OVR_Atomic.h
Go to the documentation of this file.
00001 /************************************************************************************
00002 
00003 PublicHeader:   OVR.h
00004 Filename    :   OVR_Atomic.h
00005 Content     :   Contains atomic operations and inline fastest locking
00006                 functionality. Will contain #ifdefs for OS efficiency.
00007                 Have non-thread-safe implementaion if not available.
00008 Created     :   September 19, 2012
00009 Notes       : 
00010 
00011 Copyright   :   Copyright 2012 Oculus VR, Inc. All Rights reserved.
00012 
00013 Use of this software is subject to the terms of the Oculus license
00014 agreement provided at the time of installation or download, or which
00015 otherwise accompanies this software in either electronic or hard copy form.
00016 
00017 ************************************************************************************/
00018 #ifndef OVR_Atomic_h
00019 #define OVR_Atomic_h
00020 
00021 #include "OVR_Types.h"
00022 
00023 // Include System thread functionality.
00024 #if defined(OVR_OS_WIN32)
00025 #include <windows.h>
00026 #else
00027 #include <pthread.h>
00028 #endif
00029 
00030 
00031 namespace OVR {
00032 
00033 
00034 // ****** Declared classes
00035 
00036 // If there is NO thread support we implement AtomicOps and
00037 // Lock objects as no-ops. The other classes are not defined.
00038 template<class C> class AtomicOps;
00039 template<class T> class AtomicInt;
00040 template<class T> class AtomicPtr;
00041 
00042 class   Lock;
00043 
00044 
00045 //-----------------------------------------------------------------------------------
00046 // ***** AtomicOps
00047 
00048 // Atomic operations are provided by the AtomicOps templates class,
00049 // implemented through system-specific AtomicOpsRaw specializations.
00050 // It provides several fundamental operations such as Exchange, ExchangeAdd
00051 // CompareAndSet, and Store_Release. Each function includes several memory
00052 // synchronization versions, important for multiprocessing CPUs with weak
00053 // memory consistency. The following memory fencing strategies are supported:
00054 //
00055 //  - NoSync.  No memory synchronization is done for atomic op.
00056 //  - Release. All other memory writes are completed before atomic op
00057 //             writes its results.
00058 //  - Acquire. Further memory reads are forced to wait until atomic op
00059 //             executes, guaranteeing that the right values will be seen.
00060 //  - Sync.    A combination of Release and Acquire.
00061 
00062 
00063 // *** AtomicOpsRaw
00064 
00065 // AtomicOpsRaw is a specialized template that provides atomic operations 
00066 // used by AtomicOps. This class has two fundamental qualities: (1) it
00067 // defines a type T of correct size, and (2) provides operations that work
00068 // atomically, such as Exchange_Sync and CompareAndSet_Release.
00069 
00070 // AtomicOpsRawBase class contains shared constants/classes for AtomicOpsRaw.
00071 // The primary thing is does is define sync class objects, whose destructor and
00072 // constructor provide places to insert appropriate synchronization calls, on 
00073 // systems where such calls are necessary. So far, the breakdown is as follows:
00074 // 
00075 //  - X86 systems don't need custom syncs, since their exchange/atomic
00076 //    instructions are implicitly synchronized.
00077 //  - PowerPC requires lwsync/isync instructions that can use this mechanism.
00078 //  - If some other systems require a mechanism where syncing type is associated
00079 //    with a particular instruction, the default implementation (which implements
00080 //    all Sync, Acquire, and Release modes in terms of NoSync and fence) may not
00081 //    work. Ii that case it will need to be #ifdef-ed conditionally.
00082 
00083 struct AtomicOpsRawBase
00084 {
00085 #if !defined(OVR_ENABLE_THREADS) || defined(OVR_CPU_X86) || defined(OVR_OS_WIN32) || defined(OVR_OS_IPHONE)
00086     // Need to have empty constructor to avoid class 'unused' variable warning.
00087     struct FullSync { inline FullSync() { } };
00088     struct AcquireSync { inline AcquireSync() { } };
00089     struct ReleaseSync { inline ReleaseSync() { } };
00090 
00091 #elif defined(OVR_CPU_PPC64) || defined(OVR_CPU_PPC)
00092     struct FullSync { inline FullSync() { asm volatile("sync\n"); } ~FullSync() { asm volatile("isync\n"); } };
00093     struct AcquireSync { inline AcquireSync() { } ~AcquireSync() { asm volatile("isync\n"); } };
00094     struct ReleaseSync { inline ReleaseSync() { asm volatile("sync\n"); } };
00095 
00096 #elif defined(OVR_CPU_MIPS)
00097     struct FullSync { inline FullSync() { asm volatile("sync\n"); } ~FullSync() { asm volatile("sync\n"); } };
00098     struct AcquireSync { inline AcquireSync() { } ~AcquireSync() { asm volatile("sync\n"); } };
00099     struct ReleaseSync { inline ReleaseSync() { asm volatile("sync\n"); } };
00100 
00101 #elif defined(OVR_CPU_ARM)
00102     struct FullSync { inline FullSync() { asm volatile("dmb\n"); } ~FullSync() { asm volatile("dmb\n"); } };
00103     struct AcquireSync { inline AcquireSync() { } ~AcquireSync() { asm volatile("dmb\n"); } };
00104     struct ReleaseSync { inline ReleaseSync() { asm volatile("dmb\n"); } };
00105 
00106 
00107 #elif defined(OVR_CC_GNU) && (__GNUC__ >= 4)
00108     // __sync functions are already full sync
00109     struct FullSync { inline FullSync() { } };
00110     struct AcquireSync { inline AcquireSync() { } };
00111     struct ReleaseSync { inline ReleaseSync() { } };
00112 #endif
00113 };
00114 
00115 
00116 // 4-Byte raw data atomic op implementation class.
00117 struct AtomicOpsRaw_4ByteImpl : public AtomicOpsRawBase
00118 {
00119 #if !defined(OVR_ENABLE_THREADS)
00120 
00121     // Provide a type for no-thread-support cases. Used by AtomicOpsRaw_DefImpl.
00122     typedef UInt32 T;   
00123 
00124     // *** Thread - Safe Atomic Versions.
00125 
00126 #elif defined(OVR_OS_WIN32)
00127 
00128     // Use special defined for VC6, where volatile is not used and
00129     // InterlockedCompareExchange is declared incorrectly.
00130     typedef LONG T;      
00131 #if defined(OVR_CC_MSVC) && (OVR_CC_MSVC < 1300)
00132     typedef T* InterlockTPtr;
00133     typedef LPVOID ET;
00134     typedef ET* InterlockETPtr;
00135 #else
00136     typedef volatile T* InterlockTPtr;
00137     typedef T ET;
00138     typedef InterlockTPtr InterlockETPtr;
00139 #endif
00140     inline static T     Exchange_NoSync(volatile T* p, T val)            { return InterlockedExchange((InterlockTPtr)p, val); }
00141     inline static T     ExchangeAdd_NoSync(volatile T* p, T val)         { return InterlockedExchangeAdd((InterlockTPtr)p, val); }
00142     inline static bool  CompareAndSet_NoSync(volatile T* p, T c, T val)  { return InterlockedCompareExchange((InterlockETPtr)p, (ET)val, (ET)c) == (ET)c; }
00143 
00144 #elif defined(OVR_CPU_PPC64) || defined(OVR_CPU_PPC)
00145     typedef UInt32 T;
00146     static inline UInt32   Exchange_NoSync(volatile UInt32 *i, UInt32 j)
00147     {
00148         UInt32 ret;
00149 
00150         asm volatile("1:\n\t"
00151                      "lwarx  %[r],0,%[i]\n\t"
00152                      "stwcx. %[j],0,%[i]\n\t"
00153                      "bne-   1b\n"
00154                      : "+m" (*i), [r] "=&b" (ret) : [i] "b" (i), [j] "b" (j) : "cc", "memory");
00155 
00156         return ret;
00157     }
00158 
00159     static inline UInt32   ExchangeAdd_NoSync(volatile UInt32 *i, UInt32 j)
00160     {
00161         UInt32 dummy, ret;
00162 
00163         asm volatile("1:\n\t"
00164                      "lwarx  %[r],0,%[i]\n\t"
00165                      "add    %[o],%[r],%[j]\n\t"
00166                      "stwcx. %[o],0,%[i]\n\t"
00167                      "bne-   1b\n"
00168                      : "+m" (*i), [r] "=&b" (ret), [o] "=&r" (dummy) : [i] "b" (i), [j] "b" (j) : "cc", "memory");
00169 
00170         return ret;
00171     }
00172 
00173     static inline bool     CompareAndSet_NoSync(volatile UInt32 *i, UInt32 c, UInt32 value)
00174     {
00175         UInt32 ret;
00176 
00177         asm volatile("1:\n\t"
00178                      "lwarx  %[r],0,%[i]\n\t"
00179                      "cmpw   0,%[r],%[cmp]\n\t"
00180                      "mfcr   %[r]\n\t"
00181                      "bne-   2f\n\t"
00182                      "stwcx. %[val],0,%[i]\n\t"
00183                      "bne-   1b\n\t"
00184                      "2:\n"
00185                      : "+m" (*i), [r] "=&b" (ret) : [i] "b" (i), [cmp] "b" (c), [val] "b" (value) : "cc", "memory");
00186 
00187         return (ret & 0x20000000) ? 1 : 0;
00188     }
00189 
00190 #elif defined(OVR_CPU_MIPS)
00191     typedef UInt32 T;
00192 
00193     static inline UInt32   Exchange_NoSync(volatile UInt32 *i, UInt32 j)
00194     {
00195         UInt32 ret;
00196 
00197         asm volatile("1:\n\t"
00198                      "ll     %[r],0(%[i])\n\t"
00199                      "sc     %[j],0(%[i])\n\t"
00200                      "beq    %[j],$0,1b\n\t"
00201                      "nop    \n"
00202                      : "+m" (*i), [r] "=&d" (ret) : [i] "d" (i), [j] "d" (j) : "cc", "memory");
00203 
00204         return ret;
00205     }
00206 
00207     static inline UInt32   ExchangeAdd_NoSync(volatile UInt32 *i, UInt32 j)
00208     {
00209         UInt32 ret;
00210 
00211         asm volatile("1:\n\t"
00212                      "ll     %[r],0(%[i])\n\t"
00213                      "addu   %[j],%[r],%[j]\n\t"
00214                      "sc     %[j],0(%[i])\n\t"
00215                      "beq    %[j],$0,1b\n\t"
00216                      "nop    \n"
00217                      : "+m" (*i), [r] "=&d" (ret) : [i] "d" (i), [j] "d" (j) : "cc", "memory");
00218 
00219         return ret;
00220     }
00221 
00222     static inline bool     CompareAndSet_NoSync(volatile UInt32 *i, UInt32 c, UInt32 value)
00223     {
00224         UInt32 ret, dummy;
00225 
00226         asm volatile("1:\n\t"
00227                      "move   %[r],$0\n\t"
00228                      "ll     %[o],0(%[i])\n\t"
00229                      "bne    %[o],%[c],2f\n\t"
00230                      "move   %[r],%[v]\n\t"
00231                      "sc     %[r],0(%[i])\n\t"
00232                      "beq    %[r],$0,1b\n\t"
00233                      "nop    \n\t"
00234                      "2:\n"
00235                      : "+m" (*i),[r] "=&d" (ret), [o] "=&d" (dummy) : [i] "d" (i), [c] "d" (c), [v] "d" (value)
00236                      : "cc", "memory");
00237 
00238         return ret;
00239     }
00240 
00241 #elif defined(OVR_CPU_ARM) && defined(OVR_CC_ARM)
00242     typedef UInt32 T;
00243 
00244     static inline UInt32   Exchange_NoSync(volatile UInt32 *i, UInt32 j)
00245     {
00246         for(;;)
00247         {
00248             T r = __ldrex(i);
00249             if (__strex(j, i) == 0)
00250                 return r;
00251         }
00252     }
00253     static inline UInt32   ExchangeAdd_NoSync(volatile UInt32 *i, UInt32 j)
00254     {
00255         for(;;)
00256         {
00257             T r = __ldrex(i);
00258             if (__strex(r + j, i) == 0)
00259                 return r;
00260         }
00261     }
00262 
00263     static inline bool     CompareAndSet_NoSync(volatile UInt32 *i, UInt32 c, UInt32 value)
00264     {
00265         for(;;)
00266         {
00267             T r = __ldrex(i);
00268             if (r != c)
00269                 return 0;
00270             if (__strex(value, i) == 0)
00271                 return 1;
00272         }
00273     }
00274 
00275 #elif defined(OVR_CPU_ARM)
00276     typedef UInt32 T;
00277 
00278     static inline UInt32   Exchange_NoSync(volatile UInt32 *i, UInt32 j)
00279     {
00280         UInt32 ret, dummy;
00281 
00282         asm volatile("1:\n\t"
00283             "ldrex  %[r],[%[i]]\n\t"
00284             "strex  %[t],%[j],[%[i]]\n\t"
00285             "cmp    %[t],#0\n\t"
00286             "bne    1b\n\t"
00287             : "+m" (*i), [r] "=&r" (ret), [t] "=&r" (dummy) : [i] "r" (i), [j] "r" (j) : "cc", "memory");
00288 
00289         return ret;
00290     }
00291 
00292     static inline UInt32   ExchangeAdd_NoSync(volatile UInt32 *i, UInt32 j)
00293     {
00294         UInt32 ret, dummy, test;
00295 
00296         asm volatile("1:\n\t"
00297             "ldrex  %[r],[%[i]]\n\t"
00298             "add    %[o],%[r],%[j]\n\t"
00299             "strex  %[t],%[o],[%[i]]\n\t"
00300             "cmp    %[t],#0\n\t"
00301             "bne    1b\n\t"
00302             : "+m" (*i), [r] "=&r" (ret), [o] "=&r" (dummy), [t] "=&r" (test)  : [i] "r" (i), [j] "r" (j) : "cc", "memory");
00303 
00304         return ret;
00305     }
00306 
00307     static inline bool     CompareAndSet_NoSync(volatile UInt32 *i, UInt32 c, UInt32 value)
00308     {
00309         UInt32 ret = 1, dummy, test;
00310 
00311         asm volatile("1:\n\t"
00312             "ldrex  %[o],[%[i]]\n\t"
00313             "cmp    %[o],%[c]\n\t"
00314             "bne    2f\n\t"
00315             "strex  %[r],%[v],[%[i]]\n\t"
00316             "cmp    %[r],#0\n\t"
00317             "bne    1b\n\t"
00318             "2:\n"
00319             : "+m" (*i),[r] "=&r" (ret), [o] "=&r" (dummy), [t] "=&r" (test) : [i] "r" (i), [c] "r" (c), [v] "r" (value)
00320             : "cc", "memory");
00321 
00322         return !ret;
00323     }
00324 
00325 #elif defined(OVR_CPU_X86)
00326     typedef UInt32 T;
00327 
00328     static inline UInt32   Exchange_NoSync(volatile UInt32 *i, UInt32 j)
00329     {
00330         asm volatile("xchgl %1,%[i]\n"
00331                      : "+m" (*i), "=q" (j) : [i] "m" (*i), "1" (j) : "cc", "memory");
00332 
00333         return j;
00334     }
00335 
00336     static inline UInt32   ExchangeAdd_NoSync(volatile UInt32 *i, UInt32 j)
00337     {
00338         asm volatile("lock; xaddl %1,%[i]\n"
00339                      : "+m" (*i), "+q" (j) : [i] "m" (*i) : "cc", "memory");
00340 
00341         return j;
00342     }
00343 
00344     static inline bool     CompareAndSet_NoSync(volatile UInt32 *i, UInt32 c, UInt32 value)
00345     {
00346         UInt32 ret;
00347 
00348         asm volatile("lock; cmpxchgl %[v],%[i]\n"
00349                      : "+m" (*i), "=a" (ret) : [i] "m" (*i), "1" (c), [v] "q" (value) : "cc", "memory");
00350 
00351         return (ret == c);
00352     }
00353 
00354 #elif defined(OVR_CC_GNU) && (__GNUC__ >= 4 && __GNUC_MINOR__ >= 1)
00355 
00356     typedef UInt32 T;
00357 
00358     static inline T   Exchange_NoSync(volatile T *i, T j)
00359     {
00360         T v;
00361         do {
00362             v = *i;
00363         } while (!__sync_bool_compare_and_swap(i, v, j));
00364         return v;
00365     }
00366 
00367     static inline T   ExchangeAdd_NoSync(volatile T *i, T j)
00368     {
00369         return __sync_fetch_and_add(i, j);
00370     }
00371 
00372     static inline bool     CompareAndSet_NoSync(volatile T *i, T c, T value)
00373     {
00374         return __sync_bool_compare_and_swap(i, c, value);
00375     }
00376 
00377 #endif // OS
00378 };
00379 
00380 
00381 // 8-Byte raw data data atomic op implementation class.
00382 // Currently implementation is provided only on systems with 64-bit pointers.
00383 struct AtomicOpsRaw_8ByteImpl : public AtomicOpsRawBase
00384 {    
00385 #if !defined(OVR_64BIT_POINTERS) || !defined(OVR_ENABLE_THREADS)
00386 
00387     // Provide a type for no-thread-support cases. Used by AtomicOpsRaw_DefImpl.
00388     typedef UInt64 T;
00389 
00390     // *** Thread - Safe OS specific versions.
00391 #elif defined(OVR_OS_WIN32)
00392 
00393     // This is only for 64-bit systems.
00394     typedef LONG64      T;
00395     typedef volatile T* InterlockTPtr;    
00396     inline static T     Exchange_NoSync(volatile T* p, T val)            { return InterlockedExchange64((InterlockTPtr)p, val); }
00397     inline static T     ExchangeAdd_NoSync(volatile T* p, T val)         { return InterlockedExchangeAdd64((InterlockTPtr)p, val); }
00398     inline static bool  CompareAndSet_NoSync(volatile T* p, T c, T val)  { return InterlockedCompareExchange64((InterlockTPtr)p, val, c) == c; }
00399 
00400 #elif defined(OVR_CPU_PPC64)
00401  
00402     typedef UInt64 T;
00403 
00404     static inline UInt64   Exchange_NoSync(volatile UInt64 *i, UInt64 j)
00405     {
00406         UInt64 dummy, ret;
00407 
00408         asm volatile("1:\n\t"
00409                      "ldarx  %[r],0,%[i]\n\t"
00410                      "mr     %[o],%[j]\n\t"
00411                      "stdcx. %[o],0,%[i]\n\t"
00412                      "bne-   1b\n"
00413                      : "+m" (*i), [r] "=&b" (ret), [o] "=&r" (dummy) : [i] "b" (i), [j] "b" (j) : "cc");
00414 
00415         return ret;
00416     }
00417 
00418     static inline UInt64   ExchangeAdd_NoSync(volatile UInt64 *i, UInt64 j)
00419     {
00420         UInt64 dummy, ret;
00421 
00422         asm volatile("1:\n\t"
00423                      "ldarx  %[r],0,%[i]\n\t"
00424                      "add    %[o],%[r],%[j]\n\t"
00425                      "stdcx. %[o],0,%[i]\n\t"
00426                      "bne-   1b\n"
00427                      : "+m" (*i), [r] "=&b" (ret), [o] "=&r" (dummy) : [i] "b" (i), [j] "b" (j) : "cc");
00428 
00429         return ret;
00430     }
00431 
00432     static inline bool     CompareAndSet_NoSync(volatile UInt64 *i, UInt64 c, UInt64 value)
00433     {
00434         UInt64 ret, dummy;
00435 
00436         asm volatile("1:\n\t"
00437                      "ldarx  %[r],0,%[i]\n\t"
00438                      "cmpw   0,%[r],%[cmp]\n\t"
00439                      "mfcr   %[r]\n\t"
00440                      "bne-   2f\n\t"
00441                      "stdcx. %[val],0,%[i]\n\t"
00442                      "bne-   1b\n\t"
00443                      "2:\n"
00444                      : "+m" (*i), [r] "=&b" (ret), [o] "=&r" (dummy) : [i] "b" (i), [cmp] "b" (c), [val] "b" (value) : "cc");
00445 
00446         return (ret & 0x20000000) ? 1 : 0;
00447     }
00448 
00449 #elif defined(OVR_CC_GNU) && (__GNUC__ >= 4 && __GNUC_MINOR__ >= 1)
00450 
00451     typedef UInt64 T;
00452 
00453     static inline T   Exchange_NoSync(volatile T *i, T j)
00454     {
00455         T v;
00456         do {
00457             v = *i;
00458         } while (!__sync_bool_compare_and_swap(i, v, j));
00459         return v;
00460     }
00461 
00462     static inline T   ExchangeAdd_NoSync(volatile T *i, T j)
00463     {
00464         return __sync_fetch_and_add(i, j);
00465     }
00466 
00467     static inline bool     CompareAndSet_NoSync(volatile T *i, T c, T value)
00468     {
00469         return __sync_bool_compare_and_swap(i, c, value);
00470     }
00471 
00472 #endif // OS
00473 };
00474 
00475 
00476 // Default implementation for AtomicOpsRaw; provides implementation of mem-fenced
00477 // atomic operations where fencing is done with a sync object wrapped around a NoSync
00478 // operation implemented in the base class. If such implementation is not possible
00479 // on a given platform, #ifdefs can be used to disable it and then op functions can be
00480 // implemented individually in the appropriate AtomicOpsRaw<size> class.
00481 
00482 template<class O>
00483 struct AtomicOpsRaw_DefImpl : public O
00484 {
00485     typedef typename O::T O_T;
00486     typedef typename O::FullSync    O_FullSync;
00487     typedef typename O::AcquireSync O_AcquireSync;
00488     typedef typename O::ReleaseSync O_ReleaseSync;
00489 
00490     // If there is no thread support, provide the default implementation. In this case,
00491     // the base class (0) must still provide the T declaration.
00492 #ifndef OVR_ENABLE_THREADS
00493 
00494     // Atomic exchange of val with argument. Returns old val.
00495     inline static O_T   Exchange_NoSync(volatile O_T* p, O_T val)           { O_T old = *p; *p = val; return old; }
00496     // Adds a new val to argument; returns its old val.
00497     inline static O_T   ExchangeAdd_NoSync(volatile O_T* p, O_T val)        { O_T old = *p; *p += val; return old; }
00498     // Compares the argument data with 'c' val.
00499     // If succeeded, stores val int '*p' and returns true; otherwise returns false.
00500     inline static bool  CompareAndSet_NoSync(volatile O_T* p, O_T c, O_T val) { if (*p==c) { *p = val; return 1; } return 0; }
00501 
00502 #endif
00503 
00504     // If NoSync wrapped implementation may not be possible, it this block should be
00505     //  replaced with per-function implementation in O.
00506     // "AtomicOpsRaw_DefImpl<O>::" prefix in calls below.
00507     inline static O_T   Exchange_Sync(volatile O_T* p, O_T val)                { O_FullSync    sync; OVR_UNUSED(sync); return AtomicOpsRaw_DefImpl<O>::Exchange_NoSync(p, val); }
00508     inline static O_T   Exchange_Release(volatile O_T* p, O_T val)             { O_ReleaseSync sync; OVR_UNUSED(sync); return AtomicOpsRaw_DefImpl<O>::Exchange_NoSync(p, val); }
00509     inline static O_T   Exchange_Acquire(volatile O_T* p, O_T val)             { O_AcquireSync sync; OVR_UNUSED(sync); return AtomicOpsRaw_DefImpl<O>::Exchange_NoSync(p, val); }  
00510     inline static O_T   ExchangeAdd_Sync(volatile O_T* p, O_T val)             { O_FullSync    sync; OVR_UNUSED(sync); return AtomicOpsRaw_DefImpl<O>::ExchangeAdd_NoSync(p, val); }
00511     inline static O_T   ExchangeAdd_Release(volatile O_T* p, O_T val)          { O_ReleaseSync sync; OVR_UNUSED(sync); return AtomicOpsRaw_DefImpl<O>::ExchangeAdd_NoSync(p, val); }
00512     inline static O_T   ExchangeAdd_Acquire(volatile O_T* p, O_T val)          { O_AcquireSync sync; OVR_UNUSED(sync); return AtomicOpsRaw_DefImpl<O>::ExchangeAdd_NoSync(p, val); }
00513     inline static bool  CompareAndSet_Sync(volatile O_T* p, O_T c, O_T val)    { O_FullSync    sync; OVR_UNUSED(sync); return AtomicOpsRaw_DefImpl<O>::CompareAndSet_NoSync(p,c,val); }
00514     inline static bool  CompareAndSet_Release(volatile O_T* p, O_T c, O_T val) { O_ReleaseSync sync; OVR_UNUSED(sync); return AtomicOpsRaw_DefImpl<O>::CompareAndSet_NoSync(p,c,val); }
00515     inline static bool  CompareAndSet_Acquire(volatile O_T* p, O_T c, O_T val) { O_AcquireSync sync; OVR_UNUSED(sync); return AtomicOpsRaw_DefImpl<O>::CompareAndSet_NoSync(p,c,val); }
00516 
00517     // Loads and stores with memory fence. These have only the relevant versions.
00518 #ifdef OVR_CPU_X86
00519     // On X86, Store_Release is implemented as exchange. Note that we can also
00520     // consider 'sfence' in the future, although it is not as compatible with older CPUs.
00521     inline static void  Store_Release(volatile O_T* p, O_T val)  { Exchange_Release(p, val); }
00522 #else
00523     inline static void  Store_Release(volatile O_T* p, O_T val)  { O_ReleaseSync sync; OVR_UNUSED(sync); *p = val; }
00524 #endif
00525     inline static O_T   Load_Acquire(const volatile O_T* p)      { O_AcquireSync sync; OVR_UNUSED(sync); return *p; }
00526 };
00527 
00528 
00529 template<int size>
00530 struct AtomicOpsRaw : public AtomicOpsRawBase { };
00531 
00532 template<>
00533 struct AtomicOpsRaw<4> : public AtomicOpsRaw_DefImpl<AtomicOpsRaw_4ByteImpl>
00534 {   
00535     // Ensure that assigned type size is correct.
00536     AtomicOpsRaw()
00537     { OVR_COMPILER_ASSERT(sizeof(AtomicOpsRaw_DefImpl<AtomicOpsRaw_4ByteImpl>::T) == 4); }
00538 };
00539 template<>
00540 struct AtomicOpsRaw<8> : public AtomicOpsRaw_DefImpl<AtomicOpsRaw_8ByteImpl>
00541 {
00542     AtomicOpsRaw()
00543     { OVR_COMPILER_ASSERT(sizeof(AtomicOpsRaw_DefImpl<AtomicOpsRaw_8ByteImpl>::T) == 8); }
00544 };
00545 
00546 
00547 // *** AtomicOps - implementation of atomic Ops for specified class
00548 
00549 // Implements atomic ops on a class, provided that the object is either
00550 // 4 or 8 bytes in size (depending on the AtomicOpsRaw specializations
00551 // available). Relies on AtomicOpsRaw for much of implementation.
00552 
00553 template<class C>
00554 class AtomicOps
00555 {
00556     typedef AtomicOpsRaw<sizeof(C)>       Ops;
00557     typedef typename Ops::T               T;
00558     typedef volatile typename Ops::T*     PT;
00559     // We cast through unions to (1) avoid pointer size compiler warnings
00560     // and (2) ensure that there are no problems with strict pointer aliasing.
00561     union C2T_union { C c; T t; };
00562 
00563 public:
00564     // General purpose implementation for standard syncs.    
00565     inline static C     Exchange_Sync(volatile C* p, C val)             { C2T_union u; u.c = val; u.t = Ops::Exchange_Sync((PT)p, u.t); return u.c; }
00566     inline static C     Exchange_Release(volatile C* p, C val)          { C2T_union u; u.c = val; u.t = Ops::Exchange_Release((PT)p, u.t); return u.c; }
00567     inline static C     Exchange_Acquire(volatile C* p, C val)          { C2T_union u; u.c = val; u.t = Ops::Exchange_Acquire((PT)p, u.t); return u.c; }
00568     inline static C     Exchange_NoSync(volatile C* p, C val)           { C2T_union u; u.c = val; u.t = Ops::Exchange_NoSync((PT)p, u.t); return u.c; }
00569     inline static C     ExchangeAdd_Sync(volatile C* p, C val)          { C2T_union u; u.c = val; u.t = Ops::ExchangeAdd_Sync((PT)p, u.t); return u.c; }
00570     inline static C     ExchangeAdd_Release(volatile C* p, C val)       { C2T_union u; u.c = val; u.t = Ops::ExchangeAdd_Release((PT)p, u.t); return u.c; }
00571     inline static C     ExchangeAdd_Acquire(volatile C* p, C val)       { C2T_union u; u.c = val; u.t = Ops::ExchangeAdd_Acquire((PT)p, u.t); return u.c; }
00572     inline static C     ExchangeAdd_NoSync(volatile C* p, C val)        { C2T_union u; u.c = val; u.t = Ops::ExchangeAdd_NoSync((PT)p, u.t); return u.c; }
00573     inline static bool  CompareAndSet_Sync(volatile C* p, C c, C val)   { C2T_union u,cu; u.c = val; cu.c = c; return Ops::CompareAndSet_Sync((PT)p, cu.t, u.t); }
00574     inline static bool  CompareAndSet_Release(volatile C* p, C c, C val){ C2T_union u,cu; u.c = val; cu.c = c; return Ops::CompareAndSet_Release((PT)p, cu.t, u.t); }
00575     inline static bool  CompareAndSet_Relse(volatile C* p, C c, C val){ C2T_union u,cu; u.c = val; cu.c = c; return Ops::CompareAndSet_Acquire((PT)p, cu.t, u.t); }
00576     inline static bool  CompareAndSet_NoSync(volatile C* p, C c, C val) { C2T_union u,cu; u.c = val; cu.c = c; return Ops::CompareAndSet_NoSync((PT)p, cu.t, u.t); }
00577     // Loads and stores with memory fence. These have only the relevant versions.    
00578     inline static void  Store_Release(volatile C* p, C val)             { C2T_union u; u.c = val; Ops::Store_Release((PT)p, u.t); }    
00579     inline static C     Load_Acquire(const volatile C* p)               { C2T_union u; u.t = Ops::Load_Acquire((PT)p); return u.c; }
00580 };
00581 
00582 
00583 
00584 // Atomic value base class - implements operations shared for integers and pointers.
00585 template<class T>
00586 class AtomicValueBase
00587 {
00588 protected:
00589     typedef AtomicOps<T> Ops;
00590 public:
00591 
00592     volatile T  Value;
00593 
00594     inline AtomicValueBase()                  { }
00595     explicit inline AtomicValueBase(T val)    { Ops::Store_Release(&Value, val); }
00596 
00597     // Most libraries (TBB and Joshua Scholar's) library do not do Load_Acquire
00598     // here, since most algorithms do not require atomic loads. Needs some research.    
00599     inline operator T() const { return Value; }
00600 
00601     // *** Standard Atomic inlines
00602     inline T     Exchange_Sync(T val)               { return Ops::Exchange_Sync(&Value,  val); }
00603     inline T     Exchange_Release(T val)            { return Ops::Exchange_Release(&Value, val); }
00604     inline T     Exchange_Acquire(T val)            { return Ops::Exchange_Acquire(&Value, val); }
00605     inline T     Exchange_NoSync(T val)             { return Ops::Exchange_NoSync(&Value, val); }
00606     inline bool  CompareAndSet_Sync(T c, T val)     { return Ops::CompareAndSet_Sync(&Value, c, val); }
00607     inline bool  CompareAndSet_Release(T c, T val)  { return Ops::CompareAndSet_Release(&Value, c, val); }
00608     inline bool  CompareAndSet_Acquire(T c, T val)  { return Ops::CompareAndSet_Relse(&Value, c, val); }
00609     inline bool  CompareAndSet_NoSync(T c, T val)   { return Ops::CompareAndSet_NoSync(&Value, c, val); }
00610     // Load & Store.
00611     inline void  Store_Release(T val)               { Ops::Store_Release(&Value, val); }
00612     inline T     Load_Acquire() const               { return Ops::Load_Acquire(&Value);  }
00613 };
00614 
00615 
00616 // ***** AtomicPtr - Atomic pointer template
00617 
00618 // This pointer class supports atomic assignments with release,
00619 // increment / decrement operations, and conditional compare + set.
00620 
00621 template<class T>
00622 class AtomicPtr : public AtomicValueBase<T*>
00623 {
00624     typedef typename AtomicValueBase<T*>::Ops Ops;
00625 
00626 public:
00627     // Initialize pointer value to 0 by default; use Store_Release only with explicit constructor.
00628     inline AtomicPtr() : AtomicValueBase<T*>()                     { this->Value = 0; }
00629     explicit inline AtomicPtr(T* val) : AtomicValueBase<T*>(val)   { }
00630         
00631     // Pointer access.
00632     inline T* operator -> () const     { return this->Load_Acquire(); }
00633 
00634     // It looks like it is convenient to have Load_Acquire characteristics
00635     // for this, since that is convenient for algorithms such as linked
00636     // list traversals that can be added to bu another thread.
00637     inline operator T* () const        { return this->Load_Acquire(); }
00638 
00639 
00640     // *** Standard Atomic inlines (applicable to pointers)
00641 
00642     // ExhangeAdd considers pointer size for pointers.
00643     template<class I>
00644     inline T*     ExchangeAdd_Sync(I incr)      { return Ops::ExchangeAdd_Sync(&this->Value, ((T*)0) + incr); }
00645     template<class I>
00646     inline T*     ExchangeAdd_Release(I incr)   { return Ops::ExchangeAdd_Release(&this->Value, ((T*)0) + incr); }
00647     template<class I>
00648     inline T*     ExchangeAdd_Acquire(I incr)   { return Ops::ExchangeAdd_Acquire(&this->Value, ((T*)0) + incr); }
00649     template<class I>
00650     inline T*     ExchangeAdd_NoSync(I incr)    { return Ops::ExchangeAdd_NoSync(&this->Value, ((T*)0) + incr); }
00651 
00652     // *** Atomic Operators
00653 
00654     inline T* operator = (T* val)  { this->Store_Release(val); return val; }
00655 
00656     template<class I>
00657     inline T* operator += (I val) { return ExchangeAdd_Sync(val) + val; }
00658     template<class I>
00659     inline T* operator -= (I val) { return operator += (-val); }
00660 
00661     inline T* operator ++ ()      { return ExchangeAdd_Sync(1) + 1; }
00662     inline T* operator -- ()      { return ExchangeAdd_Sync(-1) - 1; }
00663     inline T* operator ++ (int)   { return ExchangeAdd_Sync(1); }
00664     inline T* operator -- (int)   { return ExchangeAdd_Sync(-1); }
00665 };
00666 
00667 
00668 // ***** AtomicInt - Atomic integer template
00669 
00670 // Implements an atomic integer type; the exact type to use is provided 
00671 // as an argument. Supports atomic Acquire / Release semantics, atomic
00672 // arithmetic operations, and atomic conditional compare + set.
00673 
00674 template<class T>
00675 class AtomicInt : public AtomicValueBase<T>
00676 {
00677     typedef typename AtomicValueBase<T>::Ops Ops;
00678 
00679 public:
00680     inline AtomicInt() : AtomicValueBase<T>()                     { }
00681     explicit inline AtomicInt(T val) : AtomicValueBase<T>(val)    { }
00682 
00683 
00684     // *** Standard Atomic inlines (applicable to int)   
00685     inline T     ExchangeAdd_Sync(T val)            { return Ops::ExchangeAdd_Sync(&this->Value, val); }
00686     inline T     ExchangeAdd_Release(T val)         { return Ops::ExchangeAdd_Release(&this->Value, val); }
00687     inline T     ExchangeAdd_Acquire(T val)         { return Ops::ExchangeAdd_Acquire(&this->Value, val); }
00688     inline T     ExchangeAdd_NoSync(T val)          { return Ops::ExchangeAdd_NoSync(&this->Value, val); }
00689     // These increments could be more efficient because they don't return a value.
00690     inline void  Increment_Sync()                   { ExchangeAdd_Sync((T)1); }
00691     inline void  Increment_Release()                { ExchangeAdd_Release((T)1); }
00692     inline void  Increment_Acquire()                { ExchangeAdd_Acquire((T)1); }    
00693     inline void  Increment_NoSync()                 { ExchangeAdd_NoSync((T)1); }
00694 
00695     // *** Atomic Operators
00696 
00697     inline T operator = (T val)  { this->Store_Release(val); return val; }
00698     inline T operator += (T val) { return ExchangeAdd_Sync(val) + val; }
00699     inline T operator -= (T val) { return ExchangeAdd_Sync(0 - val) - val; }
00700 
00701     inline T operator ++ ()      { return ExchangeAdd_Sync((T)1) + 1; }
00702     inline T operator -- ()      { return ExchangeAdd_Sync(((T)0)-1) - 1; }
00703     inline T operator ++ (int)   { return ExchangeAdd_Sync((T)1); }
00704     inline T operator -- (int)   { return ExchangeAdd_Sync(((T)0)-1); }
00705 
00706     // More complex atomic operations. Leave it to compiler whether to optimize them or not.
00707     T operator &= (T arg)
00708     {
00709         T comp, newVal;
00710         do {
00711             comp   = this->Value;
00712             newVal = comp & arg;
00713         } while(!this->CompareAndSet_Sync(comp, newVal));
00714         return newVal;
00715     }
00716 
00717     T operator |= (T arg)
00718     {
00719         T comp, newVal;
00720         do {
00721             comp   = this->Value;
00722             newVal = comp | arg;
00723         } while(!this->CompareAndSet_Sync(comp, newVal));
00724         return newVal;
00725     }
00726 
00727     T operator ^= (T arg)
00728     {
00729         T comp, newVal;
00730         do {
00731             comp   = this->Value;
00732             newVal = comp ^ arg;
00733         } while(!this->CompareAndSet_Sync(comp, newVal));
00734         return newVal;
00735     }
00736 
00737     T operator *= (T arg)
00738     {
00739         T comp, newVal;
00740         do {
00741             comp   = this->Value;
00742             newVal = comp * arg;
00743         } while(!this->CompareAndSet_Sync(comp, newVal));
00744         return newVal;
00745     }
00746 
00747     T operator /= (T arg)
00748     {
00749         T comp, newVal;
00750         do {
00751             comp   = this->Value;
00752             newVal = comp / arg;
00753         } while(!CompareAndSet_Sync(comp, newVal));
00754         return newVal;
00755     }
00756 
00757     T operator >>= (unsigned bits)
00758     {
00759         T comp, newVal;
00760         do {
00761             comp   = this->Value;
00762             newVal = comp >> bits;
00763         } while(!CompareAndSet_Sync(comp, newVal));
00764         return newVal;
00765     }
00766 
00767     T operator <<= (unsigned bits)
00768     {
00769         T comp, newVal;
00770         do {
00771             comp   = this->Value;
00772             newVal = comp << bits;
00773         } while(!this->CompareAndSet_Sync(comp, newVal));
00774         return newVal;
00775     }
00776 };
00777 
00778 
00779 
00780 //-----------------------------------------------------------------------------------
00781 // ***** Lock
00782 
00783 // Lock is a simplest and most efficient mutual-exclusion lock class.
00784 // Unlike Mutex, it cannot be waited on.
00785 
00786 class Lock
00787 {
00788     // NOTE: Locks are not allocatable and they themselves should not allocate 
00789     // memory by standard means. This is the case because StandardAllocator
00790     // relies on this class.
00791     // Make 'delete' private. Don't do this for 'new' since it can be redefined.  
00792     void    operator delete(void*) {}
00793 
00794 
00795     // *** Lock implementation for various platforms.
00796     
00797 #if !defined(OVR_ENABLE_THREADS)
00798 
00799 public:
00800     // With no thread support, lock does nothing.
00801     inline Lock() { }
00802     inline Lock(unsigned) { }
00803     inline ~Lock() { }    
00804     inline void DoLock() { }
00805     inline void Unlock() { }
00806 
00807    // Windows.   
00808 #elif defined(OVR_OS_WIN32)
00809 
00810     CRITICAL_SECTION cs;
00811 public:   
00812     Lock(unsigned spinCount = 0);      
00813     ~Lock();
00814     // Locking functions.
00815     inline void DoLock()    { ::EnterCriticalSection(&cs); }
00816     inline void Unlock()    { ::LeaveCriticalSection(&cs); }
00817 
00818 #else
00819     pthread_mutex_t mutex;
00820 
00821 public:
00822     static pthread_mutexattr_t RecursiveAttr;
00823     static bool                RecursiveAttrInit;
00824 
00825     Lock (unsigned dummy = 0)
00826     {
00827         if (!RecursiveAttrInit)
00828         {
00829             pthread_mutexattr_init(&RecursiveAttr);
00830             pthread_mutexattr_settype(&RecursiveAttr, PTHREAD_MUTEX_RECURSIVE);
00831             RecursiveAttrInit = 1;
00832         }
00833         pthread_mutex_init(&mutex,&RecursiveAttr);
00834     }
00835     ~Lock ()                { pthread_mutex_destroy(&mutex); }
00836     inline void DoLock()    { pthread_mutex_lock(&mutex); }
00837     inline void Unlock()    { pthread_mutex_unlock(&mutex); }
00838 
00839 #endif // OVR_ENABLE_THREDS
00840 
00841 
00842 public:
00843     // Locker class, used for automatic locking
00844     class Locker
00845     {
00846     public:     
00847         Lock *pLock;
00848         inline Locker(Lock *plock)
00849         { pLock = plock; pLock->DoLock(); }
00850         inline ~Locker()
00851         { pLock->Unlock();  }
00852     };
00853 };
00854 
00855 
00856 
00857 } // OVR
00858 
00859 #endif


oculus_sdk
Author(s):
autogenerated on Mon Oct 6 2014 03:01:18