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

Comparing libeio/ecb.h (file contents):
Revision 1.15 by root, Wed Jul 25 16:12:28 2012 UTC vs.
Revision 1.27 by root, Mon Jan 18 11:53:09 2016 UTC

1/* 1/*
2 * libecb - http://software.schmorp.de/pkg/libecb 2 * libecb - http://software.schmorp.de/pkg/libecb
3 * 3 *
4 * Copyright (©) 2009-2012 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:
23 * PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; 23 * PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS;
24 * OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, 24 * OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY,
25 * WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTH- 25 * WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTH-
26 * ERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED 26 * ERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED
27 * OF THE POSSIBILITY OF SUCH DAMAGE. 27 * OF THE POSSIBILITY OF SUCH DAMAGE.
28 *
29 * Alternatively, the contents of this file may be used under the terms of
30 * the GNU General Public License ("GPL") version 2 or any later version,
31 * in which case the provisions of the GPL are applicable instead of
32 * the above. If you wish to allow the use of your version of this file
33 * only under the terms of the GPL and not to allow others to use your
34 * version of this file under the BSD license, indicate your decision
35 * by deleting the provisions above and replace them with the notice
36 * and other provisions required by the GPL. If you do not delete the
37 * provisions above, a recipient may use your version of this file under
38 * either the BSD or the GPL.
28 */ 39 */
29 40
30#ifndef ECB_H 41#ifndef ECB_H
31#define ECB_H 42#define ECB_H
32 43
33/* 16 bits major, 16 bits minor */ 44/* 16 bits major, 16 bits minor */
34#define ECB_VERSION 0x00010002 45#define ECB_VERSION 0x00010005
35 46
36#ifdef _WIN32 47#ifdef _WIN32
37 typedef signed char int8_t; 48 typedef signed char int8_t;
38 typedef unsigned char uint8_t; 49 typedef unsigned char uint8_t;
39 typedef signed short int16_t; 50 typedef signed short int16_t;
56 typedef uint32_t uintptr_t; 67 typedef uint32_t uintptr_t;
57 typedef int32_t intptr_t; 68 typedef int32_t intptr_t;
58 #endif 69 #endif
59#else 70#else
60 #include <inttypes.h> 71 #include <inttypes.h>
61 #if UINTMAX_MAX > 0xffffffffU 72 #if (defined INTPTR_MAX ? INTPTR_MAX : ULONG_MAX) > 0xffffffffU
62 #define ECB_PTRSIZE 8 73 #define ECB_PTRSIZE 8
63 #else 74 #else
64 #define ECB_PTRSIZE 4 75 #define ECB_PTRSIZE 4
76 #endif
77#endif
78
79#define ECB_GCC_AMD64 (__amd64 || __amd64__ || __x86_64 || __x86_64__)
80#define ECB_MSVC_AMD64 (_M_AMD64 || _M_X64)
81
82/* work around x32 idiocy by defining proper macros */
83#if ECB_GCC_AMD64 || ECB_MSVC_AMD64
84 #if _ILP32
85 #define ECB_AMD64_X32 1
86 #else
87 #define ECB_AMD64 1
65 #endif 88 #endif
66#endif 89#endif
67 90
68/* many compilers define _GNUC_ to some versions but then only implement 91/* many compilers define _GNUC_ to some versions but then only implement
69 * what their idiot authors think are the "more important" extensions, 92 * what their idiot authors think are the "more important" extensions,
70 * causing enormous grief in return for some better fake benchmark numbers. 93 * causing enormous grief in return for some better fake benchmark numbers.
71 * or so. 94 * or so.
72 * 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
73 * 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.
74 */ 97 */
75#ifndef ECB_GCC_VERSION
76 #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__
77 #define ECB_GCC_VERSION(major,minor) 0 99 #define ECB_GCC_VERSION(major,minor) 0
78 #else 100#else
79 #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)))
80 #endif 102#endif
81#endif
82 103
83#define ECB_C (__STDC__+0) /* this assumes that __STDC__ is either empty or a number */ 104#define ECB_CLANG_VERSION(major,minor) (__clang_major__ > (major) || (__clang_major__ == (major) && __clang_minor__ >= (minor)))
84#define ECB_C99 (__STDC_VERSION__ >= 199901L) 105
85#define ECB_C11 (__STDC_VERSION__ >= 201112L) 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
116#endif
117
86#define ECB_CPP (__cplusplus+0) 118#define ECB_CPP (__cplusplus+0)
87#define ECB_CPP11 (__cplusplus >= 201103L) 119#define ECB_CPP11 (__cplusplus >= 201103L)
120
121#if ECB_CPP
122 #define ECB_C 0
123 #define ECB_STDC_VERSION 0
124#else
125 #define ECB_C 1
126 #define ECB_STDC_VERSION __STDC_VERSION__
127#endif
128
129#define ECB_C99 (ECB_STDC_VERSION >= 199901L)
130#define ECB_C11 (ECB_STDC_VERSION >= 201112L)
88 131
89#if ECB_CPP 132#if ECB_CPP
90 #define ECB_EXTERN_C extern "C" 133 #define ECB_EXTERN_C extern "C"
91 #define ECB_EXTERN_C_BEG ECB_EXTERN_C { 134 #define ECB_EXTERN_C_BEG ECB_EXTERN_C {
92 #define ECB_EXTERN_C_END } 135 #define ECB_EXTERN_C_END }
105 #define ECB_NO_SMP 1 148 #define ECB_NO_SMP 1
106#endif 149#endif
107 150
108#if ECB_NO_SMP 151#if ECB_NO_SMP
109 #define ECB_MEMORY_FENCE do { } while (0) 152 #define ECB_MEMORY_FENCE do { } while (0)
153#endif
154
155/* http://www-01.ibm.com/support/knowledgecenter/SSGH3R_13.1.0/com.ibm.xlcpp131.aix.doc/compiler_ref/compiler_builtins.html */
156#if __xlC__ && ECB_CPP
157 #include <builtins.h>
158#endif
159
160#if 1400 <= _MSC_VER
161 #include <intrin.h> /* fence functions _ReadBarrier, also bit search functions _BitScanReverse */
110#endif 162#endif
111 163
112#ifndef ECB_MEMORY_FENCE 164#ifndef ECB_MEMORY_FENCE
113 #if ECB_GCC_VERSION(2,5) || defined __INTEL_COMPILER || (__llvm__ && __GNUC__) || __SUNPRO_C >= 0x5110 || __SUNPRO_CC >= 0x5110 165 #if ECB_GCC_VERSION(2,5) || defined __INTEL_COMPILER || (__llvm__ && __GNUC__) || __SUNPRO_C >= 0x5110 || __SUNPRO_CC >= 0x5110
114 #if __i386 || __i386__ 166 #if __i386 || __i386__
115 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("lock; orb $0, -1(%%esp)" : : : "memory") 167 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("lock; orb $0, -1(%%esp)" : : : "memory")
116 #define ECB_MEMORY_FENCE_ACQUIRE __asm__ __volatile__ ("" : : : "memory") 168 #define ECB_MEMORY_FENCE_ACQUIRE __asm__ __volatile__ ("" : : : "memory")
117 #define ECB_MEMORY_FENCE_RELEASE __asm__ __volatile__ ("") 169 #define ECB_MEMORY_FENCE_RELEASE __asm__ __volatile__ ("")
118 #elif __amd64 || __amd64__ || __x86_64 || __x86_64__ 170 #elif ECB_GCC_AMD64
119 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("mfence" : : : "memory") 171 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("mfence" : : : "memory")
120 #define ECB_MEMORY_FENCE_ACQUIRE __asm__ __volatile__ ("" : : : "memory") 172 #define ECB_MEMORY_FENCE_ACQUIRE __asm__ __volatile__ ("" : : : "memory")
121 #define ECB_MEMORY_FENCE_RELEASE __asm__ __volatile__ ("") 173 #define ECB_MEMORY_FENCE_RELEASE __asm__ __volatile__ ("")
122 #elif __powerpc__ || __ppc__ || __powerpc64__ || __ppc64__ 174 #elif __powerpc__ || __ppc__ || __powerpc64__ || __ppc64__
123 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("sync" : : : "memory") 175 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("sync" : : : "memory")
176 #elif defined __ARM_ARCH_2__ \
177 || defined __ARM_ARCH_3__ || defined __ARM_ARCH_3M__ \
178 || defined __ARM_ARCH_4__ || defined __ARM_ARCH_4T__ \
179 || defined __ARM_ARCH_5__ || defined __ARM_ARCH_5E__ \
180 || defined __ARM_ARCH_5T__ || defined __ARM_ARCH_5TE__ \
181 || defined __ARM_ARCH_5TEJ__
182 /* should not need any, unless running old code on newer cpu - arm doesn't support that */
124 #elif defined __ARM_ARCH_6__ || defined __ARM_ARCH_6J__ \ 183 #elif defined __ARM_ARCH_6__ || defined __ARM_ARCH_6J__ \
125 || defined __ARM_ARCH_6K__ || defined __ARM_ARCH_6ZK__ 184 || defined __ARM_ARCH_6K__ || defined __ARM_ARCH_6ZK__ \
185 || defined __ARM_ARCH_6T2__
126 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("mcr p15,0,%0,c7,c10,5" : : "r" (0) : "memory") 186 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("mcr p15,0,%0,c7,c10,5" : : "r" (0) : "memory")
127 #elif defined __ARM_ARCH_7__ || defined __ARM_ARCH_7A__ \ 187 #elif defined __ARM_ARCH_7__ || defined __ARM_ARCH_7A__ \
128 || defined __ARM_ARCH_7M__ || defined __ARM_ARCH_7R__ 188 || defined __ARM_ARCH_7R__ || defined __ARM_ARCH_7M__
129 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("dmb" : : : "memory") 189 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("dmb" : : : "memory")
130 #elif __sparc || __sparc__ 190 #elif __aarch64__
191 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("dmb ish" : : : "memory")
192 #elif (__sparc || __sparc__) && !(__sparc_v8__ || defined __sparcv8)
131 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("membar #LoadStore | #LoadLoad | #StoreStore | #StoreLoad" : : : "memory") 193 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("membar #LoadStore | #LoadLoad | #StoreStore | #StoreLoad" : : : "memory")
132 #define ECB_MEMORY_FENCE_ACQUIRE __asm__ __volatile__ ("membar #LoadStore | #LoadLoad" : : : "memory") 194 #define ECB_MEMORY_FENCE_ACQUIRE __asm__ __volatile__ ("membar #LoadStore | #LoadLoad" : : : "memory")
133 #define ECB_MEMORY_FENCE_RELEASE __asm__ __volatile__ ("membar #LoadStore | #StoreStore") 195 #define ECB_MEMORY_FENCE_RELEASE __asm__ __volatile__ ("membar #LoadStore | #StoreStore")
134 #elif defined __s390__ || defined __s390x__ 196 #elif defined __s390__ || defined __s390x__
135 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("bcr 15,0" : : : "memory") 197 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("bcr 15,0" : : : "memory")
136 #elif defined __mips__ 198 #elif defined __mips__
199 /* GNU/Linux emulates sync on mips1 architectures, so we force its use */
200 /* anybody else who still uses mips1 is supposed to send in their version, with detection code. */
137 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("sync" : : : "memory") 201 #define ECB_MEMORY_FENCE __asm__ __volatile__ (".set mips2; sync; .set mips0" : : : "memory")
138 #elif defined __alpha__ 202 #elif defined __alpha__
139 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("mb" : : : "memory") 203 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("mb" : : : "memory")
140 #elif defined __hppa__ 204 #elif defined __hppa__
141 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("" : : : "memory") 205 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("" : : : "memory")
142 #define ECB_MEMORY_FENCE_RELEASE __asm__ __volatile__ ("") 206 #define ECB_MEMORY_FENCE_RELEASE __asm__ __volatile__ ("")
143 #elif defined __ia64__ 207 #elif defined __ia64__
144 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("mf" : : : "memory") 208 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("mf" : : : "memory")
209 #elif defined __m68k__
210 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("" : : : "memory")
211 #elif defined __m88k__
212 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("tb1 0,%%r0,128" : : : "memory")
213 #elif defined __sh__
214 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("" : : : "memory")
145 #endif 215 #endif
146 #endif 216 #endif
147#endif 217#endif
148 218
149#ifndef ECB_MEMORY_FENCE 219#ifndef ECB_MEMORY_FENCE
150 #if ECB_GCC_VERSION(4,7) 220 #if ECB_GCC_VERSION(4,7)
151 /* see comment below (stdatomic.h) about the C11 memory model. */ 221 /* see comment below (stdatomic.h) about the C11 memory model. */
152 #define ECB_MEMORY_FENCE __atomic_thread_fence (__ATOMIC_SEQ_CST) 222 #define ECB_MEMORY_FENCE __atomic_thread_fence (__ATOMIC_SEQ_CST)
223 #define ECB_MEMORY_FENCE_ACQUIRE __atomic_thread_fence (__ATOMIC_ACQUIRE)
224 #define ECB_MEMORY_FENCE_RELEASE __atomic_thread_fence (__ATOMIC_RELEASE)
153 225
154 /* The __has_feature syntax from clang is so misdesigned that we cannot use it 226 #elif ECB_CLANG_EXTENSION(c_atomic)
155 * without risking compile time errors with other compilers. We *could*
156 * define our own ecb_clang_has_feature, but I just can't be bothered to work
157 * around * this shit time and again.
158 * #elif defined __clang && __has_feature (cxx_atomic)
159 * // see comment below (stdatomic.h) about the C11 memory model. 227 /* see comment below (stdatomic.h) about the C11 memory model. */
160 * #define ECB_MEMORY_FENCE __c11_atomic_thread_fence (__ATOMIC_SEQ_CST) 228 #define ECB_MEMORY_FENCE __c11_atomic_thread_fence (__ATOMIC_SEQ_CST)
161 */ 229 #define ECB_MEMORY_FENCE_ACQUIRE __c11_atomic_thread_fence (__ATOMIC_ACQUIRE)
230 #define ECB_MEMORY_FENCE_RELEASE __c11_atomic_thread_fence (__ATOMIC_RELEASE)
162 231
163 #elif ECB_GCC_VERSION(4,4) || defined __INTEL_COMPILER || defined __clang__ 232 #elif ECB_GCC_VERSION(4,4) || defined __INTEL_COMPILER || defined __clang__
164 #define ECB_MEMORY_FENCE __sync_synchronize () 233 #define ECB_MEMORY_FENCE __sync_synchronize ()
234 #elif _MSC_VER >= 1500 /* VC++ 2008 */
235 /* apparently, microsoft broke all the memory barrier stuff in Visual Studio 2008... */
236 #pragma intrinsic(_ReadBarrier,_WriteBarrier,_ReadWriteBarrier)
237 #define ECB_MEMORY_FENCE _ReadWriteBarrier (); MemoryBarrier()
238 #define ECB_MEMORY_FENCE_ACQUIRE _ReadWriteBarrier (); MemoryBarrier() /* according to msdn, _ReadBarrier is not a load fence */
239 #define ECB_MEMORY_FENCE_RELEASE _WriteBarrier (); MemoryBarrier()
165 #elif _MSC_VER >= 1400 /* VC++ 2005 */ 240 #elif _MSC_VER >= 1400 /* VC++ 2005 */
166 #pragma intrinsic(_ReadBarrier,_WriteBarrier,_ReadWriteBarrier) 241 #pragma intrinsic(_ReadBarrier,_WriteBarrier,_ReadWriteBarrier)
167 #define ECB_MEMORY_FENCE _ReadWriteBarrier () 242 #define ECB_MEMORY_FENCE _ReadWriteBarrier ()
168 #define ECB_MEMORY_FENCE_ACQUIRE _ReadWriteBarrier () /* according to msdn, _ReadBarrier is not a load fence */ 243 #define ECB_MEMORY_FENCE_ACQUIRE _ReadWriteBarrier () /* according to msdn, _ReadBarrier is not a load fence */
169 #define ECB_MEMORY_FENCE_RELEASE _WriteBarrier () 244 #define ECB_MEMORY_FENCE_RELEASE _WriteBarrier ()
189 /* any fence other than seq_cst, which isn't very efficient for us. */ 264 /* any fence other than seq_cst, which isn't very efficient for us. */
190 /* Why that is, we don't know - either the C11 memory model is quite useless */ 265 /* Why that is, we don't know - either the C11 memory model is quite useless */
191 /* for most usages, or gcc and clang have a bug */ 266 /* for most usages, or gcc and clang have a bug */
192 /* I *currently* lean towards the latter, and inefficiently implement */ 267 /* I *currently* lean towards the latter, and inefficiently implement */
193 /* all three of ecb's fences as a seq_cst fence */ 268 /* all three of ecb's fences as a seq_cst fence */
269 /* Update, gcc-4.8 generates mfence for all c++ fences, but nothing */
270 /* for all __atomic_thread_fence's except seq_cst */
194 #define ECB_MEMORY_FENCE atomic_thread_fence (memory_order_seq_cst) 271 #define ECB_MEMORY_FENCE atomic_thread_fence (memory_order_seq_cst)
195 #endif 272 #endif
196#endif 273#endif
197 274
198#ifndef ECB_MEMORY_FENCE 275#ifndef ECB_MEMORY_FENCE
221 #define ECB_MEMORY_FENCE_RELEASE ECB_MEMORY_FENCE 298 #define ECB_MEMORY_FENCE_RELEASE ECB_MEMORY_FENCE
222#endif 299#endif
223 300
224/*****************************************************************************/ 301/*****************************************************************************/
225 302
226#if __cplusplus 303#if ECB_CPP
227 #define ecb_inline static inline 304 #define ecb_inline static inline
228#elif ECB_GCC_VERSION(2,5) 305#elif ECB_GCC_VERSION(2,5)
229 #define ecb_inline static __inline__ 306 #define ecb_inline static __inline__
230#elif ECB_C99 307#elif ECB_C99
231 #define ecb_inline static inline 308 #define ecb_inline static inline
245 322
246#define ECB_CONCAT_(a, b) a ## b 323#define ECB_CONCAT_(a, b) a ## b
247#define ECB_CONCAT(a, b) ECB_CONCAT_(a, b) 324#define ECB_CONCAT(a, b) ECB_CONCAT_(a, b)
248#define ECB_STRINGIFY_(a) # a 325#define ECB_STRINGIFY_(a) # a
249#define ECB_STRINGIFY(a) ECB_STRINGIFY_(a) 326#define ECB_STRINGIFY(a) ECB_STRINGIFY_(a)
327#define ECB_STRINGIFY_EXPR(expr) ((expr), ECB_STRINGIFY_ (expr))
250 328
251#define ecb_function_ ecb_inline 329#define ecb_function_ ecb_inline
252 330
253#if ECB_GCC_VERSION(3,1) 331#if ECB_GCC_VERSION(3,1) || ECB_CLANG_VERSION(2,8)
254 #define ecb_attribute(attrlist) __attribute__(attrlist) 332 #define ecb_attribute(attrlist) __attribute__ (attrlist)
333#else
334 #define ecb_attribute(attrlist)
335#endif
336
337#if ECB_GCC_VERSION(3,1) || ECB_CLANG_BUILTIN(__builtin_constant_p)
255 #define ecb_is_constant(expr) __builtin_constant_p (expr) 338 #define ecb_is_constant(expr) __builtin_constant_p (expr)
339#else
340 /* possible C11 impl for integral types
341 typedef struct ecb_is_constant_struct ecb_is_constant_struct;
342 #define ecb_is_constant(expr) _Generic ((1 ? (struct ecb_is_constant_struct *)0 : (void *)((expr) - (expr)), ecb_is_constant_struct *: 0, default: 1)) */
343
344 #define ecb_is_constant(expr) 0
345#endif
346
347#if ECB_GCC_VERSION(3,1) || ECB_CLANG_BUILTIN(__builtin_expect)
256 #define ecb_expect(expr,value) __builtin_expect ((expr),(value)) 348 #define ecb_expect(expr,value) __builtin_expect ((expr),(value))
349#else
350 #define ecb_expect(expr,value) (expr)
351#endif
352
353#if ECB_GCC_VERSION(3,1) || ECB_CLANG_BUILTIN(__builtin_prefetch)
257 #define ecb_prefetch(addr,rw,locality) __builtin_prefetch (addr, rw, locality) 354 #define ecb_prefetch(addr,rw,locality) __builtin_prefetch (addr, rw, locality)
258#else 355#else
259 #define ecb_attribute(attrlist)
260 #define ecb_is_constant(expr) 0
261 #define ecb_expect(expr,value) (expr)
262 #define ecb_prefetch(addr,rw,locality) 356 #define ecb_prefetch(addr,rw,locality)
263#endif 357#endif
264 358
265/* no emulation for ecb_decltype */ 359/* no emulation for ecb_decltype */
266#if ECB_GCC_VERSION(4,5) 360#if ECB_CPP11
361 // older implementations might have problems with decltype(x)::type, work around it
362 template<class T> struct ecb_decltype_t { typedef T type; };
267 #define ecb_decltype(x) __decltype(x) 363 #define ecb_decltype(x) ecb_decltype_t<decltype (x)>::type
268#elif ECB_GCC_VERSION(3,0) 364#elif ECB_GCC_VERSION(3,0) || ECB_CLANG_VERSION(2,8)
269 #define ecb_decltype(x) __typeof(x) 365 #define ecb_decltype(x) __typeof__ (x)
270#endif 366#endif
271 367
368#if _MSC_VER >= 1300
369 #define ecb_deprecated __declspec (deprecated)
370#else
371 #define ecb_deprecated ecb_attribute ((__deprecated__))
372#endif
373
374#if _MSC_VER >= 1500
375 #define ecb_deprecated_message(msg) __declspec (deprecated (msg))
376#elif ECB_GCC_VERSION(4,5)
377 #define ecb_deprecated_message(msg) ecb_attribute ((__deprecated__ (msg))
378#else
379 #define ecb_deprecated_message(msg) ecb_deprecated
380#endif
381
382#if _MSC_VER >= 1400
383 #define ecb_noinline __declspec (noinline)
384#else
272#define ecb_noinline ecb_attribute ((__noinline__)) 385 #define ecb_noinline ecb_attribute ((__noinline__))
386#endif
387
273#define ecb_unused ecb_attribute ((__unused__)) 388#define ecb_unused ecb_attribute ((__unused__))
274#define ecb_const ecb_attribute ((__const__)) 389#define ecb_const ecb_attribute ((__const__))
275#define ecb_pure ecb_attribute ((__pure__)) 390#define ecb_pure ecb_attribute ((__pure__))
276 391
277#if ECB_C11 392#if ECB_C11 || __IBMC_NORETURN
393 /* http://www-01.ibm.com/support/knowledgecenter/SSGH3R_13.1.0/com.ibm.xlcpp131.aix.doc/language_ref/noreturn.html */
278 #define ecb_noreturn _Noreturn 394 #define ecb_noreturn _Noreturn
395#elif ECB_CPP11
396 #define ecb_noreturn [[noreturn]]
397#elif _MSC_VER >= 1200
398 /* http://msdn.microsoft.com/en-us/library/k6ktzx3s.aspx */
399 #define ecb_noreturn __declspec (noreturn)
279#else 400#else
280 #define ecb_noreturn ecb_attribute ((__noreturn__)) 401 #define ecb_noreturn ecb_attribute ((__noreturn__))
281#endif 402#endif
282 403
283#if ECB_GCC_VERSION(4,3) 404#if ECB_GCC_VERSION(4,3)
298/* for compatibility to the rest of the world */ 419/* for compatibility to the rest of the world */
299#define ecb_likely(expr) ecb_expect_true (expr) 420#define ecb_likely(expr) ecb_expect_true (expr)
300#define ecb_unlikely(expr) ecb_expect_false (expr) 421#define ecb_unlikely(expr) ecb_expect_false (expr)
301 422
302/* count trailing zero bits and count # of one bits */ 423/* count trailing zero bits and count # of one bits */
303#if ECB_GCC_VERSION(3,4) 424#if ECB_GCC_VERSION(3,4) \
425 || (ECB_CLANG_BUILTIN(__builtin_clz) && ECB_CLANG_BUILTIN(__builtin_clzll) \
426 && ECB_CLANG_BUILTIN(__builtin_ctz) && ECB_CLANG_BUILTIN(__builtin_ctzll) \
427 && ECB_CLANG_BUILTIN(__builtin_popcount))
304 /* we assume int == 32 bit, long == 32 or 64 bit and long long == 64 bit */ 428 /* we assume int == 32 bit, long == 32 or 64 bit and long long == 64 bit */
305 #define ecb_ld32(x) (__builtin_clz (x) ^ 31) 429 #define ecb_ld32(x) (__builtin_clz (x) ^ 31)
306 #define ecb_ld64(x) (__builtin_clzll (x) ^ 63) 430 #define ecb_ld64(x) (__builtin_clzll (x) ^ 63)
307 #define ecb_ctz32(x) __builtin_ctz (x) 431 #define ecb_ctz32(x) __builtin_ctz (x)
308 #define ecb_ctz64(x) __builtin_ctzll (x) 432 #define ecb_ctz64(x) __builtin_ctzll (x)
309 #define ecb_popcount32(x) __builtin_popcount (x) 433 #define ecb_popcount32(x) __builtin_popcount (x)
310 /* no popcountll */ 434 /* no popcountll */
311#else 435#else
312 ecb_function_ int ecb_ctz32 (uint32_t x) ecb_const; 436 ecb_function_ ecb_const int ecb_ctz32 (uint32_t x);
313 ecb_function_ int 437 ecb_function_ ecb_const int
314 ecb_ctz32 (uint32_t x) 438 ecb_ctz32 (uint32_t x)
315 { 439 {
440#if 1400 <= _MSC_VER && (_M_IX86 || _M_X64 || _M_IA64 || _M_ARM)
441 unsigned long r;
442 _BitScanForward (&r, x);
443 return (int)r;
444#else
316 int r = 0; 445 int r = 0;
317 446
318 x &= ~x + 1; /* this isolates the lowest bit */ 447 x &= ~x + 1; /* this isolates the lowest bit */
319 448
320#if ECB_branchless_on_i386 449#if ECB_branchless_on_i386
330 if (x & 0xff00ff00) r += 8; 459 if (x & 0xff00ff00) r += 8;
331 if (x & 0xffff0000) r += 16; 460 if (x & 0xffff0000) r += 16;
332#endif 461#endif
333 462
334 return r; 463 return r;
464#endif
335 } 465 }
336 466
337 ecb_function_ int ecb_ctz64 (uint64_t x) ecb_const; 467 ecb_function_ ecb_const int ecb_ctz64 (uint64_t x);
338 ecb_function_ int 468 ecb_function_ ecb_const int
339 ecb_ctz64 (uint64_t x) 469 ecb_ctz64 (uint64_t x)
340 { 470 {
471#if 1400 <= _MSC_VER && (_M_X64 || _M_IA64 || _M_ARM)
472 unsigned long r;
473 _BitScanForward64 (&r, x);
474 return (int)r;
475#else
341 int shift = x & 0xffffffffU ? 0 : 32; 476 int shift = x & 0xffffffff ? 0 : 32;
342 return ecb_ctz32 (x >> shift) + shift; 477 return ecb_ctz32 (x >> shift) + shift;
478#endif
343 } 479 }
344 480
345 ecb_function_ int ecb_popcount32 (uint32_t x) ecb_const; 481 ecb_function_ ecb_const int ecb_popcount32 (uint32_t x);
346 ecb_function_ int 482 ecb_function_ ecb_const int
347 ecb_popcount32 (uint32_t x) 483 ecb_popcount32 (uint32_t x)
348 { 484 {
349 x -= (x >> 1) & 0x55555555; 485 x -= (x >> 1) & 0x55555555;
350 x = ((x >> 2) & 0x33333333) + (x & 0x33333333); 486 x = ((x >> 2) & 0x33333333) + (x & 0x33333333);
351 x = ((x >> 4) + x) & 0x0f0f0f0f; 487 x = ((x >> 4) + x) & 0x0f0f0f0f;
352 x *= 0x01010101; 488 x *= 0x01010101;
353 489
354 return x >> 24; 490 return x >> 24;
355 } 491 }
356 492
357 ecb_function_ int ecb_ld32 (uint32_t x) ecb_const; 493 ecb_function_ ecb_const int ecb_ld32 (uint32_t x);
358 ecb_function_ int ecb_ld32 (uint32_t x) 494 ecb_function_ ecb_const int ecb_ld32 (uint32_t x)
359 { 495 {
496#if 1400 <= _MSC_VER && (_M_IX86 || _M_X64 || _M_IA64 || _M_ARM)
497 unsigned long r;
498 _BitScanReverse (&r, x);
499 return (int)r;
500#else
360 int r = 0; 501 int r = 0;
361 502
362 if (x >> 16) { x >>= 16; r += 16; } 503 if (x >> 16) { x >>= 16; r += 16; }
363 if (x >> 8) { x >>= 8; r += 8; } 504 if (x >> 8) { x >>= 8; r += 8; }
364 if (x >> 4) { x >>= 4; r += 4; } 505 if (x >> 4) { x >>= 4; r += 4; }
365 if (x >> 2) { x >>= 2; r += 2; } 506 if (x >> 2) { x >>= 2; r += 2; }
366 if (x >> 1) { r += 1; } 507 if (x >> 1) { r += 1; }
367 508
368 return r; 509 return r;
510#endif
369 } 511 }
370 512
371 ecb_function_ int ecb_ld64 (uint64_t x) ecb_const; 513 ecb_function_ ecb_const int ecb_ld64 (uint64_t x);
372 ecb_function_ int ecb_ld64 (uint64_t x) 514 ecb_function_ ecb_const int ecb_ld64 (uint64_t x)
373 { 515 {
516#if 1400 <= _MSC_VER && (_M_X64 || _M_IA64 || _M_ARM)
517 unsigned long r;
518 _BitScanReverse64 (&r, x);
519 return (int)r;
520#else
374 int r = 0; 521 int r = 0;
375 522
376 if (x >> 32) { x >>= 32; r += 32; } 523 if (x >> 32) { x >>= 32; r += 32; }
377 524
378 return r + ecb_ld32 (x); 525 return r + ecb_ld32 (x);
379 }
380#endif 526#endif
527 }
528#endif
381 529
382ecb_function_ ecb_bool ecb_is_pot32 (uint32_t x) ecb_const; 530ecb_function_ ecb_const ecb_bool ecb_is_pot32 (uint32_t x);
383ecb_function_ ecb_bool ecb_is_pot32 (uint32_t x) { return !(x & (x - 1)); } 531ecb_function_ ecb_const ecb_bool ecb_is_pot32 (uint32_t x) { return !(x & (x - 1)); }
384ecb_function_ ecb_bool ecb_is_pot64 (uint64_t x) ecb_const; 532ecb_function_ ecb_const ecb_bool ecb_is_pot64 (uint64_t x);
385ecb_function_ ecb_bool ecb_is_pot64 (uint64_t x) { return !(x & (x - 1)); } 533ecb_function_ ecb_const ecb_bool ecb_is_pot64 (uint64_t x) { return !(x & (x - 1)); }
386 534
387ecb_function_ uint8_t ecb_bitrev8 (uint8_t x) ecb_const; 535ecb_function_ ecb_const uint8_t ecb_bitrev8 (uint8_t x);
388ecb_function_ uint8_t ecb_bitrev8 (uint8_t x) 536ecb_function_ ecb_const uint8_t ecb_bitrev8 (uint8_t x)
389{ 537{
390 return ( (x * 0x0802U & 0x22110U) 538 return ( (x * 0x0802U & 0x22110U)
391 | (x * 0x8020U & 0x88440U)) * 0x10101U >> 16; 539 | (x * 0x8020U & 0x88440U)) * 0x10101U >> 16;
392} 540}
393 541
394ecb_function_ uint16_t ecb_bitrev16 (uint16_t x) ecb_const; 542ecb_function_ ecb_const uint16_t ecb_bitrev16 (uint16_t x);
395ecb_function_ uint16_t ecb_bitrev16 (uint16_t x) 543ecb_function_ ecb_const uint16_t ecb_bitrev16 (uint16_t x)
396{ 544{
397 x = ((x >> 1) & 0x5555) | ((x & 0x5555) << 1); 545 x = ((x >> 1) & 0x5555) | ((x & 0x5555) << 1);
398 x = ((x >> 2) & 0x3333) | ((x & 0x3333) << 2); 546 x = ((x >> 2) & 0x3333) | ((x & 0x3333) << 2);
399 x = ((x >> 4) & 0x0f0f) | ((x & 0x0f0f) << 4); 547 x = ((x >> 4) & 0x0f0f) | ((x & 0x0f0f) << 4);
400 x = ( x >> 8 ) | ( x << 8); 548 x = ( x >> 8 ) | ( x << 8);
401 549
402 return x; 550 return x;
403} 551}
404 552
405ecb_function_ uint32_t ecb_bitrev32 (uint32_t x) ecb_const; 553ecb_function_ ecb_const uint32_t ecb_bitrev32 (uint32_t x);
406ecb_function_ uint32_t ecb_bitrev32 (uint32_t x) 554ecb_function_ ecb_const uint32_t ecb_bitrev32 (uint32_t x)
407{ 555{
408 x = ((x >> 1) & 0x55555555) | ((x & 0x55555555) << 1); 556 x = ((x >> 1) & 0x55555555) | ((x & 0x55555555) << 1);
409 x = ((x >> 2) & 0x33333333) | ((x & 0x33333333) << 2); 557 x = ((x >> 2) & 0x33333333) | ((x & 0x33333333) << 2);
410 x = ((x >> 4) & 0x0f0f0f0f) | ((x & 0x0f0f0f0f) << 4); 558 x = ((x >> 4) & 0x0f0f0f0f) | ((x & 0x0f0f0f0f) << 4);
411 x = ((x >> 8) & 0x00ff00ff) | ((x & 0x00ff00ff) << 8); 559 x = ((x >> 8) & 0x00ff00ff) | ((x & 0x00ff00ff) << 8);
414 return x; 562 return x;
415} 563}
416 564
417/* popcount64 is only available on 64 bit cpus as gcc builtin */ 565/* popcount64 is only available on 64 bit cpus as gcc builtin */
418/* so for this version we are lazy */ 566/* so for this version we are lazy */
419ecb_function_ int ecb_popcount64 (uint64_t x) ecb_const; 567ecb_function_ ecb_const int ecb_popcount64 (uint64_t x);
420ecb_function_ int 568ecb_function_ ecb_const int
421ecb_popcount64 (uint64_t x) 569ecb_popcount64 (uint64_t x)
422{ 570{
423 return ecb_popcount32 (x) + ecb_popcount32 (x >> 32); 571 return ecb_popcount32 (x) + ecb_popcount32 (x >> 32);
424} 572}
425 573
426ecb_inline uint8_t ecb_rotl8 (uint8_t x, unsigned int count) ecb_const; 574ecb_inline ecb_const uint8_t ecb_rotl8 (uint8_t x, unsigned int count);
427ecb_inline uint8_t ecb_rotr8 (uint8_t x, unsigned int count) ecb_const; 575ecb_inline ecb_const uint8_t ecb_rotr8 (uint8_t x, unsigned int count);
428ecb_inline uint16_t ecb_rotl16 (uint16_t x, unsigned int count) ecb_const; 576ecb_inline ecb_const uint16_t ecb_rotl16 (uint16_t x, unsigned int count);
429ecb_inline uint16_t ecb_rotr16 (uint16_t x, unsigned int count) ecb_const; 577ecb_inline ecb_const uint16_t ecb_rotr16 (uint16_t x, unsigned int count);
430ecb_inline uint32_t ecb_rotl32 (uint32_t x, unsigned int count) ecb_const; 578ecb_inline ecb_const uint32_t ecb_rotl32 (uint32_t x, unsigned int count);
431ecb_inline uint32_t ecb_rotr32 (uint32_t x, unsigned int count) ecb_const; 579ecb_inline ecb_const uint32_t ecb_rotr32 (uint32_t x, unsigned int count);
432ecb_inline uint64_t ecb_rotl64 (uint64_t x, unsigned int count) ecb_const; 580ecb_inline ecb_const uint64_t ecb_rotl64 (uint64_t x, unsigned int count);
433ecb_inline uint64_t ecb_rotr64 (uint64_t x, unsigned int count) ecb_const; 581ecb_inline ecb_const uint64_t ecb_rotr64 (uint64_t x, unsigned int count);
434 582
435ecb_inline uint8_t ecb_rotl8 (uint8_t x, unsigned int count) { return (x >> ( 8 - count)) | (x << count); } 583ecb_inline ecb_const uint8_t ecb_rotl8 (uint8_t x, unsigned int count) { return (x >> ( 8 - count)) | (x << count); }
436ecb_inline uint8_t ecb_rotr8 (uint8_t x, unsigned int count) { return (x << ( 8 - count)) | (x >> count); } 584ecb_inline ecb_const uint8_t ecb_rotr8 (uint8_t x, unsigned int count) { return (x << ( 8 - count)) | (x >> count); }
437ecb_inline uint16_t ecb_rotl16 (uint16_t x, unsigned int count) { return (x >> (16 - count)) | (x << count); } 585ecb_inline ecb_const uint16_t ecb_rotl16 (uint16_t x, unsigned int count) { return (x >> (16 - count)) | (x << count); }
438ecb_inline uint16_t ecb_rotr16 (uint16_t x, unsigned int count) { return (x << (16 - count)) | (x >> count); } 586ecb_inline ecb_const uint16_t ecb_rotr16 (uint16_t x, unsigned int count) { return (x << (16 - count)) | (x >> count); }
439ecb_inline uint32_t ecb_rotl32 (uint32_t x, unsigned int count) { return (x >> (32 - count)) | (x << count); } 587ecb_inline ecb_const uint32_t ecb_rotl32 (uint32_t x, unsigned int count) { return (x >> (32 - count)) | (x << count); }
440ecb_inline uint32_t ecb_rotr32 (uint32_t x, unsigned int count) { return (x << (32 - count)) | (x >> count); } 588ecb_inline ecb_const uint32_t ecb_rotr32 (uint32_t x, unsigned int count) { return (x << (32 - count)) | (x >> count); }
441ecb_inline uint64_t ecb_rotl64 (uint64_t x, unsigned int count) { return (x >> (64 - count)) | (x << count); } 589ecb_inline ecb_const uint64_t ecb_rotl64 (uint64_t x, unsigned int count) { return (x >> (64 - count)) | (x << count); }
442ecb_inline uint64_t ecb_rotr64 (uint64_t x, unsigned int count) { return (x << (64 - count)) | (x >> count); } 590ecb_inline ecb_const uint64_t ecb_rotr64 (uint64_t x, unsigned int count) { return (x << (64 - count)) | (x >> count); }
443 591
444#if ECB_GCC_VERSION(4,3) 592#if ECB_GCC_VERSION(4,3) || (ECB_CLANG_BUILTIN(__builtin_bswap32) && ECB_CLANG_BUILTIN(__builtin_bswap64))
593 #if ECB_GCC_VERSION(4,8) || ECB_CLANG_BUILTIN(__builtin_bswap16)
594 #define ecb_bswap16(x) __builtin_bswap16 (x)
595 #else
445 #define ecb_bswap16(x) (__builtin_bswap32 (x) >> 16) 596 #define ecb_bswap16(x) (__builtin_bswap32 (x) >> 16)
597 #endif
446 #define ecb_bswap32(x) __builtin_bswap32 (x) 598 #define ecb_bswap32(x) __builtin_bswap32 (x)
447 #define ecb_bswap64(x) __builtin_bswap64 (x) 599 #define ecb_bswap64(x) __builtin_bswap64 (x)
600#elif _MSC_VER
601 #include <stdlib.h>
602 #define ecb_bswap16(x) ((uint16_t)_byteswap_ushort ((uint16_t)(x)))
603 #define ecb_bswap32(x) ((uint32_t)_byteswap_ulong ((uint32_t)(x)))
604 #define ecb_bswap64(x) ((uint64_t)_byteswap_uint64 ((uint64_t)(x)))
448#else 605#else
449 ecb_function_ uint16_t ecb_bswap16 (uint16_t x) ecb_const; 606 ecb_function_ ecb_const uint16_t ecb_bswap16 (uint16_t x);
450 ecb_function_ uint16_t 607 ecb_function_ ecb_const uint16_t
451 ecb_bswap16 (uint16_t x) 608 ecb_bswap16 (uint16_t x)
452 { 609 {
453 return ecb_rotl16 (x, 8); 610 return ecb_rotl16 (x, 8);
454 } 611 }
455 612
456 ecb_function_ uint32_t ecb_bswap32 (uint32_t x) ecb_const; 613 ecb_function_ ecb_const uint32_t ecb_bswap32 (uint32_t x);
457 ecb_function_ uint32_t 614 ecb_function_ ecb_const uint32_t
458 ecb_bswap32 (uint32_t x) 615 ecb_bswap32 (uint32_t x)
459 { 616 {
460 return (((uint32_t)ecb_bswap16 (x)) << 16) | ecb_bswap16 (x >> 16); 617 return (((uint32_t)ecb_bswap16 (x)) << 16) | ecb_bswap16 (x >> 16);
461 } 618 }
462 619
463 ecb_function_ uint64_t ecb_bswap64 (uint64_t x) ecb_const; 620 ecb_function_ ecb_const uint64_t ecb_bswap64 (uint64_t x);
464 ecb_function_ uint64_t 621 ecb_function_ ecb_const uint64_t
465 ecb_bswap64 (uint64_t x) 622 ecb_bswap64 (uint64_t x)
466 { 623 {
467 return (((uint64_t)ecb_bswap32 (x)) << 32) | ecb_bswap32 (x >> 32); 624 return (((uint64_t)ecb_bswap32 (x)) << 32) | ecb_bswap32 (x >> 32);
468 } 625 }
469#endif 626#endif
470 627
471#if ECB_GCC_VERSION(4,5) 628#if ECB_GCC_VERSION(4,5) || ECB_CLANG_BUILTIN(__builtin_unreachable)
472 #define ecb_unreachable() __builtin_unreachable () 629 #define ecb_unreachable() __builtin_unreachable ()
473#else 630#else
474 /* this seems to work fine, but gcc always emits a warning for it :/ */ 631 /* this seems to work fine, but gcc always emits a warning for it :/ */
475 ecb_inline void ecb_unreachable (void) ecb_noreturn; 632 ecb_inline ecb_noreturn void ecb_unreachable (void);
476 ecb_inline void ecb_unreachable (void) { } 633 ecb_inline ecb_noreturn void ecb_unreachable (void) { }
477#endif 634#endif
478 635
479/* try to tell the compiler that some condition is definitely true */ 636/* try to tell the compiler that some condition is definitely true */
480#define ecb_assume(cond) if (!(cond)) ecb_unreachable (); else 0 637#define ecb_assume(cond) if (!(cond)) ecb_unreachable (); else 0
481 638
482ecb_inline unsigned char ecb_byteorder_helper (void) ecb_const; 639ecb_inline ecb_const uint32_t ecb_byteorder_helper (void);
483ecb_inline unsigned char 640ecb_inline ecb_const uint32_t
484ecb_byteorder_helper (void) 641ecb_byteorder_helper (void)
485{ 642{
486 /* the union code still generates code under pressure in gcc, */ 643 /* the union code still generates code under pressure in gcc, */
487 /* but less than using pointers, and always seem to */ 644 /* but less than using pointers, and always seems to */
488 /* successfully return a constant. */ 645 /* successfully return a constant. */
489 /* the reason why we have this horrible preprocessor mess */ 646 /* the reason why we have this horrible preprocessor mess */
490 /* is to avoid it in all cases, at least on common architectures */ 647 /* is to avoid it in all cases, at least on common architectures */
491 /* and yes, gcc defines __BYTE_ORDER__, g++ does not */ 648 /* or when using a recent enough gcc version (>= 4.6) */
492#if __i386 || __i386__ || _M_X86 || __amd64 || __amd64__ || _M_X64
493 return 0x44;
494#elif __BYTE_ORDER__ && __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__ 649#if (defined __BYTE_ORDER__ && __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__) \
650 || ((__i386 || __i386__ || _M_IX86 || ECB_GCC_AMD64 || ECB_MSVC_AMD64) && !__VOS__)
651 #define ECB_LITTLE_ENDIAN 1
495 return 0x44; 652 return 0x44332211;
496#elif __BYTE_ORDER__ && __BYTE_ORDER__ == __ORDER_BIG_ENDIAN__ 653#elif (defined __BYTE_ORDER__ && __BYTE_ORDER__ == __ORDER_BIG_ENDIAN__) \
654 || ((__AARCH64EB__ || __MIPSEB__ || __ARMEB__) && !__VOS__)
655 #define ECB_BIG_ENDIAN 1
497 retrurn 0x11; 656 return 0x11223344;
498#else 657#else
499 union 658 union
500 { 659 {
660 uint8_t c[4];
501 uint32_t i; 661 uint32_t u;
502 uint8_t c;
503 } u = { 0x11223344 }; 662 } u = { 0x11, 0x22, 0x33, 0x44 };
504 return u.c; 663 return u.u;
505#endif 664#endif
506} 665}
507 666
508ecb_inline ecb_bool ecb_big_endian (void) ecb_const; 667ecb_inline ecb_const ecb_bool ecb_big_endian (void);
509ecb_inline ecb_bool ecb_big_endian (void) { return ecb_byteorder_helper () == 0x11; } 668ecb_inline ecb_const ecb_bool ecb_big_endian (void) { return ecb_byteorder_helper () == 0x11223344; }
510ecb_inline ecb_bool ecb_little_endian (void) ecb_const; 669ecb_inline ecb_const ecb_bool ecb_little_endian (void);
511ecb_inline ecb_bool ecb_little_endian (void) { return ecb_byteorder_helper () == 0x44; } 670ecb_inline ecb_const ecb_bool ecb_little_endian (void) { return ecb_byteorder_helper () == 0x44332211; }
512 671
513#if ECB_GCC_VERSION(3,0) || ECB_C99 672#if ECB_GCC_VERSION(3,0) || ECB_C99
514 #define ecb_mod(m,n) ((m) % (n) + ((m) % (n) < 0 ? (n) : 0)) 673 #define ecb_mod(m,n) ((m) % (n) + ((m) % (n) < 0 ? (n) : 0))
515#else 674#else
516 #define ecb_mod(m,n) ((m) < 0 ? ((n) - 1 - ((-1 - (m)) % (n))) : ((m) % (n))) 675 #define ecb_mod(m,n) ((m) < 0 ? ((n) - 1 - ((-1 - (m)) % (n))) : ((m) % (n)))
517#endif 676#endif
518 677
519#if __cplusplus 678#if ECB_CPP
520 template<typename T> 679 template<typename T>
521 static inline T ecb_div_rd (T val, T div) 680 static inline T ecb_div_rd (T val, T div)
522 { 681 {
523 return val < 0 ? - ((-val + div - 1) / div) : (val ) / div; 682 return val < 0 ? - ((-val + div - 1) / div) : (val ) / div;
524 } 683 }
541 } 700 }
542#else 701#else
543 #define ecb_array_length(name) (sizeof (name) / sizeof (name [0])) 702 #define ecb_array_length(name) (sizeof (name) / sizeof (name [0]))
544#endif 703#endif
545 704
705ecb_function_ ecb_const uint32_t ecb_binary16_to_binary32 (uint32_t x);
706ecb_function_ ecb_const uint32_t
707ecb_binary16_to_binary32 (uint32_t x)
708{
709 unsigned int s = (x & 0x8000) << (31 - 15);
710 int e = (x >> 10) & 0x001f;
711 unsigned int m = x & 0x03ff;
712
713 if (ecb_expect_false (e == 31))
714 /* infinity or NaN */
715 e = 255 - (127 - 15);
716 else if (ecb_expect_false (!e))
717 {
718 if (ecb_expect_true (!m))
719 /* zero, handled by code below by forcing e to 0 */
720 e = 0 - (127 - 15);
721 else
722 {
723 /* subnormal, renormalise */
724 unsigned int s = 10 - ecb_ld32 (m);
725
726 m = (m << s) & 0x3ff; /* mask implicit bit */
727 e -= s - 1;
728 }
729 }
730
731 /* e and m now are normalised, or zero, (or inf or nan) */
732 e += 127 - 15;
733
734 return s | (e << 23) | (m << (23 - 10));
735}
736
737ecb_function_ ecb_const uint16_t ecb_binary32_to_binary16 (uint32_t x);
738ecb_function_ ecb_const uint16_t
739ecb_binary32_to_binary16 (uint32_t x)
740{
741 unsigned int s = (x >> 16) & 0x00008000; /* sign bit, the easy part */
742 unsigned int e = ((x >> 23) & 0x000000ff) - (127 - 15); /* the desired exponent */
743 unsigned int m = x & 0x007fffff;
744
745 x &= 0x7fffffff;
746
747 /* if it's within range of binary16 normals, use fast path */
748 if (ecb_expect_true (0x38800000 <= x && x <= 0x477fefff))
749 {
750 /* mantissa round-to-even */
751 m += 0x00000fff + ((m >> (23 - 10)) & 1);
752
753 /* handle overflow */
754 if (ecb_expect_false (m >= 0x00800000))
755 {
756 m >>= 1;
757 e += 1;
758 }
759
760 return s | (e << 10) | (m >> (23 - 10));
761 }
762
763 /* handle large numbers and infinity */
764 if (ecb_expect_true (0x477fefff < x && x <= 0x7f800000))
765 return s | 0x7c00;
766
767 /* handle zero, subnormals and small numbers */
768 if (ecb_expect_true (x < 0x38800000))
769 {
770 /* zero */
771 if (ecb_expect_true (!x))
772 return s;
773
774 /* handle subnormals */
775
776 /* too small, will be zero */
777 if (e < (14 - 24)) /* might not be sharp, but is good enough */
778 return s;
779
780 m |= 0x00800000; /* make implicit bit explicit */
781
782 /* very tricky - we need to round to the nearest e (+10) bit value */
783 {
784 unsigned int bits = 14 - e;
785 unsigned int half = (1 << (bits - 1)) - 1;
786 unsigned int even = (m >> bits) & 1;
787
788 /* if this overflows, we will end up with a normalised number */
789 m = (m + half + even) >> bits;
790 }
791
792 return s | m;
793 }
794
795 /* handle NaNs, preserve leftmost nan bits, but make sure we don't turn them into infinities */
796 m >>= 13;
797
798 return s | 0x7c00 | m | !m;
799}
800
546/*******************************************************************************/ 801/*******************************************************************************/
547/* floating point stuff, can be disabled by defining ECB_NO_LIBM */ 802/* floating point stuff, can be disabled by defining ECB_NO_LIBM */
548 803
549/* basically, everything uses "ieee pure-endian" floating point numbers */ 804/* basically, everything uses "ieee pure-endian" floating point numbers */
550/* the only noteworthy exception is ancient armle, which uses order 43218765 */ 805/* the only noteworthy exception is ancient armle, which uses order 43218765 */
551#if 0 \ 806#if 0 \
552 || __i386 || __i386__ \ 807 || __i386 || __i386__ \
553 || __amd64 || __amd64__ || __x86_64 || __x86_64__ \ 808 || ECB_GCC_AMD64 \
554 || __powerpc__ || __ppc__ || __powerpc64__ || __ppc64__ \ 809 || __powerpc__ || __ppc__ || __powerpc64__ || __ppc64__ \
555 || defined __arm__ && defined __ARM_EABI__ \
556 || defined __s390__ || defined __s390x__ \ 810 || defined __s390__ || defined __s390x__ \
557 || defined __mips__ \ 811 || defined __mips__ \
558 || defined __alpha__ \ 812 || defined __alpha__ \
559 || defined __hppa__ \ 813 || defined __hppa__ \
560 || defined __ia64__ \ 814 || defined __ia64__ \
815 || defined __m68k__ \
816 || defined __m88k__ \
817 || defined __sh__ \
561 || defined _M_IX86 || defined _M_AMD64 || defined _M_IA64 818 || defined _M_IX86 || defined ECB_MSVC_AMD64 || defined _M_IA64 \
819 || (defined __arm__ && (defined __ARM_EABI__ || defined __EABI__ || defined __VFP_FP__ || defined _WIN32_WCE || defined __ANDROID__)) \
820 || defined __aarch64__
562 #define ECB_STDFP 1 821 #define ECB_STDFP 1
563 #include <string.h> /* for memcpy */ 822 #include <string.h> /* for memcpy */
564#else 823#else
565 #define ECB_STDFP 0 824 #define ECB_STDFP 0
566 #include <math.h> /* for frexp*, ldexp* */
567#endif 825#endif
568 826
569#ifndef ECB_NO_LIBM 827#ifndef ECB_NO_LIBM
570 828
829 #include <math.h> /* for frexp*, ldexp*, INFINITY, NAN */
830
831 /* only the oldest of old doesn't have this one. solaris. */
832 #ifdef INFINITY
833 #define ECB_INFINITY INFINITY
834 #else
835 #define ECB_INFINITY HUGE_VAL
836 #endif
837
838 #ifdef NAN
839 #define ECB_NAN NAN
840 #else
841 #define ECB_NAN ECB_INFINITY
842 #endif
843
844 #if ECB_C99 || _XOPEN_VERSION >= 600 || _POSIX_VERSION >= 200112L
845 #define ecb_ldexpf(x,e) ldexpf ((x), (e))
846 #define ecb_frexpf(x,e) frexpf ((x), (e))
847 #else
848 #define ecb_ldexpf(x,e) (float) ldexp ((double) (x), (e))
849 #define ecb_frexpf(x,e) (float) frexp ((double) (x), (e))
850 #endif
851
571 /* convert a float to ieee single/binary32 */ 852 /* convert a float to ieee single/binary32 */
572 ecb_function_ uint32_t ecb_float_to_binary32 (float x) ecb_const; 853 ecb_function_ ecb_const uint32_t ecb_float_to_binary32 (float x);
573 ecb_function_ uint32_t 854 ecb_function_ ecb_const uint32_t
574 ecb_float_to_binary32 (float x) 855 ecb_float_to_binary32 (float x)
575 { 856 {
576 uint32_t r; 857 uint32_t r;
577 858
578 #if ECB_STDFP 859 #if ECB_STDFP
585 if (x == 0e0f ) return 0x00000000U; 866 if (x == 0e0f ) return 0x00000000U;
586 if (x > +3.40282346638528860e+38f) return 0x7f800000U; 867 if (x > +3.40282346638528860e+38f) return 0x7f800000U;
587 if (x < -3.40282346638528860e+38f) return 0xff800000U; 868 if (x < -3.40282346638528860e+38f) return 0xff800000U;
588 if (x != x ) return 0x7fbfffffU; 869 if (x != x ) return 0x7fbfffffU;
589 870
590 m = frexpf (x, &e) * 0x1000000U; 871 m = ecb_frexpf (x, &e) * 0x1000000U;
591 872
592 r = m & 0x80000000U; 873 r = m & 0x80000000U;
593 874
594 if (r) 875 if (r)
595 m = -m; 876 m = -m;
607 888
608 return r; 889 return r;
609 } 890 }
610 891
611 /* converts an ieee single/binary32 to a float */ 892 /* converts an ieee single/binary32 to a float */
612 ecb_function_ float ecb_binary32_to_float (uint32_t x) ecb_const; 893 ecb_function_ ecb_const float ecb_binary32_to_float (uint32_t x);
613 ecb_function_ float 894 ecb_function_ ecb_const float
614 ecb_binary32_to_float (uint32_t x) 895 ecb_binary32_to_float (uint32_t x)
615 { 896 {
616 float r; 897 float r;
617 898
618 #if ECB_STDFP 899 #if ECB_STDFP
628 x |= 0x800000U; 909 x |= 0x800000U;
629 else 910 else
630 e = 1; 911 e = 1;
631 912
632 /* we distrust ldexpf a bit and do the 2**-24 scaling by an extra multiply */ 913 /* we distrust ldexpf a bit and do the 2**-24 scaling by an extra multiply */
633 r = ldexpf (x * (0.5f / 0x800000U), e - 126); 914 r = ecb_ldexpf (x * (0.5f / 0x800000U), e - 126);
634 915
635 r = neg ? -r : r; 916 r = neg ? -r : r;
636 #endif 917 #endif
637 918
638 return r; 919 return r;
639 } 920 }
640 921
641 /* convert a double to ieee double/binary64 */ 922 /* convert a double to ieee double/binary64 */
642 ecb_function_ uint64_t ecb_double_to_binary64 (double x) ecb_const; 923 ecb_function_ ecb_const uint64_t ecb_double_to_binary64 (double x);
643 ecb_function_ uint64_t 924 ecb_function_ ecb_const uint64_t
644 ecb_double_to_binary64 (double x) 925 ecb_double_to_binary64 (double x)
645 { 926 {
646 uint64_t r; 927 uint64_t r;
647 928
648 #if ECB_STDFP 929 #if ECB_STDFP
677 958
678 return r; 959 return r;
679 } 960 }
680 961
681 /* converts an ieee double/binary64 to a double */ 962 /* converts an ieee double/binary64 to a double */
682 ecb_function_ double ecb_binary64_to_double (uint64_t x) ecb_const; 963 ecb_function_ ecb_const double ecb_binary64_to_double (uint64_t x);
683 ecb_function_ double 964 ecb_function_ ecb_const double
684 ecb_binary64_to_double (uint64_t x) 965 ecb_binary64_to_double (uint64_t x)
685 { 966 {
686 double r; 967 double r;
687 968
688 #if ECB_STDFP 969 #if ECB_STDFP
706 #endif 987 #endif
707 988
708 return r; 989 return r;
709 } 990 }
710 991
711#endif 992 /* convert a float to ieee half/binary16 */
993 ecb_function_ ecb_const uint16_t ecb_float_to_binary16 (float x);
994 ecb_function_ ecb_const uint16_t
995 ecb_float_to_binary16 (float x)
996 {
997 return ecb_binary32_to_binary16 (ecb_float_to_binary32 (x));
998 }
712 999
713#endif 1000 /* convert an ieee half/binary16 to float */
1001 ecb_function_ ecb_const float ecb_binary16_to_float (uint16_t x);
1002 ecb_function_ ecb_const float
1003 ecb_binary16_to_float (uint16_t x)
1004 {
1005 return ecb_binary32_to_float (ecb_binary16_to_binary32 (x));
1006 }
714 1007
1008#endif
1009
1010#endif
1011

Diff Legend

Removed lines
+ Added lines
< Changed lines
> Changed lines