Compile-time memory reuse in logic programming languages through update in place

Abstract

Standard implementation techniques for single-assignment languages modify a data structure without destroying the original, which may subsequently be accessed. Instead a variant structure is created by using newly allocated cells to represent the changed portion and to replace any cell that references a newly allocated cell. The rest of the original structure is shared by the variant. The effort required to leave the original uncorrupted is unnecessary when the program will never reference the original again. This effort includes allocating and initializing new cells, as well as garbage collecting replaced cells. This article specifies a transformation system that introduces update-in-place operations, making Prolog programs update recursive data structures much as an imperative program would. The article introduces the notion of a reuse map, which formalizes reallocation decisions. Because optimal memory reuse is intractable, a heuristical method is presented that performs well in practice. Small Prolog programs that manipulate recursive data structures have their speed increased up to about five times (naive reverse), not counting any speedup obtained by avoiding garbage collection. Quicksort is about three times as fast, merge-sort about one and a half, matrix transposition about twice, and Gaussian elimination is about 1.2 times as fast.