Just as your code can benefit from complete unit tests, so can your grammars. Here's an example of developing a simple grammar for CSV using Java, ANTLR and TestNG.

CSV defined

CSV (Comma Separated Values) is a file format commonly used for exchanging spreadsheet data and which generally follows these rules:

A record consists of multiple fields seperated by commas. Each record ends with a line feed (a/k/a newline) or carriage return + line feed.
A field may not contain spaces, commas, double-quotes, line feeds, or carriage returns unless the field itself is wrapped in double-quotes on each end.
To insert a double quote into a quoted field, double it (i.e. use "")
Spaces are ignored immediately before and after a comma.
White space at the front or end of the record is not allowed unless part of a quoted field.

Prerequisites for this tutorial

You have ANTLR and TestNG installed. (You can use the same ideas with other unit testing frameworks such as JUnit)
You can run ANTLR to generate Java sources from a grammar.
You can build and run Java code.
(optional) You have JDK 5 or later installed. If not, you won't be able to use some of the constructs in this example and will need to translate back to 1.4 or earlier.

Develop a basic CSV parser

Basic setup

Every CSV line ends in a newline (line feed) or carriage return + line feed. Let's set that up as our basic grammar.

CSV.g

grammar CSV;

line :	NEWLINE;

NEWLINE	:	'\r'? '\n';

Create the corresponding test harness.

CSVTests.java

public class CSVTests {
    @Test
    public void testNewline() throws IOException, RecognitionException {
        CSVParser parser = createParser("\n");
        parser.line();
    }

    @Test
    public void testCRLF() throws IOException, RecognitionException {
        CSVParser parser = createParser("\r\n");
        parser.line();
    }

    private CSVParser createParser(String testString) throws IOException {
    	CharStream stream = new ANTLRStringStream(testString);
    	CSVLexer lexer = new CSVLexer(stream);
        CommonTokenStream tokens = new CommonTokenStream(lexer);
        CSVParser parser = new CSVParser(tokens);
        return parser;
    }

}

Generate the Java files for this grammar, then run the test.

The test should pass.

Return a result

We'll want to get a list of strings back from each line of CSV. In true TDD form, we'll alter the test then update the grammar to make the test correct again.

CSVTests.java fragment

@Test
    public void testNewline() throws IOException, RecognitionException {
        CSVParser parser = createParser("\n");
        List<String> result = parser.line(); // final public void line...
        assert result.isEmpty() : "Nothing to return";
    }

    @Test
    public void testCRLF() throws IOException, RecognitionException {
        CSVParser parser = createParser("\r\n");
        List<String> result = parser.line();
        assert result.isEmpty() : "Nothing to return";
    }

As expected, this fails to compile: line is declared as final public void line

Now update the grammar so this compiles.

CSV.g fragment

line returns [List<String> result]
	:	NEWLINE;

Regenerate the grammar. The test compiles, but you get a NullPointerException. Look at the generated code in CSVParser.java:

Generated code...

public final List<String> line() throws RecognitionException {
        List<String> result = null;

Oops! We need to initialize the result. Edit your grammar again:

CSV.g fragment

line returns [List<String> result]
@init {
    result = new ArrayList<String>();
}
	:	NEWLINE;

NEWLINE	:	'\r'? '\n';

Run ANTLR to regenerate your files and run the test again. The test should pass.

Extract a single word

A record in CSV is a series of fields separated by commas and ending in a newline or CRLF. We'll start by testing a single field in isolation then building back up to testing a whole record's worth of fields.

Start by adding a new test:

CSVTests.java

@Test
public void testSingleWord() throws IOException, RecognitionException {
    CSVParser parser = createParser("Red");
    String result = parser.field();
    assert result.equals("Red") : "Expected Red, found " + result;
}

Of course, this won't compile. We need to define field in the grammar, doing just enough to keep the tests working:

CSV.g

grammar CSV;

line returns [List<String> result]
@init {
	result = new ArrayList<String>();
}
	:	field NEWLINE;

field returns [String parsedItem]
	:	f=FIELD { $parsedItem = $f.text;}
	|   // nothing
	;

NEWLINE	:	'\r'? '\n';

FIELD	:	NONBREAKING* ;

// Anything except a line-breaking character is allowed.
fragement NONBREAKING
	:	~('\r' | '\n');

What's that funny character?

"~" means "not" and is used to match any item that's not in a set.

See The Definitive ANTLR Reference, page 95.

Support multiple fields

Let's try multiple simple fields:

CSVTests.java fragment

@Test
    public void testMultipleWords() throws IOException, RecognitionException {
        CSVParser parser = createParser("Red,Green,,Blue\n");
        List<String> result = parser.line();
        assert result.size() == 4: "Expected 4 items";
        assert result.get(0).equals("Red") : "Expected Red";
        assert result.get(1).equals("Green") : "Expected Green";
        assert result.get(2).equals("") : "Expected empty";
        assert result.get(3).equals("Blue") : "Expected Blue";
    }

This is going to take a few more changes than just defining a line as field,field,field...

We need to add each field's value to the line.
A record such as a,,b should write out an empty field in the middle.
Testing shows a potential nondeterminism between <empty field>NEWLINE and NEWLINE.

CSV.g

grammar CSV;

line returns [List<String> result]
@init {
	result = new ArrayList<String>();
}
	: (NEWLINE) => NEWLINE
	| (
		fieldResult=field { result.add(fieldResult); }
		( COMMA fieldResult=field {result.add(fieldResult);} )*
	 	NEWLINE
	  )
	;

field returns [String parsedItem]
@init {
	parsedItem = "";
}
	: f=FIELD {$parsedItem=$f.text;}
	| // nothing
	;

NEWLINE	:	'\r'? '\n';

COMMA	:	',';

FIELD:	NONBREAKING+;

// Anything except a line-breaking character is allowed.
fragment NONBREAKING
	:	~('\r' | '\n' | ',');

This works, but line is getting cluttered.

Introducing scopes

We've been passing a value back from field to line, but there's another way to pass information between rules: dynamic scopes. This is a good place to see how they work.

CSV.g with scope

grammar CSV;

/* Old definitions commented out:
line returns [List<String> result]
@init { result = new ArrayList<String>(); }
	: (NEWLINE) => NEWLINE
	| (
		fieldResult=field { result.add(fieldResult); }
		( COMMA fieldResult=field {result.add(fieldResult);} )*
	 	NEWLINE
	  )
	;

field returns [String parsedItem]
@init { parsedItem = ""; }
	: (f=FIELD {$parsedItem=$f.text;}
	  | // nothing
	  )
 	{ fields.add($parsedItem); }
	;

*/

// New definitions:
line returns [List<String> result]
scope { List fields; }
@init { $line::fields = new ArrayList(); }
	: (
	    (NEWLINE) => NEWLINE
	    | field (COMMA  field)* NEWLINE
	  )
	  { $result = $line::fields; }
	;

field
	: ( f=FIELD
	  | // nothing
	  )
 	{ $line::fields.add($f.text); }
	;

NEWLINE	:	'\r'? '\n';

COMMA	:	',';

FIELD:	NONBREAKING+;

// Anything except a line-breaking character is allowed.
fragment NONBREAKING
	:	~('\r' | '\n' | ',');

Since field no longer returns a string, we'll need to alter the test to pass the value through line and add a newline to the end of the line:

CSVTests.java, new field test via line

@Test
public void testSingleWord() throws IOException, RecognitionException {
    String result = parseField("Red");
    assert result.equals("Red") : "Expected Red, but found " + result;
}

private String parseField(String testString) throws IOException, RecognitionException {
    CSVParser parser = createParser(testString + "\n");
    List<String> result = parser.line();
    return result.get(0);
}

Quoting, part 1

CSV requires that fields that contain special characters (newline, return, double-quote, comma, space) be surrounded by double quotes.

CSVTests.java fragment

@Test
public void testQuotedString() throws IOException, RecognitionException {
    String result = parseField("\"Red, White, and Blue\"");
    assert result.equals("Red, White, and Blue") : "Expected <<Red, White, and Blue>>, but found <<" + result + ">>";
}

You can treat this like a multi-line comment, grabbing all characters between the opening quote and the first closing quote found ("nongreedy" behavior):

CSV.g

grammar CSV;

line returns [List<String> result]
scope { List fields; }
@init { $line::fields = new ArrayList(); }
	: (
	    (NEWLINE) => NEWLINE
	    | field (COMMA  field)* NEWLINE
	  )
	  { $result = $line::fields; }
	;

field
	: ( f=QUOTED
	  | f=UNQUOTED
	  | // nothing
	  )
 	{ $line::fields.add(($f == null) ? "" : $f.text); }
	;

NEWLINE	:	'\r'? '\n';

COMMA	:	',';

QUOTED	: '"' ( options {greedy=false;} : . )* '"'
	  {
	  	// Strip the surrounding quotes
	  	String txt = getText();
	  	setText(txt.substring(1, txt.length() -1));
	  };

UNQUOTED	:	~('\r' | '\n' | ',' | ' ' | '"')+;

This gets the job done.

Quoting, part 2

Let's allow quotes inside quoted fields. CSV uses "" to represent " in the final output.

CSVTests.java fragment

@Test
public void testQuoteEscaping() throws IOException, RecognitionException {
    String result = parseField("\"Before\"\"After\"");
    assert result.equals("Before\"After") : "Expected <<Before\"After>>, but found <<" + result + ">>";
}

The QUOTED lexer rule is clearly the place to put this. But think about how you would do it. [Seriously, go try some solutions before using mine. – RDC]

Here's one solution, including a bit of post-processing code to convert "" to ":

CSV.g fragment

QUOTED	: ('"' ( options {greedy=false;}: . )+ '"')+
	  {
	  	StringBuffer txt = new StringBuffer(getText());
	  	// Remove first and last double-quote
	  	txt.deleteCharAt(0);
	  	txt.deleteCharAt(txt.length()-1);
	  	// "" -> "
	  	int probe;
	  	while ((probe = txt.lastIndexOf("\"\"")) >= 0) {
	  		txt.deleteCharAt(probe);
	  	}
	  	setText(txt.toString());
	  };

Remove spaces around commas

CSV ignores leading and trailing spaces, so we should do the same. Here's the test case:

CSVTests.java fragment

@Test
public void testSpaceRemoval() throws IOException, RecognitionException {
    CSVParser parser = createParser("Red  ,   Green, ,Blue\n");
    List<String> result = parser.line();
    assert result.size() == 4 : "Expected 4 items";
    assert result.get(0).equals("Red") : "Expected Red";
    assert result.get(1).equals("Green") : "Expected Green";
    assert result.get(2).equals("") : "Expected empty";
    assert result.get(3).equals("Blue") : "Expected Blue";
}

Surprise – it passes! This is unexpected, but a look at the error output shows what's happening:

line 1:3 no viable alternative at character ' '
line 1:4 no viable alternative at character ' '
line 1:6 no viable alternative at character ' '
line 1:7 no viable alternative at character ' '
line 1:8 no viable alternative at character ' '
line 1:15 no viable alternative at character ' '

ANTLR 3's error recovery is taking care of this by skipping the space characters (which are unrecognized by the UNQUOTED rule. If we want to catch the error instead, we'll have to alter the error recovery mechanism.

Do you use continuous builds?

Consider disabling automatic error recovery while developing and testing your grammar. Otherwise, you may miss problems that crop up during automated unit testing (e.g. when using CruiseControl).

Intercepting errors

Since the last test passed when it should have failed, but printed a series of warning messages, let's intercept the error information and retain it for later use.

Start by altering the test to extract stored error messages from the lexer. This also requires setting the code that creates the parser and lexer:

CSVTests.java fragment

@Test
public void testSpaceRemoval() throws IOException, RecognitionException {
    CSVLexer lexer = createLexer("Red  ,   Green, ,Blue\n");
    CSVParser parser = createParser(lexer);
    List<String> result = parser.line();
    assert result.size() == 4 : "Expected 4 items";
    assert result.get(0).equals("Red") : "Expected Red";
    assert result.get(1).equals("Green") : "Expected Green";
    assert result.get(2).equals("") : "Expected empty";
    assert result.get(3).equals("Blue") : "Expected Blue";
    // Now make sure we didn't have any lexing errors, which were failing silently earlier
    // The parser drives the lexer, so check for exceptions after
    // parsing.
    List<RecognitionException> lexerExceptions = lexer.getExceptions();
    assert lexerExceptions.isEmpty() : "Lexer threw exceptions -- see output";
}

// and later...

private CSVParser createParser(CSVLexer lexer) throws IOException {
    CommonTokenStream tokens = new CommonTokenStream(lexer);
    CSVParser parser = new CSVParser(tokens);
    return parser;
}

private CSVParser createParser(String testString) throws IOException {
    CSVLexer lexer = createLexer(testString);
    CommonTokenStream tokens = new CommonTokenStream(lexer);
    CSVParser parser = new CSVParser(tokens);
    return parser;
}

private CSVLexer createLexer(String testString) throws IOException {
	CharStream stream = new ANTLRStringStream(testString);
	CSVLexer lexer = new CSVLexer(stream);
    return lexer;
}

Of course, this doesn't compile: CSVLexer doesn't implement getExceptions.

A little work with a debugger shows that the lexer calls reportError when it encounters an error:

public void reportError(RecognitionException e)

Override this method in the lexer and add the exception to a list (all of the changes are in @lexer::members):

CSV.g

grammar CSV;

@lexer::members {
	List<RecognitionException> exceptions = new ArrayList<RecognitionException>();

	public List<RecognitionException> getExceptions() {
  		return exceptions;
	}

	@Override
	public void reportError(RecognitionException e) {
  		super.reportError(e);
  		exceptions.add(e);
	}
}
line returns [List<String> result]
scope { List fields; }
@init { $line::fields = new ArrayList(); }
	: (
	    (NEWLINE) => NEWLINE
	    | field (COMMA  field)* NEWLINE
	  )
	  { $result = $line::fields; }
	;

field
	: ( f=QUOTED
	  | f=UNQUOTED
	  | // nothing
	  )
 	{ $line::fields.add(($f == null) ? "" : $f.text); }
	;

NEWLINE	:	'\r'? '\n';

COMMA	:	',';

QUOTED	: ('"' ( options {greedy=false;}: . )+ '"')+
	  {
	  	StringBuffer txt = new StringBuffer(getText());
	  	// Remove first and last double-quote
	  	txt.deleteCharAt(0);
	  	txt.deleteCharAt(txt.length()-1);
	  	// "" -> "
	  	int probe;
	  	while ((probe = txt.lastIndexOf("\"\"")) >= 0) {
	  		txt.deleteCharAt(probe);
	  	}
	  	setText(txt.toString());
	  };

// Anything except a line-breaking character is allowed.
UNQUOTED
	:	~('\r' | '\n' | ',' | ' ')+;

Regenerate the lexer and parser, then run the tests.

The test fails, as we expected.

Ignore whitespace around commas (we mean it this time...)

Now we can alter the grammar to ignore leading and trailing spaces around commas. The answer turns out to be ridiculously simple:

CSV.g fragment

COMMA	:	( ' '* ',' ' '*);

The test passes.