On Wed, Aug 12, 2026 at 7:10 PM Jeffrey Law <[email protected]> wrote: > > > > On 8/6/2026 11:55 PM, Kyrylo Tkachov wrote: > > > >> On 6 Aug 2026, at 07:45, Jeffrey Law <[email protected]> wrote: > >> > >> > >> > >> On 8/5/2026 11:06 PM, Kyrylo Tkachov wrote: > >>>> Like most of these, I'm curious where this showed up :-) > >>> I asked an AI agent to mine SPEC2026 and a few other workloads that are > >>> of interest to me to find missing folds. > >>> It turned out to be a quite productive exercise. I can recommend it as a > >>> way to constructively use these agents for things other than writing GCC > >>> code. I don’t think this particular fold is on a hot path or anything, > >>> but it is obviously a beneficial one anyway (it reduces the complexity of > >>> the IR, as match.pd folds intend) > >> THanks. We should probably talk further given that info. I've proposed > >> a talk for Cauldron, one section of which would be using LLMs to help > >> discover code sequences we can improve. What we've been experimenting > >> with is taking hot block information from QEMU (in the form of assembly > >> blocks with headers indicating how hot the block is), then having the LLM > >> mine those blocks for improved instruction sequences. Then we (of course) > >> validate what the LLM finds and determine if it's worth implementing. > >> We've found several things using this approach for RISC-V. It has the > >> nice property that when the LLM flags something, you know it's hit at > >> runtime and you can evaluate if it's hot enough to be likely worth the > >> time to fix. > >> > >> If you wanted to co-present what you're doing, I can certainly make room. > >> Or if you're already proposing a session, I'll have to make sure to attend. > > I’d be happy to. > > I was thinking of proposing a session on using AI tools for GCC development > > in general: finding opportunities, bugs, orchestration etc. I’m happy to > > co-present at your session as it seems we’re thinking about the same topics. > THen let's go with that as a plan. I haven't heard anything yet, but > figure we'll likely have an hour. I'll probably want to cover a bit of > the context around the AI policy. Then say 15-20 for the 3 technical > topics. > > 1. Your work using LLMs to mine benchmarks (or sources in general) > for missed folds. > 2. Using LLMs to mine assembly sequences for potential improvements > 3. Using LLMs to help with patch review
I think we're also seeing a lot of (unattributed) LLM assisted patches lately. I have used LLMs to do offline bug analysis with moderate success and I might want to play with LLMs taking up the daily churn of going through bugzilla, adding missing classification, etc. - but that would make LLM modify bugs .... > > Jeff
