#include <stddef.h>

#define ALIGNED(p) (!((long)p & (sizeof(long) - 1)))
#define QUADBLOCKSIZE (sizeof(long) << 2)
#define BLOCKSIZE (sizeof(long))

void *memcpy(void *dest, const void *src, size_t sz)
{
    /* Only try word-writing if the pointers are aligned and
     * the buffer is big enough */
    if (sz >= BLOCKSIZE && ALIGNED(dest) && ALIGNED(src)) {
        long *dest_word = dest;
        const long *src_word = src;

        /* First try copying 4x blocks per iteration */
        while (sz >= QUADBLOCKSIZE) {
            *dest_word++ = *src_word++;
            *dest_word++ = *src_word++;
            *dest_word++ = *src_word++;
            *dest_word++ = *src_word++;
            sz -= QUADBLOCKSIZE;
        }

        /* Then try copying 1x words per iteration */
        while (sz >= BLOCKSIZE) {
            *dest_word++ = *src_word++;
            sz -= BLOCKSIZE;
        }

        /* Then go back to the byte-level copying */
        dest = dest_word;
        src = src_word;
    }

    char *dest0 = dest;
    const char *src0 = src;

    while (sz-- > 0) {
        *dest0++ = *src0++;
    }

    return dest;
}