Register Allocation using Graph Neural Networks

Anonymous

Keywords: Register allocation, Graph Coloring, Graph Neural Networks

TL;DR: In this paper we propose a solution for register allocation problem of compiler using Deep Learning networks based on graphs called as graph neural networks (GNN)

Abstract:

The execution lifecycle of a program goes through various optimizations and code generation. Register allocation is critical part of optimization phase of compiler. Graph coloring is one of the NP hard problem and it applies to various applications such as register allocation. Deep Learning methods has various state-of-art techniques in domains like computer vision, natural language processing.

In this paper we propose a solution for register allocation problem of compiler using Deep Learning networks based on graphs called as graph neural networks (GNN). We start with basic idea that two connected nodes with temporaries will have different registers allocated.