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

Comparing libeio/ecb.h (file contents):
Revision 1.24 by root, Sat Jun 13 00:06:13 2015 UTC vs.
Revision 1.28 by root, Thu Nov 29 19:47:33 2018 UTC

… …
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
… …
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)
116#define ECB_CPP11 (__cplusplus >= 201103L) 119#define ECB_CPP11 (__cplusplus >= 201103L)
120#define ECB_CPP14 (__cplusplus >= 201402L)
121#define ECB_CPP17 (__cplusplus >= 201703L)
117 122
118#if ECB_CPP 123#if ECB_CPP
119 #define ECB_C 0 124 #define ECB_C 0
120 #define ECB_STDC_VERSION 0 125 #define ECB_STDC_VERSION 0
121#else 126#else
… …
123 #define ECB_STDC_VERSION __STDC_VERSION__ 128 #define ECB_STDC_VERSION __STDC_VERSION__
124#endif 129#endif
125 130
126#define ECB_C99 (ECB_STDC_VERSION >= 199901L) 131#define ECB_C99 (ECB_STDC_VERSION >= 199901L)
127#define ECB_C11 (ECB_STDC_VERSION >= 201112L) 132#define ECB_C11 (ECB_STDC_VERSION >= 201112L)
133#define ECB_C17 (ECB_STDC_VERSION >= 201710L)
128 134
129#if ECB_CPP 135#if ECB_CPP
130 #define ECB_EXTERN_C extern "C" 136 #define ECB_EXTERN_C extern "C"
131 #define ECB_EXTERN_C_BEG ECB_EXTERN_C { 137 #define ECB_EXTERN_C_BEG ECB_EXTERN_C {
132 #define ECB_EXTERN_C_END } 138 #define ECB_EXTERN_C_END }
… …
145 #define ECB_NO_SMP 1 151 #define ECB_NO_SMP 1
146#endif 152#endif
147 153
148#if ECB_NO_SMP 154#if ECB_NO_SMP
149 #define ECB_MEMORY_FENCE do { } while (0) 155 #define ECB_MEMORY_FENCE do { } while (0)
156#endif
157
158/* http://www-01.ibm.com/support/knowledgecenter/SSGH3R_13.1.0/com.ibm.xlcpp131.aix.doc/compiler_ref/compiler_builtins.html */
159#if __xlC__ && ECB_CPP
160 #include <builtins.h>
161#endif
162
163#if 1400 <= _MSC_VER
164 #include <intrin.h> /* fence functions _ReadBarrier, also bit search functions _BitScanReverse */
150#endif 165#endif
151 166
152#ifndef ECB_MEMORY_FENCE 167#ifndef ECB_MEMORY_FENCE
153 #if ECB_GCC_VERSION(2,5) || defined __INTEL_COMPILER || (__llvm__ && __GNUC__) || __SUNPRO_C >= 0x5110 || __SUNPRO_CC >= 0x5110 168 #if ECB_GCC_VERSION(2,5) || defined __INTEL_COMPILER || (__llvm__ && __GNUC__) || __SUNPRO_C >= 0x5110 || __SUNPRO_CC >= 0x5110
154 #if __i386 || __i386__ 169 #if __i386 || __i386__
155 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("lock; orb $0, -1(%%esp)" : : : "memory") 170 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("lock; orb $0, -1(%%esp)" : : : "memory")
156 #define ECB_MEMORY_FENCE_ACQUIRE __asm__ __volatile__ ("" : : : "memory") 171 #define ECB_MEMORY_FENCE_ACQUIRE __asm__ __volatile__ ("" : : : "memory")
157 #define ECB_MEMORY_FENCE_RELEASE __asm__ __volatile__ ("") 172 #define ECB_MEMORY_FENCE_RELEASE __asm__ __volatile__ ("" : : : "memory")
158 #elif __amd64 || __amd64__ || __x86_64 || __x86_64__ 173 #elif ECB_GCC_AMD64
159 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("mfence" : : : "memory") 174 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("mfence" : : : "memory")
160 #define ECB_MEMORY_FENCE_ACQUIRE __asm__ __volatile__ ("" : : : "memory") 175 #define ECB_MEMORY_FENCE_ACQUIRE __asm__ __volatile__ ("" : : : "memory")
161 #define ECB_MEMORY_FENCE_RELEASE __asm__ __volatile__ ("") 176 #define ECB_MEMORY_FENCE_RELEASE __asm__ __volatile__ ("" : : : "memory")
162 #elif __powerpc__ || __ppc__ || __powerpc64__ || __ppc64__ 177 #elif __powerpc__ || __ppc__ || __powerpc64__ || __ppc64__
163 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("sync" : : : "memory") 178 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("sync" : : : "memory")
179 #elif defined __ARM_ARCH_2__ \
180 || defined __ARM_ARCH_3__ || defined __ARM_ARCH_3M__ \
181 || defined __ARM_ARCH_4__ || defined __ARM_ARCH_4T__ \
182 || defined __ARM_ARCH_5__ || defined __ARM_ARCH_5E__ \
183 || defined __ARM_ARCH_5T__ || defined __ARM_ARCH_5TE__ \
184 || defined __ARM_ARCH_5TEJ__
185 /* should not need any, unless running old code on newer cpu - arm doesn't support that */
164 #elif defined __ARM_ARCH_6__ || defined __ARM_ARCH_6J__ \ 186 #elif defined __ARM_ARCH_6__ || defined __ARM_ARCH_6J__ \
165 || defined __ARM_ARCH_6K__ || defined __ARM_ARCH_6ZK__ 187 || defined __ARM_ARCH_6K__ || defined __ARM_ARCH_6ZK__ \
188 || defined __ARM_ARCH_6T2__
166 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("mcr p15,0,%0,c7,c10,5" : : "r" (0) : "memory") 189 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("mcr p15,0,%0,c7,c10,5" : : "r" (0) : "memory")
167 #elif defined __ARM_ARCH_7__ || defined __ARM_ARCH_7A__ \ 190 #elif defined __ARM_ARCH_7__ || defined __ARM_ARCH_7A__ \
168 || defined __ARM_ARCH_7M__ || defined __ARM_ARCH_7R__ 191 || defined __ARM_ARCH_7R__ || defined __ARM_ARCH_7M__
169 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("dmb" : : : "memory") 192 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("dmb" : : : "memory")
170 #elif __aarch64__ 193 #elif __aarch64__
171 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("dmb ish" : : : "memory") 194 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("dmb ish" : : : "memory")
172 #elif (__sparc || __sparc__) && !__sparcv8 195 #elif (__sparc || __sparc__) && !(__sparc_v8__ || defined __sparcv8)
173 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("membar #LoadStore | #LoadLoad | #StoreStore | #StoreLoad" : : : "memory") 196 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("membar #LoadStore | #LoadLoad | #StoreStore | #StoreLoad" : : : "memory")
174 #define ECB_MEMORY_FENCE_ACQUIRE __asm__ __volatile__ ("membar #LoadStore | #LoadLoad" : : : "memory") 197 #define ECB_MEMORY_FENCE_ACQUIRE __asm__ __volatile__ ("membar #LoadStore | #LoadLoad" : : : "memory")
175 #define ECB_MEMORY_FENCE_RELEASE __asm__ __volatile__ ("membar #LoadStore | #StoreStore") 198 #define ECB_MEMORY_FENCE_RELEASE __asm__ __volatile__ ("membar #LoadStore | #StoreStore")
176 #elif defined __s390__ || defined __s390x__ 199 #elif defined __s390__ || defined __s390x__
177 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("bcr 15,0" : : : "memory") 200 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("bcr 15,0" : : : "memory")
… …
349 #define ecb_deprecated __declspec (deprecated) 372 #define ecb_deprecated __declspec (deprecated)
350#else 373#else
351 #define ecb_deprecated ecb_attribute ((__deprecated__)) 374 #define ecb_deprecated ecb_attribute ((__deprecated__))
352#endif 375#endif
353 376
354#if __MSC_VER >= 1500 377#if _MSC_VER >= 1500
355 #define ecb_deprecated_message(msg) __declspec (deprecated (msg)) 378 #define ecb_deprecated_message(msg) __declspec (deprecated (msg))
356#elif ECB_GCC_VERSION(4,5) 379#elif ECB_GCC_VERSION(4,5)
357 #define ecb_deprecated_message(msg) ecb_attribute ((__deprecated__ (msg)) 380 #define ecb_deprecated_message(msg) ecb_attribute ((__deprecated__ (msg))
358#else 381#else
359 #define ecb_deprecated_message(msg) ecb_deprecated 382 #define ecb_deprecated_message(msg) ecb_deprecated
… …
367 390
368#define ecb_unused ecb_attribute ((__unused__)) 391#define ecb_unused ecb_attribute ((__unused__))
369#define ecb_const ecb_attribute ((__const__)) 392#define ecb_const ecb_attribute ((__const__))
370#define ecb_pure ecb_attribute ((__pure__)) 393#define ecb_pure ecb_attribute ((__pure__))
371 394
372/* TODO http://msdn.microsoft.com/en-us/library/k6ktzx3s.aspx __declspec(noreturn) */
373#if ECB_C11 || __IBMC_NORETURN 395#if ECB_C11 || __IBMC_NORETURN
374 /* http://pic.dhe.ibm.com/infocenter/compbg/v121v141/topic/com.ibm.xlcpp121.bg.doc/language_ref/noreturn.html */ 396 /* http://www-01.ibm.com/support/knowledgecenter/SSGH3R_13.1.0/com.ibm.xlcpp131.aix.doc/language_ref/noreturn.html */
375 #define ecb_noreturn _Noreturn 397 #define ecb_noreturn _Noreturn
376#elif ECB_CPP11 398#elif ECB_CPP11
377 #define ecb_noreturn [[noreturn]] 399 #define ecb_noreturn [[noreturn]]
378#elif _MSC_VER >= 1200 400#elif _MSC_VER >= 1200
401 /* http://msdn.microsoft.com/en-us/library/k6ktzx3s.aspx */
379 #define ecb_noreturn __declspec (noreturn) 402 #define ecb_noreturn __declspec (noreturn)
380#else 403#else
381 #define ecb_noreturn ecb_attribute ((__noreturn__)) 404 #define ecb_noreturn ecb_attribute ((__noreturn__))
382#endif 405#endif
383 406
… …
415#else 438#else
416 ecb_function_ ecb_const int ecb_ctz32 (uint32_t x); 439 ecb_function_ ecb_const int ecb_ctz32 (uint32_t x);
417 ecb_function_ ecb_const int 440 ecb_function_ ecb_const int
418 ecb_ctz32 (uint32_t x) 441 ecb_ctz32 (uint32_t x)
419 { 442 {
443#if 1400 <= _MSC_VER && (_M_IX86 || _M_X64 || _M_IA64 || _M_ARM)
444 unsigned long r;
445 _BitScanForward (&r, x);
446 return (int)r;
447#else
420 int r = 0; 448 int r = 0;
421 449
422 x &= ~x + 1; /* this isolates the lowest bit */ 450 x &= ~x + 1; /* this isolates the lowest bit */
423 451
424#if ECB_branchless_on_i386 452#if ECB_branchless_on_i386
… …
434 if (x & 0xff00ff00) r += 8; 462 if (x & 0xff00ff00) r += 8;
435 if (x & 0xffff0000) r += 16; 463 if (x & 0xffff0000) r += 16;
436#endif 464#endif
437 465
438 return r; 466 return r;
467#endif
439 } 468 }
440 469
441 ecb_function_ ecb_const int ecb_ctz64 (uint64_t x); 470 ecb_function_ ecb_const int ecb_ctz64 (uint64_t x);
442 ecb_function_ ecb_const int 471 ecb_function_ ecb_const int
443 ecb_ctz64 (uint64_t x) 472 ecb_ctz64 (uint64_t x)
444 { 473 {
474#if 1400 <= _MSC_VER && (_M_X64 || _M_IA64 || _M_ARM)
475 unsigned long r;
476 _BitScanForward64 (&r, x);
477 return (int)r;
478#else
445 int shift = x & 0xffffffffU ? 0 : 32; 479 int shift = x & 0xffffffff ? 0 : 32;
446 return ecb_ctz32 (x >> shift) + shift; 480 return ecb_ctz32 (x >> shift) + shift;
481#endif
447 } 482 }
448 483
449 ecb_function_ ecb_const int ecb_popcount32 (uint32_t x); 484 ecb_function_ ecb_const int ecb_popcount32 (uint32_t x);
450 ecb_function_ ecb_const int 485 ecb_function_ ecb_const int
451 ecb_popcount32 (uint32_t x) 486 ecb_popcount32 (uint32_t x)
… …
459 } 494 }
460 495
461 ecb_function_ ecb_const int ecb_ld32 (uint32_t x); 496 ecb_function_ ecb_const int ecb_ld32 (uint32_t x);
462 ecb_function_ ecb_const int ecb_ld32 (uint32_t x) 497 ecb_function_ ecb_const int ecb_ld32 (uint32_t x)
463 { 498 {
499#if 1400 <= _MSC_VER && (_M_IX86 || _M_X64 || _M_IA64 || _M_ARM)
500 unsigned long r;
501 _BitScanReverse (&r, x);
502 return (int)r;
503#else
464 int r = 0; 504 int r = 0;
465 505
466 if (x >> 16) { x >>= 16; r += 16; } 506 if (x >> 16) { x >>= 16; r += 16; }
467 if (x >> 8) { x >>= 8; r += 8; } 507 if (x >> 8) { x >>= 8; r += 8; }
468 if (x >> 4) { x >>= 4; r += 4; } 508 if (x >> 4) { x >>= 4; r += 4; }
469 if (x >> 2) { x >>= 2; r += 2; } 509 if (x >> 2) { x >>= 2; r += 2; }
470 if (x >> 1) { r += 1; } 510 if (x >> 1) { r += 1; }
471 511
472 return r; 512 return r;
513#endif
473 } 514 }
474 515
475 ecb_function_ ecb_const int ecb_ld64 (uint64_t x); 516 ecb_function_ ecb_const int ecb_ld64 (uint64_t x);
476 ecb_function_ ecb_const int ecb_ld64 (uint64_t x) 517 ecb_function_ ecb_const int ecb_ld64 (uint64_t x)
477 { 518 {
519#if 1400 <= _MSC_VER && (_M_X64 || _M_IA64 || _M_ARM)
520 unsigned long r;
521 _BitScanReverse64 (&r, x);
522 return (int)r;
523#else
478 int r = 0; 524 int r = 0;
479 525
480 if (x >> 32) { x >>= 32; r += 32; } 526 if (x >> 32) { x >>= 32; r += 32; }
481 527
482 return r + ecb_ld32 (x); 528 return r + ecb_ld32 (x);
529#endif
483 } 530 }
484#endif 531#endif
485 532
486ecb_function_ ecb_const ecb_bool ecb_is_pot32 (uint32_t x); 533ecb_function_ ecb_const ecb_bool ecb_is_pot32 (uint32_t x);
487ecb_function_ ecb_const ecb_bool ecb_is_pot32 (uint32_t x) { return !(x & (x - 1)); } 534ecb_function_ ecb_const ecb_bool ecb_is_pot32 (uint32_t x) { return !(x & (x - 1)); }
… …
544ecb_inline ecb_const uint32_t ecb_rotr32 (uint32_t x, unsigned int count) { return (x << (32 - count)) | (x >> count); } 591ecb_inline ecb_const uint32_t ecb_rotr32 (uint32_t x, unsigned int count) { return (x << (32 - count)) | (x >> count); }
545ecb_inline ecb_const uint64_t ecb_rotl64 (uint64_t x, unsigned int count) { return (x >> (64 - count)) | (x << count); } 592ecb_inline ecb_const uint64_t ecb_rotl64 (uint64_t x, unsigned int count) { return (x >> (64 - count)) | (x << count); }
546ecb_inline ecb_const uint64_t ecb_rotr64 (uint64_t x, unsigned int count) { return (x << (64 - count)) | (x >> count); } 593ecb_inline ecb_const uint64_t ecb_rotr64 (uint64_t x, unsigned int count) { return (x << (64 - count)) | (x >> count); }
547 594
548#if ECB_GCC_VERSION(4,3) || (ECB_CLANG_BUILTIN(__builtin_bswap32) && ECB_CLANG_BUILTIN(__builtin_bswap64)) 595#if ECB_GCC_VERSION(4,3) || (ECB_CLANG_BUILTIN(__builtin_bswap32) && ECB_CLANG_BUILTIN(__builtin_bswap64))
596 #if ECB_GCC_VERSION(4,8) || ECB_CLANG_BUILTIN(__builtin_bswap16)
597 #define ecb_bswap16(x) __builtin_bswap16 (x)
598 #else
549 #define ecb_bswap16(x) (__builtin_bswap32 (x) >> 16) 599 #define ecb_bswap16(x) (__builtin_bswap32 (x) >> 16)
600 #endif
550 #define ecb_bswap32(x) __builtin_bswap32 (x) 601 #define ecb_bswap32(x) __builtin_bswap32 (x)
551 #define ecb_bswap64(x) __builtin_bswap64 (x) 602 #define ecb_bswap64(x) __builtin_bswap64 (x)
603#elif _MSC_VER
604 #include <stdlib.h>
605 #define ecb_bswap16(x) ((uint16_t)_byteswap_ushort ((uint16_t)(x)))
606 #define ecb_bswap32(x) ((uint32_t)_byteswap_ulong ((uint32_t)(x)))
607 #define ecb_bswap64(x) ((uint64_t)_byteswap_uint64 ((uint64_t)(x)))
552#else 608#else
553 ecb_function_ ecb_const uint16_t ecb_bswap16 (uint16_t x); 609 ecb_function_ ecb_const uint16_t ecb_bswap16 (uint16_t x);
554 ecb_function_ ecb_const uint16_t 610 ecb_function_ ecb_const uint16_t
555 ecb_bswap16 (uint16_t x) 611 ecb_bswap16 (uint16_t x)
556 { 612 {
… …
581#endif 637#endif
582 638
583/* try to tell the compiler that some condition is definitely true */ 639/* try to tell the compiler that some condition is definitely true */
584#define ecb_assume(cond) if (!(cond)) ecb_unreachable (); else 0 640#define ecb_assume(cond) if (!(cond)) ecb_unreachable (); else 0
585 641
586ecb_inline ecb_const unsigned char ecb_byteorder_helper (void); 642ecb_inline ecb_const uint32_t ecb_byteorder_helper (void);
587ecb_inline ecb_const unsigned char 643ecb_inline ecb_const uint32_t
588ecb_byteorder_helper (void) 644ecb_byteorder_helper (void)
589{ 645{
590 /* the union code still generates code under pressure in gcc, */ 646 /* the union code still generates code under pressure in gcc, */
591 /* but less than using pointers, and always seems to */ 647 /* but less than using pointers, and always seems to */
592 /* successfully return a constant. */ 648 /* successfully return a constant. */
593 /* the reason why we have this horrible preprocessor mess */ 649 /* the reason why we have this horrible preprocessor mess */
594 /* is to avoid it in all cases, at least on common architectures */ 650 /* is to avoid it in all cases, at least on common architectures */
595 /* or when using a recent enough gcc version (>= 4.6) */ 651 /* or when using a recent enough gcc version (>= 4.6) */
596#if __i386 || __i386__ || _M_X86 || __amd64 || __amd64__ || _M_X64
597 return 0x44;
598#elif __BYTE_ORDER__ && __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__ 652#if (defined __BYTE_ORDER__ && __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__) \
653 || ((__i386 || __i386__ || _M_IX86 || ECB_GCC_AMD64 || ECB_MSVC_AMD64) && !__VOS__)
654 #define ECB_LITTLE_ENDIAN 1
599 return 0x44; 655 return 0x44332211;
600#elif __BYTE_ORDER__ && __BYTE_ORDER__ == __ORDER_BIG_ENDIAN__ 656#elif (defined __BYTE_ORDER__ && __BYTE_ORDER__ == __ORDER_BIG_ENDIAN__) \
657 || ((__AARCH64EB__ || __MIPSEB__ || __ARMEB__) && !__VOS__)
658 #define ECB_BIG_ENDIAN 1
601 return 0x11; 659 return 0x11223344;
602#else 660#else
603 union 661 union
604 { 662 {
663 uint8_t c[4];
605 uint32_t i; 664 uint32_t u;
606 uint8_t c;
607 } u = { 0x11223344 }; 665 } u = { 0x11, 0x22, 0x33, 0x44 };
608 return u.c; 666 return u.u;
609#endif 667#endif
610} 668}
611 669
612ecb_inline ecb_const ecb_bool ecb_big_endian (void); 670ecb_inline ecb_const ecb_bool ecb_big_endian (void);
613ecb_inline ecb_const ecb_bool ecb_big_endian (void) { return ecb_byteorder_helper () == 0x11; } 671ecb_inline ecb_const ecb_bool ecb_big_endian (void) { return ecb_byteorder_helper () == 0x11223344; }
614ecb_inline ecb_const ecb_bool ecb_little_endian (void); 672ecb_inline ecb_const ecb_bool ecb_little_endian (void);
615ecb_inline ecb_const ecb_bool ecb_little_endian (void) { return ecb_byteorder_helper () == 0x44; } 673ecb_inline ecb_const ecb_bool ecb_little_endian (void) { return ecb_byteorder_helper () == 0x44332211; }
616 674
617#if ECB_GCC_VERSION(3,0) || ECB_C99 675#if ECB_GCC_VERSION(3,0) || ECB_C99
618 #define ecb_mod(m,n) ((m) % (n) + ((m) % (n) < 0 ? (n) : 0)) 676 #define ecb_mod(m,n) ((m) % (n) + ((m) % (n) < 0 ? (n) : 0))
619#else 677#else
620 #define ecb_mod(m,n) ((m) < 0 ? ((n) - 1 - ((-1 - (m)) % (n))) : ((m) % (n))) 678 #define ecb_mod(m,n) ((m) < 0 ? ((n) - 1 - ((-1 - (m)) % (n))) : ((m) % (n)))
… …
645 } 703 }
646#else 704#else
647 #define ecb_array_length(name) (sizeof (name) / sizeof (name [0])) 705 #define ecb_array_length(name) (sizeof (name) / sizeof (name [0]))
648#endif 706#endif
649 707
708ecb_function_ ecb_const uint32_t ecb_binary16_to_binary32 (uint32_t x);
709ecb_function_ ecb_const uint32_t
710ecb_binary16_to_binary32 (uint32_t x)
711{
712 unsigned int s = (x & 0x8000) << (31 - 15);
713 int e = (x >> 10) & 0x001f;
714 unsigned int m = x & 0x03ff;
715
716 if (ecb_expect_false (e == 31))
717 /* infinity or NaN */
718 e = 255 - (127 - 15);
719 else if (ecb_expect_false (!e))
720 {
721 if (ecb_expect_true (!m))
722 /* zero, handled by code below by forcing e to 0 */
723 e = 0 - (127 - 15);
724 else
725 {
726 /* subnormal, renormalise */
727 unsigned int s = 10 - ecb_ld32 (m);
728
729 m = (m << s) & 0x3ff; /* mask implicit bit */
730 e -= s - 1;
731 }
732 }
733
734 /* e and m now are normalised, or zero, (or inf or nan) */
735 e += 127 - 15;
736
737 return s | (e << 23) | (m << (23 - 10));
738}
739
740ecb_function_ ecb_const uint16_t ecb_binary32_to_binary16 (uint32_t x);
741ecb_function_ ecb_const uint16_t
742ecb_binary32_to_binary16 (uint32_t x)
743{
744 unsigned int s = (x >> 16) & 0x00008000; /* sign bit, the easy part */
745 unsigned int e = ((x >> 23) & 0x000000ff) - (127 - 15); /* the desired exponent */
746 unsigned int m = x & 0x007fffff;
747
748 x &= 0x7fffffff;
749
750 /* if it's within range of binary16 normals, use fast path */
751 if (ecb_expect_true (0x38800000 <= x && x <= 0x477fefff))
752 {
753 /* mantissa round-to-even */
754 m += 0x00000fff + ((m >> (23 - 10)) & 1);
755
756 /* handle overflow */
757 if (ecb_expect_false (m >= 0x00800000))
758 {
759 m >>= 1;
760 e += 1;
761 }
762
763 return s | (e << 10) | (m >> (23 - 10));
764 }
765
766 /* handle large numbers and infinity */
767 if (ecb_expect_true (0x477fefff < x && x <= 0x7f800000))
768 return s | 0x7c00;
769
770 /* handle zero, subnormals and small numbers */
771 if (ecb_expect_true (x < 0x38800000))
772 {
773 /* zero */
774 if (ecb_expect_true (!x))
775 return s;
776
777 /* handle subnormals */
778
779 /* too small, will be zero */
780 if (e < (14 - 24)) /* might not be sharp, but is good enough */
781 return s;
782
783 m |= 0x00800000; /* make implicit bit explicit */
784
785 /* very tricky - we need to round to the nearest e (+10) bit value */
786 {
787 unsigned int bits = 14 - e;
788 unsigned int half = (1 << (bits - 1)) - 1;
789 unsigned int even = (m >> bits) & 1;
790
791 /* if this overflows, we will end up with a normalised number */
792 m = (m + half + even) >> bits;
793 }
794
795 return s | m;
796 }
797
798 /* handle NaNs, preserve leftmost nan bits, but make sure we don't turn them into infinities */
799 m >>= 13;
800
801 return s | 0x7c00 | m | !m;
802}
803
650/*******************************************************************************/ 804/*******************************************************************************/
651/* floating point stuff, can be disabled by defining ECB_NO_LIBM */ 805/* floating point stuff, can be disabled by defining ECB_NO_LIBM */
652 806
653/* basically, everything uses "ieee pure-endian" floating point numbers */ 807/* basically, everything uses "ieee pure-endian" floating point numbers */
654/* the only noteworthy exception is ancient armle, which uses order 43218765 */ 808/* the only noteworthy exception is ancient armle, which uses order 43218765 */
655#if 0 \ 809#if 0 \
656 || __i386 || __i386__ \ 810 || __i386 || __i386__ \
657 || __amd64 || __amd64__ || __x86_64 || __x86_64__ \ 811 || ECB_GCC_AMD64 \
658 || __powerpc__ || __ppc__ || __powerpc64__ || __ppc64__ \ 812 || __powerpc__ || __ppc__ || __powerpc64__ || __ppc64__ \
659 || defined __s390__ || defined __s390x__ \ 813 || defined __s390__ || defined __s390x__ \
660 || defined __mips__ \ 814 || defined __mips__ \
661 || defined __alpha__ \ 815 || defined __alpha__ \
662 || defined __hppa__ \ 816 || defined __hppa__ \
663 || defined __ia64__ \ 817 || defined __ia64__ \
664 || defined __m68k__ \ 818 || defined __m68k__ \
665 || defined __m88k__ \ 819 || defined __m88k__ \
666 || defined __sh__ \ 820 || defined __sh__ \
667 || defined _M_IX86 || defined _M_AMD64 || defined _M_IA64 \ 821 || defined _M_IX86 || defined ECB_MSVC_AMD64 || defined _M_IA64 \
668 || (defined __arm__ && (defined __ARM_EABI__ || defined __EABI__ || defined __VFP_FP__ || defined _WIN32_WCE || defined __ANDROID__)) \ 822 || (defined __arm__ && (defined __ARM_EABI__ || defined __EABI__ || defined __VFP_FP__ || defined _WIN32_WCE || defined __ANDROID__)) \
669 || defined __aarch64__ 823 || defined __aarch64__
670 #define ECB_STDFP 1 824 #define ECB_STDFP 1
671 #include <string.h> /* for memcpy */ 825 #include <string.h> /* for memcpy */
672#else 826#else
… …
690 #define ECB_NAN ECB_INFINITY 844 #define ECB_NAN ECB_INFINITY
691 #endif 845 #endif
692 846
693 #if ECB_C99 || _XOPEN_VERSION >= 600 || _POSIX_VERSION >= 200112L 847 #if ECB_C99 || _XOPEN_VERSION >= 600 || _POSIX_VERSION >= 200112L
694 #define ecb_ldexpf(x,e) ldexpf ((x), (e)) 848 #define ecb_ldexpf(x,e) ldexpf ((x), (e))
849 #define ecb_frexpf(x,e) frexpf ((x), (e))
695 #else 850 #else
696 #define ecb_ldexpf(x,e) (float) ldexp ((x), (e)) 851 #define ecb_ldexpf(x,e) (float) ldexp ((double) (x), (e))
852 #define ecb_frexpf(x,e) (float) frexp ((double) (x), (e))
697 #endif 853 #endif
698
699 /* converts an ieee half/binary16 to a float */
700 ecb_function_ ecb_const float ecb_binary16_to_float (uint16_t x);
701 ecb_function_ ecb_const float
702 ecb_binary16_to_float (uint16_t x)
703 {
704 int e = (x >> 10) & 0x1f;
705 int m = x & 0x3ff;
706 float r;
707
708 if (!e ) r = ecb_ldexpf (m , -24);
709 else if (e != 31) r = ecb_ldexpf (m + 0x400, e - 25);
710 else if (m ) r = ECB_NAN;
711 else r = ECB_INFINITY;
712
713 return x & 0x8000 ? -r : r;
714 }
715 854
716 /* convert a float to ieee single/binary32 */ 855 /* convert a float to ieee single/binary32 */
717 ecb_function_ ecb_const uint32_t ecb_float_to_binary32 (float x); 856 ecb_function_ ecb_const uint32_t ecb_float_to_binary32 (float x);
718 ecb_function_ ecb_const uint32_t 857 ecb_function_ ecb_const uint32_t
719 ecb_float_to_binary32 (float x) 858 ecb_float_to_binary32 (float x)
… …
730 if (x == 0e0f ) return 0x00000000U; 869 if (x == 0e0f ) return 0x00000000U;
731 if (x > +3.40282346638528860e+38f) return 0x7f800000U; 870 if (x > +3.40282346638528860e+38f) return 0x7f800000U;
732 if (x < -3.40282346638528860e+38f) return 0xff800000U; 871 if (x < -3.40282346638528860e+38f) return 0xff800000U;
733 if (x != x ) return 0x7fbfffffU; 872 if (x != x ) return 0x7fbfffffU;
734 873
735 m = frexpf (x, &e) * 0x1000000U; 874 m = ecb_frexpf (x, &e) * 0x1000000U;
736 875
737 r = m & 0x80000000U; 876 r = m & 0x80000000U;
738 877
739 if (r) 878 if (r)
740 m = -m; 879 m = -m;
… …
851 #endif 990 #endif
852 991
853 return r; 992 return r;
854 } 993 }
855 994
856#endif 995 /* convert a float to ieee half/binary16 */
996 ecb_function_ ecb_const uint16_t ecb_float_to_binary16 (float x);
997 ecb_function_ ecb_const uint16_t
998 ecb_float_to_binary16 (float x)
999 {
1000 return ecb_binary32_to_binary16 (ecb_float_to_binary32 (x));
1001 }
857 1002
858#endif 1003 /* convert an ieee half/binary16 to float */
1004 ecb_function_ ecb_const float ecb_binary16_to_float (uint16_t x);
1005 ecb_function_ ecb_const float
1006 ecb_binary16_to_float (uint16_t x)
1007 {
1008 return ecb_binary32_to_float (ecb_binary16_to_binary32 (x));
1009 }
859 1010
1011#endif
1012
1013#endif
1014

Diff Legend

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