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

Comparing libecb/ecb.h (file contents):
Revision 1.148 by root, Fri Oct 24 04:50:13 2014 UTC vs.
Revision 1.174 by root, Wed Nov 25 02:36:53 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;
… …
67 typedef uint32_t uintptr_t; 67 typedef uint32_t uintptr_t;
68 typedef int32_t intptr_t; 68 typedef int32_t intptr_t;
69 #endif 69 #endif
70#else 70#else
71 #include <inttypes.h> 71 #include <inttypes.h>
72 #if UINTMAX_MAX > 0xffffffffU 72 #if (defined INTPTR_MAX ? INTPTR_MAX : ULONG_MAX) > 0xffffffffU
73 #define ECB_PTRSIZE 8 73 #define ECB_PTRSIZE 8
74 #else 74 #else
75 #define ECB_PTRSIZE 4 75 #define ECB_PTRSIZE 4
76 #endif 76 #endif
77#endif 77#endif
78 78
79#define ECB_GCC_AMD64 (__amd64 || __amd64__ || __x86_64 || __x86_64__)
80#define ECB_MSVC_AMD64 (_M_AMD64 || _M_X64)
81
79/* work around x32 idiocy by defining proper macros */ 82/* work around x32 idiocy by defining proper macros */
80#if __amd64 || __x86_64 || _M_AMD64 || _M_X64 83#if ECB_GCC_AMD64 || ECB_MSVC_AMD64
81 #if _ILP32 84 #if _ILP32
82 #define ECB_AMD64_X32 1 85 #define ECB_AMD64_X32 1
83 #else 86 #else
84 #define ECB_AMD64 1 87 #define ECB_AMD64 1
85 #endif 88 #endif
… …
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)
150#endif 153#endif
151 154
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 */
162#endif
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)
… …
343#elif ECB_GCC_VERSION(3,0) || ECB_CLANG_VERSION(2,8) 356#elif ECB_GCC_VERSION(3,0) || ECB_CLANG_VERSION(2,8)
344 #define ecb_decltype(x) __typeof__ (x) 357 #define ecb_decltype(x) __typeof__ (x)
345#endif 358#endif
346 359
347#if _MSC_VER >= 1300 360#if _MSC_VER >= 1300
348 #define ecb_deprecated __declspec(deprecated) 361 #define ecb_deprecated __declspec (deprecated)
349#else 362#else
350 #define ecb_deprecated ecb_attribute ((__deprecated__)) 363 #define ecb_deprecated ecb_attribute ((__deprecated__))
351#endif 364#endif
352 365
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
353#define ecb_noinline ecb_attribute ((__noinline__)) 377 #define ecb_noinline ecb_attribute ((__noinline__))
378#endif
379
354#define ecb_unused ecb_attribute ((__unused__)) 380#define ecb_unused ecb_attribute ((__unused__))
355#define ecb_const ecb_attribute ((__const__)) 381#define ecb_const ecb_attribute ((__const__))
356#define ecb_pure ecb_attribute ((__pure__)) 382#define ecb_pure ecb_attribute ((__pure__))
357 383
358/* TODO http://msdn.microsoft.com/en-us/library/k6ktzx3s.aspx __declspec(noreturn) */
359#if ECB_C11 || __IBMC_NORETURN 384#if ECB_C11 || __IBMC_NORETURN
360 /* http://pic.dhe.ibm.com/infocenter/compbg/v121v141/topic/com.ibm.xlcpp121.bg.doc/language_ref/noreturn.html */ 385 /* 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 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)
362#else 392#else
363 #define ecb_noreturn ecb_attribute ((__noreturn__)) 393 #define ecb_noreturn ecb_attribute ((__noreturn__))
364#endif 394#endif
365 395
366#if ECB_GCC_VERSION(4,3) 396#if ECB_GCC_VERSION(4,3)
… …
393 #define ecb_ctz32(x) __builtin_ctz (x) 423 #define ecb_ctz32(x) __builtin_ctz (x)
394 #define ecb_ctz64(x) __builtin_ctzll (x) 424 #define ecb_ctz64(x) __builtin_ctzll (x)
395 #define ecb_popcount32(x) __builtin_popcount (x) 425 #define ecb_popcount32(x) __builtin_popcount (x)
396 /* no popcountll */ 426 /* no popcountll */
397#else 427#else
398 ecb_function_ int ecb_ctz32 (uint32_t x) ecb_const; 428 ecb_function_ ecb_const int ecb_ctz32 (uint32_t x);
399 ecb_function_ int 429 ecb_function_ ecb_const int
400 ecb_ctz32 (uint32_t x) 430 ecb_ctz32 (uint32_t x)
401 { 431 {
432#if 1400 <= _MSC_VER && (_M_IX86 || _M_X64 || _M_IA64 || _M_ARM)
433 unsigned long r;
434 _BitScanForward (&r, x);
435 return (int)r;
436#else
402 int r = 0; 437 int r = 0;
403 438
404 x &= ~x + 1; /* this isolates the lowest bit */ 439 x &= ~x + 1; /* this isolates the lowest bit */
405 440
406#if ECB_branchless_on_i386 441#if ECB_branchless_on_i386
… …
416 if (x & 0xff00ff00) r += 8; 451 if (x & 0xff00ff00) r += 8;
417 if (x & 0xffff0000) r += 16; 452 if (x & 0xffff0000) r += 16;
418#endif 453#endif
419 454
420 return r; 455 return r;
456#endif
421 } 457 }
422 458
423 ecb_function_ int ecb_ctz64 (uint64_t x) ecb_const; 459 ecb_function_ ecb_const int ecb_ctz64 (uint64_t x);
424 ecb_function_ int 460 ecb_function_ ecb_const int
425 ecb_ctz64 (uint64_t x) 461 ecb_ctz64 (uint64_t x)
426 { 462 {
463#if 1400 <= _MSC_VER && (_M_X64 || _M_IA64 || _M_ARM)
464 unsigned long r;
465 _BitScanForward64 (&r, x);
466 return (int)r;
467#else
427 int shift = x & 0xffffffffU ? 0 : 32; 468 int shift = x & 0xffffffff ? 0 : 32;
428 return ecb_ctz32 (x >> shift) + shift; 469 return ecb_ctz32 (x >> shift) + shift;
470#endif
429 } 471 }
430 472
431 ecb_function_ int ecb_popcount32 (uint32_t x) ecb_const; 473 ecb_function_ ecb_const int ecb_popcount32 (uint32_t x);
432 ecb_function_ int 474 ecb_function_ ecb_const int
433 ecb_popcount32 (uint32_t x) 475 ecb_popcount32 (uint32_t x)
434 { 476 {
435 x -= (x >> 1) & 0x55555555; 477 x -= (x >> 1) & 0x55555555;
436 x = ((x >> 2) & 0x33333333) + (x & 0x33333333); 478 x = ((x >> 2) & 0x33333333) + (x & 0x33333333);
437 x = ((x >> 4) + x) & 0x0f0f0f0f; 479 x = ((x >> 4) + x) & 0x0f0f0f0f;
438 x *= 0x01010101; 480 x *= 0x01010101;
439 481
440 return x >> 24; 482 return x >> 24;
441 } 483 }
442 484
443 ecb_function_ int ecb_ld32 (uint32_t x) ecb_const; 485 ecb_function_ ecb_const int ecb_ld32 (uint32_t x);
444 ecb_function_ int ecb_ld32 (uint32_t x) 486 ecb_function_ ecb_const int ecb_ld32 (uint32_t x)
445 { 487 {
488#if 1400 <= _MSC_VER && (_M_IX86 || _M_X64 || _M_IA64 || _M_ARM)
489 unsigned long r;
490 _BitScanReverse (&r, x);
491 return (int)r;
492#else
446 int r = 0; 493 int r = 0;
447 494
448 if (x >> 16) { x >>= 16; r += 16; } 495 if (x >> 16) { x >>= 16; r += 16; }
449 if (x >> 8) { x >>= 8; r += 8; } 496 if (x >> 8) { x >>= 8; r += 8; }
450 if (x >> 4) { x >>= 4; r += 4; } 497 if (x >> 4) { x >>= 4; r += 4; }
451 if (x >> 2) { x >>= 2; r += 2; } 498 if (x >> 2) { x >>= 2; r += 2; }
452 if (x >> 1) { r += 1; } 499 if (x >> 1) { r += 1; }
453 500
454 return r; 501 return r;
502#endif
455 } 503 }
456 504
457 ecb_function_ int ecb_ld64 (uint64_t x) ecb_const; 505 ecb_function_ ecb_const int ecb_ld64 (uint64_t x);
458 ecb_function_ int ecb_ld64 (uint64_t x) 506 ecb_function_ ecb_const int ecb_ld64 (uint64_t x)
459 { 507 {
508#if 1400 <= _MSC_VER && (_M_X64 || _M_IA64 || _M_ARM)
509 unsigned long r;
510 _BitScanReverse64 (&r, x);
511 return (int)r;
512#else
460 int r = 0; 513 int r = 0;
461 514
462 if (x >> 32) { x >>= 32; r += 32; } 515 if (x >> 32) { x >>= 32; r += 32; }
463 516
464 return r + ecb_ld32 (x); 517 return r + ecb_ld32 (x);
465 }
466#endif 518#endif
519 }
520#endif
467 521
468ecb_function_ ecb_bool ecb_is_pot32 (uint32_t x) ecb_const; 522ecb_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)); } 523ecb_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; 524ecb_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)); } 525ecb_function_ ecb_const ecb_bool ecb_is_pot64 (uint64_t x) { return !(x & (x - 1)); }
472 526
473ecb_function_ uint8_t ecb_bitrev8 (uint8_t x) ecb_const; 527ecb_function_ ecb_const uint8_t ecb_bitrev8 (uint8_t x);
474ecb_function_ uint8_t ecb_bitrev8 (uint8_t x) 528ecb_function_ ecb_const uint8_t ecb_bitrev8 (uint8_t x)
475{ 529{
476 return ( (x * 0x0802U & 0x22110U) 530 return ( (x * 0x0802U & 0x22110U)
477 | (x * 0x8020U & 0x88440U)) * 0x10101U >> 16; 531 | (x * 0x8020U & 0x88440U)) * 0x10101U >> 16;
478} 532}
479 533
480ecb_function_ uint16_t ecb_bitrev16 (uint16_t x) ecb_const; 534ecb_function_ ecb_const uint16_t ecb_bitrev16 (uint16_t x);
481ecb_function_ uint16_t ecb_bitrev16 (uint16_t x) 535ecb_function_ ecb_const uint16_t ecb_bitrev16 (uint16_t x)
482{ 536{
483 x = ((x >> 1) & 0x5555) | ((x & 0x5555) << 1); 537 x = ((x >> 1) & 0x5555) | ((x & 0x5555) << 1);
484 x = ((x >> 2) & 0x3333) | ((x & 0x3333) << 2); 538 x = ((x >> 2) & 0x3333) | ((x & 0x3333) << 2);
485 x = ((x >> 4) & 0x0f0f) | ((x & 0x0f0f) << 4); 539 x = ((x >> 4) & 0x0f0f) | ((x & 0x0f0f) << 4);
486 x = ( x >> 8 ) | ( x << 8); 540 x = ( x >> 8 ) | ( x << 8);
487 541
488 return x; 542 return x;
489} 543}
490 544
491ecb_function_ uint32_t ecb_bitrev32 (uint32_t x) ecb_const; 545ecb_function_ ecb_const uint32_t ecb_bitrev32 (uint32_t x);
492ecb_function_ uint32_t ecb_bitrev32 (uint32_t x) 546ecb_function_ ecb_const uint32_t ecb_bitrev32 (uint32_t x)
493{ 547{
494 x = ((x >> 1) & 0x55555555) | ((x & 0x55555555) << 1); 548 x = ((x >> 1) & 0x55555555) | ((x & 0x55555555) << 1);
495 x = ((x >> 2) & 0x33333333) | ((x & 0x33333333) << 2); 549 x = ((x >> 2) & 0x33333333) | ((x & 0x33333333) << 2);
496 x = ((x >> 4) & 0x0f0f0f0f) | ((x & 0x0f0f0f0f) << 4); 550 x = ((x >> 4) & 0x0f0f0f0f) | ((x & 0x0f0f0f0f) << 4);
497 x = ((x >> 8) & 0x00ff00ff) | ((x & 0x00ff00ff) << 8); 551 x = ((x >> 8) & 0x00ff00ff) | ((x & 0x00ff00ff) << 8);
… …
500 return x; 554 return x;
501} 555}
502 556
503/* popcount64 is only available on 64 bit cpus as gcc builtin */ 557/* popcount64 is only available on 64 bit cpus as gcc builtin */
504/* so for this version we are lazy */ 558/* so for this version we are lazy */
505ecb_function_ int ecb_popcount64 (uint64_t x) ecb_const; 559ecb_function_ ecb_const int ecb_popcount64 (uint64_t x);
506ecb_function_ int 560ecb_function_ ecb_const int
507ecb_popcount64 (uint64_t x) 561ecb_popcount64 (uint64_t x)
508{ 562{
509 return ecb_popcount32 (x) + ecb_popcount32 (x >> 32); 563 return ecb_popcount32 (x) + ecb_popcount32 (x >> 32);
510} 564}
511 565
512ecb_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);
513ecb_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);
514ecb_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);
515ecb_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);
516ecb_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);
517ecb_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);
518ecb_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);
519ecb_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);
520 574
521ecb_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); }
522ecb_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); }
523ecb_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); }
524ecb_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); }
525ecb_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); }
526ecb_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); }
527ecb_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); }
528ecb_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); }
529 583
530#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
531 #define ecb_bswap16(x) (__builtin_bswap32 (x) >> 16) 588 #define ecb_bswap16(x) (__builtin_bswap32 (x) >> 16)
589 #endif
532 #define ecb_bswap32(x) __builtin_bswap32 (x) 590 #define ecb_bswap32(x) __builtin_bswap32 (x)
533 #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)))
534#else 597#else
535 ecb_function_ uint16_t ecb_bswap16 (uint16_t x) ecb_const; 598 ecb_function_ ecb_const uint16_t ecb_bswap16 (uint16_t x);
536 ecb_function_ uint16_t 599 ecb_function_ ecb_const uint16_t
537 ecb_bswap16 (uint16_t x) 600 ecb_bswap16 (uint16_t x)
538 { 601 {
539 return ecb_rotl16 (x, 8); 602 return ecb_rotl16 (x, 8);
540 } 603 }
541 604
542 ecb_function_ uint32_t ecb_bswap32 (uint32_t x) ecb_const; 605 ecb_function_ ecb_const uint32_t ecb_bswap32 (uint32_t x);
543 ecb_function_ uint32_t 606 ecb_function_ ecb_const uint32_t
544 ecb_bswap32 (uint32_t x) 607 ecb_bswap32 (uint32_t x)
545 { 608 {
546 return (((uint32_t)ecb_bswap16 (x)) << 16) | ecb_bswap16 (x >> 16); 609 return (((uint32_t)ecb_bswap16 (x)) << 16) | ecb_bswap16 (x >> 16);
547 } 610 }
548 611
549 ecb_function_ uint64_t ecb_bswap64 (uint64_t x) ecb_const; 612 ecb_function_ ecb_const uint64_t ecb_bswap64 (uint64_t x);
550 ecb_function_ uint64_t 613 ecb_function_ ecb_const uint64_t
551 ecb_bswap64 (uint64_t x) 614 ecb_bswap64 (uint64_t x)
552 { 615 {
553 return (((uint64_t)ecb_bswap32 (x)) << 32) | ecb_bswap32 (x >> 32); 616 return (((uint64_t)ecb_bswap32 (x)) << 32) | ecb_bswap32 (x >> 32);
554 } 617 }
555#endif 618#endif
556 619
557#if ECB_GCC_VERSION(4,5) || ECB_CLANG_BUILTIN(__builtin_unreachable) 620#if ECB_GCC_VERSION(4,5) || ECB_CLANG_BUILTIN(__builtin_unreachable)
558 #define ecb_unreachable() __builtin_unreachable () 621 #define ecb_unreachable() __builtin_unreachable ()
559#else 622#else
560 /* 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 :/ */
561 ecb_inline void ecb_unreachable (void) ecb_noreturn; 624 ecb_inline ecb_noreturn void ecb_unreachable (void);
562 ecb_inline void ecb_unreachable (void) { } 625 ecb_inline ecb_noreturn void ecb_unreachable (void) { }
563#endif 626#endif
564 627
565/* try to tell the compiler that some condition is definitely true */ 628/* try to tell the compiler that some condition is definitely true */
566#define ecb_assume(cond) if (!(cond)) ecb_unreachable (); else 0 629#define ecb_assume(cond) if (!(cond)) ecb_unreachable (); else 0
567 630
568ecb_inline unsigned char ecb_byteorder_helper (void) ecb_const; 631ecb_inline ecb_const uint32_t ecb_byteorder_helper (void);
569ecb_inline unsigned char 632ecb_inline ecb_const uint32_t
570ecb_byteorder_helper (void) 633ecb_byteorder_helper (void)
571{ 634{
572 /* the union code still generates code under pressure in gcc, */ 635 /* the union code still generates code under pressure in gcc, */
573 /* but less than using pointers, and always seems to */ 636 /* but less than using pointers, and always seems to */
574 /* successfully return a constant. */ 637 /* successfully return a constant. */
575 /* the reason why we have this horrible preprocessor mess */ 638 /* the reason why we have this horrible preprocessor mess */
576 /* 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 */
577 /* or when using a recent enough gcc version (>= 4.6) */ 640 /* 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__ 641#if (defined __BYTE_ORDER__ && __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__) \
642 || ((__i386 || __i386__ || _M_IX86 || ECB_GCC_AMD64 || ECB_MSVC_AMD64) && !__VOS__)
643 #define ECB_LITTLE_ENDIAN 1
581 return 0x44; 644 return 0x44332211;
582#elif __BYTE_ORDER__ && __BYTE_ORDER__ == __ORDER_BIG_ENDIAN__ 645#elif (defined __BYTE_ORDER__ && __BYTE_ORDER__ == __ORDER_BIG_ENDIAN__) \
646 || ((__AARCH64EB__ || __MIPSEB__ || __ARMEB__) && !__VOS__)
647 #define ECB_BIG_ENDIAN 1
583 return 0x11; 648 return 0x11223344;
584#else 649#else
585 union 650 union
586 { 651 {
652 uint8_t c[4];
587 uint32_t i; 653 uint32_t u;
588 uint8_t c;
589 } u = { 0x11223344 }; 654 } u = { 0x11, 0x22, 0x33, 0x44 };
590 return u.c; 655 return u.u;
591#endif 656#endif
592} 657}
593 658
594ecb_inline ecb_bool ecb_big_endian (void) ecb_const; 659ecb_inline ecb_const ecb_bool ecb_big_endian (void);
595ecb_inline ecb_bool ecb_big_endian (void) { return ecb_byteorder_helper () == 0x11; } 660ecb_inline ecb_const ecb_bool ecb_big_endian (void) { return ecb_byteorder_helper () == 0x11223344; }
596ecb_inline ecb_bool ecb_little_endian (void) ecb_const; 661ecb_inline ecb_const ecb_bool ecb_little_endian (void);
597ecb_inline ecb_bool ecb_little_endian (void) { return ecb_byteorder_helper () == 0x44; } 662ecb_inline ecb_const ecb_bool ecb_little_endian (void) { return ecb_byteorder_helper () == 0x44332211; }
598 663
599#if ECB_GCC_VERSION(3,0) || ECB_C99 664#if ECB_GCC_VERSION(3,0) || ECB_C99
600 #define ecb_mod(m,n) ((m) % (n) + ((m) % (n) < 0 ? (n) : 0)) 665 #define ecb_mod(m,n) ((m) % (n) + ((m) % (n) < 0 ? (n) : 0))
601#else 666#else
602 #define ecb_mod(m,n) ((m) < 0 ? ((n) - 1 - ((-1 - (m)) % (n))) : ((m) % (n))) 667 #define ecb_mod(m,n) ((m) < 0 ? ((n) - 1 - ((-1 - (m)) % (n))) : ((m) % (n)))
603#endif 668#endif
604 669
605#if __cplusplus 670#if ECB_CPP
606 template<typename T> 671 template<typename T>
607 static inline T ecb_div_rd (T val, T div) 672 static inline T ecb_div_rd (T val, T div)
608 { 673 {
609 return val < 0 ? - ((-val + div - 1) / div) : (val ) / div; 674 return val < 0 ? - ((-val + div - 1) / div) : (val ) / div;
610 } 675 }
… …
627 } 692 }
628#else 693#else
629 #define ecb_array_length(name) (sizeof (name) / sizeof (name [0])) 694 #define ecb_array_length(name) (sizeof (name) / sizeof (name [0]))
630#endif 695#endif
631 696
697ecb_function_ ecb_const uint32_t ecb_binary16_to_binary32 (uint32_t x);
698ecb_function_ ecb_const uint32_t
699ecb_binary16_to_binary32 (uint32_t x)
700{
701 unsigned int s = (x & 0x8000) << (31 - 15);
702 int e = (x >> 10) & 0x001f;
703 unsigned int m = x & 0x03ff;
704
705 if (ecb_expect_false (e == 31))
706 /* infinity or NaN */
707 e = 255 - (127 - 15);
708 else if (ecb_expect_false (!e))
709 {
710 if (ecb_expect_true (!m))
711 /* zero, handled by code below by forcing e to 0 */
712 e = 0 - (127 - 15);
713 else
714 {
715 /* subnormal, renormalise */
716 unsigned int s = 10 - ecb_ld32 (m);
717
718 m = (m << s) & 0x3ff; /* mask implicit bit */
719 e -= s - 1;
720 }
721 }
722
723 /* e and m now are normalised, or zero, (or inf or nan) */
724 e += 127 - 15;
725
726 return s | (e << 23) | (m << (23 - 10));
727}
728
729ecb_function_ ecb_const uint16_t ecb_binary32_to_binary16 (uint32_t x);
730ecb_function_ ecb_const uint16_t
731ecb_binary32_to_binary16 (uint32_t x)
732{
733 unsigned int s = (x >> 16) & 0x00008000; /* sign bit, the easy part */
734 unsigned int e = ((x >> 23) & 0x000000ff) - (127 - 15); /* the desired exponent */
735 unsigned int m = x & 0x007fffff;
736
737 x &= 0x7fffffff;
738
739 /* if it's within range of binary16 normals, use fast path */
740 if (ecb_expect_true (0x38800000 <= x && x <= 0x477fefff))
741 {
742 /* mantissa round-to-even */
743 m += 0x00000fff + ((m >> (23 - 10)) & 1);
744
745 /* handle overflow */
746 if (ecb_expect_false (m >= 0x00800000))
747 {
748 m >>= 1;
749 e += 1;
750 }
751
752 return s | (e << 10) | (m >> (23 - 10));
753 }
754
755 /* handle large numbers and infinity */
756 if (ecb_expect_true (0x477fefff < x && x <= 0x7f800000))
757 return s | 0x7c00;
758
759 /* handle zero, subnormals and small numbers */
760 if (ecb_expect_true (x < 0x38800000))
761 {
762 /* zero */
763 if (ecb_expect_true (!x))
764 return s;
765
766 /* handle subnormals */
767
768 /* too small, will be zero */
769 if (e < (14 - 24)) /* might not be sharp, but is good enough */
770 return s;
771
772 m |= 0x00800000; /* make implicit bit explicit */
773
774 /* very tricky - we need to round to the nearest e (+10) bit value */
775 {
776 unsigned int bits = 14 - e;
777 unsigned int half = (1 << (bits - 1)) - 1;
778 unsigned int even = (m >> bits) & 1;
779
780 /* if this overflows, we will end up with a normalised number */
781 m = (m + half + even) >> bits;
782 }
783
784 return s | m;
785 }
786
787 /* handle NaNs, preserve leftmost nan bits, but make sure we don't turn them into infinities */
788 m >>= 13;
789
790 return s | 0x7c00 | m | !m;
791}
792
632/*******************************************************************************/ 793/*******************************************************************************/
633/* floating point stuff, can be disabled by defining ECB_NO_LIBM */ 794/* floating point stuff, can be disabled by defining ECB_NO_LIBM */
634 795
635/* basically, everything uses "ieee pure-endian" floating point numbers */ 796/* basically, everything uses "ieee pure-endian" floating point numbers */
636/* the only noteworthy exception is ancient armle, which uses order 43218765 */ 797/* the only noteworthy exception is ancient armle, which uses order 43218765 */
637#if 0 \ 798#if 0 \
638 || __i386 || __i386__ \ 799 || __i386 || __i386__ \
639 || __amd64 || __amd64__ || __x86_64 || __x86_64__ \ 800 || ECB_GCC_AMD64 \
640 || __powerpc__ || __ppc__ || __powerpc64__ || __ppc64__ \ 801 || __powerpc__ || __ppc__ || __powerpc64__ || __ppc64__ \
641 || defined __s390__ || defined __s390x__ \ 802 || defined __s390__ || defined __s390x__ \
642 || defined __mips__ \ 803 || defined __mips__ \
643 || defined __alpha__ \ 804 || defined __alpha__ \
644 || defined __hppa__ \ 805 || defined __hppa__ \
645 || defined __ia64__ \ 806 || defined __ia64__ \
646 || defined __m68k__ \ 807 || defined __m68k__ \
647 || defined __m88k__ \ 808 || defined __m88k__ \
648 || defined __sh__ \ 809 || defined __sh__ \
649 || defined _M_IX86 || defined _M_AMD64 || defined _M_IA64 \ 810 || 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__)) \ 811 || (defined __arm__ && (defined __ARM_EABI__ || defined __EABI__ || defined __VFP_FP__ || defined _WIN32_WCE || defined __ANDROID__)) \
651 || defined __aarch64__ 812 || defined __aarch64__
652 #define ECB_STDFP 1 813 #define ECB_STDFP 1
653 #include <string.h> /* for memcpy */ 814 #include <string.h> /* for memcpy */
654#else 815#else
… …
672 #define ECB_NAN ECB_INFINITY 833 #define ECB_NAN ECB_INFINITY
673 #endif 834 #endif
674 835
675 #if ECB_C99 || _XOPEN_VERSION >= 600 || _POSIX_VERSION >= 200112L 836 #if ECB_C99 || _XOPEN_VERSION >= 600 || _POSIX_VERSION >= 200112L
676 #define ecb_ldexpf(x,e) ldexpf ((x), (e)) 837 #define ecb_ldexpf(x,e) ldexpf ((x), (e))
838 #define ecb_frexpf(x,e) frexpf ((x), (e))
677 #else 839 #else
678 #define ecb_ldexpf(x,e) (float) ldexp ((double)(x), (e)) 840 #define ecb_ldexpf(x,e) (float) ldexp ((double) (x), (e))
841 #define ecb_frexpf(x,e) (float) frexp ((double) (x), (e))
679 #endif 842 #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 }
697 843
698 /* convert a float to ieee single/binary32 */ 844 /* convert a float to ieee single/binary32 */
699 ecb_function_ uint32_t ecb_float_to_binary32 (float x) ecb_const; 845 ecb_function_ ecb_const uint32_t ecb_float_to_binary32 (float x);
700 ecb_function_ uint32_t 846 ecb_function_ ecb_const uint32_t
701 ecb_float_to_binary32 (float x) 847 ecb_float_to_binary32 (float x)
702 { 848 {
703 uint32_t r; 849 uint32_t r;
704 850
705 #if ECB_STDFP 851 #if ECB_STDFP
… …
712 if (x == 0e0f ) return 0x00000000U; 858 if (x == 0e0f ) return 0x00000000U;
713 if (x > +3.40282346638528860e+38f) return 0x7f800000U; 859 if (x > +3.40282346638528860e+38f) return 0x7f800000U;
714 if (x < -3.40282346638528860e+38f) return 0xff800000U; 860 if (x < -3.40282346638528860e+38f) return 0xff800000U;
715 if (x != x ) return 0x7fbfffffU; 861 if (x != x ) return 0x7fbfffffU;
716 862
717 m = frexpf (x, &e) * 0x1000000U; 863 m = ecb_frexpf (x, &e) * 0x1000000U;
718 864
719 r = m & 0x80000000U; 865 r = m & 0x80000000U;
720 866
721 if (r) 867 if (r)
722 m = -m; 868 m = -m;
… …
734 880
735 return r; 881 return r;
736 } 882 }
737 883
738 /* converts an ieee single/binary32 to a float */ 884 /* converts an ieee single/binary32 to a float */
739 ecb_function_ float ecb_binary32_to_float (uint32_t x) ecb_const; 885 ecb_function_ ecb_const float ecb_binary32_to_float (uint32_t x);
740 ecb_function_ float 886 ecb_function_ ecb_const float
741 ecb_binary32_to_float (uint32_t x) 887 ecb_binary32_to_float (uint32_t x)
742 { 888 {
743 float r; 889 float r;
744 890
745 #if ECB_STDFP 891 #if ECB_STDFP
… …
764 910
765 return r; 911 return r;
766 } 912 }
767 913
768 /* convert a double to ieee double/binary64 */ 914 /* convert a double to ieee double/binary64 */
769 ecb_function_ uint64_t ecb_double_to_binary64 (double x) ecb_const; 915 ecb_function_ ecb_const uint64_t ecb_double_to_binary64 (double x);
770 ecb_function_ uint64_t 916 ecb_function_ ecb_const uint64_t
771 ecb_double_to_binary64 (double x) 917 ecb_double_to_binary64 (double x)
772 { 918 {
773 uint64_t r; 919 uint64_t r;
774 920
775 #if ECB_STDFP 921 #if ECB_STDFP
… …
804 950
805 return r; 951 return r;
806 } 952 }
807 953
808 /* converts an ieee double/binary64 to a double */ 954 /* converts an ieee double/binary64 to a double */
809 ecb_function_ double ecb_binary64_to_double (uint64_t x) ecb_const; 955 ecb_function_ ecb_const double ecb_binary64_to_double (uint64_t x);
810 ecb_function_ double 956 ecb_function_ ecb_const double
811 ecb_binary64_to_double (uint64_t x) 957 ecb_binary64_to_double (uint64_t x)
812 { 958 {
813 double r; 959 double r;
814 960
815 #if ECB_STDFP 961 #if ECB_STDFP
… …
833 #endif 979 #endif
834 980
835 return r; 981 return r;
836 } 982 }
837 983
838#endif 984 /* convert a float to ieee half/binary16 */
985 ecb_function_ ecb_const uint16_t ecb_float_to_binary16 (float x);
986 ecb_function_ ecb_const uint16_t
987 ecb_float_to_binary16 (float x)
988 {
989 return ecb_binary32_to_binary16 (ecb_float_to_binary32 (x));
990 }
839 991
840#endif 992 /* convert an ieee half/binary16 to float */
993 ecb_function_ ecb_const float ecb_binary16_to_float (uint16_t x);
994 ecb_function_ ecb_const float
995 ecb_binary16_to_float (uint16_t x)
996 {
997 return ecb_binary32_to_float (ecb_binary16_to_binary32 (x));
998 }
841 999
1000#endif
1001
1002#endif
1003

Diff Legend

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