[Python-Dev] RE: Possible bug (was Re: numpy, overflow, inf, ieee, and rich comparison)

[Tim Peters]

[Mark Favas]
> Just a little unreality check - the whole world is not (yet) Linux - or
> even MS. The behaviour of math.exp(-746) on my flavours of 1.5.2 on
> Solaris boxes and DEC Alphas (Tru64 Unix) is the same as the behaviour
> of 2.0 - OverflowError: math range error.

This is unfortunate.  Guido never intended for Python to raise exceptions on
underflow (and especially not *Over*flowError(!)), but he apparently didn't
run on any platforms where this happened (I never have!), and AFAIK nobody
ever mentioned this before.

> Thus, on these platform, the proposed fixes _will_ change behaviour.

See "unfortunate" above <0.5 wink>.

> But that's cool - I like Tim's proposed, um, steps to "rationalizing"
> fp behaviour, and will do what I can to help in the future.

Want to fund a crack group of 754 experts for a year?  Didn't think so.
Constructive complaints will help a lot.  So will beta testing.  For
example, I'm very keen to know whether the current CVS version of Python
(2.0c1) still raises OverflowError for math.exp(-746) on your Solaris and/or
Tru64 Unix boxes.  Running the (current CVS) std test suite will tell you:
test_libm will pass or fail accordingly.  If it's still busted on one of
your systems, we've only got about a day to fix it (2.0 final is shipping
early next week, and we have to let the codebase settle down for a couple
days before building the final release).

but-since-it's-been-a-random-mess-for-a-decade-already-i'm-not-
    losing-much-sleep-over-it-ly y'rs  - tim