Dataflow analysis for energy-efficient scratch-pad memory management

Abstract

Scratch-pad memories (SPMs) are a serious alternative to conventional cache memories in embedded computing since they allow software to manage data flowing from and into memory components, resulting in a predictable behavior at runtime. The prior studies considered compiler-directed SPM management using both static and dynamic approaches. One of the assumptions under which most of the proposed approaches to data SPM management operate is that the application code is structured with regular loop nests with little or no control flow within the loops. This assumption, while it makes data SPM management relatively easy to implement, limits the applicability of those approaches to the codes involve conditional execution and complex control flows. To address this problem, this paper proposes a novel data SPM management strategy based on dataflow analysis. This analysis operates on a representation that reflects the conditional execution flow of the application and, consequently, it is applicable to a large class of embedded applications, including those with complex control flows.