Warehouse-Scale Interprocedural Code Layout Optimization with AlphaEvolve [pdf] (arxiv.org)

A new approach called AI-PROPELLER has been introduced to enhance performance optimization in warehouse-scale applications through interprocedural code layout. Traditional post-link optimizers have focused primarily on intraprocedural techniques, but AI-PROPELLER leverages an agentic workflow known as Magellan to evolve compiler heuristics and perform fine-grained interprocedural optimization. By generating layout variants and executing them on actual hardware, the system moves away from static cost models, resulting in more accurate performance measurements and significant optimizations, yielding improvements between 0.23% and 1.6% against state-of-the-art frameworks like FDO and PLO. This advancement is significant as it represents a breakthrough in optimizing large-scale industrial applications, a domain that has historically seen limited exploration due to the complexity of interprocedural layouts. By successfully applying fine-grained optimization methods to real-world binaries, AI-PROPELLER opens the door for further enhancements in software performance, potentially influencing how future compilers and optimization techniques are developed in the AI and machine learning communities.