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

Comparing libeio/ecb.h (file contents):
Revision 1.25 by root, Thu Jun 25 15:59:57 2015 UTC vs.
Revision 1.29 by root, Fri Aug 23 07:20:21 2019 UTC

… …
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 0x00010004 45#define ECB_VERSION 0x00010006
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;
… …
67 typedef uint32_t uintptr_t; 67 typedef uint32_t uintptr_t;
68 typedef int32_t intptr_t; 68 typedef int32_t intptr_t;
69 #endif 69 #endif
70#else 70#else
71 #include <inttypes.h> 71 #include <inttypes.h>
72 #if UINTMAX_MAX > 0xffffffffU 72 #if (defined INTPTR_MAX ? INTPTR_MAX : ULONG_MAX) > 0xffffffffU
73 #define ECB_PTRSIZE 8 73 #define ECB_PTRSIZE 8
74 #else 74 #else
75 #define ECB_PTRSIZE 4 75 #define ECB_PTRSIZE 4
76 #endif 76 #endif
77#endif 77#endif
… …
115 #define ECB_CLANG_EXTENSION(x) 0 115 #define ECB_CLANG_EXTENSION(x) 0
116#endif 116#endif
117 117
118#define ECB_CPP (__cplusplus+0) 118#define ECB_CPP (__cplusplus+0)
119#define ECB_CPP11 (__cplusplus >= 201103L) 119#define ECB_CPP11 (__cplusplus >= 201103L)
120#define ECB_CPP14 (__cplusplus >= 201402L)
121#define ECB_CPP17 (__cplusplus >= 201703L)
120 122
121#if ECB_CPP 123#if ECB_CPP
122 #define ECB_C 0 124 #define ECB_C 0
123 #define ECB_STDC_VERSION 0 125 #define ECB_STDC_VERSION 0
124#else 126#else
… …
126 #define ECB_STDC_VERSION __STDC_VERSION__ 128 #define ECB_STDC_VERSION __STDC_VERSION__
127#endif 129#endif
128 130
129#define ECB_C99 (ECB_STDC_VERSION >= 199901L) 131#define ECB_C99 (ECB_STDC_VERSION >= 199901L)
130#define ECB_C11 (ECB_STDC_VERSION >= 201112L) 132#define ECB_C11 (ECB_STDC_VERSION >= 201112L)
133#define ECB_C17 (ECB_STDC_VERSION >= 201710L)
131 134
132#if ECB_CPP 135#if ECB_CPP
133 #define ECB_EXTERN_C extern "C" 136 #define ECB_EXTERN_C extern "C"
134 #define ECB_EXTERN_C_BEG ECB_EXTERN_C { 137 #define ECB_EXTERN_C_BEG ECB_EXTERN_C {
135 #define ECB_EXTERN_C_END } 138 #define ECB_EXTERN_C_END }
… …
155/* http://www-01.ibm.com/support/knowledgecenter/SSGH3R_13.1.0/com.ibm.xlcpp131.aix.doc/compiler_ref/compiler_builtins.html */ 158/* 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 159#if __xlC__ && ECB_CPP
157 #include <builtins.h> 160 #include <builtins.h>
158#endif 161#endif
159 162
163#if 1400 <= _MSC_VER
164 #include <intrin.h> /* fence functions _ReadBarrier, also bit search functions _BitScanReverse */
165#endif
166
160#ifndef ECB_MEMORY_FENCE 167#ifndef ECB_MEMORY_FENCE
161 #if ECB_GCC_VERSION(2,5) || defined __INTEL_COMPILER || (__llvm__ && __GNUC__) || __SUNPRO_C >= 0x5110 || __SUNPRO_CC >= 0x5110 168 #if ECB_GCC_VERSION(2,5) || defined __INTEL_COMPILER || (__llvm__ && __GNUC__) || __SUNPRO_C >= 0x5110 || __SUNPRO_CC >= 0x5110
169 #define ECB_MEMORY_FENCE_RELAXED __asm__ __volatile__ ("" : : : "memory")
162 #if __i386 || __i386__ 170 #if __i386 || __i386__
163 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("lock; orb $0, -1(%%esp)" : : : "memory") 171 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("lock; orb $0, -1(%%esp)" : : : "memory")
164 #define ECB_MEMORY_FENCE_ACQUIRE __asm__ __volatile__ ("" : : : "memory") 172 #define ECB_MEMORY_FENCE_ACQUIRE __asm__ __volatile__ ("" : : : "memory")
165 #define ECB_MEMORY_FENCE_RELEASE __asm__ __volatile__ ("") 173 #define ECB_MEMORY_FENCE_RELEASE __asm__ __volatile__ ("" : : : "memory")
166 #elif ECB_GCC_AMD64 174 #elif ECB_GCC_AMD64
167 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("mfence" : : : "memory") 175 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("mfence" : : : "memory")
168 #define ECB_MEMORY_FENCE_ACQUIRE __asm__ __volatile__ ("" : : : "memory") 176 #define ECB_MEMORY_FENCE_ACQUIRE __asm__ __volatile__ ("" : : : "memory")
169 #define ECB_MEMORY_FENCE_RELEASE __asm__ __volatile__ ("") 177 #define ECB_MEMORY_FENCE_RELEASE __asm__ __volatile__ ("" : : : "memory")
170 #elif __powerpc__ || __ppc__ || __powerpc64__ || __ppc64__ 178 #elif __powerpc__ || __ppc__ || __powerpc64__ || __ppc64__
171 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("sync" : : : "memory") 179 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("sync" : : : "memory")
180 #elif defined __ARM_ARCH_2__ \
181 || defined __ARM_ARCH_3__ || defined __ARM_ARCH_3M__ \
182 || defined __ARM_ARCH_4__ || defined __ARM_ARCH_4T__ \
183 || defined __ARM_ARCH_5__ || defined __ARM_ARCH_5E__ \
184 || defined __ARM_ARCH_5T__ || defined __ARM_ARCH_5TE__ \
185 || defined __ARM_ARCH_5TEJ__
186 /* should not need any, unless running old code on newer cpu - arm doesn't support that */
172 #elif defined __ARM_ARCH_6__ || defined __ARM_ARCH_6J__ \ 187 #elif defined __ARM_ARCH_6__ || defined __ARM_ARCH_6J__ \
173 || defined __ARM_ARCH_6K__ || defined __ARM_ARCH_6ZK__ 188 || defined __ARM_ARCH_6K__ || defined __ARM_ARCH_6ZK__ \
189 || defined __ARM_ARCH_6T2__
174 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("mcr p15,0,%0,c7,c10,5" : : "r" (0) : "memory") 190 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("mcr p15,0,%0,c7,c10,5" : : "r" (0) : "memory")
175 #elif defined __ARM_ARCH_7__ || defined __ARM_ARCH_7A__ \ 191 #elif defined __ARM_ARCH_7__ || defined __ARM_ARCH_7A__ \
176 || defined __ARM_ARCH_7M__ || defined __ARM_ARCH_7R__ 192 || defined __ARM_ARCH_7R__ || defined __ARM_ARCH_7M__
177 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("dmb" : : : "memory") 193 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("dmb" : : : "memory")
178 #elif __aarch64__ 194 #elif __aarch64__
179 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("dmb ish" : : : "memory") 195 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("dmb ish" : : : "memory")
180 #elif (__sparc || __sparc__) && !__sparcv8 196 #elif (__sparc || __sparc__) && !(__sparc_v8__ || defined __sparcv8)
181 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("membar #LoadStore | #LoadLoad | #StoreStore | #StoreLoad" : : : "memory") 197 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("membar #LoadStore | #LoadLoad | #StoreStore | #StoreLoad" : : : "memory")
182 #define ECB_MEMORY_FENCE_ACQUIRE __asm__ __volatile__ ("membar #LoadStore | #LoadLoad" : : : "memory") 198 #define ECB_MEMORY_FENCE_ACQUIRE __asm__ __volatile__ ("membar #LoadStore | #LoadLoad" : : : "memory")
183 #define ECB_MEMORY_FENCE_RELEASE __asm__ __volatile__ ("membar #LoadStore | #StoreStore") 199 #define ECB_MEMORY_FENCE_RELEASE __asm__ __volatile__ ("membar #LoadStore | #StoreStore")
184 #elif defined __s390__ || defined __s390x__ 200 #elif defined __s390__ || defined __s390x__
185 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("bcr 15,0" : : : "memory") 201 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("bcr 15,0" : : : "memory")
… …
208 #if ECB_GCC_VERSION(4,7) 224 #if ECB_GCC_VERSION(4,7)
209 /* see comment below (stdatomic.h) about the C11 memory model. */ 225 /* see comment below (stdatomic.h) about the C11 memory model. */
210 #define ECB_MEMORY_FENCE __atomic_thread_fence (__ATOMIC_SEQ_CST) 226 #define ECB_MEMORY_FENCE __atomic_thread_fence (__ATOMIC_SEQ_CST)
211 #define ECB_MEMORY_FENCE_ACQUIRE __atomic_thread_fence (__ATOMIC_ACQUIRE) 227 #define ECB_MEMORY_FENCE_ACQUIRE __atomic_thread_fence (__ATOMIC_ACQUIRE)
212 #define ECB_MEMORY_FENCE_RELEASE __atomic_thread_fence (__ATOMIC_RELEASE) 228 #define ECB_MEMORY_FENCE_RELEASE __atomic_thread_fence (__ATOMIC_RELEASE)
229 #define ECB_MEMORY_FENCE_RELAXED __atomic_thread_fence (__ATOMIC_RELAXED)
213 230
214 #elif ECB_CLANG_EXTENSION(c_atomic) 231 #elif ECB_CLANG_EXTENSION(c_atomic)
215 /* see comment below (stdatomic.h) about the C11 memory model. */ 232 /* see comment below (stdatomic.h) about the C11 memory model. */
216 #define ECB_MEMORY_FENCE __c11_atomic_thread_fence (__ATOMIC_SEQ_CST) 233 #define ECB_MEMORY_FENCE __c11_atomic_thread_fence (__ATOMIC_SEQ_CST)
217 #define ECB_MEMORY_FENCE_ACQUIRE __c11_atomic_thread_fence (__ATOMIC_ACQUIRE) 234 #define ECB_MEMORY_FENCE_ACQUIRE __c11_atomic_thread_fence (__ATOMIC_ACQUIRE)
218 #define ECB_MEMORY_FENCE_RELEASE __c11_atomic_thread_fence (__ATOMIC_RELEASE) 235 #define ECB_MEMORY_FENCE_RELEASE __c11_atomic_thread_fence (__ATOMIC_RELEASE)
236 #define ECB_MEMORY_FENCE_RELAXED __c11_atomic_thread_fence (__ATOMIC_RELAXED)
219 237
220 #elif ECB_GCC_VERSION(4,4) || defined __INTEL_COMPILER || defined __clang__ 238 #elif ECB_GCC_VERSION(4,4) || defined __INTEL_COMPILER || defined __clang__
221 #define ECB_MEMORY_FENCE __sync_synchronize () 239 #define ECB_MEMORY_FENCE __sync_synchronize ()
222 #elif _MSC_VER >= 1500 /* VC++ 2008 */ 240 #elif _MSC_VER >= 1500 /* VC++ 2008 */
223 /* apparently, microsoft broke all the memory barrier stuff in Visual Studio 2008... */ 241 /* apparently, microsoft broke all the memory barrier stuff in Visual Studio 2008... */
… …
233 #elif defined _WIN32 251 #elif defined _WIN32
234 #include <WinNT.h> 252 #include <WinNT.h>
235 #define ECB_MEMORY_FENCE MemoryBarrier () /* actually just xchg on x86... scary */ 253 #define ECB_MEMORY_FENCE MemoryBarrier () /* actually just xchg on x86... scary */
236 #elif __SUNPRO_C >= 0x5110 || __SUNPRO_CC >= 0x5110 254 #elif __SUNPRO_C >= 0x5110 || __SUNPRO_CC >= 0x5110
237 #include <mbarrier.h> 255 #include <mbarrier.h>
238 #define ECB_MEMORY_FENCE __machine_rw_barrier () 256 #define ECB_MEMORY_FENCE __machine_rw_barrier ()
239 #define ECB_MEMORY_FENCE_ACQUIRE __machine_r_barrier () 257 #define ECB_MEMORY_FENCE_ACQUIRE __machine_acq_barrier ()
240 #define ECB_MEMORY_FENCE_RELEASE __machine_w_barrier () 258 #define ECB_MEMORY_FENCE_RELEASE __machine_rel_barrier ()
259 #define ECB_MEMORY_FENCE_RELAXED __compiler_barrier ()
241 #elif __xlC__ 260 #elif __xlC__
242 #define ECB_MEMORY_FENCE __sync () 261 #define ECB_MEMORY_FENCE __sync ()
243 #endif 262 #endif
244#endif 263#endif
245 264
246#ifndef ECB_MEMORY_FENCE 265#ifndef ECB_MEMORY_FENCE
247 #if ECB_C11 && !defined __STDC_NO_ATOMICS__ 266 #if ECB_C11 && !defined __STDC_NO_ATOMICS__
248 /* we assume that these memory fences work on all variables/all memory accesses, */ 267 /* we assume that these memory fences work on all variables/all memory accesses, */
249 /* not just C11 atomics and atomic accesses */ 268 /* not just C11 atomics and atomic accesses */
250 #include <stdatomic.h> 269 #include <stdatomic.h>
251 /* Unfortunately, neither gcc 4.7 nor clang 3.1 generate any instructions for */
252 /* any fence other than seq_cst, which isn't very efficient for us. */
253 /* Why that is, we don't know - either the C11 memory model is quite useless */
254 /* for most usages, or gcc and clang have a bug */
255 /* I *currently* lean towards the latter, and inefficiently implement */
256 /* all three of ecb's fences as a seq_cst fence */
257 /* Update, gcc-4.8 generates mfence for all c++ fences, but nothing */
258 /* for all __atomic_thread_fence's except seq_cst */
259 #define ECB_MEMORY_FENCE atomic_thread_fence (memory_order_seq_cst) 270 #define ECB_MEMORY_FENCE atomic_thread_fence (memory_order_seq_cst)
271 #define ECB_MEMORY_FENCE_ACQUIRE atomic_thread_fence (memory_order_acquire)
272 #define ECB_MEMORY_FENCE_RELEASE atomic_thread_fence (memory_order_release)
260 #endif 273 #endif
261#endif 274#endif
262 275
263#ifndef ECB_MEMORY_FENCE 276#ifndef ECB_MEMORY_FENCE
264 #if !ECB_AVOID_PTHREADS 277 #if !ECB_AVOID_PTHREADS
… …
284 297
285#if !defined ECB_MEMORY_FENCE_RELEASE && defined ECB_MEMORY_FENCE 298#if !defined ECB_MEMORY_FENCE_RELEASE && defined ECB_MEMORY_FENCE
286 #define ECB_MEMORY_FENCE_RELEASE ECB_MEMORY_FENCE 299 #define ECB_MEMORY_FENCE_RELEASE ECB_MEMORY_FENCE
287#endif 300#endif
288 301
302#if !defined ECB_MEMORY_FENCE_RELAXED && defined ECB_MEMORY_FENCE
303 #define ECB_MEMORY_FENCE_RELAXED ECB_MEMORY_FENCE /* very heavy-handed */
304#endif
305
289/*****************************************************************************/ 306/*****************************************************************************/
290 307
291#if ECB_CPP 308#if ECB_CPP
292 #define ecb_inline static inline 309 #define ecb_inline static inline
293#elif ECB_GCC_VERSION(2,5) 310#elif ECB_GCC_VERSION(2,5)
… …
423#else 440#else
424 ecb_function_ ecb_const int ecb_ctz32 (uint32_t x); 441 ecb_function_ ecb_const int ecb_ctz32 (uint32_t x);
425 ecb_function_ ecb_const int 442 ecb_function_ ecb_const int
426 ecb_ctz32 (uint32_t x) 443 ecb_ctz32 (uint32_t x)
427 { 444 {
445#if 1400 <= _MSC_VER && (_M_IX86 || _M_X64 || _M_IA64 || _M_ARM)
446 unsigned long r;
447 _BitScanForward (&r, x);
448 return (int)r;
449#else
428 int r = 0; 450 int r = 0;
429 451
430 x &= ~x + 1; /* this isolates the lowest bit */ 452 x &= ~x + 1; /* this isolates the lowest bit */
431 453
432#if ECB_branchless_on_i386 454#if ECB_branchless_on_i386
… …
442 if (x & 0xff00ff00) r += 8; 464 if (x & 0xff00ff00) r += 8;
443 if (x & 0xffff0000) r += 16; 465 if (x & 0xffff0000) r += 16;
444#endif 466#endif
445 467
446 return r; 468 return r;
469#endif
447 } 470 }
448 471
449 ecb_function_ ecb_const int ecb_ctz64 (uint64_t x); 472 ecb_function_ ecb_const int ecb_ctz64 (uint64_t x);
450 ecb_function_ ecb_const int 473 ecb_function_ ecb_const int
451 ecb_ctz64 (uint64_t x) 474 ecb_ctz64 (uint64_t x)
452 { 475 {
476#if 1400 <= _MSC_VER && (_M_X64 || _M_IA64 || _M_ARM)
477 unsigned long r;
478 _BitScanForward64 (&r, x);
479 return (int)r;
480#else
453 int shift = x & 0xffffffffU ? 0 : 32; 481 int shift = x & 0xffffffff ? 0 : 32;
454 return ecb_ctz32 (x >> shift) + shift; 482 return ecb_ctz32 (x >> shift) + shift;
483#endif
455 } 484 }
456 485
457 ecb_function_ ecb_const int ecb_popcount32 (uint32_t x); 486 ecb_function_ ecb_const int ecb_popcount32 (uint32_t x);
458 ecb_function_ ecb_const int 487 ecb_function_ ecb_const int
459 ecb_popcount32 (uint32_t x) 488 ecb_popcount32 (uint32_t x)
… …
467 } 496 }
468 497
469 ecb_function_ ecb_const int ecb_ld32 (uint32_t x); 498 ecb_function_ ecb_const int ecb_ld32 (uint32_t x);
470 ecb_function_ ecb_const int ecb_ld32 (uint32_t x) 499 ecb_function_ ecb_const int ecb_ld32 (uint32_t x)
471 { 500 {
501#if 1400 <= _MSC_VER && (_M_IX86 || _M_X64 || _M_IA64 || _M_ARM)
502 unsigned long r;
503 _BitScanReverse (&r, x);
504 return (int)r;
505#else
472 int r = 0; 506 int r = 0;
473 507
474 if (x >> 16) { x >>= 16; r += 16; } 508 if (x >> 16) { x >>= 16; r += 16; }
475 if (x >> 8) { x >>= 8; r += 8; } 509 if (x >> 8) { x >>= 8; r += 8; }
476 if (x >> 4) { x >>= 4; r += 4; } 510 if (x >> 4) { x >>= 4; r += 4; }
477 if (x >> 2) { x >>= 2; r += 2; } 511 if (x >> 2) { x >>= 2; r += 2; }
478 if (x >> 1) { r += 1; } 512 if (x >> 1) { r += 1; }
479 513
480 return r; 514 return r;
515#endif
481 } 516 }
482 517
483 ecb_function_ ecb_const int ecb_ld64 (uint64_t x); 518 ecb_function_ ecb_const int ecb_ld64 (uint64_t x);
484 ecb_function_ ecb_const int ecb_ld64 (uint64_t x) 519 ecb_function_ ecb_const int ecb_ld64 (uint64_t x)
485 { 520 {
521#if 1400 <= _MSC_VER && (_M_X64 || _M_IA64 || _M_ARM)
522 unsigned long r;
523 _BitScanReverse64 (&r, x);
524 return (int)r;
525#else
486 int r = 0; 526 int r = 0;
487 527
488 if (x >> 32) { x >>= 32; r += 32; } 528 if (x >> 32) { x >>= 32; r += 32; }
489 529
490 return r + ecb_ld32 (x); 530 return r + ecb_ld32 (x);
531#endif
491 } 532 }
492#endif 533#endif
493 534
494ecb_function_ ecb_const ecb_bool ecb_is_pot32 (uint32_t x); 535ecb_function_ ecb_const ecb_bool ecb_is_pot32 (uint32_t x);
495ecb_function_ ecb_const ecb_bool ecb_is_pot32 (uint32_t x) { return !(x & (x - 1)); } 536ecb_function_ ecb_const ecb_bool ecb_is_pot32 (uint32_t x) { return !(x & (x - 1)); }
… …
598#endif 639#endif
599 640
600/* try to tell the compiler that some condition is definitely true */ 641/* try to tell the compiler that some condition is definitely true */
601#define ecb_assume(cond) if (!(cond)) ecb_unreachable (); else 0 642#define ecb_assume(cond) if (!(cond)) ecb_unreachable (); else 0
602 643
603ecb_inline ecb_const unsigned char ecb_byteorder_helper (void); 644ecb_inline ecb_const uint32_t ecb_byteorder_helper (void);
604ecb_inline ecb_const unsigned char 645ecb_inline ecb_const uint32_t
605ecb_byteorder_helper (void) 646ecb_byteorder_helper (void)
606{ 647{
607 /* the union code still generates code under pressure in gcc, */ 648 /* the union code still generates code under pressure in gcc, */
608 /* but less than using pointers, and always seems to */ 649 /* but less than using pointers, and always seems to */
609 /* successfully return a constant. */ 650 /* successfully return a constant. */
610 /* the reason why we have this horrible preprocessor mess */ 651 /* the reason why we have this horrible preprocessor mess */
611 /* is to avoid it in all cases, at least on common architectures */ 652 /* is to avoid it in all cases, at least on common architectures */
612 /* or when using a recent enough gcc version (>= 4.6) */ 653 /* or when using a recent enough gcc version (>= 4.6) */
613#if ((__i386 || __i386__) && !__VOS__) || _M_IX86 || ECB_GCC_AMD64 || ECB_MSVC_AMD64
614 return 0x44;
615#elif __BYTE_ORDER__ && __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__ 654#if (defined __BYTE_ORDER__ && __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__) \
655 || ((__i386 || __i386__ || _M_IX86 || ECB_GCC_AMD64 || ECB_MSVC_AMD64) && !__VOS__)
656 #define ECB_LITTLE_ENDIAN 1
616 return 0x44; 657 return 0x44332211;
617#elif __BYTE_ORDER__ && __BYTE_ORDER__ == __ORDER_BIG_ENDIAN__ 658#elif (defined __BYTE_ORDER__ && __BYTE_ORDER__ == __ORDER_BIG_ENDIAN__) \
659 || ((__AARCH64EB__ || __MIPSEB__ || __ARMEB__) && !__VOS__)
660 #define ECB_BIG_ENDIAN 1
618 return 0x11; 661 return 0x11223344;
619#else 662#else
620 union 663 union
621 { 664 {
665 uint8_t c[4];
622 uint32_t i; 666 uint32_t u;
623 uint8_t c;
624 } u = { 0x11223344 }; 667 } u = { 0x11, 0x22, 0x33, 0x44 };
625 return u.c; 668 return u.u;
626#endif 669#endif
627} 670}
628 671
629ecb_inline ecb_const ecb_bool ecb_big_endian (void); 672ecb_inline ecb_const ecb_bool ecb_big_endian (void);
630ecb_inline ecb_const ecb_bool ecb_big_endian (void) { return ecb_byteorder_helper () == 0x11; } 673ecb_inline ecb_const ecb_bool ecb_big_endian (void) { return ecb_byteorder_helper () == 0x11223344; }
631ecb_inline ecb_const ecb_bool ecb_little_endian (void); 674ecb_inline ecb_const ecb_bool ecb_little_endian (void);
632ecb_inline ecb_const ecb_bool ecb_little_endian (void) { return ecb_byteorder_helper () == 0x44; } 675ecb_inline ecb_const ecb_bool ecb_little_endian (void) { return ecb_byteorder_helper () == 0x44332211; }
633 676
634#if ECB_GCC_VERSION(3,0) || ECB_C99 677#if ECB_GCC_VERSION(3,0) || ECB_C99
635 #define ecb_mod(m,n) ((m) % (n) + ((m) % (n) < 0 ? (n) : 0)) 678 #define ecb_mod(m,n) ((m) % (n) + ((m) % (n) < 0 ? (n) : 0))
636#else 679#else
637 #define ecb_mod(m,n) ((m) < 0 ? ((n) - 1 - ((-1 - (m)) % (n))) : ((m) % (n))) 680 #define ecb_mod(m,n) ((m) < 0 ? ((n) - 1 - ((-1 - (m)) % (n))) : ((m) % (n)))
… …
661 return N; 704 return N;
662 } 705 }
663#else 706#else
664 #define ecb_array_length(name) (sizeof (name) / sizeof (name [0])) 707 #define ecb_array_length(name) (sizeof (name) / sizeof (name [0]))
665#endif 708#endif
709
710ecb_function_ ecb_const uint32_t ecb_binary16_to_binary32 (uint32_t x);
711ecb_function_ ecb_const uint32_t
712ecb_binary16_to_binary32 (uint32_t x)
713{
714 unsigned int s = (x & 0x8000) << (31 - 15);
715 int e = (x >> 10) & 0x001f;
716 unsigned int m = x & 0x03ff;
717
718 if (ecb_expect_false (e == 31))
719 /* infinity or NaN */
720 e = 255 - (127 - 15);
721 else if (ecb_expect_false (!e))
722 {
723 if (ecb_expect_true (!m))
724 /* zero, handled by code below by forcing e to 0 */
725 e = 0 - (127 - 15);
726 else
727 {
728 /* subnormal, renormalise */
729 unsigned int s = 10 - ecb_ld32 (m);
730
731 m = (m << s) & 0x3ff; /* mask implicit bit */
732 e -= s - 1;
733 }
734 }
735
736 /* e and m now are normalised, or zero, (or inf or nan) */
737 e += 127 - 15;
738
739 return s | (e << 23) | (m << (23 - 10));
740}
741
742ecb_function_ ecb_const uint16_t ecb_binary32_to_binary16 (uint32_t x);
743ecb_function_ ecb_const uint16_t
744ecb_binary32_to_binary16 (uint32_t x)
745{
746 unsigned int s = (x >> 16) & 0x00008000; /* sign bit, the easy part */
747 unsigned int e = ((x >> 23) & 0x000000ff) - (127 - 15); /* the desired exponent */
748 unsigned int m = x & 0x007fffff;
749
750 x &= 0x7fffffff;
751
752 /* if it's within range of binary16 normals, use fast path */
753 if (ecb_expect_true (0x38800000 <= x && x <= 0x477fefff))
754 {
755 /* mantissa round-to-even */
756 m += 0x00000fff + ((m >> (23 - 10)) & 1);
757
758 /* handle overflow */
759 if (ecb_expect_false (m >= 0x00800000))
760 {
761 m >>= 1;
762 e += 1;
763 }
764
765 return s | (e << 10) | (m >> (23 - 10));
766 }
767
768 /* handle large numbers and infinity */
769 if (ecb_expect_true (0x477fefff < x && x <= 0x7f800000))
770 return s | 0x7c00;
771
772 /* handle zero, subnormals and small numbers */
773 if (ecb_expect_true (x < 0x38800000))
774 {
775 /* zero */
776 if (ecb_expect_true (!x))
777 return s;
778
779 /* handle subnormals */
780
781 /* too small, will be zero */
782 if (e < (14 - 24)) /* might not be sharp, but is good enough */
783 return s;
784
785 m |= 0x00800000; /* make implicit bit explicit */
786
787 /* very tricky - we need to round to the nearest e (+10) bit value */
788 {
789 unsigned int bits = 14 - e;
790 unsigned int half = (1 << (bits - 1)) - 1;
791 unsigned int even = (m >> bits) & 1;
792
793 /* if this overflows, we will end up with a normalised number */
794 m = (m + half + even) >> bits;
795 }
796
797 return s | m;
798 }
799
800 /* handle NaNs, preserve leftmost nan bits, but make sure we don't turn them into infinities */
801 m >>= 13;
802
803 return s | 0x7c00 | m | !m;
804}
666 805
667/*******************************************************************************/ 806/*******************************************************************************/
668/* floating point stuff, can be disabled by defining ECB_NO_LIBM */ 807/* floating point stuff, can be disabled by defining ECB_NO_LIBM */
669 808
670/* basically, everything uses "ieee pure-endian" floating point numbers */ 809/* basically, everything uses "ieee pure-endian" floating point numbers */
… …
713 #else 852 #else
714 #define ecb_ldexpf(x,e) (float) ldexp ((double) (x), (e)) 853 #define ecb_ldexpf(x,e) (float) ldexp ((double) (x), (e))
715 #define ecb_frexpf(x,e) (float) frexp ((double) (x), (e)) 854 #define ecb_frexpf(x,e) (float) frexp ((double) (x), (e))
716 #endif 855 #endif
717 856
718 /* converts an ieee half/binary16 to a float */
719 ecb_function_ ecb_const float ecb_binary16_to_float (uint16_t x);
720 ecb_function_ ecb_const float
721 ecb_binary16_to_float (uint16_t x)
722 {
723 int e = (x >> 10) & 0x1f;
724 int m = x & 0x3ff;
725 float r;
726
727 if (!e ) r = ecb_ldexpf (m , -24);
728 else if (e != 31) r = ecb_ldexpf (m + 0x400, e - 25);
729 else if (m ) r = ECB_NAN;
730 else r = ECB_INFINITY;
731
732 return x & 0x8000 ? -r : r;
733 }
734
735 /* convert a float to ieee single/binary32 */ 857 /* convert a float to ieee single/binary32 */
736 ecb_function_ ecb_const uint32_t ecb_float_to_binary32 (float x); 858 ecb_function_ ecb_const uint32_t ecb_float_to_binary32 (float x);
737 ecb_function_ ecb_const uint32_t 859 ecb_function_ ecb_const uint32_t
738 ecb_float_to_binary32 (float x) 860 ecb_float_to_binary32 (float x)
739 { 861 {
… …
870 #endif 992 #endif
871 993
872 return r; 994 return r;
873 } 995 }
874 996
875#endif 997 /* convert a float to ieee half/binary16 */
998 ecb_function_ ecb_const uint16_t ecb_float_to_binary16 (float x);
999 ecb_function_ ecb_const uint16_t
1000 ecb_float_to_binary16 (float x)
1001 {
1002 return ecb_binary32_to_binary16 (ecb_float_to_binary32 (x));
1003 }
876 1004
877#endif 1005 /* convert an ieee half/binary16 to float */
1006 ecb_function_ ecb_const float ecb_binary16_to_float (uint16_t x);
1007 ecb_function_ ecb_const float
1008 ecb_binary16_to_float (uint16_t x)
1009 {
1010 return ecb_binary32_to_float (ecb_binary16_to_binary32 (x));
1011 }
878 1012
1013#endif
1014
1015#endif
1016

Diff Legend

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