ITR: Building Practical Compilers Based on Adaptive Search

William Marsh Rice University, Houston TX

Investigators

Abstract

The information age is built on software. Web browsers and servers, office productivity tools, anti-lock brake systems, cellular telephones, and online trading systems are all implemented in software. Most of that code is prepared for execution using a compiler. For four decades, compilers have applied a fixed sequence of transformations to the code. The fact that different transformations, applied in different orders, can produce different results has been known, but the techniques for picking effective transformation sequences have not. The computational power available with modern processors makes it possible to experimentally discover good transformation orders. This project will systematically explore the problems of choosing transformation sequences. This five-year program will develop practical techniques for building adaptive compilers - compilers that change their behavior in response to the input program and the end-user's stated goals for optimization. This will require experimentation, algorithm development, application of techniques from machine learning, and basic software engineering. This project will produce the tools, techniques, and insights needed to make adaptive compilation both practical and productive. The resulting compilers will give users greater control over the run-time characteristics of their programs and will reduce the performance variability that plagues compiled code today.