Permalink
2337 lines (1865 sloc) 66.2 KB

Trompeloeil cook book

Integrating with unit test frame works

By default, Trompeloeil reports violations by throwing an exception, explaining the problem in the what() string.

Depending on your test frame work and your runtime environment, this may, or may not, suffice.

Trompeloeil offers support for adaptation to any test frame work. Some sample adaptations are:

There are two mechanisms for adapting to a testing frame work. The compile time adapter and the run time adapter. The compile time adapter is easier to use if you do not implement main() yourself.

Compile time adapter

Compile time adaptation to unit test frame works is done by specializing the trompeloeil::reporter<trompeloeil::specialized> struct.

Somewhere in global namespace, in one of your translation units, enter the following code:

namespace trompeloeil
{
  template <>
  void reporter<specialized>::send(
    severity s,
    const char* file,
    unsigned long line,
    const char* message)
  {
    // your adaptation here
  }
}

In all other translation units, add the following extern declaration in global namespace instead:

extern template struct trompeloeil::reporter<trompeloeil::specialized>;

It is important to understand the first parameter trompeloeil::severity. It is an enum with the values trompeloeil::severity::fatal and trompeloeil::severity::nonfatal. The value severity::nonfatal is used when reporting violations during stack rollback, typically during the destruction of an expectation. In this case it is vital that no exception is thrown, or the process will terminate. If the value is severity::fatal, it is instead imperative that the function does not return. It may throw or abort.

NOTE! There are some violations that cannot be attributed to a source code location. An example is an unexpected call to a mock function for which there are no expectations. In these cases file will be "" string and line == 0.

Run time adapter

Run time adaptation to unit test frame works is done with this function:

using reporter_func = std::function<void(
  severity,
  char const *file,
  unsigned long line,
  std::string const &msg)>;
reporter_func trompeloeil::set_reporter(reporter_func new_reporter)

Call it with the adapter to your test frame work. The return value is the old adapter.

It is important to understand the first parameter trompeloeil::severity. It is an enum with the values trompeloeil::severity::fatal and trompeloeil::severity::nonfatal. The value severity::nonfatal is used when reporting violations during stack rollback, typically during the destruction of an expectation. In this case it is vital that no exception is thrown, or the process will terminate. If the value is severity::fatal, it is instead imperative that the function does not return. It may throw or abort.

NOTE! There are some violations that cannot be attributed to a source code location. An example is an unexpected call to a mock function for which there are no expectations. In these cases file will be "" string and line == 0.

Use Trompeloeil with Catch!

Paste the following code snippet in global namespace in one of your translation units:

  namespace trompeloeil
  {
    template <>
    void reporter<specialized>::send(
      severity s,
      const char* file,
      unsigned long line,
      const char* msg)
    {
      std::ostringstream os;
      if (line) os << file << ':' << line << '\n';
      os << msg;
      auto failure = os.str();
      if (s == severity::fatal)
      {
        FAIL(failure);
      }
      else
      {
        CAPTURE(failure);
        CHECK(failure.empty());
      }
    }
  }

If you have several translation units, add the following extern declaration in the others:

extern template struct trompeloeil::reporter<trompeloeil::specialized>;

If you roll your own main(), you may prefer a runtime adapter instead. Before running any tests, make sure to call:

  trompeloeil::set_reporter([](
    trompeloeil::severity s,
    const char* file,
    unsigned long line,
    const char* msg)
  {
    std::ostringstream os;
    if (line) os << file << ':' << line << '\n';
    os << msg;
    auto failure = os.str();
    if (s == trompeloeil::severity::fatal)
    {
      FAIL(failure);
    }
    else
    {
      CAPTURE(failure);
      CHECK(failure.empty());
    }
  });

Use Trompeloeil with crpcut

Paste the following code snippet in global namespace in one of the translation units:

  namespace trompeloeil
  {
    template <>
    void reporter<specialized>::send(
      severity,
      char const *file,
      unsigned long line,
      const char* msg)
    {
      std::ostringstream os;
      os << file << ':' << line;
      auto loc = os.str();
      auto location = line == 0U
        ? ::crpcut::crpcut_test_monitor::current_test()->get_location()
        : ::crpcut::datatypes::fixed_string::make(loc.c_str(), loc.length());
      ::crpcut::comm::report(::crpcut::comm::exit_fail,
                             std::ostringstream(msg),
                             location);
    }
  }

In all other translation units, add the extern declaration

extern template struct trompeloeil::reporter<trompeloeil::specialized>;

If you instead prefer a runtime adapter, make sure to call

  trompeloeil::set_reporter([](
    trompeloeil::severity,
    const char* file,
    unsigned long line,
    const char* msg)
  {
    std::ostringstream os;
    os << file << ':' << line;
    auto loc = os.str();
    auto location = line == 0U
      ? ::crpcut::crpcut_test_monitor::current_test()->get_location()
      : ::crpcut::datatypes::fixed_string::make(loc.c_str(), loc.length());
    ::crpcut::comm::report(::crpcut::comm::exit_fail,
                           std::ostringstream(msg),
                           location);
  });

before any tests are run.

Use Trompeloeil with doctest

doctest 1.2 or newer

Paste the following code snippet in global namespace in one of your translation units:

  namespace trompeloeil
  {
    template <>
    void reporter<specialized>::send(
      severity s,
      const char* file,
      unsigned long line,
      const char* msg)
    {
      auto f = line ? file : "[file/line unavailable]";
      if (s == severity::fatal)
      {
        ADD_FAIL_AT(f, line, msg);
      }
      else
      {
        ADD_FAIL_CHECK_AT(f, line, msg);
      }
    }
  }

If you have several translation units, add the following extern declaration in the others:

extern template struct trompeloeil::reporter<trompeloeil::specialized>;

If you roll your own main(), you may prefer a runtime adapter instead. Before running any tests, make sure to call:

  trompeloeil::set_reporter([](
    trompeloeil::severity s,
    const char* file,
    unsigned long line,
    const char* msg)
  {
    auto f = line ? file : "[file/line unavailable]";
    if (s == severity::fatal)
    {
      ADD_FAIL_AT(f, line, msg);
    }
    else
    {
      ADD_FAIL_CHECK_AT(f, line, msg);
    }
  });

doctest < 1.2

Paste the following code snippet in global namespace in one of your translation units:

  struct doctest_violation : std::ostringstream
  {
    friend std::ostream& operator<<(std::ostream& os, doctest_violation const& v)
    {
      return os << v.str();
    }
  };

  namespace trompeloeil
  {
    template <>
    void reporter<specialized>::send(
      severity s,
      const char* file,
      unsigned long line,
      const char* msg)
    {
      ::doctest_violation violation;
      if (line) violation << file << ':' << line << '\n';
      violation << msg;
      if (s == severity::fatal)
      {
        REQUIRE_FALSE(violation);
      }
      else
      {
        CHECK_FALSE(violation);
      }
    }
  }

If you have several translation units, add the following extern declaration in the others:

extern template struct trompeloeil::reporter<trompeloeil::specialized>;

If you roll your own main(), you may prefer a runtime adapter instead. As above, create a simple doctest_violation type by pasting the below code into the file containing main().

  struct doctest_violation : std::ostringstream
  {
    friend std::ostream& operator<<(std::ostream& os, doctest_violation const& v)
    {
      return os << v.str();
    }
  };

Then, before running any tests, make sure to call:

  trompeloeil::set_reporter([](
    trompeloeil::severity s,
    const char* file,
    unsigned long line,
    const char* msg)
  {
    ::doctest_violation violation;
    if (line) violation << file << ':' << line << '\n';
    violation << msg;
    if (s == trompeloeil::severity::fatal)
    {
      REQUIRE_FALSE(violation);
    }
    else
    {
       CHECK_FALSE(violation);
    }
  });

Use Trompeloeil with gtest

Paste the following code snippet in global namespace in one of the translation units:

  namespace trompeloeil
  {
    template <>
    void reporter<specialized>::send(
      severity s,
      char const *file,
      unsigned long line,
      const char* msg)
    {
      if (s == severity::fatal)
      {
        std::ostringstream os;
        if (line != 0U)
        {
          os << file << ':' << line << '\n';
        }
        throw expectation_violation(os.str() + msg);
      }

      ADD_FAILURE_AT(file, line) << msg;
    }
  }

In all other translation units, add the following extern declaration:

extern template struct trompeloeil::reporter<trompeloeil::specialized>;

If you instead prefer a runtime adapter, make sure to call

  trompeloeil::set_reporter([](
    trompeloeil::severity s,
    const char* file,
    unsigned long line,
    const char* msg)
  {
    if (s == trompeloeil::severity::fatal)
    {
      std::ostringstream os;
      if (line != 0U)
      {
        os << file << ':' << line << '\n';
      }
      throw trompeloeil::expectation_violation(os.str() + msg);
    }

    ADD_FAILURE_AT(file, line) << msg;
  });

before running any tests.

Use Trompeloeil with lest

With lest, you always provide your own main(). In it, provide a runtime adapter like the one below.

int main(int argc, char *argv[])
{
  std::ostream& stream = std::cout;

  trompeloeil::set_reporter([&stream](
    trompeloeil::severity s,
    const char* file,
    unsigned long line,
    std::string const& msg)
  {
    if (s == trompeloeil::severity::fatal)
    {
      throw lest::message{"", lest::location{ line ? file : "[file/line unavailable]", int(line) }, "", msg };
    }
    else
    {
      stream << lest::location{ line ? file : "[file/line unavailable]", int(line) } << ": " << msg;
    }
  });

  return lest::run(specification, argc, argv, stream);
}

Use Trompeloeil with boost Unit Test Framework

Paste the following code snippet in global namespace in one of your translation units:

  namespace trompeloeil
  {
    template <>
    void reporter<specialized>::send(
      severity s,
      char const *file,
      unsigned long line,
      const char* msg)
    {
      std::ostringstream os;
      if (line != 0U) os << file << ':' << line << '\n';
      auto text = os.str() + msg;
      if (s == severity::fatal)
        BOOST_FAIL(text);
      else
        BOOST_ERROR(text);
    }
  }

In all other translation units, add the following extern declaration:

extern template struct trompeloeil::reporter<trompeloeil::specialized>;

If you instead prefer a runtime adapter, make sure to call

  trompeloeil::set_reporter([](
    trompeloeil::severity s,
    const char* file,
    unsigned long line,
    const char* msg)
  {
    std::ostringstream os;
    if (line != 0U) os << file << ':' << line << '\n';
    auto text = os.str() + msg;
    if (s == trompeloeil::severity::fatal)
      BOOST_FAIL(text);
    else
      BOOST_ERROR(text);
  });

before running any tests.

Use Trompeloeil with MSTest

Place the below code snippet in, for example, your TEST_CLASS_INITIALIZE(...)

  using namespace trompeloeil;
  set_reporter([](
    severity,
    char const* file,
    unsigned long line,
    std::string const& msg)
  {
    std::wstring wideMsg(msg.begin(), msg.end());
    std::wstring wfile;
    if (line > 0) wfile.append(file, file + strlen(file));
    __LineInfo loc(wfile.c_str(), "", line);
    Assert::Fail(wideMsg.c_str(), line == 0 ? nullptr : &loc);
  });

Creating Mock Classes

A Mock class is any class that mocks member functions.

There are two ways to create mocks. A very frequently seen situation is when inheriting from an interface (i.e. an abstract base class with pure virtual functions). When this is the case, the easiest route is to inherit the interface via trompeloeil::mock_interface<T> and implement the mock functions with the macros IMPLEMENT_MOCKn(...) and IMPLEMENT_CONST_MOCKn(...). These only work when implementing to an interface, do not handle multiple inheritance and do not handle overloads.

A more generic technique is to implement free mocks as members of any struct or class using the macros MAKE_MOCKn and MAKE_CONST_MOCKn, where n is the number of parameters in the function.

Example:

class Dictionary
{
public:
  virtual ~Dictionary() = default;
  virtual std::string& lookup(int n) const = 0;
  virtual void add(int n, std::string&&) = 0;
};

class MockDictionary : public trompeloeil::mock_interface<Dictionary>
{
  IMPLEMENT_CONST_MOCK1(lookup);
  IMPLEMENT_MOCK2(add);
};

struct Logger
{
  MAKE_MOCK2(log, void(int severity, const std::string& msg));
};

In the example above, MockDictionary is, as the name implies, a mock class for the pure virtual class Dictionary.

The line IMPLEMENT_CONST_MOCK1(lookup); implements the function std::string& lookup(int) const and the line IMPLEMENT_MOCK2(add); implements the function void add(int, std::string&&).

The line MAKE_MOCK2(log, void(int severity, const std::string& msg)) creates a mock function void Logger::log(int, const std::string&). If MAKE_MOCKn(...) or MAKE_CONST_MOCKn(...) are used to implement a virtual function from a base class, it is always recommended to add a third macro parameter override since it gives the compiler an ability to complain about mistakes.

Mocking private or protected member functions

Mocking private or protected member functions is no different from mocking public member functions. Just make them public in the mock class. It may seem strange that you can change access rights of a member function through inheritance, but C++ allows it.

Example:

class Base
{
private:
  virtual void secret(int);
};

class Mock : public Base
{
public:
  MAKE_MOCK1(secret, void(int), override); // not so secret now
};

The mock functions must be public for you to be able to set expectations on them, but there is nothing preventing a public function from implementing a private virtual function in a base class.

Mocking overloaded member functions

Trompeloeil matches mock functions by their name and their signature, so there is nothing special about adding several overloads of mocked functions.

Example:

class Mock
{
public:
  MAKE_MOCK1(overload, void(int));
  MAKE_MOCK1(overload, int(const std::string&));
  MAKE_MOCK2(overload, int(const char*, size_t));
};

Above there are three mock functions named overload, with different signatures.

See Matching calls to overloaded member functions for how to place expectations on them.

NOTE! Overloaded member functions cannot be mocked using the macros IMPLEMENT_MOCKn(...) or IMPLEMENT_CONST_MOCKn(...)`.

Mocking a class template

Unlike some C++ mocking frame works, Trompeloeil does not make a distinction between mocks in class templates and mocks in concrete classes.

Example:

template <typename T>
class Mock
{
public:
  MAKE_MOCK1(func, void(int));
  MAKE_MOCK2(tfunc, int(const T&, size_t));
};

Above, Mock<T> is a mock class template with two member functions. The member function void func(int) does not depend on the template parameter, whereas the member function int tfunc(const T&, size_t) does. This will work for any type T.

Mocking non-virtual member functions

While it is often the case that mocks are used to implement interfaces, there is no such requirement. Just add the mock functions that are needed.

Example:

class ConcreteMock
{
public:
  MAKE_MOCK2(func, bool(size_t, const char*));
};

Above ConcreteMock is a mock class that implements a non-virtual mock function bool func(size_t, const char*).

Mocking free functions

Free functions on their own cannot be mocked, the calls to them needs to be dispatched to mock objects. Often there are several free functions that together form an API, and then it makes sense to implement one mock class for the API, with mock functions for each.

Example, assume a simple C-API

// C-API.h

#ifdef __cplusplus
extern "C" {
#endif

struct c_api_cookie;

struct c_api_cookie* c_api_init();

int c_api_func1(struct c_api_cookie* cookie, const char* str, size_t len);

void c_api_end(struct c_api_cookie*);

#ifdef __cplusplus
}
#endif
// unit-test-C-API.h

#include "C-API.h"

class API
{
public:
  MAKE_MOCK0(c_api_init, c_api_cookie*());
  MAKE_MOCK3(c_api_func1, int(c_api_cookie*, const char*, size_t));
  MAKE_MOCK1(c_api_end, void(c_api_cookie*));
};

extern API c_api_mock;

Then implement the functions in a test version of the API, which uses the mock.

// unit-test_c_api.cpp
#include "unit-test-C-API.h"

API c_api_mock;

extern "C" {
  c_api_cookie c_api_init()
  {
    return api_mock.c_api_init();
  }

  int c_api_func1(c_api_cookie* cookie, const char* str, size_t len)
  {
    return api_mock.c_api_func1(cookie, str, len);
  }

  void c_api_end(c_api_cookie* cookie)
  {
    api_mock.c_api_end(cookie);
  }
}

A test program can place expectations on the mock object, and the tested functionality calls the C-API functions which dispatch to the mock object.

#include "unit-test-C-API.h"

void a_test()
{
  REQUIRE_CALL(c_api_mock, create())
    .RETURN(nullptr);

  REQUIRE_CALL(c_api_mock, c_api_end(nullptr));

  function_under_test();
}

Setting Expectations

It is with expectations you define the behaviour of your test. By default all calls to mock functions are illegal and will be reported as violations. You use expectations, long or short lived, wide or narrow, to make some calls legal and define what happens.

There are three basic types of expectations.

ALLOW_CALL(...) is often used for a default. It can match any number of times.

REQUIRE_CALL(...) is stricter and defaults to match exactly once, although you can change that and control exactly how many times you want the expectation to match.

FORBID_CALL(...) may seem unnecessary since calls are forbidden by default, but it is useful in combination with ALLOW_CALL(...) or REQUIRE_CALL(...) to forbid something that would otherwise be accepted.

If several expectations match a call, it is the last matching expectation created that is used. ALLOW_CALL(...), REQUIRE_CALL(...) and FORBID_CALL(...) are active until the end of the scope. This means that you can place a wide default, and use temporary special expectations in local scopes, for example to temporarily forbid a call that is otherwise allowed.

If the scoped lifetime rules are unsuitable, there are also thee named versions of the expectations.

These do the same, but they create a std::unique_ptr<trompeloeil::expectation>, which you can bind to variables that you control the life time of.

Matching exact values

The simplest expectations are for calls with exact expected parameter values. You just provide the expected values in the parameter list of the expectation.

Example:

class Mock
{
public:
  MAKE_MOCK1(func, void(int));
  MAKE_MOCK2(func, void(const char*));
};

void test()
{
  Mock m;
  ALLOW_CALL(m, func(1));         // int version any number of times
  REQUIRE_CALL(m, func(nullptr)); // const char * version exactly once
  func(&m);
  // expectations must be met before end of scope
}

Matching values with conditions

Instead of using exact values of parameters to match calls with, Trompeloeil provides a set of matchers. Simple value matchers are:

  • eq( value ) matches value equal (using operator==())
  • ne( value ) matches value not equal (using operator!=())
  • gt( value ) matches value greater than (using operator>())
  • ge( value ) matches value greater than or equal (using operator>=())
  • lt( value ) matches value less than (using operator<())
  • le( value ) matches value less than or equal (using operator<=())

By default, the matchers are duck typed, i.e. they match a parameter that supports the operation. If disambiguation is necessary to resolve overloads, an explicit type can be specified.

Example:

class Mock
{
public:
  MAKE_MOCK1(func, void(int));
  MAKE_MOCK1(func, void(const char*));
  MAKE_MOCK1(func, void(const std::string&))
};

void test()
{
  Mock m;
  ALLOW_CALL(m, func(trompeloeil::gt(1))); // int version any number of times
  REQUIRE_CALL(m, func(trompeloeil::ne<std::string>(""))); // const std::string& version once
  func(&m);
  // expectations must be met before end of scope
}

Matching strings with regular expressions

Matching string parameters to regular expressions is convenient with Trompeloeil re( expression ) regular expression matchers.

Example:

class Mock
{
public:
  MAKE_MOCK1(func, void(const char*));
};

void test()
{
  Mock m;
  REQUIRE_CALL(m, func(trompeloeil::re("^begin.*end$")));
  func(&m);
  // expectation must be met before end of scope
}

TIP! Using C++ raw string literals can massively help getting regular expression escapes right.

Matching pointers to values

All matchers can be converted to a pointer matcher by using the dereference prefix operator *. This works for smart pointers too. These pointer matchers fail if the pointer parameter is nullptr.

Example:

class Mock
{
public:
  MAKE_MOCK1(func, void(int*));
  MAKE_MOCK2(func, void(std::unique_ptr<short>*));
};

using trompeloeil::eq;
using trompeloeil::gt;

void test()
{
  Mock m;
  ALLOW_CALL(m, func(*eq(1))); // pointer to int value 1 any number of times
  REQUIRE_CALL(m, func(*gt<short>(5))); // unique_ptr<short> to >5 once
  func(&m);
  // expectations must be met before end of scope
}

Matching the opposite of a matcher

All matchers can be negated, allowing what the matcher disallows and disallowing what the matcher allows, using the operator ! on the matcher.

Example:

struct Mock {
  MAKE_MOCK1(func, void(const std::string&));
};

using trompeloeil::re; // matching regular expressions

TEST(atest)
{
  Mock m;
  REQUIRE_CALL(m, func(!re("^foo")));
  func(&m);
  // m.func() must've been called with a string not beginning with "foo"
}

Matching calls with conditions depending on several parameters

Some times a matching call cannot be judged for individual parameter values alone, but together they work. Assume for example a C-string API where you have a const char* and a length.

Example:

class Mock
{
public:
  MAKE_MOCK2(func, void(const char*, size_t len));
};

using trompeloeil::ne;
using trompeloeil::_;

void test()
{
  Mock m;
  REQUIRE_CALL(m, func(ne(nullptr), _)))  // once
    .WITH(std::string(_1, _2) == "meow"));
  func(&m);
  // expectations must be met before end of scope
}

_ is a special matcher that matches everything. .WITH(...) is a construction used for when simple matchers aren't enough. If a call is made which matches the values given in the REQUIRE_CALL(...), the selection process continues in .WITH(std::string(_1, _2) == "meow").

_1 and _2 are the parameters to the call, so in this case a std::string is constructed using the non-null const char* and the length, and its value is compared with "meow".

The expression in .WITH(...) can be anything at all that returns a boolean value. It can refer to global variables, for example.

It is important to understand that .WITH(...) accesses any local variable used in the expression as a copy. If you want to refer to a local variable by reference, use .LR_WITH(...) instead (LR_ for "local reference").

Matching std::unique_ptr<T> and other non-copyable values

Matching parameter values that you cannot copy, or do not want to copy, requires a bit of thought.

The wildcards _ and ANY(...) works. For std::unique_ptr<T> and std::shared_ptr<T>, the matcher ne(nullptr) also works.

If you want to be more specific, you will need to use .WITH(...) or .LR_WITH(...)

Example:

class Mock
{
public:
  MAKE_MOCK1(func, void(std::unique_ptr<int>));
};

using trompeloeil::ne;

void test()
{
  Mock m;
  REQUIRE_CALL(m, func(ne(nullptr)))
    .WITH(*_1 == 3);
  func(&m);
  // expectations must be met before end of scope
}

Above there is a requirement that the function is called with a non-null std::unique_ptr<int>, which points to a value of 3.

Matching calls to overloaded member functions

Distinguishing between overloads is simple when using exact values to match since the type follows the values. It is more difficult when you want to use wildcards and other matchers.

One useful matcher is ANY(...), which behaves like the open wildcard _, but has a type. It is also possible to specify types in the matchers.

Example:

class Mock
{
public:
  MAKE_MOCK1(func, void(int*));
  MAKE_MOCK1(func, void(char*));
};

using namespace trompeloeil;

void test()
{
  Mock m;

  REQUIRE_CALL(m, func(ANY(int*)));
  REQUIRE_CALL(m, func(ne<char*>(nullptr)));

  func(&m);
}

Above, each of the func overloads must be called once, the int* version with any pointer value at all, and the char* version with a non-null value.

Define side effects for matching calls

A side effect, in Trompeloeil parlance, is something that is done after a match has been made for an expectation, and before returning (or throwing).

Typical side effects are:

  • Setting out parameters
  • Capturing in parameters
  • Calling other functions

Example:

class Dispatcher
{
public:
  MAKE_MOCK1(subscribe, void(std::function<void(const std::string&)>));
};

using trompeloeil::_;

void test()
{
  Dispatcher d;

  std::vector<std::function<void(const std::string&)>> clients;

  {
    REQUIRE_CALL(d, subscribe(_))
      .LR_SIDE_EFFECT(clients.push_back(std::move(_1)))
      .TIMES(AT_LEAST(1));

    func(&d);
  }
  for (auto& cb : clients) cb("meow");
}

Above, any call to d.subscribe(...) will have the side effect that the parameter value is stored in the local vector clients.

The test then goes on to call all subscribers.

LR_SIDE_EFFECT(...) accesses references to local variables. There is also SIDE_EFFECT(...), which accesses copies of local variables.

Return values from matching calls

An expectation on a non-void function must return something or throw an exception. There are no default values. Returning is easy, however. Just use a .RETURN(...) or .LR_RETURN(...) with an expression of the right type.

Example:

class Dictionary
{
public:
  using id_t = size_t;
  MAKE_MOCK1(lookup, std::string(id_t));
};

using trompeloeil::ge; // greater than or equal
using trompeloeil::lt; // less than

void test()
{
  Dictionary d;
  std::vector<std::string> dict{...};

  ALLOW_CALL(d, lookup(ge(dict.size())))
    .RETURN("");                          // create std::string from ""
  ALLOW_CALL(d, lookup(lt(dict.size())))
    .LR_RETURN(dict[_1]);                 // access element in vector
  func(&d);
}

Above, the matchers lt(...) and ge(...) are used to ensure that the indexing in the local variable dict can be made safely. Note that the first expectation does not match the return type exactly, but is something that can be implicitly converted.

LR_RETURN(...) is used in the second to avoid copying the vector, since RETURN(...) always accesses copies of local variables.

Return references from matching calls

Returning references from matching expectations exposes some peculiarities in the language. Specifically, it is not allowed to return a captured local variable as a reference in RETURN(...), and in LR_RETURN(...) a returned variable must be decorated to ensure that a reference is intended.

Example:

class Dictionary
{
public:
  using id_t = size_t;
  MAKE_MOCK1(lookup, const std::string&(id_t));
};

using trompeloeil::gt; // greater than or equal
using trompeloeil::lt; // less than

std::string global_empty;

void test()
{
  Dictionary d;
  std::vector<std::string> dict{...};

  std::string empty;

  ALLOW_CALL(d, lookup(gt(dict.size())))
    .LR_RETURN((empty));                // extra () -> reference to local variable
  ALLOW_CALL(d, lookup(dict.size()))
    .LR_RETURN(std::ref(empty));        // reference to local variable
  ALLOW_CALL(d, lookup(lt(dict.size())))
    .LR_RETURN(dict[_1]);               // result of function call
  ALLOW_CALL(d, lookup(0))
    .RETURN(std::ref(global_empty));    // reference to global variable
  func(&d);
}

Captured variables that are returned as references must either be enclosed in extra parenthesis, or std::ref().

Returning a reference obtained from a function call, however, does not require any extra decoration, as the third expectation above, which looks up values in dict shows.

Throwing exceptions from matching calls

To throw an exception, just add a .THROW(...) or .LR_THROW(...), with the value to throw. For non-void functions, .LR_THROW(...) and .THROW(...) takes the place of a .RETURN(...) or .LR_RETURN(...).

Example:

class Dictionary
{
public:
  using id_t = size_t;
  MAKE_CONST_MOCK1(lookup, const std::string&(id_t));
};

using trompeloeil::_; // matches anything

void test()
{
  Dictionary d;
  std::vector<std::string> dict{...};

  ALLOW_CALL(d, lookup(_))
    .LR_WITH(_1 >= dict.size())
    .THROW(std::out_of_range("index too large for dictionary"));

  ALLOW_CALL(d, lookup(_))
    .LR_WITH(_1 < dict.size())
    .LR_RETURN(dict[_1]);

  func(&d);
}

Above, any call to d.lookup(...) with an index within the size of the vector will return the string reference, while any call with an index outside the size of the vector will throw a std::out_of_range exception.

Allowing any call

By default it is illegal to call any mock function and you provide narrow specific expectations according to the needs of your test. However, sometimes it makes sense to have a wide-open default. That is done with the exceptations ALLOW_CALL(...) and NAMED_ALLOW_CALL(...). The difference between them is that ALLOW_CALL is local in nature and is only valid until the end of the scope, while NAMED_ALLOW_CALL(...) can be bound to a std::unique_ptr<trompeloeil::expectation>, which you can control the lifetime of.

Example:

template <typename T>
class Allocator
{
public:
  MAKE_MOCK1(allocate, T*(size_t));
  MAKE_MOCK1(deallocate, void(T*));
};

using trompeloeil::_;

void test_no_mem()
{
  Allocator<int> ai;

  ALLOW_CALL(ai, allocate(_))
    .RETURN(nullptr);

  ALLOW_CALL(ai, deallocate(nullptr));

  hairy_int_job(&ai);
}

The simplistic allocator above is rigged to allow any attempts to allocate memory, but always return nullptr, and only allow deallocation of nullptr.

Temporarily disallowing matching calls

Just as it is sometimes convenient to provide a blanket default behaviour, it is sometimes desirable to temporarily ban calls.

Example:

#include "hairy_job.h"

template <typename T>
class Allocator
{
public:
  MAKE_MOCK1(allocate, T*(size_t));
  MAKE_MOCK1(deallocate, void(T*));
};

using trompeloeil::_;

void test_restricted_mem()
{
  Allocator<int> ai;

  ALLOW_CALL(ai, allocate(_))
    .RETURN(new int[_1]);

  ALLOW_CALL(ai, deallocate(_))
    .SIDE_EFFECT(delete[] _1);

  hairy_job<int, Allocator<int>> job(ai, initial_data);

  {
    FORBID_CALL(ai, allocate(_));

    job.churn(); // must not allocate memory
  }

  job.get_result(); // may allocate memory
}

Above we see a simplistic Allocator that by default allocates and deallocates arrays.

The hairy_job uses the Allocator for its setup, and is expected to allocate all memory it needs for churn() in its constructor.

That churn() doesn't use the allocator is ensured by the local scope, in which all calls to allocate(...) are forbidden.

This pattern is quite common when writing tests with Trompeloeil. Use wide defaults in the scope of the test case (or in a fixture), and use local scopes with specifics, be they forbidden or exact requirements.

Expecting several matching calls in some sequences

By default all expectations are equal, and the only sequencing relationship is that if several match a call, the one last created is the one matched.

This means that expectations that do not compete for matching the same call have no ordering relationship at all, they are logically parallel.

Often this is exactly what you want. When you poke an object, you want this and that thing to happen and the order between them is irrelevant. For example, if calling callbacks stored in a hash table, you don't want to impose an order of those calls.

There are two very different reasons for using sequence control with Trompeloeil.

One is hinted at above, to impose an order between expectations that are logically parallel. The other is to set an exact order of indistinguishable expectations. The latter can be achieved by setting them up in reverse order of matching, but this can make the test code very difficult to read.

First example. Impose an order between logically parallel calls:

class FileOps
{
public:
  using handle = int;
  MAKE_MOCK1(open, handle(const std::string&));
  MAKE_MOCK3(write, size_t(handle, const char*, size_t));
  MAKE_MOCK1(close, void(handle));
};

using trompeloeil::ne;

void test()
{
  FileOps ops;

  trompeloeil::sequence seq;

  int handle = 4711;

  REQUIRE_CALL(ops, open("name"))
    .RETURN(handle)
    .IN_SEQUENCE(seq);

  REQUIRE_CALL(ops, write(handle, ne(nullptr), ne(0)))
    .RETURN(_3)
    .IN_SEQUENCE(seq);

  REQUIRE_CALL(ops, close(handle))
    .IN_SEQUENCE(seq);

  test_writes(&ops);
}

Without the use of trompeloeil::sequence above, all three expectations would be logically parallel and all permutations of matches would be considered equally correct.

By imposing an order between them, there is now only one legal sequence of calls.

The other example is to provide an order between equally matching calls. Suppose we want the write function above to first return 0 once and then give the desired result:

class FileOps
{
public:
  using handle = int;
  MAKE_MOCK1(open, handle(const std::string&));
  MAKE_MOCK3(write, size_t(handle, const char*, size_t));
  MAKE_MOCK1(close, void(handle));
};

using trompeloeil::ne;

void test()
{
  FileOps ops;

  trompeloeil::sequence seq;

  int handle = 4711;

  REQUIRE_CALL(ops, open("name"))
    .RETURN(handle)
    .IN_SEQUENCE(seq);

  REQUIRE_CALL(ops, write(handle, ne(nullptr), ne(0)))
    .RETURN(0)                                         // indicate failure
    .IN_SEQUENCE(seq);

  REQUIRE_CALL(ops, write(handle, ne(nullptr), ne(0)))
    .RETURN(_3)                                        // successful retry
    .IN_SEQUENCE(seq);

  REQUIRE_CALL(ops, close(handle))
    .IN_SEQUENCE(seq);

  test_writes(&ops);
}

Here the two calls to write are supposed to be made with exactly the same parameters, so they cannot be distinguished that way. We want the first call to indicate intermittent failure, and to be followed by a retry that will succeed.

.IN_SEQUENCE(...) can refer to several sequence objects, which is a way to allow some variation in order, without being too lax. For a more thorough walk through, see the blog post Sequence control with the Trompeloeil C++14 Mocking Framework

.IN_SEQUNECE(...) can also be used on REQUIRE_DESTRUCTION(...) and NAMED_REQUIRE_DESTRUCTION(...).

Expecting matching calls a certain number of times

By default REQUIRE_CALL(...) needs exactly one matching call, otherwise a violation is reported. Sometimes the need is for something else. A modifier TIMES(...) is used to change that. You can either specify an exact number of times matching calls must be made, or a range of numbers.

Example:

class Mock
{
public:
  MAKE_MOCK1(func, void(int));
};

void some_test()
{
  Mock m;

  REQUIRE_CALL(m, func(0))
    .TIMES(2);

  REQUIRE_CALL(m, func(1))
    .TIMES(3, 5);

  REQUIRE_CALL(m, func(2))
    .TIMES(AT_LEAST(3));

  REQUIRE_CALL(m, func(3))
    .TIMES(AT_MOST(4));

  func(&m);
}

Above, m.func(0) must be called exactly twice. m.func(1) must be called three, four or five times. The call m.func(2) must be made three or more times. Finally m.func(4) must not be called more than four times.

Controlling lifetime of mock objects

If you test a case where you hand over ownership of a mock object, you may want to test that the mock object is destroyed when intended. For this there is a modifier class template trompeloeil::deathwatched<T> and the macros REQUIRE_DESTRUCTION(...) and NAMED_REQUIRE_DESTRUCTION(...).

Example:

class Mock
{
public:
  virtual ~Mock() {}          // virtual destructor required for deathwatched<>
  MAKE_MOCK1(func, void(int));
}

template <typename T>
class consumer
{
public:
  consumer(T&&);
  void poke(int n);
private:
  ...
};

void consume_test()
{
  auto owner = std::make_unique<trompeloeil::deathwatched<Mock>>();

  auto mock = owner.get(); // use raw unowned pointer

  consumer<Mock> c(std::move(owner));

  {
    REQUIRE_CALL(*mock, func(3));

    c.poke(3);
  }
  {
    REQUIRE_CALL(*mock, func(-1));
    REQUIRE_DESTRUCTION(*mock);

    c.poke(0);
  }
}

Above, the constructor of object c takes ownership of the mock object.

Since the mock object is on deathwatch, destruction is reported as a violation. Thus we can be sure that if the constructor destroys the mock object, the test will fail. Likewise if the call c.poke(3) would destroy the mock object.

The local scope afterwards has a requirement that the mock object is destroyed. If the call c.poke(0) does not destroy the mock, a violation will be reported and fail the test. There is an implied order that the mock function func(-1) is called before the destruction of the mock object, since destroying any mock object that still has expectations is reported as a violation. It is also possible to be explicit with the sequencing by using IN_SEQUENCE(...) on both REQUIRE_CALL(...) and REQUIRE_DESTRUCTION(...), as below:

  {
    trompeloeil::sequence s;
    REQUIRE_CALL(*mock, func(-1));
      .IN_SEQUENCE(s);
    REQUIRE_DESTRUCTION(*mock);
      .IN_SEQUENCE(s);

    c.poke(0);
  }

Customize output format of values

When tracing or printing parameter values in violation reports, the values are printed using their stream insertion operators, if available, or hexadecimal dumps otherwise. If this is not what you want, you can provide your own output formatting used solely for testing.

The simple way to do this is to specialize a function template print(std::ostream&, const T&) in namespace trompeloeil for your type T.

Example:

class char_buff : public std::vector<char>
{
  ...
};

namespace trompeloeil {
  template <>
  void print(std::ostream& os, const char_buff& b)
  {
    os << b.size() << "#{ ";
    for (auto v : b) { os << int(v) << " ";
    os << "}";
  }
}

Any reports involving the char_buff above will be printed using the trompeloeil::print<char_buff>(...) function, showing the size and integer values.

Tracing mocks

Trompeloeil offers tracing as a way of manually following the calls of mocks. In pure TDD this is hardly ever needed, but if you are in the undesirable situation of exploring the behaviour of code written without tests, tracing can vastly simplify your job.

Simply put, tracing is exposing which mocks are called with which values.

Trompeloeil offers a stream_tracer, which outputs all calls to a std::ostream, but you can also write your own custom tracer.

Using trompeloeil::stream_tracer

stream_tracer is a mechanism used to find out how mock functions are called, by simply printing the calls with their parameter values on a std::ostream like std::cout.

There is no requirement from Trompeloeil on the expectations placed on the mocks, but open blanket ALLOW_CALL(...) can be a good start until more detailed tests can be written.

Example:

class Mock
{
public:
  MAKE_MOCK1(create, int(const std::string&));
  MAKE_MOCK1(func, std::string(int));
};

using trompeloeil::_;

void tracing_test()
{
  trompeloeil::stream_tracer tracer{std::cout};

  Mock m;

  ALLOW_CALL(m, create(_))
    .RETURN(3);

  ALLOW_CALL(m, func(_))
    .RETURN("value");

  weird_func(&m);
}

Running the above test will print on std::cout all calls made. A sample output may be:

/tmp/t.cpp:33
m.create(_) with.
  param  _1 = hello

/tmp/t.cpp:36
m.func(_) with.
  param  _1 = 3

/tmp/t.cpp:36
m.func(_) with.
  param  _1 = 2

/tmp/t.cpp:36
m.func(_) with.
  param  _1 = 1

Writing custom tracers

If tracing is important, but the trompeloeil::stream_tracer for some reason does not satisfy your needs, you can easily write your own tracer.

There is a base class:

namespace trompeloeil {

class tracer {
public:
  tracer();
  virtual ~tracer();
  virtual void trace(const char* file,
                     unsigned long line,
                     const std::string& call) = 0;
};

}

Write your own class inheriting from trompeloeil::tracer, and implement the member function trace, to do what you need, and you're done.

Writing custom matchers

If you need additional matchers over the ones provided by Trompeloeil (eq(...), ne(...), lt(...), le(...), gt(...) or ge(...), and re(...)), you can easily do so.

Matchers are created using the aptly named function template trompeloeil::make_matcher<Type>(...), which takes a predicate lambda to check the condition, a print lambda for error messages, and any number of stored values.

All matchers, including your own custom designed matchers, can be used as pointer matchers by using the unary prefix * dereference operator.

Typed matcher

The simplest matcher is a typed matcher. As an example of a typed matcher, an any_of matcher is shown, checking if a value is included in a range of values. It is implemented using the standard library algorithm std::any_of, allowing a parameter to match any of a set of values.

To create a matcher, you provide a function that calls trompeleil::make_matcher<Type>(...).

Below is the code for the function any_of(std::initializer_list<int>) which creates the matcher.

  inline auto any_of(std::initializer_list<int> elements)
  {
    return trompeloeil::make_matcher<int>( // matcher of int

      // predicate lambda that checks the condition
      [](int value, std::vector<int> const & alternatives) {
        return std::any_of(std::begin(alternatives), std::end(alternatives),
                           [&value](int element) { return value == element; });
      },

      // print lambda for error message
      [](std::ostream& os, std::vector<int> const& alternatives) {
        os << " matching any_of({";
        char const* prefix=" ";
        for (auto& element : alternatives)
        {
          os << prefix << element;
          prefix = ", ";
        }
        os << " }";
      },

      // stored value
      std::vector<int>(elements)
    )
  }

The predicate lambda is called with the value to check, and the stored values in order.

The print lambda is called with an ostream&, and the stored values in order.

You can capture values in the lambdas instead of storing in the matcher, but capturing them twice wastes memory, and what's in the lambda capture for the predicate lambda is not accessible in the print lambda.

Example usage:

class Mock
{
public:
  MAKE_MOCK1(func, void(int));
};

void test()
{
  Mock m;
  REQUIRE_CALL(m, func(any_of({1, 2, 4, 8})));

  m.func(7);
}

The print lambda is only called if a failure is reported. The report in the above example will look like:

No match for call of m.func with signature void(int) with.
  param  _1 = 7

Tried m.func(any_of({1, 2, 4, 8}) at file.cpp:12
  Expected _1 matching any_of({ 1, 2, 4, 8 });

Where everything after Expected _1 is the output from the print lambda.

Extending the example above to work with any type, using a template, is straight forward:

  template <typename T>
  inline auto any_of(std::initializer_list<T> elements)
  {
    return trompeloeil::make_matcher<T>( // matcher of T

      // predicate lambda that checks the condition
      [](T const& value, std::vector<T> const & alternatives) {
        return std::any_of(std::begin(alternatives), std::end(alternatives),
                           [&value](T const& element) { return value == element; });
      },

      // print lambda for error message
      [](std::ostream& os, std::vector<T> const& alternatives) {
        os << " matching any_of({";
        char const* prefix=" ";
        for (auto& element : alternatives)
        {
          os << prefix;
          ::trompeloeil::print(os, element);
          prefix = ", ";
        }
        os << " }";
      },

      // stored value
      std::vector<T>(elements)
    )
  }

The only difference compared to the int version, is that the predicate lambda accepts values by const& instead of by value, since T might be expensive to copy, and that the print lambda uses trompeloeil::print(...) to print the elements.

Duck-typed matcher

A duck-typed matcher accepts any type that matches a required set of operations. An example of a duck-typed matcher is a not_empty() matcher, requiring that a .empty() member function of the parameter returns false. Another example is an is_clamped(min, max) matcher, that ensures min <= value && value <= max.

A duck-typed matcher is created by specifying trompeloeil::wildcard as the type to to trompeloeil::make_matcher<Type>(...).

It is also important that the predicate lambda uses a trailing return type specifier, which uses the required operations, in order to filter out calls that would not compile.

A not_empty() matcher

Here's an implementation of a not_empty() matcher.

  inline auto not_empty()
  {
    return trompeloeil::make_matcher<trompeloeil::wildcard>( // duck typed

      // predicate lambda that checks the condition
      [](auto const& value) -> decltype(!value.empty()) {
        return !value.empty();
      },

      // print lambda for error message
      [](std::ostream& os) {
        os << " is not empty";
      }

      // no stored values
    );
  }

It is unfortunate that the !value.empty() condition is expressed twice, but those are the rules of the language.

Here's an example of the usage.

  struct C
  {
    MAKE_MOCK1(func, void(int));
    MAKE_MOCK1(func, void(std::string&&));
    MAKE_MOCK1(func2, void(std::vector<int> const&);
  };

  void test()
  {
    C obj;
    REQUIRE_CALL(obj, func(not_empty()));  // std::string&&
    REQUIRE_CALL(obj, func2(not_empty())); // std::vector<int> const&
    func_under_test(&obj);
  }

The expectation placed on func() is not ambiguous. While func() is overloaded on both int and std::string&&, the trailing return type specification on the predicate lambda causes SFINAE to kick in and chose only the std::string&& overload, since .empty() on an int would not compile.

If you make a mistake and place an expectation with a duck-typed matcher that cannot be used, the SFINAE on the trailing return type specification of the predicate lambda, ensures a compilation error at the site of use (REQUIRE_CALL(), ALLOW_CALL() or FORBID_CALL().)

TIP! The expectation on func() in the example above is not ambiguous, as explained, but what if func2 had been yet an overload of func() instead? You can easily make your matchers typed or duck-typed at the user's discretion. Alter the not_empty() to be a function template, with trompeloeil::wildcard as the default.

  template <typename Type = trompeloeil::wildcard>
  inline auto not_empty()
  {
    return trompeloeil::make_matcher<Type>( // typed or duck typed

      // predicate lambda that checks the condition
      [](auto const& value) -> decltype(!value.empty()) {
        return !value.empty();
      },

      // print lambda for error message
      [](std::ostream& os) {
        os << " is not empty";
      }

      // no stored values
    );
  }

Now, if the user writes EXPECT_CALL(obj, func(not_empty())), it is duck-typed, but if the user writes EXPECT_CALL(obj, func<std::string&&>() it will only match a call with a std::string&& parameter.

An is_clamped(min, max) matcher

Here's an implementation of an is_clamped(min, max) matcher.

  template <typename Type = trompeloeil::wildcard, typename T, typename U>
  inline auto is_clamped(T const& min, U const& max)
  {
    return trompeloeil::make_matcher<Type>( // typed or duck typed

      // predicate lambda that checks the condition
      [](auto const& value, auto const& lower, auto const& upper)
       -> decltype(lower <= value && value <= upper)
      {
        return !trompeloeil::is_null(value) && lower <= value && value <= upper;
      },

      // print lambda for error message
      [](std::ostream& os, auto const& lower, auto const& upper) {
        os << " clamped by [";
        trompeloeil::print(os, lower);
        os << ", ";
        trompeloeil::print(os, upper);
        os << "]";
      }

      // stored values
      min,
      max
    );
  }

The trompeloeil::is_null(value) in the predicate lambda is there to prevent against e.g. clamp checks for const char* between two std::strings, where the const char* may be null. The is_null() check is omitted in the trailing return specification, because it does not add anything to it - it always returns a bool and it works for all types.

By allowing min and max to be different types, it becomes possible to, e.g. check that a std::string_view is clamped by a std::string and a const char*.

NOTE! There is a bug in GCC versions 5.3 and lower, that does not allow trailing return type specifications in lambdas expressed in template functions. The work around is annoying but simple:

  inline auto is_clamped_predicate()
  {
    return [](auto const& value, auto const& lower, auto const& upper)
           -> decltype(lower <= value && value <= upper) {
             return !trompeloeil::is_null(value) && lower <= value && value <= upper;
           };
  }

  template <typename Type = trompeloeil::wildcard, typename T, typename U>
  inline auto is_clamped(T const& min, U const% max)
  {
    return trompeloeil::make_matcher<Type>( // duck typed

      // predicate lambda that checks the condition
      is_clamped_predicate(),
      ...

NOTE! There is also a bug in VisualStudio 2015 Update 3, which does not respect the trailing return type specifications of lambdas in the context of template deduction. The work around is annoying but simple - use a struct instead:

  struct is_clamped_predicate
  {
    template <typename T, typename L, typename U>
    auto operator()(T const& value, L const& lower, U const& upper)
    -> decltype(lower <= value && value <= upper)
    {
      return !trompeloeil::is_null(value) && lower <= value && value <= upper;
    }
  };

  template <typename Type = trompeloeil::wildcard, typename T, typename U>
  inline auto is_clamped(T const& min, U const% max)
  {
    return trompeloeil::make_matcher<Type>( // duck typed

      // predicate lambda that checks the condition
      is_clamped_predicate(),
      ...

Legacy Matchers

Before trompeloeil::make_matcher<Type>(...) was introduced in Trompeloeil v18, writing matchers was more elaborate. This section is here for those who need to maintain old matcher code.

All legacy matchers

  • inherit from trompeloeil::matcher or trompeloeil::typed_matcher<T>
  • implement a bool matches(parameter_value) const member function
  • implement an output stream insertion operator

All legacy matchers can be used as pointer matchers by using the unary prefix * dereference operator.

Typed legacy matcher

Typed legacy matchers are relatively easy to understand. As an example of a typed matcher, an any_of matcher is shown, mimicking the behaviour of the standard library algorithm std::any_of, allowing a parameter to match any of a set of values.

For templated matchers, it is often convenient to provide a function that creates the matcher object. Below is the code for any_of_t<T>, which is the matcher created by the any_of(std::vector<T>) function template.

  template <typename T>
  class any_of_t : public trompeloeil::typed_matcher<T>
  {
  public:
    any_of_t(std::initializer_list<T> elements)
    : alternatives(elements)
    {
    }
    bool matches(T const& t) const
    {
      return std::any_of(std::begin(alternatives), std::end(alternatives),
                         [&t](T const& val) { return t == val; });
    }
    friend std::ostream& operator<<(std::ostream& os, any_of_t<T> const& t)
    {
      os << " matching any_of({";
      char const* prefix=" ";
      for (auto& n : t.alternatives)
      {
        os << prefix;
        trompeloeil::print(os, n);
        prefix = ", ";
      }
      return os << " })";
    }
  private:
    std::vector<T> alternatives;
  };

  template <typename T>
  auto any_of(std::initializer_list<T> elements)
  {
    return any_of_t<T>(elements);
  }

The matches member function at accepts the parameter and returns true if the value is in the specified set, in this case if it is any of the values stored in the alternatives vector, otherwise false.

Example usage:

class Mock
{
public:
  MAKE_MOCK1(func, void(int));
};

void test()
{
  Mock m;
  REQUIRE_CALL(m, func(any_of({1, 2, 4, 8}));

  m.func(7);
}

The output stream insertion operator is only called if a failure is reported. The report in the above example will look like:

No match for call of m.func with signature void(int) with.
  param  _1 = 7

Tried m.func(any_of({1, 2, 4, 8}) at file.cpp:12
  Expected _1 matching any_of({ 1, 2, 4, 8 });

Where everything after Expected _1 is the output from the stream insertion operator.

Duck-typed legacy matcher

A duck-typed matcher accepts any type that matches a required set of operations. Duck-typed legacy matchers have a type conversion operator that selects which types it can operate on. The conversion operator is never implemented, but the signature must be available since it is used at compile time to select overload.

As an example of a duck-typed matcher is a not_empty matcher, requiring that a .empty() member function of the parameter returns false.

First the restricting SFINAE predicate used to match only types that has a .empty() member function.

  template <typename T>
  class has_empty
  {
    template <typename U>
    static constexpr std::false_type func(...) { return {}; }
    template <typename U>
    static constexpr auto func(U const* u) -> decltype(u->empty(),std::true_type{})
    {
      return {};
    }
  public:
    static const bool value = func<T>(nullptr);
  };

Here has_empty<T>::value is true only for types T that has a .empty() member function callable on const objects.

  class not_empty : public trompeloeil::matcher
  {
  public:
    template <typename T,
              typename = std::enable_if_t<has_empty<T>::value>>
    operator T() const;            //1
    template <typename T>
    bool matches(T const& t) const //2
    {
      return !t.empty();
    }
    friend std::ostream& operator<<(std::ostream& os, not_empty const&)
    {
      return os << " is not empty";
    }
  };

At //1 the type conversion operator selects for types that has a .empty() member function. std::enable_if_t<> ensures that no calls to mismatching types will occur, and that if no matching call can be found, a compilation error is generated at the site of use (REQUIRE_CALL(), ALLOW_CALL() or FORBID_CALL().)

The matches(T const&) member function at //2 becomes very simple. It does not need the SFINAE std::enable_if_t<> to select valid types, since a type mismatch gives a compilation error on the type conversion operator at //1.

The output stream insertion operator is neither more or less tricky than with typed matchers. Making violation reports readable may require some thought, however.