The Boyer benchmark meets linear logic

Of the Gabriel Lisp Benchmarks, the Boyer Benchmark ("Boyer") is the most representative of real AI applications, because it performs Prolog-like rule-directed rewriting, and because it relies heavily on garbage collection (GC) for the recovery of storage. We investigated the notion that such programs are unsuitable for explicit storage management---e.g., by means of a "linear" programming style in which every bound name dynamically occurs exactly once. We programmed Boyer in a "linear" fragment of Lisp in both interpretive-rule and compiled-rule versions, using both true linear (unshared) and reference count methods.We found that since the intermediate result of rewrite is unshared, the linear interpreted version is slower than the non-linear interpreted version, while the linear compiled version is slightly faster than the non-linear compiled version. When sharing is allowed, requiring reference counts for the linear versions, the linear shared versions are not competitive with the non-linear versions, except when "anchored pointers" are used. The anchored pointer reference count version, which reclaims storage, is still 1.25X slower than the non-linear version, which reclaims no storage.