Towards intelligent compiler optimization

The future of computation is massively parallel and heterogeneous with specialized accelerator devices and instruction sets in both edge- and cluster-computing. However, software development is bound to become the bottleneck. To extract the potential of hardware wonders, the software would have to solve the following problems: heterogeneous device mapping, capability discovery, parallelization, adaptation to new ISAs, and many others. This systematic complexity will be impossible to manually tame for human developers. These problems need to be offloaded to intelligent compilers. In this paper, we present the current research that utilizes deep learning, polyhedral optimization, reinforcement learning, etc. We envision the future of compilers as consisting of empirical testing, automatic statistics collection, continual learning, device capability discovery, multiphase compiling – precompiling and JIT tuning, and classification of workloads. We devise a simple classification experiment to demonstrate the power of simple graph neural networks (GNNs) paired with program graphs. The test performance demonstrates the effectiveness and representational appropriateness of GNNs for compiler optimizations in heterogeneous systems. The benefits of intelligent compilers are time savings for the economy, energy savings for the environment, and greater democratization of software development.