UC San Diego

Recently, microprocessor architects have redirected their attention from improving clock frequency to exploiting large-scale parallelism on multi-core processors as a means of continuing Moore's Law. Tiled multi-core processors are one such class of multi-core processors that facilitate compilers' automatic parallelization by providing low-latency communication over on-chip Scalar Operand Networks (SON). Though parallelizing compilers for academic tiled architectures such as RAW and Wavescalar have shown their potential, they are not portable solutions because they are closely coupled with the source language (typically C) and their own target architecture. Furthermore, their source languages are not suitable for exploiting the full parallelizing power of tiled architectures. In order to provide portable compiler infrastructure for the class of tiled architecture, we implemented a compiler infrastructure with separate front- end and back-end components. Because MATLAB is the source language, the compiler will benefit from an inherent abundance of parallelism in source codes and conduct an easier analysis without pointers in both performance and complexity.The thesis presents a complete compiler front- end specifically optimized for MATLAB. On top of conventional front-end tasks, it performs static type and shape inference by using an inference engine, MAGICA. To overcome the limitations of MAGICA, the front-end provides an extended MATLAB format which enables programmers to define type- and dimension-aware MATLAB libraries in a high-level MATLAB. Performance evaluation measures the effect of function inlining performed in the front-end; function inlining significantly improved performance by 2.38x on average by allowing the inference engine to produce more exact type and shape information