compynator 0.1.2

A pure Python implementation of parser combinators with asymptotically best performance and support for context-sensitive or ambiguous grammars.

Introduction

Compynator is a tiny (~400 SLOCs), pure Python implementation of parser combinators. With this library, one can build up a complex parser from primitive parsers such as “get one token” (the One parser), or compose parsers together such as “if this parser fails, try that parser” (the Or combinator), etc. The mental model is a binary tree of execution nodes through which a sequence of input tokens flows.

The implementation in this package supports optional memoization and curtailment as described in [Frost2007]. This allows a memoized parser to achieve asymptotically best performance, and support ambiguous grammars.

Compynator is not an official Google product.

Examples

An example that solves https://leetcode.com/problems/parsing-a-boolean-expression:

t = Terminal('t').value(True)
f = Terminal('f').value(False)

e = Forward()
n = Terminal('!(').then(e).value(lambda x: not x).skip(')')

a_empty = Succeed(True)
a_e_tail = Forward()
a_e_tail.is_(
    Terminal(',').then(e).then(
    a_e_tail, lambda x, y: x and y) | a_empty)
a = Terminal('&(').then(e).then(
    a_e_tail, lambda x, y: x and y).skip(')')

o_empty = Succeed(False)
o_e_tail = Forward()
o_e_tail.is_(
    Terminal(',').then(e).then(
    o_e_tail, lambda x, y: x or y) | o_empty)
o = Terminal('|(').then(e).then(
    o_e_tail, lambda x, y: x or y).skip(')')

e.is_(t | f | n | a | o)

self.assertEqual(e('!(f)'), {Result(True, '')})
self.assertEqual(e('|(f,t)'), {Result(True, '')})
self.assertEqual(e('&(t,f)'), {Result(False, '')})
self.assertEqual(e('|(&(t,f,t),!(t))'), {Result(False, '')})

See test_benchmark.py for a literal translation of the ABNF rules for URI parsing as given in [RFC3986] into a structured tree:

Uri
    Scheme http
    HierPart
        Authority
            UserInfo None
            Host www.ics.uci.edu
            Port None
        Path /pub/ietf/uri/
    Query query
    Fragment fragment

There are more examples in the tests directory.

Translating Augmented Backus-Naur Form (ABNF)

Care should be taken to ensure that repeat takes all possible results instead of greedily parsing for the most. For example, the ABNF *3"a" "aa" cannot be simply translated as Terminal('a').repeat(0, 3) + 'aa'. This translation will fail to parse aaaa because it greedily matches the first 3 a’s, then fails to find the remaining 2 a’s. The correct translation should be Terminal('a').repeat(0, 3, take_all=True) + 'aa'.

ABNF	Compynator
Terminal	Terminal
Rule	Parser
Concatenation	p1 + p2
Alternative	p1 | p2
Sequence group	Normal use of parentheses
Variable repetition	p.repeat(lower, upper, take_all)
Specific repetition	p.repeat(x, x)
Optional	p.repeat(0, 1, take_all)

Translating Parsing Expression Grammar (PEG)

Unlike ABNF, PEG operators are always greedy. When translating PEG, we do not need to worry about backtracking the repetitions with take_all.

PEG	Compynator
Terminal	Terminal
Nonterminal	Parser
Epsilon	Empty
Sequence	p1 + p2
Ordered choice	p1 | p2
Zero or more	p.repeat()
One or more	p.repeat(1)
Optional	p.repeat(0, 1)
And predicate	Lookahead(p)
Not predicate	Lookahead(p, take_if=False)

Combinator vs Generator

Advantages

Advantages of parser combinators versus parser generators are:

1. Readability. A grammar can be expressed in a very similar form as its BNF. The code can be considered an executable specification of the grammar.

2. Simple setup. The code is the grammar. There is no need to run a generator to regenerate code when the grammar changes.

3. Understandability. Each parser is generally short and simple that its correctness can be easily verified. There is no need to look into generated code, or the code of the parser generator.

4. Parser combinators support context-sensitive grammars. For example, to parse an XML body, assuming start parses a start tag, body parses the body, and end parses a specified end tag:

   xml_tag = start.then(lambda tag_name: body.skip(end(tag_name)))

5. Combination of lexing and parsing. Most parser generators perform their lexing and parsing phases separately. Parser combinators combine these phases together. Hence they are not limited to string inputs. The example (in test_core.py) below takes a tokenized sequence.

   NUM, OP, TERMINAL = 0, 1, 2
   tokens = [(NUM, 2), (OP, operator.add), (NUM, 10),
             (OP, operator.mul), (NUM, 4)]
   num = One.where(lambda c: c[0] == NUM)
   op = One.where(lambda c: c[0] == OP).value(lambda c: c[1])
   mult_div = op.where(lambda c: c in (operator.mul, operator.truediv))
   add_sub = op.where(lambda c: c in (operator.add, operator.sub))
   left_paren = One.where(lambda c: c[0] == TERMINAL and c[1] == '(')
   right_paren = One.where(lambda c: c[0] == TERMINAL and c[1] == ')')
   expr = Forward()
   factor = (
       num.value(lambda t: t[1]) |
       left_paren.then(expr).skip(right_paren)
   )
   def do_op(left, op, right):
       return op(left, right)
   term = Forward()
   term.is_((
       Collect(term, mult_div, factor).value(lambda v: do_op(*v)) ^
       factor
   ).memoize())
   expr.is_((
       Collect(expr, add_sub, term).value(lambda v: do_op(*v)) ^
       term
   ).memoize())
   calc = expr.filter(lambda r: not r.remain)
   self.assertEqual(
           set(expr(tokens)),
           {
               Result(value=42, remain=[]),
               Result(value=12, remain=tokens[3:]),
               Result(value=2, remain=tokens[1:]),
           })
   self.assertEqual(calc(tokens), Succeed(42)([]))

Disadvantages

Disadvantages of parser combinators are:

1. Familiarity. Most textbooks write about parser generators and traditional parsing techniques such as LL, LR, etc. Parser combinators are more common in functional and logic programming communities, as popularized by [Wadler1985] and [Hutton1992].

2. Coupling of code and grammar. The downside of simple setup is a tight coupling of code and grammar, which might make it difficult to understand.

3. As it is implemented here, performance might be impacted due to composition overhead. See test_benchmark.py for details. On the same machine, the result for URI parsing could be ~70 times slower:

   t.test_parse_uri() 903.5961110000001 usec per run
   t.test_urlparse() 13.704007000000074 usec per run

4. All the advantages and disadvantages of scannerless parsing apply too.

Limitations

Currently, this library does not implement:

1. Source context such as line and column number of the token.

2. “Greedy” matching in the same sense as in regular expression (i.e. longest match). The greedy operation in this library is in the “greedy algorithm” sense, i.e. the first rule that matches will be taken.

3. Space treatments. Spaces have to be explicitly taken care of in grammars.

References

[Wadler1985]

Wadler, Philip. (1985). “How to replace failure by a list of successes”. Proc. conference on functional programming and computer architecture. Springer–Verlag.

[Hutton1992]

Hutton, Graham. (1992). “Higher-order functions for parsing”. Journal of functional programming, 2(3), 323–343.

[Frost2007]

Frost R.A., Hafiz R., Callaghan P. (2007) “Parser Combinators for Ambiguous Left-Recursive Grammars”. In: Hudak P., Warren D.S. (eds) “Practical Aspects of Declarative Languages”. PADL 2008. Lecture Notes in Computer Science, vol 4902. Springer, Berlin, Heidelberg

[RFC3986]

Berners-Lee, T., Fielding, R., and L. Masinter, “Uniform Resource Identifier (URI): Generic Syntax”, STD 66, RFC 3986, January 2005.