_____________________________________________________________________________ Immediately - comparison test is probably chewing up most of the time - Check if the only remaining stack is lame (hopeful/nothopeful) - write a testcase for this - circular gramars s = A A = A | "b" - foo.add(x) foo.add(y.andnot(x)) ==> this is broken - Annotation Tutorial .................................................. - serializable parse tables? - Treewalker code compiler? ______________________________________________________________________________ v1.1 - precedes restrictions ("<-") - MUST HAVE BETTER ERROR MESSAGES - use for developing java15.g - java15.g - once this is ready, do big announcement - broader regression testing (for stuff like error messages, etc) - More topology untangling [later] - grammar highlighting? - Forest needs a "manual access" API - the unwrap bit in Forest makes it really hard to expose an API for forests ______________________________________________________________________________ v1.2 - finalize metagrammar and rdp-op's - write some grammars - Java grammar - TeX (math?) - URL (RFC) - RFC2822 (email message/headers) - clean up the whole Walk situation (?) ______________________________________________________________________________ Soon - serialization of parse tables - "ambiguity modulo dropped fragments"? - can this be checked statically? - eliminated statically? - substring parsing for better error messages - Parameterized LR - "Regular Right Part" grammars (NP Chapman, etc) - Attribute unification - inference of rejections for literals - "prefer whitespace higher up" (?) - Labeled edges on trees (associate a label with each slot in the child array in Forest.Body? might make equality tough) -- equivalent to Feature Structures. Colon-labeling. ______________________________________________________________________________ Later - understand and implement the RNGLR "kernel state" optimization. The _Practical Early Parsing_ paper may help. - Partly-Linear-PATR? (O(n^6) unification grammar) - Implement a k-token peek buffer (for each state, see if it "dead ends" during the next k Phases based solely on state -- ignoring result SPPF) - Arrange for the SPPF corresponding to dropped subtrees to never be generated (or merged, etc) - Is there any way we can avoid creating a GSS.Node instance for nodes which are transient in the sense that they have only one eligible reduction? - Re-read Rekers, particularly the stuff on optimal sharing - Isolate the Element objects from Parse.Table/GSS so we can move towards compilation. - consider allowing a Forest.Body to represent some other Tree whose Body's should be [recursively] considered part of this Forest. - perhaps not: right now we have a nice situation where Forest.Ref instances become immutable once iterator()ed. This also gives us a strong place to to culling with the certainty that we won't throw out a Body which would later be salvaged by some yet-to-be-added dependency. - Figure out if there is a way to: - allow unwrapping of children other than the very last one. - fold repetitions into an array form in Forest, before conversion to Tree. The major problem here is that multiple tree-arrays are possible, all of different lengths. Worse, even if they're all the same length, not all elements belong in the same "possibility vector" as all others. You essentially need a GSS to represent the array, which perhaps is what the unfolded form was in the first place. - Wikipedia grammar (needs to be both lexerless and boolean) - Boolean Parsing => Ordered Choice (";" operator) - bring back in parse-table phase resolution of precedence (just like associativity). This can be inferred from the use of ">" when the rules are in one of these special forms: E ::= E _ > _ E E ::= _ E > E _ E E ::= E _ E > E _ E where "_" is anything and "E" is the defining nonterminal. Essentially what we're looking for is the situation where the leftmost portion of one rule produces another rule, and the rightmost portion of the latter produces the former. I'm not 100% certain that this is as "strong" as the prefer/avoid form (try to prove this, you probably can), but it's "what people intend" most of the time. - implement Johnstone's algorithm for "reduced, resolved LR tables" to eliminate superfluous reductions on epsilon-transitions. ______________________________________________________________________________ Neat Ideas - Rekers & Koorn note that GLR Substring Parsing can be used to do really elegant and generalized "autocompletion". ______________________________________________________________________________ Ideas for the Future - Incremental parse table construction - "lazy GLR" and "lazy trees" -> language with first-class CF matching - perhaps linear boolean grammars instead? (linear time, quad space) - Forest parsing => chained parsers - unification parsing, attributes, etc - RRP grammars? - Take another stab at maximal-match? Nonterminal not-followed-by is too strong. - Error recovery based on substring parsing