gmp and its friends mpfr and mpc are used for all numerical calculations. You could perhaps set bignum-precision to 64 to mimic C doubles, and get faster float calculations.