1==========================
2Source-based Code Coverage
3==========================
4
5.. contents::
6   :local:
7
8Introduction
9============
10
11This document explains how to use clang's source-based code coverage feature.
12It's called "source-based" because it operates on AST and preprocessor
13information directly. This allows it to generate very precise coverage data.
14
15Clang ships two other code coverage implementations:
16
17* :doc:`SanitizerCoverage` - A low-overhead tool meant for use alongside the
18  various sanitizers. It can provide up to edge-level coverage.
19
20* gcov - A GCC-compatible coverage implementation which operates on DebugInfo.
21  This is enabled by ``-ftest-coverage`` or ``--coverage``.
22
23From this point onwards "code coverage" will refer to the source-based kind.
24
25The code coverage workflow
26==========================
27
28The code coverage workflow consists of three main steps:
29
30* Compiling with coverage enabled.
31
32* Running the instrumented program.
33
34* Creating coverage reports.
35
36The next few sections work through a complete, copy-'n-paste friendly example
37based on this program:
38
39.. code-block:: cpp
40
41    % cat <<EOF > foo.cc
42    #define BAR(x) ((x) || (x))
43    template <typename T> void foo(T x) {
44      for (unsigned I = 0; I < 10; ++I) { BAR(I); }
45    }
46    int main() {
47      foo<int>(0);
48      foo<float>(0);
49      return 0;
50    }
51    EOF
52
53Compiling with coverage enabled
54===============================
55
56To compile code with coverage enabled, pass ``-fprofile-instr-generate
57-fcoverage-mapping`` to the compiler:
58
59.. code-block:: console
60
61    # Step 1: Compile with coverage enabled.
62    % clang++ -fprofile-instr-generate -fcoverage-mapping foo.cc -o foo
63
64Note that linking together code with and without coverage instrumentation is
65supported. Uninstrumented code simply won't be accounted for in reports.
66
67Running the instrumented program
68================================
69
70The next step is to run the instrumented program. When the program exits it
71will write a **raw profile** to the path specified by the ``LLVM_PROFILE_FILE``
72environment variable. If that variable does not exist, the profile is written
73to ``default.profraw`` in the current directory of the program. If
74``LLVM_PROFILE_FILE`` contains a path to a non-existent directory, the missing
75directory structure will be created.  Additionally, the following special
76**pattern strings** are rewritten:
77
78* "%p" expands out to the process ID.
79
80* "%h" expands out to the hostname of the machine running the program.
81
82* "%t" expands out to the value of the ``TMPDIR`` environment variable. On
83  Darwin, this is typically set to a temporary scratch directory.
84
85* "%Nm" expands out to the instrumented binary's signature. When this pattern
86  is specified, the runtime creates a pool of N raw profiles which are used for
87  on-line profile merging. The runtime takes care of selecting a raw profile
88  from the pool, locking it, and updating it before the program exits.  If N is
89  not specified (i.e the pattern is "%m"), it's assumed that ``N = 1``. N must
90  be between 1 and 9. The merge pool specifier can only occur once per filename
91  pattern.
92
93* "%c" expands out to nothing, but enables a mode in which profile counter
94  updates are continuously synced to a file. This means that if the
95  instrumented program crashes, or is killed by a signal, perfect coverage
96  information can still be recovered. Continuous mode does not support value
97  profiling for PGO, and is only supported on Darwin at the moment. Support for
98  Linux may be mostly complete but requires testing, and support for Windows
99  may require more extensive changes: please get involved if you are interested
100  in porting this feature.
101
102.. code-block:: console
103
104    # Step 2: Run the program.
105    % LLVM_PROFILE_FILE="foo.profraw" ./foo
106
107Note that continuous mode is also used on Fuchsia where it's the only supported
108mode, but the implementation is different. The Darwin and Linux implementation
109relies on padding and the ability to map a file over the existing memory
110mapping which is generally only available on POSIX systems and isn't suitable
111for other platforms.
112
113On Fuchsia, we rely on the ability to relocate counters at runtime using a
114level of indirection. On every counter access, we add a bias to the counter
115address. This bias is stored in ``__llvm_profile_counter_bias`` symbol that's
116provided by the profile runtime and is initially set to zero, meaning no
117relocation. The runtime can map the profile into memory at arbitrary locations,
118and set bias to the offset between the original and the new counter location,
119at which point every subsequent counter access will be to the new location,
120which allows updating profile directly akin to the continuous mode.
121
122The advantage of this approach is that doesn't require any special OS support.
123The disadvantage is the extra overhead due to additional instructions required
124for each counter access (overhead both in terms of binary size and performance)
125plus duplication of counters (i.e. one copy in the binary itself and another
126copy that's mapped into memory). This implementation can be also enabled for
127other platforms by passing the ``-runtime-counter-relocation`` option to the
128backend during compilation.
129
130.. code-block:: console
131
132    % clang++ -fprofile-instr-generate -fcoverage-mapping -mllvm -runtime-counter-relocation foo.cc -o foo
133
134Creating coverage reports
135=========================
136
137Raw profiles have to be **indexed** before they can be used to generate
138coverage reports. This is done using the "merge" tool in ``llvm-profdata``
139(which can combine multiple raw profiles and index them at the same time):
140
141.. code-block:: console
142
143    # Step 3(a): Index the raw profile.
144    % llvm-profdata merge -sparse foo.profraw -o foo.profdata
145
146There are multiple different ways to render coverage reports. The simplest
147option is to generate a line-oriented report:
148
149.. code-block:: console
150
151    # Step 3(b): Create a line-oriented coverage report.
152    % llvm-cov show ./foo -instr-profile=foo.profdata
153
154This report includes a summary view as well as dedicated sub-views for
155templated functions and their instantiations. For our example program, we get
156distinct views for ``foo<int>(...)`` and ``foo<float>(...)``.  If
157``-show-line-counts-or-regions`` is enabled, ``llvm-cov`` displays sub-line
158region counts (even in macro expansions):
159
160.. code-block:: none
161
162        1|   20|#define BAR(x) ((x) || (x))
163                               ^20     ^2
164        2|    2|template <typename T> void foo(T x) {
165        3|   22|  for (unsigned I = 0; I < 10; ++I) { BAR(I); }
166                                       ^22     ^20  ^20^20
167        4|    2|}
168    ------------------
169    | void foo<int>(int):
170    |      2|    1|template <typename T> void foo(T x) {
171    |      3|   11|  for (unsigned I = 0; I < 10; ++I) { BAR(I); }
172    |                                     ^11     ^10  ^10^10
173    |      4|    1|}
174    ------------------
175    | void foo<float>(int):
176    |      2|    1|template <typename T> void foo(T x) {
177    |      3|   11|  for (unsigned I = 0; I < 10; ++I) { BAR(I); }
178    |                                     ^11     ^10  ^10^10
179    |      4|    1|}
180    ------------------
181
182If ``--show-branches=count`` and ``--show-expansions`` are also enabled, the
183sub-views will show detailed branch coverage information in addition to the
184region counts:
185
186.. code-block:: none
187
188    ------------------
189    | void foo<float>(int):
190    |      2|    1|template <typename T> void foo(T x) {
191    |      3|   11|  for (unsigned I = 0; I < 10; ++I) { BAR(I); }
192    |                                     ^11     ^10  ^10^10
193    |  ------------------
194    |  |  |    1|     10|#define BAR(x) ((x) || (x))
195    |  |  |                             ^10     ^1
196    |  |  |  ------------------
197    |  |  |  |  Branch (1:17): [True: 9, False: 1]
198    |  |  |  |  Branch (1:24): [True: 0, False: 1]
199    |  |  |  ------------------
200    |  ------------------
201    |  |  Branch (3:23): [True: 10, False: 1]
202    |  ------------------
203    |      4|    1|}
204    ------------------
205
206
207To generate a file-level summary of coverage statistics instead of a
208line-oriented report, try:
209
210.. code-block:: console
211
212    # Step 3(c): Create a coverage summary.
213    % llvm-cov report ./foo -instr-profile=foo.profdata
214    Filename           Regions    Missed Regions     Cover   Functions  Missed Functions  Executed       Lines      Missed Lines     Cover     Branches    Missed Branches     Cover
215    --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
216    /tmp/foo.cc             13                 0   100.00%           3                 0   100.00%          13                 0   100.00%           12                  2    83.33%
217    --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
218    TOTAL                   13                 0   100.00%           3                 0   100.00%          13                 0   100.00%           12                  2    83.33%
219
220The ``llvm-cov`` tool supports specifying a custom demangler, writing out
221reports in a directory structure, and generating html reports. For the full
222list of options, please refer to the `command guide
223<https://llvm.org/docs/CommandGuide/llvm-cov.html>`_.
224
225A few final notes:
226
227* The ``-sparse`` flag is optional but can result in dramatically smaller
228  indexed profiles. This option should not be used if the indexed profile will
229  be reused for PGO.
230
231* Raw profiles can be discarded after they are indexed. Advanced use of the
232  profile runtime library allows an instrumented program to merge profiling
233  information directly into an existing raw profile on disk. The details are
234  out of scope.
235
236* The ``llvm-profdata`` tool can be used to merge together multiple raw or
237  indexed profiles. To combine profiling data from multiple runs of a program,
238  try e.g:
239
240  .. code-block:: console
241
242      % llvm-profdata merge -sparse foo1.profraw foo2.profdata -o foo3.profdata
243
244Exporting coverage data
245=======================
246
247Coverage data can be exported into JSON using the ``llvm-cov export``
248sub-command. There is a comprehensive reference which defines the structure of
249the exported data at a high level in the llvm-cov source code.
250
251Interpreting reports
252====================
253
254There are five statistics tracked in a coverage summary:
255
256* Function coverage is the percentage of functions which have been executed at
257  least once. A function is considered to be executed if any of its
258  instantiations are executed.
259
260* Instantiation coverage is the percentage of function instantiations which
261  have been executed at least once. Template functions and static inline
262  functions from headers are two kinds of functions which may have multiple
263  instantiations. This statistic is hidden by default in reports, but can be
264  enabled via the ``-show-instantiation-summary`` option.
265
266* Line coverage is the percentage of code lines which have been executed at
267  least once. Only executable lines within function bodies are considered to be
268  code lines.
269
270* Region coverage is the percentage of code regions which have been executed at
271  least once. A code region may span multiple lines (e.g in a large function
272  body with no control flow). However, it's also possible for a single line to
273  contain multiple code regions (e.g in "return x || y && z").
274
275* Branch coverage is the percentage of "true" and "false" branches that have
276  been taken at least once. Each branch is tied to individual conditions in the
277  source code that may each evaluate to either "true" or "false".  These
278  conditions may comprise larger boolean expressions linked by boolean logical
279  operators. For example, "x = (y == 2) || (z < 10)" is a boolean expression
280  that is comprised of two individual conditions, each of which evaluates to
281  either true or false, producing four total branch outcomes.
282
283Of these five statistics, function coverage is usually the least granular while
284branch coverage is the most granular. 100% branch coverage for a function
285implies 100% region coverage for a function. The project-wide totals for each
286statistic are listed in the summary.
287
288Format compatibility guarantees
289===============================
290
291* There are no backwards or forwards compatibility guarantees for the raw
292  profile format. Raw profiles may be dependent on the specific compiler
293  revision used to generate them. It's inadvisable to store raw profiles for
294  long periods of time.
295
296* Tools must retain **backwards** compatibility with indexed profile formats.
297  These formats are not forwards-compatible: i.e, a tool which uses format
298  version X will not be able to understand format version (X+k).
299
300* Tools must also retain **backwards** compatibility with the format of the
301  coverage mappings emitted into instrumented binaries. These formats are not
302  forwards-compatible.
303
304* The JSON coverage export format has a (major, minor, patch) version triple.
305  Only a major version increment indicates a backwards-incompatible change. A
306  minor version increment is for added functionality, and patch version
307  increments are for bugfixes.
308
309Impact of llvm optimizations on coverage reports
310================================================
311
312llvm optimizations (such as inlining or CFG simplification) should have no
313impact on coverage report quality. This is due to the fact that the mapping
314from source regions to profile counters is immutable, and is generated before
315the llvm optimizer kicks in. The optimizer can't prove that profile counter
316instrumentation is safe to delete (because it's not: it affects the profile the
317program emits), and so leaves it alone.
318
319Note that this coverage feature does not rely on information that can degrade
320during the course of optimization, such as debug info line tables.
321
322Using the profiling runtime without static initializers
323=======================================================
324
325By default the compiler runtime uses a static initializer to determine the
326profile output path and to register a writer function. To collect profiles
327without using static initializers, do this manually:
328
329* Export a ``int __llvm_profile_runtime`` symbol from each instrumented shared
330  library and executable. When the linker finds a definition of this symbol, it
331  knows to skip loading the object which contains the profiling runtime's
332  static initializer.
333
334* Forward-declare ``void __llvm_profile_initialize_file(void)`` and call it
335  once from each instrumented executable. This function parses
336  ``LLVM_PROFILE_FILE``, sets the output path, and truncates any existing files
337  at that path. To get the same behavior without truncating existing files,
338  pass a filename pattern string to ``void __llvm_profile_set_filename(char
339  *)``.  These calls can be placed anywhere so long as they precede all calls
340  to ``__llvm_profile_write_file``.
341
342* Forward-declare ``int __llvm_profile_write_file(void)`` and call it to write
343  out a profile. This function returns 0 when it succeeds, and a non-zero value
344  otherwise. Calling this function multiple times appends profile data to an
345  existing on-disk raw profile.
346
347In C++ files, declare these as ``extern "C"``.
348
349Collecting coverage reports for the llvm project
350================================================
351
352To prepare a coverage report for llvm (and any of its sub-projects), add
353``-DLLVM_BUILD_INSTRUMENTED_COVERAGE=On`` to the cmake configuration. Raw
354profiles will be written to ``$BUILD_DIR/profiles/``. To prepare an html
355report, run ``llvm/utils/prepare-code-coverage-artifact.py``.
356
357To specify an alternate directory for raw profiles, use
358``-DLLVM_PROFILE_DATA_DIR``. To change the size of the profile merge pool, use
359``-DLLVM_PROFILE_MERGE_POOL_SIZE``.
360
361Drawbacks and limitations
362=========================
363
364* Prior to version 2.26, the GNU binutils BFD linker is not able link programs
365  compiled with ``-fcoverage-mapping`` in its ``--gc-sections`` mode.  Possible
366  workarounds include disabling ``--gc-sections``, upgrading to a newer version
367  of BFD, or using the Gold linker.
368
369* Code coverage does not handle unpredictable changes in control flow or stack
370  unwinding in the presence of exceptions precisely. Consider the following
371  function:
372
373  .. code-block:: cpp
374
375      int f() {
376        may_throw();
377        return 0;
378      }
379
380  If the call to ``may_throw()`` propagates an exception into ``f``, the code
381  coverage tool may mark the ``return`` statement as executed even though it is
382  not. A call to ``longjmp()`` can have similar effects.
383
384Clang implementation details
385============================
386
387This section may be of interest to those wishing to understand or improve
388the clang code coverage implementation.
389
390Gap regions
391-----------
392
393Gap regions are source regions with counts. A reporting tool cannot set a line
394execution count to the count from a gap region unless that region is the only
395one on a line.
396
397Gap regions are used to eliminate unnatural artifacts in coverage reports, such
398as red "unexecuted" highlights present at the end of an otherwise covered line,
399or blue "executed" highlights present at the start of a line that is otherwise
400not executed.
401
402Branch regions
403--------------
404When viewing branch coverage details in source-based file-level sub-views using
405``--show-branches``, it is recommended that users show all macro expansions
406(using option ``--show-expansions``) since macros may contain hidden branch
407conditions.  The coverage summary report will always include these macro-based
408boolean expressions in the overall branch coverage count for a function or
409source file.
410
411Branch coverage is not tracked for constant folded branch conditions since
412branches are not generated for these cases.  In the source-based file-level
413sub-view, these branches will simply be shown as ``[Folded - Ignored]`` so that
414users are informed about what happened.
415
416Branch coverage is tied directly to branch-generating conditions in the source
417code.  Users should not see hidden branches that aren't actually tied to the
418source code.
419
420
421Switch statements
422-----------------
423
424The region mapping for a switch body consists of a gap region that covers the
425entire body (starting from the '{' in 'switch (...) {', and terminating where the
426last case ends). This gap region has a zero count: this causes "gap" areas in
427between case statements, which contain no executable code, to appear uncovered.
428
429When a switch case is visited, the parent region is extended: if the parent
430region has no start location, its start location becomes the start of the case.
431This is used to support switch statements without a ``CompoundStmt`` body, in
432which the switch body and the single case share a count.
433
434For switches with ``CompoundStmt`` bodies, a new region is created at the start
435of each switch case.
436
437Branch regions are also generated for each switch case, including the default
438case. If there is no explicitly defined default case in the source code, a
439branch region is generated to correspond to the implicit default case that is
440generated by the compiler.  The implicit branch region is tied to the line and
441column number of the switch statement condition since no source code for the
442implicit case exists.
443