Exploring Compiler Optimization: A Survey of ML, DL and RL Techniques

The past few years, traditional compiler optimization methods have been found to be further enhanced by machine learning (ML), deep learning (DL) and reinforcement learning (RL). These differ from classical techniques that often use rule of thumb based decision making. Rather, ML/DL/RL based approaches provide a means for learning from data thus improving performance in different dimensions such as code generation, resource allocation and runtime. In this paper we give an overview of current research and methodologies utilizing ML, DL and RL for compiler optimization purposes. We analyze the major models in terms of their employed learning strategies and desired optimizations within a compiler framework. Moreover, we highlight some of the difficulties faced when these compilers are embedded with these learning models such as adaptability, generalization and overhead trade-offs. Additionally, our survey presents case studies demonstrating Quantitative improvements on well-known benchmarks mainly focusing on models' adaptability to different architectures and their role in supporting the decision-making process of compilers. We conclude outlining open research questions as well as possible future directions for further investigations into this emerging interdisciplinary field.