1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
|
// SPDX-License-Identifier: 0BSD
///////////////////////////////////////////////////////////////////////////////
//
/// \file crc32.c
/// \brief CRC32 calculation
//
// Authors: Lasse Collin
// Ilya Kurdyukov
// Hans Jansen
//
///////////////////////////////////////////////////////////////////////////////
#include "check.h"
#include "crc_common.h"
#if defined(CRC_X86_CLMUL)
# define BUILDING_CRC32_CLMUL
# include "crc_x86_clmul.h"
#elif defined(CRC32_ARM64)
# include "crc32_arm64.h"
#endif
#ifdef CRC32_GENERIC
///////////////////
// Generic CRC32 //
///////////////////
static uint32_t
crc32_generic(const uint8_t *buf, size_t size, uint32_t crc)
{
crc = ~crc;
#ifdef WORDS_BIGENDIAN
crc = byteswap32(crc);
#endif
if (size > 8) {
// Fix the alignment, if needed. The if statement above
// ensures that this won't read past the end of buf[].
while ((uintptr_t)(buf) & 7) {
crc = lzma_crc32_table[0][*buf++ ^ A(crc)] ^ S8(crc);
--size;
}
// Calculate the position where to stop.
const uint8_t *const limit = buf + (size & ~(size_t)(7));
// Calculate how many bytes must be calculated separately
// before returning the result.
size &= (size_t)(7);
// Calculate the CRC32 using the slice-by-eight algorithm.
while (buf < limit) {
crc ^= aligned_read32ne(buf);
buf += 4;
crc = lzma_crc32_table[7][A(crc)]
^ lzma_crc32_table[6][B(crc)]
^ lzma_crc32_table[5][C(crc)]
^ lzma_crc32_table[4][D(crc)];
const uint32_t tmp = aligned_read32ne(buf);
buf += 4;
// At least with some compilers, it is critical for
// performance, that the crc variable is XORed
// between the two table-lookup pairs.
crc = lzma_crc32_table[3][A(tmp)]
^ lzma_crc32_table[2][B(tmp)]
^ crc
^ lzma_crc32_table[1][C(tmp)]
^ lzma_crc32_table[0][D(tmp)];
}
}
while (size-- != 0)
crc = lzma_crc32_table[0][*buf++ ^ A(crc)] ^ S8(crc);
#ifdef WORDS_BIGENDIAN
crc = byteswap32(crc);
#endif
return ~crc;
}
#endif
#if defined(CRC32_GENERIC) && defined(CRC32_ARCH_OPTIMIZED)
//////////////////////////
// Function dispatching //
//////////////////////////
// If both the generic and arch-optimized implementations are built, then
// the function to use is selected at runtime because the system running
// the binary might not have the arch-specific instruction set extension(s)
// available. The dispatch methods in order of priority:
//
// 1. Constructor. This method uses __attribute__((__constructor__)) to
// set crc32_func at load time. This avoids extra computation (and any
// unlikely threading bugs) on the first call to lzma_crc32() to decide
// which implementation should be used.
//
// 2. First Call Resolution. On the very first call to lzma_crc32(), the
// call will be directed to crc32_dispatch() instead. This will set the
// appropriate implementation function and will not be called again.
// This method does not use any kind of locking but is safe because if
// multiple threads run the dispatcher simultaneously then they will all
// set crc32_func to the same value.
typedef uint32_t (*crc32_func_type)(
const uint8_t *buf, size_t size, uint32_t crc);
// This resolver is shared between all dispatch methods.
static crc32_func_type
crc32_resolve(void)
{
return is_arch_extension_supported()
? &crc32_arch_optimized : &crc32_generic;
}
#ifdef HAVE_FUNC_ATTRIBUTE_CONSTRUCTOR
// Constructor method.
# define CRC32_SET_FUNC_ATTR __attribute__((__constructor__))
static crc32_func_type crc32_func;
#else
// First Call Resolution method.
# define CRC32_SET_FUNC_ATTR
static uint32_t crc32_dispatch(const uint8_t *buf, size_t size, uint32_t crc);
static crc32_func_type crc32_func = &crc32_dispatch;
#endif
CRC32_SET_FUNC_ATTR
static void
crc32_set_func(void)
{
crc32_func = crc32_resolve();
return;
}
#ifndef HAVE_FUNC_ATTRIBUTE_CONSTRUCTOR
static uint32_t
crc32_dispatch(const uint8_t *buf, size_t size, uint32_t crc)
{
// When __attribute__((__constructor__)) isn't supported, set the
// function pointer without any locking. If multiple threads run
// the detection code in parallel, they will all end up setting
// the pointer to the same value. This avoids the use of
// mythread_once() on every call to lzma_crc32() but this likely
// isn't strictly standards compliant. Let's change it if it breaks.
crc32_set_func();
return crc32_func(buf, size, crc);
}
#endif
#endif
extern LZMA_API(uint32_t)
lzma_crc32(const uint8_t *buf, size_t size, uint32_t crc)
{
#if defined(CRC32_GENERIC) && defined(CRC32_ARCH_OPTIMIZED)
// On x86-64, if CLMUL is available, it is the best for non-tiny
// inputs, being over twice as fast as the generic slice-by-four
// version. However, for size <= 16 it's different. In the extreme
// case of size == 1 the generic version can be five times faster.
// At size >= 8 the CLMUL starts to become reasonable. It
// varies depending on the alignment of buf too.
//
// The above doesn't include the overhead of mythread_once().
// At least on x86-64 GNU/Linux, pthread_once() is very fast but
// it still makes lzma_crc32(buf, 1, crc) 50-100 % slower. When
// size reaches 12-16 bytes the overhead becomes negligible.
//
// So using the generic version for size <= 16 may give better
// performance with tiny inputs but if such inputs happen rarely
// it's not so obvious because then the lookup table of the
// generic version may not be in the processor cache.
#ifdef CRC_USE_GENERIC_FOR_SMALL_INPUTS
if (size <= 16)
return crc32_generic(buf, size, crc);
#endif
/*
#ifndef HAVE_FUNC_ATTRIBUTE_CONSTRUCTOR
// See crc32_dispatch(). This would be the alternative which uses
// locking and doesn't use crc32_dispatch(). Note that on Windows
// this method needs Vista threads.
mythread_once(crc64_set_func);
#endif
*/
return crc32_func(buf, size, crc);
#elif defined(CRC32_ARCH_OPTIMIZED)
return crc32_arch_optimized(buf, size, crc);
#else
return crc32_generic(buf, size, crc);
#endif
}
|
