The LookAheadSet data structure is currently implemented as a list of lists. This is pretty ineffective, as look-ahead token searches are run with O(n*m) performance, where n is the number of look-ahead sets and m the maximum length of the look-ahead sequences. The current representation is also wasteful of resources, as roughly similar sequences (differing only in the last token) are completely duplicated.

By using an efficient tree representation using hash tables (or similar) at each node, the look-ahead tree sets could use less data while also performing better. Ideally O(m) performance could be reached, where m is the maximum length of the look-ahead sequences.

This change is not trivial, however. Probably, the easiest way would be to first create an intermediary solution where the external API of LookAheadSet is kept intact. This means that the current Sequence class would be kept, although no longer used internally inside LookAheadSet for storing the actual tokens.