source: vbox/trunk/src/libs/liblzma-5.4.1/check/crc32_fast.c

///////////////////////////////////////////////////////////////////////////////
//
/// \file       crc32.c
/// \brief      CRC32 calculation
///
/// Calculate the CRC32 using the slice-by-eight algorithm.
/// It is explained in this document:
/// http://www.intel.com/technology/comms/perfnet/download/CRC_generators.pdf
/// The code in this file is not the same as in Intel's paper, but
/// the basic principle is identical.
//
//  Author:     Lasse Collin
//
//  This file has been put into the public domain.
//  You can do whatever you want with this file.
//
///////////////////////////////////////////////////////////////////////////////

#include "check.h"
#include "crc_macros.h"


// If you make any changes, do some benchmarking! Seemingly unrelated
// changes can very easily ruin the performance (and very probably is
// very compiler dependent).
extern LZMA_API(uint32_t)
lzma_crc32(const uint8_t *buf, size_t size, uint32_t crc)
{
        crc = ~crc;

#ifdef WORDS_BIGENDIAN
        crc = bswap32(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 = bswap32(crc);
#endif

        return ~crc;
}