Add some perf tests to compare Mojo overhead.

ipc/run_all_perftests.cc

Index: ipc/run_all_perftests.cc
diff --git a/ipc/run_all_perftests.cc b/ipc/run_all_perftests.cc
index a942b8b57496f6401f833bbea96b52168ceb8138..59865ec5f547c9b15063b2c9bf80c1f63b9c1a37 100644
--- a/ipc/run_all_perftests.cc
+++ b/ipc/run_all_perftests.cc
@@ -6,12 +6,21 @@
 
 #include "base/command_line.h"
 #include "base/test/perf_test_suite.h"
+#include "base/test/test_io_thread.h"
 #include "mojo/edk/embedder/embedder.h"
+#include "mojo/edk/test/scoped_ipc_support.h"
+#include "mojo/edk/test/test_support_impl.h"
 
 int main(int argc, char** argv) {
   base::PerfTestSuite test(argc, argv);
 
   mojo::edk::Init();
+  base::TestIOThread test_io_thread(base::TestIOThread::kAutoStart);
+  // Leak this because its destructor calls mojo::edk::ShutdownIPCSupport which
+  // really does nothing in the new EDK but does depend on the current message
+  // loop, which is destructed inside base::LaunchUnitTests.

[Ken Rockot(use gerrit already), 2017/01/06 18:15:07]

This is actually not true anymore. ShutdownIPCSupport does useful things,
particularly in multiprocess. It might not affect these tests (unclear), but it
doesn't matter: ScopedIPCSupport can safely be constructed on the stack here and
its destructor will work with or without a MessageLoop.

[jam, 2017/01/06 18:24:18]

ah ok, I had copied that from the other run_all_perftests files. I've confirmed
this isn't needed, so updated both.

+  new mojo::edk::test::ScopedIPCSupport(test_io_thread.task_runner());
+  mojo::test::TestSupport::Init(new mojo::edk::test::TestSupportImpl());
 
   return test.Run();
 }