Enclosure performance of pow vs. timing performance. #1
Description
This might be a bit anecdotal, but I suspect that the current implementation of pow might be trading away smaller enclosures for reduced run time.
Right now, I've been evaluating a large number of 10-15 degree polynomials. In all cases that I've considered, I've found that rewriting the polynomial to multiply out expressions rather than using pow(double, int) produces a smaller enclosure.
I think this might be because pow is using a non-obvious algorithm (as noted in the changelog); in particular, involving the sqrt function is (I suspect) suboptimal for enclosures than standard multiplication.
This isn't necessarily a bug as the library is free to prioritize speed over enclosure size, but I think it is at least worth talking about.