Mike,

Can you propose an implementation, for those platforms that haven't yet 
caught up with C99?  Something comparable to the implementation of log1p 
in Python/pymath.c would be appropriate.  Ideally, such an 
implementation would:

- be accurate to within a few ulps across the whole domain,
- be not too long, and not too slow
- have a decent chance of working with strange floating-point formats 
(Python doesn't assume IEEE 754)
- handle IEEE 754 values 'correctly' (i.e., as recommended by Annex F to 
the C99 standard)

It's too late to get this into Python 2.6/3.0, but patches aimed at 2.7 
or 3.1 would be welcome.