Generation of distributed supervisors for parallel compilers

Abstract

This paper presents a new approach towards solving the combination and communication problems between different compiler tasks. As optimizations may generate as well as destroy application conditions of other tasks a carefully chosen application order is important for the effectiveness of the compiler system. Each task is solved by exactly one implementation (engine) and is characterized by its input-output behaviour and an optional heuristics. The specification of all tasks in this manner allows the generation of distributed supervisors for the whole compilation system. The result is a clear semantics of the compiler behaviour during compilation and the separation of algorithm and communication. Software engineering advantages are the easy integration of independently developed parts and the reusability of code. The flexibility of such a compiler system results in high portability even across hardware architectures and topologies.