Run media_unittests on android test bots.

media/base/run_all_unittests.cc

Index: media/base/run_all_unittests.cc
diff --git a/media/base/run_all_unittests.cc b/media/base/run_all_unittests.cc
index 4dc422450c8634f3fd164e28bd3dcb172c81d2d4..d4d92b6111a8f27dd1d7c68c6501cee618828f81 100644
--- a/media/base/run_all_unittests.cc
+++ b/media/base/run_all_unittests.cc
@@ -26,7 +26,13 @@ int main(int argc, char** argv) {
       logging::DONT_LOCK_LOG_FILE,
       logging::APPEND_TO_OLD_LOG_FILE,
       logging::ENABLE_DCHECK_FOR_NON_OFFICIAL_RELEASE_BUILDS));
+
+  // On Android, AtExitManager is created in
+  // testing/android/native_test_wrapper.cc before main() is called.

[Ami GONE FROM CHROMIUM, 2012/07/24 23:37:35]

Do you mean native_test_launcher.cc?

Is this pattern really the right way to go?  ISTM you'll end up having to have
an equivalent of this code in every unittests target.  Why not instead omit the
AEM from n_t_l.cc and let the main()'s continue to own their AtExitManagers?

[nilesh, 2012/07/25 00:19:45]

Yes. Corrected.
There is code executed in n_t_l.cc:RunTests (before main in called) which needs
an AtExitManager. Unfortunately this means that we need this #ifdef in all
unittest targets. I don't have a cleaner solution.

[nilesh, 2012/07/25 00:23:11]

I can look into having multiple AtExitManager not result in a crash. The DCHECK
in AtExitManager is skipped for COMPONENET_BUILD.

[Ami GONE FROM CHROMIUM, 2012/07/25 17:09:59]

Can you make the pre-main() code not need an AEM?  Or better yet not execute
code pre-main?  For example, does it make sense to have a
base::InitializeAndroidForTesting() mimicking
media::InitializeMediaLibraryForTesting() below?
This would still require modifying each test target's main(), but at least it'd
be obviouser why, and you wouldn't run into this sort of double-AEM-init
problem.

My real worry about this approach is that it sets the precedent that each main()
needs to account for execution outside of itself, when pretty much every main()
ever written is written with the assumption that the process starts at the top
of main().
As an example of this assumption, the CommandLine::Init() above doesn't inspect
its return code (and is a no-op when n_t_l.cc calls CL::Init() itself).

[nilesh, 2012/07/25 21:56:32]

I take back my comment that we need similar code in all unittest targets. Most
of these can reply on the exitmanager being created (or not) in
base/test/test_suite.cc
Like what I have tried in this CL. 
AtExitManager is needed because we need to setup the global application context
(in base::android::InitApplicationContext).   g_application_context in
jni_android.cc is a LazyInstance and it needs an exitmanager.
This approach is cleaner, but the issue with this is that we can't pass
arguments to main. The java side provides an application context which needs to
be set. We can export another method (not main) which can take arguments.

A hacky way would be to make base::InitializeAndroidForTesting() call back
statically into JAVA to get the application context.

+  // The same thing is also done in base/test/test_suite.cc

[Ami GONE FROM CHROMIUM, 2012/07/24 23:37:35]

This line isn't useful to commit IMO.

[nilesh, 2012/07/25 00:19:45]

Removed.

+#if !defined(OS_ANDROID)
   base::AtExitManager exit_manager;
+#endif
 
   media::InitializeMediaLibraryForTesting();