Browse Theses

This thesis discusses several problems facing compiler design for supercomputers. One problem is that of building a comprehensive yet efficient means of representing the flow of data in a program. We propose a means of computing a data dependence graph and labelling the arcs with "direction vectors" which show how the flow of data corresponds to the loop structure of the program. We use these data dependence direction vectors in several high level compiler loop optimizations, namely, loop vectorization, loop fission, loop fusion and loop interchanging. We show how to perform these transformations and how to use them to optimize programs for a wide range of supercomputers. We also discuss the problems of recurrence relations. We study arithmetic recurrences with IF statements, and recurrences involving both data and control dependence relations in a cycle. The wavefront method of solving recurrences is also treated. Another problem for a compiler is the management of compiler temporary variables; supercomputer compilers must manage temporary arrays, and we discuss ways to make this problem more tractable. Finally we discuss WHILE loops and see how they relate to iterative DO loops, and show several methods to execute WHILE loops on a machine. We also give the general structure of an optimizing compiler for supercomputers, developed from our experience with a test-bed compiler.