Abstract

Computer systems are increasingly parallel and heterogeneous, while programs are still largely written in sequential languages like C/C++ assuming a single processor connected to uniform memory. The obvious suggestion that the compiler should automatically distribute a sequential program across the system usually fails in practice because the dependence analysis in the presence of aliasing can be over-pessimistic.

This talk will explore how the compiler can rely on the programmer’s knowledge about dependences in the code to enable automatic parallelisation. Practically speaking, by enclosing a code fragment inside a special block the programmer (solemnly) swears to the compiler that the code fragment is free of dependences on all or some memory locations. This essentially creates a treaty-point between what is easy for the programmer to guarantee and what the compiler is capable of refactoring.