Parallelizing Compiler Translation Validation Using Happens-Before and Task-Set

Program optimization passes performed by an automatically parallelizing compiler require highly sophisticated transformations of a source program. The compiler must preserve the behavior of the source program with the parallelized program. However, the implementation of such optimization passes may introduce bugs, and detecting them is usually difficult. In this paper, based on a task graph obtained from a source program, we provide a translation validation technique, which validates that the task related program properties, such as partial order among tasks and task-set to be executed, are preserved before and after compiler transformations. These properties are derived from edges in the graph consisting of both data and control dependencies. The implemented translation validator is evaluated with the OSCAR parallelizing compiler using four C programs generated by Embedded Coder from MATLAB/Simulink models to show the feasibility of the validation.