1 |
735 |
jeremybenn |
/* Copyright (C) 2005, 2009 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 ORDERED construct. */
|
26 |
|
|
|
27 |
|
|
#include "libgomp.h"
|
28 |
|
|
|
29 |
|
|
|
30 |
|
|
/* This function is called when first allocating an iteration block. That
|
31 |
|
|
is, the thread is not currently on the queue. The work-share lock must
|
32 |
|
|
be held on entry. */
|
33 |
|
|
|
34 |
|
|
void
|
35 |
|
|
gomp_ordered_first (void)
|
36 |
|
|
{
|
37 |
|
|
struct gomp_thread *thr = gomp_thread ();
|
38 |
|
|
struct gomp_team *team = thr->ts.team;
|
39 |
|
|
struct gomp_work_share *ws = thr->ts.work_share;
|
40 |
|
|
unsigned index;
|
41 |
|
|
|
42 |
|
|
/* Work share constructs can be orphaned. */
|
43 |
|
|
if (team == NULL || team->nthreads == 1)
|
44 |
|
|
return;
|
45 |
|
|
|
46 |
|
|
index = ws->ordered_cur + ws->ordered_num_used;
|
47 |
|
|
if (index >= team->nthreads)
|
48 |
|
|
index -= team->nthreads;
|
49 |
|
|
ws->ordered_team_ids[index] = thr->ts.team_id;
|
50 |
|
|
|
51 |
|
|
/* If this is the first and only thread in the queue, then there is
|
52 |
|
|
no one to release us when we get to our ordered section. Post to
|
53 |
|
|
our own release queue now so that we won't block later. */
|
54 |
|
|
if (ws->ordered_num_used++ == 0)
|
55 |
|
|
gomp_sem_post (team->ordered_release[thr->ts.team_id]);
|
56 |
|
|
}
|
57 |
|
|
|
58 |
|
|
/* This function is called when completing the last iteration block. That
|
59 |
|
|
is, there are no more iterations to perform and so the thread should be
|
60 |
|
|
removed from the queue entirely. Because of the way ORDERED blocks are
|
61 |
|
|
managed, it follows that we currently own access to the ORDERED block,
|
62 |
|
|
and should now pass it on to the next thread. The work-share lock must
|
63 |
|
|
be held on entry. */
|
64 |
|
|
|
65 |
|
|
void
|
66 |
|
|
gomp_ordered_last (void)
|
67 |
|
|
{
|
68 |
|
|
struct gomp_thread *thr = gomp_thread ();
|
69 |
|
|
struct gomp_team *team = thr->ts.team;
|
70 |
|
|
struct gomp_work_share *ws = thr->ts.work_share;
|
71 |
|
|
unsigned next_id;
|
72 |
|
|
|
73 |
|
|
/* Work share constructs can be orphaned. */
|
74 |
|
|
if (team == NULL || team->nthreads == 1)
|
75 |
|
|
return;
|
76 |
|
|
|
77 |
|
|
/* We're no longer the owner. */
|
78 |
|
|
ws->ordered_owner = -1;
|
79 |
|
|
|
80 |
|
|
/* If we're not the last thread in the queue, then wake the next. */
|
81 |
|
|
if (--ws->ordered_num_used > 0)
|
82 |
|
|
{
|
83 |
|
|
unsigned next = ws->ordered_cur + 1;
|
84 |
|
|
if (next == team->nthreads)
|
85 |
|
|
next = 0;
|
86 |
|
|
ws->ordered_cur = next;
|
87 |
|
|
|
88 |
|
|
next_id = ws->ordered_team_ids[next];
|
89 |
|
|
gomp_sem_post (team->ordered_release[next_id]);
|
90 |
|
|
}
|
91 |
|
|
}
|
92 |
|
|
|
93 |
|
|
|
94 |
|
|
/* This function is called when allocating a subsequent allocation block.
|
95 |
|
|
That is, we're done with the current iteration block and we're allocating
|
96 |
|
|
another. This is the logical combination of a call to gomp_ordered_last
|
97 |
|
|
followed by a call to gomp_ordered_first. The work-share lock must be
|
98 |
|
|
held on entry. */
|
99 |
|
|
|
100 |
|
|
void
|
101 |
|
|
gomp_ordered_next (void)
|
102 |
|
|
{
|
103 |
|
|
struct gomp_thread *thr = gomp_thread ();
|
104 |
|
|
struct gomp_team *team = thr->ts.team;
|
105 |
|
|
struct gomp_work_share *ws = thr->ts.work_share;
|
106 |
|
|
unsigned index, next_id;
|
107 |
|
|
|
108 |
|
|
/* Work share constructs can be orphaned. */
|
109 |
|
|
if (team == NULL || team->nthreads == 1)
|
110 |
|
|
return;
|
111 |
|
|
|
112 |
|
|
/* We're no longer the owner. */
|
113 |
|
|
ws->ordered_owner = -1;
|
114 |
|
|
|
115 |
|
|
/* If there's only one thread in the queue, that must be us. */
|
116 |
|
|
if (ws->ordered_num_used == 1)
|
117 |
|
|
{
|
118 |
|
|
/* We have a similar situation as in gomp_ordered_first
|
119 |
|
|
where we need to post to our own release semaphore. */
|
120 |
|
|
gomp_sem_post (team->ordered_release[thr->ts.team_id]);
|
121 |
|
|
return;
|
122 |
|
|
}
|
123 |
|
|
|
124 |
|
|
/* If the queue is entirely full, then we move ourself to the end of
|
125 |
|
|
the queue merely by incrementing ordered_cur. Only if it's not
|
126 |
|
|
full do we have to write our id. */
|
127 |
|
|
if (ws->ordered_num_used < team->nthreads)
|
128 |
|
|
{
|
129 |
|
|
index = ws->ordered_cur + ws->ordered_num_used;
|
130 |
|
|
if (index >= team->nthreads)
|
131 |
|
|
index -= team->nthreads;
|
132 |
|
|
ws->ordered_team_ids[index] = thr->ts.team_id;
|
133 |
|
|
}
|
134 |
|
|
|
135 |
|
|
index = ws->ordered_cur + 1;
|
136 |
|
|
if (index == team->nthreads)
|
137 |
|
|
index = 0;
|
138 |
|
|
ws->ordered_cur = index;
|
139 |
|
|
|
140 |
|
|
next_id = ws->ordered_team_ids[index];
|
141 |
|
|
gomp_sem_post (team->ordered_release[next_id]);
|
142 |
|
|
}
|
143 |
|
|
|
144 |
|
|
|
145 |
|
|
/* This function is called when a statically scheduled loop is first
|
146 |
|
|
being created. */
|
147 |
|
|
|
148 |
|
|
void
|
149 |
|
|
gomp_ordered_static_init (void)
|
150 |
|
|
{
|
151 |
|
|
struct gomp_thread *thr = gomp_thread ();
|
152 |
|
|
struct gomp_team *team = thr->ts.team;
|
153 |
|
|
|
154 |
|
|
if (team == NULL || team->nthreads == 1)
|
155 |
|
|
return;
|
156 |
|
|
|
157 |
|
|
gomp_sem_post (team->ordered_release[0]);
|
158 |
|
|
}
|
159 |
|
|
|
160 |
|
|
/* This function is called when a statically scheduled loop is moving to
|
161 |
|
|
the next allocation block. Static schedules are not first come first
|
162 |
|
|
served like the others, so we're to move to the numerically next thread,
|
163 |
|
|
not the next thread on a list. The work-share lock should *not* be held
|
164 |
|
|
on entry. */
|
165 |
|
|
|
166 |
|
|
void
|
167 |
|
|
gomp_ordered_static_next (void)
|
168 |
|
|
{
|
169 |
|
|
struct gomp_thread *thr = gomp_thread ();
|
170 |
|
|
struct gomp_team *team = thr->ts.team;
|
171 |
|
|
struct gomp_work_share *ws = thr->ts.work_share;
|
172 |
|
|
unsigned id = thr->ts.team_id;
|
173 |
|
|
|
174 |
|
|
if (team == NULL || team->nthreads == 1)
|
175 |
|
|
return;
|
176 |
|
|
|
177 |
|
|
ws->ordered_owner = -1;
|
178 |
|
|
|
179 |
|
|
/* This thread currently owns the lock. Increment the owner. */
|
180 |
|
|
if (++id == team->nthreads)
|
181 |
|
|
id = 0;
|
182 |
|
|
ws->ordered_team_ids[0] = id;
|
183 |
|
|
gomp_sem_post (team->ordered_release[id]);
|
184 |
|
|
}
|
185 |
|
|
|
186 |
|
|
/* This function is called when we need to assert that the thread owns the
|
187 |
|
|
ordered section. Due to the problem of posted-but-not-waited semaphores,
|
188 |
|
|
this needs to happen before completing a loop iteration. */
|
189 |
|
|
|
190 |
|
|
void
|
191 |
|
|
gomp_ordered_sync (void)
|
192 |
|
|
{
|
193 |
|
|
struct gomp_thread *thr = gomp_thread ();
|
194 |
|
|
struct gomp_team *team = thr->ts.team;
|
195 |
|
|
struct gomp_work_share *ws = thr->ts.work_share;
|
196 |
|
|
|
197 |
|
|
/* Work share constructs can be orphaned. But this clearly means that
|
198 |
|
|
we are the only thread, and so we automatically own the section. */
|
199 |
|
|
if (team == NULL || team->nthreads == 1)
|
200 |
|
|
return;
|
201 |
|
|
|
202 |
|
|
/* ??? I believe it to be safe to access this data without taking the
|
203 |
|
|
ws->lock. The only presumed race condition is with the previous
|
204 |
|
|
thread on the queue incrementing ordered_cur such that it points
|
205 |
|
|
to us, concurrently with our check below. But our team_id is
|
206 |
|
|
already present in the queue, and the other thread will always
|
207 |
|
|
post to our release semaphore. So the two cases are that we will
|
208 |
|
|
either win the race an momentarily block on the semaphore, or lose
|
209 |
|
|
the race and find the semaphore already unlocked and so not block.
|
210 |
|
|
Either way we get correct results.
|
211 |
|
|
However, there is an implicit flush on entry to an ordered region,
|
212 |
|
|
so we do need to have a barrier here. If we were taking a lock
|
213 |
|
|
this could be MEMMODEL_RELEASE since the acquire would be coverd
|
214 |
|
|
by the lock. */
|
215 |
|
|
|
216 |
|
|
__atomic_thread_fence (MEMMODEL_ACQ_REL);
|
217 |
|
|
if (ws->ordered_owner != thr->ts.team_id)
|
218 |
|
|
{
|
219 |
|
|
gomp_sem_wait (team->ordered_release[thr->ts.team_id]);
|
220 |
|
|
ws->ordered_owner = thr->ts.team_id;
|
221 |
|
|
}
|
222 |
|
|
}
|
223 |
|
|
|
224 |
|
|
/* This function is called by user code when encountering the start of an
|
225 |
|
|
ORDERED block. We must check to see if the current thread is at the
|
226 |
|
|
head of the queue, and if not, block. */
|
227 |
|
|
|
228 |
|
|
#ifdef HAVE_ATTRIBUTE_ALIAS
|
229 |
|
|
extern void GOMP_ordered_start (void)
|
230 |
|
|
__attribute__((alias ("gomp_ordered_sync")));
|
231 |
|
|
#else
|
232 |
|
|
void
|
233 |
|
|
GOMP_ordered_start (void)
|
234 |
|
|
{
|
235 |
|
|
gomp_ordered_sync ();
|
236 |
|
|
}
|
237 |
|
|
#endif
|
238 |
|
|
|
239 |
|
|
/* This function is called by user code when encountering the end of an
|
240 |
|
|
ORDERED block. With the current ORDERED implementation there's nothing
|
241 |
|
|
for us to do.
|
242 |
|
|
|
243 |
|
|
However, the current implementation has a flaw in that it does not allow
|
244 |
|
|
the next thread into the ORDERED section immediately after the current
|
245 |
|
|
thread exits the ORDERED section in its last iteration. The existance
|
246 |
|
|
of this function allows the implementation to change. */
|
247 |
|
|
|
248 |
|
|
void
|
249 |
|
|
GOMP_ordered_end (void)
|
250 |
|
|
{
|
251 |
|
|
}
|