Sonic Partial Deduction

Abstract

Abstract. The current state of the art for ensuring finite unfolding of logic programs consists of a number of online techniques where unfolding decisions are made at specialisation time. Introduction of a static termination analysis phase into a partial deduction algorithm permits unfolding decisions to be made offline, before the actual specialisation phase itself. This separation improves specialisation time and facilitates the automatic construction of compilers and compiler generators. The main contribution of this paper is how this separation may be achieved in the context of logic programming, while providing non-trivial support for partially static datastructures. The paper establishes a solid link between the fields of static termination analysis and partial deduction enabling existing termination analyses to be used to ensure finiteness of the unfolding process. This is the first offline technique which allows arbitrarily partially instantiated goals to be sufficiently unfolded to achieve good specialisation results. Furthermore, it is demonstrated that an offline technique such as this one can be implemented very efficiently and, surprisingly, yield even better specialisation than a (pure) online technique. It is also, to our knowledge, the first offline approach which passes the KMP test (i.e., obtaining an efficient Knuth-Morris-Pratt pattern matcher by specialising a naive one).

Offline specialisation in Prolog using a hand-written compiler generator

“…mathitkmp in Table 2. (To be able to pass the KMP-test, more sophisticated local control would be required, see (Martin and Leuschel 1999) and the discussion below.) Again, to be fair, both ecce and mixtus are fully automatic systems guaranteeing termination, while for logen sufficient specialisation and termination had to be manually ensured by the user via the BTC .…”

“…The local control component of our generating extensions is still rather limited: either a call is always reducible or never reducible. To remedy this problem, and to allow any kind of partially instantiated data, an extension of our cogen approach has been developed in (Martin and Leuschel 1999). This approach uses a sounding analysis (at specialisation time) to measure the minimum depth of partially instantiated terms.…”

“…In terms of speed of the specialisation process, it is hence slower than our fully offline cogen approach (but still much faster than online systems such as mixtus or ecce). Also, (Martin and Leuschel 1999) only addresses the local control component and it is still unclear how it can be extended for the global control (the prototype in (Martin and Leuschel 1999) uses the online ecce system for global control; to this end trace terms were built up in the generating extension like in (Gallagher and Lafave 1996)).…”

Partial Evaluation of Pattern Matching in Strings, revisited

“…The same is true for Generalized Partial Computation [6], where a theorem prover is used to derive additional information from the truth or falsity of enclosing conditional tests, and for Queinnec and Geffroy's intelligent backtracking system [19], where abstract descriptions of the matched and unmatched patterns are maintained across success and failure continuations. Other partial evaluators that pass the KMP-test include partial deduction [18] and partial evaluators for functional logic programs [1,17].…”

The Abstraction and Instantiation of String-Matching Programs

“…Such partial evaluators include Futamura's Generalized Partial Computation [26], Smith's partial evaluator for constraint logic programming languages [48], Queinnec and Geffroy's intelligent backtracking [46], supercompilation [28,29,49,50,51], partial deduction [44], partial evaluators for functional logic programs [4,40], and the composition of a memoizing interpreter and a standard partial evaluator [27].…”