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);
44 /* When Intel CET is enabled, include <cet.h> in assembly code to mark
53 * On some systems, the functions need to be prefixed. The prefix is
54 * usually an underscore.
56 #ifndef __USER_LABEL_PREFIX__
57 # define __USER_LABEL_PREFIX__
59 #define MAKE_SYM_CAT(prefix, sym) prefix ## sym
60 #define MAKE_SYM(prefix, sym) MAKE_SYM_CAT(prefix, sym)
61 #define LZMA_CRC64 MAKE_SYM(__USER_LABEL_PREFIX__, lzma_crc64)
62 #define LZMA_CRC64_TABLE MAKE_SYM(__USER_LABEL_PREFIX__, lzma_crc64_table)
65 * Solaris assembler doesn't have .p2align, and Darwin uses .align
66 * differently than GNU/Linux and Solaris.
68 #if defined(__APPLE__) || defined(__MSDOS__)
69 # define ALIGN(pow2, abs) .align pow2
71 # define ALIGN(pow2, abs) .align abs
77 #if !defined(__APPLE__) && !defined(_WIN32) && !defined(__CYGWIN__) \
78 && !defined(__MSDOS__)
79 .type LZMA_CRC64, @function
90 * %edi size or buf + size
91 * %ebx lzma_crc64_table
99 movl 0x14(%esp), %esi /* buf */
100 movl 0x18(%esp), %edi /* size */
101 movl 0x1C(%esp), %eax /* crc LSB */
102 movl 0x20(%esp), %edx /* crc MSB */
105 * Store the address of lzma_crc64_table to %ebx. This is needed to
106 * get position-independent code (PIC).
108 * The PIC macro is defined by libtool, while __PIC__ is defined
109 * by GCC but only on some systems. Testing for both makes it simpler
110 * to test this code without libtool, and keeps the code working also
111 * when built with libtool but using something else than GCC.
113 * I understood that libtool may define PIC on Windows even though
114 * the code in Windows DLLs is not PIC in sense that it is in ELF
115 * binaries, so we need a separate check to always use the non-PIC
118 #if (!defined(PIC) && !defined(__PIC__)) \
119 || (defined(_WIN32) || defined(__CYGWIN__))
121 movl $ LZMA_CRC64_TABLE, %ebx
122 #elif defined(__APPLE__)
126 leal .L_lzma_crc64_table$non_lazy_ptr-.L_pic(%ebx), %ebx
131 addl $_GLOBAL_OFFSET_TABLE_, %ebx
132 movl LZMA_CRC64_TABLE@GOT(%ebx), %ebx
135 /* Complement the initial value. */
141 * Check if there is enough input to use slicing-by-four.
142 * We need eight bytes, because the loop pre-reads four bytes.
147 /* Check if we have reached alignment of four bytes. */
151 /* Calculate CRC of the next input byte. */
157 xorl (%ebx, %ebp, 8), %eax
159 xorl 4(%ebx, %ebp, 8), %edx
165 * If we get here, there's at least eight bytes of aligned input
166 * available. Make %edi multiple of four bytes. Store the possible
167 * remainder over the "size" variable in the argument stack.
169 movl %edi, 0x18(%esp)
171 subl %edi, 0x18(%esp)
174 * Let %edi be buf + size - 4 while running the main loop. This way
175 * we can compare for equality to determine when exit the loop.
180 /* Read in the first four aligned bytes. */
186 movl 0x1800(%ebx, %ebp, 8), %eax
188 movl 0x1804(%ebx, %ebp, 8), %edx
190 xorl 0x1000(%ebx, %ebp, 8), %eax
191 xorl 0x1004(%ebx, %ebp, 8), %edx
194 xorl 0x0800(%ebx, %ebp, 8), %eax
195 xorl 0x0804(%ebx, %ebp, 8), %edx
198 xorl (%ebx, %ebp, 8), %eax
199 xorl 4(%ebx, %ebp, 8), %edx
201 /* Check for end of aligned input. */
205 * Copy the next input byte to %ecx. It is slightly faster to
206 * read it here than at the top of the loop.
212 * Process the remaining four bytes, which we have already
217 movl 0x1800(%ebx, %ebp, 8), %eax
219 movl 0x1804(%ebx, %ebp, 8), %edx
221 xorl 0x1000(%ebx, %ebp, 8), %eax
222 xorl 0x1004(%ebx, %ebp, 8), %edx
225 xorl 0x0800(%ebx, %ebp, 8), %eax
226 xorl 0x0804(%ebx, %ebp, 8), %edx
229 xorl (%ebx, %ebp, 8), %eax
230 xorl 4(%ebx, %ebp, 8), %edx
232 /* Copy the number of remaining bytes to %edi. */
233 movl 0x18(%esp), %edi
236 /* Check for end of input. */
240 /* Calculate CRC of the next input byte. */
246 xorl (%ebx, %ebp, 8), %eax
248 xorl 4(%ebx, %ebp, 8), %edx
253 /* Complement the final value. */
263 #if defined(PIC) || defined(__PIC__)
270 #if defined(__APPLE__) && (defined(PIC) || defined(__PIC__))
272 .section __IMPORT,__pointers,non_lazy_symbol_pointers
273 .L_lzma_crc64_table$non_lazy_ptr:
274 .indirect_symbol LZMA_CRC64_TABLE
277 #elif defined(_WIN32) || defined(__CYGWIN__)
279 /* This is equivalent of __declspec(dllexport). */
281 .ascii " -export:lzma_crc64"
284 #elif !defined(__MSDOS__)
286 .size LZMA_CRC64, .-LZMA_CRC64
290 * This is needed to support non-executable stack. It's ugly to
291 * use __FreeBSD__ and __linux__ here, but I don't know a way to detect when
292 * we are using GNU assembler.
294 #if defined(__ELF__) && (defined(__FreeBSD__) || defined(__linux__))
295 .section .note.GNU-stack,"",@progbits