[PATCH 00/12] Support gcov instrumentation

[Chris Johns]

On 1/7/2022 3:07 pm, Sebastian Huber wrote:
> On 01.07.22 03:19, Chris Johns wrote:
>> On 29/6/2022 4:30 pm, Sebastian Huber wrote:
>>> This patch set adds support to build the RTEMS libraries with gcov
>>> instrumentation to get code and branch coverage.  There are some improvements
>>> necessary in the build system to support different compiler flags for libraries
>>> and tests.  In general, code coverage is enabled by the new RTEMS_COVERAGE
>>> build configuration option.
>>
>> Does RTEMS_COVERAGE effect covoar and the coverage testing that tool and flow
>> provides?
> 
> The covoar uses uninstrumented code, so these two approaches are orthogonal. You
> could even use both approaches at the same time. This makes little sense, but
> you could do it.

Agreed, it does not make sense. Someone new to RTEMS would not know what this
all means and may think they need to enable the option for any type of coverage.

>> If this enable is specific to GCC's gcov then maybe we should clearly label it
>> as that and avoid overloading the term and any possible confusion that might
>> bring. For example RTEMS_GCOV_CONVERGE?
> 
> It is not per se specific to GCC. Maybe we should name it
> RTEMS_COVERAGE_INSTRUMENTATION.

What if another approach to instrumented coverage is developed?

Why the reluctance to RTEMS_GCOV_COVERAGE? I like it because it is direct.

>>> The user can fine tune the build through the
>>> LIBRARY_OPTIMIZATION_FLAGS, TEST_OPTIMIZATION_FLAGS, and
>>> COVERAGE_COMPILER_FLAGS,
>>> COVERAGE_LINKER_FLAGS options.
>>
>> How would the LTO fat options be added? Would the pattern for
>> COVERAGE_COMPILER_FLAGS flags be copied to add LTO_COMPILER_FLAGS?
> 
> Do you have to specify the compiler flags also during linking if you use LTO?

You add `-flto -ffat-lto-objects` to build the object files so they then contain
the data the lto pass in the linker needs to optimise across the object files.

I am encouraged by some of the generated code I have reviewed. I like what it does.

>> I am a little lost with the difference between LIBRARY_OPTIMIZATION_FLAGS and
>> OPTIMIZATION_FLAGS? RTEMS is always a library so what is difference between
>> them? Are the OPTIMIZATION_FLAGS the "-O" ones and still something I can set
>> from an INI file?
> 
> Currently the OPTIMIZATION_FLAGS are used for all code (code for libraries and
> code for tests).  With this change you can build the libraries with -O0 and the
> tests with -O2 for example.

Oh so if I understand this the LIBRARY_OPTIMIZATION_FLAGS are per library?

>> I think the changes to cppflags, cflags and cxxflags are nice and a great
>> progression.
>>
>>>   The gcov information is dumped after each test
>>> case in a base64 encoded gcfn and gcda data stream.  It looks like this:
>>>
>>> *** BEGIN OF GCOV INFO BASE64 ***
>>> bmZjZyAxMkJSAAAAL3RtcC9zaC9iLXJ0ZW1zL2FybS94aWxpbnhfenlucV9hOV9xZW11L2NwdWtp
>>> dC9saWJjc3VwcG9ydC9zcmMvX19nZXR0b2QuYy42NS5nY2RhAGFkY2cgMTJCPoT6qo6mcBUAAAAB
>>> DAAAAOOrF1pxp0aIPrK7wAAAoQHw////AAAAAQwAAADTfDdNQYCYwJ9ZRL0AAKEBIAAAAAQAAAAA
>>> ...
>>> srvAAAChAfD///8AAAABDAAAAMXOxS0Rhzqx6Old2wAAoQH4////AAAAAQwAAACYh8U9/rgJ5hMc
>>> ig8AAKEB2P///wAAAAEMAAAAN3+9YAwMW8gTHIoPAAChAdj///8AAAABDAAAADAwMx4TuamFPrK7
>>> wAAAoQHw////AAAAAQwAAACvassJzDNWd/lKs3wAAKEB8P///wAAAAA=
>>> *** END OF GCOV INFO BASE64 ***
>>>
>>> On the host you can decode the block and pipe it through the gcov-tool to
>>> produce the *.gcda files.  Example Python code:
>>>
>>> block = some regex stuff to get the block between the BEGIN/END OF ...
>>> stream = base64.b64decode(block)
>>> subprocess.run(["arm-rtems6-gcov-tool", "merge-stream"], input=stream)
>>
>> Nice. Could  rtems-test detect this and then run that command if available?
> 
> Yes, this could be an option. Another is to produce a JSON report and then run a
> second script.

Sure.

Chris