Meeting 07: Unit Testing using Catch

The goal of today's meeting it to learn about a very important part of programming, testing.

Unit tests
Integration tests
Testing functional requirements
Error and Recovery testing
Performance testing
Usability testing
How to test
Unit Testing using Catch
Test coverage

Links:

Slides
Notes on Testing

Exercise 07: Testing unordered_map

GitHub Invitation URL: exercise07

Steps:

Create a working directory somewhere on your computer, then change to that directory.
Clone the assignment for today after accepting the GitHub invitation at the link above.
```
git clone https://github.com/VTECE3574/exercise07-USER.git
```
where USER is your GitHub username. You may have to enter your GitHub username and password.
Examine the repository. There is a single file called catch.hpp. That is all that is needed to use Catch.
Write some unit tests for the standard library container std::unordered_map in a file named test_unordered_map.cpp in the top-level directory of the repository. Compile and run your tests.
Now, use git to commit the source file you added to the local repository.
Finally, use git push to synchronize the repository with that on GitHub

You have completed the Exercise.

Some Notes on Testing

Suppose that we wanted to write a C++ class, Puzzle, that models an eight-tile sliding puzzle. You have probably seen these, a square array of tiles with numbers, letters, or part of an image printed on them, and one blank space. The tiles can slide left-right and up-down within the puzzle, exchanging positions with the empty location. The goal is, from a scrambled state, slide the tiles around until the tiles show a particular image or spell some text.

Consider a 3 by 3 puzzle with eight tiles, addressed by row and column with 0-based indexing, with labels 'A' through 'H' and an empty spot denoted by the label ' '(space). For example:

ABC
DEF
GH

is empty at position (2,2), has tile 'D' in position (1,0), tile 'B' on position (0,1), etc.

Our goal here is not to write a program to solve the puzzle (although it is fun to do), but to just model the puzzle itself, a component of a larger puzzle solver program. But we want to ensure that the code works, perhaps before we write either the puzzle class or the solver, so we need to write some unit tests, code that tests individual units of other code, in this case the Puzzle class.

Lets begin by defining how our class should behave, it's specification. Puzzle should support:

Construction of a default puzzle instance with the layout in the example above,
a move method taking two position arguments from and to, throwing an exception if either position or the move is invalid
and a get method taking a position argument and returning the tile label at that position, throwing an exception if the position is invalid.

This specification is pretty detailed but it still has some missing information. For example what types should the position, labels, and exceptions be? We can nail down the specification further and define a set of tests that tell us how well we are doing implementing Puzzle by writing a test before we write the Puzzle class. This is called Test-Driven-Development or TDD.

In it's simplest form unit tests are just a program that tries to use the code being tested. So we might write a file puzzle_test.cpp:

#include "puzzle.h"

void run_tests();

int main()
{
  run_tests();

  return 0;
}

where the function run_tests has yet to be implemented and the puzzle.h file does not exist yet.

Lets implement the first version of our test by appending the following to puzzle_test.cpp:

void run_tests()
{
  Puzzle p;
}

All this function does (at this point) is attempt to create an instance (an object/variable named p) of type Puzzle. If we try to compile this we get the error along the lines of

puzzle_test.cpp:1:10: fatal error: 'puzzle.h' file not found
#include "puzzle.h"
         ^
1 error generated.

Congratulations!, we have written our first failing test (it will not even compile!).

So let us fix the problem. We clearly need to create a file name ``puzzle.h'' defining a type Puzzle, like so:

class Puzzle {};

Now if we compile puzzle_test.cpp it gives no errors, and it even runs. But clearly the test is not very good, we say it does not cover the functionality of the Puzzle specification. Notice we have started to define the puzzle class, but it is what we call a stub, it is just a placeholder to get the tests to at least compile.

Improving the tests

Ok, lets get more serious about our tests. In order to test the specification we need test code that calls and checks the constructor, the get method, and the move method. To test the constructor we can adapt our simple test above to check, or assert, the instantiated object p has the correct default contents:

void run_tests()
{
  Puzzle p;
  
  assert(p.get(0,0) == Puzzle::A);
  assert(p.get(0,1) == Puzzle::B);
  assert(p.get(0,2) == Puzzle::C);
  assert(p.get(1,0) == Puzzle::D);
  assert(p.get(1,1) == Puzzle::E);
  assert(p.get(1,2) == Puzzle::F);
  assert(p.get(2,0) == Puzzle::G);
  assert(p.get(2,1) == Puzzle::H);
  assert(p.get(2,2) == Puzzle::EMPTY);
}

This requires adding the include cassert at the top of the puzzle_test.cpp file to use assert from the standard library. To get this to compile we will need to extend our stub to define the type for the label and the default values. We can use an enum for this:

class Puzzle
{
public:

  enum LabelType {A,B,C,D,E,F,G,H,EMPTY};

  LabelType get(int row, int col)
  {
    return A;
  }
};

Compiling and running this gives us what we expect:

Assertion failed: (p.get(0,1) == Puzzle::B), function run_tests, file puzzle_test.cpp, line 19.

since the stub always returns the label A. It may seem odd but we are not concerned with the tests passing at this point just writing the tests to cover the functionality desired and the minimal stub necessary to get the tests to compile and run.

Now lets test our move function by making a legal move and checking that it actually occurred. So that our testing code does not get too messy, lets refactor the tests into separate test functions with more meaningful names

void test_constructor()
{
  Puzzle p;

  assert(p.get(0,0) == Puzzle::A);
  assert(p.get(0,1) == Puzzle::B);
  assert(p.get(0,2) == Puzzle::C);
  assert(p.get(1,0) == Puzzle::D);
  assert(p.get(1,1) == Puzzle::E);
  assert(p.get(1,2) == Puzzle::F);
  assert(p.get(2,0) == Puzzle::G);
  assert(p.get(2,1) == Puzzle::H);
  assert(p.get(2,2) == Puzzle::EMPTY);
}

void test_move()
{
  Puzzle p;

  p.move(2,1,2,2); // slide H to the right
  assert(p.get(2,1) == Puzzle::EMPTY);
  assert(p.get(2,2) == Puzzle::H);
}

int main()
{
  test_constructor();
  test_move();
  
  return 0;
}

To get this to compile requires defining a stub method inside the Puzzle class of puzzle.h

  void move(int from_row, int from_col, int to_row, int to_col)
  {
    // do nothing
  }

So now we have some basic unit tests, most of which fail (the first test in test_constructor passes by accident). This gives us a few important things:

We have a goal to work toward, namely to implement the methods of Puzzle so that all the tests pass.
When all the tests pass, we have a reasonable belief the code is correct. There is a saying, often attributed to Fred Brooks, "UNTESTED code is BROKEN code".
Further, there are automated tools that can check how many of the statements we write in our implementations are covered by the tests. This is called the test code coverage. Particularly in critical applications e.g. avionics, medical devices, 100% code coverage is needed.
We can spot potential design flaws in the detailed design of the code early on in the process.

An example of the latter is in the move test, where the call to slide tile H to the right looks like p.move(2,1,2,2);. A function/method call with that many arguments is a code smell, one of many we will see during the course. Unless I am looking at the definition of the method or some documentation how do I remember if it is from then to, to to then from, and if the row is first or the column? Such code is ripe for being used incorrectly and causing a bug. So we might add a separate type defining position and modify the call. Note this happens early in the coding process, before we have put even minimal effort into implementing the methods. That does not mean that design flaws do not show up later, but we catch them earlier this way, when refactoring is easier. Paradoxically, the time put into writing test and the stub saves time overall. This process of test writing, stub writing, refactoring continues until all the functionality is covered and the design is in good shape. The result is a complete specification. Only then do we worry about implementing the methods to get the tests to pass.

Our testing approach using assert is not very fancy. It stops at the first test that fails and does not give us very good diagnostic information about which test is failing. We could print out additional information and write our own version of assert that does not abort right away, however this is so common a task that there are many libraries that provide support for this. The one we will use is called Catch.