Pickling objects by name without __module__ is suboptimal

[serhiy-storchaka]

When the object is pickled by name (classes, functions, and other enum-like objects) has no __module__ attribute or it is None, the code searches the object in all imported modules: tries to resolve its name relatively to the module and check that the result is the same object.

Whether the module was get from the __module__ attribute or found in sys.modules, the code checks it again:
Then it checks again: tries to resolve its name relatively to the module and check that the result is the same object.

So in some the work is repeated twice. The following PR makes it only be performed once. It also contains few other minor optimizations and simplifications.

Linked PRs

• gh-122459: Optimize pickling by name objects without __module__ #122460

[serhiy-storchaka]

I did not expect significant effect from this change, but it make pickling a function without __module__ with the Python implementation twice faster:

$ ./python -m timeit -s "import sys; sys.modules['_pickle'] = None; import pickle; global func" -s "def func(): pass" -s "func.__module__ = None" -- "pickle.dumps(func)"

Before: 2000 loops, best of 5: 192 usec per loop
After:  2000 loops, best of 5: 92.3 usec per loop

This affects also pickling Ellipsis and NotImplemented singletons:

$ ./python -m timeit -s "import sys; sys.modules['_pickle'] = None; import pickle" -- "pickle.dumps(...)"

Before: 20000 loops, best of 5: 14.1 usec per loop
After:  20000 loops, best of 5: 10.6 usec per loop

The main purpose is of course cleaning up the code.