Training Large Language Models to Comprehend LLVM IR via Feedback-Driven Optimization

LLVM Intermediate Representation (LLVM IR) serves as a crucial bridge between source code and machine code, yet its complexity presents challenges for both human developers and AI-driven analysis tools. We propose an approach to train Large Language Models (LLMs) to better comprehend and optimize LLVM IR by leveraging diverse optimization feedback. Instead of relying on a single optimization strategy like -Oz, our framework incorporates multiple types of optimizations, combining (1) static supervision from various compiler-defined passes with (2) interactive, feedback-driven refinement learning based on preferences (potentially using reinforcement learning). This dual-feedback mechanism is designed to enable LLMs to process and analyze LLVM IR more effectively, thereby facilitating downstream software engineering tasks such as code summarization, optimization recommendation, and vulnerability detection.