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

Comparing CBOR-XS/ecb.h (file contents):
Revision 1.11 by root, Mon Apr 27 20:21:53 2015 UTC vs.
Revision 1.14 by root, Thu Nov 15 19:52:41 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
… …
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__ ("" : : : "memory")
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__ ("" : : : "memory")
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")
176 #elif defined __ARM_ARCH_2__ \
177 || defined __ARM_ARCH_3__ || defined __ARM_ARCH_3M__ \
178 || defined __ARM_ARCH_4__ || defined __ARM_ARCH_4T__ \
179 || defined __ARM_ARCH_5__ || defined __ARM_ARCH_5E__ \
180 || defined __ARM_ARCH_5T__ || defined __ARM_ARCH_5TE__ \
181 || defined __ARM_ARCH_5TEJ__
182 /* 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__ \ 183 #elif defined __ARM_ARCH_6__ || defined __ARM_ARCH_6J__ \
165 || defined __ARM_ARCH_6K__ || defined __ARM_ARCH_6ZK__ 184 || defined __ARM_ARCH_6K__ || defined __ARM_ARCH_6ZK__ \
185 || defined __ARM_ARCH_6T2__
166 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("mcr p15,0,%0,c7,c10,5" : : "r" (0) : "memory") 186 #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__ \ 187 #elif defined __ARM_ARCH_7__ || defined __ARM_ARCH_7A__ \
168 || defined __ARM_ARCH_7M__ || defined __ARM_ARCH_7R__ 188 || defined __ARM_ARCH_7R__ || defined __ARM_ARCH_7M__
169 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("dmb" : : : "memory") 189 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("dmb" : : : "memory")
170 #elif __aarch64__ 190 #elif __aarch64__
171 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("dmb ish" : : : "memory") 191 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("dmb ish" : : : "memory")
172 #elif (__sparc || __sparc__) && !__sparcv8 192 #elif (__sparc || __sparc__) && !(__sparc_v8__ || defined __sparcv8)
173 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("membar #LoadStore | #LoadLoad | #StoreStore | #StoreLoad" : : : "memory") 193 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("membar #LoadStore | #LoadLoad | #StoreStore | #StoreLoad" : : : "memory")
174 #define ECB_MEMORY_FENCE_ACQUIRE __asm__ __volatile__ ("membar #LoadStore | #LoadLoad" : : : "memory") 194 #define ECB_MEMORY_FENCE_ACQUIRE __asm__ __volatile__ ("membar #LoadStore | #LoadLoad" : : : "memory")
175 #define ECB_MEMORY_FENCE_RELEASE __asm__ __volatile__ ("membar #LoadStore | #StoreStore") 195 #define ECB_MEMORY_FENCE_RELEASE __asm__ __volatile__ ("membar #LoadStore | #StoreStore")
176 #elif defined __s390__ || defined __s390x__ 196 #elif defined __s390__ || defined __s390x__
177 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("bcr 15,0" : : : "memory") 197 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("bcr 15,0" : : : "memory")
… …
349 #define ecb_deprecated __declspec (deprecated) 369 #define ecb_deprecated __declspec (deprecated)
350#else 370#else
351 #define ecb_deprecated ecb_attribute ((__deprecated__)) 371 #define ecb_deprecated ecb_attribute ((__deprecated__))
352#endif 372#endif
353 373
354#if __MSC_VER >= 1500 374#if _MSC_VER >= 1500
355 #define ecb_deprecated_message(msg) __declspec (deprecated (msg)) 375 #define ecb_deprecated_message(msg) __declspec (deprecated (msg))
356#elif ECB_GCC_VERSION(4,5) 376#elif ECB_GCC_VERSION(4,5)
357 #define ecb_deprecated_message(msg) ecb_attribute ((__deprecated__ (msg)) 377 #define ecb_deprecated_message(msg) ecb_attribute ((__deprecated__ (msg))
358#else 378#else
359 #define ecb_deprecated_message(msg) ecb_deprecated 379 #define ecb_deprecated_message(msg) ecb_deprecated
… …
367 387
368#define ecb_unused ecb_attribute ((__unused__)) 388#define ecb_unused ecb_attribute ((__unused__))
369#define ecb_const ecb_attribute ((__const__)) 389#define ecb_const ecb_attribute ((__const__))
370#define ecb_pure ecb_attribute ((__pure__)) 390#define ecb_pure ecb_attribute ((__pure__))
371 391
372/* TODO http://msdn.microsoft.com/en-us/library/k6ktzx3s.aspx __declspec(noreturn) */
373#if ECB_C11 || __IBMC_NORETURN 392#if ECB_C11 || __IBMC_NORETURN
374 /* http://pic.dhe.ibm.com/infocenter/compbg/v121v141/topic/com.ibm.xlcpp121.bg.doc/language_ref/noreturn.html */ 393 /* 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 394 #define ecb_noreturn _Noreturn
376#elif ECB_CPP11 395#elif ECB_CPP11
377 #define ecb_noreturn [[noreturn]] 396 #define ecb_noreturn [[noreturn]]
378#elif _MSC_VER >= 1200 397#elif _MSC_VER >= 1200
398 /* http://msdn.microsoft.com/en-us/library/k6ktzx3s.aspx */
379 #define ecb_noreturn __declspec (noreturn) 399 #define ecb_noreturn __declspec (noreturn)
380#else 400#else
381 #define ecb_noreturn ecb_attribute ((__noreturn__)) 401 #define ecb_noreturn ecb_attribute ((__noreturn__))
382#endif 402#endif
383 403
… …
415#else 435#else
416 ecb_function_ ecb_const int ecb_ctz32 (uint32_t x); 436 ecb_function_ ecb_const int ecb_ctz32 (uint32_t x);
417 ecb_function_ ecb_const int 437 ecb_function_ ecb_const int
418 ecb_ctz32 (uint32_t x) 438 ecb_ctz32 (uint32_t x)
419 { 439 {
440#if 1400 <= _MSC_VER && (_M_IX86 || _M_X64 || _M_IA64 || _M_ARM)
441 unsigned long r;
442 _BitScanForward (&r, x);
443 return (int)r;
444#else
420 int r = 0; 445 int r = 0;
421 446
422 x &= ~x + 1; /* this isolates the lowest bit */ 447 x &= ~x + 1; /* this isolates the lowest bit */
423 448
424#if ECB_branchless_on_i386 449#if ECB_branchless_on_i386
… …
434 if (x & 0xff00ff00) r += 8; 459 if (x & 0xff00ff00) r += 8;
435 if (x & 0xffff0000) r += 16; 460 if (x & 0xffff0000) r += 16;
436#endif 461#endif
437 462
438 return r; 463 return r;
464#endif
439 } 465 }
440 466
441 ecb_function_ ecb_const int ecb_ctz64 (uint64_t x); 467 ecb_function_ ecb_const int ecb_ctz64 (uint64_t x);
442 ecb_function_ ecb_const int 468 ecb_function_ ecb_const int
443 ecb_ctz64 (uint64_t x) 469 ecb_ctz64 (uint64_t x)
444 { 470 {
471#if 1400 <= _MSC_VER && (_M_X64 || _M_IA64 || _M_ARM)
472 unsigned long r;
473 _BitScanForward64 (&r, x);
474 return (int)r;
475#else
445 int shift = x & 0xffffffffU ? 0 : 32; 476 int shift = x & 0xffffffff ? 0 : 32;
446 return ecb_ctz32 (x >> shift) + shift; 477 return ecb_ctz32 (x >> shift) + shift;
478#endif
447 } 479 }
448 480
449 ecb_function_ ecb_const int ecb_popcount32 (uint32_t x); 481 ecb_function_ ecb_const int ecb_popcount32 (uint32_t x);
450 ecb_function_ ecb_const int 482 ecb_function_ ecb_const int
451 ecb_popcount32 (uint32_t x) 483 ecb_popcount32 (uint32_t x)
… …
459 } 491 }
460 492
461 ecb_function_ ecb_const int ecb_ld32 (uint32_t x); 493 ecb_function_ ecb_const int ecb_ld32 (uint32_t x);
462 ecb_function_ ecb_const int ecb_ld32 (uint32_t x) 494 ecb_function_ ecb_const int ecb_ld32 (uint32_t x)
463 { 495 {
496#if 1400 <= _MSC_VER && (_M_IX86 || _M_X64 || _M_IA64 || _M_ARM)
497 unsigned long r;
498 _BitScanReverse (&r, x);
499 return (int)r;
500#else
464 int r = 0; 501 int r = 0;
465 502
466 if (x >> 16) { x >>= 16; r += 16; } 503 if (x >> 16) { x >>= 16; r += 16; }
467 if (x >> 8) { x >>= 8; r += 8; } 504 if (x >> 8) { x >>= 8; r += 8; }
468 if (x >> 4) { x >>= 4; r += 4; } 505 if (x >> 4) { x >>= 4; r += 4; }
469 if (x >> 2) { x >>= 2; r += 2; } 506 if (x >> 2) { x >>= 2; r += 2; }
470 if (x >> 1) { r += 1; } 507 if (x >> 1) { r += 1; }
471 508
472 return r; 509 return r;
510#endif
473 } 511 }
474 512
475 ecb_function_ ecb_const int ecb_ld64 (uint64_t x); 513 ecb_function_ ecb_const int ecb_ld64 (uint64_t x);
476 ecb_function_ ecb_const int ecb_ld64 (uint64_t x) 514 ecb_function_ ecb_const int ecb_ld64 (uint64_t x)
477 { 515 {
516#if 1400 <= _MSC_VER && (_M_X64 || _M_IA64 || _M_ARM)
517 unsigned long r;
518 _BitScanReverse64 (&r, x);
519 return (int)r;
520#else
478 int r = 0; 521 int r = 0;
479 522
480 if (x >> 32) { x >>= 32; r += 32; } 523 if (x >> 32) { x >>= 32; r += 32; }
481 524
482 return r + ecb_ld32 (x); 525 return r + ecb_ld32 (x);
526#endif
483 } 527 }
484#endif 528#endif
485 529
486ecb_function_ ecb_const ecb_bool ecb_is_pot32 (uint32_t x); 530ecb_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)); } 531ecb_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); } 588ecb_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); } 589ecb_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); } 590ecb_inline ecb_const uint64_t ecb_rotr64 (uint64_t x, unsigned int count) { return (x << (64 - count)) | (x >> count); }
547 591
548#if ECB_GCC_VERSION(4,3) || (ECB_CLANG_BUILTIN(__builtin_bswap32) && ECB_CLANG_BUILTIN(__builtin_bswap64)) 592#if ECB_GCC_VERSION(4,3) || (ECB_CLANG_BUILTIN(__builtin_bswap32) && ECB_CLANG_BUILTIN(__builtin_bswap64))
593 #if ECB_GCC_VERSION(4,8) || ECB_CLANG_BUILTIN(__builtin_bswap16)
594 #define ecb_bswap16(x) __builtin_bswap16 (x)
595 #else
549 #define ecb_bswap16(x) (__builtin_bswap32 (x) >> 16) 596 #define ecb_bswap16(x) (__builtin_bswap32 (x) >> 16)
597 #endif
550 #define ecb_bswap32(x) __builtin_bswap32 (x) 598 #define ecb_bswap32(x) __builtin_bswap32 (x)
551 #define ecb_bswap64(x) __builtin_bswap64 (x) 599 #define ecb_bswap64(x) __builtin_bswap64 (x)
600#elif _MSC_VER
601 #include <stdlib.h>
602 #define ecb_bswap16(x) ((uint16_t)_byteswap_ushort ((uint16_t)(x)))
603 #define ecb_bswap32(x) ((uint32_t)_byteswap_ulong ((uint32_t)(x)))
604 #define ecb_bswap64(x) ((uint64_t)_byteswap_uint64 ((uint64_t)(x)))
552#else 605#else
553 ecb_function_ ecb_const uint16_t ecb_bswap16 (uint16_t x); 606 ecb_function_ ecb_const uint16_t ecb_bswap16 (uint16_t x);
554 ecb_function_ ecb_const uint16_t 607 ecb_function_ ecb_const uint16_t
555 ecb_bswap16 (uint16_t x) 608 ecb_bswap16 (uint16_t x)
556 { 609 {
… …
581#endif 634#endif
582 635
583/* try to tell the compiler that some condition is definitely true */ 636/* try to tell the compiler that some condition is definitely true */
584#define ecb_assume(cond) if (!(cond)) ecb_unreachable (); else 0 637#define ecb_assume(cond) if (!(cond)) ecb_unreachable (); else 0
585 638
586ecb_inline ecb_const unsigned char ecb_byteorder_helper (void); 639ecb_inline ecb_const uint32_t ecb_byteorder_helper (void);
587ecb_inline ecb_const unsigned char 640ecb_inline ecb_const uint32_t
588ecb_byteorder_helper (void) 641ecb_byteorder_helper (void)
589{ 642{
590 /* the union code still generates code under pressure in gcc, */ 643 /* the union code still generates code under pressure in gcc, */
591 /* but less than using pointers, and always seems to */ 644 /* but less than using pointers, and always seems to */
592 /* successfully return a constant. */ 645 /* successfully return a constant. */
593 /* the reason why we have this horrible preprocessor mess */ 646 /* the reason why we have this horrible preprocessor mess */
594 /* is to avoid it in all cases, at least on common architectures */ 647 /* is to avoid it in all cases, at least on common architectures */
595 /* or when using a recent enough gcc version (>= 4.6) */ 648 /* 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__ 649#if (defined __BYTE_ORDER__ && __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__) \
650 || ((__i386 || __i386__ || _M_IX86 || ECB_GCC_AMD64 || ECB_MSVC_AMD64) && !__VOS__)
651 #define ECB_LITTLE_ENDIAN 1
599 return 0x44; 652 return 0x44332211;
600#elif __BYTE_ORDER__ && __BYTE_ORDER__ == __ORDER_BIG_ENDIAN__ 653#elif (defined __BYTE_ORDER__ && __BYTE_ORDER__ == __ORDER_BIG_ENDIAN__) \
654 || ((__AARCH64EB__ || __MIPSEB__ || __ARMEB__) && !__VOS__)
655 #define ECB_BIG_ENDIAN 1
601 return 0x11; 656 return 0x11223344;
602#else 657#else
603 union 658 union
604 { 659 {
660 uint8_t c[4];
605 uint32_t i; 661 uint32_t u;
606 uint8_t c;
607 } u = { 0x11223344 }; 662 } u = { 0x11, 0x22, 0x33, 0x44 };
608 return u.c; 663 return u.u;
609#endif 664#endif
610} 665}
611 666
612ecb_inline ecb_const ecb_bool ecb_big_endian (void); 667ecb_inline ecb_const ecb_bool ecb_big_endian (void);
613ecb_inline ecb_const ecb_bool ecb_big_endian (void) { return ecb_byteorder_helper () == 0x11; } 668ecb_inline ecb_const ecb_bool ecb_big_endian (void) { return ecb_byteorder_helper () == 0x11223344; }
614ecb_inline ecb_const ecb_bool ecb_little_endian (void); 669ecb_inline ecb_const ecb_bool ecb_little_endian (void);
615ecb_inline ecb_const ecb_bool ecb_little_endian (void) { return ecb_byteorder_helper () == 0x44; } 670ecb_inline ecb_const ecb_bool ecb_little_endian (void) { return ecb_byteorder_helper () == 0x44332211; }
616 671
617#if ECB_GCC_VERSION(3,0) || ECB_C99 672#if ECB_GCC_VERSION(3,0) || ECB_C99
618 #define ecb_mod(m,n) ((m) % (n) + ((m) % (n) < 0 ? (n) : 0)) 673 #define ecb_mod(m,n) ((m) % (n) + ((m) % (n) < 0 ? (n) : 0))
619#else 674#else
620 #define ecb_mod(m,n) ((m) < 0 ? ((n) - 1 - ((-1 - (m)) % (n))) : ((m) % (n))) 675 #define ecb_mod(m,n) ((m) < 0 ? ((n) - 1 - ((-1 - (m)) % (n))) : ((m) % (n)))
… …
645 } 700 }
646#else 701#else
647 #define ecb_array_length(name) (sizeof (name) / sizeof (name [0])) 702 #define ecb_array_length(name) (sizeof (name) / sizeof (name [0]))
648#endif 703#endif
649 704
705ecb_function_ ecb_const uint32_t ecb_binary16_to_binary32 (uint32_t x);
706ecb_function_ ecb_const uint32_t
707ecb_binary16_to_binary32 (uint32_t x)
708{
709 unsigned int s = (x & 0x8000) << (31 - 15);
710 int e = (x >> 10) & 0x001f;
711 unsigned int m = x & 0x03ff;
712
713 if (ecb_expect_false (e == 31))
714 /* infinity or NaN */
715 e = 255 - (127 - 15);
716 else if (ecb_expect_false (!e))
717 {
718 if (ecb_expect_true (!m))
719 /* zero, handled by code below by forcing e to 0 */
720 e = 0 - (127 - 15);
721 else
722 {
723 /* subnormal, renormalise */
724 unsigned int s = 10 - ecb_ld32 (m);
725
726 m = (m << s) & 0x3ff; /* mask implicit bit */
727 e -= s - 1;
728 }
729 }
730
731 /* e and m now are normalised, or zero, (or inf or nan) */
732 e += 127 - 15;
733
734 return s | (e << 23) | (m << (23 - 10));
735}
736
737ecb_function_ ecb_const uint16_t ecb_binary32_to_binary16 (uint32_t x);
738ecb_function_ ecb_const uint16_t
739ecb_binary32_to_binary16 (uint32_t x)
740{
741 unsigned int s = (x >> 16) & 0x00008000; /* sign bit, the easy part */
742 unsigned int e = ((x >> 23) & 0x000000ff) - (127 - 15); /* the desired exponent */
743 unsigned int m = x & 0x007fffff;
744
745 x &= 0x7fffffff;
746
747 /* if it's within range of binary16 normals, use fast path */
748 if (ecb_expect_true (0x38800000 <= x && x <= 0x477fefff))
749 {
750 /* mantissa round-to-even */
751 m += 0x00000fff + ((m >> (23 - 10)) & 1);
752
753 /* handle overflow */
754 if (ecb_expect_false (m >= 0x00800000))
755 {
756 m >>= 1;
757 e += 1;
758 }
759
760 return s | (e << 10) | (m >> (23 - 10));
761 }
762
763 /* handle large numbers and infinity */
764 if (ecb_expect_true (0x477fefff < x && x <= 0x7f800000))
765 return s | 0x7c00;
766
767 /* handle zero, subnormals and small numbers */
768 if (ecb_expect_true (x < 0x38800000))
769 {
770 /* zero */
771 if (ecb_expect_true (!x))
772 return s;
773
774 /* handle subnormals */
775
776 /* too small, will be zero */
777 if (e < (14 - 24)) /* might not be sharp, but is good enough */
778 return s;
779
780 m |= 0x00800000; /* make implicit bit explicit */
781
782 /* very tricky - we need to round to the nearest e (+10) bit value */
783 {
784 unsigned int bits = 14 - e;
785 unsigned int half = (1 << (bits - 1)) - 1;
786 unsigned int even = (m >> bits) & 1;
787
788 /* if this overflows, we will end up with a normalised number */
789 m = (m + half + even) >> bits;
790 }
791
792 return s | m;
793 }
794
795 /* handle NaNs, preserve leftmost nan bits, but make sure we don't turn them into infinities */
796 m >>= 13;
797
798 return s | 0x7c00 | m | !m;
799}
800
650/*******************************************************************************/ 801/*******************************************************************************/
651/* floating point stuff, can be disabled by defining ECB_NO_LIBM */ 802/* floating point stuff, can be disabled by defining ECB_NO_LIBM */
652 803
653/* basically, everything uses "ieee pure-endian" floating point numbers */ 804/* basically, everything uses "ieee pure-endian" floating point numbers */
654/* the only noteworthy exception is ancient armle, which uses order 43218765 */ 805/* the only noteworthy exception is ancient armle, which uses order 43218765 */
655#if 0 \ 806#if 0 \
656 || __i386 || __i386__ \ 807 || __i386 || __i386__ \
657 || __amd64 || __amd64__ || __x86_64 || __x86_64__ \ 808 || ECB_GCC_AMD64 \
658 || __powerpc__ || __ppc__ || __powerpc64__ || __ppc64__ \ 809 || __powerpc__ || __ppc__ || __powerpc64__ || __ppc64__ \
659 || defined __s390__ || defined __s390x__ \ 810 || defined __s390__ || defined __s390x__ \
660 || defined __mips__ \ 811 || defined __mips__ \
661 || defined __alpha__ \ 812 || defined __alpha__ \
662 || defined __hppa__ \ 813 || defined __hppa__ \
663 || defined __ia64__ \ 814 || defined __ia64__ \
664 || defined __m68k__ \ 815 || defined __m68k__ \
665 || defined __m88k__ \ 816 || defined __m88k__ \
666 || defined __sh__ \ 817 || defined __sh__ \
667 || defined _M_IX86 || defined _M_AMD64 || defined _M_IA64 \ 818 || 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__)) \ 819 || (defined __arm__ && (defined __ARM_EABI__ || defined __EABI__ || defined __VFP_FP__ || defined _WIN32_WCE || defined __ANDROID__)) \
669 || defined __aarch64__ 820 || defined __aarch64__
670 #define ECB_STDFP 1 821 #define ECB_STDFP 1
671 #include <string.h> /* for memcpy */ 822 #include <string.h> /* for memcpy */
672#else 823#else
… …
690 #define ECB_NAN ECB_INFINITY 841 #define ECB_NAN ECB_INFINITY
691 #endif 842 #endif
692 843
693 #if ECB_C99 || _XOPEN_VERSION >= 600 || _POSIX_VERSION >= 200112L 844 #if ECB_C99 || _XOPEN_VERSION >= 600 || _POSIX_VERSION >= 200112L
694 #define ecb_ldexpf(x,e) ldexpf ((x), (e)) 845 #define ecb_ldexpf(x,e) ldexpf ((x), (e))
846 #define ecb_frexpf(x,e) frexpf ((x), (e))
695 #else 847 #else
696 #define ecb_ldexpf(x,e) (float) ldexp ((x), (e)) 848 #define ecb_ldexpf(x,e) (float) ldexp ((double) (x), (e))
849 #define ecb_frexpf(x,e) (float) frexp ((double) (x), (e))
697 #endif 850 #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 }
715 851
716 /* convert a float to ieee single/binary32 */ 852 /* convert a float to ieee single/binary32 */
717 ecb_function_ ecb_const uint32_t ecb_float_to_binary32 (float x); 853 ecb_function_ ecb_const uint32_t ecb_float_to_binary32 (float x);
718 ecb_function_ ecb_const uint32_t 854 ecb_function_ ecb_const uint32_t
719 ecb_float_to_binary32 (float x) 855 ecb_float_to_binary32 (float x)
… …
730 if (x == 0e0f ) return 0x00000000U; 866 if (x == 0e0f ) return 0x00000000U;
731 if (x > +3.40282346638528860e+38f) return 0x7f800000U; 867 if (x > +3.40282346638528860e+38f) return 0x7f800000U;
732 if (x < -3.40282346638528860e+38f) return 0xff800000U; 868 if (x < -3.40282346638528860e+38f) return 0xff800000U;
733 if (x != x ) return 0x7fbfffffU; 869 if (x != x ) return 0x7fbfffffU;
734 870
735 m = frexpf (x, &e) * 0x1000000U; 871 m = ecb_frexpf (x, &e) * 0x1000000U;
736 872
737 r = m & 0x80000000U; 873 r = m & 0x80000000U;
738 874
739 if (r) 875 if (r)
740 m = -m; 876 m = -m;
… …
851 #endif 987 #endif
852 988
853 return r; 989 return r;
854 } 990 }
855 991
856#endif 992 /* convert a float to ieee half/binary16 */
993 ecb_function_ ecb_const uint16_t ecb_float_to_binary16 (float x);
994 ecb_function_ ecb_const uint16_t
995 ecb_float_to_binary16 (float x)
996 {
997 return ecb_binary32_to_binary16 (ecb_float_to_binary32 (x));
998 }
857 999
858#endif 1000 /* convert an ieee half/binary16 to float */
1001 ecb_function_ ecb_const float ecb_binary16_to_float (uint16_t x);
1002 ecb_function_ ecb_const float
1003 ecb_binary16_to_float (uint16_t x)
1004 {
1005 return ecb_binary32_to_float (ecb_binary16_to_binary32 (x));
1006 }
859 1007
1008#endif
1009
1010#endif
1011

Diff Legend

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