Fix for #3789, segfault in _tri #3797
Conversation
|
@mdboom: It looks like I spoke too soon about knowing what I was doing with |
|
Could the differences with respect to numpy be due to numpy/numpy#5203 which relates to some strange sharing of masks in Numpy between 1.7.0 and 1.9.0 (fixed in numpy 1.9.1) We have been bitten by this before in a different context. |
|
@jenshnielsen: I don't think so. Although the problem arrays are known as 'masks', they are standalone boolean arrays rather than masks of numpy MaskedArrays. Plus it looks like the problems occurring on travis all used numpy 1.6.1. But I will be keeping an eye on this in case I am wrong. |
|
@ianthomas23 your probably right. It just crossed my mind when I saw the changes. |
|
I've checked all of the travis logs and this PR does indeed fix the problem, so is ready to be merged. |
|
Great |
Fix for #3789, segfault in _tri
Candidate fix for issue #3789. Segfaults were occurring when travis was running
test_tritools, but only under python 3. This is an esoteric one - my local tests all ran fine under python 3 with both old and new gcc and numpy, so I could only debug by pushing candidate fixes to my github mpl repository and running travis on them there. That's a new form of remote debugging for me.The problem occurs accessing const/non-const boolean arrays wrapped in the new numpy::array_view classes. It shouldn't occur, but evidently some combination of hardware, gcc and python versions behave differently.
I will need to go through the travis logs with a fine toothcomb before this is merged.
@jenshnielsen: Thanks for reporting this.