Efficient Compiler Option Tuning: Exploiting Objective Interdependencies for Enhanced Performance

The process of managing numerous optimization options in a compiler like GCC, aimed at achieving various runtime performance goals such as file size, compilation time, and execution time, is quite challenging. Existing methods, both meta-heuristic and data-driven, often overlook the crucial interdependencies among these objectives, leading to suboptimal performance in optimization algorithms. Our research introduces a cost-effective approach that not only recognizes but also effectively exploits these interdependencies for more efficient compiler option tuning. This approach involves sequentially developing three Gaussian process surrogate models, each tailored to a specific optimization goal but closely interlinked to ensure a comprehensive optimization strategy. Additionally, we utilize a clustering-based selection mechanism that leverages the connections and diversity within the Pareto set. Our evaluation with GCC and the cBench benchmark suite demonstrates superior performance in Pareto front diversity compared to current methods, highlighting the importance of considering objective interdependencies in optimization algorithm efficiency.