TypeGuard behavior does not match python semantics in async functions

[BobTheBuidler]

I'm not sure if this should go here or in the mypyc issue tracker. I believe its a bug in mypy itself which only causes notable issues when you compile code with mypyc.

Bug Report

TypeGuard behavior does not match python semantics in async functions. The block which a TypeGuard applies to should end at the next await, assuming the TypeGuard is on an instance attribute or some other mutable nonlocal var.

This causes unnecessary TypeErrors in mypyc-compiled code.

To Reproduce

https://mypy-play.net/?mypy=latest&python=3.12&gist=50cf9499e4e5be34d3faa754bd6052c0

Note: This code snippet does not produce any mypy errors, the bug is only exposed when compiling similar code with mypyc.

Expected Behavior

The TypeGuard should only guarantee the type of self._db up to the next await/async with/async for, because after that point the type is no longer guaranteed.

Actual Behavior

The TypeGuard "guarantees" the type of self._db longer than it is logically able to. Mypyc will raise a TypeError on line 12 if another task sets self._db while our coroutine waits for the lock.

Your Environment

• Mypy version used: 1.17.0
• Python version used: 3.12

[sterliakov]

I think this is very similar to #17537 - retaining narrowed attribute types is not always safe. The main reason is calling mutating methods between accesses (as in that report), but any await point should also invalidate such narrowing. Unfortunately, we cannot just disable such narrowing of attributes, and tracking mutating methods is difficult.

[JelleZijlstra]

I don't think async code is really special here. Narrowing class attributes is basically never safe, as code that modifies the attribute could be running at any time as a result of a method call, or even code running in another thread. However, narrowing class attributes is very important in practice for making type checkers practically useful, so we shouldn't drop it entirely, though perhaps we can come up with better heuristics to drop the narrowing in some cases where it makes sense.

For mypyc, this means that we can't rely on the type that was narrowed by a TypeGuard. (Which we already can't, since we don't verify that the TypeGuard's behavior is right.)

[JelleZijlstra]

(Sorry didn't mean to close, my browser suddenly started hanging.)