ITR: An Optimizing Compiler for Languages with Programmable Memory Models

University Of Illinois At Urbana-Champaign, Urbana IL

Investigators

Abstract

The memory model for a programming language specifies the relationship between the order in which data accesses appear in a program and the order seen during execution by the different program components With the advent of parallel programming environments like Posix threads, Java, and OpenMP, multi- threaded explicitly parallel programs have become much more frequent. This increases the need for memory models that are easy to understand and efficient so that correct pro- grams can be developed and still give good performance. Unfortunately, the usability of memory models, their impact on performance, and the compiler technology needed to per- form optimizations of parallel programs are poorly under- stood. The result is that current memory models tend to favor performance over usability by restricting the programs that can be written with them, or by being difficult to understand. The objective of this project is to study com- piler techniques to optimize explicitly parallel programs by using optimizations and analysis algorithms structured to handle a broad class of consistency models. The techniques studied will be implemented in a compiler that will serve as a testbed for prototyping and studying programming language memory models and for studying the optimization and analysis of explicitly parallel programs.