How to write a library

This is merely a convention, but you are advised to follow it, especially if you are working in a large project.

Let’s say you have a foo library.

You have the following files:

  • foo.hpp: the public header of the foo library. It contains the public API, and include the least possible number of other headers. (Use of forward declarations, and PIMPL implementations are recommanded)
  • foo.cpp : implementation of the foo.hpp functions
  • foo_private.hpp: private header of the foo library. This one may include third-party headers (say zeromq.h), without having the foo.hpp header depending on zeromq.h, which is nice for the users of your library. If you link statically with zeromq, users of foo won’t even need to know about zeromq (well, this is true if foo is a dynamic library, but that’s an other topic)
  • foo_private.cpp : private implementation.
  • test_foo.cpp : You would not dare writing a library without unit tests, would you?

Proposed layout

This is what your layout should look like:

|__ libfoo
    | CMakeLists.txt
    |__ foo
    |   |__ foo.hpp
    |__ src
    |   |__ foo.cpp
    |   |__ foo_private.hpp
    |   |__ foo_private.cpp
    |__ test
        |__ CMakeLists.txt
        |__ foo_test.cpp
  • The full path to the public header is libfoo/foo/foo.hpp. Note that the name of the root directory is fooproject
  • The private code is put in a src sub-directory. Private and public directories are separated, it’s easy to search only in public headers.

Note: you can download an archive containing the foo project here:


Here’s what the CMakeLists.txt should look like


  SRC foo/foo.hpp


qi_install_header(KEEP_RELATIVE_PATHS foo/foo.hpp)



Please note that the location of the CMake list file matters.


You will note that:

  • The only time we call include_directories() is when we are staging the foo library.
  • The foo.hpp header is in a directory named foo, and will be installed to foo/foo.hpp. It’s advised you use the same name for the target and the subdirectory.
  • Everything that need a foo header must use
#include <foo/...>

This way, we are sure that the code we use can be re-distributed when the headers are installed, and that the path to find the headers while in the source tree does not differ from the paths to find the installed headers. This works because:

  • We have put foo.hpp in a foo subdirectory.
  • We have used qi_install_header with the KEEP_RELATIVE_PATHS argument. You could also have used it with a SUBFOLDER argument, like this:
qi_install_header(foo/foo.hpp SUBFOLDER foo)
  • Let’s assume you have two libraries, foo and bar, and a foobar executable that needs code from foo and bar.

With the proposed layout, you have something like:

|__ libfoo
|    |__ foo
|         |__ foo.hpp
|__ libbar
|    |__ bar
|         |__ bar.hpp
|__ foobar
     |__ foobar.cpp

You may want to get rid of the libfoo/foo, libbar/bar redundancy and do this instead:

|__ foo
|   |__ foo.hpp
|__ bar
|   |__ bar.hpp
|__ foobar
    |__ foobar.cpp

But, let’s assume you make a mistake, and write

qi_use_lib(foobar foo)

instead of

qi_use_lib(foobar foo bar)

In the first layout, you will have an error during compile time, looking like:

bar/bar.hpp : no such file or directory

because the include directory that has been staged for foo is different from the include directory that has been staged for bar. Using the second layout, you will have an error during link time, looking like:

undefined reference to `bar_func'

because the include directory that was staged was always the same: fooproject. The additional nesting level helps you catch this king of errors early.


For large libraries, also consider using submodules. The documentation can be found here