| Line | Branch | Exec | Source |
|---|---|---|---|
| 1 | /* | ||
| 2 | * Copyright (c) 2000-2022 Inria | ||
| 3 | * All rights reserved. | ||
| 4 | * | ||
| 5 | * Redistribution and use in source and binary forms, with or without | ||
| 6 | * modification, are permitted provided that the following conditions are met: | ||
| 7 | * | ||
| 8 | * * Redistributions of source code must retain the above copyright notice, | ||
| 9 | * this list of conditions and the following disclaimer. | ||
| 10 | * * Redistributions in binary form must reproduce the above copyright notice, | ||
| 11 | * this list of conditions and the following disclaimer in the documentation | ||
| 12 | * and/or other materials provided with the distribution. | ||
| 13 | * * Neither the name of the ALICE Project-Team nor the names of its | ||
| 14 | * contributors may be used to endorse or promote products derived from this | ||
| 15 | * software without specific prior written permission. | ||
| 16 | * | ||
| 17 | * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" | ||
| 18 | * AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE | ||
| 19 | * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE | ||
| 20 | * ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE | ||
| 21 | * LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR | ||
| 22 | * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF | ||
| 23 | * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS | ||
| 24 | * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN | ||
| 25 | * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) | ||
| 26 | * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE | ||
| 27 | * POSSIBILITY OF SUCH DAMAGE. | ||
| 28 | * | ||
| 29 | * Contact: Bruno Levy | ||
| 30 | * | ||
| 31 | * https://www.inria.fr/fr/bruno-levy | ||
| 32 | * | ||
| 33 | * Inria, | ||
| 34 | * Domaine de Voluceau, | ||
| 35 | * 78150 Le Chesnay - Rocquencourt | ||
| 36 | * FRANCE | ||
| 37 | * | ||
| 38 | */ | ||
| 39 | |||
| 40 | #ifndef GEOGRAM_BASIC_THREAD_SYNC | ||
| 41 | #define GEOGRAM_BASIC_THREAD_SYNC | ||
| 42 | |||
| 43 | /** | ||
| 44 | * \file geogram/basic/thread_sync.h | ||
| 45 | * \brief Functions and classes for process manipulation | ||
| 46 | */ | ||
| 47 | |||
| 48 | #include <geogram/basic/common.h> | ||
| 49 | #include <geogram/basic/numeric.h> | ||
| 50 | #include <geogram/basic/assert.h> | ||
| 51 | #include <geogram/basic/argused.h> | ||
| 52 | #include <vector> | ||
| 53 | #include <atomic> | ||
| 54 | |||
| 55 | // On Windows/MSCV, we need to use a special implementation | ||
| 56 | // of spinlocks because std::atomic_flag in MSVC's stl does | ||
| 57 | // not fully implement the norm (lacks a constructor). | ||
| 58 | #ifdef GEO_OS_WINDOWS | ||
| 59 | #include <windows.h> | ||
| 60 | #include <intrin.h> | ||
| 61 | #pragma intrinsic(_InterlockedCompareExchange16) | ||
| 62 | #pragma intrinsic(_WriteBarrier) | ||
| 63 | #endif | ||
| 64 | |||
| 65 | // On MacOS, I get many warnings with atomic_flag initialization, | ||
| 66 | // such as std::atomic_flag f = ATOMIC_FLAG_INIT | ||
| 67 | #if defined(__clang__) | ||
| 68 | #pragma GCC diagnostic ignored "-Wbraced-scalar-init" | ||
| 69 | #endif | ||
| 70 | |||
| 71 | /** | ||
| 72 | * \brief executes the pause instruction | ||
| 73 | * \details should be called when a spinlock is spinning | ||
| 74 | */ | ||
| 75 | 122530318 | inline void geo_pause() { | |
| 76 | #ifdef GEO_OS_WINDOWS | ||
| 77 | YieldProcessor(); | ||
| 78 | #else | ||
| 79 | # ifdef GEO_PROCESSOR_X86 | ||
| 80 | # ifdef __ICC | ||
| 81 | _mm_pause(); | ||
| 82 | # else | ||
| 83 | 122530318 | __builtin_ia32_pause(); | |
| 84 | # endif | ||
| 85 | # endif | ||
| 86 | #endif | ||
| 87 | 122530318 | } | |
| 88 | |||
| 89 | /*******************************************************************************/ | ||
| 90 | |||
| 91 | #ifdef GEO_OS_WINDOWS | ||
| 92 | |||
| 93 | // Windows-specific spinlock implementation. | ||
| 94 | // I'd have prefered to use std::atomic_flag for everybody, | ||
| 95 | // unfortunately atomic_flag's constructor is not implemented in MSCV's stl, | ||
| 96 | // so we reimplement them using atomic compare-exchange functions... | ||
| 97 | |||
| 98 | namespace GEO { | ||
| 99 | namespace Process { | ||
| 100 | /** A lightweight synchronization structure. */ | ||
| 101 | typedef short spinlock; | ||
| 102 | |||
| 103 | /** The initialization value of a spin lock. */ | ||
| 104 | # define GEOGRAM_SPINLOCK_INIT 0 | ||
| 105 | inline void acquire_spinlock(volatile spinlock& x) { | ||
| 106 | while(_InterlockedCompareExchange16(&x, 1, 0) == 1) { | ||
| 107 | // Intel recommends to have a PAUSE asm instruction | ||
| 108 | // in the spinlock loop. Under MSVC/Windows, | ||
| 109 | // YieldProcessor() is a macro that calls the | ||
| 110 | // (undocumented) _mm_pause() intrinsic function | ||
| 111 | // that generates a PAUSE opcode. | ||
| 112 | YieldProcessor(); | ||
| 113 | } | ||
| 114 | // We do not need _ReadBarrier() here since | ||
| 115 | // _InterlockedCompareExchange16 | ||
| 116 | // "acts as a full barrier in VC2005" according to the doc | ||
| 117 | } | ||
| 118 | |||
| 119 | inline void release_spinlock(volatile spinlock& x) { | ||
| 120 | _WriteBarrier(); // prevents compiler reordering | ||
| 121 | x = 0; | ||
| 122 | } | ||
| 123 | |||
| 124 | } | ||
| 125 | } | ||
| 126 | |||
| 127 | /*******************************************************************************/ | ||
| 128 | |||
| 129 | #else | ||
| 130 | |||
| 131 | namespace GEO { | ||
| 132 | namespace Process { | ||
| 133 | |||
| 134 | /** The initialization value of a spinlock. */ | ||
| 135 | // Note: C++20 does not need it anymore, in C++20 | ||
| 136 | // std::atomic_flag's constructor initializes it, | ||
| 137 | // we keep it because | ||
| 138 | // - we are using C++17 | ||
| 139 | // - the Windows implementation that uses integers rather than | ||
| 140 | // std::atomic_flag needs an initialization value. | ||
| 141 | #define GEOGRAM_SPINLOCK_INIT ATOMIC_FLAG_INIT | ||
| 142 | |||
| 143 | /** | ||
| 144 | * \brief A lightweight synchronization structure. | ||
| 145 | * \details See | ||
| 146 | * - https://rigtorp.se/spinlock/ | ||
| 147 | * - https://www.sobyte.net/post/2022-06/cpp-memory-order/ | ||
| 148 | */ | ||
| 149 | typedef std::atomic_flag spinlock; | ||
| 150 | |||
| 151 | /** | ||
| 152 | * \brief Loops until \p x is available then reserves it. | ||
| 153 | * \param[in] x a spinlock | ||
| 154 | */ | ||
| 155 | 2121837 | inline void acquire_spinlock(volatile spinlock& x) { | |
| 156 | for (;;) { | ||
| 157 |
2/2✓ Branch 0 taken 2121837 times.
✓ Branch 1 taken 122530150 times.
|
247182137 | if (!x.test_and_set(std::memory_order_acquire)) { |
| 158 | 2121837 | break; | |
| 159 | } | ||
| 160 | // If compiling in C++20 we can be slightly more efficient when spinning | ||
| 161 | // (avoid unrequired atomic operations, just "peek" the flag) | ||
| 162 | #if defined(__cpp_lib_atomic_flag_test) | ||
| 163 | while (x.test(std::memory_order_relaxed)) | ||
| 164 | #endif | ||
| 165 | 122530150 | geo_pause(); | |
| 166 | } | ||
| 167 | 2121837 | } | |
| 168 | |||
| 169 | /** | ||
| 170 | * \brief Makes \p x available to other threads. | ||
| 171 | * \param[in] x a spinlock | ||
| 172 | */ | ||
| 173 | 2121837 | inline void release_spinlock(volatile spinlock& x) { | |
| 174 | x.clear(std::memory_order_release); | ||
| 175 | 2121837 | } | |
| 176 | |||
| 177 | } | ||
| 178 | } | ||
| 179 | #endif | ||
| 180 | |||
| 181 | /****************************************************************************/ | ||
| 182 | |||
| 183 | namespace GEO { | ||
| 184 | namespace Process { | ||
| 185 | |||
| 186 | /** | ||
| 187 | * \brief An array of light-weight synchronisation | ||
| 188 | * primitives (spinlocks). | ||
| 189 | * | ||
| 190 | * \details This is the reference implementation, that uses | ||
| 191 | * an array of spinlock. There is also a more memory-efficient | ||
| 192 | * implementation, CompactSpinLockArray, to be used for a very | ||
| 193 | * large number of spinlocks. | ||
| 194 | * | ||
| 195 | * \see acquire_spinlock(), release_spinlock() | ||
| 196 | */ | ||
| 197 | class BasicSpinLockArray { | ||
| 198 | public: | ||
| 199 | /** | ||
| 200 | * \brief Constructs a new BasicSpinLockArray of size 0. | ||
| 201 | */ | ||
| 202 | BasicSpinLockArray() : spinlocks_(nullptr), size_(0) { | ||
| 203 | } | ||
| 204 | |||
| 205 | /** | ||
| 206 | * \brief Constructs a new BasicSpinLockArray of size \p size_in. | ||
| 207 | * \param[in] size_in number of spinlocks in the array. | ||
| 208 | */ | ||
| 209 | BasicSpinLockArray(index_t size_in) : spinlocks_(nullptr), size_(0) { | ||
| 210 | resize(size_in); | ||
| 211 | } | ||
| 212 | |||
| 213 | /** | ||
| 214 | * \brief Forbids copy | ||
| 215 | */ | ||
| 216 | BasicSpinLockArray(const BasicSpinLockArray& rhs) = delete; | ||
| 217 | |||
| 218 | /** | ||
| 219 | * \brief Forbids copy | ||
| 220 | */ | ||
| 221 | BasicSpinLockArray& operator=( | ||
| 222 | const BasicSpinLockArray& rhs | ||
| 223 | ) = delete; | ||
| 224 | |||
| 225 | /** | ||
| 226 | * \brief Resizes a BasicSpinLockArray. | ||
| 227 | * \details All the spinlocks are reset to 0. | ||
| 228 | * \param[in] size_in The desired new size. | ||
| 229 | */ | ||
| 230 | void resize(index_t size_in) { | ||
| 231 | delete[] spinlocks_; | ||
| 232 | spinlocks_ = new spinlock[size_in]; | ||
| 233 | size_ = size_in; | ||
| 234 | // Need to initialize the spinlocks to false (dirty !) | ||
| 235 | // (maybe use placement new on each item..., to be tested) | ||
| 236 | for(index_t i=0; i<size_; ++i) { | ||
| 237 | Process::release_spinlock(spinlocks_[i]); | ||
| 238 | } | ||
| 239 | } | ||
| 240 | |||
| 241 | /** | ||
| 242 | * \brief Resets size to 0 and clears all the memory. | ||
| 243 | */ | ||
| 244 | void clear() { | ||
| 245 | delete[] spinlocks_; | ||
| 246 | spinlocks_ = nullptr; | ||
| 247 | } | ||
| 248 | |||
| 249 | /** | ||
| 250 | * \brief Gets the number of spinlocks in this array. | ||
| 251 | */ | ||
| 252 | index_t size() const { | ||
| 253 | return size_; | ||
| 254 | } | ||
| 255 | |||
| 256 | /** | ||
| 257 | * \brief Acquires a spinlock at a given index | ||
| 258 | * \details Loops until spinlock at index \p i is available then | ||
| 259 | * reserve it. | ||
| 260 | * \param[in] i index of the spinlock | ||
| 261 | */ | ||
| 262 | void acquire_spinlock(index_t i) { | ||
| 263 | geo_debug_assert(i < size()); | ||
| 264 | GEO::Process::acquire_spinlock(spinlocks_[i]); | ||
| 265 | } | ||
| 266 | |||
| 267 | /** | ||
| 268 | * \brief Releases a spinlock at a given index | ||
| 269 | * \details Makes spinlock at index \p i available to other threads. | ||
| 270 | * \param[in] i index of the spinlock | ||
| 271 | */ | ||
| 272 | void release_spinlock(index_t i) { | ||
| 273 | geo_debug_assert(i < size()); | ||
| 274 | GEO::Process::release_spinlock(spinlocks_[i]); | ||
| 275 | } | ||
| 276 | |||
| 277 | private: | ||
| 278 | // Cannot use a std::vector because std::atomic_flag does not | ||
| 279 | // have copy ctor nor assignment operator. | ||
| 280 | spinlock* spinlocks_; | ||
| 281 | index_t size_; | ||
| 282 | }; | ||
| 283 | } | ||
| 284 | } | ||
| 285 | |||
| 286 | /*******************************************************************************/ | ||
| 287 | |||
| 288 | namespace GEO { | ||
| 289 | namespace Process { | ||
| 290 | |||
| 291 | /** | ||
| 292 | * \brief An array of light-weight synchronisation | ||
| 293 | * primitives (spinlocks). | ||
| 294 | * | ||
| 295 | * \details In this implementation, storage is optimized so that | ||
| 296 | * a single bit per spinlock is used. This implementation uses | ||
| 297 | * std::atomic<uint32_t> | ||
| 298 | * | ||
| 299 | * \see acquire_spinlock(), release_spinlock() | ||
| 300 | */ | ||
| 301 | class CompactSpinLockArray { | ||
| 302 | public: | ||
| 303 | /** | ||
| 304 | * \brief Constructs a new SpinLockArray of size 0. | ||
| 305 | */ | ||
| 306 | 161 | CompactSpinLockArray() : spinlocks_(nullptr), size_(0) { | |
| 307 | 161 | } | |
| 308 | |||
| 309 | /** | ||
| 310 | * \brief Constructs a new CompactSpinLockArray of size \p size_in. | ||
| 311 | * \param[in] size_in number of spinlocks in the array. | ||
| 312 | */ | ||
| 313 | CompactSpinLockArray(index_t size_in) : spinlocks_(nullptr),size_(0){ | ||
| 314 | resize(size_in); | ||
| 315 | } | ||
| 316 | |||
| 317 | /** | ||
| 318 | * \brief CompactSpinLockArray destructor | ||
| 319 | */ | ||
| 320 | 161 | ~CompactSpinLockArray() { | |
| 321 | 161 | clear(); | |
| 322 | 161 | } | |
| 323 | |||
| 324 | /** | ||
| 325 | * \brief Forbids copy | ||
| 326 | */ | ||
| 327 | CompactSpinLockArray(const CompactSpinLockArray& rhs) = delete; | ||
| 328 | |||
| 329 | /** | ||
| 330 | * \brief Forbids copy | ||
| 331 | */ | ||
| 332 | CompactSpinLockArray& operator=( | ||
| 333 | const CompactSpinLockArray& rhs | ||
| 334 | ) = delete; | ||
| 335 | |||
| 336 | /** | ||
| 337 | * \brief Resizes a CompactSpinLockArray. | ||
| 338 | * \details All the spinlocks are reset to 0. | ||
| 339 | * \param[in] size_in The desired new size. | ||
| 340 | */ | ||
| 341 | 531 | void resize(index_t size_in) { | |
| 342 |
2/2✓ Branch 0 taken 75 times.
✓ Branch 1 taken 456 times.
|
531 | if(size_ != size_in) { |
| 343 | 75 | size_ = size_in; | |
| 344 | 75 | index_t nb_words = (size_ >> 5) + 1; | |
| 345 |
1/2✗ Branch 0 not taken.
✓ Branch 1 taken 75 times.
|
75 | delete[] spinlocks_; |
| 346 | // Each entry is initialized with 0 -------------v | ||
| 347 |
4/6✓ Branch 0 taken 75 times.
✗ Branch 1 not taken.
✓ Branch 4 taken 3181 times.
✓ Branch 5 taken 75 times.
✗ Branch 6 not taken.
✓ Branch 7 taken 75 times.
|
3256 | spinlocks_ = new std::atomic<uint32_t>[nb_words]{}; |
| 348 | } | ||
| 349 | |||
| 350 | // Test at compile time that we are using atomic | ||
| 351 | // uint32_t operations (and not using an additional | ||
| 352 | // lock which would be catastrophic in terms of | ||
| 353 | // performance) | ||
| 354 | static_assert(std::atomic<uint32_t>::is_always_lock_free); | ||
| 355 | 531 | } | |
| 356 | |||
| 357 | /** | ||
| 358 | * \brief Gets the number of spinlocks in this array. | ||
| 359 | */ | ||
| 360 | 36217314 | index_t size() const { | |
| 361 | 36217314 | return size_; | |
| 362 | } | ||
| 363 | |||
| 364 | /** | ||
| 365 | * \brief Resets size to 0 and clears all the memory. | ||
| 366 | */ | ||
| 367 | 161 | void clear() { | |
| 368 |
2/2✓ Branch 0 taken 75 times.
✓ Branch 1 taken 86 times.
|
161 | delete[] spinlocks_; |
| 369 | 161 | size_ = 0; | |
| 370 | 161 | } | |
| 371 | |||
| 372 | /** | ||
| 373 | * \brief Acquires a spinlock at a given index | ||
| 374 | * \details Loops until spinlock at index \p i is available then | ||
| 375 | * reserve it. | ||
| 376 | * \param[in] i index of the spinlock | ||
| 377 | */ | ||
| 378 | 18108657 | void acquire_spinlock(index_t i) { | |
| 379 |
1/6✗ Branch 1 not taken.
✓ Branch 2 taken 18108657 times.
✗ Branch 4 not taken.
✗ Branch 5 not taken.
✗ Branch 7 not taken.
✗ Branch 8 not taken.
|
18108657 | geo_debug_assert(i < size()); |
| 380 | 18108657 | index_t w = i >> 5; | |
| 381 | 18108657 | uint32_t b = uint32_t(i & 31); | |
| 382 | 18108657 | uint32_t mask = (1u << b); | |
| 383 | 18108657 | while( | |
| 384 | 18108825 | (spinlocks_[w].fetch_or( | |
| 385 | mask, std::memory_order_acquire | ||
| 386 |
2/2✓ Branch 0 taken 168 times.
✓ Branch 1 taken 18108657 times.
|
18108825 | ) & mask) != 0 |
| 387 | ) { | ||
| 388 | 168 | geo_pause(); | |
| 389 | } | ||
| 390 | 18108657 | } | |
| 391 | |||
| 392 | /** | ||
| 393 | * \brief Releases a spinlock at a given index | ||
| 394 | * \details Makes spinlock at index \p i available to other threads. | ||
| 395 | * \param[in] i index of the spinlock | ||
| 396 | */ | ||
| 397 | 18108657 | void release_spinlock(index_t i) { | |
| 398 |
1/6✗ Branch 1 not taken.
✓ Branch 2 taken 18108657 times.
✗ Branch 4 not taken.
✗ Branch 5 not taken.
✗ Branch 7 not taken.
✗ Branch 8 not taken.
|
18108657 | geo_debug_assert(i < size()); |
| 399 | 18108657 | index_t w = i >> 5; | |
| 400 | 18108657 | uint32_t b = uint32_t(i & 31); | |
| 401 | 18108657 | uint32_t mask = ~(1u << b); | |
| 402 | 18108657 | spinlocks_[w].fetch_and(mask, std::memory_order_release); | |
| 403 | 18108657 | } | |
| 404 | |||
| 405 | private: | ||
| 406 | // Cannot use a std::vector because std::atomic<> does not | ||
| 407 | // have copy ctor nor assignment operator. | ||
| 408 | std::atomic<uint32_t>* spinlocks_; | ||
| 409 | index_t size_; | ||
| 410 | }; | ||
| 411 | |||
| 412 | } | ||
| 413 | } | ||
| 414 | |||
| 415 | /*******************************************************************************/ | ||
| 416 | |||
| 417 | namespace GEO { | ||
| 418 | namespace Process { | ||
| 419 | typedef CompactSpinLockArray SpinLockArray; | ||
| 420 | } | ||
| 421 | } | ||
| 422 | |||
| 423 | /*******************************************************************************/ | ||
| 424 | |||
| 425 | #endif | ||
| 426 |