]>
Commit | Line | Data |
---|---|---|
818ab71a | 1 | /* Copyright (C) 2005-2016 Free Software Foundation, Inc. |
13c41b2e SH |
2 | Contributed by Sebastian Huber <sebastian.huber@embedded-brains.de>. |
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 is the RTEMS implementation of a barrier synchronization | |
26 | mechanism for libgomp. It is identical to the Linux implementation, except | |
27 | that the futex API is slightly different. This type is private to the | |
28 | library. */ | |
29 | ||
30 | #include "libgomp.h" | |
31 | #include "bar.h" | |
32 | #include <limits.h> | |
33 | ||
34 | static gomp_barrier_t * | |
35 | generation_to_barrier (int *addr) | |
36 | { | |
37 | return (gomp_barrier_t *) | |
38 | ((char *) addr - __builtin_offsetof (gomp_barrier_t, generation)); | |
39 | } | |
40 | ||
41 | static void | |
42 | futex_wait (int *addr, int val) | |
43 | { | |
44 | gomp_barrier_t *bar = generation_to_barrier (addr); | |
45 | _Futex_Wait (&bar->futex, addr, val); | |
46 | } | |
47 | ||
48 | static void | |
49 | futex_wake (int *addr, int count) | |
50 | { | |
51 | gomp_barrier_t *bar = generation_to_barrier (addr); | |
52 | _Futex_Wake (&bar->futex, count); | |
53 | } | |
54 | ||
55 | static int | |
56 | do_spin (int *addr, int val) | |
57 | { | |
58 | unsigned long long i, count = gomp_spin_count_var; | |
59 | ||
60 | if (__builtin_expect (gomp_managed_threads > gomp_available_cpus, 0)) | |
61 | count = gomp_throttled_spin_count_var; | |
62 | for (i = 0; i < count; i++) | |
63 | if (__builtin_expect (__atomic_load_n (addr, MEMMODEL_RELAXED) != val, 0)) | |
64 | return 0; | |
65 | return 1; | |
66 | } | |
67 | ||
68 | static void | |
69 | do_wait (int *addr, int val) | |
70 | { | |
71 | if (do_spin (addr, val)) | |
72 | futex_wait (addr, val); | |
73 | } | |
74 | ||
75 | /* Everything below this point should be identical to the Linux | |
76 | implementation. */ | |
77 | ||
78 | void | |
79 | gomp_barrier_wait_end (gomp_barrier_t *bar, gomp_barrier_state_t state) | |
80 | { | |
81 | if (__builtin_expect (state & BAR_WAS_LAST, 0)) | |
82 | { | |
83 | /* Next time we'll be awaiting TOTAL threads again. */ | |
84 | bar->awaited = bar->total; | |
85 | __atomic_store_n (&bar->generation, bar->generation + BAR_INCR, | |
86 | MEMMODEL_RELEASE); | |
87 | futex_wake ((int *) &bar->generation, INT_MAX); | |
88 | } | |
89 | else | |
90 | { | |
91 | do | |
92 | do_wait ((int *) &bar->generation, state); | |
93 | while (__atomic_load_n (&bar->generation, MEMMODEL_ACQUIRE) == state); | |
94 | } | |
95 | } | |
96 | ||
97 | void | |
98 | gomp_barrier_wait (gomp_barrier_t *bar) | |
99 | { | |
100 | gomp_barrier_wait_end (bar, gomp_barrier_wait_start (bar)); | |
101 | } | |
102 | ||
103 | /* Like gomp_barrier_wait, except that if the encountering thread | |
104 | is not the last one to hit the barrier, it returns immediately. | |
105 | The intended usage is that a thread which intends to gomp_barrier_destroy | |
106 | this barrier calls gomp_barrier_wait, while all other threads | |
107 | call gomp_barrier_wait_last. When gomp_barrier_wait returns, | |
108 | the barrier can be safely destroyed. */ | |
109 | ||
110 | void | |
111 | gomp_barrier_wait_last (gomp_barrier_t *bar) | |
112 | { | |
113 | gomp_barrier_state_t state = gomp_barrier_wait_start (bar); | |
114 | if (state & BAR_WAS_LAST) | |
115 | gomp_barrier_wait_end (bar, state); | |
116 | } | |
117 | ||
118 | void | |
119 | gomp_team_barrier_wake (gomp_barrier_t *bar, int count) | |
120 | { | |
121 | futex_wake ((int *) &bar->generation, count == 0 ? INT_MAX : count); | |
122 | } | |
123 | ||
124 | void | |
125 | gomp_team_barrier_wait_end (gomp_barrier_t *bar, gomp_barrier_state_t state) | |
126 | { | |
127 | unsigned int generation, gen; | |
128 | ||
129 | if (__builtin_expect (state & BAR_WAS_LAST, 0)) | |
130 | { | |
131 | /* Next time we'll be awaiting TOTAL threads again. */ | |
132 | struct gomp_thread *thr = gomp_thread (); | |
133 | struct gomp_team *team = thr->ts.team; | |
134 | ||
135 | bar->awaited = bar->total; | |
136 | team->work_share_cancelled = 0; | |
137 | if (__builtin_expect (team->task_count, 0)) | |
138 | { | |
139 | gomp_barrier_handle_tasks (state); | |
140 | state &= ~BAR_WAS_LAST; | |
141 | } | |
142 | else | |
143 | { | |
144 | state &= ~BAR_CANCELLED; | |
145 | state += BAR_INCR - BAR_WAS_LAST; | |
146 | __atomic_store_n (&bar->generation, state, MEMMODEL_RELEASE); | |
147 | futex_wake ((int *) &bar->generation, INT_MAX); | |
148 | return; | |
149 | } | |
150 | } | |
151 | ||
152 | generation = state; | |
153 | state &= ~BAR_CANCELLED; | |
154 | do | |
155 | { | |
156 | do_wait ((int *) &bar->generation, generation); | |
157 | gen = __atomic_load_n (&bar->generation, MEMMODEL_ACQUIRE); | |
158 | if (__builtin_expect (gen & BAR_TASK_PENDING, 0)) | |
159 | { | |
160 | gomp_barrier_handle_tasks (state); | |
161 | gen = __atomic_load_n (&bar->generation, MEMMODEL_ACQUIRE); | |
162 | } | |
163 | generation |= gen & BAR_WAITING_FOR_TASK; | |
164 | } | |
165 | while (gen != state + BAR_INCR); | |
166 | } | |
167 | ||
168 | void | |
169 | gomp_team_barrier_wait (gomp_barrier_t *bar) | |
170 | { | |
171 | gomp_team_barrier_wait_end (bar, gomp_barrier_wait_start (bar)); | |
172 | } | |
173 | ||
174 | void | |
175 | gomp_team_barrier_wait_final (gomp_barrier_t *bar) | |
176 | { | |
177 | gomp_barrier_state_t state = gomp_barrier_wait_final_start (bar); | |
178 | if (__builtin_expect (state & BAR_WAS_LAST, 0)) | |
179 | bar->awaited_final = bar->total; | |
180 | gomp_team_barrier_wait_end (bar, state); | |
181 | } | |
182 | ||
183 | bool | |
184 | gomp_team_barrier_wait_cancel_end (gomp_barrier_t *bar, | |
185 | gomp_barrier_state_t state) | |
186 | { | |
187 | unsigned int generation, gen; | |
188 | ||
189 | if (__builtin_expect (state & BAR_WAS_LAST, 0)) | |
190 | { | |
191 | /* Next time we'll be awaiting TOTAL threads again. */ | |
192 | /* BAR_CANCELLED should never be set in state here, because | |
193 | cancellation means that at least one of the threads has been | |
194 | cancelled, thus on a cancellable barrier we should never see | |
195 | all threads to arrive. */ | |
196 | struct gomp_thread *thr = gomp_thread (); | |
197 | struct gomp_team *team = thr->ts.team; | |
198 | ||
199 | bar->awaited = bar->total; | |
200 | team->work_share_cancelled = 0; | |
201 | if (__builtin_expect (team->task_count, 0)) | |
202 | { | |
203 | gomp_barrier_handle_tasks (state); | |
204 | state &= ~BAR_WAS_LAST; | |
205 | } | |
206 | else | |
207 | { | |
208 | state += BAR_INCR - BAR_WAS_LAST; | |
209 | __atomic_store_n (&bar->generation, state, MEMMODEL_RELEASE); | |
210 | futex_wake ((int *) &bar->generation, INT_MAX); | |
211 | return false; | |
212 | } | |
213 | } | |
214 | ||
215 | if (__builtin_expect (state & BAR_CANCELLED, 0)) | |
216 | return true; | |
217 | ||
218 | generation = state; | |
219 | do | |
220 | { | |
221 | do_wait ((int *) &bar->generation, generation); | |
222 | gen = __atomic_load_n (&bar->generation, MEMMODEL_ACQUIRE); | |
223 | if (__builtin_expect (gen & BAR_CANCELLED, 0)) | |
224 | return true; | |
225 | if (__builtin_expect (gen & BAR_TASK_PENDING, 0)) | |
226 | { | |
227 | gomp_barrier_handle_tasks (state); | |
228 | gen = __atomic_load_n (&bar->generation, MEMMODEL_ACQUIRE); | |
229 | } | |
230 | generation |= gen & BAR_WAITING_FOR_TASK; | |
231 | } | |
232 | while (gen != state + BAR_INCR); | |
233 | ||
234 | return false; | |
235 | } | |
236 | ||
237 | bool | |
238 | gomp_team_barrier_wait_cancel (gomp_barrier_t *bar) | |
239 | { | |
240 | return gomp_team_barrier_wait_cancel_end (bar, gomp_barrier_wait_start (bar)); | |
241 | } | |
242 | ||
243 | void | |
244 | gomp_team_barrier_cancel (struct gomp_team *team) | |
245 | { | |
246 | gomp_mutex_lock (&team->task_lock); | |
247 | if (team->barrier.generation & BAR_CANCELLED) | |
248 | { | |
249 | gomp_mutex_unlock (&team->task_lock); | |
250 | return; | |
251 | } | |
252 | team->barrier.generation |= BAR_CANCELLED; | |
253 | gomp_mutex_unlock (&team->task_lock); | |
254 | futex_wake ((int *) &team->barrier.generation, INT_MAX); | |
255 | } |