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 0x00010002 |
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; |
… | |
… | |
63 | #else |
63 | #else |
64 | #define ECB_PTRSIZE 4 |
64 | #define ECB_PTRSIZE 4 |
65 | #endif |
65 | #endif |
66 | #endif |
66 | #endif |
67 | |
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 |
|
|
74 | #endif |
|
|
75 | #endif |
|
|
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, |
70 | * causing enormous grief in return for some better fake benchmark numbers. |
79 | * causing enormous grief in return for some better fake benchmark numbers. |
71 | * or so. |
80 | * or so. |
72 | * we try to detect these and simply assume they are not gcc - if they have |
81 | * we try to detect these and simply assume they are not gcc - if they have |
… | |
… | |
78 | #else |
87 | #else |
79 | #define ECB_GCC_VERSION(major,minor) (__GNUC__ > (major) || (__GNUC__ == (major) && __GNUC_MINOR__ >= (minor))) |
88 | #define ECB_GCC_VERSION(major,minor) (__GNUC__ > (major) || (__GNUC__ == (major) && __GNUC_MINOR__ >= (minor))) |
80 | #endif |
89 | #endif |
81 | #endif |
90 | #endif |
82 | |
91 | |
83 | #define ECB_C (__STDC__+0) /* this assumes that __STDC__ is either empty or a number */ |
|
|
84 | #define ECB_C99 (__STDC_VERSION__ >= 199901L) |
|
|
85 | #define ECB_C11 (__STDC_VERSION__ >= 201112L) |
|
|
86 | #define ECB_CPP (__cplusplus+0) |
92 | #define ECB_CPP (__cplusplus+0) |
87 | #define ECB_CPP11 (__cplusplus >= 201103L) |
93 | #define ECB_CPP11 (__cplusplus >= 201103L) |
|
|
94 | |
|
|
95 | #if ECB_CPP |
|
|
96 | #define ECB_C 0 |
|
|
97 | #define ECB_STDC_VERSION 0 |
|
|
98 | #else |
|
|
99 | #define ECB_C 1 |
|
|
100 | #define ECB_STDC_VERSION __STDC_VERSION__ |
|
|
101 | #endif |
|
|
102 | |
|
|
103 | #define ECB_C99 (ECB_STDC_VERSION >= 199901L) |
|
|
104 | #define ECB_C11 (ECB_STDC_VERSION >= 201112L) |
88 | |
105 | |
89 | #if ECB_CPP |
106 | #if ECB_CPP |
90 | #define ECB_EXTERN_C extern "C" |
107 | #define ECB_EXTERN_C extern "C" |
91 | #define ECB_EXTERN_C_BEG ECB_EXTERN_C { |
108 | #define ECB_EXTERN_C_BEG ECB_EXTERN_C { |
92 | #define ECB_EXTERN_C_END } |
109 | #define ECB_EXTERN_C_END } |
… | |
… | |
125 | || defined __ARM_ARCH_6K__ || defined __ARM_ARCH_6ZK__ |
142 | || 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") |
143 | #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__ \ |
144 | #elif defined __ARM_ARCH_7__ || defined __ARM_ARCH_7A__ \ |
128 | || defined __ARM_ARCH_7M__ || defined __ARM_ARCH_7R__ |
145 | || defined __ARM_ARCH_7M__ || defined __ARM_ARCH_7R__ |
129 | #define ECB_MEMORY_FENCE __asm__ __volatile__ ("dmb" : : : "memory") |
146 | #define ECB_MEMORY_FENCE __asm__ __volatile__ ("dmb" : : : "memory") |
130 | #elif __sparc || __sparc__ |
147 | #elif (__sparc || __sparc__) && !__sparcv8 |
131 | #define ECB_MEMORY_FENCE __asm__ __volatile__ ("membar #LoadStore | #LoadLoad | #StoreStore | #StoreLoad" : : : "memory") |
148 | #define ECB_MEMORY_FENCE __asm__ __volatile__ ("membar #LoadStore | #LoadLoad | #StoreStore | #StoreLoad" : : : "memory") |
132 | #define ECB_MEMORY_FENCE_ACQUIRE __asm__ __volatile__ ("membar #LoadStore | #LoadLoad" : : : "memory") |
149 | #define ECB_MEMORY_FENCE_ACQUIRE __asm__ __volatile__ ("membar #LoadStore | #LoadLoad" : : : "memory") |
133 | #define ECB_MEMORY_FENCE_RELEASE __asm__ __volatile__ ("membar #LoadStore | #StoreStore") |
150 | #define ECB_MEMORY_FENCE_RELEASE __asm__ __volatile__ ("membar #LoadStore | #StoreStore") |
134 | #elif defined __s390__ || defined __s390x__ |
151 | #elif defined __s390__ || defined __s390x__ |
135 | #define ECB_MEMORY_FENCE __asm__ __volatile__ ("bcr 15,0" : : : "memory") |
152 | #define ECB_MEMORY_FENCE __asm__ __volatile__ ("bcr 15,0" : : : "memory") |
136 | #elif defined __mips__ |
153 | #elif defined __mips__ |
|
|
154 | /* GNU/Linux emulates sync on mips1 architectures, so we force its use */ |
|
|
155 | /* 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") |
156 | #define ECB_MEMORY_FENCE __asm__ __volatile__ (".set mips2; sync; .set mips0" : : : "memory") |
138 | #elif defined __alpha__ |
157 | #elif defined __alpha__ |
139 | #define ECB_MEMORY_FENCE __asm__ __volatile__ ("mb" : : : "memory") |
158 | #define ECB_MEMORY_FENCE __asm__ __volatile__ ("mb" : : : "memory") |
140 | #elif defined __hppa__ |
159 | #elif defined __hppa__ |
141 | #define ECB_MEMORY_FENCE __asm__ __volatile__ ("" : : : "memory") |
160 | #define ECB_MEMORY_FENCE __asm__ __volatile__ ("" : : : "memory") |
142 | #define ECB_MEMORY_FENCE_RELEASE __asm__ __volatile__ ("") |
161 | #define ECB_MEMORY_FENCE_RELEASE __asm__ __volatile__ ("") |
143 | #elif defined __ia64__ |
162 | #elif defined __ia64__ |
144 | #define ECB_MEMORY_FENCE __asm__ __volatile__ ("mf" : : : "memory") |
163 | #define ECB_MEMORY_FENCE __asm__ __volatile__ ("mf" : : : "memory") |
|
|
164 | #elif defined __m68k__ |
|
|
165 | #define ECB_MEMORY_FENCE __asm__ __volatile__ ("" : : : "memory") |
|
|
166 | #elif defined __m88k__ |
|
|
167 | #define ECB_MEMORY_FENCE __asm__ __volatile__ ("tb1 0,%%r0,128" : : : "memory") |
|
|
168 | #elif defined __sh__ |
|
|
169 | #define ECB_MEMORY_FENCE __asm__ __volatile__ ("" : : : "memory") |
145 | #endif |
170 | #endif |
146 | #endif |
171 | #endif |
147 | #endif |
172 | #endif |
148 | |
173 | |
149 | #ifndef ECB_MEMORY_FENCE |
174 | #ifndef ECB_MEMORY_FENCE |
… | |
… | |
160 | * #define ECB_MEMORY_FENCE __c11_atomic_thread_fence (__ATOMIC_SEQ_CST) |
185 | * #define ECB_MEMORY_FENCE __c11_atomic_thread_fence (__ATOMIC_SEQ_CST) |
161 | */ |
186 | */ |
162 | |
187 | |
163 | #elif ECB_GCC_VERSION(4,4) || defined __INTEL_COMPILER || defined __clang__ |
188 | #elif ECB_GCC_VERSION(4,4) || defined __INTEL_COMPILER || defined __clang__ |
164 | #define ECB_MEMORY_FENCE __sync_synchronize () |
189 | #define ECB_MEMORY_FENCE __sync_synchronize () |
|
|
190 | #elif _MSC_VER >= 1500 /* VC++ 2008 */ |
|
|
191 | /* apparently, microsoft broke all the memory barrier stuff in Visual Studio 2008... */ |
|
|
192 | #pragma intrinsic(_ReadBarrier,_WriteBarrier,_ReadWriteBarrier) |
|
|
193 | #define ECB_MEMORY_FENCE _ReadWriteBarrier (); MemoryBarrier() |
|
|
194 | #define ECB_MEMORY_FENCE_ACQUIRE _ReadWriteBarrier (); MemoryBarrier() /* according to msdn, _ReadBarrier is not a load fence */ |
|
|
195 | #define ECB_MEMORY_FENCE_RELEASE _WriteBarrier (); MemoryBarrier() |
165 | #elif _MSC_VER >= 1400 /* VC++ 2005 */ |
196 | #elif _MSC_VER >= 1400 /* VC++ 2005 */ |
166 | #pragma intrinsic(_ReadBarrier,_WriteBarrier,_ReadWriteBarrier) |
197 | #pragma intrinsic(_ReadBarrier,_WriteBarrier,_ReadWriteBarrier) |
167 | #define ECB_MEMORY_FENCE _ReadWriteBarrier () |
198 | #define ECB_MEMORY_FENCE _ReadWriteBarrier () |
168 | #define ECB_MEMORY_FENCE_ACQUIRE _ReadWriteBarrier () /* according to msdn, _ReadBarrier is not a load fence */ |
199 | #define ECB_MEMORY_FENCE_ACQUIRE _ReadWriteBarrier () /* according to msdn, _ReadBarrier is not a load fence */ |
169 | #define ECB_MEMORY_FENCE_RELEASE _WriteBarrier () |
200 | #define ECB_MEMORY_FENCE_RELEASE _WriteBarrier () |
… | |
… | |
255 | #define ecb_is_constant(expr) __builtin_constant_p (expr) |
286 | #define ecb_is_constant(expr) __builtin_constant_p (expr) |
256 | #define ecb_expect(expr,value) __builtin_expect ((expr),(value)) |
287 | #define ecb_expect(expr,value) __builtin_expect ((expr),(value)) |
257 | #define ecb_prefetch(addr,rw,locality) __builtin_prefetch (addr, rw, locality) |
288 | #define ecb_prefetch(addr,rw,locality) __builtin_prefetch (addr, rw, locality) |
258 | #else |
289 | #else |
259 | #define ecb_attribute(attrlist) |
290 | #define ecb_attribute(attrlist) |
|
|
291 | |
|
|
292 | /* possible C11 impl for integral types |
|
|
293 | typedef struct ecb_is_constant_struct ecb_is_constant_struct; |
|
|
294 | #define ecb_is_constant(expr) _Generic ((1 ? (struct ecb_is_constant_struct *)0 : (void *)((expr) - (expr)), ecb_is_constant_struct *: 0, default: 1)) */ |
|
|
295 | |
260 | #define ecb_is_constant(expr) 0 |
296 | #define ecb_is_constant(expr) 0 |
261 | #define ecb_expect(expr,value) (expr) |
297 | #define ecb_expect(expr,value) (expr) |
262 | #define ecb_prefetch(addr,rw,locality) |
298 | #define ecb_prefetch(addr,rw,locality) |
263 | #endif |
299 | #endif |
264 | |
300 | |
… | |
… | |
556 | || defined __s390__ || defined __s390x__ \ |
592 | || defined __s390__ || defined __s390x__ \ |
557 | || defined __mips__ \ |
593 | || defined __mips__ \ |
558 | || defined __alpha__ \ |
594 | || defined __alpha__ \ |
559 | || defined __hppa__ \ |
595 | || defined __hppa__ \ |
560 | || defined __ia64__ \ |
596 | || defined __ia64__ \ |
|
|
597 | || defined __m68k__ \ |
|
|
598 | || defined __m88k__ \ |
|
|
599 | || defined __sh__ \ |
561 | || defined _M_IX86 || defined _M_AMD64 || defined _M_IA64 |
600 | || defined _M_IX86 || defined _M_AMD64 || defined _M_IA64 |
562 | #define ECB_STDFP 1 |
601 | #define ECB_STDFP 1 |
563 | #include <string.h> /* for memcpy */ |
602 | #include <string.h> /* for memcpy */ |
564 | #else |
603 | #else |
565 | #define ECB_STDFP 0 |
604 | #define ECB_STDFP 0 |
566 | #include <math.h> /* for frexp*, ldexp* */ |
|
|
567 | #endif |
605 | #endif |
568 | |
606 | |
569 | #ifndef ECB_NO_LIBM |
607 | #ifndef ECB_NO_LIBM |
|
|
608 | |
|
|
609 | #include <math.h> /* for frexp*, ldexp*, INFINITY, NAN */ |
|
|
610 | |
|
|
611 | /* only the oldest of old doesn't have this one. solaris. */ |
|
|
612 | #ifdef INFINITY |
|
|
613 | #define ECB_INFINITY INFINITY |
|
|
614 | #else |
|
|
615 | #define ECB_INFINITY HUGE_VAL |
|
|
616 | #endif |
|
|
617 | |
|
|
618 | #ifdef NAN |
|
|
619 | #define ECB_NAN NAN |
|
|
620 | #else |
|
|
621 | #define ECB_NAN ECB_INFINITY |
|
|
622 | #endif |
|
|
623 | |
|
|
624 | /* converts an ieee half/binary16 to a float */ |
|
|
625 | ecb_function_ float ecb_binary16_to_float (uint16_t x) ecb_const; |
|
|
626 | ecb_function_ float |
|
|
627 | ecb_binary16_to_float (uint16_t x) |
|
|
628 | { |
|
|
629 | int e = (x >> 10) & 0x1f; |
|
|
630 | int m = x & 0x3ff; |
|
|
631 | float r; |
|
|
632 | |
|
|
633 | if (!e ) r = ldexpf (m , -24); |
|
|
634 | else if (e != 31) r = ldexpf (m + 0x400, e - 25); |
|
|
635 | else if (m ) r = ECB_NAN; |
|
|
636 | else r = ECB_INFINITY; |
|
|
637 | |
|
|
638 | return x & 0x8000 ? -r : r; |
|
|
639 | } |
570 | |
640 | |
571 | /* convert a float to ieee single/binary32 */ |
641 | /* convert a float to ieee single/binary32 */ |
572 | ecb_function_ uint32_t ecb_float_to_binary32 (float x) ecb_const; |
642 | ecb_function_ uint32_t ecb_float_to_binary32 (float x) ecb_const; |
573 | ecb_function_ uint32_t |
643 | ecb_function_ uint32_t |
574 | ecb_float_to_binary32 (float x) |
644 | ecb_float_to_binary32 (float x) |