--- libecb/ecb.h 2012/06/29 00:56:30 1.107 +++ libecb/ecb.h 2013/12/03 15:35:14 1.124 @@ -1,7 +1,7 @@ /* * libecb - http://software.schmorp.de/pkg/libecb * - * Copyright (©) 2009-2012 Marc Alexander Lehmann + * Copyright (©) 2009-2013 Marc Alexander Lehmann * Copyright (©) 2011 Emanuele Giaquinta * All rights reserved. * @@ -31,7 +31,7 @@ #define ECB_H /* 16 bits major, 16 bits minor */ -#define ECB_VERSION 0x00010001 +#define ECB_VERSION 0x00010003 #ifdef _WIN32 typedef signed char int8_t; @@ -65,6 +65,15 @@ #endif #endif +/* work around x32 idiocy by defining proper macros */ +#if __x86_64 || _M_AMD64 + #if _ILP32 + #define ECB_AMD64_X32 1 + #else + #define ECB_AMD64 1 + #endif +#endif + /* many compilers define _GNUC_ to some versions but then only implement * what their idiot authors think are the "more important" extensions, * causing enormous grief in return for some better fake benchmark numbers. @@ -127,14 +136,16 @@ #elif defined __ARM_ARCH_7__ || defined __ARM_ARCH_7A__ \ || defined __ARM_ARCH_7M__ || defined __ARM_ARCH_7R__ #define ECB_MEMORY_FENCE __asm__ __volatile__ ("dmb" : : : "memory") - #elif __sparc || __sparc__ + #elif (__sparc || __sparc__) && !__sparcv8 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("membar #LoadStore | #LoadLoad | #StoreStore | #StoreLoad" : : : "memory") #define ECB_MEMORY_FENCE_ACQUIRE __asm__ __volatile__ ("membar #LoadStore | #LoadLoad" : : : "memory") #define ECB_MEMORY_FENCE_RELEASE __asm__ __volatile__ ("membar #LoadStore | #StoreStore") #elif defined __s390__ || defined __s390x__ #define ECB_MEMORY_FENCE __asm__ __volatile__ ("bcr 15,0" : : : "memory") #elif defined __mips__ - #define ECB_MEMORY_FENCE __asm__ __volatile__ ("sync" : : : "memory") + /* GNU/Linux emulates sync on mips1 architectures, so we force its use */ + /* anybody else who still uses mips1 is supposed to send in their version, with detection code. */ + #define ECB_MEMORY_FENCE __asm__ __volatile__ (".set mips2; sync; .set mips0" : : : "memory") #elif defined __alpha__ #define ECB_MEMORY_FENCE __asm__ __volatile__ ("mb" : : : "memory") #elif defined __hppa__ @@ -142,6 +153,12 @@ #define ECB_MEMORY_FENCE_RELEASE __asm__ __volatile__ ("") #elif defined __ia64__ #define ECB_MEMORY_FENCE __asm__ __volatile__ ("mf" : : : "memory") + #elif defined __m68k__ + #define ECB_MEMORY_FENCE __asm__ __volatile__ ("" : : : "memory") + #elif defined __m88k__ + #define ECB_MEMORY_FENCE __asm__ __volatile__ ("tb1 0,%%r0,128" : : : "memory") + #elif defined __sh__ + #define ECB_MEMORY_FENCE __asm__ __volatile__ ("" : : : "memory") #endif #endif #endif @@ -150,9 +167,16 @@ #if ECB_GCC_VERSION(4,7) /* see comment below (stdatomic.h) about the C11 memory model. */ #define ECB_MEMORY_FENCE __atomic_thread_fence (__ATOMIC_SEQ_CST) - /*#elif defined __clang && __has_feature (cxx_atomic)*/ - /* see comment below (stdatomic.h) about the C11 memory model. */ - /*#define ECB_MEMORY_FENCE __c11_atomic_thread_fence (__ATOMIC_SEQ_CST)*/ + + /* The __has_feature syntax from clang is so misdesigned that we cannot use it + * without risking compile time errors with other compilers. We *could* + * define our own ecb_clang_has_feature, but I just can't be bothered to work + * around this shit time and again. + * #elif defined __clang && __has_feature (cxx_atomic) + * // see comment below (stdatomic.h) about the C11 memory model. + * #define ECB_MEMORY_FENCE __c11_atomic_thread_fence (__ATOMIC_SEQ_CST) + */ + #elif ECB_GCC_VERSION(4,4) || defined __INTEL_COMPILER || defined __clang__ #define ECB_MEMORY_FENCE __sync_synchronize () #elif _MSC_VER >= 1400 /* VC++ 2005 */ @@ -477,17 +501,17 @@ ecb_byteorder_helper (void) { /* the union code still generates code under pressure in gcc, */ - /* but less than using pointers, and always seem to */ + /* but less than using pointers, and always seems to */ /* successfully return a constant. */ /* the reason why we have this horrible preprocessor mess */ /* is to avoid it in all cases, at least on common architectures */ - /* and yes, gcc defines __BYTE_ORDER__, g++ does not */ + /* or when using a recent enough gcc version (>= 4.6) */ #if __i386 || __i386__ || _M_X86 || __amd64 || __amd64__ || _M_X64 return 0x44; #elif __BYTE_ORDER__ && __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__ return 0x44; #elif __BYTE_ORDER__ && __BYTE_ORDER__ == __ORDER_BIG_ENDIAN__ - retrurn 0x11; + return 0x11; #else union { @@ -551,6 +575,9 @@ || defined __alpha__ \ || defined __hppa__ \ || defined __ia64__ \ + || defined __m68k__ \ + || defined __m88k__ \ + || defined __sh__ \ || defined _M_IX86 || defined _M_AMD64 || defined _M_IA64 #define ECB_STDFP 1 #include /* for memcpy */ @@ -560,6 +587,38 @@ #ifndef ECB_NO_LIBM + #include /* for frexp*, ldexp*, INFINITY, NAN */ + + /* only the oldest of old doesn't have this one. solaris. */ + #ifdef INFINITY + #define ECB_INFINITY INFINITY + #else + #define ECB_INFINITY HUGE_VAL + #endif + + #ifdef NAN + #define ECB_NAN NAN + #else + #define ECB_NAN ECB_INFINITY + #endif + + /* converts an ieee half/binary16 to a float */ + ecb_function_ float ecb_binary16_to_float (uint16_t x) ecb_const; + ecb_function_ float + ecb_binary16_to_float (uint16_t x) + { + int e = (x >> 10) & 0x1f; + int m = x & 0x3ff; + float r; + + if (!e ) r = ldexpf (m , -24); + else if (e != 31) r = ldexpf (m + 0x400, e - 25); + else if (m ) r = ECB_NAN; + else r = ECB_INFINITY; + + return x & 0x8000 ? -r : r; + } + /* convert a float to ieee single/binary32 */ ecb_function_ uint32_t ecb_float_to_binary32 (float x) ecb_const; ecb_function_ uint32_t @@ -571,11 +630,10 @@ memcpy (&r, &x, 4); #else /* slow emulation, works for anything but -0 */ - ECB_EXTERN_C float frexpf (float v, int *e); uint32_t m; int e; - if (x == 0e0f ) return 0; + if (x == 0e0f ) return 0x00000000U; if (x > +3.40282346638528860e+38f) return 0x7f800000U; if (x < -3.40282346638528860e+38f) return 0xff800000U; if (x != x ) return 0x7fbfffffU; @@ -587,7 +645,7 @@ if (r) m = -m; - if (e < -125) + if (e <= -126) { m &= 0xffffffU; m >>= (-125 - e); @@ -612,8 +670,6 @@ memcpy (&r, &x, 4); #else /* emulation, only works for normals and subnormals and +0 */ - ECB_EXTERN_C float ldexpf (float x, int e); - int neg = x >> 31; int e = (x >> 23) & 0xffU; @@ -625,7 +681,7 @@ e = 1; /* we distrust ldexpf a bit and do the 2**-24 scaling by an extra multiply */ - r = ldexpf (x * (1.f / 0x1000000U), e - 126); + r = ldexpf (x * (0.5f / 0x800000U), e - 126); r = neg ? -r : r; #endif @@ -644,11 +700,10 @@ memcpy (&r, &x, 8); #else /* slow emulation, works for anything but -0 */ - ECB_EXTERN_C double frexp (double v, int *e); uint64_t m; int e; - if (x == 0e0 ) return 0; + if (x == 0e0 ) return 0x0000000000000000U; if (x > +1.79769313486231470e+308) return 0x7ff0000000000000U; if (x < -1.79769313486231470e+308) return 0xfff0000000000000U; if (x != x ) return 0X7ff7ffffffffffffU; @@ -660,7 +715,7 @@ if (r) m = -m; - if (e < -1021) + if (e <= -1022) { m &= 0x1fffffffffffffU; m >>= (-1021 - e); @@ -685,8 +740,6 @@ memcpy (&r, &x, 8); #else /* emulation, only works for normals and subnormals and +0 */ - ECB_EXTERN_C double ldexp (double x, int e); - int neg = x >> 63; int e = (x >> 52) & 0x7ffU; @@ -698,7 +751,7 @@ e = 1; /* we distrust ldexp a bit and do the 2**-53 scaling by an extra multiply */ - r = ldexp (x * (1. / 0x20000000000000U), e - 1022); + r = ldexp (x * (0.5 / 0x10000000000000U), e - 1022); r = neg ? -r : r; #endif