| 1 | // Licensed to the .NET Foundation under one or more agreements. |
| 2 | // The .NET Foundation licenses this file to you under the MIT license. |
| 3 | // See the LICENSE file in the project root for more information. |
| 4 | // |
| 5 | // Volatile.h |
| 6 | // |
| 7 | |
| 8 | // |
| 9 | // Defines the Volatile<T> type, which provides uniform volatile-ness on |
| 10 | // Visual C++ and GNU C++. |
| 11 | // |
| 12 | // Visual C++ treats accesses to volatile variables as follows: no read or write |
| 13 | // can be removed by the compiler, no global memory access can be moved backwards past |
| 14 | // a volatile read, and no global memory access can be moved forward past a volatile |
| 15 | // write. |
| 16 | // |
| 17 | // The GCC volatile semantic is straight out of the C standard: the compiler is not |
| 18 | // allowed to remove accesses to volatile variables, and it is not allowed to reorder |
| 19 | // volatile accesses relative to other volatile accesses. It is allowed to freely |
| 20 | // reorder non-volatile accesses relative to volatile accesses. |
| 21 | // |
| 22 | // We have lots of code that assumes that ordering of non-volatile accesses will be |
| 23 | // constrained relative to volatile accesses. For example, this pattern appears all |
| 24 | // over the place: |
| 25 | // |
| 26 | // static volatile int lock = 0; |
| 27 | // |
| 28 | // while (InterlockedCompareExchange(&lock, 0, 1)) |
| 29 | // { |
| 30 | // //spin |
| 31 | // } |
| 32 | // |
| 33 | // //read and write variables protected by the lock |
| 34 | // |
| 35 | // lock = 0; |
| 36 | // |
| 37 | // This depends on the reads and writes in the critical section not moving past the |
| 38 | // final statement, which releases the lock. If this should happen, then you have an |
| 39 | // unintended race. |
| 40 | // |
| 41 | // The solution is to ban the use of the "volatile" keyword, and instead define our |
| 42 | // own type Volatile<T>, which acts like a variable of type T except that accesses to |
| 43 | // the variable are always given VC++'s volatile semantics. |
| 44 | // |
| 45 | // (NOTE: The code above is not intended to be an example of how a spinlock should be |
| 46 | // implemented; it has many flaws, and should not be used. This code is intended only |
| 47 | // to illustrate where we might get into trouble with GCC's volatile semantics.) |
| 48 | // |
| 49 | // @TODO: many of the variables marked volatile in the CLR do not actually need to be |
| 50 | // volatile. For example, if a variable is just always passed to Interlocked functions |
| 51 | // (such as a refcount variable), there is no need for it to be volatile. A future |
| 52 | // cleanup task should be to examine each volatile variable and make them non-volatile |
| 53 | // if possible. |
| 54 | // |
| 55 | // @TODO: link to a "Memory Models for CLR Devs" doc here (this doc does not yet exist). |
| 56 | // |
| 57 | |
| 58 | #ifndef _VOLATILE_H_ |
| 59 | #define _VOLATILE_H_ |
| 60 | |
| 61 | // |
| 62 | // This code is extremely compiler- and CPU-specific, and will need to be altered to |
| 63 | // support new compilers and/or CPUs. Here we enforce that we can only compile using |
| 64 | // VC++, or GCC on x86 or AMD64. |
| 65 | // |
| 66 | #if !defined(_MSC_VER) && !defined(__GNUC__) |
| 67 | #error The Volatile type is currently only defined for Visual C++ and GNU C++ |
| 68 | #endif |
| 69 | |
| 70 | #if defined(__GNUC__) && !defined(_X86_) && !defined(_AMD64_) && !defined(_ARM_) && !defined(_ARM64_) |
| 71 | #error The Volatile type is currently only defined for GCC when targeting x86, AMD64, ARM or ARM64 CPUs |
| 72 | #endif |
| 73 | |
| 74 | #if defined(__GNUC__) |
| 75 | #if defined(_ARM_) || defined(_ARM64_) |
| 76 | // This is functionally equivalent to the MemoryBarrier() macro used on ARM on Windows. |
| 77 | #define VOLATILE_MEMORY_BARRIER() asm volatile ("dmb ish" : : : "memory") |
| 78 | #else |
| 79 | // |
| 80 | // For GCC, we prevent reordering by the compiler by inserting the following after a volatile |
| 81 | // load (to prevent subsequent operations from moving before the read), and before a volatile |
| 82 | // write (to prevent prior operations from moving past the write). We don't need to do anything |
| 83 | // special to prevent CPU reorderings, because the x86 and AMD64 architectures are already |
| 84 | // sufficiently constrained for our purposes. If we ever need to run on weaker CPU architectures |
| 85 | // (such as PowerPC), then we will need to do more work. |
| 86 | // |
| 87 | // Please do not use this macro outside of this file. It is subject to change or removal without |
| 88 | // notice. |
| 89 | // |
| 90 | #define VOLATILE_MEMORY_BARRIER() asm volatile ("" : : : "memory") |
| 91 | #endif // _ARM_ || _ARM64_ |
| 92 | #elif (defined(_ARM_) || defined(_ARM64_)) && _ISO_VOLATILE |
| 93 | // ARM & ARM64 have a very weak memory model and very few tools to control that model. We're forced to perform a full |
| 94 | // memory barrier to preserve the volatile semantics. Technically this is only necessary on MP systems but we |
| 95 | // currently don't have a cheap way to determine the number of CPUs from this header file. Revisit this if it |
| 96 | // turns out to be a performance issue for the uni-proc case. |
| 97 | #define VOLATILE_MEMORY_BARRIER() MemoryBarrier() |
| 98 | #else |
| 99 | // |
| 100 | // On VC++, reorderings at the compiler and machine level are prevented by the use of the |
| 101 | // "volatile" keyword in VolatileLoad and VolatileStore. This should work on any CPU architecture |
| 102 | // targeted by VC++ with /iso_volatile-. |
| 103 | // |
| 104 | #define VOLATILE_MEMORY_BARRIER() |
| 105 | #endif // __GNUC__ |
| 106 | |
| 107 | template<typename T> |
| 108 | struct RemoveVolatile |
| 109 | { |
| 110 | typedef T type; |
| 111 | }; |
| 112 | |
| 113 | template<typename T> |
| 114 | struct RemoveVolatile<volatile T> |
| 115 | { |
| 116 | typedef T type; |
| 117 | }; |
| 118 | |
| 119 | |
| 120 | // |
| 121 | // VolatileLoad loads a T from a pointer to T. It is guaranteed that this load will not be optimized |
| 122 | // away by the compiler, and that any operation that occurs after this load, in program order, will |
| 123 | // not be moved before this load. In general it is not guaranteed that the load will be atomic, though |
| 124 | // this is the case for most aligned scalar data types. If you need atomic loads or stores, you need |
| 125 | // to consult the compiler and CPU manuals to find which circumstances allow atomicity. |
| 126 | // |
| 127 | // Starting at version 3.8, clang errors out on initializing of type int * to volatile int *. To fix this, we add two templates to cast away volatility |
| 128 | // Helper structures for casting away volatileness |
| 129 | |
| 130 | |
| 131 | template<typename T> |
| 132 | inline |
| 133 | T VolatileLoad(T const * pt) |
| 134 | { |
| 135 | #ifndef DACCESS_COMPILE |
| 136 | #if defined(_ARM64_) && defined(__GNUC__) |
| 137 | T val; |
| 138 | static const unsigned lockFreeAtomicSizeMask = (1 << 1) | (1 << 2) | (1 << 4) | (1 << 8); |
| 139 | if((1 << sizeof(T)) & lockFreeAtomicSizeMask) |
| 140 | { |
| 141 | __atomic_load((T const *)pt, const_cast<typename RemoveVolatile<T>::type *>(&val), __ATOMIC_ACQUIRE); |
| 142 | } |
| 143 | else |
| 144 | { |
| 145 | val = *(T volatile const *)pt; |
| 146 | asm volatile ("dmb ishld" : : : "memory" ); |
| 147 | } |
| 148 | #else |
| 149 | T val = *(T volatile const *)pt; |
| 150 | VOLATILE_MEMORY_BARRIER(); |
| 151 | #endif |
| 152 | #else |
| 153 | T val = *pt; |
| 154 | #endif |
| 155 | return val; |
| 156 | } |
| 157 | |
| 158 | template<typename T> |
| 159 | inline |
| 160 | T VolatileLoadWithoutBarrier(T const * pt) |
| 161 | { |
| 162 | #ifndef DACCESS_COMPILE |
| 163 | T val = *(T volatile const *)pt; |
| 164 | #else |
| 165 | T val = *pt; |
| 166 | #endif |
| 167 | return val; |
| 168 | } |
| 169 | |
| 170 | template <typename T> class Volatile; |
| 171 | |
| 172 | template<typename T> |
| 173 | inline |
| 174 | T VolatileLoad(Volatile<T> const * pt) |
| 175 | { |
| 176 | return pt->Load(); |
| 177 | } |
| 178 | |
| 179 | // |
| 180 | // VolatileStore stores a T into the target of a pointer to T. Is is guaranteed that this store will |
| 181 | // not be optimized away by the compiler, and that any operation that occurs before this store, in program |
| 182 | // order, will not be moved after this store. In general, it is not guaranteed that the store will be |
| 183 | // atomic, though this is the case for most aligned scalar data types. If you need atomic loads or stores, |
| 184 | // you need to consult the compiler and CPU manuals to find which circumstances allow atomicity. |
| 185 | // |
| 186 | template<typename T> |
| 187 | inline |
| 188 | void VolatileStore(T* pt, T val) |
| 189 | { |
| 190 | #ifndef DACCESS_COMPILE |
| 191 | #if defined(_ARM64_) && defined(__GNUC__) |
| 192 | static const unsigned lockFreeAtomicSizeMask = (1 << 1) | (1 << 2) | (1 << 4) | (1 << 8); |
| 193 | if((1 << sizeof(T)) & lockFreeAtomicSizeMask) |
| 194 | { |
| 195 | __atomic_store((T volatile *)pt, &val, __ATOMIC_RELEASE); |
| 196 | } |
| 197 | else |
| 198 | { |
| 199 | VOLATILE_MEMORY_BARRIER(); |
| 200 | *(T volatile *)pt = val; |
| 201 | } |
| 202 | #else |
| 203 | VOLATILE_MEMORY_BARRIER(); |
| 204 | *(T volatile *)pt = val; |
| 205 | #endif |
| 206 | #else |
| 207 | *pt = val; |
| 208 | #endif |
| 209 | } |
| 210 | |
| 211 | template<typename T> |
| 212 | inline |
| 213 | void VolatileStoreWithoutBarrier(T* pt, T val) |
| 214 | { |
| 215 | #ifndef DACCESS_COMPILE |
| 216 | *(T volatile *)pt = val; |
| 217 | #else |
| 218 | *pt = val; |
| 219 | #endif |
| 220 | } |
| 221 | |
| 222 | // |
| 223 | // Volatile<T> implements accesses with our volatile semantics over a variable of type T. |
| 224 | // Wherever you would have used a "volatile Foo" or, equivalently, "Foo volatile", use Volatile<Foo> |
| 225 | // instead. If Foo is a pointer type, use VolatilePtr. |
| 226 | // |
| 227 | // Note that there are still some things that don't work with a Volatile<T>, |
| 228 | // that would have worked with a "volatile T". For example, you can't cast a Volatile<int> to a float. |
| 229 | // You must instead cast to an int, then to a float. Or you can call Load on the Volatile<int>, and |
| 230 | // cast the result to a float. In general, calling Load or Store explicitly will work around |
| 231 | // any problems that can't be solved by operator overloading. |
| 232 | // |
| 233 | // @TODO: it's not clear that we actually *want* any operator overloading here. It's in here primarily |
| 234 | // to ease the task of converting all of the old uses of the volatile keyword, but in the long |
| 235 | // run it's probably better if users of this class are forced to call Load() and Store() explicitly. |
| 236 | // This would make it much more clear where the memory barriers are, and which operations are actually |
| 237 | // being performed, but it will have to wait for another cleanup effort. |
| 238 | // |
| 239 | template <typename T> |
| 240 | class Volatile |
| 241 | { |
| 242 | private: |
| 243 | // |
| 244 | // The data which we are treating as volatile |
| 245 | // |
| 246 | T m_val; |
| 247 | |
| 248 | public: |
| 249 | // |
| 250 | // Default constructor. Results in an unitialized value! |
| 251 | // |
| 252 | inline Volatile() |
| 253 | { |
| 254 | } |
| 255 | |
| 256 | // |
| 257 | // Allow initialization of Volatile<T> from a T |
| 258 | // |
| 259 | inline Volatile(const T& val) |
| 260 | { |
| 261 | ((volatile T &)m_val) = val; |
| 262 | } |
| 263 | |
| 264 | // |
| 265 | // Copy constructor |
| 266 | // |
| 267 | inline Volatile(const Volatile<T>& other) |
| 268 | { |
| 269 | ((volatile T &)m_val) = other.Load(); |
| 270 | } |
| 271 | |
| 272 | // |
| 273 | // Loads the value of the volatile variable. See code:VolatileLoad for the semantics of this operation. |
| 274 | // |
| 275 | inline T Load() const |
| 276 | { |
| 277 | return VolatileLoad(&m_val); |
| 278 | } |
| 279 | |
| 280 | // |
| 281 | // Loads the value of the volatile variable atomically without erecting the memory barrier. |
| 282 | // |
| 283 | inline T LoadWithoutBarrier() const |
| 284 | { |
| 285 | return ((volatile T &)m_val); |
| 286 | } |
| 287 | |
| 288 | // |
| 289 | // Stores a new value to the volatile variable. See code:VolatileStore for the semantics of this |
| 290 | // operation. |
| 291 | // |
| 292 | inline void Store(const T& val) |
| 293 | { |
| 294 | VolatileStore(&m_val, val); |
| 295 | } |
| 296 | |
| 297 | |
| 298 | // |
| 299 | // Stores a new value to the volatile variable atomically without erecting the memory barrier. |
| 300 | // |
| 301 | inline void StoreWithoutBarrier(const T& val) const |
| 302 | { |
| 303 | ((volatile T &)m_val) = val; |
| 304 | } |
| 305 | |
| 306 | |
| 307 | // |
| 308 | // Gets a pointer to the volatile variable. This is dangerous, as it permits the variable to be |
| 309 | // accessed without using Load and Store, but it is necessary for passing Volatile<T> to APIs like |
| 310 | // InterlockedIncrement. |
| 311 | // |
| 312 | inline volatile T* GetPointer() { return (volatile T*)&m_val; } |
| 313 | |
| 314 | |
| 315 | // |
| 316 | // Gets the raw value of the variable. This is dangerous, as it permits the variable to be |
| 317 | // accessed without using Load and Store |
| 318 | // |
| 319 | inline T& RawValue() { return m_val; } |
| 320 | |
| 321 | // |
| 322 | // Allow casts from Volatile<T> to T. Note that this allows implicit casts, so you can |
| 323 | // pass a Volatile<T> directly to a method that expects a T. |
| 324 | // |
| 325 | inline operator T() const |
| 326 | { |
| 327 | return this->Load(); |
| 328 | } |
| 329 | |
| 330 | // |
| 331 | // Assignment from T |
| 332 | // |
| 333 | inline Volatile<T>& operator=(T val) {Store(val); return *this;} |
| 334 | |
| 335 | // |
| 336 | // Get the address of the volatile variable. This is dangerous, as it allows the value of the |
| 337 | // volatile variable to be accessed directly, without going through Load and Store, but it is |
| 338 | // necessary for passing Volatile<T> to APIs like InterlockedIncrement. Note that we are returning |
| 339 | // a pointer to a volatile T here, so we cannot accidentally pass this pointer to an API that |
| 340 | // expects a normal pointer. |
| 341 | // |
| 342 | inline T volatile * operator&() {return this->GetPointer();} |
| 343 | inline T volatile const * operator&() const {return this->GetPointer();} |
| 344 | |
| 345 | // |
| 346 | // Comparison operators |
| 347 | // |
| 348 | template<typename TOther> |
| 349 | inline bool operator==(const TOther& other) const {return this->Load() == other;} |
| 350 | |
| 351 | template<typename TOther> |
| 352 | inline bool operator!=(const TOther& other) const {return this->Load() != other;} |
| 353 | |
| 354 | // |
| 355 | // Miscellaneous operators. Add more as necessary. |
| 356 | // |
| 357 | inline Volatile<T>& operator+=(T val) {Store(this->Load() + val); return *this;} |
| 358 | inline Volatile<T>& operator-=(T val) {Store(this->Load() - val); return *this;} |
| 359 | inline Volatile<T>& operator|=(T val) {Store(this->Load() | val); return *this;} |
| 360 | inline Volatile<T>& operator&=(T val) {Store(this->Load() & val); return *this;} |
| 361 | inline bool operator!() const { return !this->Load();} |
| 362 | |
| 363 | // |
| 364 | // Prefix increment |
| 365 | // |
| 366 | inline Volatile& operator++() {this->Store(this->Load()+1); return *this;} |
| 367 | |
| 368 | // |
| 369 | // Postfix increment |
| 370 | // |
| 371 | inline T operator++(int) {T val = this->Load(); this->Store(val+1); return val;} |
| 372 | |
| 373 | // |
| 374 | // Prefix decrement |
| 375 | // |
| 376 | inline Volatile& operator--() {this->Store(this->Load()-1); return *this;} |
| 377 | |
| 378 | // |
| 379 | // Postfix decrement |
| 380 | // |
| 381 | inline T operator--(int) {T val = this->Load(); this->Store(val-1); return val;} |
| 382 | }; |
| 383 | |
| 384 | // |
| 385 | // A VolatilePtr builds on Volatile<T> by adding operators appropriate to pointers. |
| 386 | // Wherever you would have used "Foo * volatile", use "VolatilePtr<Foo>" instead. |
| 387 | // |
| 388 | // VolatilePtr also allows the substution of other types for the underlying pointer. This |
| 389 | // allows you to wrap a VolatilePtr around a custom type that looks like a pointer. For example, |
| 390 | // if what you want is a "volatile DPTR<Foo>", use "VolatilePtr<Foo, DPTR<Foo>>". |
| 391 | // |
| 392 | template <typename T, typename P = T*> |
| 393 | class VolatilePtr : public Volatile<P> |
| 394 | { |
| 395 | public: |
| 396 | // |
| 397 | // Default constructor. Results in an uninitialized pointer! |
| 398 | // |
| 399 | inline VolatilePtr() |
| 400 | { |
| 401 | } |
| 402 | |
| 403 | // |
| 404 | // Allow assignment from the pointer type. |
| 405 | // |
| 406 | inline VolatilePtr(P val) : Volatile<P>(val) |
| 407 | { |
| 408 | } |
| 409 | |
| 410 | // |
| 411 | // Copy constructor |
| 412 | // |
| 413 | inline VolatilePtr(const VolatilePtr& other) : Volatile<P>(other) |
| 414 | { |
| 415 | } |
| 416 | |
| 417 | // |
| 418 | // Cast to the pointer type |
| 419 | // |
| 420 | inline operator P() const |
| 421 | { |
| 422 | return (P)this->Load(); |
| 423 | } |
| 424 | |
| 425 | // |
| 426 | // Member access |
| 427 | // |
| 428 | inline P operator->() const |
| 429 | { |
| 430 | return (P)this->Load(); |
| 431 | } |
| 432 | |
| 433 | // |
| 434 | // Dereference the pointer |
| 435 | // |
| 436 | inline T& operator*() const |
| 437 | { |
| 438 | return *(P)this->Load(); |
| 439 | } |
| 440 | |
| 441 | // |
| 442 | // Access the pointer as an array |
| 443 | // |
| 444 | template <typename TIndex> |
| 445 | inline T& operator[](TIndex index) |
| 446 | { |
| 447 | return ((P)this->Load())[index]; |
| 448 | } |
| 449 | }; |
| 450 | |
| 451 | #define VOLATILE(T) Volatile<T> |
| 452 | |
| 453 | #endif //_VOLATILE_H_ |
| 454 | |