Optimization Space Learning: A Lightweight, Noniterative Technique for Compiler Autotuning
Tamim Burgstaller, Damian Garber, Viet-Man Le, Alexander Felfernig · 2024
Compilers are highly configurable systems. One can influence the performance of a compiled program by activating and deactivating selected compiler optimizations. However, automatically finding well-performing configurations is a challenging task. We consider expensive iteration, paired with recompilation of the program to optimize, as one of the main shortcomings of state-of-the-art approaches. Therefore, we propose Optimization Space Learning, a lightweight and noniterative technique. It exploits concepts known from configuration space learning and recommender systems to discover well-performing compiler configurations. This reduces the overhead induced by the approach significantly, compared to existing approaches. The process of finding a well-performing configuration is 800k times faster than with the state-of-the-art techniques.