Retargeting a compiler's back end to a new architecture is a time-consuming process. This becomes an evident problem in the area of programmable graphics hardware (graphics processing units, GPUs) or embedded processors, where architectural changes are faster than elsewhere. We propose the object-oriented rewrite system OORS to overcome this problem.
Using the OORS language, a compiler developer can express the code generation and optimization phase in terms of cost-annotated rewrite rules supporting complex non-linear matching and replacing patterns. Retargetability is achieved by organizing rules into profiles, one for each supported target architecture. Featuring a rule and profile inheritance mechanism, OORS makes the reuse of existing specification possible. This is an improvement regarding traditional approaches. Altogether OORS increases the maintainability of the compiler's back end and thus both decreases the complexity and reduces the effort of the retargeting process. To show the potential of this approach, we have implemented a code generation and a code optimization pattern matcher supporting different target architectures using the OORS language and introduced them in a GPU compiler.
@conference(OORS-WAPL07,
author={Gernot Gebhard and Philipp Lucas},
title={{OORS}: An Object-Oriented Rewrite System with Applications in
Retargetable Code Generation and Optimization},
booktitle={Proceedings of the 1st Workshop on Advances in Programming
Languages (WAPL'07)},
editor={Marjan Mernik},
pages={1057--1069},
year=2007,
month={October}
)
An extended version is available as a journal paper.
For more information about OORS, please visit the OORS homepage.
Valid XHTML 1.0 Strict, valid CSS.
Author: Philipp Lucas.