Track test coverage

[briansmith]

• [ ] Measure code coverage
  • [x] x86 and x86-64 Linux.
  • [ ] armv6, armv7+neon, and AAarch64: Some progress, but blocked on toolchain issues (see comments below).
  • [ ] Windows: No progress yet. This article makes it look like it might not be too painful to do.
• [ ] Report code coverage. I guess this is easiest to do with coveralls. I guess maybe we need to add the coveralls badge to the README or somewhere else where people can find it.

This is related to #131.

See also https://boringssl-review.googlesource.com/#/c/7251/ and related BoringSSL work.

[pietro]

@briansmith I'm trying to get aarch64-unknown-linux-gnu builds going on native arm on travis. If I get it working I'll and kcov to them. hopefully it won't be too hard.

[briansmith]

PR #1124 adds code coverage measurement of the test suite to GitHub Actions for {x86_64,i686}-unknown-linux-musl and aarch64-unknown-linux-gnu (via QEMU) using Rust Nightly's new source-basd branch coverage. I integrated codecov.io into GitHub so that code reviews can make use of the code coverage information.

In the future, we should make the following improvements:

• [ ] Add 32-bit ARM code coverage. As of 2020-11-25, I don't know of a 32-bit ARM target for Rust that has profiler_builtins available to do this.
• [ ] Switch any -gnu targets for profiling to instead be -musl targets.
• [ ] Enable code coverage measurement for the assembly language code.
• [ ] Measure coverage for the Windows targets.
• [ ] Measure coverage for Apple targets.
• [ ] Enable branch coverage measurement. Branch coverage measurement is being added to LLVM: https://reviews.llvm.org/D84467.
• [ ] Organize the coverage reports into a more useful form. Right now, all the code coverage information for running the test suite is combined into a single report. Ideally we'd separate things out so that we can verify that, for example, aead_tests provides complete code coverage for the ring::aead module, regardless of what other tests do. As another example, if would be useful to be able to look at code coverage on a per-target basis.
• [ ] Automatically enforce the rules that new PRs don't land without test coverage for the changed/new code.