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 #include <stdint.h>
16
17 /* Favor OS-provided implementations.
18 *
19 * Where no OS-provided implementation is available, fall back to
20 * locally coded assembly, compiler intrinsic or ultimately a
21 * mutex-based implementation.
22 */
23 #if defined(__sun)
24 #define PIPE_ATOMIC_OS_SOLARIS
25 #elif defined(_MSC_VER)
26 #define PIPE_ATOMIC_MSVC_INTRINSIC
27 #elif defined(__GNUC__)
28 #define PIPE_ATOMIC_GCC_INTRINSIC
29 #else
30 #error "Unsupported platform"
31 #endif
32
33
34 /* Implementation using GCC-provided synchronization intrinsics
35 */
36 #if defined(PIPE_ATOMIC_GCC_INTRINSIC)
37
38 #define PIPE_ATOMIC "GCC Sync Intrinsics"
39
40 #if defined(USE_GCC_ATOMIC_BUILTINS)
41
42 /* The builtins with explicit memory model are available since GCC 4.7. */
43 #define p_atomic_set(_v, _i) __atomic_store_n((_v), (_i), __ATOMIC_RELEASE)
44 #define p_atomic_read(_v) __atomic_load_n((_v), __ATOMIC_ACQUIRE)
45 #define p_atomic_read_relaxed(_v) __atomic_load_n((_v), __ATOMIC_RELAXED)
46 #define p_atomic_dec_zero(v) (__atomic_sub_fetch((v), 1, __ATOMIC_ACQ_REL) == 0)
47 #define p_atomic_inc(v) (void) __atomic_add_fetch((v), 1, __ATOMIC_ACQ_REL)
48 #define p_atomic_dec(v) (void) __atomic_sub_fetch((v), 1, __ATOMIC_ACQ_REL)
49 #define p_atomic_add(v, i) (void) __atomic_add_fetch((v), (i), __ATOMIC_ACQ_REL)
50 #define p_atomic_inc_return(v) __atomic_add_fetch((v), 1, __ATOMIC_ACQ_REL)
51 #define p_atomic_dec_return(v) __atomic_sub_fetch((v), 1, __ATOMIC_ACQ_REL)
52 #define p_atomic_add_return(v, i) __atomic_add_fetch((v), (i), __ATOMIC_ACQ_REL)
53 #define p_atomic_xchg(v, i) __atomic_exchange_n((v), (i), __ATOMIC_ACQ_REL)
54 #define PIPE_NATIVE_ATOMIC_XCHG
55
56 #else
57
58 #define p_atomic_set(_v, _i) (*(_v) = (_i))
59 #define p_atomic_read(_v) (*(_v))
60 #define p_atomic_read_relaxed(_v) (*(_v))
61 #define p_atomic_dec_zero(v) (__sync_sub_and_fetch((v), 1) == 0)
62 #define p_atomic_inc(v) (void) __sync_add_and_fetch((v), 1)
63 #define p_atomic_dec(v) (void) __sync_sub_and_fetch((v), 1)
64 #define p_atomic_add(v, i) (void) __sync_add_and_fetch((v), (i))
65 #define p_atomic_inc_return(v) __sync_add_and_fetch((v), 1)
66 #define p_atomic_dec_return(v) __sync_sub_and_fetch((v), 1)
67 #define p_atomic_add_return(v, i) __sync_add_and_fetch((v), (i))
68
69 #endif
70
71 /* There is no __atomic_* compare and exchange that returns the current value.
72 * Also, GCC 5.4 seems unable to optimize a compound statement expression that
73 * uses an additional stack variable with __atomic_compare_exchange[_n].
74 */
75 #define p_atomic_cmpxchg(v, old, _new) \
76 __sync_val_compare_and_swap((v), (old), (_new))
77
78 #endif
79
80
81
82 /* Unlocked version for single threaded environments, such as some
83 * windows kernel modules.
84 */
85 #if defined(PIPE_ATOMIC_OS_UNLOCKED)
86
87 #define PIPE_ATOMIC "Unlocked"
88
89 #define p_atomic_set(_v, _i) (*(_v) = (_i))
90 #define p_atomic_read(_v) (*(_v))
91 #define p_atomic_read_relaxed(_v) (*(_v))
92 #define p_atomic_dec_zero(_v) (p_atomic_dec_return(_v) == 0)
93 #define p_atomic_inc(_v) ((void) p_atomic_inc_return(_v))
94 #define p_atomic_dec(_v) ((void) p_atomic_dec_return(_v))
95 #define p_atomic_add(_v, _i) ((void) p_atomic_add_return((_v), (_i)))
96 #define p_atomic_inc_return(_v) (++(*(_v)))
97 #define p_atomic_dec_return(_v) (--(*(_v)))
98 #define p_atomic_add_return(_v, _i) (*(_v) = *(_v) + (_i))
99 #define p_atomic_cmpxchg(_v, _old, _new) (*(_v) == (_old) ? (*(_v) = (_new), (_old)) : *(_v))
100
101 #endif
102
103
104 #if defined(PIPE_ATOMIC_MSVC_INTRINSIC)
105
106 #define PIPE_ATOMIC "MSVC Intrinsics"
107
108 /* We use the Windows header's Interlocked*64 functions instead of the
109 * _Interlocked*64 intrinsics wherever we can, as support for the latter varies
110 * with target CPU, whereas Windows headers take care of all portability
111 * issues: using intrinsics where available, falling back to library
112 * implementations where not.
113 */
114 #ifndef WIN32_LEAN_AND_MEAN
115 #define WIN32_LEAN_AND_MEAN 1
116 #endif
117 #include <windows.h>
118 #include <intrin.h>
119 #include <assert.h>
120
121 /* MSVC supports decltype keyword, but it's only supported on C++ and doesn't
122 * quite work here; and if a C++-only solution is worthwhile, then it would be
123 * better to use templates / function overloading, instead of decltype magic.
124 * Therefore, we rely on implicit casting to LONGLONG for the functions that return
125 */
126
127 #define p_atomic_set(_v, _i) (*(_v) = (_i))
128 #define p_atomic_read(_v) (*(_v))
129 #define p_atomic_read_relaxed(_v) (*(_v))
130
131 #define p_atomic_dec_zero(_v) \
132 (p_atomic_dec_return(_v) == 0)
133
134 #define p_atomic_inc(_v) \
135 ((void) p_atomic_inc_return(_v))
136
137 #define p_atomic_inc_return(_v) (\
138 sizeof *(_v) == sizeof(short) ? _InterlockedIncrement16((short *) (_v)) : \
139 sizeof *(_v) == sizeof(long) ? _InterlockedIncrement ((long *) (_v)) : \
140 sizeof *(_v) == sizeof(__int64) ? InterlockedIncrement64 ((__int64 *)(_v)) : \
141 (assert(!"should not get here"), 0))
142
143 #define p_atomic_dec(_v) \
144 ((void) p_atomic_dec_return(_v))
145
146 #define p_atomic_dec_return(_v) (\
147 sizeof *(_v) == sizeof(short) ? _InterlockedDecrement16((short *) (_v)) : \
148 sizeof *(_v) == sizeof(long) ? _InterlockedDecrement ((long *) (_v)) : \
149 sizeof *(_v) == sizeof(__int64) ? InterlockedDecrement64 ((__int64 *)(_v)) : \
150 (assert(!"should not get here"), 0))
151
152 #define p_atomic_add(_v, _i) \
153 ((void) p_atomic_add_return((_v), (_i)))
154
155 #define p_atomic_add_return(_v, _i) (\
156 sizeof *(_v) == sizeof(char) ? _InterlockedExchangeAdd8 ((char *) (_v), (_i)) : \
157 sizeof *(_v) == sizeof(short) ? _InterlockedExchangeAdd16((short *) (_v), (_i)) : \
158 sizeof *(_v) == sizeof(long) ? _InterlockedExchangeAdd ((long *) (_v), (_i)) : \
159 sizeof *(_v) == sizeof(__int64) ? InterlockedExchangeAdd64((__int64 *)(_v), (_i)) : \
160 (assert(!"should not get here"), 0))
161
162 #define p_atomic_cmpxchg(_v, _old, _new) (\
163 sizeof *(_v) == sizeof(char) ? _InterlockedCompareExchange8 ((char *) (_v), (char) (_new), (char) (_old)) : \
164 sizeof *(_v) == sizeof(short) ? _InterlockedCompareExchange16((short *) (_v), (short) (_new), (short) (_old)) : \
165 sizeof *(_v) == sizeof(long) ? _InterlockedCompareExchange ((long *) (_v), (long) (_new), (long) (_old)) : \
166 sizeof *(_v) == sizeof(__int64) ? InterlockedCompareExchange64 ((__int64 *)(_v), (__int64)(_new), (__int64)(_old)) : \
167 (assert(!"should not get here"), 0))
168
169 #endif
170
171 #if defined(PIPE_ATOMIC_OS_SOLARIS)
172
173 #define PIPE_ATOMIC "Solaris OS atomic functions"
174
175 #include <atomic.h>
176 #include <assert.h>
177
178 #define p_atomic_set(_v, _i) (*(_v) = (_i))
179 #define p_atomic_read(_v) (*(_v))
180
181 #define p_atomic_dec_zero(v) (\
182 sizeof(*v) == sizeof(uint8_t) ? atomic_dec_8_nv ((uint8_t *)(v)) == 0 : \
183 sizeof(*v) == sizeof(uint16_t) ? atomic_dec_16_nv((uint16_t *)(v)) == 0 : \
184 sizeof(*v) == sizeof(uint32_t) ? atomic_dec_32_nv((uint32_t *)(v)) == 0 : \
185 sizeof(*v) == sizeof(uint64_t) ? atomic_dec_64_nv((uint64_t *)(v)) == 0 : \
186 (assert(!"should not get here"), 0))
187
188 #define p_atomic_inc(v) (void) (\
189 sizeof(*v) == sizeof(uint8_t) ? atomic_inc_8 ((uint8_t *)(v)) : \
190 sizeof(*v) == sizeof(uint16_t) ? atomic_inc_16((uint16_t *)(v)) : \
191 sizeof(*v) == sizeof(uint32_t) ? atomic_inc_32((uint32_t *)(v)) : \
192 sizeof(*v) == sizeof(uint64_t) ? atomic_inc_64((uint64_t *)(v)) : \
193 (assert(!"should not get here"), 0))
194
195 #define p_atomic_inc_return(v) (__typeof(*v))( \
196 sizeof(*v) == sizeof(uint8_t) ? atomic_inc_8_nv ((uint8_t *)(v)) : \
197 sizeof(*v) == sizeof(uint16_t) ? atomic_inc_16_nv((uint16_t *)(v)) : \
198 sizeof(*v) == sizeof(uint32_t) ? atomic_inc_32_nv((uint32_t *)(v)) : \
199 sizeof(*v) == sizeof(uint64_t) ? atomic_inc_64_nv((uint64_t *)(v)) : \
200 (assert(!"should not get here"), 0))
201
202 #define p_atomic_dec(v) (void) ( \
203 sizeof(*v) == sizeof(uint8_t) ? atomic_dec_8 ((uint8_t *)(v)) : \
204 sizeof(*v) == sizeof(uint16_t) ? atomic_dec_16((uint16_t *)(v)) : \
205 sizeof(*v) == sizeof(uint32_t) ? atomic_dec_32((uint32_t *)(v)) : \
206 sizeof(*v) == sizeof(uint64_t) ? atomic_dec_64((uint64_t *)(v)) : \
207 (assert(!"should not get here"), 0))
208
209 #define p_atomic_dec_return(v) (__typeof(*v))( \
210 sizeof(*v) == sizeof(uint8_t) ? atomic_dec_8_nv ((uint8_t *)(v)) : \
211 sizeof(*v) == sizeof(uint16_t) ? atomic_dec_16_nv((uint16_t *)(v)) : \
212 sizeof(*v) == sizeof(uint32_t) ? atomic_dec_32_nv((uint32_t *)(v)) : \
213 sizeof(*v) == sizeof(uint64_t) ? atomic_dec_64_nv((uint64_t *)(v)) : \
214 (assert(!"should not get here"), 0))
215
216 #define p_atomic_add(v, i) (void) ( \
217 sizeof(*v) == sizeof(uint8_t) ? atomic_add_8 ((uint8_t *)(v), (i)) : \
218 sizeof(*v) == sizeof(uint16_t) ? atomic_add_16((uint16_t *)(v), (i)) : \
219 sizeof(*v) == sizeof(uint32_t) ? atomic_add_32((uint32_t *)(v), (i)) : \
220 sizeof(*v) == sizeof(uint64_t) ? atomic_add_64((uint64_t *)(v), (i)) : \
221 (assert(!"should not get here"), 0))
222
223 #define p_atomic_add_return(v, i) (__typeof(*v)) ( \
224 sizeof(*v) == sizeof(uint8_t) ? atomic_add_8_nv ((uint8_t *)(v), (i)) : \
225 sizeof(*v) == sizeof(uint16_t) ? atomic_add_16_nv((uint16_t *)(v), (i)) : \
226 sizeof(*v) == sizeof(uint32_t) ? atomic_add_32_nv((uint32_t *)(v), (i)) : \
227 sizeof(*v) == sizeof(uint64_t) ? atomic_add_64_nv((uint64_t *)(v), (i)) : \
228 (assert(!"should not get here"), 0))
229
230 #define p_atomic_cmpxchg(v, old, _new) (__typeof(*v))( \
231 sizeof(*v) == sizeof(uint8_t) ? atomic_cas_8 ((uint8_t *)(v), (uint8_t )(old), (uint8_t )(_new)) : \
232 sizeof(*v) == sizeof(uint16_t) ? atomic_cas_16((uint16_t *)(v), (uint16_t)(old), (uint16_t)(_new)) : \
233 sizeof(*v) == sizeof(uint32_t) ? atomic_cas_32((uint32_t *)(v), (uint32_t)(old), (uint32_t)(_new)) : \
234 sizeof(*v) == sizeof(uint64_t) ? atomic_cas_64((uint64_t *)(v), (uint64_t)(old), (uint64_t)(_new)) : \
235 (assert(!"should not get here"), 0))
236
237 #endif
238
239 #ifndef PIPE_ATOMIC
240 #error "No pipe_atomic implementation selected"
241 #endif
242
243 #ifndef PIPE_NATIVE_ATOMIC_XCHG
p_atomic_xchg_32(uint32_t * v,uint32_t i)244 static inline uint32_t p_atomic_xchg_32(uint32_t *v, uint32_t i)
245 {
246 uint32_t actual = p_atomic_read(v);
247 uint32_t expected;
248 do {
249 expected = actual;
250 actual = p_atomic_cmpxchg(v, expected, i);
251 } while (expected != actual);
252 return actual;
253 }
254
p_atomic_xchg_64(uint64_t * v,uint64_t i)255 static inline uint64_t p_atomic_xchg_64(uint64_t *v, uint64_t i)
256 {
257 uint64_t actual = p_atomic_read(v);
258 uint64_t expected;
259 do {
260 expected = actual;
261 actual = p_atomic_cmpxchg(v, expected, i);
262 } while (expected != actual);
263 return actual;
264 }
265
266 #define p_atomic_xchg(v, i) (__typeof(*(v)))( \
267 sizeof(*(v)) == sizeof(uint32_t) ? p_atomic_xchg_32((uint32_t *)(v), (uint32_t)(i)) : \
268 sizeof(*(v)) == sizeof(uint64_t) ? p_atomic_xchg_64((uint64_t *)(v), (uint64_t)(i)) : \
269 (assert(!"should not get here"), 0))
270 #endif
271
272 #endif /* U_ATOMIC_H */
273