You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
PEP 670 converted macros to static inline functions in Python 3.11. I propose to implement the second phase of PEP 670: convert static inline functions to regular functions in Python 3.12.
Static inline functions are causing C++ warnings. I had to add _Py_CAST() and _Py_NULL macros to Python 3.11 to limit the number of C++ warnings. The _Py_CAST() macro has a complicated history. @serge-sans-paille helped me to write a correct C++ implementation of _Py_CAST(), but it had to be reverted to fix the issue #94731. See:
As explained in PEP 670, static inline functions cannot be used in programming languages other than C, and these functions cannot be loaded by loading the symbol from the Python library (libpython). For example, the vim text editor loads uses the Python C API by loading symbols.
Converting static inline functions also hides implementation details. When using regular function calls, at the ABI level, it's just a regular function call, the machine code doesn't access directly structure members. So we have more freedom to changes Python internals (like the structure).
This issue is not about changing Python internals. It's just about convert static inline functions to regular functions.
The text was updated successfully, but these errors were encountered:
Converting PyBytes_AS_STRING() and PyBytes_GET_SIZE() static inline functions to regular functions is a prerequirement to hide PyBytesObject members from the public C API.
Exposing members of this structure causes two issues:
PyBytesObject.ob_shash is deprecated in Python 3.11: PyObject_Hash() must be called instead
Using regular function calls rather than accessing directly structure member avoids these issues.
For the flexible array, we should replace the last member with char ob_sval[0]; (C99) or char ob_sval[]; (GNU extension?). It can cause issues on projects using sizeof(PyBytesObject) and making assumption on the memory layout (that the structure always contains 1 item as part of the structure size).
Converting PyBytes_AS_STRING() and PyBytes_GET_SIZE() static inline functions to regular functions makes them in average 1.25x slower: +2.7 ns per function call, in the worst case, according to a micro-benchmark: #97017 (comment)
In Python core, PyBytes_AS_STRING() and PyBytes_GET_SIZE() function calls should still be inlined thanks to LTO. But in C extensions built as shared libraries (ex: array, _asyncio, _csv, _json, _struct, _datetime, _codecs_jp, _socket, etc.), they become (slower) regular function calls.
If performance is an issue in some shared libraries, we can provide static inline functions in the internal C API and use explicitly this faster flavor.
PEP 670 converted macros to static inline functions in Python 3.11. I propose to implement the second phase of PEP 670: convert static inline functions to regular functions in Python 3.12.
Static inline functions are causing C++ warnings. I had to add _Py_CAST() and _Py_NULL macros to Python 3.11 to limit the number of C++ warnings. The _Py_CAST() macro has a complicated history. @serge-sans-paille helped me to write a correct C++ implementation of _Py_CAST(), but it had to be reverted to fix the issue #94731. See:
As explained in PEP 670, static inline functions cannot be used in programming languages other than C, and these functions cannot be loaded by loading the symbol from the Python library (libpython). For example, the vim text editor loads uses the Python C API by loading symbols.
Converting static inline functions also hides implementation details. When using regular function calls, at the ABI level, it's just a regular function call, the machine code doesn't access directly structure members. So we have more freedom to changes Python internals (like the structure).
This issue is not about changing Python internals. It's just about convert static inline functions to regular functions.
The text was updated successfully, but these errors were encountered: