1 // SPDX-License-Identifier: 0BSD
2
3 ///////////////////////////////////////////////////////////////////////////////
4 //
5 /// \file crc64.c
6 /// \brief CRC64 calculation
7 //
8 // Authors: Lasse Collin
9 // Ilya Kurdyukov
10 //
11 ///////////////////////////////////////////////////////////////////////////////
12
13 #include "check.h"
14 #include "crc_common.h"
15
16 #if defined(CRC_X86_CLMUL)
17 # define BUILDING_CRC_CLMUL 64
18 # include "crc_x86_clmul.h"
19 #endif
20
21
22 #ifdef CRC64_GENERIC
23
24 /////////////////////////////////
25 // Generic slice-by-four CRC64 //
26 /////////////////////////////////
27
28 #if defined(WORDS_BIGENDIAN)
29 # include "crc64_table_be.h"
30 #else
31 # include "crc64_table_le.h"
32 #endif
33
34
35 #ifdef HAVE_CRC_X86_ASM
36 extern uint64_t lzma_crc64_generic(
37 const uint8_t *buf, size_t size, uint64_t crc);
38 #else
39
40 #ifdef WORDS_BIGENDIAN
41 # define A1(x) ((x) >> 56)
42 #else
43 # define A1 A
44 #endif
45
46
47 // See the comments in crc32_fast.c. They aren't duplicated here.
48 static uint64_t
lzma_crc64_generic(const uint8_t * buf,size_t size,uint64_t crc)49 lzma_crc64_generic(const uint8_t *buf, size_t size, uint64_t crc)
50 {
51 crc = ~crc;
52
53 #ifdef WORDS_BIGENDIAN
54 crc = byteswap64(crc);
55 #endif
56
57 if (size > 4) {
58 while ((uintptr_t)(buf) & 3) {
59 crc = lzma_crc64_table[0][*buf++ ^ A1(crc)] ^ S8(crc);
60 --size;
61 }
62
63 const uint8_t *const limit = buf + (size & ~(size_t)(3));
64 size &= (size_t)(3);
65
66 while (buf < limit) {
67 #ifdef WORDS_BIGENDIAN
68 const uint32_t tmp = (uint32_t)(crc >> 32)
69 ^ aligned_read32ne(buf);
70 #else
71 const uint32_t tmp = (uint32_t)crc
72 ^ aligned_read32ne(buf);
73 #endif
74 buf += 4;
75
76 crc = lzma_crc64_table[3][A(tmp)]
77 ^ lzma_crc64_table[2][B(tmp)]
78 ^ S32(crc)
79 ^ lzma_crc64_table[1][C(tmp)]
80 ^ lzma_crc64_table[0][D(tmp)];
81 }
82 }
83
84 while (size-- != 0)
85 crc = lzma_crc64_table[0][*buf++ ^ A1(crc)] ^ S8(crc);
86
87 #ifdef WORDS_BIGENDIAN
88 crc = byteswap64(crc);
89 #endif
90
91 return ~crc;
92 }
93 #endif // HAVE_CRC_X86_ASM
94 #endif // CRC64_GENERIC
95
96
97 #if defined(CRC64_GENERIC) && defined(CRC64_ARCH_OPTIMIZED)
98
99 //////////////////////////
100 // Function dispatching //
101 //////////////////////////
102
103 // If both the generic and arch-optimized implementations are usable, then
104 // the function that is used is selected at runtime. See crc32_fast.c.
105
106 typedef uint64_t (*crc64_func_type)(
107 const uint8_t *buf, size_t size, uint64_t crc);
108
109 static crc64_func_type
crc64_resolve(void)110 crc64_resolve(void)
111 {
112 return is_arch_extension_supported()
113 ? &crc64_arch_optimized : &lzma_crc64_generic;
114 }
115
116 #ifdef HAVE_FUNC_ATTRIBUTE_CONSTRUCTOR
117 # define CRC64_SET_FUNC_ATTR __attribute__((__constructor__))
118 static crc64_func_type crc64_func;
119 #else
120 # define CRC64_SET_FUNC_ATTR
121 static uint64_t crc64_dispatch(const uint8_t *buf, size_t size, uint64_t crc);
122 static crc64_func_type crc64_func = &crc64_dispatch;
123 #endif
124
125
126 CRC64_SET_FUNC_ATTR
127 static void
crc64_set_func(void)128 crc64_set_func(void)
129 {
130 crc64_func = crc64_resolve();
131 return;
132 }
133
134
135 #ifndef HAVE_FUNC_ATTRIBUTE_CONSTRUCTOR
136 static uint64_t
crc64_dispatch(const uint8_t * buf,size_t size,uint64_t crc)137 crc64_dispatch(const uint8_t *buf, size_t size, uint64_t crc)
138 {
139 crc64_set_func();
140 return crc64_func(buf, size, crc);
141 }
142 #endif
143 #endif
144
145
146 extern LZMA_API(uint64_t)
lzma_crc64(const uint8_t * buf,size_t size,uint64_t crc)147 lzma_crc64(const uint8_t *buf, size_t size, uint64_t crc)
148 {
149 #if defined(_MSC_VER) && !defined(__INTEL_COMPILER) && !defined(__clang__) \
150 && defined(_M_IX86) && defined(CRC64_ARCH_OPTIMIZED)
151 // VS2015-2022 might corrupt the ebx register on 32-bit x86 when
152 // the CLMUL code is enabled. This hack forces MSVC to store and
153 // restore ebx. This is only needed here, not in lzma_crc32().
154 __asm mov ebx, ebx
155 #endif
156
157 #if defined(CRC64_GENERIC) && defined(CRC64_ARCH_OPTIMIZED)
158 return crc64_func(buf, size, crc);
159
160 #elif defined(CRC64_ARCH_OPTIMIZED)
161 // If arch-optimized version is used unconditionally without runtime
162 // CPU detection then omitting the generic version and its 8 KiB
163 // lookup table makes the library smaller.
164 return crc64_arch_optimized(buf, size, crc);
165
166 #else
167 return lzma_crc64_generic(buf, size, crc);
168 #endif
169 }
170