#define bit_for_each64(pos, tmp, ul)                                    \
    for (tmp = ul, pos = ffs64(tmp); pos; tmp &= (tmp - 1),  pos = ffs64(tmp))
    /**/
#define bit_for_each64(pos, tmp, ul)                                    \
    for (tmp = ul, pos = ffs64(tmp); pos; tmp &= (tmp - 1),  pos = ffs64(tmp))