Basic-block Reordering Using Neural Networks

Basic-block reordering is a compiler optimization technique which has the effect of reducing branch cost and I-Cache misses by rearranging code layout. In this paper, we present our basic-block reordering method which detects typical structures in the control-flow graph. It uses the architecture-specific branch cost model and execution possibilities of control-flow edges to estimate the possible layout costs of specific sub-structures. The layout with the minimal cost estimation would be chosen. We further investigate a new approach to use neural network to predict execution possibility for each edge. We choose a set of programs and record particular static information of the edges in the typical structures. These data includes the knowledge about the relationship between static program features and dynamic behavior, and is fed to train the neural network. In this paper, we adopted an improved back propagation neural network. The algorithm has been implemented based on a 5-stage pipeline UniCore architecture. The experiments show that it improves programs' performance well, and execution possibility of edges may be predicted using machine learning techniques.