See also NewRegressionFramework.
Running the M5 regression tests is the recommended way to make sure that any changes to the simulator still comply with M5. Regression Testing is performed using SCons to help guide which tests are run. All current tests are located in the: tests/ directory. The command for running regression tests has the following format:
% scons build/<arch>/tests/<binary_type>/<test_directory>/<mode>
Here is an example of running all of the 'quick' regression tests for the ALPHA architecture in syscall-emulation (SE) mode.
% scons build/ALPHA/tests/debug/quick/se ... **** build/ALPHA/tests/debug/quick/se/00.hello/alpha/linux/o3-timing PASSED! ... **** build/ALPHA/tests/debug/quick/se/00.hello/alpha/linux/simple-atomic PASSED! ... **** build/ALPHA/tests/debug/quick/se/00.hello/alpha/linux/simple-timing PASSED! ... **** build/ALPHA/tests/debug/quick/se/00.hello/alpha/tru64/o3-timing PASSED! ...
The script util/regress can be used for running regression tests. Run the script with the option '-h' or '--help' for the options can be passed to the script.
Regression scripts execution description
The regression scripts are complicated. Below, we describe the interaction and execution order of all of the scripts.
- util/regress executes scons with varying options depending on its parameters. This scripts executes both compiling (scons build/<build_opts (ISA)>/gem5.<compile-variants (opt,debug,etc.)>) and running regressions (scons build/<ISA or build_opts>/tests/<test-variants (opt,debug,etc.)>/<quick/long>/<modes (se/fs)>
- scons is executed. For compile tests, this is the same way you compile gem5. For regression tests it is a little different. There are XXX different parameters to each regression test.
- What build to use (e.g., ARM, X86, ALPHA_MOESI_CMP_token)
- Which binary to use (opt, debug, fast)
- Which test to run
- quick vs long
- fs vs se mode
- The primary configuration script to execute
- The secondary configuration script to execute
- Overall, the scons command looks like this:
% scons build/<build to use>/tests/<binary name>/<quick or long>/<se or fs>/<location of secondary script>/<ISA>/linux/<primary config script>
- Note: You don't need to specify every subdirectory. If you leave out some subdirectories, all of the tests in the directory specified will run.
- scons executes builds/<build to use>/gem5.<binary name> tests/run.py <crazy path that encodes everything>
- Control is now transferred to tests/run.py
- run.py execute the primary configuration script (the last directory on the scons command line)
- The way scons knows which tests to run is by looking in tests/<quick or long>/<se or fs>/<secondary script>/ref/<ISA>/linux/
- The name of this directory should match the primary script (in tests/configs) that you want to execute.
- tests/run.py executes the python file tests/configs/<name of dir>.py
- After the file in tests/configs executes to completion, run.py executes the secondary config file, or the file specific to this test
- The file is found in tests/<quick or long>/<se or fs>/<secondary script>/test.py
- This file is usually used to specify the binary to simulate
- tests/run.py sets up some other parameters, like initializing CPUs.
- tests/run.py executes the function run_test(root)
- Note: At some point in the files that tests/run.py executes, you must specify a variable "root"
- You can overload run_test(root) by specifying it in one of the executed files. These files are executed in python as if their code were contained in run.py.
- Usually, in run_test(root), gem5 simulation is started.
- The output is written to build/<build to use>/tests/<binary name>/<quick or long>/<se or fs>/<secondary script>/<ISA>/linux/<primary script>
- The output is compared to the reference in tests/<quick or long>/<se or fs>/<secondary script>/ref/<ISA>/linux/<primary script>
Creating Your Own Regressions
Adding a regression test is done by adding a new directory to the 'm5/tests' directory and filling that directory up with the appropriate reference files. Every regression test needs these files in order to compare the current changes of M5 with the 'known working state:
- simout - Standard Output Stream
- simerr - Standard Error Stream
- config.ini - M5 Configuration File
- stats.txt - Statistics From M5 Simulation
One needs to run the M5 with the appropriate binary in order to get the aforementioned simulation files. For example, a command line to collect simulation data for a hello world regression test might be:
% build/ALPHA/gem5.fast -re configs/example/se.py --cmd=tests/test-progs/hello/bin/alpha/hello
Creating the 'hello world' regression test directory and copying over the reference output might then follow this command line sequence:
% mkdir tests/quick/00.hello % mkdir tests/quick/00.hello/ref % cp config.ini tests/quick/00.hello/ref % cp stats.txt tests/quick/00.hello/ref % cp simout tests/quick/00.hello/ref % cp simerr tests/quick/00.hello/ref
The last thing one needs to do is add a 'test.py' configuration file that specifies the workload for the regression test. For this 'hello world' example, the 'test.py' file would just contain:
root.system.cpu.workload = LiveProcess(cmd='hello', executable = binpath('hello'))
Once the test.py is copied to the 'hello world' regression test directory, the individual regression test can be run:
% cp test.py tests/quick/00.hello % scons build/ALPHA/tests/debug/quick/00.hello
Updating Regression Tests
Once you've made and verified changes to the simulator, you can update any regression tests with the "update" option:
% scons --update-ref build/ALPHA/tests/debug/quick/00.hello