1 | /* |
1 | /* |
2 | * libecb - http://software.schmorp.de/pkg/libecb |
2 | * libecb - http://software.schmorp.de/pkg/libecb |
3 | * |
3 | * |
4 | * Copyright (©) 2009-2011 Marc Alexander Lehmann <libecb@schmorp.de> |
4 | * Copyright (©) 2009-2012 Marc Alexander Lehmann <libecb@schmorp.de> |
5 | * Copyright (©) 2011 Emanuele Giaquinta |
5 | * Copyright (©) 2011 Emanuele Giaquinta |
6 | * All rights reserved. |
6 | * All rights reserved. |
7 | * |
7 | * |
8 | * Redistribution and use in source and binary forms, with or without modifica- |
8 | * Redistribution and use in source and binary forms, with or without modifica- |
9 | * tion, are permitted provided that the following conditions are met: |
9 | * tion, are permitted provided that the following conditions are met: |
… | |
… | |
67 | |
67 | |
68 | /* ECB_NO_THREADS - ecb is not used by multiple threads, ever */ |
68 | /* 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 */ |
69 | /* ECB_NO_SMP - ecb might be used in multiple threads, but only on a single cpu */ |
70 | |
70 | |
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 | #define ECB_MEMORY_FENCE_ACQUIRE do { } while (0) |
|
|
74 | #define ECB_MEMORY_FENCE_RELEASE do { } while (0) |
|
|
75 | #endif |
73 | #endif |
76 | |
74 | |
77 | #ifndef ECB_MEMORY_FENCE |
75 | #ifndef ECB_MEMORY_FENCE |
78 | #if ECB_GCC_VERSION(2,5) |
76 | #if ECB_GCC_VERSION(2,5) || defined(__INTEL_COMPILER) || (__llvm__ && __GNUC__) || __SUNPRO_C >= 0x5110 || __SUNPRO_CC >= 0x5110 |
79 | #if __x86 |
77 | #if __i386 || __i386__ |
80 | #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") |
81 | #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 */ |
82 | #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 */ |
83 | #elif __amd64 |
81 | #elif __amd64 || __amd64__ || __x86_64 || __x86_64__ |
84 | #define ECB_MEMORY_FENCE __asm__ __volatile__ ("mfence" : : : "memory") |
82 | #define ECB_MEMORY_FENCE __asm__ __volatile__ ("mfence" : : : "memory") |
85 | #define ECB_MEMORY_FENCE_ACQUIRE __asm__ __volatile__ ("lfence" : : : "memory") |
83 | #define ECB_MEMORY_FENCE_ACQUIRE __asm__ __volatile__ ("lfence" : : : "memory") |
86 | #define ECB_MEMORY_FENCE_RELEASE __asm__ __volatile__ ("sfence") /* play safe - not needed in any current cpu */ |
84 | #define ECB_MEMORY_FENCE_RELEASE __asm__ __volatile__ ("sfence") /* play safe - not needed in any current cpu */ |
|
|
85 | #elif __powerpc__ || __ppc__ || __powerpc64__ || __ppc64__ |
|
|
86 | #define ECB_MEMORY_FENCE __asm__ __volatile__ ("sync" : : : "memory") |
|
|
87 | #elif defined(__ARM_ARCH_6__ ) || defined(__ARM_ARCH_6J__ ) \ |
|
|
88 | || 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") |
|
|
90 | #elif defined(__ARM_ARCH_7__ ) || defined(__ARM_ARCH_7A__ ) \ |
|
|
91 | || defined(__ARM_ARCH_7M__) || defined(__ARM_ARCH_7R__ ) |
|
|
92 | #define ECB_MEMORY_FENCE __asm__ __volatile__ ("dmb" : : : "memory") |
|
|
93 | #elif __sparc || __sparc__ |
|
|
94 | #define ECB_MEMORY_FENCE __asm__ __volatile__ ("membar #LoadStore | #LoadLoad | #StoreStore | #StoreLoad | " : : : "memory") |
|
|
95 | #define ECB_MEMORY_FENCE_ACQUIRE __asm__ __volatile__ ("membar #LoadStore | #LoadLoad" : : : "memory") |
|
|
96 | #define ECB_MEMORY_FENCE_RELEASE __asm__ __volatile__ ("membar #LoadStore | #StoreStore") |
|
|
97 | #elif defined(__s390__) || defined(__s390x__) |
|
|
98 | #define ECB_MEMORY_FENCE __asm__ __volatile__ ("bcr 15,0" : : : "memory") |
87 | #endif |
99 | #endif |
88 | #endif |
100 | #endif |
89 | #endif |
101 | #endif |
90 | |
102 | |
91 | #ifndef ECB_MEMORY_FENCE |
103 | #ifndef ECB_MEMORY_FENCE |
92 | #if ECB_GCC_VERSION(4,4) || defined(__INTEL_COMPILER) |
104 | #if ECB_GCC_VERSION(4,4) || defined(__INTEL_COMPILER) || defined(__clang__) |
93 | #define ECB_MEMORY_FENCE __sync_synchronize () |
105 | #define ECB_MEMORY_FENCE __sync_synchronize () |
94 | #define ECB_MEMORY_FENCE_ACQUIRE __sync_synchronize () /* ({ char dummy = 0; __sync_lock_test_and_set (&dummy, 1); }) */ |
106 | /*#define ECB_MEMORY_FENCE_ACQUIRE ({ char dummy = 0; __sync_lock_test_and_set (&dummy, 1); }) */ |
95 | #define ECB_MEMORY_FENCE_RELEASE __sync_synchronize () /* ({ char dummy = 1; __sync_lock_release (&dummy ); }) */ |
107 | /*#define ECB_MEMORY_FENCE_RELEASE ({ char dummy = 1; __sync_lock_release (&dummy ); }) */ |
96 | #elif _MSC_VER >= 1400 /* VC++ 2005 */ |
108 | #elif _MSC_VER >= 1400 /* VC++ 2005 */ |
97 | #pragma intrinsic(_ReadBarrier,_WriteBarrier,_ReadWriteBarrier) |
109 | #pragma intrinsic(_ReadBarrier,_WriteBarrier,_ReadWriteBarrier) |
98 | #define ECB_MEMORY_FENCE _ReadWriteBarrier () |
110 | #define ECB_MEMORY_FENCE _ReadWriteBarrier () |
99 | #define ECB_MEMORY_FENCE_ACQUIRE _ReadWriteBarrier () /* according to msdn, _ReadBarrier is not a load fence */ |
111 | #define ECB_MEMORY_FENCE_ACQUIRE _ReadWriteBarrier () /* according to msdn, _ReadBarrier is not a load fence */ |
100 | #define ECB_MEMORY_FENCE_RELEASE _WriteBarrier () |
112 | #define ECB_MEMORY_FENCE_RELEASE _WriteBarrier () |
101 | #elif defined(_WIN32) |
113 | #elif defined(_WIN32) |
102 | #include <WinNT.h> |
114 | #include <WinNT.h> |
103 | #define ECB_MEMORY_FENCE MemoryBarrier () /* actually just xchg on x86... scary */ |
115 | #define ECB_MEMORY_FENCE MemoryBarrier () /* actually just xchg on x86... scary */ |
104 | #define ECB_MEMORY_FENCE_ACQUIRE ECB_MEMORY_FENCE |
116 | #elif __SUNPRO_C >= 0x5110 || __SUNPRO_CC >= 0x5110 |
105 | #define ECB_MEMORY_FENCE_RELEASE ECB_MEMORY_FENCE |
117 | #include <mbarrier.h> |
|
|
118 | #define ECB_MEMORY_FENCE __machine_rw_barrier () |
|
|
119 | #define ECB_MEMORY_FENCE_ACQUIRE __machine_r_barrier () |
|
|
120 | #define ECB_MEMORY_FENCE_RELEASE __machine_w_barrier () |
106 | #endif |
121 | #endif |
107 | #endif |
122 | #endif |
108 | |
123 | |
109 | #ifndef ECB_MEMORY_FENCE |
124 | #ifndef ECB_MEMORY_FENCE |
|
|
125 | #if !ECB_AVOID_PTHREADS |
110 | /* |
126 | /* |
111 | * if you get undefined symbol references to pthread_mutex_lock, |
127 | * if you get undefined symbol references to pthread_mutex_lock, |
112 | * or failure to find pthread.h, then you should implement |
128 | * or failure to find pthread.h, then you should implement |
113 | * the ECB_MEMORY_FENCE operations for your cpu/compiler |
129 | * the ECB_MEMORY_FENCE operations for your cpu/compiler |
114 | * OR provide pthread.h and link against the posix thread library |
130 | * OR provide pthread.h and link against the posix thread library |
115 | * of your system. |
131 | * of your system. |
116 | */ |
132 | */ |
117 | #include <pthread.h> |
133 | #include <pthread.h> |
118 | #define ECB_NEEDS_PTHREADS 1 |
134 | #define ECB_NEEDS_PTHREADS 1 |
119 | #define ECB_MEMORY_FENCE_NEEDS_PTHREADS 1 |
135 | #define ECB_MEMORY_FENCE_NEEDS_PTHREADS 1 |
120 | |
136 | |
121 | static pthread_mutex_t ecb_mf_lock = PTHREAD_MUTEX_INITIALIZER; |
137 | static pthread_mutex_t ecb_mf_lock = PTHREAD_MUTEX_INITIALIZER; |
122 | #define ECB_MEMORY_FENCE do { pthread_mutex_lock (&ecb_mf_lock); pthread_mutex_unlock (&ecb_mf_lock); } while (0) |
138 | #define ECB_MEMORY_FENCE do { pthread_mutex_lock (&ecb_mf_lock); pthread_mutex_unlock (&ecb_mf_lock); } while (0) |
|
|
139 | #endif |
|
|
140 | #endif |
|
|
141 | |
|
|
142 | #if !defined(ECB_MEMORY_FENCE_ACQUIRE) && defined(ECB_MEMORY_FENCE) |
123 | #define ECB_MEMORY_FENCE_ACQUIRE ECB_MEMORY_FENCE |
143 | #define ECB_MEMORY_FENCE_ACQUIRE ECB_MEMORY_FENCE |
|
|
144 | #endif |
|
|
145 | |
|
|
146 | #if !defined(ECB_MEMORY_FENCE_RELEASE) && defined(ECB_MEMORY_FENCE) |
124 | #define ECB_MEMORY_FENCE_RELEASE ECB_MEMORY_FENCE |
147 | #define ECB_MEMORY_FENCE_RELEASE ECB_MEMORY_FENCE |
125 | #endif |
148 | #endif |
126 | |
149 | |
127 | /*****************************************************************************/ |
150 | /*****************************************************************************/ |
128 | |
151 | |
… | |
… | |
277 | |
300 | |
278 | return r + ecb_ld32 (x); |
301 | return r + ecb_ld32 (x); |
279 | } |
302 | } |
280 | #endif |
303 | #endif |
281 | |
304 | |
|
|
305 | ecb_function_ uint8_t ecb_bitrev8 (uint8_t x) ecb_const; |
|
|
306 | ecb_function_ uint8_t ecb_bitrev8 (uint8_t x) |
|
|
307 | { |
|
|
308 | return ( (x * 0x0802U & 0x22110U) |
|
|
309 | | (x * 0x8020U & 0x88440U)) * 0x10101U >> 16; |
|
|
310 | } |
|
|
311 | |
|
|
312 | ecb_function_ uint16_t ecb_bitrev16 (uint16_t x) ecb_const; |
|
|
313 | ecb_function_ uint16_t ecb_bitrev16 (uint16_t x) |
|
|
314 | { |
|
|
315 | x = ((x >> 1) & 0x5555) | ((x & 0x5555) << 1); |
|
|
316 | x = ((x >> 2) & 0x3333) | ((x & 0x3333) << 2); |
|
|
317 | x = ((x >> 4) & 0x0f0f) | ((x & 0x0f0f) << 4); |
|
|
318 | x = ( x >> 8 ) | ( x << 8); |
|
|
319 | |
|
|
320 | return x; |
|
|
321 | } |
|
|
322 | |
|
|
323 | ecb_function_ uint32_t ecb_bitrev32 (uint32_t x) ecb_const; |
|
|
324 | ecb_function_ uint32_t ecb_bitrev32 (uint32_t x) |
|
|
325 | { |
|
|
326 | x = ((x >> 1) & 0x55555555) | ((x & 0x55555555) << 1); |
|
|
327 | x = ((x >> 2) & 0x33333333) | ((x & 0x33333333) << 2); |
|
|
328 | x = ((x >> 4) & 0x0f0f0f0f) | ((x & 0x0f0f0f0f) << 4); |
|
|
329 | x = ((x >> 8) & 0x00ff00ff) | ((x & 0x00ff00ff) << 8); |
|
|
330 | x = ( x >> 16 ) | ( x << 16); |
|
|
331 | |
|
|
332 | return x; |
|
|
333 | } |
|
|
334 | |
282 | /* popcount64 is only available on 64 bit cpus as gcc builtin */ |
335 | /* popcount64 is only available on 64 bit cpus as gcc builtin */ |
283 | /* so for this version we are lazy */ |
336 | /* so for this version we are lazy */ |
284 | ecb_function_ int ecb_popcount64 (uint64_t x) ecb_const; |
337 | ecb_function_ int ecb_popcount64 (uint64_t x) ecb_const; |
285 | ecb_function_ int |
338 | ecb_function_ int |
286 | ecb_popcount64 (uint64_t x) |
339 | ecb_popcount64 (uint64_t x) |
… | |
… | |
335 | |
388 | |
336 | #if ECB_GCC_VERSION(4,5) |
389 | #if ECB_GCC_VERSION(4,5) |
337 | #define ecb_unreachable() __builtin_unreachable () |
390 | #define ecb_unreachable() __builtin_unreachable () |
338 | #else |
391 | #else |
339 | /* this seems to work fine, but gcc always emits a warning for it :/ */ |
392 | /* this seems to work fine, but gcc always emits a warning for it :/ */ |
340 | ecb_function_ void ecb_unreachable (void) ecb_noreturn; |
393 | ecb_inline void ecb_unreachable (void) ecb_noreturn; |
341 | ecb_function_ void ecb_unreachable (void) { } |
394 | ecb_inline void ecb_unreachable (void) { } |
342 | #endif |
395 | #endif |
343 | |
396 | |
344 | /* try to tell the compiler that some condition is definitely true */ |
397 | /* try to tell the compiler that some condition is definitely true */ |
345 | #define ecb_assume(cond) do { if (!(cond)) ecb_unreachable (); } while (0) |
398 | #define ecb_assume(cond) do { if (!(cond)) ecb_unreachable (); } while (0) |
346 | |
399 | |
347 | ecb_function_ unsigned char ecb_byteorder_helper (void) ecb_const; |
400 | ecb_inline unsigned char ecb_byteorder_helper (void) ecb_const; |
348 | ecb_function_ unsigned char |
401 | ecb_inline unsigned char |
349 | ecb_byteorder_helper (void) |
402 | ecb_byteorder_helper (void) |
350 | { |
403 | { |
351 | const uint32_t u = 0x11223344; |
404 | const uint32_t u = 0x11223344; |
352 | return *(unsigned char *)&u; |
405 | return *(unsigned char *)&u; |
353 | } |
406 | } |
354 | |
407 | |
355 | ecb_function_ ecb_bool ecb_big_endian (void) ecb_const; |
408 | ecb_inline ecb_bool ecb_big_endian (void) ecb_const; |
356 | ecb_function_ ecb_bool ecb_big_endian (void) { return ecb_byteorder_helper () == 0x11; } |
409 | ecb_inline ecb_bool ecb_big_endian (void) { return ecb_byteorder_helper () == 0x11; } |
357 | ecb_function_ ecb_bool ecb_little_endian (void) ecb_const; |
410 | ecb_inline ecb_bool ecb_little_endian (void) ecb_const; |
358 | ecb_function_ ecb_bool ecb_little_endian (void) { return ecb_byteorder_helper () == 0x44; } |
411 | ecb_inline ecb_bool ecb_little_endian (void) { return ecb_byteorder_helper () == 0x44; } |
359 | |
412 | |
360 | #if ECB_GCC_VERSION(3,0) || ECB_C99 |
413 | #if ECB_GCC_VERSION(3,0) || ECB_C99 |
361 | #define ecb_mod(m,n) ((m) % (n) + ((m) % (n) < 0 ? (n) : 0)) |
414 | #define ecb_mod(m,n) ((m) % (n) + ((m) % (n) < 0 ? (n) : 0)) |
362 | #else |
415 | #else |
363 | #define ecb_mod(m,n) ((m) < 0 ? ((n) - 1 - ((-1 - (m)) % (n))) : ((m) % (n))) |
416 | #define ecb_mod(m,n) ((m) < 0 ? ((n) - 1 - ((-1 - (m)) % (n))) : ((m) % (n))) |
|
|
417 | #endif |
|
|
418 | |
|
|
419 | #if __cplusplus |
|
|
420 | template<typename T> |
|
|
421 | static inline T ecb_div_rd (T val, T div) |
|
|
422 | { |
|
|
423 | return val < 0 ? - ((-val + div - 1) / div) : (val ) / div; |
|
|
424 | } |
|
|
425 | template<typename T> |
|
|
426 | static inline T ecb_div_ru (T val, T div) |
|
|
427 | { |
|
|
428 | return val < 0 ? - ((-val ) / div) : (val + div - 1) / div; |
|
|
429 | } |
|
|
430 | #else |
|
|
431 | #define ecb_div_rd(val,div) ((val) < 0 ? - ((-(val) + (div) - 1) / (div)) : ((val) ) / (div)) |
|
|
432 | #define ecb_div_ru(val,div) ((val) < 0 ? - ((-(val) ) / (div)) : ((val) + (div) - 1) / (div)) |
364 | #endif |
433 | #endif |
365 | |
434 | |
366 | #if ecb_cplusplus_does_not_suck |
435 | #if ecb_cplusplus_does_not_suck |
367 | /* does not work for local types (http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2008/n2657.htm) */ |
436 | /* does not work for local types (http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2008/n2657.htm) */ |
368 | template<typename T, int N> |
437 | template<typename T, int N> |