Lexers and Unicode text

Prior to ANTLR 4.7, generated lexers in most targets only supported part of the Unicode standard (code points up to U+FFFF). As of ANTLR 4.7, the lexers in all language runtimes support the full range of Unicode code points up to U+10FFFF.

C++, Python, Go, and Swift APIs didn't need any API changes to support Unicode code points, so we decided to leave those class interfaces as-is.

Java, C#, and JavaScript runtimes required changes and, rather than break the previous interface, we deprecated them. (The Java-target deprecated ANTLRInputStream and ANTLRFileStream APIs only support Unicode code points up to U+FFFF.) Now, those targets must create CharStreams from input using CharStreams.fromPath(), CharStreams.fromFileName(), etc...

A big shout out to Ben Hamilton (github bhamiltoncx) for his superhuman efforts across all targets to get true support for U+10FFFF code points.

Example

The Java, C#, and JavaScript runtimes use the new factory style stream creation interface. For example, here is some sample Java code that uses CharStreams.fromPath():

public static void main(String[] args) {
  CharStream charStream = CharStreams.fromPath(Paths.get(args[0]));
  Lexer lexer = new UnicodeLexer(charStream);
  CommonTokenStream tokens = new CommonTokenStream(lexer);
  tokens.fill();
  for (Token token : tokens.getTokens()) {
    System.out.println("Got token: " + token.toString());
  }
}

Unicode Code Points in Lexer Grammars

To refer to Unicode code points in lexer grammars, use the \u string escape plus up to 4 hex digits. For example, to create a lexer rule for a single Cyrillic character by creating a range from U+0400 to U+04FF:

CYRILLIC : '\u0400'..'\u04FF' ; // or [\u0400-\u04FF] without quotes

Unicode literals larger than U+FFFF must use the extended \u{12345} syntax. For example, to create a lexer rule for a selection of smiley faces from the Emoticons Unicode block:

EMOTICONS : ('\u{1F600}' | '\u{1F602}' | '\u{1F615}') ; // or [\u{1F600}\u{1F602}\u{1F615}]

Finally, lexer char sets can include Unicode properties. Each Unicode code point has at least one property that describes the type group to which it belongs (e.g. alpha, number, punctuation). Other properties can be the language script or special binary properties and Unicode code blocks. That means however, that a property specifies a group of code points, hence they are only allowed in lexer char sets.

EMOJI : [\p{Emoji}] ;
JAPANESE : [\p{Script=Hiragana}\p{Script=Katakana}\p{Script=Han}] ;
NOT_CYRILLIC : [\P{Script=Cyrillic}] ;

See lexer-rules.md for more detail on Unicode escapes in lexer rules.

Migration

Code for 4.6 looked like this:

CharStream input = new ANTLRFileStream("myinputfile");
JavaLexer lexer = new JavaLexer(input);
CommonTokenStream tokens = new CommonTokenStream(lexer);

(It didn't use UTF-8 by default, despite the documentation saying so previously; it actually depended on the calling environments default.)

Code for 4.7 assumes UTF-8 by default and looks like this:

CharStream input = CharStreams.fromFileName("inputfile");
JavaLexer lexer = new JavaLexer(input);
CommonTokenStream tokens = new CommonTokenStream(lexer);

Or, if you'd like to specify the file encoding:

CharStream input = CharStreams.fromFileName("inputfile", Charset.forName("windows-1252"));

Motivation

After a lively discussion, I (parrt) decided not to simply gut the 4.6 ANTLRFileStream and ANTLRInputStream to incorporate the new U+10FFFF functionality. I decided to deprecate the old interface and recommend use of the new interface to prevent confusion. My reasoning is summarized as:

  • I didn't like the idea of breaking all 4.6 code. To get the previous streams to properly support > 16 bit Unicode would require a lot of changes to the method signatures.
  • Using int buffer element types would double the size of memory required to hold streams in memory, given that we buffer everything (and I didn't want to change that aspect of the streams).
  • The new factory-style interface supports creation of the smallest possible code point buffer element size according to the Unicode code points found in the input stream. This means using half as much memory as the old {@link ANTLRFileStream}, which assumed 16-bit characters, for ASCII text.
  • Through some serious testing and performance tweaking, the new streams perform as fast or faster than the 4.6 streams.

WARNING. You should avoid using both the deprecated and the new streams in the same application because you will see a nontrivial performance degradation. This speed hit is because the Lexer‘s internal code goes from a monomorphic to megamorphic dynamic dispatch to get characters from the input stream. Java’s on-the-fly compiler (JIT) is unable to perform the same optimizations so stick with either the old or the new streams, if performance is a primary concern. See the extreme debugging and spelunking needed to identify this issue in our timing rig.

Legacy grammar using surrogate code units

Legacy grammars that did their own UTF-16 surrogate code unit matching will need to continue to use ANTLRInputStream (Java target) until the parser-application code can upgrade to CharStreams interface. Then the surrogate code unit matching should be removed from the grammar in favor of letting the new streams do the decoding.

Prior to 4.7, application code could directly pass Token.getStartIndex() and Token.getStopIndex() to Java and C# String APIs (because both used UTF-16 code units as the fundamental unit of length). With the new streams, clients will have to convert from code point indices to UTF-16 code unit indices. Here is some (Java) code to show you the necessary logic:

public final class CodePointCounter {
  private final String input;
  public int inputIndex = 0;
  public int codePointIndex = 0;
  
  public int advanceToIndex(int newCodePointIndex) {
    assert newCodePointIndex >= codePointIndex;
    while (codePointIndex < newCodePointOffset) {
        int codePoint = Character.codePointAt(input, inputIndex);
        inputIndex += Character.charCount(codePoint);
        codePointIndex++;
    }
    return inputIndex;
  }
}

Character Buffering, Unbuffered streams

The ANTLR character streams still buffer all the input when you create the stream, as they have done for ~20 years.

If you need unbuffered access, please note that it becomes challenging to create parse trees. The parse tree has to point to tokens which will either point into a stale location in an unbuffered stream or you have to copy the characters out of the buffer into the token. That defeats the purpose of unbuffered input. See the ANTLR 4 book “13.8 Unbuffered Character and Token Streams”. Unbuffered streams are primarily useful for processing infinite streams during the parse and require that you manually buffer characters. Use UnbufferedCharStream and UnbufferedTokenStream.

CharStream input = new UnbufferedCharStream(is);
CSVLexer lex = new CSVLexer(input); // copy text out of sliding buffer and store in tokens
lex.setTokenFactory(new CommonTokenFactory(true));
TokenStream tokens = new UnbufferedTokenStream<CommonToken>(lex);
CSVParser parser = new CSVParser(tokens);
parser.setBuildParseTree(false);
parser.file();

Your grammar that needs to have embedded actions that access the tokens as they are created, but before they disappear and are garbage collected. For example,

data : a=INT {int x = Integer.parseInt($a.text);} ;

From the code comments of CommonTokenFactory:

That true in new CommonTokenFactory(true) indicates whether CommonToken.setText should be called after constructing tokens to explicitly set the text. This is useful for cases where the input stream might not be able to provide arbitrary substrings of text from the input after the lexer creates a token (e.g. the implementation of CharStream.getText in UnbufferedCharStream throws an UnsupportedOperationException). Explicitly setting the token text allows Token.getText to be called at any time regardless of the input stream implementation.

Currently, only Java, C++, and C# have these unbuffered streams implemented.