tools/lldb-perf/README

 The lldb-perf infrastructure for LLDB performance testing
===========================================================

lldb-perf is an infrastructure meant to simplify the creation of performance tests for the LLDB debugger
It is contained in liblldbperf.a which is part of the standard opensource checkout of LLDB
Its main concepts are:
- Gauges: a gauge is a thing that takes a sample (e.g. it samples the time elapsed, the memory used, the energy consumed)
- Metrics: a metric is a collection of samples that knows how to do statistics (for now, sum() and average() but can be extended as need be)
- Measurements: a measurement is the thing that stores an action, a gauge and a metric. essentially, you define measurements as in “take the time to run this function”, “take the memory to run this block of code”, and then after you invoke it, your stats will be automagically there
- Tests: a test is a sequence of steps and measurements

Tests cases should be added as targets to the lldbperf.xcodeproj project.  It is probably easiest to duplicate one of the existing targets.

In order to write a test based on lldb-perf, you need to subclass lldb_perf::TestCase, as in:

using namespace lldb_perf;

class FormattersTest : public TestCase
{

Usually, you will define measurements as variables of your test case class:
private:
    // C++ formatters
    TimeMeasurement<std::function<void(SBValue)>> m_dump_std_vector_measurement;
    TimeMeasurement<std::function<void(SBValue)>> m_dump_std_list_measurement;
    TimeMeasurement<std::function<void(SBValue)>> m_dump_std_map_measurement;
    TimeMeasurement<std::function<void(SBValue)>> m_dump_std_string_measurement;

    // Cocoa formatters
    TimeMeasurement<std::function<void(SBValue)>> m_dump_nsstring_measurement;
    TimeMeasurement<std::function<void(SBValue)>> m_dump_nsarray_measurement;
    TimeMeasurement<std::function<void(SBValue)>> m_dump_nsdictionary_measurement;
    TimeMeasurement<std::function<void(SBValue)>> m_dump_nsset_measurement;
    TimeMeasurement<std::function<void(SBValue)>> m_dump_nsbundle_measurement;
    TimeMeasurement<std::function<void(SBValue)>> m_dump_nsdate_measurement;

A TimeMeasurement is, obviously, a class that measures “how much time to run this block of code”. the block of code is passed as an std::function (you can construct with a lambda!) - you need however to give the prototype of your block of code. in this example, we run blocks of code that take an SBValue and return nothing. other things are possible.
These blocks look like:
        m_dump_std_vector_measurement = CreateTimeMeasurement([] (SBValue value) -> void {
            lldb_perf::Xcode::FetchVariable (value,1,false);
        }, "std-vector", "time to dump an std::vector");
Here we are saying: make me a measurement named “std-vector”, whose description is “time to dump an std::vector” and that takes the time required to call lldb_perf::Xcode::FetchVariable(value,1,false)

The Xcode class is a collection of utility functions that replicate common Xcode patterns (FetchVariable unsurprisingly calls API functions that Xcode could use when populating a variables view entry - the 1 means “expand 1 level of depth” and the 
false means “do not dump the data to stdout”)

A full constructor for a TestCase looks like:
    FormattersTest () : TestCase()
    {
        m_dump_std_vector_measurement = CreateTimeMeasurement([] (SBValue value) -> void {
            lldb_perf::Xcode::FetchVariable (value,1,false);
        }, "std-vector", "time to dump an std::vector");
        m_dump_std_list_measurement = CreateTimeMeasurement([] (SBValue value) -> void {
            lldb_perf::Xcode::FetchVariable (value,1,false);
        }, "std-list", "time to dump an std::list");
        m_dump_std_map_measurement = CreateTimeMeasurement([] (SBValue value) -> void {
            lldb_perf::Xcode::FetchVariable (value,1,false);
        }, "std-map", "time to dump an std::map");
        m_dump_std_string_measurement = CreateTimeMeasurement([] (SBValue value) -> void {
            lldb_perf::Xcode::FetchVariable (value,1,false);
        }, "std-string", "time to dump an std::string");
        
        m_dump_nsstring_measurement = CreateTimeMeasurement([] (SBValue value) -> void {
            lldb_perf::Xcode::FetchVariable (value,0,false);
        }, "ns-string", "time to dump an NSString");
        
        m_dump_nsarray_measurement = CreateTimeMeasurement([] (SBValue value) -> void {
            lldb_perf::Xcode::FetchVariable (value,1,false);
        }, "ns-array", "time to dump an NSArray");
        
        m_dump_nsdictionary_measurement = CreateTimeMeasurement([] (SBValue value) -> void {
            lldb_perf::Xcode::FetchVariable (value,1,false);
        }, "ns-dictionary", "time to dump an NSDictionary");
        
        m_dump_nsset_measurement = CreateTimeMeasurement([] (SBValue value) -> void {
            lldb_perf::Xcode::FetchVariable (value,1,false);
        }, "ns-set", "time to dump an NSSet");
        
        m_dump_nsbundle_measurement = CreateTimeMeasurement([] (SBValue value) -> void {
            lldb_perf::Xcode::FetchVariable (value,1,false);
        }, "ns-bundle", "time to dump an NSBundle");
        
        m_dump_nsdate_measurement = CreateTimeMeasurement([] (SBValue value) -> void {
            lldb_perf::Xcode::FetchVariable (value,0,false);
        }, "ns-date", "time to dump an NSDate");
    }

Once your test case is constructed, Setup() is called on it:
    virtual void
	Setup (int argc, const char** argv)
    {
        m_app_path.assign(argv[1]);
        m_out_path.assign(argv[2]);
        m_target = m_debugger.CreateTarget(m_app_path.c_str());
        m_target.BreakpointCreateByName("main");
        
        Launch (NULL,".");
    }
(I am considering moving Setup to the constructor, to be honest, but have not done so yet because it might also make sense to have Setup() return a bool or an Error object. stay tuned for more here!)
In Setup() you process command-line arguments and create a target to run
The last thing you want to do in setup is call Launch():
	bool
	Launch (const char** args, const char* cwd);
This ensures your target is now alive. Make sure to have a breakpoint created :)
Once you launched, the event loop is entered. The event loop waits for stops, and when it gets one, it calls your test case’s 
    virtual void
	TestStep (int counter, ActionWanted &next_action)

the counter is the step id (a monotonically increasing counter). In TestStep() you will essentially run your measurements and then return what you want the driver to do.
Possible options are:
- continue process          ActionWanted::Continue()
- kill process              ActionWanted::Kill()
- finish on a thread        ActionWanted::Finish(SBThread)
- step-over on a thread.    ActionWanted::Next(SBThread)

If you use ActionWanted::Next() or ActionWanted::Finish(), define a SelectMyThread() that gives you an SBThread. I usually use a filename as key:
    SBThread
	SelectMyThread (const char* file_name)
	{
		auto threads_count = m_process.GetNumThreads();
		for (auto thread_num = 0; thread_num < threads_count; thread_num++)
		{
			SBThread thread(m_process.GetThreadAtIndex(thread_num));
			auto local_file_name = thread.GetFrameAtIndex(0).GetCompileUnit().GetFileSpec().GetFilename();
			if (!local_file_name)
				continue;
			if (strcmp(local_file_name,file_name))
				continue;
			return thread;
		}
		Xcode::RunCommand(m_debugger,"bt all",true);
		assert(false);
	}
you might want to use some other logic in your selection.
For your convenience TestCase has m_debugger, m_target and m_process as member variables. There is also an m_thread, but that is for you to set if you need/want to use it!

An example:
    virtual void
	TestStep (int counter, ActionWanted &next_action)
    {
        case 0:
            m_target.BreakpointCreateByLocation("fmts_tester.mm", 68);
            next_action.Continue();
            break;
        case 1:
            DoTest ();
            next_action.Continue();
            break;
        case 2:
            DoTest ();
            next_action.Continue();
            break;

DoTest() is a function I define in my own class that calls the measurements:
    void
    DoTest ()
    {
        SBThread thread_main(SelectMyThread("fmts_tester.mm"));
        SBFrame frame_zero(thread_main.GetFrameAtIndex(0));
        
        m_dump_nsarray_measurement(frame_zero.FindVariable("nsarray", lldb::eDynamicCanRunTarget));
        m_dump_nsarray_measurement(frame_zero.FindVariable("nsmutablearray", lldb::eDynamicCanRunTarget));

        m_dump_nsdictionary_measurement(frame_zero.FindVariable("nsdictionary", lldb::eDynamicCanRunTarget));
        m_dump_nsdictionary_measurement(frame_zero.FindVariable("nsmutabledictionary", lldb::eDynamicCanRunTarget));
        
        m_dump_nsstring_measurement(frame_zero.FindVariable("str0", lldb::eDynamicCanRunTarget));
        m_dump_nsstring_measurement(frame_zero.FindVariable("str1", lldb::eDynamicCanRunTarget));
        m_dump_nsstring_measurement(frame_zero.FindVariable("str2", lldb::eDynamicCanRunTarget));
        m_dump_nsstring_measurement(frame_zero.FindVariable("str3", lldb::eDynamicCanRunTarget));
        m_dump_nsstring_measurement(frame_zero.FindVariable("str4", lldb::eDynamicCanRunTarget));
        
        m_dump_nsdate_measurement(frame_zero.FindVariable("me", lldb::eDynamicCanRunTarget));
        m_dump_nsdate_measurement(frame_zero.FindVariable("cutie", lldb::eDynamicCanRunTarget));
        m_dump_nsdate_measurement(frame_zero.FindVariable("mom", lldb::eDynamicCanRunTarget));
        m_dump_nsdate_measurement(frame_zero.FindVariable("dad", lldb::eDynamicCanRunTarget));
        m_dump_nsdate_measurement(frame_zero.FindVariable("today", lldb::eDynamicCanRunTarget));
        
        m_dump_nsbundle_measurement(frame_zero.FindVariable("bundles", lldb::eDynamicCanRunTarget));
        m_dump_nsbundle_measurement(frame_zero.FindVariable("frameworks", lldb::eDynamicCanRunTarget));
        
        m_dump_nsset_measurement(frame_zero.FindVariable("nsset", lldb::eDynamicCanRunTarget));
        m_dump_nsset_measurement(frame_zero.FindVariable("nsmutableset", lldb::eDynamicCanRunTarget));
        
        m_dump_std_vector_measurement(frame_zero.FindVariable("vector", lldb::eDynamicCanRunTarget));
        m_dump_std_list_measurement(frame_zero.FindVariable("list", lldb::eDynamicCanRunTarget));
        m_dump_std_map_measurement(frame_zero.FindVariable("map", lldb::eDynamicCanRunTarget));

        m_dump_std_string_measurement(frame_zero.FindVariable("sstr0", lldb::eDynamicCanRunTarget));
        m_dump_std_string_measurement(frame_zero.FindVariable("sstr1", lldb::eDynamicCanRunTarget));
        m_dump_std_string_measurement(frame_zero.FindVariable("sstr2", lldb::eDynamicCanRunTarget));
        m_dump_std_string_measurement(frame_zero.FindVariable("sstr3", lldb::eDynamicCanRunTarget));
        m_dump_std_string_measurement(frame_zero.FindVariable("sstr4", lldb::eDynamicCanRunTarget));
    }

Essentially, you call your measurements as if they were functions, passing them arguments and all, and they will do the right thing with gathering stats.
The last step is usually to KILL the inferior and bail out:
    virtual ActionWanted
	TestStep (int counter)
    {
...     
        case 9:
            DoTest ();
            next_action.Continue();
            break;
        case 10:
            DoTest ();
            next_action.Continue();
            break;
        default:
            next_action.Kill();
            break;
    }


At the end, you define a Results() function:
    void
    Results ()
    {
        CFCMutableArray array;
        m_dump_std_vector_measurement.Write(array);
        m_dump_std_list_measurement.Write(array);
        m_dump_std_map_measurement.Write(array);
        m_dump_std_string_measurement.Write(array);

        m_dump_nsstring_measurement.Write(array);
        m_dump_nsarray_measurement.Write(array);
        m_dump_nsdictionary_measurement.Write(array);
        m_dump_nsset_measurement.Write(array);
        m_dump_nsbundle_measurement.Write(array);
        m_dump_nsdate_measurement.Write(array);

        CFDataRef xmlData = CFPropertyListCreateData(kCFAllocatorDefault, array.get(), kCFPropertyListXMLFormat_v1_0, 0, NULL);
        
        CFURLRef file = CFURLCreateFromFileSystemRepresentation(NULL, (const UInt8*)m_out_path.c_str(), m_out_path.size(), FALSE);
        
        CFURLWriteDataAndPropertiesToResource(file,xmlData,NULL,NULL);
    }

For now, pretty much copy this and just call Write() on all your measurements. I plan to move this higher in the hierarchy (e.g. make a TestCase::Write(filename) fairly soon).

Your main() will look like:
int main(int argc, const char * argv[])
{
    MyTest test;
    TestCase::Run(test,argc,argv);
    return 0;
}

If you are debugging your test, before Run() call
    test.SetVerbose(true);

Feel free to send any questions and ideas for improvements.