src/util/u_atomic.h

   1 /**
   2  * Many similar implementations exist. See for example libwsbm
   3  * or the linux kernel include/atomic.h
   4  *
   5  * No copyright claimed on this file.
   6  *
   7  */
   8
   9 #include "no_extern_c.h"
  10
  11 #ifndef U_ATOMIC_H
  12 #define U_ATOMIC_H
  13
  14 #include <stdbool.h>
  15
  16 /* Favor OS-provided implementations.
  17  *
  18  * Where no OS-provided implementation is available, fall back to
  19  * locally coded assembly, compiler intrinsic or ultimately a
  20  * mutex-based implementation.
  21  */
  22 #if defined(__sun)
  23 #define PIPE_ATOMIC_OS_SOLARIS
  24 #elif defined(_MSC_VER)
  25 #define PIPE_ATOMIC_MSVC_INTRINSIC
  26 #elif defined(__GNUC__)
  27 #define PIPE_ATOMIC_GCC_INTRINSIC
  28 #else
  29 #error "Unsupported platform"
  30 #endif
  31
  32
  33 /* Implementation using GCC-provided synchronization intrinsics
  34  */
  35 #if defined(PIPE_ATOMIC_GCC_INTRINSIC)
  36
  37 #define PIPE_ATOMIC "GCC Sync Intrinsics"
  38
  39 #define p_atomic_set(_v, _i) (*(_v) = (_i))
  40 #define p_atomic_read(_v) (*(_v))
  41 #define p_atomic_dec_zero(v) (__sync_sub_and_fetch((v), 1) == 0)
  42 #define p_atomic_inc(v) (void) __sync_add_and_fetch((v), 1)
  43 #define p_atomic_dec(v) (void) __sync_sub_and_fetch((v), 1)
  44 #define p_atomic_add(v, i) (void) __sync_add_and_fetch((v), (i))
  45 #define p_atomic_inc_return(v) __sync_add_and_fetch((v), 1)
  46 #define p_atomic_dec_return(v) __sync_sub_and_fetch((v), 1)
  47 #define p_atomic_cmpxchg(v, old, _new) \
  48    __sync_val_compare_and_swap((v), (old), (_new))
  49
  50 #endif
  51
  52
  53
  54 /* Unlocked version for single threaded environments, such as some
  55  * windows kernel modules.
  56  */
  57 #if defined(PIPE_ATOMIC_OS_UNLOCKED)
  58
  59 #define PIPE_ATOMIC "Unlocked"
  60
  61 #define p_atomic_set(_v, _i) (*(_v) = (_i))
  62 #define p_atomic_read(_v) (*(_v))
  63 #define p_atomic_dec_zero(_v) (p_atomic_dec_return(_v) == 0)
  64 #define p_atomic_inc(_v) ((void) p_atomic_inc_return(_v))
  65 #define p_atomic_dec(_v) ((void) p_atomic_dec_return(_v))
  66 #define p_atomic_add(_v, _i) (*(_v) = *(_v) + (_i))
  67 #define p_atomic_inc_return(_v) (++(*(_v)))
  68 #define p_atomic_dec_return(_v) (--(*(_v)))
  69 #define p_atomic_cmpxchg(_v, _old, _new) (*(_v) == (_old) ? (*(_v) = (_new), (_old)) : *(_v))
  70
  71 #endif
  72
  73
  74 #if defined(PIPE_ATOMIC_MSVC_INTRINSIC)
  75
  76 #define PIPE_ATOMIC "MSVC Intrinsics"
  77
  78 /* We use the Windows header's Interlocked*64 functions instead of the
  79  * _Interlocked*64 intrinsics wherever we can, as support for the latter varies
  80  * with target CPU, whereas Windows headers take care of all portability
  81  * issues: using intrinsics where available, falling back to library
  82  * implementations where not.
  83  */
  84 #ifndef WIN32_LEAN_AND_MEAN
  85 #define WIN32_LEAN_AND_MEAN 1
  86 #endif
  87 #include <windows.h>
  88 #include <intrin.h>
  89 #include <assert.h>
  90
  91 #if _MSC_VER < 1600
  92
  93 /* Implement _InterlockedCompareExchange8 in terms of _InterlockedCompareExchange16 */
  94 static __inline char
  95 _InterlockedCompareExchange8(char volatile *destination8, char exchange8, char comparand8)
  96 {
  97    INT_PTR destinationAddr = (INT_PTR)destination8;
  98    short volatile *destination16 = (short volatile *)(destinationAddr & ~1);
  99    const short shift8 = (destinationAddr & 1) * 8;
 100    const short mask8 = 0xff << shift8;
 101    short initial16 = *destination16;
 102    char initial8 = initial16 >> shift8;
 103    while (initial8 == comparand8) {
 104       /* initial *destination8 matches, so try exchange it while keeping the
 105        * neighboring byte untouched */
 106       short exchange16 = (initial16 & ~mask8) | ((short)exchange8 << shift8);
 107       short comparand16 = initial16;
 108       short initial16 = _InterlockedCompareExchange16(destination16, exchange16, comparand16);
 109       if (initial16 == comparand16) {
 110          /* succeeded */
 111          return comparand8;
 112       }
 113       /* something changed, retry with the new initial value */
 114       initial8 = initial16 >> shift8;
 115    }
 116    return initial8;
 117 }
 118
 119 /* Implement _InterlockedExchangeAdd16 in terms of _InterlockedCompareExchange16 */
 120 static __inline short
 121 _InterlockedExchangeAdd16(short volatile *addend, short value)
 122 {
 123    short initial = *addend;
 124    short comparand;
 125    do {
 126       short exchange = initial + value;
 127       comparand = initial;
 128       /* if *addend==comparand then *addend=exchange, return original *addend */
 129       initial = _InterlockedCompareExchange16(addend, exchange, comparand);
 130    } while(initial != comparand);
 131    return comparand;
 132 }
 133
 134 /* Implement _InterlockedExchangeAdd8 in terms of _InterlockedCompareExchange8 */
 135 static __inline char
 136 _InterlockedExchangeAdd8(char volatile *addend, char value)
 137 {
 138    char initial = *addend;
 139    char comparand;
 140    do {
 141       char exchange = initial + value;
 142       comparand = initial;
 143       initial = _InterlockedCompareExchange8(addend, exchange, comparand);
 144    } while(initial != comparand);
 145    return comparand;
 146 }
 147
 148 #endif /* _MSC_VER < 1600 */
 149
 150 /* MSVC supports decltype keyword, but it's only supported on C++ and doesn't
 151  * quite work here; and if a C++-only solution is worthwhile, then it would be
 152  * better to use templates / function overloading, instead of decltype magic.
 153  * Therefore, we rely on implicit casting to LONGLONG for the functions that return
 154  */
 155
 156 #define p_atomic_set(_v, _i) (*(_v) = (_i))
 157 #define p_atomic_read(_v) (*(_v))
 158
 159 #define p_atomic_dec_zero(_v) \
 160    (p_atomic_dec_return(_v) == 0)
 161
 162 #define p_atomic_inc(_v) \
 163    ((void) p_atomic_inc_return(_v))
 164
 165 #define p_atomic_inc_return(_v) (\
 166    sizeof *(_v) == sizeof(short)   ? _InterlockedIncrement16((short *)  (_v)) : \
 167    sizeof *(_v) == sizeof(long)    ? _InterlockedIncrement  ((long *)   (_v)) : \
 168    sizeof *(_v) == sizeof(__int64) ? InterlockedIncrement64 ((__int64 *)(_v)) : \
 169                                      (assert(!"should not get here"), 0))
 170
 171 #define p_atomic_dec(_v) \
 172    ((void) p_atomic_dec_return(_v))
 173
 174 #define p_atomic_dec_return(_v) (\
 175    sizeof *(_v) == sizeof(short)   ? _InterlockedDecrement16((short *)  (_v)) : \
 176    sizeof *(_v) == sizeof(long)    ? _InterlockedDecrement  ((long *)   (_v)) : \
 177    sizeof *(_v) == sizeof(__int64) ? InterlockedDecrement64 ((__int64 *)(_v)) : \
 178                                      (assert(!"should not get here"), 0))
 179
 180 #define p_atomic_add(_v, _i) (\
 181    sizeof *(_v) == sizeof(char)    ? _InterlockedExchangeAdd8 ((char *)   (_v), (_i)) : \
 182    sizeof *(_v) == sizeof(short)   ? _InterlockedExchangeAdd16((short *)  (_v), (_i)) : \
 183    sizeof *(_v) == sizeof(long)    ? _InterlockedExchangeAdd  ((long *)   (_v), (_i)) : \
 184    sizeof *(_v) == sizeof(__int64) ? InterlockedExchangeAdd64((__int64 *)(_v), (_i)) : \
 185                                      (assert(!"should not get here"), 0))
 186
 187 #define p_atomic_cmpxchg(_v, _old, _new) (\
 188    sizeof *(_v) == sizeof(char)    ? _InterlockedCompareExchange8 ((char *)   (_v), (char)   (_new), (char)   (_old)) : \
 189    sizeof *(_v) == sizeof(short)   ? _InterlockedCompareExchange16((short *)  (_v), (short)  (_new), (short)  (_old)) : \
 190    sizeof *(_v) == sizeof(long)    ? _InterlockedCompareExchange  ((long *)   (_v), (long)   (_new), (long)   (_old)) : \
 191    sizeof *(_v) == sizeof(__int64) ? InterlockedCompareExchange64 ((__int64 *)(_v), (__int64)(_new), (__int64)(_old)) : \
 192                                      (assert(!"should not get here"), 0))
 193
 194 #endif
 195
 196 #if defined(PIPE_ATOMIC_OS_SOLARIS)
 197
 198 #define PIPE_ATOMIC "Solaris OS atomic functions"
 199
 200 #include <atomic.h>
 201 #include <assert.h>
 202
 203 #define p_atomic_set(_v, _i) (*(_v) = (_i))
 204 #define p_atomic_read(_v) (*(_v))
 205
 206 #define p_atomic_dec_zero(v) (\
 207    sizeof(*v) == sizeof(uint8_t)  ? atomic_dec_8_nv ((uint8_t  *)(v)) == 0 : \
 208    sizeof(*v) == sizeof(uint16_t) ? atomic_dec_16_nv((uint16_t *)(v)) == 0 : \
 209    sizeof(*v) == sizeof(uint32_t) ? atomic_dec_32_nv((uint32_t *)(v)) == 0 : \
 210    sizeof(*v) == sizeof(uint64_t) ? atomic_dec_64_nv((uint64_t *)(v)) == 0 : \
 211                                     (assert(!"should not get here"), 0))
 212
 213 #define p_atomic_inc(v) (void) (\
 214    sizeof(*v) == sizeof(uint8_t)  ? atomic_inc_8 ((uint8_t  *)(v)) : \
 215    sizeof(*v) == sizeof(uint16_t) ? atomic_inc_16((uint16_t *)(v)) : \
 216    sizeof(*v) == sizeof(uint32_t) ? atomic_inc_32((uint32_t *)(v)) : \
 217    sizeof(*v) == sizeof(uint64_t) ? atomic_inc_64((uint64_t *)(v)) : \
 218                                     (assert(!"should not get here"), 0))
 219
 220 #define p_atomic_inc_return(v) ((__typeof(*v)) \
 221    sizeof(*v) == sizeof(uint8_t)  ? atomic_inc_8_nv ((uint8_t  *)(v)) : \
 222    sizeof(*v) == sizeof(uint16_t) ? atomic_inc_16_nv((uint16_t *)(v)) : \
 223    sizeof(*v) == sizeof(uint32_t) ? atomic_inc_32_nv((uint32_t *)(v)) : \
 224    sizeof(*v) == sizeof(uint64_t) ? atomic_inc_64_nv((uint64_t *)(v)) : \
 225                                     (assert(!"should not get here"), 0))
 226
 227 #define p_atomic_dec(v) ((void) \
 228    sizeof(*v) == sizeof(uint8_t)  ? atomic_dec_8 ((uint8_t  *)(v)) : \
 229    sizeof(*v) == sizeof(uint16_t) ? atomic_dec_16((uint16_t *)(v)) : \
 230    sizeof(*v) == sizeof(uint32_t) ? atomic_dec_32((uint32_t *)(v)) : \
 231    sizeof(*v) == sizeof(uint64_t) ? atomic_dec_64((uint64_t *)(v)) : \
 232                                     (assert(!"should not get here"), 0))
 233
 234 #define p_atomic_dec_return(v) ((__typeof(*v)) \
 235    sizeof(*v) == sizeof(uint8_t)  ? atomic_dec_8_nv ((uint8_t  *)(v)) : \
 236    sizeof(*v) == sizeof(uint16_t) ? atomic_dec_16_nv((uint16_t *)(v)) : \
 237    sizeof(*v) == sizeof(uint32_t) ? atomic_dec_32_nv((uint32_t *)(v)) : \
 238    sizeof(*v) == sizeof(uint64_t) ? atomic_dec_64_nv((uint64_t *)(v)) : \
 239                                     (assert(!"should not get here"), 0))
 240
 241 #define p_atomic_add(v, i) ((void)                                   \
 242    sizeof(*v) == sizeof(uint8_t)  ? atomic_add_8 ((uint8_t  *)(v), (i)) : \
 243    sizeof(*v) == sizeof(uint16_t) ? atomic_add_16((uint16_t *)(v), (i)) : \
 244    sizeof(*v) == sizeof(uint32_t) ? atomic_add_32((uint32_t *)(v), (i)) : \
 245    sizeof(*v) == sizeof(uint64_t) ? atomic_add_64((uint64_t *)(v), (i)) : \
 246                                     (assert(!"should not get here"), 0))
 247
 248 #define p_atomic_cmpxchg(v, old, _new) ((__typeof(*v)) \
 249    sizeof(*v) == sizeof(uint8_t)  ? atomic_cas_8 ((uint8_t  *)(v), (uint8_t )(old), (uint8_t )(_new)) : \
 250    sizeof(*v) == sizeof(uint16_t) ? atomic_cas_16((uint16_t *)(v), (uint16_t)(old), (uint16_t)(_new)) : \
 251    sizeof(*v) == sizeof(uint32_t) ? atomic_cas_32((uint32_t *)(v), (uint32_t)(old), (uint32_t)(_new)) : \
 252    sizeof(*v) == sizeof(uint64_t) ? atomic_cas_64((uint64_t *)(v), (uint64_t)(old), (uint64_t)(_new)) : \
 253                                     (assert(!"should not get here"), 0))
 254
 255 #endif
 256
 257 #ifndef PIPE_ATOMIC
 258 #error "No pipe_atomic implementation selected"
 259 #endif
 260
 261
 262
 263 #endif /* U_ATOMIC_H */