Question: Can he improve the floating point computational accuracy? Are there any compiler options etc.?

Answer: c40 has 24 bits of mantissa in floating pt. The compiler cannot go beyond this.