Parsing Expression Grammars

Janet 0.4.0 Documentation
Special Forms
String Library
Data Structures
Object Oriented Programming
Abstract Machine

A common programming task is recognizing patterns in text, be it filtering emails from a list or extracting data from a CSV file. Programming languages and libraries usually offer a number of tools for this, including prebuilt parsers, simple operations on strings (splitting a string on commas), and regular expressions. The pre-built or custom-built parser is usually the most robust solution, but can be very complex to maintain and may not exist for many languages. String functions are not powerful enough for a large class of languages, and regular expressions can be hard to read (which characters are escaped?) and under-powered (don't parse HTML with regex!).

PEGs, or Parsing Expression Grammars, are another formalism for recognizing languages that are easier to write than a custom parser and more powerful than regular expressions. They also can produce grammars that are easily understandable and fast. PEGs can also be compiled to a bytecode format that can be reused. Janet offers the peg module for writing and evaluating PEGs.

Janet's peg module borrows syntax and ideas from both LPeg and REBOL/Red parse module. Janet has no built in regex module because PEGs offer a superset of regex's functionality.

Below is a simple example for checking if a string is a valid IP address. Notice how the grammar is descriptive enough that you can read it even if you don't know the peg syntax (example is translated from a RED language blog post).

(def ip-address
 '{:dig (range "09")
   :0-4 (range "04")
   :0-5 (range "05")
   :byte (choice
           (sequence "25" :0-5)
           (sequence "2" :0-4 :dig)
           (sequence "1" :dig :dig)
           (between 1 2 :dig))
   :main (sequence :byte "." :byte "." :byte "." :byte)})

(peg/match ip-address "") # -> @[]
(peg/match ip-address "elephant") # -> nil
(peg/match ip-address "") # -> nil


The peg module has few functions because the complexity is exposed through the pattern syntax. Note that there is only one match function, peg/match. Variations on matching, such as parsing or searching, can be implemented inside patterns. PEGs can also be compiled ahead of time with peg/compile if a PEG will be reused many times.

(peg/match peg text [,start=0] & arguments)

Match a peg against some text. Returns an array of captured data if the text matches, or nil if there is no match. The caller can provide an optional start index to begin matching the text at, otherwise the PEG starts on the first character of text. A peg can either a compile PEG object or peg source.

(peg/compile peg)

Compiles a peg source data structure into a new PEG. Throws an error if there are problems with the peg code.

Primitive Patterns

Larger patterns are built up with primitive patterns, which recognize individual characters, string literals, or a given number of characters. A character in Janet is considered a byte, so PEGs will work on any string of bytes. No special meaning is given to the 0 byte, or the string terminator in many languages.

Pattern Signature What it Matches
string ("cat") The literal string.
integer (3) Matches a number of characters, and advances that many characters. If negative, matches if not that many characters and does not advance. For example, -1 will match the end of a string
(range "az" "AZ") Matches characters in a range and advances 1 character. Multiple ranges can be combined together.
(set "abcd") Match any character in the argument string. Advances 1 character.

Primitive patterns are not that useful by themselves, but can be passed to peg/match and peg/compile like any other pattern.

(peg/match "hello" "hello") # -> @[]
(peg/match "hello" "hi") # -> nil
(peg/match 1 "hi") # -> @[]
(peg/match 1 "") # -> nil
(peg/match '(range "AZ") "F") # -> @[]
(peg/match '(range "AZ") "-") # -> nil
(peg/match '(set "AZ") "F") # -> nil
(peg/match '(set "ABCDEFGHIJKLMNOPQRSTUVWXYZ") "F") # -> @[]

Combining Patterns

These primitive patterns can be combined with several combinators to match a wide number of languages. These combinators can be thought of as the looping and branching forms in a traditional language (that is how they are implemented when compiled to bytecode).

Pattern Signature What it matches
(choice a b c ...) Tries to match a, then b, and so on. Will succeed on the first successful match, and fails if none of the arguments match the text.
(+ a b c ...) Alias for (choice a b c ...)
(sequence a b c) Tries to match a, b, c and so on in sequence. If any of these arguments fail to match the text, the whole pattern fails.
(* a b c ...) Alias for (sequence a b c ...)
(any x) Matches 0 or more repetitions of x.
(some x) Matches 1 or more repetitions of x.
(between min max x) Matches between min and max (inclusive) repetitions of x.
(at-least n x) Matches at least n repetitions of x.
(at-most n x) Matches at most n repetitions of x.
(if cond patt) Tries to match patt only if cond matches as well. cond will not produce any captures.
(if-not cond patt) Tries to match only if cond does not match. cond will not produce any captures.
(not patt) Matches only if patt does not match. Will not produce captures or advance any characters.
(! patt) Alias for (not patt)
(look offset patt) Matches only if patt matches at a fixed offset. offset can be any integer. patt will not produce captures and the peg will not advance any characters.
(> offset patt) Alias for (look offset patt)
(opt patt) Alias for (between 0 1 patt)
(? patt) Alias for (between 0 1 patt)

PEGs try to match an input text with a pattern in a greedy manner. This means that if a rule fails to match, that rule will fail and not try again. The only backtracking provided in a peg is provided by the (choice x y z ...) special, which will try rules in order until one succeeds, and the whole pattern succeeds. If no sub pattern succeeds, then the whole pattern fails. Note that this means that the order of x y z in choice DOES matter. If y matches everything that z matches, z will never succeed.


So far we have only been concerned with "does this text match this language?". This is useful, but it is often more useful to extract data from text if it does match a peg. The peg module uses that concept of a capture stack to extract data from text. As the PEG is trying to match a piece of text, some forms may push Janet values onto the capture stack as a side effect. If the text matches the main peg language, (peg/match) will return the final capture stack as an array.

Capture specials will only push captures to the capture stack if their child pattern matches the text. Most captures specials will match the same text as their first argument pattern. Also most specials that produce captures can take an optional argument tag that applies a keyword tag to the capture. These tagged captures can then be recaptured via the (backref tag) special in subsequent matches. Tagged captures, when combined with the (cmt) special, provide a powerful form of look-behind that can make many grammars simpler.

Pattern Signature What it captures
(capture patt ?tag) Captures all of the text in patt if patt matches, If patt contains any captures, then those captures will be pushed to the capture stack before the total text.
(<- patt ?tag) Alias for (capture patt ?tag)
(quote patt ?tag) Another alias for (capture patt ?tag). This allows code like 'patt to capture a pattern.
(group patt ?tag) Captures an array of all of the captures in patt.
(replace patt subst ?tag) Replaces the captures produced by patt by applying subst to them. If subst is a table or struct, will push (get subst last-capture) to the capture stack after removing the old captures. If a subst is a function, will call subst with the captures of patt as arguments and push the result to the capture stack. Otherwise, will push subst literally to the capture stack.
(/ patt subst ?tag) Alias for (replace patt subst ?tag)
(constant k ?tag) Captures a constant value and advances no characters.
(argument n ?tag) Captures the nth extra argument to the match function and does not advance.
(position ?tag) Captures the current index into the text and advances no input.
($ ?tag) Alias for (position ?tag).
(accumulate patt ?tag) Capture a string that is the concatenation of all captures in patt. This will try to be efficient and not create intermediate strings if possible.
(% patt ?tag) Alias for (accumulate patt ?tag)
(cmt patt fun ?tag) Invokes fun with all of the captures of patt as arguments (if patt matches). If the result is truthy, then captures the result. The whole expression fails if fun returns false or nil.
(backref tag ?tag) Duplicates the last capture with the tag tag. If no such capture exists then the match fails.
(-> tag ?tag) Alias for (backref tag).
(error patt) Throws a Janet error if patt matches. The error thrown will be the last capture of patt, or a generic error if patt produces no captures.
(drop patt) Ignores (drops) all captures from patt.

Grammars and Recursion

The feature that makes PEGs so much more powerful than pattern matching solutions like (vanilla) regex is mutual recursion. To do recursion in a peg, you can wrap multiple patterns in a grammar, which is a Janet struct. The patterns must be named by keywords, which can then be used in all sub-patterns in the grammar.

Each grammar, defined by a struct, must also have a main rule, called :main, that is the pattern that the entire grammar is defined by.

An example grammar that uses mutual recursion:

(def my-grammar
 '{:a (* "a" :b "a")
   :b (* "b" (+ :a 0) "b")
   :main (* "(" :b ")")})

(peg/match my-grammar "(bb)") # -> @[]
(peg/match my-grammar "(babbab)") # -> @[]
(peg/match my-grammar "(baab)") # -> nil
(peg/match my-grammar "(babaabab)") # -> nil

Keep in mind that recursion is implemented with a stack, meaning that very recursive grammars can overflow the stack. The compiler is able to turn some recursion into iteration via tail call optimization, but some patterns may fail on large inputs. It is also possible to construct (very poorly written) patterns that will result in long loops and be very slow in general.

String Searching and other Idioms

Although all pattern matching is done in anchored mode, operations like global substitution and searching can be implemented with the peg module. A simple Janet function that produces PEGs that search for strings shows how captures and looping specials can composed, and how quasiquoting can be used to embed values in patterns.

(defn finder
 "Creates a peg that finds all locations of str in the text."
 (peg/compile ~(any (+ (* ($) ,str) 1))))

(def where-are-the-dogs? (finder "dog"))

(peg/match where-are-the-dogs? "dog dog cat dog") # -> @[0 4 12]

# Our finder function also works any pattern, not just strings.

(def find-cats (finder '(* "c" (some "a") "t")))

(peg/match find-cats "cat ct caat caaaaat cat") # -> @[0 7 12 20]

We can also wrap a peg to turn it into a global substituion grammar with the accumulate special (%).

(defn replacer
 "Creates a peg that replaces instances of patt with subst."
 [patt subst]
 (peg/compile ~(% (any (+ (/ (<- ,patt) ,subst) (<- 1))))))