As of 1.19 it is not possible to accurately measure unit test coverage

[ctz]

1.19 prevents access to -Zno-landing-pads. I was using that to get accurate test coverage; without it all landing pads appear as uncovered code at the start and end of every function.

-Cpanic=abort is intended to be the stable replacement for -Zno-landing-pads but unfortunately that isn't supported for unit tests (apparently, because libtest requires unwinding support).

I'd like a way to do any of the following on stable:

• restore no-landing-pads behaviour, or,
• have panic=abort work for unit tests. Note that I don't much care that should_panic tests wouldn't work.

Is there some combination of cargo/rustc options that might achieve that?

[hanna-kruppe]

#[should_panic] tests not working seems like a pretty big downside. Even if you don't care, the usefulness of coverage info seems entirely independent from whether you have #[should_panic] tests. Perhaps there's a better solution, e.g.,

• Not having landing pads mess up coverage information (how do C++ tools solve this?)
• Having a mode in libtest that runs tests in a child process, to be able to detect and handle panics

[kennytm]

@rkruppe

C++ tools solve this by:

1. In GCC, the generated *.gcno will encode whether an edge in the control-flow graph is exceptional or not, via the GCOV_ARC_FAKE flag. Coverage analyzer can choose to ignore these edges. LLVM's insert-gcov-profiling pass does not provide any of these information.

2. clang uses instrprof (Implement support for LLVMs code coverage instrumentation #34701) instead of GCOV.

[marco-c]

I guess we could try to add support for GCOV_ARC_FAKE in LLVM's gcov instrumentation.

[marco-c]

The tracking issue for -Zprofile is #42524.

I think in the meantime, until this is fixed in LLVM, you can keep using -Zno-landing-pads as a workaround.

[richardwhiuk]

See also #55352 which proposes some alternative solutions here.

[ctz]

Note #43410 removes -Zno-landing-pads

[sanmai-NL]

@ctz: that's this issue, I think you mean a PR.

[ctz]

Derp. I meant #70175

[marco-c]

Another workaround can be mozilla/grcov#427 (comment):

I managed to run tests for pyo3 in using -Cpanic=abort -Zpanic_abort_tests in RUSTFLAGS.
The doc tests were failing with that so adding an extra RUSTDOCFLAGS="-Cpanic=abort" fixes failures.

[rivy]

Alternatively, pinning the rust version to nightly-2020-04-29 for code coverage also works, without altering flags or resultant coverage.