Unedfined symbol pybind11::object::borrowed

[uentity]

Hi devs!

Faced topic issue when building modules with Python bindings. Seems that this is related to my custom array class that is inherited from pybind11::array_t. It's a templated type that imports (inherits) all constructors from array_t, but couple of methods are only declared. Module1 contains explicit specializations of these methods (from my custom array) and have no problems.
Then, I use array in Module2 and it fails to load because of the issue above. I think the same would apply for pybind11::object::stolen, but I just didn't use it -- my array code contains reinterpret_borrow<>.

Right now I have a situation when release build does not have undefined symbol, but debug build does. In some other Module3 both release and debug builds have undefined symbol. Tested with gcc 6.3.1 and clang 3.9.1 -- results are the same.

Any ideas how to fix this?

[jagerman]

Can you be a bit more specific as to the missing symbols, or a (small) code sample that isn't working (so we can add a test for it)? It's a bit hard to guess what specifically might be wrong from a general description.

[jagerman]

Oh, never mind, I didn't read the issue subject!

[jagerman]

Cc @dean0x7d

I don't see an easy fix to this (aside from changing all object::borrowed to object::borrowed_t{}): borrowed/stolen are constexpr declarations without definitions which can (typically when compiling without optimizations and/or in debug mode) trigger a linkage failure. In C++17 we could add a inline const object::borrowed_t object::borrowed; definition, but in earlier C++ standards that's not allowed. We could add a templated class (with a dummy template parameter) in between object and handle to allow an inline definition, but that's a bit messy.

[jagerman]

@uentity - I can't seem to reproduce this easily (I'd still appreciate some sample test code!), but I think #771 may fix this; can you give it a try?

[dean0x7d]

@uentity How did you define the borrow and steal constructors? I'm guessing the definition and declaration are split between header and source files. If that's the case, moving just those constructors into the header should work around the issue. (The underlying issue should still be resolved in pybind11.)

[uentity]

@jagerman yes, #771 fixes my undefined symbol issue.
@dean0x7d - I didn't define borrow and steal constructors. I just imported them (inherited) from base array_t class with using base_t::base_t, where base_t is array_t. Class itself is a template, so there is no split, all definitions are provided in header except for a couple of methods.

[uentity]

Btw, I've read a little on constexpr subject and don't understand why compiler got confused here. Everything seems pretty legit. There is no single place in my code where I mention object::borrowed or object::borrowed_t explicitly. pybind11 internals also seems to be fine at a glance - nobody tries to take an address of borrowed.

[jagerman]

It's more common to see this sort of thing on non-optimizing debug builds. Compilers don't have to replace constexpr values with the literal value, but it's an easy optimization to make, so most do. But under -O0 (i.e. no optimization), the compiler sometimes references rather than substituting the constexpr value.

[jagerman]

I updated #771 with an alternate fix that stops using borrowed in pybind11, so your inherited constructors will no longer reference it at all. It could still be a potential issue if you try to use object::borrowed directly, but as long as you aren't you should be fine (and you'll get deprecation warnings if you try).

[uentity]

@jagerman I already checked out your work. No issues so far.
Great project and great responsiveness, thank you guys!