12 August, 2005
Thinking further on the problem of stateful parsing (from yesterday’s article), the way we’ve done it in the past (essentially by using Scheme’s parameter mechanism) is a special case of a transactional system, with rollback on error and commit after each toplevel expression parsed.
This suggests that if Scheme had an implementation of a Software Transactional Memory, then it would be a good fit for a stateful packrat parser. The symbol-table (or whatever state needs to be speculatively propagated along a line of parsing) would simply be placed under transactional control. Each time an attempt is made to match against a particular nonterminal a nested transaction would be entered. If the nonterminal matched the input successfully, the nested transaction would be committed; otherwise it would be rolled back, and any alternative options would be tried in their own nested transactions.