• python2.7
• py-yaml
• cheetah (for target language templates)
• libclang, from clang 3.1

If using macports, you can easily install all requirements above.

sudo port install python27 py27-yaml py27-cheetah clang-3.1 llvm-3.1

You should also modify the generator script, to point to the directory where libclang.dyld is located.

Usage: generator.py [options] {configfile}

Options:
  -h, --help   show this help message and exit
  -s SECTION   sets a specific section to be converted
  -t TARGET    specifies the target vm. Will search for TARGET.yaml

Basically, you specify a target vm (spidermonkey is the only current target vm) and the section from the .ini file you want to generate code for.

The .ini file is a simple text file specifying the settings for the code generator. Here's the default one, used for cocos2d-x

[cocos2d-x]
prefix = cocos2dx
events  = CCNode#onEnter CCNode#onExit
extra_arguments = -I../../cocos2dx/include -I../../cocos2dx/platform -I../../cocos2dx/platform/ios -I../../cocos2dx -I../../cocos2dx/kazmath/include -arch i386 -DTARGET_OS_IPHONE -isysroot /Applications/Xcode.app/Contents/Developer/Platforms/iPhoneSimulator.platform/Developer/SDKs/iPhoneSimulator5.1.sdk -x c++
headers = ../../cocos2dx/include/cocos2d.h
classes = CCSprite
functions = my_free_function

• prefix: the prefix for the project. Must be a valid identifier for the language of the target vm. Most of the time, the name will be intermixed between the class name and the function name, since all generated (probably) will be free functions, we do that in order to avoid name collition. The script will generate ${prefix}.cpp and ${prefix}.hpp as a result.
• events: a list of identifiers in the form of ClassName#functionName that are events to be called from the native world to the target vm.
• extra_arguments: extra arguments to pass to the clang interface. Basically you can think of this as the arguments to pass to the "compiler", so add as many as you need here. If you're targetting C++, make sure you add "-x c++" as the last argument to force C++ mode on a ".h" file. Otherwise, name your header files as ".hpp".
• headers: list of headers to parse. Usually you add a single header that in turn #includes the rest of the files.
• classes: the classes that will be parsed. Right not is just a string, but it will accept regular expressions
• functions: space-separated list of free functions to be binded. Same as with classes, it will support regular expressions.
• skip: a space-separated list of Classes::functions or just functions to not generate any code.

The generator is using Cheetah templates to create a more flexible generator. The way it was thought, is that for every target environment, you should provide with a way to generate the same C/C++ functionality. Every template has access to the proper meta information for the code or generator (function, classes, etc.)

Right now it's separated in the following set of templates:

• prelude.c/.h: The header of the generated files.
• ifunction.c/.h: The template for an instance function
• ifunction_overloaded.c: The template for the implementation of an overloaded function. An overloaded function is exactly the same as a function, but it has an array of functions sharing the same name. The current implementation for spidermonkey only works if the overloading is with different number of arguments.
• sfunction.c/.h: The template for a static function
• sfunction_overloaded.c: The template for an overloaded static function
• register.c: Here you should add the constructor/finalizer, the registration function (if needed) and the footer of the header file. This is the last chunk being generated

Templates are stored in the templates/${target} directory and follow the naming specified above.

One final part of the puzzle is the ${target}.yaml file, that contains specific type conversion snippets to be used by the templates. For instance, for spidermonkey, this is the place where we specify the conversion routines for the native types (to and from int, float, string, etc.)

Currently the generator is leveraging clang in order to get information about the C/C++ code, so we can only get as much information as clang give us. Known list of things that won't work:

• variable number of arguments. Solution: write a manual wrapper
• default arguments. Solution: write another function that actually calls the real function with the default argument.