Leveraging Large Language Models to Detect Missed Peephole Optimizations

By replacing small, suboptimal instruction sequences within programs with a more efficient equivalent, peephole optimization can not only directly optimize code size and performance, but also potentially enables further transformations in the subsequent optimization pipeline. Although peephole optimization is a critical class of compiler optimizations, discovering new and effective peephole optimizations is challenging as the instruction sets can be extremely complex and diverse. Previous methods either do not scale well or can only capture a limited subset of peephole optimizations. In this work, we leverage Large Language Models (LLMs) to detect missed peephole optimizations. We propose Lampo, a novel automated framework that synergistically combines the creative but unreliable code optimization ability of LLMs with rigorous correctness verification performed by translation validation tools, integrated in a feedback-driven iterative process. Through a comprehensive evaluation within LLVM ecosystems, we show that Lampo can successfully detect up to 17 out of 25 previously reported missed optimizations in LLVM on average, and that 22 out of 25 can potentially be found by Lampo with different LLMs. For comparison, the state-of-the-art superoptimizer for LLVM, Souper, identified 15 of them. Moreover, within seven months of development and intermittent experiments, Lampo found 26 missed peephole optimizations, 15 of which have been confirmed and 6 already fixed. These results demonstrate Lampo's strong potential in continuously detecting missed peephole optimizations.