ViewVC Help
View File | Revision Log | Show Annotations | Download File
/cvs/libev/ev.c
(Generate patch)

Comparing libev/ev.c (file contents):
Revision 1.472 by root, Tue Sep 9 13:24:13 2014 UTC vs.
Revision 1.479 by root, Sun Dec 20 01:31:17 2015 UTC

… …
491/* the following is ecb.h embedded into libev - use update_ev_c to update from an external copy */ 491/* the following is ecb.h embedded into libev - use update_ev_c to update from an external copy */
492/* ECB.H BEGIN */ 492/* ECB.H BEGIN */
493/* 493/*
494 * libecb - http://software.schmorp.de/pkg/libecb 494 * libecb - http://software.schmorp.de/pkg/libecb
495 * 495 *
496 * Copyright (©) 2009-2014 Marc Alexander Lehmann <libecb@schmorp.de> 496 * Copyright (©) 2009-2015 Marc Alexander Lehmann <libecb@schmorp.de>
497 * Copyright (©) 2011 Emanuele Giaquinta 497 * Copyright (©) 2011 Emanuele Giaquinta
498 * All rights reserved. 498 * All rights reserved.
499 * 499 *
500 * Redistribution and use in source and binary forms, with or without modifica- 500 * Redistribution and use in source and binary forms, with or without modifica-
501 * tion, are permitted provided that the following conditions are met: 501 * tion, are permitted provided that the following conditions are met:
… …
532 532
533#ifndef ECB_H 533#ifndef ECB_H
534#define ECB_H 534#define ECB_H
535 535
536/* 16 bits major, 16 bits minor */ 536/* 16 bits major, 16 bits minor */
537#define ECB_VERSION 0x00010003 537#define ECB_VERSION 0x00010005
538 538
539#ifdef _WIN32 539#ifdef _WIN32
540 typedef signed char int8_t; 540 typedef signed char int8_t;
541 typedef unsigned char uint8_t; 541 typedef unsigned char uint8_t;
542 typedef signed short int16_t; 542 typedef signed short int16_t;
… …
559 typedef uint32_t uintptr_t; 559 typedef uint32_t uintptr_t;
560 typedef int32_t intptr_t; 560 typedef int32_t intptr_t;
561 #endif 561 #endif
562#else 562#else
563 #include <inttypes.h> 563 #include <inttypes.h>
564 #if UINTMAX_MAX > 0xffffffffU 564 #if (defined INTPTR_MAX ? INTPTR_MAX : ULONG_MAX) > 0xffffffffU
565 #define ECB_PTRSIZE 8 565 #define ECB_PTRSIZE 8
566 #else 566 #else
567 #define ECB_PTRSIZE 4 567 #define ECB_PTRSIZE 4
568 #endif 568 #endif
569#endif 569#endif
570 570
571#define ECB_GCC_AMD64 (__amd64 || __amd64__ || __x86_64 || __x86_64__)
572#define ECB_MSVC_AMD64 (_M_AMD64 || _M_X64)
573
571/* work around x32 idiocy by defining proper macros */ 574/* work around x32 idiocy by defining proper macros */
572#if __amd64 || __x86_64 || _M_AMD64 || _M_X64 575#if ECB_GCC_AMD64 || ECB_MSVC_AMD64
573 #if _ILP32 576 #if _ILP32
574 #define ECB_AMD64_X32 1 577 #define ECB_AMD64_X32 1
575 #else 578 #else
576 #define ECB_AMD64 1 579 #define ECB_AMD64 1
577 #endif 580 #endif
… …
582 * causing enormous grief in return for some better fake benchmark numbers. 585 * causing enormous grief in return for some better fake benchmark numbers.
583 * or so. 586 * or so.
584 * we try to detect these and simply assume they are not gcc - if they have 587 * we try to detect these and simply assume they are not gcc - if they have
585 * an issue with that they should have done it right in the first place. 588 * an issue with that they should have done it right in the first place.
586 */ 589 */
587#ifndef ECB_GCC_VERSION
588 #if !defined __GNUC_MINOR__ || defined __INTEL_COMPILER || defined __SUNPRO_C || defined __SUNPRO_CC || defined __llvm__ || defined __clang__ 590#if !defined __GNUC_MINOR__ || defined __INTEL_COMPILER || defined __SUNPRO_C || defined __SUNPRO_CC || defined __llvm__ || defined __clang__
589 #define ECB_GCC_VERSION(major,minor) 0 591 #define ECB_GCC_VERSION(major,minor) 0
590 #else 592#else
591 #define ECB_GCC_VERSION(major,minor) (__GNUC__ > (major) || (__GNUC__ == (major) && __GNUC_MINOR__ >= (minor))) 593 #define ECB_GCC_VERSION(major,minor) (__GNUC__ > (major) || (__GNUC__ == (major) && __GNUC_MINOR__ >= (minor)))
592 #endif 594#endif
595
596#define ECB_CLANG_VERSION(major,minor) (__clang_major__ > (major) || (__clang_major__ == (major) && __clang_minor__ >= (minor)))
597
598#if __clang__ && defined __has_builtin
599 #define ECB_CLANG_BUILTIN(x) __has_builtin (x)
600#else
601 #define ECB_CLANG_BUILTIN(x) 0
602#endif
603
604#if __clang__ && defined __has_extension
605 #define ECB_CLANG_EXTENSION(x) __has_extension (x)
606#else
607 #define ECB_CLANG_EXTENSION(x) 0
593#endif 608#endif
594 609
595#define ECB_CPP (__cplusplus+0) 610#define ECB_CPP (__cplusplus+0)
596#define ECB_CPP11 (__cplusplus >= 201103L) 611#define ECB_CPP11 (__cplusplus >= 201103L)
597 612
… …
625 #define ECB_NO_SMP 1 640 #define ECB_NO_SMP 1
626#endif 641#endif
627 642
628#if ECB_NO_SMP 643#if ECB_NO_SMP
629 #define ECB_MEMORY_FENCE do { } while (0) 644 #define ECB_MEMORY_FENCE do { } while (0)
645#endif
646
647/* http://www-01.ibm.com/support/knowledgecenter/SSGH3R_13.1.0/com.ibm.xlcpp131.aix.doc/compiler_ref/compiler_builtins.html */
648#if __xlC__ && ECB_CPP
649 #include <builtins.h>
650#endif
651
652#if 1400 <= _MSC_VER
653 #include <intrin.h> /* fence functions _ReadBarrier, also bit search functions _BitScanReverse */
630#endif 654#endif
631 655
632#ifndef ECB_MEMORY_FENCE 656#ifndef ECB_MEMORY_FENCE
633 #if ECB_GCC_VERSION(2,5) || defined __INTEL_COMPILER || (__llvm__ && __GNUC__) || __SUNPRO_C >= 0x5110 || __SUNPRO_CC >= 0x5110 657 #if ECB_GCC_VERSION(2,5) || defined __INTEL_COMPILER || (__llvm__ && __GNUC__) || __SUNPRO_C >= 0x5110 || __SUNPRO_CC >= 0x5110
634 #if __i386 || __i386__ 658 #if __i386 || __i386__
635 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("lock; orb $0, -1(%%esp)" : : : "memory") 659 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("lock; orb $0, -1(%%esp)" : : : "memory")
636 #define ECB_MEMORY_FENCE_ACQUIRE __asm__ __volatile__ ("" : : : "memory") 660 #define ECB_MEMORY_FENCE_ACQUIRE __asm__ __volatile__ ("" : : : "memory")
637 #define ECB_MEMORY_FENCE_RELEASE __asm__ __volatile__ ("") 661 #define ECB_MEMORY_FENCE_RELEASE __asm__ __volatile__ ("")
638 #elif __amd64 || __amd64__ || __x86_64 || __x86_64__ 662 #elif ECB_GCC_AMD64
639 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("mfence" : : : "memory") 663 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("mfence" : : : "memory")
640 #define ECB_MEMORY_FENCE_ACQUIRE __asm__ __volatile__ ("" : : : "memory") 664 #define ECB_MEMORY_FENCE_ACQUIRE __asm__ __volatile__ ("" : : : "memory")
641 #define ECB_MEMORY_FENCE_RELEASE __asm__ __volatile__ ("") 665 #define ECB_MEMORY_FENCE_RELEASE __asm__ __volatile__ ("")
642 #elif __powerpc__ || __ppc__ || __powerpc64__ || __ppc64__ 666 #elif __powerpc__ || __ppc__ || __powerpc64__ || __ppc64__
643 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("sync" : : : "memory") 667 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("sync" : : : "memory")
668 #elif defined __ARM_ARCH_2__ \
669 || defined __ARM_ARCH_3__ || defined __ARM_ARCH_3M__ \
670 || defined __ARM_ARCH_4__ || defined __ARM_ARCH_4T__ \
671 || defined __ARM_ARCH_5__ || defined __ARM_ARCH_5E__ \
672 || defined __ARM_ARCH_5T__ || defined __ARM_ARCH_5TE__ \
673 || defined __ARM_ARCH_5TEJ__
674 /* should not need any, unless running old code on newer cpu - arm doesn't support that */
644 #elif defined __ARM_ARCH_6__ || defined __ARM_ARCH_6J__ \ 675 #elif defined __ARM_ARCH_6__ || defined __ARM_ARCH_6J__ \
645 || defined __ARM_ARCH_6K__ || defined __ARM_ARCH_6ZK__ 676 || defined __ARM_ARCH_6K__ || defined __ARM_ARCH_6ZK__ \
677 || defined __ARM_ARCH_6T2__
646 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("mcr p15,0,%0,c7,c10,5" : : "r" (0) : "memory") 678 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("mcr p15,0,%0,c7,c10,5" : : "r" (0) : "memory")
647 #elif defined __ARM_ARCH_7__ || defined __ARM_ARCH_7A__ \ 679 #elif defined __ARM_ARCH_7__ || defined __ARM_ARCH_7A__ \
648 || defined __ARM_ARCH_7M__ || defined __ARM_ARCH_7R__ 680 || defined __ARM_ARCH_7R__ || defined __ARM_ARCH_7M__
649 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("dmb" : : : "memory") 681 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("dmb" : : : "memory")
650 #elif __aarch64__ 682 #elif __aarch64__
651 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("dmb ish" : : : "memory") 683 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("dmb ish" : : : "memory")
652 #elif (__sparc || __sparc__) && !__sparcv8 684 #elif (__sparc || __sparc__) && !(__sparc_v8__ || defined __sparcv8)
653 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("membar #LoadStore | #LoadLoad | #StoreStore | #StoreLoad" : : : "memory") 685 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("membar #LoadStore | #LoadLoad | #StoreStore | #StoreLoad" : : : "memory")
654 #define ECB_MEMORY_FENCE_ACQUIRE __asm__ __volatile__ ("membar #LoadStore | #LoadLoad" : : : "memory") 686 #define ECB_MEMORY_FENCE_ACQUIRE __asm__ __volatile__ ("membar #LoadStore | #LoadLoad" : : : "memory")
655 #define ECB_MEMORY_FENCE_RELEASE __asm__ __volatile__ ("membar #LoadStore | #StoreStore") 687 #define ECB_MEMORY_FENCE_RELEASE __asm__ __volatile__ ("membar #LoadStore | #StoreStore")
656 #elif defined __s390__ || defined __s390x__ 688 #elif defined __s390__ || defined __s390x__
657 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("bcr 15,0" : : : "memory") 689 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("bcr 15,0" : : : "memory")
… …
681 /* see comment below (stdatomic.h) about the C11 memory model. */ 713 /* see comment below (stdatomic.h) about the C11 memory model. */
682 #define ECB_MEMORY_FENCE __atomic_thread_fence (__ATOMIC_SEQ_CST) 714 #define ECB_MEMORY_FENCE __atomic_thread_fence (__ATOMIC_SEQ_CST)
683 #define ECB_MEMORY_FENCE_ACQUIRE __atomic_thread_fence (__ATOMIC_ACQUIRE) 715 #define ECB_MEMORY_FENCE_ACQUIRE __atomic_thread_fence (__ATOMIC_ACQUIRE)
684 #define ECB_MEMORY_FENCE_RELEASE __atomic_thread_fence (__ATOMIC_RELEASE) 716 #define ECB_MEMORY_FENCE_RELEASE __atomic_thread_fence (__ATOMIC_RELEASE)
685 717
686 /* The __has_feature syntax from clang is so misdesigned that we cannot use it 718 #elif ECB_CLANG_EXTENSION(c_atomic)
687 * without risking compile time errors with other compilers. We *could*
688 * define our own ecb_clang_has_feature, but I just can't be bothered to work
689 * around this shit time and again.
690 * #elif defined __clang && __has_feature (cxx_atomic)
691 * // see comment below (stdatomic.h) about the C11 memory model. 719 /* see comment below (stdatomic.h) about the C11 memory model. */
692 * #define ECB_MEMORY_FENCE __c11_atomic_thread_fence (__ATOMIC_SEQ_CST) 720 #define ECB_MEMORY_FENCE __c11_atomic_thread_fence (__ATOMIC_SEQ_CST)
693 * #define ECB_MEMORY_FENCE_ACQUIRE __c11_atomic_thread_fence (__ATOMIC_ACQUIRE) 721 #define ECB_MEMORY_FENCE_ACQUIRE __c11_atomic_thread_fence (__ATOMIC_ACQUIRE)
694 * #define ECB_MEMORY_FENCE_RELEASE __c11_atomic_thread_fence (__ATOMIC_RELEASE) 722 #define ECB_MEMORY_FENCE_RELEASE __c11_atomic_thread_fence (__ATOMIC_RELEASE)
695 */
696 723
697 #elif ECB_GCC_VERSION(4,4) || defined __INTEL_COMPILER || defined __clang__ 724 #elif ECB_GCC_VERSION(4,4) || defined __INTEL_COMPILER || defined __clang__
698 #define ECB_MEMORY_FENCE __sync_synchronize () 725 #define ECB_MEMORY_FENCE __sync_synchronize ()
699 #elif _MSC_VER >= 1500 /* VC++ 2008 */ 726 #elif _MSC_VER >= 1500 /* VC++ 2008 */
700 /* apparently, microsoft broke all the memory barrier stuff in Visual Studio 2008... */ 727 /* apparently, microsoft broke all the memory barrier stuff in Visual Studio 2008... */
… …
763 #define ECB_MEMORY_FENCE_RELEASE ECB_MEMORY_FENCE 790 #define ECB_MEMORY_FENCE_RELEASE ECB_MEMORY_FENCE
764#endif 791#endif
765 792
766/*****************************************************************************/ 793/*****************************************************************************/
767 794
768#if __cplusplus 795#if ECB_CPP
769 #define ecb_inline static inline 796 #define ecb_inline static inline
770#elif ECB_GCC_VERSION(2,5) 797#elif ECB_GCC_VERSION(2,5)
771 #define ecb_inline static __inline__ 798 #define ecb_inline static __inline__
772#elif ECB_C99 799#elif ECB_C99
773 #define ecb_inline static inline 800 #define ecb_inline static inline
… …
787 814
788#define ECB_CONCAT_(a, b) a ## b 815#define ECB_CONCAT_(a, b) a ## b
789#define ECB_CONCAT(a, b) ECB_CONCAT_(a, b) 816#define ECB_CONCAT(a, b) ECB_CONCAT_(a, b)
790#define ECB_STRINGIFY_(a) # a 817#define ECB_STRINGIFY_(a) # a
791#define ECB_STRINGIFY(a) ECB_STRINGIFY_(a) 818#define ECB_STRINGIFY(a) ECB_STRINGIFY_(a)
819#define ECB_STRINGIFY_EXPR(expr) ((expr), ECB_STRINGIFY_ (expr))
792 820
793#define ecb_function_ ecb_inline 821#define ecb_function_ ecb_inline
794 822
795#if ECB_GCC_VERSION(3,1) 823#if ECB_GCC_VERSION(3,1) || ECB_CLANG_VERSION(2,8)
796 #define ecb_attribute(attrlist) __attribute__(attrlist) 824 #define ecb_attribute(attrlist) __attribute__ (attrlist)
825#else
826 #define ecb_attribute(attrlist)
827#endif
828
829#if ECB_GCC_VERSION(3,1) || ECB_CLANG_BUILTIN(__builtin_constant_p)
797 #define ecb_is_constant(expr) __builtin_constant_p (expr) 830 #define ecb_is_constant(expr) __builtin_constant_p (expr)
798 #define ecb_expect(expr,value) __builtin_expect ((expr),(value))
799 #define ecb_prefetch(addr,rw,locality) __builtin_prefetch (addr, rw, locality)
800#else 831#else
801 #define ecb_attribute(attrlist)
802
803 /* possible C11 impl for integral types 832 /* possible C11 impl for integral types
804 typedef struct ecb_is_constant_struct ecb_is_constant_struct; 833 typedef struct ecb_is_constant_struct ecb_is_constant_struct;
805 #define ecb_is_constant(expr) _Generic ((1 ? (struct ecb_is_constant_struct *)0 : (void *)((expr) - (expr)), ecb_is_constant_struct *: 0, default: 1)) */ 834 #define ecb_is_constant(expr) _Generic ((1 ? (struct ecb_is_constant_struct *)0 : (void *)((expr) - (expr)), ecb_is_constant_struct *: 0, default: 1)) */
806 835
807 #define ecb_is_constant(expr) 0 836 #define ecb_is_constant(expr) 0
837#endif
838
839#if ECB_GCC_VERSION(3,1) || ECB_CLANG_BUILTIN(__builtin_expect)
840 #define ecb_expect(expr,value) __builtin_expect ((expr),(value))
841#else
808 #define ecb_expect(expr,value) (expr) 842 #define ecb_expect(expr,value) (expr)
843#endif
844
845#if ECB_GCC_VERSION(3,1) || ECB_CLANG_BUILTIN(__builtin_prefetch)
846 #define ecb_prefetch(addr,rw,locality) __builtin_prefetch (addr, rw, locality)
847#else
809 #define ecb_prefetch(addr,rw,locality) 848 #define ecb_prefetch(addr,rw,locality)
810#endif 849#endif
811 850
812/* no emulation for ecb_decltype */ 851/* no emulation for ecb_decltype */
813#if ECB_GCC_VERSION(4,5) 852#if ECB_CPP11
853 // older implementations might have problems with decltype(x)::type, work around it
854 template<class T> struct ecb_decltype_t { typedef T type; };
814 #define ecb_decltype(x) __decltype(x) 855 #define ecb_decltype(x) ecb_decltype_t<decltype (x)>::type
815#elif ECB_GCC_VERSION(3,0) 856#elif ECB_GCC_VERSION(3,0) || ECB_CLANG_VERSION(2,8)
816 #define ecb_decltype(x) __typeof(x) 857 #define ecb_decltype(x) __typeof__ (x)
817#endif 858#endif
818 859
819#if _MSC_VER >= 1300 860#if _MSC_VER >= 1300
820 #define ecb_deprecated __declspec(deprecated) 861 #define ecb_deprecated __declspec (deprecated)
821#else 862#else
822 #define ecb_deprecated ecb_attribute ((__deprecated__)) 863 #define ecb_deprecated ecb_attribute ((__deprecated__))
823#endif 864#endif
824 865
866#if _MSC_VER >= 1500
867 #define ecb_deprecated_message(msg) __declspec (deprecated (msg))
868#elif ECB_GCC_VERSION(4,5)
869 #define ecb_deprecated_message(msg) ecb_attribute ((__deprecated__ (msg))
870#else
871 #define ecb_deprecated_message(msg) ecb_deprecated
872#endif
873
874#if _MSC_VER >= 1400
875 #define ecb_noinline __declspec (noinline)
876#else
825#define ecb_noinline ecb_attribute ((__noinline__)) 877 #define ecb_noinline ecb_attribute ((__noinline__))
878#endif
879
826#define ecb_unused ecb_attribute ((__unused__)) 880#define ecb_unused ecb_attribute ((__unused__))
827#define ecb_const ecb_attribute ((__const__)) 881#define ecb_const ecb_attribute ((__const__))
828#define ecb_pure ecb_attribute ((__pure__)) 882#define ecb_pure ecb_attribute ((__pure__))
829 883
830/* http://msdn.microsoft.com/en-us/library/k6ktzx3s.aspx __declspec(noreturn) */ 884#if ECB_C11 || __IBMC_NORETURN
831#if ECB_C11 885 /* http://www-01.ibm.com/support/knowledgecenter/SSGH3R_13.1.0/com.ibm.xlcpp131.aix.doc/language_ref/noreturn.html */
832 #define ecb_noreturn _Noreturn 886 #define ecb_noreturn _Noreturn
887#elif ECB_CPP11
888 #define ecb_noreturn [[noreturn]]
889#elif _MSC_VER >= 1200
890 /* http://msdn.microsoft.com/en-us/library/k6ktzx3s.aspx */
891 #define ecb_noreturn __declspec (noreturn)
833#else 892#else
834 #define ecb_noreturn ecb_attribute ((__noreturn__)) 893 #define ecb_noreturn ecb_attribute ((__noreturn__))
835#endif 894#endif
836 895
837#if ECB_GCC_VERSION(4,3) 896#if ECB_GCC_VERSION(4,3)
… …
852/* for compatibility to the rest of the world */ 911/* for compatibility to the rest of the world */
853#define ecb_likely(expr) ecb_expect_true (expr) 912#define ecb_likely(expr) ecb_expect_true (expr)
854#define ecb_unlikely(expr) ecb_expect_false (expr) 913#define ecb_unlikely(expr) ecb_expect_false (expr)
855 914
856/* count trailing zero bits and count # of one bits */ 915/* count trailing zero bits and count # of one bits */
857#if ECB_GCC_VERSION(3,4) 916#if ECB_GCC_VERSION(3,4) \
917 || (ECB_CLANG_BUILTIN(__builtin_clz) && ECB_CLANG_BUILTIN(__builtin_clzll) \
918 && ECB_CLANG_BUILTIN(__builtin_ctz) && ECB_CLANG_BUILTIN(__builtin_ctzll) \
919 && ECB_CLANG_BUILTIN(__builtin_popcount))
858 /* we assume int == 32 bit, long == 32 or 64 bit and long long == 64 bit */ 920 /* we assume int == 32 bit, long == 32 or 64 bit and long long == 64 bit */
859 #define ecb_ld32(x) (__builtin_clz (x) ^ 31) 921 #define ecb_ld32(x) (__builtin_clz (x) ^ 31)
860 #define ecb_ld64(x) (__builtin_clzll (x) ^ 63) 922 #define ecb_ld64(x) (__builtin_clzll (x) ^ 63)
861 #define ecb_ctz32(x) __builtin_ctz (x) 923 #define ecb_ctz32(x) __builtin_ctz (x)
862 #define ecb_ctz64(x) __builtin_ctzll (x) 924 #define ecb_ctz64(x) __builtin_ctzll (x)
863 #define ecb_popcount32(x) __builtin_popcount (x) 925 #define ecb_popcount32(x) __builtin_popcount (x)
864 /* no popcountll */ 926 /* no popcountll */
865#else 927#else
866 ecb_function_ int ecb_ctz32 (uint32_t x) ecb_const; 928 ecb_function_ ecb_const int ecb_ctz32 (uint32_t x);
867 ecb_function_ int 929 ecb_function_ ecb_const int
868 ecb_ctz32 (uint32_t x) 930 ecb_ctz32 (uint32_t x)
869 { 931 {
932#if 1400 <= _MSC_VER && (_M_IX86 || _M_X64 || _M_IA64 || _M_ARM)
933 unsigned long r;
934 _BitScanForward (&r, x);
935 return (int)r;
936#else
870 int r = 0; 937 int r = 0;
871 938
872 x &= ~x + 1; /* this isolates the lowest bit */ 939 x &= ~x + 1; /* this isolates the lowest bit */
873 940
874#if ECB_branchless_on_i386 941#if ECB_branchless_on_i386
… …
884 if (x & 0xff00ff00) r += 8; 951 if (x & 0xff00ff00) r += 8;
885 if (x & 0xffff0000) r += 16; 952 if (x & 0xffff0000) r += 16;
886#endif 953#endif
887 954
888 return r; 955 return r;
956#endif
889 } 957 }
890 958
891 ecb_function_ int ecb_ctz64 (uint64_t x) ecb_const; 959 ecb_function_ ecb_const int ecb_ctz64 (uint64_t x);
892 ecb_function_ int 960 ecb_function_ ecb_const int
893 ecb_ctz64 (uint64_t x) 961 ecb_ctz64 (uint64_t x)
894 { 962 {
963#if 1400 <= _MSC_VER && (_M_X64 || _M_IA64 || _M_ARM)
964 unsigned long r;
965 _BitScanForward64 (&r, x);
966 return (int)r;
967#else
895 int shift = x & 0xffffffffU ? 0 : 32; 968 int shift = x & 0xffffffff ? 0 : 32;
896 return ecb_ctz32 (x >> shift) + shift; 969 return ecb_ctz32 (x >> shift) + shift;
970#endif
897 } 971 }
898 972
899 ecb_function_ int ecb_popcount32 (uint32_t x) ecb_const; 973 ecb_function_ ecb_const int ecb_popcount32 (uint32_t x);
900 ecb_function_ int 974 ecb_function_ ecb_const int
901 ecb_popcount32 (uint32_t x) 975 ecb_popcount32 (uint32_t x)
902 { 976 {
903 x -= (x >> 1) & 0x55555555; 977 x -= (x >> 1) & 0x55555555;
904 x = ((x >> 2) & 0x33333333) + (x & 0x33333333); 978 x = ((x >> 2) & 0x33333333) + (x & 0x33333333);
905 x = ((x >> 4) + x) & 0x0f0f0f0f; 979 x = ((x >> 4) + x) & 0x0f0f0f0f;
906 x *= 0x01010101; 980 x *= 0x01010101;
907 981
908 return x >> 24; 982 return x >> 24;
909 } 983 }
910 984
911 ecb_function_ int ecb_ld32 (uint32_t x) ecb_const; 985 ecb_function_ ecb_const int ecb_ld32 (uint32_t x);
912 ecb_function_ int ecb_ld32 (uint32_t x) 986 ecb_function_ ecb_const int ecb_ld32 (uint32_t x)
913 { 987 {
988#if 1400 <= _MSC_VER && (_M_IX86 || _M_X64 || _M_IA64 || _M_ARM)
989 unsigned long r;
990 _BitScanReverse (&r, x);
991 return (int)r;
992#else
914 int r = 0; 993 int r = 0;
915 994
916 if (x >> 16) { x >>= 16; r += 16; } 995 if (x >> 16) { x >>= 16; r += 16; }
917 if (x >> 8) { x >>= 8; r += 8; } 996 if (x >> 8) { x >>= 8; r += 8; }
918 if (x >> 4) { x >>= 4; r += 4; } 997 if (x >> 4) { x >>= 4; r += 4; }
919 if (x >> 2) { x >>= 2; r += 2; } 998 if (x >> 2) { x >>= 2; r += 2; }
920 if (x >> 1) { r += 1; } 999 if (x >> 1) { r += 1; }
921 1000
922 return r; 1001 return r;
1002#endif
923 } 1003 }
924 1004
925 ecb_function_ int ecb_ld64 (uint64_t x) ecb_const; 1005 ecb_function_ ecb_const int ecb_ld64 (uint64_t x);
926 ecb_function_ int ecb_ld64 (uint64_t x) 1006 ecb_function_ ecb_const int ecb_ld64 (uint64_t x)
927 { 1007 {
1008#if 1400 <= _MSC_VER && (_M_X64 || _M_IA64 || _M_ARM)
1009 unsigned long r;
1010 _BitScanReverse64 (&r, x);
1011 return (int)r;
1012#else
928 int r = 0; 1013 int r = 0;
929 1014
930 if (x >> 32) { x >>= 32; r += 32; } 1015 if (x >> 32) { x >>= 32; r += 32; }
931 1016
932 return r + ecb_ld32 (x); 1017 return r + ecb_ld32 (x);
1018#endif
933 } 1019 }
934#endif 1020#endif
935 1021
936ecb_function_ ecb_bool ecb_is_pot32 (uint32_t x) ecb_const; 1022ecb_function_ ecb_const ecb_bool ecb_is_pot32 (uint32_t x);
937ecb_function_ ecb_bool ecb_is_pot32 (uint32_t x) { return !(x & (x - 1)); } 1023ecb_function_ ecb_const ecb_bool ecb_is_pot32 (uint32_t x) { return !(x & (x - 1)); }
938ecb_function_ ecb_bool ecb_is_pot64 (uint64_t x) ecb_const; 1024ecb_function_ ecb_const ecb_bool ecb_is_pot64 (uint64_t x);
939ecb_function_ ecb_bool ecb_is_pot64 (uint64_t x) { return !(x & (x - 1)); } 1025ecb_function_ ecb_const ecb_bool ecb_is_pot64 (uint64_t x) { return !(x & (x - 1)); }
940 1026
941ecb_function_ uint8_t ecb_bitrev8 (uint8_t x) ecb_const; 1027ecb_function_ ecb_const uint8_t ecb_bitrev8 (uint8_t x);
942ecb_function_ uint8_t ecb_bitrev8 (uint8_t x) 1028ecb_function_ ecb_const uint8_t ecb_bitrev8 (uint8_t x)
943{ 1029{
944 return ( (x * 0x0802U & 0x22110U) 1030 return ( (x * 0x0802U & 0x22110U)
945 | (x * 0x8020U & 0x88440U)) * 0x10101U >> 16; 1031 | (x * 0x8020U & 0x88440U)) * 0x10101U >> 16;
946} 1032}
947 1033
948ecb_function_ uint16_t ecb_bitrev16 (uint16_t x) ecb_const; 1034ecb_function_ ecb_const uint16_t ecb_bitrev16 (uint16_t x);
949ecb_function_ uint16_t ecb_bitrev16 (uint16_t x) 1035ecb_function_ ecb_const uint16_t ecb_bitrev16 (uint16_t x)
950{ 1036{
951 x = ((x >> 1) & 0x5555) | ((x & 0x5555) << 1); 1037 x = ((x >> 1) & 0x5555) | ((x & 0x5555) << 1);
952 x = ((x >> 2) & 0x3333) | ((x & 0x3333) << 2); 1038 x = ((x >> 2) & 0x3333) | ((x & 0x3333) << 2);
953 x = ((x >> 4) & 0x0f0f) | ((x & 0x0f0f) << 4); 1039 x = ((x >> 4) & 0x0f0f) | ((x & 0x0f0f) << 4);
954 x = ( x >> 8 ) | ( x << 8); 1040 x = ( x >> 8 ) | ( x << 8);
955 1041
956 return x; 1042 return x;
957} 1043}
958 1044
959ecb_function_ uint32_t ecb_bitrev32 (uint32_t x) ecb_const; 1045ecb_function_ ecb_const uint32_t ecb_bitrev32 (uint32_t x);
960ecb_function_ uint32_t ecb_bitrev32 (uint32_t x) 1046ecb_function_ ecb_const uint32_t ecb_bitrev32 (uint32_t x)
961{ 1047{
962 x = ((x >> 1) & 0x55555555) | ((x & 0x55555555) << 1); 1048 x = ((x >> 1) & 0x55555555) | ((x & 0x55555555) << 1);
963 x = ((x >> 2) & 0x33333333) | ((x & 0x33333333) << 2); 1049 x = ((x >> 2) & 0x33333333) | ((x & 0x33333333) << 2);
964 x = ((x >> 4) & 0x0f0f0f0f) | ((x & 0x0f0f0f0f) << 4); 1050 x = ((x >> 4) & 0x0f0f0f0f) | ((x & 0x0f0f0f0f) << 4);
965 x = ((x >> 8) & 0x00ff00ff) | ((x & 0x00ff00ff) << 8); 1051 x = ((x >> 8) & 0x00ff00ff) | ((x & 0x00ff00ff) << 8);
… …
968 return x; 1054 return x;
969} 1055}
970 1056
971/* popcount64 is only available on 64 bit cpus as gcc builtin */ 1057/* popcount64 is only available on 64 bit cpus as gcc builtin */
972/* so for this version we are lazy */ 1058/* so for this version we are lazy */
973ecb_function_ int ecb_popcount64 (uint64_t x) ecb_const; 1059ecb_function_ ecb_const int ecb_popcount64 (uint64_t x);
974ecb_function_ int 1060ecb_function_ ecb_const int
975ecb_popcount64 (uint64_t x) 1061ecb_popcount64 (uint64_t x)
976{ 1062{
977 return ecb_popcount32 (x) + ecb_popcount32 (x >> 32); 1063 return ecb_popcount32 (x) + ecb_popcount32 (x >> 32);
978} 1064}
979 1065
980ecb_inline uint8_t ecb_rotl8 (uint8_t x, unsigned int count) ecb_const; 1066ecb_inline ecb_const uint8_t ecb_rotl8 (uint8_t x, unsigned int count);
981ecb_inline uint8_t ecb_rotr8 (uint8_t x, unsigned int count) ecb_const; 1067ecb_inline ecb_const uint8_t ecb_rotr8 (uint8_t x, unsigned int count);
982ecb_inline uint16_t ecb_rotl16 (uint16_t x, unsigned int count) ecb_const; 1068ecb_inline ecb_const uint16_t ecb_rotl16 (uint16_t x, unsigned int count);
983ecb_inline uint16_t ecb_rotr16 (uint16_t x, unsigned int count) ecb_const; 1069ecb_inline ecb_const uint16_t ecb_rotr16 (uint16_t x, unsigned int count);
984ecb_inline uint32_t ecb_rotl32 (uint32_t x, unsigned int count) ecb_const; 1070ecb_inline ecb_const uint32_t ecb_rotl32 (uint32_t x, unsigned int count);
985ecb_inline uint32_t ecb_rotr32 (uint32_t x, unsigned int count) ecb_const; 1071ecb_inline ecb_const uint32_t ecb_rotr32 (uint32_t x, unsigned int count);
986ecb_inline uint64_t ecb_rotl64 (uint64_t x, unsigned int count) ecb_const; 1072ecb_inline ecb_const uint64_t ecb_rotl64 (uint64_t x, unsigned int count);
987ecb_inline uint64_t ecb_rotr64 (uint64_t x, unsigned int count) ecb_const; 1073ecb_inline ecb_const uint64_t ecb_rotr64 (uint64_t x, unsigned int count);
988 1074
989ecb_inline uint8_t ecb_rotl8 (uint8_t x, unsigned int count) { return (x >> ( 8 - count)) | (x << count); } 1075ecb_inline ecb_const uint8_t ecb_rotl8 (uint8_t x, unsigned int count) { return (x >> ( 8 - count)) | (x << count); }
990ecb_inline uint8_t ecb_rotr8 (uint8_t x, unsigned int count) { return (x << ( 8 - count)) | (x >> count); } 1076ecb_inline ecb_const uint8_t ecb_rotr8 (uint8_t x, unsigned int count) { return (x << ( 8 - count)) | (x >> count); }
991ecb_inline uint16_t ecb_rotl16 (uint16_t x, unsigned int count) { return (x >> (16 - count)) | (x << count); } 1077ecb_inline ecb_const uint16_t ecb_rotl16 (uint16_t x, unsigned int count) { return (x >> (16 - count)) | (x << count); }
992ecb_inline uint16_t ecb_rotr16 (uint16_t x, unsigned int count) { return (x << (16 - count)) | (x >> count); } 1078ecb_inline ecb_const uint16_t ecb_rotr16 (uint16_t x, unsigned int count) { return (x << (16 - count)) | (x >> count); }
993ecb_inline uint32_t ecb_rotl32 (uint32_t x, unsigned int count) { return (x >> (32 - count)) | (x << count); } 1079ecb_inline ecb_const uint32_t ecb_rotl32 (uint32_t x, unsigned int count) { return (x >> (32 - count)) | (x << count); }
994ecb_inline uint32_t ecb_rotr32 (uint32_t x, unsigned int count) { return (x << (32 - count)) | (x >> count); } 1080ecb_inline ecb_const uint32_t ecb_rotr32 (uint32_t x, unsigned int count) { return (x << (32 - count)) | (x >> count); }
995ecb_inline uint64_t ecb_rotl64 (uint64_t x, unsigned int count) { return (x >> (64 - count)) | (x << count); } 1081ecb_inline ecb_const uint64_t ecb_rotl64 (uint64_t x, unsigned int count) { return (x >> (64 - count)) | (x << count); }
996ecb_inline uint64_t ecb_rotr64 (uint64_t x, unsigned int count) { return (x << (64 - count)) | (x >> count); } 1082ecb_inline ecb_const uint64_t ecb_rotr64 (uint64_t x, unsigned int count) { return (x << (64 - count)) | (x >> count); }
997 1083
998#if ECB_GCC_VERSION(4,3) 1084#if ECB_GCC_VERSION(4,3) || (ECB_CLANG_BUILTIN(__builtin_bswap32) && ECB_CLANG_BUILTIN(__builtin_bswap64))
1085 #if ECB_GCC_VERSION(4,8) || ECB_CLANG_BUILTIN(__builtin_bswap16)
1086 #define ecb_bswap16(x) __builtin_bswap16 (x)
1087 #else
999 #define ecb_bswap16(x) (__builtin_bswap32 (x) >> 16) 1088 #define ecb_bswap16(x) (__builtin_bswap32 (x) >> 16)
1089 #endif
1000 #define ecb_bswap32(x) __builtin_bswap32 (x) 1090 #define ecb_bswap32(x) __builtin_bswap32 (x)
1001 #define ecb_bswap64(x) __builtin_bswap64 (x) 1091 #define ecb_bswap64(x) __builtin_bswap64 (x)
1092#elif _MSC_VER
1093 #include <stdlib.h>
1094 #define ecb_bswap16(x) ((uint16_t)_byteswap_ushort ((uint16_t)(x)))
1095 #define ecb_bswap32(x) ((uint32_t)_byteswap_ulong ((uint32_t)(x)))
1096 #define ecb_bswap64(x) ((uint64_t)_byteswap_uint64 ((uint64_t)(x)))
1002#else 1097#else
1003 ecb_function_ uint16_t ecb_bswap16 (uint16_t x) ecb_const; 1098 ecb_function_ ecb_const uint16_t ecb_bswap16 (uint16_t x);
1004 ecb_function_ uint16_t 1099 ecb_function_ ecb_const uint16_t
1005 ecb_bswap16 (uint16_t x) 1100 ecb_bswap16 (uint16_t x)
1006 { 1101 {
1007 return ecb_rotl16 (x, 8); 1102 return ecb_rotl16 (x, 8);
1008 } 1103 }
1009 1104
1010 ecb_function_ uint32_t ecb_bswap32 (uint32_t x) ecb_const; 1105 ecb_function_ ecb_const uint32_t ecb_bswap32 (uint32_t x);
1011 ecb_function_ uint32_t 1106 ecb_function_ ecb_const uint32_t
1012 ecb_bswap32 (uint32_t x) 1107 ecb_bswap32 (uint32_t x)
1013 { 1108 {
1014 return (((uint32_t)ecb_bswap16 (x)) << 16) | ecb_bswap16 (x >> 16); 1109 return (((uint32_t)ecb_bswap16 (x)) << 16) | ecb_bswap16 (x >> 16);
1015 } 1110 }
1016 1111
1017 ecb_function_ uint64_t ecb_bswap64 (uint64_t x) ecb_const; 1112 ecb_function_ ecb_const uint64_t ecb_bswap64 (uint64_t x);
1018 ecb_function_ uint64_t 1113 ecb_function_ ecb_const uint64_t
1019 ecb_bswap64 (uint64_t x) 1114 ecb_bswap64 (uint64_t x)
1020 { 1115 {
1021 return (((uint64_t)ecb_bswap32 (x)) << 32) | ecb_bswap32 (x >> 32); 1116 return (((uint64_t)ecb_bswap32 (x)) << 32) | ecb_bswap32 (x >> 32);
1022 } 1117 }
1023#endif 1118#endif
1024 1119
1025#if ECB_GCC_VERSION(4,5) 1120#if ECB_GCC_VERSION(4,5) || ECB_CLANG_BUILTIN(__builtin_unreachable)
1026 #define ecb_unreachable() __builtin_unreachable () 1121 #define ecb_unreachable() __builtin_unreachable ()
1027#else 1122#else
1028 /* this seems to work fine, but gcc always emits a warning for it :/ */ 1123 /* this seems to work fine, but gcc always emits a warning for it :/ */
1029 ecb_inline void ecb_unreachable (void) ecb_noreturn; 1124 ecb_inline ecb_noreturn void ecb_unreachable (void);
1030 ecb_inline void ecb_unreachable (void) { } 1125 ecb_inline ecb_noreturn void ecb_unreachable (void) { }
1031#endif 1126#endif
1032 1127
1033/* try to tell the compiler that some condition is definitely true */ 1128/* try to tell the compiler that some condition is definitely true */
1034#define ecb_assume(cond) if (!(cond)) ecb_unreachable (); else 0 1129#define ecb_assume(cond) if (!(cond)) ecb_unreachable (); else 0
1035 1130
1036ecb_inline unsigned char ecb_byteorder_helper (void) ecb_const; 1131ecb_inline ecb_const uint32_t ecb_byteorder_helper (void);
1037ecb_inline unsigned char 1132ecb_inline ecb_const uint32_t
1038ecb_byteorder_helper (void) 1133ecb_byteorder_helper (void)
1039{ 1134{
1040 /* the union code still generates code under pressure in gcc, */ 1135 /* the union code still generates code under pressure in gcc, */
1041 /* but less than using pointers, and always seems to */ 1136 /* but less than using pointers, and always seems to */
1042 /* successfully return a constant. */ 1137 /* successfully return a constant. */
1043 /* the reason why we have this horrible preprocessor mess */ 1138 /* the reason why we have this horrible preprocessor mess */
1044 /* is to avoid it in all cases, at least on common architectures */ 1139 /* is to avoid it in all cases, at least on common architectures */
1045 /* or when using a recent enough gcc version (>= 4.6) */ 1140 /* or when using a recent enough gcc version (>= 4.6) */
1046#if __i386 || __i386__ || _M_X86 || __amd64 || __amd64__ || _M_X64
1047 return 0x44;
1048#elif __BYTE_ORDER__ && __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__ 1141#if (defined __BYTE_ORDER__ && __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__) \
1142 || ((__i386 || __i386__ || _M_IX86 || ECB_GCC_AMD64 || ECB_MSVC_AMD64) && !__VOS__)
1143 #define ECB_LITTLE_ENDIAN 1
1049 return 0x44; 1144 return 0x44332211;
1050#elif __BYTE_ORDER__ && __BYTE_ORDER__ == __ORDER_BIG_ENDIAN__ 1145#elif (defined __BYTE_ORDER__ && __BYTE_ORDER__ == __ORDER_BIG_ENDIAN__) \
1146 || ((__AARCH64EB__ || __MIPSEB__ || __ARMEB__) && !__VOS__)
1147 #define ECB_BIG_ENDIAN 1
1051 return 0x11; 1148 return 0x11223344;
1052#else 1149#else
1053 union 1150 union
1054 { 1151 {
1152 uint8_t c[4];
1055 uint32_t i; 1153 uint32_t u;
1056 uint8_t c;
1057 } u = { 0x11223344 }; 1154 } u = { 0x11, 0x22, 0x33, 0x44 };
1058 return u.c; 1155 return u.u;
1059#endif 1156#endif
1060} 1157}
1061 1158
1062ecb_inline ecb_bool ecb_big_endian (void) ecb_const; 1159ecb_inline ecb_const ecb_bool ecb_big_endian (void);
1063ecb_inline ecb_bool ecb_big_endian (void) { return ecb_byteorder_helper () == 0x11; } 1160ecb_inline ecb_const ecb_bool ecb_big_endian (void) { return ecb_byteorder_helper () == 0x11223344; }
1064ecb_inline ecb_bool ecb_little_endian (void) ecb_const; 1161ecb_inline ecb_const ecb_bool ecb_little_endian (void);
1065ecb_inline ecb_bool ecb_little_endian (void) { return ecb_byteorder_helper () == 0x44; } 1162ecb_inline ecb_const ecb_bool ecb_little_endian (void) { return ecb_byteorder_helper () == 0x44332211; }
1066 1163
1067#if ECB_GCC_VERSION(3,0) || ECB_C99 1164#if ECB_GCC_VERSION(3,0) || ECB_C99
1068 #define ecb_mod(m,n) ((m) % (n) + ((m) % (n) < 0 ? (n) : 0)) 1165 #define ecb_mod(m,n) ((m) % (n) + ((m) % (n) < 0 ? (n) : 0))
1069#else 1166#else
1070 #define ecb_mod(m,n) ((m) < 0 ? ((n) - 1 - ((-1 - (m)) % (n))) : ((m) % (n))) 1167 #define ecb_mod(m,n) ((m) < 0 ? ((n) - 1 - ((-1 - (m)) % (n))) : ((m) % (n)))
1071#endif 1168#endif
1072 1169
1073#if __cplusplus 1170#if ECB_CPP
1074 template<typename T> 1171 template<typename T>
1075 static inline T ecb_div_rd (T val, T div) 1172 static inline T ecb_div_rd (T val, T div)
1076 { 1173 {
1077 return val < 0 ? - ((-val + div - 1) / div) : (val ) / div; 1174 return val < 0 ? - ((-val + div - 1) / div) : (val ) / div;
1078 } 1175 }
… …
1095 } 1192 }
1096#else 1193#else
1097 #define ecb_array_length(name) (sizeof (name) / sizeof (name [0])) 1194 #define ecb_array_length(name) (sizeof (name) / sizeof (name [0]))
1098#endif 1195#endif
1099 1196
1197ecb_function_ ecb_const uint32_t ecb_binary16_to_binary32 (uint32_t x);
1198ecb_function_ ecb_const uint32_t
1199ecb_binary16_to_binary32 (uint32_t x)
1200{
1201 unsigned int s = (x & 0x8000) << (31 - 15);
1202 int e = (x >> 10) & 0x001f;
1203 unsigned int m = x & 0x03ff;
1204
1205 if (ecb_expect_false (e == 31))
1206 /* infinity or NaN */
1207 e = 255 - (127 - 15);
1208 else if (ecb_expect_false (!e))
1209 {
1210 if (ecb_expect_true (!m))
1211 /* zero, handled by code below by forcing e to 0 */
1212 e = 0 - (127 - 15);
1213 else
1214 {
1215 /* subnormal, renormalise */
1216 unsigned int s = 10 - ecb_ld32 (m);
1217
1218 m = (m << s) & 0x3ff; /* mask implicit bit */
1219 e -= s - 1;
1220 }
1221 }
1222
1223 /* e and m now are normalised, or zero, (or inf or nan) */
1224 e += 127 - 15;
1225
1226 return s | (e << 23) | (m << (23 - 10));
1227}
1228
1229ecb_function_ ecb_const uint16_t ecb_binary32_to_binary16 (uint32_t x);
1230ecb_function_ ecb_const uint16_t
1231ecb_binary32_to_binary16 (uint32_t x)
1232{
1233 unsigned int s = (x >> 16) & 0x00008000; /* sign bit, the easy part */
1234 unsigned int e = ((x >> 23) & 0x000000ff) - (127 - 15); /* the desired exponent */
1235 unsigned int m = x & 0x007fffff;
1236
1237 x &= 0x7fffffff;
1238
1239 /* if it's within range of binary16 normals, use fast path */
1240 if (ecb_expect_true (0x38800000 <= x && x <= 0x477fefff))
1241 {
1242 /* mantissa round-to-even */
1243 m += 0x00000fff + ((m >> (23 - 10)) & 1);
1244
1245 /* handle overflow */
1246 if (ecb_expect_false (m >= 0x00800000))
1247 {
1248 m >>= 1;
1249 e += 1;
1250 }
1251
1252 return s | (e << 10) | (m >> (23 - 10));
1253 }
1254
1255 /* handle large numbers and infinity */
1256 if (ecb_expect_true (0x477fefff < x && x <= 0x7f800000))
1257 return s | 0x7c00;
1258
1259 /* handle zero, subnormals and small numbers */
1260 if (ecb_expect_true (x < 0x38800000))
1261 {
1262 /* zero */
1263 if (ecb_expect_true (!x))
1264 return s;
1265
1266 /* handle subnormals */
1267
1268 /* too small, will be zero */
1269 if (e < (14 - 24)) /* might not be sharp, but is good enough */
1270 return s;
1271
1272 m |= 0x00800000; /* make implicit bit explicit */
1273
1274 /* very tricky - we need to round to the nearest e (+10) bit value */
1275 {
1276 unsigned int bits = 14 - e;
1277 unsigned int half = (1 << (bits - 1)) - 1;
1278 unsigned int even = (m >> bits) & 1;
1279
1280 /* if this overflows, we will end up with a normalised number */
1281 m = (m + half + even) >> bits;
1282 }
1283
1284 return s | m;
1285 }
1286
1287 /* handle NaNs, preserve leftmost nan bits, but make sure we don't turn them into infinities */
1288 m >>= 13;
1289
1290 return s | 0x7c00 | m | !m;
1291}
1292
1100/*******************************************************************************/ 1293/*******************************************************************************/
1101/* floating point stuff, can be disabled by defining ECB_NO_LIBM */ 1294/* floating point stuff, can be disabled by defining ECB_NO_LIBM */
1102 1295
1103/* basically, everything uses "ieee pure-endian" floating point numbers */ 1296/* basically, everything uses "ieee pure-endian" floating point numbers */
1104/* the only noteworthy exception is ancient armle, which uses order 43218765 */ 1297/* the only noteworthy exception is ancient armle, which uses order 43218765 */
1105#if 0 \ 1298#if 0 \
1106 || __i386 || __i386__ \ 1299 || __i386 || __i386__ \
1107 || __amd64 || __amd64__ || __x86_64 || __x86_64__ \ 1300 || ECB_GCC_AMD64 \
1108 || __powerpc__ || __ppc__ || __powerpc64__ || __ppc64__ \ 1301 || __powerpc__ || __ppc__ || __powerpc64__ || __ppc64__ \
1109 || defined __s390__ || defined __s390x__ \ 1302 || defined __s390__ || defined __s390x__ \
1110 || defined __mips__ \ 1303 || defined __mips__ \
1111 || defined __alpha__ \ 1304 || defined __alpha__ \
1112 || defined __hppa__ \ 1305 || defined __hppa__ \
1113 || defined __ia64__ \ 1306 || defined __ia64__ \
1114 || defined __m68k__ \ 1307 || defined __m68k__ \
1115 || defined __m88k__ \ 1308 || defined __m88k__ \
1116 || defined __sh__ \ 1309 || defined __sh__ \
1117 || defined _M_IX86 || defined _M_AMD64 || defined _M_IA64 \ 1310 || defined _M_IX86 || defined ECB_MSVC_AMD64 || defined _M_IA64 \
1118 || (defined __arm__ && (defined __ARM_EABI__ || defined __EABI__ || defined __VFP_FP__ || defined _WIN32_WCE || defined __ANDROID__)) \ 1311 || (defined __arm__ && (defined __ARM_EABI__ || defined __EABI__ || defined __VFP_FP__ || defined _WIN32_WCE || defined __ANDROID__)) \
1119 || defined __aarch64__ 1312 || defined __aarch64__
1120 #define ECB_STDFP 1 1313 #define ECB_STDFP 1
1121 #include <string.h> /* for memcpy */ 1314 #include <string.h> /* for memcpy */
1122#else 1315#else
… …
1138 #define ECB_NAN NAN 1331 #define ECB_NAN NAN
1139 #else 1332 #else
1140 #define ECB_NAN ECB_INFINITY 1333 #define ECB_NAN ECB_INFINITY
1141 #endif 1334 #endif
1142 1335
1143 /* converts an ieee half/binary16 to a float */ 1336 #if ECB_C99 || _XOPEN_VERSION >= 600 || _POSIX_VERSION >= 200112L
1144 ecb_function_ float ecb_binary16_to_float (uint16_t x) ecb_const; 1337 #define ecb_ldexpf(x,e) ldexpf ((x), (e))
1145 ecb_function_ float 1338 #define ecb_frexpf(x,e) frexpf ((x), (e))
1146 ecb_binary16_to_float (uint16_t x) 1339 #else
1147 { 1340 #define ecb_ldexpf(x,e) (float) ldexp ((double) (x), (e))
1148 int e = (x >> 10) & 0x1f; 1341 #define ecb_frexpf(x,e) (float) frexp ((double) (x), (e))
1149 int m = x & 0x3ff; 1342 #endif
1150 float r;
1151
1152 if (!e ) r = ldexpf (m , -24);
1153 else if (e != 31) r = ldexpf (m + 0x400, e - 25);
1154 else if (m ) r = ECB_NAN;
1155 else r = ECB_INFINITY;
1156
1157 return x & 0x8000 ? -r : r;
1158 }
1159 1343
1160 /* convert a float to ieee single/binary32 */ 1344 /* convert a float to ieee single/binary32 */
1161 ecb_function_ uint32_t ecb_float_to_binary32 (float x) ecb_const; 1345 ecb_function_ ecb_const uint32_t ecb_float_to_binary32 (float x);
1162 ecb_function_ uint32_t 1346 ecb_function_ ecb_const uint32_t
1163 ecb_float_to_binary32 (float x) 1347 ecb_float_to_binary32 (float x)
1164 { 1348 {
1165 uint32_t r; 1349 uint32_t r;
1166 1350
1167 #if ECB_STDFP 1351 #if ECB_STDFP
… …
1174 if (x == 0e0f ) return 0x00000000U; 1358 if (x == 0e0f ) return 0x00000000U;
1175 if (x > +3.40282346638528860e+38f) return 0x7f800000U; 1359 if (x > +3.40282346638528860e+38f) return 0x7f800000U;
1176 if (x < -3.40282346638528860e+38f) return 0xff800000U; 1360 if (x < -3.40282346638528860e+38f) return 0xff800000U;
1177 if (x != x ) return 0x7fbfffffU; 1361 if (x != x ) return 0x7fbfffffU;
1178 1362
1179 m = frexpf (x, &e) * 0x1000000U; 1363 m = ecb_frexpf (x, &e) * 0x1000000U;
1180 1364
1181 r = m & 0x80000000U; 1365 r = m & 0x80000000U;
1182 1366
1183 if (r) 1367 if (r)
1184 m = -m; 1368 m = -m;
… …
1196 1380
1197 return r; 1381 return r;
1198 } 1382 }
1199 1383
1200 /* converts an ieee single/binary32 to a float */ 1384 /* converts an ieee single/binary32 to a float */
1201 ecb_function_ float ecb_binary32_to_float (uint32_t x) ecb_const; 1385 ecb_function_ ecb_const float ecb_binary32_to_float (uint32_t x);
1202 ecb_function_ float 1386 ecb_function_ ecb_const float
1203 ecb_binary32_to_float (uint32_t x) 1387 ecb_binary32_to_float (uint32_t x)
1204 { 1388 {
1205 float r; 1389 float r;
1206 1390
1207 #if ECB_STDFP 1391 #if ECB_STDFP
… …
1217 x |= 0x800000U; 1401 x |= 0x800000U;
1218 else 1402 else
1219 e = 1; 1403 e = 1;
1220 1404
1221 /* we distrust ldexpf a bit and do the 2**-24 scaling by an extra multiply */ 1405 /* we distrust ldexpf a bit and do the 2**-24 scaling by an extra multiply */
1222 r = ldexpf (x * (0.5f / 0x800000U), e - 126); 1406 r = ecb_ldexpf (x * (0.5f / 0x800000U), e - 126);
1223 1407
1224 r = neg ? -r : r; 1408 r = neg ? -r : r;
1225 #endif 1409 #endif
1226 1410
1227 return r; 1411 return r;
1228 } 1412 }
1229 1413
1230 /* convert a double to ieee double/binary64 */ 1414 /* convert a double to ieee double/binary64 */
1231 ecb_function_ uint64_t ecb_double_to_binary64 (double x) ecb_const; 1415 ecb_function_ ecb_const uint64_t ecb_double_to_binary64 (double x);
1232 ecb_function_ uint64_t 1416 ecb_function_ ecb_const uint64_t
1233 ecb_double_to_binary64 (double x) 1417 ecb_double_to_binary64 (double x)
1234 { 1418 {
1235 uint64_t r; 1419 uint64_t r;
1236 1420
1237 #if ECB_STDFP 1421 #if ECB_STDFP
… …
1266 1450
1267 return r; 1451 return r;
1268 } 1452 }
1269 1453
1270 /* converts an ieee double/binary64 to a double */ 1454 /* converts an ieee double/binary64 to a double */
1271 ecb_function_ double ecb_binary64_to_double (uint64_t x) ecb_const; 1455 ecb_function_ ecb_const double ecb_binary64_to_double (uint64_t x);
1272 ecb_function_ double 1456 ecb_function_ ecb_const double
1273 ecb_binary64_to_double (uint64_t x) 1457 ecb_binary64_to_double (uint64_t x)
1274 { 1458 {
1275 double r; 1459 double r;
1276 1460
1277 #if ECB_STDFP 1461 #if ECB_STDFP
… …
1293 1477
1294 r = neg ? -r : r; 1478 r = neg ? -r : r;
1295 #endif 1479 #endif
1296 1480
1297 return r; 1481 return r;
1482 }
1483
1484 /* convert a float to ieee half/binary16 */
1485 ecb_function_ ecb_const uint16_t ecb_float_to_binary16 (float x);
1486 ecb_function_ ecb_const uint16_t
1487 ecb_float_to_binary16 (float x)
1488 {
1489 return ecb_binary32_to_binary16 (ecb_float_to_binary32 (x));
1490 }
1491
1492 /* convert an ieee half/binary16 to float */
1493 ecb_function_ ecb_const float ecb_binary16_to_float (uint16_t x);
1494 ecb_function_ ecb_const float
1495 ecb_binary16_to_float (uint16_t x)
1496 {
1497 return ecb_binary32_to_float (ecb_binary16_to_binary32 (x));
1298 } 1498 }
1299 1499
1300#endif 1500#endif
1301 1501
1302#endif 1502#endif
… …
2613{ 2813{
2614 invoke_cb = invoke_pending_cb; 2814 invoke_cb = invoke_pending_cb;
2615} 2815}
2616 2816
2617void 2817void
2618ev_set_loop_release_cb (EV_P_ ev_loop_callback EV_THROW release, ev_loop_callback EV_THROW acquire) EV_THROW 2818ev_set_loop_release_cb (EV_P_ void (*release)(EV_P) EV_THROW, void (*acquire)(EV_P) EV_THROW) EV_THROW
2619{ 2819{
2620 release_cb = release; 2820 release_cb = release;
2621 acquire_cb = acquire; 2821 acquire_cb = acquire;
2622} 2822}
2623#endif 2823#endif
… …
2849#if EV_USE_INOTIFY 3049#if EV_USE_INOTIFY
2850 infy_fork (EV_A); 3050 infy_fork (EV_A);
2851#endif 3051#endif
2852 3052
2853#if EV_SIGNAL_ENABLE || EV_ASYNC_ENABLE 3053#if EV_SIGNAL_ENABLE || EV_ASYNC_ENABLE
2854 if (ev_is_active (&pipe_w)) 3054 if (ev_is_active (&pipe_w) && postfork != 2)
2855 { 3055 {
2856 /* pipe_write_wanted must be false now, so modifying fd vars should be safe */ 3056 /* pipe_write_wanted must be false now, so modifying fd vars should be safe */
2857 3057
2858 ev_ref (EV_A); 3058 ev_ref (EV_A);
2859 ev_io_stop (EV_A_ &pipe_w); 3059 ev_io_stop (EV_A_ &pipe_w);

Diff Legend

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