Abstract

This paper introduces an annotation language and a compiler that together can customize a library implementation for specific application needs. Our approach is distinguished by its ability to exploit high level, domain-specific information in the customization process. In particular, the annotations provide semantic information that enables our compiler to analyze and optimize library operations as if they were primitives of a domain-specific language. Thus, our approach yields many of the performance benefits of domain-specific languages, without the effort of developing a new compiler for each domain.

This paper presents the annotation language, describes its role in optimization, and illustrates the benefits of the overall approach. Using a partially implemented compiler, we show how our system can significantly improve the performance of two applications written using the PLAPACK parallel linear algebra library.