Math.signum(double d) and Math.signum(float f) are ~1.2-1.5x slower on positive and negative numbers than on zero and NaN. The case for majority of numbers can be improved with reasonable degradation in special cases.