bpo-36904: Optimize _PyStack_UnpackDict by jdemeyer · Pull Request #14517 · python/cpython

the-knights-who-say-ni

Optimize, clean up and fix _PyStack_UnpackDict:

_PyStack_UnpackDict is made a static function inside call.c.
The API is simplified: the new argument vector is returned directly instead of being passed via a pointer.
Move the case of no keywords outside of _PyStack_UnpackDict, so _PyStack_UnpackDict is now only used when there actually are keyword arguments. This speeds up the case of no keywords and it simplifies the code.
Allocate the argument vector such that it supports PY_VECTORCALL_ARGUMENTS_OFFSET. Tests indicate that _PyStack_UnpackDict is responsible for about 60% of cases where method_vectorcall needs to do an allocation. With this change, these allocations are no longer needed.
Move the freeing code (which is not trivial) to a new function _PyStack_UnpackDict_Free.
Fix the overflow check by using Py_ssize_t instead of size_t. In the old code, the overflow check could theoretically overflow (the right hand side could be negative).