock);
 905 #endif
 906       /* This barrier has gomp_barrier_wait_last counterparts
 907          and ensures the team can be safely destroyed.  */
 908       gomp_barrier_wait (&team->barrier);
 909     }
 910
 911   if (__builtin_expect (team->work_shares[0].next_alloc != NULL, 0))
 912     {
 913       struct gomp_work_share *ws = team->work_shares[0].next_alloc;
 914       do
 915         {
 916           struct gomp_work_share *next_ws = ws->next_alloc;
 917           free (ws);
 918           ws = next_ws;
 919         }
 920       while (ws != NULL);
 921     }
 922   gomp_sem_destroy (&team->master_release);
 923
 924   if (__builtin_expect (thr->ts.team != NULL, 0)
 925       || __builtin_expect (team->nthreads == 1, 0))
 926     free_team (team);
 927   else
 928     {
 929       struct gomp_thread_pool *pool = thr->thread_pool;
 930       if (pool->last_team)
 931         free_team (pool->last_team);
 932       pool->last_team = team;
 933       gomp_release_thread_pool (pool);
 934     }
 935 }
 936
 937 #ifdef LIBGOMP_USE_PTHREADS
 938
 939 /* Constructors for this file.  */
 940
 941 static void __attribute__((constructor))
 942 initialize_team (void)
 943 {
 944 #if !defined HAVE_TLS && !defined USE_EMUTLS
 945   static struct gomp_thread initial_thread_tls_data;
 946
 947   pthread_key_create (&gomp_tls_key, NULL);
 948   pthread_setspecific (gomp_tls_key, &initial_thread_tls_data);
 949 #endif
 950
 951   if (pthread_key_create (&gomp_thread_destructor, gomp_free_thread) != 0)
 952     gomp_fatal ("could not create thread pool destructor.");
 953 }
 954
 955 static void __attribute__((destructor))
 956 team_destructor (void)
 957 {
 958   /* Without this dlclose on libgomp could lead to subsequent
 959      crashes.  */
 960   pthread_key_delete (gomp_thread_destructor);
 961 }
 962 #endif
 963
 964 struct gomp_task_icv *
 965 gomp_new_icv (void)
 966 {
 967   struct gomp_thread *thr = gomp_thread ();
 968   struct gomp_task *task = gomp_malloc (sizeof (struct gomp_task));
 969   gomp_init_task (task, NULL, &gomp_global_icv);
 970   thr->task = task;
 971 #ifdef LIBGOMP_USE_PTHREADS
 972   pthread_setspecific (gomp_thread_destructor, thr);
 973 #endif
 974   return &task->icv;
 975 }