MileStone: A Multi-Objective Compiler Phase Ordering Framework (arxiv.org)

Researchers have introduced MileStone, a groundbreaking framework designed to optimize compiler phase ordering through a multi-objective approach. This initiative addresses challenges in finding effective optimization sequences, which often struggle with conflicting requirements related to execution time, code size, and energy consumption. MileStone treats compiler phase ordering as a multi-objective optimization problem, utilizing graph-based representations of programs and employing a combination of graph neural networks and reinforcement-learning agents to explore optimal pass sequences according to user-defined constraints. The significance of MileStone for the AI/ML community lies in its innovative integration of machine learning techniques to enhance compiler performance. The system builds a self-evolving database that tracks compiler transformations, continuously refining its predictive capabilities. Experimental results indicate that MileStone yields strong Pareto-optimal solutions, outperforming traditional LLVM optimization levels by reducing execution time by up to 45 percent within the same energy constraints. This framework presents a scalable and effective solution to the complex problem of multi-objective compiler optimization, marking a significant advancement in the intersection of compiler technology and AI methodologies.