Hi,

On OpenBSD 4.4, the test_math.py regression test fails with the following:

Traceback (most recent call last):
  File "Lib/test/test_math.py", line 419, in testLog
    self.assertRaises(ValueError, math.log, NINF)
AssertionError: ValueError not raised

This is because libm's log function does not return NaN when passed
-INFINITY. It returns -INFINITY and sets EDOM. This behaviour seems to
be permitted by the C99 spec (ISO/IEC 9899:1999):

> 7.12.1 Treatment of error conditions
> paragraph2:
> ... On a domain error, the function returns an implementation-defined
> value; if the integer expression math_errhandling & MATH_ERRNO is 
> nonzero, the integer expression errno acquires the value EDOM; 

in mathmodule.c:math_1() errno seems to be deliberately ignored when the
result is infinite. The attached patch modified math_1() to retain EDOM
errors for infinite results.

It's possible that this patch would cause breakage on other systems: 
there's a reason that errno is currently ignored, which is that it can't
be trusted on all systems (notably Linux: on one Linux system I've used
of three different evaluations, all of log singularity type (log(0),
log1p(-1), atanh(1), ...), one gave errno=EDOM, one gave errno=ERANGE,
and one didn't set errno at all).

So I'm reluctant to mess with math_1, especially this close to a release.

An alternative solution would be to check special cases for log directly
within the mathmodule.c code, only passing positive nonspecial arguments
to the system log function.  This would have a much lower risk of
causing breakage on other systems.  Note that Solaris with Sun's
compiler also has some problems with log:  depending on compiler
options, etc., log(-1.0) can give -inf instead of nan.  So having Python
deal with log special cases directly seems like a good thing to do there
as well.