Source-based Code Coverage


This document explains how to use clang’s source-based code coverage feature. It’s called “source-based” because it operates on AST and preprocessor information directly. This allows it to generate very precise coverage data.

Clang ships two other code coverage implementations:

  • SanitizerCoverage - A low-overhead tool meant for use alongside the various sanitizers. It can provide up to edge-level coverage.

  • gcov - A GCC-compatible coverage implementation which operates on DebugInfo. This is enabled by -ftest-coverage or --coverage.

From this point onwards “code coverage” will refer to the source-based kind.

The code coverage workflow

The code coverage workflow consists of three main steps:

  • Compiling with coverage enabled.

  • Running the instrumented program.

  • Creating coverage reports.

The next few sections work through a complete, copy-‘n-paste friendly example based on this program:

% cat <<EOF >
#define BAR(x) ((x) || (x))
template <typename T> void foo(T x) {
  for (unsigned I = 0; I < 10; ++I) { BAR(I); }
int main() {
  return 0;

Compiling with coverage enabled

To compile code with coverage enabled, pass -fprofile-instr-generate -fcoverage-mapping to the compiler:

# Step 1: Compile with coverage enabled.
% clang++ -fprofile-instr-generate -fcoverage-mapping -o foo

Note that linking together code with and without coverage instrumentation is supported. Uninstrumented code simply won’t be accounted for in reports.

Running the instrumented program

The next step is to run the instrumented program. When the program exits it will write a raw profile to the path specified by the LLVM_PROFILE_FILE environment variable. If that variable does not exist, the profile is written to default.profraw in the current directory of the program. 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” expands out to the process ID.

  • “%h” expands out to the hostname of the machine running the program.

  • “%Nm” expands out to the instrumented binary’s signature. When this pattern is specified, the runtime creates a pool of N raw profiles which are 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. If N is not specified (i.e the pattern is “%m”), it’s assumed that N = 1. N must be between 1 and 9. The merge pool specifier can only occur once per filename pattern.

  • “%c” expands out to nothing, but enables a mode 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. Continuous mode does not support value profiling for PGO, and is only supported on Darwin at the moment. Support for Linux may be mostly complete but requires testing, and support for Windows may require more extensive changes: please get involved if you are interested in porting this feature.

# Step 2: Run the program.
% LLVM_PROFILE_FILE="foo.profraw" ./foo

Note that continuous mode is also used on Fuchsia where it’s the only supported mode, but the implementation is different. The Darwin and Linux implementation relies on padding and the ability to map a file over the existing memory mapping which is generally only available on POSIX systems and isn’t suitable for other platforms.

On Fuchsia, we rely on the ability to relocate counters at runtime using a level of indirection. On every counter access, we add a bias to the counter address. This bias is stored in __llvm_profile_counter_bias symbol that’s provided by the profile runtime and is initially set to zero, meaning no relocation. The runtime can map the profile into memory at arbitrary locations, and set bias to the offset between the original and the new counter location, at which point every subsequent counter access will be to the new location, which allows updating profile directly akin to the continuous mode.

The advantage of this approach is that doesn’t require any special OS support. The disadvantage is the extra overhead due to additional instructions required for each counter access (overhead both in terms of binary size and performance) plus duplication of counters (i.e. one copy in the binary itself and another copy that’s mapped into memory). This implementation can be also enabled for other platforms by passing the -runtime-counter-relocation option to the backend during compilation.

% clang++ -fprofile-instr-generate -fcoverage-mapping -mllvm -runtime-counter-relocation -o foo

Creating coverage reports

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

# Step 3(a): Index the raw profile.
% llvm-profdata merge -sparse foo.profraw -o foo.profdata

There are multiple different ways to render coverage reports. The simplest option is to generate a line-oriented report:

# Step 3(b): Create a line-oriented coverage report.
% llvm-cov show ./foo -instr-profile=foo.profdata

This report includes a summary view as well as dedicated sub-views for templated functions and their instantiations. For our example program, we get distinct views for foo<int>(...) and foo<float>(...). If -show-line-counts-or-regions is enabled, llvm-cov displays sub-line region counts (even in macro expansions):

    1|   20|#define BAR(x) ((x) || (x))
                           ^20     ^2
    2|    2|template <typename T> void foo(T x) {
    3|   22|  for (unsigned I = 0; I < 10; ++I) { BAR(I); }
                                   ^22     ^20  ^20^20
    4|    2|}
| void foo<int>(int):
|      2|    1|template <typename T> void foo(T x) {
|      3|   11|  for (unsigned I = 0; I < 10; ++I) { BAR(I); }
|                                     ^11     ^10  ^10^10
|      4|    1|}
| void foo<float>(int):
|      2|    1|template <typename T> void foo(T x) {
|      3|   11|  for (unsigned I = 0; I < 10; ++I) { BAR(I); }
|                                     ^11     ^10  ^10^10
|      4|    1|}

To generate a file-level summary of coverage statistics instead of a line-oriented report, try:

# Step 3(c): Create a coverage summary.
% llvm-cov report ./foo -instr-profile=foo.profdata
Filename           Regions    Missed Regions     Cover   Functions  Missed Functions  Executed       Lines      Missed Lines     Cover
/tmp/             13                 0   100.00%           3                 0   100.00%          13                 0   100.00%
TOTAL                   13                 0   100.00%           3                 0   100.00%          13                 0   100.00%

The llvm-cov tool supports specifying a custom demangler, writing out reports in a directory structure, and generating html reports. For the full list of options, please refer to the command guide.

A few final notes:

  • The -sparse flag is optional but can result in dramatically smaller indexed profiles. This option should not be used if the indexed profile will be reused for PGO.

  • Raw profiles can be discarded after they are indexed. Advanced use of the profile runtime library allows an instrumented program to merge profiling information directly into an existing raw profile on disk. The details are out of scope.

  • The llvm-profdata tool can be used to merge together multiple raw or indexed profiles. To combine profiling data from multiple runs of a program, try e.g:

    % llvm-profdata merge -sparse foo1.profraw foo2.profdata -o foo3.profdata

Exporting coverage data

Coverage data can be exported into JSON using the llvm-cov export sub-command. There is a comprehensive reference which defines the structure of the exported data at a high level in the llvm-cov source code.

Interpreting reports

There are four statistics tracked in a coverage summary:

  • Function coverage is the percentage of functions which 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 which have been executed at least once. Template functions and static inline functions from headers are two kinds of functions which may have multiple instantiations.

  • Line coverage is the percentage of code lines which 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 which have been executed at least once. A code region may span multiple lines (e.g in a large function body with no control flow). However, it’s also possible for a single line to contain multiple code regions (e.g in “return x || y && z”).

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.

Format compatibility guarantees

  • There are no backwards or forwards compatibility guarantees for the raw profile format. Raw profiles may be dependent on the specific compiler revision used to generate them. It’s inadvisable to store raw profiles for long periods of time.

  • Tools must retain backwards compatibility with indexed profile formats. These formats are not forwards-compatible: i.e, a tool which uses format version X will not be able to understand format version (X+k).

  • Tools must also retain backwards compatibility with the format of the coverage mappings emitted into instrumented binaries. These formats are not forwards-compatible.

  • The JSON coverage export format has a (major, minor, patch) version triple. Only a major version increment indicates a backwards-incompatible change. A minor version increment is for added functionality, and patch version increments are for bugfixes.

Using the profiling runtime without static initializers

By default the compiler runtime uses a static initializer to determine the profile output path and to register a writer function. To collect profiles without using static initializers, do this manually:

  • Export a int __llvm_profile_runtime symbol from each instrumented shared library and executable. When the linker finds a definition of this symbol, it knows to skip loading the object which contains the profiling runtime’s static initializer.

  • Forward-declare void __llvm_profile_initialize_file(void) and call it once from each instrumented executable. This function parses LLVM_PROFILE_FILE, sets the output path, and truncates any existing files at that path. To get the same behavior without truncating existing files, pass a filename pattern string to void __llvm_profile_set_filename(char *). These calls can be placed anywhere so long as they precede all calls to __llvm_profile_write_file.

  • Forward-declare int __llvm_profile_write_file(void) and call it to write out a profile. This function returns 0 when it succeeds, and a non-zero value otherwise. Calling this function multiple times appends profile data to an existing on-disk raw profile.

In C++ files, declare these as extern "C".

Collecting coverage reports for the llvm project

To prepare a coverage report for llvm (and any of its sub-projects), add -DLLVM_BUILD_INSTRUMENTED_COVERAGE=On to the cmake configuration. Raw profiles will be written to $BUILD_DIR/profiles/. To prepare an html report, run llvm/utils/

To specify an alternate directory for raw profiles, use -DLLVM_PROFILE_DATA_DIR. To change the size of the profile merge pool, use -DLLVM_PROFILE_MERGE_POOL_SIZE.

Drawbacks and limitations

  • Prior to version 2.26, the GNU binutils BFD linker is not able link programs compiled with -fcoverage-mapping in its --gc-sections mode. Possible workarounds include disabling --gc-sections, upgrading to a newer version of BFD, or using the Gold linker.

  • Code coverage does not handle unpredictable changes in control flow or stack unwinding in the presence of exceptions precisely. Consider the following function:

    int f() {
      return 0;

    If the call to may_throw() propagates an exception into f, the code coverage tool may mark the return statement as executed even though it is not. A call to longjmp() can have similar effects.

Clang implementation details

This section may be of interest to those wishing to understand or improve the clang code coverage implementation.

Gap regions

Gap regions are source regions with counts. A reporting tool cannot set a line execution count to the count from a gap region unless that region is the only one on a line.

Gap regions are used to eliminate unnatural artifacts in coverage reports, such as red “unexecuted” highlights present at the end of an otherwise covered line, or blue “executed” highlights present at the start of a line that is otherwise not executed.

Switch statements

The region mapping for a switch body consists of a gap region that covers the entire body (starting from the ‘{‘ in ‘switch (…) {‘, and terminating where the last case ends). This gap region has a zero count: this causes “gap” areas in between case statements, which contain no executable code, to appear uncovered.

When a switch case is visited, the parent region is extended: if the parent region has no start location, its start location becomes the start of the case. This is used to support switch statements without a CompoundStmt body, in which the switch body and the single case share a count.

For switches with CompoundStmt bodies, a new region is created at the start of each switch case.