worst cases for exp() in the decimal64 format

[Paul.Zimmermann@xxxxxxxx (Paul Zimmermann)]

[Sorry to be off-topic with respect to the current rush for the final version.]

Together with Vincent Lefèvre and Damien Stehlé, we have conducted a full
search for worst cases of exp() in the (hopefully-not-to-be-changed) decimal64
format. The full results, including bad cases at a distance less than 10^(-15)
ulp, are available from the paper (to appear in a LNCS special issue):

     http://www.loria.fr/~zimmerma/papers/index.html#decimal64

This will enable to efficiently implement correct rounding of exp() in the
decimal64 format, for example in libraries such as decNumber.

Of course, the method can be applied to other functions. It is mainly a matter
of computer and human time.

The main remaining problem (for the computation of worst cases) is sin(BIG).
We are working on this.

Paul Zimmermann