Runtime functions in V8

[dmonji]

1) In v8, ic.cc, i see functions like  RUNTIME_FUNCTION(Runtime_KeyedLoadIC_Miss). 

    I don't understand how are they invoked. Is there some connection with   __ TailCallRuntime(Runtime::kKeyedLoadIC_Miss, arg_count, 1);

    from ic-<arch>.cc? 

2) what is the difference between Runtime_KeyedLoadIC_Miss and Runtime_LoadIC_Miss ?

3)  Can I profile these runtime functions using some c++ profiler? I want to see how much time is spent on subset of these runtime functions.

     

[Jakob Kummerow]

On Thu, Sep 24, 2015 at 1:58 PM, dmonji <monika...@gmail.com> wrote:

1) In v8, ic.cc, i see functions like  RUNTIME_FUNCTION(Runtime_KeyedLoadIC_Miss). 

    I don't understand how are they invoked. Is there some connection with   __ TailCallRuntime(Runtime::kKeyedLoadIC_Miss, arg_count, 1);

    from ic-<arch>.cc? 

Yes, that's precisely how they're invoked.

 

2) what is the difference between Runtime_KeyedLoadIC_Miss and Runtime_LoadIC_Miss ?

Essentially, foo.bar is a load and foo[bar] is a keyed load.

 

3)  Can I profile these runtime functions using some c++ profiler? I want to see how much time is spent on subset of these runtime functions.

Sure, C++ profilers can profile C++ functions. 

However, profiling short-running functions is notoriously hard because of the overhead that profiling introduces.

For example, V8's timeline plotter reports time spent handling IC misses; however getting the required timestamps is so expensive compared to what the functions actually do that it tends to over-report them by about 2x.

The sampling profiler's output has a section with "C++ entry points", which include IC miss handlers among other things. It has less (but still non-zero) overhead, and can be rather flaky due to the inherent randomness of sampling.