I think decref adds up to about 20% of mypy runtime on Python 3.11 when doing import torch (python/mypy#17919), so this could be a nice win.

Do you know off the top of your head if there's other low hanging fruit in mypyc regarding refcounting? Skimming the C code, I think I do see some redundant incref+decref pairs...

I think much of decref cost is from freeing objects. Having per-type freelists should help with it (e.g. #1018, but many classes could benefit). Mypy allocates a lot of temporary objects, and if we'd have small per-type freelists for selected types, we could avoid a lot of expensive allocation/free operations.

Another idea would be to identify the top N most expensive compiled functions in a CPU profile, and manually look for redundant incref/decref operations by inspecting the generated source or IR. Any redundant operations you find are likely to be worth fixing, or at least worth investigating.

We could possibly avoid some incref/decref pairs if we'd support borrowing of final attributes. Example where this would help:

from typing import Final

class C:
   def __init__(self, s: str) -> None:
      self.s: Final = s

def foo(s: str) -> None: ...

def bar(c: C) -> None:
   foo(c.s)  # Redundant incref/decref, since c.s can't be freed during the call

We'd also need to make various attributes (e.g. in mypy.types) final to benefit from this.

I've been working on upgrading the mypyc benchmarks runner to use Python 3.13 (previously it was using 3.8), and I noticed that some benchmarks are clearly slower on 3.13, and at least much of the impact seems to be from immortality checks. For example, the richards benchmark went from 0.0022s on 3.11 to 0.0038s on 3.12 (around 70% increase in execution time).

This now seems high priority, even if the impact to self check is not high. Also, we should try to avoid the overhead by default, instead of putting it behind a compiler flag.

I have a basic draft implementation that skips immortality checks for native classes and some mutable built-in types that can't be safely shared between subinterpreters. This speeds up richards by about 30%, and self check by about 1.9% 1.6%. It looks like this may give us back about half of the perf regression from immortal objects, while still allowing the necessary object sharing between subinterpreters (I think e.g. None must be immortal). I'll try to get this included in the next mypy public release.