This tutorial describes how to produce test-coverage metrics for C/C++ projects. We start by instrumenting an autotools build to produce gcov output with gcc; we then generate test coverage artifacts by running our test suite, and finally we explore our results using lcov.

Introduction

We’d like to assess the quality of the existing test suite for each Product Strategy project. A measurement of test coverage will tell us what part of the project’s code is “covered” or exercised by its tests. 50% is a start; 80% would be a good goal for a neglected project; one rarely encounters 100% test coverage, but we’d like to get as close as we can. Initially we’ll use these findings to gain an overview of the test-quality of each project; ultimately these metrics will guide the improvement of our codebase, and enable us to monitor our progress using Jenkins and associated open-source tools.

A Three-Part Process in Several Steps

We’ll enable test-coverage of a particular C/C++ project in a three-part process:

• enabling a special build
• running the tests
• studying the output

Our first step will be to enable a special build. Ultimately this just means adding a few flags to our gcc invocation, but it never seems so straightforward with autotools projects ;) . The Product Strategy Quality team has made available a set of files to facilitate this build–please check these out now if you’d like to follow along.

bzr co lp:coverage-tutorial

Inspecting this archive you’ll find

• gcov.m4: an autoconf macro which will check for relevant tools
• Makefile.am.coverage, which includes our coverage-enabled automake targets
• an old revision of dbus-test-runner
• a copy of this tutorial

Before we start let’s make sure we have lcov installed. lcov incoporates the GNU tool gcov and produces pretty HTML reports of our coverage results.

sudo apt-get lcov

(Note that the Debian lcov package includes genhtml.) We’re ready to begin.

1. Branch a project of interest

I heard you like testing, so for this example we’ll test dbus-test-runner which is a runner which runs tests against the dbus messaging system. An old revision is included in our coverage-tutorial archive; if you want to start fresh (or on a different project), you would

bzr branch lp:dbus-test-runner

2. Install the gcov.m4 autoconf macro and invoke it

Open gcov.m4: here’s where we’re defining the compiler flags and checking for necessary tools. Let’s put this file where autoconf can find it.

Your project may already have a directory for m4 macros:

$ grep AC_CONFIG_MACRO_DIR *
configure.ac:AC_CONFIG_MACRO_DIR([m4])

If you find such a directive, simply copy the gcov.m4 file into the named directory. If not, create an m4 directory and copy the gcov.m4 file there–and don’t forget to include a reference to the directory in the body of configure.ac (it’ll look like the grep result above).

3. Massage configure.ac to include our necessary coverage flags.

This is the essential move of our special build. gcc supports coverage reporting with the addition of a few compiler flags. Here they are in the gcov.m4 file:

# Remove all optimization flags from CFLAGS
changequote({,})
CFLAGS=`echo "$CFLAGS" | $SED -e 's/-O[0-9]*//g'`
changequote([,])
# Add the special gcc flags
COVERAGE_CFLAGS="-O0 -fprofile-arcs -ftest-coverage"
COVERAGE_CXXFLAGS="-O0 -fprofile-arcs -ftest-coverage"
COVERAGE_LDFLAGS="-lgcov"

When the tests are run, --fprofile-arcs produces a tally of the execution of each arc of the code–one .gcda file for each source file. The --ftest-coverage flag produces .gcno files, which link an arc to a source line so that we can see which lines were touched. We’ll watch these files being produced in a little bit; you can read about the flags at length in the GNU documentation. Note that we also remove optimization with the -O0 flag to get more precise results.

Now here’s the step which requires knowledge both of your code and little bit of autoconf. We’ve included the flags above (in step 3), but we’ll now need to edit our configure.ac to make the flags available to our build. For dbus-test-runner the diff looks like this:

=== modified file configure.ac
--- configure.ac 2010-12-08 02:35:12 +0000
+++ configure.ac 2011-12-06 21:42:04 +0000
@@ -45,6 +45,16 @@
 AM_GLIB_GNU_GETTEXT
 ###########################
+# gcov coverage reporting
+###########################
+
+m4_include([m4/gcov.m4])
+AC_TDD_GCOV
+AC_SUBST(COVERAGE_CFLAGS)
+AC_SUBST(COVERAGE_CXXFLAGS)
+AC_SUBST(COVERAGE_LDFLAGS)
+
+###########################
 # Files #
 ##########################

And then having added these flags to the build process, we need to actually actually add them to the build proper:

=== modified file src/Makefile.am
--- src/Makefile.am 2009-12-07 21:00:43 +0000
+++ src/Makefile.am 2011-12-06 21:42:04 +0000
@@ -3,6 +3,8 @@
 dbus_test_runner_SOURCES = dbus-test-runner.c
 dbus_test_runner_CFLAGS = $(DBUS_TEST_RUNNER_CFLAGS) \
+ $(COVERAGE_CFLAGS) \
-DDEFAULT_SESSION_CONF="\"$(datadir)/dbus-test-runner/session.conf\"" \
 -Wall -Werror
 dbus_test_runner_LDADD = $(DBUS_TEST_RUNNER_LIBS)
+dbus_test_runner_LDFLAGS = $(COVERAGE_LDFLAGS)

Here we risk running afoul of the autotools “ancient ones”. This is an uncomplicated project–if you’re not getting the results you want in the next step, I can recommend The Goat Book as a decent tutorial on how these macros work.

4. Verify that autoconf.sh --enable-gcov generates .gcno files

Having enabled our special build with the gcc compiler flags, let’s verify that autoconf is generating the .gcno files we’ve asked for:

$ ./autogen.sh --enable-gcov
...
$ find -name *.gcno
./src/dbus_test_runner-dbus-test-runner.gcno

If all goes well we’ll find a .gcno for each .o file compiled. If not, then our flags aren’t being respected–maybe something’s wrong with our configure.ac addition in step 3. It’s important to have these checkpoints along the way to divide the autotools process into testable pieces.

5. Install Makefile.am.coverage

So now that we have the instrumentation we need, we’ll use a special build to run the tests. Copy Makefile.am.coverage into the top-level directory.

cp ../Makefile.am.coverage .

Inspection shows that this file defines some extra make targets to generate coverage results using lcov. We’ll alert automake to this by simply including the file–here’s the diff for dbus-test-runner:

=== modified file Makefile.am
--- Makefile.am 2009-04-23 21:19:56 +0000
+++ Makefile.am 2011-12-19 18:00:38 +0000
@@ -1,1 +1,3 @@
 SUBDIRS = data src tests po
+
+include $(top_srcdir)/Makefile.am.coverage

6. Verify that make check generates .gcda files

Now we’ll actually run the tests.

$ make coverage-html
...
$ find -name *.gcda
./src/dbus_test_runner-dbus-test-runner.gcda

Again these .gcda files represent a tally of ‘arcs’ through the code. Note that we can generate these files not just while running tests, but also during normal execution–you can see the kinship with profiling. lcov (via gcov) has used these tallies with our .gcno files to produce line-by-line results which we’ll study in a moment. Here’s what the test coverage looks like for dbus-test-runner:

Overall coverage rate:
lines......: 78.8% (215 of 273 lines)
functions..: 86.4% (19 of 22 functions)
branches...: 61.2% (60 of 98 branches)

For a small project these figures show that we have a good test-suite to build on. We’re curious about what we’re missing, though, so we’ll investigate below. Before you register a comment about “no coverage results”, recognize that you may actually have zero code coverage. But at least you’re able to measure it, riiight?

7. Have a look at the lcov pages

lcov has produced a coverage-results directory at the top-level of our project; open index.html with a web browser and explore.

At a glance our line coverage is 78.8%, and our “test-coverage progress bar” is yellow. Be aware of the difference between the offered metrics:

• line coverage: how many lines have our tests touched
• function coverage: how many functions have our tests touched
• branch coverage: for the graph which describes all possible paths of control through this file–especially through conditionals, e.g.–what percentage has our tests touched

In our opinion the killer feature of the lcov output is the source-file display, which shows which lines weren’t touched by tests. Drill into the source directory to see the results for a particular file.

Here in dbus-test-runner.c line 93 we’ve hit the line that tests the status of G_IO_STATUS_NORMAL 61 times, but never taken the path in the line below, for which G_IO_STATUS_NORMAL is false.

Conclusion

We’ve presented an introduction to test coverage, featuring the gcc toolchain and autotools–as demonstrated with a set of utilities we’ve found useful here in the Quality group.

I hope it’s obvious that this tutorial is just the beginning–not just of applying these methods to our code (we and you), but also of understanding what test coverage means to the improvement of Quality. Our team has witnessed that having an observable measure like the lcov green bars encourages developers to increase test coverage. However while chasing into a particularly narrow corner of their code, our developers have produced some confounding results which have caused us to doubt the accuracy of the available tools–I hope that Thomas Voss will follow with a post on some of his findings in the future.

Regardless, test coverage will be an important metric for our group for the coming cycles, and the bars will continue to grow greener as Spring approaches. . . . Meanwhile I’ll be interested to learn about your autotools/gcc/coverage experiences:

• What lcov alternatives have you tried?
• Which hallowed autotools rituals have I profaned in preparing this tutorial?
• What’s the highest coverage percentage you’ve witnessed in a GPL project?

Last week was my first one at Canonical and I spent it in Budapest where Ubuntu Engineering and Product Strategy teams had their rally. This is the first company where I was asked to travel on my first day and it did feel strange at the beginning :) After being a Canonical employee for 1+ week I think it is the best way how to start in a new job.

When talking about quality these are probably the key observations I have made from my new perspective of a Product Strategy Quality Lead:

With Canonical’s growing engagement with test driven development (TDD), we recently started introducing Google Test to a number of our upstream projects. Google Test is a unit testing framework targeted towards C & C++. It is lightweight and offers a load of features that ease unit testing tremendously. Most importantly, it supports automatic test case registration and the notions of test suite environments and fixtures. Here, a developer’s efficiency is our main focus. We acknowledge that providing a decent test coverage requires time and effort and we would like to free a developer from writing boiler plate code for test case setup, teardown and for acutally assembling his test cases.

Recently, we started working on an extension to Google test that introduces a dummy (read:headless) X server environment by means of a custom Google test environment and custom Google test fixtures. In summary, the extension starts and stops a dummy xserver and makes sure that a connection to the server is established correctly as part of the test setup procedure. On tear down, the connection to the X server is terminated before the server instance is stopped. Jump over to launchpad for the actual project and take a look at utouch-frame for a real-world application of the extension.

Test Case Result Integration with Jenkins CI

Continuous integration is a hot topic at Canonical and automatic execution of unit test suites is an integral part of our development model. For this reason, we store all unit test results in Google test’s XML dialect and export it to Jenkins to provide user, developers and QA with a summary of a project’s build status.

Conclusions & Outlook

With the introduction of Google test & Jenkins CI to Canonical’s testing landscape we want to provide the plumbing layer for a TDD environment that is loved by developers. It is the first step towards supporting engineers with meaningful metrics to further understand our software ecosystem and identify blank spots on our software map. Thus, stay tuned for Allan Lesage‘s work on code coverage calculation going to land in the not too distant future :-)