include/SDL_atomic.h
author Sam Lantinga <slouken@libsdl.org>
Wed, 10 Jul 2013 20:17:20 -0700
changeset 7394 38dc4961ab15
parent 7393 358696c354a8
child 7677 871d43c6968a
permissions -rw-r--r--
Added PowerPC and ARM versions of the memory barrier functions.
     1 /*
     2   Simple DirectMedia Layer
     3   Copyright (C) 1997-2013 Sam Lantinga <slouken@libsdl.org>
     4 
     5   This software is provided 'as-is', without any express or implied
     6   warranty.  In no event will the authors be held liable for any damages
     7   arising from the use of this software.
     8 
     9   Permission is granted to anyone to use this software for any purpose,
    10   including commercial applications, and to alter it and redistribute it
    11   freely, subject to the following restrictions:
    12 
    13   1. The origin of this software must not be misrepresented; you must not
    14      claim that you wrote the original software. If you use this software
    15      in a product, an acknowledgment in the product documentation would be
    16      appreciated but is not required.
    17   2. Altered source versions must be plainly marked as such, and must not be
    18      misrepresented as being the original software.
    19   3. This notice may not be removed or altered from any source distribution.
    20 */
    21 
    22 /**
    23  * \file SDL_atomic.h
    24  *
    25  * Atomic operations.
    26  *
    27  * IMPORTANT:
    28  * If you are not an expert in concurrent lockless programming, you should
    29  * only be using the atomic lock and reference counting functions in this
    30  * file.  In all other cases you should be protecting your data structures
    31  * with full mutexes.
    32  *
    33  * The list of "safe" functions to use are:
    34  *  SDL_AtomicLock()
    35  *  SDL_AtomicUnlock()
    36  *  SDL_AtomicIncRef()
    37  *  SDL_AtomicDecRef()
    38  *
    39  * Seriously, here be dragons!
    40  * ^^^^^^^^^^^^^^^^^^^^^^^^^^^
    41  *
    42  * You can find out a little more about lockless programming and the
    43  * subtle issues that can arise here:
    44  * http://msdn.microsoft.com/en-us/library/ee418650%28v=vs.85%29.aspx
    45  *
    46  * There's also lots of good information here:
    47  * http://www.1024cores.net/home/lock-free-algorithms
    48  * http://preshing.com/
    49  *
    50  * These operations may or may not actually be implemented using
    51  * processor specific atomic operations. When possible they are
    52  * implemented as true processor specific atomic operations. When that
    53  * is not possible the are implemented using locks that *do* use the
    54  * available atomic operations.
    55  *
    56  * All of the atomic operations that modify memory are full memory barriers.
    57  */
    58 
    59 #ifndef _SDL_atomic_h_
    60 #define _SDL_atomic_h_
    61 
    62 #include "SDL_stdinc.h"
    63 #include "SDL_platform.h"
    64 
    65 #include "begin_code.h"
    66 
    67 /* Need to do this here because intrin.h has C++ code in it */
    68 /* Visual Studio 2005 has a bug where intrin.h conflicts with winnt.h */
    69 #if defined(_MSC_VER) && (_MSC_VER >= 1500)
    70 #include <intrin.h>
    71 #define HAVE_MSC_ATOMICS 1
    72 #endif
    73 
    74 /* Set up for C function definitions, even when using C++ */
    75 #ifdef __cplusplus
    76 extern "C" {
    77 #endif
    78 
    79 /**
    80  * \name SDL AtomicLock
    81  *
    82  * The atomic locks are efficient spinlocks using CPU instructions,
    83  * but are vulnerable to starvation and can spin forever if a thread
    84  * holding a lock has been terminated.  For this reason you should
    85  * minimize the code executed inside an atomic lock and never do
    86  * expensive things like API or system calls while holding them.
    87  *
    88  * The atomic locks are not safe to lock recursively.
    89  *
    90  * Porting Note:
    91  * The spin lock functions and type are required and can not be
    92  * emulated because they are used in the atomic emulation code.
    93  */
    94 /*@{*/
    95 
    96 typedef int SDL_SpinLock;
    97 
    98 /**
    99  * \brief Try to lock a spin lock by setting it to a non-zero value.
   100  *
   101  * \param lock Points to the lock.
   102  *
   103  * \return SDL_TRUE if the lock succeeded, SDL_FALSE if the lock is already held.
   104  */
   105 extern DECLSPEC SDL_bool SDLCALL SDL_AtomicTryLock(SDL_SpinLock *lock);
   106 
   107 /**
   108  * \brief Lock a spin lock by setting it to a non-zero value.
   109  *
   110  * \param lock Points to the lock.
   111  */
   112 extern DECLSPEC void SDLCALL SDL_AtomicLock(SDL_SpinLock *lock);
   113 
   114 /**
   115  * \brief Unlock a spin lock by setting it to 0. Always returns immediately
   116  *
   117  * \param lock Points to the lock.
   118  */
   119 extern DECLSPEC void SDLCALL SDL_AtomicUnlock(SDL_SpinLock *lock);
   120 
   121 /*@}*//*SDL AtomicLock*/
   122 
   123 
   124 /**
   125  * The compiler barrier prevents the compiler from reordering
   126  * reads and writes to globally visible variables across the call.
   127  */
   128 #if defined(_MSC_VER) && (_MSC_VER > 1200)
   129 void _ReadWriteBarrier(void);
   130 #pragma intrinsic(_ReadWriteBarrier)
   131 #define SDL_CompilerBarrier()   _ReadWriteBarrier()
   132 #elif defined(__GNUC__)
   133 #define SDL_CompilerBarrier()   __asm__ __volatile__ ("" : : : "memory")
   134 #else
   135 #define SDL_CompilerBarrier()   \
   136 { SDL_SpinLock _tmp = 0; SDL_AtomicLock(&_tmp); SDL_AtomicUnlock(&_tmp); }
   137 #endif
   138 
   139 /**
   140  * Memory barriers are designed to prevent reads and writes from being
   141  * reordered by the compiler and being seen out of order on multi-core CPUs.
   142  *
   143  * A typical pattern would be for thread A to write some data and a flag,
   144  * and for thread B to read the flag and get the data. In this case you
   145  * would insert a release barrier between writing the data and the flag,
   146  * guaranteeing that the data write completes no later than the flag is
   147  * written, and you would insert an acquire barrier between reading the
   148  * flag and reading the data, to ensure that all the reads associated
   149  * with the flag have completed.
   150  *
   151  * In this pattern you should always see a release barrier paired with
   152  * an acquire barrier and you should gate the data reads/writes with a
   153  * single flag variable.
   154  *
   155  * For more information on these semantics, take a look at the blog post:
   156  * http://preshing.com/20120913/acquire-and-release-semantics
   157  */
   158 #if defined(__GNUC__) && (defined(__powerpc__) || defined(__ppc__))
   159 #define SDL_MemoryBarrierRelease()   __asm__ __volatile__ ("lwsync" : : : "memory")
   160 #define SDL_MemoryBarrierAcquire()   __asm__ __volatile__ ("lwsync" : : : "memory")
   161 #elif defined(__GNUC__) && defined(__arm__)
   162 #if defined(__ARM_ARCH_7__) || defined(__ARM_ARCH_7A__) || defined(__ARM_ARCH_7EM__) || defined(__ARM_ARCH_7R__) || defined(__ARM_ARCH_7M__) || defined(__ARM_ARCH_7S__)
   163 #define SDL_MemoryBarrierRelease()   __asm__ __volatile__ ("dmb ish" : : : "memory")
   164 #define SDL_MemoryBarrierAcquire()   __asm__ __volatile__ ("dmb ish" : : : "memory")
   165 #elif defined(__ARM_ARCH_6__) || defined(__ARM_ARCH_6J__) || defined(__ARM_ARCH_6K__) || defined(__ARM_ARCH_6T2__) || defined(__ARM_ARCH_6Z__) || defined(__ARM_ARCH_6ZK__)
   166 #ifdef __thumb__
   167 /* The mcr instruction isn't available in thumb mode, use real functions */
   168 extern DECLSPEC void SDLCALL SDL_MemoryBarrierRelease();
   169 extern DECLSPEC void SDLCALL SDL_MemoryBarrierAcquire();
   170 #else
   171 #define SDL_MemoryBarrierRelease()   __asm__ __volatile__ ("mcr p15, 0, %0, c7, c10, 5" : : "r"(0) : "memory")
   172 #define SDL_MemoryBarrierAcquire()   __asm__ __volatile__ ("mcr p15, 0, %0, c7, c10, 5" : : "r"(0) : "memory")
   173 #endif /* __thumb__ */
   174 #else
   175 #define SDL_MemoryBarrierRelease()   __asm__ __volatile__ ("" : : : "memory")
   176 #define SDL_MemoryBarrierAcquire()   __asm__ __volatile__ ("" : : : "memory")
   177 #endif /* __GNUC__ && __arm__ */
   178 #else
   179 /* This is correct for the x86 and x64 CPUs, and we'll expand this over time. */
   180 #define SDL_MemoryBarrierRelease()  SDL_CompilerBarrier()
   181 #define SDL_MemoryBarrierAcquire()  SDL_CompilerBarrier()
   182 #endif
   183 
   184 
   185 /* Platform specific optimized versions of the atomic functions,
   186  * you can disable these by defining SDL_DISABLE_ATOMIC_INLINE
   187  */
   188 #if defined(SDL_ATOMIC_DISABLED) && SDL_ATOMIC_DISABLED
   189 #define SDL_DISABLE_ATOMIC_INLINE
   190 #endif
   191 #ifndef SDL_DISABLE_ATOMIC_INLINE
   192 
   193 #ifdef HAVE_MSC_ATOMICS
   194 
   195 #define SDL_AtomicSet(a, v)     _InterlockedExchange((long*)&(a)->value, (v))
   196 #define SDL_AtomicAdd(a, v)     _InterlockedExchangeAdd((long*)&(a)->value, (v))
   197 #define SDL_AtomicCAS(a, oldval, newval) (_InterlockedCompareExchange((long*)&(a)->value, (newval), (oldval)) == (oldval))
   198 #define SDL_AtomicSetPtr(a, v)  _InterlockedExchangePointer((a), (v))
   199 #if _M_IX86
   200 #define SDL_AtomicCASPtr(a, oldval, newval) (_InterlockedCompareExchange((long*)(a), (long)(newval), (long)(oldval)) == (long)(oldval))
   201 #else
   202 #define SDL_AtomicCASPtr(a, oldval, newval) (_InterlockedCompareExchangePointer((a), (newval), (oldval)) == (oldval))
   203 #endif
   204 
   205 #elif defined(__MACOSX__)
   206 #include <libkern/OSAtomic.h>
   207 
   208 #define SDL_AtomicCAS(a, oldval, newval) OSAtomicCompareAndSwap32Barrier((oldval), (newval), &(a)->value)
   209 #ifdef __LP64__
   210 #define SDL_AtomicCASPtr(a, oldval, newval) OSAtomicCompareAndSwap64Barrier((int64_t)(oldval), (int64_t)(newval), (int64_t*)(a))
   211 #else
   212 #define SDL_AtomicCASPtr(a, oldval, newval) OSAtomicCompareAndSwap32Barrier((int32_t)(oldval), (int32_t)(newval), (int32_t*)(a))
   213 #endif
   214 
   215 #elif defined(HAVE_GCC_ATOMICS)
   216 
   217 #define SDL_AtomicSet(a, v)     __sync_lock_test_and_set(&(a)->value, v)
   218 #define SDL_AtomicAdd(a, v)     __sync_fetch_and_add(&(a)->value, v)
   219 #define SDL_AtomicSetPtr(a, v)  __sync_lock_test_and_set(a, v)
   220 #define SDL_AtomicCAS(a, oldval, newval) __sync_bool_compare_and_swap(&(a)->value, oldval, newval)
   221 #define SDL_AtomicCASPtr(a, oldval, newval) __sync_bool_compare_and_swap(a, oldval, newval)
   222 
   223 #endif
   224 
   225 #endif /* !SDL_DISABLE_ATOMIC_INLINE */
   226 
   227 
   228 /**
   229  * \brief A type representing an atomic integer value.  It is a struct
   230  *        so people don't accidentally use numeric operations on it.
   231  */
   232 #ifndef SDL_atomic_t_defined
   233 typedef struct { int value; } SDL_atomic_t;
   234 #endif
   235 
   236 /**
   237  * \brief Set an atomic variable to a new value if it is currently an old value.
   238  *
   239  * \return SDL_TRUE if the atomic variable was set, SDL_FALSE otherwise.
   240  *
   241  * \note If you don't know what this function is for, you shouldn't use it!
   242 */
   243 #ifndef SDL_AtomicCAS
   244 extern DECLSPEC SDL_bool SDLCALL SDL_AtomicCAS(SDL_atomic_t *a, int oldval, int newval);
   245 #endif
   246 
   247 /**
   248  * \brief Set an atomic variable to a value.
   249  *
   250  * \return The previous value of the atomic variable.
   251  */
   252 #ifndef SDL_AtomicSet
   253 SDL_FORCE_INLINE int SDL_AtomicSet(SDL_atomic_t *a, int v)
   254 {
   255     int value;
   256     do {
   257         value = a->value;
   258     } while (!SDL_AtomicCAS(a, value, v));
   259     return value;
   260 }
   261 #endif
   262 
   263 /**
   264  * \brief Get the value of an atomic variable
   265  */
   266 #ifndef SDL_AtomicGet
   267 SDL_FORCE_INLINE int SDL_AtomicGet(SDL_atomic_t *a)
   268 {
   269     int value = a->value;
   270     SDL_CompilerBarrier();
   271     return value;
   272 }
   273 #endif
   274 
   275 /**
   276  * \brief Add to an atomic variable.
   277  *
   278  * \return The previous value of the atomic variable.
   279  *
   280  * \note This same style can be used for any number operation
   281  */
   282 #ifndef SDL_AtomicAdd
   283 SDL_FORCE_INLINE int SDL_AtomicAdd(SDL_atomic_t *a, int v)
   284 {
   285     int value;
   286     do {
   287         value = a->value;
   288     } while (!SDL_AtomicCAS(a, value, (value + v)));
   289     return value;
   290 }
   291 #endif
   292 
   293 /**
   294  * \brief Increment an atomic variable used as a reference count.
   295  */
   296 #ifndef SDL_AtomicIncRef
   297 #define SDL_AtomicIncRef(a)    SDL_AtomicAdd(a, 1)
   298 #endif
   299 
   300 /**
   301  * \brief Decrement an atomic variable used as a reference count.
   302  *
   303  * \return SDL_TRUE if the variable reached zero after decrementing,
   304  *         SDL_FALSE otherwise
   305  */
   306 #ifndef SDL_AtomicDecRef
   307 #define SDL_AtomicDecRef(a)    (SDL_AtomicAdd(a, -1) == 1)
   308 #endif
   309 
   310 /**
   311  * \brief Set a pointer to a new value if it is currently an old value.
   312  *
   313  * \return SDL_TRUE if the pointer was set, SDL_FALSE otherwise.
   314  *
   315  * \note If you don't know what this function is for, you shouldn't use it!
   316 */
   317 #ifndef SDL_AtomicCASPtr
   318 extern DECLSPEC SDL_bool SDLCALL SDL_AtomicCASPtr(void* *a, void *oldval, void *newval);
   319 #endif
   320 
   321 /**
   322  * \brief Set a pointer to a value atomically.
   323  *
   324  * \return The previous value of the pointer.
   325  */
   326 #ifndef SDL_AtomicSetPtr
   327 SDL_FORCE_INLINE void* SDL_AtomicSetPtr(void* *a, void* v)
   328 {
   329     void* value;
   330     do {
   331         value = *a;
   332     } while (!SDL_AtomicCASPtr(a, value, v));
   333     return value;
   334 }
   335 #endif
   336 
   337 /**
   338  * \brief Get the value of a pointer atomically.
   339  */
   340 #ifndef SDL_AtomicGetPtr
   341 SDL_FORCE_INLINE void* SDL_AtomicGetPtr(void* *a)
   342 {
   343     void* value = *a;
   344     SDL_CompilerBarrier();
   345     return value;
   346 }
   347 #endif
   348 
   349 
   350 /* Ends C function definitions when using C++ */
   351 #ifdef __cplusplus
   352 }
   353 #endif
   354 
   355 #include "close_code.h"
   356 
   357 #endif /* _SDL_atomic_h_ */
   358 
   359 /* vi: set ts=4 sw=4 expandtab: */