Generation Flexibility versus Performance

Defining domain specific abstractions for generator systems leads to a quandary between choosing abstractions that exhibit powerful programming amplification through the combinatorial opportunities provided by composition, and choosing abstractions that can be easily transformed into high performance code. Most generation systems opt for high levels of abstraction to achieve programming amplification and as an added benefit get safety, understandability, and several other -ilities. As a consequence, the performance of their generated code is often compromised. My hypothesis is that a new generator architecture is needed to achieve both high levels of abstraction and high performance code. A generator based on such an architecture has been implemented in Common LISP. It is called the Anticipatory Optimization based generator because it allows the component and transform writers to anticipate the kinds of optimization opportunities that might arise and to prepare an abstract, distributed plan that attempts to achieve them. Based on a small number of examples that we have tested, the approach appears promising, allowing high levels of abstraction, flexibility, and performance. (See Biggerstaff98a-b.)