Skip to Content

Parglare Python parser

With this post I’m announcing availability of a new Python parser based on LR(1) and GLR algorithms.

For years I’ve been working with and developing PEG based parsers (see Arpeggio and textX). PEGs have some really nice properties: they represent a concise description of top-down parsers and are easy to understand and debug. Furthermore, with choice operator being ordered PEG grammars can’t be ambiguous. Backtracking is natural and easy to implement giving it unlimited lookahead. A composition of PEG grammars will give a valid PEG grammar, although it might not be what you would expect.

But, there are some drawbacks.

PEG drawbacks

First, PEG based parser are naturally implemented as top-down parsers which is troublesome for left-recursive grammar rules. This leads to unnatural grammar specification for some languages which would be much easier to describe using left recursion.

Second, PEG parsers deal with ambiguity by implicitly resolving it with left-to-right ordered choice operator. This hides ambiguity and forces you to think imperatively during grammar construction. There are even situations when there exists no valid order of the rules for the language one wish to describe. This leads to a problem where you can’t be sure what language your grammar describes.

Third, the most straight-forward way to implement PEG parser is by mutually recursive function calls. Parser state is thus kept on the program stack which is not easy to manipulate. This makes error recovery hard to implement. In order to recover from an error you would have to restore the state where you consumed the most of the input and try to skip erroneous input or create expected tokens. Recovering that state would mean to recover the program stack. Another technique would be to implement explicit stack but that would kill much of the PEG attractiveness.

Parglare

With parglare I wanted to keep most of the good stuff from Arpeggio/textX but to use more robust algorithm with a solid scientific theory and to alleviate some of the PEG’s drawbacks.

LR parsing is invented by Donald Knuth in 1965. It is based on Context-Free Grammars and Push-Down Automata (PDA) construction process which if successful proves that the grammar is indeed LR(1) and that any input that correspond to the grammar can be parsed deterministically.

I wanted parglare to be a scannerless parser. Separate tokenization/scanning phase leads to a lot of grief. Parglare will recognize tokens on the fly and given the right parsing context this eliminates a lot of potential source of lexical ambiguities.

Parsing is a process of transforming a flat sequence of tokens to some other form, tree being a most common. This doesn’t imply that the input must be text at all. It can be sequence of any objects. Parglare indeed can parse anything by introducing the concept of recognizers. You can also see in this test how a list of integer is being parsed.

Both Arpeggio and textX provide a handy visualization features which can really help in understanding the grammar and parsing process. Parglare also provides pglr command which can be used to visualize LR automata and GLR parsing trace.

Grammar language is based on a pure BNF without any extensions at the moment. This might lead to a more verbose grammar but the actions are really straight-forward to write. In the future a syntactic sugar will be added for some common idioms.

Actions can be called on-the-fly during parsing or applied on the parser tree afterwards. The parse tree is built using default actions. Calling actions on the fly is again something that can hardly be done with PEG parsers because of backtracking.

Error recovery was much easier to implement as the parser state is explicit. In parglare you can use built-in error recovery strategy or write your own.

From the same grammar Parglare can construct both LR(1) (modified LALR(1), or SLR tables) and GLR parser. GLR could be used if the parser needs more than one token of lookahead to deduce next step or if the language is inherently ambiguous. In case of ambiguous parse, all trees will be returned by the GLR parser (a.k.a. the parse forest). Common sub-tree sharing is used which reduce memory usage significantly. Some common pitfalls with GLR algorithm are taken care of which is tested using published grammars that are specifically crafted to confuse GLR parser.

Describing everything that can be done with parglare would make this post too long, so if I’ve got you interested, you can find more info in the docs.

Where is it heading?

There are a lot of plans for parglare. First I would like to build some optional syntactic sugar for grammar language that would reduce boiler-plate for some common idioms. I’m also thinking about providing some implicit transformation to AST similar to textX. Lexical ambiguity detection for overlapping regexes would be nice too. And, of course, performance optimization is something that needs to be tackled from time to time.

But I’ll definitely try to keep the core simple and all additions optional.

If you have ideas, advices or any other feedback please use the discussion section on GitHub.