An NSF Proposal

The objectives of this research are to improve software productivity, reliability, and performance of complex systems. The approach combines program transformations, sometimes in reflective ways, to turn very high level perspicuous specifications into efficient implementations. These transformations will be implemented in a meta-transformational system, which itself will be transformed from an executable specification into efficient code. Experiments will be conducted to assess the research objectives in scaled up applications targetted to systems that perform complex program analysis and translation. The transformations to be used include dominated convergence (for implementing fixed points efficiently), finite differencing (for replacing costly repeated calculations by less expensive incremental counterparts), data structure selection (for simulating associative access on a RAM in real time), and partial evaluation (for eliminating interpretive overhead and simplification). Correctness of these transformations, of user-defined transformations, and of the transformational system itself will be addressed in part. Both the partial evaluator and components of the transformational system that perform inference and conditional rewriting will be derived by transformation from high level specifications. Other transformations will be specified in terms of Datalog-like inference and conditional rewriting rules that should be amenable to various forms of rule induction. Previously, [Cai and Paige 93] used an ideal model of productivity free from all human factors in order to demonstrate experimentally how a transformation from a low level specification language into C could be used to obtain a five-fold increase in the productivity of efficient algorithm implementation in C in comparison to hand-coded C. However, only small-scale examples were considered. The proposed research includes a plan to expand this model of productivity to involve other specification languages (and their transformation to C), and to conduct experiments to demonstrate how to obtain a similar five-fold improvement in productivity for large-scale examples of C programs that might exceed 100,000 lines. The proposal lays out extensive evidence to support the approach, which will be evaluated together with its theoretical underpinnings through substantial experiments. If successful, the results are expected to have important scientific and economic impact. They are also expected to make interesting, new pedagogical connections between the areas of programming languages, software engineering, databases, artificial intelligence, and algorithms.