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

Comparing libecb/ecb.h (file contents):
Revision 1.150 by root, Thu Oct 30 07:21:42 2014 UTC vs.
Revision 1.181 by root, Mon Jan 20 13:16:01 2020 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 0x00010004 45#define ECB_VERSION 0x00010008
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 char int_fast8_t;
51 typedef unsigned char uint_fast8_t;
50 typedef signed short int16_t; 52 typedef signed short int16_t;
51 typedef unsigned short uint16_t; 53 typedef unsigned short uint16_t;
54 typedef signed int int_fast16_t;
55 typedef unsigned int uint_fast16_t;
52 typedef signed int int32_t; 56 typedef signed int int32_t;
53 typedef unsigned int uint32_t; 57 typedef unsigned int uint32_t;
58 typedef signed int int_fast32_t;
59 typedef unsigned int uint_fast32_t;
54 #if __GNUC__ 60 #if __GNUC__
55 typedef signed long long int64_t; 61 typedef signed long long int64_t;
56 typedef unsigned long long uint64_t; 62 typedef unsigned long long uint64_t;
57 #else /* _MSC_VER || __BORLANDC__ */ 63 #else /* _MSC_VER || __BORLANDC__ */
58 typedef signed __int64 int64_t; 64 typedef signed __int64 int64_t;
59 typedef unsigned __int64 uint64_t; 65 typedef unsigned __int64 uint64_t;
60 #endif 66 #endif
67 typedef int64_t int_fast64_t;
68 typedef uint64_t uint_fast64_t;
61 #ifdef _WIN64 69 #ifdef _WIN64
62 #define ECB_PTRSIZE 8 70 #define ECB_PTRSIZE 8
63 typedef uint64_t uintptr_t; 71 typedef uint64_t uintptr_t;
64 typedef int64_t intptr_t; 72 typedef int64_t intptr_t;
65 #else 73 #else
… …
67 typedef uint32_t uintptr_t; 75 typedef uint32_t uintptr_t;
68 typedef int32_t intptr_t; 76 typedef int32_t intptr_t;
69 #endif 77 #endif
70#else 78#else
71 #include <inttypes.h> 79 #include <inttypes.h>
72 #if UINTMAX_MAX > 0xffffffffU 80 #if (defined INTPTR_MAX ? INTPTR_MAX : ULONG_MAX) > 0xffffffffU
73 #define ECB_PTRSIZE 8 81 #define ECB_PTRSIZE 8
74 #else 82 #else
75 #define ECB_PTRSIZE 4 83 #define ECB_PTRSIZE 4
76 #endif 84 #endif
77#endif 85#endif
78 86
87#define ECB_GCC_AMD64 (__amd64 || __amd64__ || __x86_64 || __x86_64__)
88#define ECB_MSVC_AMD64 (_M_AMD64 || _M_X64)
89
90#ifndef ECB_OPTIMIZE_SIZE
91 #if __OPTIMIZE_SIZE__
92 #define ECB_OPTIMIZE_SIZE 1
93 #else
94 #define ECB_OPTIMIZE_SIZE 0
95 #endif
96#endif
97
79/* work around x32 idiocy by defining proper macros */ 98/* work around x32 idiocy by defining proper macros */
80#if __amd64 || __x86_64 || _M_AMD64 || _M_X64 99#if ECB_GCC_AMD64 || ECB_MSVC_AMD64
81 #if _ILP32 100 #if _ILP32
82 #define ECB_AMD64_X32 1 101 #define ECB_AMD64_X32 1
83 #else 102 #else
84 #define ECB_AMD64 1 103 #define ECB_AMD64 1
85 #endif 104 #endif
… …
112 #define ECB_CLANG_EXTENSION(x) 0 131 #define ECB_CLANG_EXTENSION(x) 0
113#endif 132#endif
114 133
115#define ECB_CPP (__cplusplus+0) 134#define ECB_CPP (__cplusplus+0)
116#define ECB_CPP11 (__cplusplus >= 201103L) 135#define ECB_CPP11 (__cplusplus >= 201103L)
136#define ECB_CPP14 (__cplusplus >= 201402L)
137#define ECB_CPP17 (__cplusplus >= 201703L)
117 138
118#if ECB_CPP 139#if ECB_CPP
119 #define ECB_C 0 140 #define ECB_C 0
120 #define ECB_STDC_VERSION 0 141 #define ECB_STDC_VERSION 0
121#else 142#else
… …
123 #define ECB_STDC_VERSION __STDC_VERSION__ 144 #define ECB_STDC_VERSION __STDC_VERSION__
124#endif 145#endif
125 146
126#define ECB_C99 (ECB_STDC_VERSION >= 199901L) 147#define ECB_C99 (ECB_STDC_VERSION >= 199901L)
127#define ECB_C11 (ECB_STDC_VERSION >= 201112L) 148#define ECB_C11 (ECB_STDC_VERSION >= 201112L)
149#define ECB_C17 (ECB_STDC_VERSION >= 201710L)
128 150
129#if ECB_CPP 151#if ECB_CPP
130 #define ECB_EXTERN_C extern "C" 152 #define ECB_EXTERN_C extern "C"
131 #define ECB_EXTERN_C_BEG ECB_EXTERN_C { 153 #define ECB_EXTERN_C_BEG ECB_EXTERN_C {
132 #define ECB_EXTERN_C_END } 154 #define ECB_EXTERN_C_END }
… …
147 169
148#if ECB_NO_SMP 170#if ECB_NO_SMP
149 #define ECB_MEMORY_FENCE do { } while (0) 171 #define ECB_MEMORY_FENCE do { } while (0)
150#endif 172#endif
151 173
174/* http://www-01.ibm.com/support/knowledgecenter/SSGH3R_13.1.0/com.ibm.xlcpp131.aix.doc/compiler_ref/compiler_builtins.html */
175#if __xlC__ && ECB_CPP
176 #include <builtins.h>
177#endif
178
179#if 1400 <= _MSC_VER
180 #include <intrin.h> /* fence functions _ReadBarrier, also bit search functions _BitScanReverse */
181#endif
182
152#ifndef ECB_MEMORY_FENCE 183#ifndef ECB_MEMORY_FENCE
153 #if ECB_GCC_VERSION(2,5) || defined __INTEL_COMPILER || (__llvm__ && __GNUC__) || __SUNPRO_C >= 0x5110 || __SUNPRO_CC >= 0x5110 184 #if ECB_GCC_VERSION(2,5) || defined __INTEL_COMPILER || (__llvm__ && __GNUC__) || __SUNPRO_C >= 0x5110 || __SUNPRO_CC >= 0x5110
185 #define ECB_MEMORY_FENCE_RELAXED __asm__ __volatile__ ("" : : : "memory")
154 #if __i386 || __i386__ 186 #if __i386 || __i386__
155 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("lock; orb $0, -1(%%esp)" : : : "memory") 187 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("lock; orb $0, -1(%%esp)" : : : "memory")
156 #define ECB_MEMORY_FENCE_ACQUIRE __asm__ __volatile__ ("" : : : "memory") 188 #define ECB_MEMORY_FENCE_ACQUIRE __asm__ __volatile__ ("" : : : "memory")
157 #define ECB_MEMORY_FENCE_RELEASE __asm__ __volatile__ ("") 189 #define ECB_MEMORY_FENCE_RELEASE __asm__ __volatile__ ("" : : : "memory")
158 #elif __amd64 || __amd64__ || __x86_64 || __x86_64__ 190 #elif ECB_GCC_AMD64
159 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("mfence" : : : "memory") 191 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("mfence" : : : "memory")
160 #define ECB_MEMORY_FENCE_ACQUIRE __asm__ __volatile__ ("" : : : "memory") 192 #define ECB_MEMORY_FENCE_ACQUIRE __asm__ __volatile__ ("" : : : "memory")
161 #define ECB_MEMORY_FENCE_RELEASE __asm__ __volatile__ ("") 193 #define ECB_MEMORY_FENCE_RELEASE __asm__ __volatile__ ("" : : : "memory")
162 #elif __powerpc__ || __ppc__ || __powerpc64__ || __ppc64__ 194 #elif __powerpc__ || __ppc__ || __powerpc64__ || __ppc64__
163 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("sync" : : : "memory") 195 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("sync" : : : "memory")
196 #elif defined __ARM_ARCH_2__ \
197 || defined __ARM_ARCH_3__ || defined __ARM_ARCH_3M__ \
198 || defined __ARM_ARCH_4__ || defined __ARM_ARCH_4T__ \
199 || defined __ARM_ARCH_5__ || defined __ARM_ARCH_5E__ \
200 || defined __ARM_ARCH_5T__ || defined __ARM_ARCH_5TE__ \
201 || defined __ARM_ARCH_5TEJ__
202 /* should not need any, unless running old code on newer cpu - arm doesn't support that */
164 #elif defined __ARM_ARCH_6__ || defined __ARM_ARCH_6J__ \ 203 #elif defined __ARM_ARCH_6__ || defined __ARM_ARCH_6J__ \
165 || defined __ARM_ARCH_6K__ || defined __ARM_ARCH_6ZK__ 204 || defined __ARM_ARCH_6K__ || defined __ARM_ARCH_6ZK__ \
205 || defined __ARM_ARCH_6T2__
166 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("mcr p15,0,%0,c7,c10,5" : : "r" (0) : "memory") 206 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("mcr p15,0,%0,c7,c10,5" : : "r" (0) : "memory")
167 #elif defined __ARM_ARCH_7__ || defined __ARM_ARCH_7A__ \ 207 #elif defined __ARM_ARCH_7__ || defined __ARM_ARCH_7A__ \
168 || defined __ARM_ARCH_7M__ || defined __ARM_ARCH_7R__ 208 || defined __ARM_ARCH_7R__ || defined __ARM_ARCH_7M__
169 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("dmb" : : : "memory") 209 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("dmb" : : : "memory")
170 #elif __aarch64__ 210 #elif __aarch64__
171 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("dmb ish" : : : "memory") 211 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("dmb ish" : : : "memory")
172 #elif (__sparc || __sparc__) && !__sparcv8 212 #elif (__sparc || __sparc__) && !(__sparc_v8__ || defined __sparcv8)
173 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("membar #LoadStore | #LoadLoad | #StoreStore | #StoreLoad" : : : "memory") 213 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("membar #LoadStore | #LoadLoad | #StoreStore | #StoreLoad" : : : "memory")
174 #define ECB_MEMORY_FENCE_ACQUIRE __asm__ __volatile__ ("membar #LoadStore | #LoadLoad" : : : "memory") 214 #define ECB_MEMORY_FENCE_ACQUIRE __asm__ __volatile__ ("membar #LoadStore | #LoadLoad" : : : "memory")
175 #define ECB_MEMORY_FENCE_RELEASE __asm__ __volatile__ ("membar #LoadStore | #StoreStore") 215 #define ECB_MEMORY_FENCE_RELEASE __asm__ __volatile__ ("membar #LoadStore | #StoreStore")
176 #elif defined __s390__ || defined __s390x__ 216 #elif defined __s390__ || defined __s390x__
177 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("bcr 15,0" : : : "memory") 217 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("bcr 15,0" : : : "memory")
… …
200 #if ECB_GCC_VERSION(4,7) 240 #if ECB_GCC_VERSION(4,7)
201 /* see comment below (stdatomic.h) about the C11 memory model. */ 241 /* see comment below (stdatomic.h) about the C11 memory model. */
202 #define ECB_MEMORY_FENCE __atomic_thread_fence (__ATOMIC_SEQ_CST) 242 #define ECB_MEMORY_FENCE __atomic_thread_fence (__ATOMIC_SEQ_CST)
203 #define ECB_MEMORY_FENCE_ACQUIRE __atomic_thread_fence (__ATOMIC_ACQUIRE) 243 #define ECB_MEMORY_FENCE_ACQUIRE __atomic_thread_fence (__ATOMIC_ACQUIRE)
204 #define ECB_MEMORY_FENCE_RELEASE __atomic_thread_fence (__ATOMIC_RELEASE) 244 #define ECB_MEMORY_FENCE_RELEASE __atomic_thread_fence (__ATOMIC_RELEASE)
245 #define ECB_MEMORY_FENCE_RELAXED __atomic_thread_fence (__ATOMIC_RELAXED)
205 246
206 #elif ECB_CLANG_EXTENSION(c_atomic) 247 #elif ECB_CLANG_EXTENSION(c_atomic)
207 /* see comment below (stdatomic.h) about the C11 memory model. */ 248 /* see comment below (stdatomic.h) about the C11 memory model. */
208 #define ECB_MEMORY_FENCE __c11_atomic_thread_fence (__ATOMIC_SEQ_CST) 249 #define ECB_MEMORY_FENCE __c11_atomic_thread_fence (__ATOMIC_SEQ_CST)
209 #define ECB_MEMORY_FENCE_ACQUIRE __c11_atomic_thread_fence (__ATOMIC_ACQUIRE) 250 #define ECB_MEMORY_FENCE_ACQUIRE __c11_atomic_thread_fence (__ATOMIC_ACQUIRE)
210 #define ECB_MEMORY_FENCE_RELEASE __c11_atomic_thread_fence (__ATOMIC_RELEASE) 251 #define ECB_MEMORY_FENCE_RELEASE __c11_atomic_thread_fence (__ATOMIC_RELEASE)
252 #define ECB_MEMORY_FENCE_RELAXED __c11_atomic_thread_fence (__ATOMIC_RELAXED)
211 253
212 #elif ECB_GCC_VERSION(4,4) || defined __INTEL_COMPILER || defined __clang__ 254 #elif ECB_GCC_VERSION(4,4) || defined __INTEL_COMPILER || defined __clang__
213 #define ECB_MEMORY_FENCE __sync_synchronize () 255 #define ECB_MEMORY_FENCE __sync_synchronize ()
214 #elif _MSC_VER >= 1500 /* VC++ 2008 */ 256 #elif _MSC_VER >= 1500 /* VC++ 2008 */
215 /* apparently, microsoft broke all the memory barrier stuff in Visual Studio 2008... */ 257 /* apparently, microsoft broke all the memory barrier stuff in Visual Studio 2008... */
… …
225 #elif defined _WIN32 267 #elif defined _WIN32
226 #include <WinNT.h> 268 #include <WinNT.h>
227 #define ECB_MEMORY_FENCE MemoryBarrier () /* actually just xchg on x86... scary */ 269 #define ECB_MEMORY_FENCE MemoryBarrier () /* actually just xchg on x86... scary */
228 #elif __SUNPRO_C >= 0x5110 || __SUNPRO_CC >= 0x5110 270 #elif __SUNPRO_C >= 0x5110 || __SUNPRO_CC >= 0x5110
229 #include <mbarrier.h> 271 #include <mbarrier.h>
230 #define ECB_MEMORY_FENCE __machine_rw_barrier () 272 #define ECB_MEMORY_FENCE __machine_rw_barrier ()
231 #define ECB_MEMORY_FENCE_ACQUIRE __machine_r_barrier () 273 #define ECB_MEMORY_FENCE_ACQUIRE __machine_acq_barrier ()
232 #define ECB_MEMORY_FENCE_RELEASE __machine_w_barrier () 274 #define ECB_MEMORY_FENCE_RELEASE __machine_rel_barrier ()
275 #define ECB_MEMORY_FENCE_RELAXED __compiler_barrier ()
233 #elif __xlC__ 276 #elif __xlC__
234 #define ECB_MEMORY_FENCE __sync () 277 #define ECB_MEMORY_FENCE __sync ()
235 #endif 278 #endif
236#endif 279#endif
237 280
238#ifndef ECB_MEMORY_FENCE 281#ifndef ECB_MEMORY_FENCE
239 #if ECB_C11 && !defined __STDC_NO_ATOMICS__ 282 #if ECB_C11 && !defined __STDC_NO_ATOMICS__
240 /* we assume that these memory fences work on all variables/all memory accesses, */ 283 /* we assume that these memory fences work on all variables/all memory accesses, */
241 /* not just C11 atomics and atomic accesses */ 284 /* not just C11 atomics and atomic accesses */
242 #include <stdatomic.h> 285 #include <stdatomic.h>
243 /* Unfortunately, neither gcc 4.7 nor clang 3.1 generate any instructions for */
244 /* any fence other than seq_cst, which isn't very efficient for us. */
245 /* Why that is, we don't know - either the C11 memory model is quite useless */
246 /* for most usages, or gcc and clang have a bug */
247 /* I *currently* lean towards the latter, and inefficiently implement */
248 /* all three of ecb's fences as a seq_cst fence */
249 /* Update, gcc-4.8 generates mfence for all c++ fences, but nothing */
250 /* for all __atomic_thread_fence's except seq_cst */
251 #define ECB_MEMORY_FENCE atomic_thread_fence (memory_order_seq_cst) 286 #define ECB_MEMORY_FENCE atomic_thread_fence (memory_order_seq_cst)
287 #define ECB_MEMORY_FENCE_ACQUIRE atomic_thread_fence (memory_order_acquire)
288 #define ECB_MEMORY_FENCE_RELEASE atomic_thread_fence (memory_order_release)
252 #endif 289 #endif
253#endif 290#endif
254 291
255#ifndef ECB_MEMORY_FENCE 292#ifndef ECB_MEMORY_FENCE
256 #if !ECB_AVOID_PTHREADS 293 #if !ECB_AVOID_PTHREADS
… …
276 313
277#if !defined ECB_MEMORY_FENCE_RELEASE && defined ECB_MEMORY_FENCE 314#if !defined ECB_MEMORY_FENCE_RELEASE && defined ECB_MEMORY_FENCE
278 #define ECB_MEMORY_FENCE_RELEASE ECB_MEMORY_FENCE 315 #define ECB_MEMORY_FENCE_RELEASE ECB_MEMORY_FENCE
279#endif 316#endif
280 317
318#if !defined ECB_MEMORY_FENCE_RELAXED && defined ECB_MEMORY_FENCE
319 #define ECB_MEMORY_FENCE_RELAXED ECB_MEMORY_FENCE /* very heavy-handed */
320#endif
321
281/*****************************************************************************/ 322/*****************************************************************************/
282 323
283#if ECB_CPP 324#if ECB_CPP
284 #define ecb_inline static inline 325 #define ecb_inline static inline
285#elif ECB_GCC_VERSION(2,5) 326#elif ECB_GCC_VERSION(2,5)
… …
302 343
303#define ECB_CONCAT_(a, b) a ## b 344#define ECB_CONCAT_(a, b) a ## b
304#define ECB_CONCAT(a, b) ECB_CONCAT_(a, b) 345#define ECB_CONCAT(a, b) ECB_CONCAT_(a, b)
305#define ECB_STRINGIFY_(a) # a 346#define ECB_STRINGIFY_(a) # a
306#define ECB_STRINGIFY(a) ECB_STRINGIFY_(a) 347#define ECB_STRINGIFY(a) ECB_STRINGIFY_(a)
348#define ECB_STRINGIFY_EXPR(expr) ((expr), ECB_STRINGIFY_ (expr))
307 349
308#define ecb_function_ ecb_inline 350#define ecb_function_ ecb_inline
309 351
310#if ECB_GCC_VERSION(3,1) || ECB_CLANG_VERSION(2,8) 352#if ECB_GCC_VERSION(3,1) || ECB_CLANG_VERSION(2,8)
311 #define ecb_attribute(attrlist) __attribute__ (attrlist) 353 #define ecb_attribute(attrlist) __attribute__ (attrlist)
… …
348 #define ecb_deprecated __declspec (deprecated) 390 #define ecb_deprecated __declspec (deprecated)
349#else 391#else
350 #define ecb_deprecated ecb_attribute ((__deprecated__)) 392 #define ecb_deprecated ecb_attribute ((__deprecated__))
351#endif 393#endif
352 394
395#if _MSC_VER >= 1500
396 #define ecb_deprecated_message(msg) __declspec (deprecated (msg))
397#elif ECB_GCC_VERSION(4,5)
398 #define ecb_deprecated_message(msg) ecb_attribute ((__deprecated__ (msg))
399#else
400 #define ecb_deprecated_message(msg) ecb_deprecated
401#endif
402
403#if _MSC_VER >= 1400
404 #define ecb_noinline __declspec (noinline)
405#else
353#define ecb_noinline ecb_attribute ((__noinline__)) 406 #define ecb_noinline ecb_attribute ((__noinline__))
407#endif
408
354#define ecb_unused ecb_attribute ((__unused__)) 409#define ecb_unused ecb_attribute ((__unused__))
355#define ecb_const ecb_attribute ((__const__)) 410#define ecb_const ecb_attribute ((__const__))
356#define ecb_pure ecb_attribute ((__pure__)) 411#define ecb_pure ecb_attribute ((__pure__))
357 412
358/* TODO http://msdn.microsoft.com/en-us/library/k6ktzx3s.aspx __declspec(noreturn) */
359#if ECB_C11 || __IBMC_NORETURN 413#if ECB_C11 || __IBMC_NORETURN
360 /* http://pic.dhe.ibm.com/infocenter/compbg/v121v141/topic/com.ibm.xlcpp121.bg.doc/language_ref/noreturn.html */ 414 /* http://www-01.ibm.com/support/knowledgecenter/SSGH3R_13.1.0/com.ibm.xlcpp131.aix.doc/language_ref/noreturn.html */
361 #define ecb_noreturn _Noreturn 415 #define ecb_noreturn _Noreturn
416#elif ECB_CPP11
417 #define ecb_noreturn [[noreturn]]
418#elif _MSC_VER >= 1200
419 /* http://msdn.microsoft.com/en-us/library/k6ktzx3s.aspx */
420 #define ecb_noreturn __declspec (noreturn)
362#else 421#else
363 #define ecb_noreturn ecb_attribute ((__noreturn__)) 422 #define ecb_noreturn ecb_attribute ((__noreturn__))
364#endif 423#endif
365 424
366#if ECB_GCC_VERSION(4,3) 425#if ECB_GCC_VERSION(4,3)
… …
393 #define ecb_ctz32(x) __builtin_ctz (x) 452 #define ecb_ctz32(x) __builtin_ctz (x)
394 #define ecb_ctz64(x) __builtin_ctzll (x) 453 #define ecb_ctz64(x) __builtin_ctzll (x)
395 #define ecb_popcount32(x) __builtin_popcount (x) 454 #define ecb_popcount32(x) __builtin_popcount (x)
396 /* no popcountll */ 455 /* no popcountll */
397#else 456#else
398 ecb_function_ int ecb_ctz32 (uint32_t x) ecb_const; 457 ecb_function_ ecb_const int ecb_ctz32 (uint32_t x);
399 ecb_function_ int 458 ecb_function_ ecb_const int
400 ecb_ctz32 (uint32_t x) 459 ecb_ctz32 (uint32_t x)
401 { 460 {
461#if 1400 <= _MSC_VER && (_M_IX86 || _M_X64 || _M_IA64 || _M_ARM)
462 unsigned long r;
463 _BitScanForward (&r, x);
464 return (int)r;
465#else
402 int r = 0; 466 int r = 0;
403 467
404 x &= ~x + 1; /* this isolates the lowest bit */ 468 x &= ~x + 1; /* this isolates the lowest bit */
405 469
406#if ECB_branchless_on_i386 470#if ECB_branchless_on_i386
… …
416 if (x & 0xff00ff00) r += 8; 480 if (x & 0xff00ff00) r += 8;
417 if (x & 0xffff0000) r += 16; 481 if (x & 0xffff0000) r += 16;
418#endif 482#endif
419 483
420 return r; 484 return r;
485#endif
421 } 486 }
422 487
423 ecb_function_ int ecb_ctz64 (uint64_t x) ecb_const; 488 ecb_function_ ecb_const int ecb_ctz64 (uint64_t x);
424 ecb_function_ int 489 ecb_function_ ecb_const int
425 ecb_ctz64 (uint64_t x) 490 ecb_ctz64 (uint64_t x)
426 { 491 {
492#if 1400 <= _MSC_VER && (_M_X64 || _M_IA64 || _M_ARM)
493 unsigned long r;
494 _BitScanForward64 (&r, x);
495 return (int)r;
496#else
427 int shift = x & 0xffffffffU ? 0 : 32; 497 int shift = x & 0xffffffff ? 0 : 32;
428 return ecb_ctz32 (x >> shift) + shift; 498 return ecb_ctz32 (x >> shift) + shift;
499#endif
429 } 500 }
430 501
431 ecb_function_ int ecb_popcount32 (uint32_t x) ecb_const; 502 ecb_function_ ecb_const int ecb_popcount32 (uint32_t x);
432 ecb_function_ int 503 ecb_function_ ecb_const int
433 ecb_popcount32 (uint32_t x) 504 ecb_popcount32 (uint32_t x)
434 { 505 {
435 x -= (x >> 1) & 0x55555555; 506 x -= (x >> 1) & 0x55555555;
436 x = ((x >> 2) & 0x33333333) + (x & 0x33333333); 507 x = ((x >> 2) & 0x33333333) + (x & 0x33333333);
437 x = ((x >> 4) + x) & 0x0f0f0f0f; 508 x = ((x >> 4) + x) & 0x0f0f0f0f;
438 x *= 0x01010101; 509 x *= 0x01010101;
439 510
440 return x >> 24; 511 return x >> 24;
441 } 512 }
442 513
443 ecb_function_ int ecb_ld32 (uint32_t x) ecb_const; 514 ecb_function_ ecb_const int ecb_ld32 (uint32_t x);
444 ecb_function_ int ecb_ld32 (uint32_t x) 515 ecb_function_ ecb_const int ecb_ld32 (uint32_t x)
445 { 516 {
517#if 1400 <= _MSC_VER && (_M_IX86 || _M_X64 || _M_IA64 || _M_ARM)
518 unsigned long r;
519 _BitScanReverse (&r, x);
520 return (int)r;
521#else
446 int r = 0; 522 int r = 0;
447 523
448 if (x >> 16) { x >>= 16; r += 16; } 524 if (x >> 16) { x >>= 16; r += 16; }
449 if (x >> 8) { x >>= 8; r += 8; } 525 if (x >> 8) { x >>= 8; r += 8; }
450 if (x >> 4) { x >>= 4; r += 4; } 526 if (x >> 4) { x >>= 4; r += 4; }
451 if (x >> 2) { x >>= 2; r += 2; } 527 if (x >> 2) { x >>= 2; r += 2; }
452 if (x >> 1) { r += 1; } 528 if (x >> 1) { r += 1; }
453 529
454 return r; 530 return r;
531#endif
455 } 532 }
456 533
457 ecb_function_ int ecb_ld64 (uint64_t x) ecb_const; 534 ecb_function_ ecb_const int ecb_ld64 (uint64_t x);
458 ecb_function_ int ecb_ld64 (uint64_t x) 535 ecb_function_ ecb_const int ecb_ld64 (uint64_t x)
459 { 536 {
537#if 1400 <= _MSC_VER && (_M_X64 || _M_IA64 || _M_ARM)
538 unsigned long r;
539 _BitScanReverse64 (&r, x);
540 return (int)r;
541#else
460 int r = 0; 542 int r = 0;
461 543
462 if (x >> 32) { x >>= 32; r += 32; } 544 if (x >> 32) { x >>= 32; r += 32; }
463 545
464 return r + ecb_ld32 (x); 546 return r + ecb_ld32 (x);
465 }
466#endif 547#endif
548 }
549#endif
467 550
468ecb_function_ ecb_bool ecb_is_pot32 (uint32_t x) ecb_const; 551ecb_function_ ecb_const ecb_bool ecb_is_pot32 (uint32_t x);
469ecb_function_ ecb_bool ecb_is_pot32 (uint32_t x) { return !(x & (x - 1)); } 552ecb_function_ ecb_const ecb_bool ecb_is_pot32 (uint32_t x) { return !(x & (x - 1)); }
470ecb_function_ ecb_bool ecb_is_pot64 (uint64_t x) ecb_const; 553ecb_function_ ecb_const ecb_bool ecb_is_pot64 (uint64_t x);
471ecb_function_ ecb_bool ecb_is_pot64 (uint64_t x) { return !(x & (x - 1)); } 554ecb_function_ ecb_const ecb_bool ecb_is_pot64 (uint64_t x) { return !(x & (x - 1)); }
472 555
473ecb_function_ uint8_t ecb_bitrev8 (uint8_t x) ecb_const; 556ecb_function_ ecb_const uint8_t ecb_bitrev8 (uint8_t x);
474ecb_function_ uint8_t ecb_bitrev8 (uint8_t x) 557ecb_function_ ecb_const uint8_t ecb_bitrev8 (uint8_t x)
475{ 558{
476 return ( (x * 0x0802U & 0x22110U) 559 return ( (x * 0x0802U & 0x22110U)
477 | (x * 0x8020U & 0x88440U)) * 0x10101U >> 16; 560 | (x * 0x8020U & 0x88440U)) * 0x10101U >> 16;
478} 561}
479 562
480ecb_function_ uint16_t ecb_bitrev16 (uint16_t x) ecb_const; 563ecb_function_ ecb_const uint16_t ecb_bitrev16 (uint16_t x);
481ecb_function_ uint16_t ecb_bitrev16 (uint16_t x) 564ecb_function_ ecb_const uint16_t ecb_bitrev16 (uint16_t x)
482{ 565{
483 x = ((x >> 1) & 0x5555) | ((x & 0x5555) << 1); 566 x = ((x >> 1) & 0x5555) | ((x & 0x5555) << 1);
484 x = ((x >> 2) & 0x3333) | ((x & 0x3333) << 2); 567 x = ((x >> 2) & 0x3333) | ((x & 0x3333) << 2);
485 x = ((x >> 4) & 0x0f0f) | ((x & 0x0f0f) << 4); 568 x = ((x >> 4) & 0x0f0f) | ((x & 0x0f0f) << 4);
486 x = ( x >> 8 ) | ( x << 8); 569 x = ( x >> 8 ) | ( x << 8);
487 570
488 return x; 571 return x;
489} 572}
490 573
491ecb_function_ uint32_t ecb_bitrev32 (uint32_t x) ecb_const; 574ecb_function_ ecb_const uint32_t ecb_bitrev32 (uint32_t x);
492ecb_function_ uint32_t ecb_bitrev32 (uint32_t x) 575ecb_function_ ecb_const uint32_t ecb_bitrev32 (uint32_t x)
493{ 576{
494 x = ((x >> 1) & 0x55555555) | ((x & 0x55555555) << 1); 577 x = ((x >> 1) & 0x55555555) | ((x & 0x55555555) << 1);
495 x = ((x >> 2) & 0x33333333) | ((x & 0x33333333) << 2); 578 x = ((x >> 2) & 0x33333333) | ((x & 0x33333333) << 2);
496 x = ((x >> 4) & 0x0f0f0f0f) | ((x & 0x0f0f0f0f) << 4); 579 x = ((x >> 4) & 0x0f0f0f0f) | ((x & 0x0f0f0f0f) << 4);
497 x = ((x >> 8) & 0x00ff00ff) | ((x & 0x00ff00ff) << 8); 580 x = ((x >> 8) & 0x00ff00ff) | ((x & 0x00ff00ff) << 8);
… …
500 return x; 583 return x;
501} 584}
502 585
503/* popcount64 is only available on 64 bit cpus as gcc builtin */ 586/* popcount64 is only available on 64 bit cpus as gcc builtin */
504/* so for this version we are lazy */ 587/* so for this version we are lazy */
505ecb_function_ int ecb_popcount64 (uint64_t x) ecb_const; 588ecb_function_ ecb_const int ecb_popcount64 (uint64_t x);
506ecb_function_ int 589ecb_function_ ecb_const int
507ecb_popcount64 (uint64_t x) 590ecb_popcount64 (uint64_t x)
508{ 591{
509 return ecb_popcount32 (x) + ecb_popcount32 (x >> 32); 592 return ecb_popcount32 (x) + ecb_popcount32 (x >> 32);
510} 593}
511 594
512ecb_inline uint8_t ecb_rotl8 (uint8_t x, unsigned int count) ecb_const; 595ecb_inline ecb_const uint8_t ecb_rotl8 (uint8_t x, unsigned int count);
513ecb_inline uint8_t ecb_rotr8 (uint8_t x, unsigned int count) ecb_const; 596ecb_inline ecb_const uint8_t ecb_rotr8 (uint8_t x, unsigned int count);
514ecb_inline uint16_t ecb_rotl16 (uint16_t x, unsigned int count) ecb_const; 597ecb_inline ecb_const uint16_t ecb_rotl16 (uint16_t x, unsigned int count);
515ecb_inline uint16_t ecb_rotr16 (uint16_t x, unsigned int count) ecb_const; 598ecb_inline ecb_const uint16_t ecb_rotr16 (uint16_t x, unsigned int count);
516ecb_inline uint32_t ecb_rotl32 (uint32_t x, unsigned int count) ecb_const; 599ecb_inline ecb_const uint32_t ecb_rotl32 (uint32_t x, unsigned int count);
517ecb_inline uint32_t ecb_rotr32 (uint32_t x, unsigned int count) ecb_const; 600ecb_inline ecb_const uint32_t ecb_rotr32 (uint32_t x, unsigned int count);
518ecb_inline uint64_t ecb_rotl64 (uint64_t x, unsigned int count) ecb_const; 601ecb_inline ecb_const uint64_t ecb_rotl64 (uint64_t x, unsigned int count);
519ecb_inline uint64_t ecb_rotr64 (uint64_t x, unsigned int count) ecb_const; 602ecb_inline ecb_const uint64_t ecb_rotr64 (uint64_t x, unsigned int count);
520 603
521ecb_inline uint8_t ecb_rotl8 (uint8_t x, unsigned int count) { return (x >> ( 8 - count)) | (x << count); } 604ecb_inline ecb_const uint8_t ecb_rotl8 (uint8_t x, unsigned int count) { return (x >> ( 8 - count)) | (x << count); }
522ecb_inline uint8_t ecb_rotr8 (uint8_t x, unsigned int count) { return (x << ( 8 - count)) | (x >> count); } 605ecb_inline ecb_const uint8_t ecb_rotr8 (uint8_t x, unsigned int count) { return (x << ( 8 - count)) | (x >> count); }
523ecb_inline uint16_t ecb_rotl16 (uint16_t x, unsigned int count) { return (x >> (16 - count)) | (x << count); } 606ecb_inline ecb_const uint16_t ecb_rotl16 (uint16_t x, unsigned int count) { return (x >> (16 - count)) | (x << count); }
524ecb_inline uint16_t ecb_rotr16 (uint16_t x, unsigned int count) { return (x << (16 - count)) | (x >> count); } 607ecb_inline ecb_const uint16_t ecb_rotr16 (uint16_t x, unsigned int count) { return (x << (16 - count)) | (x >> count); }
525ecb_inline uint32_t ecb_rotl32 (uint32_t x, unsigned int count) { return (x >> (32 - count)) | (x << count); } 608ecb_inline ecb_const uint32_t ecb_rotl32 (uint32_t x, unsigned int count) { return (x >> (32 - count)) | (x << count); }
526ecb_inline uint32_t ecb_rotr32 (uint32_t x, unsigned int count) { return (x << (32 - count)) | (x >> count); } 609ecb_inline ecb_const uint32_t ecb_rotr32 (uint32_t x, unsigned int count) { return (x << (32 - count)) | (x >> count); }
527ecb_inline uint64_t ecb_rotl64 (uint64_t x, unsigned int count) { return (x >> (64 - count)) | (x << count); } 610ecb_inline ecb_const uint64_t ecb_rotl64 (uint64_t x, unsigned int count) { return (x >> (64 - count)) | (x << count); }
528ecb_inline uint64_t ecb_rotr64 (uint64_t x, unsigned int count) { return (x << (64 - count)) | (x >> count); } 611ecb_inline ecb_const uint64_t ecb_rotr64 (uint64_t x, unsigned int count) { return (x << (64 - count)) | (x >> count); }
529 612
530#if ECB_GCC_VERSION(4,3) || (ECB_CLANG_BUILTIN(__builtin_bswap32) && ECB_CLANG_BUILTIN(__builtin_bswap64)) 613#if ECB_GCC_VERSION(4,3) || (ECB_CLANG_BUILTIN(__builtin_bswap32) && ECB_CLANG_BUILTIN(__builtin_bswap64))
614 #if ECB_GCC_VERSION(4,8) || ECB_CLANG_BUILTIN(__builtin_bswap16)
615 #define ecb_bswap16(x) __builtin_bswap16 (x)
616 #else
531 #define ecb_bswap16(x) (__builtin_bswap32 (x) >> 16) 617 #define ecb_bswap16(x) (__builtin_bswap32 (x) >> 16)
618 #endif
532 #define ecb_bswap32(x) __builtin_bswap32 (x) 619 #define ecb_bswap32(x) __builtin_bswap32 (x)
533 #define ecb_bswap64(x) __builtin_bswap64 (x) 620 #define ecb_bswap64(x) __builtin_bswap64 (x)
621#elif _MSC_VER
622 #include <stdlib.h>
623 #define ecb_bswap16(x) ((uint16_t)_byteswap_ushort ((uint16_t)(x)))
624 #define ecb_bswap32(x) ((uint32_t)_byteswap_ulong ((uint32_t)(x)))
625 #define ecb_bswap64(x) ((uint64_t)_byteswap_uint64 ((uint64_t)(x)))
534#else 626#else
535 ecb_function_ uint16_t ecb_bswap16 (uint16_t x) ecb_const; 627 ecb_function_ ecb_const uint16_t ecb_bswap16 (uint16_t x);
536 ecb_function_ uint16_t 628 ecb_function_ ecb_const uint16_t
537 ecb_bswap16 (uint16_t x) 629 ecb_bswap16 (uint16_t x)
538 { 630 {
539 return ecb_rotl16 (x, 8); 631 return ecb_rotl16 (x, 8);
540 } 632 }
541 633
542 ecb_function_ uint32_t ecb_bswap32 (uint32_t x) ecb_const; 634 ecb_function_ ecb_const uint32_t ecb_bswap32 (uint32_t x);
543 ecb_function_ uint32_t 635 ecb_function_ ecb_const uint32_t
544 ecb_bswap32 (uint32_t x) 636 ecb_bswap32 (uint32_t x)
545 { 637 {
546 return (((uint32_t)ecb_bswap16 (x)) << 16) | ecb_bswap16 (x >> 16); 638 return (((uint32_t)ecb_bswap16 (x)) << 16) | ecb_bswap16 (x >> 16);
547 } 639 }
548 640
549 ecb_function_ uint64_t ecb_bswap64 (uint64_t x) ecb_const; 641 ecb_function_ ecb_const uint64_t ecb_bswap64 (uint64_t x);
550 ecb_function_ uint64_t 642 ecb_function_ ecb_const uint64_t
551 ecb_bswap64 (uint64_t x) 643 ecb_bswap64 (uint64_t x)
552 { 644 {
553 return (((uint64_t)ecb_bswap32 (x)) << 32) | ecb_bswap32 (x >> 32); 645 return (((uint64_t)ecb_bswap32 (x)) << 32) | ecb_bswap32 (x >> 32);
554 } 646 }
555#endif 647#endif
556 648
557#if ECB_GCC_VERSION(4,5) || ECB_CLANG_BUILTIN(__builtin_unreachable) 649#if ECB_GCC_VERSION(4,5) || ECB_CLANG_BUILTIN(__builtin_unreachable)
558 #define ecb_unreachable() __builtin_unreachable () 650 #define ecb_unreachable() __builtin_unreachable ()
559#else 651#else
560 /* this seems to work fine, but gcc always emits a warning for it :/ */ 652 /* this seems to work fine, but gcc always emits a warning for it :/ */
561 ecb_inline void ecb_unreachable (void) ecb_noreturn; 653 ecb_inline ecb_noreturn void ecb_unreachable (void);
562 ecb_inline void ecb_unreachable (void) { } 654 ecb_inline ecb_noreturn void ecb_unreachable (void) { }
563#endif 655#endif
564 656
565/* try to tell the compiler that some condition is definitely true */ 657/* try to tell the compiler that some condition is definitely true */
566#define ecb_assume(cond) if (!(cond)) ecb_unreachable (); else 0 658#define ecb_assume(cond) if (!(cond)) ecb_unreachable (); else 0
567 659
568ecb_inline unsigned char ecb_byteorder_helper (void) ecb_const; 660ecb_inline ecb_const uint32_t ecb_byteorder_helper (void);
569ecb_inline unsigned char 661ecb_inline ecb_const uint32_t
570ecb_byteorder_helper (void) 662ecb_byteorder_helper (void)
571{ 663{
572 /* the union code still generates code under pressure in gcc, */ 664 /* the union code still generates code under pressure in gcc, */
573 /* but less than using pointers, and always seems to */ 665 /* but less than using pointers, and always seems to */
574 /* successfully return a constant. */ 666 /* successfully return a constant. */
575 /* the reason why we have this horrible preprocessor mess */ 667 /* the reason why we have this horrible preprocessor mess */
576 /* is to avoid it in all cases, at least on common architectures */ 668 /* is to avoid it in all cases, at least on common architectures */
577 /* or when using a recent enough gcc version (>= 4.6) */ 669 /* or when using a recent enough gcc version (>= 4.6) */
578#if __i386 || __i386__ || _M_X86 || __amd64 || __amd64__ || _M_X64
579 return 0x44;
580#elif __BYTE_ORDER__ && __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__ 670#if (defined __BYTE_ORDER__ && __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__) \
671 || ((__i386 || __i386__ || _M_IX86 || ECB_GCC_AMD64 || ECB_MSVC_AMD64) && !__VOS__)
672 #define ECB_LITTLE_ENDIAN 1
581 return 0x44; 673 return 0x44332211;
582#elif __BYTE_ORDER__ && __BYTE_ORDER__ == __ORDER_BIG_ENDIAN__ 674#elif (defined __BYTE_ORDER__ && __BYTE_ORDER__ == __ORDER_BIG_ENDIAN__) \
675 || ((__AARCH64EB__ || __MIPSEB__ || __ARMEB__) && !__VOS__)
676 #define ECB_BIG_ENDIAN 1
583 return 0x11; 677 return 0x11223344;
584#else 678#else
585 union 679 union
586 { 680 {
681 uint8_t c[4];
587 uint32_t i; 682 uint32_t u;
588 uint8_t c;
589 } u = { 0x11223344 }; 683 } u = { 0x11, 0x22, 0x33, 0x44 };
590 return u.c; 684 return u.u;
591#endif 685#endif
592} 686}
593 687
594ecb_inline ecb_bool ecb_big_endian (void) ecb_const; 688ecb_inline ecb_const ecb_bool ecb_big_endian (void);
595ecb_inline ecb_bool ecb_big_endian (void) { return ecb_byteorder_helper () == 0x11; } 689ecb_inline ecb_const ecb_bool ecb_big_endian (void) { return ecb_byteorder_helper () == 0x11223344; }
596ecb_inline ecb_bool ecb_little_endian (void) ecb_const; 690ecb_inline ecb_const ecb_bool ecb_little_endian (void);
597ecb_inline ecb_bool ecb_little_endian (void) { return ecb_byteorder_helper () == 0x44; } 691ecb_inline ecb_const ecb_bool ecb_little_endian (void) { return ecb_byteorder_helper () == 0x44332211; }
692
693/*****************************************************************************/
694/* unaligned load/store */
695
696ecb_inline uint_fast16_t ecb_be_u16_to_host (uint_fast16_t v) { return ecb_little_endian () ? ecb_bswap16 (v) : v; }
697ecb_inline uint_fast32_t ecb_be_u32_to_host (uint_fast32_t v) { return ecb_little_endian () ? ecb_bswap32 (v) : v; }
698ecb_inline uint_fast64_t ecb_be_u64_to_host (uint_fast64_t v) { return ecb_little_endian () ? ecb_bswap64 (v) : v; }
699
700ecb_inline uint_fast16_t ecb_le_u16_to_host (uint_fast16_t v) { return ecb_big_endian () ? ecb_bswap16 (v) : v; }
701ecb_inline uint_fast32_t ecb_le_u32_to_host (uint_fast32_t v) { return ecb_big_endian () ? ecb_bswap32 (v) : v; }
702ecb_inline uint_fast64_t ecb_le_u64_to_host (uint_fast64_t v) { return ecb_big_endian () ? ecb_bswap64 (v) : v; }
703
704ecb_inline uint_fast16_t ecb_peek_u16_u (const void *ptr) { uint16_t v; memcpy (&v, ptr, sizeof (v)); return v; }
705ecb_inline uint_fast32_t ecb_peek_u32_u (const void *ptr) { uint32_t v; memcpy (&v, ptr, sizeof (v)); return v; }
706ecb_inline uint_fast64_t ecb_peek_u64_u (const void *ptr) { uint64_t v; memcpy (&v, ptr, sizeof (v)); return v; }
707
708ecb_inline uint_fast16_t ecb_peek_be_u16_u (const void *ptr) { return ecb_be_u16_to_host (ecb_peek_u16_u (ptr)); }
709ecb_inline uint_fast32_t ecb_peek_be_u32_u (const void *ptr) { return ecb_be_u32_to_host (ecb_peek_u32_u (ptr)); }
710ecb_inline uint_fast64_t ecb_peek_be_u64_u (const void *ptr) { return ecb_be_u64_to_host (ecb_peek_u64_u (ptr)); }
711
712ecb_inline uint_fast16_t ecb_peek_le_u16_u (const void *ptr) { return ecb_le_u16_to_host (ecb_peek_u16_u (ptr)); }
713ecb_inline uint_fast32_t ecb_peek_le_u32_u (const void *ptr) { return ecb_le_u32_to_host (ecb_peek_u32_u (ptr)); }
714ecb_inline uint_fast64_t ecb_peek_le_u64_u (const void *ptr) { return ecb_le_u64_to_host (ecb_peek_u64_u (ptr)); }
715
716ecb_inline uint_fast16_t ecb_host_to_be_u16 (uint_fast16_t v) { return ecb_little_endian () ? ecb_bswap16 (v) : v; }
717ecb_inline uint_fast32_t ecb_host_to_be_u32 (uint_fast32_t v) { return ecb_little_endian () ? ecb_bswap32 (v) : v; }
718ecb_inline uint_fast64_t ecb_host_to_be_u64 (uint_fast64_t v) { return ecb_little_endian () ? ecb_bswap64 (v) : v; }
719
720ecb_inline uint_fast16_t ecb_host_to_le_u16 (uint_fast16_t v) { return ecb_big_endian () ? ecb_bswap16 (v) : v; }
721ecb_inline uint_fast32_t ecb_host_to_le_u32 (uint_fast32_t v) { return ecb_big_endian () ? ecb_bswap32 (v) : v; }
722ecb_inline uint_fast64_t ecb_host_to_le_u64 (uint_fast64_t v) { return ecb_big_endian () ? ecb_bswap64 (v) : v; }
723
724ecb_inline void ecb_poke_u16_u (void *ptr, uint16_t v) { memcpy (ptr, &v, sizeof (v)); }
725ecb_inline void ecb_poke_u32_u (void *ptr, uint32_t v) { memcpy (ptr, &v, sizeof (v)); }
726ecb_inline void ecb_poke_u64_u (void *ptr, uint64_t v) { memcpy (ptr, &v, sizeof (v)); }
727
728ecb_inline void ecb_poke_be_u16_u (void *ptr, uint_fast16_t v) { ecb_poke_u16_u (ptr, ecb_host_to_be_u16 (v)); }
729ecb_inline void ecb_poke_be_u32_u (void *ptr, uint_fast32_t v) { ecb_poke_u32_u (ptr, ecb_host_to_be_u32 (v)); }
730ecb_inline void ecb_poke_be_u64_u (void *ptr, uint_fast64_t v) { ecb_poke_u64_u (ptr, ecb_host_to_be_u64 (v)); }
731
732ecb_inline void ecb_poke_le_u16_u (void *ptr, uint_fast16_t v) { ecb_poke_u16_u (ptr, ecb_host_to_le_u16 (v)); }
733ecb_inline void ecb_poke_le_u32_u (void *ptr, uint_fast32_t v) { ecb_poke_u32_u (ptr, ecb_host_to_le_u32 (v)); }
734ecb_inline void ecb_poke_le_u64_u (void *ptr, uint_fast64_t v) { ecb_poke_u64_u (ptr, ecb_host_to_le_u64 (v)); }
735
736#ifdef ECB_CPP
737
738inline uint8_t ecb_bswap (uint8_t v) { return v; }
739inline uint16_t ecb_bswap (uint16_t v) { return ecb_bswap16 (v); }
740inline uint32_t ecb_bswap (uint32_t v) { return ecb_bswap32 (v); }
741inline uint64_t ecb_bswap (uint64_t v) { return ecb_bswap64 (v); }
742
743template<typename T> inline T ecb_be_to_host (T v) { return ecb_little_endian () ? ecb_bswap (v) : v; }
744template<typename T> inline T ecb_le_to_host (T v) { return ecb_big_endian () ? ecb_bswap (v) : v; }
745template<typename T> inline T ecb_peek (const void *ptr) { return *(const T *)ptr; }
746template<typename T> inline T ecb_peek_be (const void *ptr) { return ecb_be_to_host (ecb_peek <T> (ptr)); }
747template<typename T> inline T ecb_peek_le (const void *ptr) { return ecb_le_to_host (ecb_peek <T> (ptr)); }
748template<typename T> inline T ecb_peek_u (const void *ptr) { T v; std::memcpy (&v, ptr, sizeof (v)); return v; }
749template<typename T> inline T ecb_peek_be_u (const void *ptr) { return ecb_be_to_host (ecb_peek_u<T> (ptr)); }
750template<typename T> inline T ecb_peek_le_u (const void *ptr) { return ecb_le_to_host (ecb_peek_u<T> (ptr)); }
751
752template<typename T> inline T ecb_host_to_be (T v) { return ecb_little_endian () ? ecb_bswap (v) : v; }
753template<typename T> inline T ecb_host_to_le (T v) { return ecb_big_endian () ? ecb_bswap (v) : v; }
754template<typename T> inline void ecb_poke (void *ptr, T v) { *(T *)ptr = v; }
755template<typename T> inline void ecb_poke_be (void *ptr, T v) { return ecb_poke <T> (ptr, ecb_host_to_be (v)); }
756template<typename T> inline void ecb_poke_le (void *ptr, T v) { return ecb_poke <T> (ptr, ecb_host_to_le (v)); }
757template<typename T> inline void ecb_poke_u (void *ptr, T v) { std::memcpy (ptr, &v, sizeof (v)); }
758template<typename T> inline void ecb_poke_be_u (void *ptr, T v) { return ecb_poke_u<T> (ptr, ecb_host_to_be (v)); }
759template<typename T> inline void ecb_poke_le_u (void *ptr, T v) { return ecb_poke_u<T> (ptr, ecb_host_to_le (v)); }
760
761#endif
762
763/*****************************************************************************/
598 764
599#if ECB_GCC_VERSION(3,0) || ECB_C99 765#if ECB_GCC_VERSION(3,0) || ECB_C99
600 #define ecb_mod(m,n) ((m) % (n) + ((m) % (n) < 0 ? (n) : 0)) 766 #define ecb_mod(m,n) ((m) % (n) + ((m) % (n) < 0 ? (n) : 0))
601#else 767#else
602 #define ecb_mod(m,n) ((m) < 0 ? ((n) - 1 - ((-1 - (m)) % (n))) : ((m) % (n))) 768 #define ecb_mod(m,n) ((m) < 0 ? ((n) - 1 - ((-1 - (m)) % (n))) : ((m) % (n)))
… …
627 } 793 }
628#else 794#else
629 #define ecb_array_length(name) (sizeof (name) / sizeof (name [0])) 795 #define ecb_array_length(name) (sizeof (name) / sizeof (name [0]))
630#endif 796#endif
631 797
798/*****************************************************************************/
799
800ecb_function_ ecb_const uint32_t ecb_binary16_to_binary32 (uint32_t x);
801ecb_function_ ecb_const uint32_t
802ecb_binary16_to_binary32 (uint32_t x)
803{
804 unsigned int s = (x & 0x8000) << (31 - 15);
805 int e = (x >> 10) & 0x001f;
806 unsigned int m = x & 0x03ff;
807
808 if (ecb_expect_false (e == 31))
809 /* infinity or NaN */
810 e = 255 - (127 - 15);
811 else if (ecb_expect_false (!e))
812 {
813 if (ecb_expect_true (!m))
814 /* zero, handled by code below by forcing e to 0 */
815 e = 0 - (127 - 15);
816 else
817 {
818 /* subnormal, renormalise */
819 unsigned int s = 10 - ecb_ld32 (m);
820
821 m = (m << s) & 0x3ff; /* mask implicit bit */
822 e -= s - 1;
823 }
824 }
825
826 /* e and m now are normalised, or zero, (or inf or nan) */
827 e += 127 - 15;
828
829 return s | (e << 23) | (m << (23 - 10));
830}
831
832ecb_function_ ecb_const uint16_t ecb_binary32_to_binary16 (uint32_t x);
833ecb_function_ ecb_const uint16_t
834ecb_binary32_to_binary16 (uint32_t x)
835{
836 unsigned int s = (x >> 16) & 0x00008000; /* sign bit, the easy part */
837 unsigned int e = ((x >> 23) & 0x000000ff) - (127 - 15); /* the desired exponent */
838 unsigned int m = x & 0x007fffff;
839
840 x &= 0x7fffffff;
841
842 /* if it's within range of binary16 normals, use fast path */
843 if (ecb_expect_true (0x38800000 <= x && x <= 0x477fefff))
844 {
845 /* mantissa round-to-even */
846 m += 0x00000fff + ((m >> (23 - 10)) & 1);
847
848 /* handle overflow */
849 if (ecb_expect_false (m >= 0x00800000))
850 {
851 m >>= 1;
852 e += 1;
853 }
854
855 return s | (e << 10) | (m >> (23 - 10));
856 }
857
858 /* handle large numbers and infinity */
859 if (ecb_expect_true (0x477fefff < x && x <= 0x7f800000))
860 return s | 0x7c00;
861
862 /* handle zero, subnormals and small numbers */
863 if (ecb_expect_true (x < 0x38800000))
864 {
865 /* zero */
866 if (ecb_expect_true (!x))
867 return s;
868
869 /* handle subnormals */
870
871 /* too small, will be zero */
872 if (e < (14 - 24)) /* might not be sharp, but is good enough */
873 return s;
874
875 m |= 0x00800000; /* make implicit bit explicit */
876
877 /* very tricky - we need to round to the nearest e (+10) bit value */
878 {
879 unsigned int bits = 14 - e;
880 unsigned int half = (1 << (bits - 1)) - 1;
881 unsigned int even = (m >> bits) & 1;
882
883 /* if this overflows, we will end up with a normalised number */
884 m = (m + half + even) >> bits;
885 }
886
887 return s | m;
888 }
889
890 /* handle NaNs, preserve leftmost nan bits, but make sure we don't turn them into infinities */
891 m >>= 13;
892
893 return s | 0x7c00 | m | !m;
894}
895
632/*******************************************************************************/ 896/*******************************************************************************/
633/* floating point stuff, can be disabled by defining ECB_NO_LIBM */ 897/* floating point stuff, can be disabled by defining ECB_NO_LIBM */
634 898
635/* basically, everything uses "ieee pure-endian" floating point numbers */ 899/* basically, everything uses "ieee pure-endian" floating point numbers */
636/* the only noteworthy exception is ancient armle, which uses order 43218765 */ 900/* the only noteworthy exception is ancient armle, which uses order 43218765 */
637#if 0 \ 901#if 0 \
638 || __i386 || __i386__ \ 902 || __i386 || __i386__ \
639 || __amd64 || __amd64__ || __x86_64 || __x86_64__ \ 903 || ECB_GCC_AMD64 \
640 || __powerpc__ || __ppc__ || __powerpc64__ || __ppc64__ \ 904 || __powerpc__ || __ppc__ || __powerpc64__ || __ppc64__ \
641 || defined __s390__ || defined __s390x__ \ 905 || defined __s390__ || defined __s390x__ \
642 || defined __mips__ \ 906 || defined __mips__ \
643 || defined __alpha__ \ 907 || defined __alpha__ \
644 || defined __hppa__ \ 908 || defined __hppa__ \
645 || defined __ia64__ \ 909 || defined __ia64__ \
646 || defined __m68k__ \ 910 || defined __m68k__ \
647 || defined __m88k__ \ 911 || defined __m88k__ \
648 || defined __sh__ \ 912 || defined __sh__ \
649 || defined _M_IX86 || defined _M_AMD64 || defined _M_IA64 \ 913 || defined _M_IX86 || defined ECB_MSVC_AMD64 || defined _M_IA64 \
650 || (defined __arm__ && (defined __ARM_EABI__ || defined __EABI__ || defined __VFP_FP__ || defined _WIN32_WCE || defined __ANDROID__)) \ 914 || (defined __arm__ && (defined __ARM_EABI__ || defined __EABI__ || defined __VFP_FP__ || defined _WIN32_WCE || defined __ANDROID__)) \
651 || defined __aarch64__ 915 || defined __aarch64__
652 #define ECB_STDFP 1 916 #define ECB_STDFP 1
653 #include <string.h> /* for memcpy */ 917 #include <string.h> /* for memcpy */
654#else 918#else
… …
672 #define ECB_NAN ECB_INFINITY 936 #define ECB_NAN ECB_INFINITY
673 #endif 937 #endif
674 938
675 #if ECB_C99 || _XOPEN_VERSION >= 600 || _POSIX_VERSION >= 200112L 939 #if ECB_C99 || _XOPEN_VERSION >= 600 || _POSIX_VERSION >= 200112L
676 #define ecb_ldexpf(x,e) ldexpf ((x), (e)) 940 #define ecb_ldexpf(x,e) ldexpf ((x), (e))
941 #define ecb_frexpf(x,e) frexpf ((x), (e))
677 #else 942 #else
678 #define ecb_ldexpf(x,e) (float) ldexp ((x), (e)) 943 #define ecb_ldexpf(x,e) (float) ldexp ((double) (x), (e))
944 #define ecb_frexpf(x,e) (float) frexp ((double) (x), (e))
679 #endif 945 #endif
680
681 /* converts an ieee half/binary16 to a float */
682 ecb_function_ float ecb_binary16_to_float (uint16_t x) ecb_const;
683 ecb_function_ float
684 ecb_binary16_to_float (uint16_t x)
685 {
686 int e = (x >> 10) & 0x1f;
687 int m = x & 0x3ff;
688 float r;
689
690 if (!e ) r = ecb_ldexpf (m , -24);
691 else if (e != 31) r = ecb_ldexpf (m + 0x400, e - 25);
692 else if (m ) r = ECB_NAN;
693 else r = ECB_INFINITY;
694
695 return x & 0x8000 ? -r : r;
696 }
697 946
698 /* convert a float to ieee single/binary32 */ 947 /* convert a float to ieee single/binary32 */
699 ecb_function_ uint32_t ecb_float_to_binary32 (float x) ecb_const; 948 ecb_function_ ecb_const uint32_t ecb_float_to_binary32 (float x);
700 ecb_function_ uint32_t 949 ecb_function_ ecb_const uint32_t
701 ecb_float_to_binary32 (float x) 950 ecb_float_to_binary32 (float x)
702 { 951 {
703 uint32_t r; 952 uint32_t r;
704 953
705 #if ECB_STDFP 954 #if ECB_STDFP
… …
712 if (x == 0e0f ) return 0x00000000U; 961 if (x == 0e0f ) return 0x00000000U;
713 if (x > +3.40282346638528860e+38f) return 0x7f800000U; 962 if (x > +3.40282346638528860e+38f) return 0x7f800000U;
714 if (x < -3.40282346638528860e+38f) return 0xff800000U; 963 if (x < -3.40282346638528860e+38f) return 0xff800000U;
715 if (x != x ) return 0x7fbfffffU; 964 if (x != x ) return 0x7fbfffffU;
716 965
717 m = frexpf (x, &e) * 0x1000000U; 966 m = ecb_frexpf (x, &e) * 0x1000000U;
718 967
719 r = m & 0x80000000U; 968 r = m & 0x80000000U;
720 969
721 if (r) 970 if (r)
722 m = -m; 971 m = -m;
… …
734 983
735 return r; 984 return r;
736 } 985 }
737 986
738 /* converts an ieee single/binary32 to a float */ 987 /* converts an ieee single/binary32 to a float */
739 ecb_function_ float ecb_binary32_to_float (uint32_t x) ecb_const; 988 ecb_function_ ecb_const float ecb_binary32_to_float (uint32_t x);
740 ecb_function_ float 989 ecb_function_ ecb_const float
741 ecb_binary32_to_float (uint32_t x) 990 ecb_binary32_to_float (uint32_t x)
742 { 991 {
743 float r; 992 float r;
744 993
745 #if ECB_STDFP 994 #if ECB_STDFP
… …
764 1013
765 return r; 1014 return r;
766 } 1015 }
767 1016
768 /* convert a double to ieee double/binary64 */ 1017 /* convert a double to ieee double/binary64 */
769 ecb_function_ uint64_t ecb_double_to_binary64 (double x) ecb_const; 1018 ecb_function_ ecb_const uint64_t ecb_double_to_binary64 (double x);
770 ecb_function_ uint64_t 1019 ecb_function_ ecb_const uint64_t
771 ecb_double_to_binary64 (double x) 1020 ecb_double_to_binary64 (double x)
772 { 1021 {
773 uint64_t r; 1022 uint64_t r;
774 1023
775 #if ECB_STDFP 1024 #if ECB_STDFP
… …
804 1053
805 return r; 1054 return r;
806 } 1055 }
807 1056
808 /* converts an ieee double/binary64 to a double */ 1057 /* converts an ieee double/binary64 to a double */
809 ecb_function_ double ecb_binary64_to_double (uint64_t x) ecb_const; 1058 ecb_function_ ecb_const double ecb_binary64_to_double (uint64_t x);
810 ecb_function_ double 1059 ecb_function_ ecb_const double
811 ecb_binary64_to_double (uint64_t x) 1060 ecb_binary64_to_double (uint64_t x)
812 { 1061 {
813 double r; 1062 double r;
814 1063
815 #if ECB_STDFP 1064 #if ECB_STDFP
… …
833 #endif 1082 #endif
834 1083
835 return r; 1084 return r;
836 } 1085 }
837 1086
838#endif 1087 /* convert a float to ieee half/binary16 */
1088 ecb_function_ ecb_const uint16_t ecb_float_to_binary16 (float x);
1089 ecb_function_ ecb_const uint16_t
1090 ecb_float_to_binary16 (float x)
1091 {
1092 return ecb_binary32_to_binary16 (ecb_float_to_binary32 (x));
1093 }
839 1094
840#endif 1095 /* convert an ieee half/binary16 to float */
1096 ecb_function_ ecb_const float ecb_binary16_to_float (uint16_t x);
1097 ecb_function_ ecb_const float
1098 ecb_binary16_to_float (uint16_t x)
1099 {
1100 return ecb_binary32_to_float (ecb_binary16_to_binary32 (x));
1101 }
841 1102
1103#endif
1104
1105#endif
1106

Diff Legend

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