2 * Speed-optimized CRC64 using slicing-by-four algorithm
4 * This uses only i386 instructions, but it is optimized for i686 and later
5 * (including e.g. Pentium II/III/IV, Athlon XP, and Core 2).
7 * Authors: Igor Pavlov (original CRC32 assembly code)
8 * Lasse Collin (CRC64 adaptation of the modified CRC32 code)
10 * This file has been put into the public domain.
11 * You can do whatever you want with this file.
13 * This code needs lzma_crc64_table, which can be created using the
16 uint64_t lzma_crc64_table[4][256];
22 static const uint64_t poly64 = UINT64_C(0xC96C5795D7870F42);
24 for (size_t s = 0; s < 4; ++s) {
25 for (size_t b = 0; b < 256; ++b) {
26 uint64_t r = s == 0 ? b : lzma_crc64_table[s - 1][b];
28 for (size_t i = 0; i < 8; ++i) {
30 r = (r >> 1) ^ poly64;
35 lzma_crc64_table[s][b] = r;
40 * The prototype of the CRC64 function:
41 * extern uint64_t lzma_crc64(const uint8_t *buf, size_t size, uint64_t crc);
45 * On some systems, the functions need to be prefixed. The prefix is
46 * usually an underscore.
48 #ifndef __USER_LABEL_PREFIX__
49 # define __USER_LABEL_PREFIX__
51 #define MAKE_SYM_CAT(prefix, sym) prefix ## sym
52 #define MAKE_SYM(prefix, sym) MAKE_SYM_CAT(prefix, sym)
53 #define LZMA_CRC64 MAKE_SYM(__USER_LABEL_PREFIX__, lzma_crc64)
54 #define LZMA_CRC64_TABLE MAKE_SYM(__USER_LABEL_PREFIX__, lzma_crc64_table)
57 * Solaris assembler doesn't have .p2align, and Darwin uses .align
58 * differently than GNU/Linux and Solaris.
60 #if defined(__APPLE__) || defined(__MSDOS__)
61 # define ALIGN(pow2, abs) .align pow2
63 # define ALIGN(pow2, abs) .align abs
69 #if !defined(__APPLE__) && !defined(_WIN32) && !defined(__CYGWIN__) \
70 && !defined(__MSDOS__)
71 .type LZMA_CRC64, @function
81 * %edi size or buf + size
82 * %ebx lzma_crc64_table
90 movl 0x14(%esp), %esi /* buf */
91 movl 0x18(%esp), %edi /* size */
92 movl 0x1C(%esp), %eax /* crc LSB */
93 movl 0x20(%esp), %edx /* crc MSB */
96 * Store the address of lzma_crc64_table to %ebx. This is needed to
97 * get position-independent code (PIC).
99 * The PIC macro is defined by libtool, while __PIC__ is defined
100 * by GCC but only on some systems. Testing for both makes it simpler
101 * to test this code without libtool, and keeps the code working also
102 * when built with libtool but using something else than GCC.
104 * I understood that libtool may define PIC on Windows even though
105 * the code in Windows DLLs is not PIC in sense that it is in ELF
106 * binaries, so we need a separate check to always use the non-PIC
109 #if (!defined(PIC) && !defined(__PIC__)) \
110 || (defined(_WIN32) || defined(__CYGWIN__))
112 movl $ LZMA_CRC64_TABLE, %ebx
113 #elif defined(__APPLE__)
117 leal .L_lzma_crc64_table$non_lazy_ptr-.L_pic(%ebx), %ebx
122 addl $_GLOBAL_OFFSET_TABLE_, %ebx
123 movl LZMA_CRC64_TABLE@GOT(%ebx), %ebx
126 /* Complement the initial value. */
132 * Check if there is enough input to use slicing-by-four.
133 * We need eight bytes, because the loop pre-reads four bytes.
138 /* Check if we have reached alignment of four bytes. */
142 /* Calculate CRC of the next input byte. */
148 xorl (%ebx, %ebp, 8), %eax
150 xorl 4(%ebx, %ebp, 8), %edx
156 * If we get here, there's at least eight bytes of aligned input
157 * available. Make %edi multiple of four bytes. Store the possible
158 * remainder over the "size" variable in the argument stack.
160 movl %edi, 0x18(%esp)
162 subl %edi, 0x18(%esp)
165 * Let %edi be buf + size - 4 while running the main loop. This way
166 * we can compare for equality to determine when exit the loop.
171 /* Read in the first four aligned bytes. */
177 movl 0x1800(%ebx, %ebp, 8), %eax
179 movl 0x1804(%ebx, %ebp, 8), %edx
181 xorl 0x1000(%ebx, %ebp, 8), %eax
182 xorl 0x1004(%ebx, %ebp, 8), %edx
185 xorl 0x0800(%ebx, %ebp, 8), %eax
186 xorl 0x0804(%ebx, %ebp, 8), %edx
189 xorl (%ebx, %ebp, 8), %eax
190 xorl 4(%ebx, %ebp, 8), %edx
192 /* Check for end of aligned input. */
196 * Copy the next input byte to %ecx. It is slightly faster to
197 * read it here than at the top of the loop.
203 * Process the remaining four bytes, which we have already
208 movl 0x1800(%ebx, %ebp, 8), %eax
210 movl 0x1804(%ebx, %ebp, 8), %edx
212 xorl 0x1000(%ebx, %ebp, 8), %eax
213 xorl 0x1004(%ebx, %ebp, 8), %edx
216 xorl 0x0800(%ebx, %ebp, 8), %eax
217 xorl 0x0804(%ebx, %ebp, 8), %edx
220 xorl (%ebx, %ebp, 8), %eax
221 xorl 4(%ebx, %ebp, 8), %edx
223 /* Copy the number of remaining bytes to %edi. */
224 movl 0x18(%esp), %edi
227 /* Check for end of input. */
231 /* Calculate CRC of the next input byte. */
237 xorl (%ebx, %ebp, 8), %eax
239 xorl 4(%ebx, %ebp, 8), %edx
244 /* Complement the final value. */
254 #if defined(PIC) || defined(__PIC__)
261 #if defined(__APPLE__) && (defined(PIC) || defined(__PIC__))
263 .section __IMPORT,__pointers,non_lazy_symbol_pointers
264 .L_lzma_crc64_table$non_lazy_ptr:
265 .indirect_symbol LZMA_CRC64_TABLE
268 #elif defined(_WIN32) || defined(__CYGWIN__)
270 /* This is equivalent of __declspec(dllexport). */
272 .ascii " -export:lzma_crc64"
275 #elif !defined(__MSDOS__)
277 .size LZMA_CRC64, .-LZMA_CRC64
281 * This is needed to support non-executable stack. It's ugly to
282 * use __linux__ here, but I don't know a way to detect when
283 * we are using GNU assembler.
285 #if defined(__ELF__) && defined(__linux__)
286 .section .note.GNU-stack,"",@progbits