Currently the full-text search engine uses TREE (tree.c) in a few places. This implies O(log N) access time. Replacing trees with hashes may help here. Using move-to-first hash optimization may help even more, (and there's no concurrent accesses here)
