AutoPass: Evidence-Guided LLM Agents for Compiler Performance Tuning

AutoPass is a multi-agent framework designed for compiler performance tuning that leverages evidence from both the compiler and runtime to guide optimization decisions made by Large Language Models (LLMs). It allows LLMs to interact with the compiler's internal states and analyze intermediate representations without requiring offline training or fine-tuning, making it adaptable to various benchmarks and platforms. Evaluated on the LLVM compiler, AutoPass achieved geometric-mean speedups of 1.043x on x86-64 and 1.117x on ARM64 compared to LLVM's -O3, demonstrating its effectiveness over traditional auto-tuning methods.