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

Comparing libeio/ecb.h (file contents):
Revision 1.21 by root, Fri May 16 15:15:23 2014 UTC vs.
Revision 1.26 by root, Mon Sep 21 08:01:00 2015 UTC

1/* 1/*
2 * libecb - http://software.schmorp.de/pkg/libecb 2 * libecb - http://software.schmorp.de/pkg/libecb
3 * 3 *
4 * Copyright (©) 2009-2014 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:
… …
40 40
41#ifndef ECB_H 41#ifndef ECB_H
42#define ECB_H 42#define ECB_H
43 43
44/* 16 bits major, 16 bits minor */ 44/* 16 bits major, 16 bits minor */
45#define ECB_VERSION 0x00010003 45#define ECB_VERSION 0x00010004
46 46
47#ifdef _WIN32 47#ifdef _WIN32
48 typedef signed char int8_t; 48 typedef signed char int8_t;
49 typedef unsigned char uint8_t; 49 typedef unsigned char uint8_t;
50 typedef signed short int16_t; 50 typedef signed short int16_t;
… …
74 #else 74 #else
75 #define ECB_PTRSIZE 4 75 #define ECB_PTRSIZE 4
76 #endif 76 #endif
77#endif 77#endif
78 78
79#define ECB_GCC_AMD64 (__amd64 || __amd64__ || __x86_64 || __x86_64__)
80#define ECB_MSVC_AMD64 (_M_AMD64 || _M_X64)
81
79/* work around x32 idiocy by defining proper macros */ 82/* work around x32 idiocy by defining proper macros */
80#if __amd64 || __x86_64 || _M_AMD64 || _M_X64 83#if ECB_GCC_AMD64 || ECB_MSVC_AMD64
81 #if _ILP32 84 #if _ILP32
82 #define ECB_AMD64_X32 1 85 #define ECB_AMD64_X32 1
83 #else 86 #else
84 #define ECB_AMD64 1 87 #define ECB_AMD64 1
85 #endif 88 #endif
… …
90 * causing enormous grief in return for some better fake benchmark numbers. 93 * causing enormous grief in return for some better fake benchmark numbers.
91 * or so. 94 * or so.
92 * we try to detect these and simply assume they are not gcc - if they have 95 * we try to detect these and simply assume they are not gcc - if they have
93 * an issue with that they should have done it right in the first place. 96 * an issue with that they should have done it right in the first place.
94 */ 97 */
95#ifndef ECB_GCC_VERSION
96 #if !defined __GNUC_MINOR__ || defined __INTEL_COMPILER || defined __SUNPRO_C || defined __SUNPRO_CC || defined __llvm__ || defined __clang__ 98#if !defined __GNUC_MINOR__ || defined __INTEL_COMPILER || defined __SUNPRO_C || defined __SUNPRO_CC || defined __llvm__ || defined __clang__
97 #define ECB_GCC_VERSION(major,minor) 0 99 #define ECB_GCC_VERSION(major,minor) 0
98 #else 100#else
99 #define ECB_GCC_VERSION(major,minor) (__GNUC__ > (major) || (__GNUC__ == (major) && __GNUC_MINOR__ >= (minor))) 101 #define ECB_GCC_VERSION(major,minor) (__GNUC__ > (major) || (__GNUC__ == (major) && __GNUC_MINOR__ >= (minor)))
100 #endif 102#endif
103
104#define ECB_CLANG_VERSION(major,minor) (__clang_major__ > (major) || (__clang_major__ == (major) && __clang_minor__ >= (minor)))
105
106#if __clang__ && defined __has_builtin
107 #define ECB_CLANG_BUILTIN(x) __has_builtin (x)
108#else
109 #define ECB_CLANG_BUILTIN(x) 0
110#endif
111
112#if __clang__ && defined __has_extension
113 #define ECB_CLANG_EXTENSION(x) __has_extension (x)
114#else
115 #define ECB_CLANG_EXTENSION(x) 0
101#endif 116#endif
102 117
103#define ECB_CPP (__cplusplus+0) 118#define ECB_CPP (__cplusplus+0)
104#define ECB_CPP11 (__cplusplus >= 201103L) 119#define ECB_CPP11 (__cplusplus >= 201103L)
105 120
… …
133 #define ECB_NO_SMP 1 148 #define ECB_NO_SMP 1
134#endif 149#endif
135 150
136#if ECB_NO_SMP 151#if ECB_NO_SMP
137 #define ECB_MEMORY_FENCE do { } while (0) 152 #define ECB_MEMORY_FENCE do { } while (0)
153#endif
154
155/* http://www-01.ibm.com/support/knowledgecenter/SSGH3R_13.1.0/com.ibm.xlcpp131.aix.doc/compiler_ref/compiler_builtins.html */
156#if __xlC__ && ECB_CPP
157 #include <builtins.h>
138#endif 158#endif
139 159
140#ifndef ECB_MEMORY_FENCE 160#ifndef ECB_MEMORY_FENCE
141 #if ECB_GCC_VERSION(2,5) || defined __INTEL_COMPILER || (__llvm__ && __GNUC__) || __SUNPRO_C >= 0x5110 || __SUNPRO_CC >= 0x5110 161 #if ECB_GCC_VERSION(2,5) || defined __INTEL_COMPILER || (__llvm__ && __GNUC__) || __SUNPRO_C >= 0x5110 || __SUNPRO_CC >= 0x5110
142 #if __i386 || __i386__ 162 #if __i386 || __i386__
143 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("lock; orb $0, -1(%%esp)" : : : "memory") 163 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("lock; orb $0, -1(%%esp)" : : : "memory")
144 #define ECB_MEMORY_FENCE_ACQUIRE __asm__ __volatile__ ("" : : : "memory") 164 #define ECB_MEMORY_FENCE_ACQUIRE __asm__ __volatile__ ("" : : : "memory")
145 #define ECB_MEMORY_FENCE_RELEASE __asm__ __volatile__ ("") 165 #define ECB_MEMORY_FENCE_RELEASE __asm__ __volatile__ ("")
146 #elif __amd64 || __amd64__ || __x86_64 || __x86_64__ 166 #elif ECB_GCC_AMD64
147 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("mfence" : : : "memory") 167 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("mfence" : : : "memory")
148 #define ECB_MEMORY_FENCE_ACQUIRE __asm__ __volatile__ ("" : : : "memory") 168 #define ECB_MEMORY_FENCE_ACQUIRE __asm__ __volatile__ ("" : : : "memory")
149 #define ECB_MEMORY_FENCE_RELEASE __asm__ __volatile__ ("") 169 #define ECB_MEMORY_FENCE_RELEASE __asm__ __volatile__ ("")
150 #elif __powerpc__ || __ppc__ || __powerpc64__ || __ppc64__ 170 #elif __powerpc__ || __ppc__ || __powerpc64__ || __ppc64__
151 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("sync" : : : "memory") 171 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("sync" : : : "memory")
… …
155 #elif defined __ARM_ARCH_7__ || defined __ARM_ARCH_7A__ \ 175 #elif defined __ARM_ARCH_7__ || defined __ARM_ARCH_7A__ \
156 || defined __ARM_ARCH_7M__ || defined __ARM_ARCH_7R__ 176 || defined __ARM_ARCH_7M__ || defined __ARM_ARCH_7R__
157 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("dmb" : : : "memory") 177 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("dmb" : : : "memory")
158 #elif __aarch64__ 178 #elif __aarch64__
159 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("dmb ish" : : : "memory") 179 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("dmb ish" : : : "memory")
160 #elif (__sparc || __sparc__) && !__sparcv8 180 #elif (__sparc || __sparc__) && !(__sparc_v8__ || defined __sparcv8)
161 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("membar #LoadStore | #LoadLoad | #StoreStore | #StoreLoad" : : : "memory") 181 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("membar #LoadStore | #LoadLoad | #StoreStore | #StoreLoad" : : : "memory")
162 #define ECB_MEMORY_FENCE_ACQUIRE __asm__ __volatile__ ("membar #LoadStore | #LoadLoad" : : : "memory") 182 #define ECB_MEMORY_FENCE_ACQUIRE __asm__ __volatile__ ("membar #LoadStore | #LoadLoad" : : : "memory")
163 #define ECB_MEMORY_FENCE_RELEASE __asm__ __volatile__ ("membar #LoadStore | #StoreStore") 183 #define ECB_MEMORY_FENCE_RELEASE __asm__ __volatile__ ("membar #LoadStore | #StoreStore")
164 #elif defined __s390__ || defined __s390x__ 184 #elif defined __s390__ || defined __s390x__
165 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("bcr 15,0" : : : "memory") 185 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("bcr 15,0" : : : "memory")
… …
189 /* see comment below (stdatomic.h) about the C11 memory model. */ 209 /* see comment below (stdatomic.h) about the C11 memory model. */
190 #define ECB_MEMORY_FENCE __atomic_thread_fence (__ATOMIC_SEQ_CST) 210 #define ECB_MEMORY_FENCE __atomic_thread_fence (__ATOMIC_SEQ_CST)
191 #define ECB_MEMORY_FENCE_ACQUIRE __atomic_thread_fence (__ATOMIC_ACQUIRE) 211 #define ECB_MEMORY_FENCE_ACQUIRE __atomic_thread_fence (__ATOMIC_ACQUIRE)
192 #define ECB_MEMORY_FENCE_RELEASE __atomic_thread_fence (__ATOMIC_RELEASE) 212 #define ECB_MEMORY_FENCE_RELEASE __atomic_thread_fence (__ATOMIC_RELEASE)
193 213
194 /* The __has_feature syntax from clang is so misdesigned that we cannot use it 214 #elif ECB_CLANG_EXTENSION(c_atomic)
195 * without risking compile time errors with other compilers. We *could*
196 * define our own ecb_clang_has_feature, but I just can't be bothered to work
197 * around this shit time and again.
198 * #elif defined __clang && __has_feature (cxx_atomic)
199 * // see comment below (stdatomic.h) about the C11 memory model. 215 /* see comment below (stdatomic.h) about the C11 memory model. */
200 * #define ECB_MEMORY_FENCE __c11_atomic_thread_fence (__ATOMIC_SEQ_CST) 216 #define ECB_MEMORY_FENCE __c11_atomic_thread_fence (__ATOMIC_SEQ_CST)
201 * #define ECB_MEMORY_FENCE_ACQUIRE __c11_atomic_thread_fence (__ATOMIC_ACQUIRE) 217 #define ECB_MEMORY_FENCE_ACQUIRE __c11_atomic_thread_fence (__ATOMIC_ACQUIRE)
202 * #define ECB_MEMORY_FENCE_RELEASE __c11_atomic_thread_fence (__ATOMIC_RELEASE) 218 #define ECB_MEMORY_FENCE_RELEASE __c11_atomic_thread_fence (__ATOMIC_RELEASE)
203 */
204 219
205 #elif ECB_GCC_VERSION(4,4) || defined __INTEL_COMPILER || defined __clang__ 220 #elif ECB_GCC_VERSION(4,4) || defined __INTEL_COMPILER || defined __clang__
206 #define ECB_MEMORY_FENCE __sync_synchronize () 221 #define ECB_MEMORY_FENCE __sync_synchronize ()
207 #elif _MSC_VER >= 1500 /* VC++ 2008 */ 222 #elif _MSC_VER >= 1500 /* VC++ 2008 */
208 /* apparently, microsoft broke all the memory barrier stuff in Visual Studio 2008... */ 223 /* apparently, microsoft broke all the memory barrier stuff in Visual Studio 2008... */
… …
271 #define ECB_MEMORY_FENCE_RELEASE ECB_MEMORY_FENCE 286 #define ECB_MEMORY_FENCE_RELEASE ECB_MEMORY_FENCE
272#endif 287#endif
273 288
274/*****************************************************************************/ 289/*****************************************************************************/
275 290
276#if __cplusplus 291#if ECB_CPP
277 #define ecb_inline static inline 292 #define ecb_inline static inline
278#elif ECB_GCC_VERSION(2,5) 293#elif ECB_GCC_VERSION(2,5)
279 #define ecb_inline static __inline__ 294 #define ecb_inline static __inline__
280#elif ECB_C99 295#elif ECB_C99
281 #define ecb_inline static inline 296 #define ecb_inline static inline
… …
295 310
296#define ECB_CONCAT_(a, b) a ## b 311#define ECB_CONCAT_(a, b) a ## b
297#define ECB_CONCAT(a, b) ECB_CONCAT_(a, b) 312#define ECB_CONCAT(a, b) ECB_CONCAT_(a, b)
298#define ECB_STRINGIFY_(a) # a 313#define ECB_STRINGIFY_(a) # a
299#define ECB_STRINGIFY(a) ECB_STRINGIFY_(a) 314#define ECB_STRINGIFY(a) ECB_STRINGIFY_(a)
315#define ECB_STRINGIFY_EXPR(expr) ((expr), ECB_STRINGIFY_ (expr))
300 316
301#define ecb_function_ ecb_inline 317#define ecb_function_ ecb_inline
302 318
303#if ECB_GCC_VERSION(3,1) 319#if ECB_GCC_VERSION(3,1) || ECB_CLANG_VERSION(2,8)
304 #define ecb_attribute(attrlist) __attribute__(attrlist) 320 #define ecb_attribute(attrlist) __attribute__ (attrlist)
321#else
322 #define ecb_attribute(attrlist)
323#endif
324
325#if ECB_GCC_VERSION(3,1) || ECB_CLANG_BUILTIN(__builtin_constant_p)
305 #define ecb_is_constant(expr) __builtin_constant_p (expr) 326 #define ecb_is_constant(expr) __builtin_constant_p (expr)
306 #define ecb_expect(expr,value) __builtin_expect ((expr),(value))
307 #define ecb_prefetch(addr,rw,locality) __builtin_prefetch (addr, rw, locality)
308#else 327#else
309 #define ecb_attribute(attrlist)
310
311 /* possible C11 impl for integral types 328 /* possible C11 impl for integral types
312 typedef struct ecb_is_constant_struct ecb_is_constant_struct; 329 typedef struct ecb_is_constant_struct ecb_is_constant_struct;
313 #define ecb_is_constant(expr) _Generic ((1 ? (struct ecb_is_constant_struct *)0 : (void *)((expr) - (expr)), ecb_is_constant_struct *: 0, default: 1)) */ 330 #define ecb_is_constant(expr) _Generic ((1 ? (struct ecb_is_constant_struct *)0 : (void *)((expr) - (expr)), ecb_is_constant_struct *: 0, default: 1)) */
314 331
315 #define ecb_is_constant(expr) 0 332 #define ecb_is_constant(expr) 0
333#endif
334
335#if ECB_GCC_VERSION(3,1) || ECB_CLANG_BUILTIN(__builtin_expect)
336 #define ecb_expect(expr,value) __builtin_expect ((expr),(value))
337#else
316 #define ecb_expect(expr,value) (expr) 338 #define ecb_expect(expr,value) (expr)
339#endif
340
341#if ECB_GCC_VERSION(3,1) || ECB_CLANG_BUILTIN(__builtin_prefetch)
342 #define ecb_prefetch(addr,rw,locality) __builtin_prefetch (addr, rw, locality)
343#else
317 #define ecb_prefetch(addr,rw,locality) 344 #define ecb_prefetch(addr,rw,locality)
318#endif 345#endif
319 346
320/* no emulation for ecb_decltype */ 347/* no emulation for ecb_decltype */
321#if ECB_GCC_VERSION(4,5) 348#if ECB_CPP11
349 // older implementations might have problems with decltype(x)::type, work around it
350 template<class T> struct ecb_decltype_t { typedef T type; };
322 #define ecb_decltype(x) __decltype(x) 351 #define ecb_decltype(x) ecb_decltype_t<decltype (x)>::type
323#elif ECB_GCC_VERSION(3,0) 352#elif ECB_GCC_VERSION(3,0) || ECB_CLANG_VERSION(2,8)
324 #define ecb_decltype(x) __typeof(x) 353 #define ecb_decltype(x) __typeof__ (x)
325#endif 354#endif
326 355
356#if _MSC_VER >= 1300
357 #define ecb_deprecated __declspec (deprecated)
358#else
359 #define ecb_deprecated ecb_attribute ((__deprecated__))
360#endif
361
362#if _MSC_VER >= 1500
363 #define ecb_deprecated_message(msg) __declspec (deprecated (msg))
364#elif ECB_GCC_VERSION(4,5)
365 #define ecb_deprecated_message(msg) ecb_attribute ((__deprecated__ (msg))
366#else
367 #define ecb_deprecated_message(msg) ecb_deprecated
368#endif
369
370#if _MSC_VER >= 1400
371 #define ecb_noinline __declspec (noinline)
372#else
327#define ecb_noinline ecb_attribute ((__noinline__)) 373 #define ecb_noinline ecb_attribute ((__noinline__))
374#endif
375
328#define ecb_unused ecb_attribute ((__unused__)) 376#define ecb_unused ecb_attribute ((__unused__))
329#define ecb_const ecb_attribute ((__const__)) 377#define ecb_const ecb_attribute ((__const__))
330#define ecb_pure ecb_attribute ((__pure__)) 378#define ecb_pure ecb_attribute ((__pure__))
331 379
332#if ECB_C11 380#if ECB_C11 || __IBMC_NORETURN
381 /* http://www-01.ibm.com/support/knowledgecenter/SSGH3R_13.1.0/com.ibm.xlcpp131.aix.doc/language_ref/noreturn.html */
333 #define ecb_noreturn _Noreturn 382 #define ecb_noreturn _Noreturn
383#elif ECB_CPP11
384 #define ecb_noreturn [[noreturn]]
385#elif _MSC_VER >= 1200
386 /* http://msdn.microsoft.com/en-us/library/k6ktzx3s.aspx */
387 #define ecb_noreturn __declspec (noreturn)
334#else 388#else
335 #define ecb_noreturn ecb_attribute ((__noreturn__)) 389 #define ecb_noreturn ecb_attribute ((__noreturn__))
336#endif 390#endif
337 391
338#if ECB_GCC_VERSION(4,3) 392#if ECB_GCC_VERSION(4,3)
… …
353/* for compatibility to the rest of the world */ 407/* for compatibility to the rest of the world */
354#define ecb_likely(expr) ecb_expect_true (expr) 408#define ecb_likely(expr) ecb_expect_true (expr)
355#define ecb_unlikely(expr) ecb_expect_false (expr) 409#define ecb_unlikely(expr) ecb_expect_false (expr)
356 410
357/* count trailing zero bits and count # of one bits */ 411/* count trailing zero bits and count # of one bits */
358#if ECB_GCC_VERSION(3,4) 412#if ECB_GCC_VERSION(3,4) \
413 || (ECB_CLANG_BUILTIN(__builtin_clz) && ECB_CLANG_BUILTIN(__builtin_clzll) \
414 && ECB_CLANG_BUILTIN(__builtin_ctz) && ECB_CLANG_BUILTIN(__builtin_ctzll) \
415 && ECB_CLANG_BUILTIN(__builtin_popcount))
359 /* we assume int == 32 bit, long == 32 or 64 bit and long long == 64 bit */ 416 /* we assume int == 32 bit, long == 32 or 64 bit and long long == 64 bit */
360 #define ecb_ld32(x) (__builtin_clz (x) ^ 31) 417 #define ecb_ld32(x) (__builtin_clz (x) ^ 31)
361 #define ecb_ld64(x) (__builtin_clzll (x) ^ 63) 418 #define ecb_ld64(x) (__builtin_clzll (x) ^ 63)
362 #define ecb_ctz32(x) __builtin_ctz (x) 419 #define ecb_ctz32(x) __builtin_ctz (x)
363 #define ecb_ctz64(x) __builtin_ctzll (x) 420 #define ecb_ctz64(x) __builtin_ctzll (x)
364 #define ecb_popcount32(x) __builtin_popcount (x) 421 #define ecb_popcount32(x) __builtin_popcount (x)
365 /* no popcountll */ 422 /* no popcountll */
366#else 423#else
367 ecb_function_ int ecb_ctz32 (uint32_t x) ecb_const; 424 ecb_function_ ecb_const int ecb_ctz32 (uint32_t x);
368 ecb_function_ int 425 ecb_function_ ecb_const int
369 ecb_ctz32 (uint32_t x) 426 ecb_ctz32 (uint32_t x)
370 { 427 {
371 int r = 0; 428 int r = 0;
372 429
373 x &= ~x + 1; /* this isolates the lowest bit */ 430 x &= ~x + 1; /* this isolates the lowest bit */
… …
387#endif 444#endif
388 445
389 return r; 446 return r;
390 } 447 }
391 448
392 ecb_function_ int ecb_ctz64 (uint64_t x) ecb_const; 449 ecb_function_ ecb_const int ecb_ctz64 (uint64_t x);
393 ecb_function_ int 450 ecb_function_ ecb_const int
394 ecb_ctz64 (uint64_t x) 451 ecb_ctz64 (uint64_t x)
395 { 452 {
396 int shift = x & 0xffffffffU ? 0 : 32; 453 int shift = x & 0xffffffffU ? 0 : 32;
397 return ecb_ctz32 (x >> shift) + shift; 454 return ecb_ctz32 (x >> shift) + shift;
398 } 455 }
399 456
400 ecb_function_ int ecb_popcount32 (uint32_t x) ecb_const; 457 ecb_function_ ecb_const int ecb_popcount32 (uint32_t x);
401 ecb_function_ int 458 ecb_function_ ecb_const int
402 ecb_popcount32 (uint32_t x) 459 ecb_popcount32 (uint32_t x)
403 { 460 {
404 x -= (x >> 1) & 0x55555555; 461 x -= (x >> 1) & 0x55555555;
405 x = ((x >> 2) & 0x33333333) + (x & 0x33333333); 462 x = ((x >> 2) & 0x33333333) + (x & 0x33333333);
406 x = ((x >> 4) + x) & 0x0f0f0f0f; 463 x = ((x >> 4) + x) & 0x0f0f0f0f;
407 x *= 0x01010101; 464 x *= 0x01010101;
408 465
409 return x >> 24; 466 return x >> 24;
410 } 467 }
411 468
412 ecb_function_ int ecb_ld32 (uint32_t x) ecb_const; 469 ecb_function_ ecb_const int ecb_ld32 (uint32_t x);
413 ecb_function_ int ecb_ld32 (uint32_t x) 470 ecb_function_ ecb_const int ecb_ld32 (uint32_t x)
414 { 471 {
415 int r = 0; 472 int r = 0;
416 473
417 if (x >> 16) { x >>= 16; r += 16; } 474 if (x >> 16) { x >>= 16; r += 16; }
418 if (x >> 8) { x >>= 8; r += 8; } 475 if (x >> 8) { x >>= 8; r += 8; }
… …
421 if (x >> 1) { r += 1; } 478 if (x >> 1) { r += 1; }
422 479
423 return r; 480 return r;
424 } 481 }
425 482
426 ecb_function_ int ecb_ld64 (uint64_t x) ecb_const; 483 ecb_function_ ecb_const int ecb_ld64 (uint64_t x);
427 ecb_function_ int ecb_ld64 (uint64_t x) 484 ecb_function_ ecb_const int ecb_ld64 (uint64_t x)
428 { 485 {
429 int r = 0; 486 int r = 0;
430 487
431 if (x >> 32) { x >>= 32; r += 32; } 488 if (x >> 32) { x >>= 32; r += 32; }
432 489
433 return r + ecb_ld32 (x); 490 return r + ecb_ld32 (x);
434 } 491 }
435#endif 492#endif
436 493
437ecb_function_ ecb_bool ecb_is_pot32 (uint32_t x) ecb_const; 494ecb_function_ ecb_const ecb_bool ecb_is_pot32 (uint32_t x);
438ecb_function_ ecb_bool ecb_is_pot32 (uint32_t x) { return !(x & (x - 1)); } 495ecb_function_ ecb_const ecb_bool ecb_is_pot32 (uint32_t x) { return !(x & (x - 1)); }
439ecb_function_ ecb_bool ecb_is_pot64 (uint64_t x) ecb_const; 496ecb_function_ ecb_const ecb_bool ecb_is_pot64 (uint64_t x);
440ecb_function_ ecb_bool ecb_is_pot64 (uint64_t x) { return !(x & (x - 1)); } 497ecb_function_ ecb_const ecb_bool ecb_is_pot64 (uint64_t x) { return !(x & (x - 1)); }
441 498
442ecb_function_ uint8_t ecb_bitrev8 (uint8_t x) ecb_const; 499ecb_function_ ecb_const uint8_t ecb_bitrev8 (uint8_t x);
443ecb_function_ uint8_t ecb_bitrev8 (uint8_t x) 500ecb_function_ ecb_const uint8_t ecb_bitrev8 (uint8_t x)
444{ 501{
445 return ( (x * 0x0802U & 0x22110U) 502 return ( (x * 0x0802U & 0x22110U)
446 | (x * 0x8020U & 0x88440U)) * 0x10101U >> 16; 503 | (x * 0x8020U & 0x88440U)) * 0x10101U >> 16;
447} 504}
448 505
449ecb_function_ uint16_t ecb_bitrev16 (uint16_t x) ecb_const; 506ecb_function_ ecb_const uint16_t ecb_bitrev16 (uint16_t x);
450ecb_function_ uint16_t ecb_bitrev16 (uint16_t x) 507ecb_function_ ecb_const uint16_t ecb_bitrev16 (uint16_t x)
451{ 508{
452 x = ((x >> 1) & 0x5555) | ((x & 0x5555) << 1); 509 x = ((x >> 1) & 0x5555) | ((x & 0x5555) << 1);
453 x = ((x >> 2) & 0x3333) | ((x & 0x3333) << 2); 510 x = ((x >> 2) & 0x3333) | ((x & 0x3333) << 2);
454 x = ((x >> 4) & 0x0f0f) | ((x & 0x0f0f) << 4); 511 x = ((x >> 4) & 0x0f0f) | ((x & 0x0f0f) << 4);
455 x = ( x >> 8 ) | ( x << 8); 512 x = ( x >> 8 ) | ( x << 8);
456 513
457 return x; 514 return x;
458} 515}
459 516
460ecb_function_ uint32_t ecb_bitrev32 (uint32_t x) ecb_const; 517ecb_function_ ecb_const uint32_t ecb_bitrev32 (uint32_t x);
461ecb_function_ uint32_t ecb_bitrev32 (uint32_t x) 518ecb_function_ ecb_const uint32_t ecb_bitrev32 (uint32_t x)
462{ 519{
463 x = ((x >> 1) & 0x55555555) | ((x & 0x55555555) << 1); 520 x = ((x >> 1) & 0x55555555) | ((x & 0x55555555) << 1);
464 x = ((x >> 2) & 0x33333333) | ((x & 0x33333333) << 2); 521 x = ((x >> 2) & 0x33333333) | ((x & 0x33333333) << 2);
465 x = ((x >> 4) & 0x0f0f0f0f) | ((x & 0x0f0f0f0f) << 4); 522 x = ((x >> 4) & 0x0f0f0f0f) | ((x & 0x0f0f0f0f) << 4);
466 x = ((x >> 8) & 0x00ff00ff) | ((x & 0x00ff00ff) << 8); 523 x = ((x >> 8) & 0x00ff00ff) | ((x & 0x00ff00ff) << 8);
… …
469 return x; 526 return x;
470} 527}
471 528
472/* popcount64 is only available on 64 bit cpus as gcc builtin */ 529/* popcount64 is only available on 64 bit cpus as gcc builtin */
473/* so for this version we are lazy */ 530/* so for this version we are lazy */
474ecb_function_ int ecb_popcount64 (uint64_t x) ecb_const; 531ecb_function_ ecb_const int ecb_popcount64 (uint64_t x);
475ecb_function_ int 532ecb_function_ ecb_const int
476ecb_popcount64 (uint64_t x) 533ecb_popcount64 (uint64_t x)
477{ 534{
478 return ecb_popcount32 (x) + ecb_popcount32 (x >> 32); 535 return ecb_popcount32 (x) + ecb_popcount32 (x >> 32);
479} 536}
480 537
481ecb_inline uint8_t ecb_rotl8 (uint8_t x, unsigned int count) ecb_const; 538ecb_inline ecb_const uint8_t ecb_rotl8 (uint8_t x, unsigned int count);
482ecb_inline uint8_t ecb_rotr8 (uint8_t x, unsigned int count) ecb_const; 539ecb_inline ecb_const uint8_t ecb_rotr8 (uint8_t x, unsigned int count);
483ecb_inline uint16_t ecb_rotl16 (uint16_t x, unsigned int count) ecb_const; 540ecb_inline ecb_const uint16_t ecb_rotl16 (uint16_t x, unsigned int count);
484ecb_inline uint16_t ecb_rotr16 (uint16_t x, unsigned int count) ecb_const; 541ecb_inline ecb_const uint16_t ecb_rotr16 (uint16_t x, unsigned int count);
485ecb_inline uint32_t ecb_rotl32 (uint32_t x, unsigned int count) ecb_const; 542ecb_inline ecb_const uint32_t ecb_rotl32 (uint32_t x, unsigned int count);
486ecb_inline uint32_t ecb_rotr32 (uint32_t x, unsigned int count) ecb_const; 543ecb_inline ecb_const uint32_t ecb_rotr32 (uint32_t x, unsigned int count);
487ecb_inline uint64_t ecb_rotl64 (uint64_t x, unsigned int count) ecb_const; 544ecb_inline ecb_const uint64_t ecb_rotl64 (uint64_t x, unsigned int count);
488ecb_inline uint64_t ecb_rotr64 (uint64_t x, unsigned int count) ecb_const; 545ecb_inline ecb_const uint64_t ecb_rotr64 (uint64_t x, unsigned int count);
489 546
490ecb_inline uint8_t ecb_rotl8 (uint8_t x, unsigned int count) { return (x >> ( 8 - count)) | (x << count); } 547ecb_inline ecb_const uint8_t ecb_rotl8 (uint8_t x, unsigned int count) { return (x >> ( 8 - count)) | (x << count); }
491ecb_inline uint8_t ecb_rotr8 (uint8_t x, unsigned int count) { return (x << ( 8 - count)) | (x >> count); } 548ecb_inline ecb_const uint8_t ecb_rotr8 (uint8_t x, unsigned int count) { return (x << ( 8 - count)) | (x >> count); }
492ecb_inline uint16_t ecb_rotl16 (uint16_t x, unsigned int count) { return (x >> (16 - count)) | (x << count); } 549ecb_inline ecb_const uint16_t ecb_rotl16 (uint16_t x, unsigned int count) { return (x >> (16 - count)) | (x << count); }
493ecb_inline uint16_t ecb_rotr16 (uint16_t x, unsigned int count) { return (x << (16 - count)) | (x >> count); } 550ecb_inline ecb_const uint16_t ecb_rotr16 (uint16_t x, unsigned int count) { return (x << (16 - count)) | (x >> count); }
494ecb_inline uint32_t ecb_rotl32 (uint32_t x, unsigned int count) { return (x >> (32 - count)) | (x << count); } 551ecb_inline ecb_const uint32_t ecb_rotl32 (uint32_t x, unsigned int count) { return (x >> (32 - count)) | (x << count); }
495ecb_inline uint32_t ecb_rotr32 (uint32_t x, unsigned int count) { return (x << (32 - count)) | (x >> count); } 552ecb_inline ecb_const uint32_t ecb_rotr32 (uint32_t x, unsigned int count) { return (x << (32 - count)) | (x >> count); }
496ecb_inline uint64_t ecb_rotl64 (uint64_t x, unsigned int count) { return (x >> (64 - count)) | (x << count); } 553ecb_inline ecb_const uint64_t ecb_rotl64 (uint64_t x, unsigned int count) { return (x >> (64 - count)) | (x << count); }
497ecb_inline uint64_t ecb_rotr64 (uint64_t x, unsigned int count) { return (x << (64 - count)) | (x >> count); } 554ecb_inline ecb_const uint64_t ecb_rotr64 (uint64_t x, unsigned int count) { return (x << (64 - count)) | (x >> count); }
498 555
499#if ECB_GCC_VERSION(4,3) 556#if ECB_GCC_VERSION(4,3) || (ECB_CLANG_BUILTIN(__builtin_bswap32) && ECB_CLANG_BUILTIN(__builtin_bswap64))
557 #if ECB_GCC_VERSION(4,8) || ECB_CLANG_BUILTIN(__builtin_bswap16)
558 #define ecb_bswap16(x) __builtin_bswap16 (x)
559 #else
500 #define ecb_bswap16(x) (__builtin_bswap32 (x) >> 16) 560 #define ecb_bswap16(x) (__builtin_bswap32 (x) >> 16)
561 #endif
501 #define ecb_bswap32(x) __builtin_bswap32 (x) 562 #define ecb_bswap32(x) __builtin_bswap32 (x)
502 #define ecb_bswap64(x) __builtin_bswap64 (x) 563 #define ecb_bswap64(x) __builtin_bswap64 (x)
564#elif _MSC_VER
565 #include <stdlib.h>
566 #define ecb_bswap16(x) ((uint16_t)_byteswap_ushort ((uint16_t)(x)))
567 #define ecb_bswap32(x) ((uint32_t)_byteswap_ulong ((uint32_t)(x)))
568 #define ecb_bswap64(x) ((uint64_t)_byteswap_uint64 ((uint64_t)(x)))
503#else 569#else
504 ecb_function_ uint16_t ecb_bswap16 (uint16_t x) ecb_const; 570 ecb_function_ ecb_const uint16_t ecb_bswap16 (uint16_t x);
505 ecb_function_ uint16_t 571 ecb_function_ ecb_const uint16_t
506 ecb_bswap16 (uint16_t x) 572 ecb_bswap16 (uint16_t x)
507 { 573 {
508 return ecb_rotl16 (x, 8); 574 return ecb_rotl16 (x, 8);
509 } 575 }
510 576
511 ecb_function_ uint32_t ecb_bswap32 (uint32_t x) ecb_const; 577 ecb_function_ ecb_const uint32_t ecb_bswap32 (uint32_t x);
512 ecb_function_ uint32_t 578 ecb_function_ ecb_const uint32_t
513 ecb_bswap32 (uint32_t x) 579 ecb_bswap32 (uint32_t x)
514 { 580 {
515 return (((uint32_t)ecb_bswap16 (x)) << 16) | ecb_bswap16 (x >> 16); 581 return (((uint32_t)ecb_bswap16 (x)) << 16) | ecb_bswap16 (x >> 16);
516 } 582 }
517 583
518 ecb_function_ uint64_t ecb_bswap64 (uint64_t x) ecb_const; 584 ecb_function_ ecb_const uint64_t ecb_bswap64 (uint64_t x);
519 ecb_function_ uint64_t 585 ecb_function_ ecb_const uint64_t
520 ecb_bswap64 (uint64_t x) 586 ecb_bswap64 (uint64_t x)
521 { 587 {
522 return (((uint64_t)ecb_bswap32 (x)) << 32) | ecb_bswap32 (x >> 32); 588 return (((uint64_t)ecb_bswap32 (x)) << 32) | ecb_bswap32 (x >> 32);
523 } 589 }
524#endif 590#endif
525 591
526#if ECB_GCC_VERSION(4,5) 592#if ECB_GCC_VERSION(4,5) || ECB_CLANG_BUILTIN(__builtin_unreachable)
527 #define ecb_unreachable() __builtin_unreachable () 593 #define ecb_unreachable() __builtin_unreachable ()
528#else 594#else
529 /* this seems to work fine, but gcc always emits a warning for it :/ */ 595 /* this seems to work fine, but gcc always emits a warning for it :/ */
530 ecb_inline void ecb_unreachable (void) ecb_noreturn; 596 ecb_inline ecb_noreturn void ecb_unreachable (void);
531 ecb_inline void ecb_unreachable (void) { } 597 ecb_inline ecb_noreturn void ecb_unreachable (void) { }
532#endif 598#endif
533 599
534/* try to tell the compiler that some condition is definitely true */ 600/* try to tell the compiler that some condition is definitely true */
535#define ecb_assume(cond) if (!(cond)) ecb_unreachable (); else 0 601#define ecb_assume(cond) if (!(cond)) ecb_unreachable (); else 0
536 602
537ecb_inline unsigned char ecb_byteorder_helper (void) ecb_const; 603ecb_inline ecb_const unsigned char ecb_byteorder_helper (void);
538ecb_inline unsigned char 604ecb_inline ecb_const unsigned char
539ecb_byteorder_helper (void) 605ecb_byteorder_helper (void)
540{ 606{
541 /* the union code still generates code under pressure in gcc, */ 607 /* the union code still generates code under pressure in gcc, */
542 /* but less than using pointers, and always seems to */ 608 /* but less than using pointers, and always seems to */
543 /* successfully return a constant. */ 609 /* successfully return a constant. */
544 /* the reason why we have this horrible preprocessor mess */ 610 /* the reason why we have this horrible preprocessor mess */
545 /* is to avoid it in all cases, at least on common architectures */ 611 /* is to avoid it in all cases, at least on common architectures */
546 /* or when using a recent enough gcc version (>= 4.6) */ 612 /* or when using a recent enough gcc version (>= 4.6) */
547#if __i386 || __i386__ || _M_X86 || __amd64 || __amd64__ || _M_X64 613#if ((__i386 || __i386__) && !__VOS__) || _M_IX86 || ECB_GCC_AMD64 || ECB_MSVC_AMD64
548 return 0x44; 614 return 0x44;
549#elif __BYTE_ORDER__ && __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__ 615#elif __BYTE_ORDER__ && __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__
550 return 0x44; 616 return 0x44;
551#elif __BYTE_ORDER__ && __BYTE_ORDER__ == __ORDER_BIG_ENDIAN__ 617#elif __BYTE_ORDER__ && __BYTE_ORDER__ == __ORDER_BIG_ENDIAN__
552 return 0x11; 618 return 0x11;
… …
558 } u = { 0x11223344 }; 624 } u = { 0x11223344 };
559 return u.c; 625 return u.c;
560#endif 626#endif
561} 627}
562 628
563ecb_inline ecb_bool ecb_big_endian (void) ecb_const; 629ecb_inline ecb_const ecb_bool ecb_big_endian (void);
564ecb_inline ecb_bool ecb_big_endian (void) { return ecb_byteorder_helper () == 0x11; } 630ecb_inline ecb_const ecb_bool ecb_big_endian (void) { return ecb_byteorder_helper () == 0x11; }
565ecb_inline ecb_bool ecb_little_endian (void) ecb_const; 631ecb_inline ecb_const ecb_bool ecb_little_endian (void);
566ecb_inline ecb_bool ecb_little_endian (void) { return ecb_byteorder_helper () == 0x44; } 632ecb_inline ecb_const ecb_bool ecb_little_endian (void) { return ecb_byteorder_helper () == 0x44; }
567 633
568#if ECB_GCC_VERSION(3,0) || ECB_C99 634#if ECB_GCC_VERSION(3,0) || ECB_C99
569 #define ecb_mod(m,n) ((m) % (n) + ((m) % (n) < 0 ? (n) : 0)) 635 #define ecb_mod(m,n) ((m) % (n) + ((m) % (n) < 0 ? (n) : 0))
570#else 636#else
571 #define ecb_mod(m,n) ((m) < 0 ? ((n) - 1 - ((-1 - (m)) % (n))) : ((m) % (n))) 637 #define ecb_mod(m,n) ((m) < 0 ? ((n) - 1 - ((-1 - (m)) % (n))) : ((m) % (n)))
572#endif 638#endif
573 639
574#if __cplusplus 640#if ECB_CPP
575 template<typename T> 641 template<typename T>
576 static inline T ecb_div_rd (T val, T div) 642 static inline T ecb_div_rd (T val, T div)
577 { 643 {
578 return val < 0 ? - ((-val + div - 1) / div) : (val ) / div; 644 return val < 0 ? - ((-val + div - 1) / div) : (val ) / div;
579 } 645 }
… …
603 669
604/* basically, everything uses "ieee pure-endian" floating point numbers */ 670/* basically, everything uses "ieee pure-endian" floating point numbers */
605/* the only noteworthy exception is ancient armle, which uses order 43218765 */ 671/* the only noteworthy exception is ancient armle, which uses order 43218765 */
606#if 0 \ 672#if 0 \
607 || __i386 || __i386__ \ 673 || __i386 || __i386__ \
608 || __amd64 || __amd64__ || __x86_64 || __x86_64__ \ 674 || ECB_GCC_AMD64 \
609 || __powerpc__ || __ppc__ || __powerpc64__ || __ppc64__ \ 675 || __powerpc__ || __ppc__ || __powerpc64__ || __ppc64__ \
610 || defined __s390__ || defined __s390x__ \ 676 || defined __s390__ || defined __s390x__ \
611 || defined __mips__ \ 677 || defined __mips__ \
612 || defined __alpha__ \ 678 || defined __alpha__ \
613 || defined __hppa__ \ 679 || defined __hppa__ \
614 || defined __ia64__ \ 680 || defined __ia64__ \
615 || defined __m68k__ \ 681 || defined __m68k__ \
616 || defined __m88k__ \ 682 || defined __m88k__ \
617 || defined __sh__ \ 683 || defined __sh__ \
618 || defined _M_IX86 || defined _M_AMD64 || defined _M_IA64 \ 684 || defined _M_IX86 || defined ECB_MSVC_AMD64 || defined _M_IA64 \
619 || (defined __arm__ && (defined __ARM_EABI__ || defined __EABI__ || defined __VFP_FP__ || defined _WIN32_WCE || defined __ANDROID__)) \ 685 || (defined __arm__ && (defined __ARM_EABI__ || defined __EABI__ || defined __VFP_FP__ || defined _WIN32_WCE || defined __ANDROID__)) \
620 || defined __aarch64__ 686 || defined __aarch64__
621 #define ECB_STDFP 1 687 #define ECB_STDFP 1
622 #include <string.h> /* for memcpy */ 688 #include <string.h> /* for memcpy */
623#else 689#else
… …
639 #define ECB_NAN NAN 705 #define ECB_NAN NAN
640 #else 706 #else
641 #define ECB_NAN ECB_INFINITY 707 #define ECB_NAN ECB_INFINITY
642 #endif 708 #endif
643 709
710 #if ECB_C99 || _XOPEN_VERSION >= 600 || _POSIX_VERSION >= 200112L
711 #define ecb_ldexpf(x,e) ldexpf ((x), (e))
712 #define ecb_frexpf(x,e) frexpf ((x), (e))
713 #else
714 #define ecb_ldexpf(x,e) (float) ldexp ((double) (x), (e))
715 #define ecb_frexpf(x,e) (float) frexp ((double) (x), (e))
716 #endif
717
644 /* converts an ieee half/binary16 to a float */ 718 /* converts an ieee half/binary16 to a float */
645 ecb_function_ float ecb_binary16_to_float (uint16_t x) ecb_const; 719 ecb_function_ ecb_const float ecb_binary16_to_float (uint16_t x);
646 ecb_function_ float 720 ecb_function_ ecb_const float
647 ecb_binary16_to_float (uint16_t x) 721 ecb_binary16_to_float (uint16_t x)
648 { 722 {
649 int e = (x >> 10) & 0x1f; 723 int e = (x >> 10) & 0x1f;
650 int m = x & 0x3ff; 724 int m = x & 0x3ff;
651 float r; 725 float r;
652 726
653 if (!e ) r = ldexpf (m , -24); 727 if (!e ) r = ecb_ldexpf (m , -24);
654 else if (e != 31) r = ldexpf (m + 0x400, e - 25); 728 else if (e != 31) r = ecb_ldexpf (m + 0x400, e - 25);
655 else if (m ) r = ECB_NAN; 729 else if (m ) r = ECB_NAN;
656 else r = ECB_INFINITY; 730 else r = ECB_INFINITY;
657 731
658 return x & 0x8000 ? -r : r; 732 return x & 0x8000 ? -r : r;
659 } 733 }
660 734
661 /* convert a float to ieee single/binary32 */ 735 /* convert a float to ieee single/binary32 */
662 ecb_function_ uint32_t ecb_float_to_binary32 (float x) ecb_const; 736 ecb_function_ ecb_const uint32_t ecb_float_to_binary32 (float x);
663 ecb_function_ uint32_t 737 ecb_function_ ecb_const uint32_t
664 ecb_float_to_binary32 (float x) 738 ecb_float_to_binary32 (float x)
665 { 739 {
666 uint32_t r; 740 uint32_t r;
667 741
668 #if ECB_STDFP 742 #if ECB_STDFP
… …
675 if (x == 0e0f ) return 0x00000000U; 749 if (x == 0e0f ) return 0x00000000U;
676 if (x > +3.40282346638528860e+38f) return 0x7f800000U; 750 if (x > +3.40282346638528860e+38f) return 0x7f800000U;
677 if (x < -3.40282346638528860e+38f) return 0xff800000U; 751 if (x < -3.40282346638528860e+38f) return 0xff800000U;
678 if (x != x ) return 0x7fbfffffU; 752 if (x != x ) return 0x7fbfffffU;
679 753
680 m = frexpf (x, &e) * 0x1000000U; 754 m = ecb_frexpf (x, &e) * 0x1000000U;
681 755
682 r = m & 0x80000000U; 756 r = m & 0x80000000U;
683 757
684 if (r) 758 if (r)
685 m = -m; 759 m = -m;
… …
697 771
698 return r; 772 return r;
699 } 773 }
700 774
701 /* converts an ieee single/binary32 to a float */ 775 /* converts an ieee single/binary32 to a float */
702 ecb_function_ float ecb_binary32_to_float (uint32_t x) ecb_const; 776 ecb_function_ ecb_const float ecb_binary32_to_float (uint32_t x);
703 ecb_function_ float 777 ecb_function_ ecb_const float
704 ecb_binary32_to_float (uint32_t x) 778 ecb_binary32_to_float (uint32_t x)
705 { 779 {
706 float r; 780 float r;
707 781
708 #if ECB_STDFP 782 #if ECB_STDFP
… …
718 x |= 0x800000U; 792 x |= 0x800000U;
719 else 793 else
720 e = 1; 794 e = 1;
721 795
722 /* we distrust ldexpf a bit and do the 2**-24 scaling by an extra multiply */ 796 /* we distrust ldexpf a bit and do the 2**-24 scaling by an extra multiply */
723 r = ldexpf (x * (0.5f / 0x800000U), e - 126); 797 r = ecb_ldexpf (x * (0.5f / 0x800000U), e - 126);
724 798
725 r = neg ? -r : r; 799 r = neg ? -r : r;
726 #endif 800 #endif
727 801
728 return r; 802 return r;
729 } 803 }
730 804
731 /* convert a double to ieee double/binary64 */ 805 /* convert a double to ieee double/binary64 */
732 ecb_function_ uint64_t ecb_double_to_binary64 (double x) ecb_const; 806 ecb_function_ ecb_const uint64_t ecb_double_to_binary64 (double x);
733 ecb_function_ uint64_t 807 ecb_function_ ecb_const uint64_t
734 ecb_double_to_binary64 (double x) 808 ecb_double_to_binary64 (double x)
735 { 809 {
736 uint64_t r; 810 uint64_t r;
737 811
738 #if ECB_STDFP 812 #if ECB_STDFP
… …
767 841
768 return r; 842 return r;
769 } 843 }
770 844
771 /* converts an ieee double/binary64 to a double */ 845 /* converts an ieee double/binary64 to a double */
772 ecb_function_ double ecb_binary64_to_double (uint64_t x) ecb_const; 846 ecb_function_ ecb_const double ecb_binary64_to_double (uint64_t x);
773 ecb_function_ double 847 ecb_function_ ecb_const double
774 ecb_binary64_to_double (uint64_t x) 848 ecb_binary64_to_double (uint64_t x)
775 { 849 {
776 double r; 850 double r;
777 851
778 #if ECB_STDFP 852 #if ECB_STDFP

Diff Legend

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