Skip to content

Conversation

meheff
Copy link
Collaborator

@meheff meheff commented Sep 12, 2025

The implicit activation bit was improperly generated inside for loops in some cases. This resulted in spurious asserts or other side-effecting operations firing improperly. This manifested as a divergence in assert behavior in a test between DSLX interpreter and IR interpreter .

The implicit activation bit was improperly generated inside for loops in some cases. This resulted in spurious asserts or other side-effecting operations firing improperly.
@ericastor
Copy link
Collaborator

This is a great find! We may need to try to figure out why the fuzzing never found this...

@meheff
Copy link
Collaborator Author

meheff commented Sep 12, 2025

It was a tough one to suss out as the failing test was large. Most of the time was spent building infrastructure for dumping call traces in both interpreters (last bit is #3057) which lead me to believe it had something to do with the implicit activation. Then I handed the AI (gpt5 in cursor) the call trace, user source code, and my vague notion of what might be going wrong and it one shotted the fix which was impressive.

@allight
Copy link
Collaborator

allight commented Sep 12, 2025

@copybara-service copybara-service bot merged commit 3646c63 into google:main Sep 12, 2025
6 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Projects
None yet
Development

Successfully merging this pull request may close these issues.

3 participants