Multi-task Ada code generation from synchronous dataflow programs on multi-core: Approach and industrial study

Special issue edited by Osman Hasan, Frédéric MalletInternational audienceThe growing trend to use multi-core processors to get more performance is increasingly present in safety-critical systems. Synchronous dataflow programming is naturally well-suited to parallel execution, thanks to the fact that all data dependencies are always explicit. MiniSIGNAL is a multi-task code generation tool for the synchronous dataflow language SIGNAL. The existing MiniSIGNAL code generation strategies mainly consider coarse-grained parallelism based on Ada multi-task model. However, when we applied it to industrial case studies, this code generation scheme has revealed inefficient: architecture aspects of the target platform have to be taken into account to achieve fine-grained parallelism. To generate more efficient target code from industrial cases, this paper presents a new multi-task code generation method for MiniSIGNAL. Starting at the level of synchronous clocked guarded actions (S-CGA) which is an intermediate language for the compilation process of MiniSIGNAL, the transformation consists of two parts: at the platform-independent level, transforming the S-CGA representation to an abstract multi-task structure (called Virtual Multi-Tasks, VMT); at the platform-dependent level, adopting the thread pool pattern concurrent JobQueue to support fine-grained parallel Ada code generation from the VMT structure. Moreover, the formal syntax and the operational semantics of VMT are mechanized in the proof assistant Coq. Finally, the effectiveness of our approach is illustrated by an application of the real-world Guidance, Navigation and Control system