Print this page
5044 define static inlines for most often used atomic functions

Split Close
Expand all
Collapse all
          --- old/usr/src/uts/intel/asm/atomic.h
          +++ new/usr/src/uts/intel/asm/atomic.h
↓ open down ↓ 28 lines elided ↑ open up ↑
  29   29  
  30   30  #include <sys/ccompile.h>
  31   31  #include <sys/types.h>
  32   32  
  33   33  #ifdef  __cplusplus
  34   34  extern "C" {
  35   35  #endif
  36   36  
  37   37  #if !defined(__lint) && defined(__GNUC__)
  38   38  
  39      -#if defined(__amd64)
       39 +/*
       40 + * This file contains a number of static inline functions implementing
       41 + * various atomic variable functions.  Note that these are *not* all of the
       42 + * atomic_* functions as defined in usr/src/uts/common/sys/atomic.h.  All
       43 + * possible atomic_* functions are implemented in usr/src/common/atomic in
       44 + * pure assembly.  In the absence of an identically named function in this
       45 + * header file, any use of the function will result in the compiler emitting
       46 + * a function call as usual.  On the other hand, if an identically named
       47 + * function exists in this header as a static inline, the compiler will
       48 + * inline its contents and the linker never sees the symbol reference.  We
       49 + * use this to avoid implementing some of the more complex and less used
       50 + * functions and instead falling back to function calls.  Note that in some
       51 + * cases (e.g., atomic_inc_64) we implement a static inline only on AMD64
       52 + * but not i386.
       53 + */
       54 +
       55 +/*
       56 + * Instruction suffixes for various operand sizes (assuming AMD64)
       57 + */
       58 +#define SUF_8           "b"
       59 +#define SUF_16          "w"
       60 +#define SUF_32          "l"
       61 +#define SUF_64          "q"
  40   62  
       63 +#if defined(__amd64)
       64 +#define SUF_LONG        SUF_64
       65 +#define SUF_PTR         SUF_64
       66 +#define __ATOMIC_OP64(...)      __ATOMIC_OPXX(__VA_ARGS__)
  41   67  #elif defined(__i386)
       68 +#define SUF_LONG        SUF_32
       69 +#define SUF_PTR         SUF_32
       70 +#define __ATOMIC_OP64(...)
       71 +#else
       72 +#error "port me"
       73 +#endif
       74 +
       75 +#if defined(__amd64) || defined(__i386)
       76 +
       77 +#define __ATOMIC_OPXX(fxn, type, op)                                    \
       78 +extern __GNU_INLINE void                                                \
       79 +fxn(volatile type *target)                                              \
       80 +{                                                                       \
       81 +        __asm__ __volatile__(                                           \
       82 +            "lock; " op " %0"                                           \
       83 +            : "+m" (*target));                                          \
       84 +}
       85 +
       86 +__ATOMIC_OPXX(atomic_inc_8,      uint8_t,  "inc" SUF_8)
       87 +__ATOMIC_OPXX(atomic_inc_16,     uint16_t, "inc" SUF_16)
       88 +__ATOMIC_OPXX(atomic_inc_32,     uint32_t, "inc" SUF_32)
       89 +__ATOMIC_OP64(atomic_inc_64,     uint64_t, "inc" SUF_64)
       90 +__ATOMIC_OPXX(atomic_inc_uchar,  uchar_t,  "inc" SUF_8)
       91 +__ATOMIC_OPXX(atomic_inc_ushort, ushort_t, "inc" SUF_16)
       92 +__ATOMIC_OPXX(atomic_inc_uint,   uint_t,   "inc" SUF_32)
       93 +__ATOMIC_OPXX(atomic_inc_ulong,  ulong_t,  "inc" SUF_LONG)
       94 +
       95 +__ATOMIC_OPXX(atomic_dec_8,      uint8_t,  "dec" SUF_8)
       96 +__ATOMIC_OPXX(atomic_dec_16,     uint16_t, "dec" SUF_16)
       97 +__ATOMIC_OPXX(atomic_dec_32,     uint32_t, "dec" SUF_32)
       98 +__ATOMIC_OP64(atomic_dec_64,     uint64_t, "dec" SUF_64)
       99 +__ATOMIC_OPXX(atomic_dec_uchar,  uchar_t,  "dec" SUF_8)
      100 +__ATOMIC_OPXX(atomic_dec_ushort, ushort_t, "dec" SUF_16)
      101 +__ATOMIC_OPXX(atomic_dec_uint,   uint_t,   "dec" SUF_32)
      102 +__ATOMIC_OPXX(atomic_dec_ulong,  ulong_t,  "dec" SUF_LONG)
      103 +
      104 +#undef __ATOMIC_OPXX
      105 +
      106 +#define __ATOMIC_OPXX(fxn, type1, type2, op)                            \
      107 +extern __GNU_INLINE void                                                \
      108 +fxn(volatile type1 *target, type2 delta)                                \
      109 +{                                                                       \
      110 +        __asm__ __volatile__(                                           \
      111 +            "lock; " op " %1,%0"                                        \
      112 +            : "+m" (*target)                                            \
      113 +            : "ir" (delta));                                            \
      114 +}
      115 +
      116 +__ATOMIC_OPXX(atomic_add_8,     uint8_t,  int8_t,      "add" SUF_8)
      117 +__ATOMIC_OPXX(atomic_add_16,    uint16_t, int16_t,     "add" SUF_16)
      118 +__ATOMIC_OPXX(atomic_add_32,    uint32_t, int32_t,     "add" SUF_32)
      119 +__ATOMIC_OP64(atomic_add_64,    uint64_t, int64_t,     "add" SUF_64)
      120 +__ATOMIC_OPXX(atomic_add_char,  uchar_t,  signed char, "add" SUF_8)
      121 +__ATOMIC_OPXX(atomic_add_short, ushort_t, short,       "add" SUF_16)
      122 +__ATOMIC_OPXX(atomic_add_int,   uint_t,   int,         "add" SUF_32)
      123 +__ATOMIC_OPXX(atomic_add_long,  ulong_t,  long,        "add" SUF_LONG)
      124 +
      125 +/*
      126 + * We don't use the above macro here because atomic_add_ptr has an
      127 + * inconsistent type.  The first argument should really be a 'volatile void
      128 + * **'.
      129 + */
      130 +extern __GNU_INLINE void
      131 +atomic_add_ptr(volatile void *target, ssize_t delta)
      132 +{
      133 +        volatile void **tmp = (volatile void **)target;
      134 +
      135 +        __asm__ __volatile__(
      136 +            "lock; add" SUF_PTR " %1,%0"
      137 +            : "+m" (*tmp)
      138 +            : "ir" (delta));
      139 +}
      140 +
      141 +__ATOMIC_OPXX(atomic_or_8,       uint8_t,  uint8_t,  "or" SUF_8)
      142 +__ATOMIC_OPXX(atomic_or_16,      uint16_t, uint16_t, "or" SUF_16)
      143 +__ATOMIC_OPXX(atomic_or_32,      uint32_t, uint32_t, "or" SUF_32)
      144 +__ATOMIC_OP64(atomic_or_64,      uint64_t, uint64_t, "or" SUF_64)
      145 +__ATOMIC_OPXX(atomic_or_uchar,   uchar_t,  uchar_t,  "or" SUF_8)
      146 +__ATOMIC_OPXX(atomic_or_ushort,  ushort_t, ushort_t, "or" SUF_16)
      147 +__ATOMIC_OPXX(atomic_or_uint,    uint_t,   uint_t,   "or" SUF_32)
      148 +__ATOMIC_OPXX(atomic_or_ulong,   ulong_t,  ulong_t,  "or" SUF_LONG)
      149 +
      150 +__ATOMIC_OPXX(atomic_and_8,      uint8_t,  uint8_t,  "and" SUF_8)
      151 +__ATOMIC_OPXX(atomic_and_16,     uint16_t, uint16_t, "and" SUF_16)
      152 +__ATOMIC_OPXX(atomic_and_32,     uint32_t, uint32_t, "and" SUF_32)
      153 +__ATOMIC_OP64(atomic_and_64,     uint64_t, uint64_t, "and" SUF_64)
      154 +__ATOMIC_OPXX(atomic_and_uchar,  uchar_t,  uchar_t,  "and" SUF_8)
      155 +__ATOMIC_OPXX(atomic_and_ushort, ushort_t, ushort_t, "and" SUF_16)
      156 +__ATOMIC_OPXX(atomic_and_uint,   uint_t,   uint_t,   "and" SUF_32)
      157 +__ATOMIC_OPXX(atomic_and_ulong,  ulong_t,  ulong_t,  "and" SUF_LONG)
      158 +
      159 +#undef __ATOMIC_OPXX
      160 +
      161 +#define __ATOMIC_OPXX(fxn, type, op, reg)                               \
      162 +extern __GNU_INLINE type                                                \
      163 +fxn(volatile type *target, type cmp, type new)                          \
      164 +{                                                                       \
      165 +        type ret;                                                       \
      166 +        __asm__ __volatile__(                                           \
      167 +            "lock; " op " %2,%0"                                        \
      168 +            : "+m" (*target), "=a" (ret)                                \
      169 +            : reg (new), "1" (cmp)                                      \
      170 +            : "cc");                                                    \
      171 +        return (ret);                                                   \
      172 +}
      173 +
      174 +__ATOMIC_OPXX(atomic_cas_8,      uint8_t,  "cmpxchg" SUF_8,    "q")
      175 +__ATOMIC_OPXX(atomic_cas_16,     uint16_t, "cmpxchg" SUF_16,   "r")
      176 +__ATOMIC_OPXX(atomic_cas_32,     uint32_t, "cmpxchg" SUF_32,   "r")
      177 +__ATOMIC_OP64(atomic_cas_64,     uint64_t, "cmpxchg" SUF_64,   "r")
      178 +__ATOMIC_OPXX(atomic_cas_uchar,  uchar_t,  "cmpxchg" SUF_8,    "q")
      179 +__ATOMIC_OPXX(atomic_cas_ushort, ushort_t, "cmpxchg" SUF_16,   "r")
      180 +__ATOMIC_OPXX(atomic_cas_uint,   uint_t,   "cmpxchg" SUF_32,   "r")
      181 +__ATOMIC_OPXX(atomic_cas_ulong,  ulong_t,  "cmpxchg" SUF_LONG, "r")
      182 +
      183 +#undef __ATOMIC_OPXX
      184 +
      185 +/*
      186 + * We don't use the above macro here because atomic_cas_ptr has an
      187 + * inconsistent type.  The first argument should really be a 'volatile void
      188 + * **'.
      189 + */
      190 +extern __GNU_INLINE void *
      191 +atomic_cas_ptr(volatile void *target, void *cmp, void *new)
      192 +{
      193 +        volatile void **tmp = (volatile void **)target;
      194 +        void *ret;
      195 +
      196 +        __asm__ __volatile__(
      197 +            "lock; cmpxchg" SUF_PTR " %2,%0"
      198 +            : "+m" (*tmp), "=a" (ret)
      199 +            : "r" (new), "1" (cmp)
      200 +            : "cc");
      201 +
      202 +        return (ret);
      203 +}
      204 +
      205 +#define __ATOMIC_OPXX(fxn, type, op, reg)                               \
      206 +extern __GNU_INLINE type                                                \
      207 +fxn(volatile type *target, type val)                                    \
      208 +{                                                                       \
      209 +        __asm__ __volatile__(                                           \
      210 +            op " %1,%0"                                                 \
      211 +            : "+m" (*target), "+" reg (val));                           \
      212 +        return (val);                                                   \
      213 +}
      214 +
      215 +__ATOMIC_OPXX(atomic_swap_8,      uint8_t,  "xchg" SUF_8,    "q")
      216 +__ATOMIC_OPXX(atomic_swap_16,     uint16_t, "xchg" SUF_16,   "r")
      217 +__ATOMIC_OPXX(atomic_swap_32,     uint32_t, "xchg" SUF_32,   "r")
      218 +__ATOMIC_OP64(atomic_swap_64,     uint64_t, "xchg" SUF_64,   "r")
      219 +__ATOMIC_OPXX(atomic_swap_uchar,  uchar_t,  "xchg" SUF_8,    "q")
      220 +__ATOMIC_OPXX(atomic_swap_ushort, ushort_t, "xchg" SUF_16,   "r")
      221 +__ATOMIC_OPXX(atomic_swap_uint,   uint_t,   "xchg" SUF_32,   "r")
      222 +__ATOMIC_OPXX(atomic_swap_ulong,  ulong_t,  "xchg" SUF_LONG, "r")
      223 +
      224 +#undef __ATOMIC_OPXX
      225 +
      226 +/*
      227 + * We don't use the above macro here because atomic_swap_ptr has an
      228 + * inconsistent type.  The first argument should really be a 'volatile void
      229 + * **'.
      230 + */
      231 +extern __GNU_INLINE void *
      232 +atomic_swap_ptr(volatile void *target, void *val)
      233 +{
      234 +        volatile void **tmp = (volatile void **)target;
      235 +
      236 +        __asm__ __volatile__(
      237 +            "xchg" SUF_PTR " %1,%0"
      238 +            : "+m" (*tmp), "+r" (val));
      239 +
      240 +        return (val);
      241 +}
  42  242  
  43  243  #else
  44  244  #error  "port me"
  45  245  #endif
  46  246  
      247 +#undef SUF_8
      248 +#undef SUF_16
      249 +#undef SUF_32
      250 +#undef SUF_64
      251 +#undef SUF_LONG
      252 +#undef SUF_PTR
      253 +
      254 +#undef __ATOMIC_OP64
      255 +
  47  256  #endif /* !__lint && __GNUC__ */
  48  257  
  49  258  #ifdef __cplusplus
  50  259  }
  51  260  #endif
  52  261  
  53  262  #endif  /* _ASM_ATOMIC_H */
    
XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX