Poly-Controlled Partial Evaluation and its Application to Resource-Aware Program Specialization PhD Thesis

Partial Evaluation is an automatic technique for program optimization. The aim of partial evaluation is to specialize a program with respect to part of its input, which is known as the static data. Existing algorithms for on-line partial evaluation of logic programs, given an initial program and a description of runtime queries, deterministically produce a specialized program. The quality of the code generated by partial evaluation of logic programs greatly depends on the control strategy used. Unfortunately, the existence of sophisticated control rules which behave (almost) optimally for all programs is still far from reality. The main contribution of this thesis is the development of Poly-Controlled Partial Evaluation, a novel framework for partial evaluation which is polycontrolled in that it can take into account repertoires of control strategies instead of a single, predetermined combination (as done by traditional partial evaluation). This approach is more flexible than existing ones since it allows assigning different control strategies to different call patterns, thus generating results that cannot be obtained using traditional partial evaluation. As a result, sets of candidate specialized programs can be generated, instead of a single one. In order to make the algorithm fully automatic, it requires the use of self-tuning techniques which allow automatically measuring the quality of the different candidate specialized programs. Our approach is resource aware in that every solution obtained by poly-controlled partial evaluation is assessed by means of fitness functions, which can consider multiple factors such as run-time and code size for the specialized programs. The framework has been implemented in the CiaoPP system, and tested on numerous benchmarks. Experimental results show that our proposal obtains better specializations than those achieved using traditional partial evaluation, especially in the context of resource-aware program specialization. Another main contribution of this thesis is the presentation of a unifying view