11706 vs ([lno] Canonical iv creation)

[Paolo Carlini]

Paolo Carlini wrote:

> It would be an improvement, a huge improvement.

Perhaps now I see a situation you care about: non-constant integer. In 
that case, ::pow
issues a library call (irrespective of -ffast-math), whereas currently 
std::pow uses the
binary algorithm which *may* be faster.

The price we pay for this possible - only possible, I haven't really 
looked at the
assembler or benchmarked, yet - better performance in this case is that 
we use the
binary algorithm *always* not only when -ffast-math.

And I see now where a __builtin_powi would be the optimal solution.

In the current situation, I still believe that all-in-all using simply 
the builtins, for
pow(T, n) too is better, but I'll try to come up with some numbers.

Paolo.