Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

LPO: Discovering Missed Peephole Optimizations with Large Language Models

Zhenyang Xu, Hongxu Xu, Yongqiang Tian, Xintong Zhou, Chengnian Sun

Abstract

Peephole optimization is an essential class of compiler optimizations that targets small, inefficient instruction sequences within programs. By replacing such suboptimal instructions with refined and more optimal sequences, these optimizations not only directly optimize code size and performance, but also enable more transformations in the subsequent optimization pipeline. Despite their importance, discovering new and effective peephole optimizations remains challenging due to the complexity and breadth of instruction sets. Prior approaches either lack scalability or have significant restrictions on the peephole optimizations that they can find. This paper introduces LPO, a novel automated framework to discover missed peephole optimizations. Our key insight is that, Large Language Models (LLMs) are effective at creative exploration but susceptible to hallucinations; conversely, formal verification techniques provide rigorous guarantees but struggle with creative discovery. By synergistically combining the strengths of LLMs and formal verifiers in a closed-loop feedback mechanism, LPO can effectively discover verified peephole optimizations that were previously missed. We comprehensively evaluated LPO within LLVM ecosystems. Our evaluation shows that LPO can successfully identify up to 22 out of 25 previously reported missed optimizations in LLVM. In contrast, the recently proposed superoptimizers for LLVM, Souper and Minotaur detected 15 and 3 of them, respectively. More importantly, within eleven months of development and intermittent testing, LPO found 62 missed peephole optimizations, of which 28 were confirmed and an additional 13 had already been fixed in LLVM. These results demonstrate LPO's strong potential to continuously uncover new optimizations as LLMs' reasoning improves.

LPO: Discovering Missed Peephole Optimizations with Large Language Models

Abstract

Peephole optimization is an essential class of compiler optimizations that targets small, inefficient instruction sequences within programs. By replacing such suboptimal instructions with refined and more optimal sequences, these optimizations not only directly optimize code size and performance, but also enable more transformations in the subsequent optimization pipeline. Despite their importance, discovering new and effective peephole optimizations remains challenging due to the complexity and breadth of instruction sets. Prior approaches either lack scalability or have significant restrictions on the peephole optimizations that they can find. This paper introduces LPO, a novel automated framework to discover missed peephole optimizations. Our key insight is that, Large Language Models (LLMs) are effective at creative exploration but susceptible to hallucinations; conversely, formal verification techniques provide rigorous guarantees but struggle with creative discovery. By synergistically combining the strengths of LLMs and formal verifiers in a closed-loop feedback mechanism, LPO can effectively discover verified peephole optimizations that were previously missed. We comprehensively evaluated LPO within LLVM ecosystems. Our evaluation shows that LPO can successfully identify up to 22 out of 25 previously reported missed optimizations in LLVM. In contrast, the recently proposed superoptimizers for LLVM, Souper and Minotaur detected 15 and 3 of them, respectively. More importantly, within eleven months of development and intermittent testing, LPO found 62 missed peephole optimizations, of which 28 were confirmed and an additional 13 had already been fixed in LLVM. These results demonstrate LPO's strong potential to continuously uncover new optimizations as LLMs' reasoning improves.

Paper Structure

This paper contains 21 sections, 5 figures, 5 tables, 2 algorithms.

Table of Contents

  1. Introduction
  2. Background
  3. LLVM Intermediate Representation
  4. Peephole Optimization in LLVM
  5. Superoptimization
  6. Translation Validation
  7. Methodology
  8. Illustrative Example
  9. Extracting Instruction Sequences
  10. Verifying Outputs of the LLM
  11. Evaluation
  12. Experimental Setup
  13. RQ1: Detecting Previously Reported Missed Optimizations
  14. RQ2: Discovering New Missed Optimization
  15. RQ3: Throughput and Cost of LPO
  16. ...and 6 more sections

Figures (5)

  • Figure 1: (\ref{['fig:rust-clamp']}) and (\ref{['fig:llvm-ir']}) show a Rust function clamp revealing a missed optimization in LLVM, and its corresponding LLVM IR module. (\ref{['subfig:alive2-input-src']}) and (\ref{['subfig:alive2-input-tgt']}) are a pair of simplified LLVM IR functions illustrate the essence of the optimization.
  • Figure 2: The overall workflow of LPO. The verification phase is outlined by the dashed line.
  • Figure 3: An illustrative example demonstrating the workflow of LPO.
  • Figure 4: Three examples of confirmed missed optimizations found by LPO that Souper and Minotaur fail to detect.
  • Figure 5: Geometric mean of speedup on SPEC CPU2017 Integer for selected patches, as well as for two LLVM versions (f3501d7 on Dec. 1, 2025 vs. 017c75b on Dec. 1, 2024).