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

Comparing libecb/ecb.h (file contents):
Revision 1.143 by root, Fri Oct 17 10:56:48 2014 UTC vs.
Revision 1.171 by root, Tue Nov 24 19:46:20 2015 UTC

1/* 1/*
2 * libecb - http://software.schmorp.de/pkg/libecb 2 * libecb - http://software.schmorp.de/pkg/libecb
3 * 3 *
4 * Copyright (©) 2009-2014 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 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;
… …
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
… …
98 #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)))
99#endif 102#endif
100 103
101#define ECB_CLANG_VERSION(major,minor) (__clang_major__ > (major) || (__clang_major__ == (major) && __clang_minor__ >= (minor))) 104#define ECB_CLANG_VERSION(major,minor) (__clang_major__ > (major) || (__clang_major__ == (major) && __clang_minor__ >= (minor)))
102 105
103#if __clang__ && defined(__has_builtin) 106#if __clang__ && defined __has_builtin
104 #define ECB_CLANG_BUILTIN(x) __has_builtin(x) 107 #define ECB_CLANG_BUILTIN(x) __has_builtin (x)
105#else 108#else
106 #define ECB_CLANG_BUILTIN(x) 0 109 #define ECB_CLANG_BUILTIN(x) 0
107#endif 110#endif
108 111
109#if __clang__ && defined(__has_extension) 112#if __clang__ && defined __has_extension
110 #define ECB_CLANG_EXTENSION(x) __has_extension(x) 113 #define ECB_CLANG_EXTENSION(x) __has_extension (x)
111#else 114#else
112 #define ECB_CLANG_EXTENSION(x) 0 115 #define ECB_CLANG_EXTENSION(x) 0
113#endif 116#endif
114 117
115#define ECB_CPP (__cplusplus+0) 118#define ECB_CPP (__cplusplus+0)
… …
145 #define ECB_NO_SMP 1 148 #define ECB_NO_SMP 1
146#endif 149#endif
147 150
148#if ECB_NO_SMP 151#if ECB_NO_SMP
149 #define ECB_MEMORY_FENCE do { } while (0) 152 #define ECB_MEMORY_FENCE do { } while (0)
153#endif
154
155/* http://www-01.ibm.com/support/knowledgecenter/SSGH3R_13.1.0/com.ibm.xlcpp131.aix.doc/compiler_ref/compiler_builtins.html */
156#if __xlC__ && ECB_CPP
157 #include <builtins.h>
158#endif
159
160#if 1400 <= _MSC_VER
161 #include <intrin.h> /* fence functions _ReadBarrier, also bit search functions _BitScanReverse */
150#endif 162#endif
151 163
152#ifndef ECB_MEMORY_FENCE 164#ifndef ECB_MEMORY_FENCE
153 #if ECB_GCC_VERSION(2,5) || defined __INTEL_COMPILER || (__llvm__ && __GNUC__) || __SUNPRO_C >= 0x5110 || __SUNPRO_CC >= 0x5110 165 #if ECB_GCC_VERSION(2,5) || defined __INTEL_COMPILER || (__llvm__ && __GNUC__) || __SUNPRO_C >= 0x5110 || __SUNPRO_CC >= 0x5110
154 #if __i386 || __i386__ 166 #if __i386 || __i386__
155 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("lock; orb $0, -1(%%esp)" : : : "memory") 167 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("lock; orb $0, -1(%%esp)" : : : "memory")
156 #define ECB_MEMORY_FENCE_ACQUIRE __asm__ __volatile__ ("" : : : "memory") 168 #define ECB_MEMORY_FENCE_ACQUIRE __asm__ __volatile__ ("" : : : "memory")
157 #define ECB_MEMORY_FENCE_RELEASE __asm__ __volatile__ ("") 169 #define ECB_MEMORY_FENCE_RELEASE __asm__ __volatile__ ("")
158 #elif __amd64 || __amd64__ || __x86_64 || __x86_64__ 170 #elif ECB_GCC_AMD64
159 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("mfence" : : : "memory") 171 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("mfence" : : : "memory")
160 #define ECB_MEMORY_FENCE_ACQUIRE __asm__ __volatile__ ("" : : : "memory") 172 #define ECB_MEMORY_FENCE_ACQUIRE __asm__ __volatile__ ("" : : : "memory")
161 #define ECB_MEMORY_FENCE_RELEASE __asm__ __volatile__ ("") 173 #define ECB_MEMORY_FENCE_RELEASE __asm__ __volatile__ ("")
162 #elif __powerpc__ || __ppc__ || __powerpc64__ || __ppc64__ 174 #elif __powerpc__ || __ppc__ || __powerpc64__ || __ppc64__
163 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("sync" : : : "memory") 175 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("sync" : : : "memory")
… …
167 #elif defined __ARM_ARCH_7__ || defined __ARM_ARCH_7A__ \ 179 #elif defined __ARM_ARCH_7__ || defined __ARM_ARCH_7A__ \
168 || defined __ARM_ARCH_7M__ || defined __ARM_ARCH_7R__ 180 || defined __ARM_ARCH_7M__ || defined __ARM_ARCH_7R__
169 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("dmb" : : : "memory") 181 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("dmb" : : : "memory")
170 #elif __aarch64__ 182 #elif __aarch64__
171 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("dmb ish" : : : "memory") 183 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("dmb ish" : : : "memory")
172 #elif (__sparc || __sparc__) && !__sparcv8 184 #elif (__sparc || __sparc__) && !(__sparc_v8__ || defined __sparcv8)
173 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("membar #LoadStore | #LoadLoad | #StoreStore | #StoreLoad" : : : "memory") 185 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("membar #LoadStore | #LoadLoad | #StoreStore | #StoreLoad" : : : "memory")
174 #define ECB_MEMORY_FENCE_ACQUIRE __asm__ __volatile__ ("membar #LoadStore | #LoadLoad" : : : "memory") 186 #define ECB_MEMORY_FENCE_ACQUIRE __asm__ __volatile__ ("membar #LoadStore | #LoadLoad" : : : "memory")
175 #define ECB_MEMORY_FENCE_RELEASE __asm__ __volatile__ ("membar #LoadStore | #StoreStore") 187 #define ECB_MEMORY_FENCE_RELEASE __asm__ __volatile__ ("membar #LoadStore | #StoreStore")
176 #elif defined __s390__ || defined __s390x__ 188 #elif defined __s390__ || defined __s390x__
177 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("bcr 15,0" : : : "memory") 189 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("bcr 15,0" : : : "memory")
… …
278 #define ECB_MEMORY_FENCE_RELEASE ECB_MEMORY_FENCE 290 #define ECB_MEMORY_FENCE_RELEASE ECB_MEMORY_FENCE
279#endif 291#endif
280 292
281/*****************************************************************************/ 293/*****************************************************************************/
282 294
283#if __cplusplus 295#if ECB_CPP
284 #define ecb_inline static inline 296 #define ecb_inline static inline
285#elif ECB_GCC_VERSION(2,5) 297#elif ECB_GCC_VERSION(2,5)
286 #define ecb_inline static __inline__ 298 #define ecb_inline static __inline__
287#elif ECB_C99 299#elif ECB_C99
288 #define ecb_inline static inline 300 #define ecb_inline static inline
… …
302 314
303#define ECB_CONCAT_(a, b) a ## b 315#define ECB_CONCAT_(a, b) a ## b
304#define ECB_CONCAT(a, b) ECB_CONCAT_(a, b) 316#define ECB_CONCAT(a, b) ECB_CONCAT_(a, b)
305#define ECB_STRINGIFY_(a) # a 317#define ECB_STRINGIFY_(a) # a
306#define ECB_STRINGIFY(a) ECB_STRINGIFY_(a) 318#define ECB_STRINGIFY(a) ECB_STRINGIFY_(a)
319#define ECB_STRINGIFY_EXPR(expr) ((expr), ECB_STRINGIFY_ (expr))
307 320
308#define ecb_function_ ecb_inline 321#define ecb_function_ ecb_inline
309 322
310#if ECB_GCC_VERSION(3,1) || ECB_CLANG_VERSION(2,8) 323#if ECB_GCC_VERSION(3,1) || ECB_CLANG_VERSION(2,8)
311 #define ecb_attribute(attrlist) __attribute__ (attrlist) 324 #define ecb_attribute(attrlist) __attribute__ (attrlist)
… …
335 #define ecb_prefetch(addr,rw,locality) 348 #define ecb_prefetch(addr,rw,locality)
336#endif 349#endif
337 350
338/* no emulation for ecb_decltype */ 351/* no emulation for ecb_decltype */
339#if ECB_CPP11 352#if ECB_CPP11
353 // older implementations might have problems with decltype(x)::type, work around it
354 template<class T> struct ecb_decltype_t { typedef T type; };
340 #define ecb_decltype(x) decltype (x) 355 #define ecb_decltype(x) ecb_decltype_t<decltype (x)>::type
341#elif ECB_GCC_VERSION(3,0) || ECB_CLANG_VERSION(2,8) 356#elif ECB_GCC_VERSION(3,0) || ECB_CLANG_VERSION(2,8)
342 #define ecb_decltype(x) __typeof__ (x) 357 #define ecb_decltype(x) __typeof__ (x)
343#endif 358#endif
344 359
345#if _MSC_VER >= 1300 360#if _MSC_VER >= 1300
346 #define ecb_deprecated __declspec(deprecated) 361 #define ecb_deprecated __declspec (deprecated)
347#else 362#else
348 #define ecb_deprecated ecb_attribute ((__deprecated__)) 363 #define ecb_deprecated ecb_attribute ((__deprecated__))
349#endif 364#endif
350 365
366#if _MSC_VER >= 1500
367 #define ecb_deprecated_message(msg) __declspec (deprecated (msg))
368#elif ECB_GCC_VERSION(4,5)
369 #define ecb_deprecated_message(msg) ecb_attribute ((__deprecated__ (msg))
370#else
371 #define ecb_deprecated_message(msg) ecb_deprecated
372#endif
373
374#if _MSC_VER >= 1400
375 #define ecb_noinline __declspec (noinline)
376#else
351#define ecb_noinline ecb_attribute ((__noinline__)) 377 #define ecb_noinline ecb_attribute ((__noinline__))
378#endif
379
352#define ecb_unused ecb_attribute ((__unused__)) 380#define ecb_unused ecb_attribute ((__unused__))
353#define ecb_const ecb_attribute ((__const__)) 381#define ecb_const ecb_attribute ((__const__))
354#define ecb_pure ecb_attribute ((__pure__)) 382#define ecb_pure ecb_attribute ((__pure__))
355 383
356/* http://msdn.microsoft.com/en-us/library/k6ktzx3s.aspx __declspec(noreturn) */ 384#if ECB_C11 || __IBMC_NORETURN
357#if ECB_C11 385 /* http://www-01.ibm.com/support/knowledgecenter/SSGH3R_13.1.0/com.ibm.xlcpp131.aix.doc/language_ref/noreturn.html */
358 #define ecb_noreturn _Noreturn 386 #define ecb_noreturn _Noreturn
387#elif ECB_CPP11
388 #define ecb_noreturn [[noreturn]]
389#elif _MSC_VER >= 1200
390 /* http://msdn.microsoft.com/en-us/library/k6ktzx3s.aspx */
391 #define ecb_noreturn __declspec (noreturn)
359#else 392#else
360 #define ecb_noreturn ecb_attribute ((__noreturn__)) 393 #define ecb_noreturn ecb_attribute ((__noreturn__))
361#endif 394#endif
362 395
363#if ECB_GCC_VERSION(4,3) 396#if ECB_GCC_VERSION(4,3)
… …
390 #define ecb_ctz32(x) __builtin_ctz (x) 423 #define ecb_ctz32(x) __builtin_ctz (x)
391 #define ecb_ctz64(x) __builtin_ctzll (x) 424 #define ecb_ctz64(x) __builtin_ctzll (x)
392 #define ecb_popcount32(x) __builtin_popcount (x) 425 #define ecb_popcount32(x) __builtin_popcount (x)
393 /* no popcountll */ 426 /* no popcountll */
394#else 427#else
395 ecb_function_ int ecb_ctz32 (uint32_t x) ecb_const; 428 ecb_function_ ecb_const int ecb_ctz32 (uint32_t x);
396 ecb_function_ int 429 ecb_function_ ecb_const int
397 ecb_ctz32 (uint32_t x) 430 ecb_ctz32 (uint32_t x)
398 { 431 {
432#if 1400 <= _MSC_VER && (_M_IX86 || _M_X64 || _M_IA64 || _M_ARM)
433 unsigned long r = 0;
434 _BitScanForward (&r, x);
435 return (int)r;
436#else
399 int r = 0; 437 int r = 0;
400 438
401 x &= ~x + 1; /* this isolates the lowest bit */ 439 x &= ~x + 1; /* this isolates the lowest bit */
402 440
403#if ECB_branchless_on_i386 441#if ECB_branchless_on_i386
… …
413 if (x & 0xff00ff00) r += 8; 451 if (x & 0xff00ff00) r += 8;
414 if (x & 0xffff0000) r += 16; 452 if (x & 0xffff0000) r += 16;
415#endif 453#endif
416 454
417 return r; 455 return r;
456#endif
418 } 457 }
419 458
420 ecb_function_ int ecb_ctz64 (uint64_t x) ecb_const; 459 ecb_function_ ecb_const int ecb_ctz64 (uint64_t x);
421 ecb_function_ int 460 ecb_function_ ecb_const int
422 ecb_ctz64 (uint64_t x) 461 ecb_ctz64 (uint64_t x)
423 { 462 {
463#if 1400 <= _MSC_VER && (_M_X64 || _M_IA64 || _M_ARM)
464 unsigned long r = 0;
465 _BitScanForward64 (&r, x);
466 return (int)r;
467#else
424 int shift = x & 0xffffffffU ? 0 : 32; 468 int shift = x & 0xffffffff ? 0 : 32;
425 return ecb_ctz32 (x >> shift) + shift; 469 return ecb_ctz32 (x >> shift) + shift;
470#endif
426 } 471 }
427 472
428 ecb_function_ int ecb_popcount32 (uint32_t x) ecb_const; 473 ecb_function_ ecb_const int ecb_popcount32 (uint32_t x);
429 ecb_function_ int 474 ecb_function_ ecb_const int
430 ecb_popcount32 (uint32_t x) 475 ecb_popcount32 (uint32_t x)
431 { 476 {
432 x -= (x >> 1) & 0x55555555; 477 x -= (x >> 1) & 0x55555555;
433 x = ((x >> 2) & 0x33333333) + (x & 0x33333333); 478 x = ((x >> 2) & 0x33333333) + (x & 0x33333333);
434 x = ((x >> 4) + x) & 0x0f0f0f0f; 479 x = ((x >> 4) + x) & 0x0f0f0f0f;
435 x *= 0x01010101; 480 x *= 0x01010101;
436 481
437 return x >> 24; 482 return x >> 24;
438 } 483 }
439 484
440 ecb_function_ int ecb_ld32 (uint32_t x) ecb_const; 485 ecb_function_ ecb_const int ecb_ld32 (uint32_t x);
441 ecb_function_ int ecb_ld32 (uint32_t x) 486 ecb_function_ ecb_const int ecb_ld32 (uint32_t x)
442 { 487 {
488#if 1400 <= _MSC_VER && (_M_IX86 || _M_X64 || _M_IA64 || _M_ARM)
489 unsigned long r = 0;
490 _BitScanReverse (&r, x);
491 return (int)r;
492#else
443 int r = 0; 493 int r = 0;
444 494
445 if (x >> 16) { x >>= 16; r += 16; } 495 if (x >> 16) { x >>= 16; r += 16; }
446 if (x >> 8) { x >>= 8; r += 8; } 496 if (x >> 8) { x >>= 8; r += 8; }
447 if (x >> 4) { x >>= 4; r += 4; } 497 if (x >> 4) { x >>= 4; r += 4; }
448 if (x >> 2) { x >>= 2; r += 2; } 498 if (x >> 2) { x >>= 2; r += 2; }
449 if (x >> 1) { r += 1; } 499 if (x >> 1) { r += 1; }
450 500
451 return r; 501 return r;
502#endif
452 } 503 }
453 504
454 ecb_function_ int ecb_ld64 (uint64_t x) ecb_const; 505 ecb_function_ ecb_const int ecb_ld64 (uint64_t x);
455 ecb_function_ int ecb_ld64 (uint64_t x) 506 ecb_function_ ecb_const int ecb_ld64 (uint64_t x)
456 { 507 {
508#if 1400 <= _MSC_VER && (_M_X64 || _M_IA64 || _M_ARM)
509 unsigned long r = 0;
510 _BitScanReverse64 (&r, x);
511 return (int)r;
512#else
457 int r = 0; 513 int r = 0;
458 514
459 if (x >> 32) { x >>= 32; r += 32; } 515 if (x >> 32) { x >>= 32; r += 32; }
460 516
461 return r + ecb_ld32 (x); 517 return r + ecb_ld32 (x);
462 }
463#endif 518#endif
519 }
520#endif
464 521
465ecb_function_ ecb_bool ecb_is_pot32 (uint32_t x) ecb_const; 522ecb_function_ ecb_const ecb_bool ecb_is_pot32 (uint32_t x);
466ecb_function_ ecb_bool ecb_is_pot32 (uint32_t x) { return !(x & (x - 1)); } 523ecb_function_ ecb_const ecb_bool ecb_is_pot32 (uint32_t x) { return !(x & (x - 1)); }
467ecb_function_ ecb_bool ecb_is_pot64 (uint64_t x) ecb_const; 524ecb_function_ ecb_const ecb_bool ecb_is_pot64 (uint64_t x);
468ecb_function_ ecb_bool ecb_is_pot64 (uint64_t x) { return !(x & (x - 1)); } 525ecb_function_ ecb_const ecb_bool ecb_is_pot64 (uint64_t x) { return !(x & (x - 1)); }
469 526
470ecb_function_ uint8_t ecb_bitrev8 (uint8_t x) ecb_const; 527ecb_function_ ecb_const uint8_t ecb_bitrev8 (uint8_t x);
471ecb_function_ uint8_t ecb_bitrev8 (uint8_t x) 528ecb_function_ ecb_const uint8_t ecb_bitrev8 (uint8_t x)
472{ 529{
473 return ( (x * 0x0802U & 0x22110U) 530 return ( (x * 0x0802U & 0x22110U)
474 | (x * 0x8020U & 0x88440U)) * 0x10101U >> 16; 531 | (x * 0x8020U & 0x88440U)) * 0x10101U >> 16;
475} 532}
476 533
477ecb_function_ uint16_t ecb_bitrev16 (uint16_t x) ecb_const; 534ecb_function_ ecb_const uint16_t ecb_bitrev16 (uint16_t x);
478ecb_function_ uint16_t ecb_bitrev16 (uint16_t x) 535ecb_function_ ecb_const uint16_t ecb_bitrev16 (uint16_t x)
479{ 536{
480 x = ((x >> 1) & 0x5555) | ((x & 0x5555) << 1); 537 x = ((x >> 1) & 0x5555) | ((x & 0x5555) << 1);
481 x = ((x >> 2) & 0x3333) | ((x & 0x3333) << 2); 538 x = ((x >> 2) & 0x3333) | ((x & 0x3333) << 2);
482 x = ((x >> 4) & 0x0f0f) | ((x & 0x0f0f) << 4); 539 x = ((x >> 4) & 0x0f0f) | ((x & 0x0f0f) << 4);
483 x = ( x >> 8 ) | ( x << 8); 540 x = ( x >> 8 ) | ( x << 8);
484 541
485 return x; 542 return x;
486} 543}
487 544
488ecb_function_ uint32_t ecb_bitrev32 (uint32_t x) ecb_const; 545ecb_function_ ecb_const uint32_t ecb_bitrev32 (uint32_t x);
489ecb_function_ uint32_t ecb_bitrev32 (uint32_t x) 546ecb_function_ ecb_const uint32_t ecb_bitrev32 (uint32_t x)
490{ 547{
491 x = ((x >> 1) & 0x55555555) | ((x & 0x55555555) << 1); 548 x = ((x >> 1) & 0x55555555) | ((x & 0x55555555) << 1);
492 x = ((x >> 2) & 0x33333333) | ((x & 0x33333333) << 2); 549 x = ((x >> 2) & 0x33333333) | ((x & 0x33333333) << 2);
493 x = ((x >> 4) & 0x0f0f0f0f) | ((x & 0x0f0f0f0f) << 4); 550 x = ((x >> 4) & 0x0f0f0f0f) | ((x & 0x0f0f0f0f) << 4);
494 x = ((x >> 8) & 0x00ff00ff) | ((x & 0x00ff00ff) << 8); 551 x = ((x >> 8) & 0x00ff00ff) | ((x & 0x00ff00ff) << 8);
… …
497 return x; 554 return x;
498} 555}
499 556
500/* popcount64 is only available on 64 bit cpus as gcc builtin */ 557/* popcount64 is only available on 64 bit cpus as gcc builtin */
501/* so for this version we are lazy */ 558/* so for this version we are lazy */
502ecb_function_ int ecb_popcount64 (uint64_t x) ecb_const; 559ecb_function_ ecb_const int ecb_popcount64 (uint64_t x);
503ecb_function_ int 560ecb_function_ ecb_const int
504ecb_popcount64 (uint64_t x) 561ecb_popcount64 (uint64_t x)
505{ 562{
506 return ecb_popcount32 (x) + ecb_popcount32 (x >> 32); 563 return ecb_popcount32 (x) + ecb_popcount32 (x >> 32);
507} 564}
508 565
509ecb_inline uint8_t ecb_rotl8 (uint8_t x, unsigned int count) ecb_const; 566ecb_inline ecb_const uint8_t ecb_rotl8 (uint8_t x, unsigned int count);
510ecb_inline uint8_t ecb_rotr8 (uint8_t x, unsigned int count) ecb_const; 567ecb_inline ecb_const uint8_t ecb_rotr8 (uint8_t x, unsigned int count);
511ecb_inline uint16_t ecb_rotl16 (uint16_t x, unsigned int count) ecb_const; 568ecb_inline ecb_const uint16_t ecb_rotl16 (uint16_t x, unsigned int count);
512ecb_inline uint16_t ecb_rotr16 (uint16_t x, unsigned int count) ecb_const; 569ecb_inline ecb_const uint16_t ecb_rotr16 (uint16_t x, unsigned int count);
513ecb_inline uint32_t ecb_rotl32 (uint32_t x, unsigned int count) ecb_const; 570ecb_inline ecb_const uint32_t ecb_rotl32 (uint32_t x, unsigned int count);
514ecb_inline uint32_t ecb_rotr32 (uint32_t x, unsigned int count) ecb_const; 571ecb_inline ecb_const uint32_t ecb_rotr32 (uint32_t x, unsigned int count);
515ecb_inline uint64_t ecb_rotl64 (uint64_t x, unsigned int count) ecb_const; 572ecb_inline ecb_const uint64_t ecb_rotl64 (uint64_t x, unsigned int count);
516ecb_inline uint64_t ecb_rotr64 (uint64_t x, unsigned int count) ecb_const; 573ecb_inline ecb_const uint64_t ecb_rotr64 (uint64_t x, unsigned int count);
517 574
518ecb_inline uint8_t ecb_rotl8 (uint8_t x, unsigned int count) { return (x >> ( 8 - count)) | (x << count); } 575ecb_inline ecb_const uint8_t ecb_rotl8 (uint8_t x, unsigned int count) { return (x >> ( 8 - count)) | (x << count); }
519ecb_inline uint8_t ecb_rotr8 (uint8_t x, unsigned int count) { return (x << ( 8 - count)) | (x >> count); } 576ecb_inline ecb_const uint8_t ecb_rotr8 (uint8_t x, unsigned int count) { return (x << ( 8 - count)) | (x >> count); }
520ecb_inline uint16_t ecb_rotl16 (uint16_t x, unsigned int count) { return (x >> (16 - count)) | (x << count); } 577ecb_inline ecb_const uint16_t ecb_rotl16 (uint16_t x, unsigned int count) { return (x >> (16 - count)) | (x << count); }
521ecb_inline uint16_t ecb_rotr16 (uint16_t x, unsigned int count) { return (x << (16 - count)) | (x >> count); } 578ecb_inline ecb_const uint16_t ecb_rotr16 (uint16_t x, unsigned int count) { return (x << (16 - count)) | (x >> count); }
522ecb_inline uint32_t ecb_rotl32 (uint32_t x, unsigned int count) { return (x >> (32 - count)) | (x << count); } 579ecb_inline ecb_const uint32_t ecb_rotl32 (uint32_t x, unsigned int count) { return (x >> (32 - count)) | (x << count); }
523ecb_inline uint32_t ecb_rotr32 (uint32_t x, unsigned int count) { return (x << (32 - count)) | (x >> count); } 580ecb_inline ecb_const uint32_t ecb_rotr32 (uint32_t x, unsigned int count) { return (x << (32 - count)) | (x >> count); }
524ecb_inline uint64_t ecb_rotl64 (uint64_t x, unsigned int count) { return (x >> (64 - count)) | (x << count); } 581ecb_inline ecb_const uint64_t ecb_rotl64 (uint64_t x, unsigned int count) { return (x >> (64 - count)) | (x << count); }
525ecb_inline uint64_t ecb_rotr64 (uint64_t x, unsigned int count) { return (x << (64 - count)) | (x >> count); } 582ecb_inline ecb_const uint64_t ecb_rotr64 (uint64_t x, unsigned int count) { return (x << (64 - count)) | (x >> count); }
526 583
527#if ECB_GCC_VERSION(4,3) || (ECB_CLANG_BUILTIN(__builtin_bswap32) && ECB_CLANG_BUILTIN(__builtin_bswap64)) 584#if ECB_GCC_VERSION(4,3) || (ECB_CLANG_BUILTIN(__builtin_bswap32) && ECB_CLANG_BUILTIN(__builtin_bswap64))
585 #if ECB_GCC_VERSION(4,8) || ECB_CLANG_BUILTIN(__builtin_bswap16)
586 #define ecb_bswap16(x) __builtin_bswap16 (x)
587 #else
528 #define ecb_bswap16(x) (__builtin_bswap32 (x) >> 16) 588 #define ecb_bswap16(x) (__builtin_bswap32 (x) >> 16)
589 #endif
529 #define ecb_bswap32(x) __builtin_bswap32 (x) 590 #define ecb_bswap32(x) __builtin_bswap32 (x)
530 #define ecb_bswap64(x) __builtin_bswap64 (x) 591 #define ecb_bswap64(x) __builtin_bswap64 (x)
592#elif _MSC_VER
593 #include <stdlib.h>
594 #define ecb_bswap16(x) ((uint16_t)_byteswap_ushort ((uint16_t)(x)))
595 #define ecb_bswap32(x) ((uint32_t)_byteswap_ulong ((uint32_t)(x)))
596 #define ecb_bswap64(x) ((uint64_t)_byteswap_uint64 ((uint64_t)(x)))
531#else 597#else
532 ecb_function_ uint16_t ecb_bswap16 (uint16_t x) ecb_const; 598 ecb_function_ ecb_const uint16_t ecb_bswap16 (uint16_t x);
533 ecb_function_ uint16_t 599 ecb_function_ ecb_const uint16_t
534 ecb_bswap16 (uint16_t x) 600 ecb_bswap16 (uint16_t x)
535 { 601 {
536 return ecb_rotl16 (x, 8); 602 return ecb_rotl16 (x, 8);
537 } 603 }
538 604
539 ecb_function_ uint32_t ecb_bswap32 (uint32_t x) ecb_const; 605 ecb_function_ ecb_const uint32_t ecb_bswap32 (uint32_t x);
540 ecb_function_ uint32_t 606 ecb_function_ ecb_const uint32_t
541 ecb_bswap32 (uint32_t x) 607 ecb_bswap32 (uint32_t x)
542 { 608 {
543 return (((uint32_t)ecb_bswap16 (x)) << 16) | ecb_bswap16 (x >> 16); 609 return (((uint32_t)ecb_bswap16 (x)) << 16) | ecb_bswap16 (x >> 16);
544 } 610 }
545 611
546 ecb_function_ uint64_t ecb_bswap64 (uint64_t x) ecb_const; 612 ecb_function_ ecb_const uint64_t ecb_bswap64 (uint64_t x);
547 ecb_function_ uint64_t 613 ecb_function_ ecb_const uint64_t
548 ecb_bswap64 (uint64_t x) 614 ecb_bswap64 (uint64_t x)
549 { 615 {
550 return (((uint64_t)ecb_bswap32 (x)) << 32) | ecb_bswap32 (x >> 32); 616 return (((uint64_t)ecb_bswap32 (x)) << 32) | ecb_bswap32 (x >> 32);
551 } 617 }
552#endif 618#endif
553 619
554#if ECB_GCC_VERSION(4,5) || ECB_CLANG_BUILTIN(__builtin_unreachable) 620#if ECB_GCC_VERSION(4,5) || ECB_CLANG_BUILTIN(__builtin_unreachable)
555 #define ecb_unreachable() __builtin_unreachable () 621 #define ecb_unreachable() __builtin_unreachable ()
556#else 622#else
557 /* this seems to work fine, but gcc always emits a warning for it :/ */ 623 /* this seems to work fine, but gcc always emits a warning for it :/ */
558 ecb_inline void ecb_unreachable (void) ecb_noreturn; 624 ecb_inline ecb_noreturn void ecb_unreachable (void);
559 ecb_inline void ecb_unreachable (void) { } 625 ecb_inline ecb_noreturn void ecb_unreachable (void) { }
560#endif 626#endif
561 627
562/* try to tell the compiler that some condition is definitely true */ 628/* try to tell the compiler that some condition is definitely true */
563#define ecb_assume(cond) if (!(cond)) ecb_unreachable (); else 0 629#define ecb_assume(cond) if (!(cond)) ecb_unreachable (); else 0
564 630
565ecb_inline unsigned char ecb_byteorder_helper (void) ecb_const; 631ecb_inline ecb_const unsigned char ecb_byteorder_helper (void);
566ecb_inline unsigned char 632ecb_inline ecb_const unsigned char
567ecb_byteorder_helper (void) 633ecb_byteorder_helper (void)
568{ 634{
569 /* the union code still generates code under pressure in gcc, */ 635 /* the union code still generates code under pressure in gcc, */
570 /* but less than using pointers, and always seems to */ 636 /* but less than using pointers, and always seems to */
571 /* successfully return a constant. */ 637 /* successfully return a constant. */
572 /* the reason why we have this horrible preprocessor mess */ 638 /* the reason why we have this horrible preprocessor mess */
573 /* is to avoid it in all cases, at least on common architectures */ 639 /* is to avoid it in all cases, at least on common architectures */
574 /* or when using a recent enough gcc version (>= 4.6) */ 640 /* or when using a recent enough gcc version (>= 4.6) */
575#if __i386 || __i386__ || _M_X86 || __amd64 || __amd64__ || _M_X64 641#if ((__i386 || __i386__) && !__VOS__) || _M_IX86 || ECB_GCC_AMD64 || ECB_MSVC_AMD64
576 return 0x44; 642 return 0x44;
577#elif __BYTE_ORDER__ && __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__ 643#elif __BYTE_ORDER__ && __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__
578 return 0x44; 644 return 0x44;
579#elif __BYTE_ORDER__ && __BYTE_ORDER__ == __ORDER_BIG_ENDIAN__ 645#elif __BYTE_ORDER__ && __BYTE_ORDER__ == __ORDER_BIG_ENDIAN__
580 return 0x11; 646 return 0x11;
… …
586 } u = { 0x11223344 }; 652 } u = { 0x11223344 };
587 return u.c; 653 return u.c;
588#endif 654#endif
589} 655}
590 656
591ecb_inline ecb_bool ecb_big_endian (void) ecb_const; 657ecb_inline ecb_const ecb_bool ecb_big_endian (void);
592ecb_inline ecb_bool ecb_big_endian (void) { return ecb_byteorder_helper () == 0x11; } 658ecb_inline ecb_const ecb_bool ecb_big_endian (void) { return ecb_byteorder_helper () == 0x11; }
593ecb_inline ecb_bool ecb_little_endian (void) ecb_const; 659ecb_inline ecb_const ecb_bool ecb_little_endian (void);
594ecb_inline ecb_bool ecb_little_endian (void) { return ecb_byteorder_helper () == 0x44; } 660ecb_inline ecb_const ecb_bool ecb_little_endian (void) { return ecb_byteorder_helper () == 0x44; }
595 661
596#if ECB_GCC_VERSION(3,0) || ECB_C99 662#if ECB_GCC_VERSION(3,0) || ECB_C99
597 #define ecb_mod(m,n) ((m) % (n) + ((m) % (n) < 0 ? (n) : 0)) 663 #define ecb_mod(m,n) ((m) % (n) + ((m) % (n) < 0 ? (n) : 0))
598#else 664#else
599 #define ecb_mod(m,n) ((m) < 0 ? ((n) - 1 - ((-1 - (m)) % (n))) : ((m) % (n))) 665 #define ecb_mod(m,n) ((m) < 0 ? ((n) - 1 - ((-1 - (m)) % (n))) : ((m) % (n)))
600#endif 666#endif
601 667
602#if __cplusplus 668#if ECB_CPP
603 template<typename T> 669 template<typename T>
604 static inline T ecb_div_rd (T val, T div) 670 static inline T ecb_div_rd (T val, T div)
605 { 671 {
606 return val < 0 ? - ((-val + div - 1) / div) : (val ) / div; 672 return val < 0 ? - ((-val + div - 1) / div) : (val ) / div;
607 } 673 }
… …
624 } 690 }
625#else 691#else
626 #define ecb_array_length(name) (sizeof (name) / sizeof (name [0])) 692 #define ecb_array_length(name) (sizeof (name) / sizeof (name [0]))
627#endif 693#endif
628 694
695ecb_function_ ecb_const uint32_t ecb_binary16_to_binary32 (uint32_t x);
696ecb_function_ ecb_const uint32_t
697ecb_binary16_to_binary32 (uint32_t x)
698{
699 unsigned int s = (x & 0x8000) << (31 - 15);
700 int e = (x >> 10) & 0x001f;
701 unsigned int m = x & 0x03ff;
702
703 if (ecb_expect_false (e == 31))
704 /* infinity or NaN */
705 e = 255 - (127 - 15);
706 else if (ecb_expect_false (!e))
707 {
708 if (ecb_expect_true (!m))
709 /* zero, handled by code below by forcing e to 0 */
710 e = 0 - (127 - 15);
711 else
712 {
713 /* subnormal, renormalise */
714 unsigned int s = 10 - ecb_ld32 (m);
715
716 m = (m << s) & 0x3ff; /* mask implicit bit */
717 e -= s - 1;
718 }
719 }
720
721 /* e and m now are normalised, or zero, (or inf or nan) */
722 e += 127 - 15;
723
724 return s | (e << 23) | (m << (23 - 10));
725}
726
727ecb_function_ ecb_const uint16_t ecb_binary32_to_binary16 (uint32_t x);
728ecb_function_ ecb_const uint16_t
729ecb_binary32_to_binary16 (uint32_t x)
730{
731 unsigned int s = (x >> 16) & 0x00008000; /* sign bit, the easy part */
732 unsigned int e = ((x >> 23) & 0x000000ff) - (127 - 15); /* the desired exponent */
733 unsigned int m = x & 0x007fffff;
734
735 x &= 0x7fffffff;
736
737 /* if it's within range of binary16 normals, use fast path */
738 if (ecb_expect_true (0x38800000 <= x && x <= 0x477fefff))
739 {
740 /* mantissa round-to-even */
741 m += 0x00000fff + ((m >> (23 - 10)) & 1);
742
743 /* handle overflow */
744 if (ecb_expect_false (m >= 0x00800000))
745 {
746 m >>= 1;
747 e += 1;
748 }
749
750 return s | (e << 10) | (m >> (23 - 10));
751 }
752
753 /* handle large numbers and infinity */
754 if (ecb_expect_true (0x477fefff < x && x <= 0x7f800000))
755 return s | 0x7c00;
756
757 /* handle zero, subnormals and small numbers */
758 if (ecb_expect_true (x < 0x38800000))
759 {
760 /* zero */
761 if (ecb_expect_true (!x))
762 return s;
763
764 /* handle subnormals */
765
766 /* too small, will be zero */
767 if (e < (14 - 24)) /* might not be sharp, but is good enough */
768 return s;
769
770 m |= 0x00800000; /* make implicit bit explicit */
771
772 /* very tricky - we need to round to the nearest e (+10) bit value */
773 {
774 unsigned int bits = 14 - e;
775 unsigned int half = (1 << (bits - 1)) - 1;
776 unsigned int even = (m >> bits) & 1;
777
778 /* if this overflows, we will end up with a normalised number */
779 m = (m + half + even) >> bits;
780 }
781
782 return s | m;
783 }
784
785 /* handle NaNs, preserve leftmost nan bits, but make sure we don't turn them into infinities */
786 m >>= 13;
787
788 return s | 0x7c00 | m | !m;
789}
790
629/*******************************************************************************/ 791/*******************************************************************************/
630/* floating point stuff, can be disabled by defining ECB_NO_LIBM */ 792/* floating point stuff, can be disabled by defining ECB_NO_LIBM */
631 793
632/* basically, everything uses "ieee pure-endian" floating point numbers */ 794/* basically, everything uses "ieee pure-endian" floating point numbers */
633/* the only noteworthy exception is ancient armle, which uses order 43218765 */ 795/* the only noteworthy exception is ancient armle, which uses order 43218765 */
634#if 0 \ 796#if 0 \
635 || __i386 || __i386__ \ 797 || __i386 || __i386__ \
636 || __amd64 || __amd64__ || __x86_64 || __x86_64__ \ 798 || ECB_GCC_AMD64 \
637 || __powerpc__ || __ppc__ || __powerpc64__ || __ppc64__ \ 799 || __powerpc__ || __ppc__ || __powerpc64__ || __ppc64__ \
638 || defined __s390__ || defined __s390x__ \ 800 || defined __s390__ || defined __s390x__ \
639 || defined __mips__ \ 801 || defined __mips__ \
640 || defined __alpha__ \ 802 || defined __alpha__ \
641 || defined __hppa__ \ 803 || defined __hppa__ \
642 || defined __ia64__ \ 804 || defined __ia64__ \
643 || defined __m68k__ \ 805 || defined __m68k__ \
644 || defined __m88k__ \ 806 || defined __m88k__ \
645 || defined __sh__ \ 807 || defined __sh__ \
646 || defined _M_IX86 || defined _M_AMD64 || defined _M_IA64 \ 808 || defined _M_IX86 || defined ECB_MSVC_AMD64 || defined _M_IA64 \
647 || (defined __arm__ && (defined __ARM_EABI__ || defined __EABI__ || defined __VFP_FP__ || defined _WIN32_WCE || defined __ANDROID__)) \ 809 || (defined __arm__ && (defined __ARM_EABI__ || defined __EABI__ || defined __VFP_FP__ || defined _WIN32_WCE || defined __ANDROID__)) \
648 || defined __aarch64__ 810 || defined __aarch64__
649 #define ECB_STDFP 1 811 #define ECB_STDFP 1
650 #include <string.h> /* for memcpy */ 812 #include <string.h> /* for memcpy */
651#else 813#else
… …
667 #define ECB_NAN NAN 829 #define ECB_NAN NAN
668 #else 830 #else
669 #define ECB_NAN ECB_INFINITY 831 #define ECB_NAN ECB_INFINITY
670 #endif 832 #endif
671 833
672 /* converts an ieee half/binary16 to a float */ 834 #if ECB_C99 || _XOPEN_VERSION >= 600 || _POSIX_VERSION >= 200112L
673 ecb_function_ float ecb_binary16_to_float (uint16_t x) ecb_const; 835 #define ecb_ldexpf(x,e) ldexpf ((x), (e))
674 ecb_function_ float 836 #define ecb_frexpf(x,e) frexpf ((x), (e))
675 ecb_binary16_to_float (uint16_t x) 837 #else
676 { 838 #define ecb_ldexpf(x,e) (float) ldexp ((double) (x), (e))
677 int e = (x >> 10) & 0x1f; 839 #define ecb_frexpf(x,e) (float) frexp ((double) (x), (e))
678 int m = x & 0x3ff; 840 #endif
679 float r;
680
681 if (!e ) r = ldexpf (m , -24);
682 else if (e != 31) r = ldexpf (m + 0x400, e - 25);
683 else if (m ) r = ECB_NAN;
684 else r = ECB_INFINITY;
685
686 return x & 0x8000 ? -r : r;
687 }
688 841
689 /* convert a float to ieee single/binary32 */ 842 /* convert a float to ieee single/binary32 */
690 ecb_function_ uint32_t ecb_float_to_binary32 (float x) ecb_const; 843 ecb_function_ ecb_const uint32_t ecb_float_to_binary32 (float x);
691 ecb_function_ uint32_t 844 ecb_function_ ecb_const uint32_t
692 ecb_float_to_binary32 (float x) 845 ecb_float_to_binary32 (float x)
693 { 846 {
694 uint32_t r; 847 uint32_t r;
695 848
696 #if ECB_STDFP 849 #if ECB_STDFP
… …
703 if (x == 0e0f ) return 0x00000000U; 856 if (x == 0e0f ) return 0x00000000U;
704 if (x > +3.40282346638528860e+38f) return 0x7f800000U; 857 if (x > +3.40282346638528860e+38f) return 0x7f800000U;
705 if (x < -3.40282346638528860e+38f) return 0xff800000U; 858 if (x < -3.40282346638528860e+38f) return 0xff800000U;
706 if (x != x ) return 0x7fbfffffU; 859 if (x != x ) return 0x7fbfffffU;
707 860
708 m = frexpf (x, &e) * 0x1000000U; 861 m = ecb_frexpf (x, &e) * 0x1000000U;
709 862
710 r = m & 0x80000000U; 863 r = m & 0x80000000U;
711 864
712 if (r) 865 if (r)
713 m = -m; 866 m = -m;
… …
725 878
726 return r; 879 return r;
727 } 880 }
728 881
729 /* converts an ieee single/binary32 to a float */ 882 /* converts an ieee single/binary32 to a float */
730 ecb_function_ float ecb_binary32_to_float (uint32_t x) ecb_const; 883 ecb_function_ ecb_const float ecb_binary32_to_float (uint32_t x);
731 ecb_function_ float 884 ecb_function_ ecb_const float
732 ecb_binary32_to_float (uint32_t x) 885 ecb_binary32_to_float (uint32_t x)
733 { 886 {
734 float r; 887 float r;
735 888
736 #if ECB_STDFP 889 #if ECB_STDFP
… …
746 x |= 0x800000U; 899 x |= 0x800000U;
747 else 900 else
748 e = 1; 901 e = 1;
749 902
750 /* we distrust ldexpf a bit and do the 2**-24 scaling by an extra multiply */ 903 /* we distrust ldexpf a bit and do the 2**-24 scaling by an extra multiply */
751 r = ldexpf (x * (0.5f / 0x800000U), e - 126); 904 r = ecb_ldexpf (x * (0.5f / 0x800000U), e - 126);
752 905
753 r = neg ? -r : r; 906 r = neg ? -r : r;
754 #endif 907 #endif
755 908
756 return r; 909 return r;
757 } 910 }
758 911
759 /* convert a double to ieee double/binary64 */ 912 /* convert a double to ieee double/binary64 */
760 ecb_function_ uint64_t ecb_double_to_binary64 (double x) ecb_const; 913 ecb_function_ ecb_const uint64_t ecb_double_to_binary64 (double x);
761 ecb_function_ uint64_t 914 ecb_function_ ecb_const uint64_t
762 ecb_double_to_binary64 (double x) 915 ecb_double_to_binary64 (double x)
763 { 916 {
764 uint64_t r; 917 uint64_t r;
765 918
766 #if ECB_STDFP 919 #if ECB_STDFP
… …
795 948
796 return r; 949 return r;
797 } 950 }
798 951
799 /* converts an ieee double/binary64 to a double */ 952 /* converts an ieee double/binary64 to a double */
800 ecb_function_ double ecb_binary64_to_double (uint64_t x) ecb_const; 953 ecb_function_ ecb_const double ecb_binary64_to_double (uint64_t x);
801 ecb_function_ double 954 ecb_function_ ecb_const double
802 ecb_binary64_to_double (uint64_t x) 955 ecb_binary64_to_double (uint64_t x)
803 { 956 {
804 double r; 957 double r;
805 958
806 #if ECB_STDFP 959 #if ECB_STDFP
… …
824 #endif 977 #endif
825 978
826 return r; 979 return r;
827 } 980 }
828 981
829#endif 982 /* convert a float to ieee half/binary16 */
983 ecb_function_ ecb_const uint16_t ecb_float_to_binary16 (float x);
984 ecb_function_ ecb_const uint16_t
985 ecb_float_to_binary16 (float x)
986 {
987 return ecb_binary32_to_binary16 (ecb_float_to_binary32 (x));
988 }
830 989
831#endif 990 /* convert an ieee half/binary16 to float */
991 ecb_function_ ecb_const float ecb_binary16_to_float (uint16_t x);
992 ecb_function_ ecb_const float
993 ecb_binary16_to_float (uint16_t x)
994 {
995 return ecb_binary32_to_float (ecb_binary16_to_binary32 (x));
996 }
832 997
998#endif
999
1000#endif
1001

Diff Legend

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