D
Dik T. Winter
> But you could test it for a few numbers within a certain limit of float
> and declare it works in that range. If you suggest to me there might be
> a subtle bug only on 2.339058309583905890 then we might as well give
> up....
Well, you know, actually such things *do* happen. You will probably know
about Intels floating poit bug which was just such kind of bug. I know
of a bug in Cray hardware multiplication where for some pairs of non-zero
input numbers the half-precision product would be zero. Partial testing
does not prove it works also with the non-tested inputs.