Parsimonious aims to be the fastest arbitrary-lookahead parser written in pure Python—and the most usable. It's based on parsing expression grammars (PEGs), which means you feed it a simplified sort of EBNF notation. Parsimonious was designed to undergird a MediaWiki parser that wouldn't take 5 seconds or a GB of RAM to do one page, but it's applicable to all sorts of languages.
- Frugal RAM use
- Minimalistic, understandable, idiomatic Python code
- Readable grammars
- Extensible grammars
- Complete test coverage
- Separation of concerns. Some Python parsing kits mix recognition with instructions about how to turn the resulting tree into some kind of other representation. This is limiting when you want to do several different things with a tree: for example, render wiki markup to HTML or to text.
- Good error reporting. I want the parser to work with me as I develop a grammar.
Here's how to build a simple grammar:
>>> from parsimonious.grammar import Grammar >>> grammar = Grammar( ... """ ... bold_text = bold_open text bold_close ... text = ~"[A-Z 0-9]*"i ... bold_open = "((" ... bold_close = "))" ... """)
You can have forward references and even right recursion; it's all taken care of by the grammar compiler. The first rule is taken to be the default start symbol, but you can override that.
Next, let's parse something and get an abstract syntax tree:
>>> print grammar.parse('((bold stuff))') <Node called "bold_text" matching "((bold stuff))"> <Node called "bold_open" matching "(("> <RegexNode called "text" matching "bold stuff"> <Node called "bold_close" matching "))">
You'd typically then use a
nodes.NodeVisitor subclass (see below) to walk
the tree and do something useful with it.
- Everything that exists works. Test coverage is good.
- I don't plan on making any backward-incompatible changes to the rule syntax in the future, so you can write grammars with confidence.
- It may be slow and use a lot of RAM; I haven't measured either yet. However, I have yet to begin optimizing in earnest.
- Error reporting is now in place.
reprmethods of expressions, grammars, and nodes are clear and helpful as well. The
Grammarones are even round-trippable!
- The grammar extensibility story is underdeveloped at the moment. You should be able to extend a grammar by simply concatening more rules onto the existing ones; later rules of the same name should override previous ones. However, this is untested and may not be the final story.
- Sphinx docs are coming, but the docstrings are quite useful now.
- Note that there may be API changes until we get to 1.0, so be sure to pin to the version you're using.
- Optimizations to make Parsimonious worthy of its name
- Tighter RAM use
- Better-thought-out grammar extensibility story
- Amazing grammar debugging
A Little About PEG Parsers
PEG parsers don't draw a distinction between lexing and parsing; everything is done at once. As a result, there is no lookahead limit, as there is with, for instance, Yacc. And, due to both of these properties, PEG grammars are easier to write: they're basically just a more practical dialect of EBNF. With caching, they take O(grammar size * text length) memory (though I plan to do better), but they run in O(text length) time.
PEGs can describe a superset of LL(k) languages, any deterministic LR(k)
language, and many others—including some that aren't context-free
(http://www.brynosaurus.com/pub/lang/peg.pdf). They can also deal with what
would be ambiguous languages if described in canonical EBNF. They do this by
| alternation operator for the
/ operator, which works the
same except that it makes priority explicit:
a / b / c first tries matching
a. If that fails, it tries
b, and, failing that, moves on to
Thus, ambiguity is resolved by always yielding the first successful recognition.
Grammars are defined by a series of rules. The syntax should be familiar to anyone who uses regexes or reads programming language manuals. An example will serve best:
my_grammar = Grammar(r""" styled_text = bold_text / italic_text bold_text = "((" text "))" italic_text = "''" text "''" text = ~"[A-Z 0-9]*"i """)
You can wrap a rule across multiple lines if you like; the syntax is very forgiving.
||Used to quote literals. Backslash escaping and Python conventions for "raw" and Unicode strings help support fiddly characters.|
|[space]||Sequences are made out of space- or tab-delimited
||Alternatives. The first to succeed of
||An optional expression. This is greedy, always consuming
||A lookahead assertion. Ensures
||A negative lookahead assertion. Matches if
||Zero or more things. This is greedy, always consuming as many repetitions as it can.|
||One or more things. This is greedy, always consuming as many repetitions as it can.|
||Parentheses are used for grouping, like in every other language.|
Don't Repeat Expressions
If you need a
~"[a-z0-9]"i at two points in your grammar, don't type it
twice. Make it a rule of its own, and reference it from wherever you need it.
You'll get the most out of the caching this way, since cache lookups are by
expression object identity (for speed).
Even if you have an expression that's very simple, not repeating it will save RAM, as there can, at worst, be a cached int for every char in the text you're parsing. In the future, we may identify repeated subexpressions automatically and factor them up while building the grammar.
How much should you shove into one regex, versus how much should you break them up to not repeat yourself? That's a fine balance and worthy of benchmarking. More stuff jammed into a regex will execute faster, because it doesn't have to run any Python between pieces, but a broken-up one will give better cache performance if the individual pieces are re-used elsewhere. If the pieces of a regex aren't used anywhere else, by all means keep the whole thing together.
* quantifiers up to the highest level you
can. Otherwise, lower-level patterns could succeed but be empty and put a bunch
of useless nodes in your tree that didn't really match anything.
Processing Parse Trees
A parse tree has a node for each expression matched, even if it matched a
zero-length string, like
NodeVisitor class provides an inversion-of-control framework for
walking a tree and returning a new construct (tree, string, or whatever) based
on it. For now, have a look at its docstrings for more detail. There's also a
good example in
grammar.RuleVisitor. Notice how we take advantage of nodes'
iterability by using tuple unpacks in the formal parameter lists:
def visit_or_term(self, or_term, (slash, _, term)): ...
For reference, here is the production the above unpacks:
or_term = "/" _ term
When something goes wrong in your visitor, you get a nice error like this:
[normal traceback here...] VisitationException: 'Node' object has no attribute 'foo' Parse tree: <Node called "rules" matching "number = ~"[0-9]+""> <-- *** We were here. *** <Node matching "number = ~"[0-9]+""> <Node called "rule" matching "number = ~"[0-9]+""> <Node matching ""> <Node called "label" matching "number"> <Node matching " "> <Node called "_" matching " "> <Node matching "="> <Node matching " "> <Node called "_" matching " "> <Node called "rhs" matching "~"[0-9]+""> <Node called "term" matching "~"[0-9]+""> <Node called "atom" matching "~"[0-9]+""> <Node called "regex" matching "~"[0-9]+""> <Node matching "~"> <Node called "literal" matching ""[0-9]+""> <Node matching ""> <Node matching ""> <Node called "eol" matching " "> <Node matching "">
The parse tree is tacked onto the exception, and the node whose visitor method raised the error is pointed out.
Why No Streaming Tree Processing?
Some have asked why we don't process the tree as we go, SAX-style. There are two main reasons:
- It wouldn't work. With a PEG parser, no parsing decision is final until the whole text is parsed. If we had to change a decision, we'd have to backtrack and redo the SAX-style interpretation as well, which would involve reconstituting part of the AST and quite possibly scuttling whatever you were doing with the streaming output. (Note that some bursty SAX-style processing may be possible in the future if we use cuts.)
- It interferes with the ability to derive multiple representations from the AST: for example, turning wiki markup into first HTML and then text.
Rule Syntax Changes
- Maybe support left-recursive rules like PyMeta, if anybody cares.
- Ultimately, I'd like to get rid of explicit regexes and break them into more atomic things like character classes. Then we can dynamically compile bits of the grammar into regexes as necessary to boost speed.
- Make RAM use almost constant by automatically inserting "cuts", as described in http://ialab.cs.tsukuba.ac.jp/~mizusima/publications/paste513-mizushima.pdf. This would also improve error reporting, as we wouldn't backtrack out of everything informative before finally failing.
- Find all the distinct subexpressions, and unify duplicates for a better cache hit ratio.
- Think about having the user (optionally) provide some representative input along with a grammar. We can then profile against it, see which expressions are worth caching, and annotate the grammar. Perhaps there will even be positions at which a given expression is more worth caching. Or we could keep a count of how many times each cache entry has been used and evict the most useless ones as RAM use grows.
- We could possibly compile the grammar into VM instructions, like in "A parsing machine for PEGs" by Medeiros.
- If the recursion gets too deep in practice, use trampolining to dodge it.
- Pijnu has a raft of tree manipulators. I don't think I want all of them, but a judicious subset might be nice. Don't get into mixing formatting with tree manipulation. https://github.com/erikrose/pijnu/blob/master/library/node.py#L333. PyPy's parsing lib exposes a sane subset: http://doc.pypy.org/en/latest/rlib.html#tree-transformations.
- Make Grammar iteration ordered, making the
__repr__more like the original input. (Lucas Wiman)
- Improve text representation and error messages for anonymous subexpressions. (Lucas Wiman)
- Expose BadGrammar and VisitationError as top-level imports.
- No longer crash when you try to compare a Node to an instance of a different class. (Esben Sonne)
sixat 1.9.0 to ensure we have
python_2_unicode_compatible. (Sam Raker)
- Drop Python 2.6 support.
- Make Grammar iteration ordered, making the
- Add experimental token-based parsing, via TokenGrammar class, for those operating on pre-lexed streams of tokens. This can, for example, help parse indentation-sensitive languages that use the "off-side rule", like Python. (Erik Rose)
- Common codebase for Python 2 and 3: no more 2to3 translation step (Mattias Urlichs, Lucas Wiman)
- Drop Python 3.1 and 3.2 support.
- Fix a bug in
Grammar.__repr__which fails to work on Python 3 since the string_escape codec is gone in Python 3. (Lucas Wiman)
- Don't lose parentheses when printing representations of expressions. (Michael Kelly)
- Make Grammar an immutable mapping (until we add automatic recompilation). (Michael Kelly)
- Make grammar compilation 100x faster. Thanks to dmoisset for the initial patch.
- Fix bug which made the default rule of a grammar invalid when it contained a forward reference.
This release makes backward-incompatible changes:
default_rulearg to Grammar's constructor has been replaced with a method,
some_grammar.default('rule_name'), which returns a new grammar just like the old except with its default rule changed. This is to free up the constructor kwargs for custom rules.
UndefinedLabelis no longer a subclass of
VisitationError. This matters only in the unlikely case that you were catching
VisitationErrorexceptions and expecting to thus also catch
- Add support for "custom rules" in Grammars. These provide a hook for simple
custom parsing hooks spelled as Python lambdas. For heavy-duty needs,
you can put in Compound Expressions with LazyReferences as subexpressions,
and the Grammar will hook them up for optimal efficiency--no calling
__getitem__on Grammar at parse time.
- Allow grammars without a default rule (in cases where there are no string rules), which leads to also allowing empty grammars. Perhaps someone building up grammars dynamically will find that useful.
@ruledecorator, allowing grammars to be constructed out of notations on
NodeVisitormethods. This saves looking back and forth between the visitor and the grammar when there is only one visitor per grammar.
match()convenience methods to
NodeVisitor. This makes the common case of parsing a string and applying exactly one visitor to the AST shorter and simpler.
- Improve exception message when you forget to declare a visitor method.
NodeVisitor, letting you name certain exceptions which propagate out of visitors without being wrapped by
- Expose much more of the library in
__init__, making your imports shorter.
- Drastically simplify reference resolution machinery. (Vladimir Keleshev)
This release makes some backward-incompatible changes. See below.
- Add alpha-quality error reporting. Now, rather than returning
ParseErrorif they don't succeed. This makes more sense, since you'd rarely attempt to parse something and not care if it succeeds. It was too easy before to forget to check for a
ParseErrorgives you a human-readable unicode representation as well as some attributes that let you construct your own custom presentation.
- Grammar construction now raises
BadGrammarif it can't parse your rules.
parse()now takes an optional
- Make the
UndefinedLabelreturn the right type.
- Support splitting rules across multiple lines, interleaving comments, putting multiple rules on one line (but don't do that) and all sorts of other horrific behavior.
- Tolerate whitespace after opening parens.
- Add support for single-quoted literals.
- Add alpha-quality error reporting. Now, rather than returning
- Support Python 3.
- Rewrite grammar compiler so right-recursive rules can be compiled and parsing no longer fails in some cases with forward rule references.
- Support comments, the
!("not") operator, and parentheses in grammar definition syntax.
- Change the
&operator to a prefix operator to conform to the original PEG syntax. The version in Parsing Techniques was infix, and that's what I used as a reference. However, the unary version is more convenient, as it lets you spell
AB & Aas simply
- Take the
- Give Node an evaluate-able
- Support comments, the
- Support matching of prefixes and other not-to-the-end slices of strings by
match()public and able to initialize a new cache. Add
match()callthrough method to
- Report a
BadGrammarexception (rather than crashing) when there are mistakes in a grammar definition.
- Simplify grammar compilation internals: get rid of superfluous visitor methods and factor up repetitive ones. Simplify rule grammar as well.
VisitationErrorfor consistency with the standard Python exception hierarchy.
strvalues for grammars and expressions. Now they both look like rule syntax. Grammars are even round-trippable! This fixes a unicode encoding error when printing nodes that had parsed unicode text.
- Add tox for testing. Stop advertising Python 2.5 support, which never worked (and won't unless somebody cares a lot, since it makes Python 3 support harder).
- Settle (hopefully) on the term "rule" to mean "the string representation of a production". Get rid of the vague, mysterious "DSL".
- Support matching of prefixes and other not-to-the-end slices of strings by making
- A rough but useable preview release
Thanks to Wiki Loves Monuments Panama for showing their support with a generous gift.