Question: interest in additional reproducible fixes from GPT-assisted review?

[janrysavy]

Hi Pierre,

we have been using FastMM for a long time, and recently tried running a focused GPT-5.5 Pro-assisted review of FastMM5. We use the same workflow successfully on our own code: the AI review is only a starting point, and we only act on findings that we can understand, reproduce, and validate locally.

The review output is here:
https://chatgpt.com/share/69f06ff0-caa8-83eb-9b96-d9287a7e4cf7

As a first sample, we prepared two small, reproducible PRs:

• Fix debug AllocMem zero-fill header lookup #67 fixes debug AllocMem zero-filling for medium debug blocks near the medium/large threshold.
• Fix debug block size overflow checks #68 fixes unchecked debug allocation-size arithmetic so invalid/overflowing debug sizes fail cleanly instead of corrupting debug metadata.

Both PRs include the reasoning, reproduction scenario, and local validation details. The standalone reproducers are intentionally kept out of the repository to avoid introducing a larger test structure unless that is something you would want.

There are a few additional findings that look potentially worth investigating, for example:

• event log handle / filename synchronization during concurrent diagnostic logging,
• FastMM_WalkBlocks callback restrictions while allocator locks are held,
• concurrency contract for mutable public configuration,
• FastMM_CurrentAllocationGroup being global rather than thread-local,
• possible UTF-8 encoding issue for non-BMP diagnostic text.

Before continuing with more PRs, we wanted to ask whether this kind of contribution is useful to you.

Would you prefer us to:

1. continue sending small, focused PRs with reproduction notes,
2. first open issues for each remaining finding before preparing fixes,
3. limit this to the two already-submitted PRs,
4. or avoid this review-driven work if it does not match how you want FastMM5 maintained?

No pressure either way. We want to make sure this is useful upstream and not just adding review noise.

Best regards,
Jan

[pleriche]

Hi Jan,

Thank you very much, this is very helpful. Coincidentally the repository was recently run through Claude Code and it missed both the issues that ChatGPT picked up. I have pushed fixes for those, and I will now investigate the additional findings in the ChatGPT report.

As to future reports, a bug report should be sufficient in most cases. I don't always agree with the extent of changes and/or formatting of the suggested patch from the AI, so being able to reproduce the issue is usually all I need. The test cases are very helpful for this, so if the AI already produces those then if you could paste it in the bug report it would save me the time to create a test case myself.

Thanks again,
Pierre

[janrysavy]

OK, thank you. Please let me know here once you have finished your own fixes, and I can run the updated upstream version through a second GPT-5.5 Pro review pass.

Thanks again for looking into these issues.

[pleriche]

Ok, I have fixed the UTF-8 issue and also added thread safety for the event log. I've gone through the report by ChatGPT and I think this covers all the significant issues that it reported.

I have noted its concerns around FastMM_WalkBlocks and FastMM_CurrentAllocationGroup, but those are by design.

[janrysavy]

Pierre, It looks like GPT-5.5 PRO is no longer finding additional issues beyond the ones we have just published. Some of them may of course turn out to be by design or not worth changing from your perspective.

Please let me know once you have reviewed them, and I can run one final review pass over the updated upstream code.

[pleriche]

Thank you very much for all your effort! This has been very helpful. I have addressed most of the issues that you reported. There are a couple that I am still mulling over, but I think I have fixed all the critical ones.

[janrysavy]

Final review: https://chatgpt.com/share/69f24674-e0a4-83eb-9248-bc08b7a31962

Final Verdict

No new high-value public-contract issues found in current HEAD.

Thank you for your time...

[pleriche]

Great, thank you very much for all your effort! I'm very interested to see what the AIs will pick up 6 months or a year from now. The speed of progress has been quite impressive.