Analysis and Compilation of Parallel Programming Languages

Parallel programming is increasingly important for embedded systems as well as scientific computing because of the hardware evolution toward heterogeneous multiprocessors. However developers of parallel applications are still stuck with three main, relatively primitive approaches: low-level multithreading, OpenMP annotations for shared-memory multiprocessors, and message passing in MPI for distributed platforms. No recent parallel language has emerged as a serious contender. The current compiler technology for such programming models breaks down with every kind of parallel constructs, because concurrency in these low-level models is always expressed as procedure calls, with unknown side effects. Even recent parallel languages, such as PGAS ones, are first translated into C or C++ by converting parallel constructs into procedure calls. So many, if not all, optimization passes are disabled, and no theoretical foundation or common practice exist for the construction of optimizing compilers for such languages. The ACOPAL project seeks to establish solid formal foundations for high- and low-level parallel internal representations (PIR, HPIR, LPIR) able to support 1) multiple parallel paradigms and languages, 2) analyses of both parallel and sequential codes, and 3) optimizing transformations legal for both parallel and sequential codes, and to demonstrate its usefulness by implementing such a PIR in at least one existing tool. Parallel internal representations are key to the evolution of production compilers such as GCC or LLVM, and to the construction of static analyzers for parallel programs. The cooperation between ENS and MINES ParisTech will lead to a PSL center of excellence in static program analysis and compilation