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

Comparing libeio/ecb.h (file contents):
Revision 1.21 by root, Fri May 16 15:15:23 2014 UTC vs.
Revision 1.28 by root, Thu Nov 29 19:47:33 2018 UTC

1/* 1/*
2 * libecb - http://software.schmorp.de/pkg/libecb 2 * libecb - http://software.schmorp.de/pkg/libecb
3 * 3 *
4 * Copyright (©) 2009-2014 Marc Alexander Lehmann <libecb@schmorp.de> 4 * Copyright (©) 2009-2015 Marc Alexander Lehmann <libecb@schmorp.de>
5 * Copyright (©) 2011 Emanuele Giaquinta 5 * Copyright (©) 2011 Emanuele Giaquinta
6 * All rights reserved. 6 * All rights reserved.
7 * 7 *
8 * Redistribution and use in source and binary forms, with or without modifica- 8 * Redistribution and use in source and binary forms, with or without modifica-
9 * tion, are permitted provided that the following conditions are met: 9 * tion, are permitted provided that the following conditions are met:
… …
40 40
41#ifndef ECB_H 41#ifndef ECB_H
42#define ECB_H 42#define ECB_H
43 43
44/* 16 bits major, 16 bits minor */ 44/* 16 bits major, 16 bits minor */
45#define ECB_VERSION 0x00010003 45#define ECB_VERSION 0x00010005
46 46
47#ifdef _WIN32 47#ifdef _WIN32
48 typedef signed char int8_t; 48 typedef signed char int8_t;
49 typedef unsigned char uint8_t; 49 typedef unsigned char uint8_t;
50 typedef signed short int16_t; 50 typedef signed short int16_t;
… …
67 typedef uint32_t uintptr_t; 67 typedef uint32_t uintptr_t;
68 typedef int32_t intptr_t; 68 typedef int32_t intptr_t;
69 #endif 69 #endif
70#else 70#else
71 #include <inttypes.h> 71 #include <inttypes.h>
72 #if UINTMAX_MAX > 0xffffffffU 72 #if (defined INTPTR_MAX ? INTPTR_MAX : ULONG_MAX) > 0xffffffffU
73 #define ECB_PTRSIZE 8 73 #define ECB_PTRSIZE 8
74 #else 74 #else
75 #define ECB_PTRSIZE 4 75 #define ECB_PTRSIZE 4
76 #endif 76 #endif
77#endif 77#endif
78 78
79#define ECB_GCC_AMD64 (__amd64 || __amd64__ || __x86_64 || __x86_64__)
80#define ECB_MSVC_AMD64 (_M_AMD64 || _M_X64)
81
79/* work around x32 idiocy by defining proper macros */ 82/* work around x32 idiocy by defining proper macros */
80#if __amd64 || __x86_64 || _M_AMD64 || _M_X64 83#if ECB_GCC_AMD64 || ECB_MSVC_AMD64
81 #if _ILP32 84 #if _ILP32
82 #define ECB_AMD64_X32 1 85 #define ECB_AMD64_X32 1
83 #else 86 #else
84 #define ECB_AMD64 1 87 #define ECB_AMD64 1
85 #endif 88 #endif
… …
90 * causing enormous grief in return for some better fake benchmark numbers. 93 * causing enormous grief in return for some better fake benchmark numbers.
91 * or so. 94 * or so.
92 * we try to detect these and simply assume they are not gcc - if they have 95 * we try to detect these and simply assume they are not gcc - if they have
93 * an issue with that they should have done it right in the first place. 96 * an issue with that they should have done it right in the first place.
94 */ 97 */
95#ifndef ECB_GCC_VERSION
96 #if !defined __GNUC_MINOR__ || defined __INTEL_COMPILER || defined __SUNPRO_C || defined __SUNPRO_CC || defined __llvm__ || defined __clang__ 98#if !defined __GNUC_MINOR__ || defined __INTEL_COMPILER || defined __SUNPRO_C || defined __SUNPRO_CC || defined __llvm__ || defined __clang__
97 #define ECB_GCC_VERSION(major,minor) 0 99 #define ECB_GCC_VERSION(major,minor) 0
98 #else 100#else
99 #define ECB_GCC_VERSION(major,minor) (__GNUC__ > (major) || (__GNUC__ == (major) && __GNUC_MINOR__ >= (minor))) 101 #define ECB_GCC_VERSION(major,minor) (__GNUC__ > (major) || (__GNUC__ == (major) && __GNUC_MINOR__ >= (minor)))
100 #endif 102#endif
103
104#define ECB_CLANG_VERSION(major,minor) (__clang_major__ > (major) || (__clang_major__ == (major) && __clang_minor__ >= (minor)))
105
106#if __clang__ && defined __has_builtin
107 #define ECB_CLANG_BUILTIN(x) __has_builtin (x)
108#else
109 #define ECB_CLANG_BUILTIN(x) 0
110#endif
111
112#if __clang__ && defined __has_extension
113 #define ECB_CLANG_EXTENSION(x) __has_extension (x)
114#else
115 #define ECB_CLANG_EXTENSION(x) 0
101#endif 116#endif
102 117
103#define ECB_CPP (__cplusplus+0) 118#define ECB_CPP (__cplusplus+0)
104#define ECB_CPP11 (__cplusplus >= 201103L) 119#define ECB_CPP11 (__cplusplus >= 201103L)
120#define ECB_CPP14 (__cplusplus >= 201402L)
121#define ECB_CPP17 (__cplusplus >= 201703L)
105 122
106#if ECB_CPP 123#if ECB_CPP
107 #define ECB_C 0 124 #define ECB_C 0
108 #define ECB_STDC_VERSION 0 125 #define ECB_STDC_VERSION 0
109#else 126#else
… …
111 #define ECB_STDC_VERSION __STDC_VERSION__ 128 #define ECB_STDC_VERSION __STDC_VERSION__
112#endif 129#endif
113 130
114#define ECB_C99 (ECB_STDC_VERSION >= 199901L) 131#define ECB_C99 (ECB_STDC_VERSION >= 199901L)
115#define ECB_C11 (ECB_STDC_VERSION >= 201112L) 132#define ECB_C11 (ECB_STDC_VERSION >= 201112L)
133#define ECB_C17 (ECB_STDC_VERSION >= 201710L)
116 134
117#if ECB_CPP 135#if ECB_CPP
118 #define ECB_EXTERN_C extern "C" 136 #define ECB_EXTERN_C extern "C"
119 #define ECB_EXTERN_C_BEG ECB_EXTERN_C { 137 #define ECB_EXTERN_C_BEG ECB_EXTERN_C {
120 #define ECB_EXTERN_C_END } 138 #define ECB_EXTERN_C_END }
… …
133 #define ECB_NO_SMP 1 151 #define ECB_NO_SMP 1
134#endif 152#endif
135 153
136#if ECB_NO_SMP 154#if ECB_NO_SMP
137 #define ECB_MEMORY_FENCE do { } while (0) 155 #define ECB_MEMORY_FENCE do { } while (0)
156#endif
157
158/* http://www-01.ibm.com/support/knowledgecenter/SSGH3R_13.1.0/com.ibm.xlcpp131.aix.doc/compiler_ref/compiler_builtins.html */
159#if __xlC__ && ECB_CPP
160 #include <builtins.h>
161#endif
162
163#if 1400 <= _MSC_VER
164 #include <intrin.h> /* fence functions _ReadBarrier, also bit search functions _BitScanReverse */
138#endif 165#endif
139 166
140#ifndef ECB_MEMORY_FENCE 167#ifndef ECB_MEMORY_FENCE
141 #if ECB_GCC_VERSION(2,5) || defined __INTEL_COMPILER || (__llvm__ && __GNUC__) || __SUNPRO_C >= 0x5110 || __SUNPRO_CC >= 0x5110 168 #if ECB_GCC_VERSION(2,5) || defined __INTEL_COMPILER || (__llvm__ && __GNUC__) || __SUNPRO_C >= 0x5110 || __SUNPRO_CC >= 0x5110
142 #if __i386 || __i386__ 169 #if __i386 || __i386__
143 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("lock; orb $0, -1(%%esp)" : : : "memory") 170 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("lock; orb $0, -1(%%esp)" : : : "memory")
144 #define ECB_MEMORY_FENCE_ACQUIRE __asm__ __volatile__ ("" : : : "memory") 171 #define ECB_MEMORY_FENCE_ACQUIRE __asm__ __volatile__ ("" : : : "memory")
145 #define ECB_MEMORY_FENCE_RELEASE __asm__ __volatile__ ("") 172 #define ECB_MEMORY_FENCE_RELEASE __asm__ __volatile__ ("" : : : "memory")
146 #elif __amd64 || __amd64__ || __x86_64 || __x86_64__ 173 #elif ECB_GCC_AMD64
147 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("mfence" : : : "memory") 174 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("mfence" : : : "memory")
148 #define ECB_MEMORY_FENCE_ACQUIRE __asm__ __volatile__ ("" : : : "memory") 175 #define ECB_MEMORY_FENCE_ACQUIRE __asm__ __volatile__ ("" : : : "memory")
149 #define ECB_MEMORY_FENCE_RELEASE __asm__ __volatile__ ("") 176 #define ECB_MEMORY_FENCE_RELEASE __asm__ __volatile__ ("" : : : "memory")
150 #elif __powerpc__ || __ppc__ || __powerpc64__ || __ppc64__ 177 #elif __powerpc__ || __ppc__ || __powerpc64__ || __ppc64__
151 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("sync" : : : "memory") 178 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("sync" : : : "memory")
179 #elif defined __ARM_ARCH_2__ \
180 || defined __ARM_ARCH_3__ || defined __ARM_ARCH_3M__ \
181 || defined __ARM_ARCH_4__ || defined __ARM_ARCH_4T__ \
182 || defined __ARM_ARCH_5__ || defined __ARM_ARCH_5E__ \
183 || defined __ARM_ARCH_5T__ || defined __ARM_ARCH_5TE__ \
184 || defined __ARM_ARCH_5TEJ__
185 /* should not need any, unless running old code on newer cpu - arm doesn't support that */
152 #elif defined __ARM_ARCH_6__ || defined __ARM_ARCH_6J__ \ 186 #elif defined __ARM_ARCH_6__ || defined __ARM_ARCH_6J__ \
153 || defined __ARM_ARCH_6K__ || defined __ARM_ARCH_6ZK__ 187 || defined __ARM_ARCH_6K__ || defined __ARM_ARCH_6ZK__ \
188 || defined __ARM_ARCH_6T2__
154 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("mcr p15,0,%0,c7,c10,5" : : "r" (0) : "memory") 189 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("mcr p15,0,%0,c7,c10,5" : : "r" (0) : "memory")
155 #elif defined __ARM_ARCH_7__ || defined __ARM_ARCH_7A__ \ 190 #elif defined __ARM_ARCH_7__ || defined __ARM_ARCH_7A__ \
156 || defined __ARM_ARCH_7M__ || defined __ARM_ARCH_7R__ 191 || defined __ARM_ARCH_7R__ || defined __ARM_ARCH_7M__
157 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("dmb" : : : "memory") 192 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("dmb" : : : "memory")
158 #elif __aarch64__ 193 #elif __aarch64__
159 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("dmb ish" : : : "memory") 194 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("dmb ish" : : : "memory")
160 #elif (__sparc || __sparc__) && !__sparcv8 195 #elif (__sparc || __sparc__) && !(__sparc_v8__ || defined __sparcv8)
161 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("membar #LoadStore | #LoadLoad | #StoreStore | #StoreLoad" : : : "memory") 196 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("membar #LoadStore | #LoadLoad | #StoreStore | #StoreLoad" : : : "memory")
162 #define ECB_MEMORY_FENCE_ACQUIRE __asm__ __volatile__ ("membar #LoadStore | #LoadLoad" : : : "memory") 197 #define ECB_MEMORY_FENCE_ACQUIRE __asm__ __volatile__ ("membar #LoadStore | #LoadLoad" : : : "memory")
163 #define ECB_MEMORY_FENCE_RELEASE __asm__ __volatile__ ("membar #LoadStore | #StoreStore") 198 #define ECB_MEMORY_FENCE_RELEASE __asm__ __volatile__ ("membar #LoadStore | #StoreStore")
164 #elif defined __s390__ || defined __s390x__ 199 #elif defined __s390__ || defined __s390x__
165 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("bcr 15,0" : : : "memory") 200 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("bcr 15,0" : : : "memory")
… …
189 /* see comment below (stdatomic.h) about the C11 memory model. */ 224 /* see comment below (stdatomic.h) about the C11 memory model. */
190 #define ECB_MEMORY_FENCE __atomic_thread_fence (__ATOMIC_SEQ_CST) 225 #define ECB_MEMORY_FENCE __atomic_thread_fence (__ATOMIC_SEQ_CST)
191 #define ECB_MEMORY_FENCE_ACQUIRE __atomic_thread_fence (__ATOMIC_ACQUIRE) 226 #define ECB_MEMORY_FENCE_ACQUIRE __atomic_thread_fence (__ATOMIC_ACQUIRE)
192 #define ECB_MEMORY_FENCE_RELEASE __atomic_thread_fence (__ATOMIC_RELEASE) 227 #define ECB_MEMORY_FENCE_RELEASE __atomic_thread_fence (__ATOMIC_RELEASE)
193 228
194 /* The __has_feature syntax from clang is so misdesigned that we cannot use it 229 #elif ECB_CLANG_EXTENSION(c_atomic)
195 * without risking compile time errors with other compilers. We *could*
196 * define our own ecb_clang_has_feature, but I just can't be bothered to work
197 * around this shit time and again.
198 * #elif defined __clang && __has_feature (cxx_atomic)
199 * // see comment below (stdatomic.h) about the C11 memory model. 230 /* see comment below (stdatomic.h) about the C11 memory model. */
200 * #define ECB_MEMORY_FENCE __c11_atomic_thread_fence (__ATOMIC_SEQ_CST) 231 #define ECB_MEMORY_FENCE __c11_atomic_thread_fence (__ATOMIC_SEQ_CST)
201 * #define ECB_MEMORY_FENCE_ACQUIRE __c11_atomic_thread_fence (__ATOMIC_ACQUIRE) 232 #define ECB_MEMORY_FENCE_ACQUIRE __c11_atomic_thread_fence (__ATOMIC_ACQUIRE)
202 * #define ECB_MEMORY_FENCE_RELEASE __c11_atomic_thread_fence (__ATOMIC_RELEASE) 233 #define ECB_MEMORY_FENCE_RELEASE __c11_atomic_thread_fence (__ATOMIC_RELEASE)
203 */
204 234
205 #elif ECB_GCC_VERSION(4,4) || defined __INTEL_COMPILER || defined __clang__ 235 #elif ECB_GCC_VERSION(4,4) || defined __INTEL_COMPILER || defined __clang__
206 #define ECB_MEMORY_FENCE __sync_synchronize () 236 #define ECB_MEMORY_FENCE __sync_synchronize ()
207 #elif _MSC_VER >= 1500 /* VC++ 2008 */ 237 #elif _MSC_VER >= 1500 /* VC++ 2008 */
208 /* apparently, microsoft broke all the memory barrier stuff in Visual Studio 2008... */ 238 /* apparently, microsoft broke all the memory barrier stuff in Visual Studio 2008... */
… …
271 #define ECB_MEMORY_FENCE_RELEASE ECB_MEMORY_FENCE 301 #define ECB_MEMORY_FENCE_RELEASE ECB_MEMORY_FENCE
272#endif 302#endif
273 303
274/*****************************************************************************/ 304/*****************************************************************************/
275 305
276#if __cplusplus 306#if ECB_CPP
277 #define ecb_inline static inline 307 #define ecb_inline static inline
278#elif ECB_GCC_VERSION(2,5) 308#elif ECB_GCC_VERSION(2,5)
279 #define ecb_inline static __inline__ 309 #define ecb_inline static __inline__
280#elif ECB_C99 310#elif ECB_C99
281 #define ecb_inline static inline 311 #define ecb_inline static inline
… …
295 325
296#define ECB_CONCAT_(a, b) a ## b 326#define ECB_CONCAT_(a, b) a ## b
297#define ECB_CONCAT(a, b) ECB_CONCAT_(a, b) 327#define ECB_CONCAT(a, b) ECB_CONCAT_(a, b)
298#define ECB_STRINGIFY_(a) # a 328#define ECB_STRINGIFY_(a) # a
299#define ECB_STRINGIFY(a) ECB_STRINGIFY_(a) 329#define ECB_STRINGIFY(a) ECB_STRINGIFY_(a)
330#define ECB_STRINGIFY_EXPR(expr) ((expr), ECB_STRINGIFY_ (expr))
300 331
301#define ecb_function_ ecb_inline 332#define ecb_function_ ecb_inline
302 333
303#if ECB_GCC_VERSION(3,1) 334#if ECB_GCC_VERSION(3,1) || ECB_CLANG_VERSION(2,8)
304 #define ecb_attribute(attrlist) __attribute__(attrlist) 335 #define ecb_attribute(attrlist) __attribute__ (attrlist)
336#else
337 #define ecb_attribute(attrlist)
338#endif
339
340#if ECB_GCC_VERSION(3,1) || ECB_CLANG_BUILTIN(__builtin_constant_p)
305 #define ecb_is_constant(expr) __builtin_constant_p (expr) 341 #define ecb_is_constant(expr) __builtin_constant_p (expr)
306 #define ecb_expect(expr,value) __builtin_expect ((expr),(value))
307 #define ecb_prefetch(addr,rw,locality) __builtin_prefetch (addr, rw, locality)
308#else 342#else
309 #define ecb_attribute(attrlist)
310
311 /* possible C11 impl for integral types 343 /* possible C11 impl for integral types
312 typedef struct ecb_is_constant_struct ecb_is_constant_struct; 344 typedef struct ecb_is_constant_struct ecb_is_constant_struct;
313 #define ecb_is_constant(expr) _Generic ((1 ? (struct ecb_is_constant_struct *)0 : (void *)((expr) - (expr)), ecb_is_constant_struct *: 0, default: 1)) */ 345 #define ecb_is_constant(expr) _Generic ((1 ? (struct ecb_is_constant_struct *)0 : (void *)((expr) - (expr)), ecb_is_constant_struct *: 0, default: 1)) */
314 346
315 #define ecb_is_constant(expr) 0 347 #define ecb_is_constant(expr) 0
348#endif
349
350#if ECB_GCC_VERSION(3,1) || ECB_CLANG_BUILTIN(__builtin_expect)
351 #define ecb_expect(expr,value) __builtin_expect ((expr),(value))
352#else
316 #define ecb_expect(expr,value) (expr) 353 #define ecb_expect(expr,value) (expr)
354#endif
355
356#if ECB_GCC_VERSION(3,1) || ECB_CLANG_BUILTIN(__builtin_prefetch)
357 #define ecb_prefetch(addr,rw,locality) __builtin_prefetch (addr, rw, locality)
358#else
317 #define ecb_prefetch(addr,rw,locality) 359 #define ecb_prefetch(addr,rw,locality)
318#endif 360#endif
319 361
320/* no emulation for ecb_decltype */ 362/* no emulation for ecb_decltype */
321#if ECB_GCC_VERSION(4,5) 363#if ECB_CPP11
364 // older implementations might have problems with decltype(x)::type, work around it
365 template<class T> struct ecb_decltype_t { typedef T type; };
322 #define ecb_decltype(x) __decltype(x) 366 #define ecb_decltype(x) ecb_decltype_t<decltype (x)>::type
323#elif ECB_GCC_VERSION(3,0) 367#elif ECB_GCC_VERSION(3,0) || ECB_CLANG_VERSION(2,8)
324 #define ecb_decltype(x) __typeof(x) 368 #define ecb_decltype(x) __typeof__ (x)
325#endif 369#endif
326 370
371#if _MSC_VER >= 1300
372 #define ecb_deprecated __declspec (deprecated)
373#else
374 #define ecb_deprecated ecb_attribute ((__deprecated__))
375#endif
376
377#if _MSC_VER >= 1500
378 #define ecb_deprecated_message(msg) __declspec (deprecated (msg))
379#elif ECB_GCC_VERSION(4,5)
380 #define ecb_deprecated_message(msg) ecb_attribute ((__deprecated__ (msg))
381#else
382 #define ecb_deprecated_message(msg) ecb_deprecated
383#endif
384
385#if _MSC_VER >= 1400
386 #define ecb_noinline __declspec (noinline)
387#else
327#define ecb_noinline ecb_attribute ((__noinline__)) 388 #define ecb_noinline ecb_attribute ((__noinline__))
389#endif
390
328#define ecb_unused ecb_attribute ((__unused__)) 391#define ecb_unused ecb_attribute ((__unused__))
329#define ecb_const ecb_attribute ((__const__)) 392#define ecb_const ecb_attribute ((__const__))
330#define ecb_pure ecb_attribute ((__pure__)) 393#define ecb_pure ecb_attribute ((__pure__))
331 394
332#if ECB_C11 395#if ECB_C11 || __IBMC_NORETURN
396 /* http://www-01.ibm.com/support/knowledgecenter/SSGH3R_13.1.0/com.ibm.xlcpp131.aix.doc/language_ref/noreturn.html */
333 #define ecb_noreturn _Noreturn 397 #define ecb_noreturn _Noreturn
398#elif ECB_CPP11
399 #define ecb_noreturn [[noreturn]]
400#elif _MSC_VER >= 1200
401 /* http://msdn.microsoft.com/en-us/library/k6ktzx3s.aspx */
402 #define ecb_noreturn __declspec (noreturn)
334#else 403#else
335 #define ecb_noreturn ecb_attribute ((__noreturn__)) 404 #define ecb_noreturn ecb_attribute ((__noreturn__))
336#endif 405#endif
337 406
338#if ECB_GCC_VERSION(4,3) 407#if ECB_GCC_VERSION(4,3)
… …
353/* for compatibility to the rest of the world */ 422/* for compatibility to the rest of the world */
354#define ecb_likely(expr) ecb_expect_true (expr) 423#define ecb_likely(expr) ecb_expect_true (expr)
355#define ecb_unlikely(expr) ecb_expect_false (expr) 424#define ecb_unlikely(expr) ecb_expect_false (expr)
356 425
357/* count trailing zero bits and count # of one bits */ 426/* count trailing zero bits and count # of one bits */
358#if ECB_GCC_VERSION(3,4) 427#if ECB_GCC_VERSION(3,4) \
428 || (ECB_CLANG_BUILTIN(__builtin_clz) && ECB_CLANG_BUILTIN(__builtin_clzll) \
429 && ECB_CLANG_BUILTIN(__builtin_ctz) && ECB_CLANG_BUILTIN(__builtin_ctzll) \
430 && ECB_CLANG_BUILTIN(__builtin_popcount))
359 /* we assume int == 32 bit, long == 32 or 64 bit and long long == 64 bit */ 431 /* we assume int == 32 bit, long == 32 or 64 bit and long long == 64 bit */
360 #define ecb_ld32(x) (__builtin_clz (x) ^ 31) 432 #define ecb_ld32(x) (__builtin_clz (x) ^ 31)
361 #define ecb_ld64(x) (__builtin_clzll (x) ^ 63) 433 #define ecb_ld64(x) (__builtin_clzll (x) ^ 63)
362 #define ecb_ctz32(x) __builtin_ctz (x) 434 #define ecb_ctz32(x) __builtin_ctz (x)
363 #define ecb_ctz64(x) __builtin_ctzll (x) 435 #define ecb_ctz64(x) __builtin_ctzll (x)
364 #define ecb_popcount32(x) __builtin_popcount (x) 436 #define ecb_popcount32(x) __builtin_popcount (x)
365 /* no popcountll */ 437 /* no popcountll */
366#else 438#else
367 ecb_function_ int ecb_ctz32 (uint32_t x) ecb_const; 439 ecb_function_ ecb_const int ecb_ctz32 (uint32_t x);
368 ecb_function_ int 440 ecb_function_ ecb_const int
369 ecb_ctz32 (uint32_t x) 441 ecb_ctz32 (uint32_t x)
370 { 442 {
443#if 1400 <= _MSC_VER && (_M_IX86 || _M_X64 || _M_IA64 || _M_ARM)
444 unsigned long r;
445 _BitScanForward (&r, x);
446 return (int)r;
447#else
371 int r = 0; 448 int r = 0;
372 449
373 x &= ~x + 1; /* this isolates the lowest bit */ 450 x &= ~x + 1; /* this isolates the lowest bit */
374 451
375#if ECB_branchless_on_i386 452#if ECB_branchless_on_i386
… …
385 if (x & 0xff00ff00) r += 8; 462 if (x & 0xff00ff00) r += 8;
386 if (x & 0xffff0000) r += 16; 463 if (x & 0xffff0000) r += 16;
387#endif 464#endif
388 465
389 return r; 466 return r;
467#endif
390 } 468 }
391 469
392 ecb_function_ int ecb_ctz64 (uint64_t x) ecb_const; 470 ecb_function_ ecb_const int ecb_ctz64 (uint64_t x);
393 ecb_function_ int 471 ecb_function_ ecb_const int
394 ecb_ctz64 (uint64_t x) 472 ecb_ctz64 (uint64_t x)
395 { 473 {
474#if 1400 <= _MSC_VER && (_M_X64 || _M_IA64 || _M_ARM)
475 unsigned long r;
476 _BitScanForward64 (&r, x);
477 return (int)r;
478#else
396 int shift = x & 0xffffffffU ? 0 : 32; 479 int shift = x & 0xffffffff ? 0 : 32;
397 return ecb_ctz32 (x >> shift) + shift; 480 return ecb_ctz32 (x >> shift) + shift;
481#endif
398 } 482 }
399 483
400 ecb_function_ int ecb_popcount32 (uint32_t x) ecb_const; 484 ecb_function_ ecb_const int ecb_popcount32 (uint32_t x);
401 ecb_function_ int 485 ecb_function_ ecb_const int
402 ecb_popcount32 (uint32_t x) 486 ecb_popcount32 (uint32_t x)
403 { 487 {
404 x -= (x >> 1) & 0x55555555; 488 x -= (x >> 1) & 0x55555555;
405 x = ((x >> 2) & 0x33333333) + (x & 0x33333333); 489 x = ((x >> 2) & 0x33333333) + (x & 0x33333333);
406 x = ((x >> 4) + x) & 0x0f0f0f0f; 490 x = ((x >> 4) + x) & 0x0f0f0f0f;
407 x *= 0x01010101; 491 x *= 0x01010101;
408 492
409 return x >> 24; 493 return x >> 24;
410 } 494 }
411 495
412 ecb_function_ int ecb_ld32 (uint32_t x) ecb_const; 496 ecb_function_ ecb_const int ecb_ld32 (uint32_t x);
413 ecb_function_ int ecb_ld32 (uint32_t x) 497 ecb_function_ ecb_const int ecb_ld32 (uint32_t x)
414 { 498 {
499#if 1400 <= _MSC_VER && (_M_IX86 || _M_X64 || _M_IA64 || _M_ARM)
500 unsigned long r;
501 _BitScanReverse (&r, x);
502 return (int)r;
503#else
415 int r = 0; 504 int r = 0;
416 505
417 if (x >> 16) { x >>= 16; r += 16; } 506 if (x >> 16) { x >>= 16; r += 16; }
418 if (x >> 8) { x >>= 8; r += 8; } 507 if (x >> 8) { x >>= 8; r += 8; }
419 if (x >> 4) { x >>= 4; r += 4; } 508 if (x >> 4) { x >>= 4; r += 4; }
420 if (x >> 2) { x >>= 2; r += 2; } 509 if (x >> 2) { x >>= 2; r += 2; }
421 if (x >> 1) { r += 1; } 510 if (x >> 1) { r += 1; }
422 511
423 return r; 512 return r;
513#endif
424 } 514 }
425 515
426 ecb_function_ int ecb_ld64 (uint64_t x) ecb_const; 516 ecb_function_ ecb_const int ecb_ld64 (uint64_t x);
427 ecb_function_ int ecb_ld64 (uint64_t x) 517 ecb_function_ ecb_const int ecb_ld64 (uint64_t x)
428 { 518 {
519#if 1400 <= _MSC_VER && (_M_X64 || _M_IA64 || _M_ARM)
520 unsigned long r;
521 _BitScanReverse64 (&r, x);
522 return (int)r;
523#else
429 int r = 0; 524 int r = 0;
430 525
431 if (x >> 32) { x >>= 32; r += 32; } 526 if (x >> 32) { x >>= 32; r += 32; }
432 527
433 return r + ecb_ld32 (x); 528 return r + ecb_ld32 (x);
434 }
435#endif 529#endif
530 }
531#endif
436 532
437ecb_function_ ecb_bool ecb_is_pot32 (uint32_t x) ecb_const; 533ecb_function_ ecb_const ecb_bool ecb_is_pot32 (uint32_t x);
438ecb_function_ ecb_bool ecb_is_pot32 (uint32_t x) { return !(x & (x - 1)); } 534ecb_function_ ecb_const ecb_bool ecb_is_pot32 (uint32_t x) { return !(x & (x - 1)); }
439ecb_function_ ecb_bool ecb_is_pot64 (uint64_t x) ecb_const; 535ecb_function_ ecb_const ecb_bool ecb_is_pot64 (uint64_t x);
440ecb_function_ ecb_bool ecb_is_pot64 (uint64_t x) { return !(x & (x - 1)); } 536ecb_function_ ecb_const ecb_bool ecb_is_pot64 (uint64_t x) { return !(x & (x - 1)); }
441 537
442ecb_function_ uint8_t ecb_bitrev8 (uint8_t x) ecb_const; 538ecb_function_ ecb_const uint8_t ecb_bitrev8 (uint8_t x);
443ecb_function_ uint8_t ecb_bitrev8 (uint8_t x) 539ecb_function_ ecb_const uint8_t ecb_bitrev8 (uint8_t x)
444{ 540{
445 return ( (x * 0x0802U & 0x22110U) 541 return ( (x * 0x0802U & 0x22110U)
446 | (x * 0x8020U & 0x88440U)) * 0x10101U >> 16; 542 | (x * 0x8020U & 0x88440U)) * 0x10101U >> 16;
447} 543}
448 544
449ecb_function_ uint16_t ecb_bitrev16 (uint16_t x) ecb_const; 545ecb_function_ ecb_const uint16_t ecb_bitrev16 (uint16_t x);
450ecb_function_ uint16_t ecb_bitrev16 (uint16_t x) 546ecb_function_ ecb_const uint16_t ecb_bitrev16 (uint16_t x)
451{ 547{
452 x = ((x >> 1) & 0x5555) | ((x & 0x5555) << 1); 548 x = ((x >> 1) & 0x5555) | ((x & 0x5555) << 1);
453 x = ((x >> 2) & 0x3333) | ((x & 0x3333) << 2); 549 x = ((x >> 2) & 0x3333) | ((x & 0x3333) << 2);
454 x = ((x >> 4) & 0x0f0f) | ((x & 0x0f0f) << 4); 550 x = ((x >> 4) & 0x0f0f) | ((x & 0x0f0f) << 4);
455 x = ( x >> 8 ) | ( x << 8); 551 x = ( x >> 8 ) | ( x << 8);
456 552
457 return x; 553 return x;
458} 554}
459 555
460ecb_function_ uint32_t ecb_bitrev32 (uint32_t x) ecb_const; 556ecb_function_ ecb_const uint32_t ecb_bitrev32 (uint32_t x);
461ecb_function_ uint32_t ecb_bitrev32 (uint32_t x) 557ecb_function_ ecb_const uint32_t ecb_bitrev32 (uint32_t x)
462{ 558{
463 x = ((x >> 1) & 0x55555555) | ((x & 0x55555555) << 1); 559 x = ((x >> 1) & 0x55555555) | ((x & 0x55555555) << 1);
464 x = ((x >> 2) & 0x33333333) | ((x & 0x33333333) << 2); 560 x = ((x >> 2) & 0x33333333) | ((x & 0x33333333) << 2);
465 x = ((x >> 4) & 0x0f0f0f0f) | ((x & 0x0f0f0f0f) << 4); 561 x = ((x >> 4) & 0x0f0f0f0f) | ((x & 0x0f0f0f0f) << 4);
466 x = ((x >> 8) & 0x00ff00ff) | ((x & 0x00ff00ff) << 8); 562 x = ((x >> 8) & 0x00ff00ff) | ((x & 0x00ff00ff) << 8);
… …
469 return x; 565 return x;
470} 566}
471 567
472/* popcount64 is only available on 64 bit cpus as gcc builtin */ 568/* popcount64 is only available on 64 bit cpus as gcc builtin */
473/* so for this version we are lazy */ 569/* so for this version we are lazy */
474ecb_function_ int ecb_popcount64 (uint64_t x) ecb_const; 570ecb_function_ ecb_const int ecb_popcount64 (uint64_t x);
475ecb_function_ int 571ecb_function_ ecb_const int
476ecb_popcount64 (uint64_t x) 572ecb_popcount64 (uint64_t x)
477{ 573{
478 return ecb_popcount32 (x) + ecb_popcount32 (x >> 32); 574 return ecb_popcount32 (x) + ecb_popcount32 (x >> 32);
479} 575}
480 576
481ecb_inline uint8_t ecb_rotl8 (uint8_t x, unsigned int count) ecb_const; 577ecb_inline ecb_const uint8_t ecb_rotl8 (uint8_t x, unsigned int count);
482ecb_inline uint8_t ecb_rotr8 (uint8_t x, unsigned int count) ecb_const; 578ecb_inline ecb_const uint8_t ecb_rotr8 (uint8_t x, unsigned int count);
483ecb_inline uint16_t ecb_rotl16 (uint16_t x, unsigned int count) ecb_const; 579ecb_inline ecb_const uint16_t ecb_rotl16 (uint16_t x, unsigned int count);
484ecb_inline uint16_t ecb_rotr16 (uint16_t x, unsigned int count) ecb_const; 580ecb_inline ecb_const uint16_t ecb_rotr16 (uint16_t x, unsigned int count);
485ecb_inline uint32_t ecb_rotl32 (uint32_t x, unsigned int count) ecb_const; 581ecb_inline ecb_const uint32_t ecb_rotl32 (uint32_t x, unsigned int count);
486ecb_inline uint32_t ecb_rotr32 (uint32_t x, unsigned int count) ecb_const; 582ecb_inline ecb_const uint32_t ecb_rotr32 (uint32_t x, unsigned int count);
487ecb_inline uint64_t ecb_rotl64 (uint64_t x, unsigned int count) ecb_const; 583ecb_inline ecb_const uint64_t ecb_rotl64 (uint64_t x, unsigned int count);
488ecb_inline uint64_t ecb_rotr64 (uint64_t x, unsigned int count) ecb_const; 584ecb_inline ecb_const uint64_t ecb_rotr64 (uint64_t x, unsigned int count);
489 585
490ecb_inline uint8_t ecb_rotl8 (uint8_t x, unsigned int count) { return (x >> ( 8 - count)) | (x << count); } 586ecb_inline ecb_const uint8_t ecb_rotl8 (uint8_t x, unsigned int count) { return (x >> ( 8 - count)) | (x << count); }
491ecb_inline uint8_t ecb_rotr8 (uint8_t x, unsigned int count) { return (x << ( 8 - count)) | (x >> count); } 587ecb_inline ecb_const uint8_t ecb_rotr8 (uint8_t x, unsigned int count) { return (x << ( 8 - count)) | (x >> count); }
492ecb_inline uint16_t ecb_rotl16 (uint16_t x, unsigned int count) { return (x >> (16 - count)) | (x << count); } 588ecb_inline ecb_const uint16_t ecb_rotl16 (uint16_t x, unsigned int count) { return (x >> (16 - count)) | (x << count); }
493ecb_inline uint16_t ecb_rotr16 (uint16_t x, unsigned int count) { return (x << (16 - count)) | (x >> count); } 589ecb_inline ecb_const uint16_t ecb_rotr16 (uint16_t x, unsigned int count) { return (x << (16 - count)) | (x >> count); }
494ecb_inline uint32_t ecb_rotl32 (uint32_t x, unsigned int count) { return (x >> (32 - count)) | (x << count); } 590ecb_inline ecb_const uint32_t ecb_rotl32 (uint32_t x, unsigned int count) { return (x >> (32 - count)) | (x << count); }
495ecb_inline uint32_t ecb_rotr32 (uint32_t x, unsigned int count) { return (x << (32 - count)) | (x >> count); } 591ecb_inline ecb_const uint32_t ecb_rotr32 (uint32_t x, unsigned int count) { return (x << (32 - count)) | (x >> count); }
496ecb_inline uint64_t ecb_rotl64 (uint64_t x, unsigned int count) { return (x >> (64 - count)) | (x << count); } 592ecb_inline ecb_const uint64_t ecb_rotl64 (uint64_t x, unsigned int count) { return (x >> (64 - count)) | (x << count); }
497ecb_inline uint64_t ecb_rotr64 (uint64_t x, unsigned int count) { return (x << (64 - count)) | (x >> count); } 593ecb_inline ecb_const uint64_t ecb_rotr64 (uint64_t x, unsigned int count) { return (x << (64 - count)) | (x >> count); }
498 594
499#if ECB_GCC_VERSION(4,3) 595#if ECB_GCC_VERSION(4,3) || (ECB_CLANG_BUILTIN(__builtin_bswap32) && ECB_CLANG_BUILTIN(__builtin_bswap64))
596 #if ECB_GCC_VERSION(4,8) || ECB_CLANG_BUILTIN(__builtin_bswap16)
597 #define ecb_bswap16(x) __builtin_bswap16 (x)
598 #else
500 #define ecb_bswap16(x) (__builtin_bswap32 (x) >> 16) 599 #define ecb_bswap16(x) (__builtin_bswap32 (x) >> 16)
600 #endif
501 #define ecb_bswap32(x) __builtin_bswap32 (x) 601 #define ecb_bswap32(x) __builtin_bswap32 (x)
502 #define ecb_bswap64(x) __builtin_bswap64 (x) 602 #define ecb_bswap64(x) __builtin_bswap64 (x)
603#elif _MSC_VER
604 #include <stdlib.h>
605 #define ecb_bswap16(x) ((uint16_t)_byteswap_ushort ((uint16_t)(x)))
606 #define ecb_bswap32(x) ((uint32_t)_byteswap_ulong ((uint32_t)(x)))
607 #define ecb_bswap64(x) ((uint64_t)_byteswap_uint64 ((uint64_t)(x)))
503#else 608#else
504 ecb_function_ uint16_t ecb_bswap16 (uint16_t x) ecb_const; 609 ecb_function_ ecb_const uint16_t ecb_bswap16 (uint16_t x);
505 ecb_function_ uint16_t 610 ecb_function_ ecb_const uint16_t
506 ecb_bswap16 (uint16_t x) 611 ecb_bswap16 (uint16_t x)
507 { 612 {
508 return ecb_rotl16 (x, 8); 613 return ecb_rotl16 (x, 8);
509 } 614 }
510 615
511 ecb_function_ uint32_t ecb_bswap32 (uint32_t x) ecb_const; 616 ecb_function_ ecb_const uint32_t ecb_bswap32 (uint32_t x);
512 ecb_function_ uint32_t 617 ecb_function_ ecb_const uint32_t
513 ecb_bswap32 (uint32_t x) 618 ecb_bswap32 (uint32_t x)
514 { 619 {
515 return (((uint32_t)ecb_bswap16 (x)) << 16) | ecb_bswap16 (x >> 16); 620 return (((uint32_t)ecb_bswap16 (x)) << 16) | ecb_bswap16 (x >> 16);
516 } 621 }
517 622
518 ecb_function_ uint64_t ecb_bswap64 (uint64_t x) ecb_const; 623 ecb_function_ ecb_const uint64_t ecb_bswap64 (uint64_t x);
519 ecb_function_ uint64_t 624 ecb_function_ ecb_const uint64_t
520 ecb_bswap64 (uint64_t x) 625 ecb_bswap64 (uint64_t x)
521 { 626 {
522 return (((uint64_t)ecb_bswap32 (x)) << 32) | ecb_bswap32 (x >> 32); 627 return (((uint64_t)ecb_bswap32 (x)) << 32) | ecb_bswap32 (x >> 32);
523 } 628 }
524#endif 629#endif
525 630
526#if ECB_GCC_VERSION(4,5) 631#if ECB_GCC_VERSION(4,5) || ECB_CLANG_BUILTIN(__builtin_unreachable)
527 #define ecb_unreachable() __builtin_unreachable () 632 #define ecb_unreachable() __builtin_unreachable ()
528#else 633#else
529 /* this seems to work fine, but gcc always emits a warning for it :/ */ 634 /* this seems to work fine, but gcc always emits a warning for it :/ */
530 ecb_inline void ecb_unreachable (void) ecb_noreturn; 635 ecb_inline ecb_noreturn void ecb_unreachable (void);
531 ecb_inline void ecb_unreachable (void) { } 636 ecb_inline ecb_noreturn void ecb_unreachable (void) { }
532#endif 637#endif
533 638
534/* try to tell the compiler that some condition is definitely true */ 639/* try to tell the compiler that some condition is definitely true */
535#define ecb_assume(cond) if (!(cond)) ecb_unreachable (); else 0 640#define ecb_assume(cond) if (!(cond)) ecb_unreachable (); else 0
536 641
537ecb_inline unsigned char ecb_byteorder_helper (void) ecb_const; 642ecb_inline ecb_const uint32_t ecb_byteorder_helper (void);
538ecb_inline unsigned char 643ecb_inline ecb_const uint32_t
539ecb_byteorder_helper (void) 644ecb_byteorder_helper (void)
540{ 645{
541 /* the union code still generates code under pressure in gcc, */ 646 /* the union code still generates code under pressure in gcc, */
542 /* but less than using pointers, and always seems to */ 647 /* but less than using pointers, and always seems to */
543 /* successfully return a constant. */ 648 /* successfully return a constant. */
544 /* the reason why we have this horrible preprocessor mess */ 649 /* the reason why we have this horrible preprocessor mess */
545 /* is to avoid it in all cases, at least on common architectures */ 650 /* is to avoid it in all cases, at least on common architectures */
546 /* or when using a recent enough gcc version (>= 4.6) */ 651 /* or when using a recent enough gcc version (>= 4.6) */
547#if __i386 || __i386__ || _M_X86 || __amd64 || __amd64__ || _M_X64
548 return 0x44;
549#elif __BYTE_ORDER__ && __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__ 652#if (defined __BYTE_ORDER__ && __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__) \
653 || ((__i386 || __i386__ || _M_IX86 || ECB_GCC_AMD64 || ECB_MSVC_AMD64) && !__VOS__)
654 #define ECB_LITTLE_ENDIAN 1
550 return 0x44; 655 return 0x44332211;
551#elif __BYTE_ORDER__ && __BYTE_ORDER__ == __ORDER_BIG_ENDIAN__ 656#elif (defined __BYTE_ORDER__ && __BYTE_ORDER__ == __ORDER_BIG_ENDIAN__) \
657 || ((__AARCH64EB__ || __MIPSEB__ || __ARMEB__) && !__VOS__)
658 #define ECB_BIG_ENDIAN 1
552 return 0x11; 659 return 0x11223344;
553#else 660#else
554 union 661 union
555 { 662 {
663 uint8_t c[4];
556 uint32_t i; 664 uint32_t u;
557 uint8_t c;
558 } u = { 0x11223344 }; 665 } u = { 0x11, 0x22, 0x33, 0x44 };
559 return u.c; 666 return u.u;
560#endif 667#endif
561} 668}
562 669
563ecb_inline ecb_bool ecb_big_endian (void) ecb_const; 670ecb_inline ecb_const ecb_bool ecb_big_endian (void);
564ecb_inline ecb_bool ecb_big_endian (void) { return ecb_byteorder_helper () == 0x11; } 671ecb_inline ecb_const ecb_bool ecb_big_endian (void) { return ecb_byteorder_helper () == 0x11223344; }
565ecb_inline ecb_bool ecb_little_endian (void) ecb_const; 672ecb_inline ecb_const ecb_bool ecb_little_endian (void);
566ecb_inline ecb_bool ecb_little_endian (void) { return ecb_byteorder_helper () == 0x44; } 673ecb_inline ecb_const ecb_bool ecb_little_endian (void) { return ecb_byteorder_helper () == 0x44332211; }
567 674
568#if ECB_GCC_VERSION(3,0) || ECB_C99 675#if ECB_GCC_VERSION(3,0) || ECB_C99
569 #define ecb_mod(m,n) ((m) % (n) + ((m) % (n) < 0 ? (n) : 0)) 676 #define ecb_mod(m,n) ((m) % (n) + ((m) % (n) < 0 ? (n) : 0))
570#else 677#else
571 #define ecb_mod(m,n) ((m) < 0 ? ((n) - 1 - ((-1 - (m)) % (n))) : ((m) % (n))) 678 #define ecb_mod(m,n) ((m) < 0 ? ((n) - 1 - ((-1 - (m)) % (n))) : ((m) % (n)))
572#endif 679#endif
573 680
574#if __cplusplus 681#if ECB_CPP
575 template<typename T> 682 template<typename T>
576 static inline T ecb_div_rd (T val, T div) 683 static inline T ecb_div_rd (T val, T div)
577 { 684 {
578 return val < 0 ? - ((-val + div - 1) / div) : (val ) / div; 685 return val < 0 ? - ((-val + div - 1) / div) : (val ) / div;
579 } 686 }
… …
596 } 703 }
597#else 704#else
598 #define ecb_array_length(name) (sizeof (name) / sizeof (name [0])) 705 #define ecb_array_length(name) (sizeof (name) / sizeof (name [0]))
599#endif 706#endif
600 707
708ecb_function_ ecb_const uint32_t ecb_binary16_to_binary32 (uint32_t x);
709ecb_function_ ecb_const uint32_t
710ecb_binary16_to_binary32 (uint32_t x)
711{
712 unsigned int s = (x & 0x8000) << (31 - 15);
713 int e = (x >> 10) & 0x001f;
714 unsigned int m = x & 0x03ff;
715
716 if (ecb_expect_false (e == 31))
717 /* infinity or NaN */
718 e = 255 - (127 - 15);
719 else if (ecb_expect_false (!e))
720 {
721 if (ecb_expect_true (!m))
722 /* zero, handled by code below by forcing e to 0 */
723 e = 0 - (127 - 15);
724 else
725 {
726 /* subnormal, renormalise */
727 unsigned int s = 10 - ecb_ld32 (m);
728
729 m = (m << s) & 0x3ff; /* mask implicit bit */
730 e -= s - 1;
731 }
732 }
733
734 /* e and m now are normalised, or zero, (or inf or nan) */
735 e += 127 - 15;
736
737 return s | (e << 23) | (m << (23 - 10));
738}
739
740ecb_function_ ecb_const uint16_t ecb_binary32_to_binary16 (uint32_t x);
741ecb_function_ ecb_const uint16_t
742ecb_binary32_to_binary16 (uint32_t x)
743{
744 unsigned int s = (x >> 16) & 0x00008000; /* sign bit, the easy part */
745 unsigned int e = ((x >> 23) & 0x000000ff) - (127 - 15); /* the desired exponent */
746 unsigned int m = x & 0x007fffff;
747
748 x &= 0x7fffffff;
749
750 /* if it's within range of binary16 normals, use fast path */
751 if (ecb_expect_true (0x38800000 <= x && x <= 0x477fefff))
752 {
753 /* mantissa round-to-even */
754 m += 0x00000fff + ((m >> (23 - 10)) & 1);
755
756 /* handle overflow */
757 if (ecb_expect_false (m >= 0x00800000))
758 {
759 m >>= 1;
760 e += 1;
761 }
762
763 return s | (e << 10) | (m >> (23 - 10));
764 }
765
766 /* handle large numbers and infinity */
767 if (ecb_expect_true (0x477fefff < x && x <= 0x7f800000))
768 return s | 0x7c00;
769
770 /* handle zero, subnormals and small numbers */
771 if (ecb_expect_true (x < 0x38800000))
772 {
773 /* zero */
774 if (ecb_expect_true (!x))
775 return s;
776
777 /* handle subnormals */
778
779 /* too small, will be zero */
780 if (e < (14 - 24)) /* might not be sharp, but is good enough */
781 return s;
782
783 m |= 0x00800000; /* make implicit bit explicit */
784
785 /* very tricky - we need to round to the nearest e (+10) bit value */
786 {
787 unsigned int bits = 14 - e;
788 unsigned int half = (1 << (bits - 1)) - 1;
789 unsigned int even = (m >> bits) & 1;
790
791 /* if this overflows, we will end up with a normalised number */
792 m = (m + half + even) >> bits;
793 }
794
795 return s | m;
796 }
797
798 /* handle NaNs, preserve leftmost nan bits, but make sure we don't turn them into infinities */
799 m >>= 13;
800
801 return s | 0x7c00 | m | !m;
802}
803
601/*******************************************************************************/ 804/*******************************************************************************/
602/* floating point stuff, can be disabled by defining ECB_NO_LIBM */ 805/* floating point stuff, can be disabled by defining ECB_NO_LIBM */
603 806
604/* basically, everything uses "ieee pure-endian" floating point numbers */ 807/* basically, everything uses "ieee pure-endian" floating point numbers */
605/* the only noteworthy exception is ancient armle, which uses order 43218765 */ 808/* the only noteworthy exception is ancient armle, which uses order 43218765 */
606#if 0 \ 809#if 0 \
607 || __i386 || __i386__ \ 810 || __i386 || __i386__ \
608 || __amd64 || __amd64__ || __x86_64 || __x86_64__ \ 811 || ECB_GCC_AMD64 \
609 || __powerpc__ || __ppc__ || __powerpc64__ || __ppc64__ \ 812 || __powerpc__ || __ppc__ || __powerpc64__ || __ppc64__ \
610 || defined __s390__ || defined __s390x__ \ 813 || defined __s390__ || defined __s390x__ \
611 || defined __mips__ \ 814 || defined __mips__ \
612 || defined __alpha__ \ 815 || defined __alpha__ \
613 || defined __hppa__ \ 816 || defined __hppa__ \
614 || defined __ia64__ \ 817 || defined __ia64__ \
615 || defined __m68k__ \ 818 || defined __m68k__ \
616 || defined __m88k__ \ 819 || defined __m88k__ \
617 || defined __sh__ \ 820 || defined __sh__ \
618 || defined _M_IX86 || defined _M_AMD64 || defined _M_IA64 \ 821 || defined _M_IX86 || defined ECB_MSVC_AMD64 || defined _M_IA64 \
619 || (defined __arm__ && (defined __ARM_EABI__ || defined __EABI__ || defined __VFP_FP__ || defined _WIN32_WCE || defined __ANDROID__)) \ 822 || (defined __arm__ && (defined __ARM_EABI__ || defined __EABI__ || defined __VFP_FP__ || defined _WIN32_WCE || defined __ANDROID__)) \
620 || defined __aarch64__ 823 || defined __aarch64__
621 #define ECB_STDFP 1 824 #define ECB_STDFP 1
622 #include <string.h> /* for memcpy */ 825 #include <string.h> /* for memcpy */
623#else 826#else
… …
639 #define ECB_NAN NAN 842 #define ECB_NAN NAN
640 #else 843 #else
641 #define ECB_NAN ECB_INFINITY 844 #define ECB_NAN ECB_INFINITY
642 #endif 845 #endif
643 846
644 /* converts an ieee half/binary16 to a float */ 847 #if ECB_C99 || _XOPEN_VERSION >= 600 || _POSIX_VERSION >= 200112L
645 ecb_function_ float ecb_binary16_to_float (uint16_t x) ecb_const; 848 #define ecb_ldexpf(x,e) ldexpf ((x), (e))
646 ecb_function_ float 849 #define ecb_frexpf(x,e) frexpf ((x), (e))
647 ecb_binary16_to_float (uint16_t x) 850 #else
648 { 851 #define ecb_ldexpf(x,e) (float) ldexp ((double) (x), (e))
649 int e = (x >> 10) & 0x1f; 852 #define ecb_frexpf(x,e) (float) frexp ((double) (x), (e))
650 int m = x & 0x3ff; 853 #endif
651 float r;
652
653 if (!e ) r = ldexpf (m , -24);
654 else if (e != 31) r = ldexpf (m + 0x400, e - 25);
655 else if (m ) r = ECB_NAN;
656 else r = ECB_INFINITY;
657
658 return x & 0x8000 ? -r : r;
659 }
660 854
661 /* convert a float to ieee single/binary32 */ 855 /* convert a float to ieee single/binary32 */
662 ecb_function_ uint32_t ecb_float_to_binary32 (float x) ecb_const; 856 ecb_function_ ecb_const uint32_t ecb_float_to_binary32 (float x);
663 ecb_function_ uint32_t 857 ecb_function_ ecb_const uint32_t
664 ecb_float_to_binary32 (float x) 858 ecb_float_to_binary32 (float x)
665 { 859 {
666 uint32_t r; 860 uint32_t r;
667 861
668 #if ECB_STDFP 862 #if ECB_STDFP
… …
675 if (x == 0e0f ) return 0x00000000U; 869 if (x == 0e0f ) return 0x00000000U;
676 if (x > +3.40282346638528860e+38f) return 0x7f800000U; 870 if (x > +3.40282346638528860e+38f) return 0x7f800000U;
677 if (x < -3.40282346638528860e+38f) return 0xff800000U; 871 if (x < -3.40282346638528860e+38f) return 0xff800000U;
678 if (x != x ) return 0x7fbfffffU; 872 if (x != x ) return 0x7fbfffffU;
679 873
680 m = frexpf (x, &e) * 0x1000000U; 874 m = ecb_frexpf (x, &e) * 0x1000000U;
681 875
682 r = m & 0x80000000U; 876 r = m & 0x80000000U;
683 877
684 if (r) 878 if (r)
685 m = -m; 879 m = -m;
… …
697 891
698 return r; 892 return r;
699 } 893 }
700 894
701 /* converts an ieee single/binary32 to a float */ 895 /* converts an ieee single/binary32 to a float */
702 ecb_function_ float ecb_binary32_to_float (uint32_t x) ecb_const; 896 ecb_function_ ecb_const float ecb_binary32_to_float (uint32_t x);
703 ecb_function_ float 897 ecb_function_ ecb_const float
704 ecb_binary32_to_float (uint32_t x) 898 ecb_binary32_to_float (uint32_t x)
705 { 899 {
706 float r; 900 float r;
707 901
708 #if ECB_STDFP 902 #if ECB_STDFP
… …
718 x |= 0x800000U; 912 x |= 0x800000U;
719 else 913 else
720 e = 1; 914 e = 1;
721 915
722 /* we distrust ldexpf a bit and do the 2**-24 scaling by an extra multiply */ 916 /* we distrust ldexpf a bit and do the 2**-24 scaling by an extra multiply */
723 r = ldexpf (x * (0.5f / 0x800000U), e - 126); 917 r = ecb_ldexpf (x * (0.5f / 0x800000U), e - 126);
724 918
725 r = neg ? -r : r; 919 r = neg ? -r : r;
726 #endif 920 #endif
727 921
728 return r; 922 return r;
729 } 923 }
730 924
731 /* convert a double to ieee double/binary64 */ 925 /* convert a double to ieee double/binary64 */
732 ecb_function_ uint64_t ecb_double_to_binary64 (double x) ecb_const; 926 ecb_function_ ecb_const uint64_t ecb_double_to_binary64 (double x);
733 ecb_function_ uint64_t 927 ecb_function_ ecb_const uint64_t
734 ecb_double_to_binary64 (double x) 928 ecb_double_to_binary64 (double x)
735 { 929 {
736 uint64_t r; 930 uint64_t r;
737 931
738 #if ECB_STDFP 932 #if ECB_STDFP
… …
767 961
768 return r; 962 return r;
769 } 963 }
770 964
771 /* converts an ieee double/binary64 to a double */ 965 /* converts an ieee double/binary64 to a double */
772 ecb_function_ double ecb_binary64_to_double (uint64_t x) ecb_const; 966 ecb_function_ ecb_const double ecb_binary64_to_double (uint64_t x);
773 ecb_function_ double 967 ecb_function_ ecb_const double
774 ecb_binary64_to_double (uint64_t x) 968 ecb_binary64_to_double (uint64_t x)
775 { 969 {
776 double r; 970 double r;
777 971
778 #if ECB_STDFP 972 #if ECB_STDFP
… …
796 #endif 990 #endif
797 991
798 return r; 992 return r;
799 } 993 }
800 994
801#endif 995 /* convert a float to ieee half/binary16 */
996 ecb_function_ ecb_const uint16_t ecb_float_to_binary16 (float x);
997 ecb_function_ ecb_const uint16_t
998 ecb_float_to_binary16 (float x)
999 {
1000 return ecb_binary32_to_binary16 (ecb_float_to_binary32 (x));
1001 }
802 1002
803#endif 1003 /* convert an ieee half/binary16 to float */
1004 ecb_function_ ecb_const float ecb_binary16_to_float (uint16_t x);
1005 ecb_function_ ecb_const float
1006 ecb_binary16_to_float (uint16_t x)
1007 {
1008 return ecb_binary32_to_float (ecb_binary16_to_binary32 (x));
1009 }
804 1010
1011#endif
1012
1013#endif
1014

Diff Legend

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