Static Neural Compiler Optimization via Deep Reinforcement Learning