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

Comparing libecb/ecb.h (file contents):
Revision 1.105 by root, Thu Jun 28 19:42:30 2012 UTC vs.
Revision 1.126 by root, Sun Jan 5 02:58:29 2014 UTC

1/* 1/*
2 * libecb - http://software.schmorp.de/pkg/libecb 2 * libecb - http://software.schmorp.de/pkg/libecb
3 * 3 *
4 * Copyright (©) 2009-2012 Marc Alexander Lehmann <libecb@schmorp.de> 4 * Copyright (©) 2009-2013 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:
… …
29 29
30#ifndef ECB_H 30#ifndef ECB_H
31#define ECB_H 31#define ECB_H
32 32
33/* 16 bits major, 16 bits minor */ 33/* 16 bits major, 16 bits minor */
34#define ECB_VERSION 0x00010001 34#define ECB_VERSION 0x00010003
35 35
36#ifdef _WIN32 36#ifdef _WIN32
37 typedef signed char int8_t; 37 typedef signed char int8_t;
38 typedef unsigned char uint8_t; 38 typedef unsigned char uint8_t;
39 typedef signed short int16_t; 39 typedef signed short int16_t;
… …
60 #include <inttypes.h> 60 #include <inttypes.h>
61 #if UINTMAX_MAX > 0xffffffffU 61 #if UINTMAX_MAX > 0xffffffffU
62 #define ECB_PTRSIZE 8 62 #define ECB_PTRSIZE 8
63 #else 63 #else
64 #define ECB_PTRSIZE 4 64 #define ECB_PTRSIZE 4
65 #endif
66#endif
67
68/* work around x32 idiocy by defining proper macros */
69#if __amd64 || __x86_64 || _M_AMD64 || _M_X64
70 #if _ILP32
71 #define ECB_AMD64_X32 1
72 #else
73 #define ECB_AMD64 1
65 #endif 74 #endif
66#endif 75#endif
67 76
68/* many compilers define _GNUC_ to some versions but then only implement 77/* many compilers define _GNUC_ to some versions but then only implement
69 * what their idiot authors think are the "more important" extensions, 78 * what their idiot authors think are the "more important" extensions,
… …
125 || defined __ARM_ARCH_6K__ || defined __ARM_ARCH_6ZK__ 134 || defined __ARM_ARCH_6K__ || defined __ARM_ARCH_6ZK__
126 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("mcr p15,0,%0,c7,c10,5" : : "r" (0) : "memory") 135 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("mcr p15,0,%0,c7,c10,5" : : "r" (0) : "memory")
127 #elif defined __ARM_ARCH_7__ || defined __ARM_ARCH_7A__ \ 136 #elif defined __ARM_ARCH_7__ || defined __ARM_ARCH_7A__ \
128 || defined __ARM_ARCH_7M__ || defined __ARM_ARCH_7R__ 137 || defined __ARM_ARCH_7M__ || defined __ARM_ARCH_7R__
129 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("dmb" : : : "memory") 138 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("dmb" : : : "memory")
130 #elif __sparc || __sparc__ 139 #elif (__sparc || __sparc__) && !__sparcv8
131 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("membar #LoadStore | #LoadLoad | #StoreStore | #StoreLoad" : : : "memory") 140 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("membar #LoadStore | #LoadLoad | #StoreStore | #StoreLoad" : : : "memory")
132 #define ECB_MEMORY_FENCE_ACQUIRE __asm__ __volatile__ ("membar #LoadStore | #LoadLoad" : : : "memory") 141 #define ECB_MEMORY_FENCE_ACQUIRE __asm__ __volatile__ ("membar #LoadStore | #LoadLoad" : : : "memory")
133 #define ECB_MEMORY_FENCE_RELEASE __asm__ __volatile__ ("membar #LoadStore | #StoreStore") 142 #define ECB_MEMORY_FENCE_RELEASE __asm__ __volatile__ ("membar #LoadStore | #StoreStore")
134 #elif defined __s390__ || defined __s390x__ 143 #elif defined __s390__ || defined __s390x__
135 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("bcr 15,0" : : : "memory") 144 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("bcr 15,0" : : : "memory")
136 #elif defined __mips__ 145 #elif defined __mips__
146 /* GNU/Linux emulates sync on mips1 architectures, so we force its use */
147 /* anybody else who still uses mips1 is supposed to send in their version, with detection code. */
137 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("sync" : : : "memory") 148 #define ECB_MEMORY_FENCE __asm__ __volatile__ (".set mips2; sync; .set mips0" : : : "memory")
138 #elif defined __alpha__ 149 #elif defined __alpha__
139 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("mb" : : : "memory") 150 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("mb" : : : "memory")
140 #elif defined __hppa__ 151 #elif defined __hppa__
141 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("" : : : "memory") 152 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("" : : : "memory")
142 #define ECB_MEMORY_FENCE_RELEASE __asm__ __volatile__ ("") 153 #define ECB_MEMORY_FENCE_RELEASE __asm__ __volatile__ ("")
143 #elif defined __ia64__ 154 #elif defined __ia64__
144 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("mf" : : : "memory") 155 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("mf" : : : "memory")
156 #elif defined __m68k__
157 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("" : : : "memory")
158 #elif defined __m88k__
159 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("tb1 0,%%r0,128" : : : "memory")
160 #elif defined __sh__
161 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("" : : : "memory")
145 #endif 162 #endif
146 #endif 163 #endif
147#endif 164#endif
148 165
149#ifndef ECB_MEMORY_FENCE 166#ifndef ECB_MEMORY_FENCE
150 #if ECB_GCC_VERSION(4,7) 167 #if ECB_GCC_VERSION(4,7)
151 /* see comment below (stdatomic.h) about the C11 memory model. */ 168 /* see comment below (stdatomic.h) about the C11 memory model. */
152 #define ECB_MEMORY_FENCE __atomic_thread_fence (__ATOMIC_SEQ_CST) 169 #define ECB_MEMORY_FENCE __atomic_thread_fence (__ATOMIC_SEQ_CST)
170
171 /* The __has_feature syntax from clang is so misdesigned that we cannot use it
172 * without risking compile time errors with other compilers. We *could*
173 * define our own ecb_clang_has_feature, but I just can't be bothered to work
174 * around this shit time and again.
153 /*#elif defined __clang && __has_feature (cxx_atomic)*/ 175 * #elif defined __clang && __has_feature (cxx_atomic)
154 /* see comment below (stdatomic.h) about the C11 memory model. */ 176 * // see comment below (stdatomic.h) about the C11 memory model.
155 /*#define ECB_MEMORY_FENCE __c11_atomic_thread_fence (__ATOMIC_SEQ_CST)*/ 177 * #define ECB_MEMORY_FENCE __c11_atomic_thread_fence (__ATOMIC_SEQ_CST)
178 */
179
156 #elif ECB_GCC_VERSION(4,4) || defined __INTEL_COMPILER || defined __clang__ 180 #elif ECB_GCC_VERSION(4,4) || defined __INTEL_COMPILER || defined __clang__
157 #define ECB_MEMORY_FENCE __sync_synchronize () 181 #define ECB_MEMORY_FENCE __sync_synchronize ()
182 #elif _MSC_VER >= 1500 /* VC++ 2008 */
183 /* apparently, microsoft broke all the memory barrier stuff in Visual Studio 2008... */
184 #pragma intrinsic(_ReadBarrier,_WriteBarrier,_ReadWriteBarrier)
185 #define ECB_MEMORY_FENCE _ReadWriteBarrier (); MemoryBarrier()
186 #define ECB_MEMORY_FENCE_ACQUIRE _ReadWriteBarrier (); MemoryBarrier() /* according to msdn, _ReadBarrier is not a load fence */
187 #define ECB_MEMORY_FENCE_RELEASE _WriteBarrier (); MemoryBarrier()
158 #elif _MSC_VER >= 1400 /* VC++ 2005 */ 188 #elif _MSC_VER >= 1400 /* VC++ 2005 */
159 #pragma intrinsic(_ReadBarrier,_WriteBarrier,_ReadWriteBarrier) 189 #pragma intrinsic(_ReadBarrier,_WriteBarrier,_ReadWriteBarrier)
160 #define ECB_MEMORY_FENCE _ReadWriteBarrier () 190 #define ECB_MEMORY_FENCE _ReadWriteBarrier ()
161 #define ECB_MEMORY_FENCE_ACQUIRE _ReadWriteBarrier () /* according to msdn, _ReadBarrier is not a load fence */ 191 #define ECB_MEMORY_FENCE_ACQUIRE _ReadWriteBarrier () /* according to msdn, _ReadBarrier is not a load fence */
162 #define ECB_MEMORY_FENCE_RELEASE _WriteBarrier () 192 #define ECB_MEMORY_FENCE_RELEASE _WriteBarrier ()
… …
475ecb_inline unsigned char ecb_byteorder_helper (void) ecb_const; 505ecb_inline unsigned char ecb_byteorder_helper (void) ecb_const;
476ecb_inline unsigned char 506ecb_inline unsigned char
477ecb_byteorder_helper (void) 507ecb_byteorder_helper (void)
478{ 508{
479 /* the union code still generates code under pressure in gcc, */ 509 /* the union code still generates code under pressure in gcc, */
480 /* but less than using pointers, and always seem to */ 510 /* but less than using pointers, and always seems to */
481 /* successfully return a constant. */ 511 /* successfully return a constant. */
482 /* the reason why we have this horrible preprocessor mess */ 512 /* the reason why we have this horrible preprocessor mess */
483 /* is to avoid it in all cases, at least on common architectures */ 513 /* is to avoid it in all cases, at least on common architectures */
484 /* and yes, gcc defines __BYTE_ORDER__, g++ does not */ 514 /* or when using a recent enough gcc version (>= 4.6) */
485#if __i386 || __i386__ || _M_X86 || __amd64 || __amd64__ || _M_X64 515#if __i386 || __i386__ || _M_X86 || __amd64 || __amd64__ || _M_X64
486 return 0x44; 516 return 0x44;
487#elif __BYTE_ORDER__ && __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__ 517#elif __BYTE_ORDER__ && __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__
488 return 0x44; 518 return 0x44;
489#elif __BYTE_ORDER__ && __BYTE_ORDER__ == __ORDER_BIG_ENDIAN__ 519#elif __BYTE_ORDER__ && __BYTE_ORDER__ == __ORDER_BIG_ENDIAN__
490 retrurn 0x11; 520 return 0x11;
491#else 521#else
492 union 522 union
493 { 523 {
494 uint32_t i; 524 uint32_t i;
495 uint8_t c; 525 uint8_t c;
… …
549 || defined __s390__ || defined __s390x__ \ 579 || defined __s390__ || defined __s390x__ \
550 || defined __mips__ \ 580 || defined __mips__ \
551 || defined __alpha__ \ 581 || defined __alpha__ \
552 || defined __hppa__ \ 582 || defined __hppa__ \
553 || defined __ia64__ \ 583 || defined __ia64__ \
584 || defined __m68k__ \
585 || defined __m88k__ \
586 || defined __sh__ \
554 || defined _M_IX86 || defined _M_AMD64 || defined _M_IA64 587 || defined _M_IX86 || defined _M_AMD64 || defined _M_IA64
555 #define ECB_STDFP 1 588 #define ECB_STDFP 1
556 #include <string.h> /* for memcpy */ 589 #include <string.h> /* for memcpy */
557#else 590#else
558 #define ECB_STDFP 0 591 #define ECB_STDFP 0
559#endif 592#endif
560 593
561#ifndef ECB_NO_LIBM 594#ifndef ECB_NO_LIBM
562 595
563 #if __STDC_IEC_559__ || ECB_STDFP 596 #include <math.h> /* for frexp*, ldexp*, INFINITY, NAN */
564 /* we assume this is defined for most C and many C++ compilers */ 597
565 ecb_inline ecb_bool ecb_float_ieee (void) ecb_const; 598 /* only the oldest of old doesn't have this one. solaris. */
566 ecb_inline ecb_bool ecb_float_ieee (void) { return 1; } 599 #ifdef INFINITY
567 ecb_inline ecb_bool ecb_double_ieee (void) ecb_const; 600 #define ECB_INFINITY INFINITY
568 ecb_inline ecb_bool ecb_double_ieee (void) { return 1; }
569 #elif ECB_CPP
570 #include <limits>
571 ecb_inline ecb_bool ecb_float_ieee (void) ecb_const;
572 ecb_inline ecb_bool ecb_float_ieee (void) { return std::numeric_limits<float >::is_iec559; }
573 ecb_inline ecb_bool ecb_double_ieee (void) ecb_const;
574 ecb_inline ecb_bool ecb_double_ieee (void) { return std::numeric_limits<double>::is_iec559; }
575 #else 601 #else
576 ecb_inline ecb_bool ecb_float_ieee (void) ecb_const; 602 #define ECB_INFINITY HUGE_VAL
577 ecb_inline ecb_bool ecb_float_ieee (void) { return 0; }
578 ecb_inline ecb_bool ecb_double_ieee (void) ecb_const;
579 ecb_inline ecb_bool ecb_double_ieee (void) { return 0; }
580 #endif 603 #endif
604
605 #ifdef NAN
606 #define ECB_NAN NAN
607 #else
608 #define ECB_NAN ECB_INFINITY
609 #endif
610
611 /* converts an ieee half/binary16 to a float */
612 ecb_function_ float ecb_binary16_to_float (uint16_t x) ecb_const;
613 ecb_function_ float
614 ecb_binary16_to_float (uint16_t x)
615 {
616 int e = (x >> 10) & 0x1f;
617 int m = x & 0x3ff;
618 float r;
619
620 if (!e ) r = ldexpf (m , -24);
621 else if (e != 31) r = ldexpf (m + 0x400, e - 25);
622 else if (m ) r = ECB_NAN;
623 else r = ECB_INFINITY;
624
625 return x & 0x8000 ? -r : r;
626 }
581 627
582 /* convert a float to ieee single/binary32 */ 628 /* convert a float to ieee single/binary32 */
583 ecb_function_ uint32_t ecb_float_to_binary32 (float x) ecb_const; 629 ecb_function_ uint32_t ecb_float_to_binary32 (float x) ecb_const;
584 ecb_function_ uint32_t 630 ecb_function_ uint32_t
585 ecb_float_to_binary32 (float x) 631 ecb_float_to_binary32 (float x)
… …
588 634
589 #if ECB_STDFP 635 #if ECB_STDFP
590 memcpy (&r, &x, 4); 636 memcpy (&r, &x, 4);
591 #else 637 #else
592 /* slow emulation, works for anything but -0 */ 638 /* slow emulation, works for anything but -0 */
593 ECB_EXTERN_C float frexpf (float v, int *e);
594 uint32_t m; 639 uint32_t m;
595 int e; 640 int e;
596 641
597 if (x == 0e0f ) return 0; 642 if (x == 0e0f ) return 0x00000000U;
598 if (x > +3.40282346638528860e+38f) return 0x7f800000U; 643 if (x > +3.40282346638528860e+38f) return 0x7f800000U;
599 if (x < -3.40282346638528860e+38f) return 0xff800000U; 644 if (x < -3.40282346638528860e+38f) return 0xff800000U;
600 if (x != x ) return 0x7fbfffffU; 645 if (x != x ) return 0x7fbfffffU;
601 646
602 m = frexpf (x, &e) * 0x1000000U; 647 m = frexpf (x, &e) * 0x1000000U;
… …
604 r = m & 0x80000000U; 649 r = m & 0x80000000U;
605 650
606 if (r) 651 if (r)
607 m = -m; 652 m = -m;
608 653
609 if (e < -125) 654 if (e <= -126)
610 { 655 {
611 m &= 0xffffffU; 656 m &= 0xffffffU;
612 m >>= (-125 - e); 657 m >>= (-125 - e);
613 e = -126; 658 e = -126;
614 } 659 }
… …
629 674
630 #if ECB_STDFP 675 #if ECB_STDFP
631 memcpy (&r, &x, 4); 676 memcpy (&r, &x, 4);
632 #else 677 #else
633 /* emulation, only works for normals and subnormals and +0 */ 678 /* emulation, only works for normals and subnormals and +0 */
634 ECB_EXTERN_C float ldexpf (float x, int e);
635
636 int neg = x >> 31; 679 int neg = x >> 31;
637 int e = (x >> 23) & 0xffU; 680 int e = (x >> 23) & 0xffU;
638 681
639 x &= 0x7fffffU; 682 x &= 0x7fffffU;
640 683
… …
642 x |= 0x800000U; 685 x |= 0x800000U;
643 else 686 else
644 e = 1; 687 e = 1;
645 688
646 /* we distrust ldexpf a bit and do the 2**-24 scaling by an extra multiply */ 689 /* we distrust ldexpf a bit and do the 2**-24 scaling by an extra multiply */
647 r = ldexpf (x * (1.f / 0x1000000U), e - 126); 690 r = ldexpf (x * (0.5f / 0x800000U), e - 126);
648 691
649 r = neg ? -r : r; 692 r = neg ? -r : r;
650 #endif 693 #endif
651 694
652 return r; 695 return r;
… …
661 704
662 #if ECB_STDFP 705 #if ECB_STDFP
663 memcpy (&r, &x, 8); 706 memcpy (&r, &x, 8);
664 #else 707 #else
665 /* slow emulation, works for anything but -0 */ 708 /* slow emulation, works for anything but -0 */
666 ECB_EXTERN_C double frexp (double v, int *e);
667 uint64_t m; 709 uint64_t m;
668 int e; 710 int e;
669 711
670 if (x == 0e0 ) return 0; 712 if (x == 0e0 ) return 0x0000000000000000U;
671 if (x > +1.79769313486231470e+308) return 0x7ff0000000000000U; 713 if (x > +1.79769313486231470e+308) return 0x7ff0000000000000U;
672 if (x < -1.79769313486231470e+308) return 0xfff0000000000000U; 714 if (x < -1.79769313486231470e+308) return 0xfff0000000000000U;
673 if (x != x ) return 0X7ff7ffffffffffffU; 715 if (x != x ) return 0X7ff7ffffffffffffU;
674 716
675 m = frexpf (x, &e) * 0x1000000U;
676 m = frexp (x, &e) * 0x20000000000000U; 717 m = frexp (x, &e) * 0x20000000000000U;
677 718
678 r = m & 0x8000000000000000;; 719 r = m & 0x8000000000000000;;
679 720
680 if (r) 721 if (r)
681 m = -m; 722 m = -m;
682 723
683 if (e < -1021) 724 if (e <= -1022)
684 { 725 {
685 m &= 0x1fffffffffffffU; 726 m &= 0x1fffffffffffffU;
686 m >>= (-1021 - e); 727 m >>= (-1021 - e);
687 e = -1022; 728 e = -1022;
688 } 729 }
… …
703 744
704 #if ECB_STDFP 745 #if ECB_STDFP
705 memcpy (&r, &x, 8); 746 memcpy (&r, &x, 8);
706 #else 747 #else
707 /* emulation, only works for normals and subnormals and +0 */ 748 /* emulation, only works for normals and subnormals and +0 */
708 ECB_EXTERN_C double ldexp (double x, int e);
709
710 int neg = x >> 63; 749 int neg = x >> 63;
711 int e = (x >> 52) & 0x7ffU; 750 int e = (x >> 52) & 0x7ffU;
712 751
713 x &= 0xfffffffffffffU; 752 x &= 0xfffffffffffffU;
714 753
715 if (e) 754 if (e)
716 x |= 0x10000000000000U; 755 x |= 0x10000000000000U;
717 else 756 else
718 e = 1; 757 e = 1;
719 758
720 /* we distrust ldexpf a bit and do the 2**-53 scaling by an extra multiply */ 759 /* we distrust ldexp a bit and do the 2**-53 scaling by an extra multiply */
721 r = ldexp (x * (1. / 0x20000000000000U), e - 1022); 760 r = ldexp (x * (0.5 / 0x10000000000000U), e - 1022);
722 761
723 r = neg ? -r : r; 762 r = neg ? -r : r;
724 #endif 763 #endif
725 764
726 return r; 765 return r;

Diff Legend

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