LLMs Direct Compiler Optimizations in Closed-Loop System

Global: LLMs Guide Compiler Optimizations in Closed-Loop Framework

Researchers have introduced a new framework that uses off‑the‑shelf large language models (LLMs) to guide code optimization for complex loop nests. The system creates a closed‑loop interaction with a compiler, allowing the model to propose transformations, receive legality and performance feedback, and iteratively refine its strategy. Experiments on the PolyBench benchmark suite demonstrated notable speedups without any task‑specific fine‑tuning of the LLM.

Framework Overview

The framework, called ComPilot, treats the LLM as an interactive optimization agent. It receives the source code of a loop nest, generates candidate transformation directives, and passes them to a standard compiler interface. No additional training data or model adaptation is required; the approach relies on the LLM’s zero‑shot reasoning capabilities.

Feedback Loop Mechanism

After the compiler attempts a transformation, it returns two pieces of information: a legality verdict indicating whether the transformation preserves program semantics, and an empirical runtime measurement that quantifies speedup or slowdown. The LLM incorporates this concrete feedback to adjust subsequent proposals, effectively learning from the compiler’s responses during a single optimization session.

Performance Evaluation

Using the PolyBench suite, the authors report a geometric mean speedup of 2.66× for a single run of the optimizer, and 3.54× when selecting the best result from five independent runs. These figures are derived from direct comparisons with the original unoptimized code.

Comparison with Existing Optimizers

The study also compares ComPilot’s results with those of the Pluto polyhedral optimizer, a state‑of‑the‑art tool for loop transformations. In many benchmark cases, ComPilot outperformed Pluto, highlighting the potential of LLM‑driven optimization to rival specialized compiler techniques.

Implications and Future Work

The findings suggest that general‑purpose LLMs can serve as effective agents for low‑level code optimization when anchored by concrete compiler feedback. The authors propose further investigation into scaling the approach to larger code bases, integrating additional performance metrics, and exploring multi‑model ensembles to enhance robustness.

This report is based on information from arXiv, licensed under Academic Preprint / Open Access. Based on the abstract of the research paper. Full text available via ArXiv.