Pybind11 Weaver: Python Binding Code Generator

Pybind11 Weaver is a powerful code generator designed to automate the generation of pybind11 code from C++ header files. It streamlines the process of creating Python bindings, enabling users to focus on writing critical pybind11 code and offloading the tedious work to Pybind11 Weaver.

This tool takes a c_lib.h file and transforms it into a _binding.cc.inc file using cfg.yaml as a guide. Following the binding with a single line auto update_guard = DeclFn(m); in binding.cc, all elements from the header file become accessible in Python as demonstrated in this example.

A more pragmatic example is available in pylibclang, a comprehensive Python wrapper for libclang that uses Pybind11 Weaver to generate the binding code.

1. Its practicality stems from the fact that Pybind11 Weaver operates on it as well. Indeed, Pybind11 Weaver is self-hosted and generates the binding code for its own use.
2. Approximately 30k lines of C++ code are generated from a mere 10 lines of cfg.yaml.
3. Some binding code is manually crafted to handle special cases and integrates seamlessly with the generated code.

pylibtooling is a much more advanced example that uses Pybind11 Weaver to generate the binding code for libtooling, and will be used to demonstrate the capabilities of Pybind11 Weaver when working with large C++ only libraries.

Docs

Check https://github.com/edimetia3d/pybind11_weaver/wiki

Key Features

1. Highly Customizable: While the default configuration is super simple and suitable for most cases, it allows for high customization.
2. Ease of Use: As a pure Python package, a simple pip install gets it ready to work.
3. Versatility: All generated code is under your control, you can easily modify/enhance/disable any part of generated code, and all generated code will work with your hand-written code seamlessly.
4. Structure Preservation: It retains the module structure of the original C++ code.

Features & Roadmap

• Binding for Enum
• Binding for Namespace (as submodule)
• Binding for Function, with support of function overloading
• Binding for C style function pointer (usually used as callback functions)
• Binding for opaque pointer and pointer to incomplete type
• Binding for Operator overloading
• Binding for Class method, method overloading, static method, static method overloading, constructor, constructor overloading, class field
• Trampoline class for virtual function
• Binding for concreate template instance, that includes: implicit(explicit) class(struct) template instantiation, full class(struct) template specialization, extern function template instance declaration.
• Support class inheritance hierarchy
• Auto ignore symbols by : Linkage (e.g. static), visibility (e.g. visibility=hidden), member access control (e.g. private, protected)
• Docstring generation from c++ doxygen style comment
• Namespace hierarchy to Python module hierarchy
• Dynamic update/disable binding by API call.
• Static update/disable binding by define macro (Mainly used to disable wrong binding code to avoid compilation error)
• Auto snake case

Background & Recommendations

This project originated from an internal project aimed at creating a Python binding for a LARGE developing C++ library. This posed significant challenges:

1. The C++ library interface contained a vast number of classes, functions, and enums. Creating bindings for all these elements was not only tedious but also error-prone.
2. Because the C++ library was under active development, staying updated with daily additions and frequent code modifications was a maintenance challenge.
3. Some aspects of the C++ library, due to historical reasons, were incompatible with Python conventions, necessitating hand-written binding codes.
4. The sheer size of the library added to the complexity, making it difficult to develop a generator smart enough to handle everything, hence the need for manual binding code writing.

In light of these challenges, I designed Pybind11 Weaver as a tool to generate the majority of the binding code, leaving users to handcraft the remaining parts as needed. If this approach suits your needs, this tool will be a valuable asset.

Typical workflow:

Though most features should work out of the box, the more your API looks like "C With Class", the higher chance Pybind11 Weaver will do all the work for you. If you use too many advanced C++ features, you may need to write some binding code by yourself.

1. Create a cfg.yaml file, mainly to tell the generator which files to parse.
2. Use Pybind11 Weaver to generate files, like pybind11-weaver --config cfg.yaml.
3. Create a binding.cc, include the generated files, and call the binding code.
4. Disable some generated binding code by define some macro, if there is any compilation error.
5. Add some custom code to replace part of the generated code, or adding some new binding that generator had not exported.
6. Compile all code into a pybind11 module.
7. Optionally, use pybind11-stubgen to generate .pyi stub files, enhancing readability for both humans and MYPY in a static way.
8. Test the module in Python, find bugs, and go to step 5 to fix them.

Also, if you encountered too many problems, you are welcome to open an issue at github, or create a PR to fix it.

How it works

The Pybind11 Weaver operates under the hood by utilizing libclang, a library that parses C++ header files. This enables us to obtain all APIs from the header file, which are then used to generate the binding code on your behalf.

Notably, only header files are required, as we need declarations, not definitions. However, to ensure accurate parsing of the code, some compiler flags, especially for macros, are necessary.

The code generated is structured into a struct:

1. During the construction of the struct, it creates some Pybind11 objects, such as pybind11::class_ or pybind11::enum_.
2. When the Update() API is invoked, the Pybind11 object experiences an update.

The use of a struct permits us to:

• Separate the processes of object creation and updates, ensuring that Pybind11 consistently acknowledges all exported classes, which aids in the generation of accurate documentation.
• Increase the readability of the generated code, making it simpler to debug.
• Simplify customization, as you can easily inherit the struct and override or reimplement necessary elements.