--- libecb/ecb.h 2012/06/07 21:28:36 1.99 +++ libecb/ecb.h 2014/01/05 15:54:09 1.127 @@ -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; @@ -56,7 +56,6 @@ typedef uint32_t uintptr_t; typedef int32_t intptr_t; #endif - typedef intptr_t ptrdiff_t; #else #include #if UINTMAX_MAX > 0xffffffffU @@ -66,6 +65,15 @@ #endif #endif +/* work around x32 idiocy by defining proper macros */ +#if __amd64 || __x86_64 || _M_AMD64 || _M_X64 + #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. @@ -81,12 +89,30 @@ #endif #endif -#define ECB_C (__STDC__+0) /* this assumes that __STDC__ is either empty or a number */ -#define ECB_C99 (__STDC_VERSION__ >= 199901L) -#define ECB_C11 (__STDC_VERSION__ >= 201112L) #define ECB_CPP (__cplusplus+0) #define ECB_CPP11 (__cplusplus >= 201103L) +#if ECB_CPP + #define ECB_C 0 + #define ECB_STDC_VERSION 0 +#else + #define ECB_C 1 + #define ECB_STDC_VERSION __STDC_VERSION__ +#endif + +#define ECB_C99 (ECB_STDC_VERSION >= 199901L) +#define ECB_C11 (ECB_STDC_VERSION >= 201112L) + +#if ECB_CPP + #define ECB_EXTERN_C extern "C" + #define ECB_EXTERN_C_BEG ECB_EXTERN_C { + #define ECB_EXTERN_C_END } +#else + #define ECB_EXTERN_C extern + #define ECB_EXTERN_C_BEG + #define ECB_EXTERN_C_END +#endif + /*****************************************************************************/ /* ECB_NO_THREADS - ecb is not used by multiple threads, ever */ @@ -118,14 +144,16 @@ #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__ @@ -133,6 +161,12 @@ #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 @@ -141,11 +175,24 @@ #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 >= 1500 /* VC++ 2008 */ + /* apparently, microsoft broke all the memory barrier stuff in Visual Studio 2008... */ + #pragma intrinsic(_ReadBarrier,_WriteBarrier,_ReadWriteBarrier) + #define ECB_MEMORY_FENCE _ReadWriteBarrier (); MemoryBarrier() + #define ECB_MEMORY_FENCE_ACQUIRE _ReadWriteBarrier (); MemoryBarrier() /* according to msdn, _ReadBarrier is not a load fence */ + #define ECB_MEMORY_FENCE_RELEASE _WriteBarrier (); MemoryBarrier() #elif _MSC_VER >= 1400 /* VC++ 2005 */ #pragma intrinsic(_ReadBarrier,_WriteBarrier,_ReadWriteBarrier) #define ECB_MEMORY_FENCE _ReadWriteBarrier () @@ -241,6 +288,11 @@ #define ecb_prefetch(addr,rw,locality) __builtin_prefetch (addr, rw, locality) #else #define ecb_attribute(attrlist) + + /* possible C11 impl for integral types + typedef struct ecb_is_constant_struct ecb_is_constant_struct; + #define ecb_is_constant(expr) _Generic ((1 ? (struct ecb_is_constant_struct *)0 : (void *)((expr) - (expr)), ecb_is_constant_struct *: 0, default: 1)) */ + #define ecb_is_constant(expr) 0 #define ecb_expect(expr,value) (expr) #define ecb_prefetch(addr,rw,locality) @@ -461,24 +513,24 @@ #endif /* try to tell the compiler that some condition is definitely true */ -#define ecb_assume(cond) do { if (!(cond)) ecb_unreachable (); } while (0) +#define ecb_assume(cond) if (!(cond)) ecb_unreachable (); else 0 ecb_inline unsigned char ecb_byteorder_helper (void) ecb_const; ecb_inline unsigned char 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 { @@ -527,5 +579,206 @@ #define ecb_array_length(name) (sizeof (name) / sizeof (name [0])) #endif +/*******************************************************************************/ +/* floating point stuff, can be disabled by defining ECB_NO_LIBM */ + +/* basically, everything uses "ieee pure-endian" floating point numbers */ +/* the only noteworthy exception is ancient armle, which uses order 43218765 */ +#if 0 \ + || __i386 || __i386__ \ + || __amd64 || __amd64__ || __x86_64 || __x86_64__ \ + || __powerpc__ || __ppc__ || __powerpc64__ || __ppc64__ \ + || defined __arm__ && defined __ARM_EABI__ \ + || defined __s390__ || defined __s390x__ \ + || defined __mips__ \ + || 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 */ +#else + #define ECB_STDFP 0 +#endif + +#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 + ecb_float_to_binary32 (float x) + { + uint32_t r; + + #if ECB_STDFP + memcpy (&r, &x, 4); + #else + /* slow emulation, works for anything but -0 */ + uint32_t m; + int e; + + if (x == 0e0f ) return 0x00000000U; + if (x > +3.40282346638528860e+38f) return 0x7f800000U; + if (x < -3.40282346638528860e+38f) return 0xff800000U; + if (x != x ) return 0x7fbfffffU; + + m = frexpf (x, &e) * 0x1000000U; + + r = m & 0x80000000U; + + if (r) + m = -m; + + if (e <= -126) + { + m &= 0xffffffU; + m >>= (-125 - e); + e = -126; + } + + r |= (e + 126) << 23; + r |= m & 0x7fffffU; + #endif + + return r; + } + + /* converts an ieee single/binary32 to a float */ + ecb_function_ float ecb_binary32_to_float (uint32_t x) ecb_const; + ecb_function_ float + ecb_binary32_to_float (uint32_t x) + { + float r; + + #if ECB_STDFP + memcpy (&r, &x, 4); + #else + /* emulation, only works for normals and subnormals and +0 */ + int neg = x >> 31; + int e = (x >> 23) & 0xffU; + + x &= 0x7fffffU; + + if (e) + x |= 0x800000U; + else + e = 1; + + /* we distrust ldexpf a bit and do the 2**-24 scaling by an extra multiply */ + r = ldexpf (x * (0.5f / 0x800000U), e - 126); + + r = neg ? -r : r; + #endif + + return r; + } + + /* convert a double to ieee double/binary64 */ + ecb_function_ uint64_t ecb_double_to_binary64 (double x) ecb_const; + ecb_function_ uint64_t + ecb_double_to_binary64 (double x) + { + uint64_t r; + + #if ECB_STDFP + memcpy (&r, &x, 8); + #else + /* slow emulation, works for anything but -0 */ + uint64_t m; + int e; + + if (x == 0e0 ) return 0x0000000000000000U; + if (x > +1.79769313486231470e+308) return 0x7ff0000000000000U; + if (x < -1.79769313486231470e+308) return 0xfff0000000000000U; + if (x != x ) return 0X7ff7ffffffffffffU; + + m = frexp (x, &e) * 0x20000000000000U; + + r = m & 0x8000000000000000;; + + if (r) + m = -m; + + if (e <= -1022) + { + m &= 0x1fffffffffffffU; + m >>= (-1021 - e); + e = -1022; + } + + r |= ((uint64_t)(e + 1022)) << 52; + r |= m & 0xfffffffffffffU; + #endif + + return r; + } + + /* converts an ieee double/binary64 to a double */ + ecb_function_ double ecb_binary64_to_double (uint64_t x) ecb_const; + ecb_function_ double + ecb_binary64_to_double (uint64_t x) + { + double r; + + #if ECB_STDFP + memcpy (&r, &x, 8); + #else + /* emulation, only works for normals and subnormals and +0 */ + int neg = x >> 63; + int e = (x >> 52) & 0x7ffU; + + x &= 0xfffffffffffffU; + + if (e) + x |= 0x10000000000000U; + else + e = 1; + + /* we distrust ldexp a bit and do the 2**-53 scaling by an extra multiply */ + r = ldexp (x * (0.5 / 0x10000000000000U), e - 1022); + + r = neg ? -r : r; + #endif + + return r; + } + +#endif + #endif