Parsing expression grammar (PEG) was introduced by Bryan Ford in 20041. Its characteristic is its output is a single, non-ambiguous parse tree.

The PEG grammar is \((N, \Sigma, P, e_S)\), where

  • \(N\): a finite set nonterminal symbols
  • \(\Sigma\): a finite set of terminal symbols that is disjoint from \(N\)
  • \(P\): a finite set parsing rules
  • \(e_S\): an expression termed the starting expression

The parsing expression is any terminal symbol, any nonterminal symbol, the empty string \(\epsilon\) (these three: atomic parsing expression), or any compound expressions:

  • Sequence: e1 e2
  • Ordered choice: e1 / e2
  • Zero-or-more: e*
  • One-or-more: e+
  • Optional: e?
  • And-predicate: &e
  • Not-predicate: !e

It is quite intuitive from the simularity in structure that the PEG grammar is context free. Indeed, the only difference I can see between PEG and CFG is the ordered choice vs unordered choice, and the two predicates. This should not give PEG any more power than the CFG.

However, because of this, it is known that PEG can’t parse the indentation level in Python code2. Indeed, many computer languages are not context free and not parsable by PEG. In my experiment to implement a parser for Markdown (CommonMark to be exact) using neg, I found we cannot parse the very simple structure like the following:

#{n} [^#]+ #{m}

which m is no less than n and the grammar should consume all m pound sign. This is an example of context sensitivity: we need to refer to the beginning of the line to check what is n so we can determine if this is valid.

To solve problems like this, I extended neg to include a subparser:

class SequenceParser < CompositeParser
  def do_parse(i, opts)
    results = []
    @children.each do |c|
      #results << c.parse(i, opts)
      if c.respond_to? :call then
        # semantic predicates
        results << [, results.last[1],, nil, [] ]
        # parser
        results << c.parse(i, opts)

      break unless results.last[2]
    [ results.last[2], nil, results ]

# Predicate will not modify the results nor consume input but instead, return true or false to indicate whether to accept it
class PredicateParser < SubParser
  attr_reader :name
  @@scoreboard = {}
  def initialize(name='', &block)
    @name = name
    @predicate = block_given? ? block : nil
  def call(results)
    status =
    if @name && status
      @@scoreboard[@name] = results # remember result if predicate is true
  def self.scoreboard
  def to_s(parent=nil)

The SequenceParser above replaced the c.parse() call with an if block to allow to call a function whenever it exists and the Predicate subparser is to include a function block in the middle of a parse rule. Then we can use this to implement the fencing in CommonMark:

fence_block_closed == \
    fence + Predicate('fence'){true} + \
    info_string + nl + \
    (code_line + Predicate(){|x|
        ofence = PredicateParser.scoreboard['fence'].first[3].strip
        thisline = x.first[3]
        indent = thisline[/\A */].size
        unindented = thisline[indent,thisline.length]
        (indent > 3 || !unindented.start_with?(ofence)) # closing fence should be at least as long as opening
    }) * 0 + \
    fence + space * 0 + eol

The Predicate rule in the example above make use of the scoreboard feature to remember states, which empowered the CFG to learn some context.

Indeed, this is an example of using external state in PEG. People argue that this can be dangerous if used not carefully, as PEG is backtracking: when PEG backtracks, the external state does not rollback and hence it could be out of sync with the parsing system. In parsimonious, such shortcoming of PEG drove the implementation of TokenGrammar, but for a different approach: instead of producing a parse tree in one shot, it first parse input into tokens and use a second set of grammar to parse the result. Below is an example from

>>> from parsimonious.utils import Token
>>> from parsimonious import TokenGrammar
>>> grammar = TokenGrammar("""
... bold_text = bold_open text bold_close
... text = "TEXT"
... bold_open = "BOLD_OPEN"
... bold_close = "BOLD_CLOSE"
... """)
>>> tl = []
>>> tl.append(Token(type='BOLD_OPEN'))
>>> tl.append(Token(type='TEXT'))
>>> tl.append(Token(type='BOLD_CLOSE'))
>>> print(grammar.parse(tl))
<Node called "bold_text" matching "[<Token "BOLD_OPEN">, <Token "TEXT">, <Token "BOLD_CLOSE">]">
    <Node called "bold_open" matching "[<Token "BOLD_OPEN">]">
    <Node called "text" matching "[<Token "TEXT">]">
    <Node called "bold_close" matching "[<Token "BOLD_CLOSE">]">

Considering the two parser that I looked at, the neg and parsimonious, I see that the parse tree are generated progressively with one node per rule (with children nodes created in the sub-rules of grammar). I believe the right way is to extend the sequence construct and create a “look behind” predicate such that:

  • the “look behind” only works in sequence and fails everywhere else to make the implementation simple
  • the predicate is not a grammar rule but a small, immutable function defined elsewhere with no side effects
  • when used in the middle of a sequence, it gives what is parsed so far (i.e. a node with arbitrary number of children) to the function and the function will determine if the node passed in is acceptable or not

Using the fencing example above, this predicate can look for children in the node that matches the opening and closing fence and verify the string length matches the rule of \(m\ge n\).

  1. Bryan Ford (2004). “Parsing Expression Grammars: A Recognition Based Syntactic Foundation”. Proceedings of the 31st ACM SIGPLAN-SIGACT Symposium on Principles of Programming Languages. ACM. 

  2. StackOverflow - Parse indentation level with peg.js