John Regehr1d ago

some nice followup to my group's (non-LLM) work on generalizing peephole optimizations. the non-LLM side of program synthesis really does seem to be dead at this point.

https://arxiv.org/pdf/2603.18477

Show threadChris Leary
@regehr I’m not an expert so if this is too ignorant feel free to just ignore, but I thought with CEGIS kind of techniques LLMs just fill in a “previously missing piece” of how to reasonably pick something to inject to make progress — template based injection was very bespoke and domain specific.
Mar 24 at 4:40pmWeb
Show threadSam1d ago
@cdleary @regehr that was "sketches". CEGIS is just "here's a counterexample for what i asked for, make it work on this case too"
Show threadJohn Regehr1d ago
@lenary @cdleary right. imo, sketches were a nice idea for certain domains, but in compiler domains where I work, sketches were never even slightly useful.
Show threadChris Leary1d ago
@regehr @lenary got it, IIRC the peephole work was more “find existence of an improvement and generalize as much as possible” which I imagine is a totally different angle, thanks for the correction, I remember thinking it was super cool!
Show threadJohn Regehr1d ago
@cdleary @lenary separately, I've done a few experiments on my own, using a current-generation coding agent to generalize peephole rewrites and they're *very* good at this
Show threadSam1d ago
@regehr @cdleary you had to do very specific sketches with very specific IRs and it was fun but also a nightmare. I would say in some domains there continues to be less pressure for compilers to be fully general, but for sure not LLVM verification :)