| 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 (©) 2009-2013 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; |
| … | |
… | |
| 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 |
| 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 enormous 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 */ |
|
|
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 | |
| 66 | /*****************************************************************************/ |
108 | /*****************************************************************************/ |
| 67 | |
109 | |
| 68 | /* ECB_NO_THREADS - ecb is not used by multiple threads, ever */ |
110 | /* 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 */ |
111 | /* ECB_NO_SMP - ecb might be used in multiple threads, but only on a single cpu */ |
| 70 | |
112 | |
| 71 | #if ECB_NO_THREADS || ECB_NO_SMP |
113 | #if ECB_NO_THREADS |
|
|
114 | #define ECB_NO_SMP 1 |
|
|
115 | #endif |
|
|
116 | |
|
|
117 | #if ECB_NO_SMP |
| 72 | #define ECB_MEMORY_FENCE do { } while (0) |
118 | #define ECB_MEMORY_FENCE do { } while (0) |
| 73 | #endif |
119 | #endif |
| 74 | |
120 | |
| 75 | #ifndef ECB_MEMORY_FENCE |
121 | #ifndef ECB_MEMORY_FENCE |
| 76 | #if ECB_GCC_VERSION(2,5) |
122 | #if ECB_GCC_VERSION(2,5) || defined __INTEL_COMPILER || (__llvm__ && __GNUC__) || __SUNPRO_C >= 0x5110 || __SUNPRO_CC >= 0x5110 |
| 77 | #if __x86 |
123 | #if __i386 || __i386__ |
| 78 | #define ECB_MEMORY_FENCE __asm__ __volatile__ ("lock; orb $0, -1(%%esp)" : : : "memory") |
124 | #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 */ |
125 | #define ECB_MEMORY_FENCE_ACQUIRE __asm__ __volatile__ ("" : : : "memory") |
| 80 | #define ECB_MEMORY_FENCE_RELEASE do { } while (0) /* unlikely to change in future cpus */ |
126 | #define ECB_MEMORY_FENCE_RELEASE __asm__ __volatile__ ("") |
| 81 | #elif __amd64 |
127 | #elif __amd64 || __amd64__ || __x86_64 || __x86_64__ |
| 82 | #define ECB_MEMORY_FENCE __asm__ __volatile__ ("mfence" : : : "memory") |
128 | #define ECB_MEMORY_FENCE __asm__ __volatile__ ("mfence" : : : "memory") |
| 83 | #define ECB_MEMORY_FENCE_ACQUIRE __asm__ __volatile__ ("lfence" : : : "memory") |
129 | #define ECB_MEMORY_FENCE_ACQUIRE __asm__ __volatile__ ("" : : : "memory") |
| 84 | #define ECB_MEMORY_FENCE_RELEASE __asm__ __volatile__ ("sfence") /* play safe - not needed in any current cpu */ |
130 | #define ECB_MEMORY_FENCE_RELEASE __asm__ __volatile__ ("") |
| 85 | #elif __powerpc__ || __ppc__ || __powerpc64__ || __ppc64__ |
131 | #elif __powerpc__ || __ppc__ || __powerpc64__ || __ppc64__ |
| 86 | #define ECB_MEMORY_FENCE __asm__ __volatile__ ("sync" : : : "memory") |
132 | #define ECB_MEMORY_FENCE __asm__ __volatile__ ("sync" : : : "memory") |
| 87 | #elif defined(__ARM_ARCH_6__ ) || defined(__ARM_ARCH_6J__ ) \ |
133 | #elif defined __ARM_ARCH_6__ || defined __ARM_ARCH_6J__ \ |
| 88 | || defined(__ARM_ARCH_6K__) || defined(__ARM_ARCH_6ZK__) |
134 | || 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") |
135 | #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__ ) \ |
136 | #elif defined __ARM_ARCH_7__ || defined __ARM_ARCH_7A__ \ |
| 91 | || defined(__ARM_ARCH_7M__) || defined(__ARM_ARCH_7R__ ) |
137 | || defined __ARM_ARCH_7M__ || defined __ARM_ARCH_7R__ |
| 92 | #define ECB_MEMORY_FENCE __asm__ __volatile__ ("dmb" : : : "memory") |
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 its 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 | #elif defined __m68k__ |
|
|
157 | #define ECB_MEMORY_FENCE __asm__ __volatile__ ("" : : : "memory") |
|
|
158 | #elif defined __m88k__ |
|
|
159 | #define ECB_MEMORY_FENCE __asm__ __volatile__ ("tb1 0,%%r0,128" : : : "memory") |
|
|
160 | #elif defined __sh__ |
|
|
161 | #define ECB_MEMORY_FENCE __asm__ __volatile__ ("" : : : "memory") |
| 93 | #endif |
162 | #endif |
| 94 | #endif |
163 | #endif |
| 95 | #endif |
164 | #endif |
| 96 | |
165 | |
| 97 | #ifndef ECB_MEMORY_FENCE |
166 | #ifndef ECB_MEMORY_FENCE |
|
|
167 | #if ECB_GCC_VERSION(4,7) |
|
|
168 | /* see comment below (stdatomic.h) about the C11 memory model. */ |
|
|
169 | #define ECB_MEMORY_FENCE __atomic_thread_fence (__ATOMIC_SEQ_CST) |
|
|
170 | |
|
|
171 | /* The __has_feature syntax from clang is so misdesigned that we cannot use it |
|
|
172 | * without risking compile time errors with other compilers. We *could* |
|
|
173 | * define our own ecb_clang_has_feature, but I just can't be bothered to work |
|
|
174 | * around this shit time and again. |
|
|
175 | * #elif defined __clang && __has_feature (cxx_atomic) |
|
|
176 | * // see comment below (stdatomic.h) about the C11 memory model. |
|
|
177 | * #define ECB_MEMORY_FENCE __c11_atomic_thread_fence (__ATOMIC_SEQ_CST) |
|
|
178 | */ |
|
|
179 | |
| 98 | #if ECB_GCC_VERSION(4,4) || defined(__INTEL_COMPILER) |
180 | #elif ECB_GCC_VERSION(4,4) || defined __INTEL_COMPILER || defined __clang__ |
| 99 | #define ECB_MEMORY_FENCE __sync_synchronize () |
181 | #define ECB_MEMORY_FENCE __sync_synchronize () |
| 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 ); }) */ |
|
|
| 102 | #elif _MSC_VER >= 1400 /* VC++ 2005 */ |
182 | #elif _MSC_VER >= 1400 /* VC++ 2005 */ |
| 103 | #pragma intrinsic(_ReadBarrier,_WriteBarrier,_ReadWriteBarrier) |
183 | #pragma intrinsic(_ReadBarrier,_WriteBarrier,_ReadWriteBarrier) |
| 104 | #define ECB_MEMORY_FENCE _ReadWriteBarrier () |
184 | #define ECB_MEMORY_FENCE _ReadWriteBarrier () |
| 105 | #define ECB_MEMORY_FENCE_ACQUIRE _ReadWriteBarrier () /* according to msdn, _ReadBarrier is not a load fence */ |
185 | #define ECB_MEMORY_FENCE_ACQUIRE _ReadWriteBarrier () /* according to msdn, _ReadBarrier is not a load fence */ |
| 106 | #define ECB_MEMORY_FENCE_RELEASE _WriteBarrier () |
186 | #define ECB_MEMORY_FENCE_RELEASE _WriteBarrier () |
| 107 | #elif defined(_WIN32) |
187 | #elif defined _WIN32 |
| 108 | #include <WinNT.h> |
188 | #include <WinNT.h> |
| 109 | #define ECB_MEMORY_FENCE MemoryBarrier () /* actually just xchg on x86... scary */ |
189 | #define ECB_MEMORY_FENCE MemoryBarrier () /* actually just xchg on x86... scary */ |
|
|
190 | #elif __SUNPRO_C >= 0x5110 || __SUNPRO_CC >= 0x5110 |
|
|
191 | #include <mbarrier.h> |
|
|
192 | #define ECB_MEMORY_FENCE __machine_rw_barrier () |
|
|
193 | #define ECB_MEMORY_FENCE_ACQUIRE __machine_r_barrier () |
|
|
194 | #define ECB_MEMORY_FENCE_RELEASE __machine_w_barrier () |
|
|
195 | #elif __xlC__ |
|
|
196 | #define ECB_MEMORY_FENCE __sync () |
|
|
197 | #endif |
|
|
198 | #endif |
|
|
199 | |
|
|
200 | #ifndef ECB_MEMORY_FENCE |
|
|
201 | #if ECB_C11 && !defined __STDC_NO_ATOMICS__ |
|
|
202 | /* we assume that these memory fences work on all variables/all memory accesses, */ |
|
|
203 | /* not just C11 atomics and atomic accesses */ |
|
|
204 | #include <stdatomic.h> |
|
|
205 | /* Unfortunately, neither gcc 4.7 nor clang 3.1 generate any instructions for */ |
|
|
206 | /* any fence other than seq_cst, which isn't very efficient for us. */ |
|
|
207 | /* Why that is, we don't know - either the C11 memory model is quite useless */ |
|
|
208 | /* for most usages, or gcc and clang have a bug */ |
|
|
209 | /* I *currently* lean towards the latter, and inefficiently implement */ |
|
|
210 | /* all three of ecb's fences as a seq_cst fence */ |
|
|
211 | #define ECB_MEMORY_FENCE atomic_thread_fence (memory_order_seq_cst) |
| 110 | #endif |
212 | #endif |
| 111 | #endif |
213 | #endif |
| 112 | |
214 | |
| 113 | #ifndef ECB_MEMORY_FENCE |
215 | #ifndef ECB_MEMORY_FENCE |
| 114 | #if !ECB_AVOID_PTHREADS |
216 | #if !ECB_AVOID_PTHREADS |
| … | |
… | |
| 126 | static pthread_mutex_t ecb_mf_lock = PTHREAD_MUTEX_INITIALIZER; |
228 | static pthread_mutex_t ecb_mf_lock = PTHREAD_MUTEX_INITIALIZER; |
| 127 | #define ECB_MEMORY_FENCE do { pthread_mutex_lock (&ecb_mf_lock); pthread_mutex_unlock (&ecb_mf_lock); } while (0) |
229 | #define ECB_MEMORY_FENCE do { pthread_mutex_lock (&ecb_mf_lock); pthread_mutex_unlock (&ecb_mf_lock); } while (0) |
| 128 | #endif |
230 | #endif |
| 129 | #endif |
231 | #endif |
| 130 | |
232 | |
| 131 | #if !defined(ECB_MEMORY_FENCE_ACQUIRE) && defined(ECB_MEMORY_FENCE) |
233 | #if !defined ECB_MEMORY_FENCE_ACQUIRE && defined ECB_MEMORY_FENCE |
| 132 | #define ECB_MEMORY_FENCE_ACQUIRE ECB_MEMORY_FENCE |
234 | #define ECB_MEMORY_FENCE_ACQUIRE ECB_MEMORY_FENCE |
| 133 | #endif |
235 | #endif |
| 134 | |
236 | |
| 135 | #if !defined(ECB_MEMORY_FENCE_RELEASE) && defined(ECB_MEMORY_FENCE) |
237 | #if !defined ECB_MEMORY_FENCE_RELEASE && defined ECB_MEMORY_FENCE |
| 136 | #define ECB_MEMORY_FENCE_RELEASE ECB_MEMORY_FENCE |
238 | #define ECB_MEMORY_FENCE_RELEASE ECB_MEMORY_FENCE |
| 137 | #endif |
239 | #endif |
| 138 | |
240 | |
| 139 | /*****************************************************************************/ |
241 | /*****************************************************************************/ |
| 140 | |
|
|
| 141 | #define ECB_C99 (__STDC_VERSION__ >= 199901L) |
|
|
| 142 | |
242 | |
| 143 | #if __cplusplus |
243 | #if __cplusplus |
| 144 | #define ecb_inline static inline |
244 | #define ecb_inline static inline |
| 145 | #elif ECB_GCC_VERSION(2,5) |
245 | #elif ECB_GCC_VERSION(2,5) |
| 146 | #define ecb_inline static __inline__ |
246 | #define ecb_inline static __inline__ |
| … | |
… | |
| 185 | #elif ECB_GCC_VERSION(3,0) |
285 | #elif ECB_GCC_VERSION(3,0) |
| 186 | #define ecb_decltype(x) __typeof(x) |
286 | #define ecb_decltype(x) __typeof(x) |
| 187 | #endif |
287 | #endif |
| 188 | |
288 | |
| 189 | #define ecb_noinline ecb_attribute ((__noinline__)) |
289 | #define ecb_noinline ecb_attribute ((__noinline__)) |
| 190 | #define ecb_noreturn ecb_attribute ((__noreturn__)) |
|
|
| 191 | #define ecb_unused ecb_attribute ((__unused__)) |
290 | #define ecb_unused ecb_attribute ((__unused__)) |
| 192 | #define ecb_const ecb_attribute ((__const__)) |
291 | #define ecb_const ecb_attribute ((__const__)) |
| 193 | #define ecb_pure ecb_attribute ((__pure__)) |
292 | #define ecb_pure ecb_attribute ((__pure__)) |
|
|
293 | |
|
|
294 | #if ECB_C11 |
|
|
295 | #define ecb_noreturn _Noreturn |
|
|
296 | #else |
|
|
297 | #define ecb_noreturn ecb_attribute ((__noreturn__)) |
|
|
298 | #endif |
| 194 | |
299 | |
| 195 | #if ECB_GCC_VERSION(4,3) |
300 | #if ECB_GCC_VERSION(4,3) |
| 196 | #define ecb_artificial ecb_attribute ((__artificial__)) |
301 | #define ecb_artificial ecb_attribute ((__artificial__)) |
| 197 | #define ecb_hot ecb_attribute ((__hot__)) |
302 | #define ecb_hot ecb_attribute ((__hot__)) |
| 198 | #define ecb_cold ecb_attribute ((__cold__)) |
303 | #define ecb_cold ecb_attribute ((__cold__)) |
| … | |
… | |
| 289 | |
394 | |
| 290 | return r + ecb_ld32 (x); |
395 | return r + ecb_ld32 (x); |
| 291 | } |
396 | } |
| 292 | #endif |
397 | #endif |
| 293 | |
398 | |
|
|
399 | ecb_function_ ecb_bool ecb_is_pot32 (uint32_t x) ecb_const; |
|
|
400 | ecb_function_ ecb_bool ecb_is_pot32 (uint32_t x) { return !(x & (x - 1)); } |
|
|
401 | ecb_function_ ecb_bool ecb_is_pot64 (uint64_t x) ecb_const; |
|
|
402 | ecb_function_ ecb_bool ecb_is_pot64 (uint64_t x) { return !(x & (x - 1)); } |
|
|
403 | |
|
|
404 | ecb_function_ uint8_t ecb_bitrev8 (uint8_t x) ecb_const; |
|
|
405 | ecb_function_ uint8_t ecb_bitrev8 (uint8_t x) |
|
|
406 | { |
|
|
407 | return ( (x * 0x0802U & 0x22110U) |
|
|
408 | | (x * 0x8020U & 0x88440U)) * 0x10101U >> 16; |
|
|
409 | } |
|
|
410 | |
|
|
411 | ecb_function_ uint16_t ecb_bitrev16 (uint16_t x) ecb_const; |
|
|
412 | ecb_function_ uint16_t ecb_bitrev16 (uint16_t x) |
|
|
413 | { |
|
|
414 | x = ((x >> 1) & 0x5555) | ((x & 0x5555) << 1); |
|
|
415 | x = ((x >> 2) & 0x3333) | ((x & 0x3333) << 2); |
|
|
416 | x = ((x >> 4) & 0x0f0f) | ((x & 0x0f0f) << 4); |
|
|
417 | x = ( x >> 8 ) | ( x << 8); |
|
|
418 | |
|
|
419 | return x; |
|
|
420 | } |
|
|
421 | |
|
|
422 | ecb_function_ uint32_t ecb_bitrev32 (uint32_t x) ecb_const; |
|
|
423 | ecb_function_ uint32_t ecb_bitrev32 (uint32_t x) |
|
|
424 | { |
|
|
425 | x = ((x >> 1) & 0x55555555) | ((x & 0x55555555) << 1); |
|
|
426 | x = ((x >> 2) & 0x33333333) | ((x & 0x33333333) << 2); |
|
|
427 | x = ((x >> 4) & 0x0f0f0f0f) | ((x & 0x0f0f0f0f) << 4); |
|
|
428 | x = ((x >> 8) & 0x00ff00ff) | ((x & 0x00ff00ff) << 8); |
|
|
429 | x = ( x >> 16 ) | ( x << 16); |
|
|
430 | |
|
|
431 | return x; |
|
|
432 | } |
|
|
433 | |
| 294 | /* popcount64 is only available on 64 bit cpus as gcc builtin */ |
434 | /* popcount64 is only available on 64 bit cpus as gcc builtin */ |
| 295 | /* so for this version we are lazy */ |
435 | /* so for this version we are lazy */ |
| 296 | ecb_function_ int ecb_popcount64 (uint64_t x) ecb_const; |
436 | ecb_function_ int ecb_popcount64 (uint64_t x) ecb_const; |
| 297 | ecb_function_ int |
437 | ecb_function_ int |
| 298 | ecb_popcount64 (uint64_t x) |
438 | ecb_popcount64 (uint64_t x) |
| … | |
… | |
| 347 | |
487 | |
| 348 | #if ECB_GCC_VERSION(4,5) |
488 | #if ECB_GCC_VERSION(4,5) |
| 349 | #define ecb_unreachable() __builtin_unreachable () |
489 | #define ecb_unreachable() __builtin_unreachable () |
| 350 | #else |
490 | #else |
| 351 | /* this seems to work fine, but gcc always emits a warning for it :/ */ |
491 | /* this seems to work fine, but gcc always emits a warning for it :/ */ |
| 352 | ecb_function_ void ecb_unreachable (void) ecb_noreturn; |
492 | ecb_inline void ecb_unreachable (void) ecb_noreturn; |
| 353 | ecb_function_ void ecb_unreachable (void) { } |
493 | ecb_inline void ecb_unreachable (void) { } |
| 354 | #endif |
494 | #endif |
| 355 | |
495 | |
| 356 | /* try to tell the compiler that some condition is definitely true */ |
496 | /* try to tell the compiler that some condition is definitely true */ |
| 357 | #define ecb_assume(cond) do { if (!(cond)) ecb_unreachable (); } while (0) |
497 | #define ecb_assume(cond) if (!(cond)) ecb_unreachable (); else 0 |
| 358 | |
498 | |
| 359 | ecb_function_ unsigned char ecb_byteorder_helper (void) ecb_const; |
499 | ecb_inline unsigned char ecb_byteorder_helper (void) ecb_const; |
| 360 | ecb_function_ unsigned char |
500 | ecb_inline unsigned char |
| 361 | ecb_byteorder_helper (void) |
501 | ecb_byteorder_helper (void) |
| 362 | { |
502 | { |
| 363 | const uint32_t u = 0x11223344; |
503 | /* the union code still generates code under pressure in gcc, */ |
| 364 | return *(unsigned char *)&u; |
504 | /* but less than using pointers, and always seems to */ |
|
|
505 | /* successfully return a constant. */ |
|
|
506 | /* the reason why we have this horrible preprocessor mess */ |
|
|
507 | /* is to avoid it in all cases, at least on common architectures */ |
|
|
508 | /* or when using a recent enough gcc version (>= 4.6) */ |
|
|
509 | #if __i386 || __i386__ || _M_X86 || __amd64 || __amd64__ || _M_X64 |
|
|
510 | return 0x44; |
|
|
511 | #elif __BYTE_ORDER__ && __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__ |
|
|
512 | return 0x44; |
|
|
513 | #elif __BYTE_ORDER__ && __BYTE_ORDER__ == __ORDER_BIG_ENDIAN__ |
|
|
514 | return 0x11; |
|
|
515 | #else |
|
|
516 | union |
|
|
517 | { |
|
|
518 | uint32_t i; |
|
|
519 | uint8_t c; |
|
|
520 | } u = { 0x11223344 }; |
|
|
521 | return u.c; |
|
|
522 | #endif |
| 365 | } |
523 | } |
| 366 | |
524 | |
| 367 | ecb_function_ ecb_bool ecb_big_endian (void) ecb_const; |
525 | ecb_inline ecb_bool ecb_big_endian (void) ecb_const; |
| 368 | ecb_function_ ecb_bool ecb_big_endian (void) { return ecb_byteorder_helper () == 0x11; } |
526 | ecb_inline ecb_bool ecb_big_endian (void) { return ecb_byteorder_helper () == 0x11; } |
| 369 | ecb_function_ ecb_bool ecb_little_endian (void) ecb_const; |
527 | ecb_inline ecb_bool ecb_little_endian (void) ecb_const; |
| 370 | ecb_function_ ecb_bool ecb_little_endian (void) { return ecb_byteorder_helper () == 0x44; } |
528 | ecb_inline ecb_bool ecb_little_endian (void) { return ecb_byteorder_helper () == 0x44; } |
| 371 | |
529 | |
| 372 | #if ECB_GCC_VERSION(3,0) || ECB_C99 |
530 | #if ECB_GCC_VERSION(3,0) || ECB_C99 |
| 373 | #define ecb_mod(m,n) ((m) % (n) + ((m) % (n) < 0 ? (n) : 0)) |
531 | #define ecb_mod(m,n) ((m) % (n) + ((m) % (n) < 0 ? (n) : 0)) |
| 374 | #else |
532 | #else |
| 375 | #define ecb_mod(m,n) ((m) < 0 ? ((n) - 1 - ((-1 - (m)) % (n))) : ((m) % (n))) |
533 | #define ecb_mod(m,n) ((m) < 0 ? ((n) - 1 - ((-1 - (m)) % (n))) : ((m) % (n))) |
|
|
534 | #endif |
|
|
535 | |
|
|
536 | #if __cplusplus |
|
|
537 | template<typename T> |
|
|
538 | static inline T ecb_div_rd (T val, T div) |
|
|
539 | { |
|
|
540 | return val < 0 ? - ((-val + div - 1) / div) : (val ) / div; |
|
|
541 | } |
|
|
542 | template<typename T> |
|
|
543 | static inline T ecb_div_ru (T val, T div) |
|
|
544 | { |
|
|
545 | return val < 0 ? - ((-val ) / div) : (val + div - 1) / div; |
|
|
546 | } |
|
|
547 | #else |
|
|
548 | #define ecb_div_rd(val,div) ((val) < 0 ? - ((-(val) + (div) - 1) / (div)) : ((val) ) / (div)) |
|
|
549 | #define ecb_div_ru(val,div) ((val) < 0 ? - ((-(val) ) / (div)) : ((val) + (div) - 1) / (div)) |
| 376 | #endif |
550 | #endif |
| 377 | |
551 | |
| 378 | #if ecb_cplusplus_does_not_suck |
552 | #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) */ |
553 | /* 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> |
554 | template<typename T, int N> |
| … | |
… | |
| 384 | } |
558 | } |
| 385 | #else |
559 | #else |
| 386 | #define ecb_array_length(name) (sizeof (name) / sizeof (name [0])) |
560 | #define ecb_array_length(name) (sizeof (name) / sizeof (name [0])) |
| 387 | #endif |
561 | #endif |
| 388 | |
562 | |
|
|
563 | /*******************************************************************************/ |
|
|
564 | /* floating point stuff, can be disabled by defining ECB_NO_LIBM */ |
|
|
565 | |
|
|
566 | /* basically, everything uses "ieee pure-endian" floating point numbers */ |
|
|
567 | /* the only noteworthy exception is ancient armle, which uses order 43218765 */ |
|
|
568 | #if 0 \ |
|
|
569 | || __i386 || __i386__ \ |
|
|
570 | || __amd64 || __amd64__ || __x86_64 || __x86_64__ \ |
|
|
571 | || __powerpc__ || __ppc__ || __powerpc64__ || __ppc64__ \ |
|
|
572 | || defined __arm__ && defined __ARM_EABI__ \ |
|
|
573 | || defined __s390__ || defined __s390x__ \ |
|
|
574 | || defined __mips__ \ |
|
|
575 | || defined __alpha__ \ |
|
|
576 | || defined __hppa__ \ |
|
|
577 | || defined __ia64__ \ |
|
|
578 | || defined __m68k__ \ |
|
|
579 | || defined __m88k__ \ |
|
|
580 | || defined __sh__ \ |
|
|
581 | || defined _M_IX86 || defined _M_AMD64 || defined _M_IA64 |
|
|
582 | #define ECB_STDFP 1 |
|
|
583 | #include <string.h> /* for memcpy */ |
|
|
584 | #else |
|
|
585 | #define ECB_STDFP 0 |
|
|
586 | #endif |
|
|
587 | |
|
|
588 | #ifndef ECB_NO_LIBM |
|
|
589 | |
|
|
590 | #include <math.h> /* for frexp*, ldexp*, INFINITY, NAN */ |
|
|
591 | |
|
|
592 | /* only the oldest of old doesn't have this one. solaris. */ |
|
|
593 | #ifdef INFINITY |
|
|
594 | #define ECB_INFINITY INFINITY |
|
|
595 | #else |
|
|
596 | #define ECB_INFINITY HUGE_VAL |
| 389 | #endif |
597 | #endif |
| 390 | |
598 | |
|
|
599 | #ifdef NAN |
|
|
600 | #define ECB_NAN NAN |
|
|
601 | #else |
|
|
602 | #define ECB_NAN ECB_INFINITY |
|
|
603 | #endif |
|
|
604 | |
|
|
605 | /* converts an ieee half/binary16 to a float */ |
|
|
606 | ecb_function_ float ecb_binary16_to_float (uint16_t x) ecb_const; |
|
|
607 | ecb_function_ float |
|
|
608 | ecb_binary16_to_float (uint16_t x) |
|
|
609 | { |
|
|
610 | int e = (x >> 10) & 0x1f; |
|
|
611 | int m = x & 0x3ff; |
|
|
612 | float r; |
|
|
613 | |
|
|
614 | if (!e ) r = ldexpf (m , -24); |
|
|
615 | else if (e != 31) r = ldexpf (m + 0x400, e - 25); |
|
|
616 | else if (m ) r = ECB_NAN; |
|
|
617 | else r = ECB_INFINITY; |
|
|
618 | |
|
|
619 | return x & 0x8000 ? -r : r; |
|
|
620 | } |
|
|
621 | |
|
|
622 | /* convert a float to ieee single/binary32 */ |
|
|
623 | ecb_function_ uint32_t ecb_float_to_binary32 (float x) ecb_const; |
|
|
624 | ecb_function_ uint32_t |
|
|
625 | ecb_float_to_binary32 (float x) |
|
|
626 | { |
|
|
627 | uint32_t r; |
|
|
628 | |
|
|
629 | #if ECB_STDFP |
|
|
630 | memcpy (&r, &x, 4); |
|
|
631 | #else |
|
|
632 | /* slow emulation, works for anything but -0 */ |
|
|
633 | uint32_t m; |
|
|
634 | int e; |
|
|
635 | |
|
|
636 | if (x == 0e0f ) return 0x00000000U; |
|
|
637 | if (x > +3.40282346638528860e+38f) return 0x7f800000U; |
|
|
638 | if (x < -3.40282346638528860e+38f) return 0xff800000U; |
|
|
639 | if (x != x ) return 0x7fbfffffU; |
|
|
640 | |
|
|
641 | m = frexpf (x, &e) * 0x1000000U; |
|
|
642 | |
|
|
643 | r = m & 0x80000000U; |
|
|
644 | |
|
|
645 | if (r) |
|
|
646 | m = -m; |
|
|
647 | |
|
|
648 | if (e <= -126) |
|
|
649 | { |
|
|
650 | m &= 0xffffffU; |
|
|
651 | m >>= (-125 - e); |
|
|
652 | e = -126; |
|
|
653 | } |
|
|
654 | |
|
|
655 | r |= (e + 126) << 23; |
|
|
656 | r |= m & 0x7fffffU; |
|
|
657 | #endif |
|
|
658 | |
|
|
659 | return r; |
|
|
660 | } |
|
|
661 | |
|
|
662 | /* converts an ieee single/binary32 to a float */ |
|
|
663 | ecb_function_ float ecb_binary32_to_float (uint32_t x) ecb_const; |
|
|
664 | ecb_function_ float |
|
|
665 | ecb_binary32_to_float (uint32_t x) |
|
|
666 | { |
|
|
667 | float r; |
|
|
668 | |
|
|
669 | #if ECB_STDFP |
|
|
670 | memcpy (&r, &x, 4); |
|
|
671 | #else |
|
|
672 | /* emulation, only works for normals and subnormals and +0 */ |
|
|
673 | int neg = x >> 31; |
|
|
674 | int e = (x >> 23) & 0xffU; |
|
|
675 | |
|
|
676 | x &= 0x7fffffU; |
|
|
677 | |
|
|
678 | if (e) |
|
|
679 | x |= 0x800000U; |
|
|
680 | else |
|
|
681 | e = 1; |
|
|
682 | |
|
|
683 | /* we distrust ldexpf a bit and do the 2**-24 scaling by an extra multiply */ |
|
|
684 | r = ldexpf (x * (0.5f / 0x800000U), e - 126); |
|
|
685 | |
|
|
686 | r = neg ? -r : r; |
|
|
687 | #endif |
|
|
688 | |
|
|
689 | return r; |
|
|
690 | } |
|
|
691 | |
|
|
692 | /* convert a double to ieee double/binary64 */ |
|
|
693 | ecb_function_ uint64_t ecb_double_to_binary64 (double x) ecb_const; |
|
|
694 | ecb_function_ uint64_t |
|
|
695 | ecb_double_to_binary64 (double x) |
|
|
696 | { |
|
|
697 | uint64_t r; |
|
|
698 | |
|
|
699 | #if ECB_STDFP |
|
|
700 | memcpy (&r, &x, 8); |
|
|
701 | #else |
|
|
702 | /* slow emulation, works for anything but -0 */ |
|
|
703 | uint64_t m; |
|
|
704 | int e; |
|
|
705 | |
|
|
706 | if (x == 0e0 ) return 0x0000000000000000U; |
|
|
707 | if (x > +1.79769313486231470e+308) return 0x7ff0000000000000U; |
|
|
708 | if (x < -1.79769313486231470e+308) return 0xfff0000000000000U; |
|
|
709 | if (x != x ) return 0X7ff7ffffffffffffU; |
|
|
710 | |
|
|
711 | m = frexp (x, &e) * 0x20000000000000U; |
|
|
712 | |
|
|
713 | r = m & 0x8000000000000000;; |
|
|
714 | |
|
|
715 | if (r) |
|
|
716 | m = -m; |
|
|
717 | |
|
|
718 | if (e <= -1022) |
|
|
719 | { |
|
|
720 | m &= 0x1fffffffffffffU; |
|
|
721 | m >>= (-1021 - e); |
|
|
722 | e = -1022; |
|
|
723 | } |
|
|
724 | |
|
|
725 | r |= ((uint64_t)(e + 1022)) << 52; |
|
|
726 | r |= m & 0xfffffffffffffU; |
|
|
727 | #endif |
|
|
728 | |
|
|
729 | return r; |
|
|
730 | } |
|
|
731 | |
|
|
732 | /* converts an ieee double/binary64 to a double */ |
|
|
733 | ecb_function_ double ecb_binary64_to_double (uint64_t x) ecb_const; |
|
|
734 | ecb_function_ double |
|
|
735 | ecb_binary64_to_double (uint64_t x) |
|
|
736 | { |
|
|
737 | double r; |
|
|
738 | |
|
|
739 | #if ECB_STDFP |
|
|
740 | memcpy (&r, &x, 8); |
|
|
741 | #else |
|
|
742 | /* emulation, only works for normals and subnormals and +0 */ |
|
|
743 | int neg = x >> 63; |
|
|
744 | int e = (x >> 52) & 0x7ffU; |
|
|
745 | |
|
|
746 | x &= 0xfffffffffffffU; |
|
|
747 | |
|
|
748 | if (e) |
|
|
749 | x |= 0x10000000000000U; |
|
|
750 | else |
|
|
751 | e = 1; |
|
|
752 | |
|
|
753 | /* we distrust ldexp a bit and do the 2**-53 scaling by an extra multiply */ |
|
|
754 | r = ldexp (x * (0.5 / 0x10000000000000U), e - 1022); |
|
|
755 | |
|
|
756 | r = neg ? -r : r; |
|
|
757 | #endif |
|
|
758 | |
|
|
759 | return r; |
|
|
760 | } |
|
|
761 | |
|
|
762 | #endif |
|
|
763 | |
|
|
764 | #endif |
|
|
765 | |
