a DSL for creating total and truly incremental parsers in Clojure
Switch branches/tags
Nothing to show
Pull request Compare This branch is 27 commits behind cgrand:master.
Fetching latest commit…
Cannot retrieve the latest commit at this time.
Failed to load latest commit information.




Parsley generates total and truly incremental parsers.

Total: a Parsley parser yields a parse-tree for any input string.

Truly incremental: a Parsley parser can operate as a text buffer, in best cases recomputing the parse-tree after a sequence of edits happens in logarithmic time (worst case: it behaves like a restartable parser).

Parsley parsers have no separate lexer, this allows for better compositionality of grammars.

For now Parsley uses the same technique (for lexer-less parsing) as described in this paper: Context-Aware Scanning for Parsing Extensible Languages http://www.umsec.umn.edu/publications/Context-Aware-Scanning-Parsing-Extensible-Language

(I independently rediscovered this technique and dubbed it LR+.)

Without a separate lexer, a language is entirely defined by its grammar. A grammar is an alternation of keywords (non-terminal names) and other values. A keyword and another value form a production rule.

Specifying grammars

A simple grammar is:

:expr #{"x" ["(" :expr* ")"]}

x () (xx) x)( are recognized by this grammar.

By default the main production of a grammar is the first one.

A production right value is a combination of:

  • strings and regexes (terminals — the set of terminal types is broader and even open, more later)

  • keywords (non-terminals) which can be suffixed by *, + or ? to denote repetitions or options.

  • sets to denote an alternative

  • vectors to denote a sequence. Inside vectors :*, :+ and :? are postfix unary operators. That is ["ab" :+] denotes a non-empty repetition of the ab string

A production left value is always a keyword. If this keyword is suffixed by -, no node will be generated in the parse-tree for this rule, its child nodes are inlined in the parent node. Rules with such names are called anonymous rules. An anonymous rule must be referred to by its base name (without the -).

These two grammars specify the same language but the resulting parse-trees will be different (additional :expr-rep nodes):

:expr #{"x" ["(" :expr* ")"]}
:expr #{"x" :expr-rep}
:expr-rep ["(" :expr* ")"]

These two grammars specify the same language and the same parse-trees:

:expr #{"x" ["(" :expr* ")"]}
:expr #{"x" :expr-rep}
:expr-rep- ["(" :expr* ")"]

Creating parsers

A parser is created using the parser or make-parser functions.

(require '[net.cgrand.parsley :as p])
(defn p (p/parser :expr #{"x" ["(" :expr* ")"]}))
(pprint (p "(x(x))"))

{:tag :net.cgrand.parsley/root,
 [{:tag :expr,
    {:tag :expr, :content ["x"]}
    {:tag :expr, :content ["(" {:tag :expr, :content ["x"]} ")"]}

; running on malformed input with garbage
(pprint (p "a(zldxn(dez)"))

{:tag :net.cgrand.parsley/unfinished,
 [{:tag :net.cgrand.parsley/unexpected, :content ["a"]}
  {:tag :net.cgrand.parsley/unfinished,
    {:tag :net.cgrand.parsley/unexpected, :content ["zld"]}
    {:tag :expr, :content ["x"]}
    {:tag :net.cgrand.parsley/unexpected, :content ["n"]}
    {:tag :expr,
      {:tag :net.cgrand.parsley/unexpected, :content ["dez"]}

Creating buffers

Creating a buffer, editing it and getting its resulting parse-tree:

(-> p p/incremental-buffer (p/edit 0 0 "(") (p/edit 1 0 "(x)") p/parse-tree pprint)

{:tag :net.cgrand.parsley/unfinished,
 [{:tag :net.cgrand.parsley/unfinished,
    {:tag :expr, :content ["(" {:tag :expr, :content ["x"]} ")"]}]}]}

Incremental parsing at work:

=> (def p (p/parser :expr #{"x" "\n" ["(" :expr* ")"]}))
=> (let [line (apply str "\n" (repeat 10 "((x))"))
         input (str "(" (apply str (repeat 1000 line)) ")")
         buf (p/incremental-buffer p)
         buf (p/edit buf 0 0 input)]
     (time (p/parse-tree buf))
     (time (p/parse-tree (-> buf (p/edit 2 0 "(") (p/edit 51002 0 ")"))))
"Elapsed time: 508.834 msecs"
"Elapsed time: 86.038 msecs"

Hence, reparsing the buffer only took a fraction of the original time despite the buffer having been modified at the start and at the end.

Incremental parsing

The input string is split into chunks (lines by default) and chunks are always reparsed as a whole, so don’t experiment with incremental parsing with 1-line inputs!

Let’s look at a bit more complex example:

=> (def p (p/parser
            {:main :expr*
             :space :ws?
             :make-node (fn [tag content] {:tag tag :content content :id (gensym)})}
            :ws #"\s+"
            :expr #{#"\w+" ["(" :expr* ")"]}))

This example introduces the option map: if the first arg to `parser` is a map
(instead of a keyword), it's a map of options. See <<options>> for more.

The important option here is that we redefine how nodes of the parse-tree are
constructed (via the `make-node` option). We add a unique identifier to each

Now let's create a 3-line input and parse it:

⇒ (def buf (→ p incremental-buffer (edit 0 0 "a)\n(b)\n(c"))) ⇒ (→ buf parse-tree pprint) nil {:tag :net.cgrand.parsley/root, :content [{:tag :expr, :content ["(" {:tag :expr, :content ["(" {:tag :expr, :content ["a"], :id G1806} ")"], :id G1807} {:tag :ws, :content ["\n"], :id G1808} {:tag :expr, :content ["(" {:tag :expr, :content ["b"], :id G1809} ")"], :id G1810} {:tag :ws, :content ["\n"], :id G1811} {:tag :expr, :content ["(" {:tag :expr, :content ["c"], :id G1812} ")"], :id G1813} ")"], :id G1814}], :id G1815}

Now, let's modify this "B" in "BOO" and parse the buffer again:

⇒ (→ buf (edit 6 1 "BOO") parse-tree pprint) nil {:tag :net.cgrand.parsley/root, :content [{:tag :expr, :content ["(" {:tag :expr, :content ["(" {:tag :expr, :content ["a"], :id G1806} ")"], :id G1807} {:tag :ws, :content ["\n"], :id G1818} {:tag :expr, :content ["(" {:tag :expr, :content ["BOO"], :id G1819} ")"], :id G1820} {:tag :ws, :content ["\n"], :id G1811} {:tag :expr, :content ["(" {:tag :expr, :content ["c"], :id G1812} ")"], :id G1813} ")"], :id G1821}], :id G1822}

We can spot that 5 out of the 10 nodes are shared with the previous parse-tree.

== Options ==

`:main` specifies the root production, by default this is the the first
production of the grammar.

`:root-tag` specifies the tag name to use for the root node
(`:net.cgrand.parsley/root` by default).

`:space` specifies a production which will be interspersed between every symbol
(terminal or not) *except in a sequence created with `unspaced`.*

`:make-node` specifies a function whose arglist is `[tag children-vec]` which
returns a new node. By default create instances the Node record with keys `tag`
and `content`.

`:make-unexpected` specifies a 1-arg function which converts a string (of
unexpected characters) to a node. By defaut delegates to `:make-node`.

`:make-leaf` specifies a 1-arg function which converts a string (token) to a
node, by default behaves like identity.