Revolutionizing Compiler Performance Tuning: Meet AutoPass, the Evidence-Guided LLM Agent

In a groundbreaking study by Zepeng Li and his team, a novel framework named AutoPass is unveiled, promising to reshape the landscape of compiler performance tuning. Traditionally, optimizing code for better performance has been a daunting task due to the intricate interplay of microarchitectural impacts and unpredictable runtime measurements. However, AutoPass harnesses the power of large language models (LLMs) to make this process not only more efficient but also more intelligent.

The Challenge of Compiler Optimization

Compiler optimization is essential for maximizing software performance on contemporary systems, especially given the multitude of optimization passes available in compilers like LLVM. Developers typically rely on predefined optimization levels, but a one-size-fits-all approach often falls short. The issue lies in the "phase ordering problem" and the numerous parameters that must be tuned for individual passes. With hundreds of potential configurations, finding the ideal combination for a specific program can be computationally expensive and time-consuming.

Introducing AutoPass: A Multi-Agent Framework

AutoPass steps in to address these challenges by integrating LLMs with compiler feedback mechanisms. Unlike previous auto-tuning methods that operated as "black boxes," AutoPass allows the LLM to access internal optimization states and analyze intermediate representations of code. This openness facilitates a guided optimization process where decisions are backed by real-time performance feedback.

The AutoPass framework consists of multiple agents that work collaboratively. It employs a "Score Agent" to identify critical hotspots in the code, an "Analysis Agent" to extract pertinent features, a "Reasoning Agent" to make optimization decisions, and an "Evaluation Agent" that verifies these decisions through benchmarking. This holistic approach ensures that the optimization pipeline remains responsive to actual performance data.

Impressive Results Across Platforms

Implemented on the LLVM compiler and evaluated on various systems, AutoPass demonstrated impressive results, surpassing traditional methods like Profile-Guided Optimization (PGO). Initial tests showed geometric mean speedups of 1.043x and 1.117x over the LLVM -O3 baseline for x86-64 and ARM64 platforms, respectively. These outcomes underline the efficacy of AutoPass in adapting to different architectures and workloads, showcasing its potential for real-world application.

Why AutoPass Matters

Beyond mere performance enhancements, AutoPass introduces a new paradigm in compiler optimization that emphasizes adaptability and interpretability. By grounding decisions in empirical evidence, this multi-agent framework enhances the transparency of the optimization process, making it easier for developers to understand how decisions are made. The implications of this research extend beyond compiler optimization, potentially influencing how AI can assist in various aspects of software development.

In summary, AutoPass not only streamlines the optimization process but also paves the way for a future where LLMs play a critical role in enhancing software performance. As we continue to push the boundaries of what AI can achieve, frameworks like AutoPass are vital for bridging the gap between complex software systems and efficient execution.

Authors: Zepeng Li, Jie Ren, Zhanyong Tang, Jie Zheng, Zheng Wang