Better coverage?

[jzmaddock]

I've just briefly looked at the codecov reports (many thanks for doing this @mborland !) to see what needs improving, and I'm seeing a lot of false flags, for example in https://app.codecov.io/gh/boostorg/math/pull/1067/blob/include/boost/math/special_functions/lanczos.hpp line 870 only the final initializer in the list is marked as covered, even though they must have all been executed if one has! Is there anything we can do about this?

[ckormanyos]

Numerous return lines are, in fact, marked as covered. Notwithstanding that, however, the tool seems to be experiencing problems with BOOST_MATH_BIG_CONSTANT. We could see what happens on some explicit local tests as time allows.

Personally, I have found coverage on highly-templated C++ code to be somewhat challenging.

I also recommend to cut the optimization level way down to -O1 or even -O0 when running gcov/lcov. I have not looked at the make system. But that is also something can be tuned. I almost never change code for the sake of covering templates, but try to tune the environment or code-annotations to better reflect the honest situation.

[ckormanyos]

That being said, my personal experience is that it can take a comparable amount of time to attain $100\%$ line coverage as it took or takes to write the code in the first place. I find that testing effort is often underestimated on the road to quality. Somewhere a while back I mentioned increasing line coverage very briefly and suspect we need a dedicated effort to reach (via specially contrived test-cases) more hard-to-reach lines.

In the case mentioned here, however, I do believe that lcov might have representation problems with some of those big macros.

[jzmaddock]

It took a bit of detective work, but yes, it's using lcov, and lcov has exclusion macros we can use. I did spot a few genuine cases not covered while browsing through, so maybe I'll see what I can squash.

[ckormanyos]

Sometimes, well we will ultimately just know better than lcov, which my opinion is, can't fully handle modern C++, but sure gives a good indication of what's going on.

Then I use (in justified cases - whatever that is decided to mean for our projects)

lcov has exclusion macros we can use

For multipple lines:

// LCOV_EXCL_START
my_uncovered_line();
my_other_uncovered_line();
// LCOV_EXCL_STOP

Or for single line(s)

my_uncovered_line(); // LCOV_EXCL_LINE
my_other_uncovered_line(); // LCOV_EXCL_LINE

[ckormanyos]

See also:

• 1084 tgamma()/lgamma() edge cases for $250$ decimal digits.

• 1090 achieve rudimentary octonion coverage.

• 1096 actually run (in addition to instantiate) the code in instantiate.hpp.