Converting fixed point numbers

Hello,

madplay uses the following in audio_linear_round

  /* quantize and scale */
  return sample >> (MAD_F_FRACBITS + 1 - bits);

if bits is 16 (unsigned short e,g) we will shift 13 bits, whats about 
the 3 MSB bits ???

With this I think only the first digit of the whole part is taken, what 
is with the sign and the other 2 digits from the whole part, are they 
never used?

Thanks

Reiner