Skip to content

Latest commit

 

History

History
768 lines (598 loc) · 31.7 KB

test-framework.rst

File metadata and controls

768 lines (598 loc) · 31.7 KB
Raw

SCons Testing Framework

Introduction

SCons uses extensive automated tests to ensure quality. The primary goal is that users be able to upgrade from version to version without any surprise changes in behavior.

In general, no change goes into SCons unless it has one or more new or modified tests that demonstrably exercise the bug being fixed or the feature being added. There are exceptions to this guideline, but they should be just that, exceptions. When in doubt, make sure it's tested.

Test organization

There are three types of SCons tests:

End-to-End Tests
End-to-end tests of SCons are Python scripts (*.py) underneath the test/ subdirectory. They use the test infrastructure modules in the testing/framework subdirectory. They set up small complete projects and call SCons to execute them, checking that the behavior is as expected.
Unit Tests
Unit tests for individual SCons modules live underneath the SCons/ subdirectory and are the same base name as the module to be tested, with Tests appended to the basename. For example, the unit tests for the Builder.py module are in the BuilderTests.py script. Unit tests tend to be based on assertions.
External Tests
For the support of external Tools (in the form of packages, preferably), the testing framework is extended so it can run in standalone mode. You can start it from the top-level directory of your Tool's source tree, where it then finds all Python scripts (*.py) underneath the local test/ directory. This implies that Tool tests have to be kept in a directory named test, like for the SCons core.

Contrasting end-to-end and unit tests

In general, end-to-end tests verify hardened parts of the public interface: interfaces documented in the manpage that a user might use in their project. These cannot be broken (of course, errors can be corrected, though sometimes a transition period may be required). Unit tests are now considered for testing internal interfaces, which do not themselves directly have API guarantees. An example could be using and end-to-end test to verify that things added by env.Append() actually appear correctly in issued command lines, while unit tests assure correct behavior given various inputs of internal routines that Append() may make use of. If a reported error can be tested by adding a new case to an existing unit test, by all means, do that, as it tends to be simpler and cleaner. On the other hand, reported problems that come with a reproducer are by their nature more like an e2e test - this is something a user has tried in their SConscripts that didn't have the expected result.

End-to-end tests are by their nature harder to debug. For the unit tests, you're running a test program directly, so you can drop straight into the Python debugger by calling runtest.py with the -d / --debug option and setting breakpoints to help examine the internal state as the test is running. The e2e tests are each mini SCons projects executed by an instance of scons in a subprocess, and the Python debugger isn't particularly useful in this context. There's a separate section of this document on that topic: see Debugging end-to-end tests.

Naming conventions

The end-to-end tests, more or less, follow this naming convention:

  1. All tests end with a .py suffix.

  2. In the General form we use

    Feature.py

    for the test of a specified feature; try to keep this description reasonably short

    Feature-x.py

    for the test of a specified feature using option x

  3. The command line option tests take the form

    option-x.py

    for a lower-case single-letter option

    option--X.py

    upper-case single-letter option (with an extra hyphen, so the file names will be unique on case-insensitive systems)

    option--lo.py

    long option; you can abbreviate the long option name to a few characters (the abbreviation must be unique, of course).

  4. Use a suitably named subdirectory if there's a whole group of related test files.

Testing Architecture

The test framework provides a lot of useful functions for use within a test program. This includes test setup, parameterization, running tests, looking at results and reporting outcomes. You can run a particular test directly by making sure the Python interpreter can find the framework:

$ PYTHON_PATH=testing/framework python SCons/ActionTests.py

The framework does not provide facilities for handling a collection of test programs. For that, SCons provides a driver script runtest.py. Help is available through the -h option:

$ python runtest.py -h

You run tests from the top-level source directory. To simply run all the tests, use the -a option:

$ python runtest.py -a

You may specifically list one or more tests to be run. runtest considers all arguments it doesn't recognize as options to be part of the test list:

$ python runtest.py SCons/BuilderTests.py
$ python runtest.py -t test/option/option-j.py test/option/option-p.py

Folder names work in the test list as well, so you can do:

$ python runtest.py test/SWIG

to run all SWIG tests (and no others).

You can also use the -f option to execute just the tests listed in a test list file:

$ cat testlist.txt
test/option/option-j.py
test/option/option-p.py
$ python runtest.py -f testlist.txt

List one test file per line. Lines that begin with the comment mark # will be ignored (this lets you quickly change the test list by commenting out a few tests in the testlist file).

If more than one test is run, the runtest.py script prints a summary and count of tests that failed or yielded no result (skips). Skipped tests do not count towards considering the overall run to have failed, unless the --no-ignore-skips option is used. Passed tests can be listed using the --passed option, though this tends to make the result section at the end quite noisy, which is why it's off by default. Also by default, runtest.py prints a running count and completion percentage message for each test case as it finishes, along with the name of the test file. You can quiet this output: have a look at the -q, -s and -k options.

Since a test run can produce a lot of output that you may want to examine later, there is an option -o FILE to save the same output that went to the screen to a file named by FILE. There is also an option to save the results in a custom XML format.

The above invocations all test against the SCons files in the current directory (that is, in ./SCons, and do not require that a packaging build of SCons be performed first. This is the most common mode: make some changes, and test the effects in place. The runtest.py script supports additional options to run tests against unpacked packages in the build/test-*/ subdirectories.

If you are testing a separate Tool outside of the SCons source tree, call the runtest.py script in external (stand-alone) mode:

$ python ~/scons/runtest.py -e -a

This ensures that the testing framework doesn't try to access SCons classes needed for some of the internal test cases.

Note that as each test is run, it is executed in a temporary directory created just for that test, which is by default removed when the test is complete. This ensures that your source directories don't get clobbered with temporary files and changes from the test runs. If the test itself needs to know the directory, it can be obtained as test.workdir, or more commonly by calling test.workpath(), a function which takes a path-component argument and returns the path to that path-component in the testing directory.

The use of an ephemeral test directory means that you can't simply change into a directory to debug after a test has gone wrong. For a way around this, check out the PRESERVE environment variable. It can be seen in action in How to convert old tests to use fixtures below.

Not running tests

If you simply want to check which tests would get executed, you can call the runtest.py script with the -l option combined with whichever test selection options (see below) you intend to use. Example:

$ python runtest.py -l test/scons-time

runtest.py also has a -n option, which prints the command line for each test which would have been run, but doesn't actually run them:

$ python runtest.py -n -a

Selecting tests

When started in standard mode:

$ python runtest.py -a

runtest.py assumes that it is run from the SCons top-level source directory. It then dives into the SCons and test directories, where it tries to find filenames

*Test.py
for the SCons directory (unit tests)
*.py
for the test directory (end-to-end tests)

When using fixtures, you may end up in a situation where you have supporting Python script files in a subdirectory which shouldn't be picked up as test scripts of their own. There are two options here:

  1. Add a file with the name sconstest.skip to your subdirectory. This tells runtest.py to skip the contents of the directory completely.
  2. Create a file .exclude_tests in each directory in question, and in it list line-by-line the files to exclude from testing.

The same rules apply when testing external Tools when using the -e option.

Example End-to-End test script

To illustrate how the end-to-end test scripts work, let's walk through a simple Hello, world! example:

#!python
import TestSCons

test = TestSCons.TestSCons()

test.write('SConstruct', """\
Program('hello.c')
""")

test.write('hello.c', """\
#include <stdio.h>

int
main(int argc, char *argv[])
{
    printf("Hello, world!\\n");
    exit (0);
}
""")

test.run()

test.run(program='./hello', stdout="Hello, world!\n")

test.pass_test()

Explanation

import TestSCons
Imports the main infrastructure for SCons tests. This is normally the only part of the infrastructure that needs importing. Sometimes other Python modules are necessary or helpful, and get imported before this line.
test = TestSCons.TestSCons()

Initializes an object for testing. A fair amount happens under the covers when the object is created, including:

  • A temporary directory is created for all the in-line files that will get created.
  • The temporary directory's removal is arranged for when the test is finished.
  • The test does os.chdir() to the temporary directory.
test.write('SConstruct', ...)
This line creates an SConstruct file in the temporary directory, to be used as input to the scons run(s) that we're testing. Note the use of the Python triple-quoted string for the contents of the SConstruct file (and see the next section for an alternative approach).
test.write('hello.c', ...)
This line creates an hello.c file in the temporary directory. Note that we have to escape the newline in the "Hello, world!\\n" string so that it ends up as a single backslash in the hello.c file on disk.
test.run()

This actually runs SCons. Like the object initialization, things happen under the covers:

  • The exit status is verified; the test exits with a failure if the exit status is not zero.
  • The error output is examined, and the test exits with a failure if there is any.
test.run(program='./hello', stdout="Hello, world!\n")
This shows use of the TestSCons.run() method to execute a program other than scons, in this case the hello program we just built. The stdout= keyword argument also tells the TestSCons.run() method to fail if the program output does not match the expected string "Hello, world!\n". Like the previous test.run() line, it will also fail the test if the exit status is non-zero, or there is any error output.
test.pass_test()
This is always the last line in a test script. If we get to this line, it means we haven't bailed out on a failure or skip, so the result was good. It prints PASSED on the screen and makes sure we exit with a 0 status to indicate the test passed. As a side effect of destroying the test object, the created temporary directory will be removed.

Working with fixtures

In the simple example above, the files to set up the test are created on the fly by the test program. We give a filename to the TestSCons.write() method, plus a string holding its contents, and it gets written to the test directory right before starting.

This simple technique can be seen throughout most of the end-to-end tests as it was the original technique provided to test developers, but it is no longer the preferred way to write a new test. To develop this way, you first need to create the necessary files and get them to work, then convert them to an embedded string form, which may involve lots of extra escaping. These embedded files are then tricky to maintain. As a test grows multiple steps, it becomes less easy to read, since many if the embedded strings aren't quite the final files, and the volume of test code obscures the flow of the testing steps. Additionally, as SCons moves more to the use of automated code checkers and formatters to detect problems and keep a standard coding style for better readability, note that such tools don't look inside strings for code, so the effect is lost on them.

In testing parlance, a fixture is a repeatable test setup. The SCons test harness allows the use of saved files or directories to be used in that sense: the fixture for this test is foo, instead of writing a whole bunch of strings to create files. Since these setups can be reusable across multiple tests, the fixture terminology applies well.

Note: fixtures must not be treated by SCons as runnable tests. To exclude them, see instructions in the above section named Selecting tests.

Directory fixtures

The test harness method dir_fixture(srcdir, [dstdir]) copies the contents of the specified directory srcdir from the directory of the called test script to the current temporary test directory. The srcdir name may be a list, in which case the elements are concatenated into a path first. The optional dstdir is used as a destination path under the temporary working directory. distdir is created automatically, if it does not already exist.

If srcdir represents an absolute path, it is used as-is. Otherwise, if the harness was invoked with the environment variable FIXTURE_DIRS set (which runtest.py does by default), the test instance will present that list of directories to search as self.fixture_dirs, each of these are additionally searched for a directory with the name of srcdir.

A short syntax example:

test = TestSCons.TestSCons()
test.dir_fixture('image')
test.run()

would copy all files and subdirectories from the local image directory to the temporary directory for the current test, then run it.

To see a real example for this in action, refer to the test named test/packaging/convenience-functions/convenience-functions.py.

File fixtures

The method file_fixture(srcfile, [dstfile]) copies the file srcfile from the directory of the called script to the temporary test directory. The optional dstfile is used as a destination file name under the temporary working directory, unless it is an absolute path name. If dstfile includes directory elements, they are created automatically if they don't already exist. The srcfile and dstfile parameters may each be a list, which will be concatenated into a path.

If srcfile represents an absolute path, it is used as-is. Otherwise, any passed in fixture directories are used as additional places to search for the fixture file, as for the dir_fixture case.

With the following code:

test = TestSCons.TestSCons()
test.file_fixture('SConstruct')
test.file_fixture(['src', 'main.cpp'], ['src', 'main.cpp'])
test.run()

The files SConstruct and src/main.cpp are copied to the temporary test directory. Notice the second file_fixture call preserves the path of the original, otherwise main.cpp would have been placed in the top level of the test directory.

Again, a reference example can be found in the current revision of SCons, see test/packaging/sandbox-test/sandbox-test.py.

For even more examples you should check out one of the external Tools, e.g. the Qt5 Tool at https://github.com/SCons/scons-contrib/tree/master/sconscontrib/SCons/Tool/qt5. There are many other tools in the contrib repository, and you can also visit the SCons Tools Index at https://github.com/SCons/scons/wiki/ToolsIndex for a complete list of available Tools, though not all may have tests yet.

How to convert old tests to use fixtures

Tests using the inline TestSCons.write() method can fairly easily be converted to the fixture based approach. For this, we need to get at the files as they are written to each temporary test directory, which we can do by taking advantage of an existing debugging aid, namely that runtest.py checks for the existence of an environment variable named PRESERVE. If it is set to a non-zero value, the testing framework preserves the test directory instead of deleting it, and prints a message about its name to the screen.

So, you should be able to give the commands:

$ PRESERVE=1 python runtest.py test/packaging/sandbox-test.py

assuming Linux and a bash-like shell. For a Windows cmd shell, use set PRESERVE=1 (that will leave it set for the duration of the cmd session, unless manually cleared).

The output will then look something like this:

1/1 (100.00%) /usr/bin/python test/packaging/sandbox-test.py
PASSED
preserved directory /tmp/testcmd.4060.twlYNI

You can now copy the files from that directory to your new fixture directory. Then, in the test script you simply remove all the tedious TestSCons.write() statements and replace them with a single TestSCons.dir_fixture() call.

For more complex testing scenarios you can use file_fixture with the optional second argument (or the keyword arg dstfile) to assign a name to the file being copied. For example, some tests need to write multiple SConstruct files across the full run. These files can be given different names in the source (perhaps using a suffix to distinguish them), and then be successively copied to the final name as needed:

test.file_fixture('fixture/SConstruct.part1', 'SConstruct')
# more setup, then run test
test.file_fixture('fixture/SConstruct.part2', 'SConstruct')
# run new test

When not to use a fixture

Note that some files are not appropriate for use in a fixture as-is: fixture files should be static. If the creation of the file involves interpolating data discovered during the run of the test script, that process should stay in the script. Here is an example of this kind of usage that does not lend itself easily to a fixture:

import TestSCons
_python_ = TestSCons._python_

test.write('SConstruct', f"""
cc = Environment().Dictionary('CC')
env = Environment(
    LINK=r'{_python_} mylink.py',
    LINKFLAGS=[],
    CC=r'{_python_} mycc.py',
    CXX=cc,
    CXXFLAGS=[],
)
env.Program(target='test1', source='test1.c')
"""

Here the value of _python_ from the test program is pasted in via f-string formatting. A fixture would be hard to use here because we don't know the value of _python_ until runtime (also note that as it will be an absolute pathname, it's entered using Python raw string notation to avoid interpretation problems on Windows, where the path separator is a backslash).

The other files created in this test may still be candidates for use as fixture files, however.

Debugging end-to-end tests

The end-to-end tests are hand-crafted SCons projects, so testing involves running an instance of scons with those inputs. The tests treat the SCons invocation as a black box, usually looking for external effects of the test - targets are created, created files have expected contents, files properly removed on clean, etc. They often also look for the flow of messages from SCons.

Simple tricks like inserting print statements in the SCons code itself don't really help as they end up disrupting those external effects (e.g. test.run(stdout="Some text"), but with the print, stdout contains the extra print output and the result doesn't match).

Even more irritatingly, added text can cause other tests to fail and obscure the error you're looking for. Say you have three different tests in a script exercising different code paths for the same feature, and the third one is unexpectedly failing. You add some debug prints to the affected part of scons, and now the first test of the three starts failing, aborting the test run before it even gets to the third test - the one you were trying to debug.

Still, there are some techniques to help debugging.

The first step should be to run the tests so the harness emits more information, without forcing more information into the test stdout/stderr which will confuse result evaluation. runtest.py has several levels of verbosity which can be used for this purpose:

$ python runtest.py --verbose=2 test/foo.py

You can also use the internal SCons.Debug.Trace() function, which prints output to /dev/tty on Linux/UNIX systems and con on Windows systems, so you can see what's going on, but do not contribute to the captured stdout/stderr and mess up the test expectations.

If you do need to add informational messages in scons code to debug a problem, you can use logging and send the messages to a file instead, so they don't interrupt the test expectations. Or write directly to a trace file of your choosing.

Part of the technique discussed in the section How to convert old tests to use fixtures can also be helpful for debugging purposes. If you have a failing test, try:

$ PRESERVE=1 python runtest.py test/failing-test.py

You can now go to the save directory reported from this run and invoke scons manually (with appropriate arguments matching what the test did) to see the results without the presence of the test infrastructure which would otherwise consume output you may be interested in. In this case, adding debug prints may be more useful.

There are related variables PRESERVE_PASS, PRESERVE_FAIL and PRESERVE_NORESULT that preserve the directory only if the test result was the indicated one, which is helpful if you're trying to work with multiple tests showing an unusual result.

From a Windows cmd shell, you will have to set the environment variable first, it doesn't work on a single line like the example above for POSIX-style shells.

Test infrastructure

The main end-to-end test API is defined in the TestSCons class. TestSCons is a subclass of TestCommon, which is a subclass of TestCmd. TestSCons provides the help for using an instance of SCons during the run.

The unit tests do not run an instance of SCons separately, but instead import the modules of SCons that they intend to test. Those tests should use the TestCmd class - it is intended for runnable scripts.

Those classes are defined in Python files of the same name in testing/framework. Start in testing/framework/TestCmd.py for the base API definitions, like how to create files (test.write()) and run commands (test.run()).

The match functions work like this:

TestSCons.match_re

match each line with an RE

  • Splits the lines into a list (unless they already are)
  • splits the REs at newlines (unless already a list) and puts ^..$ around each
  • then each RE must match each line. This means there must be as many REs as lines.
TestSCons.match_re_dotall

match all the lines against a single RE

  • Joins the lines with newline (unless already a string)
  • joins the REs with newline (unless it's a string) and puts ^..$ around the whole thing
  • then whole thing must match with Python re.DOTALL.

Use them in a test like this:

test.run(..., match=TestSCons.match_re, ...)

or:

test.must_match(..., match=TestSCons.match_re, ...)

Often you want to supply arguments to SCons when it is invoked to run a test, which you can do using an arguments parameter:

test.run(arguments="-O -v FOO=BAR")

One caveat here: the way the parameter is processed is unavoidably different from typing on the command line - if you need it not to be split on spaces, pre-split it yourself, and pass the list, like:

test.run(arguments=["-f", "SConstruct2", "FOO=Two Words"])

Avoiding tests based on tool existence

For many tests, if the tool being tested is backed by an external program which is not installed on the machine under test, it may not be worth proceeding with the test. For example, it's hard to test compiling code with a C compiler if no C compiler exists. In this case, the test should be skipped.

Here's a simple example for end-to-end tests:

intelc = test.detect_tool('intelc', prog='icpc')
if not intelc:
    test.skip_test("Could not load 'intelc' Tool; skipping test(s).\n")

See testing/framework/TestSCons.py for the detect_tool() method. It calls the tool's generate() method, and then looks for the given program (tool name by default) in env['ENV']['PATH'].

The where_is() method can be used to look for programs that are do not have tool specifications. The existing test code will have many samples of using either or both of these to detect if it is worth even proceeding with a test.

For the unit tests, there are decorators for conditional skipping and other actions that will produce the correct output display and statistics in abnormal situations.

@unittest.skip(reason)
Unconditionally skip the decorated test. reason should describe why the test is being skipped.
@unittest.skipIf(condition, reason)
Skip the decorated test if condition is true.
@unittest.skipUnless(condition, reason)
Skip the decorated test unless condition is true.
@unittest.expectedFailure
Mark the test as an expected failure. If the test fails it will be considered a success. If the test passes, it will be considered a failure.

You can also directly call testcase.skipTest(reason).

Note that it is usually possible to test at least part of the operation of a tool without the underlying program. Tools are responsible for setting up construction variables and having the right builders, scanners and emitters plumbed into the environment. These things can be tested by mocking the behavior of the executable. Many examples of this can be found in the test directory. See for example test/subdivide.py.

Testing DOs and DONTs

There's no question that having to write tests in order to get a change approved - even an apparently trivial change - does make it a little harder to contribute to the SCons code base - but the requirement to have features and bugfixes testable is a necessary part of ensuring SCons quality. Thinking of SCons development in terms of the red/green model from Test Driven Development should make things a little easier.

If you are working on an SCons bug, try to come up with a simple reproducer first. Bug reports (even your own!) are often like I tried to do this but it surprisingly failed, and a reproducer is normally an SConstruct along with, probably, some supporting files such as source files, data files, subsidiary SConscripts, etc. Try to make this example as simple and clean as possible. No, this isn't necessarily easy to do, but winnowing down what triggers a problem and removing the stuff that doesn't actually contribute to triggering the problem it is a step that lets you (and later readers) more clearly understand what is going on. You don't have to turn this into a formal testcase yet, but keep this reproducer around, and document with it what you expect to happen, and what actually happens. This material will help produce an E2E test later, and this is something you may be able to get help with, if the way the tests are usually written and the test harness proves too confusing. With a clean test in hand (make sure it's failing!) you can go ahead an code up a fix and make sure it passes with the fix in place. Jumping straight to a fix without working on a testcase like this will often lead to a disappointing how do I come up with a test so the maintainer will be willing to merge phase. Asking questions on a public forum can be productive here.

E2E-specific Suggestions:

  • Do not require the use of an external tool unless necessary. Usually the SCons behavior is the thing we want to test, not the behavior of the external tool. Necessary is not a precise term - sometimes it would be too time-consuming to write a script to mock a compiler with an extensive set of options, and sometimes it's not a good idea to assume you know what all those will do vs what the real tool does; there may be other good reasons for just going ahead and calling the external tool.
  • If using an external tool, be prepared to skip the test if it is unavailable.
  • Do not combine tests that need an external tool with ones that do not - split these into separate test files. There is no concept of partial skip for e2e tests, so if you successfully complete seven of eight tests, and then come to a conditional "skip if tool missing" or "skip if on Windows", and that branch is taken, then the whole test file ends up skipped, and the seven that ran will never be recorded. Some tests follow the convention of creating a second test file with the ending -live for the part that requires actually running the external tool.
  • In testing, fail fast is not always the best policy - if you can think of many scenarios that could go wrong and they are all run linearly in a single test file, then you only hear about the first one that fails. In some cases it may make sense to split them out a bit more, so you can see several fails at once, which may show a helpful failure pattern you wouldn't spot from a single fail.
  • Use test fixtures where it makes sense, and in particular, try to make use of shareable mocked tools, which, by getting lots of use, will be better debugged (that is, don't have each test produce its own myfortan.py or mylex.py etc. unless they need drastically different behaviors).

Unittest-specific hints:

  • Let the unittest module help! Lots of the existing tests just use a bare assert call for checks, which works fine, but then you are responsible for preparing the message if it fails. The base TestCase class has methods which know how to display many things, for example self.assertEqual() displays in what way the two arguments differ if they are not equal. Checking for am expected exception can be done with self.assertRaises() rather than crafting a stub of code using a try block for this situation.
  • The fail fast consideration applies here, too: try not to fail a whole testcase on the first problem, if there are more checks to go. Again, existing tests may use elaborate tricks for this, but modern unittest has a subTest context manager that can be used to wrap each distinct piece and not abort the testcase for a failing subtest (to be fair, this functionality is a recent addition, after most SCons unit tests were written - but it should be used going forward).