Effective Modern CMake

effective_modern_cmake.md

Effective Modern CMake

Getting Started

For a brief user-level introduction to CMake, watch C++ Weekly, Episode 78, Intro to CMake by Jason Turner. LLVM’s CMake Primer provides a good high-level introduction to the CMake syntax. Go read it now.

After that, watch Mathieu Ropert’s CppCon 2017 talk Using Modern CMake Patterns to Enforce a Good Modular Design (slides). It provides a thorough explanation of what modern CMake is and why it is so much better than “old school” CMake. The modular design ideas in this talk are based on the book Large-Scale C++ Software Design by John Lakos. The next video that goes more into the details of modern CMake is Daniel Pfeifer’s C++Now 2017 talk Effective CMake (slides).

This text is heavily influenced by Mathieu Ropert’s and Daniel Pfeifer’s talks.

If you are interested in the history and internal architecture of CMake, have a look at the article CMake in the book The Architecture of Open Source Applications.

General

Use at least CMake version 3.0.0.

Modern CMake is only available starting with version 3.0.0.

Treat CMake code like production code.

CMake is code. Therefore, it should be clean. Use the same principles for CMakeLists.txt and modules as for the rest of the codebase.

Define project properties globally.

For example, a project might use a common set of compiler warnings. Defining such properties globally in the top-level CMakeLists.txt file prevents scenarios where public headers of a dependent target causing a depending target not to compile because the depending target uses stricter compiler options. Defining such project properties globally makes it easier to manage the project with all its targets.

Forget the commands add_compile_options, include_directories, link_directories, link_libraries.

Those commands operate on the directory level. All targets defined on that level inherit those properties. This increases the chance of hidden dependencies. Better operate on the targets directly.

Get your hands off CMAKE_CXX_FLAGS.

Different compilers use different command-line parameter formats. Setting the C++ standard via -std=c++14 in CMAKE_CXX_FLAGS will brake in the future, because those requirements are also fulfilled in other standards like C++17 and the compiler option is not the same on old compilers. So it’s much better to tell CMake the compile features so that it can figure out the appropriate compiler option to use.

Don’t abuse usage requirements.

As an example, don’t add -Wall to the PUBLIC or INTERFACE section of target_compile_options, since it is not required to build depending targets.

Modules

Use modern find modules that declare exported targets.

Starting with CMake 3.4, more and more find modules export targets that can be used via target_link_libraries.

Use exported targets of external packages.

Don’t fall back to the old CMake style of using variables defined by external packages. Use the exported targets via target_link_libraries instead.

Use a find module for third-party libraries that do not support clients to use CMake.

CMake provides a collection of find modules for third-party libraries. For example, Boost doesn't support CMake. Instead, CMake provides a find module to use Boost in CMake.

Report it as a bug to third-party library authors if a library does not support clients to use CMake. If the library is an open-source project, consider sending a patch.

CMake dominates the industry. It’s a problem if a library author does not support CMake.

Write a find module for third-party libraries that do not support clients to use CMake.

It’s possible to retrofit a find module that properly exports targets to an external package that does not support CMake.

Export your library’s interface, if you are a library author.

See Daniel Pfeifer’s C++Now 2017 talk Effective CMake (slide 24ff.) on how to do this. Keep in mind to export the right information. Use BUILD_INTERFACE and INSTALL_INTERFACE generator expressions as filters.

Projects

Avoid custom variables in the arguments of project commands.

Keep things simple. Don't introduce unnecessary custom variables. Instead of add_library(a ${MY_HEADERS} ${MY_SOURCES}), do add_library(a b.h b.cpp).

Don't use file(GLOB) in projects.

CMake is a build system generator, not a build system. It evaluates the GLOB expression to a list of files when generating the build system. The build system then operates on this list of files. Therefore, the build system cannot detect that something changed in the file system.

CMake cannot just forward the GLOB expression to the build system, so that the expression is evaluated when building. CMake wants to be the common denominator of the supported build systems. Not all build systems support this, so CMake cannot support it neither.

Put CI-specific settings in CTest scripts, not in the project.

It just makes things simpler. See Dashboard Client via CTest Script for more information.

Follow a naming convention for test names.

This simplifies filtering by regex when running tests via CTest.

Targets and Properties

Think in terms of targets and properties.

By defining properties (i.e., compile definitions, compile options, compile features, include directories, and library dependencies) in terms of targets, it helps the developer to reason about the system at the target level. The developer does not need to understand the whole system in order to reason about a single target. The build system handles transitivity.

Imagine targets as objects.

Calling the member functions modifies the member variables of the object.

Analogy to constructors:

• add_executable
• add_library

Analogy to member variables:

• target properties (too many to list here)

Analogy to member functions:

• target_compile_definitions
• target_compile_features
• target_compile_options
• target_include_directories
• target_link_libraries
• target_sources
• get_target_property
• set_target_property

Keep internal properties PRIVATE.

If a target needs properties internally (i.e., compile definitions, compile options, compile features, include directories, and library dependencies), add them to the PRIVATE section of the target_* commands.

Declare compile definitions with target_compile_definitions.

This associates the compile definitions with their visibility (PRIVATE, PUBLIC, INTERFACE) to the target. This is better than using add_compile_definitions, which has no association with a target.

Declare compile options with target_compile_options.

This associates the compile options with their visibility (PRIVATE, PUBLIC, INTERFACE) to the target. This is better than using add_compile_options, which has no association with a target. But be careful not to declare compile options that affect the ABI. Declare those options globally. See “Don’t use target_compile_options to set options that affect the ABI.”

Declare compile features with target_compile_features.

t.b.d.

Declare include directories with target_include_directories.

This associates the include directories with their visibility (PRIVATE, PUBLIC, INTERFACE) to the target. This is better than using include_directories, which has no association with a target.

Declare direct dependencies with target_link_libraries.

This propagates usage requirements from the dependent target to the depending target. The command also resolves transitive dependencies.

Don’t use target_include_directories with a path outside the component’s directory.

Using a path outside a component’s directory is a hidden dependency. Instead, use target_include_directories to propagate include directories as usage requirements to depending targets via target_link_directories.

Always explicitly declare properties PUBLIC, PRIVATE, or INTERFACE when using target_*.

Being explicit reduces the chance to unintendedly introduce hidden dependencies.

Don’t use target_compile_options to set options that affect the ABI.

Using different compile options for multiple targets may affect ABI compatibility. The simplest solution to prevent such problems is to define compile options globally (also see “Define project properties globally.”).

Using a library defined in the same CMake tree should look the same as using an external library.

Packages defined in the same CMake tree are directly accessible. Make prebuilt libraries available via CMAKE_PREFIX_PATH. Finding a package with find_package should be a no-op if the package is defined in the same build tree. When you export target Bar into namespace Foo, also create an alias Foo::Bar via add_library(Foo::Bar ALIAS Bar). Create a variable that lists all sub-projects. Define the macro find_package to wrap the original find_package command (now accessible via _find_package). The macro inhibits calls to _find_package if the variable contains the name of the package. See Daniel Pfeifer’s C++Now 2017 talk Effective CMake (slide 31ff.) for more information.

Functions and Macros

Prefer functions over macros whenever reasonable.

In addition to directory-based scope, CMake functions have their own scope. This means variables set inside functions are not visible in the parent scope. This is not true of macros.

Use macros for defining very small bits of functionality only or to wrap commands that have output parameters. Otherwise create a function.

Functions have their own scope, macros don’t. This means variables set in macros will be visible in the calling scope.

Arguments to macros are not set as variables, instead dereferences to the parameters are resolved across the macro before executing it. This can result in unexpected behavior when using unreferenced variables. Generally speaking this issue is uncommon because it requires using non-dereferenced variables with names that overlap in the parent scope, but it is important to be aware of because it can lead to subtle bugs.

Don’t use macros that affect all targets in a directory tree, like include_directories, add_definitions, or link_libraries.

Those macros are evil. If used on the top level, all targets can use the properties defined by them. For example, all targets can use (i.e., #include) the headers defined by include_directories. If a target does not require linking (e.g., interface library, inline template), you won’t even get a compiler error in this case. It is easy to accidentally create hidden dependencies through other targets with those macros.

Arguments

Use cmake_parse_arguments as the recommended way to handle complex argument-based behaviors or optional arguments in any function.

Don’t reinvent the wheel.

Loops

Use modern foreach syntax.

• foreach(var IN ITEMS foo bar baz) ...
• foreach(var IN LISTS my_list) ...
• foreach(var IN LISTS my_list ITEMS foo bar baz) ...

Packages

Use CPack to create packages.

CPack is part of CMake and nicely integrates with it.

Write a CPackConfig.cmake that includes the one generated by CMake.

This makes it possible to set additional variables that don’t need to appear in the project.

Cross Compiling

Use toolchain files for cross compiling.

Toolchain files encapsulate toolchains for cross compilation.

Keep toolchain files simple.

It’s easier to understand and simpler to use. Don’t put logic in toolchain files. Create a single toolchain file per platform.

Warnings and Errors

Treat build errors correctly.

• Fix them.
• Reject pull requests.
• Hold off releases.

Treat warnings as errors.

To treat warnings as errors, never pass -Werror to the compiler. If you do, the compiler treats warnings as errors. You can no longer treat warnings as errors, because you no longer get any warnings. All you get is errors.

• You cannot enable -Werror unless you already reached zero warnings.
• You cannot increase the warning level unless you already fixed all warnings introduced by that level.
• You cannot upgrade your compiler unless you already fixed all new warnings that the compiler reports at your warning level.
• You cannot update your dependencies unless you already ported your code away from any symbols that are now [[deprecated]].
• You cannot [[deprecated]] your internal code as long as it is still used. But once it is no longer used, you can as well just remove it.

Treat new warnings as errors.

1. At the beginning of a development cycle (e.g., sprint), allow new warnings to be introduced.
   • Increase warning level, enable new warnings explicitly.
   • Update the compiler.
   • Update dependencies.
   • Mark symbols as [[deprecated]].
2. Burn down the number of warnings.
3. Repeat.

Static Analysis

Use more than one supported analyzer.

Using clang-tidy (<lang>_CLANG_TIDY), cpplint (<lang>_CPPLINT), include-what-you-use (<lang>_INCLUDE_WHAT_YOU_USE), and LINK_WHAT_YOU_USE help you find issues in the code. The diagnostics output of those tools will appear in the build output as well as in the IDE.

For each header file, there must be an associated source file that #includes the header file at the top, even if that source file would otherwise be empty.

Most of the analysis tools report diagnostics for the current source file plus the associated header. Header files with no associated source file will not be analyzed. You may be able to set a custom header filter, but then the headers may be analyzed multiple times.

Sources

• Intro to CMake by Jason Turner at C++ Weekly (Episode 78)
• LLVM CMake Primer
• Using Modern CMake Patterns to Enforce a Good Modular Design (slides) by Mathieu Ropert at CppCon 2017
• Effective CMake (slides) by Daniel Pfeifer at C++Now 2017
• The Architecture of Open Source Applications: CMake

@carlosgalvezp

Hi,

Great post about modern CMake! I'm having a bit of trouble understanding the following rules, they look contradicting to me:

-Define project properties globally.
-Forget the commands add_compile_options

How are we supposed to set properties globally if not via add_compile_options? Should we create an INTERFACE target that all other targets must link against? This forces the developers to remember to link against the "global options target" every time they create a new target. It would be better that this is enforced automatically.

Thanks!

I would also be interested in how to follow both of these rules. Currently, we set the following in our toolchain files:

SET (
  COMPILER_WARNING_SWITCHES
  "-Wall -Wextra -Wwrite-strings -Wunreachable-code -Wpointer-arith -Winit-self -Wredundant-decls -Wpedantic -fdiagnostics-color=always -Wno-error=deprecated-declarations -pedantic-errors"
)
SET (
  CMAKE_C_FLAGS
  "${COMPILER_WARNING_SWITCHES}"
  CACHE STRING "gcc c flags" FORCE)
SET (
  CMAKE_CXX_FLAGS
  "${COMPILER_WARNING_SWITCHES}"
  CACHE STRING "gcc cxx flags" FORCE)

Fixed some typos https://gist.github.com/serpent7776/ab6728bab406ea4ece0eb237c3c5a553

I have a question regarding custom variable in project commands.
I want to install header for a library that I need to import later using find_packages. Header are already defined in target_sources PUBLIC section.

target_sources(toto_shared
    PRIVATE
        toto_shared.cpp
    PUBLIC
        "$<BUILD_INTERFACE:${CMAKE_CURRENT_LIST_DIR}/toto_shared.h>"
    )

but does that mean I need to duplicate the list of header in the INSTALL/FILES to make sure they are effectively copied?

install(FILES ${CMAKE_CURRENT_LIST_DIR}/subdir/toto_shared.h DESTINATION include)

I would also be interested in how to follow both of these rules.

Same. It feels cumbersome to remember to link some INTERFACE "library" for all sub-directory libraries your application has.

s/will brake in the future/will break in the future/

Thanks for some valuable information. However, re, "Getting Started," please don't refer me to an hour long video; I'd rather spend ten minutes reading than watching for an hour. Reading is FAR MORE information dense than videos.

Hey guys, I need some explanation on the following, I am doing the following to set CUDA properties for a INTERFACE library;

set_target_properties(foo
    PROPERTIES
        CUDA_ARCHITECTURES 61
)

And later, I want the targets that link to this library to automatically inherit the CUDA_ARCHITECTURES, but they seem to default to an initial value of 52 in my case. How can I automatically inherit the properties in the core library in my targets later on?

I have managed to do the following, but it doesn't look like an elegant solution as I will have to do that for every single property and target;

get_target_property(FOO_CUDA_ARCHITECTURES foo CUDA_ARCHITECTURES)
set_target_properties(bar 
    PROPERTIES 
        CUDA_ARCHITECTURES ${FOO_CUDA_ARCHITECTURES}
)

Flags are automatically inherited, but I don't see properties set in the way above to function the same way.

@neoblizz Target properties are specific and their PUBLIC, PRIVATE, INTERFACE keyword is already encoded in their name. For example: INCLUDE_DIRECTORIES and INTERFACE_INCLUDE_DIRECTORIES - target_include_directories(<target> <keyword> dirs...) sets the first one if the keyword is PRIVATE, sets the second if the keyword is INTERFACE and sets both if the keyword is PUBLIC.

Unfortunately, I could not find any property named similarly to CUDA_ARCHITECTURES. Apparently it is always interface or always private. The documentation only mentions that the default is affected by CMAKE_CUDA_ARCHITECTURES variable when the target is created.

Unfortunately, I could not find any property named similarly to CUDA_ARCHITECTURES. Apparently it is always interface or always private. The documentation only mentions that the default is affected by CMAKE_CUDA_ARCHITECTURES variable when the target is created.

Thank you! That answer helps a lot, I found out that it is not in the list of properties that get inherited, yet. But, maybe with future updates, it will be in one.

Define the macro find_package to wrap the original find_package command (now accessible via _find_package).

This was recently clarified to be bad practice. I can't find the source. Something about this being an implementation detail. I think the author himself clarified it.

Don’t use target_include_directories with a path outside the component’s directory.
Using a path outside a component’s directory is a hidden dependency. Instead, use target_include_directories to propagate include directories as usage requirements to depending targets via target_link_directories.

I just.. don't understand this alternative written in italics :( Is there a more explicit rephrasing available?

For anyone interested in how to definitively do CMake right, you are welcome to take https://github.com/friendlyanon/cmake-init for a spin.

---

Define the macro find_package to wrap the original find_package command (now accessible via _find_package).

This was recently clarified to be bad practice. I can't find the source. Something about this being an implementation detail. I think the author himself clarified it.

Not only is it bad practice, but it's an undocumented debug feature left in CMake. Doing this is effectively UB. More details here https://crascit.com/2018/09/14/do-not-redefine-cmake-commands/

Thank you for all your comments. 🎉 When I created this Gist, a good reference book that teaches modern CMake didn't exist (at least not to my knowledge). Nowadays, I'd recommend the book Professional CMake: A Practical Guide by @craigscott-crascit. It's comprehensive and the quality is outstanding.

Don’t use target_include_directories with a path outside the component’s directory.
Using a path outside a component’s directory is a hidden dependency. Instead, use target_include_directories to propagate include directories as usage requirements to depending targets via target_link_directories.

I just.. don't understand this alternative written in italics :( Is there a more explicit rephrasing available?

Not quite understand the reasoning behind this. Can anyone explain it a little bit?

It's unexpected to have a directory include a dependency from above it, directly. Make an interface target at the correct level to not need .. in the include path, then add that in the list for target_link_libraries.

Basically, if you take just a subdirectory of a repo, you should be able to easily see based on its cmakelists.txt the things you need from elsewhere.

It's kind of an opinion thing but probably makes it more readable.

Don’t use target_include_directories with a path outside the component’s directory.
Using a path outside a component’s directory is a hidden dependency. Instead, use target_include_directories to propagate include directories as usage requirements to depending targets via target_link_directories.

I just.. don't understand this alternative written in italics :( Is there a more explicit rephrasing available?

I think this is a mistake. The last target_link_directories should be target_link_libraries.

From what I understand, if you need to have an include path outside of the component's directory, there should be a separate target that declares that path as an interface include directory. Then the depending target should "link" to that new target to get its usage requirement (i.e., the include path).

已收到您的邮件，我会尽快回复。