Compilation techniques for irregular problems on parallel machines

Massively parallel computers have ushered in the era of teraflop computing. Even though large and powerful machines are being built, they are used by only a fraction of the computing community. The fundamental reason for this situation is that parallel machines are difficult to program. Development of compilers that automatically parallelize programs will greatly increase the use of these machines.;A large class of scientific problems can be categorized as irregular computations. In this class of computation, the data access patterns are known only at runtime, creating significant difficulties for a parallelizing compiler to generate efficient parallel codes. Some compilers with very limited abilities to parallelize simple irregular computations exist, but the methods used by these compilers fail for any non-trivial applications code.;This research presents development of compiler transformation techniques that can be used to effectively parallelize an important class of irregular programs. A central aim of these transformation techniques is to generate codes that aggressively prefetch data. Program slicing methods are used as a part of the code generation process. In this approach, a program written in a data-parallel language, such as HPF, is transformed so that it can be executed on a distributed memory machine. An efficient compiler runtime support system has been developed that performs data movement and software caching