xref: /third_party/mesa3d/src/util/u_atomic.h (revision bf215546)
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_fetch_add(v, i) __atomic_fetch_add((v), (i), __ATOMIC_ACQ_REL)
54#define p_atomic_xchg(v, i) __atomic_exchange_n((v), (i), __ATOMIC_ACQ_REL)
55#define PIPE_NATIVE_ATOMIC_XCHG
56
57#else
58
59#define p_atomic_set(_v, _i) (*(_v) = (_i))
60#define p_atomic_read(_v) (*(_v))
61#define p_atomic_read_relaxed(_v) (*(_v))
62#define p_atomic_dec_zero(v) (__sync_sub_and_fetch((v), 1) == 0)
63#define p_atomic_inc(v) (void) __sync_add_and_fetch((v), 1)
64#define p_atomic_dec(v) (void) __sync_sub_and_fetch((v), 1)
65#define p_atomic_add(v, i) (void) __sync_add_and_fetch((v), (i))
66#define p_atomic_inc_return(v) __sync_add_and_fetch((v), 1)
67#define p_atomic_dec_return(v) __sync_sub_and_fetch((v), 1)
68#define p_atomic_add_return(v, i) __sync_add_and_fetch((v), (i))
69#define p_atomic_fetch_add(v, i) __sync_fetch_and_add((v), (i))
70
71#endif
72
73/* There is no __atomic_* compare and exchange that returns the current value.
74 * Also, GCC 5.4 seems unable to optimize a compound statement expression that
75 * uses an additional stack variable with __atomic_compare_exchange[_n].
76 */
77#define p_atomic_cmpxchg(v, old, _new) \
78   __sync_val_compare_and_swap((v), (old), (_new))
79#define p_atomic_cmpxchg_ptr(v, old, _new) p_atomic_cmpxchg(v, old, _new)
80
81#endif
82
83
84
85/* Unlocked version for single threaded environments, such as some
86 * windows kernel modules.
87 */
88#if defined(PIPE_ATOMIC_OS_UNLOCKED)
89
90#define PIPE_ATOMIC "Unlocked"
91
92#define p_atomic_set(_v, _i) (*(_v) = (_i))
93#define p_atomic_read(_v) (*(_v))
94#define p_atomic_read_relaxed(_v) (*(_v))
95#define p_atomic_dec_zero(_v) (p_atomic_dec_return(_v) == 0)
96#define p_atomic_inc(_v) ((void) p_atomic_inc_return(_v))
97#define p_atomic_dec(_v) ((void) p_atomic_dec_return(_v))
98#define p_atomic_add(_v, _i) ((void) p_atomic_add_return((_v), (_i)))
99#define p_atomic_inc_return(_v) (++(*(_v)))
100#define p_atomic_dec_return(_v) (--(*(_v)))
101#define p_atomic_add_return(_v, _i) (*(_v) = *(_v) + (_i))
102#define p_atomic_fetch_add(_v, _i) (*(_v) = *(_v) + (_i), *(_v) - (_i))
103#define p_atomic_cmpxchg(_v, _old, _new) (*(_v) == (_old) ? (*(_v) = (_new), (_old)) : *(_v))
104#define p_atomic_cmpxchg_ptr(_v, _old, _new) p_atomic_cmpxchg(_v, _old, _new)
105
106#endif
107
108
109#if defined(PIPE_ATOMIC_MSVC_INTRINSIC)
110
111#define PIPE_ATOMIC "MSVC Intrinsics"
112
113/* We use the Windows header's Interlocked*64 functions instead of the
114 * _Interlocked*64 intrinsics wherever we can, as support for the latter varies
115 * with target CPU, whereas Windows headers take care of all portability
116 * issues: using intrinsics where available, falling back to library
117 * implementations where not.
118 */
119#ifndef WIN32_LEAN_AND_MEAN
120#define WIN32_LEAN_AND_MEAN 1
121#endif
122#include <windows.h>
123#include <intrin.h>
124#include <assert.h>
125
126/* MSVC supports decltype keyword, but it's only supported on C++ and doesn't
127 * quite work here; and if a C++-only solution is worthwhile, then it would be
128 * better to use templates / function overloading, instead of decltype magic.
129 * Therefore, we rely on implicit casting to LONGLONG for the functions that return
130 */
131
132#define p_atomic_set(_v, _i) (*(_v) = (_i))
133#define p_atomic_read(_v) (*(_v))
134#define p_atomic_read_relaxed(_v) (*(_v))
135
136#define p_atomic_dec_zero(_v) \
137   (p_atomic_dec_return(_v) == 0)
138
139#define p_atomic_inc(_v) \
140   ((void) p_atomic_inc_return(_v))
141
142#define p_atomic_inc_return(_v) (\
143   sizeof *(_v) == sizeof(short)   ? _InterlockedIncrement16((short *)  (_v)) : \
144   sizeof *(_v) == sizeof(long)    ? _InterlockedIncrement  ((long *)   (_v)) : \
145   sizeof *(_v) == sizeof(__int64) ? InterlockedIncrement64 ((__int64 *)(_v)) : \
146                                     (assert(!"should not get here"), 0))
147
148#define p_atomic_dec(_v) \
149   ((void) p_atomic_dec_return(_v))
150
151#define p_atomic_dec_return(_v) (\
152   sizeof *(_v) == sizeof(short)   ? _InterlockedDecrement16((short *)  (_v)) : \
153   sizeof *(_v) == sizeof(long)    ? _InterlockedDecrement  ((long *)   (_v)) : \
154   sizeof *(_v) == sizeof(__int64) ? InterlockedDecrement64 ((__int64 *)(_v)) : \
155                                     (assert(!"should not get here"), 0))
156
157#define p_atomic_add(_v, _i) \
158   ((void) p_atomic_fetch_add((_v), (_i)))
159
160#define p_atomic_add_return(_v, _i) (\
161   sizeof *(_v) == sizeof(long)    ? InterlockedAdd  ((long *)   (_v), (_i)) : \
162   sizeof *(_v) == sizeof(__int64) ? InterlockedAdd64((__int64 *)(_v), (_i)) : \
163                                     (assert(!"should not get here"), 0))
164
165#define p_atomic_fetch_add(_v, _i) (\
166   sizeof *(_v) == sizeof(char)    ? _InterlockedExchangeAdd8 ((char *)   (_v), (_i)) : \
167   sizeof *(_v) == sizeof(short)   ? _InterlockedExchangeAdd16((short *)  (_v), (_i)) : \
168   sizeof *(_v) == sizeof(long)    ? InterlockedExchangeAdd  ((long *)   (_v), (_i)) : \
169   sizeof *(_v) == sizeof(__int64) ? InterlockedExchangeAdd64((__int64 *)(_v), (_i)) : \
170                                     (assert(!"should not get here"), 0))
171
172#define p_atomic_cmpxchg(_v, _old, _new) (\
173   sizeof *(_v) == sizeof(char)    ? _InterlockedCompareExchange8 ((char *)   (_v), (char)   (_new), (char)   (_old)) : \
174   sizeof *(_v) == sizeof(short)   ? _InterlockedCompareExchange16((short *)  (_v), (short)  (_new), (short)  (_old)) : \
175   sizeof *(_v) == sizeof(long)    ? _InterlockedCompareExchange  ((long *)   (_v), (long)   (_new), (long)   (_old)) : \
176   sizeof *(_v) == sizeof(__int64) ? InterlockedCompareExchange64 ((__int64 *)(_v), (__int64)(_new), (__int64)(_old)) : \
177                                     (assert(!"should not get here"), 0))
178
179#if defined(_WIN64)
180#define p_atomic_cmpxchg_ptr(_v, _old, _new) (void *)InterlockedCompareExchange64((__int64 *)(_v), (__int64)(_new), (__int64)(_old))
181#else
182#define p_atomic_cmpxchg_ptr(_v, _old, _new) (void *)InterlockedCompareExchange((long *)(_v), (long)(_new), (long)(_old))
183#endif
184
185#define PIPE_NATIVE_ATOMIC_XCHG
186#define p_atomic_xchg(_v, _new) (\
187   sizeof *(_v) == sizeof(long)    ? InterlockedExchange  ((long *)   (_v), (long)   (_new)) : \
188   sizeof *(_v) == sizeof(__int64) ? InterlockedExchange64((__int64 *)(_v), (__int64)(_new)) : \
189                                     (assert(!"should not get here"), 0))
190
191#endif
192
193#if defined(PIPE_ATOMIC_OS_SOLARIS)
194
195#define PIPE_ATOMIC "Solaris OS atomic functions"
196
197#include <atomic.h>
198#include <assert.h>
199
200#define p_atomic_set(_v, _i) (*(_v) = (_i))
201#define p_atomic_read(_v) (*(_v))
202
203#define p_atomic_dec_zero(v) (\
204   sizeof(*v) == sizeof(uint8_t)  ? atomic_dec_8_nv ((uint8_t  *)(v)) == 0 : \
205   sizeof(*v) == sizeof(uint16_t) ? atomic_dec_16_nv((uint16_t *)(v)) == 0 : \
206   sizeof(*v) == sizeof(uint32_t) ? atomic_dec_32_nv((uint32_t *)(v)) == 0 : \
207   sizeof(*v) == sizeof(uint64_t) ? atomic_dec_64_nv((uint64_t *)(v)) == 0 : \
208                                    (assert(!"should not get here"), 0))
209
210#define p_atomic_inc(v) (void) (\
211   sizeof(*v) == sizeof(uint8_t)  ? atomic_inc_8 ((uint8_t  *)(v)) : \
212   sizeof(*v) == sizeof(uint16_t) ? atomic_inc_16((uint16_t *)(v)) : \
213   sizeof(*v) == sizeof(uint32_t) ? atomic_inc_32((uint32_t *)(v)) : \
214   sizeof(*v) == sizeof(uint64_t) ? atomic_inc_64((uint64_t *)(v)) : \
215                                    (assert(!"should not get here"), 0))
216
217#define p_atomic_inc_return(v) (__typeof(*v))( \
218   sizeof(*v) == sizeof(uint8_t)  ? atomic_inc_8_nv ((uint8_t  *)(v)) : \
219   sizeof(*v) == sizeof(uint16_t) ? atomic_inc_16_nv((uint16_t *)(v)) : \
220   sizeof(*v) == sizeof(uint32_t) ? atomic_inc_32_nv((uint32_t *)(v)) : \
221   sizeof(*v) == sizeof(uint64_t) ? atomic_inc_64_nv((uint64_t *)(v)) : \
222                                    (assert(!"should not get here"), 0))
223
224#define p_atomic_dec(v) (void) ( \
225   sizeof(*v) == sizeof(uint8_t)  ? atomic_dec_8 ((uint8_t  *)(v)) : \
226   sizeof(*v) == sizeof(uint16_t) ? atomic_dec_16((uint16_t *)(v)) : \
227   sizeof(*v) == sizeof(uint32_t) ? atomic_dec_32((uint32_t *)(v)) : \
228   sizeof(*v) == sizeof(uint64_t) ? atomic_dec_64((uint64_t *)(v)) : \
229                                    (assert(!"should not get here"), 0))
230
231#define p_atomic_dec_return(v) (__typeof(*v))( \
232   sizeof(*v) == sizeof(uint8_t)  ? atomic_dec_8_nv ((uint8_t  *)(v)) : \
233   sizeof(*v) == sizeof(uint16_t) ? atomic_dec_16_nv((uint16_t *)(v)) : \
234   sizeof(*v) == sizeof(uint32_t) ? atomic_dec_32_nv((uint32_t *)(v)) : \
235   sizeof(*v) == sizeof(uint64_t) ? atomic_dec_64_nv((uint64_t *)(v)) : \
236                                    (assert(!"should not get here"), 0))
237
238#define p_atomic_add(v, i) (void) ( \
239   sizeof(*v) == sizeof(uint8_t)  ? atomic_add_8 ((uint8_t  *)(v), (i)) : \
240   sizeof(*v) == sizeof(uint16_t) ? atomic_add_16((uint16_t *)(v), (i)) : \
241   sizeof(*v) == sizeof(uint32_t) ? atomic_add_32((uint32_t *)(v), (i)) : \
242   sizeof(*v) == sizeof(uint64_t) ? atomic_add_64((uint64_t *)(v), (i)) : \
243                                    (assert(!"should not get here"), 0))
244
245#define p_atomic_add_return(v, i) (__typeof(*v)) ( \
246   sizeof(*v) == sizeof(uint8_t)  ? atomic_add_8_nv ((uint8_t  *)(v), (i)) : \
247   sizeof(*v) == sizeof(uint16_t) ? atomic_add_16_nv((uint16_t *)(v), (i)) : \
248   sizeof(*v) == sizeof(uint32_t) ? atomic_add_32_nv((uint32_t *)(v), (i)) : \
249   sizeof(*v) == sizeof(uint64_t) ? atomic_add_64_nv((uint64_t *)(v), (i)) : \
250                                    (assert(!"should not get here"), 0))
251
252#define p_atomic_fetch_add(v, i) (__typeof(*v)) ( \
253   sizeof(*v) == sizeof(uint8_t)  ? atomic_add_8_nv ((uint8_t  *)(v), (i)) - (i) : \
254   sizeof(*v) == sizeof(uint16_t) ? atomic_add_16_nv((uint16_t *)(v), (i)) - (i) : \
255   sizeof(*v) == sizeof(uint32_t) ? atomic_add_32_nv((uint32_t *)(v), (i)) - (i) : \
256   sizeof(*v) == sizeof(uint64_t) ? atomic_add_64_nv((uint64_t *)(v), (i)) - (i) : \
257                                    (assert(!"should not get here"), 0))
258
259#define p_atomic_cmpxchg(v, old, _new) (__typeof(*v))( \
260   sizeof(*v) == sizeof(uint8_t)  ? atomic_cas_8 ((uint8_t  *)(v), (uint8_t )(old), (uint8_t )(_new)) : \
261   sizeof(*v) == sizeof(uint16_t) ? atomic_cas_16((uint16_t *)(v), (uint16_t)(old), (uint16_t)(_new)) : \
262   sizeof(*v) == sizeof(uint32_t) ? atomic_cas_32((uint32_t *)(v), (uint32_t)(old), (uint32_t)(_new)) : \
263   sizeof(*v) == sizeof(uint64_t) ? atomic_cas_64((uint64_t *)(v), (uint64_t)(old), (uint64_t)(_new)) : \
264                                    (assert(!"should not get here"), 0))
265
266#if INTPTR_MAX == INT32_MAX
267#define p_atomic_cmpxchg_ptr(v, old, _new) (__typeof(*v))(atomic_cas_32((uint32_t *)(v), (uint32_t)(old), (uint32_t)(_new)))
268#else
269#define p_atomic_cmpxchg_ptr(v, old, _new) (__typeof(*v))(atomic_cas_64((uint64_t *)(v), (uint64_t)(old), (uint64_t)(_new)))
270#endif
271
272#endif
273
274#ifndef PIPE_ATOMIC
275#error "No pipe_atomic implementation selected"
276#endif
277
278#ifndef PIPE_NATIVE_ATOMIC_XCHG
279static inline uint32_t p_atomic_xchg_32(uint32_t *v, uint32_t i)
280{
281   uint32_t actual = p_atomic_read(v);
282   uint32_t expected;
283   do {
284      expected = actual;
285      actual = p_atomic_cmpxchg(v, expected, i);
286   } while (expected != actual);
287   return actual;
288}
289
290static inline uint64_t p_atomic_xchg_64(uint64_t *v, uint64_t i)
291{
292   uint64_t actual = p_atomic_read(v);
293   uint64_t expected;
294   do {
295      expected = actual;
296      actual = p_atomic_cmpxchg(v, expected, i);
297   } while (expected != actual);
298   return actual;
299}
300
301#define p_atomic_xchg(v, i) (__typeof(*(v)))( \
302   sizeof(*(v)) == sizeof(uint32_t) ? p_atomic_xchg_32((uint32_t *)(v), (uint32_t)(i)) : \
303   sizeof(*(v)) == sizeof(uint64_t) ? p_atomic_xchg_64((uint64_t *)(v), (uint64_t)(i)) : \
304                                      (assert(!"should not get here"), 0))
305#endif
306
307#endif /* U_ATOMIC_H */
308