Darjeeling

Darjeeling is a language-agnostic search-based program repair tool. Unlike other repair tools, such as GenProg, SPR, and Nopol, Darjeeling delegates the responsibility of generating patches, obtaining coverage, analyzing code, and executing tests to other services. Once those auxillary concerns are removed, what is left is a lightweight framework for composing and executing repair algorithms: Darjeeling.

Features

• Language Agnostic: delegates syntax transformation and static analysis to other services.
• Containerization: uses Docker to quickly and safely evaluate patches without executing arbitrary code on your machine.
• Custom Repair Templates: uses Rooibos to support rich, custom repair templates for arbitrary languages.
• Asynchronous Evaluation: accelerates patch evaluation by spreading the load across multiple threads.
• Test Redundancy Checking: uses coverage information to skip test executions that can't be affected by a given patch.
• Test Ordering: numerous test ordering schemes reduce the cost of patch evaluation by prioritizing likely failing tests.
• Equivalent Patch Detection: uses static analysis to remove duplicate transformations from the search space.

Installation

Prerequisites

To use Darjeeling, Docker must be installed on your machine, and your user account must be a member of the docker group in order to avoid problems related to insufficient privileges . Python 3.9 or greater and pip3 must also be installed; Darjeeling will not work with older versions of Python 3 nor will it work with any versions of Python 2.

Optional Extras

We strongly recommend that you use pipenv to contain your installation of Darjeeling and avoid conflicting with system packages. To install pipenv, execute the following command:

$ pip install pipenv

Darjeeling

To install Darjeeling from source via pipenv, execute the following from the root of the repository:

$ pipenv install

Usage

After installing Darjeeling via pipenv as shown above, you can drop into the newly created virtual environment by executing the following command from the root of the repository:

$ pipenv shell

To exit from the virtual environment, you can execute the following command:

(Darjeeling) $ exit

Darjeeling exposes a command-line interface (CLI) for performing program repair, as demonstrated below. The CLI provides a single command, repair, which accepts the path to a Darjeeling configuration file format, described below.

$ darjeeling repair my-config.yml

Configuration File Format (v1.0)

The Darjeeling configuration file format is written in YAML. Below is an example of a configuration file. The configuration file itself can be found in the example/gcd directory.

version: '1.0'
seed: 0
threads: 16

# provides information about the program under repair, including
# the name of the Docker image that is used to provide it, the
# location of the source code for the program within that image,
# and instructions for building and testing it.
program:
  image: darjeeling/example:gcd
  language: c
  source-directory: /experiment/source
  build-instructions:
    time-limit: 10
    steps:
      - gcc gcd.c -o gcd
    steps-for-coverage:
      - gcc gcd.c -o gcd --coverage
  tests:
    type: genprog
    workdir: /experiment
    number-of-failing-tests: 1
    number-of-passing-tests: 10
    time-limit: 5

# specifies the method/tool that should be used to obtain coverage for
# the program.
coverage:
  method:
    type: gcov
    files-to-instrument:
      - gcd.c

localization:
  type: spectrum
  metric: tarantula

algorithm:
  type: exhaustive

transformations:
  schemas:
    - type: delete-statement
    - type: replace-statement
    - type: append-statement

optimizations:
  ignore-equivalent-insertions: yes
  ignore-dead-code: yes
  ignore-string-equivalent-snippets: yes

# places a limit on the resources (e.g., wall-clock time, test executions,
# candidate patches) that may be consumed during the search for a repair.
resource-limits:
  candidates: 100

Below, we describe the top-level options exposed by the configuration file:

• version: the version of the Darjeeling configuration file format that was used to write the file.
• seed: a seed for the random number generator.
• threads: number of threads over which the repair workload should be distributed.

program

The program section is used to provide essential details about the program that should be repaired. This section contains the following properties:

• image: the name of the Docker image that provides the program under repair.
• source-directory: the absolute path of the source code for the program within the provided Docker image.
• language: the language used by the program under repair. Note that, although Darjeeling supports multiple languages, it is not currently possible to fix bugs that involve more than one language.
• build-instructions: executable instructions for (re-)building the program inside the container. Discussed below.
• tests: details of the test suite used by the program. Discussed below.

program.language

Below is a list of the languages that are fully supported by Darjeeling. Darjeeling can automatically perform static analysis and compute coverage information for each of these languages.

• C: c
• C++: cpp
• Python: python

The text option (i.e., language: text) may be used to ignore the language of the program under repair and to treat each file as a text file. When this option is used, users will need to manually provide coverage information, and static analysis will not be performed.

program.build-instructions

This section provides instructions to Darjeeling for re-building the program for purposes of (a) evaluating candidate patches, and (b) instrumenting the program for coverage collection. Below is an except of the build-instructions section from the example above.

build-instructions:
   time-limit: 10
   steps:
     - gcc gcd.c -o gcd
   steps-for-coverage:
     - gcc gcd.c -o gcd --coverage

The time-limit specifies the maximum number of seconds that Darjeeling should wait before cancelling a build attempt. The steps property provides a sequence of shell commands that are used to build the program for the purpose of patch evaluation. Similarly, the steps-for-coverage property gives a sequence of shell commands that are used to build the program with coverage instrumentation.

program.tests

This section is used to describe the test suite used by the program. Darjeeling uses the program's test suite to determine the correctness of patches and to find acceptable patches that pass all tests. Darjeeling offers a number of test suite options out of the box, specified by the type property within the tests section. We describe these below.

program.tests[type:genprog]

This type of test suite provides convenient support for GenProg-style test scripts used by benchmarks such as ManyBugs, IntroClass, and the GenProg TSE 2012 benchmarks. GenProg-style test scripts accept a single argument specifying the name of the positive or negative test case that should be executed. Positive tests correspond to tests that pass on the original, unmodified program, whereas negative tests correpond to tests that fail on the original program. The positive tests are named using the form p{k}, where {k} is replaced by the number of the positive test (starting from 1). Similarly, negative tests are named n{k}, where {k} is replaced by the number of the negative test (starting from 1).

Below is an example of a genprog test suite:

tests:
  type: genprog
  workdir: /experiment
  number-of-failing-tests: 1
  number-of-passing-tests: 10
  time-limit: 5

The time-limit property specifies the maximum number of seconds that may elapse before a test execution is aborted and declared a failure. The number-of-passing-tests and number-of-failing-tests properties state the number of passing and failing tests. The workdir property gives the absolute path of the directory that contains the test.sh for the test harness.

program.tests[type:pytest]

This test suite is used by Python programs that support the popular pytest framework. Note that pytest can run unittest and nose tests natively.

Below is an except from a configuration file that uses pytest:

tests:
  type: pytest
  workdir: /opt/flask
  tests:
    - tests/test_config.py::test_get_namespace
    - tests/test_config.py::test_config_from_pyfile
    - tests/test_config.py::test_config_from_object

The workdir directory specifies the location at which pytest should be executed. The tests property gives a list of the names of the individual tests belonging to the test suite. Each name is given the format expected by pytest. That is, the name of the file containing the test (relative to workdir), followed by :: and the name of the test method. Note that automated discovery of test cases is not currently implemented, but is planned for a future release.

coverage

The coverage section provides Darjeeling with instructions for computing test coverage for the program under repair. Below, we describe the properties contained within this section:

• method: the tool that should be used to compute coverage for the program under repair. This information is necessary since Darjeeling deals with multiple languages, and each languages may have more than one associated tool for obtaining coverage. Out of the box, Darjeeling provides support for gcov, used for C and C++ programs, and pycoverage, used for Python programs. Support for additional coverage methods may be added via Darjeeling's plugin mechanism. Further details on these two methods are provided below.
• load-from-file: optionally specifies the location of a file from which coverage should be read. An example of such a coverage file can be found in example/flask/coverage.yml.
• restrict-to-files: optionally gives a list of files to which the coverage collection should be restricted to. Files should be given as paths relative to the specified source-directory for the program. Coverage that is generated for files outside of this set will be automatically discarded by Darjeeling. Note that this property uses the same format as localization.restrict-to-files.
• restrict-to-lines: optionally gives a list of lines that the coverage coverage collection should be restricted to. Lines outside of this set will be automatically ignored. This method uses the same format as localization.restrict-to-lines, shown below.

gcov

Below is an excerpt from an example configuration that uses gcov for coverage collection.

coverage:
  method:
    type: gcov
    files-to-instrument:
      - gcd.c

This method accepts a single, optional property, files-to-instrument. This property is very important. By default, programs compiled with the appropriate --coverage option set in their CFLAGS, CXXFLAGS, and LDFLAGS will produce .gcda files at runtime. The gcov tool computes coverage by reading both those .gcda files and their associated .gcno files, generated during compilation. More specifically, programs compiled with the --coverage option will write coverage data to disk during the normal termination of the program (i.e., the program exits with code zero). If the program abruptly terminates (e.g., due to memory corruption), .gcda files will NOT be produced.

This behavior is problematic for Darjeeling. It prevents collection from being obtained for failing tests that crash the program. As a workaround, Darjeeling adds source-based instrumentation to the program (in the form of a signal handler) that causes the program to (attempt to) flush its coverage information in thee event of abrupt termination. The files-to-instrument property gives the names of the source code files that provide entrypoints to the program binaries (i.e., main functions).

localization

The localization section provides instructions for localizing the fault inside the program under repair. Currently, the configuration file format supports a single type of fault localization: spectrum-based fault localization, which assigns a suspiciousness value to each line in the program under repair based on the number of passing and failing tests that touch that line. A suspiciousness metric is used to compute individual suspiciousness values. The configuration file exposes a number of metrics via its metric property:

• tarantula
• genprog
• jaccard
• ochiai

The localization section also exposes an exclude-files property, which may be used to exclude certain files from the fault localization. Each file should be given by its location relative to the source directory for the program under repair. In the example below, the files foo.c and bar.c are excluded from the fault localization.

exclude-files:
  - foo.c
  - bar.c

Individual source code lines can also be excluded using the exclude-lines property, as shown below. The exclude-lines property states which lines should be excluded from specified files. In the example below, lines 1, 2, 3 and 4 from foo.c, and lines 4, 6, 7 from bar.c are excluded from the fault localization.

exclude-lines:
  foo.c: [1, 2, 3, 4]
  bar.c: [4, 6, 7]

The fault localization can also be restricted to only consider certain files by using the restrict-to-files property, as shown below.

restrict-to-files:
  - foo.c

Similarly, the fault localization can also be restricted to individual source code lines using the restrict-to-lines property:

restrict-to-lines:
  foo.c: [11, 14, 16]

algorithm

The algorithm section outlines the search algorithm that should be used to search the space of candidate repairs. A description of the types of search algorithm exposed by the configuration file format is given below.

• exhaustive: iterates over all single-transformation patches within the search space until the termination criteria are met.
• genetic: implements a customisable genetic algorithm, inspired by GenProg.

transformations

The transformations section describes the space of program transformations from which candidate patches should be composed. The schemas property of this section specifies a list of the program transformation schemas, along with any parameter values for those schemas, that should may be used to construct concrete program transformations. Each entry in the schemas list must specify a type.

The configuration format supports three "classical" statement-based transformation schemas based on those introduced by GenProg: delete-statement, replace-statement, and prepend-statement; swap-statement has not been implemented at the time of writing. To learn more about why Darjeeling uses prepend-statement rather than the traditional append-statement schema, see the Darjeeling design document. Below is an example of schemas property that uses all of the classical statement-based schemas.

schemas:
  - type: delete-statement
  - type: replace-statement
  - type: prepend-statement

The configuration format also supports custom repair templates via match-rewrite patterns for Rooibos. Below is an example of a simple repair template that replaces all calls to foo with calls to bar.

- type: template
  match: "foo(:[1])"
  rewrite: "bar(:[1])"

The type property is set to template to indicate that this schema represents a Rooibos-based repair template. The match and rewrite sections are used to specify match and rewrite patterns, respectively.

Darjeeling also provides support for naive line-based transformations, given below, which can be used for programs that use languages that are not fully supported (i.e., programs that use the text language).

- type: delete-line
- type: insert-line
- type: replace-line

optimizations

The optimizations section is used to toggle various optimizations available to the repair process. By default, all optimizations are enabled. Below is a list of optimizations that can be toggled by the configuration file.

• use-scope-checking: ensures that all variable and function references that occur in a given transformation are visible from the scope into which they are being inserted.
• use-syntax-scope-checking: ensures that any keywords introduced by a transformation (e.g., break and continue) are permitted by their surrounding context.
• ignore-dead-code: prevents the insertion of code that exclusively writes to dead variables.
• ignore-equivalent-insertions: uses an approach inspired by instruction scheduling to prevent equivalent insertions of code.
• ignore-untyped-returns: prevents insertion of a return statement into a context where the type of the retval is incompatible with the return type of the enclosing method or function.
• ignore-string-equivalent-snippets: transforms donor code snippets into their canonical form, thus preventing the insertion of string-equivalent snippets.
• ignore-decls: prevents transformations that are either applied to declaration statements, or else solely introduce a declaration statement.
• only-insert-executed-code: prevents the insertion of code that has not been executed by at least one test case.

resource-limits

The resource-limits section of the configuration file is used to impose limits on the resources that may be consumed during the search. The search will be terminated upon hitting any of these limits. The limits specified in this section of the configuration file may be overridden by command-line options. If a limit for a particular resource is not given in either the configuration file or as a command-line argument, then the use of that resource will be unbounded (i.e., no limit will be imposed).

Below is a list of the resource limits that may be specified in the configuration file:

• candidates: the maximum number of candidate patches that may be evaluated. May be overriden at the command line by the --max-candidates option.
• time-minutes: the maximum length of wall-clock time that may be spent searching for a patch, given in minutes. May be overriden at the command line by the --max-time-mins option.

Search Algorithms

This section describes the different search algorithms that are supported by Darjeeling.

exhaustive

The exhaustive search algorithm exhaustively searches over all legal single-transformation patches within the search space until the termination criteria are fulfilled.

genetic

The genetic search algorithm implements a genetic algorithm that is inspired by the one used by GenProg, a formative search-based program repair tool for C. Below is an excerpt from a configuration file that uses a genetic search algorithm.

algorithm:
  type: genetic
  population: 80
  generations: 20
  tournament-size: 3
  mutation-rate: 0.6
  crossover-rate: 0.1
  test-sample-size: 0.4

Below is a list of the parameters that are exposed by genetic:

• population: the size of the (initial) population. Used to control the number of individuals that are selected as parents.
• generations: the maximum number of generations.
• tournament-size: the size of the tournament when performing tournament selection to choose parents. Larger tournament sizes lead to an increased selective pressure.
• mutation-rate: the probability of an individual mutation event.
• crossover-rate: the probability of an individual crossover event between two parents.
• test-sample-size: controls test sampling. When test sampling is enabled, the fitness of an individual is computed using a randomly selected subset of the test suite, rather than the entire test suite. (More specifically, test sampling selects a subset of the passing tests whilst keeping all of the failing tests.) The value of test-sample-size is used to specify the size of the subset (or sample). If test-sample-size is given as a float, then it will be treated as a fraction. If test-sample-size is given as an integer, then its value will be used as the absolute number of (passing) tests that should be included in the sample. If test-sample-size is omitted or set to null, test sampling will be disabled.

Extending Darjeeling via Plugins

Users may extend Darjeeling's capabilities with their own plugins. Upon launch, Darjeeling will find and automatically import all installed Python packages whose name starts with darjeeling_ (e.g., darjeeling_ardupilot).

Darjeeling treats the following features as framework extension points, allowing variants to be added by plugins:

• Search algorithms
• Transformation schemas
• Test harnesses
• Coverage tools (e.g., jacoco, pycoverage, sancov)
• Spectrum-based fault localisation suspiciousness metrics