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

Comparing libecb/ecb.h (file contents):
Revision 1.134 by root, Mon Jun 9 17:31:57 2014 UTC vs.
Revision 1.160 by sf-exg, Thu Mar 19 15:43:51 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
… …
141 #if ECB_GCC_VERSION(2,5) || defined __INTEL_COMPILER || (__llvm__ && __GNUC__) || __SUNPRO_C >= 0x5110 || __SUNPRO_CC >= 0x5110 156 #if ECB_GCC_VERSION(2,5) || defined __INTEL_COMPILER || (__llvm__ && __GNUC__) || __SUNPRO_C >= 0x5110 || __SUNPRO_CC >= 0x5110
142 #if __i386 || __i386__ 157 #if __i386 || __i386__
143 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("lock; orb $0, -1(%%esp)" : : : "memory") 158 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("lock; orb $0, -1(%%esp)" : : : "memory")
144 #define ECB_MEMORY_FENCE_ACQUIRE __asm__ __volatile__ ("" : : : "memory") 159 #define ECB_MEMORY_FENCE_ACQUIRE __asm__ __volatile__ ("" : : : "memory")
145 #define ECB_MEMORY_FENCE_RELEASE __asm__ __volatile__ ("") 160 #define ECB_MEMORY_FENCE_RELEASE __asm__ __volatile__ ("")
146 #elif __amd64 || __amd64__ || __x86_64 || __x86_64__ 161 #elif ECB_GCC_AMD64
147 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("mfence" : : : "memory") 162 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("mfence" : : : "memory")
148 #define ECB_MEMORY_FENCE_ACQUIRE __asm__ __volatile__ ("" : : : "memory") 163 #define ECB_MEMORY_FENCE_ACQUIRE __asm__ __volatile__ ("" : : : "memory")
149 #define ECB_MEMORY_FENCE_RELEASE __asm__ __volatile__ ("") 164 #define ECB_MEMORY_FENCE_RELEASE __asm__ __volatile__ ("")
150 #elif __powerpc__ || __ppc__ || __powerpc64__ || __ppc64__ 165 #elif __powerpc__ || __ppc__ || __powerpc64__ || __ppc64__
151 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("sync" : : : "memory") 166 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("sync" : : : "memory")
… …
189 /* see comment below (stdatomic.h) about the C11 memory model. */ 204 /* see comment below (stdatomic.h) about the C11 memory model. */
190 #define ECB_MEMORY_FENCE __atomic_thread_fence (__ATOMIC_SEQ_CST) 205 #define ECB_MEMORY_FENCE __atomic_thread_fence (__ATOMIC_SEQ_CST)
191 #define ECB_MEMORY_FENCE_ACQUIRE __atomic_thread_fence (__ATOMIC_ACQUIRE) 206 #define ECB_MEMORY_FENCE_ACQUIRE __atomic_thread_fence (__ATOMIC_ACQUIRE)
192 #define ECB_MEMORY_FENCE_RELEASE __atomic_thread_fence (__ATOMIC_RELEASE) 207 #define ECB_MEMORY_FENCE_RELEASE __atomic_thread_fence (__ATOMIC_RELEASE)
193 208
194 /* The __has_feature syntax from clang is so misdesigned that we cannot use it 209 #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. 210 /* see comment below (stdatomic.h) about the C11 memory model. */
200 * #define ECB_MEMORY_FENCE __c11_atomic_thread_fence (__ATOMIC_SEQ_CST) 211 #define ECB_MEMORY_FENCE __c11_atomic_thread_fence (__ATOMIC_SEQ_CST)
201 * #define ECB_MEMORY_FENCE_ACQUIRE __c11_atomic_thread_fence (__ATOMIC_ACQUIRE) 212 #define ECB_MEMORY_FENCE_ACQUIRE __c11_atomic_thread_fence (__ATOMIC_ACQUIRE)
202 * #define ECB_MEMORY_FENCE_RELEASE __c11_atomic_thread_fence (__ATOMIC_RELEASE) 213 #define ECB_MEMORY_FENCE_RELEASE __c11_atomic_thread_fence (__ATOMIC_RELEASE)
203 */
204 214
205 #elif ECB_GCC_VERSION(4,4) || defined __INTEL_COMPILER || defined __clang__ 215 #elif ECB_GCC_VERSION(4,4) || defined __INTEL_COMPILER || defined __clang__
206 #define ECB_MEMORY_FENCE __sync_synchronize () 216 #define ECB_MEMORY_FENCE __sync_synchronize ()
207 #elif _MSC_VER >= 1500 /* VC++ 2008 */ 217 #elif _MSC_VER >= 1500 /* VC++ 2008 */
208 /* apparently, microsoft broke all the memory barrier stuff in Visual Studio 2008... */ 218 /* apparently, microsoft broke all the memory barrier stuff in Visual Studio 2008... */
… …
271 #define ECB_MEMORY_FENCE_RELEASE ECB_MEMORY_FENCE 281 #define ECB_MEMORY_FENCE_RELEASE ECB_MEMORY_FENCE
272#endif 282#endif
273 283
274/*****************************************************************************/ 284/*****************************************************************************/
275 285
276#if __cplusplus 286#if ECB_CPP
277 #define ecb_inline static inline 287 #define ecb_inline static inline
278#elif ECB_GCC_VERSION(2,5) 288#elif ECB_GCC_VERSION(2,5)
279 #define ecb_inline static __inline__ 289 #define ecb_inline static __inline__
280#elif ECB_C99 290#elif ECB_C99
281 #define ecb_inline static inline 291 #define ecb_inline static inline
… …
295 305
296#define ECB_CONCAT_(a, b) a ## b 306#define ECB_CONCAT_(a, b) a ## b
297#define ECB_CONCAT(a, b) ECB_CONCAT_(a, b) 307#define ECB_CONCAT(a, b) ECB_CONCAT_(a, b)
298#define ECB_STRINGIFY_(a) # a 308#define ECB_STRINGIFY_(a) # a
299#define ECB_STRINGIFY(a) ECB_STRINGIFY_(a) 309#define ECB_STRINGIFY(a) ECB_STRINGIFY_(a)
310#define ECB_STRINGIFY_EXPR(expr) ((expr), ECB_STRINGIFY_ (expr))
300 311
301#define ecb_function_ ecb_inline 312#define ecb_function_ ecb_inline
302 313
303#if ECB_GCC_VERSION(3,1) 314#if ECB_GCC_VERSION(3,1) || ECB_CLANG_VERSION(2,8)
304 #define ecb_attribute(attrlist) __attribute__(attrlist) 315 #define ecb_attribute(attrlist) __attribute__ (attrlist)
316#else
317 #define ecb_attribute(attrlist)
318#endif
319
320#if ECB_GCC_VERSION(3,1) || ECB_CLANG_BUILTIN(__builtin_constant_p)
305 #define ecb_is_constant(expr) __builtin_constant_p (expr) 321 #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 322#else
309 #define ecb_attribute(attrlist)
310
311 /* possible C11 impl for integral types 323 /* possible C11 impl for integral types
312 typedef struct ecb_is_constant_struct ecb_is_constant_struct; 324 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)) */ 325 #define ecb_is_constant(expr) _Generic ((1 ? (struct ecb_is_constant_struct *)0 : (void *)((expr) - (expr)), ecb_is_constant_struct *: 0, default: 1)) */
314 326
315 #define ecb_is_constant(expr) 0 327 #define ecb_is_constant(expr) 0
328#endif
329
330#if ECB_GCC_VERSION(3,1) || ECB_CLANG_BUILTIN(__builtin_expect)
331 #define ecb_expect(expr,value) __builtin_expect ((expr),(value))
332#else
316 #define ecb_expect(expr,value) (expr) 333 #define ecb_expect(expr,value) (expr)
334#endif
335
336#if ECB_GCC_VERSION(3,1) || ECB_CLANG_BUILTIN(__builtin_prefetch)
337 #define ecb_prefetch(addr,rw,locality) __builtin_prefetch (addr, rw, locality)
338#else
317 #define ecb_prefetch(addr,rw,locality) 339 #define ecb_prefetch(addr,rw,locality)
318#endif 340#endif
319 341
320/* no emulation for ecb_decltype */ 342/* no emulation for ecb_decltype */
321#if ECB_GCC_VERSION(4,5) 343#if ECB_CPP11
344 // older implementations might have problems with decltype(x)::type, work around it
345 template<class T> struct ecb_decltype_t { typedef T type; };
322 #define ecb_decltype(x) __decltype(x) 346 #define ecb_decltype(x) ecb_decltype_t<decltype (x)>::type
323#elif ECB_GCC_VERSION(3,0) 347#elif ECB_GCC_VERSION(3,0) || ECB_CLANG_VERSION(2,8)
324 #define ecb_decltype(x) __typeof(x) 348 #define ecb_decltype(x) __typeof__ (x)
325#endif 349#endif
326 350
351#if _MSC_VER >= 1300
352 #define ecb_deprecated __declspec (deprecated)
353#else
327#define ecb_deprecated ecb_attribute ((__deprecated__)) 354 #define ecb_deprecated ecb_attribute ((__deprecated__))
355#endif
356
357#if __MSC_VER >= 1500
358 #define ecb_deprecated_message(msg) __declspec (deprecated (msg))
359#elif ECB_GCC_VERSION(4,5)
360 #define ecb_deprecated_message(msg) ecb_attribute ((__deprecated__ (msg))
361#else
362 #define ecb_deprecated_message(msg) ecb_deprecated
363#endif
364
365#if _MSC_VER >= 1400
366 #define ecb_noinline __declspec (noinline)
367#else
328#define ecb_noinline ecb_attribute ((__noinline__)) 368 #define ecb_noinline ecb_attribute ((__noinline__))
369#endif
370
329#define ecb_unused ecb_attribute ((__unused__)) 371#define ecb_unused ecb_attribute ((__unused__))
330#define ecb_const ecb_attribute ((__const__)) 372#define ecb_const ecb_attribute ((__const__))
331#define ecb_pure ecb_attribute ((__pure__)) 373#define ecb_pure ecb_attribute ((__pure__))
332 374
333#if ECB_C11 375#if ECB_C11 || __IBMC_NORETURN
376 /* http://pic.dhe.ibm.com/infocenter/compbg/v121v141/topic/com.ibm.xlcpp121.bg.doc/language_ref/noreturn.html */
334 #define ecb_noreturn _Noreturn 377 #define ecb_noreturn _Noreturn
378#elif ECB_CPP11
379 #define ecb_noreturn [[noreturn]]
380#elif _MSC_VER >= 1200
381 /* http://msdn.microsoft.com/en-us/library/k6ktzx3s.aspx */
382 #define ecb_noreturn __declspec (noreturn)
335#else 383#else
336 #define ecb_noreturn ecb_attribute ((__noreturn__)) 384 #define ecb_noreturn ecb_attribute ((__noreturn__))
337#endif 385#endif
338 386
339#if ECB_GCC_VERSION(4,3) 387#if ECB_GCC_VERSION(4,3)
… …
354/* for compatibility to the rest of the world */ 402/* for compatibility to the rest of the world */
355#define ecb_likely(expr) ecb_expect_true (expr) 403#define ecb_likely(expr) ecb_expect_true (expr)
356#define ecb_unlikely(expr) ecb_expect_false (expr) 404#define ecb_unlikely(expr) ecb_expect_false (expr)
357 405
358/* count trailing zero bits and count # of one bits */ 406/* count trailing zero bits and count # of one bits */
359#if ECB_GCC_VERSION(3,4) 407#if ECB_GCC_VERSION(3,4) \
408 || (ECB_CLANG_BUILTIN(__builtin_clz) && ECB_CLANG_BUILTIN(__builtin_clzll) \
409 && ECB_CLANG_BUILTIN(__builtin_ctz) && ECB_CLANG_BUILTIN(__builtin_ctzll) \
410 && ECB_CLANG_BUILTIN(__builtin_popcount))
360 /* we assume int == 32 bit, long == 32 or 64 bit and long long == 64 bit */ 411 /* we assume int == 32 bit, long == 32 or 64 bit and long long == 64 bit */
361 #define ecb_ld32(x) (__builtin_clz (x) ^ 31) 412 #define ecb_ld32(x) (__builtin_clz (x) ^ 31)
362 #define ecb_ld64(x) (__builtin_clzll (x) ^ 63) 413 #define ecb_ld64(x) (__builtin_clzll (x) ^ 63)
363 #define ecb_ctz32(x) __builtin_ctz (x) 414 #define ecb_ctz32(x) __builtin_ctz (x)
364 #define ecb_ctz64(x) __builtin_ctzll (x) 415 #define ecb_ctz64(x) __builtin_ctzll (x)
365 #define ecb_popcount32(x) __builtin_popcount (x) 416 #define ecb_popcount32(x) __builtin_popcount (x)
366 /* no popcountll */ 417 /* no popcountll */
367#else 418#else
368 ecb_function_ int ecb_ctz32 (uint32_t x) ecb_const; 419 ecb_function_ ecb_const int ecb_ctz32 (uint32_t x);
369 ecb_function_ int 420 ecb_function_ ecb_const int
370 ecb_ctz32 (uint32_t x) 421 ecb_ctz32 (uint32_t x)
371 { 422 {
372 int r = 0; 423 int r = 0;
373 424
374 x &= ~x + 1; /* this isolates the lowest bit */ 425 x &= ~x + 1; /* this isolates the lowest bit */
… …
388#endif 439#endif
389 440
390 return r; 441 return r;
391 } 442 }
392 443
393 ecb_function_ int ecb_ctz64 (uint64_t x) ecb_const; 444 ecb_function_ ecb_const int ecb_ctz64 (uint64_t x);
394 ecb_function_ int 445 ecb_function_ ecb_const int
395 ecb_ctz64 (uint64_t x) 446 ecb_ctz64 (uint64_t x)
396 { 447 {
397 int shift = x & 0xffffffffU ? 0 : 32; 448 int shift = x & 0xffffffffU ? 0 : 32;
398 return ecb_ctz32 (x >> shift) + shift; 449 return ecb_ctz32 (x >> shift) + shift;
399 } 450 }
400 451
401 ecb_function_ int ecb_popcount32 (uint32_t x) ecb_const; 452 ecb_function_ ecb_const int ecb_popcount32 (uint32_t x);
402 ecb_function_ int 453 ecb_function_ ecb_const int
403 ecb_popcount32 (uint32_t x) 454 ecb_popcount32 (uint32_t x)
404 { 455 {
405 x -= (x >> 1) & 0x55555555; 456 x -= (x >> 1) & 0x55555555;
406 x = ((x >> 2) & 0x33333333) + (x & 0x33333333); 457 x = ((x >> 2) & 0x33333333) + (x & 0x33333333);
407 x = ((x >> 4) + x) & 0x0f0f0f0f; 458 x = ((x >> 4) + x) & 0x0f0f0f0f;
408 x *= 0x01010101; 459 x *= 0x01010101;
409 460
410 return x >> 24; 461 return x >> 24;
411 } 462 }
412 463
413 ecb_function_ int ecb_ld32 (uint32_t x) ecb_const; 464 ecb_function_ ecb_const int ecb_ld32 (uint32_t x);
414 ecb_function_ int ecb_ld32 (uint32_t x) 465 ecb_function_ ecb_const int ecb_ld32 (uint32_t x)
415 { 466 {
416 int r = 0; 467 int r = 0;
417 468
418 if (x >> 16) { x >>= 16; r += 16; } 469 if (x >> 16) { x >>= 16; r += 16; }
419 if (x >> 8) { x >>= 8; r += 8; } 470 if (x >> 8) { x >>= 8; r += 8; }
… …
422 if (x >> 1) { r += 1; } 473 if (x >> 1) { r += 1; }
423 474
424 return r; 475 return r;
425 } 476 }
426 477
427 ecb_function_ int ecb_ld64 (uint64_t x) ecb_const; 478 ecb_function_ ecb_const int ecb_ld64 (uint64_t x);
428 ecb_function_ int ecb_ld64 (uint64_t x) 479 ecb_function_ ecb_const int ecb_ld64 (uint64_t x)
429 { 480 {
430 int r = 0; 481 int r = 0;
431 482
432 if (x >> 32) { x >>= 32; r += 32; } 483 if (x >> 32) { x >>= 32; r += 32; }
433 484
434 return r + ecb_ld32 (x); 485 return r + ecb_ld32 (x);
435 } 486 }
436#endif 487#endif
437 488
438ecb_function_ ecb_bool ecb_is_pot32 (uint32_t x) ecb_const; 489ecb_function_ ecb_const ecb_bool ecb_is_pot32 (uint32_t x);
439ecb_function_ ecb_bool ecb_is_pot32 (uint32_t x) { return !(x & (x - 1)); } 490ecb_function_ ecb_const ecb_bool ecb_is_pot32 (uint32_t x) { return !(x & (x - 1)); }
440ecb_function_ ecb_bool ecb_is_pot64 (uint64_t x) ecb_const; 491ecb_function_ ecb_const ecb_bool ecb_is_pot64 (uint64_t x);
441ecb_function_ ecb_bool ecb_is_pot64 (uint64_t x) { return !(x & (x - 1)); } 492ecb_function_ ecb_const ecb_bool ecb_is_pot64 (uint64_t x) { return !(x & (x - 1)); }
442 493
443ecb_function_ uint8_t ecb_bitrev8 (uint8_t x) ecb_const; 494ecb_function_ ecb_const uint8_t ecb_bitrev8 (uint8_t x);
444ecb_function_ uint8_t ecb_bitrev8 (uint8_t x) 495ecb_function_ ecb_const uint8_t ecb_bitrev8 (uint8_t x)
445{ 496{
446 return ( (x * 0x0802U & 0x22110U) 497 return ( (x * 0x0802U & 0x22110U)
447 | (x * 0x8020U & 0x88440U)) * 0x10101U >> 16; 498 | (x * 0x8020U & 0x88440U)) * 0x10101U >> 16;
448} 499}
449 500
450ecb_function_ uint16_t ecb_bitrev16 (uint16_t x) ecb_const; 501ecb_function_ ecb_const uint16_t ecb_bitrev16 (uint16_t x);
451ecb_function_ uint16_t ecb_bitrev16 (uint16_t x) 502ecb_function_ ecb_const uint16_t ecb_bitrev16 (uint16_t x)
452{ 503{
453 x = ((x >> 1) & 0x5555) | ((x & 0x5555) << 1); 504 x = ((x >> 1) & 0x5555) | ((x & 0x5555) << 1);
454 x = ((x >> 2) & 0x3333) | ((x & 0x3333) << 2); 505 x = ((x >> 2) & 0x3333) | ((x & 0x3333) << 2);
455 x = ((x >> 4) & 0x0f0f) | ((x & 0x0f0f) << 4); 506 x = ((x >> 4) & 0x0f0f) | ((x & 0x0f0f) << 4);
456 x = ( x >> 8 ) | ( x << 8); 507 x = ( x >> 8 ) | ( x << 8);
457 508
458 return x; 509 return x;
459} 510}
460 511
461ecb_function_ uint32_t ecb_bitrev32 (uint32_t x) ecb_const; 512ecb_function_ ecb_const uint32_t ecb_bitrev32 (uint32_t x);
462ecb_function_ uint32_t ecb_bitrev32 (uint32_t x) 513ecb_function_ ecb_const uint32_t ecb_bitrev32 (uint32_t x)
463{ 514{
464 x = ((x >> 1) & 0x55555555) | ((x & 0x55555555) << 1); 515 x = ((x >> 1) & 0x55555555) | ((x & 0x55555555) << 1);
465 x = ((x >> 2) & 0x33333333) | ((x & 0x33333333) << 2); 516 x = ((x >> 2) & 0x33333333) | ((x & 0x33333333) << 2);
466 x = ((x >> 4) & 0x0f0f0f0f) | ((x & 0x0f0f0f0f) << 4); 517 x = ((x >> 4) & 0x0f0f0f0f) | ((x & 0x0f0f0f0f) << 4);
467 x = ((x >> 8) & 0x00ff00ff) | ((x & 0x00ff00ff) << 8); 518 x = ((x >> 8) & 0x00ff00ff) | ((x & 0x00ff00ff) << 8);
… …
470 return x; 521 return x;
471} 522}
472 523
473/* popcount64 is only available on 64 bit cpus as gcc builtin */ 524/* popcount64 is only available on 64 bit cpus as gcc builtin */
474/* so for this version we are lazy */ 525/* so for this version we are lazy */
475ecb_function_ int ecb_popcount64 (uint64_t x) ecb_const; 526ecb_function_ ecb_const int ecb_popcount64 (uint64_t x);
476ecb_function_ int 527ecb_function_ ecb_const int
477ecb_popcount64 (uint64_t x) 528ecb_popcount64 (uint64_t x)
478{ 529{
479 return ecb_popcount32 (x) + ecb_popcount32 (x >> 32); 530 return ecb_popcount32 (x) + ecb_popcount32 (x >> 32);
480} 531}
481 532
482ecb_inline uint8_t ecb_rotl8 (uint8_t x, unsigned int count) ecb_const; 533ecb_inline ecb_const uint8_t ecb_rotl8 (uint8_t x, unsigned int count);
483ecb_inline uint8_t ecb_rotr8 (uint8_t x, unsigned int count) ecb_const; 534ecb_inline ecb_const uint8_t ecb_rotr8 (uint8_t x, unsigned int count);
484ecb_inline uint16_t ecb_rotl16 (uint16_t x, unsigned int count) ecb_const; 535ecb_inline ecb_const uint16_t ecb_rotl16 (uint16_t x, unsigned int count);
485ecb_inline uint16_t ecb_rotr16 (uint16_t x, unsigned int count) ecb_const; 536ecb_inline ecb_const uint16_t ecb_rotr16 (uint16_t x, unsigned int count);
486ecb_inline uint32_t ecb_rotl32 (uint32_t x, unsigned int count) ecb_const; 537ecb_inline ecb_const uint32_t ecb_rotl32 (uint32_t x, unsigned int count);
487ecb_inline uint32_t ecb_rotr32 (uint32_t x, unsigned int count) ecb_const; 538ecb_inline ecb_const uint32_t ecb_rotr32 (uint32_t x, unsigned int count);
488ecb_inline uint64_t ecb_rotl64 (uint64_t x, unsigned int count) ecb_const; 539ecb_inline ecb_const uint64_t ecb_rotl64 (uint64_t x, unsigned int count);
489ecb_inline uint64_t ecb_rotr64 (uint64_t x, unsigned int count) ecb_const; 540ecb_inline ecb_const uint64_t ecb_rotr64 (uint64_t x, unsigned int count);
490 541
491ecb_inline uint8_t ecb_rotl8 (uint8_t x, unsigned int count) { return (x >> ( 8 - count)) | (x << count); } 542ecb_inline ecb_const uint8_t ecb_rotl8 (uint8_t x, unsigned int count) { return (x >> ( 8 - count)) | (x << count); }
492ecb_inline uint8_t ecb_rotr8 (uint8_t x, unsigned int count) { return (x << ( 8 - count)) | (x >> count); } 543ecb_inline ecb_const uint8_t ecb_rotr8 (uint8_t x, unsigned int count) { return (x << ( 8 - count)) | (x >> count); }
493ecb_inline uint16_t ecb_rotl16 (uint16_t x, unsigned int count) { return (x >> (16 - count)) | (x << count); } 544ecb_inline ecb_const uint16_t ecb_rotl16 (uint16_t x, unsigned int count) { return (x >> (16 - count)) | (x << count); }
494ecb_inline uint16_t ecb_rotr16 (uint16_t x, unsigned int count) { return (x << (16 - count)) | (x >> count); } 545ecb_inline ecb_const uint16_t ecb_rotr16 (uint16_t x, unsigned int count) { return (x << (16 - count)) | (x >> count); }
495ecb_inline uint32_t ecb_rotl32 (uint32_t x, unsigned int count) { return (x >> (32 - count)) | (x << count); } 546ecb_inline ecb_const uint32_t ecb_rotl32 (uint32_t x, unsigned int count) { return (x >> (32 - count)) | (x << count); }
496ecb_inline uint32_t ecb_rotr32 (uint32_t x, unsigned int count) { return (x << (32 - count)) | (x >> count); } 547ecb_inline ecb_const uint32_t ecb_rotr32 (uint32_t x, unsigned int count) { return (x << (32 - count)) | (x >> count); }
497ecb_inline uint64_t ecb_rotl64 (uint64_t x, unsigned int count) { return (x >> (64 - count)) | (x << count); } 548ecb_inline ecb_const uint64_t ecb_rotl64 (uint64_t x, unsigned int count) { return (x >> (64 - count)) | (x << count); }
498ecb_inline uint64_t ecb_rotr64 (uint64_t x, unsigned int count) { return (x << (64 - count)) | (x >> count); } 549ecb_inline ecb_const uint64_t ecb_rotr64 (uint64_t x, unsigned int count) { return (x << (64 - count)) | (x >> count); }
499 550
500#if ECB_GCC_VERSION(4,3) 551#if ECB_GCC_VERSION(4,3) || (ECB_CLANG_BUILTIN(__builtin_bswap32) && ECB_CLANG_BUILTIN(__builtin_bswap64))
501 #define ecb_bswap16(x) (__builtin_bswap32 (x) >> 16) 552 #define ecb_bswap16(x) (__builtin_bswap32 (x) >> 16)
502 #define ecb_bswap32(x) __builtin_bswap32 (x) 553 #define ecb_bswap32(x) __builtin_bswap32 (x)
503 #define ecb_bswap64(x) __builtin_bswap64 (x) 554 #define ecb_bswap64(x) __builtin_bswap64 (x)
504#else 555#else
505 ecb_function_ uint16_t ecb_bswap16 (uint16_t x) ecb_const; 556 ecb_function_ ecb_const uint16_t ecb_bswap16 (uint16_t x);
506 ecb_function_ uint16_t 557 ecb_function_ ecb_const uint16_t
507 ecb_bswap16 (uint16_t x) 558 ecb_bswap16 (uint16_t x)
508 { 559 {
509 return ecb_rotl16 (x, 8); 560 return ecb_rotl16 (x, 8);
510 } 561 }
511 562
512 ecb_function_ uint32_t ecb_bswap32 (uint32_t x) ecb_const; 563 ecb_function_ ecb_const uint32_t ecb_bswap32 (uint32_t x);
513 ecb_function_ uint32_t 564 ecb_function_ ecb_const uint32_t
514 ecb_bswap32 (uint32_t x) 565 ecb_bswap32 (uint32_t x)
515 { 566 {
516 return (((uint32_t)ecb_bswap16 (x)) << 16) | ecb_bswap16 (x >> 16); 567 return (((uint32_t)ecb_bswap16 (x)) << 16) | ecb_bswap16 (x >> 16);
517 } 568 }
518 569
519 ecb_function_ uint64_t ecb_bswap64 (uint64_t x) ecb_const; 570 ecb_function_ ecb_const uint64_t ecb_bswap64 (uint64_t x);
520 ecb_function_ uint64_t 571 ecb_function_ ecb_const uint64_t
521 ecb_bswap64 (uint64_t x) 572 ecb_bswap64 (uint64_t x)
522 { 573 {
523 return (((uint64_t)ecb_bswap32 (x)) << 32) | ecb_bswap32 (x >> 32); 574 return (((uint64_t)ecb_bswap32 (x)) << 32) | ecb_bswap32 (x >> 32);
524 } 575 }
525#endif 576#endif
526 577
527#if ECB_GCC_VERSION(4,5) 578#if ECB_GCC_VERSION(4,5) || ECB_CLANG_BUILTIN(__builtin_unreachable)
528 #define ecb_unreachable() __builtin_unreachable () 579 #define ecb_unreachable() __builtin_unreachable ()
529#else 580#else
530 /* this seems to work fine, but gcc always emits a warning for it :/ */ 581 /* this seems to work fine, but gcc always emits a warning for it :/ */
531 ecb_inline void ecb_unreachable (void) ecb_noreturn; 582 ecb_inline ecb_noreturn void ecb_unreachable (void);
532 ecb_inline void ecb_unreachable (void) { } 583 ecb_inline ecb_noreturn void ecb_unreachable (void) { }
533#endif 584#endif
534 585
535/* try to tell the compiler that some condition is definitely true */ 586/* try to tell the compiler that some condition is definitely true */
536#define ecb_assume(cond) if (!(cond)) ecb_unreachable (); else 0 587#define ecb_assume(cond) if (!(cond)) ecb_unreachable (); else 0
537 588
538ecb_inline unsigned char ecb_byteorder_helper (void) ecb_const; 589ecb_inline ecb_const unsigned char ecb_byteorder_helper (void);
539ecb_inline unsigned char 590ecb_inline ecb_const unsigned char
540ecb_byteorder_helper (void) 591ecb_byteorder_helper (void)
541{ 592{
542 /* the union code still generates code under pressure in gcc, */ 593 /* the union code still generates code under pressure in gcc, */
543 /* but less than using pointers, and always seems to */ 594 /* but less than using pointers, and always seems to */
544 /* successfully return a constant. */ 595 /* successfully return a constant. */
545 /* the reason why we have this horrible preprocessor mess */ 596 /* the reason why we have this horrible preprocessor mess */
546 /* is to avoid it in all cases, at least on common architectures */ 597 /* is to avoid it in all cases, at least on common architectures */
547 /* or when using a recent enough gcc version (>= 4.6) */ 598 /* or when using a recent enough gcc version (>= 4.6) */
548#if __i386 || __i386__ || _M_X86 || __amd64 || __amd64__ || _M_X64 599#if ((__i386 || __i386__) && !__VOS__) || _M_IX86 || ECB_GCC_AMD64 || ECB_MSVC_AMD64
549 return 0x44; 600 return 0x44;
550#elif __BYTE_ORDER__ && __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__ 601#elif __BYTE_ORDER__ && __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__
551 return 0x44; 602 return 0x44;
552#elif __BYTE_ORDER__ && __BYTE_ORDER__ == __ORDER_BIG_ENDIAN__ 603#elif __BYTE_ORDER__ && __BYTE_ORDER__ == __ORDER_BIG_ENDIAN__
553 return 0x11; 604 return 0x11;
… …
559 } u = { 0x11223344 }; 610 } u = { 0x11223344 };
560 return u.c; 611 return u.c;
561#endif 612#endif
562} 613}
563 614
564ecb_inline ecb_bool ecb_big_endian (void) ecb_const; 615ecb_inline ecb_const ecb_bool ecb_big_endian (void);
565ecb_inline ecb_bool ecb_big_endian (void) { return ecb_byteorder_helper () == 0x11; } 616ecb_inline ecb_const ecb_bool ecb_big_endian (void) { return ecb_byteorder_helper () == 0x11; }
566ecb_inline ecb_bool ecb_little_endian (void) ecb_const; 617ecb_inline ecb_const ecb_bool ecb_little_endian (void);
567ecb_inline ecb_bool ecb_little_endian (void) { return ecb_byteorder_helper () == 0x44; } 618ecb_inline ecb_const ecb_bool ecb_little_endian (void) { return ecb_byteorder_helper () == 0x44; }
568 619
569#if ECB_GCC_VERSION(3,0) || ECB_C99 620#if ECB_GCC_VERSION(3,0) || ECB_C99
570 #define ecb_mod(m,n) ((m) % (n) + ((m) % (n) < 0 ? (n) : 0)) 621 #define ecb_mod(m,n) ((m) % (n) + ((m) % (n) < 0 ? (n) : 0))
571#else 622#else
572 #define ecb_mod(m,n) ((m) < 0 ? ((n) - 1 - ((-1 - (m)) % (n))) : ((m) % (n))) 623 #define ecb_mod(m,n) ((m) < 0 ? ((n) - 1 - ((-1 - (m)) % (n))) : ((m) % (n)))
573#endif 624#endif
574 625
575#if __cplusplus 626#if ECB_CPP
576 template<typename T> 627 template<typename T>
577 static inline T ecb_div_rd (T val, T div) 628 static inline T ecb_div_rd (T val, T div)
578 { 629 {
579 return val < 0 ? - ((-val + div - 1) / div) : (val ) / div; 630 return val < 0 ? - ((-val + div - 1) / div) : (val ) / div;
580 } 631 }
… …
604 655
605/* basically, everything uses "ieee pure-endian" floating point numbers */ 656/* basically, everything uses "ieee pure-endian" floating point numbers */
606/* the only noteworthy exception is ancient armle, which uses order 43218765 */ 657/* the only noteworthy exception is ancient armle, which uses order 43218765 */
607#if 0 \ 658#if 0 \
608 || __i386 || __i386__ \ 659 || __i386 || __i386__ \
609 || __amd64 || __amd64__ || __x86_64 || __x86_64__ \ 660 || ECB_GCC_AMD64 \
610 || __powerpc__ || __ppc__ || __powerpc64__ || __ppc64__ \ 661 || __powerpc__ || __ppc__ || __powerpc64__ || __ppc64__ \
611 || defined __s390__ || defined __s390x__ \ 662 || defined __s390__ || defined __s390x__ \
612 || defined __mips__ \ 663 || defined __mips__ \
613 || defined __alpha__ \ 664 || defined __alpha__ \
614 || defined __hppa__ \ 665 || defined __hppa__ \
615 || defined __ia64__ \ 666 || defined __ia64__ \
616 || defined __m68k__ \ 667 || defined __m68k__ \
617 || defined __m88k__ \ 668 || defined __m88k__ \
618 || defined __sh__ \ 669 || defined __sh__ \
619 || defined _M_IX86 || defined _M_AMD64 || defined _M_IA64 \ 670 || defined _M_IX86 || defined ECB_MSVC_AMD64 || defined _M_IA64 \
620 || (defined __arm__ && (defined __ARM_EABI__ || defined __EABI__ || defined __VFP_FP__ || defined _WIN32_WCE || defined __ANDROID__)) \ 671 || (defined __arm__ && (defined __ARM_EABI__ || defined __EABI__ || defined __VFP_FP__ || defined _WIN32_WCE || defined __ANDROID__)) \
621 || defined __aarch64__ 672 || defined __aarch64__
622 #define ECB_STDFP 1 673 #define ECB_STDFP 1
623 #include <string.h> /* for memcpy */ 674 #include <string.h> /* for memcpy */
624#else 675#else
… …
640 #define ECB_NAN NAN 691 #define ECB_NAN NAN
641 #else 692 #else
642 #define ECB_NAN ECB_INFINITY 693 #define ECB_NAN ECB_INFINITY
643 #endif 694 #endif
644 695
696 #if ECB_C99 || _XOPEN_VERSION >= 600 || _POSIX_VERSION >= 200112L
697 #define ecb_ldexpf(x,e) ldexpf ((x), (e))
698 #else
699 #define ecb_ldexpf(x,e) (float) ldexp ((float) (x), (e))
700 #endif
701
645 /* converts an ieee half/binary16 to a float */ 702 /* converts an ieee half/binary16 to a float */
646 ecb_function_ float ecb_binary16_to_float (uint16_t x) ecb_const; 703 ecb_function_ ecb_const float ecb_binary16_to_float (uint16_t x);
647 ecb_function_ float 704 ecb_function_ ecb_const float
648 ecb_binary16_to_float (uint16_t x) 705 ecb_binary16_to_float (uint16_t x)
649 { 706 {
650 int e = (x >> 10) & 0x1f; 707 int e = (x >> 10) & 0x1f;
651 int m = x & 0x3ff; 708 int m = x & 0x3ff;
652 float r; 709 float r;
653 710
654 if (!e ) r = ldexpf (m , -24); 711 if (!e ) r = ecb_ldexpf (m , -24);
655 else if (e != 31) r = ldexpf (m + 0x400, e - 25); 712 else if (e != 31) r = ecb_ldexpf (m + 0x400, e - 25);
656 else if (m ) r = ECB_NAN; 713 else if (m ) r = ECB_NAN;
657 else r = ECB_INFINITY; 714 else r = ECB_INFINITY;
658 715
659 return x & 0x8000 ? -r : r; 716 return x & 0x8000 ? -r : r;
660 } 717 }
661 718
662 /* convert a float to ieee single/binary32 */ 719 /* convert a float to ieee single/binary32 */
663 ecb_function_ uint32_t ecb_float_to_binary32 (float x) ecb_const; 720 ecb_function_ ecb_const uint32_t ecb_float_to_binary32 (float x);
664 ecb_function_ uint32_t 721 ecb_function_ ecb_const uint32_t
665 ecb_float_to_binary32 (float x) 722 ecb_float_to_binary32 (float x)
666 { 723 {
667 uint32_t r; 724 uint32_t r;
668 725
669 #if ECB_STDFP 726 #if ECB_STDFP
… …
698 755
699 return r; 756 return r;
700 } 757 }
701 758
702 /* converts an ieee single/binary32 to a float */ 759 /* converts an ieee single/binary32 to a float */
703 ecb_function_ float ecb_binary32_to_float (uint32_t x) ecb_const; 760 ecb_function_ ecb_const float ecb_binary32_to_float (uint32_t x);
704 ecb_function_ float 761 ecb_function_ ecb_const float
705 ecb_binary32_to_float (uint32_t x) 762 ecb_binary32_to_float (uint32_t x)
706 { 763 {
707 float r; 764 float r;
708 765
709 #if ECB_STDFP 766 #if ECB_STDFP
… …
719 x |= 0x800000U; 776 x |= 0x800000U;
720 else 777 else
721 e = 1; 778 e = 1;
722 779
723 /* we distrust ldexpf a bit and do the 2**-24 scaling by an extra multiply */ 780 /* we distrust ldexpf a bit and do the 2**-24 scaling by an extra multiply */
724 r = ldexpf (x * (0.5f / 0x800000U), e - 126); 781 r = ecb_ldexpf (x * (0.5f / 0x800000U), e - 126);
725 782
726 r = neg ? -r : r; 783 r = neg ? -r : r;
727 #endif 784 #endif
728 785
729 return r; 786 return r;
730 } 787 }
731 788
732 /* convert a double to ieee double/binary64 */ 789 /* convert a double to ieee double/binary64 */
733 ecb_function_ uint64_t ecb_double_to_binary64 (double x) ecb_const; 790 ecb_function_ ecb_const uint64_t ecb_double_to_binary64 (double x);
734 ecb_function_ uint64_t 791 ecb_function_ ecb_const uint64_t
735 ecb_double_to_binary64 (double x) 792 ecb_double_to_binary64 (double x)
736 { 793 {
737 uint64_t r; 794 uint64_t r;
738 795
739 #if ECB_STDFP 796 #if ECB_STDFP
… …
768 825
769 return r; 826 return r;
770 } 827 }
771 828
772 /* converts an ieee double/binary64 to a double */ 829 /* converts an ieee double/binary64 to a double */
773 ecb_function_ double ecb_binary64_to_double (uint64_t x) ecb_const; 830 ecb_function_ ecb_const double ecb_binary64_to_double (uint64_t x);
774 ecb_function_ double 831 ecb_function_ ecb_const double
775 ecb_binary64_to_double (uint64_t x) 832 ecb_binary64_to_double (uint64_t x)
776 { 833 {
777 double r; 834 double r;
778 835
779 #if ECB_STDFP 836 #if ECB_STDFP

Diff Legend

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