No narrowing on type(...) is T

[superbobry]

Describe the Bug

Reproducer:

from typing import reveal_type

def f(x: int | bool):  # does not narrow
  if type(x) is bool:
    x = 42
  reveal_type(x)


def g(x: int | None):  # works OK
  if x is None:
    x = 42
  reveal_type(x)

Output:

 INFO revealed type: bool | int [reveal-type]
 --> /tmp/foo.py:6:14
  |
6 |   reveal_type(x)
  |              ---
  |
 INFO revealed type: int [reveal-type]
  --> /tmp/foo.py:12:14
   |
12 |   reveal_type(x)
   |              ---
   |
 INFO 0 errors

Sandbox Link

No response

(Only applicable for extension issues) IDE Information

No response

[yangdanny97]

This is a bug, there's logic for it here:

pyrefly/pyrefly/lib/binding/narrow.rs

Line 628 in 0d0ea90

AtomicNarrowOp::TypeEq(right.clone())

[yangdanny97]

Here's what I think is happening:

We do narrow type(x) is bool inside the if statement, but we do not do negative refinement and exclude the bool from the union for the "else" case.

This is because the inverse condition type(x) != Y does not mean that x is not a subclass of Y, it just means x is not exactly Y, so it's not safe to exclude. However, since bool is a final class, it is safe to remove bool from the union here.

So we need to tweak the logic here to narrow away the class if it's final.

pyrefly/pyrefly/lib/alt/narrow.rs

Line 1022 in 0d0ea90

AtomicNarrowOp::TypeNotEq(_) => ty.clone(),

[rchen152]

@yangdanny97 Good find! We already have an is_final check we use in a few places:

pyrefly/pyrefly/lib/alt/narrow.rs

Line 111 in 0d0ea90

fn is_final(&self, cls: &ClassType) -> bool {

, so hopefully it's as simple as adding the check there as well.

[yangdanny97]

I'll stack the fix for this on top of #2475