I tried applying the patch for this commit, which applied without complaint. But:
python3 ./test_numpy_fncs.py
Segmentation fault (core dumped)

Thread 1 "python3" received signal SIGSEGV, Segmentation fault.
void pybind11::cpp_function::initialize<pybind11_init()::{lambda(xt::pyarray<std::complex > const&)#1}, pybind11::tuple, xt::pyarray<std::complex > const&, pybind11::name, pybind11::scope, pybind11::sibling>(pybind11_init()::{lambda(xt::pyarray<std::complex > const&)#1}&&, pybind11::tuple (*)(xt::pyarray<std::complex > const&), pybind11::name const&, pybind11::scope const&, pybind11::sibling const&)::{lambda(pybind11::detail::function_call&)#3}::_FUN(pybind11::detail::function_call) [clone .lto_priv.229] ()
at /home/nbecker/.local/include/xtensor/xassign.hpp:201
��/home/nbecker/.local/include/xtensor/xassign.hpp:201:6251:beg:0x7fffe0c9f39b

@nbecker could you test with the last update ? If that doesn't work, could you tell me which compiler you use, the version of pybind11 and how you build your extension ?

I did:

1. git clone https://github.com/QuantStack/xtensor-python.git xtensor-python-johan

2. git am --signoff 30.patch
   Applying: complex benchmark
   Applying: complex arrays fixed

3. pybind11 is a bit of a mix:
   Head: fix-import-error merge
   Merge: dean0x7d/fix-import-error Fix handling of Python exceptions during module initialization
   Tag: v2.0.0 (42)
   CI: None

Unmerged into dean0x7d/fix-import-error (2)
d13f015 fix-import-error merge
2d6bc76 jagerman/array_t-overload-resolution array_t overload resolution support

4. compiled as:
   g++ -o numpy_fncs.cpython-35m-x86_64-linux-gnu.os -c -g -std=c++1z -O3 -march=native -fvisibility=hidden -flto -DNDEBUG -ffast-math -fPIC -I/home/nbecker/.local/include -I/usr/include/eigen3 -I/home/nbecker/.local/lib/python3.5/site-packages/numpy/core/include -I/usr/include/python3.5m numpy_fncs.cc

Ok, the problem seems to happen with pybind11 >2.0.1 only (current master). However, the test I've added in the PR does not trigger the bug (even with master branch of pybind11), the crash occurs after many calls.

Since this PR seems to work fine with pybind11 2.0.1, I'll merge it and I'll open a specific issue for this problem.