The Fast Lexical Analyser Generator

Copyright ©1998-2005 by Gerwin Klein

JFlex User's Manual

Version 1.4.1, July 25, 2005



JFlex is a lexical analyzer generator for Java[*]written in Java. It is also a rewrite of the very useful tool JLex [3] which was developed by Elliot Berk at Princeton University. As Vern Paxson states for his C/C++ tool flex [11]: they do not share any code though.

Design goals

The main design goals of JFlex are:

About this manual

This manual gives a brief but complete description of the tool JFlex. It assumes that you are familiar with the issue of lexical analysis. The references [1], [2], and [13] provide a good introduction to this topic.

The next section of this manual describes installation procedures for JFlex. If you never worked with JLex or just want to compare a JLex and a JFlex scanner specification you should also read Working with JFlex - an example (section 3). All options and the complete specification syntax are presented in Lexical specifications (section 4); Encodings, Platforms, and Unicode (section 5) provides information about scannig text vs. binary files. If you are interested in performance considerations and comparing JLex with JFlex speed, a few words on performance (section 6) might be just right for you. Those who want to use their old JLex specifications may want to check out section 7.1 Porting from JLex to avoid possible problems with not portable or non standard JLex behavior that has been fixed in JFlex. Section 7.2 talks about porting scanners from the Unix tools lex and flex. Interfacing JFlex scanners with the LALR parser generators CUP and BYacc/J is explained in working together (section 8). Section 9 Bugs gives a list of currently known active bugs. The manual concludes with notes about Copying and License (section 10) and references.

Installing and Running JFlex

Installing JFlex


To install JFlex on Windows 95/98/NT/XP, follow these three steps:
  1. Unzip the file you downloaded into the directory you want JFlex in (using something like WinZip). If you unzipped it to say C:\, the following directory structure should be generated:
          +--bin\                   (start scripts) 
          +--doc\                   (FAQ and manual)
                   +--binary\       (scanning binary files)
                   +--byaccj\       (calculator example for BYacc/J)
                   +--cup\          (calculator example for cup)
                   +--interpreter\  (interpreter example for cup)
                   +--java\         (Java lexer specification) 
                   +--simple\       (example scanner)
                   +--standalone\   (a simple standalone scanner) 
          +--lib\                   (the precompiled classes) 
              +--JFlex\             (source code of JFlex) 
              +--JFlex\gui          (source code of JFlex UI classes)
              +--java_cup\runtime\  (source code of cup runtime classes)

  2. Edit the file bin\jflex.bat (in the example it's C:\JFlex\bin\jflex.bat) such that

  3. Include the bin\ directory of JFlex in your path. (the one that contains the start script, in the example: C:\JFlex\bin).

Unix with tar archive

To install JFlex on a Unix system, follow these two steps:

You can verify the integrity of the downloaded file with the MD5 checksum available on the JFlex download page. If you put the checksum file in the same directory as the archive, you run:

md5sum --check jflex-1.4.1.tar.gz.md5

It should tell you

jflex-1.4.1.tar.gz: OK

Linux with RPM

You can verify the integrity of the downloaded rpm file with

rpm --checksig jflex-1.4.1-0.rpm

Running JFlex

You run JFlex with:

jflex <options> <inputfiles>

It is also possible to skip the start script in bin\ and include the file lib\JFlex.jar in your CLASSPATH environment variable instead.

Then you run JFlex with:

java JFlex.Main <options> <inputfiles>

The input files and options are in both cases optional. If you don't provide a file name on the command line, JFlex will pop up a window to ask you for one.

JFlex knows about the following options:

-d <directory>
writes the generated file to the directory <directory>

--skel <file>
uses external skeleton <file>. This is mainly for JFlex maintenance and special low level customizations. Use only when you know what you are doing! JFlex comes with a skeleton file in the src directory that reflects exactly the internal, precompiled skeleton and can be used with the -skel option.

skip the DFA minimization step during scanner generation.

tries even harder to comply to JLex interpretation of specs.

generate graphviz dot files for the NFA, DFA and minimized DFA. This feature is still in alpha status, and not fully implemented yet.

display transition tables of NFA, initial DFA, and minimized DFA

--verbose or -v
display generation progress messages (enabled by default)

--quiet or -q
display error messages only (no chatter about what JFlex is currently doing)

display time statistics about the code generation process (not very accurate)

print version number

print system and JDK information (useful if you'd like to report a problem)

use the %pack code generation method by default

use the %table code generation method by default

use the %switch code generation method by default

--help or -h
print a help message explaining options and usage of JFlex.

A simple Example: How to work with JFlex

To demonstrate what a lexical specification with JFlex looks like, this section presents a part of the specification for the Java language. The example does not describe the whole lexical structure of Java programs, but only a small and simplified part of it (some keywords, some operators, comments and only two kinds of literals). It also shows how to interface with the LALR parser generator CUP [8] and therefore uses a class sym (generated by CUP), where integer constants for the terminal tokens of the CUP grammar are declared. JFlex comes with a directory examples, where you can find a small standalone scanner that doesn't need other tools like CUP to give you a running example. The "examples" directory also contains a complete JFlex specification of the lexical structure of Java programs together with the CUP parser specification for Java by C. Scott Ananian, obtained from the CUP [8] website (it was modified to interface with the JFlex scanner). Both specifications adhere to the Java Language Specification [7].

/* JFlex example: part of Java language lexer specification */
import java_cup.runtime.*;

 * This class is a simple example lexer.
%class Lexer
  StringBuffer string = new StringBuffer();

  private Symbol symbol(int type) {
    return new Symbol(type, yyline, yycolumn);
  private Symbol symbol(int type, Object value) {
    return new Symbol(type, yyline, yycolumn, value);
LineTerminator = \r|\n|\r\n
InputCharacter = [^\r\n]
WhiteSpace     = {LineTerminator} | [ \t\f]

/* comments */
Comment = {TraditionalComment} | {EndOfLineComment} | {DocumentationComment}

TraditionalComment   = "/*" [^*] ~"*/" | "/*" "*"+ "/"
EndOfLineComment     = "//" {InputCharacter}* {LineTerminator}
DocumentationComment = "/**" {CommentContent} "*"+ "/"
CommentContent       = ( [^*] | \*+ [^/*] )*

Identifier = [:jletter:] [:jletterdigit:]*

DecIntegerLiteral = 0 | [1-9][0-9]*
%state STRING

/* keywords */
<YYINITIAL> "abstract"           { return symbol(sym.ABSTRACT); }
<YYINITIAL> "boolean"            { return symbol(sym.BOOLEAN); }
<YYINITIAL> "break"              { return symbol(sym.BREAK); }
  /* identifiers */ 
  {Identifier}                   { return symbol(sym.IDENTIFIER); }
  /* literals */
  {DecIntegerLiteral}            { return symbol(sym.INTEGER_LITERAL); }
  \"                             { string.setLength(0); yybegin(STRING); }

  /* operators */
  "="                            { return symbol(sym.EQ); }
  "=="                           { return symbol(sym.EQEQ); }
  "+"                            { return symbol(sym.PLUS); }

  /* comments */
  {Comment}                      { /* ignore */ }
  /* whitespace */
  {WhiteSpace}                   { /* ignore */ }
  \"                             { yybegin(YYINITIAL); 
                                   return symbol(sym.STRING_LITERAL, 
                                   string.toString()); }
  [^\n\r\"\\]+                   { string.append( yytext() ); }
  \\t                            { string.append('\t'); }
  \\n                            { string.append('\n'); }

  \\r                            { string.append('\r'); }
  \\\"                           { string.append('\"'); }
  \\                             { string.append('\\'); }
/* error fallback */
.|\n                             { throw new Error("Illegal character <"+
                                                    yytext()+">"); }

From this specification JFlex generates a .java file with one class that contains code for the scanner. The class will have a constructor taking a from which the input is read. The class will also have a function yylex() that runs the scanner and that can be used to get the next token from the input (in this example the function actually has the name next_token() because the specification uses the %cup switch).

As with JLex, the specification consists of three parts, divided by %%:

Code to include

Let's take a look at the first section, ``user code'': The text up to the first line starting with %% is copied verbatim to the top of the generated lexer class (before the actual class declaration). Beside package and import statements there is usually not much to do here. If the code ends with a javadoc class comment, the generated class will get this comment, if not, JFlex will generate one automatically.

Options and Macros

The second section ``options and declarations'' is more interesting. It consists of a set of options, code that is included inside the generated scanner class, lexical states and macro declarations. Each JFlex option must begin a line of the specification and starts with a %. In our example the following options are used:

The code included in %{...%} is copied verbatim into the generated lexer class source. Here you can declare member variables and functions that are used inside scanner actions. In our example we declare a StringBuffer ``string'' in which we will store parts of string literals and two helper functions ``symbol'' that create java_cup.runtime.Symbol objects with position information of the current token (see section 8.1 JFlex and CUP for how to interface with the parser generator CUP). As JFlex options, both %{ and \%} must begin a line.

The specification continues with macro declarations. Macros are abbreviations for regular expressions, used to make lexical specifications easier to read and understand. A macro declaration consists of a macro identifier followed by =, then followed by the regular expression it represents. This regular expression may itself contain macro usages. Although this allows a grammar like specification style, macros are still just abbreviations and not non terminals - they cannot be recursive or mutually recursive. Cycles in macro definitions are detected and reported at generation time by JFlex.

Here some of the example macros in more detail:

The last part of the second section in our lexical specification is a lexical state declaration: %state STRING declares a lexical state STRING that can be used in the ``lexical rules'' part of the specification. A state declaration is a line starting with %state followed by a space or comma separated list of state identifiers. There can be more than one line starting with %state.

Rules and Actions

The "lexical rules" section of a JFlex specification contains regular expressions and actions (Java code) that are executed when the scanner matches the associated regular expression. As the scanner reads its input, it keeps track of all regular expressions and activates the action of the expression that has the longest match. Our specification above for instance would with input "breaker" match the regular expression for Identifier and not the keyword "break" followed by the Identifier "er", because rule {Identifier} matches more of this input at once (i.e. it matches all of it) than any other rule in the specification. If two regular expressions both have the longest match for a certain input, the scanner chooses the action of the expression that appears first in the specification. In that way, we get for input "break" the keyword "break" and not an Identifier "break".

Additional to regular expression matches, one can use lexical states to refine a specification. A lexical state acts like a start condition. If the scanner is in lexical state STRING, only expressions that are preceded by the start condition <STRING> can be matched. A start condition of a regular expression can contain more than one lexical state. It is then matched when the lexer is in any of these lexical states. The lexical state YYINITIAL is predefined and is also the state in which the lexer begins scanning. If a regular expression has no start conditions it is matched in all lexical states.

Since you often have a bunch of expressions with the same start conditions, JFlex allows the same abbreviation as the Unix tool flex:

  expr1   { action1 }
  expr2   { action2 }
means that both expr1 and expr2 have start condition <STRING>.

The first three rules in our example demonstrate the syntax of a regular expression preceded by the start condition <YYINITIAL>.

<YYINITIAL> "abstract" { return symbol(sym.ABSTRACT); }

matches the input "abstract" only if the scanner is in its start state "YYINITIAL". When the string "abstract" is matched, the scanner function returns the CUP symbol sym.ABSTRACT. If an action does not return a value, the scanning process is resumed immediately after executing the action.

The rules enclosed in


demonstrate the abbreviated syntax and are also only matched in state YYINITIAL.

Of these rules, one may be of special interest:

\" { string.setLength(0); yybegin(STRING); }

If the scanner matches a double quote in state YYINITIAL we have recognized the start of a string literal. Therefore we clear our StringBuffer that will hold the content of this string literal and tell the scanner with yybegin(STRING) to switch into the lexical state STRING. Because we do not yet return a value to the parser, our scanner proceeds immediately.

In lexical state STRING another rule demonstrates how to refer to the input that has been matched:

[^\n\r\"]+ { string.append( yytext() ); }

The expression [^\n\r\"]+ matches all characters in the input up to the next backslash (indicating an escape sequence such as \n), double quote (indicating the end of the string), or line terminator (which must not occur in a string literal). The matched region of the input is referred to with yytext() and appended to the content of the string literal parsed so far.

The last lexical rule in the example specification is used as an error fallback. It matches any character in any state that has not been matched by another rule. It doesn't conflict with any other rule because it has the least priority (because it's the last rule) and because it matches only one character (so it can't have longest match precedence over any other rule).

How to get it going

Lexical Specifications

As shown above, a lexical specification file for JFlex consists of three parts divided by a single line starting with %%:

Options and declarations
Lexical rules

In all parts of the specification comments of the form /* comment text */ and the Java style end of line comments starting with // are permitted. JFlex comments do nest - so the number of /* and */ should be balanced.

User code

The first part contains user code that is copied verbatim into the beginning of the source file of the generated lexer before the scanner class is declared. As shown in the example above, this is the place to put package declarations and import statements. It is possible, but not considered as good Java programming style to put own helper class (such as token classes) in this section. They should get their own .java file instead.

Options and declarations

The second part of the lexical specification contains options to customize your generated lexer (JFlex directives and Java code to include in different parts of the lexer), declarations of lexical states and macro definitions for use in the third section ``Lexical rules'' of the lexical specification file.

Each JFlex directive must be situated at the beginning of a line and starts with the % character. Directives that have one or more parameters are described as follows:

%class "classname"

means that you start a line with %class followed by a space followed by the name of the class for the generated scanner (the double quotes are not to be entered, see the example specification in section 3).

Class options and user class code

These options regard name, constructor, API, and related parts of the generated scanner class.

Scanning method

This section shows how the scanning method can be customized. You can redefine the name and return type of the method and it is possible to declare exceptions that may be thrown in one of the actions of the specification. If no return type is specified, the scanning method will be declared as returning values of class Yytoken.

The end of file

There is always a default value that the scanning method will return when the end of file has been reached. You may however define a specific value to return and a specific piece of code that should be executed when the end of file is reached.

The default end of file values depends on the return type of the scanning method:

User values and code to be executed at the end of file can be defined using these directives:

Standalone scanners

CUP compatibility

You may also want to read section 8.1 JFlex and CUP if you are interested in how to interface your generated scanner with CUP.

BYacc/J compatibility

You may also want to read section 8.2 JFlex and BYacc/J if you are interested in how to interface your generated scanner with Byacc/J.

Code generation

The following options define what kind of lexical analyzer code JFlex will produce. %pack is the default setting and will be used, when no code generation method is specified.

Character sets

Line, character and column counting

Obsolete JLex options

State declarations

State declarations have the following from:

%s[tate] "state identifier" [, "state identifier", ... ] for inclusive or
%x[state] "state identifier" [, "state identifier", ... ] for exlusive states

There may be more than one line of state declarations, each starting with %state or %xstate (the first character is sufficient, %s and %x works, too). State identifiers are letters followed by a sequence of letters, digits or underscores. State identifiers can be separated by whitespace or comma.

The sequence

%state STATE1
%xstate STATE3, XYZ, STATE_10
%state ABC STATE5

declares the set of identifiers STATE1, STATE3, XYZ, STATE_10, ABC, STATE5 as lexical states, STATE1, ABC, STATE5 as inclusive, and STATE3, XYZ, STATE_10 as exclusive. See also section 4.3.3 on the way lexical states influence how the input is matched.

Macro definitions

A macro definition has the form

macroidentifier = regular expression

That means, a macro definition is a macro identifier (letter followed by a sequence of letters, digits or underscores), that can later be used to reference the macro, followed by optional whitespace, followed by an "=", followed by optional whitespace, followed by a regular expression (see section 4.3 lexical rules for more information about regular expressions).

The regular expression on the right hand side must be well formed and must not contain the ^, / or $ operators. Differently to JLex, macros are not just pieces of text that are expanded by copying - they are parsed and must be well formed.

This is a feature. It eliminates some very hard to find bugs in lexical specifications (such like not having parentheses around more complicated macros - which is not necessary with JFlex). See section 7.1 Porting from JLex for more details on the problems of JLex style macros.

Since it is allowed to have macro usages in macro definitions, it is possible to use a grammar like notation to specify the desired lexical structure. Macros however remain just abbreviations of the regular expressions they represent. They are not non terminals of a grammar and cannot be used recursively in any way. JFlex detects cycles in macro definitions and reports them at generation time. JFlex also warns you about macros that have been defined but never used in the ``lexical rules'' section of the specification.

Lexical rules

The ``lexical rules'' section of an JFlex specification contains a set of regular expressions and actions (Java code) that are executed when the scanner matches the associated regular expression.


The syntax of the "lexical rules" section is described by the following BNF grammar (terminal symbols are enclosed in 'quotes'):

LexicalRules ::= Rule+ 
Rule         ::= [StateList] ['^'] RegExp [LookAhead] Action 
               | [StateList] '<<EOF>>' Action
               | StateGroup 
StateGroup   ::= StateList '{' Rule+ '}' 
StateList    ::= '<' Identifier (',' Identifier)* '>' 
LookAhead    ::= '$' | '/' RegExp
Action       ::= '{' JavaCode '}' | '|'

RegExp       ::= RegExp '|' RegExp 
               | RegExp RegExp 
               | '(' RegExp ')'
               | ('!'|'~') RegExp
               | RegExp ('*'|'+'|'?')
               | RegExp "{" Number ["," Number] "}" 
               | '[' ['^'] (Character|Character'-'Character)* ']' 
               | PredefinedClass 
               | '{' Identifier '}' 
               | '"' StringCharacter+ '"' 
               | Character 

PredefinedClass ::= '[:jletter:]' 
                  | '[:jletterdigit:]' 
                  | '[:letter:]' 
                  | '[:digit:]' 
                  | '[:uppercase:]' 
                  | '[:lowercase:]' 
                  | '.'

The grammar uses the following terminal symbols:

Please note that the \n escape sequence stands for the ASCII LF character - not for the end of line. If you would like to match the line terminator, you should use the expression \r|\n|\r\n if you want the Java conventions, or \r|\n|\r\n|\u2028|\u2029|\u000B|\u000C|\u0085 if you want to be fully Unicode compliant (see also [5]).

As of version 1.1 of JFlex the whitespace characters " " (space) and "\t" (tab) can be used to improve the readability of regular expressions. They will be ignored by JFlex. In character classes and strings however, whitespace characters keep standing for themselves (so the string " " still matches exactly one space character and [ \n] still matches an ASCII LF or a space character).

JFlex applies the following standard operator precedences in regular expression (from highest to lowest):

So the expression a | abc | !cd* for instance is parsed as (a|(abc)) | ((!c)(d*)).


This section gives an informal description of which text is matched by a regular expression (i.e. an expression described by the RegExp production of the grammar presented above).

A regular expression that consists solely of

If a and b are regular expressions, then

a | b

is the regular expression, that matches all input that is matched by a or by b.

a b

is the regular expression, that matches the input matched by a followed by the input matched by b.

(kleene closure)

matches zero or more repetitions of the input matched by a


is equivalent to aa*


matches the empty input or the input matched by a


matches everything but the strings matched by a. Use with care: the construction of !a involves an additional, possibly exponential NFA to DFA transformation on the NFA for a. Note that with negation and union you also have (by applying DeMorgan) intersection and set difference: the intersection of a and b is !(!a|!b), the expression that matches everything of a not matched by b is !(!a|b)


matches everything up to (and including) the first occurrence of a text matched by a. The expression ~a is equivalent to !([^]* a [^]*) a. A traditional C-style comment is matched by "/*" ~"*/"


is equivalent to n times the concatenation of a. So a{4} for instance is equivalent to the expression a a a a. The decimal integer n must be positive.

is equivalent to at least n times and at most m times the concatenation of a. So a{2,4} for instance is equivalent to the expression a a a? a?. Both n and m are non negative decimal integers and m must not be smaller than n.

( a )
matches the same input as a.

In a lexical rule, a regular expression r may be preceded by a '^' (the beginning of line operator). r is then only matched at the beginning of a line in the input. A line begins after each occurrence of \r|\n|\r\n|\u2028|\u2029|\u000B|\u000C|\u0085 (see also [5]) and at the beginning of input. The preceding line terminator in the input is not consumed and can be matched by another rule.

In a lexical rule, a regular expression r may be followed by a lookahead expression. A lookahead expression is either a '$' (the end of line operator) or a '/' followed by an arbitrary regular expression. In both cases the lookahead is not consumed and not included in the matched text region, but it is considered while determining which rule has the longest match (see also 4.3.3 How the input is matched).

In the '$' case r is only matched at the end of a line in the input. The end of a line is denoted by the regular expression \r|\n|\r\n|\u2028|\u2029|\u000B|\u000C|\u0085. So a$ is equivalent to a / \r|\n|\r\n|\u2028|\u2029|\u000B|\u000C|\u0085.This is a bit different to the situation described in [5]: since in JFlex $ is a true trailing context, the end of file does not count as end of line.

For arbitrary lookahead (also called trailing context) the expression is matched only when followed by input that matches the trailing context. Unfortunately the lookahead expression is not really arbitrary: In a rule r1 / r2, either the text matched by r1 must have a fixed length (e.g. if r1 is a string) or the beginning of the trailing context r2 must not match the end of r1. So for example "abc" / "a"|"b" is ok because "abc" has a fixed length, "a"|"ab" / "x"* is ok because no prefix of "x"* matches a postfix of "a"|"ab", but "x"|"xy" / "yx" is not possible, because the postfix "y" of "x"|"xy" is also a prefix of "yx". In this case JFlex will still correctly use r1 r2 (r1 followed by r2) to determine if the rule should be matched, but it might return too many characters in yytext (it will return the longest match of r1 within r1 r2). JFlex attempts to report such cases at generation time, but it might be overeager: it also warns in cases where the lookahead is safe. The algorithm JFlex currently uses for matching trailing context expressions is the one described in [1] (leading to the deficiencies mentioned above).

As of version 1.2, JFlex allows lex/flex style «EOF» rules in lexical specifications. A rule

[StateList]  <<EOF>>    { some action code }
is very similar to the %eofval directive (section 4.2.3). The difference lies in the optional StateList that may precede the «EOF» rule. The action code will only be executed when the end of file is read and the scanner is currently in one of the lexical states listed in StateList. The same StateGroup (see section 4.3.3 How the input is matched) and precedence rules as in the ``normal'' rule case apply (i.e. if there is more than one «EOF» rule for a certain lexical state, the action of the one appearing earlier in the specification will be executed). «EOF» rules override settings of the %cup and %byaccj options and should not be mixed with the %eofval directive.

An Action consists either of a piece of Java code enclosed in curly braces or is the special | action. The | action is an abbreviation for the action of the following expression.


expression1   |
expression2   |
expression3   { some action }
is equivalent to the expanded form
expression1   { some action }
expression2   { some action }
expression3   { some action }

They are useful when you work with trailing context expressions. The expression a | (c / d) | b is not syntactically legal, but can easily be expressed using the | action:

a       |
c / d   |
b       { some action }

How the input is matched

When consuming its input, the scanner determines the regular expression that matches the longest portion of the input (longest match rule). If there is more than one regular expression that matches the longest portion of input (i.e. they all match the same input), the generated scanner chooses the expression that appears first in the specification. After determining the active regular expression, the associated action is executed. If there is no matching regular expression, the scanner terminates the program with an error message (if the %standalone directive has been used, the scanner prints the unmatched input to java.lang.System.out instead and resumes scanning).

Lexical states can be used to further restrict the set of regular expressions that match the current input.

The generated class

JFlex generates exactly one file containing one class from the specification (unless you have declared another class in the first specification section).

The generated class contains (among other things) the DFA tables, an input buffer, the lexical states of the specification, a constructor, and the scanning method with the user supplied actions.

The name of the class is by default Yylex, it is customizable with the %class directive (see also section 4.2.1). The input buffer of the lexer is connected with an input stream over the object which is passed to the lexer in the generated constructor. If you want to provide your own constructor for the lexer, you should always call the generated one in it to initialize the input buffer. The input buffer should not be accessed directly, but only over the advertised API (see also section 4.3.5). Its internal implementation may change between releases or skeleton files without notice.

The main interface to the outside world is the generated scanning method (default name yylex, default return type Yytoken). Most of its aspects are customizable (name, return type, declared exceptions etc., see also section 4.2.2). If it is called, it will consume input until one of the expressions in the specification is matched or an error occurs. If an expression is matched, the corresponding action is executed. It may return a value of the specified return type (in which case the scanning method return with this value), or if it doesn't return a value, the scanner resumes consuming input until the next expression is matched. If the end of file is reached, the scanner executes the EOF action, and (also upon each further call to the scanning method) returns the specified EOF value (see also section 4.2.3).

Scanner methods and fields accessible in actions (API)

Generated methods and member fields in JFlex scanners are prefixed with yy to indicate that they are generated and to avoid name conflicts with user code copied into the class. Since user code is part of the same class, JFlex has no language means like the private modifier to indicate which members and methods are internal and which ones belong to the API. Instead, JFlex follows a naming convention: everything starting with a zz prefix like zzStartRead is to be considered internal and subject to change without notice between JFlex releases. Methods and members of the generated class that do not have a zz prefix like yycharat belong to the API that the scanner class provides to users in action code of the specification. They will be remain stable and supported between JFlex releases as long as possible.

Currently, the API consists of the following methods and member fields:

Encodings, Platforms, and Unicode

This section tries to shed some light on the issues of Unicode and encodings, cross platform scanning, and how to deal with binary data. My thanks go to Stephen Ostermiller for his input on this topic.

The Problem

Before we dive straight into details, let's take a look at what the problem is. The problem is Java's platform independence when you want to use it. For scanners the interesting part about platform independence is character encodings and how they are handled.

If a program reads a file from disk, it gets a stream of bytes. In earlier times, when the grass was green, and the world was much simpler, everybody knew that the byte value 65 is, of course, an A. It was no problem to see which bytes meant which characters (actually these times never existed, but anyway). The normal Latin alphabet only has 26 characters, so 7 bits or 128 distinct values should surely be enough to map them, even if you allow yourself the luxury of upper and lower case. Nowadays, things are different. The world suddenly grew much larger, and all kinds of people wanted all kinds of special characters, just because they use them in their language and writing. This is were the mess starts. Since the 128 distinct values were already filled up with other stuff, people began to use all 8 bits of the byte, and extended the byte/character mappings to fit their need, and of course everybody did it differently. Some people for instance may have said ``let's use the value 213 for the German character ä''. Others may have found that 213 should much rather mean é, because they didn't need German and wrote French instead. As long as you use your program and data files only on one platform, this is no problem, as all know what means what, and everything gets used consistently.

Now Java comes into play, and wants to run everywhere (once written, that is) and now there suddenly is a problem: how do I get the same program to say ä to a certain byte when it runs in Germany and maybe é when it runs in France? And also the other way around: when I want to say é on the screen, which byte value should I send to the operating system?

Java's solution to this is to use Unicode internally. Unicode aims to be a superset of all known character sets and is therefore a perfect base for encoding things that might get used all over the world. To make things work correctly, you still have to know where you are and how to map byte values to Unicode characters and vice versa, but the important thing is, that this mapping is at least possible (you can map Kanji characters to Unicode, but you cannot map them to ASCII or iso-latin-1).

Scanning text files

Scanning text files is the standard application for scanners like JFlex. Therefore it should also be the most convenient one. Most times it is.

The following scenario works like a breeze: You work on a platform X, write your lexer specification there, can use any obscure Unicode character in it as you like, and compile the program. Your users work on any platform Y (possibly but not necessarily something different from X), they write their input files on Y and they run your program on Y. No problems.

Java does this as follows: If you want to read anything in Java that is supposed to contain text, you use a FileReader or some InputStream together with an InputStreamReader. InputStreams return the raw bytes, the InputStreamReader converts the bytes into Unicode characters with the platform's default encoding. If a text file is produced on the same platform, the platform's default encoding should do the mapping correctly. Since JFlex also uses readers and Unicode internally, this mechanism also works for the scanner specifications. If you write an A in your text editor and the editor uses the platform's encoding (say A is 65), then Java translates this into the logical Unicode A internally. If a user writes an A on a completely different platform (say A is 237 there), then Java also translates this into the logical Unicode A internally. Scanning is performed after that translation and both match.

Note that because of this mapping from bytes to characters, you should always use the %unicode switch in you lexer specification if you want to scan text files. %8bit may not be enough, even if you know that your platform only uses one byte per character. The encoding Cp1252 used on many Windows machines for instance knows 256 characters, but the character ´ with Cp1252 code \x92 has the Unicode value \u2019, which is larger than 255 and which would make your scanner throw an ArrayIndexOutOfBoundsException if it is encountered.

So for the usual case you don't have to do anything but use the %unicode switch in your lexer specification.

Things may break when you produce a text file on platform X and consume it on a different platform Y. Let's say you have a file written on a Windows PC using the encoding Cp1252. Then you move this file to a Linux PC with encoding ISO 8859-1 and there you want to run your scanner on it. Java now thinks the file is encoded in ISO 8859-1 (the platform's default encoding) while it really is encoded in Cp1252. For most characters Cp1252 and ISO 8859-1 are the same, but for the byte values \x80 to \x9f they disagree: ISO 8859-1 is undefined there. You can fix the problem by telling Java explicitly which encoding to use. When constructing the InputStreamReader, you can give the encoding as argument. The line

Reader r = new InputStreamReader(input, "Cp1252");
will do the trick.

Of course the encoding to use can also come from the data itself: for instance, when you scan a HTML page, it may have embedded information about its character encoding in the headers.

More information about encodings, which ones are supported, how they are called, and how to set them may be found in the official Java documentation in the chapter about internationalization. The link leads to an online version of this for Sun's JDK 1.3.

Scanning binaries

Scanning binaries is both easier and more difficult than scanning text files. It's easier because you want the raw bytes and not their meaning, i.e. you don't want any translation. It's more difficult because it's not so easy to get ``no translation'' when you use Java readers.

The problem (for binaries) is that JFlex scanners are designed to work on text. Therefore the interface is the Reader class (there is a constructor for InputStream instances, but it's just there for convenience and wraps an InputStreamReader around it to get characters, not bytes). You can still get a binary scanner when you write your own custom InputStreamReader class that does explicitly no translation, but just copies byte values to character codes instead. It sounds quite easy, and actually it is no big deal, but there are a few little pitfalls on the way. In the scanner specification you can only enter positive character codes (for bytes that is \x00 to \xFF). Java's byte type on the other hand is a signed 8 bit integer (-128 to 127), so you have to convert them properly in your custom Reader. Also, you should take care when you write your lexer spec: if you use text in there, it gets interpreted by an encoding first, and what scanner you get as result might depend on which platform you run JFlex on when you generate the scanner (this is what you want for text, but for binaries it gets in the way). If you are not sure, or if the development platform might change, it's probably best to use character code escapes in all places, since they don't change their meaning.

To illustrate these points, the example in examples/binary contains a very small binary scanner that tries to detect if a file is a Java class file. For that purpose it looks if the file begins with the magic number \xCAFEBABE.

A few words on performance

This section gives some empirical results about the speed of JFlex generated scanners in comparison to those generated by JLex, compares a JFlex scanner with a handwritten one, and presents some tips on how to make your specification produce a faster scanner.

Comparison of JLex and JFlex

Scanners generated by the tool JLex are quite fast. It was however possible to further improve the performance of generated scanners using JFlex. The following table shows the results that were produced by the scanner specification of a small toy programming language (the example from the JLex website). The scanner was generated using JLex 1.2.6 and JFlex version 1.3.5 with all three different JFlex code generation methods. Then it was run on a W98 system using Sun's JDK 1.3 with different sample inputs of that toy programming language. All test runs were made under the same conditions on an otherwise idle machine.

The values presented in the table denote the time from the first call to the scanning method to returning the EOF value and the speedup in percent. The tests were run both int the mixed (HotSpot) JVM mode and the pure interpreted mode. The mixed mode JVM brings about a factor of 10 performance improvement, the difference between JLex and JFlex only decreases slightly.

 KB JVM JLex %switch speedup %table speedup %pack speedup
 496 hotspot 325 ms 261 ms 24.5 % 261 ms 24.5 % 261 ms 24.5 %
 187 hotspot 127 ms 98 ms 29.6 % 94 ms 35.1 % 96 ms 32.3 %
 93 hotspot 66 ms 50 ms 32.0 % 50 ms 32.0 % 48 ms 37.5 %
 496 interpr. 4009 ms 3025 ms 32.5 % 3258 ms 23.1 % 3231 ms 24.1 %
 187 interpr. 1641 ms 1155 ms 42.1 % 1245 ms 31.8 % 1234 ms 33.0 %
 93 interpr. 817 ms 573 ms 42.6 % 617 ms 32.4 % 613 ms 33.3 %

Since the scanning time of the lexical analyzer examined in the table above includes lexical actions that often need to create new object instances, another table shows the execution time for the same specification with empty lexical actions to compare the pure scanning engines.

 KB JVM JLex %switch speedup %table speedup %pack speedup
 496 hotspot 204 ms 140 ms 45.7 % 138 ms 47.8 % 140 ms 45.7 %
 187 hotspot 83 ms 55 ms 50.9 % 52 ms 59.6 % 52 ms 59.6 %
 93 hotspot 41 ms 28 ms 46.4 % 26 ms 57.7 % 26 ms 57.7 %
 496 interpr. 2983 ms 2036 ms 46.5 % 2230 ms 33.8 % 2232 ms 33.6 %
 187 interpr. 1260 ms 793 ms 58.9 % 865 ms 45.7 % 867 ms 45.3 %
 93 interpr. 628 ms 395 ms 59.0 % 432 ms 45.4 % 432 ms 45.4 %

Execution time of single instructions depends on the platform and the implementation of the Java Virtual Machine the program is executed on. Therefore the tables above cannot be used as a reference to which code generation method of JFlex is the right one to choose in general. The following table was produced by the same lexical specification and the same input on a Linux system also using Sun's JDK 1.3.

With actions:

 KB JVM JLex %switch speedup %table speedup %pack speedup
 496 hotspot 246 ms 203 ms 21.2 % 193 ms 27.5 % 190 ms 29.5 %
 187 hotspot 99 ms 76 ms 30.3 % 69 ms 43.5 % 70 ms 41.4 %
 93 hotspot 48 ms 36 ms 33.3 % 34 ms 41.2 % 35 ms 37.1 %
 496 interpr. 3251 ms 2247 ms 44.7 % 2430 ms 33.8 % 2444 ms 33.0 %
 187 interpr. 1320 ms 848 ms 55.7 % 958 ms 37.8 % 920 ms 43.5 %
 93 interpr. 658 ms 423 ms 55.6 % 456 ms 44.3 % 452 ms 45.6 %

Without actions:

 KB JVM JLex %switch speedup %table speedup %pack speedup
 496 hotspot 136 ms 78 ms 74.4 % 76 ms 78.9 % 77 ms 76.6 %
 187 hotspot 59 ms 31 ms 90.3 % 48 ms 22.9 % 32 ms 84.4 %
 93 hotspot 28 ms 15 ms 86.7 % 15 ms 86.7 % 15 ms 86.7 %
 496 interpr. 1992 ms 1047 ms 90.3 % 1246 ms 59.9 % 1215 ms 64.0 %
 187 interpr. 859 ms 408 ms 110.5 % 479 ms 79.3 % 487 ms 76.4 %
 93 interpr. 435 ms 200 ms 117.5 % 237 ms 83.5 % 242 ms 79.8 %

Although all JFlex scanners were faster than those generated by JLex, slight differences between JFlex code generation methods show up when compared to the run on the W98 system.

The following table compares a handwritten scanner for the Java language obtained from the website of CUP with the JFlex generated scanner for Java that comes with JFlex in the examples directory. They were tested on different .java files on a Linux machine with Sun's JDK 1.3.

 lines KB JVM handwritten scanner JFlex generated scanner
 19050 496 hotspot 824 ms 248 ms 235 % faster
 6350 165 hotspot 272 ms 84 ms 232 % faster
 1270 33 hotspot 53 ms 18 ms 194 % faster
 19050 496 interpreted 5.83 s 3.85 s 51 % faster
 6350 165 interpreted 1.95 s 1.29 s 51 % faster
 1270 33 interpreted 0.38 s 0.25 s 52 % faster

Although JDK 1.3 seems to speed up the handwritten scanner if compared to JDK 1.1 or 1.2 more than the generated one, the generated scanner is still up to 3.3 times as fast as the handwritten one. One example of a handwritten scanner that is considerably slower than the equivalent generated one is surely no proof for all generated scanners being faster than handwritten. It is clearly impossible to prove something like that, since you could always write the generated scanner by hand. From a software engineering point of view however, there is no excuse for writing a scanner by hand since this task takes more time, is more difficult and therefore more error prone than writing a compact, readable and easy to change lexical specification. (I'd like to add, that I do not think, that the handwritten scanner from the CUP website used here in the test is stupid or badly written or anything like that. I actually think, Scott did a great job with it, and that for learning about lexers it is quite valuable to study it or even to write a similar one for oneself.)

How to write a faster specification

Although JFlex generated scanners show good performance without special optimizations, there are some heuristics that can make a lexical specification produce an even faster scanner. Those are (roughly in order of performance gain):

Note that writing more rules in a specification does not make the generated scanner slower (except when you have to switch to another code generation method because of the larger size).

The two main rules of optimization apply also for lexical specifications:

  1. don't do it
  2. (for experts only) don't do it yet

Some of the performance tips above contradict a readable and compact specification style. When in doubt or when requirements are not or not yet fixed: don't use them - the specification can always be optimized in a later state of the development process.

Porting Issues

Porting from JLex

JFlex was designed to read old JLex specifications unchanged and to generate a scanner which behaves exactly the same as the one generated by JLex with the only difference of being faster.

This works as expected on all well formed JLex specifications.

Since the statement above is somewhat absolute, let's take a look at what ``well formed'' means here. A JLex specification is well formed, when it

Porting from lex/flex

This section tries to give an overview of activities and possible problems when porting a lexical specification from the C/C++ tools lex and flex [11] available on most Unix systems to JFlex.

Most of the C/C++ specific features are naturally not present in JFlex, but most ``clean'' lex/flex lexical specifications can be ported to JFlex without very much work.

This section is by far not complete and is based mainly on a survey of the flex man page and very little personal experience. If you do engage in any porting activity from lex/flex to JFlex and encounter problems, have better solutions for points presented here or have just some tips you would like to share, please do contact me. I will incorporate your experiences in this manual (with all due credit to you, of course).

Basic structure

A lexical specification for flex has the following basic structure:
user code

The user code section usually contains some C code that is used in actions of the rules part of the specification. For JFlex most of this code will have to be included in the class code %{..%} directive in the options and declarations section (after translating the C code to Java, of course).

Macros and Regular Expression Syntax

The definitions section of a flex specification is quite similar to the options and declarations part of JFlex specs.

Macro definitions in flex have the form:

<identifier>  <expression>
To port them to JFlex macros, just insert a = between <identifier> and <expression>.

The syntax and semantics of regular expressions in flex are pretty much the same as in JFlex. A little attention is needed for some escape sequences present in flex (such as \a) that are not supported in JFlex. These escape sequences should be transformed into their octal or hexadecimal equivalent.

Another point are predefined character classes. Flex offers the ones directly supported by C, JFlex offers the ones supported by Java. These classes will sometimes have to be listed manually (if there is need for this feature, it may be implemented in a future JFlex version).

Lexical Rules

Since flex is mostly Unix based, the '^' (beginning of line) and '$' (end of line) operators, consider the \n character as only line terminator. This should usually cause not much problems, but you should be prepared for occurrences of \r or \r\n or one of the characters \u2028, \u2029, \u000B, \u000C, or \u0085. They are considered to be line terminators in Unicode and therefore may not be consumed when ^ or $ is present in a rule.

The trailing context algorithm of flex is better than the one used in JFlex. Therefore lookahead expressions could cause major headaches. JFlex will issue an error message at generation time, if it cannot generate a scanner for a certain lookahead expression. (sorry, I have no more tips here on that yet. If anyone knows how the flex lookahead algorithm works (or any better one) and can be efficiently implemented, again: please contact me).

Working together

JFlex and CUP

One of the main design goals of JFlex was to make interfacing with the free Java parser generator CUP [8] as easy as possibly. This has been done by giving the %cup directive a special meaning. An interface however always has two sides. This section concentrates on the CUP side of the story.

CUP version 0.10j

Since CUP version 0.10j, this has been simplified greatly by the new CUP scanner interface java_cup.runtime.Scanner. JFlex lexers now implement this interface automatically when then %cup switch is used. There are no special parser code, init code or scan with options any more that you have to provide in your CUP parser specification. You can just concentrate on your grammar.

If your generated Lexer has the class name Scanner, the parser is started from the a main program like this:

  try {
    parser p = new parser(new Scanner(new FileReader(fileName)));
    Object result = p.parse().value;
  catch (Exception e) {

Using existing JFlex/CUP specifications with CUP 0.10j

If you already have an existing specification and you would like to upgrade both JFlex and CUP to their newest version, you will probably have to adjust your specification.

The main difference between the %cup switch in JFlex 1.2.1 and lower, and the current JFlex version is, that JFlex scanners now automatically implement the java_cup.runtime.Scanner interface. This means, that the scanning function now changes its name from yylex() to next_token().

The main difference from older CUP versions to 0.10j is, that CUP now has a default constructor that accepts a java_cup.runtime.Scanner as argument and that uses this scanner as default (so no scan with code is necessary any more).

If you have an existing CUP specification, it will probably look somewhat like this:

parser code {:
  Lexer lexer;

  public parser ( input) {
    lexer = new Lexer(input);

scan with {: return lexer.yylex(); :};

To upgrade to CUP 0.10j, you could change it to look like this:

parser code {:
  public parser ( input) {
    super(new Lexer(input));

If you do not mind to change the method that is calling the parser, you could remove the constructor entirely (and if there is nothing else in it, the whole parser code section as well, of course). The calling main procedure would then construct the parser as shown in the section above.

The JFlex specification does not need to be changed.

Using older versions of CUP

For people, who like or have to use older versions of CUP, the following section explains ``the old way''. Please note, that the standard name of the scanning function with the %cup switch is not yylex(), but next_token().

If you have a scanner specification that begins like this:

package PACKAGE;
import java_cup.runtime.*;   /* this is convenience, but not necessary */
%class Lexer

then it matches a CUP specification starting like

package PACKAGE;

parser code {:
  Lexer lexer;

  public parser ( input) {
    lexer = new Lexer(input);

scan with {: return lexer.next_token(); :};


This assumes that the generated parser will get the name parser. If it doesn't, you have to adjust the constructor name.

The parser can then be started in a main routine like this:

  try {
    parser p = new parser(new FileReader(fileName));
    Object result = p.parse().value; 
  catch (Exception e) {

If you want the parser specification to be independent of the name of the generated scanner, you can instead write an interface Lexer

public interface Lexer {
  public java_cup.runtime.Symbol next_token() throws;

change the parser code to:

package PACKAGE;

parser code {:
  Lexer lexer;

  public parser (Lexer lexer) {
    this.lexer = lexer;

scan with {: return lexer.next_token(); :};


tell JFlex about the Lexer interface using the %implements directive:

%class Scanner     /* not Lexer now since that is our interface! */
%implements Lexer

and finally change the main routine to look like

  try {
    parser p = new parser(new Scanner(new FileReader(fileName)));
    Object result = p.parse().value;
  catch (Exception e) {

If you want to improve the error messages that CUP generated parsers produce, you can also override the methods report_error and report_fatal_error in the ``parser code'' section of the CUP specification. The new methods could for instance use yyline and yycolumn (stored in the left and right members of class java_cup.runtime.Symbol) to report error positions more conveniently for the user. The lexer and parser for the Java language in the examples/java directory of the JFlex distribution use this style of error reporting. These specifications also demonstrate the techniques above in action.

JFlex and BYacc/J

JFlex has builtin support for the Java extension BYacc/J [9] by Bob Jamison to the classical Berkeley Yacc parser generator. This section describes how to interface BYacc/J with JFlex. It builds on many helpful suggestions and comments from Larry Bell.

Since Yacc's architecture is a bit different from CUP's, the interface setup also works in a slightly different manner. BYacc/J expects a function int yylex() in the parser class that returns each next token. Semantic values are expected in a field yylval of type parserval where ``parser'' is the name of the generated parser class.

For a small calculator example, one could use a setup like the following on the JFlex side:



  /* store a reference to the parser object */
  private parser yyparser;

  /* constructor taking an additional parser object */
  public Yylex( r, parser yyparser) {
    this.yyparser = yyparser;

NUM = [0-9]+ ("." [0-9]+)?
NL  = \n | \r | \r\n


/* operators */
"+" | 
"(" | 
")"    { return (int) yycharat(0); }

/* newline */
{NL}   { return parser.NL; }

/* float */
{NUM}  { yyparser.yylval = new parserval(Double.parseDouble(yytext()));
         return parser.NUM; }

The lexer expects a reference to the parser in its constructor. Since Yacc allows direct use of terminal characters like '+' in its specifications, we just return the character code for single char matches (e.g. the operators in the example). Symbolic token names are stored as public static int constants in the generated parser class. They are used as in the NL token above. Finally, for some tokens, a semantic value may have to be communicated to the parser. The NUM rule demonstrates that bit.

A matching BYacc/J parser specification could look like this:

%token NL          /* newline  */
%token <dval> NUM  /* a number */

%type <dval> exp

%left '-' '+'
%right '^'         /* exponentiation */

exp:     NUM          { $$ = $1; }
       | exp '+' exp  { $$ = $1 + $3; }
       | exp '^' exp  { $$ = Math.pow($1, $3); }
       | '(' exp ')'  { $$ = $2; }

  /* a reference to the lexer object */
  private Yylex lexer;

  /* interface to the lexer */
  private int yylex () {
    int yyl_return = -1;
    try {
      yyl_return = lexer.yylex();
    catch (IOException e) {
      System.err.println("IO error :"+e);
    return yyl_return;

  /* error reporting */
  public void yyerror (String error) {
    System.err.println ("Error: " + error);

  /* lexer is created in the constructor */
  public parser(Reader r) {
    lexer = new Yylex(r, this);

  /* that's how you use the parser */
  public static void main(String args[]) throws IOException {
    parser yyparser = new parser(new FileReader(args[0]));

Here, the customized part is mostly in the user code section: We create the lexer in the constructor of the parser and store a reference to it for later use in the parser's int yylex() method. This yylex in the parser only calls int yylex() of the generated lexer and passes the result on. If something goes wrong, it returns -1 to indicate an error.

Runnable versions of the specifications above are located in the examples/byaccj directory of the JFlex distribution.

Bugs and Deficiencies


The trailing context algorithm described in [1] and used in JFlex is incorrect. It does not work, when a postfix of the regular expression matches a prefix of the trailing context and the length of the text matched by the expression does not have a fixed size. In this case JFlex will still correctly use r1 r2 (r1 followed by a lookahead r2) to determine if the rule should be matched, but it might return too many characters in yytext (it will return the longest match of r1 within r1 r2). JFlex attempts to report these cases as errors at generation time, but the warnings are overeager. A large number of safe lookaheads are reported as unsafe.


As of July 25, 2005 the following bugs are known in JFlex:

If you find new ones, please use the bugs section of the JFlex website to report them.

Copying and License

JFlex is free software, published under the terms of the GNU General Public License.

There is absolutely NO WARRANTY for JFlex, its code and its documentation.

The code generated by JFlex inherits the copyright of the specification it was produced from. If it was your specification, you may use the generated code without restriction.

See the file COPYRIGHT for more information.


A. Aho, R. Sethi, J. Ullman, Compilers: Principles, Techniques, and Tools, 1986

A. W. Appel, Modern Compiler Implementation in Java: basic techniques, 1997

E. Berk, JLex: A lexical analyser generator for Java,

K. Brouwer, W. Gellerich,E. Ploedereder, Myths and Facts about the Efficient Implementation of Finite Automata and Lexical Analysis, in: Proceedings of the 7th International Conference on Compiler Construction (CC '98), 1998

M. Davis, Unicode Regular Expression Guidelines, Unicode Technical Report #18, 2000

P. Dencker, K. Dürre, J. Henft, Optimization of Parser Tables for portable Compilers, in: ACM Transactions on Programming Languages and Systems 6(4), 1984

J. Gosling, B. Joy, G. Steele, The Java Language Specifcation, 1996,

S. E. Hudson, CUP LALR Parser Generator for Java,

B. Jamison, BYacc/J,

T. Lindholm, F. Yellin, The Java Virtual Machine Specification, 1996,

V. Paxson, flex - The fast lexical analyzer generator, 1995

R. E. Tarjan, A. Yao, Storing a Sparse Table, in: Communications of the ACM 22(11), 1979

R. Wilhelm, D. Maurer, Übersetzerbau, Berlin 19972


... Java[*]
Java is a trademark of Sun Microsystems, Inc., and refers to Sun's Java programming language. JFlex is not sponsored by or affiliated with Sun Microsystems, Inc.

Mon Jul 25 09:41:09 EST 2005, Gerwin Klein