Jeff Law via Gcc <[email protected]> writes: > On 10/21/26 8:11 AM, Richard Biener via Gcc wrote: >>> So, if I declare that I care for my wellbeing and that "I don't plan >>> to review any patches without AI pre-review", what will happen? >> I'm also using LLM review on my own patches with success (on >> company cost, not zero cost). I've been pointed to by Jeff to >> https://gcc.sashiko.dev/ which monitors gcc-patches - so that's >> only post-send LLM review, but still helpful to reviewers [cost]. > Very similar to where I am. Using company resources to review my own > work, but also using some public facing free stuff to help with older > patches not covered by Sashiko and Sashiko to help with stuff landing > on gcc-patches over the last couple weeks. > > With Sashiko starting to come up, one of the things I'm seriously > considering is an internal list for my team where they can post stuff > for review by an internal Sashiko instance using company resources. > There are also ways to do reviews of an existing commit with Sashiko > -- which can be routed to internal models or public models based on an > environment variable or two. > > Discussions in Prague repeatedly went towards using an LLM in an chat > mode where the LLM essentially acts as a fast turnaround reviewer. > It might have been Ramana that coined the term "reviewer > personality". Given my own experimenation with this, I would > strongly recommend it -- the back and forth with the LLM very much > reminded me of the review cycle (and from the direction of the > submitter since I was justifying design/implementation decisions & > providing key context to the LLM to resolve concerns raised by the > LLM). > > None of this is free, but I can help find tokens :-) > > >> I found LLMs ability to prototype changes helpful saving time to navigate >> areas of GCC I do not know thoroughly (but only know "there must be >> functionality X somewhere"); the LLM prototype itself then helps >> reviewing exactly the portion of the code that matters. >> >> On the original topic I'd not put LLM review down as absolute requirement >> but as a suggestion to augment self-learning. > Agreed. I'm not using it to prototype anything for GCC proper, but > Aldy is and he discussed it a bit in the AI/LLM discussion -- > essentially using the LLM to explore multiple avenues gather some data > to help guide Aldy towards an approach to a problem that is most > likely going to lead to a viable solution.
I now fall squarely into hobbyist with reduced time to spend on GCC, and I definitely have been using AI for prototypes and lots of other things, while still adhering to our guidelines. FWIW... When all the auditing of the DOM code I did to start my work on frange again was done by AI. It instrumented DOM and ranger, at my direction, and analyzed thousands of discrepancies, reducing them to a handful that I could then tackle. That would've taken me a couple weeks at least. In the value numbering work I've been working on, I had it prototype about 6-7 ways of solving this problem for the ranger threader, with ideas of my own, suggestions by Richi, the PRs, and other crazy ideas it came up with on its own. I had it prototype VN from the ground up, with VDEFs only, with aliasing info, etc etc. Then I had it kludge the ranger threader to use DOM's scoped tables, and even rewrite DOM with SCCVN's info, cause why not.... Ultimately, I cherry picked the one or two that gave decent numbers and implemented those. And with the FRE threader stuff, I was able to concentrate on the areas I'm more comfortable with, while the AI black boxed the SCCVN bits which make my head hurt. Again, I'm just a hobbyist here with limited time, and I'm using it to limit myself to the things that are fun, or I have some semblance of competence in, but then again, I'm not about to force it on anyone ;-). My two cents. Aldy
