Why nanobind can build this binding while pybind11 raises "use of undefined type"?

[davidlatwe]

Problem

I found this when making a Python binding for ImGui - docking branch.

The problem starts from here, ImGui's std::vector<>-like class, sizeof() is used:
https://github.com/ocornut/imgui/blob/085781f5ca5372d5fc804d7e44b5bf27a8994af7/imgui.h#L2003

// imgui.h
template<typename T>
struct ImVector {
    inline ImVector<T>& operator=(const ImVector<T>& src)   { clear(); resize(src.Size); if (src.Data) memcpy(Data, src.Data, (size_t)Size * sizeof(T)); return *this; }
}

In its internal header, ImVector is used for storing ImGuiDockNodeSettings.
https://github.com/ocornut/imgui/blob/085781f5ca5372d5fc804d7e44b5bf27a8994af7/imgui_internal.h#L1881

// imgui_internal.h
struct ImGuiDockContext {
    ImVector<ImGuiDockNodeSettings> NodesSettings;
};

But the definition of ImGuiDockNodeSettings is in imgui.cpp file:
https://github.com/ocornut/imgui/blob/085781f5ca5372d5fc804d7e44b5bf27a8994af7/imgui.cpp#L15630

So when I build the binding with pybind11, Windows is complaining:

error C2027: use of undefined type 'ImGuiDockNodeSettings'

(strange that MSVC did not mention anything about sizeof and I thought it was becasue my cmake setup was incorrect.)

And on Linux I saw:

error: invalid application of ‘sizeof’ to incomplete type ‘ImGuiDockNodeSettings’

Question

However, if I build the binding with nanobind, build was success and the binding works perfectly.

What magic did nanobind use?

I made a minimum reproducible repo for showcasing the issue: https://github.com/davidlatwe/py-binding-test

For my own ImGui binding work, I may tweak ImGui source code for pybind11, or just continue using nanobind. But I am really curious about "why this works".

P.S. The title of this thread feels not clear enough, welcome to change it.

Thanks.

[wjakob]

Sorry, I don't see how any of this is related to nanobind based on your description.

[davidlatwe]

Thank you for your reply. I'm investigating why pybind11 and nanobind yield different results when compiling the same binding code. While I understand that pybind11 is header-only and nanobind uses a support library (e.g., nanobind-static.lib), there seems to be more to it.

In my example, this is the binding code that compiles differently with nanobind and pybind11:

#ifdef USE_NANOBIND
#   include <nanobind/nanobind.h>
#   define ADD_MODULE NB_MODULE
#else
#   include <pybind11/pybind11.h>
#   define ADD_MODULE PYBIND11_MODULE
#endif

ADD_MODULE(example, m /*'m' for module*/) {
    m.def("GetContext", GetContext);
}

The function GetContext is returns a pointer to a Context struct:

struct Context {
    Vector<Node> Nodes;
};

Context* GetContext();

In this structure, the Context member Nodes is a template container Vector that holds Node objects.

The issue arises with the definition of Node, which is located in a .cpp file. When compiling with pybind11, the compiler raises an "undefined type" error, which does not occur with nanobind.

Could this discrepancy be due to differences in how nanobind and pybind11 handle template visibility or type linking? I'd appreciate any insights or suggestions on this matter.