libstdc++
barrier
Go to the documentation of this file.
1// <barrier> -*- C++ -*-
2
3// Copyright (C) 2020-2021 Free Software Foundation, Inc.
4//
5// This file is part of the GNU ISO C++ Library. This library is free
6// software; you can redistribute it and/or modify it under the
7// terms of the GNU General Public License as published by the
8// Free Software Foundation; either version 3, or (at your option)
9// any later version.
10
11// This library is distributed in the hope that it will be useful,
12// but WITHOUT ANY WARRANTY; without even the implied warranty of
13// MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
14// GNU General Public License for 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 implementation is based on libcxx/include/barrier
26//===-- barrier.h --------------------------------------------------===//
27//
28// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
29// See https://llvm.org/LICENSE.txt for license information.
30// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
31//
32//===---------------------------------------------------------------===//
33
34/** @file include/barrier
35 * This is a Standard C++ Library header.
36 */
37
38#ifndef _GLIBCXX_BARRIER
39#define _GLIBCXX_BARRIER 1
40
41#pragma GCC system_header
42
43#if __cplusplus > 201703L
44#include <bits/atomic_base.h>
45#if __cpp_lib_atomic_wait && __cpp_aligned_new
46#include <bits/std_thread.h>
47#include <bits/unique_ptr.h>
48
49#include <array>
50
51#define __cpp_lib_barrier 201907L
52
53namespace std _GLIBCXX_VISIBILITY(default)
54{
55_GLIBCXX_BEGIN_NAMESPACE_VERSION
56
57 struct __empty_completion
58 {
59 _GLIBCXX_ALWAYS_INLINE void
60 operator()() noexcept
61 { }
62 };
63
64/*
65
66The default implementation of __tree_barrier is a classic tree barrier.
67
68It looks different from literature pseudocode for two main reasons:
69 1. Threads that call into std::barrier functions do not provide indices,
70 so a numbering step is added before the actual barrier algorithm,
71 appearing as an N+1 round to the N rounds of the tree barrier.
72 2. A great deal of attention has been paid to avoid cache line thrashing
73 by flattening the tree structure into cache-line sized arrays, that
74 are indexed in an efficient way.
75
76*/
77
78 enum class __barrier_phase_t : unsigned char { };
79
80 template<typename _CompletionF>
81 class __tree_barrier
82 {
83 using __atomic_phase_ref_t = std::__atomic_ref<__barrier_phase_t>;
84 using __atomic_phase_const_ref_t = std::__atomic_ref<const __barrier_phase_t>;
85 static constexpr auto __phase_alignment =
86 __atomic_phase_ref_t::required_alignment;
87
88 using __tickets_t = std::array<__barrier_phase_t, 64>;
89 struct alignas(64) /* naturally-align the heap state */ __state_t
90 {
91 alignas(__phase_alignment) __tickets_t __tickets;
92 };
93
94 ptrdiff_t _M_expected;
95 unique_ptr<__state_t[]> _M_state;
96 __atomic_base<ptrdiff_t> _M_expected_adjustment;
97 _CompletionF _M_completion;
98
99 alignas(__phase_alignment) __barrier_phase_t _M_phase;
100
101 bool
102 _M_arrive(__barrier_phase_t __old_phase, size_t __current)
103 {
104 const auto __old_phase_val = static_cast<unsigned char>(__old_phase);
105 const auto __half_step =
106 static_cast<__barrier_phase_t>(__old_phase_val + 1);
107 const auto __full_step =
108 static_cast<__barrier_phase_t>(__old_phase_val + 2);
109
110 size_t __current_expected = _M_expected;
111 __current %= ((_M_expected + 1) >> 1);
112
113 for (int __round = 0; ; ++__round)
114 {
115 if (__current_expected <= 1)
116 return true;
117 size_t const __end_node = ((__current_expected + 1) >> 1),
118 __last_node = __end_node - 1;
119 for ( ; ; ++__current)
120 {
121 if (__current == __end_node)
122 __current = 0;
123 auto __expect = __old_phase;
124 __atomic_phase_ref_t __phase(_M_state[__current]
125 .__tickets[__round]);
126 if (__current == __last_node && (__current_expected & 1))
127 {
128 if (__phase.compare_exchange_strong(__expect, __full_step,
129 memory_order_acq_rel))
130 break; // I'm 1 in 1, go to next __round
131 }
132 else if (__phase.compare_exchange_strong(__expect, __half_step,
133 memory_order_acq_rel))
134 {
135 return false; // I'm 1 in 2, done with arrival
136 }
137 else if (__expect == __half_step)
138 {
139 if (__phase.compare_exchange_strong(__expect, __full_step,
140 memory_order_acq_rel))
141 break; // I'm 2 in 2, go to next __round
142 }
143 }
144 __current_expected = __last_node + 1;
145 __current >>= 1;
146 }
147 }
148
149 public:
150 using arrival_token = __barrier_phase_t;
151
152 static constexpr ptrdiff_t
153 max() noexcept
154 { return __PTRDIFF_MAX__; }
155
156 __tree_barrier(ptrdiff_t __expected, _CompletionF __completion)
157 : _M_expected(__expected), _M_expected_adjustment(0),
158 _M_completion(move(__completion)),
159 _M_phase(static_cast<__barrier_phase_t>(0))
160 {
161 size_t const __count = (_M_expected + 1) >> 1;
162
163 _M_state = std::make_unique<__state_t[]>(__count);
164 }
165
166 [[nodiscard]] arrival_token
167 arrive(ptrdiff_t __update)
168 {
169 std::hash<std::thread::id> __hasher;
170 size_t __current = __hasher(std::this_thread::get_id());
171 __atomic_phase_ref_t __phase(_M_phase);
172 const auto __old_phase = __phase.load(memory_order_relaxed);
173 const auto __cur = static_cast<unsigned char>(__old_phase);
174 for(; __update; --__update)
175 {
176 if(_M_arrive(__old_phase, __current))
177 {
178 _M_completion();
179 _M_expected += _M_expected_adjustment.load(memory_order_relaxed);
180 _M_expected_adjustment.store(0, memory_order_relaxed);
181 auto __new_phase = static_cast<__barrier_phase_t>(__cur + 2);
182 __phase.store(__new_phase, memory_order_release);
183 __phase.notify_all();
184 }
185 }
186 return __old_phase;
187 }
188
189 void
190 wait(arrival_token&& __old_phase) const
191 {
192 __atomic_phase_const_ref_t __phase(_M_phase);
193 auto const __test_fn = [=]
194 {
195 return __phase.load(memory_order_acquire) != __old_phase;
196 };
197 std::__atomic_wait_address(&_M_phase, __test_fn);
198 }
199
200 void
201 arrive_and_drop()
202 {
203 _M_expected_adjustment.fetch_sub(1, memory_order_relaxed);
204 (void)arrive(1);
205 }
206 };
207
208 template<typename _CompletionF = __empty_completion>
209 class barrier
210 {
211 // Note, we may introduce a "central" barrier algorithm at some point
212 // for more space constrained targets
213 using __algorithm_t = __tree_barrier<_CompletionF>;
214 __algorithm_t _M_b;
215
216 public:
217 class arrival_token final
218 {
219 public:
220 arrival_token(arrival_token&&) = default;
221 arrival_token& operator=(arrival_token&&) = default;
222 ~arrival_token() = default;
223
224 private:
225 friend class barrier;
226 using __token = typename __algorithm_t::arrival_token;
227 explicit arrival_token(__token __tok) noexcept : _M_tok(__tok) { }
228 __token _M_tok;
229 };
230
231 static constexpr ptrdiff_t
232 max() noexcept
233 { return __algorithm_t::max(); }
234
235 explicit
236 barrier(ptrdiff_t __count, _CompletionF __completion = _CompletionF())
237 : _M_b(__count, std::move(__completion))
238 { }
239
240 barrier(barrier const&) = delete;
241 barrier& operator=(barrier const&) = delete;
242
243 [[nodiscard]] arrival_token
244 arrive(ptrdiff_t __update = 1)
245 { return arrival_token{_M_b.arrive(__update)}; }
246
247 void
248 wait(arrival_token&& __phase) const
249 { _M_b.wait(std::move(__phase._M_tok)); }
250
251 void
252 arrive_and_wait()
253 { wait(arrive()); }
254
255 void
256 arrive_and_drop()
257 { _M_b.arrive_and_drop(); }
258 };
259
260_GLIBCXX_END_NAMESPACE_VERSION
261} // namespace
262#endif // __cpp_lib_atomic_wait && __cpp_aligned_new
263#endif // __cplusplus > 201703L
264#endif // _GLIBCXX_BARRIER