Tests fail on i686 and some other architectures

[susilehtola]

I'm trying to package pybind11 in Fedora, since I need it to update the Psi4 program package.

While the tests run nicely through on x86_64, armv7hl and ppc64le, they fail on i686 and ppc64. i686 build log attached.

build.txt

[jagerman]

PR #619 fixed this (on master, not in 2.0.1), at least for i386 (I don't know whether that also fixed ppc64).

[susilehtola]

It does indeed fix the issue on i686. However, now the tests crash on x86_64!

build.txt

[susilehtola]

Looks like some memory issue, the crash occurs only intermittently.

[jagerman]

Is the crash with just #619 cherry-picked into 2.0.1? It might depend on other things in master. You're probably better off just disabling the test_numpy_dtypes.py test on i386: it wasn't a bug in pybind11 itself, but rather one in the assumptions made in the test strings.

[susilehtola]

Well, given that the crash is intermittent, I can't actually say that it's related to #619. It could be some dormant issue.

I'd like to run the tests through valgrind to check for memory violations, but I'm not sure how to do that.

[jagerman]

Build normally, then: cd tests; valgrind python3 -m pytest should work.

[susilehtola]

There seems to be another problem on ppc64 in test_numpy_array.py:96.
build.txt

[jagerman]

Those failures don't look harmless. I can't reproduce this on ppc64le (via QEMU) in either 2.0.1 or current master. Perhaps an endian issue?

[susilehtola]

Probably, since it succeeds on ppc64 also on the Fedora build system. It's only ppc64le that fails.

[jagerman]

Did you mean that the other way around? My build in a QEMU ppc64le worked fine.

[susilehtola]

.. Yes.
https://koji.fedoraproject.org/koji/taskinfo?taskID=18032794

[jagerman]

I've reproduced it with a ppc64 QEMU image. The failures are bugs in the test suite, not pybind; I'll have a fix shortly.

[jagerman]

(The 20+ minute compile time for pybind running under QEMU/ppc64 is kind of a nuissance).