Add code coverage (LLVM source based coverage)

[jschwe]

This is something I've wanted to do for some time, but I haven't had the time. I'll use this issue to write down some of the issues that need to be solved. If someone has the time to address this feel free to open a PR.

Source-based coverage support via LLVM landed in the rust compiler somewhat recently. It can be enabled via the nightly RUSTFLAG -Zinstrument-coverage

Challenges specific to rusty-hermit:

RUSTC_WRAPPER

RUSTFLAGS apply to everything, including dependencies, build scripts and macros. A RUSTC_WRAPPER program should be written that filters the RUSTFLAG -Zinstrument-coverage flag and only applies it to relevant crates. In particular crates such as core must be excluded, since profiler_builtins, which is required for -Zinstrument_coverage relies on core.

It should be evaluated if we can include std, since we provide a hermit specific std library, and coverage metrics for that would be useful as well.

Optimizations

Optimizations are detrimental to code coverage so they should be disabled, i.e. by at least using the flags -Copt-level=0 -Clink-dead-code. If possible we should again use the RUSTC_WRAPPER to control this, so that crates without code coverage are optimized, while crates with code coverage are not.

Previous tests have shown that rusty-hermit may be lacking some (mathematical) functions, which are never called. -Clink-dead-code might surface such problems.

Profiler_builtins

This can be solved independently from the previous issue via the following MWE:

In rusty-hermit add profiler_builtins to the build-std configuration .cargo/config. Run cargo rustc -p rusty_demo -- -Zinstrument-coverage. This is a sort of poor mans RUSTC_WRAPPER where the -Zinstrument-coverageflag is only passed to the rusty_demo crate. Of course this means that the actual kernel is not instrumented, but i believe instrumenting the userspace application first is easier.

This will probably result in a linker error where some __llvm_profile_* functions are missing. The issue is that the llvm source code files for the profiler builtins are not shipped with the nightly and the build script (see profiler_builtins ) in the rust compiler silently ignores the missing files and simply does nothing.

To test things out you can visit rust/src and clone the linked version of the llvm-project. You can then copy the compiler-rt folder into "$HOME/.rustup/toolchains/nightly-2020-12-23-x86_64-unknown-linux-gnu/lib/rustlib/src/rust/src/llvm-project/". Replace the version in the above string to whatever version you are using (could also be simply nightly) and make sure that rustup also installed the rustsrc component. Opening an issue on rustup asking for a profiling component which installs the necessary files could also be an option.

Now when building the application (with verbose option) you should get new linker errors, that mention undefined symbols such as strlen, fopen etc. originating from the llvm profiler c files.

To solve that I'd take a look at hermit playground, but additional steps might be needed, since in the end we need the symbols in the kernel code too, if we want coverage there.

[jschwe]

@stlankes Could you add the enhancement label?

[ChrisSmith2]

Hey! I'm a computer science student at the University of Texas at Austin taking a Virtualization course and one of our assignments is to contribute to an open source project. Would I be able to work on resolving this issue?