1 /* Copyright (C) 2015-2023 Free Software Foundation, Inc.
2 Contributed by Mentor Embedded.
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 is an AMD GCN specific implementation of a barrier synchronization
27 mechanism for libgomp. This type is private to the library. This
28 implementation uses atomic instructions and s_barrier instruction. It
29 uses MEMMODEL_RELAXED here because barriers are within workgroups and
30 therefore don't need to flush caches. */
31
32 #include <limits.h>
33 #include "libgomp.h"
34
35
36 void
37 gomp_barrier_wait_end (gomp_barrier_t *bar, gomp_barrier_state_t state)
38 {
39 if (__builtin_expect (state & BAR_WAS_LAST, 0))
40 {
41 /* Next time we'll be awaiting TOTAL threads again. */
42 bar->awaited = bar->total;
43 __atomic_store_n (&bar->generation, bar->generation + BAR_INCR,
44 MEMMODEL_RELAXED);
45 }
46 if (bar->total > 1)
47 asm ("s_barrier" ::: "memory");
48 }
49
50 void
51 gomp_barrier_wait (gomp_barrier_t *bar)
52 {
53 gomp_barrier_wait_end (bar, gomp_barrier_wait_start (bar));
54 }
55
56 /* Like gomp_barrier_wait, except that if the encountering thread
57 is not the last one to hit the barrier, it returns immediately.
58 The intended usage is that a thread which intends to gomp_barrier_destroy
59 this barrier calls gomp_barrier_wait, while all other threads
60 call gomp_barrier_wait_last. When gomp_barrier_wait returns,
61 the barrier can be safely destroyed. */
62
63 void
64 gomp_barrier_wait_last (gomp_barrier_t *bar)
65 {
66 /* Deferring to gomp_barrier_wait does not use the optimization opportunity
67 allowed by the interface contract for all-but-last participants. The
68 original implementation in config/linux/bar.c handles this better. */
69 gomp_barrier_wait (bar);
70 }
71
72 void
73 gomp_team_barrier_wake (gomp_barrier_t *bar, int count)
74 {
75 if (bar->total > 1)
76 asm ("s_barrier" ::: "memory");
77 }
78
79 void
80 gomp_team_barrier_wait_end (gomp_barrier_t *bar, gomp_barrier_state_t state)
81 {
82 unsigned int generation, gen;
83
84 if (__builtin_expect (state & BAR_WAS_LAST, 0))
85 {
86 /* Next time we'll be awaiting TOTAL threads again. */
87 struct gomp_thread *thr = gomp_thread ();
88 struct gomp_team *team = thr->ts.team;
89
90 bar->awaited = bar->total;
91 team->work_share_cancelled = 0;
92 if (__builtin_expect (team->task_count, 0))
93 {
94 gomp_barrier_handle_tasks (state);
95 state &= ~BAR_WAS_LAST;
96 }
97 else
98 {
99 state &= ~BAR_CANCELLED;
100 state += BAR_INCR - BAR_WAS_LAST;
101 __atomic_store_n (&bar->generation, state, MEMMODEL_RELAXED);
102 if (bar->total > 1)
103 asm ("s_barrier" ::: "memory");
104 return;
105 }
106 }
107
108 generation = state;
109 state &= ~BAR_CANCELLED;
110 int retry = 100;
111 do
112 {
113 if (retry-- == 0)
114 {
115 /* It really shouldn't happen that barriers get out of sync, but
116 if they do then this will loop until they realign, so we need
117 to avoid an infinite loop where the thread just isn't there. */
118 const char msg[] = ("Barrier sync failed (another thread died?);"
119 " aborting.");
120 write (2, msg, sizeof (msg)-1);
121 abort();
122 }
123
124 asm ("s_barrier" ::: "memory");
125 gen = __atomic_load_n (&bar->generation, MEMMODEL_ACQUIRE);
126 if (__builtin_expect (gen & BAR_TASK_PENDING, 0))
127 {
128 gomp_barrier_handle_tasks (state);
129 gen = __atomic_load_n (&bar->generation, MEMMODEL_ACQUIRE);
130 }
131 generation |= gen & BAR_WAITING_FOR_TASK;
132 }
133 while (gen != state + BAR_INCR);
134 }
135
136 void
137 gomp_team_barrier_wait (gomp_barrier_t *bar)
138 {
139 gomp_team_barrier_wait_end (bar, gomp_barrier_wait_start (bar));
140 }
141
142 void
143 gomp_team_barrier_wait_final (gomp_barrier_t *bar)
144 {
145 gomp_barrier_state_t state = gomp_barrier_wait_final_start (bar);
146 if (__builtin_expect (state & BAR_WAS_LAST, 0))
147 bar->awaited_final = bar->total;
148 gomp_team_barrier_wait_end (bar, state);
149 }
150
151 bool
152 gomp_team_barrier_wait_cancel_end (gomp_barrier_t *bar,
153 gomp_barrier_state_t state)
154 {
155 unsigned int generation, gen;
156
157 if (__builtin_expect (state & BAR_WAS_LAST, 0))
158 {
159 /* Next time we'll be awaiting TOTAL threads again. */
160 /* BAR_CANCELLED should never be set in state here, because
161 cancellation means that at least one of the threads has been
162 cancelled, thus on a cancellable barrier we should never see
163 all threads to arrive. */
164 struct gomp_thread *thr = gomp_thread ();
165 struct gomp_team *team = thr->ts.team;
166
167 bar->awaited = bar->total;
168 team->work_share_cancelled = 0;
169 if (__builtin_expect (team->task_count, 0))
170 {
171 gomp_barrier_handle_tasks (state);
172 state &= ~BAR_WAS_LAST;
173 }
174 else
175 {
176 state += BAR_INCR - BAR_WAS_LAST;
177 __atomic_store_n (&bar->generation, state, MEMMODEL_RELAXED);
178 if (bar->total > 1)
179 asm ("s_barrier" ::: "memory");
180 return false;
181 }
182 }
183
184 if (__builtin_expect (state & BAR_CANCELLED, 0))
185 return true;
186
187 generation = state;
188 int retry = 100;
189 do
190 {
191 if (retry-- == 0)
192 {
193 /* It really shouldn't happen that barriers get out of sync, but
194 if they do then this will loop until they realign, so we need
195 to avoid an infinite loop where the thread just isn't there. */
196 const char msg[] = ("Barrier sync failed (another thread died?);"
197 " aborting.");
198 write (2, msg, sizeof (msg)-1);
199 abort();
200 }
201
202 if (bar->total > 1)
203 asm ("s_barrier" ::: "memory");
204 gen = __atomic_load_n (&bar->generation, MEMMODEL_RELAXED);
205 if (__builtin_expect (gen & BAR_CANCELLED, 0))
206 return true;
207 if (__builtin_expect (gen & BAR_TASK_PENDING, 0))
208 {
209 gomp_barrier_handle_tasks (state);
210 gen = __atomic_load_n (&bar->generation, MEMMODEL_RELAXED);
211 }
212 generation |= gen & BAR_WAITING_FOR_TASK;
213 }
214 while (gen != state + BAR_INCR);
215
216 return false;
217 }
218
219 bool
220 gomp_team_barrier_wait_cancel (gomp_barrier_t *bar)
221 {
222 return gomp_team_barrier_wait_cancel_end (bar, gomp_barrier_wait_start (bar));
223 }
224
225 void
226 gomp_team_barrier_cancel (struct gomp_team *team)
227 {
228 gomp_mutex_lock (&team->task_lock);
229 if (team->barrier.generation & BAR_CANCELLED)
230 {
231 gomp_mutex_unlock (&team->task_lock);
232 return;
233 }
234 team->barrier.generation |= BAR_CANCELLED;
235 gomp_mutex_unlock (&team->task_lock);
236 gomp_team_barrier_wake (&team->barrier, INT_MAX);
237 }