0x7FDE623822FC16E6 : a magic constant for double float reciprocal

If you want to compute lots of reciprocals, don't mind some rounding errors, and have reasonable-sized values, you can compute them with one reciprocal and a few multiplications, e.g.,

  tmp = 1 / (a * b)
  oneovera = tmp * b
  oneoverb = tmp * a