2 * Speed-optimized CRC32 using slicing-by-eight 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). For i586
6 * (e.g. Pentium), slicing-by-four would be better, and even the C version
7 * of slicing-by-eight built with gcc -march=i586 tends to be a little bit
8 * better than this. Very few probably run this code on i586 or older x86
9 * so this shouldn't be a problem in practice.
11 * Authors: Igor Pavlov (original version)
12 * Lasse Collin (AT&T syntax, PIC support, better portability)
14 * This file has been put into the public domain.
15 * You can do whatever you want with this file.
17 * This code needs lzma_crc32_table, which can be created using the
20 uint32_t lzma_crc32_table[8][256];
26 static const uint32_t poly32 = UINT32_C(0xEDB88320);
29 // static const uint32_t poly32 = UINT32_C(0x82F63B78);
32 // static const uint32_t poly32 = UINT32_C(0xEB31D82E);
34 for (size_t s = 0; s < 8; ++s) {
35 for (size_t b = 0; b < 256; ++b) {
36 uint32_t r = s == 0 ? b : lzma_crc32_table[s - 1][b];
38 for (size_t i = 0; i < 8; ++i) {
40 r = (r >> 1) ^ poly32;
45 lzma_crc32_table[s][b] = r;
50 * The prototype of the CRC32 function:
51 * extern uint32_t lzma_crc32(const uint8_t *buf, size_t size, uint32_t crc);
54 /* When Intel CET is enabled, include <cet.h> in assembly code to mark
63 * On some systems, the functions need to be prefixed. The prefix is
64 * usually an underscore.
66 #ifndef __USER_LABEL_PREFIX__
67 # define __USER_LABEL_PREFIX__
69 #define MAKE_SYM_CAT(prefix, sym) prefix ## sym
70 #define MAKE_SYM(prefix, sym) MAKE_SYM_CAT(prefix, sym)
71 #define LZMA_CRC32 MAKE_SYM(__USER_LABEL_PREFIX__, lzma_crc32)
72 #define LZMA_CRC32_TABLE MAKE_SYM(__USER_LABEL_PREFIX__, lzma_crc32_table)
75 * Solaris assembler doesn't have .p2align, and Darwin uses .align
76 * differently than GNU/Linux and Solaris.
78 #if defined(__APPLE__) || defined(__MSDOS__)
79 # define ALIGN(pow2, abs) .align pow2
81 # define ALIGN(pow2, abs) .align abs
87 #if !defined(__APPLE__) && !defined(_WIN32) && !defined(__CYGWIN__) \
88 && !defined(__MSDOS__)
89 .type LZMA_CRC32, @function
99 * %edi size or buf + size
100 * %ebx lzma_crc32_table
109 movl 0x14(%esp), %esi /* buf */
110 movl 0x18(%esp), %edi /* size */
111 movl 0x1C(%esp), %eax /* crc */
114 * Store the address of lzma_crc32_table to %ebx. This is needed to
115 * get position-independent code (PIC).
117 * The PIC macro is defined by libtool, while __PIC__ is defined
118 * by GCC but only on some systems. Testing for both makes it simpler
119 * to test this code without libtool, and keeps the code working also
120 * when built with libtool but using something else than GCC.
122 * I understood that libtool may define PIC on Windows even though
123 * the code in Windows DLLs is not PIC in sense that it is in ELF
124 * binaries, so we need a separate check to always use the non-PIC
127 #if (!defined(PIC) && !defined(__PIC__)) \
128 || (defined(_WIN32) || defined(__CYGWIN__))
130 movl $ LZMA_CRC32_TABLE, %ebx
131 #elif defined(__APPLE__)
135 leal .L_lzma_crc32_table$non_lazy_ptr-.L_pic(%ebx), %ebx
140 addl $_GLOBAL_OFFSET_TABLE_, %ebx
141 movl LZMA_CRC32_TABLE@GOT(%ebx), %ebx
144 /* Complement the initial value. */
150 * Check if there is enough input to use slicing-by-eight.
151 * We need 16 bytes, because the loop pre-reads eight bytes.
156 /* Check if we have reached alignment of eight bytes. */
160 /* Calculate CRC of the next input byte. */
166 xorl (%ebx, %ebp, 4), %eax
173 * If we get here, there's at least 16 bytes of aligned input
174 * available. Make %edi multiple of eight bytes. Store the possible
175 * remainder over the "size" variable in the argument stack.
177 movl %edi, 0x18(%esp)
179 subl %edi, 0x18(%esp)
182 * Let %edi be buf + size - 8 while running the main loop. This way
183 * we can compare for equality to determine when exit the loop.
188 /* Read in the first eight aligned bytes. */
194 movl 0x0C00(%ebx, %ebp, 4), %edx
196 xorl 0x0800(%ebx, %ebp, 4), %edx
200 xorl 0x0400(%ebx, %ebp, 4), %edx
202 xorl (%ebx, %ebp, 4), %edx
206 * Read the next four bytes, for which the CRC is calculated
207 * on the next iteration of the loop.
211 xorl 0x1C00(%ebx, %ebp, 4), %edx
214 xorl 0x1800(%ebx, %ebp, 4), %edx
217 movl 0x1400(%ebx, %eax, 4), %eax
220 xorl 0x1000(%ebx, %ebp, 4), %eax
222 /* Check for end of aligned input. */
228 * Process the remaining eight bytes, which we have already
229 * copied to %ecx and %edx.
231 movl 0x0C00(%ebx, %ebp, 4), %edx
233 xorl 0x0800(%ebx, %ebp, 4), %edx
236 xorl 0x0400(%ebx, %ebp, 4), %edx
238 xorl (%ebx, %ebp, 4), %edx
241 xorl 0x1C00(%ebx, %ebp, 4), %edx
244 xorl 0x1800(%ebx, %ebp, 4), %edx
247 movl 0x1400(%ebx, %eax, 4), %eax
250 xorl 0x1000(%ebx, %ebp, 4), %eax
252 /* Copy the number of remaining bytes to %edi. */
253 movl 0x18(%esp), %edi
256 /* Check for end of input. */
260 /* Calculate CRC of the next input byte. */
266 xorl (%ebx, %ebp, 4), %eax
271 /* Complement the final value. */
280 #if defined(PIC) || defined(__PIC__)
287 #if defined(__APPLE__) && (defined(PIC) || defined(__PIC__))
289 .section __IMPORT,__pointers,non_lazy_symbol_pointers
290 .L_lzma_crc32_table$non_lazy_ptr:
291 .indirect_symbol LZMA_CRC32_TABLE
294 #elif defined(_WIN32) || defined(__CYGWIN__)
296 /* This is equivalent of __declspec(dllexport). */
298 .ascii " -export:lzma_crc32"
301 #elif !defined(__MSDOS__)
303 .size LZMA_CRC32, .-LZMA_CRC32
307 * This is needed to support non-executable stack. It's ugly to
308 * use __FreeBSD__ and __linux__ here, but I don't know a way to detect when
309 * we are using GNU assembler.
311 #if defined(__ELF__) && (defined(__FreeBSD__) || defined(__linux__))
312 .section .note.GNU-stack,"",@progbits