… | |
… | |
28 | */ |
28 | */ |
29 | |
29 | |
30 | #ifndef ECB_H |
30 | #ifndef ECB_H |
31 | #define ECB_H |
31 | #define ECB_H |
32 | |
32 | |
|
|
33 | /* 16 bits major, 16 bits minor */ |
|
|
34 | #define ECB_VERSION 0x00010001 |
|
|
35 | |
33 | #ifdef _WIN32 |
36 | #ifdef _WIN32 |
34 | typedef signed char int8_t; |
37 | typedef signed char int8_t; |
35 | typedef unsigned char uint8_t; |
38 | typedef unsigned char uint8_t; |
36 | typedef signed short int16_t; |
39 | typedef signed short int16_t; |
37 | typedef unsigned short uint16_t; |
40 | typedef unsigned short uint16_t; |
… | |
… | |
42 | typedef unsigned long long uint64_t; |
45 | typedef unsigned long long uint64_t; |
43 | #else /* _MSC_VER || __BORLANDC__ */ |
46 | #else /* _MSC_VER || __BORLANDC__ */ |
44 | typedef signed __int64 int64_t; |
47 | typedef signed __int64 int64_t; |
45 | typedef unsigned __int64 uint64_t; |
48 | typedef unsigned __int64 uint64_t; |
46 | #endif |
49 | #endif |
|
|
50 | #ifdef _WIN64 |
|
|
51 | #define ECB_PTRSIZE 8 |
|
|
52 | typedef uint64_t uintptr_t; |
|
|
53 | typedef int64_t intptr_t; |
|
|
54 | #else |
|
|
55 | #define ECB_PTRSIZE 4 |
|
|
56 | typedef uint32_t uintptr_t; |
|
|
57 | typedef int32_t intptr_t; |
|
|
58 | #endif |
|
|
59 | typedef intptr_t ptrdiff_t; |
47 | #else |
60 | #else |
48 | #include <inttypes.h> |
61 | #include <inttypes.h> |
|
|
62 | #if UINTMAX_MAX > 0xffffffffU |
|
|
63 | #define ECB_PTRSIZE 8 |
|
|
64 | #else |
|
|
65 | #define ECB_PTRSIZE 4 |
|
|
66 | #endif |
49 | #endif |
67 | #endif |
50 | |
68 | |
51 | /* many compilers define _GNUC_ to some versions but then only implement |
69 | /* many compilers define _GNUC_ to some versions but then only implement |
52 | * what their idiot authors think are the "more important" extensions, |
70 | * what their idiot authors think are the "more important" extensions, |
53 | * causing enormous grief in return for some better fake benchmark numbers. |
71 | * causing enormous grief in return for some better fake benchmark numbers. |
54 | * or so. |
72 | * or so. |
55 | * we try to detect these and simply assume they are not gcc - if they have |
73 | * we try to detect these and simply assume they are not gcc - if they have |
56 | * an issue with that they should have done it right in the first place. |
74 | * an issue with that they should have done it right in the first place. |
57 | */ |
75 | */ |
58 | #ifndef ECB_GCC_VERSION |
76 | #ifndef ECB_GCC_VERSION |
59 | #if !defined(__GNUC_MINOR__) || defined(__INTEL_COMPILER) || defined(__SUNPRO_C) || defined(__SUNPRO_CC) || defined(__llvm__) || defined(__clang__) |
77 | #if !defined __GNUC_MINOR__ || defined __INTEL_COMPILER || defined __SUNPRO_C || defined __SUNPRO_CC || defined __llvm__ || defined __clang__ |
60 | #define ECB_GCC_VERSION(major,minor) 0 |
78 | #define ECB_GCC_VERSION(major,minor) 0 |
61 | #else |
79 | #else |
62 | #define ECB_GCC_VERSION(major,minor) (__GNUC__ > (major) || (__GNUC__ == (major) && __GNUC_MINOR__ >= (minor))) |
80 | #define ECB_GCC_VERSION(major,minor) (__GNUC__ > (major) || (__GNUC__ == (major) && __GNUC_MINOR__ >= (minor))) |
63 | #endif |
81 | #endif |
64 | #endif |
82 | #endif |
65 | |
83 | |
|
|
84 | #define ECB_C (__STDC__+0) /* this assumes that __STDC__ is either empty or a number */ |
|
|
85 | #define ECB_C99 (__STDC_VERSION__ >= 199901L) |
|
|
86 | #define ECB_C11 (__STDC_VERSION__ >= 201112L) |
|
|
87 | #define ECB_CPP (__cplusplus+0) |
|
|
88 | #define ECB_CPP98 (__cplusplus >= 199711L) |
|
|
89 | #define ECB_CPP11 (__cplusplus >= 201103L) |
|
|
90 | |
66 | /*****************************************************************************/ |
91 | /*****************************************************************************/ |
67 | |
92 | |
68 | /* ECB_NO_THREADS - ecb is not used by multiple threads, ever */ |
93 | /* ECB_NO_THREADS - ecb is not used by multiple threads, ever */ |
69 | /* ECB_NO_SMP - ecb might be used in multiple threads, but only on a single cpu */ |
94 | /* ECB_NO_SMP - ecb might be used in multiple threads, but only on a single cpu */ |
70 | |
95 | |
|
|
96 | #if ECB_NO_THREADS |
|
|
97 | # define ECB_NO_SMP 1 |
|
|
98 | #endif |
|
|
99 | |
71 | #if ECB_NO_THREADS || ECB_NO_SMP |
100 | #if ECB_NO_THREADS || ECB_NO_SMP |
72 | #define ECB_MEMORY_FENCE do { } while (0) |
101 | #define ECB_MEMORY_FENCE do { } while (0) |
73 | #endif |
102 | #endif |
74 | |
103 | |
75 | #ifndef ECB_MEMORY_FENCE |
104 | #ifndef ECB_MEMORY_FENCE |
|
|
105 | #if ECB_C11 && !defined __STDC_NO_ATOMICS__ |
|
|
106 | /* we assume that these memory fences work on all variables/all memory accesses, */ |
|
|
107 | /* not just C11 atomics and atomic accesses */ |
|
|
108 | #include <stdatomic.h> |
|
|
109 | #define ECB_MEMORY_FENCE atomic_thread_fence (memory_order_acq_rel) |
|
|
110 | #define ECB_MEMORY_FENCE_ACQUIRE atomic_thread_fence (memory_order_acquire) |
|
|
111 | #define ECB_MEMORY_FENCE_RELEASE atomic_thread_fence (memory_order_release) |
|
|
112 | #endif |
|
|
113 | #endif |
|
|
114 | |
|
|
115 | #ifndef ECB_MEMORY_FENCE |
76 | #if ECB_GCC_VERSION(2,5) || defined(__INTEL_COMPILER) || defined(__clang__) || __SUNPRO_C >= 0x5110 || __SUNPRO_CC >= 0x5110 |
116 | #if ECB_GCC_VERSION(2,5) || defined __INTEL_COMPILER || (__llvm__ && __GNUC__) || __SUNPRO_C >= 0x5110 || __SUNPRO_CC >= 0x5110 |
77 | #if __i386__ |
117 | #if __i386 || __i386__ |
78 | #define ECB_MEMORY_FENCE __asm__ __volatile__ ("lock; orb $0, -1(%%esp)" : : : "memory") |
118 | #define ECB_MEMORY_FENCE __asm__ __volatile__ ("lock; orb $0, -1(%%esp)" : : : "memory") |
79 | #define ECB_MEMORY_FENCE_ACQUIRE ECB_MEMORY_FENCE /* non-lock xchg might be enough */ |
119 | #define ECB_MEMORY_FENCE_ACQUIRE ECB_MEMORY_FENCE /* non-lock xchg might be enough */ |
80 | #define ECB_MEMORY_FENCE_RELEASE do { } while (0) /* unlikely to change in future cpus */ |
120 | #define ECB_MEMORY_FENCE_RELEASE do { } while (0) /* unlikely to change in future cpus */ |
81 | #elif __amd64 |
121 | #elif __amd64 || __amd64__ || __x86_64 || __x86_64__ |
82 | #define ECB_MEMORY_FENCE __asm__ __volatile__ ("mfence" : : : "memory") |
122 | #define ECB_MEMORY_FENCE __asm__ __volatile__ ("mfence" : : : "memory") |
83 | #define ECB_MEMORY_FENCE_ACQUIRE __asm__ __volatile__ ("lfence" : : : "memory") |
123 | #define ECB_MEMORY_FENCE_ACQUIRE __asm__ __volatile__ ("lfence" : : : "memory") |
84 | #define ECB_MEMORY_FENCE_RELEASE __asm__ __volatile__ ("sfence") /* play safe - not needed in any current cpu */ |
124 | #define ECB_MEMORY_FENCE_RELEASE __asm__ __volatile__ ("sfence") /* play safe - not needed in any current cpu */ |
85 | #elif __powerpc__ || __ppc__ || __powerpc64__ || __ppc64__ |
125 | #elif __powerpc__ || __ppc__ || __powerpc64__ || __ppc64__ |
86 | #define ECB_MEMORY_FENCE __asm__ __volatile__ ("sync" : : : "memory") |
126 | #define ECB_MEMORY_FENCE __asm__ __volatile__ ("sync" : : : "memory") |
87 | #elif defined(__ARM_ARCH_6__ ) || defined(__ARM_ARCH_6J__ ) \ |
127 | #elif defined __ARM_ARCH_6__ || defined __ARM_ARCH_6J__ \ |
88 | || defined(__ARM_ARCH_6K__) || defined(__ARM_ARCH_6ZK__) |
128 | || defined __ARM_ARCH_6K__ || defined __ARM_ARCH_6ZK__ |
89 | #define ECB_MEMORY_FENCE __asm__ __volatile__ ("mcr p15,0,%0,c7,c10,5" : : "r" (0) : "memory") |
129 | #define ECB_MEMORY_FENCE __asm__ __volatile__ ("mcr p15,0,%0,c7,c10,5" : : "r" (0) : "memory") |
90 | #elif defined(__ARM_ARCH_7__ ) || defined(__ARM_ARCH_7A__ ) \ |
130 | #elif defined __ARM_ARCH_7__ || defined __ARM_ARCH_7A__ \ |
91 | || defined(__ARM_ARCH_7M__) || defined(__ARM_ARCH_7R__ ) |
131 | || defined __ARM_ARCH_7M__ || defined __ARM_ARCH_7R__ |
92 | #define ECB_MEMORY_FENCE __asm__ __volatile__ ("dmb" : : : "memory") |
132 | #define ECB_MEMORY_FENCE __asm__ __volatile__ ("dmb" : : : "memory") |
93 | #elif defined(__sparc) |
133 | #elif __sparc || __sparc__ |
94 | #define ECB_MEMORY_FENCE __asm__ __volatile__ ("membar #LoadStore | #StoreLoad | #LoadLoad | #StoreStore" : : : "memory") |
134 | #define ECB_MEMORY_FENCE __asm__ __volatile__ ("membar #LoadStore | #LoadLoad | #StoreStore | #StoreLoad | " : : : "memory") |
95 | #define ECB_MEMORY_FENCE_ACQUIRE __asm__ __volatile__ ("membar #LoadLoad" : : : "memory") |
135 | #define ECB_MEMORY_FENCE_ACQUIRE __asm__ __volatile__ ("membar #LoadStore | #LoadLoad" : : : "memory") |
96 | #define ECB_MEMORY_FENCE_RELEASE __asm__ __volatile__ ("membar #StoreStore") |
136 | #define ECB_MEMORY_FENCE_RELEASE __asm__ __volatile__ ("membar #LoadStore | #StoreStore") |
|
|
137 | #elif defined __s390__ || defined __s390x__ |
|
|
138 | #define ECB_MEMORY_FENCE __asm__ __volatile__ ("bcr 15,0" : : : "memory") |
|
|
139 | #elif defined __mips__ |
|
|
140 | #define ECB_MEMORY_FENCE __asm__ __volatile__ ("sync" : : : "memory") |
|
|
141 | #elif defined __alpha__ |
|
|
142 | #define ECB_MEMORY_FENCE __asm__ __volatile__ ("mb" : : : "memory") |
97 | #endif |
143 | #endif |
98 | #endif |
144 | #endif |
99 | #endif |
145 | #endif |
100 | |
146 | |
101 | #ifndef ECB_MEMORY_FENCE |
147 | #ifndef ECB_MEMORY_FENCE |
102 | #if ECB_GCC_VERSION(4,4) || defined(__INTEL_COMPILER) || defined(__clang__) |
148 | #if ECB_GCC_VERSION(4,4) || defined __INTEL_COMPILER || defined __clang__ |
103 | #define ECB_MEMORY_FENCE __sync_synchronize () |
149 | #define ECB_MEMORY_FENCE __sync_synchronize () |
104 | /*#define ECB_MEMORY_FENCE_ACQUIRE ({ char dummy = 0; __sync_lock_test_and_set (&dummy, 1); }) */ |
150 | /*#define ECB_MEMORY_FENCE_ACQUIRE ({ char dummy = 0; __sync_lock_test_and_set (&dummy, 1); }) */ |
105 | /*#define ECB_MEMORY_FENCE_RELEASE ({ char dummy = 1; __sync_lock_release (&dummy ); }) */ |
151 | /*#define ECB_MEMORY_FENCE_RELEASE ({ char dummy = 1; __sync_lock_release (&dummy ); }) */ |
106 | #elif _MSC_VER >= 1400 /* VC++ 2005 */ |
152 | #elif _MSC_VER >= 1400 /* VC++ 2005 */ |
107 | #pragma intrinsic(_ReadBarrier,_WriteBarrier,_ReadWriteBarrier) |
153 | #pragma intrinsic(_ReadBarrier,_WriteBarrier,_ReadWriteBarrier) |
108 | #define ECB_MEMORY_FENCE _ReadWriteBarrier () |
154 | #define ECB_MEMORY_FENCE _ReadWriteBarrier () |
109 | #define ECB_MEMORY_FENCE_ACQUIRE _ReadWriteBarrier () /* according to msdn, _ReadBarrier is not a load fence */ |
155 | #define ECB_MEMORY_FENCE_ACQUIRE _ReadWriteBarrier () /* according to msdn, _ReadBarrier is not a load fence */ |
110 | #define ECB_MEMORY_FENCE_RELEASE _WriteBarrier () |
156 | #define ECB_MEMORY_FENCE_RELEASE _WriteBarrier () |
111 | #elif defined(_WIN32) |
157 | #elif defined _WIN32 |
112 | #include <WinNT.h> |
158 | #include <WinNT.h> |
113 | #define ECB_MEMORY_FENCE MemoryBarrier () /* actually just xchg on x86... scary */ |
159 | #define ECB_MEMORY_FENCE MemoryBarrier () /* actually just xchg on x86... scary */ |
114 | #elif __SUNPRO_C >= 0x5110 || __SUNPRO_CC >= 0x5110 |
160 | #elif __SUNPRO_C >= 0x5110 || __SUNPRO_CC >= 0x5110 |
115 | #include <mbarrier.h> |
161 | #include <mbarrier.h> |
116 | #define ECB_MEMORY_FENCE __machine_rw_barrier () |
162 | #define ECB_MEMORY_FENCE __machine_rw_barrier () |
117 | #define ECB_MEMORY_FENCE_ACQUIRE __machine_r_barrier () |
163 | #define ECB_MEMORY_FENCE_ACQUIRE __machine_r_barrier () |
118 | #define ECB_MEMORY_FENCE_RELEASE __machine_w_barrier () |
164 | #define ECB_MEMORY_FENCE_RELEASE __machine_w_barrier () |
|
|
165 | #elif __xlC__ |
|
|
166 | #define ECB_MEMORY_FENCE __sync () |
119 | #endif |
167 | #endif |
120 | #endif |
168 | #endif |
121 | |
169 | |
122 | #ifndef ECB_MEMORY_FENCE |
170 | #ifndef ECB_MEMORY_FENCE |
123 | #if !ECB_AVOID_PTHREADS |
171 | #if !ECB_AVOID_PTHREADS |
… | |
… | |
135 | static pthread_mutex_t ecb_mf_lock = PTHREAD_MUTEX_INITIALIZER; |
183 | static pthread_mutex_t ecb_mf_lock = PTHREAD_MUTEX_INITIALIZER; |
136 | #define ECB_MEMORY_FENCE do { pthread_mutex_lock (&ecb_mf_lock); pthread_mutex_unlock (&ecb_mf_lock); } while (0) |
184 | #define ECB_MEMORY_FENCE do { pthread_mutex_lock (&ecb_mf_lock); pthread_mutex_unlock (&ecb_mf_lock); } while (0) |
137 | #endif |
185 | #endif |
138 | #endif |
186 | #endif |
139 | |
187 | |
140 | #if !defined(ECB_MEMORY_FENCE_ACQUIRE) && defined(ECB_MEMORY_FENCE) |
188 | #if !defined ECB_MEMORY_FENCE_ACQUIRE && defined ECB_MEMORY_FENCE |
141 | #define ECB_MEMORY_FENCE_ACQUIRE ECB_MEMORY_FENCE |
189 | #define ECB_MEMORY_FENCE_ACQUIRE ECB_MEMORY_FENCE |
142 | #endif |
190 | #endif |
143 | |
191 | |
144 | #if !defined(ECB_MEMORY_FENCE_RELEASE) && defined(ECB_MEMORY_FENCE) |
192 | #if !defined ECB_MEMORY_FENCE_RELEASE && defined ECB_MEMORY_FENCE |
145 | #define ECB_MEMORY_FENCE_RELEASE ECB_MEMORY_FENCE |
193 | #define ECB_MEMORY_FENCE_RELEASE ECB_MEMORY_FENCE |
146 | #endif |
194 | #endif |
147 | |
195 | |
148 | /*****************************************************************************/ |
196 | /*****************************************************************************/ |
149 | |
|
|
150 | #define ECB_C99 (__STDC_VERSION__ >= 199901L) |
|
|
151 | |
197 | |
152 | #if __cplusplus |
198 | #if __cplusplus |
153 | #define ecb_inline static inline |
199 | #define ecb_inline static inline |
154 | #elif ECB_GCC_VERSION(2,5) |
200 | #elif ECB_GCC_VERSION(2,5) |
155 | #define ecb_inline static __inline__ |
201 | #define ecb_inline static __inline__ |
… | |
… | |
194 | #elif ECB_GCC_VERSION(3,0) |
240 | #elif ECB_GCC_VERSION(3,0) |
195 | #define ecb_decltype(x) __typeof(x) |
241 | #define ecb_decltype(x) __typeof(x) |
196 | #endif |
242 | #endif |
197 | |
243 | |
198 | #define ecb_noinline ecb_attribute ((__noinline__)) |
244 | #define ecb_noinline ecb_attribute ((__noinline__)) |
199 | #define ecb_noreturn ecb_attribute ((__noreturn__)) |
|
|
200 | #define ecb_unused ecb_attribute ((__unused__)) |
245 | #define ecb_unused ecb_attribute ((__unused__)) |
201 | #define ecb_const ecb_attribute ((__const__)) |
246 | #define ecb_const ecb_attribute ((__const__)) |
202 | #define ecb_pure ecb_attribute ((__pure__)) |
247 | #define ecb_pure ecb_attribute ((__pure__)) |
|
|
248 | |
|
|
249 | #if ECB_C11 |
|
|
250 | #define ecb_noreturn _Noreturn |
|
|
251 | #else |
|
|
252 | #define ecb_noreturn ecb_attribute ((__noreturn__)) |
|
|
253 | #endif |
203 | |
254 | |
204 | #if ECB_GCC_VERSION(4,3) |
255 | #if ECB_GCC_VERSION(4,3) |
205 | #define ecb_artificial ecb_attribute ((__artificial__)) |
256 | #define ecb_artificial ecb_attribute ((__artificial__)) |
206 | #define ecb_hot ecb_attribute ((__hot__)) |
257 | #define ecb_hot ecb_attribute ((__hot__)) |
207 | #define ecb_cold ecb_attribute ((__cold__)) |
258 | #define ecb_cold ecb_attribute ((__cold__)) |
… | |
… | |
298 | |
349 | |
299 | return r + ecb_ld32 (x); |
350 | return r + ecb_ld32 (x); |
300 | } |
351 | } |
301 | #endif |
352 | #endif |
302 | |
353 | |
|
|
354 | ecb_function_ ecb_bool ecb_is_pot32 (uint32_t x) ecb_const; |
|
|
355 | ecb_function_ ecb_bool ecb_is_pot32 (uint32_t x) { return !(x & (x - 1)); } |
|
|
356 | ecb_function_ ecb_bool ecb_is_pot64 (uint64_t x) ecb_const; |
|
|
357 | ecb_function_ ecb_bool ecb_is_pot64 (uint64_t x) { return !(x & (x - 1)); } |
|
|
358 | |
303 | ecb_function_ uint8_t ecb_bitrev8 (uint8_t x) ecb_const; |
359 | ecb_function_ uint8_t ecb_bitrev8 (uint8_t x) ecb_const; |
304 | ecb_function_ uint8_t ecb_bitrev8 (uint8_t x) |
360 | ecb_function_ uint8_t ecb_bitrev8 (uint8_t x) |
305 | { |
361 | { |
306 | return ( (x * 0x0802U & 0x22110U) |
362 | return ( (x * 0x0802U & 0x22110U) |
307 | | (x * 0x8020U & 0x88440U)) * 0x10101U >> 16; |
363 | | (x * 0x8020U & 0x88440U)) * 0x10101U >> 16; |
… | |
… | |
386 | |
442 | |
387 | #if ECB_GCC_VERSION(4,5) |
443 | #if ECB_GCC_VERSION(4,5) |
388 | #define ecb_unreachable() __builtin_unreachable () |
444 | #define ecb_unreachable() __builtin_unreachable () |
389 | #else |
445 | #else |
390 | /* this seems to work fine, but gcc always emits a warning for it :/ */ |
446 | /* this seems to work fine, but gcc always emits a warning for it :/ */ |
391 | ecb_function_ void ecb_unreachable (void) ecb_noreturn; |
447 | ecb_inline void ecb_unreachable (void) ecb_noreturn; |
392 | ecb_function_ void ecb_unreachable (void) { } |
448 | ecb_inline void ecb_unreachable (void) { } |
393 | #endif |
449 | #endif |
394 | |
450 | |
395 | /* try to tell the compiler that some condition is definitely true */ |
451 | /* try to tell the compiler that some condition is definitely true */ |
396 | #define ecb_assume(cond) do { if (!(cond)) ecb_unreachable (); } while (0) |
452 | #define ecb_assume(cond) do { if (!(cond)) ecb_unreachable (); } while (0) |
397 | |
453 | |
398 | ecb_function_ unsigned char ecb_byteorder_helper (void) ecb_const; |
454 | ecb_inline unsigned char ecb_byteorder_helper (void) ecb_const; |
399 | ecb_function_ unsigned char |
455 | ecb_inline unsigned char |
400 | ecb_byteorder_helper (void) |
456 | ecb_byteorder_helper (void) |
401 | { |
457 | { |
402 | const uint32_t u = 0x11223344; |
458 | const uint32_t u = 0x11223344; |
403 | return *(unsigned char *)&u; |
459 | return *(unsigned char *)&u; |
404 | } |
460 | } |
405 | |
461 | |
406 | ecb_function_ ecb_bool ecb_big_endian (void) ecb_const; |
462 | ecb_inline ecb_bool ecb_big_endian (void) ecb_const; |
407 | ecb_function_ ecb_bool ecb_big_endian (void) { return ecb_byteorder_helper () == 0x11; } |
463 | ecb_inline ecb_bool ecb_big_endian (void) { return ecb_byteorder_helper () == 0x11; } |
408 | ecb_function_ ecb_bool ecb_little_endian (void) ecb_const; |
464 | ecb_inline ecb_bool ecb_little_endian (void) ecb_const; |
409 | ecb_function_ ecb_bool ecb_little_endian (void) { return ecb_byteorder_helper () == 0x44; } |
465 | ecb_inline ecb_bool ecb_little_endian (void) { return ecb_byteorder_helper () == 0x44; } |
410 | |
466 | |
411 | #if ECB_GCC_VERSION(3,0) || ECB_C99 |
467 | #if ECB_GCC_VERSION(3,0) || ECB_C99 |
412 | #define ecb_mod(m,n) ((m) % (n) + ((m) % (n) < 0 ? (n) : 0)) |
468 | #define ecb_mod(m,n) ((m) % (n) + ((m) % (n) < 0 ? (n) : 0)) |
413 | #else |
469 | #else |
414 | #define ecb_mod(m,n) ((m) < 0 ? ((n) - 1 - ((-1 - (m)) % (n))) : ((m) % (n))) |
470 | #define ecb_mod(m,n) ((m) < 0 ? ((n) - 1 - ((-1 - (m)) % (n))) : ((m) % (n))) |