util: use GCC atomic intrinsics with explicit memory model
[mesa.git] / 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 #if defined(USE_GCC_ATOMIC_BUILTINS)
40
41 /* The builtins with explicit memory model are available since GCC 4.7. */
42 #define p_atomic_set(_v, _i) __atomic_store_n((_v), (_i), __ATOMIC_RELEASE)
43 #define p_atomic_read(_v) __atomic_load_n((_v), __ATOMIC_ACQUIRE)
44 #define p_atomic_dec_zero(v) (__atomic_sub_fetch((v), 1, __ATOMIC_ACQ_REL) == 0)
45 #define p_atomic_inc(v) (void) __atomic_add_fetch((v), 1, __ATOMIC_ACQ_REL)
46 #define p_atomic_dec(v) (void) __atomic_sub_fetch((v), 1, __ATOMIC_ACQ_REL)
47 #define p_atomic_add(v, i) (void) __atomic_add_fetch((v), (i), __ATOMIC_ACQ_REL)
48 #define p_atomic_inc_return(v) __atomic_add_fetch((v), 1, __ATOMIC_ACQ_REL)
49 #define p_atomic_dec_return(v) __atomic_sub_fetch((v), 1, __ATOMIC_ACQ_REL)
50
51 #else
52
53 #define p_atomic_set(_v, _i) (*(_v) = (_i))
54 #define p_atomic_read(_v) (*(_v))
55 #define p_atomic_dec_zero(v) (__sync_sub_and_fetch((v), 1) == 0)
56 #define p_atomic_inc(v) (void) __sync_add_and_fetch((v), 1)
57 #define p_atomic_dec(v) (void) __sync_sub_and_fetch((v), 1)
58 #define p_atomic_add(v, i) (void) __sync_add_and_fetch((v), (i))
59 #define p_atomic_inc_return(v) __sync_add_and_fetch((v), 1)
60 #define p_atomic_dec_return(v) __sync_sub_and_fetch((v), 1)
61
62 #endif
63
64 /* There is no __atomic_* compare and exchange that returns the current value.
65 * Also, GCC 5.4 seems unable to optimize a compound statement expression that
66 * uses an additional stack variable with __atomic_compare_exchange[_n].
67 */
68 #define p_atomic_cmpxchg(v, old, _new) \
69 __sync_val_compare_and_swap((v), (old), (_new))
70
71 #endif
72
73
74
75 /* Unlocked version for single threaded environments, such as some
76 * windows kernel modules.
77 */
78 #if defined(PIPE_ATOMIC_OS_UNLOCKED)
79
80 #define PIPE_ATOMIC "Unlocked"
81
82 #define p_atomic_set(_v, _i) (*(_v) = (_i))
83 #define p_atomic_read(_v) (*(_v))
84 #define p_atomic_dec_zero(_v) (p_atomic_dec_return(_v) == 0)
85 #define p_atomic_inc(_v) ((void) p_atomic_inc_return(_v))
86 #define p_atomic_dec(_v) ((void) p_atomic_dec_return(_v))
87 #define p_atomic_add(_v, _i) (*(_v) = *(_v) + (_i))
88 #define p_atomic_inc_return(_v) (++(*(_v)))
89 #define p_atomic_dec_return(_v) (--(*(_v)))
90 #define p_atomic_cmpxchg(_v, _old, _new) (*(_v) == (_old) ? (*(_v) = (_new), (_old)) : *(_v))
91
92 #endif
93
94
95 #if defined(PIPE_ATOMIC_MSVC_INTRINSIC)
96
97 #define PIPE_ATOMIC "MSVC Intrinsics"
98
99 /* We use the Windows header's Interlocked*64 functions instead of the
100 * _Interlocked*64 intrinsics wherever we can, as support for the latter varies
101 * with target CPU, whereas Windows headers take care of all portability
102 * issues: using intrinsics where available, falling back to library
103 * implementations where not.
104 */
105 #ifndef WIN32_LEAN_AND_MEAN
106 #define WIN32_LEAN_AND_MEAN 1
107 #endif
108 #include <windows.h>
109 #include <intrin.h>
110 #include <assert.h>
111
112 /* MSVC supports decltype keyword, but it's only supported on C++ and doesn't
113 * quite work here; and if a C++-only solution is worthwhile, then it would be
114 * better to use templates / function overloading, instead of decltype magic.
115 * Therefore, we rely on implicit casting to LONGLONG for the functions that return
116 */
117
118 #define p_atomic_set(_v, _i) (*(_v) = (_i))
119 #define p_atomic_read(_v) (*(_v))
120
121 #define p_atomic_dec_zero(_v) \
122 (p_atomic_dec_return(_v) == 0)
123
124 #define p_atomic_inc(_v) \
125 ((void) p_atomic_inc_return(_v))
126
127 #define p_atomic_inc_return(_v) (\
128 sizeof *(_v) == sizeof(short) ? _InterlockedIncrement16((short *) (_v)) : \
129 sizeof *(_v) == sizeof(long) ? _InterlockedIncrement ((long *) (_v)) : \
130 sizeof *(_v) == sizeof(__int64) ? InterlockedIncrement64 ((__int64 *)(_v)) : \
131 (assert(!"should not get here"), 0))
132
133 #define p_atomic_dec(_v) \
134 ((void) p_atomic_dec_return(_v))
135
136 #define p_atomic_dec_return(_v) (\
137 sizeof *(_v) == sizeof(short) ? _InterlockedDecrement16((short *) (_v)) : \
138 sizeof *(_v) == sizeof(long) ? _InterlockedDecrement ((long *) (_v)) : \
139 sizeof *(_v) == sizeof(__int64) ? InterlockedDecrement64 ((__int64 *)(_v)) : \
140 (assert(!"should not get here"), 0))
141
142 #define p_atomic_add(_v, _i) (\
143 sizeof *(_v) == sizeof(char) ? _InterlockedExchangeAdd8 ((char *) (_v), (_i)) : \
144 sizeof *(_v) == sizeof(short) ? _InterlockedExchangeAdd16((short *) (_v), (_i)) : \
145 sizeof *(_v) == sizeof(long) ? _InterlockedExchangeAdd ((long *) (_v), (_i)) : \
146 sizeof *(_v) == sizeof(__int64) ? InterlockedExchangeAdd64((__int64 *)(_v), (_i)) : \
147 (assert(!"should not get here"), 0))
148
149 #define p_atomic_cmpxchg(_v, _old, _new) (\
150 sizeof *(_v) == sizeof(char) ? _InterlockedCompareExchange8 ((char *) (_v), (char) (_new), (char) (_old)) : \
151 sizeof *(_v) == sizeof(short) ? _InterlockedCompareExchange16((short *) (_v), (short) (_new), (short) (_old)) : \
152 sizeof *(_v) == sizeof(long) ? _InterlockedCompareExchange ((long *) (_v), (long) (_new), (long) (_old)) : \
153 sizeof *(_v) == sizeof(__int64) ? InterlockedCompareExchange64 ((__int64 *)(_v), (__int64)(_new), (__int64)(_old)) : \
154 (assert(!"should not get here"), 0))
155
156 #endif
157
158 #if defined(PIPE_ATOMIC_OS_SOLARIS)
159
160 #define PIPE_ATOMIC "Solaris OS atomic functions"
161
162 #include <atomic.h>
163 #include <assert.h>
164
165 #define p_atomic_set(_v, _i) (*(_v) = (_i))
166 #define p_atomic_read(_v) (*(_v))
167
168 #define p_atomic_dec_zero(v) (\
169 sizeof(*v) == sizeof(uint8_t) ? atomic_dec_8_nv ((uint8_t *)(v)) == 0 : \
170 sizeof(*v) == sizeof(uint16_t) ? atomic_dec_16_nv((uint16_t *)(v)) == 0 : \
171 sizeof(*v) == sizeof(uint32_t) ? atomic_dec_32_nv((uint32_t *)(v)) == 0 : \
172 sizeof(*v) == sizeof(uint64_t) ? atomic_dec_64_nv((uint64_t *)(v)) == 0 : \
173 (assert(!"should not get here"), 0))
174
175 #define p_atomic_inc(v) (void) (\
176 sizeof(*v) == sizeof(uint8_t) ? atomic_inc_8 ((uint8_t *)(v)) : \
177 sizeof(*v) == sizeof(uint16_t) ? atomic_inc_16((uint16_t *)(v)) : \
178 sizeof(*v) == sizeof(uint32_t) ? atomic_inc_32((uint32_t *)(v)) : \
179 sizeof(*v) == sizeof(uint64_t) ? atomic_inc_64((uint64_t *)(v)) : \
180 (assert(!"should not get here"), 0))
181
182 #define p_atomic_inc_return(v) ((__typeof(*v)) \
183 sizeof(*v) == sizeof(uint8_t) ? atomic_inc_8_nv ((uint8_t *)(v)) : \
184 sizeof(*v) == sizeof(uint16_t) ? atomic_inc_16_nv((uint16_t *)(v)) : \
185 sizeof(*v) == sizeof(uint32_t) ? atomic_inc_32_nv((uint32_t *)(v)) : \
186 sizeof(*v) == sizeof(uint64_t) ? atomic_inc_64_nv((uint64_t *)(v)) : \
187 (assert(!"should not get here"), 0))
188
189 #define p_atomic_dec(v) ((void) \
190 sizeof(*v) == sizeof(uint8_t) ? atomic_dec_8 ((uint8_t *)(v)) : \
191 sizeof(*v) == sizeof(uint16_t) ? atomic_dec_16((uint16_t *)(v)) : \
192 sizeof(*v) == sizeof(uint32_t) ? atomic_dec_32((uint32_t *)(v)) : \
193 sizeof(*v) == sizeof(uint64_t) ? atomic_dec_64((uint64_t *)(v)) : \
194 (assert(!"should not get here"), 0))
195
196 #define p_atomic_dec_return(v) ((__typeof(*v)) \
197 sizeof(*v) == sizeof(uint8_t) ? atomic_dec_8_nv ((uint8_t *)(v)) : \
198 sizeof(*v) == sizeof(uint16_t) ? atomic_dec_16_nv((uint16_t *)(v)) : \
199 sizeof(*v) == sizeof(uint32_t) ? atomic_dec_32_nv((uint32_t *)(v)) : \
200 sizeof(*v) == sizeof(uint64_t) ? atomic_dec_64_nv((uint64_t *)(v)) : \
201 (assert(!"should not get here"), 0))
202
203 #define p_atomic_add(v, i) ((void) \
204 sizeof(*v) == sizeof(uint8_t) ? atomic_add_8 ((uint8_t *)(v), (i)) : \
205 sizeof(*v) == sizeof(uint16_t) ? atomic_add_16((uint16_t *)(v), (i)) : \
206 sizeof(*v) == sizeof(uint32_t) ? atomic_add_32((uint32_t *)(v), (i)) : \
207 sizeof(*v) == sizeof(uint64_t) ? atomic_add_64((uint64_t *)(v), (i)) : \
208 (assert(!"should not get here"), 0))
209
210 #define p_atomic_cmpxchg(v, old, _new) ((__typeof(*v)) \
211 sizeof(*v) == sizeof(uint8_t) ? atomic_cas_8 ((uint8_t *)(v), (uint8_t )(old), (uint8_t )(_new)) : \
212 sizeof(*v) == sizeof(uint16_t) ? atomic_cas_16((uint16_t *)(v), (uint16_t)(old), (uint16_t)(_new)) : \
213 sizeof(*v) == sizeof(uint32_t) ? atomic_cas_32((uint32_t *)(v), (uint32_t)(old), (uint32_t)(_new)) : \
214 sizeof(*v) == sizeof(uint64_t) ? atomic_cas_64((uint64_t *)(v), (uint64_t)(old), (uint64_t)(_new)) : \
215 (assert(!"should not get here"), 0))
216
217 #endif
218
219 #ifndef PIPE_ATOMIC
220 #error "No pipe_atomic implementation selected"
221 #endif
222
223
224
225 #endif /* U_ATOMIC_H */