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; |
… | |
… | |
63 | #else |
63 | #else |
64 | #define ECB_PTRSIZE 4 |
64 | #define ECB_PTRSIZE 4 |
65 | #endif |
65 | #endif |
66 | #endif |
66 | #endif |
67 | |
67 | |
|
|
68 | /* work around x32 idiocy by defining proper macros */ |
|
|
69 | #if __amd64 || __x86_64 || _M_AMD64 || _M_X64 |
|
|
70 | #if _ILP32 |
|
|
71 | #define ECB_AMD64_X32 1 |
|
|
72 | #else |
|
|
73 | #define ECB_AMD64 1 |
|
|
74 | #endif |
|
|
75 | #endif |
|
|
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, |
70 | * causing enormous grief in return for some better fake benchmark numbers. |
79 | * causing enormous grief in return for some better fake benchmark numbers. |
71 | * or so. |
80 | * or so. |
72 | * we try to detect these and simply assume they are not gcc - if they have |
81 | * we try to detect these and simply assume they are not gcc - if they have |
… | |
… | |
78 | #else |
87 | #else |
79 | #define ECB_GCC_VERSION(major,minor) (__GNUC__ > (major) || (__GNUC__ == (major) && __GNUC_MINOR__ >= (minor))) |
88 | #define ECB_GCC_VERSION(major,minor) (__GNUC__ > (major) || (__GNUC__ == (major) && __GNUC_MINOR__ >= (minor))) |
80 | #endif |
89 | #endif |
81 | #endif |
90 | #endif |
82 | |
91 | |
83 | #define ECB_C (__STDC__+0) /* this assumes that __STDC__ is either empty or a number */ |
|
|
84 | #define ECB_C99 (__STDC_VERSION__ >= 199901L) |
|
|
85 | #define ECB_C11 (__STDC_VERSION__ >= 201112L) |
|
|
86 | #define ECB_CPP (__cplusplus+0) |
92 | #define ECB_CPP (__cplusplus+0) |
87 | #define ECB_CPP11 (__cplusplus >= 201103L) |
93 | #define ECB_CPP11 (__cplusplus >= 201103L) |
|
|
94 | |
|
|
95 | #if ECB_CPP |
|
|
96 | #define ECB_C 0 |
|
|
97 | #define ECB_STDC_VERSION 0 |
|
|
98 | #else |
|
|
99 | #define ECB_C 1 |
|
|
100 | #define ECB_STDC_VERSION __STDC_VERSION__ |
|
|
101 | #endif |
|
|
102 | |
|
|
103 | #define ECB_C99 (ECB_STDC_VERSION >= 199901L) |
|
|
104 | #define ECB_C11 (ECB_STDC_VERSION >= 201112L) |
88 | |
105 | |
89 | #if ECB_CPP |
106 | #if ECB_CPP |
90 | #define ECB_EXTERN_C extern "C" |
107 | #define ECB_EXTERN_C extern "C" |
91 | #define ECB_EXTERN_C_BEG ECB_EXTERN_C { |
108 | #define ECB_EXTERN_C_BEG ECB_EXTERN_C { |
92 | #define ECB_EXTERN_C_END } |
109 | #define ECB_EXTERN_C_END } |
… | |
… | |
125 | || defined __ARM_ARCH_6K__ || defined __ARM_ARCH_6ZK__ |
142 | || 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") |
143 | #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__ \ |
144 | #elif defined __ARM_ARCH_7__ || defined __ARM_ARCH_7A__ \ |
128 | || defined __ARM_ARCH_7M__ || defined __ARM_ARCH_7R__ |
145 | || defined __ARM_ARCH_7M__ || defined __ARM_ARCH_7R__ |
129 | #define ECB_MEMORY_FENCE __asm__ __volatile__ ("dmb" : : : "memory") |
146 | #define ECB_MEMORY_FENCE __asm__ __volatile__ ("dmb" : : : "memory") |
130 | #elif __sparc || __sparc__ |
147 | #elif (__sparc || __sparc__) && !__sparcv8 |
131 | #define ECB_MEMORY_FENCE __asm__ __volatile__ ("membar #LoadStore | #LoadLoad | #StoreStore | #StoreLoad" : : : "memory") |
148 | #define ECB_MEMORY_FENCE __asm__ __volatile__ ("membar #LoadStore | #LoadLoad | #StoreStore | #StoreLoad" : : : "memory") |
132 | #define ECB_MEMORY_FENCE_ACQUIRE __asm__ __volatile__ ("membar #LoadStore | #LoadLoad" : : : "memory") |
149 | #define ECB_MEMORY_FENCE_ACQUIRE __asm__ __volatile__ ("membar #LoadStore | #LoadLoad" : : : "memory") |
133 | #define ECB_MEMORY_FENCE_RELEASE __asm__ __volatile__ ("membar #LoadStore | #StoreStore") |
150 | #define ECB_MEMORY_FENCE_RELEASE __asm__ __volatile__ ("membar #LoadStore | #StoreStore") |
134 | #elif defined __s390__ || defined __s390x__ |
151 | #elif defined __s390__ || defined __s390x__ |
135 | #define ECB_MEMORY_FENCE __asm__ __volatile__ ("bcr 15,0" : : : "memory") |
152 | #define ECB_MEMORY_FENCE __asm__ __volatile__ ("bcr 15,0" : : : "memory") |
136 | #elif defined __mips__ |
153 | #elif defined __mips__ |
|
|
154 | /* GNU/Linux emulates sync on mips1 architectures, so we force its use */ |
|
|
155 | /* 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") |
156 | #define ECB_MEMORY_FENCE __asm__ __volatile__ (".set mips2; sync; .set mips0" : : : "memory") |
138 | #elif defined __alpha__ |
157 | #elif defined __alpha__ |
139 | #define ECB_MEMORY_FENCE __asm__ __volatile__ ("mb" : : : "memory") |
158 | #define ECB_MEMORY_FENCE __asm__ __volatile__ ("mb" : : : "memory") |
140 | #elif defined __hppa__ |
159 | #elif defined __hppa__ |
141 | #define ECB_MEMORY_FENCE __asm__ __volatile__ ("" : : : "memory") |
160 | #define ECB_MEMORY_FENCE __asm__ __volatile__ ("" : : : "memory") |
142 | #define ECB_MEMORY_FENCE_RELEASE __asm__ __volatile__ ("") |
161 | #define ECB_MEMORY_FENCE_RELEASE __asm__ __volatile__ ("") |
143 | #elif defined __ia64__ |
162 | #elif defined __ia64__ |
144 | #define ECB_MEMORY_FENCE __asm__ __volatile__ ("mf" : : : "memory") |
163 | #define ECB_MEMORY_FENCE __asm__ __volatile__ ("mf" : : : "memory") |
|
|
164 | #elif defined __m68k__ |
|
|
165 | #define ECB_MEMORY_FENCE __asm__ __volatile__ ("" : : : "memory") |
|
|
166 | #elif defined __m88k__ |
|
|
167 | #define ECB_MEMORY_FENCE __asm__ __volatile__ ("tb1 0,%%r0,128" : : : "memory") |
|
|
168 | #elif defined __sh__ |
|
|
169 | #define ECB_MEMORY_FENCE __asm__ __volatile__ ("" : : : "memory") |
145 | #endif |
170 | #endif |
146 | #endif |
171 | #endif |
147 | #endif |
172 | #endif |
148 | |
173 | |
149 | #ifndef ECB_MEMORY_FENCE |
174 | #ifndef ECB_MEMORY_FENCE |
150 | #if ECB_GCC_VERSION(4,7) |
175 | #if ECB_GCC_VERSION(4,7) |
151 | /* see comment below (stdatomic.h) about the C11 memory model. */ |
176 | /* see comment below (stdatomic.h) about the C11 memory model. */ |
152 | #define ECB_MEMORY_FENCE __atomic_thread_fence (__ATOMIC_SEQ_CST) |
177 | #define ECB_MEMORY_FENCE __atomic_thread_fence (__ATOMIC_SEQ_CST) |
|
|
178 | #define ECB_MEMORY_FENCE_ACQUIRE __atomic_thread_fence (__ATOMIC_ACQUIRE) |
|
|
179 | #define ECB_MEMORY_FENCE_RELEASE __atomic_thread_fence (__ATOMIC_RELEASE) |
|
|
180 | |
|
|
181 | /* The __has_feature syntax from clang is so misdesigned that we cannot use it |
|
|
182 | * without risking compile time errors with other compilers. We *could* |
|
|
183 | * define our own ecb_clang_has_feature, but I just can't be bothered to work |
|
|
184 | * around this shit time and again. |
153 | /*#elif defined __clang && __has_feature (cxx_atomic)*/ |
185 | * #elif defined __clang && __has_feature (cxx_atomic) |
154 | /* see comment below (stdatomic.h) about the C11 memory model. */ |
186 | * // see comment below (stdatomic.h) about the C11 memory model. |
155 | /*#define ECB_MEMORY_FENCE __c11_atomic_thread_fence (__ATOMIC_SEQ_CST)*/ |
187 | * #define ECB_MEMORY_FENCE __c11_atomic_thread_fence (__ATOMIC_SEQ_CST) |
|
|
188 | * #define ECB_MEMORY_FENCE_ACQUIRE __c11_atomic_thread_fence (__ATOMIC_ACQUIRE) |
|
|
189 | * #define ECB_MEMORY_FENCE_RELEASE __c11_atomic_thread_fence (__ATOMIC_RELEASE) |
|
|
190 | */ |
|
|
191 | |
156 | #elif ECB_GCC_VERSION(4,4) || defined __INTEL_COMPILER || defined __clang__ |
192 | #elif ECB_GCC_VERSION(4,4) || defined __INTEL_COMPILER || defined __clang__ |
157 | #define ECB_MEMORY_FENCE __sync_synchronize () |
193 | #define ECB_MEMORY_FENCE __sync_synchronize () |
|
|
194 | #elif _MSC_VER >= 1500 /* VC++ 2008 */ |
|
|
195 | /* apparently, microsoft broke all the memory barrier stuff in Visual Studio 2008... */ |
|
|
196 | #pragma intrinsic(_ReadBarrier,_WriteBarrier,_ReadWriteBarrier) |
|
|
197 | #define ECB_MEMORY_FENCE _ReadWriteBarrier (); MemoryBarrier() |
|
|
198 | #define ECB_MEMORY_FENCE_ACQUIRE _ReadWriteBarrier (); MemoryBarrier() /* according to msdn, _ReadBarrier is not a load fence */ |
|
|
199 | #define ECB_MEMORY_FENCE_RELEASE _WriteBarrier (); MemoryBarrier() |
158 | #elif _MSC_VER >= 1400 /* VC++ 2005 */ |
200 | #elif _MSC_VER >= 1400 /* VC++ 2005 */ |
159 | #pragma intrinsic(_ReadBarrier,_WriteBarrier,_ReadWriteBarrier) |
201 | #pragma intrinsic(_ReadBarrier,_WriteBarrier,_ReadWriteBarrier) |
160 | #define ECB_MEMORY_FENCE _ReadWriteBarrier () |
202 | #define ECB_MEMORY_FENCE _ReadWriteBarrier () |
161 | #define ECB_MEMORY_FENCE_ACQUIRE _ReadWriteBarrier () /* according to msdn, _ReadBarrier is not a load fence */ |
203 | #define ECB_MEMORY_FENCE_ACQUIRE _ReadWriteBarrier () /* according to msdn, _ReadBarrier is not a load fence */ |
162 | #define ECB_MEMORY_FENCE_RELEASE _WriteBarrier () |
204 | #define ECB_MEMORY_FENCE_RELEASE _WriteBarrier () |
… | |
… | |
182 | /* any fence other than seq_cst, which isn't very efficient for us. */ |
224 | /* any fence other than seq_cst, which isn't very efficient for us. */ |
183 | /* Why that is, we don't know - either the C11 memory model is quite useless */ |
225 | /* Why that is, we don't know - either the C11 memory model is quite useless */ |
184 | /* for most usages, or gcc and clang have a bug */ |
226 | /* for most usages, or gcc and clang have a bug */ |
185 | /* I *currently* lean towards the latter, and inefficiently implement */ |
227 | /* I *currently* lean towards the latter, and inefficiently implement */ |
186 | /* all three of ecb's fences as a seq_cst fence */ |
228 | /* all three of ecb's fences as a seq_cst fence */ |
|
|
229 | /* Update, gcc-4.8 generates mfence for all c++ fences, but nothing */ |
|
|
230 | /* for all __atomic_thread_fence's except seq_cst */ |
187 | #define ECB_MEMORY_FENCE atomic_thread_fence (memory_order_seq_cst) |
231 | #define ECB_MEMORY_FENCE atomic_thread_fence (memory_order_seq_cst) |
188 | #endif |
232 | #endif |
189 | #endif |
233 | #endif |
190 | |
234 | |
191 | #ifndef ECB_MEMORY_FENCE |
235 | #ifndef ECB_MEMORY_FENCE |
… | |
… | |
248 | #define ecb_is_constant(expr) __builtin_constant_p (expr) |
292 | #define ecb_is_constant(expr) __builtin_constant_p (expr) |
249 | #define ecb_expect(expr,value) __builtin_expect ((expr),(value)) |
293 | #define ecb_expect(expr,value) __builtin_expect ((expr),(value)) |
250 | #define ecb_prefetch(addr,rw,locality) __builtin_prefetch (addr, rw, locality) |
294 | #define ecb_prefetch(addr,rw,locality) __builtin_prefetch (addr, rw, locality) |
251 | #else |
295 | #else |
252 | #define ecb_attribute(attrlist) |
296 | #define ecb_attribute(attrlist) |
|
|
297 | |
|
|
298 | /* possible C11 impl for integral types |
|
|
299 | typedef struct ecb_is_constant_struct ecb_is_constant_struct; |
|
|
300 | #define ecb_is_constant(expr) _Generic ((1 ? (struct ecb_is_constant_struct *)0 : (void *)((expr) - (expr)), ecb_is_constant_struct *: 0, default: 1)) */ |
|
|
301 | |
253 | #define ecb_is_constant(expr) 0 |
302 | #define ecb_is_constant(expr) 0 |
254 | #define ecb_expect(expr,value) (expr) |
303 | #define ecb_expect(expr,value) (expr) |
255 | #define ecb_prefetch(addr,rw,locality) |
304 | #define ecb_prefetch(addr,rw,locality) |
256 | #endif |
305 | #endif |
257 | |
306 | |
… | |
… | |
475 | ecb_inline unsigned char ecb_byteorder_helper (void) ecb_const; |
524 | ecb_inline unsigned char ecb_byteorder_helper (void) ecb_const; |
476 | ecb_inline unsigned char |
525 | ecb_inline unsigned char |
477 | ecb_byteorder_helper (void) |
526 | ecb_byteorder_helper (void) |
478 | { |
527 | { |
479 | /* the union code still generates code under pressure in gcc, */ |
528 | /* the union code still generates code under pressure in gcc, */ |
480 | /* but less than using pointers, and always seem to */ |
529 | /* but less than using pointers, and always seems to */ |
481 | /* successfully return a constant. */ |
530 | /* successfully return a constant. */ |
482 | /* the reason why we have this horrible preprocessor mess */ |
531 | /* the reason why we have this horrible preprocessor mess */ |
483 | /* is to avoid it in all cases, at least on common architectures */ |
532 | /* is to avoid it in all cases, at least on common architectures */ |
484 | /* and yes, gcc defines __BYTE_ORDER__, g++ does not */ |
533 | /* or when using a recent enough gcc version (>= 4.6) */ |
485 | #if __i386 || __i386__ || _M_X86 || __amd64 || __amd64__ || _M_X64 |
534 | #if __i386 || __i386__ || _M_X86 || __amd64 || __amd64__ || _M_X64 |
486 | return 0x44; |
535 | return 0x44; |
487 | #elif __BYTE_ORDER__ && __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__ |
536 | #elif __BYTE_ORDER__ && __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__ |
488 | return 0x44; |
537 | return 0x44; |
489 | #elif __BYTE_ORDER__ && __BYTE_ORDER__ == __ORDER_BIG_ENDIAN__ |
538 | #elif __BYTE_ORDER__ && __BYTE_ORDER__ == __ORDER_BIG_ENDIAN__ |
490 | retrurn 0x11; |
539 | return 0x11; |
491 | #else |
540 | #else |
492 | union |
541 | union |
493 | { |
542 | { |
494 | uint32_t i; |
543 | uint32_t i; |
495 | uint8_t c; |
544 | uint8_t c; |
… | |
… | |
549 | || defined __s390__ || defined __s390x__ \ |
598 | || defined __s390__ || defined __s390x__ \ |
550 | || defined __mips__ \ |
599 | || defined __mips__ \ |
551 | || defined __alpha__ \ |
600 | || defined __alpha__ \ |
552 | || defined __hppa__ \ |
601 | || defined __hppa__ \ |
553 | || defined __ia64__ \ |
602 | || defined __ia64__ \ |
|
|
603 | || defined __m68k__ \ |
|
|
604 | || defined __m88k__ \ |
|
|
605 | || defined __sh__ \ |
554 | || defined _M_IX86 || defined _M_AMD64 || defined _M_IA64 |
606 | || defined _M_IX86 || defined _M_AMD64 || defined _M_IA64 |
555 | #define ECB_STDFP 1 |
607 | #define ECB_STDFP 1 |
556 | #include <string.h> /* for memcpy */ |
608 | #include <string.h> /* for memcpy */ |
557 | #else |
609 | #else |
558 | #define ECB_STDFP 0 |
610 | #define ECB_STDFP 0 |
559 | #include <math.h> /* for frexp*, ldexp* */ |
|
|
560 | #endif |
611 | #endif |
561 | |
612 | |
562 | #ifndef ECB_NO_LIBM |
613 | #ifndef ECB_NO_LIBM |
|
|
614 | |
|
|
615 | #include <math.h> /* for frexp*, ldexp*, INFINITY, NAN */ |
|
|
616 | |
|
|
617 | /* only the oldest of old doesn't have this one. solaris. */ |
|
|
618 | #ifdef INFINITY |
|
|
619 | #define ECB_INFINITY INFINITY |
|
|
620 | #else |
|
|
621 | #define ECB_INFINITY HUGE_VAL |
|
|
622 | #endif |
|
|
623 | |
|
|
624 | #ifdef NAN |
|
|
625 | #define ECB_NAN NAN |
|
|
626 | #else |
|
|
627 | #define ECB_NAN ECB_INFINITY |
|
|
628 | #endif |
|
|
629 | |
|
|
630 | /* converts an ieee half/binary16 to a float */ |
|
|
631 | ecb_function_ float ecb_binary16_to_float (uint16_t x) ecb_const; |
|
|
632 | ecb_function_ float |
|
|
633 | ecb_binary16_to_float (uint16_t x) |
|
|
634 | { |
|
|
635 | int e = (x >> 10) & 0x1f; |
|
|
636 | int m = x & 0x3ff; |
|
|
637 | float r; |
|
|
638 | |
|
|
639 | if (!e ) r = ldexpf (m , -24); |
|
|
640 | else if (e != 31) r = ldexpf (m + 0x400, e - 25); |
|
|
641 | else if (m ) r = ECB_NAN; |
|
|
642 | else r = ECB_INFINITY; |
|
|
643 | |
|
|
644 | return x & 0x8000 ? -r : r; |
|
|
645 | } |
563 | |
646 | |
564 | /* convert a float to ieee single/binary32 */ |
647 | /* convert a float to ieee single/binary32 */ |
565 | ecb_function_ uint32_t ecb_float_to_binary32 (float x) ecb_const; |
648 | ecb_function_ uint32_t ecb_float_to_binary32 (float x) ecb_const; |
566 | ecb_function_ uint32_t |
649 | ecb_function_ uint32_t |
567 | ecb_float_to_binary32 (float x) |
650 | ecb_float_to_binary32 (float x) |