correctly-rounded elementary transcendental functions

[David Hough]

Most of the IBM work in this area was, I think, for system 370, motivated
by a desire to have identical results in scalar and vector mode.  I don't
know if any of that was applied to R6000.

As for inconsistent performance, you can run into problems occasionally
with unlucky customers - slow subnormals on Sun-3 FPA and Sun-4 
motivated me to provide a faster "nonstandard" mode - but it's seldom significant
on realistic applications.  Another reason to use those for benchmarking.

The good news about a pow() or any other function that is correctly-rounded
is that it can be composed of a bunch of unrelated algorithms for special
cases, without monotonicity problems.  
You would ordinarily have qualms about losing monotonicity if you
only used nearly-correctly-rounded algorithms.

The bad news, of course, about correctly-rounded quadruple-precision pow,
for instance, is it will be pretty slow no matter how you implement it.