Fixing file with rule F401 cause infinite loop

[qarmin]

ruff 0.5.5+353 (2f54d05 2024-07-29)

ruff check *.py --select F401 --no-cache --fix --unsafe-fixes --preview --output-format concise --isolated

file content(at the bottom should be attached raw, not formatted file - github removes some non-printable characters, so copying from here may not work):

from .main import MaµToMan

error

/tmp/tmp_folder/data/5737962126122819962.py:1:8: F401 `tima` imported but unused
Found 501 errors (500 fixed, 1 remaining).
[*] 1 fixable with the --fix option.

debug error: Failed to converge after 500 iterations in `/tmp/tmp_folder/data/5737962126122819962.py` with rule codes F401:---
import timª
---

Ruff build, that was used to reproduce problem(compiled on Ubuntu 22.04 with release mode + debug symbols + debug assertions + overflow checks) - https://github.com/qarmin/Automated-Fuzzer/releases/download/Nightly/ruff.7z

python_compressed.zip

[AlexWaygood]

Cool bug, thanks!

[AlexWaygood]

This is an issue with NFKC normalization. If you apply this diff to Ruff:

diff --git a/crates/ruff_linter/src/fix/codemods.rs b/crates/ruff_linter/src/fix/codemods.rs
index c3c769172..ed4061435 100644
--- a/crates/ruff_linter/src/fix/codemods.rs
+++ b/crates/ruff_linter/src/fix/codemods.rs
@@ -80,7 +80,7 @@ pub(crate) fn remove_imports<'a>(
     for member in member_names {
         let alias_index = aliases
             .iter()
-            .position(|alias| member == qualified_name_from_name_or_attribute(&alias.name));
+            .position(|alias| dbg!(member) == dbg!(qualified_name_from_name_or_attribute(&alias.name)));
         if let Some(index) = alias_index {
             aliases.remove(index);
         }
diff --git a/crates/ruff_linter/src/rules/pyflakes/rules/unused_import.rs b/crates/ruff_linter/src/rules/pyflakes/rules/unused_import.rs
index bfa884801..4c86744db 100644
--- a/crates/ruff_linter/src/rules/pyflakes/rules/unused_import.rs
+++ b/crates/ruff_linter/src/rules/pyflakes/rules/unused_import.rs
@@ -397,6 +397,8 @@ pub(crate) fn unused_import(checker: &Checker, scope: &Scope, diagnostics: &mut
         }
         diagnostics.push(diagnostic);
     }
+
+    panic!()
 }

Then the output from the debug calls is as follows:

[crates/ruff_linter/src/fix/codemods.rs:83:31] member = "MaμToMan"
[crates/ruff_linter/src/fix/codemods.rs:83:47] qualified_name_from_name_or_attribute(&alias.name) = "MaµToMan"

And I'll also paste a screenshot because it looks like GitHub might actually normalize these confusables as well when it renders them in an issue comment?!):

---

---

If you look very closely at that screenshot, you can see that the μ on the first line is sliiiightly narrower in size than the µ on the second line. In other words, one of the two variables has been NFKC-normalized while the other has not, meaning the comparison here evaluates to false when it should be evaluating to true:

ruff/crates/ruff_linter/src/fix/codemods.rs

Line 83 in 381bd1f

.position(|alias| member == qualified_name_from_name_or_attribute(&alias.name));