Priority Sampling of Large Language Models for Compilers