Fixed potential overflow exception in the lines.contains() method

[dhyams]

Fixed potential overflow exception in the lines.contains() method, if the application happened to have set the numpy error flags to raise exceptions in case of overflow.

[pelson]

Ok. I have no problem with the changes you have proposed, but I am interested how this manifests itself. Knowing this information, is it possible to pinpoint the problem to a single line or do both cases of the if statement need to be handled?

[dhyams]

It manifests when you have a function that has values of some huge numbers
(as can happen when you evaluate, say 1/x near zero), and if the
application has set the numpy error flags such that exceptions will be
raised on overflow.

It's the squaring of (xt-mouseevent.x) that trips the overflow exception if
the above is true.

The patch basically sets and restores the numpy error flags just during the
hit testing calculation, so that overflows are OK. It seems that
(understandably so) that matplotlib doesn't care much about overflow,
because the default numpy error flags are

{'over': 'warn', 'divide': 'warn', 'invalid': 'warn', 'under': 'ignore'}

and that's the environment that mpl operates under nearly all of the time.

I might be going off on a rabbit trail here, but in my opinion, the nose
tests should be run with more strict numpy numerical treatment:

{'over': 'raise', 'divide': 'raise', 'invalid': 'raise', 'under': 'ignore'}

I tried to get in there and do this myself to see what the results are, but
it looked like I would have to decorate every single test function in
matplotlib.tests. I might ask on the user group and/or try again later.

On Wed, Jul 18, 2012 at 2:33 PM, Phil Elson <
reply@reply.github.com

wrote:

Ok. I have no problem with the changes you have proposed, but I am
interested how this manifests itself. Knowing this information, is it
possible to pinpoint the problem to a single line or do both cases of the
if statement need to be handled?

---

Reply to this email directly or view it on GitHub:
#1002 (comment)

Daniel Hyams
dhyams@gmail.com

[dhyams]

Oh, and sorry, to answer your question about the two branches, yes both need to be covered; both involve squaring (xt-mouseevent.x) and (yt-mouseevent.y)

[WeatherGod]

This looks good to me. I see no reason not to merge. Thanks for this!