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 (©) 2009-2014 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: |
… | |
… | |
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 */ |
68 | /* work around x32 idiocy by defining proper macros */ |
69 | #if __x86_64 || _M_AMD64 |
69 | #if __amd64 || __x86_64 || _M_AMD64 || _M_X64 |
70 | #if __ILP32 |
70 | #if _ILP32 |
71 | #define ECB_AMD64_X32 1 |
71 | #define ECB_AMD64_X32 1 |
72 | #else |
72 | #else |
73 | #define ECB_AMD64 1 |
73 | #define ECB_AMD64 1 |
74 | #endif |
74 | #endif |
75 | #endif |
75 | #endif |
… | |
… | |
87 | #else |
87 | #else |
88 | #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))) |
89 | #endif |
89 | #endif |
90 | #endif |
90 | #endif |
91 | |
91 | |
92 | #define ECB_C (__STDC__+0) /* this assumes that __STDC__ is either empty or a number */ |
|
|
93 | #define ECB_C99 (__STDC_VERSION__ >= 199901L) |
|
|
94 | #define ECB_C11 (__STDC_VERSION__ >= 201112L) |
|
|
95 | #define ECB_CPP (__cplusplus+0) |
92 | #define ECB_CPP (__cplusplus+0) |
96 | #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) |
97 | |
105 | |
98 | #if ECB_CPP |
106 | #if ECB_CPP |
99 | #define ECB_EXTERN_C extern "C" |
107 | #define ECB_EXTERN_C extern "C" |
100 | #define ECB_EXTERN_C_BEG ECB_EXTERN_C { |
108 | #define ECB_EXTERN_C_BEG ECB_EXTERN_C { |
101 | #define ECB_EXTERN_C_END } |
109 | #define ECB_EXTERN_C_END } |
… | |
… | |
134 | || defined __ARM_ARCH_6K__ || defined __ARM_ARCH_6ZK__ |
142 | || defined __ARM_ARCH_6K__ || defined __ARM_ARCH_6ZK__ |
135 | #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") |
136 | #elif defined __ARM_ARCH_7__ || defined __ARM_ARCH_7A__ \ |
144 | #elif defined __ARM_ARCH_7__ || defined __ARM_ARCH_7A__ \ |
137 | || defined __ARM_ARCH_7M__ || defined __ARM_ARCH_7R__ |
145 | || defined __ARM_ARCH_7M__ || defined __ARM_ARCH_7R__ |
138 | #define ECB_MEMORY_FENCE __asm__ __volatile__ ("dmb" : : : "memory") |
146 | #define ECB_MEMORY_FENCE __asm__ __volatile__ ("dmb" : : : "memory") |
|
|
147 | #elif __aarch64__ |
|
|
148 | #define ECB_MEMORY_FENCE __asm__ __volatile__ ("dmb ish" : : : "memory") |
139 | #elif __sparc || __sparc__ |
149 | #elif (__sparc || __sparc__) && !__sparcv8 |
140 | #define ECB_MEMORY_FENCE __asm__ __volatile__ ("membar #LoadStore | #LoadLoad | #StoreStore | #StoreLoad" : : : "memory") |
150 | #define ECB_MEMORY_FENCE __asm__ __volatile__ ("membar #LoadStore | #LoadLoad | #StoreStore | #StoreLoad" : : : "memory") |
141 | #define ECB_MEMORY_FENCE_ACQUIRE __asm__ __volatile__ ("membar #LoadStore | #LoadLoad" : : : "memory") |
151 | #define ECB_MEMORY_FENCE_ACQUIRE __asm__ __volatile__ ("membar #LoadStore | #LoadLoad" : : : "memory") |
142 | #define ECB_MEMORY_FENCE_RELEASE __asm__ __volatile__ ("membar #LoadStore | #StoreStore") |
152 | #define ECB_MEMORY_FENCE_RELEASE __asm__ __volatile__ ("membar #LoadStore | #StoreStore") |
143 | #elif defined __s390__ || defined __s390x__ |
153 | #elif defined __s390__ || defined __s390x__ |
144 | #define ECB_MEMORY_FENCE __asm__ __volatile__ ("bcr 15,0" : : : "memory") |
154 | #define ECB_MEMORY_FENCE __asm__ __volatile__ ("bcr 15,0" : : : "memory") |
145 | #elif defined __mips__ |
155 | #elif defined __mips__ |
|
|
156 | /* GNU/Linux emulates sync on mips1 architectures, so we force its use */ |
|
|
157 | /* anybody else who still uses mips1 is supposed to send in their version, with detection code. */ |
146 | #define ECB_MEMORY_FENCE __asm__ __volatile__ ("sync" : : : "memory") |
158 | #define ECB_MEMORY_FENCE __asm__ __volatile__ (".set mips2; sync; .set mips0" : : : "memory") |
147 | #elif defined __alpha__ |
159 | #elif defined __alpha__ |
148 | #define ECB_MEMORY_FENCE __asm__ __volatile__ ("mb" : : : "memory") |
160 | #define ECB_MEMORY_FENCE __asm__ __volatile__ ("mb" : : : "memory") |
149 | #elif defined __hppa__ |
161 | #elif defined __hppa__ |
150 | #define ECB_MEMORY_FENCE __asm__ __volatile__ ("" : : : "memory") |
162 | #define ECB_MEMORY_FENCE __asm__ __volatile__ ("" : : : "memory") |
151 | #define ECB_MEMORY_FENCE_RELEASE __asm__ __volatile__ ("") |
163 | #define ECB_MEMORY_FENCE_RELEASE __asm__ __volatile__ ("") |
152 | #elif defined __ia64__ |
164 | #elif defined __ia64__ |
153 | #define ECB_MEMORY_FENCE __asm__ __volatile__ ("mf" : : : "memory") |
165 | #define ECB_MEMORY_FENCE __asm__ __volatile__ ("mf" : : : "memory") |
|
|
166 | #elif defined __m68k__ |
|
|
167 | #define ECB_MEMORY_FENCE __asm__ __volatile__ ("" : : : "memory") |
|
|
168 | #elif defined __m88k__ |
|
|
169 | #define ECB_MEMORY_FENCE __asm__ __volatile__ ("tb1 0,%%r0,128" : : : "memory") |
|
|
170 | #elif defined __sh__ |
|
|
171 | #define ECB_MEMORY_FENCE __asm__ __volatile__ ("" : : : "memory") |
154 | #endif |
172 | #endif |
155 | #endif |
173 | #endif |
156 | #endif |
174 | #endif |
157 | |
175 | |
158 | #ifndef ECB_MEMORY_FENCE |
176 | #ifndef ECB_MEMORY_FENCE |
159 | #if ECB_GCC_VERSION(4,7) |
177 | #if ECB_GCC_VERSION(4,7) |
160 | /* see comment below (stdatomic.h) about the C11 memory model. */ |
178 | /* see comment below (stdatomic.h) about the C11 memory model. */ |
161 | #define ECB_MEMORY_FENCE __atomic_thread_fence (__ATOMIC_SEQ_CST) |
179 | #define ECB_MEMORY_FENCE __atomic_thread_fence (__ATOMIC_SEQ_CST) |
|
|
180 | #define ECB_MEMORY_FENCE_ACQUIRE __atomic_thread_fence (__ATOMIC_ACQUIRE) |
|
|
181 | #define ECB_MEMORY_FENCE_RELEASE __atomic_thread_fence (__ATOMIC_RELEASE) |
162 | |
182 | |
163 | /* The __has_feature syntax from clang is so misdesigned that we cannot use it |
183 | /* The __has_feature syntax from clang is so misdesigned that we cannot use it |
164 | * without risking compile time errors with other compilers. We *could* |
184 | * without risking compile time errors with other compilers. We *could* |
165 | * define our own ecb_clang_has_feature, but I just can't be bothered to work |
185 | * define our own ecb_clang_has_feature, but I just can't be bothered to work |
166 | * around this shit time and again. |
186 | * around this shit time and again. |
167 | * #elif defined __clang && __has_feature (cxx_atomic) |
187 | * #elif defined __clang && __has_feature (cxx_atomic) |
168 | * // see comment below (stdatomic.h) about the C11 memory model. |
188 | * // see comment below (stdatomic.h) about the C11 memory model. |
169 | * #define ECB_MEMORY_FENCE __c11_atomic_thread_fence (__ATOMIC_SEQ_CST) |
189 | * #define ECB_MEMORY_FENCE __c11_atomic_thread_fence (__ATOMIC_SEQ_CST) |
|
|
190 | * #define ECB_MEMORY_FENCE_ACQUIRE __c11_atomic_thread_fence (__ATOMIC_ACQUIRE) |
|
|
191 | * #define ECB_MEMORY_FENCE_RELEASE __c11_atomic_thread_fence (__ATOMIC_RELEASE) |
170 | */ |
192 | */ |
171 | |
193 | |
172 | #elif ECB_GCC_VERSION(4,4) || defined __INTEL_COMPILER || defined __clang__ |
194 | #elif ECB_GCC_VERSION(4,4) || defined __INTEL_COMPILER || defined __clang__ |
173 | #define ECB_MEMORY_FENCE __sync_synchronize () |
195 | #define ECB_MEMORY_FENCE __sync_synchronize () |
|
|
196 | #elif _MSC_VER >= 1500 /* VC++ 2008 */ |
|
|
197 | /* apparently, microsoft broke all the memory barrier stuff in Visual Studio 2008... */ |
|
|
198 | #pragma intrinsic(_ReadBarrier,_WriteBarrier,_ReadWriteBarrier) |
|
|
199 | #define ECB_MEMORY_FENCE _ReadWriteBarrier (); MemoryBarrier() |
|
|
200 | #define ECB_MEMORY_FENCE_ACQUIRE _ReadWriteBarrier (); MemoryBarrier() /* according to msdn, _ReadBarrier is not a load fence */ |
|
|
201 | #define ECB_MEMORY_FENCE_RELEASE _WriteBarrier (); MemoryBarrier() |
174 | #elif _MSC_VER >= 1400 /* VC++ 2005 */ |
202 | #elif _MSC_VER >= 1400 /* VC++ 2005 */ |
175 | #pragma intrinsic(_ReadBarrier,_WriteBarrier,_ReadWriteBarrier) |
203 | #pragma intrinsic(_ReadBarrier,_WriteBarrier,_ReadWriteBarrier) |
176 | #define ECB_MEMORY_FENCE _ReadWriteBarrier () |
204 | #define ECB_MEMORY_FENCE _ReadWriteBarrier () |
177 | #define ECB_MEMORY_FENCE_ACQUIRE _ReadWriteBarrier () /* according to msdn, _ReadBarrier is not a load fence */ |
205 | #define ECB_MEMORY_FENCE_ACQUIRE _ReadWriteBarrier () /* according to msdn, _ReadBarrier is not a load fence */ |
178 | #define ECB_MEMORY_FENCE_RELEASE _WriteBarrier () |
206 | #define ECB_MEMORY_FENCE_RELEASE _WriteBarrier () |
… | |
… | |
198 | /* any fence other than seq_cst, which isn't very efficient for us. */ |
226 | /* any fence other than seq_cst, which isn't very efficient for us. */ |
199 | /* Why that is, we don't know - either the C11 memory model is quite useless */ |
227 | /* Why that is, we don't know - either the C11 memory model is quite useless */ |
200 | /* for most usages, or gcc and clang have a bug */ |
228 | /* for most usages, or gcc and clang have a bug */ |
201 | /* I *currently* lean towards the latter, and inefficiently implement */ |
229 | /* I *currently* lean towards the latter, and inefficiently implement */ |
202 | /* all three of ecb's fences as a seq_cst fence */ |
230 | /* all three of ecb's fences as a seq_cst fence */ |
|
|
231 | /* Update, gcc-4.8 generates mfence for all c++ fences, but nothing */ |
|
|
232 | /* for all __atomic_thread_fence's except seq_cst */ |
203 | #define ECB_MEMORY_FENCE atomic_thread_fence (memory_order_seq_cst) |
233 | #define ECB_MEMORY_FENCE atomic_thread_fence (memory_order_seq_cst) |
204 | #endif |
234 | #endif |
205 | #endif |
235 | #endif |
206 | |
236 | |
207 | #ifndef ECB_MEMORY_FENCE |
237 | #ifndef ECB_MEMORY_FENCE |
… | |
… | |
264 | #define ecb_is_constant(expr) __builtin_constant_p (expr) |
294 | #define ecb_is_constant(expr) __builtin_constant_p (expr) |
265 | #define ecb_expect(expr,value) __builtin_expect ((expr),(value)) |
295 | #define ecb_expect(expr,value) __builtin_expect ((expr),(value)) |
266 | #define ecb_prefetch(addr,rw,locality) __builtin_prefetch (addr, rw, locality) |
296 | #define ecb_prefetch(addr,rw,locality) __builtin_prefetch (addr, rw, locality) |
267 | #else |
297 | #else |
268 | #define ecb_attribute(attrlist) |
298 | #define ecb_attribute(attrlist) |
|
|
299 | |
|
|
300 | /* possible C11 impl for integral types |
|
|
301 | typedef struct ecb_is_constant_struct ecb_is_constant_struct; |
|
|
302 | #define ecb_is_constant(expr) _Generic ((1 ? (struct ecb_is_constant_struct *)0 : (void *)((expr) - (expr)), ecb_is_constant_struct *: 0, default: 1)) */ |
|
|
303 | |
269 | #define ecb_is_constant(expr) 0 |
304 | #define ecb_is_constant(expr) 0 |
270 | #define ecb_expect(expr,value) (expr) |
305 | #define ecb_expect(expr,value) (expr) |
271 | #define ecb_prefetch(addr,rw,locality) |
306 | #define ecb_prefetch(addr,rw,locality) |
272 | #endif |
307 | #endif |
273 | |
308 | |
… | |
… | |
565 | || defined __s390__ || defined __s390x__ \ |
600 | || defined __s390__ || defined __s390x__ \ |
566 | || defined __mips__ \ |
601 | || defined __mips__ \ |
567 | || defined __alpha__ \ |
602 | || defined __alpha__ \ |
568 | || defined __hppa__ \ |
603 | || defined __hppa__ \ |
569 | || defined __ia64__ \ |
604 | || defined __ia64__ \ |
|
|
605 | || defined __m68k__ \ |
|
|
606 | || defined __m88k__ \ |
|
|
607 | || defined __sh__ \ |
570 | || defined _M_IX86 || defined _M_AMD64 || defined _M_IA64 |
608 | || defined _M_IX86 || defined _M_AMD64 || defined _M_IA64 |
571 | #define ECB_STDFP 1 |
609 | #define ECB_STDFP 1 |
572 | #include <string.h> /* for memcpy */ |
610 | #include <string.h> /* for memcpy */ |
573 | #else |
611 | #else |
574 | #define ECB_STDFP 0 |
612 | #define ECB_STDFP 0 |
575 | #include <math.h> /* for frexp*, ldexp* */ |
|
|
576 | #endif |
613 | #endif |
577 | |
614 | |
578 | #ifndef ECB_NO_LIBM |
615 | #ifndef ECB_NO_LIBM |
|
|
616 | |
|
|
617 | #include <math.h> /* for frexp*, ldexp*, INFINITY, NAN */ |
|
|
618 | |
|
|
619 | /* only the oldest of old doesn't have this one. solaris. */ |
|
|
620 | #ifdef INFINITY |
|
|
621 | #define ECB_INFINITY INFINITY |
|
|
622 | #else |
|
|
623 | #define ECB_INFINITY HUGE_VAL |
|
|
624 | #endif |
|
|
625 | |
|
|
626 | #ifdef NAN |
|
|
627 | #define ECB_NAN NAN |
|
|
628 | #else |
|
|
629 | #define ECB_NAN ECB_INFINITY |
|
|
630 | #endif |
|
|
631 | |
|
|
632 | /* converts an ieee half/binary16 to a float */ |
|
|
633 | ecb_function_ float ecb_binary16_to_float (uint16_t x) ecb_const; |
|
|
634 | ecb_function_ float |
|
|
635 | ecb_binary16_to_float (uint16_t x) |
|
|
636 | { |
|
|
637 | int e = (x >> 10) & 0x1f; |
|
|
638 | int m = x & 0x3ff; |
|
|
639 | float r; |
|
|
640 | |
|
|
641 | if (!e ) r = ldexpf (m , -24); |
|
|
642 | else if (e != 31) r = ldexpf (m + 0x400, e - 25); |
|
|
643 | else if (m ) r = ECB_NAN; |
|
|
644 | else r = ECB_INFINITY; |
|
|
645 | |
|
|
646 | return x & 0x8000 ? -r : r; |
|
|
647 | } |
579 | |
648 | |
580 | /* convert a float to ieee single/binary32 */ |
649 | /* convert a float to ieee single/binary32 */ |
581 | ecb_function_ uint32_t ecb_float_to_binary32 (float x) ecb_const; |
650 | ecb_function_ uint32_t ecb_float_to_binary32 (float x) ecb_const; |
582 | ecb_function_ uint32_t |
651 | ecb_function_ uint32_t |
583 | ecb_float_to_binary32 (float x) |
652 | ecb_float_to_binary32 (float x) |