llvm-project/lldb/tools/lldb-perf

 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. Samples include elapsed time,
  memory used, and energy consumed.
- Metrics: a metric is a collection of samples that knows how to do statistics
  like sum() and average(). Metrics can be extended as needed.
- Measurements: a measurement is the thing that stores an action, a gauge and
  a metric. 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 automagically be 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:

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 which you
can construct with a lambda! You need to give the prototype of your block of
code. In this example, we run blocks of code that take an SBValue and return
nothing.

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 bool
	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");
        SBLaunchInfo launch_info (argv);
        return Launch (launch_info);
    }
    
Setup() returns a boolean value that indicates if setup was successful.
In Setup() you fill out a SBLaunchInfo with any needed settings for launching
your process like arguments, environment variables, working directory, and
much more.

The last thing you want to do in setup is call Launch():

	bool
	Launch (coSBLaunchInfo &launch_info);

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 TestStep() function:

    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 by filling in the ActionWanted object named "next_action".

Possible options are:
- continue process          next_action.Continue();
- kill process              next_action.Kill();
- Step-out on a thread      next_action.StepOut(SBThread)
- step-over on a thread.    next_action.StepOver(SBThread)

If you use ActionWanted::Next() or ActionWanted::Finish() you need to specify
a thread to use. By default the TestCase class will select the first thread
that had a stop reason other than eStopReasonNone and place it into the 
m_thread member variable of TestCase. This means if your test case hits a
breakpoint or steps, the thread that hit the breakpoint or finished the step
will automatically be selected in the process (m_process) and m_thread will
be set to this thread. If you have one or more threads that will stop with a
reason simultaneously, you will need to find those threads manually by 
iterating through the process list and determine what to do next.

For your convenience TestCase has m_debugger, m_target and m_process as member
variables. As state above m_thread will be filled in with the first thread 
that has a stop reason.

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.StepOver(m_thread);
            break;

DoTest() is a function I define in my own class that calls the measurements:
    void
    DoTest ()
    {
        SBThread thread_main(m_thread);
        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.