You can clone with
Get this with NumPy 1.6.2 (32 bit) on Windows 7:
>> np.seterr('raise');print np.logaddexp([-inf], [-inf])
FloatingPointError: invalid value encountered in logaddexp
The correct result log(0) = -inf is returned, but I think this is a common enough use-case that logaddexp() should account for it without raising any warnings/exceptions.
log(0) = -inf
The result is correct, but the FP exception should either be overflow rather than invalid, or be absent.
Following the logic in C99, the correct choice would be to not raise an exception.