1 | /* |
1 | /* |
2 | * libecb - http://software.schmorp.de/pkg/libecb |
2 | * libecb - http://software.schmorp.de/pkg/libecb |
3 | * |
3 | * |
4 | * Copyright (©) 2009-2012 Marc Alexander Lehmann <libecb@schmorp.de> |
4 | * Copyright (©) 2009-2013 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: |
… | |
… | |
29 | |
29 | |
30 | #ifndef ECB_H |
30 | #ifndef ECB_H |
31 | #define ECB_H |
31 | #define ECB_H |
32 | |
32 | |
33 | /* 16 bits major, 16 bits minor */ |
33 | /* 16 bits major, 16 bits minor */ |
34 | #define ECB_VERSION 0x00010001 |
34 | #define ECB_VERSION 0x00010003 |
35 | |
35 | |
36 | #ifdef _WIN32 |
36 | #ifdef _WIN32 |
37 | typedef signed char int8_t; |
37 | typedef signed char int8_t; |
38 | typedef unsigned char uint8_t; |
38 | typedef unsigned char uint8_t; |
39 | typedef signed short int16_t; |
39 | typedef signed short int16_t; |
… | |
… | |
60 | #include <inttypes.h> |
60 | #include <inttypes.h> |
61 | #if UINTMAX_MAX > 0xffffffffU |
61 | #if UINTMAX_MAX > 0xffffffffU |
62 | #define ECB_PTRSIZE 8 |
62 | #define ECB_PTRSIZE 8 |
63 | #else |
63 | #else |
64 | #define ECB_PTRSIZE 4 |
64 | #define ECB_PTRSIZE 4 |
|
|
65 | #endif |
|
|
66 | #endif |
|
|
67 | |
|
|
68 | /* work around x32 idiocy by defining proper macros */ |
|
|
69 | #if __x86_64 || _M_AMD64 |
|
|
70 | #if _ILP32 |
|
|
71 | #define ECB_AMD64_X32 1 |
|
|
72 | #else |
|
|
73 | #define ECB_AMD64 1 |
65 | #endif |
74 | #endif |
66 | #endif |
75 | #endif |
67 | |
76 | |
68 | /* many compilers define _GNUC_ to some versions but then only implement |
77 | /* many compilers define _GNUC_ to some versions but then only implement |
69 | * what their idiot authors think are the "more important" extensions, |
78 | * what their idiot authors think are the "more important" extensions, |
… | |
… | |
125 | || defined __ARM_ARCH_6K__ || defined __ARM_ARCH_6ZK__ |
134 | || defined __ARM_ARCH_6K__ || defined __ARM_ARCH_6ZK__ |
126 | #define ECB_MEMORY_FENCE __asm__ __volatile__ ("mcr p15,0,%0,c7,c10,5" : : "r" (0) : "memory") |
135 | #define ECB_MEMORY_FENCE __asm__ __volatile__ ("mcr p15,0,%0,c7,c10,5" : : "r" (0) : "memory") |
127 | #elif defined __ARM_ARCH_7__ || defined __ARM_ARCH_7A__ \ |
136 | #elif defined __ARM_ARCH_7__ || defined __ARM_ARCH_7A__ \ |
128 | || defined __ARM_ARCH_7M__ || defined __ARM_ARCH_7R__ |
137 | || defined __ARM_ARCH_7M__ || defined __ARM_ARCH_7R__ |
129 | #define ECB_MEMORY_FENCE __asm__ __volatile__ ("dmb" : : : "memory") |
138 | #define ECB_MEMORY_FENCE __asm__ __volatile__ ("dmb" : : : "memory") |
130 | #elif __sparc || __sparc__ |
139 | #elif (__sparc || __sparc__) && !__sparcv8 |
131 | #define ECB_MEMORY_FENCE __asm__ __volatile__ ("membar #LoadStore | #LoadLoad | #StoreStore | #StoreLoad" : : : "memory") |
140 | #define ECB_MEMORY_FENCE __asm__ __volatile__ ("membar #LoadStore | #LoadLoad | #StoreStore | #StoreLoad" : : : "memory") |
132 | #define ECB_MEMORY_FENCE_ACQUIRE __asm__ __volatile__ ("membar #LoadStore | #LoadLoad" : : : "memory") |
141 | #define ECB_MEMORY_FENCE_ACQUIRE __asm__ __volatile__ ("membar #LoadStore | #LoadLoad" : : : "memory") |
133 | #define ECB_MEMORY_FENCE_RELEASE __asm__ __volatile__ ("membar #LoadStore | #StoreStore") |
142 | #define ECB_MEMORY_FENCE_RELEASE __asm__ __volatile__ ("membar #LoadStore | #StoreStore") |
134 | #elif defined __s390__ || defined __s390x__ |
143 | #elif defined __s390__ || defined __s390x__ |
135 | #define ECB_MEMORY_FENCE __asm__ __volatile__ ("bcr 15,0" : : : "memory") |
144 | #define ECB_MEMORY_FENCE __asm__ __volatile__ ("bcr 15,0" : : : "memory") |
136 | #elif defined __mips__ |
145 | #elif defined __mips__ |
|
|
146 | /* GNU/Linux emulates sync on mips1 architectures, so we force its use */ |
|
|
147 | /* anybody else who still uses mips1 is supposed to send in their version, with detection code. */ |
137 | #define ECB_MEMORY_FENCE __asm__ __volatile__ ("sync" : : : "memory") |
148 | #define ECB_MEMORY_FENCE __asm__ __volatile__ (".set mips2; sync; .set mips0" : : : "memory") |
138 | #elif defined __alpha__ |
149 | #elif defined __alpha__ |
139 | #define ECB_MEMORY_FENCE __asm__ __volatile__ ("mb" : : : "memory") |
150 | #define ECB_MEMORY_FENCE __asm__ __volatile__ ("mb" : : : "memory") |
140 | #elif defined __hppa__ |
151 | #elif defined __hppa__ |
141 | #define ECB_MEMORY_FENCE __asm__ __volatile__ ("" : : : "memory") |
152 | #define ECB_MEMORY_FENCE __asm__ __volatile__ ("" : : : "memory") |
142 | #define ECB_MEMORY_FENCE_RELEASE __asm__ __volatile__ ("") |
153 | #define ECB_MEMORY_FENCE_RELEASE __asm__ __volatile__ ("") |
143 | #elif defined __ia64__ |
154 | #elif defined __ia64__ |
144 | #define ECB_MEMORY_FENCE __asm__ __volatile__ ("mf" : : : "memory") |
155 | #define ECB_MEMORY_FENCE __asm__ __volatile__ ("mf" : : : "memory") |
|
|
156 | #elif defined __m68k__ |
|
|
157 | #define ECB_MEMORY_FENCE __asm__ __volatile__ ("" : : : "memory") |
|
|
158 | #elif defined __m88k__ |
|
|
159 | #define ECB_MEMORY_FENCE __asm__ __volatile__ ("tb1 0,%%r0,128" : : : "memory") |
|
|
160 | #elif defined __sh__ |
|
|
161 | #define ECB_MEMORY_FENCE __asm__ __volatile__ ("" : : : "memory") |
145 | #endif |
162 | #endif |
146 | #endif |
163 | #endif |
147 | #endif |
164 | #endif |
148 | |
165 | |
149 | #ifndef ECB_MEMORY_FENCE |
166 | #ifndef ECB_MEMORY_FENCE |
150 | #if ECB_GCC_VERSION(4,7) |
167 | #if ECB_GCC_VERSION(4,7) |
151 | /* see comment below (stdatomic.h) about the C11 memory model. */ |
168 | /* see comment below (stdatomic.h) about the C11 memory model. */ |
152 | #define ECB_MEMORY_FENCE __atomic_thread_fence (__ATOMIC_SEQ_CST) |
169 | #define ECB_MEMORY_FENCE __atomic_thread_fence (__ATOMIC_SEQ_CST) |
|
|
170 | |
|
|
171 | /* The __has_feature syntax from clang is so misdesigned that we cannot use it |
|
|
172 | * without risking compile time errors with other compilers. We *could* |
|
|
173 | * define our own ecb_clang_has_feature, but I just can't be bothered to work |
|
|
174 | * around this shit time and again. |
153 | /*#elif defined __clang && __has_feature (cxx_atomic)*/ |
175 | * #elif defined __clang && __has_feature (cxx_atomic) |
154 | /* see comment below (stdatomic.h) about the C11 memory model. */ |
176 | * // see comment below (stdatomic.h) about the C11 memory model. |
155 | /*#define ECB_MEMORY_FENCE __c11_atomic_thread_fence (__ATOMIC_SEQ_CST)*/ |
177 | * #define ECB_MEMORY_FENCE __c11_atomic_thread_fence (__ATOMIC_SEQ_CST) |
|
|
178 | */ |
|
|
179 | |
156 | #elif ECB_GCC_VERSION(4,4) || defined __INTEL_COMPILER || defined __clang__ |
180 | #elif ECB_GCC_VERSION(4,4) || defined __INTEL_COMPILER || defined __clang__ |
157 | #define ECB_MEMORY_FENCE __sync_synchronize () |
181 | #define ECB_MEMORY_FENCE __sync_synchronize () |
158 | #elif _MSC_VER >= 1400 /* VC++ 2005 */ |
182 | #elif _MSC_VER >= 1400 /* VC++ 2005 */ |
159 | #pragma intrinsic(_ReadBarrier,_WriteBarrier,_ReadWriteBarrier) |
183 | #pragma intrinsic(_ReadBarrier,_WriteBarrier,_ReadWriteBarrier) |
160 | #define ECB_MEMORY_FENCE _ReadWriteBarrier () |
184 | #define ECB_MEMORY_FENCE _ReadWriteBarrier () |
… | |
… | |
475 | ecb_inline unsigned char ecb_byteorder_helper (void) ecb_const; |
499 | ecb_inline unsigned char ecb_byteorder_helper (void) ecb_const; |
476 | ecb_inline unsigned char |
500 | ecb_inline unsigned char |
477 | ecb_byteorder_helper (void) |
501 | ecb_byteorder_helper (void) |
478 | { |
502 | { |
479 | /* the union code still generates code under pressure in gcc, */ |
503 | /* the union code still generates code under pressure in gcc, */ |
480 | /* but less than using pointers, and always seem to */ |
504 | /* but less than using pointers, and always seems to */ |
481 | /* successfully return a constant. */ |
505 | /* successfully return a constant. */ |
482 | /* the reason why we have this horrible preprocessor mess */ |
506 | /* the reason why we have this horrible preprocessor mess */ |
483 | /* is to avoid it in all cases, at least on common architectures */ |
507 | /* is to avoid it in all cases, at least on common architectures */ |
484 | /* and yes, gcc defines __BYTE_ORDER__, g++ does not */ |
508 | /* or when using a recent enough gcc version (>= 4.6) */ |
485 | #if __i386 || __i386__ || _M_X86 || __amd64 || __amd64__ || _M_X64 |
509 | #if __i386 || __i386__ || _M_X86 || __amd64 || __amd64__ || _M_X64 |
486 | return 0x44; |
510 | return 0x44; |
487 | #elif __BYTE_ORDER__ && __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__ |
511 | #elif __BYTE_ORDER__ && __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__ |
488 | return 0x44; |
512 | return 0x44; |
489 | #elif __BYTE_ORDER__ && __BYTE_ORDER__ == __ORDER_BIG_ENDIAN__ |
513 | #elif __BYTE_ORDER__ && __BYTE_ORDER__ == __ORDER_BIG_ENDIAN__ |
490 | retrurn 0x11; |
514 | return 0x11; |
491 | #else |
515 | #else |
492 | union |
516 | union |
493 | { |
517 | { |
494 | uint32_t i; |
518 | uint32_t i; |
495 | uint8_t c; |
519 | uint8_t c; |
… | |
… | |
549 | || defined __s390__ || defined __s390x__ \ |
573 | || defined __s390__ || defined __s390x__ \ |
550 | || defined __mips__ \ |
574 | || defined __mips__ \ |
551 | || defined __alpha__ \ |
575 | || defined __alpha__ \ |
552 | || defined __hppa__ \ |
576 | || defined __hppa__ \ |
553 | || defined __ia64__ \ |
577 | || defined __ia64__ \ |
|
|
578 | || defined __m68k__ \ |
|
|
579 | || defined __m88k__ \ |
|
|
580 | || defined __sh__ \ |
554 | || defined _M_IX86 || defined _M_AMD64 || defined _M_IA64 |
581 | || defined _M_IX86 || defined _M_AMD64 || defined _M_IA64 |
555 | #define ECB_STDFP 1 |
582 | #define ECB_STDFP 1 |
556 | #include <string.h> /* for memcpy */ |
583 | #include <string.h> /* for memcpy */ |
557 | #else |
584 | #else |
558 | #define ECB_STDFP 0 |
585 | #define ECB_STDFP 0 |
559 | #include <math.h> /* for frexp*, ldexp* */ |
|
|
560 | #endif |
586 | #endif |
561 | |
587 | |
562 | #ifndef ECB_NO_LIBM |
588 | #ifndef ECB_NO_LIBM |
|
|
589 | |
|
|
590 | #include <math.h> /* for frexp*, ldexp*, INFINITY, NAN */ |
|
|
591 | |
|
|
592 | /* only the oldest of old doesn't have this one. solaris. */ |
|
|
593 | #ifdef INFINITY |
|
|
594 | #define ECB_INFINITY INFINITY |
|
|
595 | #else |
|
|
596 | #define ECB_INFINITY HUGE_VAL |
|
|
597 | #endif |
|
|
598 | |
|
|
599 | #ifdef NAN |
|
|
600 | #define ECB_NAN NAN |
|
|
601 | #else |
|
|
602 | #define ECB_NAN ECB_INFINITY |
|
|
603 | #endif |
|
|
604 | |
|
|
605 | /* converts an ieee half/binary16 to a float */ |
|
|
606 | ecb_function_ float ecb_binary16_to_float (uint16_t x) ecb_const; |
|
|
607 | ecb_function_ float |
|
|
608 | ecb_binary16_to_float (uint16_t x) |
|
|
609 | { |
|
|
610 | int e = (x >> 10) & 0x1f; |
|
|
611 | int m = x & 0x3ff; |
|
|
612 | float r; |
|
|
613 | |
|
|
614 | if (!e ) r = ldexpf (m , -24); |
|
|
615 | else if (e != 31) r = ldexpf (m + 0x400, e - 25); |
|
|
616 | else if (m ) r = ECB_NAN; |
|
|
617 | else r = ECB_INFINITY; |
|
|
618 | |
|
|
619 | return x & 0x8000 ? -r : r; |
|
|
620 | } |
563 | |
621 | |
564 | /* convert a float to ieee single/binary32 */ |
622 | /* convert a float to ieee single/binary32 */ |
565 | ecb_function_ uint32_t ecb_float_to_binary32 (float x) ecb_const; |
623 | ecb_function_ uint32_t ecb_float_to_binary32 (float x) ecb_const; |
566 | ecb_function_ uint32_t |
624 | ecb_function_ uint32_t |
567 | ecb_float_to_binary32 (float x) |
625 | ecb_float_to_binary32 (float x) |