Though I'd rather avoid it, string parsing is a crucial part of programming. Since we're more than 60 years into the use of modern computers, it seems like we should have a pretty good handle on how to build abstractions over parsing. Indeed, there are tons of great tools out there, like GNU Yacc, Haskell's Parsec, newer Packrat-based parsers like Factor's EBNF syntax for PEGs, and a bunch of other high level parsing libraries. These libraries are relatively easy to use once you understand the underlying structure (each one parses a different subset of context-free grammars), because they expose the programmer to a tree-like view of the string.
However, these incur too much overhead to be used for certain domains, like the parsing that goes on in an HTTP server or client. They're really overkill when, as in the case of HTTP interchanges, what you're dealing with is a regular language and processing can be done on-line. (I'll get back later to what I mean by those two things.) The main tools that exist to deal with this are Lex and Ragel.
Ragel seems like a really interesting solution for this domain. The entire description of parsing is eventually put in one regular expression, which is compiled to a DFA, where states and transitions can be annotated by actions. Fine-grained control is given to limit non-determinism. But the user must be acutely aware of how regular expressions correspond to DFAs in order to use the abstraction. So it is somewhat leaky. Also, it's difficult to get a tree-like view on things: actions are used purely for their side effect.
So, here's an idea: let's find a middle ground. Let's try to use regular expressions, with all of their efficiency advantages, but get an abstract tree-like view of the grammar and an ability to use parsing actions like high-level abstractions allow. Ideally, the user won't have to know about the implementation beyond two simple facts: regular languages can't use general recursion, and nondeterminism should be minimized.
This isn't something that I've implemented, but I have a pretty good idea for the design of such as system, and I wanted to share it with you. First, a little background.
DFAs and regular expressions
I'm taking a computer science class about this right now, so I'm gonna be totally pedantic. When I say regular expression, I mean an expression that describes a regular language. Perl regexes aren't regular expressions (and Larry Wall knows this). If you don't feel like putting on your theoretician's hat, this blog post will be mostly meaningless.
What's a regular language? First off, a language is a set of strings. We care about infinite sets of strings, since finite sets are trivial to represent. If a string is in the language, that means that the language matches the string, intuitively. A regular language is one which can be represented by a deterministic finite automaton (DFA) without extensions, also called a finite state machine (FSM) for some reason. Many useful languages are regular, and many are not.
The idea of a DFA is a finite number of states and a transition function, which takes the current state and a character of a string and returns the next state. The transition function is defined on all states and all characters in the alphabet. There is a set of final states, and if the string runs out when the machine is in a final state, then the string is accepted. The language of the DFA is the set of strings accepted by the DFA. For a given DFA, that DFA can be run in linear time with respect to the length of the input string and constant space. It can also be run "on-line", that is, without the whole string in advance, going incrementally.
A related construction is an NFA, or nondeterministic finite automaton. Imagine the previous idea, but instead of a transition function, there is a transition relation. That is, for any character and current state, there are zero or more next states to go to, and the NFA always picks the right one. This is called nondeterminism (at least that's what it means here). Amazingly, NFAs can accept only regular languages and nothing more, because NFAs can be translated into DFAs. Basically, you build a DFA which picks all possible states at once, given all possible paths through the NFA. Potentially, though, there's an exponential blowup in the number of states.
Every regular expression can be converted into an equivalent NFA, which can be converted into a DFA, which can then be converted back into a regular expression. They're all equivalent. So then what's a regular expression? There are different ways to define it. One is that you can build up a regular expression from the following elements: the epsilon regex (matching the empty string), the empty regex (matching nothing), single character regexes (matching just a single character), concatenation (one followed by another), disjunction (or) and the Kleene star (0 or more copies of something). Counterintuitively, it's possible to construct regexes which support negation, conjunction, lookahead and other interesting things.
The most important distinction from Perl regexes is that regular expressions cannot contain backreferences, because these are provably impossible to express in a DFA. It's impossible to parse something with backreferences in the same linear time and constant space that you get from regexes which are regular. In fact, parsing patterns with backreferences is NP-hard and not believed possible in polynomial time (with respect to the length of the input string). Since regular expressions which are regular give us such nice properties, I'm going to stick to them.
Regular expressions in practice in parsing today
The formal study of regular languages is a basically solved problem within the formalism itself: they are equivalent to DFAs, and satisfy a convenient set of properties summarized by the pumping lemmaand the Myhill-Nerode theorem. The problem is just, is the given string a member of the language? What languages are regular?
This was solved in the 1950s and 1960s, and the basic results are in most introductory compiler books. Those books use the solution to build lexers, like Lex. Lex basically takes a list of regular expressions, each associated with an action, finds one of them to match maximally with the current input, executes the associated action on the portion that matches, and then repeats with the rest of the string. This is useful to build lexers, but the programmer has very little context for things, so it's difficult to use for much else.
More recently, Ragel has been used as a way to parse more complicated things using regular expressions. Its strategy is to turn its compile-time input into one big regular expression, annotated with actions on certain states or transitions. The actions are fragments of C code, and they form the processing power of the machine. However, their behavior can get a little unintuitive if too much nondeterminism is used, so Ragel provides a bunch of tools to limit that. Also, Ragel lets you explicitly specify a DFA through transitions, which seems useful but low-level.
Group capture with regexes
One of the most useful features of Perl's regular expressions is group capture. By this, I mean how you can do something like
s/^(1*)(0*)$/$2$1/to swap ones and zeros in a string. This is different from backreferences (like the non-regular expression
/(.*)$1/) because it's only used in subsequent code, to figure out what got matched to what part of the regex. It doesn't parse any languages which aren't regular, but it's a useful tool for processing.
Curiously, this has been ignored both by academics and DFA implementors so far. I hypothesize that it's been ignored by theorists for two reasons: (1) It's easy to confuse with backreferences, which make the language non-regular, which is totally uninteresting to theorists (2) They're not part of the formalism of regular expressions as previously expressed.
Implementors of (non-Perl-based) regular expression-based parsing mechanisms tend to avoid group capture because, in the general case, it's not fast enough and can't be done on-line. Also, as far as I can tell, it hasn't been implemented any other way than interpreting an NFA, using backtracking, and keeping track of where the parser is within the regex to determine group boundaries. This would be terrible for the domain of Lex and Ragel. By "on-line" I don't mean on the internet but rather that an algorithm that can be performed incrementally, getting little pieces (characters, say) of the input and doing computation incrementally, as the input is received, without storing the whole thing and running the algorithm all at once.
So how can we do group capture on-line? Well, in general, we can't. Consider the regular expression (1*)1 where you're trying to capture the group 1*. As the input is being processed, we don't know when we've gotten to the end of the group until the entire input is over, since if there are two more 1's, then the first one must be in the group. However, in many cases group capture can in fact be done on-line, as in (0*)(1*), where the groups captured are 0* and 1*. As the regex is processing on the string, it knows that, if there is a match, the group boundary is just before the first 1. This can be formalized as a "boundary of determinism": a point where, in the subset construction to form a DFA from an NFA gets a subset of exactly one state.
I believe this can handle most cases of group capture in practice, if the regular expression is well-written, but surely not all of them. I have an idea for how to do group capture in the few remaining circumstances, but unfortunately it takes linear space and it's not online. I'll blog about it once I have a proof of correctness.
Hierarchical parsing using group capture
Using this group capture mechanism, we can build a hierarchical parsing mechanism with actions on different things, which can be built to parse regular languages in a higher-level way. Regular expressions can't use arbitrary recursion like context-free grammars can, so the parse tree will be of fixed size, but it could still be useful. In designing this, I'm thinking specifically about making a SAX-like XML parser. It'd be awkward to write everything out as one big regular expression, but split into smaller things, each with their own little steps in processing, it could be much more elegant. My goal for syntax is something like EBNF syntax, as Chris Double's PEGs library in Factor does. Here's some future pseudocode for how it could look in parsing an XML tag, simplified. (In this code, > is used like Ragel :>>, to indicate that when the expression afterwards can be matched by the regex, it is, as soon as possible (basically).)
chars = "&" entity:any* > ";" [[ entity lookup-entity ]]
string = "\"" str:chars > "\"" [[ str ]]
| "'" str:chars > "'" [[ str ]]
xml-name = name-start-char name-char*
attribute = name:xml-name ws "=" ws str:string [[ name str 2array ]]
tag = "<" ws closer?:("/"?) name:xml-name attrs:(attribute*) ws contained?:("/"?) ws ">" [[ ... ]]
Though I haven't implemented this yet, and probably shouldn't even be talking about it, I'm really excited about this idea. I even came up with a stupid little name with it: Hegel, both for High-level Ragel and because it represents the synthesis of the dialectic (as described by Hegel) of slower, higher-level parsing and fast low-level parsing into fast, high-level parsing of regular languages. I hope it works.