| 1 | /* |
1 | /* |
| 2 | * libecb |
2 | * libecb - http://software.schmorp.de/pkg/libecb |
| 3 | * |
3 | * |
| 4 | * Copyright (©) 2009-2011 Marc Alexander Lehmann |
4 | * Copyright (©) 2009-2012 Marc Alexander Lehmann <libecb@schmorp.de> |
|
|
5 | * Copyright (©) 2011 Emanuele Giaquinta |
| 5 | * All rights reserved. |
6 | * All rights reserved. |
| 6 | * |
7 | * |
| 7 | * 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- |
| 8 | * tion, are permitted provided that the following conditions are met: |
9 | * tion, are permitted provided that the following conditions are met: |
| 9 | * |
10 | * |
| … | |
… | |
| 27 | */ |
28 | */ |
| 28 | |
29 | |
| 29 | #ifndef ECB_H |
30 | #ifndef ECB_H |
| 30 | #define ECB_H |
31 | #define ECB_H |
| 31 | |
32 | |
|
|
33 | /* 16 bits major, 16 bits minor */ |
|
|
34 | #define ECB_VERSION 0x00010002 |
|
|
35 | |
|
|
36 | #ifdef _WIN32 |
|
|
37 | typedef signed char int8_t; |
|
|
38 | typedef unsigned char uint8_t; |
|
|
39 | typedef signed short int16_t; |
|
|
40 | typedef unsigned short uint16_t; |
|
|
41 | typedef signed int int32_t; |
|
|
42 | typedef unsigned int uint32_t; |
|
|
43 | #if __GNUC__ |
|
|
44 | typedef signed long long int64_t; |
|
|
45 | typedef unsigned long long uint64_t; |
|
|
46 | #else /* _MSC_VER || __BORLANDC__ */ |
|
|
47 | typedef signed __int64 int64_t; |
|
|
48 | typedef unsigned __int64 uint64_t; |
|
|
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 | #else |
|
|
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 | /* many compilers define _GNUC_ to some versions but then only implement |
|
|
69 | * what their idiot authors think are the "more important" extensions, |
|
|
70 | * causing enormous grief in return for some better fake benchmark numbers. |
|
|
71 | * or so. |
|
|
72 | * we try to detect these and simply assume they are not gcc - if they have |
|
|
73 | * an issue with that they should have done it right in the first place. |
|
|
74 | */ |
|
|
75 | #ifndef ECB_GCC_VERSION |
|
|
76 | #if !defined __GNUC_MINOR__ || defined __INTEL_COMPILER || defined __SUNPRO_C || defined __SUNPRO_CC || defined __llvm__ || defined __clang__ |
|
|
77 | #define ECB_GCC_VERSION(major,minor) 0 |
|
|
78 | #else |
| 32 | #define ECB_GCC_VERSION(major,minor) (__GNUC__ > (major) || (__GNUC__ == (major) && __GNUC_MINOR__ >= (minor))) |
79 | #define ECB_GCC_VERSION(major,minor) (__GNUC__ > (major) || (__GNUC__ == (major) && __GNUC_MINOR__ >= (minor))) |
|
|
80 | #endif |
|
|
81 | #endif |
| 33 | |
82 | |
| 34 | #ifndef __cplusplus |
83 | #define ECB_C (__STDC__+0) /* this assumes that __STDC__ is either empty or a number */ |
| 35 | # if __STDC_VERSION__ >= 199901L |
84 | #define ECB_C99 (__STDC_VERSION__ >= 199901L) |
| 36 | # define ECB_INLINE inline |
85 | #define ECB_C11 (__STDC_VERSION__ >= 201112L) |
| 37 | typedef _Bool ecb_bool; |
86 | #define ECB_CPP (__cplusplus+0) |
|
|
87 | #define ECB_CPP11 (__cplusplus >= 201103L) |
|
|
88 | |
|
|
89 | #if ECB_CPP |
|
|
90 | #define ECB_EXTERN_C extern "C" |
|
|
91 | #define ECB_EXTERN_C_BEG ECB_EXTERN_C { |
|
|
92 | #define ECB_EXTERN_C_END } |
| 38 | # else |
93 | #else |
| 39 | # define ECB_INLINE inline /* yeah! */ |
94 | #define ECB_EXTERN_C extern |
|
|
95 | #define ECB_EXTERN_C_BEG |
|
|
96 | #define ECB_EXTERN_C_END |
|
|
97 | #endif |
|
|
98 | |
|
|
99 | /*****************************************************************************/ |
|
|
100 | |
|
|
101 | /* ECB_NO_THREADS - ecb is not used by multiple threads, ever */ |
|
|
102 | /* ECB_NO_SMP - ecb might be used in multiple threads, but only on a single cpu */ |
|
|
103 | |
|
|
104 | #if ECB_NO_THREADS |
|
|
105 | #define ECB_NO_SMP 1 |
|
|
106 | #endif |
|
|
107 | |
|
|
108 | #if ECB_NO_SMP |
|
|
109 | #define ECB_MEMORY_FENCE do { } while (0) |
|
|
110 | #endif |
|
|
111 | |
|
|
112 | #ifndef ECB_MEMORY_FENCE |
|
|
113 | #if ECB_GCC_VERSION(2,5) || defined __INTEL_COMPILER || (__llvm__ && __GNUC__) || __SUNPRO_C >= 0x5110 || __SUNPRO_CC >= 0x5110 |
|
|
114 | #if __i386 || __i386__ |
|
|
115 | #define ECB_MEMORY_FENCE __asm__ __volatile__ ("lock; orb $0, -1(%%esp)" : : : "memory") |
|
|
116 | #define ECB_MEMORY_FENCE_ACQUIRE __asm__ __volatile__ ("" : : : "memory") |
|
|
117 | #define ECB_MEMORY_FENCE_RELEASE __asm__ __volatile__ ("") |
|
|
118 | #elif __amd64 || __amd64__ || __x86_64 || __x86_64__ |
|
|
119 | #define ECB_MEMORY_FENCE __asm__ __volatile__ ("mfence" : : : "memory") |
|
|
120 | #define ECB_MEMORY_FENCE_ACQUIRE __asm__ __volatile__ ("" : : : "memory") |
|
|
121 | #define ECB_MEMORY_FENCE_RELEASE __asm__ __volatile__ ("") |
|
|
122 | #elif __powerpc__ || __ppc__ || __powerpc64__ || __ppc64__ |
|
|
123 | #define ECB_MEMORY_FENCE __asm__ __volatile__ ("sync" : : : "memory") |
|
|
124 | #elif defined __ARM_ARCH_6__ || defined __ARM_ARCH_6J__ \ |
|
|
125 | || defined __ARM_ARCH_6K__ || defined __ARM_ARCH_6ZK__ |
|
|
126 | #define ECB_MEMORY_FENCE __asm__ __volatile__ ("mcr p15,0,%0,c7,c10,5" : : "r" (0) : "memory") |
|
|
127 | #elif defined __ARM_ARCH_7__ || defined __ARM_ARCH_7A__ \ |
|
|
128 | || defined __ARM_ARCH_7M__ || defined __ARM_ARCH_7R__ |
|
|
129 | #define ECB_MEMORY_FENCE __asm__ __volatile__ ("dmb" : : : "memory") |
|
|
130 | #elif __sparc || __sparc__ |
|
|
131 | #define ECB_MEMORY_FENCE __asm__ __volatile__ ("membar #LoadStore | #LoadLoad | #StoreStore | #StoreLoad" : : : "memory") |
|
|
132 | #define ECB_MEMORY_FENCE_ACQUIRE __asm__ __volatile__ ("membar #LoadStore | #LoadLoad" : : : "memory") |
|
|
133 | #define ECB_MEMORY_FENCE_RELEASE __asm__ __volatile__ ("membar #LoadStore | #StoreStore") |
|
|
134 | #elif defined __s390__ || defined __s390x__ |
|
|
135 | #define ECB_MEMORY_FENCE __asm__ __volatile__ ("bcr 15,0" : : : "memory") |
|
|
136 | #elif defined __mips__ |
|
|
137 | #define ECB_MEMORY_FENCE __asm__ __volatile__ ("sync" : : : "memory") |
|
|
138 | #elif defined __alpha__ |
|
|
139 | #define ECB_MEMORY_FENCE __asm__ __volatile__ ("mb" : : : "memory") |
|
|
140 | #elif defined __hppa__ |
|
|
141 | #define ECB_MEMORY_FENCE __asm__ __volatile__ ("" : : : "memory") |
|
|
142 | #define ECB_MEMORY_FENCE_RELEASE __asm__ __volatile__ ("") |
|
|
143 | #elif defined __ia64__ |
|
|
144 | #define ECB_MEMORY_FENCE __asm__ __volatile__ ("mf" : : : "memory") |
|
|
145 | #endif |
|
|
146 | #endif |
|
|
147 | #endif |
|
|
148 | |
|
|
149 | #ifndef ECB_MEMORY_FENCE |
|
|
150 | #if ECB_GCC_VERSION(4,7) |
|
|
151 | /* see comment below (stdatomic.h) about the C11 memory model. */ |
|
|
152 | #define ECB_MEMORY_FENCE __atomic_thread_fence (__ATOMIC_SEQ_CST) |
|
|
153 | |
|
|
154 | /* The __has_feature syntax from clang is so misdesigned that we cannot use it |
|
|
155 | * without risking compile time errors with other compilers. We *could* |
|
|
156 | * define our own ecb_clang_has_feature, but I just can't be bothered to work |
|
|
157 | * around this shit time and again. |
|
|
158 | * #elif defined __clang && __has_feature (cxx_atomic) |
|
|
159 | * // see comment below (stdatomic.h) about the C11 memory model. |
|
|
160 | * #define ECB_MEMORY_FENCE __c11_atomic_thread_fence (__ATOMIC_SEQ_CST) |
|
|
161 | */ |
|
|
162 | |
|
|
163 | #elif ECB_GCC_VERSION(4,4) || defined __INTEL_COMPILER || defined __clang__ |
|
|
164 | #define ECB_MEMORY_FENCE __sync_synchronize () |
|
|
165 | #elif _MSC_VER >= 1400 /* VC++ 2005 */ |
|
|
166 | #pragma intrinsic(_ReadBarrier,_WriteBarrier,_ReadWriteBarrier) |
|
|
167 | #define ECB_MEMORY_FENCE _ReadWriteBarrier () |
|
|
168 | #define ECB_MEMORY_FENCE_ACQUIRE _ReadWriteBarrier () /* according to msdn, _ReadBarrier is not a load fence */ |
|
|
169 | #define ECB_MEMORY_FENCE_RELEASE _WriteBarrier () |
|
|
170 | #elif defined _WIN32 |
|
|
171 | #include <WinNT.h> |
|
|
172 | #define ECB_MEMORY_FENCE MemoryBarrier () /* actually just xchg on x86... scary */ |
|
|
173 | #elif __SUNPRO_C >= 0x5110 || __SUNPRO_CC >= 0x5110 |
|
|
174 | #include <mbarrier.h> |
|
|
175 | #define ECB_MEMORY_FENCE __machine_rw_barrier () |
|
|
176 | #define ECB_MEMORY_FENCE_ACQUIRE __machine_r_barrier () |
|
|
177 | #define ECB_MEMORY_FENCE_RELEASE __machine_w_barrier () |
|
|
178 | #elif __xlC__ |
|
|
179 | #define ECB_MEMORY_FENCE __sync () |
|
|
180 | #endif |
|
|
181 | #endif |
|
|
182 | |
|
|
183 | #ifndef ECB_MEMORY_FENCE |
|
|
184 | #if ECB_C11 && !defined __STDC_NO_ATOMICS__ |
|
|
185 | /* we assume that these memory fences work on all variables/all memory accesses, */ |
|
|
186 | /* not just C11 atomics and atomic accesses */ |
|
|
187 | #include <stdatomic.h> |
|
|
188 | /* Unfortunately, neither gcc 4.7 nor clang 3.1 generate any instructions for */ |
|
|
189 | /* any fence other than seq_cst, which isn't very efficient for us. */ |
|
|
190 | /* Why that is, we don't know - either the C11 memory model is quite useless */ |
|
|
191 | /* for most usages, or gcc and clang have a bug */ |
|
|
192 | /* I *currently* lean towards the latter, and inefficiently implement */ |
|
|
193 | /* all three of ecb's fences as a seq_cst fence */ |
|
|
194 | #define ECB_MEMORY_FENCE atomic_thread_fence (memory_order_seq_cst) |
|
|
195 | #endif |
|
|
196 | #endif |
|
|
197 | |
|
|
198 | #ifndef ECB_MEMORY_FENCE |
|
|
199 | #if !ECB_AVOID_PTHREADS |
|
|
200 | /* |
|
|
201 | * if you get undefined symbol references to pthread_mutex_lock, |
|
|
202 | * or failure to find pthread.h, then you should implement |
|
|
203 | * the ECB_MEMORY_FENCE operations for your cpu/compiler |
|
|
204 | * OR provide pthread.h and link against the posix thread library |
|
|
205 | * of your system. |
|
|
206 | */ |
|
|
207 | #include <pthread.h> |
|
|
208 | #define ECB_NEEDS_PTHREADS 1 |
|
|
209 | #define ECB_MEMORY_FENCE_NEEDS_PTHREADS 1 |
|
|
210 | |
|
|
211 | static pthread_mutex_t ecb_mf_lock = PTHREAD_MUTEX_INITIALIZER; |
|
|
212 | #define ECB_MEMORY_FENCE do { pthread_mutex_lock (&ecb_mf_lock); pthread_mutex_unlock (&ecb_mf_lock); } while (0) |
|
|
213 | #endif |
|
|
214 | #endif |
|
|
215 | |
|
|
216 | #if !defined ECB_MEMORY_FENCE_ACQUIRE && defined ECB_MEMORY_FENCE |
|
|
217 | #define ECB_MEMORY_FENCE_ACQUIRE ECB_MEMORY_FENCE |
|
|
218 | #endif |
|
|
219 | |
|
|
220 | #if !defined ECB_MEMORY_FENCE_RELEASE && defined ECB_MEMORY_FENCE |
|
|
221 | #define ECB_MEMORY_FENCE_RELEASE ECB_MEMORY_FENCE |
|
|
222 | #endif |
|
|
223 | |
|
|
224 | /*****************************************************************************/ |
|
|
225 | |
|
|
226 | #if __cplusplus |
|
|
227 | #define ecb_inline static inline |
|
|
228 | #elif ECB_GCC_VERSION(2,5) |
|
|
229 | #define ecb_inline static __inline__ |
|
|
230 | #elif ECB_C99 |
|
|
231 | #define ecb_inline static inline |
|
|
232 | #else |
|
|
233 | #define ecb_inline static |
|
|
234 | #endif |
|
|
235 | |
|
|
236 | #if ECB_GCC_VERSION(3,3) |
|
|
237 | #define ecb_restrict __restrict__ |
|
|
238 | #elif ECB_C99 |
|
|
239 | #define ecb_restrict restrict |
|
|
240 | #else |
|
|
241 | #define ecb_restrict |
|
|
242 | #endif |
|
|
243 | |
| 40 | typedef int ecb_bool; |
244 | typedef int ecb_bool; |
| 41 | # endif |
|
|
| 42 | #endif |
|
|
| 43 | |
245 | |
| 44 | #define ECB_HEADER_INLINE static ECB_INLINE |
246 | #define ECB_CONCAT_(a, b) a ## b |
|
|
247 | #define ECB_CONCAT(a, b) ECB_CONCAT_(a, b) |
|
|
248 | #define ECB_STRINGIFY_(a) # a |
|
|
249 | #define ECB_STRINGIFY(a) ECB_STRINGIFY_(a) |
|
|
250 | |
|
|
251 | #define ecb_function_ ecb_inline |
| 45 | |
252 | |
| 46 | #if ECB_GCC_VERSION(3,1) |
253 | #if ECB_GCC_VERSION(3,1) |
| 47 | # define ecb_attribute(attrlist) __attribute__(attrlist) |
254 | #define ecb_attribute(attrlist) __attribute__(attrlist) |
| 48 | # define ecb_is_constant(c) __builtin_constant_p (c) |
255 | #define ecb_is_constant(expr) __builtin_constant_p (expr) |
| 49 | # define ecb_expect(expr,value) __builtin_expect ((expr),(value)) |
256 | #define ecb_expect(expr,value) __builtin_expect ((expr),(value)) |
| 50 | # define ecb_prefetch(addr,rw,locality) __builtin_prefetch (addr, rw, locality) |
257 | #define ecb_prefetch(addr,rw,locality) __builtin_prefetch (addr, rw, locality) |
| 51 | #else |
258 | #else |
| 52 | # define ecb_attribute(attrlist) |
259 | #define ecb_attribute(attrlist) |
| 53 | # define ecb_is_constant(c) 0 |
260 | #define ecb_is_constant(expr) 0 |
| 54 | # define ecb_expect(expr,value) (expr) |
261 | #define ecb_expect(expr,value) (expr) |
| 55 | # define ecb_prefetch(addr,rw,locality) |
262 | #define ecb_prefetch(addr,rw,locality) |
| 56 | #endif |
263 | #endif |
| 57 | |
264 | |
| 58 | /* no emulation for ecb_decltype */ |
265 | /* no emulation for ecb_decltype */ |
| 59 | #if ECB_GCC_VERSION(4,5) |
266 | #if ECB_GCC_VERSION(4,5) |
| 60 | # define ecb_decltype(x) __decltype(x) |
267 | #define ecb_decltype(x) __decltype(x) |
| 61 | #elif ECB_GCC_VERSION(3,0) |
268 | #elif ECB_GCC_VERSION(3,0) |
| 62 | # define ecb_decltype(x) typeof(x) |
269 | #define ecb_decltype(x) __typeof(x) |
| 63 | #endif |
270 | #endif |
| 64 | |
271 | |
| 65 | #define ecb_noinline ecb_attribute ((noinline)) |
272 | #define ecb_noinline ecb_attribute ((__noinline__)) |
| 66 | #define ecb_noreturn ecb_attribute ((noreturn)) |
|
|
| 67 | #define ecb_unused ecb_attribute ((unused)) |
273 | #define ecb_unused ecb_attribute ((__unused__)) |
| 68 | #define ecb_const ecb_attribute ((const)) |
274 | #define ecb_const ecb_attribute ((__const__)) |
| 69 | #define ecb_pure ecb_attribute ((pure)) |
275 | #define ecb_pure ecb_attribute ((__pure__)) |
| 70 | #define ecb_hot ecb_attribute ((hot)) /* 4.3 */ |
|
|
| 71 | #define ecb_cold ecb_attribute ((cold)) /* 4.3 */ |
|
|
| 72 | |
276 | |
|
|
277 | #if ECB_C11 |
|
|
278 | #define ecb_noreturn _Noreturn |
|
|
279 | #else |
|
|
280 | #define ecb_noreturn ecb_attribute ((__noreturn__)) |
|
|
281 | #endif |
|
|
282 | |
| 73 | #if ECB_GCC_VERSION(4,5) |
283 | #if ECB_GCC_VERSION(4,3) |
| 74 | # define ecb_unreachable() __builtin_unreachable () |
284 | #define ecb_artificial ecb_attribute ((__artificial__)) |
|
|
285 | #define ecb_hot ecb_attribute ((__hot__)) |
|
|
286 | #define ecb_cold ecb_attribute ((__cold__)) |
| 75 | #else |
287 | #else |
| 76 | /* this seems to work fine, but gcc always emits a warning for it :/ */ |
288 | #define ecb_artificial |
| 77 | ECB_HEADER_INLINE void ecb_unreachable () ecb_attribute ((noreturn)); |
289 | #define ecb_hot |
| 78 | ECB_HEADER_INLINE void ecb_unreachable () { } |
290 | #define ecb_cold |
| 79 | #endif |
291 | #endif |
| 80 | |
292 | |
| 81 | /* put into ifs if you are very sure that the expression */ |
293 | /* put around conditional expressions if you are very sure that the */ |
| 82 | /* is mostly true or mosty false. note that these return */ |
294 | /* expression is mostly true or mostly false. note that these return */ |
| 83 | /* booleans, not the expression. */ |
295 | /* booleans, not the expression. */ |
| 84 | #define ecb_expect_false(expr) ecb_expect ((expr) ? 1 : 0, 0) |
296 | #define ecb_expect_false(expr) ecb_expect (!!(expr), 0) |
| 85 | #define ecb_expect_true(expr) ecb_expect ((expr) ? 1 : 0, 1) |
297 | #define ecb_expect_true(expr) ecb_expect (!!(expr), 1) |
| 86 | |
298 | /* for compatibility to the rest of the world */ |
| 87 | /* try to tell the compiler that some condition is definitely true */ |
299 | #define ecb_likely(expr) ecb_expect_true (expr) |
| 88 | #define ecb_assume(cond) do { if (!(cond)) unreachable (); } while (0) |
300 | #define ecb_unlikely(expr) ecb_expect_false (expr) |
| 89 | |
301 | |
| 90 | /* count trailing zero bits and count # of one bits */ |
302 | /* count trailing zero bits and count # of one bits */ |
| 91 | #if ECB_GCC_VERSION(3,4) |
303 | #if ECB_GCC_VERSION(3,4) |
| 92 | ECB_HEADER_INLINE int ecb_ctz32 (unsigned int x) { return __builtin_ctz (x); } |
304 | /* we assume int == 32 bit, long == 32 or 64 bit and long long == 64 bit */ |
| 93 | ECB_HEADER_INLINE int ecb_popcount32 (unsigned int x) { return __builtin_popcount (x); } |
305 | #define ecb_ld32(x) (__builtin_clz (x) ^ 31) |
|
|
306 | #define ecb_ld64(x) (__builtin_clzll (x) ^ 63) |
|
|
307 | #define ecb_ctz32(x) __builtin_ctz (x) |
|
|
308 | #define ecb_ctz64(x) __builtin_ctzll (x) |
|
|
309 | #define ecb_popcount32(x) __builtin_popcount (x) |
|
|
310 | /* no popcountll */ |
| 94 | #else |
311 | #else |
| 95 | ECB_HEADER_INLINE |
312 | ecb_function_ int ecb_ctz32 (uint32_t x) ecb_const; |
| 96 | ecb_ctz32 (unsigned int x) ecb_const |
313 | ecb_function_ int |
|
|
314 | ecb_ctz32 (uint32_t x) |
|
|
315 | { |
|
|
316 | int r = 0; |
|
|
317 | |
|
|
318 | x &= ~x + 1; /* this isolates the lowest bit */ |
|
|
319 | |
|
|
320 | #if ECB_branchless_on_i386 |
|
|
321 | r += !!(x & 0xaaaaaaaa) << 0; |
|
|
322 | r += !!(x & 0xcccccccc) << 1; |
|
|
323 | r += !!(x & 0xf0f0f0f0) << 2; |
|
|
324 | r += !!(x & 0xff00ff00) << 3; |
|
|
325 | r += !!(x & 0xffff0000) << 4; |
|
|
326 | #else |
|
|
327 | if (x & 0xaaaaaaaa) r += 1; |
|
|
328 | if (x & 0xcccccccc) r += 2; |
|
|
329 | if (x & 0xf0f0f0f0) r += 4; |
|
|
330 | if (x & 0xff00ff00) r += 8; |
|
|
331 | if (x & 0xffff0000) r += 16; |
|
|
332 | #endif |
|
|
333 | |
|
|
334 | return r; |
|
|
335 | } |
|
|
336 | |
|
|
337 | ecb_function_ int ecb_ctz64 (uint64_t x) ecb_const; |
|
|
338 | ecb_function_ int |
|
|
339 | ecb_ctz64 (uint64_t x) |
|
|
340 | { |
|
|
341 | int shift = x & 0xffffffffU ? 0 : 32; |
|
|
342 | return ecb_ctz32 (x >> shift) + shift; |
|
|
343 | } |
|
|
344 | |
|
|
345 | ecb_function_ int ecb_popcount32 (uint32_t x) ecb_const; |
|
|
346 | ecb_function_ int |
|
|
347 | ecb_popcount32 (uint32_t x) |
|
|
348 | { |
|
|
349 | x -= (x >> 1) & 0x55555555; |
|
|
350 | x = ((x >> 2) & 0x33333333) + (x & 0x33333333); |
|
|
351 | x = ((x >> 4) + x) & 0x0f0f0f0f; |
|
|
352 | x *= 0x01010101; |
|
|
353 | |
|
|
354 | return x >> 24; |
|
|
355 | } |
|
|
356 | |
|
|
357 | ecb_function_ int ecb_ld32 (uint32_t x) ecb_const; |
|
|
358 | ecb_function_ int ecb_ld32 (uint32_t x) |
|
|
359 | { |
|
|
360 | int r = 0; |
|
|
361 | |
|
|
362 | if (x >> 16) { x >>= 16; r += 16; } |
|
|
363 | if (x >> 8) { x >>= 8; r += 8; } |
|
|
364 | if (x >> 4) { x >>= 4; r += 4; } |
|
|
365 | if (x >> 2) { x >>= 2; r += 2; } |
|
|
366 | if (x >> 1) { r += 1; } |
|
|
367 | |
|
|
368 | return r; |
|
|
369 | } |
|
|
370 | |
|
|
371 | ecb_function_ int ecb_ld64 (uint64_t x) ecb_const; |
|
|
372 | ecb_function_ int ecb_ld64 (uint64_t x) |
|
|
373 | { |
|
|
374 | int r = 0; |
|
|
375 | |
|
|
376 | if (x >> 32) { x >>= 32; r += 32; } |
|
|
377 | |
|
|
378 | return r + ecb_ld32 (x); |
|
|
379 | } |
|
|
380 | #endif |
|
|
381 | |
|
|
382 | ecb_function_ ecb_bool ecb_is_pot32 (uint32_t x) ecb_const; |
|
|
383 | ecb_function_ ecb_bool ecb_is_pot32 (uint32_t x) { return !(x & (x - 1)); } |
|
|
384 | ecb_function_ ecb_bool ecb_is_pot64 (uint64_t x) ecb_const; |
|
|
385 | ecb_function_ ecb_bool ecb_is_pot64 (uint64_t x) { return !(x & (x - 1)); } |
|
|
386 | |
|
|
387 | ecb_function_ uint8_t ecb_bitrev8 (uint8_t x) ecb_const; |
|
|
388 | ecb_function_ uint8_t ecb_bitrev8 (uint8_t x) |
| 97 | { |
389 | { |
| 98 | int r = 0; |
390 | return ( (x * 0x0802U & 0x22110U) |
| 99 | |
391 | | (x * 0x8020U & 0x88440U)) * 0x10101U >> 16; |
| 100 | x &= -x; /* this isolates the lowest bit */ |
|
|
| 101 | |
|
|
| 102 | if (x & 0xaaaaaaaa) r += 1; |
|
|
| 103 | if (x & 0xcccccccc) r += 2; |
|
|
| 104 | if (x & 0xf0f0f0f0) r += 4; |
|
|
| 105 | if (x & 0xff00ff00) r += 8; |
|
|
| 106 | if (x & 0xffff0000) r += 16; |
|
|
| 107 | |
|
|
| 108 | return r; |
|
|
| 109 | } |
392 | } |
| 110 | |
393 | |
| 111 | ECB_HEADER_INLINE |
394 | ecb_function_ uint16_t ecb_bitrev16 (uint16_t x) ecb_const; |
| 112 | ecb_popcount32 (unsigned int x) ecb_const |
395 | ecb_function_ uint16_t ecb_bitrev16 (uint16_t x) |
| 113 | { |
396 | { |
| 114 | x -= (x >> 1) & 0x55555555; |
397 | x = ((x >> 1) & 0x5555) | ((x & 0x5555) << 1); |
| 115 | x = ((x >> 2) & 0x33333333) + (x & 0x33333333); |
398 | x = ((x >> 2) & 0x3333) | ((x & 0x3333) << 2); |
| 116 | x = ((x >> 4) + x) & 0x0f0f0f0f; |
399 | x = ((x >> 4) & 0x0f0f) | ((x & 0x0f0f) << 4); |
| 117 | x *= 0x01010101; |
400 | x = ( x >> 8 ) | ( x << 8); |
| 118 | |
401 | |
| 119 | return x >> 24; |
402 | return x; |
| 120 | } |
403 | } |
| 121 | #endif |
|
|
| 122 | |
404 | |
| 123 | ECB_HEADER_INLINE unsigned char |
405 | ecb_function_ uint32_t ecb_bitrev32 (uint32_t x) ecb_const; |
| 124 | ecb_byteorder_helper () ecb_const |
406 | ecb_function_ uint32_t ecb_bitrev32 (uint32_t x) |
| 125 | { |
407 | { |
| 126 | const uint32_t u = 0x11223344; |
408 | x = ((x >> 1) & 0x55555555) | ((x & 0x55555555) << 1); |
| 127 | return *(unsigned char *)&u; |
409 | x = ((x >> 2) & 0x33333333) | ((x & 0x33333333) << 2); |
|
|
410 | x = ((x >> 4) & 0x0f0f0f0f) | ((x & 0x0f0f0f0f) << 4); |
|
|
411 | x = ((x >> 8) & 0x00ff00ff) | ((x & 0x00ff00ff) << 8); |
|
|
412 | x = ( x >> 16 ) | ( x << 16); |
|
|
413 | |
|
|
414 | return x; |
| 128 | } |
415 | } |
| 129 | |
416 | |
|
|
417 | /* popcount64 is only available on 64 bit cpus as gcc builtin */ |
|
|
418 | /* so for this version we are lazy */ |
|
|
419 | ecb_function_ int ecb_popcount64 (uint64_t x) ecb_const; |
|
|
420 | ecb_function_ int |
|
|
421 | ecb_popcount64 (uint64_t x) |
|
|
422 | { |
|
|
423 | return ecb_popcount32 (x) + ecb_popcount32 (x >> 32); |
|
|
424 | } |
|
|
425 | |
|
|
426 | ecb_inline uint8_t ecb_rotl8 (uint8_t x, unsigned int count) ecb_const; |
|
|
427 | ecb_inline uint8_t ecb_rotr8 (uint8_t x, unsigned int count) ecb_const; |
|
|
428 | ecb_inline uint16_t ecb_rotl16 (uint16_t x, unsigned int count) ecb_const; |
|
|
429 | ecb_inline uint16_t ecb_rotr16 (uint16_t x, unsigned int count) ecb_const; |
|
|
430 | ecb_inline uint32_t ecb_rotl32 (uint32_t x, unsigned int count) ecb_const; |
|
|
431 | ecb_inline uint32_t ecb_rotr32 (uint32_t x, unsigned int count) ecb_const; |
|
|
432 | ecb_inline uint64_t ecb_rotl64 (uint64_t x, unsigned int count) ecb_const; |
|
|
433 | ecb_inline uint64_t ecb_rotr64 (uint64_t x, unsigned int count) ecb_const; |
|
|
434 | |
|
|
435 | ecb_inline uint8_t ecb_rotl8 (uint8_t x, unsigned int count) { return (x >> ( 8 - count)) | (x << count); } |
|
|
436 | ecb_inline uint8_t ecb_rotr8 (uint8_t x, unsigned int count) { return (x << ( 8 - count)) | (x >> count); } |
|
|
437 | ecb_inline uint16_t ecb_rotl16 (uint16_t x, unsigned int count) { return (x >> (16 - count)) | (x << count); } |
|
|
438 | ecb_inline uint16_t ecb_rotr16 (uint16_t x, unsigned int count) { return (x << (16 - count)) | (x >> count); } |
|
|
439 | ecb_inline uint32_t ecb_rotl32 (uint32_t x, unsigned int count) { return (x >> (32 - count)) | (x << count); } |
|
|
440 | ecb_inline uint32_t ecb_rotr32 (uint32_t x, unsigned int count) { return (x << (32 - count)) | (x >> count); } |
|
|
441 | ecb_inline uint64_t ecb_rotl64 (uint64_t x, unsigned int count) { return (x >> (64 - count)) | (x << count); } |
|
|
442 | ecb_inline uint64_t ecb_rotr64 (uint64_t x, unsigned int count) { return (x << (64 - count)) | (x >> count); } |
|
|
443 | |
|
|
444 | #if ECB_GCC_VERSION(4,3) |
|
|
445 | #define ecb_bswap16(x) (__builtin_bswap32 (x) >> 16) |
|
|
446 | #define ecb_bswap32(x) __builtin_bswap32 (x) |
|
|
447 | #define ecb_bswap64(x) __builtin_bswap64 (x) |
|
|
448 | #else |
|
|
449 | ecb_function_ uint16_t ecb_bswap16 (uint16_t x) ecb_const; |
|
|
450 | ecb_function_ uint16_t |
|
|
451 | ecb_bswap16 (uint16_t x) |
|
|
452 | { |
|
|
453 | return ecb_rotl16 (x, 8); |
|
|
454 | } |
|
|
455 | |
|
|
456 | ecb_function_ uint32_t ecb_bswap32 (uint32_t x) ecb_const; |
|
|
457 | ecb_function_ uint32_t |
|
|
458 | ecb_bswap32 (uint32_t x) |
|
|
459 | { |
|
|
460 | return (((uint32_t)ecb_bswap16 (x)) << 16) | ecb_bswap16 (x >> 16); |
|
|
461 | } |
|
|
462 | |
|
|
463 | ecb_function_ uint64_t ecb_bswap64 (uint64_t x) ecb_const; |
|
|
464 | ecb_function_ uint64_t |
|
|
465 | ecb_bswap64 (uint64_t x) |
|
|
466 | { |
|
|
467 | return (((uint64_t)ecb_bswap32 (x)) << 32) | ecb_bswap32 (x >> 32); |
|
|
468 | } |
|
|
469 | #endif |
|
|
470 | |
|
|
471 | #if ECB_GCC_VERSION(4,5) |
|
|
472 | #define ecb_unreachable() __builtin_unreachable () |
|
|
473 | #else |
|
|
474 | /* this seems to work fine, but gcc always emits a warning for it :/ */ |
|
|
475 | ecb_inline void ecb_unreachable (void) ecb_noreturn; |
|
|
476 | ecb_inline void ecb_unreachable (void) { } |
|
|
477 | #endif |
|
|
478 | |
|
|
479 | /* try to tell the compiler that some condition is definitely true */ |
|
|
480 | #define ecb_assume(cond) if (!(cond)) ecb_unreachable (); else 0 |
|
|
481 | |
|
|
482 | ecb_inline unsigned char ecb_byteorder_helper (void) ecb_const; |
|
|
483 | ecb_inline unsigned char |
|
|
484 | ecb_byteorder_helper (void) |
|
|
485 | { |
|
|
486 | /* the union code still generates code under pressure in gcc, */ |
|
|
487 | /* but less than using pointers, and always seems to */ |
|
|
488 | /* successfully return a constant. */ |
|
|
489 | /* the reason why we have this horrible preprocessor mess */ |
|
|
490 | /* is to avoid it in all cases, at least on common architectures */ |
|
|
491 | /* or when using a recent enough gcc version (>= 4.6) */ |
|
|
492 | #if __i386 || __i386__ || _M_X86 || __amd64 || __amd64__ || _M_X64 |
|
|
493 | return 0x44; |
|
|
494 | #elif __BYTE_ORDER__ && __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__ |
|
|
495 | return 0x44; |
|
|
496 | #elif __BYTE_ORDER__ && __BYTE_ORDER__ == __ORDER_BIG_ENDIAN__ |
|
|
497 | return 0x11; |
|
|
498 | #else |
|
|
499 | union |
|
|
500 | { |
|
|
501 | uint32_t i; |
|
|
502 | uint8_t c; |
|
|
503 | } u = { 0x11223344 }; |
|
|
504 | return u.c; |
|
|
505 | #endif |
|
|
506 | } |
|
|
507 | |
|
|
508 | ecb_inline ecb_bool ecb_big_endian (void) ecb_const; |
| 130 | ECB_HEADER_INLINE ecb_bool ecb_big_endian () ecb_const { return ecb_byteorder_helper () == 0x11; }; |
509 | ecb_inline ecb_bool ecb_big_endian (void) { return ecb_byteorder_helper () == 0x11; } |
|
|
510 | ecb_inline ecb_bool ecb_little_endian (void) ecb_const; |
| 131 | ECB_HEADER_INLINE ecb_bool ecb_little_endian () ecb_const { return ecb_byteorder_helper () == 0x44; }; |
511 | ecb_inline ecb_bool ecb_little_endian (void) { return ecb_byteorder_helper () == 0x44; } |
| 132 | |
512 | |
|
|
513 | #if ECB_GCC_VERSION(3,0) || ECB_C99 |
|
|
514 | #define ecb_mod(m,n) ((m) % (n) + ((m) % (n) < 0 ? (n) : 0)) |
|
|
515 | #else |
|
|
516 | #define ecb_mod(m,n) ((m) < 0 ? ((n) - 1 - ((-1 - (m)) % (n))) : ((m) % (n))) |
|
|
517 | #endif |
|
|
518 | |
|
|
519 | #if __cplusplus |
|
|
520 | template<typename T> |
|
|
521 | static inline T ecb_div_rd (T val, T div) |
|
|
522 | { |
|
|
523 | return val < 0 ? - ((-val + div - 1) / div) : (val ) / div; |
|
|
524 | } |
|
|
525 | template<typename T> |
|
|
526 | static inline T ecb_div_ru (T val, T div) |
|
|
527 | { |
|
|
528 | return val < 0 ? - ((-val ) / div) : (val + div - 1) / div; |
|
|
529 | } |
|
|
530 | #else |
|
|
531 | #define ecb_div_rd(val,div) ((val) < 0 ? - ((-(val) + (div) - 1) / (div)) : ((val) ) / (div)) |
|
|
532 | #define ecb_div_ru(val,div) ((val) < 0 ? - ((-(val) ) / (div)) : ((val) + (div) - 1) / (div)) |
|
|
533 | #endif |
|
|
534 | |
|
|
535 | #if ecb_cplusplus_does_not_suck |
|
|
536 | /* does not work for local types (http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2008/n2657.htm) */ |
|
|
537 | template<typename T, int N> |
|
|
538 | static inline int ecb_array_length (const T (&arr)[N]) |
|
|
539 | { |
|
|
540 | return N; |
|
|
541 | } |
|
|
542 | #else |
|
|
543 | #define ecb_array_length(name) (sizeof (name) / sizeof (name [0])) |
|
|
544 | #endif |
|
|
545 | |
|
|
546 | /*******************************************************************************/ |
|
|
547 | /* floating point stuff, can be disabled by defining ECB_NO_LIBM */ |
|
|
548 | |
|
|
549 | /* basically, everything uses "ieee pure-endian" floating point numbers */ |
|
|
550 | /* the only noteworthy exception is ancient armle, which uses order 43218765 */ |
|
|
551 | #if 0 \ |
|
|
552 | || __i386 || __i386__ \ |
|
|
553 | || __amd64 || __amd64__ || __x86_64 || __x86_64__ \ |
|
|
554 | || __powerpc__ || __ppc__ || __powerpc64__ || __ppc64__ \ |
|
|
555 | || defined __arm__ && defined __ARM_EABI__ \ |
|
|
556 | || defined __s390__ || defined __s390x__ \ |
|
|
557 | || defined __mips__ \ |
|
|
558 | || defined __alpha__ \ |
|
|
559 | || defined __hppa__ \ |
|
|
560 | || defined __ia64__ \ |
|
|
561 | || defined _M_IX86 || defined _M_AMD64 || defined _M_IA64 |
|
|
562 | #define ECB_STDFP 1 |
|
|
563 | #include <string.h> /* for memcpy */ |
|
|
564 | #else |
|
|
565 | #define ECB_STDFP 0 |
|
|
566 | #include <math.h> /* for frexp*, ldexp* */ |
|
|
567 | #endif |
|
|
568 | |
|
|
569 | #ifndef ECB_NO_LIBM |
|
|
570 | |
|
|
571 | /* convert a float to ieee single/binary32 */ |
|
|
572 | ecb_function_ uint32_t ecb_float_to_binary32 (float x) ecb_const; |
|
|
573 | ecb_function_ uint32_t |
|
|
574 | ecb_float_to_binary32 (float x) |
|
|
575 | { |
|
|
576 | uint32_t r; |
|
|
577 | |
|
|
578 | #if ECB_STDFP |
|
|
579 | memcpy (&r, &x, 4); |
|
|
580 | #else |
|
|
581 | /* slow emulation, works for anything but -0 */ |
|
|
582 | uint32_t m; |
|
|
583 | int e; |
|
|
584 | |
|
|
585 | if (x == 0e0f ) return 0x00000000U; |
|
|
586 | if (x > +3.40282346638528860e+38f) return 0x7f800000U; |
|
|
587 | if (x < -3.40282346638528860e+38f) return 0xff800000U; |
|
|
588 | if (x != x ) return 0x7fbfffffU; |
|
|
589 | |
|
|
590 | m = frexpf (x, &e) * 0x1000000U; |
|
|
591 | |
|
|
592 | r = m & 0x80000000U; |
|
|
593 | |
|
|
594 | if (r) |
|
|
595 | m = -m; |
|
|
596 | |
|
|
597 | if (e <= -126) |
|
|
598 | { |
|
|
599 | m &= 0xffffffU; |
|
|
600 | m >>= (-125 - e); |
|
|
601 | e = -126; |
|
|
602 | } |
|
|
603 | |
|
|
604 | r |= (e + 126) << 23; |
|
|
605 | r |= m & 0x7fffffU; |
| 133 | #endif |
606 | #endif |
| 134 | |
607 | |
|
|
608 | return r; |
|
|
609 | } |
|
|
610 | |
|
|
611 | /* converts an ieee single/binary32 to a float */ |
|
|
612 | ecb_function_ float ecb_binary32_to_float (uint32_t x) ecb_const; |
|
|
613 | ecb_function_ float |
|
|
614 | ecb_binary32_to_float (uint32_t x) |
|
|
615 | { |
|
|
616 | float r; |
|
|
617 | |
|
|
618 | #if ECB_STDFP |
|
|
619 | memcpy (&r, &x, 4); |
|
|
620 | #else |
|
|
621 | /* emulation, only works for normals and subnormals and +0 */ |
|
|
622 | int neg = x >> 31; |
|
|
623 | int e = (x >> 23) & 0xffU; |
|
|
624 | |
|
|
625 | x &= 0x7fffffU; |
|
|
626 | |
|
|
627 | if (e) |
|
|
628 | x |= 0x800000U; |
|
|
629 | else |
|
|
630 | e = 1; |
|
|
631 | |
|
|
632 | /* we distrust ldexpf a bit and do the 2**-24 scaling by an extra multiply */ |
|
|
633 | r = ldexpf (x * (0.5f / 0x800000U), e - 126); |
|
|
634 | |
|
|
635 | r = neg ? -r : r; |
|
|
636 | #endif |
|
|
637 | |
|
|
638 | return r; |
|
|
639 | } |
|
|
640 | |
|
|
641 | /* convert a double to ieee double/binary64 */ |
|
|
642 | ecb_function_ uint64_t ecb_double_to_binary64 (double x) ecb_const; |
|
|
643 | ecb_function_ uint64_t |
|
|
644 | ecb_double_to_binary64 (double x) |
|
|
645 | { |
|
|
646 | uint64_t r; |
|
|
647 | |
|
|
648 | #if ECB_STDFP |
|
|
649 | memcpy (&r, &x, 8); |
|
|
650 | #else |
|
|
651 | /* slow emulation, works for anything but -0 */ |
|
|
652 | uint64_t m; |
|
|
653 | int e; |
|
|
654 | |
|
|
655 | if (x == 0e0 ) return 0x0000000000000000U; |
|
|
656 | if (x > +1.79769313486231470e+308) return 0x7ff0000000000000U; |
|
|
657 | if (x < -1.79769313486231470e+308) return 0xfff0000000000000U; |
|
|
658 | if (x != x ) return 0X7ff7ffffffffffffU; |
|
|
659 | |
|
|
660 | m = frexp (x, &e) * 0x20000000000000U; |
|
|
661 | |
|
|
662 | r = m & 0x8000000000000000;; |
|
|
663 | |
|
|
664 | if (r) |
|
|
665 | m = -m; |
|
|
666 | |
|
|
667 | if (e <= -1022) |
|
|
668 | { |
|
|
669 | m &= 0x1fffffffffffffU; |
|
|
670 | m >>= (-1021 - e); |
|
|
671 | e = -1022; |
|
|
672 | } |
|
|
673 | |
|
|
674 | r |= ((uint64_t)(e + 1022)) << 52; |
|
|
675 | r |= m & 0xfffffffffffffU; |
|
|
676 | #endif |
|
|
677 | |
|
|
678 | return r; |
|
|
679 | } |
|
|
680 | |
|
|
681 | /* converts an ieee double/binary64 to a double */ |
|
|
682 | ecb_function_ double ecb_binary64_to_double (uint64_t x) ecb_const; |
|
|
683 | ecb_function_ double |
|
|
684 | ecb_binary64_to_double (uint64_t x) |
|
|
685 | { |
|
|
686 | double r; |
|
|
687 | |
|
|
688 | #if ECB_STDFP |
|
|
689 | memcpy (&r, &x, 8); |
|
|
690 | #else |
|
|
691 | /* emulation, only works for normals and subnormals and +0 */ |
|
|
692 | int neg = x >> 63; |
|
|
693 | int e = (x >> 52) & 0x7ffU; |
|
|
694 | |
|
|
695 | x &= 0xfffffffffffffU; |
|
|
696 | |
|
|
697 | if (e) |
|
|
698 | x |= 0x10000000000000U; |
|
|
699 | else |
|
|
700 | e = 1; |
|
|
701 | |
|
|
702 | /* we distrust ldexp a bit and do the 2**-53 scaling by an extra multiply */ |
|
|
703 | r = ldexp (x * (0.5 / 0x10000000000000U), e - 1022); |
|
|
704 | |
|
|
705 | r = neg ? -r : r; |
|
|
706 | #endif |
|
|
707 | |
|
|
708 | return r; |
|
|
709 | } |
|
|
710 | |
|
|
711 | #endif |
|
|
712 | |
|
|
713 | #endif |
|
|
714 | |
