template <int N> struct positive_power { template <typename T> static float result(T base) { return (N%2) ? positive_power<1>::result(base) * positive_power<N-1>::result(base) : positive_power<2>::result( positive_power<N/2>::result(base) ); } };
I have no idea whether that would work better or not.
Basically, it gives the same performance without the need of a specialization with 3. But this time you have to specialize with 2 if you don't want your code to enter into an infinite loop in that case (power<2>(n) = power<2>(power<1>(n)) = power<2>(power<1>(power<1>(n))), etc....). My tests show that your solution is better for some exponents and less good for some others, so I don't really know which one to choose. I will do the same analysis on a few other platforms to make a decision. I'd also
wonder whether -ffast-math or similar flags make a difference.
Nope, there's no difference with this additional flag on my computer. Thanks Bruno