Polyhedral model based application source codes analysis for ASIP design

Application Specific Instructions-set Processors (ASIPs) are needed to handle the future demand of flexible and high performance computation in parallel real-time computing domain. Current ASIP design methodologies are mostly based on the accurate understanding of application source codes. Profilers can automatically analyse source codes and get interesting information from them to aid processor architecture design. Dynamic profilers add extra instructions to source codes and execute to generate basic block execution counts. This is often infeasible in the domain of parallel real-time computing, because of excessive execution time and the difficulty to define typical stimuli in the early architecture design phase. Static methods analyse source codes without executing them and are faster and more accurate than dynamic ones. The static analysis method proposed in this paper is based on polyhedral model. It identifies code segments that execute frequently, or named hot codes. It is usually a small part of the whole code base. This information is provided to the architecture designer as design guide. The hot codes may be implemented using custom function unit to optimize the efficiency of the whole system. A running example is used to illustrate the static method, and the results are compared to dynamic profilers. Same results are achieved by the two analysis methods.