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

Comparing libecb/ecb.h (file contents):
Revision 1.120 by root, Sat Oct 26 09:13:54 2013 UTC vs.
Revision 1.172 by root, Tue Nov 24 22:12:00 2015 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 0x00010003 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;
63 #else 74 #else
64 #define ECB_PTRSIZE 4 75 #define ECB_PTRSIZE 4
65 #endif 76 #endif
66#endif 77#endif
67 78
79#define ECB_GCC_AMD64 (__amd64 || __amd64__ || __x86_64 || __x86_64__)
80#define ECB_MSVC_AMD64 (_M_AMD64 || _M_X64)
81
68/* work around x32 idiocy by defining proper macros */ 82/* work around x32 idiocy by defining proper macros */
69#if __x86_64 || _M_AMD64 83#if ECB_GCC_AMD64 || ECB_MSVC_AMD64
70 #if _ILP32 84 #if _ILP32
71 #define ECB_AMD64_X32 1 85 #define ECB_AMD64_X32 1
72 #else 86 #else
73 #define ECB_AMD64 1 87 #define ECB_AMD64 1
74 #endif 88 #endif
79 * causing enormous grief in return for some better fake benchmark numbers. 93 * causing enormous grief in return for some better fake benchmark numbers.
80 * or so. 94 * or so.
81 * 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
82 * 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.
83 */ 97 */
84#ifndef ECB_GCC_VERSION
85 #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__
86 #define ECB_GCC_VERSION(major,minor) 0 99 #define ECB_GCC_VERSION(major,minor) 0
87 #else 100#else
88 #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)))
89 #endif 102#endif
90#endif
91 103
92#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)))
93#define ECB_C99 (__STDC_VERSION__ >= 199901L) 105
94#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
95#define ECB_CPP (__cplusplus+0) 118#define ECB_CPP (__cplusplus+0)
96#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)
97 131
98#if ECB_CPP 132#if ECB_CPP
99 #define ECB_EXTERN_C extern "C" 133 #define ECB_EXTERN_C extern "C"
100 #define ECB_EXTERN_C_BEG ECB_EXTERN_C { 134 #define ECB_EXTERN_C_BEG ECB_EXTERN_C {
101 #define ECB_EXTERN_C_END } 135 #define ECB_EXTERN_C_END }
114 #define ECB_NO_SMP 1 148 #define ECB_NO_SMP 1
115#endif 149#endif
116 150
117#if ECB_NO_SMP 151#if ECB_NO_SMP
118 #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 */
119#endif 162#endif
120 163
121#ifndef ECB_MEMORY_FENCE 164#ifndef ECB_MEMORY_FENCE
122 #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
123 #if __i386 || __i386__ 166 #if __i386 || __i386__
124 #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")
125 #define ECB_MEMORY_FENCE_ACQUIRE __asm__ __volatile__ ("" : : : "memory") 168 #define ECB_MEMORY_FENCE_ACQUIRE __asm__ __volatile__ ("" : : : "memory")
126 #define ECB_MEMORY_FENCE_RELEASE __asm__ __volatile__ ("") 169 #define ECB_MEMORY_FENCE_RELEASE __asm__ __volatile__ ("")
127 #elif __amd64 || __amd64__ || __x86_64 || __x86_64__ 170 #elif ECB_GCC_AMD64
128 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("mfence" : : : "memory") 171 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("mfence" : : : "memory")
129 #define ECB_MEMORY_FENCE_ACQUIRE __asm__ __volatile__ ("" : : : "memory") 172 #define ECB_MEMORY_FENCE_ACQUIRE __asm__ __volatile__ ("" : : : "memory")
130 #define ECB_MEMORY_FENCE_RELEASE __asm__ __volatile__ ("") 173 #define ECB_MEMORY_FENCE_RELEASE __asm__ __volatile__ ("")
131 #elif __powerpc__ || __ppc__ || __powerpc64__ || __ppc64__ 174 #elif __powerpc__ || __ppc__ || __powerpc64__ || __ppc64__
132 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("sync" : : : "memory") 175 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("sync" : : : "memory")
134 || defined __ARM_ARCH_6K__ || defined __ARM_ARCH_6ZK__ 177 || defined __ARM_ARCH_6K__ || defined __ARM_ARCH_6ZK__
135 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("mcr p15,0,%0,c7,c10,5" : : "r" (0) : "memory") 178 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("mcr p15,0,%0,c7,c10,5" : : "r" (0) : "memory")
136 #elif defined __ARM_ARCH_7__ || defined __ARM_ARCH_7A__ \ 179 #elif defined __ARM_ARCH_7__ || defined __ARM_ARCH_7A__ \
137 || defined __ARM_ARCH_7M__ || defined __ARM_ARCH_7R__ 180 || defined __ARM_ARCH_7M__ || defined __ARM_ARCH_7R__
138 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("dmb" : : : "memory") 181 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("dmb" : : : "memory")
139 #elif __sparc || __sparc__ 182 #elif __aarch64__
183 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("dmb ish" : : : "memory")
184 #elif (__sparc || __sparc__) && !(__sparc_v8__ || defined __sparcv8)
140 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("membar #LoadStore | #LoadLoad | #StoreStore | #StoreLoad" : : : "memory") 185 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("membar #LoadStore | #LoadLoad | #StoreStore | #StoreLoad" : : : "memory")
141 #define ECB_MEMORY_FENCE_ACQUIRE __asm__ __volatile__ ("membar #LoadStore | #LoadLoad" : : : "memory") 186 #define ECB_MEMORY_FENCE_ACQUIRE __asm__ __volatile__ ("membar #LoadStore | #LoadLoad" : : : "memory")
142 #define ECB_MEMORY_FENCE_RELEASE __asm__ __volatile__ ("membar #LoadStore | #StoreStore") 187 #define ECB_MEMORY_FENCE_RELEASE __asm__ __volatile__ ("membar #LoadStore | #StoreStore")
143 #elif defined __s390__ || defined __s390x__ 188 #elif defined __s390__ || defined __s390x__
144 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("bcr 15,0" : : : "memory") 189 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("bcr 15,0" : : : "memory")
165 210
166#ifndef ECB_MEMORY_FENCE 211#ifndef ECB_MEMORY_FENCE
167 #if ECB_GCC_VERSION(4,7) 212 #if ECB_GCC_VERSION(4,7)
168 /* see comment below (stdatomic.h) about the C11 memory model. */ 213 /* see comment below (stdatomic.h) about the C11 memory model. */
169 #define ECB_MEMORY_FENCE __atomic_thread_fence (__ATOMIC_SEQ_CST) 214 #define ECB_MEMORY_FENCE __atomic_thread_fence (__ATOMIC_SEQ_CST)
215 #define ECB_MEMORY_FENCE_ACQUIRE __atomic_thread_fence (__ATOMIC_ACQUIRE)
216 #define ECB_MEMORY_FENCE_RELEASE __atomic_thread_fence (__ATOMIC_RELEASE)
170 217
171 /* The __has_feature syntax from clang is so misdesigned that we cannot use it 218 #elif ECB_CLANG_EXTENSION(c_atomic)
172 * without risking compile time errors with other compilers. We *could*
173 * define our own ecb_clang_has_feature, but I just can't be bothered to work
174 * around this shit time and again.
175 * #elif defined __clang && __has_feature (cxx_atomic)
176 * // see comment below (stdatomic.h) about the C11 memory model. 219 /* see comment below (stdatomic.h) about the C11 memory model. */
177 * #define ECB_MEMORY_FENCE __c11_atomic_thread_fence (__ATOMIC_SEQ_CST) 220 #define ECB_MEMORY_FENCE __c11_atomic_thread_fence (__ATOMIC_SEQ_CST)
178 */ 221 #define ECB_MEMORY_FENCE_ACQUIRE __c11_atomic_thread_fence (__ATOMIC_ACQUIRE)
222 #define ECB_MEMORY_FENCE_RELEASE __c11_atomic_thread_fence (__ATOMIC_RELEASE)
179 223
180 #elif ECB_GCC_VERSION(4,4) || defined __INTEL_COMPILER || defined __clang__ 224 #elif ECB_GCC_VERSION(4,4) || defined __INTEL_COMPILER || defined __clang__
181 #define ECB_MEMORY_FENCE __sync_synchronize () 225 #define ECB_MEMORY_FENCE __sync_synchronize ()
226 #elif _MSC_VER >= 1500 /* VC++ 2008 */
227 /* apparently, microsoft broke all the memory barrier stuff in Visual Studio 2008... */
228 #pragma intrinsic(_ReadBarrier,_WriteBarrier,_ReadWriteBarrier)
229 #define ECB_MEMORY_FENCE _ReadWriteBarrier (); MemoryBarrier()
230 #define ECB_MEMORY_FENCE_ACQUIRE _ReadWriteBarrier (); MemoryBarrier() /* according to msdn, _ReadBarrier is not a load fence */
231 #define ECB_MEMORY_FENCE_RELEASE _WriteBarrier (); MemoryBarrier()
182 #elif _MSC_VER >= 1400 /* VC++ 2005 */ 232 #elif _MSC_VER >= 1400 /* VC++ 2005 */
183 #pragma intrinsic(_ReadBarrier,_WriteBarrier,_ReadWriteBarrier) 233 #pragma intrinsic(_ReadBarrier,_WriteBarrier,_ReadWriteBarrier)
184 #define ECB_MEMORY_FENCE _ReadWriteBarrier () 234 #define ECB_MEMORY_FENCE _ReadWriteBarrier ()
185 #define ECB_MEMORY_FENCE_ACQUIRE _ReadWriteBarrier () /* according to msdn, _ReadBarrier is not a load fence */ 235 #define ECB_MEMORY_FENCE_ACQUIRE _ReadWriteBarrier () /* according to msdn, _ReadBarrier is not a load fence */
186 #define ECB_MEMORY_FENCE_RELEASE _WriteBarrier () 236 #define ECB_MEMORY_FENCE_RELEASE _WriteBarrier ()
206 /* any fence other than seq_cst, which isn't very efficient for us. */ 256 /* any fence other than seq_cst, which isn't very efficient for us. */
207 /* Why that is, we don't know - either the C11 memory model is quite useless */ 257 /* Why that is, we don't know - either the C11 memory model is quite useless */
208 /* for most usages, or gcc and clang have a bug */ 258 /* for most usages, or gcc and clang have a bug */
209 /* I *currently* lean towards the latter, and inefficiently implement */ 259 /* I *currently* lean towards the latter, and inefficiently implement */
210 /* all three of ecb's fences as a seq_cst fence */ 260 /* all three of ecb's fences as a seq_cst fence */
261 /* Update, gcc-4.8 generates mfence for all c++ fences, but nothing */
262 /* for all __atomic_thread_fence's except seq_cst */
211 #define ECB_MEMORY_FENCE atomic_thread_fence (memory_order_seq_cst) 263 #define ECB_MEMORY_FENCE atomic_thread_fence (memory_order_seq_cst)
212 #endif 264 #endif
213#endif 265#endif
214 266
215#ifndef ECB_MEMORY_FENCE 267#ifndef ECB_MEMORY_FENCE
238 #define ECB_MEMORY_FENCE_RELEASE ECB_MEMORY_FENCE 290 #define ECB_MEMORY_FENCE_RELEASE ECB_MEMORY_FENCE
239#endif 291#endif
240 292
241/*****************************************************************************/ 293/*****************************************************************************/
242 294
243#if __cplusplus 295#if ECB_CPP
244 #define ecb_inline static inline 296 #define ecb_inline static inline
245#elif ECB_GCC_VERSION(2,5) 297#elif ECB_GCC_VERSION(2,5)
246 #define ecb_inline static __inline__ 298 #define ecb_inline static __inline__
247#elif ECB_C99 299#elif ECB_C99
248 #define ecb_inline static inline 300 #define ecb_inline static inline
262 314
263#define ECB_CONCAT_(a, b) a ## b 315#define ECB_CONCAT_(a, b) a ## b
264#define ECB_CONCAT(a, b) ECB_CONCAT_(a, b) 316#define ECB_CONCAT(a, b) ECB_CONCAT_(a, b)
265#define ECB_STRINGIFY_(a) # a 317#define ECB_STRINGIFY_(a) # a
266#define ECB_STRINGIFY(a) ECB_STRINGIFY_(a) 318#define ECB_STRINGIFY(a) ECB_STRINGIFY_(a)
319#define ECB_STRINGIFY_EXPR(expr) ((expr), ECB_STRINGIFY_ (expr))
267 320
268#define ecb_function_ ecb_inline 321#define ecb_function_ ecb_inline
269 322
270#if ECB_GCC_VERSION(3,1) 323#if ECB_GCC_VERSION(3,1) || ECB_CLANG_VERSION(2,8)
271 #define ecb_attribute(attrlist) __attribute__(attrlist) 324 #define ecb_attribute(attrlist) __attribute__ (attrlist)
325#else
326 #define ecb_attribute(attrlist)
327#endif
328
329#if ECB_GCC_VERSION(3,1) || ECB_CLANG_BUILTIN(__builtin_constant_p)
272 #define ecb_is_constant(expr) __builtin_constant_p (expr) 330 #define ecb_is_constant(expr) __builtin_constant_p (expr)
331#else
332 /* possible C11 impl for integral types
333 typedef struct ecb_is_constant_struct ecb_is_constant_struct;
334 #define ecb_is_constant(expr) _Generic ((1 ? (struct ecb_is_constant_struct *)0 : (void *)((expr) - (expr)), ecb_is_constant_struct *: 0, default: 1)) */
335
336 #define ecb_is_constant(expr) 0
337#endif
338
339#if ECB_GCC_VERSION(3,1) || ECB_CLANG_BUILTIN(__builtin_expect)
273 #define ecb_expect(expr,value) __builtin_expect ((expr),(value)) 340 #define ecb_expect(expr,value) __builtin_expect ((expr),(value))
341#else
342 #define ecb_expect(expr,value) (expr)
343#endif
344
345#if ECB_GCC_VERSION(3,1) || ECB_CLANG_BUILTIN(__builtin_prefetch)
274 #define ecb_prefetch(addr,rw,locality) __builtin_prefetch (addr, rw, locality) 346 #define ecb_prefetch(addr,rw,locality) __builtin_prefetch (addr, rw, locality)
275#else 347#else
276 #define ecb_attribute(attrlist)
277 #define ecb_is_constant(expr) 0
278 #define ecb_expect(expr,value) (expr)
279 #define ecb_prefetch(addr,rw,locality) 348 #define ecb_prefetch(addr,rw,locality)
280#endif 349#endif
281 350
282/* no emulation for ecb_decltype */ 351/* no emulation for ecb_decltype */
283#if ECB_GCC_VERSION(4,5) 352#if ECB_CPP11
353 // older implementations might have problems with decltype(x)::type, work around it
354 template<class T> struct ecb_decltype_t { typedef T type; };
284 #define ecb_decltype(x) __decltype(x) 355 #define ecb_decltype(x) ecb_decltype_t<decltype (x)>::type
285#elif ECB_GCC_VERSION(3,0) 356#elif ECB_GCC_VERSION(3,0) || ECB_CLANG_VERSION(2,8)
286 #define ecb_decltype(x) __typeof(x) 357 #define ecb_decltype(x) __typeof__ (x)
287#endif 358#endif
288 359
360#if _MSC_VER >= 1300
361 #define ecb_deprecated __declspec (deprecated)
362#else
363 #define ecb_deprecated ecb_attribute ((__deprecated__))
364#endif
365
366#if _MSC_VER >= 1500
367 #define ecb_deprecated_message(msg) __declspec (deprecated (msg))
368#elif ECB_GCC_VERSION(4,5)
369 #define ecb_deprecated_message(msg) ecb_attribute ((__deprecated__ (msg))
370#else
371 #define ecb_deprecated_message(msg) ecb_deprecated
372#endif
373
374#if _MSC_VER >= 1400
375 #define ecb_noinline __declspec (noinline)
376#else
289#define ecb_noinline ecb_attribute ((__noinline__)) 377 #define ecb_noinline ecb_attribute ((__noinline__))
378#endif
379
290#define ecb_unused ecb_attribute ((__unused__)) 380#define ecb_unused ecb_attribute ((__unused__))
291#define ecb_const ecb_attribute ((__const__)) 381#define ecb_const ecb_attribute ((__const__))
292#define ecb_pure ecb_attribute ((__pure__)) 382#define ecb_pure ecb_attribute ((__pure__))
293 383
294#if ECB_C11 384#if ECB_C11 || __IBMC_NORETURN
385 /* http://www-01.ibm.com/support/knowledgecenter/SSGH3R_13.1.0/com.ibm.xlcpp131.aix.doc/language_ref/noreturn.html */
295 #define ecb_noreturn _Noreturn 386 #define ecb_noreturn _Noreturn
387#elif ECB_CPP11
388 #define ecb_noreturn [[noreturn]]
389#elif _MSC_VER >= 1200
390 /* http://msdn.microsoft.com/en-us/library/k6ktzx3s.aspx */
391 #define ecb_noreturn __declspec (noreturn)
296#else 392#else
297 #define ecb_noreturn ecb_attribute ((__noreturn__)) 393 #define ecb_noreturn ecb_attribute ((__noreturn__))
298#endif 394#endif
299 395
300#if ECB_GCC_VERSION(4,3) 396#if ECB_GCC_VERSION(4,3)
315/* for compatibility to the rest of the world */ 411/* for compatibility to the rest of the world */
316#define ecb_likely(expr) ecb_expect_true (expr) 412#define ecb_likely(expr) ecb_expect_true (expr)
317#define ecb_unlikely(expr) ecb_expect_false (expr) 413#define ecb_unlikely(expr) ecb_expect_false (expr)
318 414
319/* count trailing zero bits and count # of one bits */ 415/* count trailing zero bits and count # of one bits */
320#if ECB_GCC_VERSION(3,4) 416#if ECB_GCC_VERSION(3,4) \
417 || (ECB_CLANG_BUILTIN(__builtin_clz) && ECB_CLANG_BUILTIN(__builtin_clzll) \
418 && ECB_CLANG_BUILTIN(__builtin_ctz) && ECB_CLANG_BUILTIN(__builtin_ctzll) \
419 && ECB_CLANG_BUILTIN(__builtin_popcount))
321 /* we assume int == 32 bit, long == 32 or 64 bit and long long == 64 bit */ 420 /* we assume int == 32 bit, long == 32 or 64 bit and long long == 64 bit */
322 #define ecb_ld32(x) (__builtin_clz (x) ^ 31) 421 #define ecb_ld32(x) (__builtin_clz (x) ^ 31)
323 #define ecb_ld64(x) (__builtin_clzll (x) ^ 63) 422 #define ecb_ld64(x) (__builtin_clzll (x) ^ 63)
324 #define ecb_ctz32(x) __builtin_ctz (x) 423 #define ecb_ctz32(x) __builtin_ctz (x)
325 #define ecb_ctz64(x) __builtin_ctzll (x) 424 #define ecb_ctz64(x) __builtin_ctzll (x)
326 #define ecb_popcount32(x) __builtin_popcount (x) 425 #define ecb_popcount32(x) __builtin_popcount (x)
327 /* no popcountll */ 426 /* no popcountll */
328#else 427#else
329 ecb_function_ int ecb_ctz32 (uint32_t x) ecb_const; 428 ecb_function_ ecb_const int ecb_ctz32 (uint32_t x);
330 ecb_function_ int 429 ecb_function_ ecb_const int
331 ecb_ctz32 (uint32_t x) 430 ecb_ctz32 (uint32_t x)
332 { 431 {
432#if 1400 <= _MSC_VER && (_M_IX86 || _M_X64 || _M_IA64 || _M_ARM)
433 unsigned long r;
434 _BitScanForward (&r, x);
435 return (int)r;
436#else
333 int r = 0; 437 int r = 0;
334 438
335 x &= ~x + 1; /* this isolates the lowest bit */ 439 x &= ~x + 1; /* this isolates the lowest bit */
336 440
337#if ECB_branchless_on_i386 441#if ECB_branchless_on_i386
347 if (x & 0xff00ff00) r += 8; 451 if (x & 0xff00ff00) r += 8;
348 if (x & 0xffff0000) r += 16; 452 if (x & 0xffff0000) r += 16;
349#endif 453#endif
350 454
351 return r; 455 return r;
456#endif
352 } 457 }
353 458
354 ecb_function_ int ecb_ctz64 (uint64_t x) ecb_const; 459 ecb_function_ ecb_const int ecb_ctz64 (uint64_t x);
355 ecb_function_ int 460 ecb_function_ ecb_const int
356 ecb_ctz64 (uint64_t x) 461 ecb_ctz64 (uint64_t x)
357 { 462 {
463#if 1400 <= _MSC_VER && (_M_X64 || _M_IA64 || _M_ARM)
464 unsigned long r;
465 _BitScanForward64 (&r, x);
466 return (int)r;
467#else
358 int shift = x & 0xffffffffU ? 0 : 32; 468 int shift = x & 0xffffffff ? 0 : 32;
359 return ecb_ctz32 (x >> shift) + shift; 469 return ecb_ctz32 (x >> shift) + shift;
470#endif
360 } 471 }
361 472
362 ecb_function_ int ecb_popcount32 (uint32_t x) ecb_const; 473 ecb_function_ ecb_const int ecb_popcount32 (uint32_t x);
363 ecb_function_ int 474 ecb_function_ ecb_const int
364 ecb_popcount32 (uint32_t x) 475 ecb_popcount32 (uint32_t x)
365 { 476 {
366 x -= (x >> 1) & 0x55555555; 477 x -= (x >> 1) & 0x55555555;
367 x = ((x >> 2) & 0x33333333) + (x & 0x33333333); 478 x = ((x >> 2) & 0x33333333) + (x & 0x33333333);
368 x = ((x >> 4) + x) & 0x0f0f0f0f; 479 x = ((x >> 4) + x) & 0x0f0f0f0f;
369 x *= 0x01010101; 480 x *= 0x01010101;
370 481
371 return x >> 24; 482 return x >> 24;
372 } 483 }
373 484
374 ecb_function_ int ecb_ld32 (uint32_t x) ecb_const; 485 ecb_function_ ecb_const int ecb_ld32 (uint32_t x);
375 ecb_function_ int ecb_ld32 (uint32_t x) 486 ecb_function_ ecb_const int ecb_ld32 (uint32_t x)
376 { 487 {
488#if 1400 <= _MSC_VER && (_M_IX86 || _M_X64 || _M_IA64 || _M_ARM)
489 unsigned long r;
490 _BitScanReverse (&r, x);
491 return (int)r;
492#else
377 int r = 0; 493 int r = 0;
378 494
379 if (x >> 16) { x >>= 16; r += 16; } 495 if (x >> 16) { x >>= 16; r += 16; }
380 if (x >> 8) { x >>= 8; r += 8; } 496 if (x >> 8) { x >>= 8; r += 8; }
381 if (x >> 4) { x >>= 4; r += 4; } 497 if (x >> 4) { x >>= 4; r += 4; }
382 if (x >> 2) { x >>= 2; r += 2; } 498 if (x >> 2) { x >>= 2; r += 2; }
383 if (x >> 1) { r += 1; } 499 if (x >> 1) { r += 1; }
384 500
385 return r; 501 return r;
502#endif
386 } 503 }
387 504
388 ecb_function_ int ecb_ld64 (uint64_t x) ecb_const; 505 ecb_function_ ecb_const int ecb_ld64 (uint64_t x);
389 ecb_function_ int ecb_ld64 (uint64_t x) 506 ecb_function_ ecb_const int ecb_ld64 (uint64_t x)
390 { 507 {
508#if 1400 <= _MSC_VER && (_M_X64 || _M_IA64 || _M_ARM)
509 unsigned long r;
510 _BitScanReverse64 (&r, x);
511 return (int)r;
512#else
391 int r = 0; 513 int r = 0;
392 514
393 if (x >> 32) { x >>= 32; r += 32; } 515 if (x >> 32) { x >>= 32; r += 32; }
394 516
395 return r + ecb_ld32 (x); 517 return r + ecb_ld32 (x);
396 }
397#endif 518#endif
519 }
520#endif
398 521
399ecb_function_ ecb_bool ecb_is_pot32 (uint32_t x) ecb_const; 522ecb_function_ ecb_const ecb_bool ecb_is_pot32 (uint32_t x);
400ecb_function_ ecb_bool ecb_is_pot32 (uint32_t x) { return !(x & (x - 1)); } 523ecb_function_ ecb_const ecb_bool ecb_is_pot32 (uint32_t x) { return !(x & (x - 1)); }
401ecb_function_ ecb_bool ecb_is_pot64 (uint64_t x) ecb_const; 524ecb_function_ ecb_const ecb_bool ecb_is_pot64 (uint64_t x);
402ecb_function_ ecb_bool ecb_is_pot64 (uint64_t x) { return !(x & (x - 1)); } 525ecb_function_ ecb_const ecb_bool ecb_is_pot64 (uint64_t x) { return !(x & (x - 1)); }
403 526
404ecb_function_ uint8_t ecb_bitrev8 (uint8_t x) ecb_const; 527ecb_function_ ecb_const uint8_t ecb_bitrev8 (uint8_t x);
405ecb_function_ uint8_t ecb_bitrev8 (uint8_t x) 528ecb_function_ ecb_const uint8_t ecb_bitrev8 (uint8_t x)
406{ 529{
407 return ( (x * 0x0802U & 0x22110U) 530 return ( (x * 0x0802U & 0x22110U)
408 | (x * 0x8020U & 0x88440U)) * 0x10101U >> 16; 531 | (x * 0x8020U & 0x88440U)) * 0x10101U >> 16;
409} 532}
410 533
411ecb_function_ uint16_t ecb_bitrev16 (uint16_t x) ecb_const; 534ecb_function_ ecb_const uint16_t ecb_bitrev16 (uint16_t x);
412ecb_function_ uint16_t ecb_bitrev16 (uint16_t x) 535ecb_function_ ecb_const uint16_t ecb_bitrev16 (uint16_t x)
413{ 536{
414 x = ((x >> 1) & 0x5555) | ((x & 0x5555) << 1); 537 x = ((x >> 1) & 0x5555) | ((x & 0x5555) << 1);
415 x = ((x >> 2) & 0x3333) | ((x & 0x3333) << 2); 538 x = ((x >> 2) & 0x3333) | ((x & 0x3333) << 2);
416 x = ((x >> 4) & 0x0f0f) | ((x & 0x0f0f) << 4); 539 x = ((x >> 4) & 0x0f0f) | ((x & 0x0f0f) << 4);
417 x = ( x >> 8 ) | ( x << 8); 540 x = ( x >> 8 ) | ( x << 8);
418 541
419 return x; 542 return x;
420} 543}
421 544
422ecb_function_ uint32_t ecb_bitrev32 (uint32_t x) ecb_const; 545ecb_function_ ecb_const uint32_t ecb_bitrev32 (uint32_t x);
423ecb_function_ uint32_t ecb_bitrev32 (uint32_t x) 546ecb_function_ ecb_const uint32_t ecb_bitrev32 (uint32_t x)
424{ 547{
425 x = ((x >> 1) & 0x55555555) | ((x & 0x55555555) << 1); 548 x = ((x >> 1) & 0x55555555) | ((x & 0x55555555) << 1);
426 x = ((x >> 2) & 0x33333333) | ((x & 0x33333333) << 2); 549 x = ((x >> 2) & 0x33333333) | ((x & 0x33333333) << 2);
427 x = ((x >> 4) & 0x0f0f0f0f) | ((x & 0x0f0f0f0f) << 4); 550 x = ((x >> 4) & 0x0f0f0f0f) | ((x & 0x0f0f0f0f) << 4);
428 x = ((x >> 8) & 0x00ff00ff) | ((x & 0x00ff00ff) << 8); 551 x = ((x >> 8) & 0x00ff00ff) | ((x & 0x00ff00ff) << 8);
431 return x; 554 return x;
432} 555}
433 556
434/* popcount64 is only available on 64 bit cpus as gcc builtin */ 557/* popcount64 is only available on 64 bit cpus as gcc builtin */
435/* so for this version we are lazy */ 558/* so for this version we are lazy */
436ecb_function_ int ecb_popcount64 (uint64_t x) ecb_const; 559ecb_function_ ecb_const int ecb_popcount64 (uint64_t x);
437ecb_function_ int 560ecb_function_ ecb_const int
438ecb_popcount64 (uint64_t x) 561ecb_popcount64 (uint64_t x)
439{ 562{
440 return ecb_popcount32 (x) + ecb_popcount32 (x >> 32); 563 return ecb_popcount32 (x) + ecb_popcount32 (x >> 32);
441} 564}
442 565
443ecb_inline uint8_t ecb_rotl8 (uint8_t x, unsigned int count) ecb_const; 566ecb_inline ecb_const uint8_t ecb_rotl8 (uint8_t x, unsigned int count);
444ecb_inline uint8_t ecb_rotr8 (uint8_t x, unsigned int count) ecb_const; 567ecb_inline ecb_const uint8_t ecb_rotr8 (uint8_t x, unsigned int count);
445ecb_inline uint16_t ecb_rotl16 (uint16_t x, unsigned int count) ecb_const; 568ecb_inline ecb_const uint16_t ecb_rotl16 (uint16_t x, unsigned int count);
446ecb_inline uint16_t ecb_rotr16 (uint16_t x, unsigned int count) ecb_const; 569ecb_inline ecb_const uint16_t ecb_rotr16 (uint16_t x, unsigned int count);
447ecb_inline uint32_t ecb_rotl32 (uint32_t x, unsigned int count) ecb_const; 570ecb_inline ecb_const uint32_t ecb_rotl32 (uint32_t x, unsigned int count);
448ecb_inline uint32_t ecb_rotr32 (uint32_t x, unsigned int count) ecb_const; 571ecb_inline ecb_const uint32_t ecb_rotr32 (uint32_t x, unsigned int count);
449ecb_inline uint64_t ecb_rotl64 (uint64_t x, unsigned int count) ecb_const; 572ecb_inline ecb_const uint64_t ecb_rotl64 (uint64_t x, unsigned int count);
450ecb_inline uint64_t ecb_rotr64 (uint64_t x, unsigned int count) ecb_const; 573ecb_inline ecb_const uint64_t ecb_rotr64 (uint64_t x, unsigned int count);
451 574
452ecb_inline uint8_t ecb_rotl8 (uint8_t x, unsigned int count) { return (x >> ( 8 - count)) | (x << count); } 575ecb_inline ecb_const uint8_t ecb_rotl8 (uint8_t x, unsigned int count) { return (x >> ( 8 - count)) | (x << count); }
453ecb_inline uint8_t ecb_rotr8 (uint8_t x, unsigned int count) { return (x << ( 8 - count)) | (x >> count); } 576ecb_inline ecb_const uint8_t ecb_rotr8 (uint8_t x, unsigned int count) { return (x << ( 8 - count)) | (x >> count); }
454ecb_inline uint16_t ecb_rotl16 (uint16_t x, unsigned int count) { return (x >> (16 - count)) | (x << count); } 577ecb_inline ecb_const uint16_t ecb_rotl16 (uint16_t x, unsigned int count) { return (x >> (16 - count)) | (x << count); }
455ecb_inline uint16_t ecb_rotr16 (uint16_t x, unsigned int count) { return (x << (16 - count)) | (x >> count); } 578ecb_inline ecb_const uint16_t ecb_rotr16 (uint16_t x, unsigned int count) { return (x << (16 - count)) | (x >> count); }
456ecb_inline uint32_t ecb_rotl32 (uint32_t x, unsigned int count) { return (x >> (32 - count)) | (x << count); } 579ecb_inline ecb_const uint32_t ecb_rotl32 (uint32_t x, unsigned int count) { return (x >> (32 - count)) | (x << count); }
457ecb_inline uint32_t ecb_rotr32 (uint32_t x, unsigned int count) { return (x << (32 - count)) | (x >> count); } 580ecb_inline ecb_const uint32_t ecb_rotr32 (uint32_t x, unsigned int count) { return (x << (32 - count)) | (x >> count); }
458ecb_inline uint64_t ecb_rotl64 (uint64_t x, unsigned int count) { return (x >> (64 - count)) | (x << count); } 581ecb_inline ecb_const uint64_t ecb_rotl64 (uint64_t x, unsigned int count) { return (x >> (64 - count)) | (x << count); }
459ecb_inline uint64_t ecb_rotr64 (uint64_t x, unsigned int count) { return (x << (64 - count)) | (x >> count); } 582ecb_inline ecb_const uint64_t ecb_rotr64 (uint64_t x, unsigned int count) { return (x << (64 - count)) | (x >> count); }
460 583
461#if ECB_GCC_VERSION(4,3) 584#if ECB_GCC_VERSION(4,3) || (ECB_CLANG_BUILTIN(__builtin_bswap32) && ECB_CLANG_BUILTIN(__builtin_bswap64))
585 #if ECB_GCC_VERSION(4,8) || ECB_CLANG_BUILTIN(__builtin_bswap16)
586 #define ecb_bswap16(x) __builtin_bswap16 (x)
587 #else
462 #define ecb_bswap16(x) (__builtin_bswap32 (x) >> 16) 588 #define ecb_bswap16(x) (__builtin_bswap32 (x) >> 16)
589 #endif
463 #define ecb_bswap32(x) __builtin_bswap32 (x) 590 #define ecb_bswap32(x) __builtin_bswap32 (x)
464 #define ecb_bswap64(x) __builtin_bswap64 (x) 591 #define ecb_bswap64(x) __builtin_bswap64 (x)
592#elif _MSC_VER
593 #include <stdlib.h>
594 #define ecb_bswap16(x) ((uint16_t)_byteswap_ushort ((uint16_t)(x)))
595 #define ecb_bswap32(x) ((uint32_t)_byteswap_ulong ((uint32_t)(x)))
596 #define ecb_bswap64(x) ((uint64_t)_byteswap_uint64 ((uint64_t)(x)))
465#else 597#else
466 ecb_function_ uint16_t ecb_bswap16 (uint16_t x) ecb_const; 598 ecb_function_ ecb_const uint16_t ecb_bswap16 (uint16_t x);
467 ecb_function_ uint16_t 599 ecb_function_ ecb_const uint16_t
468 ecb_bswap16 (uint16_t x) 600 ecb_bswap16 (uint16_t x)
469 { 601 {
470 return ecb_rotl16 (x, 8); 602 return ecb_rotl16 (x, 8);
471 } 603 }
472 604
473 ecb_function_ uint32_t ecb_bswap32 (uint32_t x) ecb_const; 605 ecb_function_ ecb_const uint32_t ecb_bswap32 (uint32_t x);
474 ecb_function_ uint32_t 606 ecb_function_ ecb_const uint32_t
475 ecb_bswap32 (uint32_t x) 607 ecb_bswap32 (uint32_t x)
476 { 608 {
477 return (((uint32_t)ecb_bswap16 (x)) << 16) | ecb_bswap16 (x >> 16); 609 return (((uint32_t)ecb_bswap16 (x)) << 16) | ecb_bswap16 (x >> 16);
478 } 610 }
479 611
480 ecb_function_ uint64_t ecb_bswap64 (uint64_t x) ecb_const; 612 ecb_function_ ecb_const uint64_t ecb_bswap64 (uint64_t x);
481 ecb_function_ uint64_t 613 ecb_function_ ecb_const uint64_t
482 ecb_bswap64 (uint64_t x) 614 ecb_bswap64 (uint64_t x)
483 { 615 {
484 return (((uint64_t)ecb_bswap32 (x)) << 32) | ecb_bswap32 (x >> 32); 616 return (((uint64_t)ecb_bswap32 (x)) << 32) | ecb_bswap32 (x >> 32);
485 } 617 }
486#endif 618#endif
487 619
488#if ECB_GCC_VERSION(4,5) 620#if ECB_GCC_VERSION(4,5) || ECB_CLANG_BUILTIN(__builtin_unreachable)
489 #define ecb_unreachable() __builtin_unreachable () 621 #define ecb_unreachable() __builtin_unreachable ()
490#else 622#else
491 /* this seems to work fine, but gcc always emits a warning for it :/ */ 623 /* this seems to work fine, but gcc always emits a warning for it :/ */
492 ecb_inline void ecb_unreachable (void) ecb_noreturn; 624 ecb_inline ecb_noreturn void ecb_unreachable (void);
493 ecb_inline void ecb_unreachable (void) { } 625 ecb_inline ecb_noreturn void ecb_unreachable (void) { }
494#endif 626#endif
495 627
496/* try to tell the compiler that some condition is definitely true */ 628/* try to tell the compiler that some condition is definitely true */
497#define ecb_assume(cond) if (!(cond)) ecb_unreachable (); else 0 629#define ecb_assume(cond) if (!(cond)) ecb_unreachable (); else 0
498 630
499ecb_inline unsigned char ecb_byteorder_helper (void) ecb_const; 631ecb_inline ecb_const unsigned char ecb_byteorder_helper (void);
500ecb_inline unsigned char 632ecb_inline ecb_const unsigned char
501ecb_byteorder_helper (void) 633ecb_byteorder_helper (void)
502{ 634{
503 /* the union code still generates code under pressure in gcc, */ 635 /* the union code still generates code under pressure in gcc, */
504 /* but less than using pointers, and always seems to */ 636 /* but less than using pointers, and always seems to */
505 /* successfully return a constant. */ 637 /* successfully return a constant. */
506 /* the reason why we have this horrible preprocessor mess */ 638 /* the reason why we have this horrible preprocessor mess */
507 /* is to avoid it in all cases, at least on common architectures */ 639 /* is to avoid it in all cases, at least on common architectures */
508 /* or when using a recent enough gcc version (>= 4.6) */ 640 /* or when using a recent enough gcc version (>= 4.6) */
509#if __i386 || __i386__ || _M_X86 || __amd64 || __amd64__ || _M_X64 641#if ((__i386 || __i386__) && !__VOS__) || _M_IX86 || ECB_GCC_AMD64 || ECB_MSVC_AMD64
510 return 0x44; 642 return 0x44;
511#elif __BYTE_ORDER__ && __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__ 643#elif __BYTE_ORDER__ && __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__
512 return 0x44; 644 return 0x44;
513#elif __BYTE_ORDER__ && __BYTE_ORDER__ == __ORDER_BIG_ENDIAN__ 645#elif __BYTE_ORDER__ && __BYTE_ORDER__ == __ORDER_BIG_ENDIAN__
514 return 0x11; 646 return 0x11;
520 } u = { 0x11223344 }; 652 } u = { 0x11223344 };
521 return u.c; 653 return u.c;
522#endif 654#endif
523} 655}
524 656
525ecb_inline ecb_bool ecb_big_endian (void) ecb_const; 657ecb_inline ecb_const ecb_bool ecb_big_endian (void);
526ecb_inline ecb_bool ecb_big_endian (void) { return ecb_byteorder_helper () == 0x11; } 658ecb_inline ecb_const ecb_bool ecb_big_endian (void) { return ecb_byteorder_helper () == 0x11; }
527ecb_inline ecb_bool ecb_little_endian (void) ecb_const; 659ecb_inline ecb_const ecb_bool ecb_little_endian (void);
528ecb_inline ecb_bool ecb_little_endian (void) { return ecb_byteorder_helper () == 0x44; } 660ecb_inline ecb_const ecb_bool ecb_little_endian (void) { return ecb_byteorder_helper () == 0x44; }
529 661
530#if ECB_GCC_VERSION(3,0) || ECB_C99 662#if ECB_GCC_VERSION(3,0) || ECB_C99
531 #define ecb_mod(m,n) ((m) % (n) + ((m) % (n) < 0 ? (n) : 0)) 663 #define ecb_mod(m,n) ((m) % (n) + ((m) % (n) < 0 ? (n) : 0))
532#else 664#else
533 #define ecb_mod(m,n) ((m) < 0 ? ((n) - 1 - ((-1 - (m)) % (n))) : ((m) % (n))) 665 #define ecb_mod(m,n) ((m) < 0 ? ((n) - 1 - ((-1 - (m)) % (n))) : ((m) % (n)))
534#endif 666#endif
535 667
536#if __cplusplus 668#if ECB_CPP
537 template<typename T> 669 template<typename T>
538 static inline T ecb_div_rd (T val, T div) 670 static inline T ecb_div_rd (T val, T div)
539 { 671 {
540 return val < 0 ? - ((-val + div - 1) / div) : (val ) / div; 672 return val < 0 ? - ((-val + div - 1) / div) : (val ) / div;
541 } 673 }
558 } 690 }
559#else 691#else
560 #define ecb_array_length(name) (sizeof (name) / sizeof (name [0])) 692 #define ecb_array_length(name) (sizeof (name) / sizeof (name [0]))
561#endif 693#endif
562 694
695ecb_function_ ecb_const uint32_t ecb_binary16_to_binary32 (uint32_t x);
696ecb_function_ ecb_const uint32_t
697ecb_binary16_to_binary32 (uint32_t x)
698{
699 unsigned int s = (x & 0x8000) << (31 - 15);
700 int e = (x >> 10) & 0x001f;
701 unsigned int m = x & 0x03ff;
702
703 if (ecb_expect_false (e == 31))
704 /* infinity or NaN */
705 e = 255 - (127 - 15);
706 else if (ecb_expect_false (!e))
707 {
708 if (ecb_expect_true (!m))
709 /* zero, handled by code below by forcing e to 0 */
710 e = 0 - (127 - 15);
711 else
712 {
713 /* subnormal, renormalise */
714 unsigned int s = 10 - ecb_ld32 (m);
715
716 m = (m << s) & 0x3ff; /* mask implicit bit */
717 e -= s - 1;
718 }
719 }
720
721 /* e and m now are normalised, or zero, (or inf or nan) */
722 e += 127 - 15;
723
724 return s | (e << 23) | (m << (23 - 10));
725}
726
727ecb_function_ ecb_const uint16_t ecb_binary32_to_binary16 (uint32_t x);
728ecb_function_ ecb_const uint16_t
729ecb_binary32_to_binary16 (uint32_t x)
730{
731 unsigned int s = (x >> 16) & 0x00008000; /* sign bit, the easy part */
732 unsigned int e = ((x >> 23) & 0x000000ff) - (127 - 15); /* the desired exponent */
733 unsigned int m = x & 0x007fffff;
734
735 x &= 0x7fffffff;
736
737 /* if it's within range of binary16 normals, use fast path */
738 if (ecb_expect_true (0x38800000 <= x && x <= 0x477fefff))
739 {
740 /* mantissa round-to-even */
741 m += 0x00000fff + ((m >> (23 - 10)) & 1);
742
743 /* handle overflow */
744 if (ecb_expect_false (m >= 0x00800000))
745 {
746 m >>= 1;
747 e += 1;
748 }
749
750 return s | (e << 10) | (m >> (23 - 10));
751 }
752
753 /* handle large numbers and infinity */
754 if (ecb_expect_true (0x477fefff < x && x <= 0x7f800000))
755 return s | 0x7c00;
756
757 /* handle zero, subnormals and small numbers */
758 if (ecb_expect_true (x < 0x38800000))
759 {
760 /* zero */
761 if (ecb_expect_true (!x))
762 return s;
763
764 /* handle subnormals */
765
766 /* too small, will be zero */
767 if (e < (14 - 24)) /* might not be sharp, but is good enough */
768 return s;
769
770 m |= 0x00800000; /* make implicit bit explicit */
771
772 /* very tricky - we need to round to the nearest e (+10) bit value */
773 {
774 unsigned int bits = 14 - e;
775 unsigned int half = (1 << (bits - 1)) - 1;
776 unsigned int even = (m >> bits) & 1;
777
778 /* if this overflows, we will end up with a normalised number */
779 m = (m + half + even) >> bits;
780 }
781
782 return s | m;
783 }
784
785 /* handle NaNs, preserve leftmost nan bits, but make sure we don't turn them into infinities */
786 m >>= 13;
787
788 return s | 0x7c00 | m | !m;
789}
790
563/*******************************************************************************/ 791/*******************************************************************************/
564/* floating point stuff, can be disabled by defining ECB_NO_LIBM */ 792/* floating point stuff, can be disabled by defining ECB_NO_LIBM */
565 793
566/* basically, everything uses "ieee pure-endian" floating point numbers */ 794/* basically, everything uses "ieee pure-endian" floating point numbers */
567/* the only noteworthy exception is ancient armle, which uses order 43218765 */ 795/* the only noteworthy exception is ancient armle, which uses order 43218765 */
568#if 0 \ 796#if 0 \
569 || __i386 || __i386__ \ 797 || __i386 || __i386__ \
570 || __amd64 || __amd64__ || __x86_64 || __x86_64__ \ 798 || ECB_GCC_AMD64 \
571 || __powerpc__ || __ppc__ || __powerpc64__ || __ppc64__ \ 799 || __powerpc__ || __ppc__ || __powerpc64__ || __ppc64__ \
572 || defined __arm__ && defined __ARM_EABI__ \
573 || defined __s390__ || defined __s390x__ \ 800 || defined __s390__ || defined __s390x__ \
574 || defined __mips__ \ 801 || defined __mips__ \
575 || defined __alpha__ \ 802 || defined __alpha__ \
576 || defined __hppa__ \ 803 || defined __hppa__ \
577 || defined __ia64__ \ 804 || defined __ia64__ \
578 || defined __m68k__ \ 805 || defined __m68k__ \
579 || defined __m88k__ \ 806 || defined __m88k__ \
580 || defined __sh__ \ 807 || defined __sh__ \
581 || defined _M_IX86 || defined _M_AMD64 || defined _M_IA64 808 || defined _M_IX86 || defined ECB_MSVC_AMD64 || defined _M_IA64 \
809 || (defined __arm__ && (defined __ARM_EABI__ || defined __EABI__ || defined __VFP_FP__ || defined _WIN32_WCE || defined __ANDROID__)) \
810 || defined __aarch64__
582 #define ECB_STDFP 1 811 #define ECB_STDFP 1
583 #include <string.h> /* for memcpy */ 812 #include <string.h> /* for memcpy */
584#else 813#else
585 #define ECB_STDFP 0 814 #define ECB_STDFP 0
586#endif 815#endif
587 816
588#ifndef ECB_NO_LIBM 817#ifndef ECB_NO_LIBM
589 818
590 #include <math.h> /* for frexp*, ldexp* */ 819 #include <math.h> /* for frexp*, ldexp*, INFINITY, NAN */
591 820
592 /* converts an ieee half/binary16 to a float */ 821 /* only the oldest of old doesn't have this one. solaris. */
593 ecb_function_ float ecb_binary16_to_float (uint16_t x) ecb_const; 822 #ifdef INFINITY
594 ecb_function_ float 823 #define ECB_INFINITY INFINITY
595 ecb_binary16_to_float (uint16_t x) 824 #else
596 { 825 #define ECB_INFINITY HUGE_VAL
597 int e = (x >> 10) & 0x1f; 826 #endif
598 int m = x & 0x3ff;
599 float r;
600 827
601 if (!e)
602 r = ldexpf (m , -24);
603 else if (e != 31)
604 r = ldexpf (m + 0x400, e - 25);
605 else if (m)
606 #ifdef NAN 828 #ifdef NAN
607 r = NAN; 829 #define ECB_NAN NAN
608 #endif
609 else 830 #else
610 r = INFINITY; 831 #define ECB_NAN ECB_INFINITY
832 #endif
611 833
612 return x & 0x8000 ? -r : r; 834 #if ECB_C99 || _XOPEN_VERSION >= 600 || _POSIX_VERSION >= 200112L
613 } 835 #define ecb_ldexpf(x,e) ldexpf ((x), (e))
836 #define ecb_frexpf(x,e) frexpf ((x), (e))
837 #else
838 #define ecb_ldexpf(x,e) (float) ldexp ((double) (x), (e))
839 #define ecb_frexpf(x,e) (float) frexp ((double) (x), (e))
840 #endif
614 841
615 /* convert a float to ieee single/binary32 */ 842 /* convert a float to ieee single/binary32 */
616 ecb_function_ uint32_t ecb_float_to_binary32 (float x) ecb_const; 843 ecb_function_ ecb_const uint32_t ecb_float_to_binary32 (float x);
617 ecb_function_ uint32_t 844 ecb_function_ ecb_const uint32_t
618 ecb_float_to_binary32 (float x) 845 ecb_float_to_binary32 (float x)
619 { 846 {
620 uint32_t r; 847 uint32_t r;
621 848
622 #if ECB_STDFP 849 #if ECB_STDFP
629 if (x == 0e0f ) return 0x00000000U; 856 if (x == 0e0f ) return 0x00000000U;
630 if (x > +3.40282346638528860e+38f) return 0x7f800000U; 857 if (x > +3.40282346638528860e+38f) return 0x7f800000U;
631 if (x < -3.40282346638528860e+38f) return 0xff800000U; 858 if (x < -3.40282346638528860e+38f) return 0xff800000U;
632 if (x != x ) return 0x7fbfffffU; 859 if (x != x ) return 0x7fbfffffU;
633 860
634 m = frexpf (x, &e) * 0x1000000U; 861 m = ecb_frexpf (x, &e) * 0x1000000U;
635 862
636 r = m & 0x80000000U; 863 r = m & 0x80000000U;
637 864
638 if (r) 865 if (r)
639 m = -m; 866 m = -m;
651 878
652 return r; 879 return r;
653 } 880 }
654 881
655 /* converts an ieee single/binary32 to a float */ 882 /* converts an ieee single/binary32 to a float */
656 ecb_function_ float ecb_binary32_to_float (uint32_t x) ecb_const; 883 ecb_function_ ecb_const float ecb_binary32_to_float (uint32_t x);
657 ecb_function_ float 884 ecb_function_ ecb_const float
658 ecb_binary32_to_float (uint32_t x) 885 ecb_binary32_to_float (uint32_t x)
659 { 886 {
660 float r; 887 float r;
661 888
662 #if ECB_STDFP 889 #if ECB_STDFP
672 x |= 0x800000U; 899 x |= 0x800000U;
673 else 900 else
674 e = 1; 901 e = 1;
675 902
676 /* we distrust ldexpf a bit and do the 2**-24 scaling by an extra multiply */ 903 /* we distrust ldexpf a bit and do the 2**-24 scaling by an extra multiply */
677 r = ldexpf (x * (0.5f / 0x800000U), e - 126); 904 r = ecb_ldexpf (x * (0.5f / 0x800000U), e - 126);
678 905
679 r = neg ? -r : r; 906 r = neg ? -r : r;
680 #endif 907 #endif
681 908
682 return r; 909 return r;
683 } 910 }
684 911
685 /* convert a double to ieee double/binary64 */ 912 /* convert a double to ieee double/binary64 */
686 ecb_function_ uint64_t ecb_double_to_binary64 (double x) ecb_const; 913 ecb_function_ ecb_const uint64_t ecb_double_to_binary64 (double x);
687 ecb_function_ uint64_t 914 ecb_function_ ecb_const uint64_t
688 ecb_double_to_binary64 (double x) 915 ecb_double_to_binary64 (double x)
689 { 916 {
690 uint64_t r; 917 uint64_t r;
691 918
692 #if ECB_STDFP 919 #if ECB_STDFP
721 948
722 return r; 949 return r;
723 } 950 }
724 951
725 /* converts an ieee double/binary64 to a double */ 952 /* converts an ieee double/binary64 to a double */
726 ecb_function_ double ecb_binary64_to_double (uint64_t x) ecb_const; 953 ecb_function_ ecb_const double ecb_binary64_to_double (uint64_t x);
727 ecb_function_ double 954 ecb_function_ ecb_const double
728 ecb_binary64_to_double (uint64_t x) 955 ecb_binary64_to_double (uint64_t x)
729 { 956 {
730 double r; 957 double r;
731 958
732 #if ECB_STDFP 959 #if ECB_STDFP
750 #endif 977 #endif
751 978
752 return r; 979 return r;
753 } 980 }
754 981
755#endif 982 /* convert a float to ieee half/binary16 */
983 ecb_function_ ecb_const uint16_t ecb_float_to_binary16 (float x);
984 ecb_function_ ecb_const uint16_t
985 ecb_float_to_binary16 (float x)
986 {
987 return ecb_binary32_to_binary16 (ecb_float_to_binary32 (x));
988 }
756 989
757#endif 990 /* convert an ieee half/binary16 to float */
991 ecb_function_ ecb_const float ecb_binary16_to_float (uint16_t x);
992 ecb_function_ ecb_const float
993 ecb_binary16_to_float (uint16_t x)
994 {
995 return ecb_binary32_to_float (ecb_binary16_to_binary32 (x));
996 }
758 997
998#endif
999
1000#endif
1001

Diff Legend

Removed lines
+ Added lines
< Changed lines
> Changed lines