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

Comparing libeio/ecb.h (file contents):
Revision 1.21 by root, Fri May 16 15:15:23 2014 UTC vs.
Revision 1.29 by root, Fri Aug 23 07:20:21 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 0x00010006
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
79/* work around x32 idiocy by defining proper macros */ 82/* work around x32 idiocy by defining proper macros */
80#if __amd64 || __x86_64 || _M_AMD64 || _M_X64 83#if ECB_GCC_AMD64 || ECB_MSVC_AMD64
81 #if _ILP32 84 #if _ILP32
82 #define ECB_AMD64_X32 1 85 #define ECB_AMD64_X32 1
83 #else 86 #else
84 #define ECB_AMD64 1 87 #define ECB_AMD64 1
85 #endif 88 #endif
… …
90 * causing enormous grief in return for some better fake benchmark numbers. 93 * causing enormous grief in return for some better fake benchmark numbers.
91 * or so. 94 * or so.
92 * we try to detect these and simply assume they are not gcc - if they have 95 * 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. 96 * an issue with that they should have done it right in the first place.
94 */ 97 */
95#ifndef ECB_GCC_VERSION
96 #if !defined __GNUC_MINOR__ || defined __INTEL_COMPILER || defined __SUNPRO_C || defined __SUNPRO_CC || defined __llvm__ || defined __clang__ 98#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 99 #define ECB_GCC_VERSION(major,minor) 0
98 #else 100#else
99 #define ECB_GCC_VERSION(major,minor) (__GNUC__ > (major) || (__GNUC__ == (major) && __GNUC_MINOR__ >= (minor))) 101 #define ECB_GCC_VERSION(major,minor) (__GNUC__ > (major) || (__GNUC__ == (major) && __GNUC_MINOR__ >= (minor)))
100 #endif 102#endif
103
104#define ECB_CLANG_VERSION(major,minor) (__clang_major__ > (major) || (__clang_major__ == (major) && __clang_minor__ >= (minor)))
105
106#if __clang__ && defined __has_builtin
107 #define ECB_CLANG_BUILTIN(x) __has_builtin (x)
108#else
109 #define ECB_CLANG_BUILTIN(x) 0
110#endif
111
112#if __clang__ && defined __has_extension
113 #define ECB_CLANG_EXTENSION(x) __has_extension (x)
114#else
115 #define ECB_CLANG_EXTENSION(x) 0
101#endif 116#endif
102 117
103#define ECB_CPP (__cplusplus+0) 118#define ECB_CPP (__cplusplus+0)
104#define ECB_CPP11 (__cplusplus >= 201103L) 119#define ECB_CPP11 (__cplusplus >= 201103L)
120#define ECB_CPP14 (__cplusplus >= 201402L)
121#define ECB_CPP17 (__cplusplus >= 201703L)
105 122
106#if ECB_CPP 123#if ECB_CPP
107 #define ECB_C 0 124 #define ECB_C 0
108 #define ECB_STDC_VERSION 0 125 #define ECB_STDC_VERSION 0
109#else 126#else
… …
111 #define ECB_STDC_VERSION __STDC_VERSION__ 128 #define ECB_STDC_VERSION __STDC_VERSION__
112#endif 129#endif
113 130
114#define ECB_C99 (ECB_STDC_VERSION >= 199901L) 131#define ECB_C99 (ECB_STDC_VERSION >= 199901L)
115#define ECB_C11 (ECB_STDC_VERSION >= 201112L) 132#define ECB_C11 (ECB_STDC_VERSION >= 201112L)
133#define ECB_C17 (ECB_STDC_VERSION >= 201710L)
116 134
117#if ECB_CPP 135#if ECB_CPP
118 #define ECB_EXTERN_C extern "C" 136 #define ECB_EXTERN_C extern "C"
119 #define ECB_EXTERN_C_BEG ECB_EXTERN_C { 137 #define ECB_EXTERN_C_BEG ECB_EXTERN_C {
120 #define ECB_EXTERN_C_END } 138 #define ECB_EXTERN_C_END }
… …
135 153
136#if ECB_NO_SMP 154#if ECB_NO_SMP
137 #define ECB_MEMORY_FENCE do { } while (0) 155 #define ECB_MEMORY_FENCE do { } while (0)
138#endif 156#endif
139 157
158/* http://www-01.ibm.com/support/knowledgecenter/SSGH3R_13.1.0/com.ibm.xlcpp131.aix.doc/compiler_ref/compiler_builtins.html */
159#if __xlC__ && ECB_CPP
160 #include <builtins.h>
161#endif
162
163#if 1400 <= _MSC_VER
164 #include <intrin.h> /* fence functions _ReadBarrier, also bit search functions _BitScanReverse */
165#endif
166
140#ifndef ECB_MEMORY_FENCE 167#ifndef ECB_MEMORY_FENCE
141 #if ECB_GCC_VERSION(2,5) || defined __INTEL_COMPILER || (__llvm__ && __GNUC__) || __SUNPRO_C >= 0x5110 || __SUNPRO_CC >= 0x5110 168 #if ECB_GCC_VERSION(2,5) || defined __INTEL_COMPILER || (__llvm__ && __GNUC__) || __SUNPRO_C >= 0x5110 || __SUNPRO_CC >= 0x5110
169 #define ECB_MEMORY_FENCE_RELAXED __asm__ __volatile__ ("" : : : "memory")
142 #if __i386 || __i386__ 170 #if __i386 || __i386__
143 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("lock; orb $0, -1(%%esp)" : : : "memory") 171 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("lock; orb $0, -1(%%esp)" : : : "memory")
144 #define ECB_MEMORY_FENCE_ACQUIRE __asm__ __volatile__ ("" : : : "memory") 172 #define ECB_MEMORY_FENCE_ACQUIRE __asm__ __volatile__ ("" : : : "memory")
145 #define ECB_MEMORY_FENCE_RELEASE __asm__ __volatile__ ("") 173 #define ECB_MEMORY_FENCE_RELEASE __asm__ __volatile__ ("" : : : "memory")
146 #elif __amd64 || __amd64__ || __x86_64 || __x86_64__ 174 #elif ECB_GCC_AMD64
147 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("mfence" : : : "memory") 175 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("mfence" : : : "memory")
148 #define ECB_MEMORY_FENCE_ACQUIRE __asm__ __volatile__ ("" : : : "memory") 176 #define ECB_MEMORY_FENCE_ACQUIRE __asm__ __volatile__ ("" : : : "memory")
149 #define ECB_MEMORY_FENCE_RELEASE __asm__ __volatile__ ("") 177 #define ECB_MEMORY_FENCE_RELEASE __asm__ __volatile__ ("" : : : "memory")
150 #elif __powerpc__ || __ppc__ || __powerpc64__ || __ppc64__ 178 #elif __powerpc__ || __ppc__ || __powerpc64__ || __ppc64__
151 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("sync" : : : "memory") 179 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("sync" : : : "memory")
180 #elif defined __ARM_ARCH_2__ \
181 || defined __ARM_ARCH_3__ || defined __ARM_ARCH_3M__ \
182 || defined __ARM_ARCH_4__ || defined __ARM_ARCH_4T__ \
183 || defined __ARM_ARCH_5__ || defined __ARM_ARCH_5E__ \
184 || defined __ARM_ARCH_5T__ || defined __ARM_ARCH_5TE__ \
185 || defined __ARM_ARCH_5TEJ__
186 /* 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__ \ 187 #elif defined __ARM_ARCH_6__ || defined __ARM_ARCH_6J__ \
153 || defined __ARM_ARCH_6K__ || defined __ARM_ARCH_6ZK__ 188 || defined __ARM_ARCH_6K__ || defined __ARM_ARCH_6ZK__ \
189 || defined __ARM_ARCH_6T2__
154 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("mcr p15,0,%0,c7,c10,5" : : "r" (0) : "memory") 190 #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__ \ 191 #elif defined __ARM_ARCH_7__ || defined __ARM_ARCH_7A__ \
156 || defined __ARM_ARCH_7M__ || defined __ARM_ARCH_7R__ 192 || defined __ARM_ARCH_7R__ || defined __ARM_ARCH_7M__
157 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("dmb" : : : "memory") 193 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("dmb" : : : "memory")
158 #elif __aarch64__ 194 #elif __aarch64__
159 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("dmb ish" : : : "memory") 195 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("dmb ish" : : : "memory")
160 #elif (__sparc || __sparc__) && !__sparcv8 196 #elif (__sparc || __sparc__) && !(__sparc_v8__ || defined __sparcv8)
161 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("membar #LoadStore | #LoadLoad | #StoreStore | #StoreLoad" : : : "memory") 197 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("membar #LoadStore | #LoadLoad | #StoreStore | #StoreLoad" : : : "memory")
162 #define ECB_MEMORY_FENCE_ACQUIRE __asm__ __volatile__ ("membar #LoadStore | #LoadLoad" : : : "memory") 198 #define ECB_MEMORY_FENCE_ACQUIRE __asm__ __volatile__ ("membar #LoadStore | #LoadLoad" : : : "memory")
163 #define ECB_MEMORY_FENCE_RELEASE __asm__ __volatile__ ("membar #LoadStore | #StoreStore") 199 #define ECB_MEMORY_FENCE_RELEASE __asm__ __volatile__ ("membar #LoadStore | #StoreStore")
164 #elif defined __s390__ || defined __s390x__ 200 #elif defined __s390__ || defined __s390x__
165 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("bcr 15,0" : : : "memory") 201 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("bcr 15,0" : : : "memory")
… …
188 #if ECB_GCC_VERSION(4,7) 224 #if ECB_GCC_VERSION(4,7)
189 /* see comment below (stdatomic.h) about the C11 memory model. */ 225 /* see comment below (stdatomic.h) about the C11 memory model. */
190 #define ECB_MEMORY_FENCE __atomic_thread_fence (__ATOMIC_SEQ_CST) 226 #define ECB_MEMORY_FENCE __atomic_thread_fence (__ATOMIC_SEQ_CST)
191 #define ECB_MEMORY_FENCE_ACQUIRE __atomic_thread_fence (__ATOMIC_ACQUIRE) 227 #define ECB_MEMORY_FENCE_ACQUIRE __atomic_thread_fence (__ATOMIC_ACQUIRE)
192 #define ECB_MEMORY_FENCE_RELEASE __atomic_thread_fence (__ATOMIC_RELEASE) 228 #define ECB_MEMORY_FENCE_RELEASE __atomic_thread_fence (__ATOMIC_RELEASE)
229 #define ECB_MEMORY_FENCE_RELAXED __atomic_thread_fence (__ATOMIC_RELAXED)
193 230
194 /* The __has_feature syntax from clang is so misdesigned that we cannot use it 231 #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. 232 /* see comment below (stdatomic.h) about the C11 memory model. */
200 * #define ECB_MEMORY_FENCE __c11_atomic_thread_fence (__ATOMIC_SEQ_CST) 233 #define ECB_MEMORY_FENCE __c11_atomic_thread_fence (__ATOMIC_SEQ_CST)
201 * #define ECB_MEMORY_FENCE_ACQUIRE __c11_atomic_thread_fence (__ATOMIC_ACQUIRE) 234 #define ECB_MEMORY_FENCE_ACQUIRE __c11_atomic_thread_fence (__ATOMIC_ACQUIRE)
202 * #define ECB_MEMORY_FENCE_RELEASE __c11_atomic_thread_fence (__ATOMIC_RELEASE) 235 #define ECB_MEMORY_FENCE_RELEASE __c11_atomic_thread_fence (__ATOMIC_RELEASE)
203 */ 236 #define ECB_MEMORY_FENCE_RELAXED __c11_atomic_thread_fence (__ATOMIC_RELAXED)
204 237
205 #elif ECB_GCC_VERSION(4,4) || defined __INTEL_COMPILER || defined __clang__ 238 #elif ECB_GCC_VERSION(4,4) || defined __INTEL_COMPILER || defined __clang__
206 #define ECB_MEMORY_FENCE __sync_synchronize () 239 #define ECB_MEMORY_FENCE __sync_synchronize ()
207 #elif _MSC_VER >= 1500 /* VC++ 2008 */ 240 #elif _MSC_VER >= 1500 /* VC++ 2008 */
208 /* apparently, microsoft broke all the memory barrier stuff in Visual Studio 2008... */ 241 /* apparently, microsoft broke all the memory barrier stuff in Visual Studio 2008... */
… …
218 #elif defined _WIN32 251 #elif defined _WIN32
219 #include <WinNT.h> 252 #include <WinNT.h>
220 #define ECB_MEMORY_FENCE MemoryBarrier () /* actually just xchg on x86... scary */ 253 #define ECB_MEMORY_FENCE MemoryBarrier () /* actually just xchg on x86... scary */
221 #elif __SUNPRO_C >= 0x5110 || __SUNPRO_CC >= 0x5110 254 #elif __SUNPRO_C >= 0x5110 || __SUNPRO_CC >= 0x5110
222 #include <mbarrier.h> 255 #include <mbarrier.h>
223 #define ECB_MEMORY_FENCE __machine_rw_barrier () 256 #define ECB_MEMORY_FENCE __machine_rw_barrier ()
224 #define ECB_MEMORY_FENCE_ACQUIRE __machine_r_barrier () 257 #define ECB_MEMORY_FENCE_ACQUIRE __machine_acq_barrier ()
225 #define ECB_MEMORY_FENCE_RELEASE __machine_w_barrier () 258 #define ECB_MEMORY_FENCE_RELEASE __machine_rel_barrier ()
259 #define ECB_MEMORY_FENCE_RELAXED __compiler_barrier ()
226 #elif __xlC__ 260 #elif __xlC__
227 #define ECB_MEMORY_FENCE __sync () 261 #define ECB_MEMORY_FENCE __sync ()
228 #endif 262 #endif
229#endif 263#endif
230 264
231#ifndef ECB_MEMORY_FENCE 265#ifndef ECB_MEMORY_FENCE
232 #if ECB_C11 && !defined __STDC_NO_ATOMICS__ 266 #if ECB_C11 && !defined __STDC_NO_ATOMICS__
233 /* we assume that these memory fences work on all variables/all memory accesses, */ 267 /* we assume that these memory fences work on all variables/all memory accesses, */
234 /* not just C11 atomics and atomic accesses */ 268 /* not just C11 atomics and atomic accesses */
235 #include <stdatomic.h> 269 #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) 270 #define ECB_MEMORY_FENCE atomic_thread_fence (memory_order_seq_cst)
271 #define ECB_MEMORY_FENCE_ACQUIRE atomic_thread_fence (memory_order_acquire)
272 #define ECB_MEMORY_FENCE_RELEASE atomic_thread_fence (memory_order_release)
245 #endif 273 #endif
246#endif 274#endif
247 275
248#ifndef ECB_MEMORY_FENCE 276#ifndef ECB_MEMORY_FENCE
249 #if !ECB_AVOID_PTHREADS 277 #if !ECB_AVOID_PTHREADS
… …
269 297
270#if !defined ECB_MEMORY_FENCE_RELEASE && defined ECB_MEMORY_FENCE 298#if !defined ECB_MEMORY_FENCE_RELEASE && defined ECB_MEMORY_FENCE
271 #define ECB_MEMORY_FENCE_RELEASE ECB_MEMORY_FENCE 299 #define ECB_MEMORY_FENCE_RELEASE ECB_MEMORY_FENCE
272#endif 300#endif
273 301
302#if !defined ECB_MEMORY_FENCE_RELAXED && defined ECB_MEMORY_FENCE
303 #define ECB_MEMORY_FENCE_RELAXED ECB_MEMORY_FENCE /* very heavy-handed */
304#endif
305
274/*****************************************************************************/ 306/*****************************************************************************/
275 307
276#if __cplusplus 308#if ECB_CPP
277 #define ecb_inline static inline 309 #define ecb_inline static inline
278#elif ECB_GCC_VERSION(2,5) 310#elif ECB_GCC_VERSION(2,5)
279 #define ecb_inline static __inline__ 311 #define ecb_inline static __inline__
280#elif ECB_C99 312#elif ECB_C99
281 #define ecb_inline static inline 313 #define ecb_inline static inline
… …
295 327
296#define ECB_CONCAT_(a, b) a ## b 328#define ECB_CONCAT_(a, b) a ## b
297#define ECB_CONCAT(a, b) ECB_CONCAT_(a, b) 329#define ECB_CONCAT(a, b) ECB_CONCAT_(a, b)
298#define ECB_STRINGIFY_(a) # a 330#define ECB_STRINGIFY_(a) # a
299#define ECB_STRINGIFY(a) ECB_STRINGIFY_(a) 331#define ECB_STRINGIFY(a) ECB_STRINGIFY_(a)
332#define ECB_STRINGIFY_EXPR(expr) ((expr), ECB_STRINGIFY_ (expr))
300 333
301#define ecb_function_ ecb_inline 334#define ecb_function_ ecb_inline
302 335
303#if ECB_GCC_VERSION(3,1) 336#if ECB_GCC_VERSION(3,1) || ECB_CLANG_VERSION(2,8)
304 #define ecb_attribute(attrlist) __attribute__(attrlist) 337 #define ecb_attribute(attrlist) __attribute__ (attrlist)
338#else
339 #define ecb_attribute(attrlist)
340#endif
341
342#if ECB_GCC_VERSION(3,1) || ECB_CLANG_BUILTIN(__builtin_constant_p)
305 #define ecb_is_constant(expr) __builtin_constant_p (expr) 343 #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 344#else
309 #define ecb_attribute(attrlist)
310
311 /* possible C11 impl for integral types 345 /* possible C11 impl for integral types
312 typedef struct ecb_is_constant_struct ecb_is_constant_struct; 346 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)) */ 347 #define ecb_is_constant(expr) _Generic ((1 ? (struct ecb_is_constant_struct *)0 : (void *)((expr) - (expr)), ecb_is_constant_struct *: 0, default: 1)) */
314 348
315 #define ecb_is_constant(expr) 0 349 #define ecb_is_constant(expr) 0
350#endif
351
352#if ECB_GCC_VERSION(3,1) || ECB_CLANG_BUILTIN(__builtin_expect)
353 #define ecb_expect(expr,value) __builtin_expect ((expr),(value))
354#else
316 #define ecb_expect(expr,value) (expr) 355 #define ecb_expect(expr,value) (expr)
356#endif
357
358#if ECB_GCC_VERSION(3,1) || ECB_CLANG_BUILTIN(__builtin_prefetch)
359 #define ecb_prefetch(addr,rw,locality) __builtin_prefetch (addr, rw, locality)
360#else
317 #define ecb_prefetch(addr,rw,locality) 361 #define ecb_prefetch(addr,rw,locality)
318#endif 362#endif
319 363
320/* no emulation for ecb_decltype */ 364/* no emulation for ecb_decltype */
321#if ECB_GCC_VERSION(4,5) 365#if ECB_CPP11
366 // older implementations might have problems with decltype(x)::type, work around it
367 template<class T> struct ecb_decltype_t { typedef T type; };
322 #define ecb_decltype(x) __decltype(x) 368 #define ecb_decltype(x) ecb_decltype_t<decltype (x)>::type
323#elif ECB_GCC_VERSION(3,0) 369#elif ECB_GCC_VERSION(3,0) || ECB_CLANG_VERSION(2,8)
324 #define ecb_decltype(x) __typeof(x) 370 #define ecb_decltype(x) __typeof__ (x)
325#endif 371#endif
326 372
373#if _MSC_VER >= 1300
374 #define ecb_deprecated __declspec (deprecated)
375#else
376 #define ecb_deprecated ecb_attribute ((__deprecated__))
377#endif
378
379#if _MSC_VER >= 1500
380 #define ecb_deprecated_message(msg) __declspec (deprecated (msg))
381#elif ECB_GCC_VERSION(4,5)
382 #define ecb_deprecated_message(msg) ecb_attribute ((__deprecated__ (msg))
383#else
384 #define ecb_deprecated_message(msg) ecb_deprecated
385#endif
386
387#if _MSC_VER >= 1400
388 #define ecb_noinline __declspec (noinline)
389#else
327#define ecb_noinline ecb_attribute ((__noinline__)) 390 #define ecb_noinline ecb_attribute ((__noinline__))
391#endif
392
328#define ecb_unused ecb_attribute ((__unused__)) 393#define ecb_unused ecb_attribute ((__unused__))
329#define ecb_const ecb_attribute ((__const__)) 394#define ecb_const ecb_attribute ((__const__))
330#define ecb_pure ecb_attribute ((__pure__)) 395#define ecb_pure ecb_attribute ((__pure__))
331 396
332#if ECB_C11 397#if ECB_C11 || __IBMC_NORETURN
398 /* http://www-01.ibm.com/support/knowledgecenter/SSGH3R_13.1.0/com.ibm.xlcpp131.aix.doc/language_ref/noreturn.html */
333 #define ecb_noreturn _Noreturn 399 #define ecb_noreturn _Noreturn
400#elif ECB_CPP11
401 #define ecb_noreturn [[noreturn]]
402#elif _MSC_VER >= 1200
403 /* http://msdn.microsoft.com/en-us/library/k6ktzx3s.aspx */
404 #define ecb_noreturn __declspec (noreturn)
334#else 405#else
335 #define ecb_noreturn ecb_attribute ((__noreturn__)) 406 #define ecb_noreturn ecb_attribute ((__noreturn__))
336#endif 407#endif
337 408
338#if ECB_GCC_VERSION(4,3) 409#if ECB_GCC_VERSION(4,3)
… …
353/* for compatibility to the rest of the world */ 424/* for compatibility to the rest of the world */
354#define ecb_likely(expr) ecb_expect_true (expr) 425#define ecb_likely(expr) ecb_expect_true (expr)
355#define ecb_unlikely(expr) ecb_expect_false (expr) 426#define ecb_unlikely(expr) ecb_expect_false (expr)
356 427
357/* count trailing zero bits and count # of one bits */ 428/* count trailing zero bits and count # of one bits */
358#if ECB_GCC_VERSION(3,4) 429#if ECB_GCC_VERSION(3,4) \
430 || (ECB_CLANG_BUILTIN(__builtin_clz) && ECB_CLANG_BUILTIN(__builtin_clzll) \
431 && ECB_CLANG_BUILTIN(__builtin_ctz) && ECB_CLANG_BUILTIN(__builtin_ctzll) \
432 && ECB_CLANG_BUILTIN(__builtin_popcount))
359 /* we assume int == 32 bit, long == 32 or 64 bit and long long == 64 bit */ 433 /* we assume int == 32 bit, long == 32 or 64 bit and long long == 64 bit */
360 #define ecb_ld32(x) (__builtin_clz (x) ^ 31) 434 #define ecb_ld32(x) (__builtin_clz (x) ^ 31)
361 #define ecb_ld64(x) (__builtin_clzll (x) ^ 63) 435 #define ecb_ld64(x) (__builtin_clzll (x) ^ 63)
362 #define ecb_ctz32(x) __builtin_ctz (x) 436 #define ecb_ctz32(x) __builtin_ctz (x)
363 #define ecb_ctz64(x) __builtin_ctzll (x) 437 #define ecb_ctz64(x) __builtin_ctzll (x)
364 #define ecb_popcount32(x) __builtin_popcount (x) 438 #define ecb_popcount32(x) __builtin_popcount (x)
365 /* no popcountll */ 439 /* no popcountll */
366#else 440#else
367 ecb_function_ int ecb_ctz32 (uint32_t x) ecb_const; 441 ecb_function_ ecb_const int ecb_ctz32 (uint32_t x);
368 ecb_function_ int 442 ecb_function_ ecb_const int
369 ecb_ctz32 (uint32_t x) 443 ecb_ctz32 (uint32_t x)
370 { 444 {
445#if 1400 <= _MSC_VER && (_M_IX86 || _M_X64 || _M_IA64 || _M_ARM)
446 unsigned long r;
447 _BitScanForward (&r, x);
448 return (int)r;
449#else
371 int r = 0; 450 int r = 0;
372 451
373 x &= ~x + 1; /* this isolates the lowest bit */ 452 x &= ~x + 1; /* this isolates the lowest bit */
374 453
375#if ECB_branchless_on_i386 454#if ECB_branchless_on_i386
… …
385 if (x & 0xff00ff00) r += 8; 464 if (x & 0xff00ff00) r += 8;
386 if (x & 0xffff0000) r += 16; 465 if (x & 0xffff0000) r += 16;
387#endif 466#endif
388 467
389 return r; 468 return r;
469#endif
390 } 470 }
391 471
392 ecb_function_ int ecb_ctz64 (uint64_t x) ecb_const; 472 ecb_function_ ecb_const int ecb_ctz64 (uint64_t x);
393 ecb_function_ int 473 ecb_function_ ecb_const int
394 ecb_ctz64 (uint64_t x) 474 ecb_ctz64 (uint64_t x)
395 { 475 {
476#if 1400 <= _MSC_VER && (_M_X64 || _M_IA64 || _M_ARM)
477 unsigned long r;
478 _BitScanForward64 (&r, x);
479 return (int)r;
480#else
396 int shift = x & 0xffffffffU ? 0 : 32; 481 int shift = x & 0xffffffff ? 0 : 32;
397 return ecb_ctz32 (x >> shift) + shift; 482 return ecb_ctz32 (x >> shift) + shift;
483#endif
398 } 484 }
399 485
400 ecb_function_ int ecb_popcount32 (uint32_t x) ecb_const; 486 ecb_function_ ecb_const int ecb_popcount32 (uint32_t x);
401 ecb_function_ int 487 ecb_function_ ecb_const int
402 ecb_popcount32 (uint32_t x) 488 ecb_popcount32 (uint32_t x)
403 { 489 {
404 x -= (x >> 1) & 0x55555555; 490 x -= (x >> 1) & 0x55555555;
405 x = ((x >> 2) & 0x33333333) + (x & 0x33333333); 491 x = ((x >> 2) & 0x33333333) + (x & 0x33333333);
406 x = ((x >> 4) + x) & 0x0f0f0f0f; 492 x = ((x >> 4) + x) & 0x0f0f0f0f;
407 x *= 0x01010101; 493 x *= 0x01010101;
408 494
409 return x >> 24; 495 return x >> 24;
410 } 496 }
411 497
412 ecb_function_ int ecb_ld32 (uint32_t x) ecb_const; 498 ecb_function_ ecb_const int ecb_ld32 (uint32_t x);
413 ecb_function_ int ecb_ld32 (uint32_t x) 499 ecb_function_ ecb_const int ecb_ld32 (uint32_t x)
414 { 500 {
501#if 1400 <= _MSC_VER && (_M_IX86 || _M_X64 || _M_IA64 || _M_ARM)
502 unsigned long r;
503 _BitScanReverse (&r, x);
504 return (int)r;
505#else
415 int r = 0; 506 int r = 0;
416 507
417 if (x >> 16) { x >>= 16; r += 16; } 508 if (x >> 16) { x >>= 16; r += 16; }
418 if (x >> 8) { x >>= 8; r += 8; } 509 if (x >> 8) { x >>= 8; r += 8; }
419 if (x >> 4) { x >>= 4; r += 4; } 510 if (x >> 4) { x >>= 4; r += 4; }
420 if (x >> 2) { x >>= 2; r += 2; } 511 if (x >> 2) { x >>= 2; r += 2; }
421 if (x >> 1) { r += 1; } 512 if (x >> 1) { r += 1; }
422 513
423 return r; 514 return r;
515#endif
424 } 516 }
425 517
426 ecb_function_ int ecb_ld64 (uint64_t x) ecb_const; 518 ecb_function_ ecb_const int ecb_ld64 (uint64_t x);
427 ecb_function_ int ecb_ld64 (uint64_t x) 519 ecb_function_ ecb_const int ecb_ld64 (uint64_t x)
428 { 520 {
521#if 1400 <= _MSC_VER && (_M_X64 || _M_IA64 || _M_ARM)
522 unsigned long r;
523 _BitScanReverse64 (&r, x);
524 return (int)r;
525#else
429 int r = 0; 526 int r = 0;
430 527
431 if (x >> 32) { x >>= 32; r += 32; } 528 if (x >> 32) { x >>= 32; r += 32; }
432 529
433 return r + ecb_ld32 (x); 530 return r + ecb_ld32 (x);
434 }
435#endif 531#endif
532 }
533#endif
436 534
437ecb_function_ ecb_bool ecb_is_pot32 (uint32_t x) ecb_const; 535ecb_function_ ecb_const ecb_bool ecb_is_pot32 (uint32_t x);
438ecb_function_ ecb_bool ecb_is_pot32 (uint32_t x) { return !(x & (x - 1)); } 536ecb_function_ ecb_const ecb_bool ecb_is_pot32 (uint32_t x) { return !(x & (x - 1)); }
439ecb_function_ ecb_bool ecb_is_pot64 (uint64_t x) ecb_const; 537ecb_function_ ecb_const ecb_bool ecb_is_pot64 (uint64_t x);
440ecb_function_ ecb_bool ecb_is_pot64 (uint64_t x) { return !(x & (x - 1)); } 538ecb_function_ ecb_const ecb_bool ecb_is_pot64 (uint64_t x) { return !(x & (x - 1)); }
441 539
442ecb_function_ uint8_t ecb_bitrev8 (uint8_t x) ecb_const; 540ecb_function_ ecb_const uint8_t ecb_bitrev8 (uint8_t x);
443ecb_function_ uint8_t ecb_bitrev8 (uint8_t x) 541ecb_function_ ecb_const uint8_t ecb_bitrev8 (uint8_t x)
444{ 542{
445 return ( (x * 0x0802U & 0x22110U) 543 return ( (x * 0x0802U & 0x22110U)
446 | (x * 0x8020U & 0x88440U)) * 0x10101U >> 16; 544 | (x * 0x8020U & 0x88440U)) * 0x10101U >> 16;
447} 545}
448 546
449ecb_function_ uint16_t ecb_bitrev16 (uint16_t x) ecb_const; 547ecb_function_ ecb_const uint16_t ecb_bitrev16 (uint16_t x);
450ecb_function_ uint16_t ecb_bitrev16 (uint16_t x) 548ecb_function_ ecb_const uint16_t ecb_bitrev16 (uint16_t x)
451{ 549{
452 x = ((x >> 1) & 0x5555) | ((x & 0x5555) << 1); 550 x = ((x >> 1) & 0x5555) | ((x & 0x5555) << 1);
453 x = ((x >> 2) & 0x3333) | ((x & 0x3333) << 2); 551 x = ((x >> 2) & 0x3333) | ((x & 0x3333) << 2);
454 x = ((x >> 4) & 0x0f0f) | ((x & 0x0f0f) << 4); 552 x = ((x >> 4) & 0x0f0f) | ((x & 0x0f0f) << 4);
455 x = ( x >> 8 ) | ( x << 8); 553 x = ( x >> 8 ) | ( x << 8);
456 554
457 return x; 555 return x;
458} 556}
459 557
460ecb_function_ uint32_t ecb_bitrev32 (uint32_t x) ecb_const; 558ecb_function_ ecb_const uint32_t ecb_bitrev32 (uint32_t x);
461ecb_function_ uint32_t ecb_bitrev32 (uint32_t x) 559ecb_function_ ecb_const uint32_t ecb_bitrev32 (uint32_t x)
462{ 560{
463 x = ((x >> 1) & 0x55555555) | ((x & 0x55555555) << 1); 561 x = ((x >> 1) & 0x55555555) | ((x & 0x55555555) << 1);
464 x = ((x >> 2) & 0x33333333) | ((x & 0x33333333) << 2); 562 x = ((x >> 2) & 0x33333333) | ((x & 0x33333333) << 2);
465 x = ((x >> 4) & 0x0f0f0f0f) | ((x & 0x0f0f0f0f) << 4); 563 x = ((x >> 4) & 0x0f0f0f0f) | ((x & 0x0f0f0f0f) << 4);
466 x = ((x >> 8) & 0x00ff00ff) | ((x & 0x00ff00ff) << 8); 564 x = ((x >> 8) & 0x00ff00ff) | ((x & 0x00ff00ff) << 8);
… …
469 return x; 567 return x;
470} 568}
471 569
472/* popcount64 is only available on 64 bit cpus as gcc builtin */ 570/* popcount64 is only available on 64 bit cpus as gcc builtin */
473/* so for this version we are lazy */ 571/* so for this version we are lazy */
474ecb_function_ int ecb_popcount64 (uint64_t x) ecb_const; 572ecb_function_ ecb_const int ecb_popcount64 (uint64_t x);
475ecb_function_ int 573ecb_function_ ecb_const int
476ecb_popcount64 (uint64_t x) 574ecb_popcount64 (uint64_t x)
477{ 575{
478 return ecb_popcount32 (x) + ecb_popcount32 (x >> 32); 576 return ecb_popcount32 (x) + ecb_popcount32 (x >> 32);
479} 577}
480 578
481ecb_inline uint8_t ecb_rotl8 (uint8_t x, unsigned int count) ecb_const; 579ecb_inline ecb_const uint8_t ecb_rotl8 (uint8_t x, unsigned int count);
482ecb_inline uint8_t ecb_rotr8 (uint8_t x, unsigned int count) ecb_const; 580ecb_inline ecb_const uint8_t ecb_rotr8 (uint8_t x, unsigned int count);
483ecb_inline uint16_t ecb_rotl16 (uint16_t x, unsigned int count) ecb_const; 581ecb_inline ecb_const uint16_t ecb_rotl16 (uint16_t x, unsigned int count);
484ecb_inline uint16_t ecb_rotr16 (uint16_t x, unsigned int count) ecb_const; 582ecb_inline ecb_const uint16_t ecb_rotr16 (uint16_t x, unsigned int count);
485ecb_inline uint32_t ecb_rotl32 (uint32_t x, unsigned int count) ecb_const; 583ecb_inline ecb_const uint32_t ecb_rotl32 (uint32_t x, unsigned int count);
486ecb_inline uint32_t ecb_rotr32 (uint32_t x, unsigned int count) ecb_const; 584ecb_inline ecb_const uint32_t ecb_rotr32 (uint32_t x, unsigned int count);
487ecb_inline uint64_t ecb_rotl64 (uint64_t x, unsigned int count) ecb_const; 585ecb_inline ecb_const uint64_t ecb_rotl64 (uint64_t x, unsigned int count);
488ecb_inline uint64_t ecb_rotr64 (uint64_t x, unsigned int count) ecb_const; 586ecb_inline ecb_const uint64_t ecb_rotr64 (uint64_t x, unsigned int count);
489 587
490ecb_inline uint8_t ecb_rotl8 (uint8_t x, unsigned int count) { return (x >> ( 8 - count)) | (x << count); } 588ecb_inline ecb_const uint8_t ecb_rotl8 (uint8_t x, unsigned int count) { return (x >> ( 8 - count)) | (x << count); }
491ecb_inline uint8_t ecb_rotr8 (uint8_t x, unsigned int count) { return (x << ( 8 - count)) | (x >> count); } 589ecb_inline ecb_const uint8_t ecb_rotr8 (uint8_t x, unsigned int count) { return (x << ( 8 - count)) | (x >> count); }
492ecb_inline uint16_t ecb_rotl16 (uint16_t x, unsigned int count) { return (x >> (16 - count)) | (x << count); } 590ecb_inline ecb_const uint16_t ecb_rotl16 (uint16_t x, unsigned int count) { return (x >> (16 - count)) | (x << count); }
493ecb_inline uint16_t ecb_rotr16 (uint16_t x, unsigned int count) { return (x << (16 - count)) | (x >> count); } 591ecb_inline ecb_const uint16_t ecb_rotr16 (uint16_t x, unsigned int count) { return (x << (16 - count)) | (x >> count); }
494ecb_inline uint32_t ecb_rotl32 (uint32_t x, unsigned int count) { return (x >> (32 - count)) | (x << count); } 592ecb_inline ecb_const uint32_t ecb_rotl32 (uint32_t x, unsigned int count) { return (x >> (32 - count)) | (x << count); }
495ecb_inline uint32_t ecb_rotr32 (uint32_t x, unsigned int count) { return (x << (32 - count)) | (x >> count); } 593ecb_inline ecb_const uint32_t ecb_rotr32 (uint32_t x, unsigned int count) { return (x << (32 - count)) | (x >> count); }
496ecb_inline uint64_t ecb_rotl64 (uint64_t x, unsigned int count) { return (x >> (64 - count)) | (x << count); } 594ecb_inline ecb_const uint64_t ecb_rotl64 (uint64_t x, unsigned int count) { return (x >> (64 - count)) | (x << count); }
497ecb_inline uint64_t ecb_rotr64 (uint64_t x, unsigned int count) { return (x << (64 - count)) | (x >> count); } 595ecb_inline ecb_const uint64_t ecb_rotr64 (uint64_t x, unsigned int count) { return (x << (64 - count)) | (x >> count); }
498 596
499#if ECB_GCC_VERSION(4,3) 597#if ECB_GCC_VERSION(4,3) || (ECB_CLANG_BUILTIN(__builtin_bswap32) && ECB_CLANG_BUILTIN(__builtin_bswap64))
598 #if ECB_GCC_VERSION(4,8) || ECB_CLANG_BUILTIN(__builtin_bswap16)
599 #define ecb_bswap16(x) __builtin_bswap16 (x)
600 #else
500 #define ecb_bswap16(x) (__builtin_bswap32 (x) >> 16) 601 #define ecb_bswap16(x) (__builtin_bswap32 (x) >> 16)
602 #endif
501 #define ecb_bswap32(x) __builtin_bswap32 (x) 603 #define ecb_bswap32(x) __builtin_bswap32 (x)
502 #define ecb_bswap64(x) __builtin_bswap64 (x) 604 #define ecb_bswap64(x) __builtin_bswap64 (x)
605#elif _MSC_VER
606 #include <stdlib.h>
607 #define ecb_bswap16(x) ((uint16_t)_byteswap_ushort ((uint16_t)(x)))
608 #define ecb_bswap32(x) ((uint32_t)_byteswap_ulong ((uint32_t)(x)))
609 #define ecb_bswap64(x) ((uint64_t)_byteswap_uint64 ((uint64_t)(x)))
503#else 610#else
504 ecb_function_ uint16_t ecb_bswap16 (uint16_t x) ecb_const; 611 ecb_function_ ecb_const uint16_t ecb_bswap16 (uint16_t x);
505 ecb_function_ uint16_t 612 ecb_function_ ecb_const uint16_t
506 ecb_bswap16 (uint16_t x) 613 ecb_bswap16 (uint16_t x)
507 { 614 {
508 return ecb_rotl16 (x, 8); 615 return ecb_rotl16 (x, 8);
509 } 616 }
510 617
511 ecb_function_ uint32_t ecb_bswap32 (uint32_t x) ecb_const; 618 ecb_function_ ecb_const uint32_t ecb_bswap32 (uint32_t x);
512 ecb_function_ uint32_t 619 ecb_function_ ecb_const uint32_t
513 ecb_bswap32 (uint32_t x) 620 ecb_bswap32 (uint32_t x)
514 { 621 {
515 return (((uint32_t)ecb_bswap16 (x)) << 16) | ecb_bswap16 (x >> 16); 622 return (((uint32_t)ecb_bswap16 (x)) << 16) | ecb_bswap16 (x >> 16);
516 } 623 }
517 624
518 ecb_function_ uint64_t ecb_bswap64 (uint64_t x) ecb_const; 625 ecb_function_ ecb_const uint64_t ecb_bswap64 (uint64_t x);
519 ecb_function_ uint64_t 626 ecb_function_ ecb_const uint64_t
520 ecb_bswap64 (uint64_t x) 627 ecb_bswap64 (uint64_t x)
521 { 628 {
522 return (((uint64_t)ecb_bswap32 (x)) << 32) | ecb_bswap32 (x >> 32); 629 return (((uint64_t)ecb_bswap32 (x)) << 32) | ecb_bswap32 (x >> 32);
523 } 630 }
524#endif 631#endif
525 632
526#if ECB_GCC_VERSION(4,5) 633#if ECB_GCC_VERSION(4,5) || ECB_CLANG_BUILTIN(__builtin_unreachable)
527 #define ecb_unreachable() __builtin_unreachable () 634 #define ecb_unreachable() __builtin_unreachable ()
528#else 635#else
529 /* this seems to work fine, but gcc always emits a warning for it :/ */ 636 /* this seems to work fine, but gcc always emits a warning for it :/ */
530 ecb_inline void ecb_unreachable (void) ecb_noreturn; 637 ecb_inline ecb_noreturn void ecb_unreachable (void);
531 ecb_inline void ecb_unreachable (void) { } 638 ecb_inline ecb_noreturn void ecb_unreachable (void) { }
532#endif 639#endif
533 640
534/* try to tell the compiler that some condition is definitely true */ 641/* try to tell the compiler that some condition is definitely true */
535#define ecb_assume(cond) if (!(cond)) ecb_unreachable (); else 0 642#define ecb_assume(cond) if (!(cond)) ecb_unreachable (); else 0
536 643
537ecb_inline unsigned char ecb_byteorder_helper (void) ecb_const; 644ecb_inline ecb_const uint32_t ecb_byteorder_helper (void);
538ecb_inline unsigned char 645ecb_inline ecb_const uint32_t
539ecb_byteorder_helper (void) 646ecb_byteorder_helper (void)
540{ 647{
541 /* the union code still generates code under pressure in gcc, */ 648 /* the union code still generates code under pressure in gcc, */
542 /* but less than using pointers, and always seems to */ 649 /* but less than using pointers, and always seems to */
543 /* successfully return a constant. */ 650 /* successfully return a constant. */
544 /* the reason why we have this horrible preprocessor mess */ 651 /* the reason why we have this horrible preprocessor mess */
545 /* is to avoid it in all cases, at least on common architectures */ 652 /* is to avoid it in all cases, at least on common architectures */
546 /* or when using a recent enough gcc version (>= 4.6) */ 653 /* or when using a recent enough gcc version (>= 4.6) */
547#if __i386 || __i386__ || _M_X86 || __amd64 || __amd64__ || _M_X64
548 return 0x44;
549#elif __BYTE_ORDER__ && __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__ 654#if (defined __BYTE_ORDER__ && __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__) \
655 || ((__i386 || __i386__ || _M_IX86 || ECB_GCC_AMD64 || ECB_MSVC_AMD64) && !__VOS__)
656 #define ECB_LITTLE_ENDIAN 1
550 return 0x44; 657 return 0x44332211;
551#elif __BYTE_ORDER__ && __BYTE_ORDER__ == __ORDER_BIG_ENDIAN__ 658#elif (defined __BYTE_ORDER__ && __BYTE_ORDER__ == __ORDER_BIG_ENDIAN__) \
659 || ((__AARCH64EB__ || __MIPSEB__ || __ARMEB__) && !__VOS__)
660 #define ECB_BIG_ENDIAN 1
552 return 0x11; 661 return 0x11223344;
553#else 662#else
554 union 663 union
555 { 664 {
665 uint8_t c[4];
556 uint32_t i; 666 uint32_t u;
557 uint8_t c;
558 } u = { 0x11223344 }; 667 } u = { 0x11, 0x22, 0x33, 0x44 };
559 return u.c; 668 return u.u;
560#endif 669#endif
561} 670}
562 671
563ecb_inline ecb_bool ecb_big_endian (void) ecb_const; 672ecb_inline ecb_const ecb_bool ecb_big_endian (void);
564ecb_inline ecb_bool ecb_big_endian (void) { return ecb_byteorder_helper () == 0x11; } 673ecb_inline ecb_const ecb_bool ecb_big_endian (void) { return ecb_byteorder_helper () == 0x11223344; }
565ecb_inline ecb_bool ecb_little_endian (void) ecb_const; 674ecb_inline ecb_const ecb_bool ecb_little_endian (void);
566ecb_inline ecb_bool ecb_little_endian (void) { return ecb_byteorder_helper () == 0x44; } 675ecb_inline ecb_const ecb_bool ecb_little_endian (void) { return ecb_byteorder_helper () == 0x44332211; }
567 676
568#if ECB_GCC_VERSION(3,0) || ECB_C99 677#if ECB_GCC_VERSION(3,0) || ECB_C99
569 #define ecb_mod(m,n) ((m) % (n) + ((m) % (n) < 0 ? (n) : 0)) 678 #define ecb_mod(m,n) ((m) % (n) + ((m) % (n) < 0 ? (n) : 0))
570#else 679#else
571 #define ecb_mod(m,n) ((m) < 0 ? ((n) - 1 - ((-1 - (m)) % (n))) : ((m) % (n))) 680 #define ecb_mod(m,n) ((m) < 0 ? ((n) - 1 - ((-1 - (m)) % (n))) : ((m) % (n)))
572#endif 681#endif
573 682
574#if __cplusplus 683#if ECB_CPP
575 template<typename T> 684 template<typename T>
576 static inline T ecb_div_rd (T val, T div) 685 static inline T ecb_div_rd (T val, T div)
577 { 686 {
578 return val < 0 ? - ((-val + div - 1) / div) : (val ) / div; 687 return val < 0 ? - ((-val + div - 1) / div) : (val ) / div;
579 } 688 }
… …
596 } 705 }
597#else 706#else
598 #define ecb_array_length(name) (sizeof (name) / sizeof (name [0])) 707 #define ecb_array_length(name) (sizeof (name) / sizeof (name [0]))
599#endif 708#endif
600 709
710ecb_function_ ecb_const uint32_t ecb_binary16_to_binary32 (uint32_t x);
711ecb_function_ ecb_const uint32_t
712ecb_binary16_to_binary32 (uint32_t x)
713{
714 unsigned int s = (x & 0x8000) << (31 - 15);
715 int e = (x >> 10) & 0x001f;
716 unsigned int m = x & 0x03ff;
717
718 if (ecb_expect_false (e == 31))
719 /* infinity or NaN */
720 e = 255 - (127 - 15);
721 else if (ecb_expect_false (!e))
722 {
723 if (ecb_expect_true (!m))
724 /* zero, handled by code below by forcing e to 0 */
725 e = 0 - (127 - 15);
726 else
727 {
728 /* subnormal, renormalise */
729 unsigned int s = 10 - ecb_ld32 (m);
730
731 m = (m << s) & 0x3ff; /* mask implicit bit */
732 e -= s - 1;
733 }
734 }
735
736 /* e and m now are normalised, or zero, (or inf or nan) */
737 e += 127 - 15;
738
739 return s | (e << 23) | (m << (23 - 10));
740}
741
742ecb_function_ ecb_const uint16_t ecb_binary32_to_binary16 (uint32_t x);
743ecb_function_ ecb_const uint16_t
744ecb_binary32_to_binary16 (uint32_t x)
745{
746 unsigned int s = (x >> 16) & 0x00008000; /* sign bit, the easy part */
747 unsigned int e = ((x >> 23) & 0x000000ff) - (127 - 15); /* the desired exponent */
748 unsigned int m = x & 0x007fffff;
749
750 x &= 0x7fffffff;
751
752 /* if it's within range of binary16 normals, use fast path */
753 if (ecb_expect_true (0x38800000 <= x && x <= 0x477fefff))
754 {
755 /* mantissa round-to-even */
756 m += 0x00000fff + ((m >> (23 - 10)) & 1);
757
758 /* handle overflow */
759 if (ecb_expect_false (m >= 0x00800000))
760 {
761 m >>= 1;
762 e += 1;
763 }
764
765 return s | (e << 10) | (m >> (23 - 10));
766 }
767
768 /* handle large numbers and infinity */
769 if (ecb_expect_true (0x477fefff < x && x <= 0x7f800000))
770 return s | 0x7c00;
771
772 /* handle zero, subnormals and small numbers */
773 if (ecb_expect_true (x < 0x38800000))
774 {
775 /* zero */
776 if (ecb_expect_true (!x))
777 return s;
778
779 /* handle subnormals */
780
781 /* too small, will be zero */
782 if (e < (14 - 24)) /* might not be sharp, but is good enough */
783 return s;
784
785 m |= 0x00800000; /* make implicit bit explicit */
786
787 /* very tricky - we need to round to the nearest e (+10) bit value */
788 {
789 unsigned int bits = 14 - e;
790 unsigned int half = (1 << (bits - 1)) - 1;
791 unsigned int even = (m >> bits) & 1;
792
793 /* if this overflows, we will end up with a normalised number */
794 m = (m + half + even) >> bits;
795 }
796
797 return s | m;
798 }
799
800 /* handle NaNs, preserve leftmost nan bits, but make sure we don't turn them into infinities */
801 m >>= 13;
802
803 return s | 0x7c00 | m | !m;
804}
805
601/*******************************************************************************/ 806/*******************************************************************************/
602/* floating point stuff, can be disabled by defining ECB_NO_LIBM */ 807/* floating point stuff, can be disabled by defining ECB_NO_LIBM */
603 808
604/* basically, everything uses "ieee pure-endian" floating point numbers */ 809/* basically, everything uses "ieee pure-endian" floating point numbers */
605/* the only noteworthy exception is ancient armle, which uses order 43218765 */ 810/* the only noteworthy exception is ancient armle, which uses order 43218765 */
606#if 0 \ 811#if 0 \
607 || __i386 || __i386__ \ 812 || __i386 || __i386__ \
608 || __amd64 || __amd64__ || __x86_64 || __x86_64__ \ 813 || ECB_GCC_AMD64 \
609 || __powerpc__ || __ppc__ || __powerpc64__ || __ppc64__ \ 814 || __powerpc__ || __ppc__ || __powerpc64__ || __ppc64__ \
610 || defined __s390__ || defined __s390x__ \ 815 || defined __s390__ || defined __s390x__ \
611 || defined __mips__ \ 816 || defined __mips__ \
612 || defined __alpha__ \ 817 || defined __alpha__ \
613 || defined __hppa__ \ 818 || defined __hppa__ \
614 || defined __ia64__ \ 819 || defined __ia64__ \
615 || defined __m68k__ \ 820 || defined __m68k__ \
616 || defined __m88k__ \ 821 || defined __m88k__ \
617 || defined __sh__ \ 822 || defined __sh__ \
618 || defined _M_IX86 || defined _M_AMD64 || defined _M_IA64 \ 823 || defined _M_IX86 || defined ECB_MSVC_AMD64 || defined _M_IA64 \
619 || (defined __arm__ && (defined __ARM_EABI__ || defined __EABI__ || defined __VFP_FP__ || defined _WIN32_WCE || defined __ANDROID__)) \ 824 || (defined __arm__ && (defined __ARM_EABI__ || defined __EABI__ || defined __VFP_FP__ || defined _WIN32_WCE || defined __ANDROID__)) \
620 || defined __aarch64__ 825 || defined __aarch64__
621 #define ECB_STDFP 1 826 #define ECB_STDFP 1
622 #include <string.h> /* for memcpy */ 827 #include <string.h> /* for memcpy */
623#else 828#else
… …
639 #define ECB_NAN NAN 844 #define ECB_NAN NAN
640 #else 845 #else
641 #define ECB_NAN ECB_INFINITY 846 #define ECB_NAN ECB_INFINITY
642 #endif 847 #endif
643 848
644 /* converts an ieee half/binary16 to a float */ 849 #if ECB_C99 || _XOPEN_VERSION >= 600 || _POSIX_VERSION >= 200112L
645 ecb_function_ float ecb_binary16_to_float (uint16_t x) ecb_const; 850 #define ecb_ldexpf(x,e) ldexpf ((x), (e))
646 ecb_function_ float 851 #define ecb_frexpf(x,e) frexpf ((x), (e))
647 ecb_binary16_to_float (uint16_t x) 852 #else
648 { 853 #define ecb_ldexpf(x,e) (float) ldexp ((double) (x), (e))
649 int e = (x >> 10) & 0x1f; 854 #define ecb_frexpf(x,e) (float) frexp ((double) (x), (e))
650 int m = x & 0x3ff; 855 #endif
651 float r;
652
653 if (!e ) r = ldexpf (m , -24);
654 else if (e != 31) r = ldexpf (m + 0x400, e - 25);
655 else if (m ) r = ECB_NAN;
656 else r = ECB_INFINITY;
657
658 return x & 0x8000 ? -r : r;
659 }
660 856
661 /* convert a float to ieee single/binary32 */ 857 /* convert a float to ieee single/binary32 */
662 ecb_function_ uint32_t ecb_float_to_binary32 (float x) ecb_const; 858 ecb_function_ ecb_const uint32_t ecb_float_to_binary32 (float x);
663 ecb_function_ uint32_t 859 ecb_function_ ecb_const uint32_t
664 ecb_float_to_binary32 (float x) 860 ecb_float_to_binary32 (float x)
665 { 861 {
666 uint32_t r; 862 uint32_t r;
667 863
668 #if ECB_STDFP 864 #if ECB_STDFP
… …
675 if (x == 0e0f ) return 0x00000000U; 871 if (x == 0e0f ) return 0x00000000U;
676 if (x > +3.40282346638528860e+38f) return 0x7f800000U; 872 if (x > +3.40282346638528860e+38f) return 0x7f800000U;
677 if (x < -3.40282346638528860e+38f) return 0xff800000U; 873 if (x < -3.40282346638528860e+38f) return 0xff800000U;
678 if (x != x ) return 0x7fbfffffU; 874 if (x != x ) return 0x7fbfffffU;
679 875
680 m = frexpf (x, &e) * 0x1000000U; 876 m = ecb_frexpf (x, &e) * 0x1000000U;
681 877
682 r = m & 0x80000000U; 878 r = m & 0x80000000U;
683 879
684 if (r) 880 if (r)
685 m = -m; 881 m = -m;
… …
697 893
698 return r; 894 return r;
699 } 895 }
700 896
701 /* converts an ieee single/binary32 to a float */ 897 /* converts an ieee single/binary32 to a float */
702 ecb_function_ float ecb_binary32_to_float (uint32_t x) ecb_const; 898 ecb_function_ ecb_const float ecb_binary32_to_float (uint32_t x);
703 ecb_function_ float 899 ecb_function_ ecb_const float
704 ecb_binary32_to_float (uint32_t x) 900 ecb_binary32_to_float (uint32_t x)
705 { 901 {
706 float r; 902 float r;
707 903
708 #if ECB_STDFP 904 #if ECB_STDFP
… …
718 x |= 0x800000U; 914 x |= 0x800000U;
719 else 915 else
720 e = 1; 916 e = 1;
721 917
722 /* we distrust ldexpf a bit and do the 2**-24 scaling by an extra multiply */ 918 /* we distrust ldexpf a bit and do the 2**-24 scaling by an extra multiply */
723 r = ldexpf (x * (0.5f / 0x800000U), e - 126); 919 r = ecb_ldexpf (x * (0.5f / 0x800000U), e - 126);
724 920
725 r = neg ? -r : r; 921 r = neg ? -r : r;
726 #endif 922 #endif
727 923
728 return r; 924 return r;
729 } 925 }
730 926
731 /* convert a double to ieee double/binary64 */ 927 /* convert a double to ieee double/binary64 */
732 ecb_function_ uint64_t ecb_double_to_binary64 (double x) ecb_const; 928 ecb_function_ ecb_const uint64_t ecb_double_to_binary64 (double x);
733 ecb_function_ uint64_t 929 ecb_function_ ecb_const uint64_t
734 ecb_double_to_binary64 (double x) 930 ecb_double_to_binary64 (double x)
735 { 931 {
736 uint64_t r; 932 uint64_t r;
737 933
738 #if ECB_STDFP 934 #if ECB_STDFP
… …
767 963
768 return r; 964 return r;
769 } 965 }
770 966
771 /* converts an ieee double/binary64 to a double */ 967 /* converts an ieee double/binary64 to a double */
772 ecb_function_ double ecb_binary64_to_double (uint64_t x) ecb_const; 968 ecb_function_ ecb_const double ecb_binary64_to_double (uint64_t x);
773 ecb_function_ double 969 ecb_function_ ecb_const double
774 ecb_binary64_to_double (uint64_t x) 970 ecb_binary64_to_double (uint64_t x)
775 { 971 {
776 double r; 972 double r;
777 973
778 #if ECB_STDFP 974 #if ECB_STDFP
… …
796 #endif 992 #endif
797 993
798 return r; 994 return r;
799 } 995 }
800 996
801#endif 997 /* convert a float to ieee half/binary16 */
998 ecb_function_ ecb_const uint16_t ecb_float_to_binary16 (float x);
999 ecb_function_ ecb_const uint16_t
1000 ecb_float_to_binary16 (float x)
1001 {
1002 return ecb_binary32_to_binary16 (ecb_float_to_binary32 (x));
1003 }
802 1004
803#endif 1005 /* convert an ieee half/binary16 to float */
1006 ecb_function_ ecb_const float ecb_binary16_to_float (uint16_t x);
1007 ecb_function_ ecb_const float
1008 ecb_binary16_to_float (uint16_t x)
1009 {
1010 return ecb_binary32_to_float (ecb_binary16_to_binary32 (x));
1011 }
804 1012
1013#endif
1014
1015#endif
1016

Diff Legend

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