make marshal output not dependent on reference count

Bug report

Currently the marshal module will emit a previously-unseen object flagged as a potential reference from later objects, unless the object has a reference count of 1. See https://github.com/python/cpython/blob/b27b57c6e44a276c8a9843fd37d4cf65b2827d5c/Python/marshal.c#L305

This is an overly-conservative heuristic -- it's easy to construct cases where an object has a reference count >1 but is not actually referenced by any other object about to be marshaled, so FLAG_REF is set when it does not need to be.

This makes marshal output unstable depending on accidents of reference counting behavior in the code calling marshal.dumps.

I ran into this because the Cinder JIT is able to reduce unnecessary increfs, and that resulted in some importlib tests failing on comparison of marshal output at https://github.com/python/cpython/blob/b27b57c6e44a276c8a9843fd37d4cf65b2827d5c/Lib/test/test_importlib/test_abc.py#L870-L871 because under Cinder JIT the reference count of code_object in that method is 1.

This previously caused issues in distutils reproducibility, resulting in a partial fix that applies only to interned strings: https://github.com/python/cpython/pull/8226

It would be better if marshal would actually determine which objects have multiple parents in the DAG and deterministically use FLAG_REF or not based on that.

Commencez par Python/marshal.c au niveau de l’heuristique de comptage des références indiquée et examinez les assertions concernées dans Lib/test/test_importlib/test_abc.py aux lignes 870-871. Suivez la manière dont marshal.dumps décide de FLAG_REF, puis mettez en place des tests montrant que la sortie est indépendante des comptes de références de l’appelant et confirmez que les tests importlib passent.