#define SL( v, o ) __shiftleft128( *((v)+1), *((v)), o )

__forceinline
U8 cmpBits( const register U64 *hay, const register U64 *pin, U8 register ohay, U8 register opin, U64 lpin ) {
    U64 register i;

    for( i=0; i < lpin/64; ++i, ++hay, ++pin ) {
        if( SL( hay, ohay ) != SL( pin, opin ) ) return 0;
    }
    {
        U8 lastbits = lpin % 64;
        U64 mask = ( ( 1ull << ( lastbits ) ) -1 ) << ( 64 - lastbits );
        if( ( SL( hay, ohay ) & mask ) != ( SL( pin, opin ) & mask ) ) return 0;
    }
    return 1;
}

U64 bitstrnstrn( const register U64 *hay, const register U64 *pin, U8 register ohay, U8 register opin, U64 lhay, U64 lpin ) {
    U64 register i;

    for( i = ohay; i < lhay+ohay; ++i ) {
        if( cmpBits( hay+(i/64), pin, i%64, opin, lpin ) ) return i;
    }
    return -1;
}