A compilation-based software estimation scheme for hardware/software co-simulation

Abstract

High-level cost and performance estimation, coupled with a fast hardware/software co-simulation framework, is a key enabler to a fast embedded system design cycle. Unfortunately, the problem of deriving such estimates without a detailed implementation available is very difficult. In this paper we focus on embedded software performance estimation. Current approaches use either behavioral simulation with (often manual) timing annotations, or a clock cycle-accurate model of instruction execution (e.g., an instruction set simulator). The former provides greater flexibility (no need to perform a detailed design) and high simulation speed, but cannot easily consider effects such as compiler optimization and processor architecture. The latter provides high accuracy, but requires a more detailed implementation model, and is much slower in general. We hence developed a hybrid approach, that incorporates some aspects of both. It provides a flexible and fast simulation platform, considering also compilation issues and processor features. The key idea is to use the GNU-C compiler (GCC) to generate "assembler-level" C code. This code can be annotated with timing information, and used as a very precise, yet fast, software simulation model. We report some experimental results that show the effectiveness of our approach, and we propose some future improvements.