Skip to content

Latest commit

 

History

History
1650 lines (1247 loc) · 62.1 KB

cov_source.rst

File metadata and controls

1650 lines (1247 loc) · 62.1 KB
Raw

Source coverage analysis with |gcvcov|

Source coverage analysis computes metrics focused on source programming language entities such as high level statements or decisions (DO178 parlance for boolean expressions), which translate as :term:`Source Coverage Obligations`, or SCOs, in |gcp| terms.

Once you have produced source or binary traces, actual analysis is then performed with |gcvcov| to generate coverage reports. Source coverage is queried by passing a specific :option:`--level` argument to |gcvcov|.

The general structure of this command line is always like:

gnatcov coverage --level=<criterion> --annotate=<format>
                 <units-of-interest> ... <traces-or-checkpoints>

The next sections describe the :ref:`available report formats <sreport-formats>`, then provide more details regarding :ref:`scov-stmt`, :ref:`scov-dc`, and :ref:`scov-mcdc`.

The :ref:`sunits` chapter describes how application units to be considered for coverage assessment are to be specified.

|gcvcov| command line

.. index::
   single: gnatcov coverage command line for source coverage

Coverage analysis with |gcp| is performed by invoking |gcvcov| for a set of critera queried via the :option:`--level` command line option. The general interface synopsis is available from |gcv| :option:`--help`:

gnatcov coverage OPTIONS TRACE_FILES

The available options are as follows:

:option:`-c`, :option:`--level` |marg|:
Tell the set of coverage criteria to be assessed. The possible values for source coverage analysis are :option:`stmt`, :option:`stmt+decision`, :option:`stmt+mcdc`, and :option:`stmt+uc_mcdc`, all explained later in this chapter.
:option:`-a`, :option:`--annotate` |marg|:
Request a specific output report format. All the criteria support xcov[+], html[+], dhtml and report formats, with interpretations that vary depending on the assessed criteria. See the corresponding documentation later in this chapter for more details.
:option:`-o` :
Request that the synthetic report produced by --annotate=report be output in the provided filname instead of standard output by default. This is just ignored for other output formats.
:option:`-t`, :option:`--target` :
State the target architecture/board/ABI for which the analyzed program was built. This corresponds to the target prefix of your compilation toolchain, for example powerpc-elf or leon-elf, and can also be specified as a Target attribute within the project file designated by :option:`-P`, if any. By default, |gcv| assumes that this target is the same as the host environment. Stating the correct target is required for correct processing of project files.
:option:`--non-coverable`:
Report about language statements for which no object code could be found in the surrounding suprogram (typically out of optimization).
:option:`-P`:
Use the indicated project file as the root project to select the units of interest for the analysis and find default options. Default options are taken only from this project. In absence of :option:`--projects` and of :option:`--no-subprojects`, the units of interest are those designated by this project and all it's transitive dependencies, minus those advertised as externally-built. The target/runtime/scenario contextual options that would be passed to build the project should also be passed for proper interpretation of the project files. See the :ref:`sunits` chapter of this manual for more details.
:option:`--projects`, |rarg|:
When using :option:`-P`, use the provided projects to select units of interest, together with their transitive dependencies unless :option:`--no-subprojects` is also provided. The projects designated by this option must all be part of the import transitive closure reachable from the root project designated by :option:`-P`.
:option:`--no-subprojects`:
Consider only the projects (and units) encompassed by the :option:`-P` / :option:`--projects` options as of-interest. Don't include any project imported from there.
:option:`--units`, |rarg|:
When using project files, override the list of units of interest for source coverage with those provided.
:option:`--subdirs`:
When using project files, look for :term:`Library Information files` in the indicated subdirectory of each project's object directory.
:option:`--scos`, |rarg|:
Provide the set of :term:`Library Information files` from which Source Coverage Obligations (SCOs) should be loaded. This low-level switch effectively overrides the project based units of interest selection.
:option:`--ignore-source-files`, |rarg|:
Provide a list of globbing patterns (as in Unix shells) of source files to be excluded from the analysis and from the output report. See the :ref:`ada_subunits` section for more information.
:option:`--dump-units-to`:
For source coverage analysis specifically, output the names of units that are considered of-interest to the requested assessment, that is, for which a report or checkpoint is going to be produced. This also outputs, for each unit of interest, the list of files that were individually ignored using the Ignored_Source_Files project attribute or corresponding command-line option. Ignored source files listed with :option:`--dump-units-to` will be either marked as always ignored, if they were ignored in all the inputs (traces or checkpoints) that were used to produce this report, or as sometimes ignored, if the source files were ignored in at least one of the inputs of this report, but not all of them. The argument may be either the name of a file, clobbered it if it already exists, or '-' to request displaying the list on standard output. In the latter case, when a report output is also requested, the list of units is displayed as an additional report section.
:option:`--save-checkpoint`:
Save the resulting coverage analysis to the named checkpoint file.
:option:`--checkpoint`, |rarg|:
Load previously saved coverage analysis checkpoint(s), and continue coverage analysis from that initial state.

A lot of options are available to control the set of units for which coverage is to be assessed. They may be combined in multiple ways and attributed within the project files are available to refine the set of units to include or exclude from each designated project. See :ref:`sunits` for extra details, and :ref:`using-gpr` for a general overview of how the project facilities operate.

Saving coverage analysis state checkpoints allows the production of consolidated results from successive runs of the coverage command. In particular this allows coverage results to be computed incrementally, and allows consolidation with different sets of units of interest, in order to avoid incidental coverage. See :ref:`checkpoints` for a discussion of these use cases.

Positional arguments on the command line (not tied to a particular option) are considered as trace file arguments. At least one trace file is required for the coverage command to operate, which may but need not be introduced with :option:`-T` or :option:`--trace`. Here are a few examples of valid command lines to illustrate. Other examples will be exposed along the course of the following sections:

gnatcov coverage --level=stmt --scos=@alis --annotate=report --trace=prog.trace
#                      (a)         (b)              (c)            (d)
# (a) Request Statement coverage assessment,
# (b) for units associated with the ALI files listed in the "alis" text file,
# (c) producing a synthetic text report on standard output (no -o option),
# (d) out of a single execution trace "prog.trace".

gnatcov coverage --level=stmt+decision --scos=@alis --annotate=html t1 t2
# Statement and Decision coverage assessments for two traces "t1" and "t2",
# producing html report files in the current directory.

gnatcov coverage --level=stmt+decision --scos=@alis --annotate=html @mytraces
# Same report, with t1 and t2 listed in the "mytraces" text file

gnatcov coverage --level=stmt -Papp.gpr --annotate=html @mytraces
# Same kind of report, for Statement coverage only, on source units owned
# by "app.gpr" and its transitive closure of project dependencies.

gnatcov coverage --level=stmt -Papp.gpr --no-subprojects --annotate=html @mytraces
# Likewise, considering only the units owned by app.gpr

Output report formats (:option:`--annotate`)

Source coverage reports may be produced in various formats, as requested with the :option:`--annotate` command line argument of |gcvcov|. The :option:`xcov`, :option:`html` and :option:`dhtml` formats produce a set of annotated source files, in the directory where |gcv| is launched unless overriden with a :option:`--output-dir` option. The :option:`report` output consists in a synthetic text report of :term:`coverage violations` with respect to the requested criteria, produced on standard output by default or in the file specified by the :option:`-o` command line option.

Later in this chapter we name output formats by the text to add to :option:`--annotate` on the command line. For example, we use "the :option:`=report` outputs" to mean "the coverage reports produced with :option:`--annotate=report`".

We will illustrate the various formats with samples extracted from outputs obtained by perfoming coverage analysis of the following example Ada application unit:

function Between (X1, X2, V : Integer) return Boolean;
--  Is V between X1 and X2, inclusive and regardless of their ordering?

function Between (X1, X2, V : Integer) return Boolean is
begin
   if X1 < X2 then
      return V >= X1 and then V <= X2;
   else
      return V >= X2 and then V <= X1;
   end if;
end Between;

Annotated sources, text (:option:`=xcov[+]`)

For source coverage criteria, |gcvcov| :option:`--annotate=xcov` produces an annotated version of each source file, in text format, named after the original source with an extra .xcov extension at the end (x.ext.xcov for a source named x.ext).

Each annotated source contains a global summary of the assessment results followed by the original source lines, all numbered and marked with a coverage annotation next to the line number. The annotation on a line always consists in a single character, which may be one of the following:

.. tabularcolumns:: cl
Annotation Meaning
. No coverage obligation is attached to the line
- Coverage obligations attached to the line, none satisfied
! Coverage obligations attached to the line, some satisfied
+ Coverage obligations attached to the line, all satisfied

Here is, to illustrate, the full statement coverage report produced for our example unit when the Between function was called so that the if control evaluated True only. The function is actually part of an Ada package, called Ranges, with an original body source file named ranges.adb:

examples/src/ranges.adb:
67% of 3 lines covered
Coverage level: stmt
  1 .: package body Ranges is
  2 .:    function Between (X1, X2, V : Integer) return Boolean is
  3 .:    begin
  4 +:       if X1 < X2 then
  5 +:          return V >= X1 and then V <= X2;
  6 .:       else
  7 -:          return V >= X2 and then V <= X1;
  8 .:       end if;
  9 .:    end;
 10 .: end;

:option:`--annotate=xcov+` (with a trailing +) works the same, only providing extra details below lines with improperly satisfied obligations. The available details consists in the list of :term:`coverage violations` diagnosed for the line, which depends on the coverage criteria involved. Here is an excerpt for our previous example, where the only improperly satisfied obligation is an uncovered statement on line 7:

7 -:          return V >= X2 and then V <= X1;
STATEMENT "return V ..." at 7:10 not executed

Annotated sources, html (:option:`=html[+]`)

For source coverage criteria, |gcvcov| :option:`--annotate=html` produces an annotated version of each source file, in html format, named after the original source with an extra .html extension at the end. Each annotated source page contains a summary of the assessment results followed by the original source lines, all numbered and marked with a coverage annotation as in the :option:`--annotate=xcov` case. Lines with obligations are colorized in green, orange or red for +, ! or - coverage respectively.

An index.html page is also produced, which contains a summary of the assessment context (assessed criteria, trace files involved, ...) and of the coverage results for all the units, with links to their annotated sources. See our :ref:`sample html index <sample_sc_html_index>` appendix for an example index page, which embeds a self-description of all the items it contains. See the :ref:`sample annotated source <sample_sc_html_unit>` appendix for a sample of html annotated source.

The page style is governed by a set of Cascading Style Sheet (CSS) parameters, fetched from a xcov.css file in the directory where |gcv| is launched. If this file is available when |gcv| starts, |gcv| uses it so users may setup a customized version if needed. If the file is not available, |gcv| creates a default one.

Similarily to the :option:`xcov` format case, :option:`--annotate=html+` (with a trailing +) adds details about improperly satisfied obligations. In the html version, these extra details are initially folded within their associated line and expanded by a mouse click on the line.

Annotated sources, dynamic html (:option:`=dhtml`)

:option:`--annotate=dhtml` produces a dynamic html output, which essentially features:

  • A more modern look & feel compared to the :option:`html` formats described earlier,
  • The ability to sort indexes by clicking on column headers, allowing for example sorts keyed on unit names or on relative coverage achievement,
  • Per-project indexes on the root page when :option:`-P` was used to designate the source units of interest.

The option produces a set of .js javascript files implementing most of the report displays and interactions, as well as an index.html root page which users should open as an entry point to the report contents.

The per-line details that differentiates :option:`html+` from :option:`html` are always produced, initially folded and available on line clicks as well.

Violations summary, text (:option:`=report`)

For source coverage criteria, |gcvcov| :option:`--annotate=report` produces a summary that lists all the :term:`coverage violations` (failure to satisfy some aspect of a coverage criterion) relevant to the set of assessed criteria.

The report features explicit start/end of report notifications and at least three sections in between: Assessment Context, Coverage Violations, and Analysis Summary. A few variations are introduced when :term:`exemption regions` are in scope. See the :ref:`exemptions` section for more details on their use and effect on the output reports.

If :option:`--dump-units-to -` is also on the command line, a UNITS OF INTEREST section is produced, which contains the list of units considered of-interest for the reported assessment, as well as the list of source files individually ignored with the Ignored_Source_Files project attribute and corresponding command-line option.

Assessment Context

The Assessment Context report section exposes the following information items:

  • Date & time when the report was produced
  • Command line and Version of |gcp| that produced the report. The set of units that the report is about is conveyed by the command line switches summarized there (:option:`--projects`, :option:`--units`, :option:`--scos`).
  • Coverage level requested to be analyzed
  • Details on the input trace files: path to binary program exercised (as recorded in the trace header), production time stamp and tag string (:option:`--tag` command line argument value).

Here is a example excerpt:

===========================
== 1. ASSESSMENT CONTEXT ==
===========================

Date and time of execution: 2011-11-24 16:33:44.00
Tool version: GNATcoverage 1.0.0w (20111119)

Command line:

gnatcov coverage -Pmytest.gpr --level=stmt+mcdc --annotate=report test_x1x2.trace

Coverage level: stmt+mcdc

Trace files:

test_x1x2.trace
  program: obj/test_x1x2
  date   : 2011-11-24 15:33:44
  tag    : sample run

Coverage Violations

The Coverage Violations report section lists and counts the coverage violations that relate to source lines not part of an exemption region. The violations are grouped in subsections, one per assessed criterion according to the :option:`--level` option:

.. tabularcolumns:: ll
:option:`--level=` Assessed criteria / Report subsections
stmt Statement Coverage
stmt+decision Statement and Decision Coverage
stmt+mcdc Statement, Decision and MCDC Coverage

All the violations are reported using a consistent format, as follows:

ranges.adb:7:10: statement not executed
  source  :sloc: violation description

source and sloc are the source file basename and the precise line:column location within that source where the violation was detected.

The following table summarizes the list of violation items that might be emitted together for each criterion:

.. tabularcolumns:: ll
Criterion Possible violations
Statement Coverage statement not executed
Decision Coverage decision outcome TRUE not covered
  decision outcome FALSE not covered
  one decision outcome not covered
MCDC Coverage all the decision coverage items, plus ...
  condition has no independent influence pair

When multiple violations apply someplace, only the most basic diagnostic is emitted, not the more precise ones corresponding to stricter criteria. For instance, if an Ada statement like X := A and then B; is not covered at all, a statement not executed violation is always emitted alone, even when assessing :option:`--level=stmt+mcdc` and we also have improper decision and conditions coverage.

Here is an output excerpt for our example with :option:`--level=stmt+mcdc`, producing one subsection for each of the three criteria requested at that level:

============================
== 2. COVERAGE VIOLATIONS ==
============================

2.1. STMT COVERAGE
------------------

ranges.adb:7:10: statement not executed

1 violation.

2.2. DECISION COVERAGE
----------------------

ranges.adb:4:10: decision outcome FALSE never exercised

1 violation.

2.3. MCDC COVERAGE
------------------

ranges.adb:5:17: condition has no independent influence pair, MC/DC not achieved

1 violation.

Analysis Summary

The Analysis Summary report section summarizes just the counts reported in each of the previous sections. For our example report so far, this would be:

=========================
== 3. ANALYSIS SUMMARY ==
=========================

1 non-exempted STMT violation.
1 non-exempted DECISION violation.
1 non-exempted MCDC violations.

This section provides a quick way to determine whether the requested coverage level is fully satisfied, with details available from the per criterion sections that precede.

Statement Coverage analysis (:option:`--level=stmt`)

|gcv| performs Statement Coverage assessments with the :option:`--level=stmt` command line option.

In synthetic :option:`=report` outputs, unexecuted source statements are listed as Statement Coverage violations in the report section dedicated to these.

In annotated source outputs, the coverage annotations convey the following indications:

.. tabularcolumns:: cl
Annotation Meaning
- At least one statement on the line, none covered
! At least one statement on the line, some covered
+ At least one statement on the line, all covered

When a single statement spans multiple lines, the coverage annotation is present on all the lines, as the two + signs for the single assignment in the following excerpt:

3 .:  -- A single assignment spanning two lines:
4 +:  Result :=
5 +:     Input1 * Input2;

For compound statements, the coverage status of the compound construct per se is reported only on the parts that embed flow control expressions. For an Ada if statement, for example, coverage is reported on the if or elsif lines only, not on the else, or end if; lines, and not on lines where inner statements reside. The lines where inner statements reside are annotated in accordance with the nature and coverage status of those statements only. For example, see the . annotations on lines 4 and 6 in:

2 +:  if This_Might_Not_Be_True then
3 -:     Result := -1;
4 .:  else
5 +:     Result := 12;
6 .:  end if;

Declarations are generally considered as statements, so are reported covered/uncovered when they have initialization code associated with them.

Finally, a statement is considered covered as soon as part of the associated machine code is executed, in particular even when the statement execution is interrupted somehow, for example by an exception occurrence. For instance, the statement below:

X := Function_That_Raises_Exception (Y) + Z;

Will be reported as covered as soon as it is reached, even if the expression evaluation never really terminates.

Note that if no executable code for a given unit can be found in any of the executables submitted to gnatcov, then all statements in the unit will be conservatively reported as not covered. This ensures that if tests for an entire unit have been omitted from a test campaign, a violation will be properly reported. Such violations can be suppressed either using exemptions, or by removing the unit from the list of units of interest.

Example program and assessments

To illustrate the just presented points further, we consider the example functional unit below, with the spec and body stored in source files named div_with_check.ads and div_with_check.adb:

function Div_With_Check (X, Y : Integer) return Integer;
--  return X / Y if Y /= 0. Raise Program_Error otherwise

function Div_With_Check (X, Y : Integer) return Integer is
begin
   if Y = 0 then
      raise Program_Error;
   else
      return X / Y;
   end if;
end;

We first exercise the function for Y = 1 only, using the following :term:`test driver` in test_div1.adb:

procedure Test_Div1  is
   X : constant Integer := 4;
begin
   Assert (Div_With_Check (X, 1) = X);
end;

From a test_div1.trace obtained with |gcvrun|, we analyze for the Statement Coverage criterion using the following |gcvcov| invocation:

gnatcov coverage --level=stmt --scos=div_with_check.ali --annotate=xcov test_div1.trace

We get an :option:`=xcov` annotated source result in text format for the functional unit on which the analysis is focused, in div_with_check.adb.xcov:

examples/src/div_with_check.adb:
67% of 3 lines covered
Coverage level: stmt
  1 .: function Div_With_Check (X, Y : Integer) return Integer is
  2 .: begin
  3 +:    if Y = 0 then
  4 -:       raise Program_Error;
  5 .:    else
  6 +:       return X / Y;
  7 .:    end if;
  8 .: end;

We can observe that:

  • Only the if line of the compound if statement is annotated, as covered since the function was called.
  • The inner raise and return statements are marked uncovered and covered respectively, as expected since the function was only called with arguments for which the if controling decision evaluates False.

As a second experiment, we exercise the function for Y = 0 only, using:

procedure Test_Div0  is
   Result : Integer
     := Div_With_Check (4, 0);
begin
   Put_Line ("R = " & Integer'Image (Result));
end;

We request results on the test driver as well this time, as it features constructs of relevance for our purposes:

gnatcov coverage --level=stmt -Pmytest.gpr --annotate=xcov test_div0.trace

The :option:`=xcov` outputs follow. First, for the functional unit, with the if statement coverage reversed compared to the previous testcase:

1 .: function Div_With_Check (X, Y : Integer) return Integer is
2 .: begin
3 +:    if Y = 0 then
4 +:       raise Program_Error;
5 .:    else
6 -:       return X / Y;
7 .:    end if;
8 .: end;
9 .:

Then, for the test driver where we can note that

  • The two lines of the local Result definition are annotated,
  • This definition is marked covered even though it was evaluated only once with an initialization expression that raised an exception, and
  • The driver body is reported uncovered, as expected since an exception triggered during the elaboration of the subprogram declarative part.
67% of 3 lines covered
Coverage level: stmt
   1 .: with Div_With_Check, Ada.Text_IO; use Ada.Text_IO;
   2 .:
   3 .: procedure Test_Div0  is
   4 +:    Result : Integer
   5 +:      := Div_With_Check (4, 0);
   6 .: begin
   7 -:    Put_Line ("R = " & Integer'Image (Result));
   8 .: end;

The corresponding synthetic report is simply obtained by running |gcvcov| again with :option:`--annotate=report` instead of :option:`--annotate=xcov`:

===========================
== 1. ASSESSMENT CONTEXT ==
===========================
...

============================
== 2. COVERAGE VIOLATIONS ==
============================

2.1. STMT COVERAGE
------------------
div_with_check.adb:6:7: statement not executed
test_div0.adb:7:4: statement not executed

2 violations.

=========================
== 3. ANALYSIS SUMMARY ==
=========================

2 STMT violations.

We can see here that the two lines marked - in the :option:`=xcov` outputs are properly reported as violations in the STMT COVERAGE section of this report, and that this section is the only one presented in the COVERAGE VIOLATIONS part, as only this criterion was to be analyzed per the :option:`--level=stmt` argument.

Decision Coverage analysis (:option:`--level=stmt+decision`)

With the :option:`--level=stmt+decision` command line option, |gcv| performs combined Statement and Decision Coverage assessments.

In this context, we consider to be a :dfn:`decision` any Boolean expression used to influence the control flow via explicit constructs in the source program, such as if statements or while loops, regardless of the type of this expression. This may be of essentially any type in C, and subtypes or types derived from the fundamental Boolean type in Ada.

A decision is said :dfn:`fully covered`, or just :dfn:`covered`, as soon as it has been evaluated at least once True and once False during the program execution. If only one of these two possible outcomes was exercised, the decision is said :dfn:`partially covered`.

A decision is also said :dfn:`partially covered` when none of the possible outcomes was exercised, which happens when the enclosing statement was not executed at all or when all the attempted evaluations were interrupted e.g. because of exceptions. In the former case, when a decision is part of a statement and the statement is not executed at all, only the statement level violation is reported. The nested decision level violations are implicit in this case and diagnosing them as well would only add redundancy.

The :option:`=report` synthetic output lists the statement and decision coverage violations in the STMT and DECISION coverage report section respectively.

For the :option:`=xcov` and :option:`=html` annotated-source oriented formats, the single annotation produced on each source line combines the statement and decision coverage indications. The following table summarizes the meaning of the possible annotations:

.. tabularcolumns:: cl
Annotation Meaning
- Statement on the line was not executed
! At least one decision partially covered on the line
+ All the statements and decisions on the line are covered

When a trailing + is added to the format passed to :option:`--annotate` (:option:`=xcov+` or :option:`=html+`), a precise description of the actual violations is available for each line in addition to the annotation.

Example program and assessments

To illustrate, we consider the example functional Ada unit below, with the spec and body stored in source files named divmod.ads and divmod.adb:

procedure Divmod
  (X, Y : Integer; Value : out Integer;
   Divides : out Boolean; Tell : Boolean);
--  Compute X / Y into VALUE and set DIVIDES to indicate
--  whether  Y divides X. Output a note to this effect when
--  requested to TELL.

procedure Divmod
  (X, Y : Integer; Value : out Integer;
   Divides : out Boolean; Tell : Boolean) is
begin
   if X mod Y = 0 then
      Divides := True;
      if Tell then
         Put_Line (Integer'Image(Y) & " divides " & Integer'Image(X));
      end if;
   else
      Divides := False;
   end if;

   Value := X / Y;
end Divmod;

We first experiment with the following test driver:

procedure Test_Divmod2  is
   Value : Integer;
   Divides : Boolean;
begin
   Divmod (X => 5, Y => 2, Value => Value,
           Divides => Divides, Tell => True);
   Assert (Divides = False);

   Divmod (X => 6, Y => 2, Value => Value,
           Divides => Divides, Tell => True);
   Assert (Divides = True);
end Test_Divmod2;

This exercises the Divmod function twice. The outer if construct executes both ways and the if Tell then test runs once only for Tell True. As a result, the only :option:`stmt+decision` violation by our driver is the Tell decision coverage, only partially achieved since we have only exercised the True case. This is confirmed by :option:`=report` excerpt below, where we find the two violations sections in accordance with the requested set of criteria:

2.1. STMT COVERAGE
------------------

No violation.

2.2. DECISION COVERAGE
----------------------

divmod.adb:14:10: decision outcome FALSE never exercised

1 violation.

For :option:`--annotate=xcov`, this translates as a single partial coverage annotation on the inner if control line:

 8 .: procedure Divmod
 9 .:   (X, Y : Integer; Value : out Integer;
10 .:    Divides : out Boolean; Tell : Boolean) is
11 .: begin
12 +:    if X mod Y = 0 then
13 +:       Divides := True;
14 !:       if Tell then
15 +:          Put_Line (Integer'Image (Y) & " divides " & Integer'Image (X));
16 .:       end if;
17 .:    else
18 +:       Divides := False;
19 .:    end if;
20 .:
21 +:    Value := X / Y;
22 .: end Divmod;

Now we exercise with another test driver:

procedure Test_Divmod0  is
   Value : Integer;
   Divides : Boolean;
begin
   Divmod (X => 5, Y => 0, Value => Value,
           Divides => Divides, Tell => True);
end Test_Divmod0;

Here we issue a single call passing 0 for the Y argument, which triggers a check failure for the mod operation.

This results in the following :option:`=xcov` output:

 8 .: procedure Divmod
 9 .:   (X, Y : Integer; Value : out Integer;
10 .:    Divides : out Boolean; Tell : Boolean) is
11 .: begin
12 !:    if X mod Y = 0 then
13 -:       Divides := True;
14 -:       if Tell then
15 -:          Put_Line (Integer'Image (Y) & " divides " & Integer'Image (X));
16 .:       end if;
17 .:    else
18 -:       Divides := False;
19 .:    end if;
20 .:
21 -:    Value := X / Y;
22 .: end Divmod;

We have an interesting situation here, where

  • The outer if statement is reached and covered (as a statement),
  • No evaluation of the X mod Y = 0 decision terminates, because the only attempted computation is interrupted by an exception, so none of the other statements is ever reached.

This gets all confirmed by the :option:`=report` output below, on which we also notice that the only diagnostic emitted for the uncovered inner if on line 14 is the statement coverage violation:

2.1. STMT COVERAGE
------------------
divmod.adb:13:7: statement not executed
divmod.adb:14:7: statement not executed
divmod.adb:15:10: statement not executed
divmod.adb:18:7: statement not executed
divmod.adb:21:4: statement not executed

5 violations.

2.2. DECISION COVERAGE
----------------------
divmod.adb:12:7: decision never evaluated

1 violation.

Modified Condition/Decision Coverage analysis (:option:`--level=stmt+mcdc`)

|gcv| performs combined Statement and Modified Condition/Decision Coverage analysis with the :option:`--level=stmt+mcdc` option passed to |gcvcov|. In addition to this particular :option:`--level` option, you also need to tell |gcvrun| the list of units on which MCDC analysis will be performed. See the :ref:`trace-control` section for more details on this aspect of the procedure.

Compared to Decision Coverage, MCDC analysis incurs two important differences:

  • In addition to expressions that pilot an explicit control-flow construct, we treat as decisions all the Boolean expressions that combine operands with short-circuit logical operators, such as the expression on the right hand side of the assignment in X := A and then B; More details on the identification of decisions, together with extra examples, are provided in the :ref:`mcdc-decisions` section of this chapter
  • For each decision in the sources of interest, testing shall demonstrate the :dfn:`independant influence` of every operand (:term:`conditions` in the DO-178 parlance) in addition to just exercising the True/False outcomes of the expression as a whole. The :ref:`mcdc-variants` section that follows expands on the notion of :dfn:`independant influence` and on possible variations of the MCDC criterion definition.

Output-wise, the source annotations for the :option:`=xcov` or :option:`=html` formats are the same as for decision coverage, with condition specific cases marked with a ! as well:

.. tabularcolumns:: cl
Annotation Meaning
- Statement on the line was not executed
! At least one decision/condition partially covered on the line
+ All the statements and decisions/conditions on the line are covered

The :option:`=report` outputs feature an extra MCDC section in the Coverage Violations segment, which holds:

  • The condition specific diagnosics (independent influence not demonstrated), as well as
  • Decision level diagnostics (such as decision outcome True not covered messages) for the Complex Boolean Expressions not directing a control-flow oriented statement and which we treat as decisions nevertheless.

There again, condition or decision related messages are only emitted when no more general diagnostic applies on the associated entity. Condition specific diagnostics, for example, are only produced in absence of enclosing statement or decision level violation.

See the :ref:`mcdc-examples` section of this chapter for a few illustrations of these points.

MCDC variants

Compared to Decision Coverage, achieving MCDC requires tests that demonstrate the independent influence of conditions in decisions. Several variants of the criterion exist.

:dfn:`Unique Cause MCDC` is the original criterion described in the DO178B reference guidelines, where independent influence of a specific condition must be demonstrated by a pair of tests where only that condition changes and the decision value toggles.

Consider the following table which exposes the 4 possible condition/decision vectors for the A and then B expression, where T stands for True, F stands for False, and the italics indicate that the condition evaluation is short-circuited:

.. tabularcolumns:: |c|cc|c|
# A B A and then B
1 T T T
2 T F F
3 F T F
4 F F F

Each line in such a table is called an :term:`evaluation vector`, and the pairs that demonstrate the independant effect of conditions are known as :term:`independence pairs`.

Evaluations 1 + 3 constitute a Unique Cause independence pair for A, where A changes, B does not, and the expression value toggles. 1 + 2 constitues a pair for B.

The closest criterion supported by |gcp| is a very minor variation where conditions that are not evaluated due to short-circuit semantics are allowed to differ as well in a pair. Indeed, their value change cannot possibly have influenced the decision toggle (since they are not even considered in the computation), so they can never invalidate the effect of another condition.

We call this variation :dfn:`Unique Cause + Short-Circuit MCDC`, activated with :option:`--level=stmt+uc_mcdc` on the command line. From the A and then B table just introduced, 4 + 1 becomes another valid independence pair for A, as B is not evaluated at all when A is False so the change on B is irrelevant in the decision switch.

:option:`--level=stmt+mcdc` actually implements another variant, known as :dfn:`Masking MCDC`, accepted as a sound alternative and offering improved support for coupled conditions.

Masking MCDC allows even further flexibility in the possible variations of conditions in an independence pair. Indeed, as soon as only short-circuit operators are involved, all the conditions that appear on the left of a given condition in the expression text are allowed to change without invalidating the said condition influence demonstration by a pair.

Example program and assessments

We reuse one of our previous examples to illustrate, with a simple functional unit to exercise:

function Between (X1, X2, V : Integer) return Boolean;
--  Whether V is between X1 and X2, inclusive and however they are ordered

function Between (X1, X2, V : Integer) return Boolean is
begin
   if X1 < X2 then
      return V >= X1 and then V <= X2;
   else
      return V >= X2 and then V <= X1;
   end if;
end Between;

First consider the following test driver, which exercises only a single case where X1 < V < X2:

procedure Test_X1VX2 is
begin
   Assert (Between (X1 => 2, X2 => 5, V => 3)); -- X1 < V < X2
end Test_X1VX2;

Performing MCDC analysis requires the execution step to be told about it, by providing both the :option:`--level` and a list of units for which analysis is to be performed to |gcvrun| (see the :ref:`trace-control` for details):

gnatcov run --level=stmt+mcdc -Pmytest.gpr test_x1vx2

We first request an :option:`=xcov+` report to get a first set of results, in the ranges.adb.xcov annotated source:

gnatcov coverage --level=stmt+mcdc -Pmytest.gpr --annotate=xcov+ test_x1vx2.trace

...
   8 .:    function Between (X1, X2, V : Integer) return Boolean is
   9 .:    begin
  10 !:       if X1 < X2 then
DECISION "X1 < X2" at 10:10 outcome FALSE never exercised
  11 !:          return V >= X1 and then V <= X2;
DECISION "V >= X1 a..." at 11:17 outcome FALSE never exercised
  12 .:       else
  13 -:          return V >= X2 and then V <= X1;
STATEMENT "return V ..." at 13:10 not executed
  14 .:       end if;
  15 .:    end Between;

This is all as expected from what the driver does, with a few points of note:

  • The diagnostic on line 11 confirms that Complex Boolean Expression are treated as decisions even when not used to direct a conditional control-flow statement. The expression is indeed used here as a straight, unconditional return statement value;
  • Only the decision level violations are emitted for lines 10 and 11. The independant influence of the conditions is not demonstrated but this is implicit from the decision partial coverage so is not notified;
  • Similarily, only the statement level violation is emitted for line 13, eventhough there are decision and condition level violations as well.

Another aspect of interest is that we have partial decision coverage on two kinds of decisions (one control-flow decision controling the if, and another one used a straight return value), and this distinction places the two decision outcome FALSE never exercised violations in distinct sections of the :option:`=report` output:

2.1. STMT COVERAGE
------------------
ranges.adb:13:10: statement not executed

2.2. DECISION COVERAGE
----------------------
ranges.adb:10:10: decision outcome FALSE never exercised

2.3. MCDC COVERAGE
------------------
ranges.adb:11:17: decision outcome FALSE never exercised

Now running another test driver which exercises two cases where X1 < X2:

procedure Test_X1VX2V is
begin
   Assert (Between (X1 => 2, X2 => 5, V => 3)); -- X1 < V < X2
   Assert (not Between (X1 => 2, X2 => 5, V => 8)); -- X1 < X2 < V
end;

The first return expression is valued both ways so we get an example of condition specific diagnostic on line 11:

   8 .:    function Between (X1, X2, V : Integer) return Boolean is
   9 .:    begin
  10 !:       if X1 < X2 then
DECISION "X1 < X2" at 10:10 outcome FALSE never exercised
  11 !:          return V >= X1 and then V <= X2;
CONDITION "V >= X1" at 11:17 has no independent influence pair, MC/DC not achieved
  ...

Indeed, looking at an evaluation table for the first return decision:

.. tabularcolumns:: |c|cc|c|c|
# A: V >= X1 B: V <= X2 A and then B Case
1 T T T X1 < V < X2
2 T F F X1 < X2 < V
3 F T F  
4 F F F  

We observe that our driver exercises vectors 1 and 2 only, where:

  • The two evaluations toggle the decision and the second condition only, so achieve decision coverage and demonstrate that condition's independant influence;
  • The first condition (V >= X1) never varies so the independant influence of this condition isn't demonstrated.

As we mentioned in the discussion on MCDC variants, adding vector 3 would achieve MCDC for this decision. Just looking at the table, adding vector 4 instead would achieve MCDC as well since the second condition is short-circuited so its value change is not relevant. The condition expressions are such that running vector 4 is not possible, however, since we can't have V both < X1 (condition 1 False) and V > X2 (condition 2 False) at the same time when X1 < X2.

Decision composition rules for MCDC

For MCDC analysis purposes, we treat as decisions two categories of expressions:

  • As for the :dfn:`decision coverage` criterion, all the expressions that directly influence control-flow constructs and which we will call :dfn:`control-flow expressions`,
  • All the expressions obtained by composition of short-circuit logical operators, and-then and or-else for Ada, && and || for C.

The most straightforward examples of non control-flow expressions treated as decisions for MCDC are the logical expressions appearing in contexts such as the right-hand side of assignments. For example:

Valid_Data := Sensor_OK and then Last_Sensor_Update_OK; -- 1 decision here
need_update = (sensor != NULL && sensor->invalid);   /* 1 decision here */

Non short-circuit binary operators, when allowed by the coding standard, are taken as regular computational devices and may either participate in the construction of operands or split an expression into multiple decisions. For instance, the following C excerpt:

return !(x & 0x3) && !(y & 0x3);  /* 1 decision here */

produces a single decision with two bitwise & operands. And the following Ada excerpt:

if ((A and then not B) == (C or else (D and then E))) then -- 3 decisions here

produces three decisions: (A and then not B), 2 operands combined with short-circuit and-then, (C or else (D and then E))), 3 operands combined with short-circuit and-then and or-else, and the whole toplevel expression controlling the if statement.

In C as in Ada, logical negation is allowed anywhere and just participates in the operands construction without influencing decision boundaries.

Non short-circuit binary operators in logical expressions might complexify the identification of decision boundaries for users. |gnat| compilers offer two devices to alleviate this for Ada:

  • The No_Direct_Boolean_Operator restriction pragma, which will trigger compilation errors on the use of non short-circuit Boolean operators and facilitates the enforcement of coding standards prohibiting such uses.
  • The Short_Circuit_And_Or pragma, which directs the compiler to translate non-short circuit and/or operators as their short-circuit counterparts.

There is no equivalent in C, where the allowed operand types are much more varied and where the restriction would make the language really much harder to use.

Ada subunits ("separates")

Subunits, declared with a separate keyword and implemented in a separate source file, are compiled as part of their parent and are not considered as units on their own. Only the parent name has an effect in the coverage analysis scope specifications and it denotes the set of sources involved in the entire unit implementation, subunit sources included.

However it is quite common to use subunits as a mean to do unit testing: a subunit is physically separated from other sources and can have access to implementation internals. Such subunits vary from one test to another and thus interfere with the consolidation process. For this specific use case, the :option:`--ignore-source-files` command-line argument for |gcvcov| makes it possible for the coverage analysis and the report production to ignore source files even though they belong to units of interest.

This option can appear multiple times on the command line. Each occurrence expects a single argument which is either a globbing pattern for the name of source file to ignore, or a :term:`@listfile argument` that contains a list of such patterns.

For instance, consider the spec and body for the Ops unit (ops.ads an ops.adb) with the body containing a subunit subprogram Ops.Test (ops-test.adb). In order to perform a coverage analysis on the Ops unit excluding the Ops.Test subunit, one must run:

gnatcov coverage [regular options] --units=ops --ignore-source-files=ops-test.adb [trace files]

In order to ignore all files whose name match *-test.adb, you can also run:

gnatcov coverage [regular options] --units=ops --ignore-source-files=*-test.adb [trace files]

Inlining & Ada generic units

In the vast majority of situations, inlining is just transparent to source coverage metrics: calls are treated as regular statements, and coverage of the inlined bodies is reported on the corresponding sources regardless of their actual inlining status. See the :ref:`optimization` section for a description of effects that might show up on rare occasions.

Generic units have a different status, where each instantiation produces a distinct, potentially unique instance of the generic code working on a different types, objects, with different helper subprograms or packages.

|gcp| offers two main kinds of coverage strategies for generic units: one where the generic source is considered as the entity of interest, to which all the instances contribute, and one where each instance is considered as a separate entity of interest.

Combined coverage on generics (default behavior)

By default, Ada generic units are also uniformly treated as single source entities, with the coverage achieved by all the instances combined and reported against the generic source only, not for each individual instance.

Consider the following functional Ada generic unit for example. It provides a simple vector type abstraction on which two operations are available; Inc adds some amount to each element of a vector, and Mult multiplies each element by some amount. The exposed type is of fixed size, provided as a parameter:

generic                               -- vops.ads
   Size : in Integer;
package Vops is
   type Vector_Type is array (1 .. Size) of Integer;

   procedure Inc (V : in out Vector_Type; Amount : Integer);
   procedure Mult (V : in out Vector_Type; Amount : Integer);
end;

package body Vops is                  -- vops.adb

   procedure Inc (V : in out Vector_Type; Amount : Integer) is
   begin
      for I in V'Range loop
         V(I) := V(I) + Amount;
      end loop;
   end;

   procedure Mult (V : in out Vector_Type; Amount : Integer) is
   begin
      for I in V'Range loop
         V(I) := V(I) * Amount;
      end loop;
   end;
end;

Now consider this test, checking operations on vectors of different sizes, from two instances of the Vops unit:

with Vops;                            -- v5.ads
package V5 is new Vops (Size => 5);

with Vops;                            -- v8.ads
package V8 is new Vops (Size => 8);

with V5, V8;                          -- test_5inc_8mult.adb
procedure Test_5inc_8mult is
   V5o : V5.Vector_Type := (others => 1);
   V8o : V8.Vector_Type := (others => 2);
begin
   V5.Inc (V5o, 3);
   V8.Mult (V8o, 2);
end;

Only the Inc subprogram is called through the V5 instance and only the Mult subprogram is called through the V8 instance. Both suprograms are nevertheless called overall, so the Vops package body is claimed fully covered by default:

gnatcov coverage -Pvops.gpr --level=stmt --annotate=xcov test_5inc_8mult.trace
...
100% of 4 lines covered
Coverage level: stmt
  1 .: package body Vops is
  2 .:
  3 .:    procedure Inc (V : in out Vector_Type; Amount : Integer) is
  4 .:    begin
  5 +:       for I in V'Range loop
  6 +:          V(I) := V(I) + Amount;
  7 .:       end loop;
  8 .:    end;
  9 .:
 10 .:    procedure Mult (V : in out Vector_Type; Amount : Integer) is
 11 .:    begin
 12 +:       for I in V'Range loop
 13 +:          V(I) := V(I) * Amount;
 14 .:       end loop;
 15 .:    end;
 16 .: end;

Per instance analysis is possible though, as part of what we refer to as :dfn:`separated coverage` facilities.

Separated coverage on generics

As described above, a single coverage analysis of any source construct is performed by default, consolidating all code copies generated by this construct. For subprograms, this means consolidation over all inlined copies. For generic units, consolidation over all instances.

A finer-grained analysis is possible, where distinct copies of the code coming from a given source construct are identified according to some criterion, and a separate coverage assessment is made for each of these copies.

In this case, coverage violations carry an additional indication of which code copy the violation is reported for, available in all but the non-extended xcov and html output formats. The non-extended xcov and html formats simply convey partial coverage achievement on a line as soon one violation get reported for an obligation on that line, regardless of which copy the violation originates from.

|gcv| supports different modes for such analyses, detailed in the following subsections.

Separation by instance (:option:`-S instance`)

In this mode, two code regions coming from the same source construct will undergo separate coverage analyses if they come from different generic instances, identified by the instanciation source location.

For our Vops example, selecting an output format where the violations detailed are exposed, this translates as:

gnatcov coverage -Pvops.gpr --annotate=report -S instance [...]
...
vops.adb:5:11: statement not executed (from v8.ads:2:1)
vops.adb:6:10: statement not executed (from v8.ads:2:1)
vops.adb:12:11: statement not executed (from v5.ads:2:1)
vops.adb:13:10: statement not executed (from v5.ads:2:1)

We do observe violations on the Vops generic body, fully covered without :option:`-S instance`. This is the outcome of an analysis conducted on the two generic instances separately, each designated by a (from <instantiation source location>) indication.

|gcv| needs to see the coverage obligations correponding to each instance in this mode. This is achieved transparently by the use of a project file in the example command lines we quoted and needs particular care when the Library Information files are provided manually with :option:`--scos` instead.

Indeed, even if we aim at getting coverage results for the vops.adb source, passing :option:`--scos=vops.ali` alone isn't enough when per instance separate analysis is desired. Separate coverage analysis for the instances entails coverage obligations for the instances, and this requires the units where the instantiations occur to be declared of interest as well. In our example, this means passing :option:`--scos=v5.ali` and :option:`--scos=v8.ali` in addition.

Separation by instance relies on specific compiler support available in the GNAT Pro toolchain since the 7.2 release. For older toolchains, another mode is available which reports separate coverage statuses for copies associated with distinct symbols of the executable file. As we will describe, this provides a good approximation of per-instance analysis in absence of inlining, and becomes inaccurate when inlining comes into play.

Separation by routine (:option:`-S routine`, obsolete)

In this mode, two code regions coming from the same source construct will undergo separate coverage analyses if they occur in different symbols of the executable file. This scheme is obsolete, unreliable in presence of inlining.

When a given subprogram is inlined in two different calling routines, each inlined copy thus undergoes a separate coverage assessment. In the absence of inlining, this will also ensure that different instances of the same generic unit will have separated coverage analyses, since the compiler generates different symbol names for different program units. For our Vops example, this would be:

gnatcov coverage -Pvops.gpr --annotate=report -S routine [...]
...
vops.adb:5:11: statement not executed (from v8__inc)
vops.adb:6:10: statement not executed (from v8__inc)
vops.adb:12:11: statement not executed (from v5__mult)
vops.adb:13:10: statement not executed (from v5__mult)

On the other hand, if two distinct instances of a generic subprogram are inlined within a single calling routine, they will undergo a single coverage analysis since they now occur in the same symbol.

Handling Ada assertions and contracts

Ada 2012 introduced sophisticated programming by contract mechanisms which generalize the base pragma Assert construct originally offered by the language. All these facilities allow asserting various properties of the program state at specific points of its execution, for example as type invariants or as Pre/Post-conditions on subprograms.

Each property is expressed a Boolean expression expected to hold True, triggering an exception otherwise. When the current assertion policy activates a given assertion construct, the associated Boolean expression is treated by |gcv| as a decision for MCDC purposes.

As assertions are by construction designed never to evaluate False, reaching proper coverage for them is non-trivial, if not entirely meaningless, for post-conditions in particular.

The simple way out consists in disabling the relevant constructs in builds intended for coverage analysis, by setting the corresponding Assertion_Policy to Disable with GNAT Pro toolchains.

Processing of C macros

For source coverage purposes, Source Coverage Obligations for C are produced after the preprocessing of sources, with two consequences of note:

  • Macro expansions leading to code with conditionals will trigger coverage violations, and multiple "calls" to the same macro just multiply these as they yield distinct expansions.
  • The source locations output by |gcv| for coverage violations within macro expansions designate preprocessed tokens at the macro expansion site, typically on the line of the macro invocation but with column numbers unrelated to what is visible in the source on this line.

Consider this C code for example:

 1  #define COND_INC(cond,x,y) \
 2    do {                     \
 3      if (cond)              \
 4        (x)++;               \
 5      else                   \
 6        (y)++;               \
 7    } while(0)
 8
 9  int main ()
10  {
11    volatile x = 0, y = 0;
12
13    COND_INC(x == 0, x, y);
14    COND_INC(x == 0, x, y);
15  }

The two macro invocations actually expand as:

13    do { if (x == 0) (x)++; else (y)++; } while(0);
14    do { if (x == 0) (x)++; else (y)++; } while(0);

The expanded version is the basis of SCO identification process, so we have one decision and two conditioned statements on line 13, likewise on line 14. Only one of each is exercised at execution time, and a :option:`stmt+decision` analysis on this program yields:

2.1. STMT COVERAGE
------------------

t.c:13:32: statement not executed
t.c:14:20: statement not executed

2 violations.

2.2. DECISION COVERAGE
----------------------

t.c:13:12: decision outcome FALSE never exercised
t.c:14:12: decision outcome TRUE never exercised

2 violations.

We do see one statement and one decision coverage violation per invocation, different in the two cases since the x == 0 test is True on the first call and False on the second one. We also observe column numbers unrelated to what the original source lines contain on line 13 and 14.

Optimization and non-coverable items

|gcp| essentially operates by relating execution traces to source entities of interest thanks to debug information mapping machine code addresses to source locations. With optimization enabled, there sometimes is no machine code attached to a given statement, for example when the statement is determined to be redundant or when the machine code for it can be factorized with the machine code for another statement. When the coverage status of a code-less statement cannot be be inferred from that of other statements around, |gcp| categorizes the statement as :dfn:`non-coverable`.

By default, nothing is said about non-coverable statements in the :option:`=report` outputs and the corresponding lines are marked with a '.' in annnotated sources, as for any other line to which no machine code is attached. Below is an example source annotated for statement coverage, where absence of code for a couple of Ada statments was triggered by constant propagation and inlining. The local Pos function is called only once, with a constant argument such that only one alternative of the if statement is taken. With :option:`-O1 -gnatn`, the compiler sees that the else part can never be entered and no code is emitted at all for this alternative:

 4 .: procedure Test_Pos1 is
 5 .:    function Pos (X : Integer) return Boolean;
 6 .:    pragma Inline (Pos);
 7 .:
 8 .:    function Pos (X : Integer) return Boolean is
 9 .:    begin
10 +:       if X > 0 then
11 +:          Put_Line ("X is positive");
12 +:          return True;
13 .:       else
14 .:          Put_Line ("X is not positive");
15 .:          return False;
16 .:       end if;
17 .:    end Pos;
18 .: begin
19 +:    Assert (Pos (1) = True);
20 .: end Test_Pos1;

A common similar case is that of debugging code inhibited on purpose for regular operation, for example with constructs like:

Debug_Mode : constant Boolean := False;    or     #define DEBUG_MODE 0
...                                               ...
if Debug_Mode then                                #if DEBUG_MODE

Another way to get this in Ada is with generic instanciations where constant parameters turn what appears to be conditional in the source into a constant value in some instances.

Back to our Test_Pos1 example, no code is emitted for the test on line 10 either. |gcv| is however able to infer the if coverage status by looking at the status of statements controlled by the decision, and the Decision coverage report remains accurate:

   8 .:    function Pos (X : Integer) return Boolean is
   9 .:    begin
  10 !:       if X > 0 then
decision "X > 0" at 10:10 outcome FALSE never exercised
  11 +:          Put_Line ("X is positive");
  12 +:          return True;

|gcvcov| features the :option:`--non-coverable` command line option to expose the non-coverable statements if needed. They are listed in an additional "NON COVERABLE ITEMS" section of the :option:`=report` outputs and the corresponding lines are flagged with a '0' mark in annotated sources, as well as a specific color in the html formats. For our example, this yields:

10 !:       if X > 0 then
11 +:          Put_Line ("X is positive");
12 +:          return True;
13 .:       else
14 0:          Put_Line ("X is not positive");
15 0:          return False;
16 .:       end if;