I'm trying to understand the theory behind a lexer with the purpose of building one (just for my own fun and experience and to compensate for not taking proper CS courses :)).

What I have yet to understand is if lexer theory is the same no matter what language you are analyzing? Is it just about splitting white space then try to figure out what each text represents or is it more than that?

In other words, does the generation of tokens depend on the language being analyzed or is it just about white space? Can you apply the same lexer rules to all programming languages? For example if I have a lexer that tokenizes Java code can I use it for Python or not, since whitespace in Python has other meaning?

For example, I found a nice Python project (Pygments) which seems to provide a working core that allows you to later plugin rules for each of your favorite language (what are the keywords, the comments etc).

  • Huh? Are you trying to write lex or something else?
    – user40980
    Commented Jan 13, 2014 at 17:06
  • 1
    @MichaelT: If I read his question correctly, I think he's asking if you can write a lexer that doesn't require you to specify the rules of the programming language. I'd say the answer is probably no. Commented Jan 13, 2014 at 17:08
  • @MichaelT: I'm thinking a tool that reads an input file and outputs the tokens. In my understanding Lex is more than that, is it not?
    – JohnDoDo
    Commented Jan 13, 2014 at 17:24
  • I think you might have an alright question here though it's unclear right now. Please edit it to dictate specifically what the behaviour/requirements of this application you refer to as a "universal lexer" is, so we can clearly know what you would call "universal" lexing. Commented Jan 13, 2014 at 17:24
  • 1
    I occasionally write simplistic regex-based lexers (obviously, with mediocre throughput). The names and definitions of tokens can be swapped out without any problem to fit a different grammar. But often, writing a generic lexer isn't possible, e.g. when your whitespace handling is unusual or when you want to use advanced parsing techniques like layout parsing or ruby slippers parsing. But sure, you can write a generic lexer once you decide on a certain feature set.
    – amon
    Commented Jan 13, 2014 at 17:24

2 Answers 2


No you can't. Here is a nonsense snippet of Lisp:

(*foo-bar* 'baz 14)

What are the tokens in this snippet? The answer is something like

SYMBOL  "*foo-bar*"
SYMBOL  "baz"

Lisps have very liberal rules what can be inside an identifer: hyphens -, asterisks * and many other characters are allowed. Expressions can be quoted which prevents their evaluation, in many dialects this is done by prepending a ' to that expression.

A lexer needs to be aware of the language that it's parsing. If we would have used a C language lexer on the above code, we would get an error because ' introduces a character literal which also requires a closing ' – none is present here. Our C tokenizer might produce:

ERROR character literal isn't terminated: expected »'« after »'b«

For each language we want to lex, we have to write special rules. Every language is unique, and many languages do have subtle differences. Here, the identifiers are so very different that the token streams have lost any resemblance of each other. However, some languages inherit syntax from each other. For a very simple snippet like foo.bar(baz), a lexer written for C++ and one for Python might produce comparable results.

But why do we actually use Lexers?

Formally, lexers are not needed. We can describe a language without them just fine. Lexers are a performance optimization. Lexers are basically very restricted preprocessors for a parser that can match the input very efficiently (e.g. implemented as a state machine). The lexer then emits tokens, which are larger building blocks (a token is a pairing of a type or ID with a string). It is then the job of the parser to combine these tokens in a hierarchical structure (generally, an abstract syntax tree). This frees the parser from doing character-level operations, which makes it more efficient and easier to maintain (I have written parser without this separation. It works, but is annoying).

And here we have another problem: A lexer is written for a specific parser. The two work in tandem and cannot be recombined willy-nilly with other lexers or parsers. In the above token streams I have called the "(" operator LPAREN. But a different grammar might require a OP_LP symbol instead.

So for a variety of reasons, it's impossible to write one lexer to rule them all.

Each language needs it's own lexer. But we can take shortcuts and use programs that generate specialized lexers for us from some specification. Often, tools like lex are used for this. If performance is not an issue, I use the regular expressions library from the host language to cobble together a simple lexer, e.g. in Perl:

# a primitive Lisp lexer
my %tokens = (
  LPAREN => qr/[(]/,
  RPAREN => qr/[)]/,
  SYMBOL => qr/[^\s()']+/,
  QUOTE  => qr/[']/,
  INTEGER => qr/[1-9][0-9]*/,
my $space = qr/\s+/;

my $input = "(*foo-bar* 'baz 14)";
my $pos = 0;
my $length = length $input;

while ($pos < $length) {
  pos($input) = $pos;
  $pos += length $1 and next if $input =~ /\G($space)/gc;

  for my $name (keys %tokens) {
    if ($input =~ /\G($tokens{$name})/gc) {
      print "$name -- $1\n";
      $pos += length $1;
      next POSITION;

  die "No token matched, just before:\n", (substr $input, $pos, 20), "\n";


SYMBOL -- *foo-bar*
QUOTE -- '
SYMBOL -- baz
SYMBOL -- 14

But we can easily change the token definitions:

# an incomplete C lexer
my %tokens = (
  LPAREN => qr/[(]/,
  RPAREN => qr/[)]/,
  IDENTIFIER => qr/[a-zA-Z_][a-zA-Z0-9_]*/,
  CHARACTER  => qr/['](?:[^'\\]|\\.)[']/,
  INTEGER => qr/[1-9][0-9]*/,
  OP_STAR => qr/[*]/,
  OP_MINUS => qr/[-]/,
my $space = qr/\s+/;

which changes the output to:

OP_STAR -- *
OP_STAR -- *
No token matched, just before:
'baz 14)

We didn't write a new lexer, we just updated the token definitions.

What to do now:

  • As a preparation, understand and implement the Shunting-Yard Algorithm. This will teach you about a simple stack machine.

  • Read up on grammars and languages in computer science. Learn what a rule is.

  • Read up on Extended Backus-Naur Form notation, even if you don't immediately understand it. This is a way to write down grammars, and grammars are a way to describe languages (e.g. programming languages).

  • Understand the implications of the Chomsky Hierarchy. Here are the three important levels:

    • Regular Languages can be matched very efficiently. Lexers usually operate on this level.
    • Context-Free Languages are the basis for the description of many programming languages. Various parsing strategies can match various subsets. You will want to read up on LL, LR, and PEG at least.
    • Context-Sensitive Languages are even more general, but they are vastly more difficult to parse. Note that many regex libraries like PCRE have context-sensitive features.

    Note that real programming languages are not context-free – semantic whitespace as in Python, here-docs as in Bash, and optional end tags in HTML are there to mess up most elegant descriptions (but tricks exist to fake a context-free grammar even then).

  • Write a simple recursive-descent parser yourself that can evaluate simple arithmetics like 1 + 2 * 3 and manages operator precedence correctly. Using a lexer is not necessary for this. Add parens to your language, then some functions like abs or sin. Post your solution on codereview to get tips on how to do it better.

  • Now that you've learned how to parse stuff yourself, you may enjoy helping tools like parser generators. As an exercise, pick a simple language like JSON and write a parser for it with your tool of choice. Return to codereview for criticism.

If you have any problems on the way, ask here on programmers or over on stackoverflow, depending on whether the question is conceptual or implementation-related.


Sure you can: split on whitespace and call it a day!

Now, that is technically a universal lexer, in that you can work through any document and create a processed, tokenized result. Which is is really all a lexer has to do...it just isn't a very useful implementation!

A lexer is all about tokenization, and the rules that apply. And you can certainly write a lexer that simply allows you to plug in new rules to modify it's functioning!

Ultimately what you are trying to do is decouple the rules from their implementation, which is a perfectly reasonable sort of thing to do. You might find it works better, if you are working from scratch, to have that idea in mind but first just get it to work with some simple rules. Then, once it works at a basic level, wrap your head around how to pull out the rules themselves and allow them to be defined elsewhere, yet still be loaded into the program and executed.

Just remember: general-use software is always, always harder to make than specialized software. The more general you want it (do you want a lexer that supports natural language processing?), the less assumptions you can make and the more if() and switch statements you are going to have all over the place, and every optimization might break some needed edge-case feature. And it gets more arduous and time-consuming to test with every generality.

Between the lexer and the parser, the lexer is the easier one to do precisely because it is pretty much context free. It doesn't care if you are missing an end parens on an if statement, for instance, because it doesn't matter to it.

But yes, lexer theory itself is ultimately the same, its just the specific language rules that are different. For instance, in English text you can assume "iscrewedup" is supposed to be one word, and when you can't find that word in a dictionary you can assume something is not right - because written English words split on spaces. But not all language is so clean, and indeed programming langauges usually aren't, because if () and if() are equally valid in most languages.

And in Python indentation means something, while most languages don't care, but at a lexer level it just means a series of spaces or tabs gets tokenized as something like "IncreaseIndent" or "DecreaseIndent", and the parser figures out what that means later on. The theory, however, ultimately remains precisely the same.

  • With the question itself reworded, I suppose my explanation makes the answer "no, every language has it's own specific rules" ...I am undecided on what to do about that as far as this Q/A combination goes...
    – BrianH
    Commented Jan 13, 2014 at 19:18
  • In early FORTRAN syntax, blanks were optional. You could write DO100I=1,2,3 or DO 100 I = 1, 2, 3. And you could even put them inside identifiers - A B C and ABC were the same identifier. Commented Jan 13, 2014 at 21:53
  • @RossPatterson Yes, and fortunately that did not survive :) But sometimes looking back is useful to see flaws in conceptions. The use of whitespace is very language specific - some languages can contain whitespace within a single string definition (wrapping). And I won't start describing white space use in Python... Commented Jan 13, 2014 at 23:52
  • @owlstead I'm pretty sure modern Fortran still accepts missing spacing. Dunno about extra spacing, but Fortran standards have tried very hard not to invalidate even ancient programs, all the way back to FORTRAN II. Commented Jan 14, 2014 at 4:09

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.