| 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: |
| … | |
… | |
| 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 0x00010003 |
|
|
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; |
| 38 | typedef signed int int32_t; |
41 | typedef signed int int32_t; |
| 39 | typedef unsigned int uint32_t; |
42 | typedef unsigned int uint32_t; |
| 40 | #if __GNUC__ |
43 | #if __GNUC__ |
| 41 | typedef signed long long int64_t; |
44 | typedef signed long long int64_t; |
| 42 | typedef unsigned long long uint64_t; |
45 | typedef unsigned long long uint64_t; |
| 43 | #else |
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 |
| 47 | #else |
59 | #else |
| 48 | #include <inttypes.h> |
60 | #include <inttypes.h> |
|
|
61 | #if UINTMAX_MAX > 0xffffffffU |
|
|
62 | #define ECB_PTRSIZE 8 |
|
|
63 | #else |
|
|
64 | #define ECB_PTRSIZE 4 |
|
|
65 | #endif |
|
|
66 | #endif |
|
|
67 | |
|
|
68 | /* work around x32 idiocy by defining proper macros */ |
|
|
69 | #if __x86_64 || _M_AMD64 |
|
|
70 | #if __ILP32 |
|
|
71 | #define ECB_AMD64_X32 1 |
|
|
72 | #else |
|
|
73 | #define ECB_AMD64 1 |
|
|
74 | #endif |
| 49 | #endif |
75 | #endif |
| 50 | |
76 | |
| 51 | /* many compilers define _GNUC_ to some versions but then only implement |
77 | /* many compilers define _GNUC_ to some versions but then only implement |
| 52 | * what their idiot authors think are the "more important" extensions, |
78 | * what their idiot authors think are the "more important" extensions, |
| 53 | * causing enourmous grief in return for some better fake benchmark numbers. |
79 | * causing enormous grief in return for some better fake benchmark numbers. |
| 54 | * or so. |
80 | * or so. |
| 55 | * we try to detect these and simply assume they are not gcc - if they have |
81 | * 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. |
82 | * an issue with that they should have done it right in the first place. |
| 57 | */ |
83 | */ |
| 58 | #ifndef ECB_GCC_VERSION |
84 | #ifndef ECB_GCC_VERSION |
| 59 | #if !defined(__GNUC_MINOR__) || defined(__INTEL_COMPILER) || defined(__SUNPRO_C) || defined(__SUNPRO_CC) || defined(__llvm__) || defined(__clang__) |
85 | #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 |
86 | #define ECB_GCC_VERSION(major,minor) 0 |
| 61 | #else |
87 | #else |
| 62 | #define ECB_GCC_VERSION(major,minor) (__GNUC__ > (major) || (__GNUC__ == (major) && __GNUC_MINOR__ >= (minor))) |
88 | #define ECB_GCC_VERSION(major,minor) (__GNUC__ > (major) || (__GNUC__ == (major) && __GNUC_MINOR__ >= (minor))) |
| 63 | #endif |
89 | #endif |
| 64 | #endif |
90 | #endif |
| 65 | |
91 | |
|
|
92 | #define ECB_C (__STDC__+0) /* this assumes that __STDC__ is either empty or a number */ |
| 66 | #define ECB_C99 (__STDC_VERSION__ >= 199901L) |
93 | #define ECB_C99 (__STDC_VERSION__ >= 199901L) |
|
|
94 | #define ECB_C11 (__STDC_VERSION__ >= 201112L) |
|
|
95 | #define ECB_CPP (__cplusplus+0) |
|
|
96 | #define ECB_CPP11 (__cplusplus >= 201103L) |
|
|
97 | |
|
|
98 | #if ECB_CPP |
|
|
99 | #define ECB_EXTERN_C extern "C" |
|
|
100 | #define ECB_EXTERN_C_BEG ECB_EXTERN_C { |
|
|
101 | #define ECB_EXTERN_C_END } |
|
|
102 | #else |
|
|
103 | #define ECB_EXTERN_C extern |
|
|
104 | #define ECB_EXTERN_C_BEG |
|
|
105 | #define ECB_EXTERN_C_END |
|
|
106 | #endif |
|
|
107 | |
|
|
108 | /*****************************************************************************/ |
|
|
109 | |
|
|
110 | /* ECB_NO_THREADS - ecb is not used by multiple threads, ever */ |
|
|
111 | /* ECB_NO_SMP - ecb might be used in multiple threads, but only on a single cpu */ |
|
|
112 | |
|
|
113 | #if ECB_NO_THREADS |
|
|
114 | #define ECB_NO_SMP 1 |
|
|
115 | #endif |
|
|
116 | |
|
|
117 | #if ECB_NO_SMP |
|
|
118 | #define ECB_MEMORY_FENCE do { } while (0) |
|
|
119 | #endif |
|
|
120 | |
|
|
121 | #ifndef ECB_MEMORY_FENCE |
|
|
122 | #if ECB_GCC_VERSION(2,5) || defined __INTEL_COMPILER || (__llvm__ && __GNUC__) || __SUNPRO_C >= 0x5110 || __SUNPRO_CC >= 0x5110 |
|
|
123 | #if __i386 || __i386__ |
|
|
124 | #define ECB_MEMORY_FENCE __asm__ __volatile__ ("lock; orb $0, -1(%%esp)" : : : "memory") |
|
|
125 | #define ECB_MEMORY_FENCE_ACQUIRE __asm__ __volatile__ ("" : : : "memory") |
|
|
126 | #define ECB_MEMORY_FENCE_RELEASE __asm__ __volatile__ ("") |
|
|
127 | #elif __amd64 || __amd64__ || __x86_64 || __x86_64__ |
|
|
128 | #define ECB_MEMORY_FENCE __asm__ __volatile__ ("mfence" : : : "memory") |
|
|
129 | #define ECB_MEMORY_FENCE_ACQUIRE __asm__ __volatile__ ("" : : : "memory") |
|
|
130 | #define ECB_MEMORY_FENCE_RELEASE __asm__ __volatile__ ("") |
|
|
131 | #elif __powerpc__ || __ppc__ || __powerpc64__ || __ppc64__ |
|
|
132 | #define ECB_MEMORY_FENCE __asm__ __volatile__ ("sync" : : : "memory") |
|
|
133 | #elif defined __ARM_ARCH_6__ || defined __ARM_ARCH_6J__ \ |
|
|
134 | || defined __ARM_ARCH_6K__ || defined __ARM_ARCH_6ZK__ |
|
|
135 | #define ECB_MEMORY_FENCE __asm__ __volatile__ ("mcr p15,0,%0,c7,c10,5" : : "r" (0) : "memory") |
|
|
136 | #elif defined __ARM_ARCH_7__ || defined __ARM_ARCH_7A__ \ |
|
|
137 | || defined __ARM_ARCH_7M__ || defined __ARM_ARCH_7R__ |
|
|
138 | #define ECB_MEMORY_FENCE __asm__ __volatile__ ("dmb" : : : "memory") |
|
|
139 | #elif __sparc || __sparc__ |
|
|
140 | #define ECB_MEMORY_FENCE __asm__ __volatile__ ("membar #LoadStore | #LoadLoad | #StoreStore | #StoreLoad" : : : "memory") |
|
|
141 | #define ECB_MEMORY_FENCE_ACQUIRE __asm__ __volatile__ ("membar #LoadStore | #LoadLoad" : : : "memory") |
|
|
142 | #define ECB_MEMORY_FENCE_RELEASE __asm__ __volatile__ ("membar #LoadStore | #StoreStore") |
|
|
143 | #elif defined __s390__ || defined __s390x__ |
|
|
144 | #define ECB_MEMORY_FENCE __asm__ __volatile__ ("bcr 15,0" : : : "memory") |
|
|
145 | #elif defined __mips__ |
|
|
146 | /* GNU/Linux emulates sync on mips1 architectures, so we force it's use */ |
|
|
147 | /* anybody else who still uses mips1 is supposed to send in their version, with detection code. */ |
|
|
148 | #define ECB_MEMORY_FENCE __asm__ __volatile__ (".set mips2; sync; .set mips0" : : : "memory") |
|
|
149 | #elif defined __alpha__ |
|
|
150 | #define ECB_MEMORY_FENCE __asm__ __volatile__ ("mb" : : : "memory") |
|
|
151 | #elif defined __hppa__ |
|
|
152 | #define ECB_MEMORY_FENCE __asm__ __volatile__ ("" : : : "memory") |
|
|
153 | #define ECB_MEMORY_FENCE_RELEASE __asm__ __volatile__ ("") |
|
|
154 | #elif defined __ia64__ |
|
|
155 | #define ECB_MEMORY_FENCE __asm__ __volatile__ ("mf" : : : "memory") |
|
|
156 | #endif |
|
|
157 | #endif |
|
|
158 | #endif |
|
|
159 | |
|
|
160 | #ifndef ECB_MEMORY_FENCE |
|
|
161 | #if ECB_GCC_VERSION(4,7) |
|
|
162 | /* see comment below (stdatomic.h) about the C11 memory model. */ |
|
|
163 | #define ECB_MEMORY_FENCE __atomic_thread_fence (__ATOMIC_SEQ_CST) |
|
|
164 | |
|
|
165 | /* The __has_feature syntax from clang is so misdesigned that we cannot use it |
|
|
166 | * without risking compile time errors with other compilers. We *could* |
|
|
167 | * define our own ecb_clang_has_feature, but I just can't be bothered to work |
|
|
168 | * around this shit time and again. |
|
|
169 | * #elif defined __clang && __has_feature (cxx_atomic) |
|
|
170 | * // see comment below (stdatomic.h) about the C11 memory model. |
|
|
171 | * #define ECB_MEMORY_FENCE __c11_atomic_thread_fence (__ATOMIC_SEQ_CST) |
|
|
172 | */ |
|
|
173 | |
|
|
174 | #elif ECB_GCC_VERSION(4,4) || defined __INTEL_COMPILER || defined __clang__ |
|
|
175 | #define ECB_MEMORY_FENCE __sync_synchronize () |
|
|
176 | #elif _MSC_VER >= 1400 /* VC++ 2005 */ |
|
|
177 | #pragma intrinsic(_ReadBarrier,_WriteBarrier,_ReadWriteBarrier) |
|
|
178 | #define ECB_MEMORY_FENCE _ReadWriteBarrier () |
|
|
179 | #define ECB_MEMORY_FENCE_ACQUIRE _ReadWriteBarrier () /* according to msdn, _ReadBarrier is not a load fence */ |
|
|
180 | #define ECB_MEMORY_FENCE_RELEASE _WriteBarrier () |
|
|
181 | #elif defined _WIN32 |
|
|
182 | #include <WinNT.h> |
|
|
183 | #define ECB_MEMORY_FENCE MemoryBarrier () /* actually just xchg on x86... scary */ |
|
|
184 | #elif __SUNPRO_C >= 0x5110 || __SUNPRO_CC >= 0x5110 |
|
|
185 | #include <mbarrier.h> |
|
|
186 | #define ECB_MEMORY_FENCE __machine_rw_barrier () |
|
|
187 | #define ECB_MEMORY_FENCE_ACQUIRE __machine_r_barrier () |
|
|
188 | #define ECB_MEMORY_FENCE_RELEASE __machine_w_barrier () |
|
|
189 | #elif __xlC__ |
|
|
190 | #define ECB_MEMORY_FENCE __sync () |
|
|
191 | #endif |
|
|
192 | #endif |
|
|
193 | |
|
|
194 | #ifndef ECB_MEMORY_FENCE |
|
|
195 | #if ECB_C11 && !defined __STDC_NO_ATOMICS__ |
|
|
196 | /* we assume that these memory fences work on all variables/all memory accesses, */ |
|
|
197 | /* not just C11 atomics and atomic accesses */ |
|
|
198 | #include <stdatomic.h> |
|
|
199 | /* Unfortunately, neither gcc 4.7 nor clang 3.1 generate any instructions for */ |
|
|
200 | /* any fence other than seq_cst, which isn't very efficient for us. */ |
|
|
201 | /* Why that is, we don't know - either the C11 memory model is quite useless */ |
|
|
202 | /* for most usages, or gcc and clang have a bug */ |
|
|
203 | /* I *currently* lean towards the latter, and inefficiently implement */ |
|
|
204 | /* all three of ecb's fences as a seq_cst fence */ |
|
|
205 | #define ECB_MEMORY_FENCE atomic_thread_fence (memory_order_seq_cst) |
|
|
206 | #endif |
|
|
207 | #endif |
|
|
208 | |
|
|
209 | #ifndef ECB_MEMORY_FENCE |
|
|
210 | #if !ECB_AVOID_PTHREADS |
|
|
211 | /* |
|
|
212 | * if you get undefined symbol references to pthread_mutex_lock, |
|
|
213 | * or failure to find pthread.h, then you should implement |
|
|
214 | * the ECB_MEMORY_FENCE operations for your cpu/compiler |
|
|
215 | * OR provide pthread.h and link against the posix thread library |
|
|
216 | * of your system. |
|
|
217 | */ |
|
|
218 | #include <pthread.h> |
|
|
219 | #define ECB_NEEDS_PTHREADS 1 |
|
|
220 | #define ECB_MEMORY_FENCE_NEEDS_PTHREADS 1 |
|
|
221 | |
|
|
222 | static pthread_mutex_t ecb_mf_lock = PTHREAD_MUTEX_INITIALIZER; |
|
|
223 | #define ECB_MEMORY_FENCE do { pthread_mutex_lock (&ecb_mf_lock); pthread_mutex_unlock (&ecb_mf_lock); } while (0) |
|
|
224 | #endif |
|
|
225 | #endif |
|
|
226 | |
|
|
227 | #if !defined ECB_MEMORY_FENCE_ACQUIRE && defined ECB_MEMORY_FENCE |
|
|
228 | #define ECB_MEMORY_FENCE_ACQUIRE ECB_MEMORY_FENCE |
|
|
229 | #endif |
|
|
230 | |
|
|
231 | #if !defined ECB_MEMORY_FENCE_RELEASE && defined ECB_MEMORY_FENCE |
|
|
232 | #define ECB_MEMORY_FENCE_RELEASE ECB_MEMORY_FENCE |
|
|
233 | #endif |
|
|
234 | |
|
|
235 | /*****************************************************************************/ |
| 67 | |
236 | |
| 68 | #if __cplusplus |
237 | #if __cplusplus |
| 69 | #define ecb_inline static inline |
238 | #define ecb_inline static inline |
| 70 | #elif ECB_GCC_VERSION(2,5) |
239 | #elif ECB_GCC_VERSION(2,5) |
| 71 | #define ecb_inline static __inline__ |
240 | #define ecb_inline static __inline__ |
| … | |
… | |
| 110 | #elif ECB_GCC_VERSION(3,0) |
279 | #elif ECB_GCC_VERSION(3,0) |
| 111 | #define ecb_decltype(x) __typeof(x) |
280 | #define ecb_decltype(x) __typeof(x) |
| 112 | #endif |
281 | #endif |
| 113 | |
282 | |
| 114 | #define ecb_noinline ecb_attribute ((__noinline__)) |
283 | #define ecb_noinline ecb_attribute ((__noinline__)) |
| 115 | #define ecb_noreturn ecb_attribute ((__noreturn__)) |
|
|
| 116 | #define ecb_unused ecb_attribute ((__unused__)) |
284 | #define ecb_unused ecb_attribute ((__unused__)) |
| 117 | #define ecb_const ecb_attribute ((__const__)) |
285 | #define ecb_const ecb_attribute ((__const__)) |
| 118 | #define ecb_pure ecb_attribute ((__pure__)) |
286 | #define ecb_pure ecb_attribute ((__pure__)) |
|
|
287 | |
|
|
288 | #if ECB_C11 |
|
|
289 | #define ecb_noreturn _Noreturn |
|
|
290 | #else |
|
|
291 | #define ecb_noreturn ecb_attribute ((__noreturn__)) |
|
|
292 | #endif |
| 119 | |
293 | |
| 120 | #if ECB_GCC_VERSION(4,3) |
294 | #if ECB_GCC_VERSION(4,3) |
| 121 | #define ecb_artificial ecb_attribute ((__artificial__)) |
295 | #define ecb_artificial ecb_attribute ((__artificial__)) |
| 122 | #define ecb_hot ecb_attribute ((__hot__)) |
296 | #define ecb_hot ecb_attribute ((__hot__)) |
| 123 | #define ecb_cold ecb_attribute ((__cold__)) |
297 | #define ecb_cold ecb_attribute ((__cold__)) |
| … | |
… | |
| 152 | { |
326 | { |
| 153 | int r = 0; |
327 | int r = 0; |
| 154 | |
328 | |
| 155 | x &= ~x + 1; /* this isolates the lowest bit */ |
329 | x &= ~x + 1; /* this isolates the lowest bit */ |
| 156 | |
330 | |
|
|
331 | #if ECB_branchless_on_i386 |
|
|
332 | r += !!(x & 0xaaaaaaaa) << 0; |
|
|
333 | r += !!(x & 0xcccccccc) << 1; |
|
|
334 | r += !!(x & 0xf0f0f0f0) << 2; |
|
|
335 | r += !!(x & 0xff00ff00) << 3; |
|
|
336 | r += !!(x & 0xffff0000) << 4; |
|
|
337 | #else |
| 157 | if (x & 0xaaaaaaaa) r += 1; |
338 | if (x & 0xaaaaaaaa) r += 1; |
| 158 | if (x & 0xcccccccc) r += 2; |
339 | if (x & 0xcccccccc) r += 2; |
| 159 | if (x & 0xf0f0f0f0) r += 4; |
340 | if (x & 0xf0f0f0f0) r += 4; |
| 160 | if (x & 0xff00ff00) r += 8; |
341 | if (x & 0xff00ff00) r += 8; |
| 161 | if (x & 0xffff0000) r += 16; |
342 | if (x & 0xffff0000) r += 16; |
|
|
343 | #endif |
| 162 | |
344 | |
| 163 | return r; |
345 | return r; |
| 164 | } |
346 | } |
| 165 | |
347 | |
| 166 | ecb_function_ int ecb_ctz64 (uint64_t x) ecb_const; |
348 | ecb_function_ int ecb_ctz64 (uint64_t x) ecb_const; |
| 167 | ecb_function_ int |
349 | ecb_function_ int |
| 168 | ecb_ctz64 (uint64_t x) |
350 | ecb_ctz64 (uint64_t x) |
| 169 | { |
351 | { |
| 170 | int shift = x & 0xffffffffU ? 0 : 32; |
352 | int shift = x & 0xffffffffU ? 0 : 32; |
| 171 | return ecb_ctz (x >> shift) + shift; |
353 | return ecb_ctz32 (x >> shift) + shift; |
| 172 | } |
354 | } |
| 173 | |
355 | |
| 174 | ecb_function_ int ecb_popcount32 (uint32_t x) ecb_const; |
356 | ecb_function_ int ecb_popcount32 (uint32_t x) ecb_const; |
| 175 | ecb_function_ int |
357 | ecb_function_ int |
| 176 | ecb_popcount32 (uint32_t x) |
358 | ecb_popcount32 (uint32_t x) |
| … | |
… | |
| 181 | x *= 0x01010101; |
363 | x *= 0x01010101; |
| 182 | |
364 | |
| 183 | return x >> 24; |
365 | return x >> 24; |
| 184 | } |
366 | } |
| 185 | |
367 | |
| 186 | /* you have the choice beetween something with a table lookup, */ |
|
|
| 187 | /* something using lots of bit arithmetic and a simple loop */ |
|
|
| 188 | /* we went for the loop */ |
|
|
| 189 | ecb_function_ int ecb_ld32 (uint32_t x) ecb_const; |
368 | ecb_function_ int ecb_ld32 (uint32_t x) ecb_const; |
| 190 | ecb_function_ int ecb_ld32 (uint32_t x) |
369 | ecb_function_ int ecb_ld32 (uint32_t x) |
| 191 | { |
370 | { |
| 192 | int r = -1; |
371 | int r = 0; |
| 193 | |
372 | |
| 194 | do |
373 | if (x >> 16) { x >>= 16; r += 16; } |
| 195 | { |
374 | if (x >> 8) { x >>= 8; r += 8; } |
| 196 | x >>= 1; |
375 | if (x >> 4) { x >>= 4; r += 4; } |
| 197 | ++r; |
376 | if (x >> 2) { x >>= 2; r += 2; } |
| 198 | } |
377 | if (x >> 1) { r += 1; } |
| 199 | while (x); |
|
|
| 200 | |
378 | |
| 201 | return r; |
379 | return r; |
| 202 | } |
380 | } |
| 203 | |
381 | |
| 204 | ecb_function_ int ecb_ld64 (uint64_t x) ecb_const; |
382 | ecb_function_ int ecb_ld64 (uint64_t x) ecb_const; |
| 205 | ecb_function_ int ecb_ld64 (uint64_t x) |
383 | ecb_function_ int ecb_ld64 (uint64_t x) |
| 206 | { |
384 | { |
| 207 | int r = -1; |
385 | int r = 0; |
| 208 | |
386 | |
| 209 | do |
387 | if (x >> 32) { x >>= 32; r += 32; } |
| 210 | { |
|
|
| 211 | x >>= 1; |
|
|
| 212 | ++r; |
|
|
| 213 | } |
|
|
| 214 | while (x); |
|
|
| 215 | |
388 | |
|
|
389 | return r + ecb_ld32 (x); |
|
|
390 | } |
|
|
391 | #endif |
|
|
392 | |
|
|
393 | ecb_function_ ecb_bool ecb_is_pot32 (uint32_t x) ecb_const; |
|
|
394 | ecb_function_ ecb_bool ecb_is_pot32 (uint32_t x) { return !(x & (x - 1)); } |
|
|
395 | ecb_function_ ecb_bool ecb_is_pot64 (uint64_t x) ecb_const; |
|
|
396 | ecb_function_ ecb_bool ecb_is_pot64 (uint64_t x) { return !(x & (x - 1)); } |
|
|
397 | |
|
|
398 | ecb_function_ uint8_t ecb_bitrev8 (uint8_t x) ecb_const; |
|
|
399 | ecb_function_ uint8_t ecb_bitrev8 (uint8_t x) |
|
|
400 | { |
|
|
401 | return ( (x * 0x0802U & 0x22110U) |
|
|
402 | | (x * 0x8020U & 0x88440U)) * 0x10101U >> 16; |
|
|
403 | } |
|
|
404 | |
|
|
405 | ecb_function_ uint16_t ecb_bitrev16 (uint16_t x) ecb_const; |
|
|
406 | ecb_function_ uint16_t ecb_bitrev16 (uint16_t x) |
|
|
407 | { |
|
|
408 | x = ((x >> 1) & 0x5555) | ((x & 0x5555) << 1); |
|
|
409 | x = ((x >> 2) & 0x3333) | ((x & 0x3333) << 2); |
|
|
410 | x = ((x >> 4) & 0x0f0f) | ((x & 0x0f0f) << 4); |
|
|
411 | x = ( x >> 8 ) | ( x << 8); |
|
|
412 | |
| 216 | return r; |
413 | return x; |
| 217 | } |
414 | } |
| 218 | #endif |
415 | |
|
|
416 | ecb_function_ uint32_t ecb_bitrev32 (uint32_t x) ecb_const; |
|
|
417 | ecb_function_ uint32_t ecb_bitrev32 (uint32_t x) |
|
|
418 | { |
|
|
419 | x = ((x >> 1) & 0x55555555) | ((x & 0x55555555) << 1); |
|
|
420 | x = ((x >> 2) & 0x33333333) | ((x & 0x33333333) << 2); |
|
|
421 | x = ((x >> 4) & 0x0f0f0f0f) | ((x & 0x0f0f0f0f) << 4); |
|
|
422 | x = ((x >> 8) & 0x00ff00ff) | ((x & 0x00ff00ff) << 8); |
|
|
423 | x = ( x >> 16 ) | ( x << 16); |
|
|
424 | |
|
|
425 | return x; |
|
|
426 | } |
| 219 | |
427 | |
| 220 | /* popcount64 is only available on 64 bit cpus as gcc builtin */ |
428 | /* popcount64 is only available on 64 bit cpus as gcc builtin */ |
| 221 | /* so for this version we are lazy */ |
429 | /* so for this version we are lazy */ |
| 222 | ecb_function_ int ecb_popcount64 (uint64_t x) ecb_const; |
430 | ecb_function_ int ecb_popcount64 (uint64_t x) ecb_const; |
| 223 | ecb_function_ int |
431 | ecb_function_ int |
| 224 | ecb_popcount64 (uint64_t x) |
432 | ecb_popcount64 (uint64_t x) |
| 225 | { |
433 | { |
| 226 | return ecb_popcount32 (x) + ecb_popcount32 (x >> 32); |
434 | return ecb_popcount32 (x) + ecb_popcount32 (x >> 32); |
| 227 | } |
435 | } |
| 228 | |
436 | |
|
|
437 | ecb_inline uint8_t ecb_rotl8 (uint8_t x, unsigned int count) ecb_const; |
|
|
438 | ecb_inline uint8_t ecb_rotr8 (uint8_t x, unsigned int count) ecb_const; |
|
|
439 | ecb_inline uint16_t ecb_rotl16 (uint16_t x, unsigned int count) ecb_const; |
|
|
440 | ecb_inline uint16_t ecb_rotr16 (uint16_t x, unsigned int count) ecb_const; |
|
|
441 | ecb_inline uint32_t ecb_rotl32 (uint32_t x, unsigned int count) ecb_const; |
|
|
442 | ecb_inline uint32_t ecb_rotr32 (uint32_t x, unsigned int count) ecb_const; |
|
|
443 | ecb_inline uint64_t ecb_rotl64 (uint64_t x, unsigned int count) ecb_const; |
|
|
444 | ecb_inline uint64_t ecb_rotr64 (uint64_t x, unsigned int count) ecb_const; |
|
|
445 | |
|
|
446 | ecb_inline uint8_t ecb_rotl8 (uint8_t x, unsigned int count) { return (x >> ( 8 - count)) | (x << count); } |
|
|
447 | ecb_inline uint8_t ecb_rotr8 (uint8_t x, unsigned int count) { return (x << ( 8 - count)) | (x >> count); } |
|
|
448 | ecb_inline uint16_t ecb_rotl16 (uint16_t x, unsigned int count) { return (x >> (16 - count)) | (x << count); } |
|
|
449 | ecb_inline uint16_t ecb_rotr16 (uint16_t x, unsigned int count) { return (x << (16 - count)) | (x >> count); } |
|
|
450 | ecb_inline uint32_t ecb_rotl32 (uint32_t x, unsigned int count) { return (x >> (32 - count)) | (x << count); } |
|
|
451 | ecb_inline uint32_t ecb_rotr32 (uint32_t x, unsigned int count) { return (x << (32 - count)) | (x >> count); } |
|
|
452 | ecb_inline uint64_t ecb_rotl64 (uint64_t x, unsigned int count) { return (x >> (64 - count)) | (x << count); } |
|
|
453 | ecb_inline uint64_t ecb_rotr64 (uint64_t x, unsigned int count) { return (x << (64 - count)) | (x >> count); } |
|
|
454 | |
| 229 | #if ECB_GCC_VERSION(4,3) |
455 | #if ECB_GCC_VERSION(4,3) |
| 230 | #define ecb_bswap16(x) (__builtin_bswap32 (x) >> 16) |
456 | #define ecb_bswap16(x) (__builtin_bswap32 (x) >> 16) |
| 231 | #define ecb_bswap32(x) __builtin_bswap32 (x) |
457 | #define ecb_bswap32(x) __builtin_bswap32 (x) |
| 232 | #define ecb_bswap64(x) __builtin_bswap64 (x) |
458 | #define ecb_bswap64(x) __builtin_bswap64 (x) |
| 233 | #else |
459 | #else |
| 234 | ecb_function_ uint32_t ecb_bswap16 (uint32_t x) ecb_const; |
460 | ecb_function_ uint16_t ecb_bswap16 (uint16_t x) ecb_const; |
| 235 | ecb_function_ uint32_t |
461 | ecb_function_ uint16_t |
| 236 | ecb_bswap16 (uint32_t x) |
462 | ecb_bswap16 (uint16_t x) |
| 237 | { |
463 | { |
| 238 | return ((x >> 8) & 0xff) |
464 | return ecb_rotl16 (x, 8); |
| 239 | | ((x << 8) & 0x00ff0000) |
|
|
| 240 | | (x << 24); |
|
|
| 241 | } |
465 | } |
| 242 | |
466 | |
| 243 | ecb_function_ uint32_t ecb_bswap32 (uint32_t x) ecb_const; |
467 | ecb_function_ uint32_t ecb_bswap32 (uint32_t x) ecb_const; |
| 244 | ecb_function_ uint32_t |
468 | ecb_function_ uint32_t |
| 245 | ecb_bswap32 (uint32_t x) |
469 | ecb_bswap32 (uint32_t x) |
| 246 | { |
470 | { |
| 247 | return (x >> 24) |
471 | return (((uint32_t)ecb_bswap16 (x)) << 16) | ecb_bswap16 (x >> 16); |
| 248 | | ((x >> 8) & 0x0000ff00) |
|
|
| 249 | | ((x << 8) & 0x00ff0000) |
|
|
| 250 | | (x << 24); |
|
|
| 251 | } |
472 | } |
| 252 | |
473 | |
| 253 | ecb_function_ uint64_t ecb_bswap64 (uint64_t x) ecb_const; |
474 | ecb_function_ uint64_t ecb_bswap64 (uint64_t x) ecb_const; |
| 254 | ecb_function_ uint64_t |
475 | ecb_function_ uint64_t |
| 255 | ecb_bswap64 (uint64_t x) |
476 | ecb_bswap64 (uint64_t x) |
| 256 | { |
477 | { |
| 257 | return (((uint64_t)ecb_bswap32 (x)) << 32) |
478 | return (((uint64_t)ecb_bswap32 (x)) << 32) | ecb_bswap32 (x >> 32); |
| 258 | | ecb_bswap32 (x >> 32); |
|
|
| 259 | } |
479 | } |
| 260 | #endif |
480 | #endif |
| 261 | |
481 | |
| 262 | #if ECB_GCC_VERSION(4,5) |
482 | #if ECB_GCC_VERSION(4,5) |
| 263 | #define ecb_unreachable() __builtin_unreachable () |
483 | #define ecb_unreachable() __builtin_unreachable () |
| 264 | #else |
484 | #else |
| 265 | /* this seems to work fine, but gcc always emits a warning for it :/ */ |
485 | /* this seems to work fine, but gcc always emits a warning for it :/ */ |
| 266 | ecb_function_ void ecb_unreachable (void) ecb_noreturn; |
486 | ecb_inline void ecb_unreachable (void) ecb_noreturn; |
| 267 | ecb_function_ void ecb_unreachable (void) { } |
487 | ecb_inline void ecb_unreachable (void) { } |
| 268 | #endif |
488 | #endif |
| 269 | |
489 | |
| 270 | /* try to tell the compiler that some condition is definitely true */ |
490 | /* try to tell the compiler that some condition is definitely true */ |
| 271 | #define ecb_assume(cond) do { if (!(cond)) ecb_unreachable (); } while (0) |
491 | #define ecb_assume(cond) if (!(cond)) ecb_unreachable (); else 0 |
| 272 | |
492 | |
| 273 | ecb_function_ unsigned char ecb_byteorder_helper (void) ecb_const; |
493 | ecb_inline unsigned char ecb_byteorder_helper (void) ecb_const; |
| 274 | ecb_function_ unsigned char |
494 | ecb_inline unsigned char |
| 275 | ecb_byteorder_helper (void) |
495 | ecb_byteorder_helper (void) |
| 276 | { |
496 | { |
| 277 | const uint32_t u = 0x11223344; |
497 | /* the union code still generates code under pressure in gcc, */ |
| 278 | return *(unsigned char *)&u; |
498 | /* but less than using pointers, and always seems to */ |
|
|
499 | /* successfully return a constant. */ |
|
|
500 | /* the reason why we have this horrible preprocessor mess */ |
|
|
501 | /* is to avoid it in all cases, at least on common architectures */ |
|
|
502 | /* or when using a recent enough gcc version (>= 4.6) */ |
|
|
503 | #if __i386 || __i386__ || _M_X86 || __amd64 || __amd64__ || _M_X64 |
|
|
504 | return 0x44; |
|
|
505 | #elif __BYTE_ORDER__ && __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__ |
|
|
506 | return 0x44; |
|
|
507 | #elif __BYTE_ORDER__ && __BYTE_ORDER__ == __ORDER_BIG_ENDIAN__ |
|
|
508 | return 0x11; |
|
|
509 | #else |
|
|
510 | union |
|
|
511 | { |
|
|
512 | uint32_t i; |
|
|
513 | uint8_t c; |
|
|
514 | } u = { 0x11223344 }; |
|
|
515 | return u.c; |
|
|
516 | #endif |
| 279 | } |
517 | } |
| 280 | |
518 | |
| 281 | ecb_function_ ecb_bool ecb_big_endian (void) ecb_const; |
519 | ecb_inline ecb_bool ecb_big_endian (void) ecb_const; |
| 282 | ecb_function_ ecb_bool ecb_big_endian (void) { return ecb_byteorder_helper () == 0x11; } |
520 | ecb_inline ecb_bool ecb_big_endian (void) { return ecb_byteorder_helper () == 0x11; } |
| 283 | ecb_function_ ecb_bool ecb_little_endian (void) ecb_const; |
521 | ecb_inline ecb_bool ecb_little_endian (void) ecb_const; |
| 284 | ecb_function_ ecb_bool ecb_little_endian (void) { return ecb_byteorder_helper () == 0x44; } |
522 | ecb_inline ecb_bool ecb_little_endian (void) { return ecb_byteorder_helper () == 0x44; } |
| 285 | |
523 | |
| 286 | #if ECB_GCC_VERSION(3,0) || ECB_C99 |
524 | #if ECB_GCC_VERSION(3,0) || ECB_C99 |
| 287 | #define ecb_mod(m,n) ((m) % (n) + ((m) % (n) < 0 ? (n) : 0)) |
525 | #define ecb_mod(m,n) ((m) % (n) + ((m) % (n) < 0 ? (n) : 0)) |
| 288 | #else |
526 | #else |
| 289 | #define ecb_mod(m,n) ((m) < 0 ? ((n) - 1 - ((-1 - (m)) % (n))) : ((m) % (n))) |
527 | #define ecb_mod(m,n) ((m) < 0 ? ((n) - 1 - ((-1 - (m)) % (n))) : ((m) % (n))) |
|
|
528 | #endif |
|
|
529 | |
|
|
530 | #if __cplusplus |
|
|
531 | template<typename T> |
|
|
532 | static inline T ecb_div_rd (T val, T div) |
|
|
533 | { |
|
|
534 | return val < 0 ? - ((-val + div - 1) / div) : (val ) / div; |
|
|
535 | } |
|
|
536 | template<typename T> |
|
|
537 | static inline T ecb_div_ru (T val, T div) |
|
|
538 | { |
|
|
539 | return val < 0 ? - ((-val ) / div) : (val + div - 1) / div; |
|
|
540 | } |
|
|
541 | #else |
|
|
542 | #define ecb_div_rd(val,div) ((val) < 0 ? - ((-(val) + (div) - 1) / (div)) : ((val) ) / (div)) |
|
|
543 | #define ecb_div_ru(val,div) ((val) < 0 ? - ((-(val) ) / (div)) : ((val) + (div) - 1) / (div)) |
| 290 | #endif |
544 | #endif |
| 291 | |
545 | |
| 292 | #if ecb_cplusplus_does_not_suck |
546 | #if ecb_cplusplus_does_not_suck |
| 293 | /* does not work for local types (http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2008/n2657.htm) */ |
547 | /* does not work for local types (http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2008/n2657.htm) */ |
| 294 | template<typename T, int N> |
548 | template<typename T, int N> |
| … | |
… | |
| 298 | } |
552 | } |
| 299 | #else |
553 | #else |
| 300 | #define ecb_array_length(name) (sizeof (name) / sizeof (name [0])) |
554 | #define ecb_array_length(name) (sizeof (name) / sizeof (name [0])) |
| 301 | #endif |
555 | #endif |
| 302 | |
556 | |
| 303 | ecb_inline uint32_t ecb_rotr32 (uint32_t x, unsigned int count) ecb_const; |
557 | /*******************************************************************************/ |
| 304 | ecb_inline uint32_t |
558 | /* floating point stuff, can be disabled by defining ECB_NO_LIBM */ |
| 305 | ecb_rotr32 (uint32_t x, unsigned int count) |
|
|
| 306 | { |
|
|
| 307 | return (x << (32 - count)) | (x >> count); |
|
|
| 308 | } |
|
|
| 309 | |
559 | |
| 310 | ecb_inline uint32_t ecb_rotl32 (uint32_t x, unsigned int count) ecb_const; |
560 | /* basically, everything uses "ieee pure-endian" floating point numbers */ |
| 311 | ecb_inline uint32_t |
561 | /* the only noteworthy exception is ancient armle, which uses order 43218765 */ |
| 312 | ecb_rotl32 (uint32_t x, unsigned int count) |
562 | #if 0 \ |
| 313 | { |
563 | || __i386 || __i386__ \ |
| 314 | return (x >> (32 - count)) | (x << count); |
564 | || __amd64 || __amd64__ || __x86_64 || __x86_64__ \ |
| 315 | } |
565 | || __powerpc__ || __ppc__ || __powerpc64__ || __ppc64__ \ |
|
|
566 | || defined __arm__ && defined __ARM_EABI__ \ |
|
|
567 | || defined __s390__ || defined __s390x__ \ |
|
|
568 | || defined __mips__ \ |
|
|
569 | || defined __alpha__ \ |
|
|
570 | || defined __hppa__ \ |
|
|
571 | || defined __ia64__ \ |
|
|
572 | || defined _M_IX86 || defined _M_AMD64 || defined _M_IA64 |
|
|
573 | #define ECB_STDFP 1 |
|
|
574 | #include <string.h> /* for memcpy */ |
|
|
575 | #else |
|
|
576 | #define ECB_STDFP 0 |
|
|
577 | #include <math.h> /* for frexp*, ldexp* */ |
|
|
578 | #endif |
| 316 | |
579 | |
| 317 | ecb_inline uint64_t ecb_rotr64 (uint64_t x, unsigned int count) ecb_const; |
580 | #ifndef ECB_NO_LIBM |
| 318 | ecb_inline uint64_t |
|
|
| 319 | ecb_rotr64 (uint64_t x, unsigned int count) |
|
|
| 320 | { |
|
|
| 321 | return (x << (64 - count)) | (x >> count); |
|
|
| 322 | } |
|
|
| 323 | |
581 | |
| 324 | ecb_inline uint64_t ecb_rotl64 (uint64_t x, unsigned int count) ecb_const; |
582 | /* convert a float to ieee single/binary32 */ |
| 325 | ecb_inline uint64_t |
583 | ecb_function_ uint32_t ecb_float_to_binary32 (float x) ecb_const; |
| 326 | ecb_rotl64 (uint64_t x, unsigned int count) |
584 | ecb_function_ uint32_t |
| 327 | { |
585 | ecb_float_to_binary32 (float x) |
| 328 | return (x >> (64 - count)) | (x << count); |
586 | { |
| 329 | } |
587 | uint32_t r; |
| 330 | |
588 | |
|
|
589 | #if ECB_STDFP |
|
|
590 | memcpy (&r, &x, 4); |
|
|
591 | #else |
|
|
592 | /* slow emulation, works for anything but -0 */ |
|
|
593 | uint32_t m; |
|
|
594 | int e; |
|
|
595 | |
|
|
596 | if (x == 0e0f ) return 0x00000000U; |
|
|
597 | if (x > +3.40282346638528860e+38f) return 0x7f800000U; |
|
|
598 | if (x < -3.40282346638528860e+38f) return 0xff800000U; |
|
|
599 | if (x != x ) return 0x7fbfffffU; |
|
|
600 | |
|
|
601 | m = frexpf (x, &e) * 0x1000000U; |
|
|
602 | |
|
|
603 | r = m & 0x80000000U; |
|
|
604 | |
|
|
605 | if (r) |
|
|
606 | m = -m; |
|
|
607 | |
|
|
608 | if (e <= -126) |
|
|
609 | { |
|
|
610 | m &= 0xffffffU; |
|
|
611 | m >>= (-125 - e); |
|
|
612 | e = -126; |
|
|
613 | } |
|
|
614 | |
|
|
615 | r |= (e + 126) << 23; |
|
|
616 | r |= m & 0x7fffffU; |
| 331 | #endif |
617 | #endif |
| 332 | |
618 | |
|
|
619 | return r; |
|
|
620 | } |
|
|
621 | |
|
|
622 | /* converts an ieee single/binary32 to a float */ |
|
|
623 | ecb_function_ float ecb_binary32_to_float (uint32_t x) ecb_const; |
|
|
624 | ecb_function_ float |
|
|
625 | ecb_binary32_to_float (uint32_t x) |
|
|
626 | { |
|
|
627 | float r; |
|
|
628 | |
|
|
629 | #if ECB_STDFP |
|
|
630 | memcpy (&r, &x, 4); |
|
|
631 | #else |
|
|
632 | /* emulation, only works for normals and subnormals and +0 */ |
|
|
633 | int neg = x >> 31; |
|
|
634 | int e = (x >> 23) & 0xffU; |
|
|
635 | |
|
|
636 | x &= 0x7fffffU; |
|
|
637 | |
|
|
638 | if (e) |
|
|
639 | x |= 0x800000U; |
|
|
640 | else |
|
|
641 | e = 1; |
|
|
642 | |
|
|
643 | /* we distrust ldexpf a bit and do the 2**-24 scaling by an extra multiply */ |
|
|
644 | r = ldexpf (x * (0.5f / 0x800000U), e - 126); |
|
|
645 | |
|
|
646 | r = neg ? -r : r; |
|
|
647 | #endif |
|
|
648 | |
|
|
649 | return r; |
|
|
650 | } |
|
|
651 | |
|
|
652 | /* convert a double to ieee double/binary64 */ |
|
|
653 | ecb_function_ uint64_t ecb_double_to_binary64 (double x) ecb_const; |
|
|
654 | ecb_function_ uint64_t |
|
|
655 | ecb_double_to_binary64 (double x) |
|
|
656 | { |
|
|
657 | uint64_t r; |
|
|
658 | |
|
|
659 | #if ECB_STDFP |
|
|
660 | memcpy (&r, &x, 8); |
|
|
661 | #else |
|
|
662 | /* slow emulation, works for anything but -0 */ |
|
|
663 | uint64_t m; |
|
|
664 | int e; |
|
|
665 | |
|
|
666 | if (x == 0e0 ) return 0x0000000000000000U; |
|
|
667 | if (x > +1.79769313486231470e+308) return 0x7ff0000000000000U; |
|
|
668 | if (x < -1.79769313486231470e+308) return 0xfff0000000000000U; |
|
|
669 | if (x != x ) return 0X7ff7ffffffffffffU; |
|
|
670 | |
|
|
671 | m = frexp (x, &e) * 0x20000000000000U; |
|
|
672 | |
|
|
673 | r = m & 0x8000000000000000;; |
|
|
674 | |
|
|
675 | if (r) |
|
|
676 | m = -m; |
|
|
677 | |
|
|
678 | if (e <= -1022) |
|
|
679 | { |
|
|
680 | m &= 0x1fffffffffffffU; |
|
|
681 | m >>= (-1021 - e); |
|
|
682 | e = -1022; |
|
|
683 | } |
|
|
684 | |
|
|
685 | r |= ((uint64_t)(e + 1022)) << 52; |
|
|
686 | r |= m & 0xfffffffffffffU; |
|
|
687 | #endif |
|
|
688 | |
|
|
689 | return r; |
|
|
690 | } |
|
|
691 | |
|
|
692 | /* converts an ieee double/binary64 to a double */ |
|
|
693 | ecb_function_ double ecb_binary64_to_double (uint64_t x) ecb_const; |
|
|
694 | ecb_function_ double |
|
|
695 | ecb_binary64_to_double (uint64_t x) |
|
|
696 | { |
|
|
697 | double r; |
|
|
698 | |
|
|
699 | #if ECB_STDFP |
|
|
700 | memcpy (&r, &x, 8); |
|
|
701 | #else |
|
|
702 | /* emulation, only works for normals and subnormals and +0 */ |
|
|
703 | int neg = x >> 63; |
|
|
704 | int e = (x >> 52) & 0x7ffU; |
|
|
705 | |
|
|
706 | x &= 0xfffffffffffffU; |
|
|
707 | |
|
|
708 | if (e) |
|
|
709 | x |= 0x10000000000000U; |
|
|
710 | else |
|
|
711 | e = 1; |
|
|
712 | |
|
|
713 | /* we distrust ldexp a bit and do the 2**-53 scaling by an extra multiply */ |
|
|
714 | r = ldexp (x * (0.5 / 0x10000000000000U), e - 1022); |
|
|
715 | |
|
|
716 | r = neg ? -r : r; |
|
|
717 | #endif |
|
|
718 | |
|
|
719 | return r; |
|
|
720 | } |
|
|
721 | |
|
|
722 | #endif |
|
|
723 | |
|
|
724 | #endif |
|
|
725 | |
