418

About a year ago I asked about header dependencies in CMake.

I realized recently that the issue seemed to be that CMake considered those header files to be external to the project. At least, when generating a Code::Blocks project the header files do not appear within the project (the source files do). It therefore seems to me that CMake consider those headers to be external to the project, and does not track them in the depends.

A quick search in the CMake tutorial only pointed to include_directories which does not seem to do what I wish...

What is the proper way to signal to CMake that a particular directory contains headers to be included, and that those headers should be tracked by the generated Makefile?

0

11 Answers 11

427

Two things must be done.

First add the directory to be included:

target_include_directories(test PRIVATE ${YOUR_DIRECTORY})

In case you are stuck with a very old CMake version (2.8.10 or older) without support for target_include_directories, you can also use the legacy include_directories instead:

include_directories(${YOUR_DIRECTORY})

Then you also must add the header files to the list of your source files for the current target, for instance:

set(SOURCES file.cpp file2.cpp ${YOUR_DIRECTORY}/file1.h ${YOUR_DIRECTORY}/file2.h)
add_executable(test ${SOURCES})

This way, the header files will appear as dependencies in the Makefile, and also for example in the generated Visual Studio project, if you generate one.

How to use those header files for several targets:

set(HEADER_FILES ${YOUR_DIRECTORY}/file1.h ${YOUR_DIRECTORY}/file2.h)

add_library(mylib libsrc.cpp ${HEADER_FILES})
target_include_directories(mylib PRIVATE ${YOUR_DIRECTORY})
add_executable(myexec execfile.cpp ${HEADER_FILES})
target_include_directories(myexec PRIVATE ${YOUR_DIRECTORY})
16
  • 2
    My question was more in the sense that I have several libraries which depend from each other: libroot, liba depends on libroot, libb depends on libroot. Can I use the LIBROOT_HEADER_FILES variable in liba/CMakefile and libb/CMakefile then ? Dec 4, 2012 at 13:37
  • 8
    This is wrong, you should never use include_directories over target_include_directories. The former sets it recursively for all targets in that directory; whereas the latter sets it for a target. Doing the former breaks the notion of a target graph in CMake, and instead relies on side effects to your file hierarchy.
    – Andy
    Aug 10, 2017 at 1:47
  • 1
    Why do you have to explicitly list the .h files in your executable's source files? Surely you should just specify the path to your headers then use #include <filename.h> from your implementation .cpp files?
    – donturner
    Dec 13, 2018 at 14:07
  • 4
    @donturner You don't have to add .h files into add_executable. But, It does have the nice benefit of making the files show up in Visual Studio projects in the expected location. Makefiles uses the internal cmake -E cmake_depends to generate dependencies from the source files (header files in add_executable are skipped). There are known issues with this the scanner. Also ` CMake's Makefile generator dependency scanner does only approximate preprocessing.` Computed header includes and that sort of thing will not work.
    – fdk1342
    Jan 1, 2019 at 7:12
  • 2
    Many times I've tried to find this answer but could only find answers to similar questions and which recommend using GLOB and I eventually give up looking. Would be super handy if this answer was included/linked on the other questions or they were phrased better to be distinctly dissimilar.
    – Chris
    Sep 6, 2021 at 1:15
110

First, you use include_directories() to tell CMake to add the directory as -I to the compilation command line. Second, you list the headers in your add_executable() or add_library() call.

As an example, if your project's sources are in src, and you need headers from include, you could do it like this:

include_directories(include)

add_executable(MyExec
  src/main.c
  src/other_source.c
  include/header1.h
  include/header2.h
)
7
  • 40
    Do you really need to add headers to add_executable? I thought CMake figured out the include file dependencies automatically. Nov 2, 2013 at 17:40
  • 80
    @ColinDBennett You don't have to list them for dependency reasons - CMake figures out build dependencies just fine if you don't. But if you list them, they are considered part of the project, and will be listed as such in IDEs (which was the topic of the question). Nov 2, 2013 at 18:57
  • At least for QtCreator it is not necessary to add class.h in case a class.cpp exists. Only lonely.h needs to be added to source. See tutorial at www.th-thielemann.de/cmake Jan 22, 2018 at 18:47
  • 1
    If you've specified "include" as one of the include directories, why would you need to prefix the header files in your source-file list with "include/"?
    – Oscar
    Jul 19, 2022 at 23:00
  • 1
    @Oscar Answered in a comment above (in reply to Colin D Bennett) Jul 22, 2022 at 9:24
60

Structure of project

.
├── CMakeLists.txt
├── external //We simulate that code is provided by an "external" library outside of src
│   ├── CMakeLists.txt
│   ├── conversion.cpp
│   ├── conversion.hpp
│   └── README.md
├── src
│   ├── CMakeLists.txt
│   ├── evolution   //propagates the system in a time step
│   │   ├── CMakeLists.txt
│   │   ├── evolution.cpp
│   │   └── evolution.hpp
│   ├── initial    //produces the initial state
│   │   ├── CMakeLists.txt
│   │   ├── initial.cpp
│   │   └── initial.hpp
│   ├── io   //contains a function to print a row
│   │   ├── CMakeLists.txt
│   │   ├── io.cpp
│   │   └── io.hpp
│   ├── main.cpp      //the main function
│   └── parser   //parses the command-line input
│       ├── CMakeLists.txt
│       ├── parser.cpp
│       └── parser.hpp
└── tests  //contains two unit tests using the Catch2 library
    ├── catch.hpp
    ├── CMakeLists.txt
    └── test.cpp

How to do it

1. The top-level CMakeLists.txt is very similar to Recipe 1, Code reuse with functions and macros

cmake_minimum_required(VERSION 3.5 FATAL_ERROR)
  
project(recipe-07 LANGUAGES CXX)

set(CMAKE_CXX_STANDARD 11)
set(CMAKE_CXX_EXTENSIONS OFF)
set(CMAKE_CXX_STANDARD_REQUIRED ON)

include(GNUInstallDirs)
set(CMAKE_ARCHIVE_OUTPUT_DIRECTORY
  ${CMAKE_BINARY_DIR}/${CMAKE_INSTALL_LIBDIR})
set(CMAKE_LIBRARY_OUTPUT_DIRECTORY
  ${CMAKE_BINARY_DIR}/${CMAKE_INSTALL_LIBDIR})
set(CMAKE_RUNTIME_OUTPUT_DIRECTORY
  ${CMAKE_BINARY_DIR}/${CMAKE_INSTALL_BINDIR})

# defines targets and sources
add_subdirectory(src)

# contains an "external" library we will link to
add_subdirectory(external)

# enable testing and define tests
enable_testing()
add_subdirectory(tests)

2.Targets and sources are defined in src/CMakeLists.txt (except the conversion target)

add_executable(automata main.cpp)
  
add_subdirectory(evolution)
add_subdirectory(initial)
add_subdirectory(io)
add_subdirectory(parser)

target_link_libraries(automata
  PRIVATE
    conversion
    evolution
    initial
    io
    parser
  )

3.The conversion library is defined in external/CMakeLists.txt

add_library(conversion "")

target_sources(conversion
  PRIVATE
    ${CMAKE_CURRENT_LIST_DIR}/conversion.cpp
  PUBLIC
    ${CMAKE_CURRENT_LIST_DIR}/conversion.hpp
  )

target_include_directories(conversion
  PUBLIC
    ${CMAKE_CURRENT_LIST_DIR}
  )

4.The src/CMakeLists.txt file adds further subdirectories, which in turn contain CMakeLists.txt files. They are all similar in structure; src/evolution/CMakeLists.txt contains the following:

add_library(evolution "")

target_sources(evolution
  PRIVATE
    ${CMAKE_CURRENT_LIST_DIR}/evolution.cpp
  PUBLIC
    ${CMAKE_CURRENT_LIST_DIR}/evolution.hpp
  )
target_include_directories(evolution
  PUBLIC
    ${CMAKE_CURRENT_LIST_DIR}
  )

5.The unit tests are registered in tests/CMakeLists.txt

add_executable(cpp_test test.cpp)

target_link_libraries(cpp_test evolution)

add_test(
  NAME
    test_evolution
  COMMAND
    $<TARGET_FILE:cpp_test>
  )

How to run it

$ mkdir -p build
$ cd build
$ cmake ..
$ cmake --build .

Refer to: https://github.com/sun1211/cmake_with_add_subdirectory

1
  • 4
    The question didn't ask anything about unit tests or using third-party projects. To me those parts of this answer just seem like unnecessary noise. Sure, it could be useful in a question that asks for that info, but that's not the case here.
    – starball
    Nov 12, 2022 at 0:39
39

Add include_directories("/your/path/here").

This will be similar to calling gcc with -I/your/path/here/ option.

Make sure you put double quotes around the path. Other people didn't mention that and it made me stuck for 2 days. So this answer is for people who are very new to CMake and very confused.

1
  • 1
    except it randomly fails half the time. =/
    – BadZen
    Mar 17, 2021 at 21:21
30

CMake is more like a script language if comparing it with other ways to create Makefile (e.g. make or qmake). It is not very cool like Python, but still.

There are no such thing like a "proper way" if looking in various opensource projects how people include directories. But there are two ways to do it.

  1. Crude include_directories will append a directory to the current project and all other descendant projects which you will append via a series of add_subdirectory commands. Sometimes people say that such approach is legacy.

  2. A more elegant way is with target_include_directories. It allows to append a directory for a specific project/target without (maybe) unnecessary inheritance or clashing of various include directories. Also allow to perform even a subtle configuration and append one of the following markers for this command.

PRIVATE - use only for this specified build target

PUBLIC - use it for specified target and for targets which links with this project

INTERFACE -- use it only for targets which links with the current project

PS:

  1. Both commands allow to mark a directory as SYSTEM to give a hint that it is not your business that specified directories will contain warnings.

  2. A similar answer is with other pairs of commands target_compile_definitions/add_definitions, target_compile_options/CMAKE_C_FLAGS

1
  • 4
    "will append a directory to the current project and all other descendant projects which you will append via a series of add_subdirectory" -- Unfortunately, that is wrong. The truth is that it applies to all targets in the same (even those that came before the include_directories call) and to targets in add_subdirectories that came after the call (not before)... and this is why we say it's legacy. DO NOT USE include_directories, EVER Aug 21, 2021 at 4:55
10

I had the same problem.

My project directory was like this:

    --project
    ---Classes
    ----Application
    -----.h and .c files
    ----OtherFolders
    --main.cpp

And what I used to include the files in all those folders:

    file(GLOB source_files CONFIGURE_DEPENDS
            "*.h"
            "*.cpp"
            "Classes/*/*.cpp"
            "Classes/*/*.h"
    )

    add_executable(Server ${source_files})

And it totally worked.

3
  • 4
    Remembering that cmake is a 'build system generator' and not a 'build system' using file glob is not a good idea in modern cmake (CMake with versions 3.0 and above) because file globs are evaluated at 'build' time and not 'build system generation' time. See link : gist.github.com/mbinna/c61dbb39bca0e4fb7d1f73b0d66a4fd1
    – ggulgulia
    Mar 21, 2020 at 10:34
  • 1
    Just add CONFIGURE_DEPENDS.
    – letmaik
    Jul 6, 2020 at 10:33
  • 1
    In CMake the use of file GLOB should be avoided every time you can, when you use it in a large project you will have a boilerplate of includes at the end. Nov 1, 2022 at 10:04
5

You have two options.

The Old:

include_directories(${PATH_TO_DIRECTORY})

and the new

target_include_directories(executable-name PRIVATE ${PATH_TO_DIRECTORY})

To use target_include_directories, You need to have your executable defined - add_executable(executable-name sourcefiles).

So your code should appear like

add_executable(executable-name sourcefiles)
target_include_directories(executable-name PRIVATE ${PATH_TO_DIRECTORY})

You can read more here https://cmake.org/cmake/help/latest/command/target_include_directories.html

1
  • What case do you suggest the scope keyword, "PRIVATE"?
    – Cloud Cho
    Jan 19 at 22:10
1

This worked for me:

set(SOURCE main.cpp)
add_executable(${PROJECT_NAME} ${SOURCE})

# target_include_directories must be added AFTER add_executable
target_include_directories(${PROJECT_NAME} PUBLIC ${INTERNAL_INCLUDES})
1

Don't forget to include ${CMAKE_CURRENT_LIST_DIR}. That's what was causing problems for me.

Example should be like this:

target_include_directories(projectname
    PUBLIC "${CMAKE_CURRENT_LIST_DIR}/include"                          
)

PUBLIC for dependencies which you want to be included by a parent project. PRIVATE for ones that you don't.

1

In all the answers to this question, there is a whole lot of "how" (to get what you want), and precious little "why" (digging into the problem that motivated the question and what the asker may have misunderstood about the ways in which different types of tools like IDEs and build tools do / do not interact and share information with each other, and what information CMake passes / needs to pass to those tools).

This question is vexing, as it is motivated by a specific behaviour of a specific IDE- Code::Blocks) and CMake, but then poses a question unrelated to that IDE and instead about Makefiles and CMake, assuming that they have done something wrong with CMake which led to a problem with Makefiles, which led to a problem with their IDE.

TL;DR CMake and Makefiles have their own way of tracking header dependencies given include directories and source files. How CMake configures the Code::Blocks IDE is a completely separate story.

What is an "external" header in CMake? (no such thing)

I realized recently that the issue seemed to be that CMake considered those header files to be external to the project. [...] It therefore seems to me that CMake consider those headers to be external to the project, and does not track them in the depends

As far as I know, there is no official or useful definition of "external header" when it comes to CMake. I have not seen that phrase used in documentation. Also note that the word "project" is a quite overloaded term. Each buildsystem generated by CMake consists of one top-level project, possibly including other external or subdirectory projects. Each project can contain multiple targets (libraries, executables, etc.). What CMake refers to as a target sometimes translates to what IDEs call projects (Ix. Visual Studio, and possibly Code::Blocks). If you had to given such a phrase a meaning, here's what would make sense to me:

In the case that the question is referring to some IDEs' sense of the word "project", which CMake calls "targets", header files are external to a project would be those that aren't intended to be accessed through any of the include directories of a target (Ex. Include directories that come from targets linked to the target in question).

In the case that the question is referring to CMake's sense of the word "project": Targets are either part of a project (defined/created by a call to the project() command, and built by the generated buildsystem), or IMPORTED, (not built by the generated buildsystem and expected to already exist, or built by some custom step added to the generated buildsystem, such as via ExternalProject_Add). Include directories of IMPORTED targets would be those headers which are external to the CMake project in question, and include directories of non-IMPORTED targets would be those that are "part of" the project.

Does CMake track header dependencies? (It depends!)

[...] CMake consider those headers to be external to the project, and does not track them in the depends

I'm not super familiar with the history of CMake, or with header dependency tracking in build tooling, but here is what I've gathered from the searching I have done on the topic.

CMake itself doesn't have much to do with any information related to header/include dependencies of implmentation files / translation units. The only way in which that information is important to CMake is if CMake needs to be the one to tell the generated buildsystem what those dependencies are. It's the generated buildsystem which wants to track changes in header file dependencies to avoid any unnecessary recompilation. For the Unix Makefiles generator in particular, before CMake 3.20, CMake would do the job of scanning header/include dependencies to tell the Makefiles buildsystem about those dependencies. Since v3.20, where supported by the compiler, CMake delegates that responsibility to the compiler by default. See the option which can be used to revert that behaviour here.

The exact details of how header/include dependency scanning differs for each supported CMake generator. For example, you can find some high-level description about the Ninja capabilities/approach on their manual. Since this question is only about Makefiles, I won't attempt to go into detail about other generators.

Notice how to get the header/include dependency information for the buildsystem, you only need to give CMake a list of a target's include directories, and a list of the implementation source files to compile? You don't need to give it a list of header files because that information can be scanned for (either by CMake or by a compiler).

Do IDEs get information about target headers by scanning? (they could)

Each IDE can display information in whatever way it wants. Problems like you are having with the IDE not showing headers usually only happen for IDE display formats of the project layout other than the filesystem layout (project headers files are usually in the same project directory as implementation files). For example, such non-filesystem layout views are available in Visual Studio and Code::Blocks.

Each IDE can get header information in whatever way it chooses. As far as I am aware (but I may be wrong for Visual Studio), both Visual Studio and Code::Blocks expect the list of project headers to be explicitly listed in the IDE project configuration files. There are other possible approaches (Ex. header dependency scanning), but it seems that many IDEs choose the explicit list approach. My guess would be because it is simple implementation-wise.

Why would scanning be burdensome for an IDE to find header files associated with a target?(Note: this is somewhat speculation, since I am not a maintainer of any such tools and have only used a couple of them) An IDE could implement the file scanning (which itself is a complicated task), but to know which headers are "in" the target, they'd either need to get information from the buildsystem about how the translation units of the target will get compiled, and that's assuming that all "not-in-target" header include paths are specified with a "system"-like flag, which doesn't have to be the case. Or, it could try to get that information from the meta-buildsystem, which here is CMake. Or it could try to do what CMake now does and try to invoke the selected compiler to scan dependencies. But in either case, they'd have to make some difficult decision about which buildsystems, meta buildsystems, and/or compilers to support, and then do the difficult work of extracting that information from whatever formats those tools store that information in, possibly without any guarantees that those formats will be the same in future tool versions (supporting a change in the format in a newer tool version could be similar to having to supporting a completely separate tool). The IDE could do all that work, or it could just ask you to give it a list of the headers belonging to each target. As you can see, there are cons to the diversity in tooling that the C/C++ ecosystem has. There are pros too, but that's outside the scope of this question.

On the bright side, CMake actually does have a mechanism to try to take some of that work off your shoulders. For such IDEs that have non-filesystem-views, it does implement a simple-heuristic to try to find header files that are associated with source files...

How does header file discovery work for the Code::Blocks IDE generator for CMake?

At least, when generating a Code::Blocks project the header files do not appear within the project (the source files do).

Here's something interesting: The Code::Blocks editor has the concept of source files and header files that are part of a project, and since CMake doesn't expect/require its users to tell it about each and every header file in the project (it only needs to know about what include directories should be associated with targets), it tries to use a certain heuristic to discover header files that are associated to implementation files. That heuristic is very basic: take the path of each source file in a project, and try changing the extenstion to be like one that is usually given to header files, and see if any such file exists. See the cmExtraCodeBlocksGenerator::CreateNewProjectFile member function in :/Source/cmExtraCodeBlocksGenerator.cxx.

In "Pitchfork Layout" terminology, it would be said that the heuristic assumes that the project uses "merged-header" placement instead of "split-header" placement, where there are separate src/ and include/ directories. So if you don't use merged-header layout, or otherwise have any target headers that don't meet that heuristic, such as utility header files, you'll need to explicitly tell CMake about those files (Ex. using target_sources) for it to pass that knowledge on to the IDE config it generates.

Further readings:

Closing Words

I'm certain there are many people who know these tools better than I do. If you are one of those people and notice that I have made a mistake, please graciously correct me in the comments or in chat, or just to edit this post.

Note that while installation of build artifacts is an important part of many projects' lifecycles and is therefore incorporated into the designs of most C/C++ buildsystems, since the question didn't explicitly ask about the configuring the installation part, I have chosen to leave it out of this answer, since it in itself is not a trivial topic to cover (just see how long the related chapters in the "Mastering CMake" book are: The chapter on installation, and the chapter on importing and exporting).

0

In newer CMake versions we can limit our include-paths to target, like:

target_include_directories(MyApp PRIVATE "${CMAKE_CURRENT_LIST_DIR}/myFolder")

I mean, if the CMakeLists.txt has multiple targets, else, the include-paths are NOT shared with other CMakeLists.txt scripts, and it's enough to do something like:

include_directories("${CMAKE_CURRENT_LIST_DIR}/myFolder")

However, maybe we can simulate what target_include_directories(...) does for CMake 2.8.10 or older versions, like:

set_property(
    TARGET MyApp
    APPEND PROPERTY
        INCLUDE_DIRECTORIES "${CMAKE_CURRENT_LIST_DIR}/myFolder"
)

All done, but seems if you want source-files to be re-compiled once any header-file they use is changed, all such header-files need to be added to each target as well, like:

set(SOURCES src/main.cpp)

set(HEADERS
    ${CMAKE_CURRENT_LIST_DIR}/myFolder/myHeaderFile.h
    ${CMAKE_CURRENT_LIST_DIR}/myFolder/myOtherHeader.h
)

add_executable(MyApp ${SOURCES} ${HEADERS})

Where with "seems" I mean that, CMake could detect such header-files automatically if it wanted, because it parses project's C/C++ files anyway.

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.