Parsing expression grammars (PEGs) are an analytic grammar formalism.
They feature the following operators:
- sequencing (e1 e2)
- unambiguous choice (e1/e2)
- (greedy) repetition (e*)
- option (e?)
- one-or-more (e+)
- positive lookahead (&e)
- negative lookahead (!e)
They are:
- fast to parse (O(n) in time and space, where n is the size of the input)
- unambiguous by construction
- closed under composition, which allows modular grammars
- more expressive than context-free grammars (they can parse L="a"^n"b"^n"c"^n)
However, they come with their share of problems:
- PEGs can't directly handle left-recursion (not much of a problem in practice)
- PEGs suffer from so called "prefix capturing": "a*a" will never match any input
- (this is what you get for defining away ambiguity)
- PEGs are bad for specifying languages: just guess what language "S <- aSa / aa" parses
- for the lazy: A string of "a"s with a length which is a power of two!
- PEGs can't generate strings, only parse them (due to lookahead operators)
- it seems to me that one could 'enforce' discovery of whatever one looked ahead to find
- Generating "(&A)B" would involve finding the intersection of A and B. This is an undecidable problem: If it would be possible to find the intersection of two PEGs we would not have issues detecting prefix capture. Prefix-capture is present in "A/B" iff the language generated by "(&A)B" is not empty. See [1] for more information.
[1] http://charybde.homeunix.org/~schmitz/pub/modular.pdf