Skip to content
Switch branches/tags

Latest commit


Git stats


Failed to load latest commit information.
Latest commit message
Commit time


ANTLRv4 grammar-based test generator

Grammarinator is a random test generator / fuzzer that creates test cases according to an input ANTLR v4 grammar. The motivation behind this grammar-based approach is to leverage the large variety of publicly available ANTLR v4 grammars.

The trophy page of the found issues is available from the wiki.


  • Python >= 3.5
  • pip and setuptools Python packages (the latter is automatically installed by pip).
  • Java SE >= 7 JRE or JDK


The quick way (to install the latest official release):

pip3 install grammarinator

Or clone the project and run setuptools (to install the freshest development revision):

python3 install


As a first step, Grammarinator takes an ANTLR v4 grammar and creates a test generator script in Python3. Such a generator can be subclassed later to customize it further if needed.

Basic command-line syntax of test generator creation:

grammarinator-process <grammar-file(s)> -o <output-directory> --no-actions


Grammarinator uses the ANTLR v4 grammar format as its input, which makes existing grammars (lexer and parser rules) easily reusable. However, because of the inherently different goals of a fuzzer and a parser, inlined code (actions and conditions, header and member blocks) are most probably not reusable, or even preventing proper execution. For first experiments with existing grammar files, grammarinator-process supports the command-line option --no-actions, which skips all such code blocks during fuzzer generation. Once inlined code is tuned for fuzzing, that option may be omitted.

After having generated and optionally customized a fuzzer, it can be executed by the grammarinator-generate script (or by manually instantiating it in a custom-written driver, of course).

Basic command-line syntax of grammarinator-generate:

grammarinator-generate <generator> -r <start-rule> -d <max-depth> \
  -o <output-pattern> -n <number-of-tests> \
  -t <transformer1> -t <transformer2>

Beside generating test cases from scratch based on the ANTLR grammar, Grammarinator is also able to recombine existing inputs or mutate only a small portion of them. To use these additional generation approaches, a population of selected test cases has to be prepared. The preparation happens with the grammarinator-parse tool, which processes the input files with an ANTLR grammar (possibly with the same one as the generator grammar) and builds grammarinator tree representations from them (with .grt extension). Having a population of such .grt files, grammarinator-generate can make use of them with the --population cli option. If the --population option is set, then Grammarinator will choose a strategy (generation, mutation, or recombination) randomly at the creation of every new test case. If any of the strategies is unwanted, they can be disabled with the --no-generate, --no-mutate or --no-recombine options.

Basic command line syntax of grammarinator-parse:

grammarinator-parse <grammar-file(s)> -r <start-rule>\
  -i <input_file> -o <output-directory>


Real-life grammars often use recursive rules to express certain patterns. However, when using such rule(s) for generation, we can easily end up in an unexpectedly deep call stack. With the --max-depth or -d options, this depth - and also the size of the generated test cases - can be controlled.

Another specialty of the ANTLR grammars is that they support so-called hidden tokens. These rules typically describe such elements of the target language that can be placed basically anywhere without breaking the syntax. The most common examples are comments or whitespaces. However, when using these grammars - which don't define explicitly where whitespace may or may not appear in rules - to generate test cases, we have to insert the missing spaces manually. This can be done by applying a serializer (with the -s option) to the tree representation of the output tests. A simple serializer - that inserts a space after every unparser rule - is provided by Grammarinator (grammarinator.runtime.simple_space_serializer).

In some cases, we may want to postprocess the output tree itself (without serializing it). For example, to enforce some logic that cannot be expressed by a context-free grammar. For this purpose the transformer mechanism can be used (with the -t option). Similarly to the serializers, it will take a tree as input, but instead of creating a string representation, it is expected to return the modified (transformed) tree object.

As a final thought, one must not forget that the original purpose of grammars is the syntax-wise validation of various inputs. As a consequence, these grammars encode syntactic expectations only and not semantic rules. If we still want to add semantic knowledge into the generated test, then we can inherit custom fuzzers from the generated ones and redefine methods corresponding to lexer or parser rules in ways that encode the required knowledge (e.g.: HTMLCustomGenerator).

Working Example

The repository contains a minimal example to generate HTML files. To give it a try, run the processor first:

grammarinator-process examples/grammars/HTMLLexer.g4 examples/grammars/HTMLParser.g4 \
  -o examples/fuzzer/

Then, use the generator to produce test cases:

grammarinator-generate HTMLCustomGenerator.HTMLCustomGenerator -r htmlDocument -d 20 \
  -o examples/tests/test_%d.html -n 100 \
  -s HTMLGenerator.html_space_serializer \
  --sys-path examples/fuzzer/


Grammarinator was tested on:

  • Linux (Ubuntu 16.04 / 18.04)
  • Mac OS X (Sierra 10.12 / High Sierra 10.13 / Mojave 10.14 / Catalina 10.15)
  • Windows (Server 2012 R2 / Server version 1809 / Windows 10)


Background on Grammarinator is published in:

  • Renata Hodovan, Akos Kiss, and Tibor Gyimothy. Grammarinator: A Grammar-Based Open Source Fuzzer. In Proceedings of the 9th ACM SIGSOFT International Workshop on Automating Test Case Design, Selection, and Evaluation (A-TEST 2018), pages 45-48, Lake Buena Vista, Florida, USA, November 2018. ACM.

Copyright and Licensing

Licensed under the BSD 3-Clause License.