libgomp/loop_ull.c

   1 /* Copyright (C) 2005-2014 Free Software Foundation, Inc.
   2    Contributed by Richard Henderson <rth@redhat.com>.
   3
   4    This file is part of the GNU OpenMP Library (libgomp).
   5
   6    Libgomp is free software; you can redistribute it and/or modify it
   7    under the terms of the GNU General Public License as published by
   8    the Free Software Foundation; either version 3, or (at your option)
   9    any later version.
  10
  11    Libgomp is distributed in the hope that it will be useful, but WITHOUT ANY
  12    WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS
  13    FOR A PARTICULAR PURPOSE.  See the GNU General Public License for
  14    more details.
  15
  16    Under Section 7 of GPL version 3, you are granted additional
  17    permissions described in the GCC Runtime Library Exception, version
  18    3.1, as published by the Free Software Foundation.
  19
  20    You should have received a copy of the GNU General Public License and
  21    a copy of the GCC Runtime Library Exception along with this program;
  22    see the files COPYING3 and COPYING.RUNTIME respectively.  If not, see
  23    <http://www.gnu.org/licenses/>.  */
  24
  25 /* This file handles the LOOP (FOR/DO) construct.  */
  26
  27 #include <limits.h>
  28 #include <stdlib.h>
  29 #include "libgomp.h"
  30
  31 typedef unsigned long long gomp_ull;
  32
  33 /* Initialize the given work share construct from the given arguments.  */
  34
  35 static inline void
  36 gomp_loop_ull_init (struct gomp_work_share *ws, bool up, gomp_ull start,
  37                     gomp_ull end, gomp_ull incr, enum gomp_schedule_type sched,
  38                     gomp_ull chunk_size)
  39 {
  40   ws->sched = sched;
  41   ws->chunk_size_ull = chunk_size;
  42   /* Canonicalize loops that have zero iterations to ->next == ->end.  */
  43   ws->end_ull = ((up && start > end) || (!up && start < end))
  44                 ? start : end;
  45   ws->incr_ull = incr;
  46   ws->next_ull = start;
  47   ws->mode = 0;
  48   if (sched == GFS_DYNAMIC)
  49     {
  50       ws->chunk_size_ull *= incr;
  51
  52 #if defined HAVE_SYNC_BUILTINS && defined __LP64__
  53       {
  54         /* For dynamic scheduling prepare things to make each iteration
  55            faster.  */
  56         struct gomp_thread *thr = gomp_thread ();
  57         struct gomp_team *team = thr->ts.team;
  58         long nthreads = team ? team->nthreads : 1;
  59
  60         if (__builtin_expect (up, 1))
  61           {
  62             /* Cheap overflow protection.  */
  63             if (__builtin_expect ((nthreads | ws->chunk_size_ull)
  64                                   < 1ULL << (sizeof (gomp_ull)
  65                                              * __CHAR_BIT__ / 2 - 1), 1))
  66               ws->mode = ws->end_ull < (__LONG_LONG_MAX__ * 2ULL + 1
  67                                         - (nthreads + 1) * ws->chunk_size_ull);
  68           }
  69         /* Cheap overflow protection.  */
  70         else if (__builtin_expect ((nthreads | -ws->chunk_size_ull)
  71                                    < 1ULL << (sizeof (gomp_ull)
  72                                               * __CHAR_BIT__ / 2 - 1), 1))
  73           ws->mode = ws->end_ull > ((nthreads + 1) * -ws->chunk_size_ull
  74                                     - (__LONG_LONG_MAX__ * 2ULL + 1));
  75       }
  76 #endif
  77     }
  78   if (!up)
  79     ws->mode |= 2;
  80 }
  81
  82 /* The *_start routines are called when first encountering a loop construct
  83    that is not bound directly to a parallel construct.  The first thread
  84    that arrives will create the work-share construct; subsequent threads
  85    will see the construct exists and allocate work from it.
  86
  87    START, END, INCR are the bounds of the loop; due to the restrictions of
  88    OpenMP, these values must be the same in every thread.  This is not
  89    verified (nor is it entirely verifiable, since START is not necessarily
  90    retained intact in the work-share data structure).  CHUNK_SIZE is the
  91    scheduling parameter; again this must be identical in all threads.
  92
  93    Returns true if there's any work for this thread to perform.  If so,
  94    *ISTART and *IEND are filled with the bounds of the iteration block
  95    allocated to this thread.  Returns false if all work was assigned to
  96    other threads prior to this thread's arrival.  */
  97
  98 static bool
  99 gomp_loop_ull_static_start (bool up, gomp_ull start, gomp_ull end,
 100                             gomp_ull incr, gomp_ull chunk_size,
 101                             gomp_ull *istart, gomp_ull *iend)
 102 {
 103   struct gomp_thread *thr = gomp_thread ();
 104
 105   thr->ts.static_trip = 0;
 106   if (gomp_work_share_start (false))
 107     {
 108       gomp_loop_ull_init (thr->ts.work_share, up, start, end, incr,
 109                           GFS_STATIC, chunk_size);
 110       gomp_work_share_init_done ();
 111     }
 112
 113   return !gomp_iter_ull_static_next (istart, iend);
 114 }
 115
 116 static bool
 117 gomp_loop_ull_dynamic_start (bool up, gomp_ull start, gomp_ull end,
 118                              gomp_ull incr, gomp_ull chunk_size,
 119                              gomp_ull *istart, gomp_ull *iend)
 120 {
 121   struct gomp_thread *thr = gomp_thread ();
 122   bool ret;
 123
 124   if (gomp_work_share_start (false))
 125     {
 126       gomp_loop_ull_init (thr->ts.work_share, up, start, end, incr,
 127                           GFS_DYNAMIC, chunk_size);
 128       gomp_work_share_init_done ();
 129     }
 130
 131 #if defined HAVE_SYNC_BUILTINS && defined __LP64__
 132   ret = gomp_iter_ull_dynamic_next (istart, iend);
 133 #else
 134   gomp_mutex_lock (&thr->ts.work_share->lock);
 135   ret = gomp_iter_ull_dynamic_next_locked (istart, iend);
 136   gomp_mutex_unlock (&thr->ts.work_share->lock);
 137 #endif
 138
 139   return ret;
 140 }
 141
 142 static bool
 143 gomp_loop_ull_guided_start (bool up, gomp_ull start, gomp_ull end,
 144                             gomp_ull incr, gomp_ull chunk_size,
 145                             gomp_ull *istart, gomp_ull *iend)
 146 {
 147   struct gomp_thread *thr = gomp_thread ();
 148   bool ret;
 149
 150   if (gomp_work_share_start (false))
 151     {
 152       gomp_loop_ull_init (thr->ts.work_share, up, start, end, incr,
 153                           GFS_GUIDED, chunk_size);
 154       gomp_work_share_init_done ();
 155     }
 156
 157 #if defined HAVE_SYNC_BUILTINS && defined __LP64__
 158   ret = gomp_iter_ull_guided_next (istart, iend);
 159 #else
 160   gomp_mutex_lock (&thr->ts.work_share->lock);
 161   ret = gomp_iter_ull_guided_next_locked (istart, iend);
 162   gomp_mutex_unlock (&thr->ts.work_share->lock);
 163 #endif
 164
 165   return ret;
 166 }
 167
 168 bool
 169 GOMP_loop_ull_runtime_start (bool up, gomp_ull start, gomp_ull end,
 170                              gomp_ull incr, gomp_ull *istart, gomp_ull *iend)
 171 {
 172   struct gomp_task_icv *icv = gomp_icv (false);
 173   switch (icv->run_sched_var)
 174     {
 175     case GFS_STATIC:
 176       return gomp_loop_ull_static_start (up, start, end, incr,
 177                                          icv->run_sched_modifier,
 178                                          istart, iend);
 179     case GFS_DYNAMIC:
 180       return gomp_loop_ull_dynamic_start (up, start, end, incr,
 181                                           icv->run_sched_modifier,
 182                                           istart, iend);
 183     case GFS_GUIDED:
 184       return gomp_loop_ull_guided_start (up, start, end, incr,
 185                                          icv->run_sched_modifier,
 186                                          istart, iend);
 187     case GFS_AUTO:
 188       /* For now map to schedule(static), later on we could play with feedback
 189          driven choice.  */
 190       return gomp_loop_ull_static_start (up, start, end, incr,
 191                                          0, istart, iend);
 192     default:
 193       abort ();
 194     }
 195 }
 196
 197 /* The *_ordered_*_start routines are similar.  The only difference is that
 198    this work-share construct is initialized to expect an ORDERED section.  */
 199
 200 static bool
 201 gomp_loop_ull_ordered_static_start (bool up, gomp_ull start, gomp_ull end,
 202                                     gomp_ull incr, gomp_ull chunk_size,
 203                                     gomp_ull *istart, gomp_ull *iend)
 204 {
 205   struct gomp_thread *thr = gomp_thread ();
 206
 207   thr->ts.static_trip = 0;
 208   if (gomp_work_share_start (true))
 209     {
 210       gomp_loop_ull_init (thr->ts.work_share, up, start, end, incr,
 211                           GFS_STATIC, chunk_size);
 212       gomp_ordered_static_init ();
 213       gomp_work_share_init_done ();
 214     }
 215
 216   return !gomp_iter_ull_static_next (istart, iend);
 217 }
 218
 219 static bool
 220 gomp_loop_ull_ordered_dynamic_start (bool up, gomp_ull start, gomp_ull end,
 221                                      gomp_ull incr, gomp_ull chunk_size,
 222                                      gomp_ull *istart, gomp_ull *iend)
 223 {
 224   struct gomp_thread *thr = gomp_thread ();
 225   bool ret;
 226
 227   if (gomp_work_share_start (true))
 228     {
 229       gomp_loop_ull_init (thr->ts.work_share, up, start, end, incr,
 230                           GFS_DYNAMIC, chunk_size);
 231       gomp_mutex_lock (&thr->ts.work_share->lock);
 232       gomp_work_share_init_done ();
 233     }
 234   else
 235     gomp_mutex_lock (&thr->ts.work_share->lock);
 236
 237   ret = gomp_iter_ull_dynamic_next_locked (istart, iend);
 238   if (ret)
 239     gomp_ordered_first ();
 240   gomp_mutex_unlock (&thr->ts.work_share->lock);
 241
 242   return ret;
 243 }
 244
 245 static bool
 246 gomp_loop_ull_ordered_guided_start (bool up, gomp_ull start, gomp_ull end,
 247                                     gomp_ull incr, gomp_ull chunk_size,
 248                                     gomp_ull *istart, gomp_ull *iend)
 249 {
 250   struct gomp_thread *thr = gomp_thread ();
 251   bool ret;
 252
 253   if (gomp_work_share_start (true))
 254     {
 255       gomp_loop_ull_init (thr->ts.work_share, up, start, end, incr,
 256                           GFS_GUIDED, chunk_size);
 257       gomp_mutex_lock (&thr->ts.work_share->lock);
 258       gomp_work_share_init_done ();
 259     }
 260   else
 261     gomp_mutex_lock (&thr->ts.work_share->lock);
 262
 263   ret = gomp_iter_ull_guided_next_locked (istart, iend);
 264   if (ret)
 265     gomp_ordered_first ();
 266   gomp_mutex_unlock (&thr->ts.work_share->lock);
 267
 268   return ret;
 269 }
 270
 271 bool
 272 GOMP_loop_ull_ordered_runtime_start (bool up, gomp_ull start, gomp_ull end,
 273                                      gomp_ull incr, gomp_ull *istart,
 274                                      gomp_ull *iend)
 275 {
 276   struct gomp_task_icv *icv = gomp_icv (false);
 277   switch (icv->run_sched_var)
 278     {
 279     case GFS_STATIC:
 280       return gomp_loop_ull_ordered_static_start (up, start, end, incr,
 281                                                  icv->run_sched_modifier,
 282                                                  istart, iend);
 283     case GFS_DYNAMIC:
 284       return gomp_loop_ull_ordered_dynamic_start (up, start, end, incr,
 285                                                   icv->run_sched_modifier,
 286                                                   istart, iend);
 287     case GFS_GUIDED:
 288       return gomp_loop_ull_ordered_guided_start (up, start, end, incr,
 289                                                  icv->run_sched_modifier,
 290                                                  istart, iend);
 291     case GFS_AUTO:
 292       /* For now map to schedule(static), later on we could play with feedback
 293          driven choice.  */
 294       return gomp_loop_ull_ordered_static_start (up, start, end, incr,
 295                                                  0, istart, iend);
 296     default:
 297       abort ();
 298     }
 299 }
 300
 301 /* The *_next routines are called when the thread completes processing of
 302    the iteration block currently assigned to it.  If the work-share
 303    construct is bound directly to a parallel construct, then the iteration
 304    bounds may have been set up before the parallel.  In which case, this
 305    may be the first iteration for the thread.
 306
 307    Returns true if there is work remaining to be performed; *ISTART and
 308    *IEND are filled with a new iteration block.  Returns false if all work
 309    has been assigned.  */
 310
 311 static bool
 312 gomp_loop_ull_static_next (gomp_ull *istart, gomp_ull *iend)
 313 {
 314   return !gomp_iter_ull_static_next (istart, iend);
 315 }
 316
 317 static bool
 318 gomp_loop_ull_dynamic_next (gomp_ull *istart, gomp_ull *iend)
 319 {
 320   bool ret;
 321
 322 #if defined HAVE_SYNC_BUILTINS && defined __LP64__
 323   ret = gomp_iter_ull_dynamic_next (istart, iend);
 324 #else
 325   struct gomp_thread *thr = gomp_thread ();
 326   gomp_mutex_lock (&thr->ts.work_share->lock);
 327   ret = gomp_iter_ull_dynamic_next_locked (istart, iend);
 328   gomp_mutex_unlock (&thr->ts.work_share->lock);
 329 #endif
 330
 331   return ret;
 332 }
 333
 334 static bool
 335 gomp_loop_ull_guided_next (gomp_ull *istart, gomp_ull *iend)
 336 {
 337   bool ret;
 338
 339 #if defined HAVE_SYNC_BUILTINS && defined __LP64__
 340   ret = gomp_iter_ull_guided_next (istart, iend);
 341 #else
 342   struct gomp_thread *thr = gomp_thread ();
 343   gomp_mutex_lock (&thr->ts.work_share->lock);
 344   ret = gomp_iter_ull_guided_next_locked (istart, iend);
 345   gomp_mutex_unlock (&thr->ts.work_share->lock);
 346 #endif
 347
 348   return ret;
 349 }
 350
 351 bool
 352 GOMP_loop_ull_runtime_next (gomp_ull *istart, gomp_ull *iend)
 353 {
 354   struct gomp_thread *thr = gomp_thread ();
 355
 356   switch (thr->ts.work_share->sched)
 357     {
 358     case GFS_STATIC:
 359     case GFS_AUTO:
 360       return gomp_loop_ull_static_next (istart, iend);
 361     case GFS_DYNAMIC:
 362       return gomp_loop_ull_dynamic_next (istart, iend);
 363     case GFS_GUIDED:
 364       return gomp_loop_ull_guided_next (istart, iend);
 365     default:
 366       abort ();
 367     }
 368 }
 369
 370 /* The *_ordered_*_next routines are called when the thread completes
 371    processing of the iteration block currently assigned to it.
 372
 373    Returns true if there is work remaining to be performed; *ISTART and
 374    *IEND are filled with a new iteration block.  Returns false if all work
 375    has been assigned.  */
 376
 377 static bool
 378 gomp_loop_ull_ordered_static_next (gomp_ull *istart, gomp_ull *iend)
 379 {
 380   struct gomp_thread *thr = gomp_thread ();
 381   int test;
 382
 383   gomp_ordered_sync ();
 384   gomp_mutex_lock (&thr->ts.work_share->lock);
 385   test = gomp_iter_ull_static_next (istart, iend);
 386   if (test >= 0)
 387     gomp_ordered_static_next ();
 388   gomp_mutex_unlock (&thr->ts.work_share->lock);
 389
 390   return test == 0;
 391 }
 392
 393 static bool
 394 gomp_loop_ull_ordered_dynamic_next (gomp_ull *istart, gomp_ull *iend)
 395 {
 396   struct gomp_thread *thr = gomp_thread ();
 397   bool ret;
 398
 399   gomp_ordered_sync ();
 400   gomp_mutex_lock (&thr->ts.work_share->lock);
 401   ret = gomp_iter_ull_dynamic_next_locked (istart, iend);
 402   if (ret)
 403     gomp_ordered_next ();
 404   else
 405     gomp_ordered_last ();
 406   gomp_mutex_unlock (&thr->ts.work_share->lock);
 407
 408   return ret;
 409 }
 410
 411 static bool
 412 gomp_loop_ull_ordered_guided_next (gomp_ull *istart, gomp_ull *iend)
 413 {
 414   struct gomp_thread *thr = gomp_thread ();
 415   bool ret;
 416
 417   gomp_ordered_sync ();
 418   gomp_mutex_lock (&thr->ts.work_share->lock);
 419   ret = gomp_iter_ull_guided_next_locked (istart, iend);
 420   if (ret)
 421     gomp_ordered_next ();
 422   else
 423     gomp_ordered_last ();
 424   gomp_mutex_unlock (&thr->ts.work_share->lock);
 425
 426   return ret;
 427 }
 428
 429 bool
 430 GOMP_loop_ull_ordered_runtime_next (gomp_ull *istart, gomp_ull *iend)
 431 {
 432   struct gomp_thread *thr = gomp_thread ();
 433
 434   switch (thr->ts.work_share->sched)
 435     {
 436     case GFS_STATIC:
 437     case GFS_AUTO:
 438       return gomp_loop_ull_ordered_static_next (istart, iend);
 439     case GFS_DYNAMIC:
 440       return gomp_loop_ull_ordered_dynamic_next (istart, iend);
 441     case GFS_GUIDED:
 442       return gomp_loop_ull_ordered_guided_next (istart, iend);
 443     default:
 444       abort ();
 445     }
 446 }
 447
 448 /* We use static functions above so that we're sure that the "runtime"
 449    function can defer to the proper routine without interposition.  We
 450    export the static function with a strong alias when possible, or with
 451    a wrapper function otherwise.  */
 452
 453 #ifdef HAVE_ATTRIBUTE_ALIAS
 454 extern __typeof(gomp_loop_ull_static_start) GOMP_loop_ull_static_start
 455         __attribute__((alias ("gomp_loop_ull_static_start")));
 456 extern __typeof(gomp_loop_ull_dynamic_start) GOMP_loop_ull_dynamic_start
 457         __attribute__((alias ("gomp_loop_ull_dynamic_start")));
 458 extern __typeof(gomp_loop_ull_guided_start) GOMP_loop_ull_guided_start
 459         __attribute__((alias ("gomp_loop_ull_guided_start")));
 460
 461 extern __typeof(gomp_loop_ull_ordered_static_start) GOMP_loop_ull_ordered_static_start
 462         __attribute__((alias ("gomp_loop_ull_ordered_static_start")));
 463 extern __typeof(gomp_loop_ull_ordered_dynamic_start) GOMP_loop_ull_ordered_dynamic_start
 464         __attribute__((alias ("gomp_loop_ull_ordered_dynamic_start")));
 465 extern __typeof(gomp_loop_ull_ordered_guided_start) GOMP_loop_ull_ordered_guided_start
 466         __attribute__((alias ("gomp_loop_ull_ordered_guided_start")));
 467
 468 extern __typeof(gomp_loop_ull_static_next) GOMP_loop_ull_static_next
 469         __attribute__((alias ("gomp_loop_ull_static_next")));
 470 extern __typeof(gomp_loop_ull_dynamic_next) GOMP_loop_ull_dynamic_next
 471         __attribute__((alias ("gomp_loop_ull_dynamic_next")));
 472 extern __typeof(gomp_loop_ull_guided_next) GOMP_loop_ull_guided_next
 473         __attribute__((alias ("gomp_loop_ull_guided_next")));
 474
 475 extern __typeof(gomp_loop_ull_ordered_static_next) GOMP_loop_ull_ordered_static_next
 476         __attribute__((alias ("gomp_loop_ull_ordered_static_next")));
 477 extern __typeof(gomp_loop_ull_ordered_dynamic_next) GOMP_loop_ull_ordered_dynamic_next
 478         __attribute__((alias ("gomp_loop_ull_ordered_dynamic_next")));
 479 extern __typeof(gomp_loop_ull_ordered_guided_next) GOMP_loop_ull_ordered_guided_next
 480         __attribute__((alias ("gomp_loop_ull_ordered_guided_next")));
 481 #else
 482 bool
 483 GOMP_loop_ull_static_start (bool up, gomp_ull start, gomp_ull end,
 484                             gomp_ull incr, gomp_ull chunk_size,
 485                             gomp_ull *istart, gomp_ull *iend)
 486 {
 487   return gomp_loop_ull_static_start (up, start, end, incr, chunk_size, istart,
 488                                      iend);
 489 }
 490
 491 bool
 492 GOMP_loop_ull_dynamic_start (bool up, gomp_ull start, gomp_ull end,
 493                              gomp_ull incr, gomp_ull chunk_size,
 494                              gomp_ull *istart, gomp_ull *iend)
 495 {
 496   return gomp_loop_ull_dynamic_start (up, start, end, incr, chunk_size, istart,
 497                                       iend);
 498 }
 499
 500 bool
 501 GOMP_loop_ull_guided_start (bool up, gomp_ull start, gomp_ull end,
 502                             gomp_ull incr, gomp_ull chunk_size,
 503                             gomp_ull *istart, gomp_ull *iend)
 504 {
 505   return gomp_loop_ull_guided_start (up, start, end, incr, chunk_size, istart,
 506                                      iend);
 507 }
 508
 509 bool
 510 GOMP_loop_ull_ordered_static_start (bool up, gomp_ull start, gomp_ull end,
 511                                     gomp_ull incr, gomp_ull chunk_size,
 512                                     gomp_ull *istart, gomp_ull *iend)
 513 {
 514   return gomp_loop_ull_ordered_static_start (up, start, end, incr, chunk_size,
 515                                              istart, iend);
 516 }
 517
 518 bool
 519 GOMP_loop_ull_ordered_dynamic_start (bool up, gomp_ull start, gomp_ull end,
 520                                      gomp_ull incr, gomp_ull chunk_size,
 521                                      gomp_ull *istart, gomp_ull *iend)
 522 {
 523   return gomp_loop_ull_ordered_dynamic_start (up, start, end, incr, chunk_size,
 524                                               istart, iend);
 525 }
 526
 527 bool
 528 GOMP_loop_ull_ordered_guided_start (bool up, gomp_ull start, gomp_ull end,
 529                                     gomp_ull incr, gomp_ull chunk_size,
 530                                     gomp_ull *istart, gomp_ull *iend)
 531 {
 532   return gomp_loop_ull_ordered_guided_start (up, start, end, incr, chunk_size,
 533                                              istart, iend);
 534 }
 535
 536 bool
 537 GOMP_loop_ull_static_next (gomp_ull *istart, gomp_ull *iend)
 538 {
 539   return gomp_loop_ull_static_next (istart, iend);
 540 }
 541
 542 bool
 543 GOMP_loop_ull_dynamic_next (gomp_ull *istart, gomp_ull *iend)
 544 {
 545   return gomp_loop_ull_dynamic_next (istart, iend);
 546 }
 547
 548 bool
 549 GOMP_loop_ull_guided_next (gomp_ull *istart, gomp_ull *iend)
 550 {
 551   return gomp_loop_ull_guided_next (istart, iend);
 552 }
 553
 554 bool
 555 GOMP_loop_ull_ordered_static_next (gomp_ull *istart, gomp_ull *iend)
 556 {
 557   return gomp_loop_ull_ordered_static_next (istart, iend);
 558 }
 559
 560 bool
 561 GOMP_loop_ull_ordered_dynamic_next (gomp_ull *istart, gomp_ull *iend)
 562 {
 563   return gomp_loop_ull_ordered_dynamic_next (istart, iend);
 564 }
 565
 566 bool
 567 GOMP_loop_ull_ordered_guided_next (gomp_ull *istart, gomp_ull *iend)
 568 {
 569   return gomp_loop_ull_ordered_guided_next (istart, iend);
 570 }
 571 #endif