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

Comparing libecb/ecb.h (file contents):
Revision 1.166 by root, Sun Aug 9 00:10:21 2015 UTC vs.
Revision 1.185 by root, Mon Jan 20 18:11:07 2020 UTC

1/* 1/*
2 * libecb - http://software.schmorp.de/pkg/libecb 2 * libecb - http://software.schmorp.de/pkg/libecb
3 * 3 *
4 * Copyright (©) 2009-2015 Marc Alexander Lehmann <libecb@schmorp.de> 4 * Copyright (©) 2009-2015,2018-2020 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
47#include <string.h> /* for memcpy */
46 48
47#ifdef _WIN32 49#ifdef _WIN32
48 typedef signed char int8_t; 50 typedef signed char int8_t;
49 typedef unsigned char uint8_t; 51 typedef unsigned char uint8_t;
52 typedef signed char int_fast8_t;
53 typedef unsigned char uint_fast8_t;
50 typedef signed short int16_t; 54 typedef signed short int16_t;
51 typedef unsigned short uint16_t; 55 typedef unsigned short uint16_t;
56 typedef signed int int_fast16_t;
57 typedef unsigned int uint_fast16_t;
52 typedef signed int int32_t; 58 typedef signed int int32_t;
53 typedef unsigned int uint32_t; 59 typedef unsigned int uint32_t;
60 typedef signed int int_fast32_t;
61 typedef unsigned int uint_fast32_t;
54 #if __GNUC__ 62 #if __GNUC__
55 typedef signed long long int64_t; 63 typedef signed long long int64_t;
56 typedef unsigned long long uint64_t; 64 typedef unsigned long long uint64_t;
57 #else /* _MSC_VER || __BORLANDC__ */ 65 #else /* _MSC_VER || __BORLANDC__ */
58 typedef signed __int64 int64_t; 66 typedef signed __int64 int64_t;
59 typedef unsigned __int64 uint64_t; 67 typedef unsigned __int64 uint64_t;
60 #endif 68 #endif
69 typedef int64_t int_fast64_t;
70 typedef uint64_t uint_fast64_t;
61 #ifdef _WIN64 71 #ifdef _WIN64
62 #define ECB_PTRSIZE 8 72 #define ECB_PTRSIZE 8
63 typedef uint64_t uintptr_t; 73 typedef uint64_t uintptr_t;
64 typedef int64_t intptr_t; 74 typedef int64_t intptr_t;
65 #else 75 #else
… …
67 typedef uint32_t uintptr_t; 77 typedef uint32_t uintptr_t;
68 typedef int32_t intptr_t; 78 typedef int32_t intptr_t;
69 #endif 79 #endif
70#else 80#else
71 #include <inttypes.h> 81 #include <inttypes.h>
72 #if UINTMAX_MAX > 0xffffffffU 82 #if (defined INTPTR_MAX ? INTPTR_MAX : ULONG_MAX) > 0xffffffffU
73 #define ECB_PTRSIZE 8 83 #define ECB_PTRSIZE 8
74 #else 84 #else
75 #define ECB_PTRSIZE 4 85 #define ECB_PTRSIZE 4
76 #endif 86 #endif
77#endif 87#endif
78 88
79#define ECB_GCC_AMD64 (__amd64 || __amd64__ || __x86_64 || __x86_64__) 89#define ECB_GCC_AMD64 (__amd64 || __amd64__ || __x86_64 || __x86_64__)
80#define ECB_MSVC_AMD64 (_M_AMD64 || _M_X64) 90#define ECB_MSVC_AMD64 (_M_AMD64 || _M_X64)
91
92#ifndef ECB_OPTIMIZE_SIZE
93 #if __OPTIMIZE_SIZE__
94 #define ECB_OPTIMIZE_SIZE 1
95 #else
96 #define ECB_OPTIMIZE_SIZE 0
97 #endif
98#endif
81 99
82/* work around x32 idiocy by defining proper macros */ 100/* work around x32 idiocy by defining proper macros */
83#if ECB_GCC_AMD64 || ECB_MSVC_AMD64 101#if ECB_GCC_AMD64 || ECB_MSVC_AMD64
84 #if _ILP32 102 #if _ILP32
85 #define ECB_AMD64_X32 1 103 #define ECB_AMD64_X32 1
… …
115 #define ECB_CLANG_EXTENSION(x) 0 133 #define ECB_CLANG_EXTENSION(x) 0
116#endif 134#endif
117 135
118#define ECB_CPP (__cplusplus+0) 136#define ECB_CPP (__cplusplus+0)
119#define ECB_CPP11 (__cplusplus >= 201103L) 137#define ECB_CPP11 (__cplusplus >= 201103L)
138#define ECB_CPP14 (__cplusplus >= 201402L)
139#define ECB_CPP17 (__cplusplus >= 201703L)
120 140
121#if ECB_CPP 141#if ECB_CPP
122 #define ECB_C 0 142 #define ECB_C 0
123 #define ECB_STDC_VERSION 0 143 #define ECB_STDC_VERSION 0
124#else 144#else
… …
126 #define ECB_STDC_VERSION __STDC_VERSION__ 146 #define ECB_STDC_VERSION __STDC_VERSION__
127#endif 147#endif
128 148
129#define ECB_C99 (ECB_STDC_VERSION >= 199901L) 149#define ECB_C99 (ECB_STDC_VERSION >= 199901L)
130#define ECB_C11 (ECB_STDC_VERSION >= 201112L) 150#define ECB_C11 (ECB_STDC_VERSION >= 201112L)
151#define ECB_C17 (ECB_STDC_VERSION >= 201710L)
131 152
132#if ECB_CPP 153#if ECB_CPP
133 #define ECB_EXTERN_C extern "C" 154 #define ECB_EXTERN_C extern "C"
134 #define ECB_EXTERN_C_BEG ECB_EXTERN_C { 155 #define ECB_EXTERN_C_BEG ECB_EXTERN_C {
135 #define ECB_EXTERN_C_END } 156 #define ECB_EXTERN_C_END }
… …
155/* http://www-01.ibm.com/support/knowledgecenter/SSGH3R_13.1.0/com.ibm.xlcpp131.aix.doc/compiler_ref/compiler_builtins.html */ 176/* http://www-01.ibm.com/support/knowledgecenter/SSGH3R_13.1.0/com.ibm.xlcpp131.aix.doc/compiler_ref/compiler_builtins.html */
156#if __xlC__ && ECB_CPP 177#if __xlC__ && ECB_CPP
157 #include <builtins.h> 178 #include <builtins.h>
158#endif 179#endif
159 180
181#if 1400 <= _MSC_VER
182 #include <intrin.h> /* fence functions _ReadBarrier, also bit search functions _BitScanReverse */
183#endif
184
160#ifndef ECB_MEMORY_FENCE 185#ifndef ECB_MEMORY_FENCE
161 #if ECB_GCC_VERSION(2,5) || defined __INTEL_COMPILER || (__llvm__ && __GNUC__) || __SUNPRO_C >= 0x5110 || __SUNPRO_CC >= 0x5110 186 #if ECB_GCC_VERSION(2,5) || defined __INTEL_COMPILER || (__llvm__ && __GNUC__) || __SUNPRO_C >= 0x5110 || __SUNPRO_CC >= 0x5110
187 #define ECB_MEMORY_FENCE_RELAXED __asm__ __volatile__ ("" : : : "memory")
162 #if __i386 || __i386__ 188 #if __i386 || __i386__
163 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("lock; orb $0, -1(%%esp)" : : : "memory") 189 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("lock; orb $0, -1(%%esp)" : : : "memory")
164 #define ECB_MEMORY_FENCE_ACQUIRE __asm__ __volatile__ ("" : : : "memory") 190 #define ECB_MEMORY_FENCE_ACQUIRE __asm__ __volatile__ ("" : : : "memory")
165 #define ECB_MEMORY_FENCE_RELEASE __asm__ __volatile__ ("") 191 #define ECB_MEMORY_FENCE_RELEASE __asm__ __volatile__ ("" : : : "memory")
166 #elif ECB_GCC_AMD64 192 #elif ECB_GCC_AMD64
167 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("mfence" : : : "memory") 193 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("mfence" : : : "memory")
168 #define ECB_MEMORY_FENCE_ACQUIRE __asm__ __volatile__ ("" : : : "memory") 194 #define ECB_MEMORY_FENCE_ACQUIRE __asm__ __volatile__ ("" : : : "memory")
169 #define ECB_MEMORY_FENCE_RELEASE __asm__ __volatile__ ("") 195 #define ECB_MEMORY_FENCE_RELEASE __asm__ __volatile__ ("" : : : "memory")
170 #elif __powerpc__ || __ppc__ || __powerpc64__ || __ppc64__ 196 #elif __powerpc__ || __ppc__ || __powerpc64__ || __ppc64__
171 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("sync" : : : "memory") 197 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("sync" : : : "memory")
198 #elif defined __ARM_ARCH_2__ \
199 || defined __ARM_ARCH_3__ || defined __ARM_ARCH_3M__ \
200 || defined __ARM_ARCH_4__ || defined __ARM_ARCH_4T__ \
201 || defined __ARM_ARCH_5__ || defined __ARM_ARCH_5E__ \
202 || defined __ARM_ARCH_5T__ || defined __ARM_ARCH_5TE__ \
203 || defined __ARM_ARCH_5TEJ__
204 /* should not need any, unless running old code on newer cpu - arm doesn't support that */
172 #elif defined __ARM_ARCH_6__ || defined __ARM_ARCH_6J__ \ 205 #elif defined __ARM_ARCH_6__ || defined __ARM_ARCH_6J__ \
173 || defined __ARM_ARCH_6K__ || defined __ARM_ARCH_6ZK__ 206 || defined __ARM_ARCH_6K__ || defined __ARM_ARCH_6ZK__ \
207 || defined __ARM_ARCH_6T2__
174 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("mcr p15,0,%0,c7,c10,5" : : "r" (0) : "memory") 208 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("mcr p15,0,%0,c7,c10,5" : : "r" (0) : "memory")
175 #elif defined __ARM_ARCH_7__ || defined __ARM_ARCH_7A__ \ 209 #elif defined __ARM_ARCH_7__ || defined __ARM_ARCH_7A__ \
176 || defined __ARM_ARCH_7M__ || defined __ARM_ARCH_7R__ 210 || defined __ARM_ARCH_7R__ || defined __ARM_ARCH_7M__
177 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("dmb" : : : "memory") 211 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("dmb" : : : "memory")
178 #elif __aarch64__ 212 #elif __aarch64__
179 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("dmb ish" : : : "memory") 213 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("dmb ish" : : : "memory")
180 #elif (__sparc || __sparc__) && !(__sparc_v8__ || defined __sparcv8) 214 #elif (__sparc || __sparc__) && !(__sparc_v8__ || defined __sparcv8)
181 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("membar #LoadStore | #LoadLoad | #StoreStore | #StoreLoad" : : : "memory") 215 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("membar #LoadStore | #LoadLoad | #StoreStore | #StoreLoad" : : : "memory")
… …
208 #if ECB_GCC_VERSION(4,7) 242 #if ECB_GCC_VERSION(4,7)
209 /* see comment below (stdatomic.h) about the C11 memory model. */ 243 /* see comment below (stdatomic.h) about the C11 memory model. */
210 #define ECB_MEMORY_FENCE __atomic_thread_fence (__ATOMIC_SEQ_CST) 244 #define ECB_MEMORY_FENCE __atomic_thread_fence (__ATOMIC_SEQ_CST)
211 #define ECB_MEMORY_FENCE_ACQUIRE __atomic_thread_fence (__ATOMIC_ACQUIRE) 245 #define ECB_MEMORY_FENCE_ACQUIRE __atomic_thread_fence (__ATOMIC_ACQUIRE)
212 #define ECB_MEMORY_FENCE_RELEASE __atomic_thread_fence (__ATOMIC_RELEASE) 246 #define ECB_MEMORY_FENCE_RELEASE __atomic_thread_fence (__ATOMIC_RELEASE)
247 #define ECB_MEMORY_FENCE_RELAXED __atomic_thread_fence (__ATOMIC_RELAXED)
213 248
214 #elif ECB_CLANG_EXTENSION(c_atomic) 249 #elif ECB_CLANG_EXTENSION(c_atomic)
215 /* see comment below (stdatomic.h) about the C11 memory model. */ 250 /* see comment below (stdatomic.h) about the C11 memory model. */
216 #define ECB_MEMORY_FENCE __c11_atomic_thread_fence (__ATOMIC_SEQ_CST) 251 #define ECB_MEMORY_FENCE __c11_atomic_thread_fence (__ATOMIC_SEQ_CST)
217 #define ECB_MEMORY_FENCE_ACQUIRE __c11_atomic_thread_fence (__ATOMIC_ACQUIRE) 252 #define ECB_MEMORY_FENCE_ACQUIRE __c11_atomic_thread_fence (__ATOMIC_ACQUIRE)
218 #define ECB_MEMORY_FENCE_RELEASE __c11_atomic_thread_fence (__ATOMIC_RELEASE) 253 #define ECB_MEMORY_FENCE_RELEASE __c11_atomic_thread_fence (__ATOMIC_RELEASE)
254 #define ECB_MEMORY_FENCE_RELAXED __c11_atomic_thread_fence (__ATOMIC_RELAXED)
219 255
220 #elif ECB_GCC_VERSION(4,4) || defined __INTEL_COMPILER || defined __clang__ 256 #elif ECB_GCC_VERSION(4,4) || defined __INTEL_COMPILER || defined __clang__
221 #define ECB_MEMORY_FENCE __sync_synchronize () 257 #define ECB_MEMORY_FENCE __sync_synchronize ()
222 #elif _MSC_VER >= 1500 /* VC++ 2008 */ 258 #elif _MSC_VER >= 1500 /* VC++ 2008 */
223 /* apparently, microsoft broke all the memory barrier stuff in Visual Studio 2008... */ 259 /* apparently, microsoft broke all the memory barrier stuff in Visual Studio 2008... */
… …
233 #elif defined _WIN32 269 #elif defined _WIN32
234 #include <WinNT.h> 270 #include <WinNT.h>
235 #define ECB_MEMORY_FENCE MemoryBarrier () /* actually just xchg on x86... scary */ 271 #define ECB_MEMORY_FENCE MemoryBarrier () /* actually just xchg on x86... scary */
236 #elif __SUNPRO_C >= 0x5110 || __SUNPRO_CC >= 0x5110 272 #elif __SUNPRO_C >= 0x5110 || __SUNPRO_CC >= 0x5110
237 #include <mbarrier.h> 273 #include <mbarrier.h>
238 #define ECB_MEMORY_FENCE __machine_rw_barrier () 274 #define ECB_MEMORY_FENCE __machine_rw_barrier ()
239 #define ECB_MEMORY_FENCE_ACQUIRE __machine_r_barrier () 275 #define ECB_MEMORY_FENCE_ACQUIRE __machine_acq_barrier ()
240 #define ECB_MEMORY_FENCE_RELEASE __machine_w_barrier () 276 #define ECB_MEMORY_FENCE_RELEASE __machine_rel_barrier ()
277 #define ECB_MEMORY_FENCE_RELAXED __compiler_barrier ()
241 #elif __xlC__ 278 #elif __xlC__
242 #define ECB_MEMORY_FENCE __sync () 279 #define ECB_MEMORY_FENCE __sync ()
243 #endif 280 #endif
244#endif 281#endif
245 282
246#ifndef ECB_MEMORY_FENCE 283#ifndef ECB_MEMORY_FENCE
247 #if ECB_C11 && !defined __STDC_NO_ATOMICS__ 284 #if ECB_C11 && !defined __STDC_NO_ATOMICS__
248 /* we assume that these memory fences work on all variables/all memory accesses, */ 285 /* we assume that these memory fences work on all variables/all memory accesses, */
249 /* not just C11 atomics and atomic accesses */ 286 /* not just C11 atomics and atomic accesses */
250 #include <stdatomic.h> 287 #include <stdatomic.h>
251 /* Unfortunately, neither gcc 4.7 nor clang 3.1 generate any instructions for */
252 /* any fence other than seq_cst, which isn't very efficient for us. */
253 /* Why that is, we don't know - either the C11 memory model is quite useless */
254 /* for most usages, or gcc and clang have a bug */
255 /* I *currently* lean towards the latter, and inefficiently implement */
256 /* all three of ecb's fences as a seq_cst fence */
257 /* Update, gcc-4.8 generates mfence for all c++ fences, but nothing */
258 /* for all __atomic_thread_fence's except seq_cst */
259 #define ECB_MEMORY_FENCE atomic_thread_fence (memory_order_seq_cst) 288 #define ECB_MEMORY_FENCE atomic_thread_fence (memory_order_seq_cst)
289 #define ECB_MEMORY_FENCE_ACQUIRE atomic_thread_fence (memory_order_acquire)
290 #define ECB_MEMORY_FENCE_RELEASE atomic_thread_fence (memory_order_release)
260 #endif 291 #endif
261#endif 292#endif
262 293
263#ifndef ECB_MEMORY_FENCE 294#ifndef ECB_MEMORY_FENCE
264 #if !ECB_AVOID_PTHREADS 295 #if !ECB_AVOID_PTHREADS
… …
284 315
285#if !defined ECB_MEMORY_FENCE_RELEASE && defined ECB_MEMORY_FENCE 316#if !defined ECB_MEMORY_FENCE_RELEASE && defined ECB_MEMORY_FENCE
286 #define ECB_MEMORY_FENCE_RELEASE ECB_MEMORY_FENCE 317 #define ECB_MEMORY_FENCE_RELEASE ECB_MEMORY_FENCE
287#endif 318#endif
288 319
320#if !defined ECB_MEMORY_FENCE_RELAXED && defined ECB_MEMORY_FENCE
321 #define ECB_MEMORY_FENCE_RELAXED ECB_MEMORY_FENCE /* very heavy-handed */
322#endif
323
289/*****************************************************************************/ 324/*****************************************************************************/
290 325
291#if ECB_CPP 326#if ECB_CPP
292 #define ecb_inline static inline 327 #define ecb_inline static inline
293#elif ECB_GCC_VERSION(2,5) 328#elif ECB_GCC_VERSION(2,5)
… …
423#else 458#else
424 ecb_function_ ecb_const int ecb_ctz32 (uint32_t x); 459 ecb_function_ ecb_const int ecb_ctz32 (uint32_t x);
425 ecb_function_ ecb_const int 460 ecb_function_ ecb_const int
426 ecb_ctz32 (uint32_t x) 461 ecb_ctz32 (uint32_t x)
427 { 462 {
463#if 1400 <= _MSC_VER && (_M_IX86 || _M_X64 || _M_IA64 || _M_ARM)
464 unsigned long r;
465 _BitScanForward (&r, x);
466 return (int)r;
467#else
428 int r = 0; 468 int r = 0;
429 469
430 x &= ~x + 1; /* this isolates the lowest bit */ 470 x &= ~x + 1; /* this isolates the lowest bit */
431 471
432#if ECB_branchless_on_i386 472#if ECB_branchless_on_i386
… …
442 if (x & 0xff00ff00) r += 8; 482 if (x & 0xff00ff00) r += 8;
443 if (x & 0xffff0000) r += 16; 483 if (x & 0xffff0000) r += 16;
444#endif 484#endif
445 485
446 return r; 486 return r;
487#endif
447 } 488 }
448 489
449 ecb_function_ ecb_const int ecb_ctz64 (uint64_t x); 490 ecb_function_ ecb_const int ecb_ctz64 (uint64_t x);
450 ecb_function_ ecb_const int 491 ecb_function_ ecb_const int
451 ecb_ctz64 (uint64_t x) 492 ecb_ctz64 (uint64_t x)
452 { 493 {
494#if 1400 <= _MSC_VER && (_M_X64 || _M_IA64 || _M_ARM)
495 unsigned long r;
496 _BitScanForward64 (&r, x);
497 return (int)r;
498#else
453 int shift = x & 0xffffffffU ? 0 : 32; 499 int shift = x & 0xffffffff ? 0 : 32;
454 return ecb_ctz32 (x >> shift) + shift; 500 return ecb_ctz32 (x >> shift) + shift;
501#endif
455 } 502 }
456 503
457 ecb_function_ ecb_const int ecb_popcount32 (uint32_t x); 504 ecb_function_ ecb_const int ecb_popcount32 (uint32_t x);
458 ecb_function_ ecb_const int 505 ecb_function_ ecb_const int
459 ecb_popcount32 (uint32_t x) 506 ecb_popcount32 (uint32_t x)
… …
467 } 514 }
468 515
469 ecb_function_ ecb_const int ecb_ld32 (uint32_t x); 516 ecb_function_ ecb_const int ecb_ld32 (uint32_t x);
470 ecb_function_ ecb_const int ecb_ld32 (uint32_t x) 517 ecb_function_ ecb_const int ecb_ld32 (uint32_t x)
471 { 518 {
519#if 1400 <= _MSC_VER && (_M_IX86 || _M_X64 || _M_IA64 || _M_ARM)
520 unsigned long r;
521 _BitScanReverse (&r, x);
522 return (int)r;
523#else
472 int r = 0; 524 int r = 0;
473 525
474 if (x >> 16) { x >>= 16; r += 16; } 526 if (x >> 16) { x >>= 16; r += 16; }
475 if (x >> 8) { x >>= 8; r += 8; } 527 if (x >> 8) { x >>= 8; r += 8; }
476 if (x >> 4) { x >>= 4; r += 4; } 528 if (x >> 4) { x >>= 4; r += 4; }
477 if (x >> 2) { x >>= 2; r += 2; } 529 if (x >> 2) { x >>= 2; r += 2; }
478 if (x >> 1) { r += 1; } 530 if (x >> 1) { r += 1; }
479 531
480 return r; 532 return r;
533#endif
481 } 534 }
482 535
483 ecb_function_ ecb_const int ecb_ld64 (uint64_t x); 536 ecb_function_ ecb_const int ecb_ld64 (uint64_t x);
484 ecb_function_ ecb_const int ecb_ld64 (uint64_t x) 537 ecb_function_ ecb_const int ecb_ld64 (uint64_t x)
485 { 538 {
539#if 1400 <= _MSC_VER && (_M_X64 || _M_IA64 || _M_ARM)
540 unsigned long r;
541 _BitScanReverse64 (&r, x);
542 return (int)r;
543#else
486 int r = 0; 544 int r = 0;
487 545
488 if (x >> 32) { x >>= 32; r += 32; } 546 if (x >> 32) { x >>= 32; r += 32; }
489 547
490 return r + ecb_ld32 (x); 548 return r + ecb_ld32 (x);
549#endif
491 } 550 }
492#endif 551#endif
493 552
494ecb_function_ ecb_const ecb_bool ecb_is_pot32 (uint32_t x); 553ecb_function_ ecb_const ecb_bool ecb_is_pot32 (uint32_t x);
495ecb_function_ ecb_const ecb_bool ecb_is_pot32 (uint32_t x) { return !(x & (x - 1)); } 554ecb_function_ ecb_const ecb_bool ecb_is_pot32 (uint32_t x) { return !(x & (x - 1)); }
… …
551ecb_inline ecb_const uint32_t ecb_rotl32 (uint32_t x, unsigned int count) { return (x >> (32 - count)) | (x << count); } 610ecb_inline ecb_const uint32_t ecb_rotl32 (uint32_t x, unsigned int count) { return (x >> (32 - count)) | (x << count); }
552ecb_inline ecb_const uint32_t ecb_rotr32 (uint32_t x, unsigned int count) { return (x << (32 - count)) | (x >> count); } 611ecb_inline ecb_const uint32_t ecb_rotr32 (uint32_t x, unsigned int count) { return (x << (32 - count)) | (x >> count); }
553ecb_inline ecb_const uint64_t ecb_rotl64 (uint64_t x, unsigned int count) { return (x >> (64 - count)) | (x << count); } 612ecb_inline ecb_const uint64_t ecb_rotl64 (uint64_t x, unsigned int count) { return (x >> (64 - count)) | (x << count); }
554ecb_inline ecb_const uint64_t ecb_rotr64 (uint64_t x, unsigned int count) { return (x << (64 - count)) | (x >> count); } 613ecb_inline ecb_const uint64_t ecb_rotr64 (uint64_t x, unsigned int count) { return (x << (64 - count)) | (x >> count); }
555 614
615#if ECB_CPP
616
617inline uint8_t ecb_ctz (uint8_t v) { return ecb_ctz32 (v); }
618inline uint16_t ecb_ctz (uint16_t v) { return ecb_ctz32 (v); }
619inline uint32_t ecb_ctz (uint32_t v) { return ecb_ctz32 (v); }
620inline uint64_t ecb_ctz (uint64_t v) { return ecb_ctz64 (v); }
621
622inline bool ecb_is_pot (uint8_t v) { return ecb_is_pot32 (v); }
623inline bool ecb_is_pot (uint16_t v) { return ecb_is_pot32 (v); }
624inline bool ecb_is_pot (uint32_t v) { return ecb_is_pot32 (v); }
625inline bool ecb_is_pot (uint64_t v) { return ecb_is_pot64 (v); }
626
627inline int ecb_ld (uint8_t v) { return ecb_ld32 (v); }
628inline int ecb_ld (uint16_t v) { return ecb_ld32 (v); }
629inline int ecb_ld (uint32_t v) { return ecb_ld32 (v); }
630inline int ecb_ld (uint64_t v) { return ecb_ld64 (v); }
631
632inline int ecb_popcount (uint8_t v) { return ecb_popcount32 (v); }
633inline int ecb_popcount (uint16_t v) { return ecb_popcount32 (v); }
634inline int ecb_popcount (uint32_t v) { return ecb_popcount32 (v); }
635inline int ecb_popcount (uint64_t v) { return ecb_popcount64 (v); }
636
637inline uint8_t ecb_bitrev (uint8_t v) { return ecb_bitrev8 (v); }
638inline uint16_t ecb_bitrev (uint16_t v) { return ecb_bitrev16 (v); }
639inline uint32_t ecb_bitrev (uint32_t v) { return ecb_bitrev32 (v); }
640
641inline uint8_t ecb_rotl (uint8_t v, unsigned int count) { return ecb_rotl8 (v, count); }
642inline uint16_t ecb_rotl (uint16_t v, unsigned int count) { return ecb_rotl16 (v, count); }
643inline uint32_t ecb_rotl (uint32_t v, unsigned int count) { return ecb_rotl32 (v, count); }
644inline uint64_t ecb_rotl (uint64_t v, unsigned int count) { return ecb_rotl64 (v, count); }
645
646inline uint8_t ecb_rotr (uint8_t v, unsigned int count) { return ecb_rotr8 (v, count); }
647inline uint16_t ecb_rotr (uint16_t v, unsigned int count) { return ecb_rotr16 (v, count); }
648inline uint32_t ecb_rotr (uint32_t v, unsigned int count) { return ecb_rotr32 (v, count); }
649inline uint64_t ecb_rotr (uint64_t v, unsigned int count) { return ecb_rotr64 (v, count); }
650
651#endif
652
556#if ECB_GCC_VERSION(4,3) || (ECB_CLANG_BUILTIN(__builtin_bswap32) && ECB_CLANG_BUILTIN(__builtin_bswap64)) 653#if ECB_GCC_VERSION(4,3) || (ECB_CLANG_BUILTIN(__builtin_bswap32) && ECB_CLANG_BUILTIN(__builtin_bswap64))
557 #if ECB_GCC_VERSION(4,8) || ECB_CLANG_BUILTIN(__builtin_bswap16) 654 #if ECB_GCC_VERSION(4,8) || ECB_CLANG_BUILTIN(__builtin_bswap16)
558 #define ecb_bswap16(x) __builtin_bswap16 (x) 655 #define ecb_bswap16(x) __builtin_bswap16 (x)
559 #else 656 #else
560 #define ecb_bswap16(x) (__builtin_bswap32 (x) >> 16) 657 #define ecb_bswap16(x) (__builtin_bswap32 (x) >> 16)
… …
598#endif 695#endif
599 696
600/* try to tell the compiler that some condition is definitely true */ 697/* try to tell the compiler that some condition is definitely true */
601#define ecb_assume(cond) if (!(cond)) ecb_unreachable (); else 0 698#define ecb_assume(cond) if (!(cond)) ecb_unreachable (); else 0
602 699
603ecb_inline ecb_const unsigned char ecb_byteorder_helper (void); 700ecb_inline ecb_const uint32_t ecb_byteorder_helper (void);
604ecb_inline ecb_const unsigned char 701ecb_inline ecb_const uint32_t
605ecb_byteorder_helper (void) 702ecb_byteorder_helper (void)
606{ 703{
607 /* the union code still generates code under pressure in gcc, */ 704 /* the union code still generates code under pressure in gcc, */
608 /* but less than using pointers, and always seems to */ 705 /* but less than using pointers, and always seems to */
609 /* successfully return a constant. */ 706 /* successfully return a constant. */
610 /* the reason why we have this horrible preprocessor mess */ 707 /* the reason why we have this horrible preprocessor mess */
611 /* is to avoid it in all cases, at least on common architectures */ 708 /* is to avoid it in all cases, at least on common architectures */
612 /* or when using a recent enough gcc version (>= 4.6) */ 709 /* or when using a recent enough gcc version (>= 4.6) */
613#if ((__i386 || __i386__) && !__VOS__) || _M_IX86 || ECB_GCC_AMD64 || ECB_MSVC_AMD64
614 return 0x44;
615#elif __BYTE_ORDER__ && __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__ 710#if (defined __BYTE_ORDER__ && __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__) \
711 || ((__i386 || __i386__ || _M_IX86 || ECB_GCC_AMD64 || ECB_MSVC_AMD64) && !__VOS__)
712 #define ECB_LITTLE_ENDIAN 1
616 return 0x44; 713 return 0x44332211;
617#elif __BYTE_ORDER__ && __BYTE_ORDER__ == __ORDER_BIG_ENDIAN__ 714#elif (defined __BYTE_ORDER__ && __BYTE_ORDER__ == __ORDER_BIG_ENDIAN__) \
715 || ((__AARCH64EB__ || __MIPSEB__ || __ARMEB__) && !__VOS__)
716 #define ECB_BIG_ENDIAN 1
618 return 0x11; 717 return 0x11223344;
619#else 718#else
620 union 719 union
621 { 720 {
721 uint8_t c[4];
622 uint32_t i; 722 uint32_t u;
623 uint8_t c;
624 } u = { 0x11223344 }; 723 } u = { 0x11, 0x22, 0x33, 0x44 };
625 return u.c; 724 return u.u;
626#endif 725#endif
627} 726}
628 727
629ecb_inline ecb_const ecb_bool ecb_big_endian (void); 728ecb_inline ecb_const ecb_bool ecb_big_endian (void);
630ecb_inline ecb_const ecb_bool ecb_big_endian (void) { return ecb_byteorder_helper () == 0x11; } 729ecb_inline ecb_const ecb_bool ecb_big_endian (void) { return ecb_byteorder_helper () == 0x11223344; }
631ecb_inline ecb_const ecb_bool ecb_little_endian (void); 730ecb_inline ecb_const ecb_bool ecb_little_endian (void);
632ecb_inline ecb_const ecb_bool ecb_little_endian (void) { return ecb_byteorder_helper () == 0x44; } 731ecb_inline ecb_const ecb_bool ecb_little_endian (void) { return ecb_byteorder_helper () == 0x44332211; }
732
733/*****************************************************************************/
734/* unaligned load/store */
735
736ecb_inline uint_fast16_t ecb_be_u16_to_host (uint_fast16_t v) { return ecb_little_endian () ? ecb_bswap16 (v) : v; }
737ecb_inline uint_fast32_t ecb_be_u32_to_host (uint_fast32_t v) { return ecb_little_endian () ? ecb_bswap32 (v) : v; }
738ecb_inline uint_fast64_t ecb_be_u64_to_host (uint_fast64_t v) { return ecb_little_endian () ? ecb_bswap64 (v) : v; }
739
740ecb_inline uint_fast16_t ecb_le_u16_to_host (uint_fast16_t v) { return ecb_big_endian () ? ecb_bswap16 (v) : v; }
741ecb_inline uint_fast32_t ecb_le_u32_to_host (uint_fast32_t v) { return ecb_big_endian () ? ecb_bswap32 (v) : v; }
742ecb_inline uint_fast64_t ecb_le_u64_to_host (uint_fast64_t v) { return ecb_big_endian () ? ecb_bswap64 (v) : v; }
743
744ecb_inline uint_fast16_t ecb_peek_u16_u (const void *ptr) { uint16_t v; memcpy (&v, ptr, sizeof (v)); return v; }
745ecb_inline uint_fast32_t ecb_peek_u32_u (const void *ptr) { uint32_t v; memcpy (&v, ptr, sizeof (v)); return v; }
746ecb_inline uint_fast64_t ecb_peek_u64_u (const void *ptr) { uint64_t v; memcpy (&v, ptr, sizeof (v)); return v; }
747
748ecb_inline uint_fast16_t ecb_peek_be_u16_u (const void *ptr) { return ecb_be_u16_to_host (ecb_peek_u16_u (ptr)); }
749ecb_inline uint_fast32_t ecb_peek_be_u32_u (const void *ptr) { return ecb_be_u32_to_host (ecb_peek_u32_u (ptr)); }
750ecb_inline uint_fast64_t ecb_peek_be_u64_u (const void *ptr) { return ecb_be_u64_to_host (ecb_peek_u64_u (ptr)); }
751
752ecb_inline uint_fast16_t ecb_peek_le_u16_u (const void *ptr) { return ecb_le_u16_to_host (ecb_peek_u16_u (ptr)); }
753ecb_inline uint_fast32_t ecb_peek_le_u32_u (const void *ptr) { return ecb_le_u32_to_host (ecb_peek_u32_u (ptr)); }
754ecb_inline uint_fast64_t ecb_peek_le_u64_u (const void *ptr) { return ecb_le_u64_to_host (ecb_peek_u64_u (ptr)); }
755
756ecb_inline uint_fast16_t ecb_host_to_be_u16 (uint_fast16_t v) { return ecb_little_endian () ? ecb_bswap16 (v) : v; }
757ecb_inline uint_fast32_t ecb_host_to_be_u32 (uint_fast32_t v) { return ecb_little_endian () ? ecb_bswap32 (v) : v; }
758ecb_inline uint_fast64_t ecb_host_to_be_u64 (uint_fast64_t v) { return ecb_little_endian () ? ecb_bswap64 (v) : v; }
759
760ecb_inline uint_fast16_t ecb_host_to_le_u16 (uint_fast16_t v) { return ecb_big_endian () ? ecb_bswap16 (v) : v; }
761ecb_inline uint_fast32_t ecb_host_to_le_u32 (uint_fast32_t v) { return ecb_big_endian () ? ecb_bswap32 (v) : v; }
762ecb_inline uint_fast64_t ecb_host_to_le_u64 (uint_fast64_t v) { return ecb_big_endian () ? ecb_bswap64 (v) : v; }
763
764ecb_inline void ecb_poke_u16_u (void *ptr, uint16_t v) { memcpy (ptr, &v, sizeof (v)); }
765ecb_inline void ecb_poke_u32_u (void *ptr, uint32_t v) { memcpy (ptr, &v, sizeof (v)); }
766ecb_inline void ecb_poke_u64_u (void *ptr, uint64_t v) { memcpy (ptr, &v, sizeof (v)); }
767
768ecb_inline void ecb_poke_be_u16_u (void *ptr, uint_fast16_t v) { ecb_poke_u16_u (ptr, ecb_host_to_be_u16 (v)); }
769ecb_inline void ecb_poke_be_u32_u (void *ptr, uint_fast32_t v) { ecb_poke_u32_u (ptr, ecb_host_to_be_u32 (v)); }
770ecb_inline void ecb_poke_be_u64_u (void *ptr, uint_fast64_t v) { ecb_poke_u64_u (ptr, ecb_host_to_be_u64 (v)); }
771
772ecb_inline void ecb_poke_le_u16_u (void *ptr, uint_fast16_t v) { ecb_poke_u16_u (ptr, ecb_host_to_le_u16 (v)); }
773ecb_inline void ecb_poke_le_u32_u (void *ptr, uint_fast32_t v) { ecb_poke_u32_u (ptr, ecb_host_to_le_u32 (v)); }
774ecb_inline void ecb_poke_le_u64_u (void *ptr, uint_fast64_t v) { ecb_poke_u64_u (ptr, ecb_host_to_le_u64 (v)); }
775
776#ifdef ECB_CPP
777
778inline uint8_t ecb_bswap (uint8_t v) { return v; }
779inline uint16_t ecb_bswap (uint16_t v) { return ecb_bswap16 (v); }
780inline uint32_t ecb_bswap (uint32_t v) { return ecb_bswap32 (v); }
781inline uint64_t ecb_bswap (uint64_t v) { return ecb_bswap64 (v); }
782
783template<typename T> inline T ecb_be_to_host (T v) { return ecb_little_endian () ? ecb_bswap (v) : v; }
784template<typename T> inline T ecb_le_to_host (T v) { return ecb_big_endian () ? ecb_bswap (v) : v; }
785template<typename T> inline T ecb_peek (const void *ptr) { return *(const T *)ptr; }
786template<typename T> inline T ecb_peek_be (const void *ptr) { return ecb_be_to_host (ecb_peek <T> (ptr)); }
787template<typename T> inline T ecb_peek_le (const void *ptr) { return ecb_le_to_host (ecb_peek <T> (ptr)); }
788template<typename T> inline T ecb_peek_u (const void *ptr) { T v; memcpy (&v, ptr, sizeof (v)); return v; }
789template<typename T> inline T ecb_peek_be_u (const void *ptr) { return ecb_be_to_host (ecb_peek_u<T> (ptr)); }
790template<typename T> inline T ecb_peek_le_u (const void *ptr) { return ecb_le_to_host (ecb_peek_u<T> (ptr)); }
791
792template<typename T> inline T ecb_host_to_be (T v) { return ecb_little_endian () ? ecb_bswap (v) : v; }
793template<typename T> inline T ecb_host_to_le (T v) { return ecb_big_endian () ? ecb_bswap (v) : v; }
794template<typename T> inline void ecb_poke (void *ptr, T v) { *(T *)ptr = v; }
795template<typename T> inline void ecb_poke_be (void *ptr, T v) { return ecb_poke <T> (ptr, ecb_host_to_be (v)); }
796template<typename T> inline void ecb_poke_le (void *ptr, T v) { return ecb_poke <T> (ptr, ecb_host_to_le (v)); }
797template<typename T> inline void ecb_poke_u (void *ptr, T v) { memcpy (ptr, &v, sizeof (v)); }
798template<typename T> inline void ecb_poke_be_u (void *ptr, T v) { return ecb_poke_u<T> (ptr, ecb_host_to_be (v)); }
799template<typename T> inline void ecb_poke_le_u (void *ptr, T v) { return ecb_poke_u<T> (ptr, ecb_host_to_le (v)); }
800
801#endif
802
803/*****************************************************************************/
633 804
634#if ECB_GCC_VERSION(3,0) || ECB_C99 805#if ECB_GCC_VERSION(3,0) || ECB_C99
635 #define ecb_mod(m,n) ((m) % (n) + ((m) % (n) < 0 ? (n) : 0)) 806 #define ecb_mod(m,n) ((m) % (n) + ((m) % (n) < 0 ? (n) : 0))
636#else 807#else
637 #define ecb_mod(m,n) ((m) < 0 ? ((n) - 1 - ((-1 - (m)) % (n))) : ((m) % (n))) 808 #define ecb_mod(m,n) ((m) < 0 ? ((n) - 1 - ((-1 - (m)) % (n))) : ((m) % (n)))
… …
661 return N; 832 return N;
662 } 833 }
663#else 834#else
664 #define ecb_array_length(name) (sizeof (name) / sizeof (name [0])) 835 #define ecb_array_length(name) (sizeof (name) / sizeof (name [0]))
665#endif 836#endif
837
838/*****************************************************************************/
839
840ecb_function_ ecb_const uint32_t ecb_binary16_to_binary32 (uint32_t x);
841ecb_function_ ecb_const uint32_t
842ecb_binary16_to_binary32 (uint32_t x)
843{
844 unsigned int s = (x & 0x8000) << (31 - 15);
845 int e = (x >> 10) & 0x001f;
846 unsigned int m = x & 0x03ff;
847
848 if (ecb_expect_false (e == 31))
849 /* infinity or NaN */
850 e = 255 - (127 - 15);
851 else if (ecb_expect_false (!e))
852 {
853 if (ecb_expect_true (!m))
854 /* zero, handled by code below by forcing e to 0 */
855 e = 0 - (127 - 15);
856 else
857 {
858 /* subnormal, renormalise */
859 unsigned int s = 10 - ecb_ld32 (m);
860
861 m = (m << s) & 0x3ff; /* mask implicit bit */
862 e -= s - 1;
863 }
864 }
865
866 /* e and m now are normalised, or zero, (or inf or nan) */
867 e += 127 - 15;
868
869 return s | (e << 23) | (m << (23 - 10));
870}
871
872ecb_function_ ecb_const uint16_t ecb_binary32_to_binary16 (uint32_t x);
873ecb_function_ ecb_const uint16_t
874ecb_binary32_to_binary16 (uint32_t x)
875{
876 unsigned int s = (x >> 16) & 0x00008000; /* sign bit, the easy part */
877 unsigned int e = ((x >> 23) & 0x000000ff) - (127 - 15); /* the desired exponent */
878 unsigned int m = x & 0x007fffff;
879
880 x &= 0x7fffffff;
881
882 /* if it's within range of binary16 normals, use fast path */
883 if (ecb_expect_true (0x38800000 <= x && x <= 0x477fefff))
884 {
885 /* mantissa round-to-even */
886 m += 0x00000fff + ((m >> (23 - 10)) & 1);
887
888 /* handle overflow */
889 if (ecb_expect_false (m >= 0x00800000))
890 {
891 m >>= 1;
892 e += 1;
893 }
894
895 return s | (e << 10) | (m >> (23 - 10));
896 }
897
898 /* handle large numbers and infinity */
899 if (ecb_expect_true (0x477fefff < x && x <= 0x7f800000))
900 return s | 0x7c00;
901
902 /* handle zero, subnormals and small numbers */
903 if (ecb_expect_true (x < 0x38800000))
904 {
905 /* zero */
906 if (ecb_expect_true (!x))
907 return s;
908
909 /* handle subnormals */
910
911 /* too small, will be zero */
912 if (e < (14 - 24)) /* might not be sharp, but is good enough */
913 return s;
914
915 m |= 0x00800000; /* make implicit bit explicit */
916
917 /* very tricky - we need to round to the nearest e (+10) bit value */
918 {
919 unsigned int bits = 14 - e;
920 unsigned int half = (1 << (bits - 1)) - 1;
921 unsigned int even = (m >> bits) & 1;
922
923 /* if this overflows, we will end up with a normalised number */
924 m = (m + half + even) >> bits;
925 }
926
927 return s | m;
928 }
929
930 /* handle NaNs, preserve leftmost nan bits, but make sure we don't turn them into infinities */
931 m >>= 13;
932
933 return s | 0x7c00 | m | !m;
934}
666 935
667/*******************************************************************************/ 936/*******************************************************************************/
668/* floating point stuff, can be disabled by defining ECB_NO_LIBM */ 937/* floating point stuff, can be disabled by defining ECB_NO_LIBM */
669 938
670/* basically, everything uses "ieee pure-endian" floating point numbers */ 939/* basically, everything uses "ieee pure-endian" floating point numbers */
… …
683 || defined __sh__ \ 952 || defined __sh__ \
684 || defined _M_IX86 || defined ECB_MSVC_AMD64 || defined _M_IA64 \ 953 || defined _M_IX86 || defined ECB_MSVC_AMD64 || defined _M_IA64 \
685 || (defined __arm__ && (defined __ARM_EABI__ || defined __EABI__ || defined __VFP_FP__ || defined _WIN32_WCE || defined __ANDROID__)) \ 954 || (defined __arm__ && (defined __ARM_EABI__ || defined __EABI__ || defined __VFP_FP__ || defined _WIN32_WCE || defined __ANDROID__)) \
686 || defined __aarch64__ 955 || defined __aarch64__
687 #define ECB_STDFP 1 956 #define ECB_STDFP 1
688 #include <string.h> /* for memcpy */
689#else 957#else
690 #define ECB_STDFP 0 958 #define ECB_STDFP 0
691#endif 959#endif
692 960
693#ifndef ECB_NO_LIBM 961#ifndef ECB_NO_LIBM
… …
713 #else 981 #else
714 #define ecb_ldexpf(x,e) (float) ldexp ((double) (x), (e)) 982 #define ecb_ldexpf(x,e) (float) ldexp ((double) (x), (e))
715 #define ecb_frexpf(x,e) (float) frexp ((double) (x), (e)) 983 #define ecb_frexpf(x,e) (float) frexp ((double) (x), (e))
716 #endif 984 #endif
717 985
718 /* converts an ieee half/binary16 to a float */
719 ecb_function_ ecb_const float ecb_binary16_to_float (uint16_t x);
720 ecb_function_ ecb_const float
721 ecb_binary16_to_float (uint16_t x)
722 {
723 int e = (x >> 10) & 0x1f;
724 int m = x & 0x3ff;
725 float r;
726
727 if (!e ) r = ecb_ldexpf (m , -24);
728 else if (e != 31) r = ecb_ldexpf (m + 0x400, e - 25);
729 else if (m ) r = ECB_NAN;
730 else r = ECB_INFINITY;
731
732 return x & 0x8000 ? -r : r;
733 }
734
735 /* convert a float to ieee single/binary32 */ 986 /* convert a float to ieee single/binary32 */
736 ecb_function_ ecb_const uint32_t ecb_float_to_binary32 (float x); 987 ecb_function_ ecb_const uint32_t ecb_float_to_binary32 (float x);
737 ecb_function_ ecb_const uint32_t 988 ecb_function_ ecb_const uint32_t
738 ecb_float_to_binary32 (float x) 989 ecb_float_to_binary32 (float x)
739 { 990 {
… …
870 #endif 1121 #endif
871 1122
872 return r; 1123 return r;
873 } 1124 }
874 1125
875#endif 1126 /* convert a float to ieee half/binary16 */
1127 ecb_function_ ecb_const uint16_t ecb_float_to_binary16 (float x);
1128 ecb_function_ ecb_const uint16_t
1129 ecb_float_to_binary16 (float x)
1130 {
1131 return ecb_binary32_to_binary16 (ecb_float_to_binary32 (x));
1132 }
876 1133
877#endif 1134 /* convert an ieee half/binary16 to float */
1135 ecb_function_ ecb_const float ecb_binary16_to_float (uint16_t x);
1136 ecb_function_ ecb_const float
1137 ecb_binary16_to_float (uint16_t x)
1138 {
1139 return ecb_binary32_to_float (ecb_binary16_to_binary32 (x));
1140 }
878 1141
1142#endif
1143
1144#endif
1145

Diff Legend

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