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 
53 namespace 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 
66 The default implementation of __tree_barrier is a classic tree barrier.
67 
68 It 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