… | |
… | |
71 | #if ECB_NO_THREADS || ECB_NO_SMP |
71 | #if ECB_NO_THREADS || ECB_NO_SMP |
72 | #define ECB_MEMORY_FENCE do { } while (0) |
72 | #define ECB_MEMORY_FENCE do { } while (0) |
73 | #endif |
73 | #endif |
74 | |
74 | |
75 | #ifndef ECB_MEMORY_FENCE |
75 | #ifndef ECB_MEMORY_FENCE |
76 | #if ECB_GCC_VERSION(2,5) |
76 | #if ECB_GCC_VERSION(2,5) || defined(__INTEL_COMPILER) || defined(__clang__) |
77 | #if __x86 |
77 | #if __i386__ |
78 | #define ECB_MEMORY_FENCE __asm__ __volatile__ ("lock; orb $0, -1(%%esp)" : : : "memory") |
78 | #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 */ |
79 | #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 */ |
80 | #define ECB_MEMORY_FENCE_RELEASE do { } while (0) /* unlikely to change in future cpus */ |
81 | #elif __amd64 |
81 | #elif __amd64 |
82 | #define ECB_MEMORY_FENCE __asm__ __volatile__ ("mfence" : : : "memory") |
82 | #define ECB_MEMORY_FENCE __asm__ __volatile__ ("mfence" : : : "memory") |
… | |
… | |
93 | #endif |
93 | #endif |
94 | #endif |
94 | #endif |
95 | #endif |
95 | #endif |
96 | |
96 | |
97 | #ifndef ECB_MEMORY_FENCE |
97 | #ifndef ECB_MEMORY_FENCE |
98 | #if ECB_GCC_VERSION(4,4) || defined(__INTEL_COMPILER) |
98 | #if ECB_GCC_VERSION(4,4) || defined(__INTEL_COMPILER) || defined(__clang__) |
99 | #define ECB_MEMORY_FENCE __sync_synchronize () |
99 | #define ECB_MEMORY_FENCE __sync_synchronize () |
100 | /*#define ECB_MEMORY_FENCE_ACQUIRE ({ char dummy = 0; __sync_lock_test_and_set (&dummy, 1); }) */ |
100 | /*#define ECB_MEMORY_FENCE_ACQUIRE ({ char dummy = 0; __sync_lock_test_and_set (&dummy, 1); }) */ |
101 | /*#define ECB_MEMORY_FENCE_RELEASE ({ char dummy = 1; __sync_lock_release (&dummy ); }) */ |
101 | /*#define ECB_MEMORY_FENCE_RELEASE ({ char dummy = 1; __sync_lock_release (&dummy ); }) */ |
102 | #elif _MSC_VER >= 1400 /* VC++ 2005 */ |
102 | #elif _MSC_VER >= 1400 /* VC++ 2005 */ |
103 | #pragma intrinsic(_ReadBarrier,_WriteBarrier,_ReadWriteBarrier) |
103 | #pragma intrinsic(_ReadBarrier,_WriteBarrier,_ReadWriteBarrier) |
… | |
… | |
289 | |
289 | |
290 | return r + ecb_ld32 (x); |
290 | return r + ecb_ld32 (x); |
291 | } |
291 | } |
292 | #endif |
292 | #endif |
293 | |
293 | |
|
|
294 | ecb_function_ uint8_t ecb_bitrev8 (uint8_t x) ecb_const; |
|
|
295 | ecb_function_ uint8_t ecb_bitrev8 (uint8_t x) |
|
|
296 | { |
|
|
297 | return ( (x * 0x0802U & 0x22110U) |
|
|
298 | | (x * 0x8020U & 0x88440U)) * 0x10101U >> 16; |
|
|
299 | } |
|
|
300 | |
|
|
301 | ecb_function_ uint16_t ecb_bitrev16 (uint16_t x) ecb_const; |
|
|
302 | ecb_function_ uint16_t ecb_bitrev16 (uint16_t x) |
|
|
303 | { |
|
|
304 | x = ((x >> 1) & 0x5555) | ((x & 0x5555) << 1); |
|
|
305 | x = ((x >> 2) & 0x3333) | ((x & 0x3333) << 2); |
|
|
306 | x = ((x >> 4) & 0x0f0f) | ((x & 0x0f0f) << 4); |
|
|
307 | x = ( x >> 8 ) | ( x << 8); |
|
|
308 | |
|
|
309 | return x; |
|
|
310 | } |
|
|
311 | |
|
|
312 | ecb_function_ uint32_t ecb_bitrev32 (uint32_t x) ecb_const; |
|
|
313 | ecb_function_ uint32_t ecb_bitrev32 (uint32_t x) |
|
|
314 | { |
|
|
315 | x = ((x >> 1) & 0x55555555) | ((x & 0x55555555) << 1); |
|
|
316 | x = ((x >> 2) & 0x33333333) | ((x & 0x33333333) << 2); |
|
|
317 | x = ((x >> 4) & 0x0f0f0f0f) | ((x & 0x0f0f0f0f) << 4); |
|
|
318 | x = ((x >> 8) & 0x00ff00ff) | ((x & 0x00ff00ff) << 8); |
|
|
319 | x = ( x >> 16 ) | ( x << 16); |
|
|
320 | |
|
|
321 | return x; |
|
|
322 | } |
|
|
323 | |
294 | /* popcount64 is only available on 64 bit cpus as gcc builtin */ |
324 | /* popcount64 is only available on 64 bit cpus as gcc builtin */ |
295 | /* so for this version we are lazy */ |
325 | /* so for this version we are lazy */ |
296 | ecb_function_ int ecb_popcount64 (uint64_t x) ecb_const; |
326 | ecb_function_ int ecb_popcount64 (uint64_t x) ecb_const; |
297 | ecb_function_ int |
327 | ecb_function_ int |
298 | ecb_popcount64 (uint64_t x) |
328 | ecb_popcount64 (uint64_t x) |
… | |
… | |
373 | #define ecb_mod(m,n) ((m) % (n) + ((m) % (n) < 0 ? (n) : 0)) |
403 | #define ecb_mod(m,n) ((m) % (n) + ((m) % (n) < 0 ? (n) : 0)) |
374 | #else |
404 | #else |
375 | #define ecb_mod(m,n) ((m) < 0 ? ((n) - 1 - ((-1 - (m)) % (n))) : ((m) % (n))) |
405 | #define ecb_mod(m,n) ((m) < 0 ? ((n) - 1 - ((-1 - (m)) % (n))) : ((m) % (n))) |
376 | #endif |
406 | #endif |
377 | |
407 | |
|
|
408 | #if __cplusplus |
|
|
409 | template<typename T> |
|
|
410 | static inline T ecb_div_rd (T val, T div) |
|
|
411 | { |
|
|
412 | return val < 0 ? - ((-val + div - 1) / div) : (val ) / div; |
|
|
413 | } |
|
|
414 | template<typename T> |
|
|
415 | static inline T ecb_div_ru (T val, T div) |
|
|
416 | { |
|
|
417 | return val < 0 ? - ((-val ) / div) : (val + div - 1) / div; |
|
|
418 | } |
|
|
419 | #else |
|
|
420 | #define ecb_div_rd(val,div) ((val) < 0 ? - ((-(val) + (div) - 1) / (div)) : ((val) ) / (div)) |
|
|
421 | #define ecb_div_ru(val,div) ((val) < 0 ? - ((-(val) ) / (div)) : ((val) + (div) - 1) / (div)) |
|
|
422 | #endif |
|
|
423 | |
378 | #if ecb_cplusplus_does_not_suck |
424 | #if ecb_cplusplus_does_not_suck |
379 | /* does not work for local types (http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2008/n2657.htm) */ |
425 | /* does not work for local types (http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2008/n2657.htm) */ |
380 | template<typename T, int N> |
426 | template<typename T, int N> |
381 | static inline int ecb_array_length (const T (&arr)[N]) |
427 | static inline int ecb_array_length (const T (&arr)[N]) |
382 | { |
428 | { |