source-based-code-coverage

The feature request for this feature is: #34701

The Major Change Proposal (MCP) for this feature is: #278


Introduction

The Rust compiler includes two code coverage implementations:

  • A GCC-compatible, gcov-based coverage implementation, enabled with -Zprofile, which operates on DebugInfo.
  • A source-based code coverage implementation, enabled with -Zinstrument-coverage, which uses LLVM's native coverage instrumentation to generate very precise coverage data.

This document describes how to enable and use the LLVM instrumentation-based coverage, via the -Zinstrument-coverage compiler flag.

How it works

When -Zinstrument-coverage is enabled, the Rust compiler enhances rust-based libraries and binaries by:

  • Automatically injecting calls to an LLVM intrinsic (llvm.instrprof.increment), at functions and branches in compiled code, to increment counters when conditional sections of code are executed.
  • Embedding additional information in the data section of each library and binary (using the LLVM Code Coverage Mapping Format), to define the code regions (start and end positions in the source code) being counted.

When running a coverage-instrumented program, the counter values are written to a profraw file at program termination. LLVM bundles tools that read the counter results, combine those results with the coverage map (embedded in the program binary), and generate coverage reports in multiple formats.

Enable coverage profiling in the Rust compiler

Rust's source-based code coverage requires the Rust "profiler runtime". Without it, compiling with -Zinstrument-coverage generates an error that the profiler runtime is missing.

The Rust nightly distribution channel should include the profiler runtime, by default.

IMPORTANT: If you are building the Rust compiler from the source distribution, the profiler runtime is not enabled in the default config.toml.example, and may not be enabled in your config.toml. Edit the config.toml file, and find the profiler feature entry. Uncomment it and set it to true:

# Build the profiler runtime (required when compiling with options that depend
# on this runtime, such as `-C profile-generate` or `-Z instrument-coverage`).
profiler = true

Then rebuild the Rust compiler (see rustc-dev-guide-how-to-build-and-run).

Building the demangler

LLVM coverage reporting tools generate results that can include function names and other symbol references, and the raw coverage results report symbols using the compiler's "mangled" version of the symbol names, which can be difficult to interpret. To work around this issue, LLVM coverage tools also support a user-specified symbol name demangler.

One option for a Rust demangler is rustfilt, which can be installed with:

cargo install rustfilt

Another option, if you are building from the Rust compiler source distribution, is to use the rust-demangler tool included in the Rust source distribution, which can be built with:

$ ./x.py build rust-demangler

Compiling with coverage enabled

Set the -Zinstrument-coverage compiler flag in order to enable LLVM source-based code coverage profiling.

With cargo, you can instrument your program binary and dependencies at the same time.

For example (if your project's Cargo.toml builds a binary by default):

$ cd your-project
$ cargo clean
$ RUSTFLAGS="-Zinstrument-coverage" cargo build

If cargo is not configured to use your profiler-enabled version of rustc, set the path explicitly via the RUSTC environment variable. Here is another example, using a stage1 build of rustc to compile an example binary (from the json5format crate):

$ RUSTC=$HOME/rust/build/x86_64-unknown-linux-gnu/stage1/bin/rustc \
    RUSTFLAGS="-Zinstrument-coverage" \
    cargo build --example formatjson5

Running the instrumented binary to generate raw coverage profiling data

In the previous example, cargo generated the coverage-instrumented binary formatjson5:

$ echo "{some: 'thing'}" | target/debug/examples/formatjson5 -
{
    some: 'thing',
}

After running this program, a new file, default.profraw, should be in the current working directory. It's often preferable to set a specific file name or path. You can change the output file using the environment variable LLVM_PROFILE_FILE:

$ echo "{some: 'thing'}" \
    | LLVM_PROFILE_FILE="formatjson5.profraw" target/debug/examples/formatjson5 -
...
$ ls formatjson5.profraw
formatjson5.profraw

If LLVM_PROFILE_FILE contains a path to a non-existent directory, the missing directory structure will be created. Additionally, the following special pattern strings are rewritten:

  • %p - The process ID.
  • %h - The hostname of the machine running the program.
  • %t - The value of the TMPDIR environment variable.
  • %Nm - the instrumented binary‚Äôs signature: The runtime creates a pool of N raw profiles, used for on-line profile merging. The runtime takes care of selecting a raw profile from the pool, locking it, and updating it before the program exits. N must be between 1 and 9, and defaults to 1 if omitted (with simply %m).
  • %c - Does not add anything to the filename, but enables a mode (on some platforms, including Darwin) in which profile counter updates are continuously synced to a file. This means that if the instrumented program crashes, or is killed by a signal, perfect coverage information can still be recovered.

Creating coverage reports

LLVM's tools to process coverage data and coverage maps have some version dependencies. If you encounter a version mismatch, try updating your LLVM tools.

If you are building the Rust compiler from source, you can optionally use the bundled LLVM tools, built from source. Those tool binaries can typically be found in your build platform directory at something like: rust/build/x86_64-unknown-linux-gnu/llvm/bin/llvm-*. (Look for llvm-profdata and llvm-cov.)

Raw profiles have to be indexed before they can be used to generate coverage reports. This is done using llvm-profdata merge (which can combine multiple raw profiles and index them at the same time):

$ llvm-profdata merge -sparse formatjson5.profraw -o formatjson5.profdata

Finally, the .profdata file is used, in combination with the coverage map (from the program binary) to generate coverage reports using llvm-cov report--for a coverage summaries--and llvm-cov show--to see detailed coverage of lines and regions (character ranges), overlaid on the original source code.

These commands have several display and filtering options. For example:

$ llvm-cov show -Xdemangler=rustfilt target/debug/examples/formatjson5 \
    -instr-profile=formatjson5.profdata \
    -show-line-counts-or-regions \
    -show-instantiations \
    -name=add_quoted_string
Screenshot of sample `llvm-cov show` result, for function add_quoted_string

Some of the more notable options in this example include:

  • --Xdemangler=rustfilt - the command name or path used to demangle Rust symbols (rustfilt in the example, but this could also be a path to the rust-demangler tool)
  • target/debug/examples/formatjson5 - the instrumented binary (from which to extract the coverage map)
  • --instr-profile=<path-to-file>.profdata - the location of the .profdata file created by llvm-profdata merge (from the .profraw file generated by the instrumented binary)
  • --name=<exact-function-name> - to show coverage for a specific function (or, consider using another filter option, such as --name-regex=<pattern>)

Interpreting reports

There are four statistics tracked in a coverage summary:

  • Function coverage is the percentage of functions that have been executed at least once. A function is considered to be executed if any of its instantiations are executed.
  • Instantiation coverage is the percentage of function instantiations that have been executed at least once. Generic functions and functions generated from macros are two kinds of functions that may have multiple instantiations.
  • Line coverage is the percentage of code lines that have been executed at least once. Only executable lines within function bodies are considered to be code lines.
  • Region coverage is the percentage of code regions that have been executed at least once. A code region may span multiple lines: for example, in a large function body with no control flow. In other cases, a single line can contain multiple code regions: return x || (y && z) has countable code regions for x (which may resolve the expression, if x is true), || (y && z) (executed only if x was false), and return (executed in either situation).

Of these four statistics, function coverage is usually the least granular while region coverage is the most granular. The project-wide totals for each statistic are listed in the summary.

Other references

Rust's implementation and workflow for source-based code coverage is based on the same library and tools used to implement source-based code coverage in Clang. (This document is partially based on the Clang guide.)