@certik commit 9ff6aed should be backported to 1.7.x.
TST: skip floating point warning test failures on ARM. Closes gh-413.
MAINT: remove painful to look at exec statement from np.errstate test.
(@certik: just to re-iterate what Ralf said in case you don't look at this closely -- he really does mean only that commit should be backported, the other commit that's included in this PR will break py24 compatibility. This is also why the Travis build failed.)
Am I right that the basic assumption here is that our FP exception checking stuff just doesn't work reliably in general, and so we might as well knownfail the relevant tests whenever this comes up? That's fine if so, but do we have a bug filed anywhere saying this? Seems like something we'd want to figure out how to fix sooner or later and have a central place to accumulate any analysis in the mean time...
Also @rgommers: If you're breaking the test suite on 2.4, can you remove it from .travis.yml as well, so the PR green/grey status remains useful?
Oh yes: actual changes look good to me.
There's gh-2350 and gh-593 at least, maybe there are others. Floating point warnings/errors are behaving different on different platforms. @charris expressed the opinion already that we can't fix this in numpy. Some improvement may be possible though.
TST: remove Python 2.4 and 2.5 testing from Travis config.
Done because for 1.8 we're not supporting these Python versions anymore,
and this PR break 2.4 compatibility.
OK, Travis build is fixed.
I think we should simply change this to test 3.1, which isn't tested by Travis anymore. I think that testing 3.1 comes pretty much for free and it can reveal possible bugs.
After seeing some recent surveys with usage statistics, I'm wondering if we shouldn't just drop 3.1 support after all. Usage was well below 2.4. If we keep it in for 1.8 you have a good point @certik.
Backported in #2883.
@njsmith let me know if I did this right. I always create a PR for the backport, especially in cases like this, where that particular commit didn't have Travis tests attached to it, and the next commit had failures. All seems to be ok.