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

Comparing libecb/ecb.h (file contents):
Revision 1.157 by root, Fri Feb 20 17:17:26 2015 UTC vs.
Revision 1.168 by root, Sat Nov 21 18:09:36 2015 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;
… …
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
152#ifndef ECB_MEMORY_FENCE 160#ifndef ECB_MEMORY_FENCE
153 #if ECB_GCC_VERSION(2,5) || defined __INTEL_COMPILER || (__llvm__ && __GNUC__) || __SUNPRO_C >= 0x5110 || __SUNPRO_CC >= 0x5110 161 #if ECB_GCC_VERSION(2,5) || defined __INTEL_COMPILER || (__llvm__ && __GNUC__) || __SUNPRO_C >= 0x5110 || __SUNPRO_CC >= 0x5110
154 #if __i386 || __i386__ 162 #if __i386 || __i386__
155 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("lock; orb $0, -1(%%esp)" : : : "memory") 163 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("lock; orb $0, -1(%%esp)" : : : "memory")
156 #define ECB_MEMORY_FENCE_ACQUIRE __asm__ __volatile__ ("" : : : "memory") 164 #define ECB_MEMORY_FENCE_ACQUIRE __asm__ __volatile__ ("" : : : "memory")
157 #define ECB_MEMORY_FENCE_RELEASE __asm__ __volatile__ ("") 165 #define ECB_MEMORY_FENCE_RELEASE __asm__ __volatile__ ("")
158 #elif __amd64 || __amd64__ || __x86_64 || __x86_64__ 166 #elif ECB_GCC_AMD64
159 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("mfence" : : : "memory") 167 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("mfence" : : : "memory")
160 #define ECB_MEMORY_FENCE_ACQUIRE __asm__ __volatile__ ("" : : : "memory") 168 #define ECB_MEMORY_FENCE_ACQUIRE __asm__ __volatile__ ("" : : : "memory")
161 #define ECB_MEMORY_FENCE_RELEASE __asm__ __volatile__ ("") 169 #define ECB_MEMORY_FENCE_RELEASE __asm__ __volatile__ ("")
162 #elif __powerpc__ || __ppc__ || __powerpc64__ || __ppc64__ 170 #elif __powerpc__ || __ppc__ || __powerpc64__ || __ppc64__
163 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("sync" : : : "memory") 171 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("sync" : : : "memory")
… …
167 #elif defined __ARM_ARCH_7__ || defined __ARM_ARCH_7A__ \ 175 #elif defined __ARM_ARCH_7__ || defined __ARM_ARCH_7A__ \
168 || defined __ARM_ARCH_7M__ || defined __ARM_ARCH_7R__ 176 || defined __ARM_ARCH_7M__ || defined __ARM_ARCH_7R__
169 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("dmb" : : : "memory") 177 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("dmb" : : : "memory")
170 #elif __aarch64__ 178 #elif __aarch64__
171 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("dmb ish" : : : "memory") 179 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("dmb ish" : : : "memory")
172 #elif (__sparc || __sparc__) && !__sparcv8 180 #elif (__sparc || __sparc__) && !(__sparc_v8__ || defined __sparcv8)
173 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("membar #LoadStore | #LoadLoad | #StoreStore | #StoreLoad" : : : "memory") 181 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("membar #LoadStore | #LoadLoad | #StoreStore | #StoreLoad" : : : "memory")
174 #define ECB_MEMORY_FENCE_ACQUIRE __asm__ __volatile__ ("membar #LoadStore | #LoadLoad" : : : "memory") 182 #define ECB_MEMORY_FENCE_ACQUIRE __asm__ __volatile__ ("membar #LoadStore | #LoadLoad" : : : "memory")
175 #define ECB_MEMORY_FENCE_RELEASE __asm__ __volatile__ ("membar #LoadStore | #StoreStore") 183 #define ECB_MEMORY_FENCE_RELEASE __asm__ __volatile__ ("membar #LoadStore | #StoreStore")
176 #elif defined __s390__ || defined __s390x__ 184 #elif defined __s390__ || defined __s390x__
177 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("bcr 15,0" : : : "memory") 185 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("bcr 15,0" : : : "memory")
… …
349 #define ecb_deprecated __declspec (deprecated) 357 #define ecb_deprecated __declspec (deprecated)
350#else 358#else
351 #define ecb_deprecated ecb_attribute ((__deprecated__)) 359 #define ecb_deprecated ecb_attribute ((__deprecated__))
352#endif 360#endif
353 361
354#if __MSC_VER >= 1500 362#if _MSC_VER >= 1500
355 #define ecb_deprecated_message(msg) __declspec (deprecated (msg)) 363 #define ecb_deprecated_message(msg) __declspec (deprecated (msg))
356#elif ECB_GCC_VERSION(4,5) 364#elif ECB_GCC_VERSION(4,5)
357 #define ecb_deprecated_message(msg) ecb_attribute ((__deprecated__ (msg)) 365 #define ecb_deprecated_message(msg) ecb_attribute ((__deprecated__ (msg))
358#else 366#else
359 #define ecb_deprecated_message(msg) ecb_deprecated 367 #define ecb_deprecated_message(msg) ecb_deprecated
… …
368#define ecb_unused ecb_attribute ((__unused__)) 376#define ecb_unused ecb_attribute ((__unused__))
369#define ecb_const ecb_attribute ((__const__)) 377#define ecb_const ecb_attribute ((__const__))
370#define ecb_pure ecb_attribute ((__pure__)) 378#define ecb_pure ecb_attribute ((__pure__))
371 379
372#if ECB_C11 || __IBMC_NORETURN 380#if ECB_C11 || __IBMC_NORETURN
373 /* http://pic.dhe.ibm.com/infocenter/compbg/v121v141/topic/com.ibm.xlcpp121.bg.doc/language_ref/noreturn.html */ 381 /* http://www-01.ibm.com/support/knowledgecenter/SSGH3R_13.1.0/com.ibm.xlcpp131.aix.doc/language_ref/noreturn.html */
374 #define ecb_noreturn _Noreturn 382 #define ecb_noreturn _Noreturn
375#elif ECB_CPP11 383#elif ECB_CPP11
376 #define ecb_noreturn [[noreturn]] 384 #define ecb_noreturn [[noreturn]]
377#elif _MSC_VER >= 1200 385#elif _MSC_VER >= 1200
378 /* http://msdn.microsoft.com/en-us/library/k6ktzx3s.aspx */ 386 /* http://msdn.microsoft.com/en-us/library/k6ktzx3s.aspx */
… …
440 448
441 ecb_function_ ecb_const int ecb_ctz64 (uint64_t x); 449 ecb_function_ ecb_const int ecb_ctz64 (uint64_t x);
442 ecb_function_ ecb_const int 450 ecb_function_ ecb_const int
443 ecb_ctz64 (uint64_t x) 451 ecb_ctz64 (uint64_t x)
444 { 452 {
445 int shift = x & 0xffffffffU ? 0 : 32; 453 int shift = x & 0xffffffff ? 0 : 32;
446 return ecb_ctz32 (x >> shift) + shift; 454 return ecb_ctz32 (x >> shift) + shift;
447 } 455 }
448 456
449 ecb_function_ ecb_const int ecb_popcount32 (uint32_t x); 457 ecb_function_ ecb_const int ecb_popcount32 (uint32_t x);
450 ecb_function_ ecb_const int 458 ecb_function_ ecb_const int
… …
544ecb_inline ecb_const uint32_t ecb_rotr32 (uint32_t x, unsigned int count) { return (x << (32 - count)) | (x >> count); } 552ecb_inline ecb_const uint32_t ecb_rotr32 (uint32_t x, unsigned int count) { return (x << (32 - count)) | (x >> count); }
545ecb_inline ecb_const uint64_t ecb_rotl64 (uint64_t x, unsigned int count) { return (x >> (64 - count)) | (x << count); } 553ecb_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); } 554ecb_inline ecb_const uint64_t ecb_rotr64 (uint64_t x, unsigned int count) { return (x << (64 - count)) | (x >> count); }
547 555
548#if ECB_GCC_VERSION(4,3) || (ECB_CLANG_BUILTIN(__builtin_bswap32) && ECB_CLANG_BUILTIN(__builtin_bswap64)) 556#if ECB_GCC_VERSION(4,3) || (ECB_CLANG_BUILTIN(__builtin_bswap32) && ECB_CLANG_BUILTIN(__builtin_bswap64))
557 #if ECB_GCC_VERSION(4,8) || ECB_CLANG_BUILTIN(__builtin_bswap16)
558 #define ecb_bswap16(x) __builtin_bswap16 (x)
559 #else
549 #define ecb_bswap16(x) (__builtin_bswap32 (x) >> 16) 560 #define ecb_bswap16(x) (__builtin_bswap32 (x) >> 16)
561 #endif
550 #define ecb_bswap32(x) __builtin_bswap32 (x) 562 #define ecb_bswap32(x) __builtin_bswap32 (x)
551 #define ecb_bswap64(x) __builtin_bswap64 (x) 563 #define ecb_bswap64(x) __builtin_bswap64 (x)
564#elif _MSC_VER
565 #include <stdlib.h>
566 #define ecb_bswap16(x) ((uint16_t)_byteswap_ushort ((uint16_t)(x)))
567 #define ecb_bswap32(x) ((uint32_t)_byteswap_ulong ((uint32_t)(x)))
568 #define ecb_bswap64(x) ((uint64_t)_byteswap_uint64 ((uint64_t)(x)))
552#else 569#else
553 ecb_function_ ecb_const uint16_t ecb_bswap16 (uint16_t x); 570 ecb_function_ ecb_const uint16_t ecb_bswap16 (uint16_t x);
554 ecb_function_ ecb_const uint16_t 571 ecb_function_ ecb_const uint16_t
555 ecb_bswap16 (uint16_t x) 572 ecb_bswap16 (uint16_t x)
556 { 573 {
… …
591 /* but less than using pointers, and always seems to */ 608 /* but less than using pointers, and always seems to */
592 /* successfully return a constant. */ 609 /* successfully return a constant. */
593 /* the reason why we have this horrible preprocessor mess */ 610 /* the reason why we have this horrible preprocessor mess */
594 /* is to avoid it in all cases, at least on common architectures */ 611 /* is to avoid it in all cases, at least on common architectures */
595 /* or when using a recent enough gcc version (>= 4.6) */ 612 /* or when using a recent enough gcc version (>= 4.6) */
596#if ((__i386 || __i386__) && !__VOS__) || _M_X86 || __amd64 || __amd64__ || _M_X64 613#if ((__i386 || __i386__) && !__VOS__) || _M_IX86 || ECB_GCC_AMD64 || ECB_MSVC_AMD64
597 return 0x44; 614 return 0x44;
598#elif __BYTE_ORDER__ && __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__ 615#elif __BYTE_ORDER__ && __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__
599 return 0x44; 616 return 0x44;
600#elif __BYTE_ORDER__ && __BYTE_ORDER__ == __ORDER_BIG_ENDIAN__ 617#elif __BYTE_ORDER__ && __BYTE_ORDER__ == __ORDER_BIG_ENDIAN__
601 return 0x11; 618 return 0x11;
… …
645 } 662 }
646#else 663#else
647 #define ecb_array_length(name) (sizeof (name) / sizeof (name [0])) 664 #define ecb_array_length(name) (sizeof (name) / sizeof (name [0]))
648#endif 665#endif
649 666
667ecb_function_ ecb_const uint32_t ecb_binary16_to_binary32 (uint16_t x);
668ecb_function_ ecb_const uint32_t
669ecb_binary16_to_binary32 (uint16_t x)
670{
671 unsigned int s = (x & 0x8000) << (31 - 15);
672 int e = (x >> 10) & 0x001f;
673 unsigned int m = x & 0x03ff;
674
675 if (ecb_expect_false (e == 31))
676 /* infinity or NaN */
677 e = 255 - (127 - 15);
678 else if (ecb_expect_false (!e))
679 {
680 if (ecb_expect_true (!m))
681 /* zero, handled by code below by forcing e to 0 */
682 e = 0 - (127 - 15);
683 else
684 {
685 /* subnormal, renormalise */
686 unsigned int s = 10 - ecb_ld32 (m);
687
688 m = (m << s) & 0x3ff; /* mask implicit bit */
689 e -= s - 1;
690 }
691 }
692
693 /* e and m now are normalised, or zero, (or inf or nan) */
694 e += 127 - 15;
695
696 return s | (e << 23) | (m << (23 - 10));
697}
698
699ecb_function_ ecb_const uint16_t ecb_binary32_to_binary16 (uint32_t x);
700ecb_function_ ecb_const uint16_t
701ecb_binary32_to_binary16 (uint32_t x)
702{
703 unsigned int s = (x >> 16) & 0x00008000; /* sign bit, the easy part */
704 unsigned int e = ((x >> 23) & 0x000000ff) - (127 - 15); /* the desired exponent */
705 unsigned int m = x & 0x007fffff;
706
707 x &= 0x7fffffff;
708
709 /* if it's within range of binary16 normals, use fast path */
710 if (ecb_expect_true (0x38800000 <= x && x <= 0x477fefff))
711 {
712 /* mantissa round-to-even */
713 m += 0x00000fff + ((m >> (23 - 10)) & 1);
714
715 /* handle overflow */
716 if (ecb_expect_false (m >= 0x00800000))
717 {
718 m >>= 1;
719 e += 1;
720 }
721
722 return s | (e << 10) | (m >> (23 - 10));
723 }
724
725 /* handle large numbers and infinity */
726 if (ecb_expect_true (0x477fefff < x && x <= 0x7f800000))
727 return s | 0x7c00;
728
729 /* handle zero and subnormals */
730 if (ecb_expect_true (x < 0x38800000))
731 {
732 /* zero */
733 if (ecb_expect_true (!x))
734 return s;
735
736 /* handle subnormals */
737
738 m |= 0x00800000; /* make implicit bit explicit */
739
740 /* very tricky - we need to round to the nearest e (+10) bit value */
741 {
742 unsigned int bits = 14 - e;
743 unsigned int half = (1 << (bits - 1)) - 1;
744 unsigned int even = (m >> bits) & 1;
745
746 /* if this overflows, we will end up with a normalised number */
747 m = (m + half + even) >> bits;
748 }
749
750 return s | m;
751 }
752
753 /* handle NaNs, preserve leftmost nan bits, but make sure we don't turn them into infinities */
754 m >>= 13;
755
756 return s | 0x7c00 | m | !m;
757}
758
650/*******************************************************************************/ 759/*******************************************************************************/
651/* floating point stuff, can be disabled by defining ECB_NO_LIBM */ 760/* floating point stuff, can be disabled by defining ECB_NO_LIBM */
652 761
653/* basically, everything uses "ieee pure-endian" floating point numbers */ 762/* basically, everything uses "ieee pure-endian" floating point numbers */
654/* the only noteworthy exception is ancient armle, which uses order 43218765 */ 763/* the only noteworthy exception is ancient armle, which uses order 43218765 */
655#if 0 \ 764#if 0 \
656 || __i386 || __i386__ \ 765 || __i386 || __i386__ \
657 || __amd64 || __amd64__ || __x86_64 || __x86_64__ \ 766 || ECB_GCC_AMD64 \
658 || __powerpc__ || __ppc__ || __powerpc64__ || __ppc64__ \ 767 || __powerpc__ || __ppc__ || __powerpc64__ || __ppc64__ \
659 || defined __s390__ || defined __s390x__ \ 768 || defined __s390__ || defined __s390x__ \
660 || defined __mips__ \ 769 || defined __mips__ \
661 || defined __alpha__ \ 770 || defined __alpha__ \
662 || defined __hppa__ \ 771 || defined __hppa__ \
663 || defined __ia64__ \ 772 || defined __ia64__ \
664 || defined __m68k__ \ 773 || defined __m68k__ \
665 || defined __m88k__ \ 774 || defined __m88k__ \
666 || defined __sh__ \ 775 || defined __sh__ \
667 || defined _M_IX86 || defined _M_AMD64 || defined _M_IA64 \ 776 || 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__)) \ 777 || (defined __arm__ && (defined __ARM_EABI__ || defined __EABI__ || defined __VFP_FP__ || defined _WIN32_WCE || defined __ANDROID__)) \
669 || defined __aarch64__ 778 || defined __aarch64__
670 #define ECB_STDFP 1 779 #define ECB_STDFP 1
671 #include <string.h> /* for memcpy */ 780 #include <string.h> /* for memcpy */
672#else 781#else
… …
690 #define ECB_NAN ECB_INFINITY 799 #define ECB_NAN ECB_INFINITY
691 #endif 800 #endif
692 801
693 #if ECB_C99 || _XOPEN_VERSION >= 600 || _POSIX_VERSION >= 200112L 802 #if ECB_C99 || _XOPEN_VERSION >= 600 || _POSIX_VERSION >= 200112L
694 #define ecb_ldexpf(x,e) ldexpf ((x), (e)) 803 #define ecb_ldexpf(x,e) ldexpf ((x), (e))
804 #define ecb_frexpf(x,e) frexpf ((x), (e))
695 #else 805 #else
696 #define ecb_ldexpf(x,e) (float) ldexp ((x), (e)) 806 #define ecb_ldexpf(x,e) (float) ldexp ((double) (x), (e))
807 #define ecb_frexpf(x,e) (float) frexp ((double) (x), (e))
697 #endif 808 #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 }
715 809
716 /* convert a float to ieee single/binary32 */ 810 /* convert a float to ieee single/binary32 */
717 ecb_function_ ecb_const uint32_t ecb_float_to_binary32 (float x); 811 ecb_function_ ecb_const uint32_t ecb_float_to_binary32 (float x);
718 ecb_function_ ecb_const uint32_t 812 ecb_function_ ecb_const uint32_t
719 ecb_float_to_binary32 (float x) 813 ecb_float_to_binary32 (float x)
… …
730 if (x == 0e0f ) return 0x00000000U; 824 if (x == 0e0f ) return 0x00000000U;
731 if (x > +3.40282346638528860e+38f) return 0x7f800000U; 825 if (x > +3.40282346638528860e+38f) return 0x7f800000U;
732 if (x < -3.40282346638528860e+38f) return 0xff800000U; 826 if (x < -3.40282346638528860e+38f) return 0xff800000U;
733 if (x != x ) return 0x7fbfffffU; 827 if (x != x ) return 0x7fbfffffU;
734 828
735 m = frexpf (x, &e) * 0x1000000U; 829 m = ecb_frexpf (x, &e) * 0x1000000U;
736 830
737 r = m & 0x80000000U; 831 r = m & 0x80000000U;
738 832
739 if (r) 833 if (r)
740 m = -m; 834 m = -m;
… …
851 #endif 945 #endif
852 946
853 return r; 947 return r;
854 } 948 }
855 949
856#endif 950 /* convert a float to ieee half/binary16 */
951 ecb_function_ ecb_const uint16_t ecb_float_to_binary16 (float x);
952 ecb_function_ ecb_const uint16_t
953 ecb_float_to_binary16 (float x)
954 {
955 return ecb_binary32_to_binary16 (ecb_float_to_binary32 (x));
956 }
857 957
858#endif 958 /* convert an ieee half/binary16 to float */
959 ecb_function_ ecb_const float ecb_binary16_to_float (uint16_t x);
960 ecb_function_ ecb_const float
961 ecb_binary16_to_float (uint16_t x)
962 {
963 return ecb_binary32_to_float (ecb_binary16_to_binary32 (x));
964 }
859 965
966#endif
967
968#endif
969

Diff Legend

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