Make intern! & py_format! strings immortal

[bschoenmaeckers]

No description provided.

[codspeed-hq]

Merging this PR will not alter performance

✅ 100 untouched benchmarks

---

_{Comparing bschoenmaeckers:intern-immortal (bb3f570) with main (46c16e4)}

[ngoldbaum]

If you're interested in making things immortal, Python 3.15 will have PyUnstable_SetImmortal, which will allow you to do this with arbitrary objects.

There's also a backport we can adapt for Python 3.13 and 3.14: python/pythoncapi-compat#164

[bschoenmaeckers]

If you're interested in making things immortal, Python 3.15 will have PyUnstable_SetImmortal, which will allow you to do this with arbitrary objects.

There's also a backport we can adapt for Python 3.13 and 3.14: python/pythoncapi-compat#164

The following docs suggest that you should not use it on strings.

https://github.com/python/cpython/blob/aa8a43d179bad5cd9fbfce63b630e2ee0bd617e4/InternalDocs/string_interning.md?plain=1#L57

[ngoldbaum]

Sorry, I wasn't trying to say it would be useful here, just as a more general comment.

[bschoenmaeckers]

Not sure if this has any positive effect on performance.

[ngoldbaum]

In my experience refcount churn isn't usually a big performance cost. The exception is if an object is shared on the free-threaded build, because inrefs and decrefs on shared objects are substantially more expensive than unshared objects.