(Soon to be) the fastest pure-Python PEG parser I could muster
Parsimonious aims to be the fastest arbitrary-lookahead parser written in pure Python. 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.
- 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:
>>> 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.
0.3 is a pretty usable release for inputs that aren’t huge. I haven’t really started optimizing yet. And note that there may be API changes until we get to 1.0.
- 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 without fear.
- It may be slow and use a lot of RAM; I haven’t measured either yet. However, I have several macro- and micro-optimizations in mind.
- Error reporting is fairly uninformative, and debugging is nonexistent. However, repr methods of expressions, grammars, and nodes are very clear and helpful. Ones of Grammar objects are even round-trippable! Huge things are planned for grammar debugging in the future.
- 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.
- Optimizations to make Parsimonious worthy of its name
- Tighter RAM use
- Better-thought-out grammar extensibility story
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 trading the | 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 c. Thus, ambiguity is resolved by always yielding the first successful recognition.
Grammars are defined by a series of rules, one per line. The syntax should be familiar to anyone who uses regexes or reads programming language manuals. An example will serve best:
styled_text = bold_text / italic_text bold_text = "((" text "))" italic_text = "''" text "''" text = ~"[A-Z 0-9]*"i
|"some literal"||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 things. a b c matches spots where those 3 terms appear in that order.|
|a / b||Alternatives. The first to succeed of a / b / c wins.|
|thing?||An optional expression. This is greedy, always consuming thing if it exists.|
|&thing||A lookahead assertion. Ensures thing matches at the current position but does not consume it.|
|!thing||A negative lookahead assertion. Matches if thing isn’t found here. Doesn’t consume any text.|
|things*||Zero or more things. This is greedy, always consuming as many repetitions as it can.|
|things+||One or more things. This is greedy, always consuming as many repetitions as it can.|
|~r"regex"ilmsux||Regexes have ~ in front and are quoted like literals. Any flags follow the end quotes as single chars. Regexes are good for representing character classes ([a-z0-9]) and optimizing for speed. The downside is that they won’t be able to take advantage of our fancy debugging, once we get that working. Ultimately, I’d like to deprecate explicit regexes and instead have Parsimonious dynamically build them out of simpler primitives.|
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.
Bring your ? and * 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 "thing"? might.
The 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)): ...
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.
- It looks like we could make an architecture-independent .o file and use LLVM to JIT it to whatever arch we’re on: https://github.com/dabeaz/bitey/. Of course, then everybody has to have LLVM, which is even harder to set up than a vanilla C toolchain.
- 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.
- 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 & A as simply A &B.
- Take the print statements out of the benchmark tests.
- Give Node an evaluate-able __repr__.
- Support matching of prefixes and other not-to-the-end slices of strings by making match() public and able to initialize a new cache. Add match() callthrough method to Grammar.
- Report a BadGrammar exception (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.
- Add NodeVisitor.lift_child convenience method.
- Rename VisitationException to VisitationError for consistency with the standard Python exception hierarchy.
- Rework repr and str values 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”.
- A rough but useable preview release