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

Comparing libecb/ecb.h (file contents):
Revision 1.91 by root, Tue May 29 14:30:44 2012 UTC vs.
Revision 1.181 by root, Mon Jan 20 13:16:01 2020 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 (©) 2009-2015 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:
23 * PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; 23 * PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS;
24 * OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, 24 * OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY,
25 * WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTH- 25 * WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTH-
26 * ERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED 26 * ERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED
27 * OF THE POSSIBILITY OF SUCH DAMAGE. 27 * OF THE POSSIBILITY OF SUCH DAMAGE.
28 *
29 * Alternatively, the contents of this file may be used under the terms of
30 * the GNU General Public License ("GPL") version 2 or any later version,
31 * in which case the provisions of the GPL are applicable instead of
32 * the above. If you wish to allow the use of your version of this file
33 * only under the terms of the GPL and not to allow others to use your
34 * version of this file under the BSD license, indicate your decision
35 * by deleting the provisions above and replace them with the notice
36 * and other provisions required by the GPL. If you do not delete the
37 * provisions above, a recipient may use your version of this file under
38 * either the BSD or the GPL.
28 */ 39 */
29 40
30#ifndef ECB_H 41#ifndef ECB_H
31#define ECB_H 42#define ECB_H
32 43
33/* 16 bits major, 16 bits minor */ 44/* 16 bits major, 16 bits minor */
34#define ECB_VERSION 0x00010001 45#define ECB_VERSION 0x00010008
35 46
36#ifdef _WIN32 47#ifdef _WIN32
37 typedef signed char int8_t; 48 typedef signed char int8_t;
38 typedef unsigned char uint8_t; 49 typedef unsigned char uint8_t;
50 typedef signed char int_fast8_t;
51 typedef unsigned char uint_fast8_t;
39 typedef signed short int16_t; 52 typedef signed short int16_t;
40 typedef unsigned short uint16_t; 53 typedef unsigned short uint16_t;
54 typedef signed int int_fast16_t;
55 typedef unsigned int uint_fast16_t;
41 typedef signed int int32_t; 56 typedef signed int int32_t;
42 typedef unsigned int uint32_t; 57 typedef unsigned int uint32_t;
58 typedef signed int int_fast32_t;
59 typedef unsigned int uint_fast32_t;
43 #if __GNUC__ 60 #if __GNUC__
44 typedef signed long long int64_t; 61 typedef signed long long int64_t;
45 typedef unsigned long long uint64_t; 62 typedef unsigned long long uint64_t;
46 #else /* _MSC_VER || __BORLANDC__ */ 63 #else /* _MSC_VER || __BORLANDC__ */
47 typedef signed __int64 int64_t; 64 typedef signed __int64 int64_t;
48 typedef unsigned __int64 uint64_t; 65 typedef unsigned __int64 uint64_t;
49 #endif 66 #endif
67 typedef int64_t int_fast64_t;
68 typedef uint64_t uint_fast64_t;
50 #ifdef _WIN64 69 #ifdef _WIN64
51 #define ECB_PTRSIZE 8 70 #define ECB_PTRSIZE 8
52 typedef uint64_t uintptr_t; 71 typedef uint64_t uintptr_t;
53 typedef int64_t intptr_t; 72 typedef int64_t intptr_t;
54 #else 73 #else
55 #define ECB_PTRSIZE 4 74 #define ECB_PTRSIZE 4
56 typedef uint32_t uintptr_t; 75 typedef uint32_t uintptr_t;
57 typedef int32_t intptr_t; 76 typedef int32_t intptr_t;
58 #endif 77 #endif
59 typedef intptr_t ptrdiff_t;
60#else 78#else
61 #include <inttypes.h> 79 #include <inttypes.h>
62 #if UINTMAX_MAX > 0xffffffffU 80 #if (defined INTPTR_MAX ? INTPTR_MAX : ULONG_MAX) > 0xffffffffU
63 #define ECB_PTRSIZE 8 81 #define ECB_PTRSIZE 8
64 #else 82 #else
65 #define ECB_PTRSIZE 4 83 #define ECB_PTRSIZE 4
84 #endif
85#endif
86
87#define ECB_GCC_AMD64 (__amd64 || __amd64__ || __x86_64 || __x86_64__)
88#define ECB_MSVC_AMD64 (_M_AMD64 || _M_X64)
89
90#ifndef ECB_OPTIMIZE_SIZE
91 #if __OPTIMIZE_SIZE__
92 #define ECB_OPTIMIZE_SIZE 1
93 #else
94 #define ECB_OPTIMIZE_SIZE 0
95 #endif
96#endif
97
98/* work around x32 idiocy by defining proper macros */
99#if ECB_GCC_AMD64 || ECB_MSVC_AMD64
100 #if _ILP32
101 #define ECB_AMD64_X32 1
102 #else
103 #define ECB_AMD64 1
66 #endif 104 #endif
67#endif 105#endif
68 106
69/* many compilers define _GNUC_ to some versions but then only implement 107/* many compilers define _GNUC_ to some versions but then only implement
70 * what their idiot authors think are the "more important" extensions, 108 * what their idiot authors think are the "more important" extensions,
71 * causing enormous grief in return for some better fake benchmark numbers. 109 * causing enormous grief in return for some better fake benchmark numbers.
72 * or so. 110 * or so.
73 * we try to detect these and simply assume they are not gcc - if they have 111 * we try to detect these and simply assume they are not gcc - if they have
74 * an issue with that they should have done it right in the first place. 112 * an issue with that they should have done it right in the first place.
75 */ 113 */
76#ifndef ECB_GCC_VERSION
77 #if !defined __GNUC_MINOR__ || defined __INTEL_COMPILER || defined __SUNPRO_C || defined __SUNPRO_CC || defined __llvm__ || defined __clang__ 114#if !defined __GNUC_MINOR__ || defined __INTEL_COMPILER || defined __SUNPRO_C || defined __SUNPRO_CC || defined __llvm__ || defined __clang__
78 #define ECB_GCC_VERSION(major,minor) 0 115 #define ECB_GCC_VERSION(major,minor) 0
79 #else 116#else
80 #define ECB_GCC_VERSION(major,minor) (__GNUC__ > (major) || (__GNUC__ == (major) && __GNUC_MINOR__ >= (minor))) 117 #define ECB_GCC_VERSION(major,minor) (__GNUC__ > (major) || (__GNUC__ == (major) && __GNUC_MINOR__ >= (minor)))
81 #endif 118#endif
82#endif
83 119
84#define ECB_C (__STDC__+0) /* this assumes that __STDC__ is either empty or a number */ 120#define ECB_CLANG_VERSION(major,minor) (__clang_major__ > (major) || (__clang_major__ == (major) && __clang_minor__ >= (minor)))
85#define ECB_C99 (__STDC_VERSION__ >= 199901L) 121
86#define ECB_C11 (__STDC_VERSION__ >= 201112L) 122#if __clang__ && defined __has_builtin
123 #define ECB_CLANG_BUILTIN(x) __has_builtin (x)
124#else
125 #define ECB_CLANG_BUILTIN(x) 0
126#endif
127
128#if __clang__ && defined __has_extension
129 #define ECB_CLANG_EXTENSION(x) __has_extension (x)
130#else
131 #define ECB_CLANG_EXTENSION(x) 0
132#endif
133
87#define ECB_CPP (__cplusplus+0) 134#define ECB_CPP (__cplusplus+0)
88#define ECB_CPP98 (__cplusplus >= 199711L)
89#define ECB_CPP11 (__cplusplus >= 201103L) 135#define ECB_CPP11 (__cplusplus >= 201103L)
136#define ECB_CPP14 (__cplusplus >= 201402L)
137#define ECB_CPP17 (__cplusplus >= 201703L)
138
139#if ECB_CPP
140 #define ECB_C 0
141 #define ECB_STDC_VERSION 0
142#else
143 #define ECB_C 1
144 #define ECB_STDC_VERSION __STDC_VERSION__
145#endif
146
147#define ECB_C99 (ECB_STDC_VERSION >= 199901L)
148#define ECB_C11 (ECB_STDC_VERSION >= 201112L)
149#define ECB_C17 (ECB_STDC_VERSION >= 201710L)
150
151#if ECB_CPP
152 #define ECB_EXTERN_C extern "C"
153 #define ECB_EXTERN_C_BEG ECB_EXTERN_C {
154 #define ECB_EXTERN_C_END }
155#else
156 #define ECB_EXTERN_C extern
157 #define ECB_EXTERN_C_BEG
158 #define ECB_EXTERN_C_END
159#endif
90 160
91/*****************************************************************************/ 161/*****************************************************************************/
92 162
93/* ECB_NO_THREADS - ecb is not used by multiple threads, ever */ 163/* ECB_NO_THREADS - ecb is not used by multiple threads, ever */
94/* ECB_NO_SMP - ecb might be used in multiple threads, but only on a single cpu */ 164/* ECB_NO_SMP - ecb might be used in multiple threads, but only on a single cpu */
95 165
96#if ECB_NO_THREADS 166#if ECB_NO_THREADS
97# define ECB_NO_SMP 1 167 #define ECB_NO_SMP 1
98#endif 168#endif
99 169
100#if ECB_NO_THREADS || ECB_NO_SMP 170#if ECB_NO_SMP
101 #define ECB_MEMORY_FENCE do { } while (0) 171 #define ECB_MEMORY_FENCE do { } while (0)
102#endif 172#endif
103 173
104#ifndef ECB_MEMORY_FENCE 174/* http://www-01.ibm.com/support/knowledgecenter/SSGH3R_13.1.0/com.ibm.xlcpp131.aix.doc/compiler_ref/compiler_builtins.html */
105 #if ECB_C11 && !defined __STDC_NO_ATOMICS__ 175#if __xlC__ && ECB_CPP
106 /* we assume that these memory fences work on all variables/all memory accesses, */ 176 #include <builtins.h>
107 /* not just C11 atomics and atomic accesses */
108 #include <stdatomic.h>
109 #define ECB_MEMORY_FENCE atomic_thread_fence (memory_order_acq_rel)
110 #define ECB_MEMORY_FENCE_ACQUIRE atomic_thread_fence (memory_order_acquire)
111 #define ECB_MEMORY_FENCE_RELEASE atomic_thread_fence (memory_order_release)
112 #endif 177#endif
178
179#if 1400 <= _MSC_VER
180 #include <intrin.h> /* fence functions _ReadBarrier, also bit search functions _BitScanReverse */
113#endif 181#endif
114 182
115#ifndef ECB_MEMORY_FENCE 183#ifndef ECB_MEMORY_FENCE
116 #if ECB_GCC_VERSION(2,5) || defined __INTEL_COMPILER || (__llvm__ && __GNUC__) || __SUNPRO_C >= 0x5110 || __SUNPRO_CC >= 0x5110 184 #if ECB_GCC_VERSION(2,5) || defined __INTEL_COMPILER || (__llvm__ && __GNUC__) || __SUNPRO_C >= 0x5110 || __SUNPRO_CC >= 0x5110
185 #define ECB_MEMORY_FENCE_RELAXED __asm__ __volatile__ ("" : : : "memory")
117 #if __i386 || __i386__ 186 #if __i386 || __i386__
118 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("lock; orb $0, -1(%%esp)" : : : "memory") 187 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("lock; orb $0, -1(%%esp)" : : : "memory")
119 #define ECB_MEMORY_FENCE_ACQUIRE ECB_MEMORY_FENCE /* non-lock xchg might be enough */ 188 #define ECB_MEMORY_FENCE_ACQUIRE __asm__ __volatile__ ("" : : : "memory")
120 #define ECB_MEMORY_FENCE_RELEASE do { } while (0) /* unlikely to change in future cpus */ 189 #define ECB_MEMORY_FENCE_RELEASE __asm__ __volatile__ ("" : : : "memory")
121 #elif __amd64 || __amd64__ || __x86_64 || __x86_64__ 190 #elif ECB_GCC_AMD64
122 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("mfence" : : : "memory") 191 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("mfence" : : : "memory")
123 #define ECB_MEMORY_FENCE_ACQUIRE __asm__ __volatile__ ("lfence" : : : "memory") 192 #define ECB_MEMORY_FENCE_ACQUIRE __asm__ __volatile__ ("" : : : "memory")
124 #define ECB_MEMORY_FENCE_RELEASE __asm__ __volatile__ ("sfence") /* play safe - not needed in any current cpu */ 193 #define ECB_MEMORY_FENCE_RELEASE __asm__ __volatile__ ("" : : : "memory")
125 #elif __powerpc__ || __ppc__ || __powerpc64__ || __ppc64__ 194 #elif __powerpc__ || __ppc__ || __powerpc64__ || __ppc64__
126 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("sync" : : : "memory") 195 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("sync" : : : "memory")
196 #elif defined __ARM_ARCH_2__ \
197 || defined __ARM_ARCH_3__ || defined __ARM_ARCH_3M__ \
198 || defined __ARM_ARCH_4__ || defined __ARM_ARCH_4T__ \
199 || defined __ARM_ARCH_5__ || defined __ARM_ARCH_5E__ \
200 || defined __ARM_ARCH_5T__ || defined __ARM_ARCH_5TE__ \
201 || defined __ARM_ARCH_5TEJ__
202 /* should not need any, unless running old code on newer cpu - arm doesn't support that */
127 #elif defined __ARM_ARCH_6__ || defined __ARM_ARCH_6J__ \ 203 #elif defined __ARM_ARCH_6__ || defined __ARM_ARCH_6J__ \
128 || defined __ARM_ARCH_6K__ || defined __ARM_ARCH_6ZK__ 204 || defined __ARM_ARCH_6K__ || defined __ARM_ARCH_6ZK__ \
205 || defined __ARM_ARCH_6T2__
129 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("mcr p15,0,%0,c7,c10,5" : : "r" (0) : "memory") 206 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("mcr p15,0,%0,c7,c10,5" : : "r" (0) : "memory")
130 #elif defined __ARM_ARCH_7__ || defined __ARM_ARCH_7A__ \ 207 #elif defined __ARM_ARCH_7__ || defined __ARM_ARCH_7A__ \
131 || defined __ARM_ARCH_7M__ || defined __ARM_ARCH_7R__ 208 || defined __ARM_ARCH_7R__ || defined __ARM_ARCH_7M__
132 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("dmb" : : : "memory") 209 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("dmb" : : : "memory")
133 #elif __sparc || __sparc__ 210 #elif __aarch64__
211 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("dmb ish" : : : "memory")
212 #elif (__sparc || __sparc__) && !(__sparc_v8__ || defined __sparcv8)
134 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("membar #LoadStore | #LoadLoad | #StoreStore | #StoreLoad | " : : : "memory") 213 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("membar #LoadStore | #LoadLoad | #StoreStore | #StoreLoad" : : : "memory")
135 #define ECB_MEMORY_FENCE_ACQUIRE __asm__ __volatile__ ("membar #LoadStore | #LoadLoad" : : : "memory") 214 #define ECB_MEMORY_FENCE_ACQUIRE __asm__ __volatile__ ("membar #LoadStore | #LoadLoad" : : : "memory")
136 #define ECB_MEMORY_FENCE_RELEASE __asm__ __volatile__ ("membar #LoadStore | #StoreStore") 215 #define ECB_MEMORY_FENCE_RELEASE __asm__ __volatile__ ("membar #LoadStore | #StoreStore")
137 #elif defined __s390__ || defined __s390x__ 216 #elif defined __s390__ || defined __s390x__
138 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("bcr 15,0" : : : "memory") 217 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("bcr 15,0" : : : "memory")
139 #elif defined __mips__ 218 #elif defined __mips__
219 /* GNU/Linux emulates sync on mips1 architectures, so we force its use */
220 /* anybody else who still uses mips1 is supposed to send in their version, with detection code. */
140 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("sync" : : : "memory") 221 #define ECB_MEMORY_FENCE __asm__ __volatile__ (".set mips2; sync; .set mips0" : : : "memory")
141 #elif defined __alpha__ 222 #elif defined __alpha__
142 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("mb" : : : "memory") 223 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("mb" : : : "memory")
224 #elif defined __hppa__
225 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("" : : : "memory")
226 #define ECB_MEMORY_FENCE_RELEASE __asm__ __volatile__ ("")
227 #elif defined __ia64__
228 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("mf" : : : "memory")
229 #elif defined __m68k__
230 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("" : : : "memory")
231 #elif defined __m88k__
232 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("tb1 0,%%r0,128" : : : "memory")
233 #elif defined __sh__
234 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("" : : : "memory")
143 #endif 235 #endif
144 #endif 236 #endif
145#endif 237#endif
146 238
147#ifndef ECB_MEMORY_FENCE 239#ifndef ECB_MEMORY_FENCE
240 #if ECB_GCC_VERSION(4,7)
241 /* see comment below (stdatomic.h) about the C11 memory model. */
242 #define ECB_MEMORY_FENCE __atomic_thread_fence (__ATOMIC_SEQ_CST)
243 #define ECB_MEMORY_FENCE_ACQUIRE __atomic_thread_fence (__ATOMIC_ACQUIRE)
244 #define ECB_MEMORY_FENCE_RELEASE __atomic_thread_fence (__ATOMIC_RELEASE)
245 #define ECB_MEMORY_FENCE_RELAXED __atomic_thread_fence (__ATOMIC_RELAXED)
246
247 #elif ECB_CLANG_EXTENSION(c_atomic)
248 /* see comment below (stdatomic.h) about the C11 memory model. */
249 #define ECB_MEMORY_FENCE __c11_atomic_thread_fence (__ATOMIC_SEQ_CST)
250 #define ECB_MEMORY_FENCE_ACQUIRE __c11_atomic_thread_fence (__ATOMIC_ACQUIRE)
251 #define ECB_MEMORY_FENCE_RELEASE __c11_atomic_thread_fence (__ATOMIC_RELEASE)
252 #define ECB_MEMORY_FENCE_RELAXED __c11_atomic_thread_fence (__ATOMIC_RELAXED)
253
148 #if ECB_GCC_VERSION(4,4) || defined __INTEL_COMPILER || defined __clang__ 254 #elif ECB_GCC_VERSION(4,4) || defined __INTEL_COMPILER || defined __clang__
149 #define ECB_MEMORY_FENCE __sync_synchronize () 255 #define ECB_MEMORY_FENCE __sync_synchronize ()
150 /*#define ECB_MEMORY_FENCE_ACQUIRE ({ char dummy = 0; __sync_lock_test_and_set (&dummy, 1); }) */ 256 #elif _MSC_VER >= 1500 /* VC++ 2008 */
151 /*#define ECB_MEMORY_FENCE_RELEASE ({ char dummy = 1; __sync_lock_release (&dummy ); }) */ 257 /* apparently, microsoft broke all the memory barrier stuff in Visual Studio 2008... */
258 #pragma intrinsic(_ReadBarrier,_WriteBarrier,_ReadWriteBarrier)
259 #define ECB_MEMORY_FENCE _ReadWriteBarrier (); MemoryBarrier()
260 #define ECB_MEMORY_FENCE_ACQUIRE _ReadWriteBarrier (); MemoryBarrier() /* according to msdn, _ReadBarrier is not a load fence */
261 #define ECB_MEMORY_FENCE_RELEASE _WriteBarrier (); MemoryBarrier()
152 #elif _MSC_VER >= 1400 /* VC++ 2005 */ 262 #elif _MSC_VER >= 1400 /* VC++ 2005 */
153 #pragma intrinsic(_ReadBarrier,_WriteBarrier,_ReadWriteBarrier) 263 #pragma intrinsic(_ReadBarrier,_WriteBarrier,_ReadWriteBarrier)
154 #define ECB_MEMORY_FENCE _ReadWriteBarrier () 264 #define ECB_MEMORY_FENCE _ReadWriteBarrier ()
155 #define ECB_MEMORY_FENCE_ACQUIRE _ReadWriteBarrier () /* according to msdn, _ReadBarrier is not a load fence */ 265 #define ECB_MEMORY_FENCE_ACQUIRE _ReadWriteBarrier () /* according to msdn, _ReadBarrier is not a load fence */
156 #define ECB_MEMORY_FENCE_RELEASE _WriteBarrier () 266 #define ECB_MEMORY_FENCE_RELEASE _WriteBarrier ()
157 #elif defined _WIN32 267 #elif defined _WIN32
158 #include <WinNT.h> 268 #include <WinNT.h>
159 #define ECB_MEMORY_FENCE MemoryBarrier () /* actually just xchg on x86... scary */ 269 #define ECB_MEMORY_FENCE MemoryBarrier () /* actually just xchg on x86... scary */
160 #elif __SUNPRO_C >= 0x5110 || __SUNPRO_CC >= 0x5110 270 #elif __SUNPRO_C >= 0x5110 || __SUNPRO_CC >= 0x5110
161 #include <mbarrier.h> 271 #include <mbarrier.h>
162 #define ECB_MEMORY_FENCE __machine_rw_barrier () 272 #define ECB_MEMORY_FENCE __machine_rw_barrier ()
163 #define ECB_MEMORY_FENCE_ACQUIRE __machine_r_barrier () 273 #define ECB_MEMORY_FENCE_ACQUIRE __machine_acq_barrier ()
164 #define ECB_MEMORY_FENCE_RELEASE __machine_w_barrier () 274 #define ECB_MEMORY_FENCE_RELEASE __machine_rel_barrier ()
275 #define ECB_MEMORY_FENCE_RELAXED __compiler_barrier ()
165 #elif __xlC__ 276 #elif __xlC__
166 #define ECB_MEMORY_FENCE __sync () 277 #define ECB_MEMORY_FENCE __sync ()
278 #endif
279#endif
280
281#ifndef ECB_MEMORY_FENCE
282 #if ECB_C11 && !defined __STDC_NO_ATOMICS__
283 /* we assume that these memory fences work on all variables/all memory accesses, */
284 /* not just C11 atomics and atomic accesses */
285 #include <stdatomic.h>
286 #define ECB_MEMORY_FENCE atomic_thread_fence (memory_order_seq_cst)
287 #define ECB_MEMORY_FENCE_ACQUIRE atomic_thread_fence (memory_order_acquire)
288 #define ECB_MEMORY_FENCE_RELEASE atomic_thread_fence (memory_order_release)
167 #endif 289 #endif
168#endif 290#endif
169 291
170#ifndef ECB_MEMORY_FENCE 292#ifndef ECB_MEMORY_FENCE
171 #if !ECB_AVOID_PTHREADS 293 #if !ECB_AVOID_PTHREADS
191 313
192#if !defined ECB_MEMORY_FENCE_RELEASE && defined ECB_MEMORY_FENCE 314#if !defined ECB_MEMORY_FENCE_RELEASE && defined ECB_MEMORY_FENCE
193 #define ECB_MEMORY_FENCE_RELEASE ECB_MEMORY_FENCE 315 #define ECB_MEMORY_FENCE_RELEASE ECB_MEMORY_FENCE
194#endif 316#endif
195 317
318#if !defined ECB_MEMORY_FENCE_RELAXED && defined ECB_MEMORY_FENCE
319 #define ECB_MEMORY_FENCE_RELAXED ECB_MEMORY_FENCE /* very heavy-handed */
320#endif
321
196/*****************************************************************************/ 322/*****************************************************************************/
197 323
198#if __cplusplus 324#if ECB_CPP
199 #define ecb_inline static inline 325 #define ecb_inline static inline
200#elif ECB_GCC_VERSION(2,5) 326#elif ECB_GCC_VERSION(2,5)
201 #define ecb_inline static __inline__ 327 #define ecb_inline static __inline__
202#elif ECB_C99 328#elif ECB_C99
203 #define ecb_inline static inline 329 #define ecb_inline static inline
217 343
218#define ECB_CONCAT_(a, b) a ## b 344#define ECB_CONCAT_(a, b) a ## b
219#define ECB_CONCAT(a, b) ECB_CONCAT_(a, b) 345#define ECB_CONCAT(a, b) ECB_CONCAT_(a, b)
220#define ECB_STRINGIFY_(a) # a 346#define ECB_STRINGIFY_(a) # a
221#define ECB_STRINGIFY(a) ECB_STRINGIFY_(a) 347#define ECB_STRINGIFY(a) ECB_STRINGIFY_(a)
348#define ECB_STRINGIFY_EXPR(expr) ((expr), ECB_STRINGIFY_ (expr))
222 349
223#define ecb_function_ ecb_inline 350#define ecb_function_ ecb_inline
224 351
225#if ECB_GCC_VERSION(3,1) 352#if ECB_GCC_VERSION(3,1) || ECB_CLANG_VERSION(2,8)
226 #define ecb_attribute(attrlist) __attribute__(attrlist) 353 #define ecb_attribute(attrlist) __attribute__ (attrlist)
354#else
355 #define ecb_attribute(attrlist)
356#endif
357
358#if ECB_GCC_VERSION(3,1) || ECB_CLANG_BUILTIN(__builtin_constant_p)
227 #define ecb_is_constant(expr) __builtin_constant_p (expr) 359 #define ecb_is_constant(expr) __builtin_constant_p (expr)
360#else
361 /* possible C11 impl for integral types
362 typedef struct ecb_is_constant_struct ecb_is_constant_struct;
363 #define ecb_is_constant(expr) _Generic ((1 ? (struct ecb_is_constant_struct *)0 : (void *)((expr) - (expr)), ecb_is_constant_struct *: 0, default: 1)) */
364
365 #define ecb_is_constant(expr) 0
366#endif
367
368#if ECB_GCC_VERSION(3,1) || ECB_CLANG_BUILTIN(__builtin_expect)
228 #define ecb_expect(expr,value) __builtin_expect ((expr),(value)) 369 #define ecb_expect(expr,value) __builtin_expect ((expr),(value))
370#else
371 #define ecb_expect(expr,value) (expr)
372#endif
373
374#if ECB_GCC_VERSION(3,1) || ECB_CLANG_BUILTIN(__builtin_prefetch)
229 #define ecb_prefetch(addr,rw,locality) __builtin_prefetch (addr, rw, locality) 375 #define ecb_prefetch(addr,rw,locality) __builtin_prefetch (addr, rw, locality)
230#else 376#else
231 #define ecb_attribute(attrlist)
232 #define ecb_is_constant(expr) 0
233 #define ecb_expect(expr,value) (expr)
234 #define ecb_prefetch(addr,rw,locality) 377 #define ecb_prefetch(addr,rw,locality)
235#endif 378#endif
236 379
237/* no emulation for ecb_decltype */ 380/* no emulation for ecb_decltype */
238#if ECB_GCC_VERSION(4,5) 381#if ECB_CPP11
382 // older implementations might have problems with decltype(x)::type, work around it
383 template<class T> struct ecb_decltype_t { typedef T type; };
239 #define ecb_decltype(x) __decltype(x) 384 #define ecb_decltype(x) ecb_decltype_t<decltype (x)>::type
240#elif ECB_GCC_VERSION(3,0) 385#elif ECB_GCC_VERSION(3,0) || ECB_CLANG_VERSION(2,8)
241 #define ecb_decltype(x) __typeof(x) 386 #define ecb_decltype(x) __typeof__ (x)
242#endif 387#endif
243 388
389#if _MSC_VER >= 1300
390 #define ecb_deprecated __declspec (deprecated)
391#else
392 #define ecb_deprecated ecb_attribute ((__deprecated__))
393#endif
394
395#if _MSC_VER >= 1500
396 #define ecb_deprecated_message(msg) __declspec (deprecated (msg))
397#elif ECB_GCC_VERSION(4,5)
398 #define ecb_deprecated_message(msg) ecb_attribute ((__deprecated__ (msg))
399#else
400 #define ecb_deprecated_message(msg) ecb_deprecated
401#endif
402
403#if _MSC_VER >= 1400
404 #define ecb_noinline __declspec (noinline)
405#else
244#define ecb_noinline ecb_attribute ((__noinline__)) 406 #define ecb_noinline ecb_attribute ((__noinline__))
407#endif
408
245#define ecb_unused ecb_attribute ((__unused__)) 409#define ecb_unused ecb_attribute ((__unused__))
246#define ecb_const ecb_attribute ((__const__)) 410#define ecb_const ecb_attribute ((__const__))
247#define ecb_pure ecb_attribute ((__pure__)) 411#define ecb_pure ecb_attribute ((__pure__))
248 412
249#if ECB_C11 413#if ECB_C11 || __IBMC_NORETURN
414 /* http://www-01.ibm.com/support/knowledgecenter/SSGH3R_13.1.0/com.ibm.xlcpp131.aix.doc/language_ref/noreturn.html */
250 #define ecb_noreturn _Noreturn 415 #define ecb_noreturn _Noreturn
416#elif ECB_CPP11
417 #define ecb_noreturn [[noreturn]]
418#elif _MSC_VER >= 1200
419 /* http://msdn.microsoft.com/en-us/library/k6ktzx3s.aspx */
420 #define ecb_noreturn __declspec (noreturn)
251#else 421#else
252 #define ecb_noreturn ecb_attribute ((__noreturn__)) 422 #define ecb_noreturn ecb_attribute ((__noreturn__))
253#endif 423#endif
254 424
255#if ECB_GCC_VERSION(4,3) 425#if ECB_GCC_VERSION(4,3)
270/* for compatibility to the rest of the world */ 440/* for compatibility to the rest of the world */
271#define ecb_likely(expr) ecb_expect_true (expr) 441#define ecb_likely(expr) ecb_expect_true (expr)
272#define ecb_unlikely(expr) ecb_expect_false (expr) 442#define ecb_unlikely(expr) ecb_expect_false (expr)
273 443
274/* count trailing zero bits and count # of one bits */ 444/* count trailing zero bits and count # of one bits */
275#if ECB_GCC_VERSION(3,4) 445#if ECB_GCC_VERSION(3,4) \
446 || (ECB_CLANG_BUILTIN(__builtin_clz) && ECB_CLANG_BUILTIN(__builtin_clzll) \
447 && ECB_CLANG_BUILTIN(__builtin_ctz) && ECB_CLANG_BUILTIN(__builtin_ctzll) \
448 && ECB_CLANG_BUILTIN(__builtin_popcount))
276 /* we assume int == 32 bit, long == 32 or 64 bit and long long == 64 bit */ 449 /* we assume int == 32 bit, long == 32 or 64 bit and long long == 64 bit */
277 #define ecb_ld32(x) (__builtin_clz (x) ^ 31) 450 #define ecb_ld32(x) (__builtin_clz (x) ^ 31)
278 #define ecb_ld64(x) (__builtin_clzll (x) ^ 63) 451 #define ecb_ld64(x) (__builtin_clzll (x) ^ 63)
279 #define ecb_ctz32(x) __builtin_ctz (x) 452 #define ecb_ctz32(x) __builtin_ctz (x)
280 #define ecb_ctz64(x) __builtin_ctzll (x) 453 #define ecb_ctz64(x) __builtin_ctzll (x)
281 #define ecb_popcount32(x) __builtin_popcount (x) 454 #define ecb_popcount32(x) __builtin_popcount (x)
282 /* no popcountll */ 455 /* no popcountll */
283#else 456#else
284 ecb_function_ int ecb_ctz32 (uint32_t x) ecb_const; 457 ecb_function_ ecb_const int ecb_ctz32 (uint32_t x);
285 ecb_function_ int 458 ecb_function_ ecb_const int
286 ecb_ctz32 (uint32_t x) 459 ecb_ctz32 (uint32_t x)
287 { 460 {
461#if 1400 <= _MSC_VER && (_M_IX86 || _M_X64 || _M_IA64 || _M_ARM)
462 unsigned long r;
463 _BitScanForward (&r, x);
464 return (int)r;
465#else
288 int r = 0; 466 int r = 0;
289 467
290 x &= ~x + 1; /* this isolates the lowest bit */ 468 x &= ~x + 1; /* this isolates the lowest bit */
291 469
292#if ECB_branchless_on_i386 470#if ECB_branchless_on_i386
302 if (x & 0xff00ff00) r += 8; 480 if (x & 0xff00ff00) r += 8;
303 if (x & 0xffff0000) r += 16; 481 if (x & 0xffff0000) r += 16;
304#endif 482#endif
305 483
306 return r; 484 return r;
485#endif
307 } 486 }
308 487
309 ecb_function_ int ecb_ctz64 (uint64_t x) ecb_const; 488 ecb_function_ ecb_const int ecb_ctz64 (uint64_t x);
310 ecb_function_ int 489 ecb_function_ ecb_const int
311 ecb_ctz64 (uint64_t x) 490 ecb_ctz64 (uint64_t x)
312 { 491 {
492#if 1400 <= _MSC_VER && (_M_X64 || _M_IA64 || _M_ARM)
493 unsigned long r;
494 _BitScanForward64 (&r, x);
495 return (int)r;
496#else
313 int shift = x & 0xffffffffU ? 0 : 32; 497 int shift = x & 0xffffffff ? 0 : 32;
314 return ecb_ctz32 (x >> shift) + shift; 498 return ecb_ctz32 (x >> shift) + shift;
499#endif
315 } 500 }
316 501
317 ecb_function_ int ecb_popcount32 (uint32_t x) ecb_const; 502 ecb_function_ ecb_const int ecb_popcount32 (uint32_t x);
318 ecb_function_ int 503 ecb_function_ ecb_const int
319 ecb_popcount32 (uint32_t x) 504 ecb_popcount32 (uint32_t x)
320 { 505 {
321 x -= (x >> 1) & 0x55555555; 506 x -= (x >> 1) & 0x55555555;
322 x = ((x >> 2) & 0x33333333) + (x & 0x33333333); 507 x = ((x >> 2) & 0x33333333) + (x & 0x33333333);
323 x = ((x >> 4) + x) & 0x0f0f0f0f; 508 x = ((x >> 4) + x) & 0x0f0f0f0f;
324 x *= 0x01010101; 509 x *= 0x01010101;
325 510
326 return x >> 24; 511 return x >> 24;
327 } 512 }
328 513
329 ecb_function_ int ecb_ld32 (uint32_t x) ecb_const; 514 ecb_function_ ecb_const int ecb_ld32 (uint32_t x);
330 ecb_function_ int ecb_ld32 (uint32_t x) 515 ecb_function_ ecb_const int ecb_ld32 (uint32_t x)
331 { 516 {
517#if 1400 <= _MSC_VER && (_M_IX86 || _M_X64 || _M_IA64 || _M_ARM)
518 unsigned long r;
519 _BitScanReverse (&r, x);
520 return (int)r;
521#else
332 int r = 0; 522 int r = 0;
333 523
334 if (x >> 16) { x >>= 16; r += 16; } 524 if (x >> 16) { x >>= 16; r += 16; }
335 if (x >> 8) { x >>= 8; r += 8; } 525 if (x >> 8) { x >>= 8; r += 8; }
336 if (x >> 4) { x >>= 4; r += 4; } 526 if (x >> 4) { x >>= 4; r += 4; }
337 if (x >> 2) { x >>= 2; r += 2; } 527 if (x >> 2) { x >>= 2; r += 2; }
338 if (x >> 1) { r += 1; } 528 if (x >> 1) { r += 1; }
339 529
340 return r; 530 return r;
531#endif
341 } 532 }
342 533
343 ecb_function_ int ecb_ld64 (uint64_t x) ecb_const; 534 ecb_function_ ecb_const int ecb_ld64 (uint64_t x);
344 ecb_function_ int ecb_ld64 (uint64_t x) 535 ecb_function_ ecb_const int ecb_ld64 (uint64_t x)
345 { 536 {
537#if 1400 <= _MSC_VER && (_M_X64 || _M_IA64 || _M_ARM)
538 unsigned long r;
539 _BitScanReverse64 (&r, x);
540 return (int)r;
541#else
346 int r = 0; 542 int r = 0;
347 543
348 if (x >> 32) { x >>= 32; r += 32; } 544 if (x >> 32) { x >>= 32; r += 32; }
349 545
350 return r + ecb_ld32 (x); 546 return r + ecb_ld32 (x);
351 }
352#endif 547#endif
548 }
549#endif
353 550
354ecb_function_ ecb_bool ecb_is_pot32 (uint32_t x) ecb_const; 551ecb_function_ ecb_const ecb_bool ecb_is_pot32 (uint32_t x);
355ecb_function_ ecb_bool ecb_is_pot32 (uint32_t x) { return !(x & (x - 1)); } 552ecb_function_ ecb_const ecb_bool ecb_is_pot32 (uint32_t x) { return !(x & (x - 1)); }
356ecb_function_ ecb_bool ecb_is_pot64 (uint64_t x) ecb_const; 553ecb_function_ ecb_const ecb_bool ecb_is_pot64 (uint64_t x);
357ecb_function_ ecb_bool ecb_is_pot64 (uint64_t x) { return !(x & (x - 1)); } 554ecb_function_ ecb_const ecb_bool ecb_is_pot64 (uint64_t x) { return !(x & (x - 1)); }
358 555
359ecb_function_ uint8_t ecb_bitrev8 (uint8_t x) ecb_const; 556ecb_function_ ecb_const uint8_t ecb_bitrev8 (uint8_t x);
360ecb_function_ uint8_t ecb_bitrev8 (uint8_t x) 557ecb_function_ ecb_const uint8_t ecb_bitrev8 (uint8_t x)
361{ 558{
362 return ( (x * 0x0802U & 0x22110U) 559 return ( (x * 0x0802U & 0x22110U)
363 | (x * 0x8020U & 0x88440U)) * 0x10101U >> 16; 560 | (x * 0x8020U & 0x88440U)) * 0x10101U >> 16;
364} 561}
365 562
366ecb_function_ uint16_t ecb_bitrev16 (uint16_t x) ecb_const; 563ecb_function_ ecb_const uint16_t ecb_bitrev16 (uint16_t x);
367ecb_function_ uint16_t ecb_bitrev16 (uint16_t x) 564ecb_function_ ecb_const uint16_t ecb_bitrev16 (uint16_t x)
368{ 565{
369 x = ((x >> 1) & 0x5555) | ((x & 0x5555) << 1); 566 x = ((x >> 1) & 0x5555) | ((x & 0x5555) << 1);
370 x = ((x >> 2) & 0x3333) | ((x & 0x3333) << 2); 567 x = ((x >> 2) & 0x3333) | ((x & 0x3333) << 2);
371 x = ((x >> 4) & 0x0f0f) | ((x & 0x0f0f) << 4); 568 x = ((x >> 4) & 0x0f0f) | ((x & 0x0f0f) << 4);
372 x = ( x >> 8 ) | ( x << 8); 569 x = ( x >> 8 ) | ( x << 8);
373 570
374 return x; 571 return x;
375} 572}
376 573
377ecb_function_ uint32_t ecb_bitrev32 (uint32_t x) ecb_const; 574ecb_function_ ecb_const uint32_t ecb_bitrev32 (uint32_t x);
378ecb_function_ uint32_t ecb_bitrev32 (uint32_t x) 575ecb_function_ ecb_const uint32_t ecb_bitrev32 (uint32_t x)
379{ 576{
380 x = ((x >> 1) & 0x55555555) | ((x & 0x55555555) << 1); 577 x = ((x >> 1) & 0x55555555) | ((x & 0x55555555) << 1);
381 x = ((x >> 2) & 0x33333333) | ((x & 0x33333333) << 2); 578 x = ((x >> 2) & 0x33333333) | ((x & 0x33333333) << 2);
382 x = ((x >> 4) & 0x0f0f0f0f) | ((x & 0x0f0f0f0f) << 4); 579 x = ((x >> 4) & 0x0f0f0f0f) | ((x & 0x0f0f0f0f) << 4);
383 x = ((x >> 8) & 0x00ff00ff) | ((x & 0x00ff00ff) << 8); 580 x = ((x >> 8) & 0x00ff00ff) | ((x & 0x00ff00ff) << 8);
386 return x; 583 return x;
387} 584}
388 585
389/* popcount64 is only available on 64 bit cpus as gcc builtin */ 586/* popcount64 is only available on 64 bit cpus as gcc builtin */
390/* so for this version we are lazy */ 587/* so for this version we are lazy */
391ecb_function_ int ecb_popcount64 (uint64_t x) ecb_const; 588ecb_function_ ecb_const int ecb_popcount64 (uint64_t x);
392ecb_function_ int 589ecb_function_ ecb_const int
393ecb_popcount64 (uint64_t x) 590ecb_popcount64 (uint64_t x)
394{ 591{
395 return ecb_popcount32 (x) + ecb_popcount32 (x >> 32); 592 return ecb_popcount32 (x) + ecb_popcount32 (x >> 32);
396} 593}
397 594
398ecb_inline uint8_t ecb_rotl8 (uint8_t x, unsigned int count) ecb_const; 595ecb_inline ecb_const uint8_t ecb_rotl8 (uint8_t x, unsigned int count);
399ecb_inline uint8_t ecb_rotr8 (uint8_t x, unsigned int count) ecb_const; 596ecb_inline ecb_const uint8_t ecb_rotr8 (uint8_t x, unsigned int count);
400ecb_inline uint16_t ecb_rotl16 (uint16_t x, unsigned int count) ecb_const; 597ecb_inline ecb_const uint16_t ecb_rotl16 (uint16_t x, unsigned int count);
401ecb_inline uint16_t ecb_rotr16 (uint16_t x, unsigned int count) ecb_const; 598ecb_inline ecb_const uint16_t ecb_rotr16 (uint16_t x, unsigned int count);
402ecb_inline uint32_t ecb_rotl32 (uint32_t x, unsigned int count) ecb_const; 599ecb_inline ecb_const uint32_t ecb_rotl32 (uint32_t x, unsigned int count);
403ecb_inline uint32_t ecb_rotr32 (uint32_t x, unsigned int count) ecb_const; 600ecb_inline ecb_const uint32_t ecb_rotr32 (uint32_t x, unsigned int count);
404ecb_inline uint64_t ecb_rotl64 (uint64_t x, unsigned int count) ecb_const; 601ecb_inline ecb_const uint64_t ecb_rotl64 (uint64_t x, unsigned int count);
405ecb_inline uint64_t ecb_rotr64 (uint64_t x, unsigned int count) ecb_const; 602ecb_inline ecb_const uint64_t ecb_rotr64 (uint64_t x, unsigned int count);
406 603
407ecb_inline uint8_t ecb_rotl8 (uint8_t x, unsigned int count) { return (x >> ( 8 - count)) | (x << count); } 604ecb_inline ecb_const uint8_t ecb_rotl8 (uint8_t x, unsigned int count) { return (x >> ( 8 - count)) | (x << count); }
408ecb_inline uint8_t ecb_rotr8 (uint8_t x, unsigned int count) { return (x << ( 8 - count)) | (x >> count); } 605ecb_inline ecb_const uint8_t ecb_rotr8 (uint8_t x, unsigned int count) { return (x << ( 8 - count)) | (x >> count); }
409ecb_inline uint16_t ecb_rotl16 (uint16_t x, unsigned int count) { return (x >> (16 - count)) | (x << count); } 606ecb_inline ecb_const uint16_t ecb_rotl16 (uint16_t x, unsigned int count) { return (x >> (16 - count)) | (x << count); }
410ecb_inline uint16_t ecb_rotr16 (uint16_t x, unsigned int count) { return (x << (16 - count)) | (x >> count); } 607ecb_inline ecb_const uint16_t ecb_rotr16 (uint16_t x, unsigned int count) { return (x << (16 - count)) | (x >> count); }
411ecb_inline uint32_t ecb_rotl32 (uint32_t x, unsigned int count) { return (x >> (32 - count)) | (x << count); } 608ecb_inline ecb_const uint32_t ecb_rotl32 (uint32_t x, unsigned int count) { return (x >> (32 - count)) | (x << count); }
412ecb_inline uint32_t ecb_rotr32 (uint32_t x, unsigned int count) { return (x << (32 - count)) | (x >> count); } 609ecb_inline ecb_const uint32_t ecb_rotr32 (uint32_t x, unsigned int count) { return (x << (32 - count)) | (x >> count); }
413ecb_inline uint64_t ecb_rotl64 (uint64_t x, unsigned int count) { return (x >> (64 - count)) | (x << count); } 610ecb_inline ecb_const uint64_t ecb_rotl64 (uint64_t x, unsigned int count) { return (x >> (64 - count)) | (x << count); }
414ecb_inline uint64_t ecb_rotr64 (uint64_t x, unsigned int count) { return (x << (64 - count)) | (x >> count); } 611ecb_inline ecb_const uint64_t ecb_rotr64 (uint64_t x, unsigned int count) { return (x << (64 - count)) | (x >> count); }
415 612
416#if ECB_GCC_VERSION(4,3) 613#if ECB_GCC_VERSION(4,3) || (ECB_CLANG_BUILTIN(__builtin_bswap32) && ECB_CLANG_BUILTIN(__builtin_bswap64))
614 #if ECB_GCC_VERSION(4,8) || ECB_CLANG_BUILTIN(__builtin_bswap16)
615 #define ecb_bswap16(x) __builtin_bswap16 (x)
616 #else
417 #define ecb_bswap16(x) (__builtin_bswap32 (x) >> 16) 617 #define ecb_bswap16(x) (__builtin_bswap32 (x) >> 16)
618 #endif
418 #define ecb_bswap32(x) __builtin_bswap32 (x) 619 #define ecb_bswap32(x) __builtin_bswap32 (x)
419 #define ecb_bswap64(x) __builtin_bswap64 (x) 620 #define ecb_bswap64(x) __builtin_bswap64 (x)
621#elif _MSC_VER
622 #include <stdlib.h>
623 #define ecb_bswap16(x) ((uint16_t)_byteswap_ushort ((uint16_t)(x)))
624 #define ecb_bswap32(x) ((uint32_t)_byteswap_ulong ((uint32_t)(x)))
625 #define ecb_bswap64(x) ((uint64_t)_byteswap_uint64 ((uint64_t)(x)))
420#else 626#else
421 ecb_function_ uint16_t ecb_bswap16 (uint16_t x) ecb_const; 627 ecb_function_ ecb_const uint16_t ecb_bswap16 (uint16_t x);
422 ecb_function_ uint16_t 628 ecb_function_ ecb_const uint16_t
423 ecb_bswap16 (uint16_t x) 629 ecb_bswap16 (uint16_t x)
424 { 630 {
425 return ecb_rotl16 (x, 8); 631 return ecb_rotl16 (x, 8);
426 } 632 }
427 633
428 ecb_function_ uint32_t ecb_bswap32 (uint32_t x) ecb_const; 634 ecb_function_ ecb_const uint32_t ecb_bswap32 (uint32_t x);
429 ecb_function_ uint32_t 635 ecb_function_ ecb_const uint32_t
430 ecb_bswap32 (uint32_t x) 636 ecb_bswap32 (uint32_t x)
431 { 637 {
432 return (((uint32_t)ecb_bswap16 (x)) << 16) | ecb_bswap16 (x >> 16); 638 return (((uint32_t)ecb_bswap16 (x)) << 16) | ecb_bswap16 (x >> 16);
433 } 639 }
434 640
435 ecb_function_ uint64_t ecb_bswap64 (uint64_t x) ecb_const; 641 ecb_function_ ecb_const uint64_t ecb_bswap64 (uint64_t x);
436 ecb_function_ uint64_t 642 ecb_function_ ecb_const uint64_t
437 ecb_bswap64 (uint64_t x) 643 ecb_bswap64 (uint64_t x)
438 { 644 {
439 return (((uint64_t)ecb_bswap32 (x)) << 32) | ecb_bswap32 (x >> 32); 645 return (((uint64_t)ecb_bswap32 (x)) << 32) | ecb_bswap32 (x >> 32);
440 } 646 }
441#endif 647#endif
442 648
443#if ECB_GCC_VERSION(4,5) 649#if ECB_GCC_VERSION(4,5) || ECB_CLANG_BUILTIN(__builtin_unreachable)
444 #define ecb_unreachable() __builtin_unreachable () 650 #define ecb_unreachable() __builtin_unreachable ()
445#else 651#else
446 /* this seems to work fine, but gcc always emits a warning for it :/ */ 652 /* this seems to work fine, but gcc always emits a warning for it :/ */
447 ecb_inline void ecb_unreachable (void) ecb_noreturn; 653 ecb_inline ecb_noreturn void ecb_unreachable (void);
448 ecb_inline void ecb_unreachable (void) { } 654 ecb_inline ecb_noreturn void ecb_unreachable (void) { }
449#endif 655#endif
450 656
451/* try to tell the compiler that some condition is definitely true */ 657/* try to tell the compiler that some condition is definitely true */
452#define ecb_assume(cond) do { if (!(cond)) ecb_unreachable (); } while (0) 658#define ecb_assume(cond) if (!(cond)) ecb_unreachable (); else 0
453 659
454ecb_inline unsigned char ecb_byteorder_helper (void) ecb_const; 660ecb_inline ecb_const uint32_t ecb_byteorder_helper (void);
455ecb_inline unsigned char 661ecb_inline ecb_const uint32_t
456ecb_byteorder_helper (void) 662ecb_byteorder_helper (void)
457{ 663{
458 const uint32_t u = 0x11223344; 664 /* the union code still generates code under pressure in gcc, */
459 return *(unsigned char *)&u; 665 /* but less than using pointers, and always seems to */
666 /* successfully return a constant. */
667 /* the reason why we have this horrible preprocessor mess */
668 /* is to avoid it in all cases, at least on common architectures */
669 /* or when using a recent enough gcc version (>= 4.6) */
670#if (defined __BYTE_ORDER__ && __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__) \
671 || ((__i386 || __i386__ || _M_IX86 || ECB_GCC_AMD64 || ECB_MSVC_AMD64) && !__VOS__)
672 #define ECB_LITTLE_ENDIAN 1
673 return 0x44332211;
674#elif (defined __BYTE_ORDER__ && __BYTE_ORDER__ == __ORDER_BIG_ENDIAN__) \
675 || ((__AARCH64EB__ || __MIPSEB__ || __ARMEB__) && !__VOS__)
676 #define ECB_BIG_ENDIAN 1
677 return 0x11223344;
678#else
679 union
680 {
681 uint8_t c[4];
682 uint32_t u;
683 } u = { 0x11, 0x22, 0x33, 0x44 };
684 return u.u;
685#endif
460} 686}
461 687
462ecb_inline ecb_bool ecb_big_endian (void) ecb_const; 688ecb_inline ecb_const ecb_bool ecb_big_endian (void);
463ecb_inline ecb_bool ecb_big_endian (void) { return ecb_byteorder_helper () == 0x11; } 689ecb_inline ecb_const ecb_bool ecb_big_endian (void) { return ecb_byteorder_helper () == 0x11223344; }
464ecb_inline ecb_bool ecb_little_endian (void) ecb_const; 690ecb_inline ecb_const ecb_bool ecb_little_endian (void);
465ecb_inline ecb_bool ecb_little_endian (void) { return ecb_byteorder_helper () == 0x44; } 691ecb_inline ecb_const ecb_bool ecb_little_endian (void) { return ecb_byteorder_helper () == 0x44332211; }
692
693/*****************************************************************************/
694/* unaligned load/store */
695
696ecb_inline uint_fast16_t ecb_be_u16_to_host (uint_fast16_t v) { return ecb_little_endian () ? ecb_bswap16 (v) : v; }
697ecb_inline uint_fast32_t ecb_be_u32_to_host (uint_fast32_t v) { return ecb_little_endian () ? ecb_bswap32 (v) : v; }
698ecb_inline uint_fast64_t ecb_be_u64_to_host (uint_fast64_t v) { return ecb_little_endian () ? ecb_bswap64 (v) : v; }
699
700ecb_inline uint_fast16_t ecb_le_u16_to_host (uint_fast16_t v) { return ecb_big_endian () ? ecb_bswap16 (v) : v; }
701ecb_inline uint_fast32_t ecb_le_u32_to_host (uint_fast32_t v) { return ecb_big_endian () ? ecb_bswap32 (v) : v; }
702ecb_inline uint_fast64_t ecb_le_u64_to_host (uint_fast64_t v) { return ecb_big_endian () ? ecb_bswap64 (v) : v; }
703
704ecb_inline uint_fast16_t ecb_peek_u16_u (const void *ptr) { uint16_t v; memcpy (&v, ptr, sizeof (v)); return v; }
705ecb_inline uint_fast32_t ecb_peek_u32_u (const void *ptr) { uint32_t v; memcpy (&v, ptr, sizeof (v)); return v; }
706ecb_inline uint_fast64_t ecb_peek_u64_u (const void *ptr) { uint64_t v; memcpy (&v, ptr, sizeof (v)); return v; }
707
708ecb_inline uint_fast16_t ecb_peek_be_u16_u (const void *ptr) { return ecb_be_u16_to_host (ecb_peek_u16_u (ptr)); }
709ecb_inline uint_fast32_t ecb_peek_be_u32_u (const void *ptr) { return ecb_be_u32_to_host (ecb_peek_u32_u (ptr)); }
710ecb_inline uint_fast64_t ecb_peek_be_u64_u (const void *ptr) { return ecb_be_u64_to_host (ecb_peek_u64_u (ptr)); }
711
712ecb_inline uint_fast16_t ecb_peek_le_u16_u (const void *ptr) { return ecb_le_u16_to_host (ecb_peek_u16_u (ptr)); }
713ecb_inline uint_fast32_t ecb_peek_le_u32_u (const void *ptr) { return ecb_le_u32_to_host (ecb_peek_u32_u (ptr)); }
714ecb_inline uint_fast64_t ecb_peek_le_u64_u (const void *ptr) { return ecb_le_u64_to_host (ecb_peek_u64_u (ptr)); }
715
716ecb_inline uint_fast16_t ecb_host_to_be_u16 (uint_fast16_t v) { return ecb_little_endian () ? ecb_bswap16 (v) : v; }
717ecb_inline uint_fast32_t ecb_host_to_be_u32 (uint_fast32_t v) { return ecb_little_endian () ? ecb_bswap32 (v) : v; }
718ecb_inline uint_fast64_t ecb_host_to_be_u64 (uint_fast64_t v) { return ecb_little_endian () ? ecb_bswap64 (v) : v; }
719
720ecb_inline uint_fast16_t ecb_host_to_le_u16 (uint_fast16_t v) { return ecb_big_endian () ? ecb_bswap16 (v) : v; }
721ecb_inline uint_fast32_t ecb_host_to_le_u32 (uint_fast32_t v) { return ecb_big_endian () ? ecb_bswap32 (v) : v; }
722ecb_inline uint_fast64_t ecb_host_to_le_u64 (uint_fast64_t v) { return ecb_big_endian () ? ecb_bswap64 (v) : v; }
723
724ecb_inline void ecb_poke_u16_u (void *ptr, uint16_t v) { memcpy (ptr, &v, sizeof (v)); }
725ecb_inline void ecb_poke_u32_u (void *ptr, uint32_t v) { memcpy (ptr, &v, sizeof (v)); }
726ecb_inline void ecb_poke_u64_u (void *ptr, uint64_t v) { memcpy (ptr, &v, sizeof (v)); }
727
728ecb_inline void ecb_poke_be_u16_u (void *ptr, uint_fast16_t v) { ecb_poke_u16_u (ptr, ecb_host_to_be_u16 (v)); }
729ecb_inline void ecb_poke_be_u32_u (void *ptr, uint_fast32_t v) { ecb_poke_u32_u (ptr, ecb_host_to_be_u32 (v)); }
730ecb_inline void ecb_poke_be_u64_u (void *ptr, uint_fast64_t v) { ecb_poke_u64_u (ptr, ecb_host_to_be_u64 (v)); }
731
732ecb_inline void ecb_poke_le_u16_u (void *ptr, uint_fast16_t v) { ecb_poke_u16_u (ptr, ecb_host_to_le_u16 (v)); }
733ecb_inline void ecb_poke_le_u32_u (void *ptr, uint_fast32_t v) { ecb_poke_u32_u (ptr, ecb_host_to_le_u32 (v)); }
734ecb_inline void ecb_poke_le_u64_u (void *ptr, uint_fast64_t v) { ecb_poke_u64_u (ptr, ecb_host_to_le_u64 (v)); }
735
736#ifdef ECB_CPP
737
738inline uint8_t ecb_bswap (uint8_t v) { return v; }
739inline uint16_t ecb_bswap (uint16_t v) { return ecb_bswap16 (v); }
740inline uint32_t ecb_bswap (uint32_t v) { return ecb_bswap32 (v); }
741inline uint64_t ecb_bswap (uint64_t v) { return ecb_bswap64 (v); }
742
743template<typename T> inline T ecb_be_to_host (T v) { return ecb_little_endian () ? ecb_bswap (v) : v; }
744template<typename T> inline T ecb_le_to_host (T v) { return ecb_big_endian () ? ecb_bswap (v) : v; }
745template<typename T> inline T ecb_peek (const void *ptr) { return *(const T *)ptr; }
746template<typename T> inline T ecb_peek_be (const void *ptr) { return ecb_be_to_host (ecb_peek <T> (ptr)); }
747template<typename T> inline T ecb_peek_le (const void *ptr) { return ecb_le_to_host (ecb_peek <T> (ptr)); }
748template<typename T> inline T ecb_peek_u (const void *ptr) { T v; std::memcpy (&v, ptr, sizeof (v)); return v; }
749template<typename T> inline T ecb_peek_be_u (const void *ptr) { return ecb_be_to_host (ecb_peek_u<T> (ptr)); }
750template<typename T> inline T ecb_peek_le_u (const void *ptr) { return ecb_le_to_host (ecb_peek_u<T> (ptr)); }
751
752template<typename T> inline T ecb_host_to_be (T v) { return ecb_little_endian () ? ecb_bswap (v) : v; }
753template<typename T> inline T ecb_host_to_le (T v) { return ecb_big_endian () ? ecb_bswap (v) : v; }
754template<typename T> inline void ecb_poke (void *ptr, T v) { *(T *)ptr = v; }
755template<typename T> inline void ecb_poke_be (void *ptr, T v) { return ecb_poke <T> (ptr, ecb_host_to_be (v)); }
756template<typename T> inline void ecb_poke_le (void *ptr, T v) { return ecb_poke <T> (ptr, ecb_host_to_le (v)); }
757template<typename T> inline void ecb_poke_u (void *ptr, T v) { std::memcpy (ptr, &v, sizeof (v)); }
758template<typename T> inline void ecb_poke_be_u (void *ptr, T v) { return ecb_poke_u<T> (ptr, ecb_host_to_be (v)); }
759template<typename T> inline void ecb_poke_le_u (void *ptr, T v) { return ecb_poke_u<T> (ptr, ecb_host_to_le (v)); }
760
761#endif
762
763/*****************************************************************************/
466 764
467#if ECB_GCC_VERSION(3,0) || ECB_C99 765#if ECB_GCC_VERSION(3,0) || ECB_C99
468 #define ecb_mod(m,n) ((m) % (n) + ((m) % (n) < 0 ? (n) : 0)) 766 #define ecb_mod(m,n) ((m) % (n) + ((m) % (n) < 0 ? (n) : 0))
469#else 767#else
470 #define ecb_mod(m,n) ((m) < 0 ? ((n) - 1 - ((-1 - (m)) % (n))) : ((m) % (n))) 768 #define ecb_mod(m,n) ((m) < 0 ? ((n) - 1 - ((-1 - (m)) % (n))) : ((m) % (n)))
471#endif 769#endif
472 770
473#if __cplusplus 771#if ECB_CPP
474 template<typename T> 772 template<typename T>
475 static inline T ecb_div_rd (T val, T div) 773 static inline T ecb_div_rd (T val, T div)
476 { 774 {
477 return val < 0 ? - ((-val + div - 1) / div) : (val ) / div; 775 return val < 0 ? - ((-val + div - 1) / div) : (val ) / div;
478 } 776 }
495 } 793 }
496#else 794#else
497 #define ecb_array_length(name) (sizeof (name) / sizeof (name [0])) 795 #define ecb_array_length(name) (sizeof (name) / sizeof (name [0]))
498#endif 796#endif
499 797
798/*****************************************************************************/
799
800ecb_function_ ecb_const uint32_t ecb_binary16_to_binary32 (uint32_t x);
801ecb_function_ ecb_const uint32_t
802ecb_binary16_to_binary32 (uint32_t x)
803{
804 unsigned int s = (x & 0x8000) << (31 - 15);
805 int e = (x >> 10) & 0x001f;
806 unsigned int m = x & 0x03ff;
807
808 if (ecb_expect_false (e == 31))
809 /* infinity or NaN */
810 e = 255 - (127 - 15);
811 else if (ecb_expect_false (!e))
812 {
813 if (ecb_expect_true (!m))
814 /* zero, handled by code below by forcing e to 0 */
815 e = 0 - (127 - 15);
816 else
817 {
818 /* subnormal, renormalise */
819 unsigned int s = 10 - ecb_ld32 (m);
820
821 m = (m << s) & 0x3ff; /* mask implicit bit */
822 e -= s - 1;
823 }
824 }
825
826 /* e and m now are normalised, or zero, (or inf or nan) */
827 e += 127 - 15;
828
829 return s | (e << 23) | (m << (23 - 10));
830}
831
832ecb_function_ ecb_const uint16_t ecb_binary32_to_binary16 (uint32_t x);
833ecb_function_ ecb_const uint16_t
834ecb_binary32_to_binary16 (uint32_t x)
835{
836 unsigned int s = (x >> 16) & 0x00008000; /* sign bit, the easy part */
837 unsigned int e = ((x >> 23) & 0x000000ff) - (127 - 15); /* the desired exponent */
838 unsigned int m = x & 0x007fffff;
839
840 x &= 0x7fffffff;
841
842 /* if it's within range of binary16 normals, use fast path */
843 if (ecb_expect_true (0x38800000 <= x && x <= 0x477fefff))
844 {
845 /* mantissa round-to-even */
846 m += 0x00000fff + ((m >> (23 - 10)) & 1);
847
848 /* handle overflow */
849 if (ecb_expect_false (m >= 0x00800000))
850 {
851 m >>= 1;
852 e += 1;
853 }
854
855 return s | (e << 10) | (m >> (23 - 10));
856 }
857
858 /* handle large numbers and infinity */
859 if (ecb_expect_true (0x477fefff < x && x <= 0x7f800000))
860 return s | 0x7c00;
861
862 /* handle zero, subnormals and small numbers */
863 if (ecb_expect_true (x < 0x38800000))
864 {
865 /* zero */
866 if (ecb_expect_true (!x))
867 return s;
868
869 /* handle subnormals */
870
871 /* too small, will be zero */
872 if (e < (14 - 24)) /* might not be sharp, but is good enough */
873 return s;
874
875 m |= 0x00800000; /* make implicit bit explicit */
876
877 /* very tricky - we need to round to the nearest e (+10) bit value */
878 {
879 unsigned int bits = 14 - e;
880 unsigned int half = (1 << (bits - 1)) - 1;
881 unsigned int even = (m >> bits) & 1;
882
883 /* if this overflows, we will end up with a normalised number */
884 m = (m + half + even) >> bits;
885 }
886
887 return s | m;
888 }
889
890 /* handle NaNs, preserve leftmost nan bits, but make sure we don't turn them into infinities */
891 m >>= 13;
892
893 return s | 0x7c00 | m | !m;
894}
895
896/*******************************************************************************/
897/* floating point stuff, can be disabled by defining ECB_NO_LIBM */
898
899/* basically, everything uses "ieee pure-endian" floating point numbers */
900/* the only noteworthy exception is ancient armle, which uses order 43218765 */
901#if 0 \
902 || __i386 || __i386__ \
903 || ECB_GCC_AMD64 \
904 || __powerpc__ || __ppc__ || __powerpc64__ || __ppc64__ \
905 || defined __s390__ || defined __s390x__ \
906 || defined __mips__ \
907 || defined __alpha__ \
908 || defined __hppa__ \
909 || defined __ia64__ \
910 || defined __m68k__ \
911 || defined __m88k__ \
912 || defined __sh__ \
913 || defined _M_IX86 || defined ECB_MSVC_AMD64 || defined _M_IA64 \
914 || (defined __arm__ && (defined __ARM_EABI__ || defined __EABI__ || defined __VFP_FP__ || defined _WIN32_WCE || defined __ANDROID__)) \
915 || defined __aarch64__
916 #define ECB_STDFP 1
917 #include <string.h> /* for memcpy */
918#else
919 #define ECB_STDFP 0
920#endif
921
922#ifndef ECB_NO_LIBM
923
924 #include <math.h> /* for frexp*, ldexp*, INFINITY, NAN */
925
926 /* only the oldest of old doesn't have this one. solaris. */
927 #ifdef INFINITY
928 #define ECB_INFINITY INFINITY
929 #else
930 #define ECB_INFINITY HUGE_VAL
500#endif 931 #endif
501 932
933 #ifdef NAN
934 #define ECB_NAN NAN
935 #else
936 #define ECB_NAN ECB_INFINITY
937 #endif
938
939 #if ECB_C99 || _XOPEN_VERSION >= 600 || _POSIX_VERSION >= 200112L
940 #define ecb_ldexpf(x,e) ldexpf ((x), (e))
941 #define ecb_frexpf(x,e) frexpf ((x), (e))
942 #else
943 #define ecb_ldexpf(x,e) (float) ldexp ((double) (x), (e))
944 #define ecb_frexpf(x,e) (float) frexp ((double) (x), (e))
945 #endif
946
947 /* convert a float to ieee single/binary32 */
948 ecb_function_ ecb_const uint32_t ecb_float_to_binary32 (float x);
949 ecb_function_ ecb_const uint32_t
950 ecb_float_to_binary32 (float x)
951 {
952 uint32_t r;
953
954 #if ECB_STDFP
955 memcpy (&r, &x, 4);
956 #else
957 /* slow emulation, works for anything but -0 */
958 uint32_t m;
959 int e;
960
961 if (x == 0e0f ) return 0x00000000U;
962 if (x > +3.40282346638528860e+38f) return 0x7f800000U;
963 if (x < -3.40282346638528860e+38f) return 0xff800000U;
964 if (x != x ) return 0x7fbfffffU;
965
966 m = ecb_frexpf (x, &e) * 0x1000000U;
967
968 r = m & 0x80000000U;
969
970 if (r)
971 m = -m;
972
973 if (e <= -126)
974 {
975 m &= 0xffffffU;
976 m >>= (-125 - e);
977 e = -126;
978 }
979
980 r |= (e + 126) << 23;
981 r |= m & 0x7fffffU;
982 #endif
983
984 return r;
985 }
986
987 /* converts an ieee single/binary32 to a float */
988 ecb_function_ ecb_const float ecb_binary32_to_float (uint32_t x);
989 ecb_function_ ecb_const float
990 ecb_binary32_to_float (uint32_t x)
991 {
992 float r;
993
994 #if ECB_STDFP
995 memcpy (&r, &x, 4);
996 #else
997 /* emulation, only works for normals and subnormals and +0 */
998 int neg = x >> 31;
999 int e = (x >> 23) & 0xffU;
1000
1001 x &= 0x7fffffU;
1002
1003 if (e)
1004 x |= 0x800000U;
1005 else
1006 e = 1;
1007
1008 /* we distrust ldexpf a bit and do the 2**-24 scaling by an extra multiply */
1009 r = ecb_ldexpf (x * (0.5f / 0x800000U), e - 126);
1010
1011 r = neg ? -r : r;
1012 #endif
1013
1014 return r;
1015 }
1016
1017 /* convert a double to ieee double/binary64 */
1018 ecb_function_ ecb_const uint64_t ecb_double_to_binary64 (double x);
1019 ecb_function_ ecb_const uint64_t
1020 ecb_double_to_binary64 (double x)
1021 {
1022 uint64_t r;
1023
1024 #if ECB_STDFP
1025 memcpy (&r, &x, 8);
1026 #else
1027 /* slow emulation, works for anything but -0 */
1028 uint64_t m;
1029 int e;
1030
1031 if (x == 0e0 ) return 0x0000000000000000U;
1032 if (x > +1.79769313486231470e+308) return 0x7ff0000000000000U;
1033 if (x < -1.79769313486231470e+308) return 0xfff0000000000000U;
1034 if (x != x ) return 0X7ff7ffffffffffffU;
1035
1036 m = frexp (x, &e) * 0x20000000000000U;
1037
1038 r = m & 0x8000000000000000;;
1039
1040 if (r)
1041 m = -m;
1042
1043 if (e <= -1022)
1044 {
1045 m &= 0x1fffffffffffffU;
1046 m >>= (-1021 - e);
1047 e = -1022;
1048 }
1049
1050 r |= ((uint64_t)(e + 1022)) << 52;
1051 r |= m & 0xfffffffffffffU;
1052 #endif
1053
1054 return r;
1055 }
1056
1057 /* converts an ieee double/binary64 to a double */
1058 ecb_function_ ecb_const double ecb_binary64_to_double (uint64_t x);
1059 ecb_function_ ecb_const double
1060 ecb_binary64_to_double (uint64_t x)
1061 {
1062 double r;
1063
1064 #if ECB_STDFP
1065 memcpy (&r, &x, 8);
1066 #else
1067 /* emulation, only works for normals and subnormals and +0 */
1068 int neg = x >> 63;
1069 int e = (x >> 52) & 0x7ffU;
1070
1071 x &= 0xfffffffffffffU;
1072
1073 if (e)
1074 x |= 0x10000000000000U;
1075 else
1076 e = 1;
1077
1078 /* we distrust ldexp a bit and do the 2**-53 scaling by an extra multiply */
1079 r = ldexp (x * (0.5 / 0x10000000000000U), e - 1022);
1080
1081 r = neg ? -r : r;
1082 #endif
1083
1084 return r;
1085 }
1086
1087 /* convert a float to ieee half/binary16 */
1088 ecb_function_ ecb_const uint16_t ecb_float_to_binary16 (float x);
1089 ecb_function_ ecb_const uint16_t
1090 ecb_float_to_binary16 (float x)
1091 {
1092 return ecb_binary32_to_binary16 (ecb_float_to_binary32 (x));
1093 }
1094
1095 /* convert an ieee half/binary16 to float */
1096 ecb_function_ ecb_const float ecb_binary16_to_float (uint16_t x);
1097 ecb_function_ ecb_const float
1098 ecb_binary16_to_float (uint16_t x)
1099 {
1100 return ecb_binary32_to_float (ecb_binary16_to_binary32 (x));
1101 }
1102
1103#endif
1104
1105#endif
1106

Diff Legend

Removed lines
+ Added lines
< Changed lines
> Changed lines