CMake

CMake seems to me like the best available cross-platform build tool. It can generate Visual Studio solutions on Windows and Make files on POSIX systems. While it's widely used and there exists thorough reference documentation for it, I feel like there aren't many resources for getting started. As a result, I've written up some notes from what I've been able to infer from a variety of project's CMake configurations as well as the scant introductory information I was able to find.

# General Usage

CMake directives are stored in files named CMakeLists.txt, and there is usually one such file present for each sub-directory in the source tree, each one conventionally containing directives pertinent to the files in that directory. These separate CMake files are then "merged together" using the add_subdirectory directive, which immediately makes the CMake interpreter load and evaluate the CMake file in the provided directory.

Variables can be set using the set directive in the following format:

set (TEST "something")

There are certain built-in CMake variables that can affect the project. A commonly set one is CMAKE_MODULE_PATH which defines where CMake modules should be looked for.

Similarly, user-modifiable options can be created with the option directive which takes a variable to associate the value with, a description string, and a default value (OFF is the default):

option (USE_FFTW "use the fftw library" OFF)

Dependencies are generally found using the find_package directive. These are backed by CMake modules---either built-in or found in the CMAKE_MODULE_PATH---and handle the logic of searching for the libraries and headers of the particular package in various common locations.

If the search was successful, these modules typically set a variable of the form PACKAGENAME_FOUND which can be tested. Further, they also set variables such as PACKAGENAME_LIBRARIES and PACKAGENAME_INCLUDE_DIRS, which are sometimes singular and sometimes plural.

In pertinent CMake files, these variables set by the find_package module are then used to resolve any dependencies in the code. For example, required headers can be added to the set of directories searched by the compiler using the include_directories directive:

include_directories (${PACKAGENAME_INCLUDE_DIR})

Similarly, a target can be linked with a library with the target_link_libraries directive:

target_link_libraries (sometarget ${PACKAGENAME_LIBRARIES})

Targets are essentially products of the build process: oftentimes this is either an executable or a library, static or shared. An executable can be created using the add_executable directive:

add_executable (sometarget file.cpp)

In the above, file.cpp refers to the file containing the main entry point. Likewise, a library can be created using the add_library directive:

add_library (sometarget STATIC ${SRC})

In the above, ${SRC} refers to the set of files whose resultant code would constitute the library. These variables are simply created using the set directive:

set (SRC somefile.cpp anotherfile.cpp)

Notice that the header files aren't necessary in the set since the C++ files already #include them. However, when using the Visual Studio generator on Windows, this means that the header files will be absent from the project. This can be resolved by adding a conditional for Windows and adding them in:

if (WIN32)
  set (SRC somefile.h anotherfile.h)
endif (WIN32)

Alternatively, the header files can be added in to the set unconditionally.

Furthermore, on Windows, the source_group directive can help organize source files into separate projects in the Visual Studio solution:

source_group (sometarget FILES ${SRC})

Libraries built in project are automatically found, there's no need for include_directories or find_library. Try to avoid link_directories.

It's possible to wire-up what would happen upon invoking make install by using the install directive:

install (TARGET sometarget
         RUNTIME DESTINATION bin
         LIBRARY DESTINATION lib)

This would install to /usr/local/bin and /usr/local/lib on POSIX.

A configuration include file can be created using configure_file:

configure_file ("Config.h.in" "Config.h")

In this configuration file, variables can be interpolated based on their values in the CMake build system:

#cmakedefine VAR // becomes #define VAR if VAR is true
#define @VAR@    // replaced with value of VAR

The add_custom_target directive creates a custom target that is always rebuilt. Similarly, add_custom_command runs a command before or after a build, or before a link.

pkg-config is discouraged.

# rpath

One problem is that on POSIX systems, libraries are usually exported as libsomething.so and are searched for in specified locations. However, it probably makes no sense to install a library used only by a single program to the system's library path. This can be circumvented by loading the file by path, including the SO's full name.

An alternative is to modify the executable's rpath using the $ORIGIN linker variable, which allows the executable to search relative to its directory, i.e. the $ORIGIN, for shared libraries by name. There are a variety of different ways to accomplish this.

The simplest way to accomplish this is to use the CMAKE_EXE_LINKER_FLAGS variable to specify the linker flag. Clang will complain if these are defined in CMAKE_CXX_FLAGS because those flags are passed to it even when it's merely compiling -c source files, to which linker flags obviously don't apply.

set (CMAKE_EXE_LINKER_FLAGS "${CMAKE_EXE_LINKER_FLAGS} -Wl,--disable-new-dtags")
set (CMAKE_EXE_LINKER_FLAGS "${CMAKE_EXE_LINKER_FLAGS} -Wl,-rpath,'$ORIGIN'")

There's also a "built-in" way to accomplish this using CMake. By default, this method only applies to installed files, which means it won't apply until the project's installation procedure is run. This is easily changed so that it applies during regular builds:

set (CMAKE_EXE_LINKER_FLAGS "${CMAKE_EXE_LINKER_FLAGS} -Wl,--disable-new-dtags")
set (CMAKE_BUILD_WITH_INSTALL_RPATH TRUE)
set (CMAKE_INSTALL_RPATH "$ORIGIN")

It's important to note that the rpath corresponds to an ELF dynamic section attribute. Historically the attribute has been DT_RPATH, but recently it has fallen out of favor for a more flexible DT_RUNPATH which honors LD_LIBRARY_PATH, allowing for greater user flexibility.

The thing with DT_RUNPATH is that it's not transitive, meaning, if your binary with a specified DT_RUNPATH links with a library that itself loads another library via dlopen(), that library load won't honor the DT_RUNPATH, by design.

In my case, the reason I wanted to set the rpath was specifically for this use case. The shared object I have exposes a game engine through a C API which is then loaded by LuaJIT's FFI system. This of course performs a dlopen() to load the shared object, so the rpath won't apply. This apparently bit other people as well, including gnome.

A quick fix for this is to revert to using the DT_RPATH attribute, which can be done by explicitly setting the --disable-new-dtags linker flag. Alternative solutions would include hard coding the library name---which would introduce system dependent naming conventions---or wrapping the binary in a script that sets LD_LIBRARY_PATH.

The solution gnome went with was to link with the shared object that's loaded after the fact, so that its dlopen() is considered to be done by the binary itself, in which case the DT_RUNPATH is honored.

It's simple to verify that the produced binary contains the attributes by running:

$ readelf -d thebinary

# SWIG Bindings

It's possible to automatically generate SWIG bindings using built-in CMake modules. The following creates a a dynamic library which a Lua script can subsequently require.

find_package (SWIG)
include (UseSWIG)

include_directories (${CMAKE_CURRENT_SOURCE_DIR} ${LUAJIT_INCLUDE_DIR})
set_source_files_properties (swig.i PROPERTIES CPLUSPLUS ON)
swig_add_module (script lua swig.i)
swig_link_libraries (script ${LUAJIT_LIBRARIES})

If instead it is desired to statically link the bindings to an existing target, one can create a custom target:

add_custom_command (
  OUTPUT ${CMAKE_CURRENT_SOURCE_DIR}/swig.cpp
  COMMAND ${SWIG_EXECUTABLE} -lua -c++
          -o ${CMAKE_CURRENT_SOURCE_DIR}/swig.cpp
          ${CMAKE_CURRENT_SOURCE_DIR}/swig.i
  DEPENDS ${CMAKE_CURRENT_SOURCE_DIR}/swig.i)

add_library (script STATIC swig.cpp)
add_dependencies (script swig)

include_directories (${LUAJIT_INCLUDE_DIR})
target_link_libraries (script ${LUAJIT_LIBRARIES})

Then the target that wants to link with it just needs to add_dependencies. Then the library should be opened using the generated luaopen_script function which the program should of course declare as extern.

It is common to place generated files in the ${CMAKE_BINARY_DIR} directory to avoid clobbering the source directory.

# Visual Studio

CMake works fine with Visual Studio but there are a few things to consider. The property for the working directory in the Debugging section should most likely be set to $(OutDir). Likewise, the start-up project must be set manually as it's set to ALL_BUILD by default. ALL_BUILD is a project that builds all projects and correctly triggers any scripts. ZERO_CHECK is a project that runs and, if any CMake files have been changed, asks to reload Visual Studio.

Furthermore, if you're making a Windows application, you should add the WIN32 parameter to add_executable to instruct the compiler to use the WinMain entry-point and WINDOWS subsystem:

add_executable (target WIN32 ${SOURCES})

# Resources

September 1, 2013