1/* Copyright (C) 2005 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 Lesser General Public License as published by
8   the Free Software Foundation; either version 2.1 of the License, or
9   (at your option) 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 Lesser General Public License for
14   more details.
15
16   You should have received a copy of the GNU Lesser General Public License
17   along with libgomp; see the file COPYING.LIB.  If not, write to the
18   Free Software Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston,
19   MA 02110-1301, USA.  */
20
21/* As a special exception, if you link this library with other files, some
22   of which are compiled with GCC, to produce an executable, this library
23   does not by itself cause the resulting executable to be covered by the
24   GNU General Public License.  This exception does not however invalidate
25   any other reasons why the executable file might be covered by the GNU
26   General Public License.  */
27
28/* This file contains routines to manage the work-share queue for a team
29   of threads.  */
30
31#include "libgomp.h"
32#include <stdlib.h>
33#include <string.h>
34
35
36/* Create a new work share structure.  */
37
38struct gomp_work_share *
39gomp_new_work_share (bool ordered, unsigned nthreads)
40{
41  struct gomp_work_share *ws;
42  size_t size;
43
44  size = sizeof (*ws);
45  if (ordered)
46    size += nthreads * sizeof (ws->ordered_team_ids[0]);
47
48  ws = gomp_malloc_cleared (size);
49  gomp_mutex_init (&ws->lock);
50  ws->ordered_owner = -1;
51
52  return ws;
53}
54
55
56/* Free a work share structure.  */
57
58static void
59free_work_share (struct gomp_work_share *ws)
60{
61  gomp_mutex_destroy (&ws->lock);
62  free (ws);
63}
64
65
66/* The current thread is ready to begin the next work sharing construct.
67   In all cases, thr->ts.work_share is updated to point to the new
68   structure.  In all cases the work_share lock is locked.  Return true
69   if this was the first thread to reach this point.  */
70
71bool
72gomp_work_share_start (bool ordered)
73{
74  struct gomp_thread *thr = gomp_thread ();
75  struct gomp_team *team = thr->ts.team;
76  struct gomp_work_share *ws;
77  unsigned ws_index, ws_gen;
78
79  /* Work sharing constructs can be orphaned.  */
80  if (team == NULL)
81    {
82      ws = gomp_new_work_share (ordered, 1);
83      thr->ts.work_share = ws;
84      thr->ts.static_trip = 0;
85      gomp_mutex_lock (&ws->lock);
86      return true;
87    }
88
89  gomp_mutex_lock (&team->work_share_lock);
90
91  /* This thread is beginning its next generation.  */
92  ws_gen = ++thr->ts.work_share_generation;
93
94  /* If this next generation is not newer than any other generation in
95     the team, then simply reference the existing construct.  */
96  if (ws_gen - team->oldest_live_gen < team->num_live_gen)
97    {
98      ws_index = ws_gen & team->generation_mask;
99      ws = team->work_shares[ws_index];
100      thr->ts.work_share = ws;
101      thr->ts.static_trip = 0;
102
103      gomp_mutex_lock (&ws->lock);
104      gomp_mutex_unlock (&team->work_share_lock);
105
106      return false;
107    }
108
109  /* Resize the work shares queue if we've run out of space.  */
110  if (team->num_live_gen++ == team->generation_mask)
111    {
112      team->work_shares = gomp_realloc (team->work_shares,
113					2 * team->num_live_gen
114					* sizeof (*team->work_shares));
115
116      /* Unless oldest_live_gen is zero, the sequence of live elements
117	 wraps around the end of the array.  If we do nothing, we break
118	 lookup of the existing elements.  Fix that by unwrapping the
119	 data from the front to the end.  */
120      if (team->oldest_live_gen > 0)
121	memcpy (team->work_shares + team->num_live_gen,
122		team->work_shares,
123		(team->oldest_live_gen & team->generation_mask)
124		* sizeof (*team->work_shares));
125
126      team->generation_mask = team->generation_mask * 2 + 1;
127    }
128
129  ws_index = ws_gen & team->generation_mask;
130  ws = gomp_new_work_share (ordered, team->nthreads);
131  thr->ts.work_share = ws;
132  thr->ts.static_trip = 0;
133  team->work_shares[ws_index] = ws;
134
135  gomp_mutex_lock (&ws->lock);
136  gomp_mutex_unlock (&team->work_share_lock);
137
138  return true;
139}
140
141
142/* The current thread is done with its current work sharing construct.
143   This version does imply a barrier at the end of the work-share.  */
144
145void
146gomp_work_share_end (void)
147{
148  struct gomp_thread *thr = gomp_thread ();
149  struct gomp_team *team = thr->ts.team;
150  struct gomp_work_share *ws = thr->ts.work_share;
151  bool last;
152
153  thr->ts.work_share = NULL;
154
155  /* Work sharing constructs can be orphaned.  */
156  if (team == NULL)
157    {
158      free_work_share (ws);
159      return;
160    }
161
162  last = gomp_barrier_wait_start (&team->barrier);
163
164  if (last)
165    {
166      unsigned ws_index;
167
168      ws_index = thr->ts.work_share_generation & team->generation_mask;
169      team->work_shares[ws_index] = NULL;
170      team->oldest_live_gen++;
171      team->num_live_gen = 0;
172
173      free_work_share (ws);
174    }
175
176  gomp_barrier_wait_end (&team->barrier, last);
177}
178
179
180/* The current thread is done with its current work sharing construct.
181   This version does NOT imply a barrier at the end of the work-share.  */
182
183void
184gomp_work_share_end_nowait (void)
185{
186  struct gomp_thread *thr = gomp_thread ();
187  struct gomp_team *team = thr->ts.team;
188  struct gomp_work_share *ws = thr->ts.work_share;
189  unsigned completed;
190
191  thr->ts.work_share = NULL;
192
193  /* Work sharing constructs can be orphaned.  */
194  if (team == NULL)
195    {
196      free_work_share (ws);
197      return;
198    }
199
200#ifdef HAVE_SYNC_BUILTINS
201  completed = __sync_add_and_fetch (&ws->threads_completed, 1);
202#else
203  gomp_mutex_lock (&ws->lock);
204  completed = ++ws->threads_completed;
205  gomp_mutex_unlock (&ws->lock);
206#endif
207
208  if (completed == team->nthreads)
209    {
210      unsigned ws_index;
211
212      gomp_mutex_lock (&team->work_share_lock);
213
214      ws_index = thr->ts.work_share_generation & team->generation_mask;
215      team->work_shares[ws_index] = NULL;
216      team->oldest_live_gen++;
217      team->num_live_gen--;
218
219      gomp_mutex_unlock (&team->work_share_lock);
220
221      free_work_share (ws);
222    }
223}
224