[LLVMdev] code coverage instrumentation

[Justin Bogner]

"Moshtaghi, Alireza" <Alireza.Moshtaghi at netapp.com> writes:
> Hi 
> Not sure if this is a clang or llvm related question so I’m sending to both
> mailing lists.
> Anyways, I have few questions regarding size and execution time of
> instrumented code:
> We are trying to run code coverage on memory limited hardware and
> investigating both (generating gcov output using —coverage and the llvm’s own
> way using -fprofile-instr-generate -fcoverage-mapping clang flags) In my
> questions I refer to the two methods as llgcov-way and llcovprof-way
> respectively. 
>
> Q1- How come size of instrumented code in llgcov-way turns bigger than
> llcovprof-way? I would imagine the other way because mapping information would
> go to *.gcno files.

I haven't measured, but I can make a guess. The gcov way isn't very
smart about minimizing the number of counters needed, so it tends to
grow the text section of the instrumented binary by quite a bit. More
counters also means that the counters get in the way of optimizations
more, so it's harder to reduce the binary size.

> Q2- The instrumented executable size of both llgcov-way and llcovprof-way
> (with and without optimization) is bloated 2x to 10x or in some cases 50x
> depending on the program. Here is output of size command for the variations on
> a simple test program that I wrote:
>    text    data     bss     dec     hex filename
>    5625     700    1696    8021    1f55 simpletest.-O0-g.llcovprof.
>   12838     776    1808   15422    3c3e simpletest.-O0-g.llgcov.
>    1481     492    1616    3589     e05 simpletest.-O0.none
>    5337     700    1696    7733    1e35 simpletest.-O1-g.llcovprof.
>   12246     776    1792   14814    39de simpletest.-O1-g.llgcov.
>    1345     492    1616    3453     d7d simpletest.-O1.none
> I was wondering if there is any suggestion for reducing the size either
> through more optimization or by compromising some feature.

The coverage data that the instrprof coverage adds to the binary isn't
actually necessary for runtime, so you could probably get away with
building the binary twice (once with -fcoverage-mapping, once with only
-fprofile-instr-generate) and using one for execution and the other for
data collection. The binaries will still be larger, due to the profiling
itself, but it might help.

Let me know if this works - it might be worth adding an option to emit
the coverage data into a separate file if this is valuable.

> Q3- in llcovprof-way since the runtime profile data is collected in a single
> file, the file system will serialize multi threaded writes, hence increasing
> the execution time. Is there a way to avoid this?

There isn't anything in place to help with this right now.