|Title:||Running extension modules using the -m option|
|Author:||Marcel Plch <gmarcel.plch at gmail.com>, Petr Viktorin <encukou at gmail.com>|
This PEP proposes implementation that allows built-in and extension modules to be executed in the __main__ namespace using the PEP 489 multi-phase initialization.
With this, a multi-phase initialization enabled module can be run using following command:
$ python3 -m _testmultiphase This is a test module named __main__.
Currently, extension modules do not support all functionality of Python source modules. Specifically, it is not possible to run extension modules as scripts using Python's -m option.
The technical groundwork to make this possible has been done for PEP 489, and enabling the -m option is listed in that PEP's “Possible Future Extensions” section. Technically, the additional changes proposed here are relatively small.
Extension modules' lack of support for the -m option has traditionally been worked around by providing a Python wrapper. For example, the _pickle module's command line interface is in the pure-Python pickle module (along with a pure-Python reimplementation).
This works well for standard library modules, as building command line interfaces using the C API is cumbersome. However, other users may want to create executable extension modules directly.
An important use case is Cython, a Python-like language that compiles to C extension modules. Cython is a (near) superset of Python, meaning that compiling a Python module with Cython will typically not change the module's functionality, allowing Cython-specific features to be added gradually. This PEP will allow Cython extension modules to behave the same as their Python counterparts when run using the -m option. Cython developers consider the feature worth implementing (see Cython issue 1715 ).
Python's -m option is handled by the function runpy._run_module_as_main.
The module specified by -m is not imported normally. Instead, it is executed in the namespace of the __main__ module, which is created quite early in interpreter initialization.
For Python source modules, running in another module's namespace is not a problem: the code is executed with locals and globals set to the existing module's __dict__. This is not the case for extension modules, whose PyInit_* entry point traditionally both created a new module object (using PyModule_Create), and initialized it.
Since Python 3.5, extension modules can use PEP 489 multi-phase initialization. In this scenario, the PyInit_* entry point returns a PyModuleDef structure: a description of how the module should be created and initialized. The extension can choose to customize creation of the module object using the Py_mod_create callback, or opt to use a normal module object by not specifying Py_mod_create. Another callback, Py_mod_exec, is then called to initialize the module object, e.g. by populating it with methods and classes.
Multi-phase initialization makes it possible to execute an extension module in another module's namespace: if a Py_mod_create callback is not specified, the __main__ module can be passed to the Py_mod_exec callback to be initialized, as if __main__ was a freshly constructed module object.
One complication in this scheme is C-level module state. Each module has a md_state pointer that points to a region of memory allocated when an extension module is created. The PyModuleDef specifies how much memory is to be allocated.
The implementation must take care that md_state memory is allocated at most once. Also, the Py_mod_exec callback should only be called once per module. The implications of multiply-initialized modules are too subtle to require expecting extension authors to reason about them. The md_state pointer itself will serve as a guard: allocating the memory and calling Py_mod_exec will always be done together, and initializing an extension module will fail if md_state is already non-NULL.
Since the __main__ module is not created as an extension module, its md_state is normally NULL. Before initializing an extension module in __main__'s context, its module state will be allocated according to the PyModuleDef of that module.
While PEP 489 was designed to make these changes generally possible, it's necessary to decouple module discovery, creation, and initialization steps for extension modules, so that another module can be used instead of a newly initialized one, and the functionality needs to be added to runpy and importlib.
A new optional method for importlib loaders will be added. This method will be called exec_in_module and will take two positional arguments: module spec and an already existing module. Any import-related attributes, such as __spec__ or __name__, already set on the module will be ignored.
The runpy._run_module_as_main function will look for this new loader method. If it is present, runpy will execute it instead of trying to load and run the module's Python code. Otherwise, runpy will act as before.
importlib's ExtensionFileLoader will get an implementation of exec_in_module that will call a new function, _imp.exec_in_module.
_imp.exec_in_module will use existing machinery to find and call an extension module's PyInit_* function.
The PyInit_* function can return either a fully initialized module (single-phase initialization) or a PyModuleDef (for PEP 489 multi-phase initialization).
In the single-phase initialization case, _imp.exec_in_module will raise ImportError.
In the multi-phase initialization case, the PyModuleDef and the module to be initialized will be passed to a new function, PyModule_ExecInModule.
This function raises ImportError if the PyModuleDef specifies a Py_mod_create slot, or if the module has already been initialized (i.e. its md_state pointer is not NULL). Otherwise, the function will initialize the module according to the PyModuleDef.
This PEP maintains backwards compatibility. It only adds new functions, and a new loader method that is added for a loader that previously did not support running modules as __main__.
This document has been placed in the public domain.