Overview
Modern "non-strict" functional programming languages are a powerful means of programming highly parallel computers, but are intrinsically difficult to compile well because decisions about ordering of subcomputations must be taken at compile time. This book represents a new technique for compiling such languages by partitioning a program into sequential threads. While the interleaving of threads can vary at run time, within each thread the order is fixed.A program is compiled by analyzing its data dependences, and developing from that a set of partitioning constraints. These practical algorithms are founded on a new theory of data dependence and ordering within functional programs, which defines dependence graphs in terms of a rewrite-rule operational semantics for the language.
By attacking the ordering problem directly, the book departs from previous approaches that obtain partitioning as a byproduct of optimizing lazy evaluation, and cleanly separates partitioning from other code generation issues. Furthermore, the method is flexible enough to produce both lazy code and also a less restrictive "lenient" variant which allows larger threads with only a slight decrease in expressive power. Code generation and optimization are explored in depth for both uniprocessor and multiprocessor targets.
Kenneth R. Traub is a researcher with the Motorola Cambridge Research Center.
Contents: Introduction. Background - Functional Language Compilers. Lenient Evaluation. Functional Quads. Code Generation. A Syntactic Theory of Data Dependence. Dependence-Based Partitioning. Conclusion.