… | |
… | |
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; |
… | |
… | |
90 | * causing enormous grief in return for some better fake benchmark numbers. |
90 | * causing enormous grief in return for some better fake benchmark numbers. |
91 | * or so. |
91 | * or so. |
92 | * we try to detect these and simply assume they are not gcc - if they have |
92 | * 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. |
93 | * an issue with that they should have done it right in the first place. |
94 | */ |
94 | */ |
95 | #ifndef ECB_GCC_VERSION |
|
|
96 | #if !defined __GNUC_MINOR__ || defined __INTEL_COMPILER || defined __SUNPRO_C || defined __SUNPRO_CC || defined __llvm__ || defined __clang__ |
95 | #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 |
96 | #define ECB_GCC_VERSION(major,minor) 0 |
98 | #else |
97 | #else |
99 | #define ECB_GCC_VERSION(major,minor) (__GNUC__ > (major) || (__GNUC__ == (major) && __GNUC_MINOR__ >= (minor))) |
98 | #define ECB_GCC_VERSION(major,minor) (__GNUC__ > (major) || (__GNUC__ == (major) && __GNUC_MINOR__ >= (minor))) |
100 | #endif |
99 | #endif |
|
|
100 | |
|
|
101 | #define ECB_CLANG_VERSION(major,minor) (__clang_major__ > (major) || (__clang_major__ == (major) && __clang_minor__ >= (minor))) |
|
|
102 | |
|
|
103 | #if __clang__ && defined __has_builtin |
|
|
104 | #define ECB_CLANG_BUILTIN(x) __has_builtin (x) |
|
|
105 | #else |
|
|
106 | #define ECB_CLANG_BUILTIN(x) 0 |
|
|
107 | #endif |
|
|
108 | |
|
|
109 | #if __clang__ && defined __has_extension |
|
|
110 | #define ECB_CLANG_EXTENSION(x) __has_extension (x) |
|
|
111 | #else |
|
|
112 | #define ECB_CLANG_EXTENSION(x) 0 |
101 | #endif |
113 | #endif |
102 | |
114 | |
103 | #define ECB_CPP (__cplusplus+0) |
115 | #define ECB_CPP (__cplusplus+0) |
104 | #define ECB_CPP11 (__cplusplus >= 201103L) |
116 | #define ECB_CPP11 (__cplusplus >= 201103L) |
105 | |
117 | |
… | |
… | |
189 | /* see comment below (stdatomic.h) about the C11 memory model. */ |
201 | /* see comment below (stdatomic.h) about the C11 memory model. */ |
190 | #define ECB_MEMORY_FENCE __atomic_thread_fence (__ATOMIC_SEQ_CST) |
202 | #define ECB_MEMORY_FENCE __atomic_thread_fence (__ATOMIC_SEQ_CST) |
191 | #define ECB_MEMORY_FENCE_ACQUIRE __atomic_thread_fence (__ATOMIC_ACQUIRE) |
203 | #define ECB_MEMORY_FENCE_ACQUIRE __atomic_thread_fence (__ATOMIC_ACQUIRE) |
192 | #define ECB_MEMORY_FENCE_RELEASE __atomic_thread_fence (__ATOMIC_RELEASE) |
204 | #define ECB_MEMORY_FENCE_RELEASE __atomic_thread_fence (__ATOMIC_RELEASE) |
193 | |
205 | |
194 | /* The __has_feature syntax from clang is so misdesigned that we cannot use it |
206 | #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. |
207 | /* see comment below (stdatomic.h) about the C11 memory model. */ |
200 | * #define ECB_MEMORY_FENCE __c11_atomic_thread_fence (__ATOMIC_SEQ_CST) |
208 | #define ECB_MEMORY_FENCE __c11_atomic_thread_fence (__ATOMIC_SEQ_CST) |
201 | * #define ECB_MEMORY_FENCE_ACQUIRE __c11_atomic_thread_fence (__ATOMIC_ACQUIRE) |
209 | #define ECB_MEMORY_FENCE_ACQUIRE __c11_atomic_thread_fence (__ATOMIC_ACQUIRE) |
202 | * #define ECB_MEMORY_FENCE_RELEASE __c11_atomic_thread_fence (__ATOMIC_RELEASE) |
210 | #define ECB_MEMORY_FENCE_RELEASE __c11_atomic_thread_fence (__ATOMIC_RELEASE) |
203 | */ |
|
|
204 | |
211 | |
205 | #elif ECB_GCC_VERSION(4,4) || defined __INTEL_COMPILER || defined __clang__ |
212 | #elif ECB_GCC_VERSION(4,4) || defined __INTEL_COMPILER || defined __clang__ |
206 | #define ECB_MEMORY_FENCE __sync_synchronize () |
213 | #define ECB_MEMORY_FENCE __sync_synchronize () |
207 | #elif _MSC_VER >= 1500 /* VC++ 2008 */ |
214 | #elif _MSC_VER >= 1500 /* VC++ 2008 */ |
208 | /* apparently, microsoft broke all the memory barrier stuff in Visual Studio 2008... */ |
215 | /* apparently, microsoft broke all the memory barrier stuff in Visual Studio 2008... */ |
… | |
… | |
271 | #define ECB_MEMORY_FENCE_RELEASE ECB_MEMORY_FENCE |
278 | #define ECB_MEMORY_FENCE_RELEASE ECB_MEMORY_FENCE |
272 | #endif |
279 | #endif |
273 | |
280 | |
274 | /*****************************************************************************/ |
281 | /*****************************************************************************/ |
275 | |
282 | |
276 | #if __cplusplus |
283 | #if ECB_CPP |
277 | #define ecb_inline static inline |
284 | #define ecb_inline static inline |
278 | #elif ECB_GCC_VERSION(2,5) |
285 | #elif ECB_GCC_VERSION(2,5) |
279 | #define ecb_inline static __inline__ |
286 | #define ecb_inline static __inline__ |
280 | #elif ECB_C99 |
287 | #elif ECB_C99 |
281 | #define ecb_inline static inline |
288 | #define ecb_inline static inline |
… | |
… | |
298 | #define ECB_STRINGIFY_(a) # a |
305 | #define ECB_STRINGIFY_(a) # a |
299 | #define ECB_STRINGIFY(a) ECB_STRINGIFY_(a) |
306 | #define ECB_STRINGIFY(a) ECB_STRINGIFY_(a) |
300 | |
307 | |
301 | #define ecb_function_ ecb_inline |
308 | #define ecb_function_ ecb_inline |
302 | |
309 | |
303 | #if ECB_GCC_VERSION(3,1) |
310 | #if ECB_GCC_VERSION(3,1) || ECB_CLANG_VERSION(2,8) |
304 | #define ecb_attribute(attrlist) __attribute__(attrlist) |
311 | #define ecb_attribute(attrlist) __attribute__ (attrlist) |
|
|
312 | #else |
|
|
313 | #define ecb_attribute(attrlist) |
|
|
314 | #endif |
|
|
315 | |
|
|
316 | #if ECB_GCC_VERSION(3,1) || ECB_CLANG_BUILTIN(__builtin_constant_p) |
305 | #define ecb_is_constant(expr) __builtin_constant_p (expr) |
317 | #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 |
318 | #else |
309 | #define ecb_attribute(attrlist) |
|
|
310 | |
|
|
311 | /* possible C11 impl for integral types |
319 | /* possible C11 impl for integral types |
312 | typedef struct ecb_is_constant_struct ecb_is_constant_struct; |
320 | 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)) */ |
321 | #define ecb_is_constant(expr) _Generic ((1 ? (struct ecb_is_constant_struct *)0 : (void *)((expr) - (expr)), ecb_is_constant_struct *: 0, default: 1)) */ |
314 | |
322 | |
315 | #define ecb_is_constant(expr) 0 |
323 | #define ecb_is_constant(expr) 0 |
|
|
324 | #endif |
|
|
325 | |
|
|
326 | #if ECB_GCC_VERSION(3,1) || ECB_CLANG_BUILTIN(__builtin_expect) |
|
|
327 | #define ecb_expect(expr,value) __builtin_expect ((expr),(value)) |
|
|
328 | #else |
316 | #define ecb_expect(expr,value) (expr) |
329 | #define ecb_expect(expr,value) (expr) |
|
|
330 | #endif |
|
|
331 | |
|
|
332 | #if ECB_GCC_VERSION(3,1) || ECB_CLANG_BUILTIN(__builtin_prefetch) |
|
|
333 | #define ecb_prefetch(addr,rw,locality) __builtin_prefetch (addr, rw, locality) |
|
|
334 | #else |
317 | #define ecb_prefetch(addr,rw,locality) |
335 | #define ecb_prefetch(addr,rw,locality) |
318 | #endif |
336 | #endif |
319 | |
337 | |
320 | /* no emulation for ecb_decltype */ |
338 | /* no emulation for ecb_decltype */ |
321 | #if ECB_GCC_VERSION(4,5) |
339 | #if ECB_CPP11 |
|
|
340 | // older implementations might have problems with decltype(x)::type, work around it |
|
|
341 | template<class T> struct ecb_decltype_t { typedef T type; }; |
322 | #define ecb_decltype(x) __decltype(x) |
342 | #define ecb_decltype(x) ecb_decltype_t<decltype (x)>::type |
323 | #elif ECB_GCC_VERSION(3,0) |
343 | #elif ECB_GCC_VERSION(3,0) || ECB_CLANG_VERSION(2,8) |
324 | #define ecb_decltype(x) __typeof(x) |
344 | #define ecb_decltype(x) __typeof__ (x) |
|
|
345 | #endif |
|
|
346 | |
|
|
347 | #if _MSC_VER >= 1300 |
|
|
348 | #define ecb_deprecated __declspec (deprecated) |
|
|
349 | #else |
|
|
350 | #define ecb_deprecated ecb_attribute ((__deprecated__)) |
325 | #endif |
351 | #endif |
326 | |
352 | |
327 | #define ecb_noinline ecb_attribute ((__noinline__)) |
353 | #define ecb_noinline ecb_attribute ((__noinline__)) |
328 | #define ecb_unused ecb_attribute ((__unused__)) |
354 | #define ecb_unused ecb_attribute ((__unused__)) |
329 | #define ecb_const ecb_attribute ((__const__)) |
355 | #define ecb_const ecb_attribute ((__const__)) |
330 | #define ecb_pure ecb_attribute ((__pure__)) |
356 | #define ecb_pure ecb_attribute ((__pure__)) |
331 | |
357 | |
332 | #if ECB_C11 |
358 | /* TODO http://msdn.microsoft.com/en-us/library/k6ktzx3s.aspx __declspec(noreturn) */ |
|
|
359 | #if ECB_C11 || __IBMC_NORETURN |
|
|
360 | /* http://pic.dhe.ibm.com/infocenter/compbg/v121v141/topic/com.ibm.xlcpp121.bg.doc/language_ref/noreturn.html */ |
333 | #define ecb_noreturn _Noreturn |
361 | #define ecb_noreturn _Noreturn |
334 | #else |
362 | #else |
335 | #define ecb_noreturn ecb_attribute ((__noreturn__)) |
363 | #define ecb_noreturn ecb_attribute ((__noreturn__)) |
336 | #endif |
364 | #endif |
337 | |
365 | |
… | |
… | |
353 | /* for compatibility to the rest of the world */ |
381 | /* for compatibility to the rest of the world */ |
354 | #define ecb_likely(expr) ecb_expect_true (expr) |
382 | #define ecb_likely(expr) ecb_expect_true (expr) |
355 | #define ecb_unlikely(expr) ecb_expect_false (expr) |
383 | #define ecb_unlikely(expr) ecb_expect_false (expr) |
356 | |
384 | |
357 | /* count trailing zero bits and count # of one bits */ |
385 | /* count trailing zero bits and count # of one bits */ |
358 | #if ECB_GCC_VERSION(3,4) |
386 | #if ECB_GCC_VERSION(3,4) \ |
|
|
387 | || (ECB_CLANG_BUILTIN(__builtin_clz) && ECB_CLANG_BUILTIN(__builtin_clzll) \ |
|
|
388 | && ECB_CLANG_BUILTIN(__builtin_ctz) && ECB_CLANG_BUILTIN(__builtin_ctzll) \ |
|
|
389 | && ECB_CLANG_BUILTIN(__builtin_popcount)) |
359 | /* we assume int == 32 bit, long == 32 or 64 bit and long long == 64 bit */ |
390 | /* we assume int == 32 bit, long == 32 or 64 bit and long long == 64 bit */ |
360 | #define ecb_ld32(x) (__builtin_clz (x) ^ 31) |
391 | #define ecb_ld32(x) (__builtin_clz (x) ^ 31) |
361 | #define ecb_ld64(x) (__builtin_clzll (x) ^ 63) |
392 | #define ecb_ld64(x) (__builtin_clzll (x) ^ 63) |
362 | #define ecb_ctz32(x) __builtin_ctz (x) |
393 | #define ecb_ctz32(x) __builtin_ctz (x) |
363 | #define ecb_ctz64(x) __builtin_ctzll (x) |
394 | #define ecb_ctz64(x) __builtin_ctzll (x) |
… | |
… | |
494 | ecb_inline uint32_t ecb_rotl32 (uint32_t x, unsigned int count) { return (x >> (32 - count)) | (x << count); } |
525 | ecb_inline uint32_t ecb_rotl32 (uint32_t x, unsigned int count) { return (x >> (32 - count)) | (x << count); } |
495 | ecb_inline uint32_t ecb_rotr32 (uint32_t x, unsigned int count) { return (x << (32 - count)) | (x >> count); } |
526 | ecb_inline uint32_t ecb_rotr32 (uint32_t x, unsigned int count) { return (x << (32 - count)) | (x >> count); } |
496 | ecb_inline uint64_t ecb_rotl64 (uint64_t x, unsigned int count) { return (x >> (64 - count)) | (x << count); } |
527 | ecb_inline uint64_t ecb_rotl64 (uint64_t x, unsigned int count) { return (x >> (64 - count)) | (x << count); } |
497 | ecb_inline uint64_t ecb_rotr64 (uint64_t x, unsigned int count) { return (x << (64 - count)) | (x >> count); } |
528 | ecb_inline uint64_t ecb_rotr64 (uint64_t x, unsigned int count) { return (x << (64 - count)) | (x >> count); } |
498 | |
529 | |
499 | #if ECB_GCC_VERSION(4,3) |
530 | #if ECB_GCC_VERSION(4,3) || (ECB_CLANG_BUILTIN(__builtin_bswap32) && ECB_CLANG_BUILTIN(__builtin_bswap64)) |
500 | #define ecb_bswap16(x) (__builtin_bswap32 (x) >> 16) |
531 | #define ecb_bswap16(x) (__builtin_bswap32 (x) >> 16) |
501 | #define ecb_bswap32(x) __builtin_bswap32 (x) |
532 | #define ecb_bswap32(x) __builtin_bswap32 (x) |
502 | #define ecb_bswap64(x) __builtin_bswap64 (x) |
533 | #define ecb_bswap64(x) __builtin_bswap64 (x) |
503 | #else |
534 | #else |
504 | ecb_function_ uint16_t ecb_bswap16 (uint16_t x) ecb_const; |
535 | ecb_function_ uint16_t ecb_bswap16 (uint16_t x) ecb_const; |
… | |
… | |
521 | { |
552 | { |
522 | return (((uint64_t)ecb_bswap32 (x)) << 32) | ecb_bswap32 (x >> 32); |
553 | return (((uint64_t)ecb_bswap32 (x)) << 32) | ecb_bswap32 (x >> 32); |
523 | } |
554 | } |
524 | #endif |
555 | #endif |
525 | |
556 | |
526 | #if ECB_GCC_VERSION(4,5) |
557 | #if ECB_GCC_VERSION(4,5) || ECB_CLANG_BUILTIN(__builtin_unreachable) |
527 | #define ecb_unreachable() __builtin_unreachable () |
558 | #define ecb_unreachable() __builtin_unreachable () |
528 | #else |
559 | #else |
529 | /* this seems to work fine, but gcc always emits a warning for it :/ */ |
560 | /* this seems to work fine, but gcc always emits a warning for it :/ */ |
530 | ecb_inline void ecb_unreachable (void) ecb_noreturn; |
561 | ecb_inline void ecb_unreachable (void) ecb_noreturn; |
531 | ecb_inline void ecb_unreachable (void) { } |
562 | ecb_inline void ecb_unreachable (void) { } |
… | |
… | |
569 | #define ecb_mod(m,n) ((m) % (n) + ((m) % (n) < 0 ? (n) : 0)) |
600 | #define ecb_mod(m,n) ((m) % (n) + ((m) % (n) < 0 ? (n) : 0)) |
570 | #else |
601 | #else |
571 | #define ecb_mod(m,n) ((m) < 0 ? ((n) - 1 - ((-1 - (m)) % (n))) : ((m) % (n))) |
602 | #define ecb_mod(m,n) ((m) < 0 ? ((n) - 1 - ((-1 - (m)) % (n))) : ((m) % (n))) |
572 | #endif |
603 | #endif |
573 | |
604 | |
574 | #if __cplusplus |
605 | #if ECB_CPP |
575 | template<typename T> |
606 | template<typename T> |
576 | static inline T ecb_div_rd (T val, T div) |
607 | static inline T ecb_div_rd (T val, T div) |
577 | { |
608 | { |
578 | return val < 0 ? - ((-val + div - 1) / div) : (val ) / div; |
609 | return val < 0 ? - ((-val + div - 1) / div) : (val ) / div; |
579 | } |
610 | } |
… | |
… | |
639 | #define ECB_NAN NAN |
670 | #define ECB_NAN NAN |
640 | #else |
671 | #else |
641 | #define ECB_NAN ECB_INFINITY |
672 | #define ECB_NAN ECB_INFINITY |
642 | #endif |
673 | #endif |
643 | |
674 | |
|
|
675 | #if ECB_C99 || _XOPEN_VERSION >= 600 || _POSIX_VERSION >= 200112L |
|
|
676 | #define ecb_ldexpf(x,e) ldexpf (x, e) |
|
|
677 | #else |
|
|
678 | #define ecb_ldexpf(x,e) (float) ldexp ((double)(x), e) |
|
|
679 | #endif |
|
|
680 | |
644 | /* converts an ieee half/binary16 to a float */ |
681 | /* converts an ieee half/binary16 to a float */ |
645 | ecb_function_ float ecb_binary16_to_float (uint16_t x) ecb_const; |
682 | ecb_function_ float ecb_binary16_to_float (uint16_t x) ecb_const; |
646 | ecb_function_ float |
683 | ecb_function_ float |
647 | ecb_binary16_to_float (uint16_t x) |
684 | ecb_binary16_to_float (uint16_t x) |
648 | { |
685 | { |
649 | int e = (x >> 10) & 0x1f; |
686 | int e = (x >> 10) & 0x1f; |
650 | int m = x & 0x3ff; |
687 | int m = x & 0x3ff; |
651 | float r; |
688 | float r; |
652 | |
689 | |
653 | if (!e ) r = ldexpf (m , -24); |
690 | if (!e ) r = ecb_ldexpf (m , -24); |
654 | else if (e != 31) r = ldexpf (m + 0x400, e - 25); |
691 | else if (e != 31) r = ecb_ldexpf (m + 0x400, e - 25); |
655 | else if (m ) r = ECB_NAN; |
692 | else if (m ) r = ECB_NAN; |
656 | else r = ECB_INFINITY; |
693 | else r = ECB_INFINITY; |
657 | |
694 | |
658 | return x & 0x8000 ? -r : r; |
695 | return x & 0x8000 ? -r : r; |
659 | } |
696 | } |
… | |
… | |
718 | x |= 0x800000U; |
755 | x |= 0x800000U; |
719 | else |
756 | else |
720 | e = 1; |
757 | e = 1; |
721 | |
758 | |
722 | /* we distrust ldexpf a bit and do the 2**-24 scaling by an extra multiply */ |
759 | /* we distrust ldexpf a bit and do the 2**-24 scaling by an extra multiply */ |
723 | r = ldexpf (x * (0.5f / 0x800000U), e - 126); |
760 | r = ecb_ldexpf (x * (0.5f / 0x800000U), e - 126); |
724 | |
761 | |
725 | r = neg ? -r : r; |
762 | r = neg ? -r : r; |
726 | #endif |
763 | #endif |
727 | |
764 | |
728 | return r; |
765 | return r; |