ViewVC Help
View File | Revision Log | Show Annotations | Download File
/cvs/libecb/ecb.h
(Generate patch)

Comparing libecb/ecb.h (file contents):
Revision 1.136 by root, Tue Jun 17 05:51:31 2014 UTC vs.
Revision 1.179 by root, Sat Dec 28 08:01:05 2019 UTC

1/* 1/*
2 * libecb - http://software.schmorp.de/pkg/libecb 2 * libecb - http://software.schmorp.de/pkg/libecb
3 * 3 *
4 * Copyright (©) 2009-2014 Marc Alexander Lehmann <libecb@schmorp.de> 4 * Copyright (©) 2009-2015 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:
… …
40 40
41#ifndef ECB_H 41#ifndef ECB_H
42#define ECB_H 42#define ECB_H
43 43
44/* 16 bits major, 16 bits minor */ 44/* 16 bits major, 16 bits minor */
45#define ECB_VERSION 0x00010003 45#define ECB_VERSION 0x00010007
46 46
47#ifdef _WIN32 47#ifdef _WIN32
48 typedef signed char int8_t; 48 typedef signed char int8_t;
49 typedef unsigned char uint8_t; 49 typedef unsigned char uint8_t;
50 typedef signed short int16_t; 50 typedef signed short int16_t;
… …
67 typedef uint32_t uintptr_t; 67 typedef uint32_t uintptr_t;
68 typedef int32_t intptr_t; 68 typedef int32_t intptr_t;
69 #endif 69 #endif
70#else 70#else
71 #include <inttypes.h> 71 #include <inttypes.h>
72 #if UINTMAX_MAX > 0xffffffffU 72 #if (defined INTPTR_MAX ? INTPTR_MAX : ULONG_MAX) > 0xffffffffU
73 #define ECB_PTRSIZE 8 73 #define ECB_PTRSIZE 8
74 #else 74 #else
75 #define ECB_PTRSIZE 4 75 #define ECB_PTRSIZE 4
76 #endif 76 #endif
77#endif 77#endif
78 78
79#define ECB_GCC_AMD64 (__amd64 || __amd64__ || __x86_64 || __x86_64__)
80#define ECB_MSVC_AMD64 (_M_AMD64 || _M_X64)
81
82#ifndef ECB_OPTIMIZE_SIZE
83 #if __OPTIMIZE_SIZE__
84 #define ECB_OPTIMIZE_SIZE 1
85 #else
86 #define ECB_OPTIMIZE_SIZE 0
87 #endif
88#endif
89
79/* work around x32 idiocy by defining proper macros */ 90/* work around x32 idiocy by defining proper macros */
80#if __amd64 || __x86_64 || _M_AMD64 || _M_X64 91#if ECB_GCC_AMD64 || ECB_MSVC_AMD64
81 #if _ILP32 92 #if _ILP32
82 #define ECB_AMD64_X32 1 93 #define ECB_AMD64_X32 1
83 #else 94 #else
84 #define ECB_AMD64 1 95 #define ECB_AMD64 1
85 #endif 96 #endif
… …
90 * causing enormous grief in return for some better fake benchmark numbers. 101 * causing enormous grief in return for some better fake benchmark numbers.
91 * or so. 102 * or so.
92 * we try to detect these and simply assume they are not gcc - if they have 103 * we try to detect these and simply assume they are not gcc - if they have
93 * an issue with that they should have done it right in the first place. 104 * an issue with that they should have done it right in the first place.
94 */ 105 */
95#ifndef ECB_GCC_VERSION
96 #if !defined __GNUC_MINOR__ || defined __INTEL_COMPILER || defined __SUNPRO_C || defined __SUNPRO_CC || defined __llvm__ || defined __clang__ 106#if !defined __GNUC_MINOR__ || defined __INTEL_COMPILER || defined __SUNPRO_C || defined __SUNPRO_CC || defined __llvm__ || defined __clang__
97 #define ECB_GCC_VERSION(major,minor) 0 107 #define ECB_GCC_VERSION(major,minor) 0
98 #else 108#else
99 #define ECB_GCC_VERSION(major,minor) (__GNUC__ > (major) || (__GNUC__ == (major) && __GNUC_MINOR__ >= (minor))) 109 #define ECB_GCC_VERSION(major,minor) (__GNUC__ > (major) || (__GNUC__ == (major) && __GNUC_MINOR__ >= (minor)))
100 #endif 110#endif
111
112#define ECB_CLANG_VERSION(major,minor) (__clang_major__ > (major) || (__clang_major__ == (major) && __clang_minor__ >= (minor)))
113
114#if __clang__ && defined __has_builtin
115 #define ECB_CLANG_BUILTIN(x) __has_builtin (x)
116#else
117 #define ECB_CLANG_BUILTIN(x) 0
118#endif
119
120#if __clang__ && defined __has_extension
121 #define ECB_CLANG_EXTENSION(x) __has_extension (x)
122#else
123 #define ECB_CLANG_EXTENSION(x) 0
101#endif 124#endif
102 125
103#define ECB_CPP (__cplusplus+0) 126#define ECB_CPP (__cplusplus+0)
104#define ECB_CPP11 (__cplusplus >= 201103L) 127#define ECB_CPP11 (__cplusplus >= 201103L)
128#define ECB_CPP14 (__cplusplus >= 201402L)
129#define ECB_CPP17 (__cplusplus >= 201703L)
105 130
106#if ECB_CPP 131#if ECB_CPP
107 #define ECB_C 0 132 #define ECB_C 0
108 #define ECB_STDC_VERSION 0 133 #define ECB_STDC_VERSION 0
109#else 134#else
… …
111 #define ECB_STDC_VERSION __STDC_VERSION__ 136 #define ECB_STDC_VERSION __STDC_VERSION__
112#endif 137#endif
113 138
114#define ECB_C99 (ECB_STDC_VERSION >= 199901L) 139#define ECB_C99 (ECB_STDC_VERSION >= 199901L)
115#define ECB_C11 (ECB_STDC_VERSION >= 201112L) 140#define ECB_C11 (ECB_STDC_VERSION >= 201112L)
141#define ECB_C17 (ECB_STDC_VERSION >= 201710L)
116 142
117#if ECB_CPP 143#if ECB_CPP
118 #define ECB_EXTERN_C extern "C" 144 #define ECB_EXTERN_C extern "C"
119 #define ECB_EXTERN_C_BEG ECB_EXTERN_C { 145 #define ECB_EXTERN_C_BEG ECB_EXTERN_C {
120 #define ECB_EXTERN_C_END } 146 #define ECB_EXTERN_C_END }
… …
135 161
136#if ECB_NO_SMP 162#if ECB_NO_SMP
137 #define ECB_MEMORY_FENCE do { } while (0) 163 #define ECB_MEMORY_FENCE do { } while (0)
138#endif 164#endif
139 165
166/* http://www-01.ibm.com/support/knowledgecenter/SSGH3R_13.1.0/com.ibm.xlcpp131.aix.doc/compiler_ref/compiler_builtins.html */
167#if __xlC__ && ECB_CPP
168 #include <builtins.h>
169#endif
170
171#if 1400 <= _MSC_VER
172 #include <intrin.h> /* fence functions _ReadBarrier, also bit search functions _BitScanReverse */
173#endif
174
140#ifndef ECB_MEMORY_FENCE 175#ifndef ECB_MEMORY_FENCE
141 #if ECB_GCC_VERSION(2,5) || defined __INTEL_COMPILER || (__llvm__ && __GNUC__) || __SUNPRO_C >= 0x5110 || __SUNPRO_CC >= 0x5110 176 #if ECB_GCC_VERSION(2,5) || defined __INTEL_COMPILER || (__llvm__ && __GNUC__) || __SUNPRO_C >= 0x5110 || __SUNPRO_CC >= 0x5110
177 #define ECB_MEMORY_FENCE_RELAXED __asm__ __volatile__ ("" : : : "memory")
142 #if __i386 || __i386__ 178 #if __i386 || __i386__
143 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("lock; orb $0, -1(%%esp)" : : : "memory") 179 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("lock; orb $0, -1(%%esp)" : : : "memory")
144 #define ECB_MEMORY_FENCE_ACQUIRE __asm__ __volatile__ ("" : : : "memory") 180 #define ECB_MEMORY_FENCE_ACQUIRE __asm__ __volatile__ ("" : : : "memory")
145 #define ECB_MEMORY_FENCE_RELEASE __asm__ __volatile__ ("") 181 #define ECB_MEMORY_FENCE_RELEASE __asm__ __volatile__ ("" : : : "memory")
146 #elif __amd64 || __amd64__ || __x86_64 || __x86_64__ 182 #elif ECB_GCC_AMD64
147 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("mfence" : : : "memory") 183 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("mfence" : : : "memory")
148 #define ECB_MEMORY_FENCE_ACQUIRE __asm__ __volatile__ ("" : : : "memory") 184 #define ECB_MEMORY_FENCE_ACQUIRE __asm__ __volatile__ ("" : : : "memory")
149 #define ECB_MEMORY_FENCE_RELEASE __asm__ __volatile__ ("") 185 #define ECB_MEMORY_FENCE_RELEASE __asm__ __volatile__ ("" : : : "memory")
150 #elif __powerpc__ || __ppc__ || __powerpc64__ || __ppc64__ 186 #elif __powerpc__ || __ppc__ || __powerpc64__ || __ppc64__
151 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("sync" : : : "memory") 187 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("sync" : : : "memory")
188 #elif defined __ARM_ARCH_2__ \
189 || defined __ARM_ARCH_3__ || defined __ARM_ARCH_3M__ \
190 || defined __ARM_ARCH_4__ || defined __ARM_ARCH_4T__ \
191 || defined __ARM_ARCH_5__ || defined __ARM_ARCH_5E__ \
192 || defined __ARM_ARCH_5T__ || defined __ARM_ARCH_5TE__ \
193 || defined __ARM_ARCH_5TEJ__
194 /* should not need any, unless running old code on newer cpu - arm doesn't support that */
152 #elif defined __ARM_ARCH_6__ || defined __ARM_ARCH_6J__ \ 195 #elif defined __ARM_ARCH_6__ || defined __ARM_ARCH_6J__ \
153 || defined __ARM_ARCH_6K__ || defined __ARM_ARCH_6ZK__ 196 || defined __ARM_ARCH_6K__ || defined __ARM_ARCH_6ZK__ \
197 || defined __ARM_ARCH_6T2__
154 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("mcr p15,0,%0,c7,c10,5" : : "r" (0) : "memory") 198 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("mcr p15,0,%0,c7,c10,5" : : "r" (0) : "memory")
155 #elif defined __ARM_ARCH_7__ || defined __ARM_ARCH_7A__ \ 199 #elif defined __ARM_ARCH_7__ || defined __ARM_ARCH_7A__ \
156 || defined __ARM_ARCH_7M__ || defined __ARM_ARCH_7R__ 200 || defined __ARM_ARCH_7R__ || defined __ARM_ARCH_7M__
157 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("dmb" : : : "memory") 201 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("dmb" : : : "memory")
158 #elif __aarch64__ 202 #elif __aarch64__
159 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("dmb ish" : : : "memory") 203 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("dmb ish" : : : "memory")
160 #elif (__sparc || __sparc__) && !__sparcv8 204 #elif (__sparc || __sparc__) && !(__sparc_v8__ || defined __sparcv8)
161 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("membar #LoadStore | #LoadLoad | #StoreStore | #StoreLoad" : : : "memory") 205 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("membar #LoadStore | #LoadLoad | #StoreStore | #StoreLoad" : : : "memory")
162 #define ECB_MEMORY_FENCE_ACQUIRE __asm__ __volatile__ ("membar #LoadStore | #LoadLoad" : : : "memory") 206 #define ECB_MEMORY_FENCE_ACQUIRE __asm__ __volatile__ ("membar #LoadStore | #LoadLoad" : : : "memory")
163 #define ECB_MEMORY_FENCE_RELEASE __asm__ __volatile__ ("membar #LoadStore | #StoreStore") 207 #define ECB_MEMORY_FENCE_RELEASE __asm__ __volatile__ ("membar #LoadStore | #StoreStore")
164 #elif defined __s390__ || defined __s390x__ 208 #elif defined __s390__ || defined __s390x__
165 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("bcr 15,0" : : : "memory") 209 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("bcr 15,0" : : : "memory")
… …
188 #if ECB_GCC_VERSION(4,7) 232 #if ECB_GCC_VERSION(4,7)
189 /* see comment below (stdatomic.h) about the C11 memory model. */ 233 /* see comment below (stdatomic.h) about the C11 memory model. */
190 #define ECB_MEMORY_FENCE __atomic_thread_fence (__ATOMIC_SEQ_CST) 234 #define ECB_MEMORY_FENCE __atomic_thread_fence (__ATOMIC_SEQ_CST)
191 #define ECB_MEMORY_FENCE_ACQUIRE __atomic_thread_fence (__ATOMIC_ACQUIRE) 235 #define ECB_MEMORY_FENCE_ACQUIRE __atomic_thread_fence (__ATOMIC_ACQUIRE)
192 #define ECB_MEMORY_FENCE_RELEASE __atomic_thread_fence (__ATOMIC_RELEASE) 236 #define ECB_MEMORY_FENCE_RELEASE __atomic_thread_fence (__ATOMIC_RELEASE)
237 #define ECB_MEMORY_FENCE_RELAXED __atomic_thread_fence (__ATOMIC_RELAXED)
193 238
194 /* The __has_feature syntax from clang is so misdesigned that we cannot use it 239 #elif ECB_CLANG_EXTENSION(c_atomic)
195 * without risking compile time errors with other compilers. We *could*
196 * define our own ecb_clang_has_feature, but I just can't be bothered to work
197 * around this shit time and again.
198 * #elif defined __clang && __has_feature (cxx_atomic)
199 * // see comment below (stdatomic.h) about the C11 memory model. 240 /* see comment below (stdatomic.h) about the C11 memory model. */
200 * #define ECB_MEMORY_FENCE __c11_atomic_thread_fence (__ATOMIC_SEQ_CST) 241 #define ECB_MEMORY_FENCE __c11_atomic_thread_fence (__ATOMIC_SEQ_CST)
201 * #define ECB_MEMORY_FENCE_ACQUIRE __c11_atomic_thread_fence (__ATOMIC_ACQUIRE) 242 #define ECB_MEMORY_FENCE_ACQUIRE __c11_atomic_thread_fence (__ATOMIC_ACQUIRE)
202 * #define ECB_MEMORY_FENCE_RELEASE __c11_atomic_thread_fence (__ATOMIC_RELEASE) 243 #define ECB_MEMORY_FENCE_RELEASE __c11_atomic_thread_fence (__ATOMIC_RELEASE)
203 */ 244 #define ECB_MEMORY_FENCE_RELAXED __c11_atomic_thread_fence (__ATOMIC_RELAXED)
204 245
205 #elif ECB_GCC_VERSION(4,4) || defined __INTEL_COMPILER || defined __clang__ 246 #elif ECB_GCC_VERSION(4,4) || defined __INTEL_COMPILER || defined __clang__
206 #define ECB_MEMORY_FENCE __sync_synchronize () 247 #define ECB_MEMORY_FENCE __sync_synchronize ()
207 #elif _MSC_VER >= 1500 /* VC++ 2008 */ 248 #elif _MSC_VER >= 1500 /* VC++ 2008 */
208 /* apparently, microsoft broke all the memory barrier stuff in Visual Studio 2008... */ 249 /* apparently, microsoft broke all the memory barrier stuff in Visual Studio 2008... */
… …
218 #elif defined _WIN32 259 #elif defined _WIN32
219 #include <WinNT.h> 260 #include <WinNT.h>
220 #define ECB_MEMORY_FENCE MemoryBarrier () /* actually just xchg on x86... scary */ 261 #define ECB_MEMORY_FENCE MemoryBarrier () /* actually just xchg on x86... scary */
221 #elif __SUNPRO_C >= 0x5110 || __SUNPRO_CC >= 0x5110 262 #elif __SUNPRO_C >= 0x5110 || __SUNPRO_CC >= 0x5110
222 #include <mbarrier.h> 263 #include <mbarrier.h>
223 #define ECB_MEMORY_FENCE __machine_rw_barrier () 264 #define ECB_MEMORY_FENCE __machine_rw_barrier ()
224 #define ECB_MEMORY_FENCE_ACQUIRE __machine_r_barrier () 265 #define ECB_MEMORY_FENCE_ACQUIRE __machine_acq_barrier ()
225 #define ECB_MEMORY_FENCE_RELEASE __machine_w_barrier () 266 #define ECB_MEMORY_FENCE_RELEASE __machine_rel_barrier ()
267 #define ECB_MEMORY_FENCE_RELAXED __compiler_barrier ()
226 #elif __xlC__ 268 #elif __xlC__
227 #define ECB_MEMORY_FENCE __sync () 269 #define ECB_MEMORY_FENCE __sync ()
228 #endif 270 #endif
229#endif 271#endif
230 272
231#ifndef ECB_MEMORY_FENCE 273#ifndef ECB_MEMORY_FENCE
232 #if ECB_C11 && !defined __STDC_NO_ATOMICS__ 274 #if ECB_C11 && !defined __STDC_NO_ATOMICS__
233 /* we assume that these memory fences work on all variables/all memory accesses, */ 275 /* we assume that these memory fences work on all variables/all memory accesses, */
234 /* not just C11 atomics and atomic accesses */ 276 /* not just C11 atomics and atomic accesses */
235 #include <stdatomic.h> 277 #include <stdatomic.h>
236 /* Unfortunately, neither gcc 4.7 nor clang 3.1 generate any instructions for */
237 /* any fence other than seq_cst, which isn't very efficient for us. */
238 /* Why that is, we don't know - either the C11 memory model is quite useless */
239 /* for most usages, or gcc and clang have a bug */
240 /* I *currently* lean towards the latter, and inefficiently implement */
241 /* all three of ecb's fences as a seq_cst fence */
242 /* Update, gcc-4.8 generates mfence for all c++ fences, but nothing */
243 /* for all __atomic_thread_fence's except seq_cst */
244 #define ECB_MEMORY_FENCE atomic_thread_fence (memory_order_seq_cst) 278 #define ECB_MEMORY_FENCE atomic_thread_fence (memory_order_seq_cst)
279 #define ECB_MEMORY_FENCE_ACQUIRE atomic_thread_fence (memory_order_acquire)
280 #define ECB_MEMORY_FENCE_RELEASE atomic_thread_fence (memory_order_release)
245 #endif 281 #endif
246#endif 282#endif
247 283
248#ifndef ECB_MEMORY_FENCE 284#ifndef ECB_MEMORY_FENCE
249 #if !ECB_AVOID_PTHREADS 285 #if !ECB_AVOID_PTHREADS
… …
269 305
270#if !defined ECB_MEMORY_FENCE_RELEASE && defined ECB_MEMORY_FENCE 306#if !defined ECB_MEMORY_FENCE_RELEASE && defined ECB_MEMORY_FENCE
271 #define ECB_MEMORY_FENCE_RELEASE ECB_MEMORY_FENCE 307 #define ECB_MEMORY_FENCE_RELEASE ECB_MEMORY_FENCE
272#endif 308#endif
273 309
310#if !defined ECB_MEMORY_FENCE_RELAXED && defined ECB_MEMORY_FENCE
311 #define ECB_MEMORY_FENCE_RELAXED ECB_MEMORY_FENCE /* very heavy-handed */
312#endif
313
274/*****************************************************************************/ 314/*****************************************************************************/
275 315
276#if __cplusplus 316#if ECB_CPP
277 #define ecb_inline static inline 317 #define ecb_inline static inline
278#elif ECB_GCC_VERSION(2,5) 318#elif ECB_GCC_VERSION(2,5)
279 #define ecb_inline static __inline__ 319 #define ecb_inline static __inline__
280#elif ECB_C99 320#elif ECB_C99
281 #define ecb_inline static inline 321 #define ecb_inline static inline
… …
295 335
296#define ECB_CONCAT_(a, b) a ## b 336#define ECB_CONCAT_(a, b) a ## b
297#define ECB_CONCAT(a, b) ECB_CONCAT_(a, b) 337#define ECB_CONCAT(a, b) ECB_CONCAT_(a, b)
298#define ECB_STRINGIFY_(a) # a 338#define ECB_STRINGIFY_(a) # a
299#define ECB_STRINGIFY(a) ECB_STRINGIFY_(a) 339#define ECB_STRINGIFY(a) ECB_STRINGIFY_(a)
340#define ECB_STRINGIFY_EXPR(expr) ((expr), ECB_STRINGIFY_ (expr))
300 341
301#define ecb_function_ ecb_inline 342#define ecb_function_ ecb_inline
302 343
303#if ECB_GCC_VERSION(3,1) 344#if ECB_GCC_VERSION(3,1) || ECB_CLANG_VERSION(2,8)
304 #define ecb_attribute(attrlist) __attribute__(attrlist) 345 #define ecb_attribute(attrlist) __attribute__ (attrlist)
346#else
347 #define ecb_attribute(attrlist)
348#endif
349
350#if ECB_GCC_VERSION(3,1) || ECB_CLANG_BUILTIN(__builtin_constant_p)
305 #define ecb_is_constant(expr) __builtin_constant_p (expr) 351 #define ecb_is_constant(expr) __builtin_constant_p (expr)
306 #define ecb_expect(expr,value) __builtin_expect ((expr),(value))
307 #define ecb_prefetch(addr,rw,locality) __builtin_prefetch (addr, rw, locality)
308#else 352#else
309 #define ecb_attribute(attrlist)
310
311 /* possible C11 impl for integral types 353 /* possible C11 impl for integral types
312 typedef struct ecb_is_constant_struct ecb_is_constant_struct; 354 typedef struct ecb_is_constant_struct ecb_is_constant_struct;
313 #define ecb_is_constant(expr) _Generic ((1 ? (struct ecb_is_constant_struct *)0 : (void *)((expr) - (expr)), ecb_is_constant_struct *: 0, default: 1)) */ 355 #define ecb_is_constant(expr) _Generic ((1 ? (struct ecb_is_constant_struct *)0 : (void *)((expr) - (expr)), ecb_is_constant_struct *: 0, default: 1)) */
314 356
315 #define ecb_is_constant(expr) 0 357 #define ecb_is_constant(expr) 0
358#endif
359
360#if ECB_GCC_VERSION(3,1) || ECB_CLANG_BUILTIN(__builtin_expect)
361 #define ecb_expect(expr,value) __builtin_expect ((expr),(value))
362#else
316 #define ecb_expect(expr,value) (expr) 363 #define ecb_expect(expr,value) (expr)
364#endif
365
366#if ECB_GCC_VERSION(3,1) || ECB_CLANG_BUILTIN(__builtin_prefetch)
367 #define ecb_prefetch(addr,rw,locality) __builtin_prefetch (addr, rw, locality)
368#else
317 #define ecb_prefetch(addr,rw,locality) 369 #define ecb_prefetch(addr,rw,locality)
318#endif 370#endif
319 371
320/* no emulation for ecb_decltype */ 372/* no emulation for ecb_decltype */
321#if ECB_GCC_VERSION(4,5) 373#if ECB_CPP11
374 // older implementations might have problems with decltype(x)::type, work around it
375 template<class T> struct ecb_decltype_t { typedef T type; };
322 #define ecb_decltype(x) __decltype(x) 376 #define ecb_decltype(x) ecb_decltype_t<decltype (x)>::type
323#elif ECB_GCC_VERSION(3,0) 377#elif ECB_GCC_VERSION(3,0) || ECB_CLANG_VERSION(2,8)
324 #define ecb_decltype(x) __typeof(x) 378 #define ecb_decltype(x) __typeof__ (x)
325#endif 379#endif
326 380
327#if _MSC_VER >= 1300 381#if _MSC_VER >= 1300
328 #define ecb_deprecated __declspec(deprecated) 382 #define ecb_deprecated __declspec (deprecated)
329#else 383#else
330 #define ecb_deprecated ecb_attribute ((__deprecated__)) 384 #define ecb_deprecated ecb_attribute ((__deprecated__))
331#endif 385#endif
332 386
387#if _MSC_VER >= 1500
388 #define ecb_deprecated_message(msg) __declspec (deprecated (msg))
389#elif ECB_GCC_VERSION(4,5)
390 #define ecb_deprecated_message(msg) ecb_attribute ((__deprecated__ (msg))
391#else
392 #define ecb_deprecated_message(msg) ecb_deprecated
393#endif
394
395#if _MSC_VER >= 1400
396 #define ecb_noinline __declspec (noinline)
397#else
333#define ecb_noinline ecb_attribute ((__noinline__)) 398 #define ecb_noinline ecb_attribute ((__noinline__))
399#endif
400
334#define ecb_unused ecb_attribute ((__unused__)) 401#define ecb_unused ecb_attribute ((__unused__))
335#define ecb_const ecb_attribute ((__const__)) 402#define ecb_const ecb_attribute ((__const__))
336#define ecb_pure ecb_attribute ((__pure__)) 403#define ecb_pure ecb_attribute ((__pure__))
337 404
338/* http://msdn.microsoft.com/en-us/library/k6ktzx3s.aspx __declspec(noreturn) */ 405#if ECB_C11 || __IBMC_NORETURN
339#if ECB_C11 406 /* http://www-01.ibm.com/support/knowledgecenter/SSGH3R_13.1.0/com.ibm.xlcpp131.aix.doc/language_ref/noreturn.html */
340 #define ecb_noreturn _Noreturn 407 #define ecb_noreturn _Noreturn
408#elif ECB_CPP11
409 #define ecb_noreturn [[noreturn]]
410#elif _MSC_VER >= 1200
411 /* http://msdn.microsoft.com/en-us/library/k6ktzx3s.aspx */
412 #define ecb_noreturn __declspec (noreturn)
341#else 413#else
342 #define ecb_noreturn ecb_attribute ((__noreturn__)) 414 #define ecb_noreturn ecb_attribute ((__noreturn__))
343#endif 415#endif
344 416
345#if ECB_GCC_VERSION(4,3) 417#if ECB_GCC_VERSION(4,3)
… …
360/* for compatibility to the rest of the world */ 432/* for compatibility to the rest of the world */
361#define ecb_likely(expr) ecb_expect_true (expr) 433#define ecb_likely(expr) ecb_expect_true (expr)
362#define ecb_unlikely(expr) ecb_expect_false (expr) 434#define ecb_unlikely(expr) ecb_expect_false (expr)
363 435
364/* count trailing zero bits and count # of one bits */ 436/* count trailing zero bits and count # of one bits */
365#if ECB_GCC_VERSION(3,4) 437#if ECB_GCC_VERSION(3,4) \
438 || (ECB_CLANG_BUILTIN(__builtin_clz) && ECB_CLANG_BUILTIN(__builtin_clzll) \
439 && ECB_CLANG_BUILTIN(__builtin_ctz) && ECB_CLANG_BUILTIN(__builtin_ctzll) \
440 && ECB_CLANG_BUILTIN(__builtin_popcount))
366 /* we assume int == 32 bit, long == 32 or 64 bit and long long == 64 bit */ 441 /* we assume int == 32 bit, long == 32 or 64 bit and long long == 64 bit */
367 #define ecb_ld32(x) (__builtin_clz (x) ^ 31) 442 #define ecb_ld32(x) (__builtin_clz (x) ^ 31)
368 #define ecb_ld64(x) (__builtin_clzll (x) ^ 63) 443 #define ecb_ld64(x) (__builtin_clzll (x) ^ 63)
369 #define ecb_ctz32(x) __builtin_ctz (x) 444 #define ecb_ctz32(x) __builtin_ctz (x)
370 #define ecb_ctz64(x) __builtin_ctzll (x) 445 #define ecb_ctz64(x) __builtin_ctzll (x)
371 #define ecb_popcount32(x) __builtin_popcount (x) 446 #define ecb_popcount32(x) __builtin_popcount (x)
372 /* no popcountll */ 447 /* no popcountll */
373#else 448#else
374 ecb_function_ int ecb_ctz32 (uint32_t x) ecb_const; 449 ecb_function_ ecb_const int ecb_ctz32 (uint32_t x);
375 ecb_function_ int 450 ecb_function_ ecb_const int
376 ecb_ctz32 (uint32_t x) 451 ecb_ctz32 (uint32_t x)
377 { 452 {
453#if 1400 <= _MSC_VER && (_M_IX86 || _M_X64 || _M_IA64 || _M_ARM)
454 unsigned long r;
455 _BitScanForward (&r, x);
456 return (int)r;
457#else
378 int r = 0; 458 int r = 0;
379 459
380 x &= ~x + 1; /* this isolates the lowest bit */ 460 x &= ~x + 1; /* this isolates the lowest bit */
381 461
382#if ECB_branchless_on_i386 462#if ECB_branchless_on_i386
… …
392 if (x & 0xff00ff00) r += 8; 472 if (x & 0xff00ff00) r += 8;
393 if (x & 0xffff0000) r += 16; 473 if (x & 0xffff0000) r += 16;
394#endif 474#endif
395 475
396 return r; 476 return r;
477#endif
397 } 478 }
398 479
399 ecb_function_ int ecb_ctz64 (uint64_t x) ecb_const; 480 ecb_function_ ecb_const int ecb_ctz64 (uint64_t x);
400 ecb_function_ int 481 ecb_function_ ecb_const int
401 ecb_ctz64 (uint64_t x) 482 ecb_ctz64 (uint64_t x)
402 { 483 {
484#if 1400 <= _MSC_VER && (_M_X64 || _M_IA64 || _M_ARM)
485 unsigned long r;
486 _BitScanForward64 (&r, x);
487 return (int)r;
488#else
403 int shift = x & 0xffffffffU ? 0 : 32; 489 int shift = x & 0xffffffff ? 0 : 32;
404 return ecb_ctz32 (x >> shift) + shift; 490 return ecb_ctz32 (x >> shift) + shift;
491#endif
405 } 492 }
406 493
407 ecb_function_ int ecb_popcount32 (uint32_t x) ecb_const; 494 ecb_function_ ecb_const int ecb_popcount32 (uint32_t x);
408 ecb_function_ int 495 ecb_function_ ecb_const int
409 ecb_popcount32 (uint32_t x) 496 ecb_popcount32 (uint32_t x)
410 { 497 {
411 x -= (x >> 1) & 0x55555555; 498 x -= (x >> 1) & 0x55555555;
412 x = ((x >> 2) & 0x33333333) + (x & 0x33333333); 499 x = ((x >> 2) & 0x33333333) + (x & 0x33333333);
413 x = ((x >> 4) + x) & 0x0f0f0f0f; 500 x = ((x >> 4) + x) & 0x0f0f0f0f;
414 x *= 0x01010101; 501 x *= 0x01010101;
415 502
416 return x >> 24; 503 return x >> 24;
417 } 504 }
418 505
419 ecb_function_ int ecb_ld32 (uint32_t x) ecb_const; 506 ecb_function_ ecb_const int ecb_ld32 (uint32_t x);
420 ecb_function_ int ecb_ld32 (uint32_t x) 507 ecb_function_ ecb_const int ecb_ld32 (uint32_t x)
421 { 508 {
509#if 1400 <= _MSC_VER && (_M_IX86 || _M_X64 || _M_IA64 || _M_ARM)
510 unsigned long r;
511 _BitScanReverse (&r, x);
512 return (int)r;
513#else
422 int r = 0; 514 int r = 0;
423 515
424 if (x >> 16) { x >>= 16; r += 16; } 516 if (x >> 16) { x >>= 16; r += 16; }
425 if (x >> 8) { x >>= 8; r += 8; } 517 if (x >> 8) { x >>= 8; r += 8; }
426 if (x >> 4) { x >>= 4; r += 4; } 518 if (x >> 4) { x >>= 4; r += 4; }
427 if (x >> 2) { x >>= 2; r += 2; } 519 if (x >> 2) { x >>= 2; r += 2; }
428 if (x >> 1) { r += 1; } 520 if (x >> 1) { r += 1; }
429 521
430 return r; 522 return r;
523#endif
431 } 524 }
432 525
433 ecb_function_ int ecb_ld64 (uint64_t x) ecb_const; 526 ecb_function_ ecb_const int ecb_ld64 (uint64_t x);
434 ecb_function_ int ecb_ld64 (uint64_t x) 527 ecb_function_ ecb_const int ecb_ld64 (uint64_t x)
435 { 528 {
529#if 1400 <= _MSC_VER && (_M_X64 || _M_IA64 || _M_ARM)
530 unsigned long r;
531 _BitScanReverse64 (&r, x);
532 return (int)r;
533#else
436 int r = 0; 534 int r = 0;
437 535
438 if (x >> 32) { x >>= 32; r += 32; } 536 if (x >> 32) { x >>= 32; r += 32; }
439 537
440 return r + ecb_ld32 (x); 538 return r + ecb_ld32 (x);
441 }
442#endif 539#endif
540 }
541#endif
443 542
444ecb_function_ ecb_bool ecb_is_pot32 (uint32_t x) ecb_const; 543ecb_function_ ecb_const ecb_bool ecb_is_pot32 (uint32_t x);
445ecb_function_ ecb_bool ecb_is_pot32 (uint32_t x) { return !(x & (x - 1)); } 544ecb_function_ ecb_const ecb_bool ecb_is_pot32 (uint32_t x) { return !(x & (x - 1)); }
446ecb_function_ ecb_bool ecb_is_pot64 (uint64_t x) ecb_const; 545ecb_function_ ecb_const ecb_bool ecb_is_pot64 (uint64_t x);
447ecb_function_ ecb_bool ecb_is_pot64 (uint64_t x) { return !(x & (x - 1)); } 546ecb_function_ ecb_const ecb_bool ecb_is_pot64 (uint64_t x) { return !(x & (x - 1)); }
448 547
449ecb_function_ uint8_t ecb_bitrev8 (uint8_t x) ecb_const; 548ecb_function_ ecb_const uint8_t ecb_bitrev8 (uint8_t x);
450ecb_function_ uint8_t ecb_bitrev8 (uint8_t x) 549ecb_function_ ecb_const uint8_t ecb_bitrev8 (uint8_t x)
451{ 550{
452 return ( (x * 0x0802U & 0x22110U) 551 return ( (x * 0x0802U & 0x22110U)
453 | (x * 0x8020U & 0x88440U)) * 0x10101U >> 16; 552 | (x * 0x8020U & 0x88440U)) * 0x10101U >> 16;
454} 553}
455 554
456ecb_function_ uint16_t ecb_bitrev16 (uint16_t x) ecb_const; 555ecb_function_ ecb_const uint16_t ecb_bitrev16 (uint16_t x);
457ecb_function_ uint16_t ecb_bitrev16 (uint16_t x) 556ecb_function_ ecb_const uint16_t ecb_bitrev16 (uint16_t x)
458{ 557{
459 x = ((x >> 1) & 0x5555) | ((x & 0x5555) << 1); 558 x = ((x >> 1) & 0x5555) | ((x & 0x5555) << 1);
460 x = ((x >> 2) & 0x3333) | ((x & 0x3333) << 2); 559 x = ((x >> 2) & 0x3333) | ((x & 0x3333) << 2);
461 x = ((x >> 4) & 0x0f0f) | ((x & 0x0f0f) << 4); 560 x = ((x >> 4) & 0x0f0f) | ((x & 0x0f0f) << 4);
462 x = ( x >> 8 ) | ( x << 8); 561 x = ( x >> 8 ) | ( x << 8);
463 562
464 return x; 563 return x;
465} 564}
466 565
467ecb_function_ uint32_t ecb_bitrev32 (uint32_t x) ecb_const; 566ecb_function_ ecb_const uint32_t ecb_bitrev32 (uint32_t x);
468ecb_function_ uint32_t ecb_bitrev32 (uint32_t x) 567ecb_function_ ecb_const uint32_t ecb_bitrev32 (uint32_t x)
469{ 568{
470 x = ((x >> 1) & 0x55555555) | ((x & 0x55555555) << 1); 569 x = ((x >> 1) & 0x55555555) | ((x & 0x55555555) << 1);
471 x = ((x >> 2) & 0x33333333) | ((x & 0x33333333) << 2); 570 x = ((x >> 2) & 0x33333333) | ((x & 0x33333333) << 2);
472 x = ((x >> 4) & 0x0f0f0f0f) | ((x & 0x0f0f0f0f) << 4); 571 x = ((x >> 4) & 0x0f0f0f0f) | ((x & 0x0f0f0f0f) << 4);
473 x = ((x >> 8) & 0x00ff00ff) | ((x & 0x00ff00ff) << 8); 572 x = ((x >> 8) & 0x00ff00ff) | ((x & 0x00ff00ff) << 8);
… …
476 return x; 575 return x;
477} 576}
478 577
479/* popcount64 is only available on 64 bit cpus as gcc builtin */ 578/* popcount64 is only available on 64 bit cpus as gcc builtin */
480/* so for this version we are lazy */ 579/* so for this version we are lazy */
481ecb_function_ int ecb_popcount64 (uint64_t x) ecb_const; 580ecb_function_ ecb_const int ecb_popcount64 (uint64_t x);
482ecb_function_ int 581ecb_function_ ecb_const int
483ecb_popcount64 (uint64_t x) 582ecb_popcount64 (uint64_t x)
484{ 583{
485 return ecb_popcount32 (x) + ecb_popcount32 (x >> 32); 584 return ecb_popcount32 (x) + ecb_popcount32 (x >> 32);
486} 585}
487 586
488ecb_inline uint8_t ecb_rotl8 (uint8_t x, unsigned int count) ecb_const; 587ecb_inline ecb_const uint8_t ecb_rotl8 (uint8_t x, unsigned int count);
489ecb_inline uint8_t ecb_rotr8 (uint8_t x, unsigned int count) ecb_const; 588ecb_inline ecb_const uint8_t ecb_rotr8 (uint8_t x, unsigned int count);
490ecb_inline uint16_t ecb_rotl16 (uint16_t x, unsigned int count) ecb_const; 589ecb_inline ecb_const uint16_t ecb_rotl16 (uint16_t x, unsigned int count);
491ecb_inline uint16_t ecb_rotr16 (uint16_t x, unsigned int count) ecb_const; 590ecb_inline ecb_const uint16_t ecb_rotr16 (uint16_t x, unsigned int count);
492ecb_inline uint32_t ecb_rotl32 (uint32_t x, unsigned int count) ecb_const; 591ecb_inline ecb_const uint32_t ecb_rotl32 (uint32_t x, unsigned int count);
493ecb_inline uint32_t ecb_rotr32 (uint32_t x, unsigned int count) ecb_const; 592ecb_inline ecb_const uint32_t ecb_rotr32 (uint32_t x, unsigned int count);
494ecb_inline uint64_t ecb_rotl64 (uint64_t x, unsigned int count) ecb_const; 593ecb_inline ecb_const uint64_t ecb_rotl64 (uint64_t x, unsigned int count);
495ecb_inline uint64_t ecb_rotr64 (uint64_t x, unsigned int count) ecb_const; 594ecb_inline ecb_const uint64_t ecb_rotr64 (uint64_t x, unsigned int count);
496 595
497ecb_inline uint8_t ecb_rotl8 (uint8_t x, unsigned int count) { return (x >> ( 8 - count)) | (x << count); } 596ecb_inline ecb_const uint8_t ecb_rotl8 (uint8_t x, unsigned int count) { return (x >> ( 8 - count)) | (x << count); }
498ecb_inline uint8_t ecb_rotr8 (uint8_t x, unsigned int count) { return (x << ( 8 - count)) | (x >> count); } 597ecb_inline ecb_const uint8_t ecb_rotr8 (uint8_t x, unsigned int count) { return (x << ( 8 - count)) | (x >> count); }
499ecb_inline uint16_t ecb_rotl16 (uint16_t x, unsigned int count) { return (x >> (16 - count)) | (x << count); } 598ecb_inline ecb_const uint16_t ecb_rotl16 (uint16_t x, unsigned int count) { return (x >> (16 - count)) | (x << count); }
500ecb_inline uint16_t ecb_rotr16 (uint16_t x, unsigned int count) { return (x << (16 - count)) | (x >> count); } 599ecb_inline ecb_const uint16_t ecb_rotr16 (uint16_t x, unsigned int count) { return (x << (16 - count)) | (x >> count); }
501ecb_inline uint32_t ecb_rotl32 (uint32_t x, unsigned int count) { return (x >> (32 - count)) | (x << count); } 600ecb_inline ecb_const uint32_t ecb_rotl32 (uint32_t x, unsigned int count) { return (x >> (32 - count)) | (x << count); }
502ecb_inline uint32_t ecb_rotr32 (uint32_t x, unsigned int count) { return (x << (32 - count)) | (x >> count); } 601ecb_inline ecb_const uint32_t ecb_rotr32 (uint32_t x, unsigned int count) { return (x << (32 - count)) | (x >> count); }
503ecb_inline uint64_t ecb_rotl64 (uint64_t x, unsigned int count) { return (x >> (64 - count)) | (x << count); } 602ecb_inline ecb_const uint64_t ecb_rotl64 (uint64_t x, unsigned int count) { return (x >> (64 - count)) | (x << count); }
504ecb_inline uint64_t ecb_rotr64 (uint64_t x, unsigned int count) { return (x << (64 - count)) | (x >> count); } 603ecb_inline ecb_const uint64_t ecb_rotr64 (uint64_t x, unsigned int count) { return (x << (64 - count)) | (x >> count); }
505 604
506#if ECB_GCC_VERSION(4,3) 605#if ECB_GCC_VERSION(4,3) || (ECB_CLANG_BUILTIN(__builtin_bswap32) && ECB_CLANG_BUILTIN(__builtin_bswap64))
606 #if ECB_GCC_VERSION(4,8) || ECB_CLANG_BUILTIN(__builtin_bswap16)
607 #define ecb_bswap16(x) __builtin_bswap16 (x)
608 #else
507 #define ecb_bswap16(x) (__builtin_bswap32 (x) >> 16) 609 #define ecb_bswap16(x) (__builtin_bswap32 (x) >> 16)
610 #endif
508 #define ecb_bswap32(x) __builtin_bswap32 (x) 611 #define ecb_bswap32(x) __builtin_bswap32 (x)
509 #define ecb_bswap64(x) __builtin_bswap64 (x) 612 #define ecb_bswap64(x) __builtin_bswap64 (x)
613#elif _MSC_VER
614 #include <stdlib.h>
615 #define ecb_bswap16(x) ((uint16_t)_byteswap_ushort ((uint16_t)(x)))
616 #define ecb_bswap32(x) ((uint32_t)_byteswap_ulong ((uint32_t)(x)))
617 #define ecb_bswap64(x) ((uint64_t)_byteswap_uint64 ((uint64_t)(x)))
510#else 618#else
511 ecb_function_ uint16_t ecb_bswap16 (uint16_t x) ecb_const; 619 ecb_function_ ecb_const uint16_t ecb_bswap16 (uint16_t x);
512 ecb_function_ uint16_t 620 ecb_function_ ecb_const uint16_t
513 ecb_bswap16 (uint16_t x) 621 ecb_bswap16 (uint16_t x)
514 { 622 {
515 return ecb_rotl16 (x, 8); 623 return ecb_rotl16 (x, 8);
516 } 624 }
517 625
518 ecb_function_ uint32_t ecb_bswap32 (uint32_t x) ecb_const; 626 ecb_function_ ecb_const uint32_t ecb_bswap32 (uint32_t x);
519 ecb_function_ uint32_t 627 ecb_function_ ecb_const uint32_t
520 ecb_bswap32 (uint32_t x) 628 ecb_bswap32 (uint32_t x)
521 { 629 {
522 return (((uint32_t)ecb_bswap16 (x)) << 16) | ecb_bswap16 (x >> 16); 630 return (((uint32_t)ecb_bswap16 (x)) << 16) | ecb_bswap16 (x >> 16);
523 } 631 }
524 632
525 ecb_function_ uint64_t ecb_bswap64 (uint64_t x) ecb_const; 633 ecb_function_ ecb_const uint64_t ecb_bswap64 (uint64_t x);
526 ecb_function_ uint64_t 634 ecb_function_ ecb_const uint64_t
527 ecb_bswap64 (uint64_t x) 635 ecb_bswap64 (uint64_t x)
528 { 636 {
529 return (((uint64_t)ecb_bswap32 (x)) << 32) | ecb_bswap32 (x >> 32); 637 return (((uint64_t)ecb_bswap32 (x)) << 32) | ecb_bswap32 (x >> 32);
530 } 638 }
531#endif 639#endif
532 640
533#if ECB_GCC_VERSION(4,5) 641#if ECB_GCC_VERSION(4,5) || ECB_CLANG_BUILTIN(__builtin_unreachable)
534 #define ecb_unreachable() __builtin_unreachable () 642 #define ecb_unreachable() __builtin_unreachable ()
535#else 643#else
536 /* this seems to work fine, but gcc always emits a warning for it :/ */ 644 /* this seems to work fine, but gcc always emits a warning for it :/ */
537 ecb_inline void ecb_unreachable (void) ecb_noreturn; 645 ecb_inline ecb_noreturn void ecb_unreachable (void);
538 ecb_inline void ecb_unreachable (void) { } 646 ecb_inline ecb_noreturn void ecb_unreachable (void) { }
539#endif 647#endif
540 648
541/* try to tell the compiler that some condition is definitely true */ 649/* try to tell the compiler that some condition is definitely true */
542#define ecb_assume(cond) if (!(cond)) ecb_unreachable (); else 0 650#define ecb_assume(cond) if (!(cond)) ecb_unreachable (); else 0
543 651
544ecb_inline unsigned char ecb_byteorder_helper (void) ecb_const; 652ecb_inline ecb_const uint32_t ecb_byteorder_helper (void);
545ecb_inline unsigned char 653ecb_inline ecb_const uint32_t
546ecb_byteorder_helper (void) 654ecb_byteorder_helper (void)
547{ 655{
548 /* the union code still generates code under pressure in gcc, */ 656 /* the union code still generates code under pressure in gcc, */
549 /* but less than using pointers, and always seems to */ 657 /* but less than using pointers, and always seems to */
550 /* successfully return a constant. */ 658 /* successfully return a constant. */
551 /* the reason why we have this horrible preprocessor mess */ 659 /* the reason why we have this horrible preprocessor mess */
552 /* is to avoid it in all cases, at least on common architectures */ 660 /* is to avoid it in all cases, at least on common architectures */
553 /* or when using a recent enough gcc version (>= 4.6) */ 661 /* or when using a recent enough gcc version (>= 4.6) */
554#if __i386 || __i386__ || _M_X86 || __amd64 || __amd64__ || _M_X64
555 return 0x44;
556#elif __BYTE_ORDER__ && __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__ 662#if (defined __BYTE_ORDER__ && __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__) \
663 || ((__i386 || __i386__ || _M_IX86 || ECB_GCC_AMD64 || ECB_MSVC_AMD64) && !__VOS__)
664 #define ECB_LITTLE_ENDIAN 1
557 return 0x44; 665 return 0x44332211;
558#elif __BYTE_ORDER__ && __BYTE_ORDER__ == __ORDER_BIG_ENDIAN__ 666#elif (defined __BYTE_ORDER__ && __BYTE_ORDER__ == __ORDER_BIG_ENDIAN__) \
667 || ((__AARCH64EB__ || __MIPSEB__ || __ARMEB__) && !__VOS__)
668 #define ECB_BIG_ENDIAN 1
559 return 0x11; 669 return 0x11223344;
560#else 670#else
561 union 671 union
562 { 672 {
673 uint8_t c[4];
563 uint32_t i; 674 uint32_t u;
564 uint8_t c;
565 } u = { 0x11223344 }; 675 } u = { 0x11, 0x22, 0x33, 0x44 };
566 return u.c; 676 return u.u;
567#endif 677#endif
568} 678}
569 679
570ecb_inline ecb_bool ecb_big_endian (void) ecb_const; 680ecb_inline ecb_const ecb_bool ecb_big_endian (void);
571ecb_inline ecb_bool ecb_big_endian (void) { return ecb_byteorder_helper () == 0x11; } 681ecb_inline ecb_const ecb_bool ecb_big_endian (void) { return ecb_byteorder_helper () == 0x11223344; }
572ecb_inline ecb_bool ecb_little_endian (void) ecb_const; 682ecb_inline ecb_const ecb_bool ecb_little_endian (void);
573ecb_inline ecb_bool ecb_little_endian (void) { return ecb_byteorder_helper () == 0x44; } 683ecb_inline ecb_const ecb_bool ecb_little_endian (void) { return ecb_byteorder_helper () == 0x44332211; }
574 684
575#if ECB_GCC_VERSION(3,0) || ECB_C99 685#if ECB_GCC_VERSION(3,0) || ECB_C99
576 #define ecb_mod(m,n) ((m) % (n) + ((m) % (n) < 0 ? (n) : 0)) 686 #define ecb_mod(m,n) ((m) % (n) + ((m) % (n) < 0 ? (n) : 0))
577#else 687#else
578 #define ecb_mod(m,n) ((m) < 0 ? ((n) - 1 - ((-1 - (m)) % (n))) : ((m) % (n))) 688 #define ecb_mod(m,n) ((m) < 0 ? ((n) - 1 - ((-1 - (m)) % (n))) : ((m) % (n)))
579#endif 689#endif
580 690
581#if __cplusplus 691#if ECB_CPP
582 template<typename T> 692 template<typename T>
583 static inline T ecb_div_rd (T val, T div) 693 static inline T ecb_div_rd (T val, T div)
584 { 694 {
585 return val < 0 ? - ((-val + div - 1) / div) : (val ) / div; 695 return val < 0 ? - ((-val + div - 1) / div) : (val ) / div;
586 } 696 }
… …
603 } 713 }
604#else 714#else
605 #define ecb_array_length(name) (sizeof (name) / sizeof (name [0])) 715 #define ecb_array_length(name) (sizeof (name) / sizeof (name [0]))
606#endif 716#endif
607 717
718ecb_function_ ecb_const uint32_t ecb_binary16_to_binary32 (uint32_t x);
719ecb_function_ ecb_const uint32_t
720ecb_binary16_to_binary32 (uint32_t x)
721{
722 unsigned int s = (x & 0x8000) << (31 - 15);
723 int e = (x >> 10) & 0x001f;
724 unsigned int m = x & 0x03ff;
725
726 if (ecb_expect_false (e == 31))
727 /* infinity or NaN */
728 e = 255 - (127 - 15);
729 else if (ecb_expect_false (!e))
730 {
731 if (ecb_expect_true (!m))
732 /* zero, handled by code below by forcing e to 0 */
733 e = 0 - (127 - 15);
734 else
735 {
736 /* subnormal, renormalise */
737 unsigned int s = 10 - ecb_ld32 (m);
738
739 m = (m << s) & 0x3ff; /* mask implicit bit */
740 e -= s - 1;
741 }
742 }
743
744 /* e and m now are normalised, or zero, (or inf or nan) */
745 e += 127 - 15;
746
747 return s | (e << 23) | (m << (23 - 10));
748}
749
750ecb_function_ ecb_const uint16_t ecb_binary32_to_binary16 (uint32_t x);
751ecb_function_ ecb_const uint16_t
752ecb_binary32_to_binary16 (uint32_t x)
753{
754 unsigned int s = (x >> 16) & 0x00008000; /* sign bit, the easy part */
755 unsigned int e = ((x >> 23) & 0x000000ff) - (127 - 15); /* the desired exponent */
756 unsigned int m = x & 0x007fffff;
757
758 x &= 0x7fffffff;
759
760 /* if it's within range of binary16 normals, use fast path */
761 if (ecb_expect_true (0x38800000 <= x && x <= 0x477fefff))
762 {
763 /* mantissa round-to-even */
764 m += 0x00000fff + ((m >> (23 - 10)) & 1);
765
766 /* handle overflow */
767 if (ecb_expect_false (m >= 0x00800000))
768 {
769 m >>= 1;
770 e += 1;
771 }
772
773 return s | (e << 10) | (m >> (23 - 10));
774 }
775
776 /* handle large numbers and infinity */
777 if (ecb_expect_true (0x477fefff < x && x <= 0x7f800000))
778 return s | 0x7c00;
779
780 /* handle zero, subnormals and small numbers */
781 if (ecb_expect_true (x < 0x38800000))
782 {
783 /* zero */
784 if (ecb_expect_true (!x))
785 return s;
786
787 /* handle subnormals */
788
789 /* too small, will be zero */
790 if (e < (14 - 24)) /* might not be sharp, but is good enough */
791 return s;
792
793 m |= 0x00800000; /* make implicit bit explicit */
794
795 /* very tricky - we need to round to the nearest e (+10) bit value */
796 {
797 unsigned int bits = 14 - e;
798 unsigned int half = (1 << (bits - 1)) - 1;
799 unsigned int even = (m >> bits) & 1;
800
801 /* if this overflows, we will end up with a normalised number */
802 m = (m + half + even) >> bits;
803 }
804
805 return s | m;
806 }
807
808 /* handle NaNs, preserve leftmost nan bits, but make sure we don't turn them into infinities */
809 m >>= 13;
810
811 return s | 0x7c00 | m | !m;
812}
813
608/*******************************************************************************/ 814/*******************************************************************************/
609/* floating point stuff, can be disabled by defining ECB_NO_LIBM */ 815/* floating point stuff, can be disabled by defining ECB_NO_LIBM */
610 816
611/* basically, everything uses "ieee pure-endian" floating point numbers */ 817/* basically, everything uses "ieee pure-endian" floating point numbers */
612/* the only noteworthy exception is ancient armle, which uses order 43218765 */ 818/* the only noteworthy exception is ancient armle, which uses order 43218765 */
613#if 0 \ 819#if 0 \
614 || __i386 || __i386__ \ 820 || __i386 || __i386__ \
615 || __amd64 || __amd64__ || __x86_64 || __x86_64__ \ 821 || ECB_GCC_AMD64 \
616 || __powerpc__ || __ppc__ || __powerpc64__ || __ppc64__ \ 822 || __powerpc__ || __ppc__ || __powerpc64__ || __ppc64__ \
617 || defined __s390__ || defined __s390x__ \ 823 || defined __s390__ || defined __s390x__ \
618 || defined __mips__ \ 824 || defined __mips__ \
619 || defined __alpha__ \ 825 || defined __alpha__ \
620 || defined __hppa__ \ 826 || defined __hppa__ \
621 || defined __ia64__ \ 827 || defined __ia64__ \
622 || defined __m68k__ \ 828 || defined __m68k__ \
623 || defined __m88k__ \ 829 || defined __m88k__ \
624 || defined __sh__ \ 830 || defined __sh__ \
625 || defined _M_IX86 || defined _M_AMD64 || defined _M_IA64 \ 831 || defined _M_IX86 || defined ECB_MSVC_AMD64 || defined _M_IA64 \
626 || (defined __arm__ && (defined __ARM_EABI__ || defined __EABI__ || defined __VFP_FP__ || defined _WIN32_WCE || defined __ANDROID__)) \ 832 || (defined __arm__ && (defined __ARM_EABI__ || defined __EABI__ || defined __VFP_FP__ || defined _WIN32_WCE || defined __ANDROID__)) \
627 || defined __aarch64__ 833 || defined __aarch64__
628 #define ECB_STDFP 1 834 #define ECB_STDFP 1
629 #include <string.h> /* for memcpy */ 835 #include <string.h> /* for memcpy */
630#else 836#else
… …
646 #define ECB_NAN NAN 852 #define ECB_NAN NAN
647 #else 853 #else
648 #define ECB_NAN ECB_INFINITY 854 #define ECB_NAN ECB_INFINITY
649 #endif 855 #endif
650 856
651 /* converts an ieee half/binary16 to a float */ 857 #if ECB_C99 || _XOPEN_VERSION >= 600 || _POSIX_VERSION >= 200112L
652 ecb_function_ float ecb_binary16_to_float (uint16_t x) ecb_const; 858 #define ecb_ldexpf(x,e) ldexpf ((x), (e))
653 ecb_function_ float 859 #define ecb_frexpf(x,e) frexpf ((x), (e))
654 ecb_binary16_to_float (uint16_t x) 860 #else
655 { 861 #define ecb_ldexpf(x,e) (float) ldexp ((double) (x), (e))
656 int e = (x >> 10) & 0x1f; 862 #define ecb_frexpf(x,e) (float) frexp ((double) (x), (e))
657 int m = x & 0x3ff; 863 #endif
658 float r;
659
660 if (!e ) r = ldexpf (m , -24);
661 else if (e != 31) r = ldexpf (m + 0x400, e - 25);
662 else if (m ) r = ECB_NAN;
663 else r = ECB_INFINITY;
664
665 return x & 0x8000 ? -r : r;
666 }
667 864
668 /* convert a float to ieee single/binary32 */ 865 /* convert a float to ieee single/binary32 */
669 ecb_function_ uint32_t ecb_float_to_binary32 (float x) ecb_const; 866 ecb_function_ ecb_const uint32_t ecb_float_to_binary32 (float x);
670 ecb_function_ uint32_t 867 ecb_function_ ecb_const uint32_t
671 ecb_float_to_binary32 (float x) 868 ecb_float_to_binary32 (float x)
672 { 869 {
673 uint32_t r; 870 uint32_t r;
674 871
675 #if ECB_STDFP 872 #if ECB_STDFP
… …
682 if (x == 0e0f ) return 0x00000000U; 879 if (x == 0e0f ) return 0x00000000U;
683 if (x > +3.40282346638528860e+38f) return 0x7f800000U; 880 if (x > +3.40282346638528860e+38f) return 0x7f800000U;
684 if (x < -3.40282346638528860e+38f) return 0xff800000U; 881 if (x < -3.40282346638528860e+38f) return 0xff800000U;
685 if (x != x ) return 0x7fbfffffU; 882 if (x != x ) return 0x7fbfffffU;
686 883
687 m = frexpf (x, &e) * 0x1000000U; 884 m = ecb_frexpf (x, &e) * 0x1000000U;
688 885
689 r = m & 0x80000000U; 886 r = m & 0x80000000U;
690 887
691 if (r) 888 if (r)
692 m = -m; 889 m = -m;
… …
704 901
705 return r; 902 return r;
706 } 903 }
707 904
708 /* converts an ieee single/binary32 to a float */ 905 /* converts an ieee single/binary32 to a float */
709 ecb_function_ float ecb_binary32_to_float (uint32_t x) ecb_const; 906 ecb_function_ ecb_const float ecb_binary32_to_float (uint32_t x);
710 ecb_function_ float 907 ecb_function_ ecb_const float
711 ecb_binary32_to_float (uint32_t x) 908 ecb_binary32_to_float (uint32_t x)
712 { 909 {
713 float r; 910 float r;
714 911
715 #if ECB_STDFP 912 #if ECB_STDFP
… …
725 x |= 0x800000U; 922 x |= 0x800000U;
726 else 923 else
727 e = 1; 924 e = 1;
728 925
729 /* we distrust ldexpf a bit and do the 2**-24 scaling by an extra multiply */ 926 /* we distrust ldexpf a bit and do the 2**-24 scaling by an extra multiply */
730 r = ldexpf (x * (0.5f / 0x800000U), e - 126); 927 r = ecb_ldexpf (x * (0.5f / 0x800000U), e - 126);
731 928
732 r = neg ? -r : r; 929 r = neg ? -r : r;
733 #endif 930 #endif
734 931
735 return r; 932 return r;
736 } 933 }
737 934
738 /* convert a double to ieee double/binary64 */ 935 /* convert a double to ieee double/binary64 */
739 ecb_function_ uint64_t ecb_double_to_binary64 (double x) ecb_const; 936 ecb_function_ ecb_const uint64_t ecb_double_to_binary64 (double x);
740 ecb_function_ uint64_t 937 ecb_function_ ecb_const uint64_t
741 ecb_double_to_binary64 (double x) 938 ecb_double_to_binary64 (double x)
742 { 939 {
743 uint64_t r; 940 uint64_t r;
744 941
745 #if ECB_STDFP 942 #if ECB_STDFP
… …
774 971
775 return r; 972 return r;
776 } 973 }
777 974
778 /* converts an ieee double/binary64 to a double */ 975 /* converts an ieee double/binary64 to a double */
779 ecb_function_ double ecb_binary64_to_double (uint64_t x) ecb_const; 976 ecb_function_ ecb_const double ecb_binary64_to_double (uint64_t x);
780 ecb_function_ double 977 ecb_function_ ecb_const double
781 ecb_binary64_to_double (uint64_t x) 978 ecb_binary64_to_double (uint64_t x)
782 { 979 {
783 double r; 980 double r;
784 981
785 #if ECB_STDFP 982 #if ECB_STDFP
… …
803 #endif 1000 #endif
804 1001
805 return r; 1002 return r;
806 } 1003 }
807 1004
808#endif 1005 /* convert a float to ieee half/binary16 */
1006 ecb_function_ ecb_const uint16_t ecb_float_to_binary16 (float x);
1007 ecb_function_ ecb_const uint16_t
1008 ecb_float_to_binary16 (float x)
1009 {
1010 return ecb_binary32_to_binary16 (ecb_float_to_binary32 (x));
1011 }
809 1012
810#endif 1013 /* convert an ieee half/binary16 to float */
1014 ecb_function_ ecb_const float ecb_binary16_to_float (uint16_t x);
1015 ecb_function_ ecb_const float
1016 ecb_binary16_to_float (uint16_t x)
1017 {
1018 return ecb_binary32_to_float (ecb_binary16_to_binary32 (x));
1019 }
811 1020
1021#endif
1022
1023#endif
1024

Diff Legend

– Removed lines
+ Added lines
< Changed lines
> Changed lines