Tree-sitter: External scanners should be able to produce tokens without calling `advance`

[movermeyer]

Introduction

I was trying to make a parser for YAML, by implementing the grammar expressed by libyaml (the same grammar is also implemented by other YAML parsers as well):

• libyaml's grammar

Notice that there is the rule:
block_sequence ::= BLOCK-SEQUENCE-START (BLOCK-ENTRY block_node?)* BLOCK-END

BLOCK-SEQUENCE-START is a zero-width token that is generated by scanners when the indentation increases. It is important to make the grammar unambiguous.

I tried to create an external scanner that produced this token. However, no matter what I did, tree-sitter parse would always return an ERROR node.
I eventually realized that this was because external scanners cannot return a token without first calling advance.

---

Example

This is a trivial grammar that should match a file containing a single dash using an external scanner:

grammar.js

module.exports = grammar({
  name: 'example',

  externals: $ => [
    $.ZERO_WIDTH_TOKEN,
  ],

  rules: {
    document: $ => seq($.ZERO_WIDTH_TOKEN, $.dash),
    dash: $ => '-'
  }
});

src/scanner.cc

#include <tree_sitter/parser.h>

namespace {

  enum TokenType {
    ZERO_WIDTH_TOKEN
  };

  struct Scanner {
    Scanner() {}

    unsigned serialize(char *buffer) {
      return 0;
    }

    void deserialize(const char *buffer, unsigned length) {
    }

    bool scan(TSLexer *lexer, const bool *valid_symbols) {
      lexer->result_symbol = ZERO_WIDTH_TOKEN;
      return true;
    }
  };

}

extern "C" {
  void *tree_sitter_example_external_scanner_create() {
    return new Scanner();
  }

  bool tree_sitter_example_external_scanner_scan(void *payload, TSLexer *lexer,
                                              const bool *valid_symbols) {
    Scanner *scanner = static_cast<Scanner *>(payload);
    return scanner->scan(lexer, valid_symbols);
  }

  unsigned tree_sitter_example_external_scanner_serialize(void *payload, char *buffer) {
    Scanner *scanner = static_cast<Scanner *>(payload);
    return scanner->serialize(buffer);
  }

  void tree_sitter_example_external_scanner_deserialize(void *payload, const char *buffer, unsigned length) {
    Scanner *scanner = static_cast<Scanner *>(payload);
    scanner->deserialize(buffer, length);
  }

  void tree_sitter_example_external_scanner_destroy(void *payload) {
    Scanner *scanner = static_cast<Scanner *>(payload);
    delete scanner;
  }
}

$ tree-sitter generate && tree-sitter parse --debug file_with_dash.txt
new_parse
process version:0, version_count:1, state:1, row:0, col:0
lex_external state:1, row:0, column:0
lex_internal state:0, row:0, column:0
  consume character:'-'
lexed_lookahead sym:dash, size:1
detect_error
resume version:0
process version:0, version_count:1, state:0, row:0, col:0
skip_token symbol:dash
process version:0, version_count:1, state:0, row:0, col:1
lex_external state:1, row:0, column:1
lex_internal state:0, row:0, column:1
  skip character:10
lexed_lookahead sym:end, size:1
recover_eof
done
(ERROR [0, 0] - [1, 0]
  (dash [0, 0] - [0, 1]))
file_with_dash.txt  0 ms    (ERROR [0, 0] - [1, 0])

I would have expected this to match fine.

Especially since it is possible to match zero-width rules if you don't use an external scanner:

grammar.js

module.exports = grammar({
  name: 'example',

  rules: {
    document: $ => seq($.zero_width_token, $.dash),
    zero_width_token: $ => '',
    dash: $ => '-'
  }
});

$ tree-sitter generate && tree-sitter parse --debug file_with_dash.txt
new_parse
process version:0, version_count:1, state:1, row:0, col:0
lex_internal state:3, row:0, column:0
lexed_lookahead sym:zero_width_token, size:0
shift state:2
process version:0, version_count:1, state:2, row:0, col:0
lex_internal state:0, row:0, column:0
  consume character:'-'
lexed_lookahead sym:dash, size:1
shift state:4
process version:0, version_count:1, state:4, row:0, col:1
lex_internal state:0, row:0, column:1
  skip character:10
lexed_lookahead sym:end, size:1
reduce sym:document, child_count:2
accept
done
(document [0, 0] - [1, 0]
  (zero_width_token [0, 0] - [0, 0])
  (dash [0, 0] - [0, 1]))

[maxbrunsfeld]

I eventually realized that this was because external scanners cannot return a token without first calling advance.

Thanks for the minimal repro steps!

We make heavy use of zero-width external tokens in the Python grammar (for indent/dedent) and the JavaScript/Typescript grammars (automatic semicolons).

It looks like there's a bug where a zero-width token isn't being allowed right at the beginning of a file. I'll fix this asap.

[movermeyer]

Thanks for resolving this so quickly! ❤️

---

Could I ask for a new release and/or docs for how to build tree-sitter?

[maxbrunsfeld]

Yeah I’ll publish a new release to npm soon. In the meantime, you just need to ‘git submodule update —init’, install rust and run ‘cargo build —release’ in the root directory. That’ll produce a binary in the ‘target/release’ folder.