> Anything that is touched by the immortal object will be leaked. This can also happen in obscure ways if reference cycles are created.

I think this is simply expected behavior if you choose to create immortal objects, and not really an issue. How could you have an immortal object that doesn't keep its strong references alive?

> this does not fully cover all cases as objects that become tracked by the GC after they are modified (for instance, dicts and tuples that only contain immutable objects). Those objects will still participate in reference counting after they start to be tracked.

I think the last sentence here is not quite right. An immortalized object will never start participating in reference counting again after it is immortalized.

There are two cases. If at the time of calling `immortalize_heap()` you have a non-GC-tracked object that is also not reachable from any GC-tracked container, then it will not be immortalized at all, so will be unaffected. This is a side effect of the PR using the GC to find objects to immortalize.

If the non-GC-tracked object is reachable from a GC-tracked object (I believe this is by far the more common case), then it will be immortalized. If it later becomes GC-tracked, it will start participating in GC (but the immortal bit causes it to appear to the GC to have a very high reference count, so GC will never collect it or any cycle it is part of), but that will not cause it to start participating in reference counting again.

> if immortal objects are handed to extension modules compiled with the other version of the macros, the reference count can be corrupted

I think the word "corrupted" makes this sound worse than it is in practice. What happens is just that the object is still effectively immortal (because the immortal bit is a very high bit), but the copy-on-write benefit is lost for the objects touched by old extensions.

> 1.17x slower on logging_silent or unpickle_pure_python is a very expensive price

Agreed. It seems the only way this makes sense is under an ifdef and off by default. CPython does a lot of that for debug features; this might be the first case of doing it for a performance feature?

> I would be more interested by an experiment to move ob_refcnt outside PyObject to solve the Copy-on-Write issue

It would certainly be interesting to see results of such an experiment. We haven't tried that for refcounts, but in the work that led to `gc.freeze()` we did try relocating the GC header to a side location. We abandoned that because the memory overhead of adding a single indirection pointer to every PyObject was too large to even consider the option further. I suspect that this memory overhead issue and/or likely cache locality problems will make moving refcounts outside PyObject look much worse for performance than this immortal-instances patch does.