gem5/TESTING.md

This file explains how to use gem5's updated testing infrastructure. Running tests before submitting a patch is incredibly important so unexpected bugs don't creep into gem5.

gem5's testing infrastructure has the following goals:

• Simple for all users to run
• Fast execution in the simple case
• High coverage of gem5 code

Running unit tests

gem5 comes with unit tests, created using the Google Test framework. These can be built through SCons.

To build and run all the unit tests:

scons build/ALL/unittests.opt

All unit tests should be run prior to creating a pull request at https://github.com/gem5/gem5/pulls/

To compile and run just one set of tests (e.g. those declared within src/base/bitunion.test.cc):

scons build/ALL/base/bitunion.test.opt
./build/ALL/base/bitunion.test.opt

To list the available test functions from a test file:

./build/ALL/base/bitunion.test.opt --gtest_list_tests

To run a specific test function (e.g., BitUnionData.NormalBitfield):

./build/ALL/base/bitunion.test.opt --gtest_filter=BitUnionData.NormalBitfield

Running system-level tests

Within the tests directory we have system-level tests. These tests run the gem5 framework against various hardware configurations, with different ISAs, then verify the simulations execute correctly. These should be seen as high-level, coarse-grained tests to compliment the unit-tests.

Below is the most common way the tests are run. This will run all of the "quick" tests for X86, ARM, and RISC-V. These tests make up our best-supported platforms and use cases. When running these tests, you will likely want to us the option -j <CPUs> where CPUs is as large as you can make it. Additionally, it is often a good idea to run longer tests (e.g., linux boot) before submitting your patch.

cd tests
./main.py run

The above is the minumum you should run before posting a pull request to https://github.com/gem5/gem5/pulls/

Running tests from multiple directories

The command line above will walk the directory tree starting from the cwd (tests), and it will run every test it encounters in its path. It is possible to specify multiple root directories by providing several positional arguments:

./main.py run <directory1> <directory2> [...]

This will load every test in directory1 and directory2 (and their subdirectories).

'quick', 'long', and 'very-long' tests

There are three categoties of tests which may be run from the "tests" directory:

1. 'quick' tests. This suite of tests are designed to finish execution in a few hours, inclusive of compilation of gem5. We run these as part of our continuous integration tests on pull requests made to our repository. These tests all utilize a binary build scons build/ALL/gem5.opt, and thus only rely on a single compilation for the tests to run.
2. 'long' tests. This suite of tests are designed to finish execution in around 12 hours. They incorporate longer running tests which are unsuitable to run as part of the 'quick' tests. We run these daily via a scheduled job.
3. 'very-long' tests. This suite of tests are designed to finish execution in days. They incorporate tests which are too long to run frequntly We run these daily via a scheduled job.

When executing ./main.py run the 'quick' tests are executed. To run the 'long' tests execute:

./main.py run --length=long

and to run the 'very-long' tests execute:

./main.py run --length=very-long

In most cases we recommend running the 'quick' tests for most changes. Only in some cases, such as contributions which significantly change the codebase, do we recommend running the 'long' or 'very-long' suite.

Running tests in batch

The testing infrastructure provides the two needed methods to run tests in batch. First, you can list all of the tests based on the same tags as above in a machine-readable format by passing the -q flag. This will list all of the suites that match the given tag(s).

./main.py list -q --suites
SuiteUID:tests/gem5/hello_se/test_hello_se.py:testhello64-static-X86-opt
SuiteUID:tests/gem5/hello_se/test_hello_se.py:testhello64-dynamic-X86-opt
SuiteUID:tests/gem5/hello_se/test_hello_se.py:testhello32-static-X86-opt
SuiteUID:tests/gem5/hello_se/test_hello_se.py:testhello64-static-ARM-opt
SuiteUID:tests/gem5/hello_se/test_hello_se.py:testhello32-static-ARM-opt
SuiteUID:tests/gem5/m5_util/test_exit.py:m5_exit_test-X86-opt
SuiteUID:tests/gem5/test_build/test_build.py:build-X86-opt
SuiteUID:tests/gem5/test_build/test_build.py:build-RISCV-opt
SuiteUID:tests/gem5/test_build/test_build.py:build-ARM-opt

Next, you can run a single suite from the command line by passing the option --uid. For instance,

./main.py run --skip-build \
    --uid SuiteUID:tests/gem5/m5_util/test_exit.py:m5_exit_test-X86-opt

With this method, you can only run a single suite at a time. If you want to run more than one uid, you must call ./main.py multiple times.

Currently, you must specify --skip-build if you want to run a single suite or run in batch mode. Otherwise, you will build gem5 for all architectures.

Rerunning failed tests

While developing software a common practice is to run tests, make a change, and assert that the tests still pass. If tests fail you'll likely want to rerun and fix those specific tests without running redundant ones. The testing infrastructure allows you to rerun tests which failed in the last execution by using the rerun command.

./main.py run
#
#  Some tests fail...
#

# Rerun only the failed test suites (not the ones which passed).
./main.py rerun

If something goes wrong

The first step is to turn up the verbosity of the output using -v. This will allow you to see what tests are running and why a test is failing.

If a test fails, the temporary directory where the gem5 output was saved is kept and the path to the directory is printed in the terminal.

Debugging the testing infrastructure

Every command takes an option for the verbosity. -v, -vv, -vvv will increase the verbosity level. If something isn't working correctly, you can start here.

Most of the code for the testing infrastructure is in ext/testlib. This code contains the base code for tests, suites, fixtures, etc. The code in tests/gem5 is gem5-specific code. For the most part, the code in tests/gem5 extends the structures in ext/testlib.

Common errors

You may see a number of lines of output during test discovery that look like the following:

    Tried to load tests from ... but failed with an exception.
    Tried to load tests from ... but failed with an exception.
    ...

The testing library searches all python files in the tests/ directory. The test library executes each python file it finds searching for tests. It's okay if the file causes an exception. This means there are no tests in that file (e.g., it's not a new-style test).

Running Tests in Parallel

Whimsy has support for parallel testing baked in. This system supports running multiple suites at the same time on the same computer. To run suites in parallel, supply the -t <number-tests> flag to the run command.

For example, to run up to three test suites at the same time::

./main.py run --skip-build -t 3

Running Tests within GitHub Actions

To run these tests within GitHub Actions, we use the format of running tests by directory as shown above in the "Running Tests from Multiple Directories" section. These tests are run within workflow files, which can be found in the .github directory of this repository. You can learn more about workflows here.

Each workflow is made up of individual jobs, where each job is a set of tests that is executed on a runner within GitHub. In each workflow, each version of gem5.opt needed is first built, and then stored as an artifact for main.py to use.

There are two sets of runners within the gem5 repository: builders and runners. The builders have more resources available to allow for a quicker compilation of gem5, while the runners have less, and are meant to only run tests.

After the gem5 artifact has been uploaded, a runner can then download the versions needed for their tests. For example, in the daily-tests.yaml, in order to run the multi_isa tests, you need artifacts of ARM, RISCV, and VEGA_X86.

Adding Tests to GitHub Actions

In order to add new tests to our GitHub Actions testing infastructure, follow the format currently shown in the existing workflows. If the new tests were added to an already existing directory (ex. A very-long test in the gem5_library_example_tests), it will automatically be included into the weekly testing, since weekly-tests.yaml already contains a job for the gem5_library_example_tests.

However, if a new directory is added to the tests, you need to manually add a new step to the GitHub workflows. This would consist of both a step to build whatever version of gem5 was required if it wasn't already included in the file, as well as a step to run main.py in the given directory after downloading the gem5 artifact.