On a tangent, I just discovered you can remote-desktop into the appveyor VM during the build... https://www.appveyor.com/docs/how-to/rdp-to-build-worker/ handy.

xref #9176

-1073740791 == c0000409 == STATUS_STACK_BUFFER_OVERRUN; I've not had much luck triggering any error in valgrind on Linux, though it's a bit hard to sort through. I added pytest-faulthandler to AppVeyor here, but I'm not sure if it'll be effective.

According to UBsan, there is a memcpy on NULL in FreeType. I am not sure if that is the cause of the problem here, but a long time ago Windows printf didn't like NULL strings so it might be the reason. (It's also triggered by a few other tests, but test_cull_markers is the first one.)

It is fixed in FreeType 2.6.3, but using that version causes 930 failures. Also, SF is not redirecting properly so we can't download it onto builders right now (it's working from cache, mostly). I'm not sure how we want to go about fixing it.

Change the download logic (setupext.py line 1181 etc.) so that on a hash mismatch the second url gets tried? Or simpler, just prefer downloading from savannah...
Although interestingly https://www.freetype.org/download.html is down too.

https://twitter.com/Slav636/status/912906308889825280 for a pointer to https://www.microsoft.com/en-us/download/confirmation.aspx?id=985 to look up details.

There's a table here for that.

The two appevyor fails on this PR are in a different place than the rest (after all of the tests passed) or the freetype issue.

https://ci.appveyor.com/project/matplotlib/matplotlib/build/1.0.3136
https://ci.appveyor.com/project/matplotlib/matplotlib/build/1.0.3149