MastringCmakeVersion31:Chapter 04

From 흡혈양파의 번역工房
Jump to navigation Jump to search

CHAPTER FOUR : WRITING CMAKELISTS FILES

This chapter will cover the basics of writing effective CMakeLists files for your software. It will cover the basic commands and issues you will need to handle most projects. It will also discuss how to convert existing UNIX or Windows projects into CMakeLists files. While CMake can handle extremely complex projects, for most projects you will find this chapter's contents will tell you all you need to know. CMake is driven by the CMakeLists.txt files written for a software project. The CMakeLists files determine everything from which options to present to users, to which source files to compile. In addition to discussing how to write a CMakeLists file, this chapter will also cover how to make them robust and maintainable. The basic syntax of a CMakeLists.txt file and key concepts of CMake have already been discussed in chapters 2 and 3. This chapter will expand on those concepts and introduce a few new ones.

CMake Language

As discussed in Chapter 2, CMakeLists files follow a simple syntax consisting of comments, commands, and whitespace. A comment is indicated using the # character and runs from that character until the end of the line. A command consists of the command name, opening parenthesis, whitespace-separated arguments, and a closing parenthesis. All whitespace (spaces, line feeds, tabs) is ignored except to separate arguments. Anything within a set of double quotes is treated as one argument, as is typical for most languages. The backslash can be used to escape characters, preventing the normal interpretation of them. The subsequent examples in this chapter will help to clear up some of these syntactic issues. You might wonder why CMake decided to have its own language instead of using an existing one such as Python, Java, or Tcl. The main reason is that we did not want to make CMake require an additional tool to run. By requiring one of these other languages, all users of CMake would be required to have that language installed, and potentially a specific version of that language. This is on top of the language extensions that would be required to do some of the CMake work, for both performance and capability reasons.

Basic Commands

While the previous chapters have already introduced many of the basic commands for CMakeLists files, this chapter will review and expand on them. The top-level CMakeLists file should call the PROJECT() (page 327) command. This command both names the project and optionally specifies which languages will be used by it:

project (projectname [C] [CXX] [Fortran] [NONE])

If no languages are specified then CMake defaults to supporting C and C++. If the NONE language is passed then CMake does not include language-specific support.

For each directory in a project where the CMakeLists.txt file invokes the project command, CMake generates a top-level IDE project file. The project will contain all targets that are in the CMakeLists.txt file and any subdirectories, as specified by the add_s ubdi rectory() (page 277) command. If the EXCLUDE_FROM_ALL (page 569) option is used in the add_subdirectory command, the generated project will not appear in the top-level Makefile or IDE project file; this is useful for generating sub-projects that do not make sense as part of the main build process. Consider that a project with a number of examples could use this feature to generate the build files for each example with one run of CMake, but not have the examples built as part of the normal build process.

The set and unset commands manipulate variables and entries in the persistent cache. The string() (page 335), list() (page 323), remove() (page 349), and separate_arguments() (page 329) com­mands offer basic manipulation of strings and lists.

The add_executable() (page 273) and add_library() (page 274) commands are the main com­mands for defining the libraries and executables to build, and which source files comprise them. For Visual Studio projects, the source files will show up in the IDE as usual, but any header files the project uses will not be. To have the header files show up, simply add them to the list of source files for the executable or library; this can be done for all generators. Any generators that do not use the header files directly (such as Makefile based generators) will simply ignore them.


Flow Control

The CMake language provides three flow control constructs:

  • Conditional statements (e.g. if() (page 313))
  • Looping constructs (e.g. foreach() (page 309) and while() (page 345))
  • Procedure definitions (e.g. macro() (page 324) and function() (page 309))

First we will consider the if command. In many ways, the if command in CMake is just like the if command in any other language. It evaluates its expression and uses it to execute the code in its body or optionally the code in the else() (page 284) clause. For example:

if (FOO)
  # do something here
else ()
  # do something here
endif ()

The condition in the if statement may optionally be repeated in the else and endif() (page 285) clauses:

if (FOO)
  # do something here
else (FOO)
  # do something here
endif (FOO)

In this book, you will see examples of both styles. When you include conditionals in the else and endif clause then they must exactly match the original conditional of the if statement. The following code would not work:

set (FOO 1)
if (${FOO})
  # do something
endif (1)
# ERROR, it dosen't match the original if conditional

CMake provides verbose error messages in cases where an if statement is not properly matched with an endif.

CMake also supports elseif() (page 284) to help sequentially test for multiple conditions. For example:

if (MSVC80)
  # do something here
elseif (MSVC90)
  # do something here
elseif (APPLE)
  # do something here
endif ()

The if command documents the many conditions it can test. Some of the more common conditions include:

if (constant) True if the constant is 1, ON, YES, TRUE, Y, or a non-zero number. False if the constant is 0, OFF, NO, FALSE, N, IGNORE; is an empty string, or ends in the suffix "-NOTFOUND." Named boolean constants are case-insensitive. If the argument is not one of these constants then it is treated as a variable. if (variable) True if the variable is defined to a value that is not a false constant. if (NOT <expression>) True if the expression is not true. if (<expr1> AND <expr2>) True if both expressions would be considered true individually. if (<expr1> OR <expr2>) True if either expression would be considered true individually. if (DEFINED variable) True if the given variable has been set, regardless of what value it was set to. if (<variablelstring> MATCHES regex) True if the given string or variable's value matches the regular given expression.

Additional binary test operators include EQUAL, LESS, and GREATER for numeric comparisons; STRLESS, STREQUAL, and STRGREATER for lexicographic comparisons; and VERSION_LESS, VERSION_EQUAL, and VERSION_GREATER to compare versions of the form major[.minor[.patch[.tweak]]].

The OR test has the lowest precedence, followed by AND, then NOT, and then any other test. Tests of the same precedence are performed from left-to-right. Expressions may be enclosed in parentheses to adjust precedence. For example, consider the following conditionals

if ((1 LESS 2) AND (3 LESS 4))
  message ("sequence of numbers")
endif ()

if (1 AND 3 AND 4)
  message ("series of true values")
endif (1 AND 3 AND 4)

if (NOT 0 AND 3 AND 4)
  message ("a false value")
endif (NOT 0 AND 3 AND 4)

if (0 OR 3 AND 4)
  message ("or statements")
endif (0 OR 3 AND 4)

if (EXISTS ${PROJECT_SOURCE_DIR}/Help.txt AND COMMAND IF)
  message ("Help exists")
endif (EXISTS ${PROJECT_SOURCE_DIR}/Help.txt AND COMMAND IF)

set (fooba 0)

if (NOT DEFINED foobar)
  message ("foobar is not defined")
endif (NOT DEFINED foobar)

if (NOT DEFINED fooba)
  message ("fooba not defined")
endif (NOT DEFINED fooba)

if (NOT 0 AND 0)
  message ("This line is never executed")
endif (NOT 0 AND 0}

if (NOT (0 AND 0))
  message ("This line is always executed")
endif (NOT (0 AND 0))

Now let us consider the other flow control commands. The foreach, while, macro, and function commands are the best way to reduce the size of your CMakeLists files and keep them maintainable. The foreach() (page 309) command enables you to execute a group of CMake commands repeatedly on the members of a list. Consider the following example adapted from VTK

foreach (tfile
        TestAnisotropicDiffusion2D
        TestButterworthLowPass
        TestButterworthHighPass
        TestCityBlockDistance
        TestConvolve
        )
  add_test (${tfile)-image $ (VTK_EXECUTABLE}
    ${VTK_SOURCE_DIR}/Tests/rtimageTest.tcl
    ${VTK_SOURCE_DIR}/Tests/${tfile}.tcl
    -D ${VTK_DATA_ROOT}
    -V Baseline/Imaging/S{tfile}.png
    -A $(VTK_SOURCE_DIR}/Wrapping/Tcl
    )
endforeach ( tfile }

The first argument of the foreach command is the name of the variable that will take on a different value with each iteration of the loop; the remaining arguments are the list of values over which to loop. In this example, the body of the foreach loop is j ust one CMake command, add_test. In the body of the foreach loop, each time the loop variable (tfile in this example) is referenced will be re­placed with the current value from the list. In the first iteration, occurrences of $ {tfile} will be re­placed with TestAnisotropicDiffusion2D. In the next iteration, ${tfile} will be rep laced with TestButterworthLowPass. The foreach loop will continue to loop until all of the arguments have been processed.

It is worth mentioning that foreach loops can be nested, and that the loop variable is replaced prior to any other variable expansion. This means that in the body of a foreach loop, you can construct variable names using the loop variable. In the code below, the loop variable tfile is expanded, and then concatenated with _TEST_RESULT. The new variable name is then expanded and tested to see if it matches FAILD.

if ($(${tfile}} TEST_RESULT) MATCHES FAILED)
  message ("Test ${tfile} failed.")
endif ()

The while() (page 345) command provides looping based on a test condition. The format for the test expression in the while command is the same as it is for the if command, as described earlier. Consider the following example, which is used by CTest. Note that CTest updates the value of CTEST_ELAPSED_TIME internally.

#####################################################
# run paraview and ctest test dashboards for 6 hours
#
while (${CTEST_ELAPSED_TIME} LESS 36000)
  set (START_TIME ${CTEST_ELAPSED_TIME})
  ctest_run_script { "dash1_ParaView_vs7icontinuous.cmake" )
  ctest_run_script { "dash1_cmake_vs7lcontinuous.cmake" )
endwhile {)

The foreach and while commands allow you to handle repetitive tasks that occur in sequence, whereas the macro and function commands support repetitive tasks that may be scattered throughout your CMakeLists files. Once a macro or function is defined, it can be used by any CMakeLists fi les processed after its definition.

A function in CMake is very much like a function in C or C++. You can pass arguments into it, and they become variables within the function. Likewise, some standard variables such as ARGC, ARGV, ARGN, and ARGV0, ARGV1, etc. are defined. Function calls have a dynamic scope. Within a function you are in a new variable scope; this is like how you drop into a subdirectory using the add_subdirectory() (page 277) command and are in a new variable scope. All the variables that were defined when the function was called remain defined, but any changes to variables or new variables only exist within the function. When the function returns, those variables will go away. Put more simply: when you invoke a function, a new variable scope is pushed; when it returns, that variable scope is popped.

The function() (page 309) command defines a new function. The first argument is the name of the function to define; all additional arguments are formal parameters to the function.

function (DetermineTime_time)
  # pass the result up to whatever invoked this
  set ($(_time) "1;23:45" PARENT_SCOPE)
endfunction()

# now use the function we just defined
DetermineTime( current_time )}

if( DEFINED current_time )
  message(STATUS "The time is now: ${current_time}")
endif ()

Note that in this example, _time is used to pass the name of the return variable. The set() (page 330) command is invoked with the value of _time, which will be current_time. Finally, the set command uses the PARENT_SCOPE option to set the variable in the caller's scope instead of the local scope.

Macros are defined and called in the same manner as functions. The main differences are that a macro does not push and pop a new variable scope, and that the arguments to a macro are not treated as variables but as strings replaced prior to execution. This is very much like the differences between a macro and a function in C or C++. The first argument is the name of the macro to create; all additional arguments are formal parameters to the macro.

# define a simple macro
macro (assert TEST COMMENT)
  if (NOT ${TEST))
    message ("Assertion failed: ${COMMENT}")
  endif (NOT ${TEST})
endmacro (assert)

# use the macro
find_library (FOO_LIB foo /usr/local/lib)
assert ( ${FOO_LIB} "Unable to find library foo")

The simple example above creates a macro called assert. The macro is defined into two arguments; the first is a value to test and the second is a comment to print out if the test fails. The body of the macro is a simple if() (page 313) command with a message() (page 326) command inside of it. The macro body ends when the endmacro() (page 285) command is found. The macro can be invoked simply by using its name as if it were a command. In the above example, if FOO_LIB was not found then a message would be displayed indicating the error condition.

The macro command also supports defining macros that take variable argument lists. This can be useful if you want to define a macro that has optional arguments or multiple signatures. Variable arguments can be referenced using ARGC and ARGV0, ARGV1, etc., instead of the formal parameters. ARGV0 represents the first argument to the macro; ARGV1 represents the next, and so forth. You can also use a mixture of formal arguments and variable arguments, as shown in the example below.

# define a macro that takes at least two arguments
# (the formal arguments) plus an optional third argument
macro (assert TEST COMMENT)
  if (NOT ${TEST})
    message ("Assertion failed: ${COMMENT}")

    # if called with three arguments then also write the
    # message to a file specified as the third argument
    if (${ARGC) MATCHES 3)
      file (APPEND ${ARGV2} "Assertion failed: ${COMMENT)")
    endif (${ARGC} MATCHES 3)

  endif (NOT ${TEST})
endmacro {assertASSERT)

# use the macro
find_library (FOO_LIB foo /usr/local/lib)
assert ( $({FOO_LIB} "Unable to find library foo" )

In this example, the two required arguments are TEST and COMMENT. These required arguments can be referenced by name, as they are in this example, or by referencing ARGV0 and ARGV1. If you want to process the arguments as a list, use the ARGV and ARGN variables. ARGV (as opposed to ARGV0, ARGV1, etc) is a list of all the arguments to the macro, while ARGN is a list of all the arguments after the formal arguments. Inside your macro, you can use the foreach command to iterate over ARGV or ARGN as desired.

CMake has two commands for interrupting the processing flow. The break() (page 279) command breaks out of a foreach or while loop before it would normally end. The return() (page 328) command returns from a function or listfile before the function or listfile has reached its end.


Regular Expressions

A few CMake commands, such as if() (page 313) and string() (page 335), make use of regular ex­pressions or can take a regular expression as an argument. In its simplest form, a regular expression is a sequence of characters used to search for exact character matches. However, many times the exact sequence to be found is unknown, or only a match at the beginning or end of a string is desired. Since there are several different conventions for specifying regular expressions, CMake's standard is described below. The descrip­tion is based on the open source regular expression class from Texas Instruments, which is used by CMake for parsing regular expressions.

Regular expressions can be specified by using combinations of standard alphanumeric characters and the following regular expression meta-characters:


^ Matches at the beginning of a line or string. $ Matches at the end of a line or string. . Matches any single character other than a new line. [ ] Matches any character(s) inside the brackets . [^ ] Matches any character(s) not inside the brackets. [ - ] Matches any character in range on either side of a dash. * Matches the preceding pattern zero-or more-times. + Matches the preceding pattern one-or-more times. ? Matches the preceding pattern zero times or once only. () Saves a matched expression and uses it in a later replacement. ( | ) Matches either the left-or-right side of the bar.


Note that more than one of these meta-characters can be used in a single regular expression in order to create complex search patterns. For example, the pattern [^ab1-9] says to match any character sequence that does not begin with the characters "a" or "b" or numbers in the series one through nine. The following examples may help clarify regular expression usage:

  • The regular expression "^hello" matches a "hello" only at the beginning of a search string. It would match "hello there," but not "hi,nhello there."
  • The regular expression "long$" matches a "long" only at the end of a search string. It would match "so long," but not "long ago."
  • The regular expression "t..t..g" will match anything that has a "t" and any two characters, followed by another "t," and any two characters, and then a "g." It would match "testing" or "test again ," but would not match "toasting."
  • The regular expression "[1-9ab]" matches any number one-through-nine, and the characters "a" and "b". It would match "hello 1" or "begin", but would not match "no-match" .
  • The regular expression "[^1-9ab]" matches any character that is not a number one-through-nine, or an "a" or "b." It would NOT match "1ab2" or "b2345a," but would match "no-match."
  • The regular expression "br* " matches something that begins with a "b" and is followed by zero-or-more "r"s, and ends in a space. It would match "brrrrr" and "b," but would not match "brrh."
  • The regular expression "br+" matche s something that begins with a "b" and is followed by one or more "r"s, and ends in a space. It would match "brrrrr," and "br," but would not match "b " or "brrh."
  • The regular expression "br?" m atches something that begins with a "b," is followed by zero-or-one "r"s, and ends in a space. It would match "br,", and "b ," but would not match "brrrr" or "brrh."
  • The regular expression "(..p)b" matches something ending with pb and beginning with the two charac­ters before the first "p" encountered in the line. For example, it would find "repb" in "rep drepaqrepb." The regular expression "(..p)a" would find "repa qrepb" in "rep drepa qrepb."
  • The regular expression "d(_p)" matches something ending with "p," beginning with "d," and having two characters in-between that are the same as the two characters before the first "p" encountered in the line. It would m atch "drepa qrepb" in "rep drepa qrepb."


Checking Versions of CMake

CMake is an evolving program and as new versions are released, new features or commands are introduced. As a result, there may be instances where you might want to use a command that is in a current version of CMake but not in previous versions. There are a couple of ways to handle this; one option is to use the if() (page 313) command to check whether a new command exists. For example:

# test if the command exists
if (COMMAND some_new_command)
  # use the command
  some_new_command ( ARGS... )
endif ()

Alternatively, one may test against the actual version of CMake that is being run by evaluating the CMAKE_VERSION (page 634) variable:

# look for newer versions of CMake
if (${CMAKE_VERSION} VERSION_GREATER 2.6.3)
  # do something special here
endif ()

When writing your CMakeLists files, you may decide that you do not want to support old versions of CMake. To do this, place the following command at the top of your CMakeLists file

cmake_minimum_required (VERSION 2.6.3)

This indicates that the person running CMake must have at least version 2.6.3. If they are running an older version of CMake, an error message will be displayed telling them that the project requires at least the specified version of CMake.

Finally, some new releases of CMake might no longer support some behavior you were using (although we try to avoid this). In these cases, use CMake policies, as discussed in the cmake-policies(7) (page 537) manual.


Using Modules

Code reuse is a valuable technique in software development and CMake has been designed to support it. Allowing CMakeLists files to make use of reusable modules enables the entire community to share reusable sections of code. For CMake, these sections are called modules and can be found in the Modules subdirectory of your installation. Modules are simply sections of CMake commands put into a file; they can then be in­cluded into other CMakeLists files using the include() (page 317) command. For example, the following commands will include the CheckTypeSize module from CMake and then use the macro it defines.

include (CheckTypeSize)
check_type_size(long SIZEOF_LONG)

A module's location can be specified using the full path to the module file, or by letting CMake find the module by itself. CMake will look for modules in the directories specified by CMAKE_MODULE_PATH (page 646); if it cannot find it there, it will look in the Modules subdirectory. This way projects can override modules that CMake provides and customize them for their needs. Modules can be broken into a few main categories:

Find Modules These modules support the find_package() (page 297) command to determine the lo­cation of software elements, such as header files or libraries, that belong to a given package. Do not include them directly. Use the find_package() (page 297) command. Each module comes with documentation describing the package it finds and the variables in which it provides results. Conven­tions used in Find modules are covered in more detail in Chapter 5.

System Introspection Modules These modules test the system to provide information about the target plat­form or compiler, such as the size of a tloat or support for ANSI C++ streams. Many of these modules have names prefixed with Test or Check, such as TestBigEndian and CheckTypeSize. Some of them try to compile code in order to determine the correct result. In these cases, the source code is typically named the same as the module, but with a c or cxx extension. System introspection modules are covered in more detail in Chapter 12.

Utility Modules These modules provide useful macros and functions implemented in the CMake language and intended for specific, common use cases. See documentation of each module for details.

Using CMake with SWIG

One example of how modules can be used is to look at wrapping your C/C++ code in another language using Simplified Wrapper and Interface Generator (SWIG; www.swig.org). SWIG is a tool that reads anno­tated C/C++ header files and creates wrapper code (glue code) to make the corresponding C/C++ libraries available to other programming languages such as Tcl, Python, or Java. CMake supports SWIG with the find_package() (page 297) command. Although it can be used from CMake with custom commands, the SWIG package provides several macros that make building SWIG projects with CMake simpler. To use the SWIG macros, you must first call the find_package command with the name SWIG. Then, include the file referenced by the variable SWIG_USE_FILE. This will define several macros and set up CMake to easily build SWIG-based projects.

Two very useful macros are SWIG_ADD_MODULE and SWIG_LINK_LIBRARIES. SWIG_ADD_MODULE works much like the add_library() (page 274) command in CMake. The command is invoked like this:

SWIG_ADD_MODULE (module_name language source1 source2 ... sourceN)

The first argument is the name of the module being created. The next argument is the target language SWIG is producing a wrapper for. The rest of the arguments consist of a list of source files used to create the shared module. The big difference is that SWIG .i interface files can be used directly as sources. The macro will create the correct custom commands to run SWIG, and generate the C or C++ wrapper code from the SWIG interface files. The sources can also be regular C or C++ files that need to be compiled in with the wrappers.

The SWIG_LINK_LIBRARIES macro is used to link support libraries to the module. This macro is used because depending on the language being wrapped by SWIG, the name of the module may be different. The actual name of the module is stored in a variable called SWIG_MODULE_${name}_REAL_NAME where ${name} is the name passed into the SWIG_ADD_MODULE macro. For example, SWIG_ADD_MODULE(foo tcl foo.i) creates a variable called SWIG_MODULE_foo_REAL_NAME, which contains the name of the actual module created.

Now consider the following example that uses the example found in SWIG under Examples/python/class.

# Find SWIG and include the use swig file
find_package (SWIG REQUIRED)
include (${SWIG_USE_FILE})

# Find python library and add include path for python headers
find_package (PythonLibs)
include_directories (${PYTHON INCLUDE_PATH) )

# set the global swig flags to empty
set (CMAKE_SWIG_PLAGS "")

# let swig know that example.i is c++ and add the -includeall
# flag to swig
set_source_files_properties (example.i PROPERTIES CPLUSPLUS ON)
set_source_files_properties (example.i
                            PROPERTIES SWIG_FLAGS "-includeall")

# Create the swig module called example
# using the example.i source and example.cxx
# swig will be used to create wrap_example.cxx from example.i
SWIG_LADD_MODULE (example python example.i example.cxx)
SWIG_LINK_LIBRARIES (example ${PYTHON_LIBRARIES})

This example first uses find_package to locate SWIG, and includes the SWIG_USE_FILE defining the SWIG CMake macros. It then finds the Python libraries and sets up CMake to build with the Python library. Notice that the SWIG input file "example.i" is used like any other source file in CMake, and the properties are set on the file tell ing SWIG that the file is C++ and that the SWIG flag -includeall should be used when running SWIG on that source file. The module is created by telling SWIG the name of the module, the target language, and the list of source files. Finally, the Python libraries are linked to the module.

Using CMake with Qt

Projects using the popular widget toolkit Qt from Nokia (qt.nokia.com) can be built with CMake. CMake supports multiple versions of Qt, including versions 3 and 4. The first step is to tell CMake which version(s) of Qt to look for. Many Qt applications are designed to work with Qt3 or Qt4, but not both. If your application is designed for Qt4, use the FindQt4 module; for Qt3 , use the FindQt3 module. If your project can work with either version of Qt then use the generic FindQt module. All of the modules provide helpful tools for building Qt projects. The following is a simple example of building a project that uses Qt4.

find_package (Qt4 REQUIRED)

include ($(QT_USE_FILE})

# what are our ui files?
set (QTUI_SRCS qtwrapping.ui)
OT4_WRAP_UI (QTUI_H_SRCS $(QTUI_SRCS})
QT4_WRAP_CPP (QT_MOC_SRCS TestMoc.h)

add_library (myqtlib ${QTUI_H_SRCS) ${QT_MOC_SRCS})
target_link_libraries (myqtlib ${QT_LIBRARIES))

add_executable (qtwrapping qtwrappingmain.cxx)
target_link_libraries (qtwrapping myqtlib)

In addition to explicitly listing Qt MOC sources. CMake also has a feature called automoc which automatically scan all source files for moc contructs and runs moc accordingly. To change the above example to use automoc, simply turn the automoc property on for the library and remove the QT4_WRAP_CPP(QT_MOC_SRCS TestMoc.h) line.

set_target_properties(foo myqtlib PROPERTIES AUTOMOC TURE)

For more information about automoc, see the documentation in the vari ables section about variables with _AUTOMOC_ in them.


Using CMake with FLTK

CMake also supports the The Fast Light Toolkit (FLTK) with special commands. The FLTK_WRAP_UI command is used to run the FLTK fluid program on a .fl file and produce a C++ source file as part of the build. The following example shows how to use FLTK with CMake.

find_package (FLTK)
if (FLTK_FOUND)
  set (FLTK_SRCS
      fltk1.fl
      )
  fltk_wrap_ui (wraplibFLTK $(FLTK_SRCS})
  add_library (wraplibFLTK ${wraplibFLTK_UI_SRCS})
endif (FLTK_FOUND)


Poicies

Occasionally a new feature or change is made to CMake that is not fully backwards compatible with older versions. This can create problems when someone tries to use an old CMakeLists file with a new version of CMake. To help both end users and developers through such issues, we have introduced policies. Policies are a mechanism for helping improve backwards compatibility and tracking compatibility issues between different versions of CMake.

Design Goals

There were four main design goals for the CMake policy mechanism:

1. Existing projects should build with newer versions of CMake than that used by the project authors.

  • Users should not need to edit code to get the projects to build.
  • Warnings may be issued but the projects should build.

2. Correctness of new interfaces or bug fixes in old interfaces should not be inhibited by compatibility requirements. Any reduction in correctness of the latest interface is not fair on new projects.

3. Every change made to CMake that may require changes to a project's CMakeLists files should be documented.

  • Each change should also have a unique identifier that can be referenced with warning and error messages.
  • The new behavior is enabled only when the project has somehow indicated it is supported.

4. We must be able to eventually remove code that implements compatibility with ancient CMake versions.

  • Such removal is necessary to keep the code clean and to allow for internal refactoring.
  • After such removal, attempts at building projects written for ancient versions must fail with an informative message.

All policies in CMake are assigned a name in the form CMPNNNN where NNNN is an integer value. Policies typically support both an old behavior that preserves compatibility with earlier versions of CMake, and a new behavior that is considered correct and preferred for use by new projects. Every policy has documentation detailing the motivation for the change, and the old and new behaviors.

Setting Policies

Projects may configure the setting of each policy to request old or new behaviors. When CMake encounters user code that may be affected by a particular policy, it checks to see whether the project has set the policy. If the policy has been set (to OLD or NEW) then CMake follows the behavior specified. If the policy has not been set then the old behavior is used, but a warning is issued telling the project author to set the policy.

There are a couple ways to set the behavior of a policy. The quickest way is to set all policies to a version that corresponds to the release version of CMake the project was written in. Setting the policy version requests the new behavior for all policies introduced in the corresponding version of CMake or earlier. Policies introduced in later versions are marked as "not set" in order to produce proper warning messages. The policy version is set using the cmake_policy() (page 280) command's VERSION signature. For example, the code

cmake_policy (VERSION 2.6)

will request the new behavior for all policies introduced in CMake 2.6 or earlier. The cmake_minimum_required() (page 280) command will also set the policy version, which is conve­nient for use at the top of projects. A project should typically begin with the lines

cmake_minimum_required (VERSION 2.6)
project (MyProject)
# ...code using CMake 2.6 policies

Of course, one should replace "2.6" with the version of CMake you are currently writing to. You can also set each policy individually if you wish; this is sometimes helpful for project authors who want to incrementally convert their projects to use a new behavior, or silence warnings about dependence on an old behavior. The cmake_policy command's SET option may be used to explicitly request old or new behavior for a particular policy.

For example, CMake 2.6 introduced the policy CMP0002 (page 538), which requires all logical target names to be globally unique (duplicate target names previously worked by accident in some cases, but were not diagnosed). Projects using duplicate target names and working accidentally will receive warnings referencing the policy. The warnings may be silenced with the code

cmake_policy (SET CMP0002 OLD)

which exp licitly tel l s CMake to use the old behavior for the policy (silently accepting duplicate target names). Another option is to use the code

cmake_policy (SET CMP0002 NEW)

to explicitly tell CMake to use new behavior and produce an error when a duplicate target is created. Once this is added to the project, it will not build until the author removes any duplicate target names.

When a new version of CMake is released, it introduces new policies that can still build old projects, because by default they do not request NEW behavior for any of the new policies. When starting a new project, one should always specify the most recent release of CMake to be supported as the policy version level. This will ensure that the project is written to work using policies from that version of CMake and not using any old behavior. If no policy version is set, CMake will warn and assume a policy version of 2.4. This allows existing projects that do not specify cmake_minimum_required to build as they would have with CMake 2.4.

The Policy Stack

Policy settings are scoped using a stack. A new level of the stack is pushed when entering a new subdirectory of the project (with add_subdirectory() (page 277)) and popped when leaving it. Therefore, setting a policy in one directory of a project will not affect parent or sibling directories, but it will affect subdirectories.

This is useful when a project contains subprojects that are maintained separately yet built inside the tree. The top-level CMakeLists file in a project may write

cmake_policy (VERSION 2.6)
project (MyProject)
add_subdirectory (OtherProject)
# ... code requiring new behavior as of CMake 2.6 ...

while the OtherProject/CMakeLists.txt file contains

cmake_policy (VERSION 2.4)
project (OtherProject)
# ... code that builds with CMake 2.4 ...

This allows a project to be updated to CMake 2.6 while subprojects, modules, and included fi les continue to build with CMake 2.4 until their maintainers update them .

User code may use the cmake_policy command to push and pop its own stack levels as long as every push is paired with a pop. Thi s is useful when temporarily requesting different behavior for a small section of code. For example, policy CMP0003 (page 539) removes extra link directories that used to be included when new behavior is used. When incrementally updating a project, it may be difficult to build a particular target with the remaining targets being OK. The code

cmake_policy (PUSH)
cmake_policy (SET CMP0003 OLD) # use old-style link for now
add_executable (myexe ...)
cmake_policy (POP)

will silence the warning and use the old behavior for that target. You can get a list of policies and help on specific policies by running CMake from the command line as follows

cmake --help-command cmake_policy
cmake --help-policies
cmake --help-policy CMP0003


Updating a Project For a New Version of CMake

When a CMake release introduces new policies, it may generate warnings for some existing projects. These warnings indicate that changes to a project may be necessary for dealing with the new policies. While old releases of a project can continue to build with the warnings, the project development tree should be updated to take the new policies into account. There are two approaches to updating a tree: one-shot and incremental. The question of which one is easier depends on the size of the project and which new policies produce warnings.

The One-Shot Approach

The simplest approach to updating a project for a new version of CMake is simply to change the policy version which is set at the top of the project. Then, try building with the new CMake version to fix problems.

For example, to update a project to build with CMake 2.8, one might write

cmake_minimum_required (VERSION 2.8)

at the beginning of the top-level CMakeLists file. This tells CMake to use the new behavior for every policy introduced in CMake 2.8 and below. When building this project with CMake 2.8, no warnings will be produced regarding policies because it knows that no policies were introduced in later versions. However, if the project was depending on the old policy behavior, it may not build since CMake is now using the new behavior without warning. It is up to the project author who added the policy version line to fix these issues.

The Incremental Approach

Another approach to updating a project for a new version of CMake is to deal with each warning one-by­-one. One advantage of this approach is that the project will continue to build throughout the process, so the changes can be made incrementally.

When CMake encounters a situation where it needs to know whether to use the old or new behavior for a policy, it checks whether the project has set the policy. If the policy is set, CMake silently uses the corre­sponding behavior. If the policy is not set, CMake uses the old behavior but warns the author that the policy is not set.

In many cases, a warning message will point to the exact line of code in the CMakeLists files that caused the warning. In some cases, the situation cannot be diagnosed until CMake is generating the native build system rules for the project, so the warning will not include explicit context information. In these cases, CMake will try to provide some information about where code may need to be changed. The documentation for these "generation-time" policies should indicate the point in the project code where the policy should be set to take effect.

In order to incrementally update a project, one warning should be addressed at a time. Several cases may occur, as described below.

Silence a Warning When the Code is Correct

Many policy warnings may be produced simply because the project has not set the policy even though the project may work correctly with the new behavior (there is no way for CMake to know the difference). For a warning about some policy, CMP<NNNN>, you can check whether this is the case by adding

cmake_policy (SET CMP<NNNN> NEW)

to the top of the project and trying to build it. If the project builds correctly with the new behavior, move on to the next policy warning. If the project does not build correctly, one of the other cases may apply.

Silence a Warning Without Updating the Code

Users can suppress all instances of a warning CMP<NNNN> by adding

to the top of a project. However, we encourage project authors to update their code to work with the new behavior for all policies. This is especially important because versions of CMake in the (distant) future may remove support for old behaviors and produce an error for projects requesting them (which tells the user to get an older versions of CMake to build the project).

Silence a Warning by Updating Code

When a project does not work correctly with the NEW behaviors for a policy, the code needs to be updated. In order to deal with a warning for some policy CMP<NNNN>,add

cmake_policy (SET CMP<NNNN> NEW)

to the top of the project and then fix the code to work with the NEW behavior.

If many instances of the warning occur fixing all of them simultaneously may be too difficult: instead, a developer may fix them one at a time by using the PUSH/POP signatures of the cmake_policy command:

cmake_policy (PUSH)
cmake_policy (SET CMP<NNNN> NEW)
# ... code updated for new policy behavior ...
cmake_policy (POP)

This will request the new behavior for a small region of code that has been fixed. Other instances of the policy warning may still appear and must be fixed separately.


Updating the Project Policy Version

After addressing all policy warnings and getting the project to build cleanly with the new CMake version one step remains. The policy version set at the top of the project should now be updated to match the new CMake version, just as in the one,shot approach described above. For example, after updating a project to build cleanly with CMake 2.8, users may update the top of the project with the line

cmake_minimum_required(VERSION 2.8)

This will set all policies introduced in CMake 2.8 or below to use the new behavior. Then users m ay sweep through the rest of the code and remove the calls that use the cmake_policy command to request the new behavior incremental ly. The end result should look the same as the one,shot approach, but could be attained step-by-step.


Supporting Multiple CMake Versions

Some projects might want to support a few releases of CMake simultaneously. The goal is to build with an older version, while also working with newer versions without warnings. In order to support both CMake 2.4 and 2.6, one may write code like

cmake_minimum_required {VERSION 2.4)
if (COMMAND cmake_policy)
  # policy settings ...
  cmake_policy (SET CMP0003 NEW)
endif (COMMAND cmake_policy)

This will set the policies to build with CMake 2.6 and to ignore them for CMake 2.4. In order to support both CMake 2.6 and some policies of CMake 2.8, one may write code like:

cmake_minimum_required (VERSION 2.6)
if (POLICY CMP1234)
  # policies not known to CMake 2.6...
  cmake_policy (SET CMP1234 NEW)
endif (POLICY CMP1234)

This will set the policies to build with CMake 2.8 and to ignore them for CMake 2.6. If it is known that the project builds with both CMake 2.6 and CMake 2.8's new policies users may write:

cmake_minimum_required (VERSION 2.6)
if (NOT ${CMAKE_VERSION} VERSION_LESS 2.8)
  cmake_policy (VERSION 2.8)
endif ()


Linking Libraries

In CMake 2.6 and later, a new approach to generating link lines for targets has been implemented. Consider these libraries:

/path/to/libfoo.a
/path/to/libfoo.so

Previously, if someone wrote

target_link_libraries (myexe /path/to/libfoo.a)

CMake would generate this code to link it:

... -L/path/to -Wl, -Bstatic -lfoo -Wl, -Bdynamic ...

This worked most of the time, but some platforms (such as Mac OS X) do not support the -Bstatic or equivalent flag. This made it impossible to link to the static version of a library without creating a symlink in another directory and using that one instead. Now CMake will generate this code:

... /path/to/libfoo.a ...

This guarantees that the correct library is chosen. However, there are some caveats to keep in mind. In the past, a project could write this (incorrect) code and it would work by accident

add_executable (myexe myexe.c)
target_link_libraries (myexe /path/to/libA.so B)

Here B is meant to link /path/to/libB.so. This code is incorrect because it asks CMake to link to B, but does not provide the proper linker search path for it. It used to work by accident because the -L/path/to would get added as part of the implementation of linking to A. The correct code would be either

link_directories (/path/to)
add_executable (myexe myexe.c)
target_link_libraries (myexe /path/to/libA.so B)

or even better

add_executable (myexe myexe.c)
target_link_libraries (myexe /path/to/libA.so /path/to/libB.so)


Linking to System Libraries

System libraries on UNIX-like systems are typically provided in /usr/lib or /lib. These directories are considered implicit linker search paths because linkers automatically search these locations, even without a flag like -L/usr/lib. Consider the code

find_library (M_LIB m)
target_link_libraries (myexe ${M_LIB})

Typically the find_library command would find the math library /usr/lib/libm.so, but some platforms provide multiple versions of libraries correesponding to different architectures. For example, on an IRIX machi ne one might find the libraries

/usr/lib/libm.so               (ELF o32)
/usr/lib32/libm.so             (ELF n32)
/usr/lib64/libm.so             (ELF 64)

On a Solaris machine one might find:

/usr/lib/libm.so               (sparcv8 architecture)
/usr/lib/sparcv9/libm.so       (sparcv9 architecture)

Unfortunately, find_library may not know about all of the architecture-specific system search paths used by the linker. In fact, when it finds /usr/lib/libm.so, it may be finding a library with the incorrect architecture. If the link computation were to produce the line

... /usr/lib/lim.so ...

the linker might complain if /usr/lib/libm.so does not match the architecture it wants. One solution to this problem is to have the link computation recognize that the library is in a system directory and ask the linker to search for the library. It could produce the link line

... -lm ...

and the linker would search through its architecture-specific implicit link directories to find the correct l ibrary. Unfortunately, this solution suffers from the original problem of distinguishing between static and shared versions. In order to ask the linker to find a static system library with the correct architecture, it must produce the link line

... -Wl,-Bstatic -lm ... -Wl,-Bshared ...

Since not all platforms support such flags, CMake compromises. Libraries that are not in implicit system locations are linked by passing the full library path to the linker. Libraries that are in implicit system locations (such as /usr/lib) are linked by passing the -l option if a flag like -Bstaticis available, and by passing the full library path to the linker otherwise.


Specifying Optimized or Debug Libraries with a Target

On Windows platforms, users are often required to link debug libraries with debug libraries, and optimized libraries with optimized libraries. CMake helps satisfy this requirement with the target_link_libraries() (page 340) command, which accepts an optional flag labeled as debug or optimized. If a library is preceded with either debug or optimized, then that library will only be linked in with the appropriate configuration type. For example

add_executable (foo foo.c)
target_link_libraries (foo debug libdebug optimized libopt)

In this case, foo will be linked against libdebug if a debug build was selected, or against libopt if an optimized build was selected.


Advanced Linking

I n CMake library dependencies are transitive by default no matter if they are dynamic or static. When a target is linked with another target it will inherit all the libraries linked to this target, and they will appear on the link line for the other target too.

This behavior can be changed by setting a targets LINK_INTERFACE_LIBRARIES (page 598) property. If set, only targets listed in LINK_INTERFACE_LIBRARIES will used as the set of transitive link depen­dencies for a target. CMake provides two convenient ways to set the LINK_INTERFACE_LIBRARIES

target_link_libraries(<target> LINK_INTERFACE_LIBRARIES
                       [[debug|optimized|general] <lib>] ...)

The LINK_INTERFACE_LIBRARIES mode the appends the libraries to LINK_INTERFACE_LIBRARIES and its per-configuration equivalent target properties in­ stead of using them for linking. Libraries specified as "debug" are appended to the LINK_INTERFACE_LIBRARIES_DEBUG property (or to the properties corresponding to configura­tions listed in the DEBUG_CONFIGURATIONS (page 563) global property if it is set). Libraries specified as "optimized" are appended to the LINK_INTERFACE_LIBRARIES property. Libraries specified as "general" (or without any keyword) are treated as if specified for both "debug" and "optimized".

target_link_libraries (<target>
                       <LINK_PRIVATE|LINK_PUBLIC>
                         [{debug|optimized|general] <lib>] ...
                       [<LINK_PRIVATE|LINK_PUBLIC>
                         [[debug|optimized|general] <lib>] ...])

The LINK_PUBLIC and LINK_PRIVATE modes can be used to specify both the link dependencies and the link interface in one command. Libraries and targets following LINK_PUBLIC are linked to, and are made part of the LINK_INTERFACE_LIBRARIES. Libraries and targets following LINK_PRIVATE are linked to, but are not made part of the LINK_INTERFACE_LIBRARIES. Using LINK_PUBLIC and LINK_PRIVATE causes all other libraries (before and after) linked to a target to be private unless they are explicitly stated to be LINK_PUBLIC.

CMake will also propagate "usage requirements" from linked library targets. Usage requirements affect compilation of sources in the <target>. They are specified by properties defined on linked targets. During generation of the build system, CMake integrates usage requirement property values with the corresponding build properties for <target>:

INTERFACE_COMPILE_DEFINITIONS (page 592): Appends to COMPILE_DEFINITIONS (page 616)

INTERFACE_INCLUDE_DIRECTORIES (page 593): Appends to INCLUDE_DIRECTORIES (page 570)

INTERFACE_POSITION_INDEPENDENT_CODE (page 594): Sets:prop_tgt: POSITION_INDEPENDENT_CODE or checked for consistency with existing value

For example to specify include directories that are required when linking to a library you would can do the following

add_library (foo foo.cxx)
set_property (TARGET foo APPEND PROPERTY
INTERFACE_INCLUDE_DIRECTORIES "${CMAKE_CURRENT_BINARY_DIR}"
"${CMAKE_CURRENT_SOURCE_DIR}")

Now anything that links to the target foo will automatically have foo's binary and source as include directo­ries. The order of the include directories brought in through "usage requirements" will match the order of the targets in the target_link_libraries call.


Object Libraries

Before version 2.8.8, CMake had no way to encapsulate numerous libraries into one combined library. Pre­viously you would have to compile each individual library and the combined library. This is okay if the compilation time was low and each library used the same preprocessor definitions, include directories, and flags.

However, large projects typically organize their source files into groups, often in separate subdirectories, that each need different include directories and preprocessor definitions. For this use case CMake has developed the concept of Object Libraries. An Object Library is a collection of source files compiled into an object file which is not linked into a library file or made into an archive.

Instead other targets created by add_library or add_executable may reference the objects using an expres­sion of the form $<TARGET_OBJECTS:name> as a source, where "name" is the target created by the add_library() (page 274) call. For example:

add_library(A OBJECT a.cpp)
add_library(B OBJECT b.cpp)
add_library(Combined $<TARGET_OBJECTS:A> $<TARGET_OBJECTS:B> )

will include A and B object files in a library called Combined. Object libraries may contain only sources (and headers) that compile to object files.

Shared Libraries and Loadable Modules

Shared libraries and loadable modules are very powerful tools for software developers. They can be used to create extension modules or plugins for off-the-shelf software, and can be used to decrease the com­pile/link/run cycles for C and C++ programs. However, despite years of use, the cross-platform creation of shared libraries and modules remains a dark art understood by only a few developers. CMake has the ability to aid developers in the creation of shared libraries and modules. CMake knows the correct tools and flags to use in order to produce shared libraries for most modern operating systems that support them. Unfortunately, CMake cannot do all the work, and developers must sometimes alter source code and understand the basic concepts and common pitfalls associated with shared libraries before they can be used effectively. This sec­tion will describe many of the considerations required for taking advantage of shared libraries and loadable modules.

A shared library should be thought of more like an executable than a static library; on most systems they actually require executable permissions to be set on the shared library file. This means that shared libraries can link to other shared libraries when they are created in the same way as an executable. Unlike a static library where the atomic unit is the object file, for shared libraries, the entire library is the atomic unit. This can cause some unexpected linker errors when converting from static to shared libraries. If an object file is part of a static library but the executable linking to the library does not use any of the symbols in that object file, then the file is simply excluded from the final linked executable. With shared libraries, all the object files that make up the library and all of the dependencies that they require come as one unit. For example, suppose you had a library with an object file defining the function DisplayOnXWindow(), which required the X11 library. If you linked an executable to that library, but did not call the DisplayOnXWindow() function, the static library version would not require X11; but the shared library version would require the X11 library. This is because a shared library has to be taken as one unit, and a static library is only an archive of object files from which linkers can choose the objects needed. This means that static linked executables can be smaller, as they only contain the object code actually used.

Another difference between shared and static libraries is library order. With static libraries the order on the link line can make a difference; this is because most linkers only use the symbols that are needed in a single pass over all the given libraries. So, the library order should go from the library that uses the most other libraries to the library that does not use any other libraries. CMake will preserve and remember the order of libraries and library dependencies in a project. This means that each library in a project should use the target_link_libraries() (page 340) command to specify all of the libraries that it directly depends on. The libraries will be linked with each other for shared builds, but not static builds; however, the link information is used in static builds when executables are linked. An executable that only links library libA will get libA plus libB and libC, as long as libA's dependency on libB and libC was properly speci fied using target_link_libraries(libA libB libC).

At this point, one m ight wonder why shared libraries would be preferred over static libraries. There are several reasons. First, shared libraries can decrease the compile/link/run cycle time because the linker does not have to do as much work as there are fewer decisions to be made about which object files to keep. Often times, the executable does not even need to be re-linked after the shared library is rebuilt; therefore developers can work on a library by compiling and linking only the small part of the program that is currently being developed, and then re-running the executable after each build of the shared library. Also, if a library is used by many different executables on a system, then there only needs to be one copy of the library on disk and often in memory too.

In addition to the concept of a software library, shared libraries can also be used on many systems as run time loadable modules. This means that at run time, a program can load and execute object code that was not part of the original software. This allows developers to create software that is both open and closed. (For more information, see Object-Oriented Software Construction by Bertrand Meyer.) Closed software is that which cannot be modified. It has been through a testing cycle and can be certified to perform specific tasks with regression tests. However, a seemingly opposite goal is sought after by developers of object-oriented software as Open software can be extended by future developers. This can be done via inheritance and polymorphism with object systems . Shared libraries that can be loaded at run time allow for these seemingly opposing goals to be achieved in the same software package. Many common applications support the idea of plugins; the most common of these applications is the web browser. Internet Explorer uses plugins to support video over the web and 3D visualization. In addition to plugins, loadable factories can be used to replace C++ objects at run time, as is done in the Visualization Toolkit (VTK).

Once it is decided that shared libraries or loadable modules are the right choice for a particular project, there are a few issues that developers need to be aware of. The first question that must be answered is "which symbols are exported by the shared library?" This may sound like a simple question, but the answer is different for each platform. On many but not all UNIX systems, the default behavior is to export all the symbols much like a static library. However, on Windows systems, developers m ust explicitly tell the linker and compiler which symbols are to be exported and imported from shared libraries. This is often a big problem for UNIX developers moving to Windows. There are two ways to tell the compiler/linker which symbols to export/import on Windows. The most common approach is to decorate the code with a Microsoft C/C++ language extension. An alternative is to create an extra file called a .def file, which is a simple ASCII file containing the names of all the symbols to be exported from a library.

The Microsoft extension uses the _declspec directive. If a symbol has _declspec(dllexport) in front of it, it will be exported; if it has _declspec (dllimport), it will be imported. Since the same file may be shared during the creation and use of a library, it must be both exported and imported in the same source file. Thi s can only be done with the preprocessor. The developer can create a macro called LIBRARY_EXPORT which is defined to dllexport when building the library, and dllimport when using the library. CMake helps this process by automatically defining ${LIBNAME}_EXPORTS when building a DLL (dynamic link library, a.k.a. a shared library) on Windows.

The following code snippet is from the VTK library, vtkCommon, and is included by all files in the vtkCom­mon library :

if defined (WIN32)

if defined (vtkCommon_EXPORTS)
define VIK_COMMON_EXPORT __declspec( dilexport )}
else
define VTIK_COMMON_EXPORT __declispec({ dllimport }
endif
else
define VTK_COMMON_EXPORT
endif

The example checks for Windows and for the vtkCommon_EXPORTS macro provided by CMake. So, on UNIX, VTK_COMMON_EXPORT is defined to nothing; on Windows during the building of vtkCommon.dll, it is defined as _declspec(dllexport); and when the file is being used by another file, it is defined to _declspec(dllimport).

More recently, Linux and other Unix systems have added linker options that allow symbols to be explicitly exported in a similar manner as Windows. CMake has a module that will allow you to use explicit symbol exports on all systems that support them. The module is GenerateExportHeader.cmake, and contains the function generate_export_header. The function will modify the CXX and C flags to turn on explicit symbol exports for the system. It will also generate a header file much like the handwritten one above, which works for Windows only. For more information, see generate_export_header in the appendix.

The second approach on Windows requires a .def file to specify the symbols to be exported. This file could be created by hand, but for a large and changing C++ library, that could be time consuming and error-prone. CMake's custom commands can be used to run a pre-link program which will create a .def file from the compiled object files automatically. In the following example, a Perl script called makedef.pl is used; the script runs the DUMPBIN program on the .obj files, extracts all of the exportable symbols, and writes a .def file with the correct exports for all the symbols in the library mylib.

----CMakeLists.txt----

cmake_minimum_required (VERSION 2.6)
project (myexe)

set (SOURCES mylib.cxx mylib2.cxx)

# create a list of all the object files
string (REGEX REPLACE "\\.cxx" ".obj" OBJECTS "${SOURCES}")

# create a shared library with the .def file
add_library (mylib SHARED ${SOURCES}
  ${CMAKE_CURRENT_BINARY_DIR} /mylib.def
  }
# set the .def file as generated
set_source_files_properties (
  ${CMAKE_CURRENT_BINARY_DIR}/mylib.def
  PROPERTIES GENERATED 1
  )

# create an executable
add_executable (myexe myexe.cxx)

# Link the executable to the dll
target_link_libraries(myexe mylib)

#convert to windows slashes
set (OUTDIR
  $(CMAKE_CURRENT_BINARY_DIR}/${CMAKE_CFG_INTDIR}
  )

string (REGEX REPLACE */* "\\\\" OUTDIR ${OUTDIR})

# create a custom pre link command that runs
# a perl script to create a .def file using dumpbin
add_custom_command {
  TARGET mylib PRE_LINK
  COMMAND perl
  ARGS $(CMAKE_CURRENT_SOURCE_DIR}/makedef.pl]
  ${CMAKE_CURRENT_BINARY_DIR}\\mylib.def mylib
  S${OUTDIR} $(OBJECTS}
  COMMENT "Create .def file"
  )

----myexe.cxx----
#include <iostream>
#include "mylib.h"
int main()
{
  std::cout << myTen() << "\n";
  std::cout << myEight() << "\n";
}

----mylib.cxx----
int myTen()
{
  return 10;
}

----mylib2.cxx----
int myEight()
{
  return 8;
}

There is a significant difference between Windows and the default linker options on UNIX systems with respect to the requirements of symbols. DLLs on Windows are required to be fully resolved, meaning that they must link every symbol at creation. UNIX systems allow shared libraries to get symbols from the executable or other shared libraries at run time. On UNIX systems that support this feature, CMake will compile with the flags that allow executable symbols to be used by shared libraries. This small difference can cause large problems. A common but hard to track with DLLs occurs with C++ template classes and static members. In these instances, two DLLs can end up with separate copies of what is supposed to be a single, global static member of a class. There are also problems with this approach on most UNIX systems; the start-up time for large applications with many symbols can be long since much of the linking is deferred to run time.

Another common pitfall occurs with C++ global objects. These objects require constructors to be called before they can be used. The main that links or loads C++ shared libraries MUST be linked with the C++ compiler, or globals like cout may not be initialized before they are used, causing strange crashes at start up time.

Since executables that link to shared libraries must be able to find the libraries at run time, special environment variables and linker flags must be used. There are tools that can be used to show which libraries an executable is actually using. On many UNIX systems there is a tool called ldd(otool -L on Mac OS X), which shows which libraries are used by an executable. On Windows, a program called depends can be used to find the same type of information. On many UNIX systems, there are also environment variables like LD_LIBRARY_PATH that tell the program where to find the libraries at run time. Where supported CMake will add run time library path information into the linked executables, so that LD_LIBRARY_PATH is not required. This feature can be turned off by setting the cache entry CMAKE_SKIP_RPATH (page 633) to false; this may be desirable for installed software that should not be looking in the build tree for shared libraries. On Windows, there is only one PATH environment variable that is used for both DLLs and finding executables.


Shared Library Versioning

When an executable is linked to a shared library, it is important that the copy of the shared library loaded at run time matches that expected by the executable. On some UNIX systems, a shared library has an associated "soname" intended to solve this problem. When an executable links against the library, its soname is copied into the executable. At run time, the dynamic linker uses this name from the executable to search for the library.

Consider a hypothetical shared library "foo" providing a few C functions that implement some functionality. The interface to foo is called an Application Programming Interface (API). If the implementation of these C functions change in a new version of foo, but the API remains the same, then executables linked against foo will still run correctly. When the API changes, old executables will no longer run with a new copy of foo; a new API version number must be associated with foo.

This can be implemented by creating the origi nal version of foo with a soname and file name such as libfoo.so.1. A symbolic link such as libfoo.so -> libfoo.so.1 will allow standard linkers to work with the library and create executables. The new version of foo can be called libfoo.so.2 and the symbolic link updated so that new executables use the new library. When an old executable runs, the dynamic linker will look for libfoo.so.1, find the old copy of the library, and run correctly. When a new executable runs, the dynamic linker will look for libfoo.so.2 and correctly load the new version.

This scheme can be expanded to handle the case of changes to foo that do not modify the APL We introduce a second set of version numbers that is totally independent of the first, which corresponds to the software version providing foo. For example, a larger project may have introduced the existence of the library foo starting in version 3.4. In this case, the file name for foo might be libfoo.so.3.4 , but the soname would still be libfoo.so.1 because the API for foo is still on its first version. A symbolic link from libfoo.so.1 -> libfoo.so.3.4 will allow executables linked against the library to run. When a bug is fixed in the software without changing the API to foo, then the new library file name might be libfoo.so.3.5, and the symbolic link can be updated to allow existing executables to run.

CMake supports this soname-based version number encoding on platforms supporting soname natively. A target property for the shared library named VERSION (page 609) specifies the version number used to create the file name for the library. This version should correspond to that of the software package providing foo. On Windows, the VERSION property is used to set the binary image number using major.minor format. Another target property named SOVERSION (page 608) specifies the version number used to create the soname for the library. Thi s version should correspond to the API version number for foo. These target properties are ignored on platforms where CMake does not support this scheme.

The following CMake code configures the version numbers of the shared library foo

set_target_properties (foo PROPERTIES VERSION 1.2 SOVERSION 4)

This results in the following library and symbolic links:

libfoo.so.1.2
libfoo.so.4 -> libfoo.so.1.2
libfoo.so -> libfoo.so.4

If only one of the two properties is specified, the other defaults to its value automatically. For example, the code

set_target_properties (foo PROPERTIES VERSION 1.2)

results in the following shared library and symbolic link:

libfoo.so.1.2
libfoo.so -> libfoo.so.1.2

CMake makes no attempt to enforce sensible version numbers. It is up to the programmer to utilize this feature in a productive manner.


Installing Files

Software is typically installed into a directory separate from the source and build trees. This allows it to be distributed in a clean form and isolates users from the details of the build process. CMake provides the install() (page 317) command to specify how a project is to be installed. This command is invoked by a project in the CMakeLists file and tells CMake how to generate installation scripts. The scripts are executed at install time to perform the actual installation of files. For Makefile generators (UNIX, NMake, Borland, MinGW, etc.), the user simply runs make install(or nmakeinstall) and the make tool will invoke CMake's installation module. With GUI based systems (Visual Studio, Xcode, etc.), the user simply builds the target called INSTALL.

Each call to the install command defines some installation rules. Within one CMakeLists file (source directory), these rules will be evaluated in the order that the corresponding commands are invoked. The order across multiple directories is not specified.

The install command has several signatures designed for common installation use cases. A particular in­ vocation of the command specifies the signature as the first argument. The signatures are TARGETS, FILES, PROGRAMS, DIRECTORY, SCRIPT, and CODE.

install (TARGETS...) Installs the binary files corresponding to targets built inside the project.

install (FILES...) General-purpose file installation, which is typically used for header files, documentation, and data files required by your software.

install (PROGRAMS...) Installs executable files not built by the project, such as shell scripts. This ar­gument is identical to install(FILES) except that the default permjssions of the installed file include the executable bit.

install (DIRECTORY...) This argument installs an entire directory tree. It may be used for installing di­rectories with resources, such as icons and images.

install (SCRIPT...) Specifies a user-provided CMake script file to be executed during installation . This is typically used to define pre-install or post-install actions for other rules.

install (CODE...) Specifies user-provided CMake code to be executed during the installation. This is sim­ilar to install(SCRIPT) but the code is provided inline in the call as a string. The TARGETS, FILES, PROGRAMS, DIRECTORY signatures are all meant to create install rules for files. The targets, files, or directories to be installed are listed immediately after the signature name argument. Addi­tional details can be specified using keyword arguments followed by corresponding values. Keyword arguments provided by most of the signatures are as follows.

DESTINATION This argument specifies the location where the installation rule will place files, and must be followed by a directory path indicating the location. If the directory is specified as a full path, it will be evaluated at install time as an absolute path. If the directory is specified as a relative path, it will be evaluated at install time relative to the installation prefix. The prefix may be set by the user through the cache variable CMAKE_INSTALL_PREFIX (page 645). A platform-specific default is provided by CMake: /usr/local on UNIX, and <SystemDrive>/ProgramFiles/<ProjectName>" on Windows, where SystemDrive is along the lines of C : and ProjectName is the name given to the top­ most PROJECT() (page 327) command.

PERMISSIONS This argument specifies file permissions to be set on the installed files. This option is needed only to override the default permissions selected by a particular INSTALL command sig­nature. Valid permissions are OWNER_READ, OWNER_WRITE, OWNER_EXECUTE, GROUP_READ, GROUP_WRITE, GROUP_EXECUTE, WORLD_READ, WORLD_WRITE, WORLD_EXECUTE, SETUID, and SETGID. Some platforms do not support all of these permissions; on such platforms those permis­sion names are ignored.

CONFIGURATIONS This argument specifies a list of build configurations for which an installation rule applies (Debug, Release, etc.). For Makefile generators, the build configuration is specified by the CMAKE_BUILD_TYPE cache variable. For Visual Studio and Xcode generators, the configuration is selected when the INSTALL target is built. An installation rule will be evaluated only if the cur­rent install configuration matches an entry in the list provided to this argument. Configuration name comparison is case-insensitive.

COMPONENT This argument specifies the installation component for which the installation rule applies. Some projects divide their installations into multiple components for separate packaging. For ex­ample, a project may define a Runt ime component that contains the files needed to run a tool; a Development component containing the files needed to build extensions to the tool; and a Documentation component containing the manual pages and other help files. The project may then package each component separately for distribution by installing only one component at a time. By default, all components are installed. Component-specific installation is an advanced feature intended for use by package maintainers. It requires manual invocation of the installation scripts with an argument defining the COMPONENT variable to name the desired component. Note that component names are not defined by CMake. Each project may define its own set of components.

OPTIONAL This argument specifies that it is not an error if the input file to be installed does not exist. If the input file exists, it will be installed as requested. If it does not exist, it will be silently not installed.

Projects typically install some of the library and executable files created during their build process. The install command provides the TARGETS signature for this purpose:

install (TARGETS targets...
        ([ARCHIVE | LIBRARY | RUNTIME | FRAMEWORK | BUNDLE |
        PRIVATE_HEADER | PUBLIC_HEADER | RESOURCE]
        [DESTINATION <dir>]
        [PERMISSIONS permissions...]
        [CONFIGURATIONS [(Debug|Release|...]]
        [COMPONENT <component >]
        [OPTIONAL]
            (EXPORT <export name>]
            [NAMELINK_ONLY|NAMELINK_SKIP]
        ] [...])

The TARGETS keyword is immediately followed by a list of the targets created using add_executable() (page 273) or add_library() (page 274), which are to be installed. One or more files corresponding to each target will be installed.

Files installed with this signature may be divided into three categories: ARCHIVE, LIBRARY, and RUNTIME. These categories are designed to group target files by typical installation destination. The corresponding keyword arguments are optional, but if present, specify that other arguments following them apply only to target files of that type. Target files are categorized as follows:

executables - "RUNTIME" Created by add_executable (.exe on Windows, no extension on UNIX)

loadable modules - "LIBRARY" Created by add_library with the MODULE option (.dll on Windows, .so on UNIX)

shared libraries - "LIBRARY" Created by add_library with the SHARED option on UNIX-like plat­forms (.so on most UNIX, .dylib on Mac)

dynamic-link libraries - "RUNTIME" Created by add_library with the SHARED option on Windows platforms (.dll)

import libraries - "ARCHIVE" A linkable file created by a dynamic-link library that exports symbols (.lib on most Windows, .dll.a on Cygwin and MinGW).

static libraries - "ARCHIVE" Created by add_library with the STATIC option (.lib on Windows, .a on UNIX, Cygwin, and MinGW)

Consider a project that defines an executable, myExecutable, which links to a shared library mySharedLib. It also provides a static library myStaticLib and a plugin module to the executable called myPlugin that also links to the shared library. The executable, static library, and plugin file may be installed individually using the commands

install (TARGETS myExecutable DESTINATION bin)
install (TARGETS myStaticLib DESTINATION lib/myproject)
install (TARGETS myPlugin DESTINATION lib)

The executable will not be able to run from the installed location until the shared library to it links to is also installed. Installation of the library requires a bit more care in order to support all platforms. It must be installed in a location searched by the dynamic linker on each platform. On UNIX-like platforms, the library is typically installed to lib, while on Windows it should be placed next to the executable in bin. An additional challenge is that the import library associated with the shared library on Windows should be treated like the static library, and installed to lib/myproject. In other words, we have three different kinds of files created with a single target name that must be installed to three different destinations! Fortunately, this problem can be solved using the category keyword arguments. The shared library may be installed using the command:

install (TARGETS myShareLib
         RUNTIME DESTINATION bin
         LIBRARY DESTINATION lib
         ARCHIVE DESTINATION lib/myproject)

This tells CMake that the RUNTIME file (.dll) should be installed to bin, the LIBRARY file (.so) should be installed to lib, and the ARCHIVE(.lib) file should be installed to lib/myproject. On UNIX, the LIBRARY file will be installed; on Windows, the RUNTIME and ARCHIVE files will be installed.

If the above sample project is to be packaged into separate run time and development components, we must assign the appropriate component to each target file installed. The executable, shared library, and plugin are required in order to run the application, so they belong in a Runtime component. Meanwhile, the import library (corresponding to the shared library on Windows) and the static library are only required to develop extensions to the application, and therefore belong in a Development component.

Component assignments may be specified by adding the COMPONENT argument to each of the commands above. You may also combine all of the installation rules into a single command invocation, which is equiv­alent to all of the above commands with components added. The files generated by each target are installed using the rule for their category.

install (TARGETS myExecutable mySharedLib myStaticLib myPlugin
        RUNTIME DESTINATION bin            COMPONENT Runtime
        LIBRARY DESTINATION lib            COMPONENT Runtime
        ARCHIVE DESTINATION lib/myproject  COMPONENT Development)

Either NAMELINK_ONLY or NAMELINK_SKIP may be specified as a LIBRARY option. On some plat­forms, a versioned shared library has a symbolic link such as

lib<name>.so -> lib<name>.so.1

where lib<name>.so.1 is the soname of the library, and lib<name>.so is a "namelink" that helps linkers to find the library when given -l<name>. The NAMELINK_ONLY option results in installation of only the namelink when a library target is installed. The NAMELINK_SKIP option causes installation of library files other than the namelink when a library target is installed. When neither option is given, both portions are installed. On platforms where versioned shared libraries do not have namelinks, or when a library is not versioned, the NAMELINK_SKIP option installs the library and the NAMELINK_ONLY option install s nothing. See the VERSION and SOVERSION target properties for details on creating versioned, shared libraries.

Projects may install files other than those that are created with add_executable or add_library, such as header files or documentation. General-purpose installation of files is specified using the FILES signature:

install (FILES files... DESTINATION <dir>
         [PERMISSIONS permissions...]
         [CONFIGURATIONS [Debug|Release|...]]
         [COMPONENT <component>]
         [RENAME <name>] [OPTIONAL])

The FILES keyword is immediately followed by a list of files to be installed. Relative paths are evaluated with respect to the current source directory. Files will be installed to the given DESTINATION directory. For example, the command

install (FILES my-api.h ${CMAKE_CURRENT_BINARY_DIR}/my-config.h
         DESTINATION include)

installs the file my-api.h from the source tree, and the file my-config.h from the build tree into the include directory under the installation prefix. By default installed files are given the permissions OWNER_WRITE, OWNER_READ, GROUP_READ, and WORLD_READ, but this may be overridden by speci­fying the PERMISSIONS option. Consider cases in which users would want to install a global configuration file on a UNIX system that is readable only by its owner (such as root). We accomplish this with the command

install (FILES my-rc DESTINATION /etc
         PERMISSIONS OWNER_WRITE OWNER_READ)

which installs the file my-rc with owner read/write permission into the absolute path /etc.

The RENAME argument specifies a name for an installed file that may be different from the original file. Renaming is allowed only when a single file is installed by the command. For example, the command

install (FILES version.h DESTINATION include RENAME my-version.h)

will install the file version.h from the source directory to include/my-version.h under the instal­ lation prefix.

Projects may also install helper programs, such as shell scripts or Python scripts that are not actually com­piled as targets. These may be installed with the FILES signature using the PERMISSIONS option to add execute permission. However, this case is common enough to justify a simpler interface. CMake provides the PROGRAMS signature for this purpose:

install (PROGRAMS files... DESTINATION <dir>
        [PERMISSIONS permissions... ]
        [CONFIGURATIONS [Debug|Release|...]]
        [COMPONENT <component>]
        [RENAME <name>] [OPTIONAL])

The PROGRAMS keyword is immediately followed by a list of scripts to be installed. This command is identical to the FILES signature, except that the default permissions additionally include OWNER_EXECUTE, GROUP_EXECUTE, and WORLD_EXECUTE. For example, we may install a Python utility script with the command

install (PROGRAMS my-util.py DESTINATION bin)

which installs my-util.py to the bin directory under the installation prefix and gives it owner, group, world read and execute permissions, plus owner write.

Projects may also provide an entire directory full of resource files, such as icons or html documentation. An entire directory may be installed using the DIRECTORY signature:

install (DIRECTORY dirs... DESTINATION <dir>
        [FILE _PERMISSIONS permissions... ]
        [DIRECTORY_PERMISSIONS permissions...]
        [USE_SOURCE_PERMISSIONS]
        [CONFIGURATIONS [Debug|Release|...]]
        [COMPONENT <component>]
        [[PATTERN <pattern> | REGEX <regex>]
        [EXCLUDE] [PERMISSIONS permissions...]] [...])

The DIRECTORY keyword is immediately followed by a list of directories to be installed. Relative paths are evaluated with respect to the current source directory. Each named directory is installed to the destination directory. The last component of each input directory name is appended to the destination directory as that directory is copied. For example, the command

install (DIRECTORY data/icons DESTINATION share/myproject)

will install the data/icons directory from the source tree into share/myproject/icons under the installation prefix. A trailing slash will leave the last component empty and install the contents of the input directory to the destination. The command

install (DIRECTORY doc/html/ DESTINATION doc/myproject)

installs the contents of doc/html from the source directory into doc/myproject under the installation prefix. If no input directory names are given, as in

install (DIRECTORY DESTINATION share/myproject/user)

the destination directory will be created but nothing will be installed into it.

Files installed by the DIRECTORY signature are given the same default permissions as the FILES signature. Directories installed by the DIRECTORY signature are given the same default permissions as the PROGRAMS signature. The FILE_PERMISSIONS and DIRECTORY_PERMISSIONS options may be used to override these defaults. Consider a case in which a directory full of example shell scripts is to be installed into a directory that is both owner and group writable. We may use the command

install (DIRECTORY data/scripts DESTINATION share/myproject
        FILE PERMISSIONS
          OWNER READ OWNER_EXECUTE OWNER WRITE
          GROUP_READ GROUP_EXECUTE
          WORLD_READ WORLD_EXECUTE
        DIRECTORY_PERMISSIONS
          OWNER READ OWNER_EXECUTE OWNER_ WRITE
          GROUP_READ GROUP_EXECUTE GROUP_WRITE
          WORLD_READ WORLD_EXECUTE)

which installs the directory data/scripts into share/myproject/scripts and sets the desired permissions. In some cases, a fully-prepared input directory created by the project may have the desired permissions already set. The USE_SOURCE_PERMISSIONS option tells CMake to use the file and directory permissions from the input directory during installation. If in the previous example the input directory were to have already been prepared with correct permissions, the following command may have been used instead:

install (DIRECTORY data/scripts DESTINATION share/myproject
        USE_SOURCE_PERMISSIONS)

If the input directory to be installed is under source management, such as CVS, there may be extra subdirecto­ries in the input that you do not wish to install. There may also be specific files that should not be installed or be installed with different permissions, while most files get the defaults. The PATTERN and REGEX options may be used for this purpose. A PATTERN option is followed first by a globbing pattern and then by an EXCLUDE or PERMISSIONS option. A REGEX option is followed first by a regular expression and then by EXCLUDE or PERMISSIONS. The EXCLUDE option skips installation of those files or directories matching the preceding pattern or expression, while the PERMISSIONS option assigns specific permissions to them.

Each input file and directory is tested against the pattern or regular expression as a full path with forward slashes. A pattern will match only complete file or directory names occurring at the end of the full path, while a regular expression may match any portion. For example, the pattern foo* will match .../foo.txt but not .../myfoo.txt or .../foo/bar.txt; however, the regular expression foo will match all of them.

Returning to the above example of installing an icons directory, consider the case in which the input directory is managed by CVS and also contains some extra text files that we do not want to install. The command

install (DIRECTORY data/icons DESTINATION share/myproject
        PATTERN *CVS" EXCLUDE
        PATTERN "*.txt" EXCLUDE)

installs the icons directory while ignoring any CVS directory or text file contained. The equivalent command using the REGEX option is

install (DIRECTORY data/icons DESTINATION share/myproject
        REGEX "/CVS$" EXCLUDE
        REGEX "/[^/]*.txt$" EXCLUDE)

which uses '/' and '$' to constrain the match in the same way as the patterns. Consider a similar case in which the input directory contains shell scripts and text files that we wish to install with different permissions than the other files. The command

install (DIRECTORY data/other/ DESTINATION share/myproject
        PATTERN "CVS" EXCLUDE
        PATTERN "*.txt"
          PERMISSIONS OWNER_READ OWNER_WRITE
        PATTERN "*.sh"
          PERMISSIONS OWNER READ OWNER_WRITE OWNER_EXECUTE)

will install the contents of data/other from the source directory to share/myproject while ignoring CVS directories and giving specific permissions to .txt and .sh files.

Project installations may need to perform tasks other than just placing files in the installation tree. Third-party packages may provide their own mechanisms for registering new plugins that must be invoked during project installation. The SCRIPT signature is provided for this purpose:

install (SCRIPT <file>)

The SCRIPT keyword is immediately followed by the name of a CMake script. CMake will execute the script during installation. If the file name given is a relative path, it will be eval uated with respect to the current source directory. A simple use case is printing a message during installation. We first write a message. cmake file containing the code

message ("Installing My Project")

and then reference this script using the command:

install (SCRIPT message.cmake)

Custom installation scripts are not executed during the main CMakeLists file processing; they are ex­ecuted during the installation process itself. Variables and macros defined in the code containing the install(SCRIPT) call will not be accessible from the script. However, there are a few variables defined during the script execution that may be used to get information about the installation. The vari­able CMAKE_INSTALL_PREFIX (page 645) is set to the actual installation prefix. This may be dif­ferent from the corresponding cache variable value, because the installation scripts may be executed by a packaging tool that uses a different prefix. An environment variable ENV{DESTDIR} may be set by the user or packaging tool. Its value is prepended to the installation prefix and to absolute installa­tion paths to determine the location where files are installed. In order to reference an install location on disk, custom script may use $ENV{DESTDIR} ${CMAKE_INSTALL_PREFIX} as the top portion of the path. The variable CMAKE_INSTALL_CONFIG_NAME is set to the name of the build configura­tion currently being installed (Debug, Release, etc.). During component-specific installation, the variable CMAKE_INSTALL_COMPONENT is set to the name of the current component.

Custom installation scripts, as simple as the message above, are more easily created with the script code placed inline in the call to the INSTALL command. The CODE signature is provided for this purpose:

install (CODE "<code>")

The CODE keyword is immediately followed by a string containing the code to place in the installation script. An install-time message may be created using the command

install (CODE "MESSAGE(\"Installing My Project\")")

which has the same effect as the message. cmake script but contains the code inline.


Installing Prerequisite Shared Libraries

Executables are frequently built using shared libraries as building blocks. When you install such an exe­cutable, you must also install its prerequisite shared libraries, called "prerequisites" because the executable requires their presence in order to load and run properly. The three main sources of shared libraries are the operating system itself, the build products of your own project, and third party libraries belonging to an external project. The ones from the operating system may be relied upon to be present without installing anything: they are on the base platform where your executable runs. The build products in your own project presumably have add_library build rules in the CMakeLists files, and so it should be straightforward to create CMake install rules for them. It is the third party libraries that frequently become a high maintenance item when there are more than a handful of them, or when the set of them fluctuates from version-to-version of the third party project. Libraries may be added, code may be reorganized, and the third party shared libraries themselves may actually have additional prerequisites that are not obvious at first glance.

CMake provides two modules to make it easier to deal with required shared libraries. The first module, GetPrerequisites.cmake, provides the get_prerequisites function to analyze and classify the prereq­uisite shared libraries upon which an executable depends. Given an executable file as input, it will produce a list of the shared libraries required to run that executable, including any prerequisites of the discovered shared libraries themselves. It uses native tools on the various underlying platforms to perform this analysis: dumpbin (Windows), otool (Mac), and ldd (Linux). The second module, BundleUtilities.cmake, provides the fixup_bundle function to copy and fix prerequisite shared libraries using well-defined locations relative to the executable. For Mac bundle applications, it embeds the libraries inside the bundle, fixing them with install_name_tool to make a self-contained unit. On Windows, it copies the libraries into the same directory with the executable since executables will search in their own directories for their required DLLs.

The fixup_bundle function helps you create relocatable install trees. Mac users appreciate self-contained bundle applications: you can drag them anywhere, double click them, and they still work. They do not rely on anything being installed in a certain location other than the operating system itself. Similarly, Windows users without administrative privileges appreciate a relocatable install tree where an executable and all required DLLs are installed in the same directory, so that it works no matter where you install it. You can even move things around after installing them and it will still work.

To use fixup_bundle, first install one of your executable targets. Then, configure a CMake script that can be called at install time. Inside the configured CMake script, simply include BundleUtilities and call the fixup_bundle function with appropriate arguments.

In CMakeLists.txt

install (TARGETS myExecutable DESTINATION bin)

# To install, for example, MSVC runtime libraries:
include (InstallRequiredSystemLibraries)

# To install other/non-system 3rd party required libraries:
configure_file (
  ${CMAKE_CURRENT_SOURCE_DIR}/FixBundle.cmake.in
  ${CMAKE_CURRENT_BINARY_DIR}/FixBundle.cmake
  @ONLY
  )

install (SCRIPT ${CMAKE_CURRENT_BINARY_DIR}/FixBundle.cmake)

In FixBundle.cmake.in:

include (BundleUtilities)

# Set bundle to the full path name of the executable already
# existing in the install tree:
set (bundle
"${CMAKE_INSTALL_PREFIX}/myExecutable@CMAKE_EXECUTABLE_SUFFIX@")

# Set other_libs to a list of full path names to additional
# libraries that cannot be reached by dependency analysis.
# (Dynamically loaded PlugIns, for example.)
set (other_libs "")

# Set dirs to a list of directories where prerequisite libraries
# may be found:
set (dirs
  "@CMAKE_RUNTIME_OUTPUT_DIRECTORY@"
  "@CMAKE_LIBRARY_OUTPUT_DIRECTORY@"
  )

fixup_bundle ("${bundle}" "${other_libs}" "${dirs}")

You are responsible for verifying that you have permission to copy and distribute the prerequisite shared libraries for your executable. Some libraries may have restrictive software licenses that prohibit making copies a la fixup_bundle.


Exporting and Importing Targets

CMake 2.6 introduced support for exporting targets from one CMake-based project and importing them into another. The main feature allowing this functionality is the notion of an IMPORTED target. Here we present imported targets and then show how CMake files may be generated by a project to export its targets for use by other projects.


Importing Targets

Imported targets are used to convert files outside of the project on disk into logical targets inside a CMake project. They are created using the IMPORTED (page 590) option to the add_executable() (page 273) and add_library() (page 274) commands. No build files are generated for imported targets. They are used simply for convenient, flexible reference to outside executables and libraries. Consider the following example which creates and uses an IMPORTED executable target

add_executable (generator IMPORTED)                         # 1
set_property (TARGET generator PROPERTY
              IMPORTED_LOCATION "/path/to/some_generator")  # 2

add_custom_command (OUTPUT generated.c
                    COMMAND generator generated.c)          # 3

add_executable (myexe src1.c src2.c generated.c)

Line # 1 creates a new CMake target called generator. Line #2 tells CMake the location of the target on disk to import. Line #3 references the target in a custom command. Once CMake is run, the generated build system will contain a command line such as

/path/to/some_generator /project/binary/dir/generated.c

in the rule to generate the source file. In a similar manner, libraries from other projects may be used through IMPORTED targets

add_library (foo IMPORTED)
set_property (TARGET foo PROPERTY
              IMPORTED_LOCATION "/path/to/libfoo.a")
add_executable (myexe src1.c src2.c)
target_link_libraries (myexe foo)

On Windows, a .dll and its .lib import library may be imported together:

add_library (bar IMPORTED)
set_property (TARGET bar PROPERTY
              IMPORTED_LOCATION "c:/path/to/bar.dll")
set_property (TARGET bar PROPERTY
              IMPORTED_IMPLIB "c:/path/to/bar.lib")
add_executable (myexe src1.c src2.c)
target_link_libraries (myexe bar)

A l ibrary with multiple configurations may be imported with a single target:

add_library (foo IMPORTED)
set_property (TARGET foo PROPERTY
              IMPORTED_LOCATION_RELEASE "c:/path/to/foo.lib")
set_property (TARGET foo PROPERTY
              IMPORTED_LOCATION_DEBUG "c:/path/to/foo_d.lib")
add_executable (myexe src1.c src2.c)
target_link_libraries (myexe foo)

The generated build system will link myexe to foo.lib when it is built in the release configuration, and foo_d.lib when built in the debug configuration.


Exporting Targets

Imported targets on their own are useful, but they still require the project that imports them to know the locations of the target files on disk. The real power of imported targets is when the project providing the target files also provides a file to help import them.

The install(TARGETS) and install(EXPORT) commands work together to install both a target and a CMake file to help import it. For example, the code

add_executable (generator generator.c)
install (TARGETS generator DESTINATION lib/myproj/generators
         EXPORT myproj-targets)
install (EXPORT myproj-targets DESTINATION lib/myproj)

will install the two files

<prefix>/lib/myproj/generators/generator
<prefix>/lib/myproj/myproj-targets.cmake

The first is the regular executable named generator. The second file, myproj-targets. cmake, is a CMake file designed to make it easy to import generator. This file contains code such as

get_filename_component (_self "${CMAKE_CURRENT_LIST_FILE}" PATH)
get_filename_component (PREFIX "${_self}/../.." ABSOLUTE)
add_executable (generator IMPORTED)
set_property (TARGET generator PROPERTY
  IMPORTED_LOCATION "${PREFIX}/lib/myproj/generators/generator")

(note that ${PREFIX} is computed relative to the file location). An outside project may now use generator as follows

include (${PREFIX}/lib/myproj/myproj-targets.cmake)  # 1
add_custom_command (OUTPUT generated.c
                    COMMAND generator generated.c) # 2
add_executable (myexe src1.c src2.c generated.c)

Line #1 loads the target import script (see section 0 to make this automatic). The script may import any number of targets. Their locations are computed relative to the script location so the install tree may be easily moved. Line #2 references the generator executable in a custom command. The resulting build system will run the executable from its installed location. Libraries may also be exported and imported

add_library (foo STATIC foo1.c)
install (TARGETS foo DESTINATION lib EXPORTS myproj-targets)
install (EXPORT myproj-targets DESTINATION lib/myproj)

This installs the library and an import file referencing it. Outside projects may simply write

include (${PREFIX}/lib/myproj/myproj-targets.cmake)
add_executable (myexe src1.c)
target_link_libraries (myexe foo)

and the executable will be linked to the library foo, exported, and installed by the original project.

Any number of target installations may be associated with the same export name. Export names are consid­ered global so any directory may contribute a target installation. Only the one for calling to the install(EXPORT) command is needed to install an import file that references all targets. Both of the examples above may be combined into a single export file, even if they are in different subdirectories of the project, as shown in the code below.

# A/CMakeLists.txt
add_executable (generator generator.c)
install (TARGETS generator DESTINATION lib/myproj/generators
         EXPORT myproj-targets)

# B/CMakeLists.txt
add_library (foo STATIC foo1.c)
install (TARGETS foo DESTINATION lib EXPORTS myproj-targets)

# Top CMakeLists.txt
add_subdirectory (A)
add_subdirectory (B)
install (EXPORT myproj-targets DESTINATION lib/myproj)

Typically projects are built and installed before being used by an outside project. However, in some cases, it is desirable to export targets directly from a build tree. The targets may then be used by an outside project that references the build tree with no installation involved. The export() (page 287) command is used to generate a file exporting targets from a project build tree. For example, the code

add_executable (generator generator.c)
export (TARGETS generator FILE myproj-exports.cmake)

will create a file in the project build tree called myproj-exports.cmake, which contains the required code to import the target. This file may be loaded by an outside project that is aware of the project build tree, in order to use the executable to generate a source file. An example application of this feature is for build­ing a generator executable on a host platform when cross-compiling. The project containing the generator executable may be built on the host platform and then the project that is being cross-compiled for another platform may load it.


Advanced Commands

There are a few commands that can be very useful, but are not typically used in writing CMakeLists files. This section will discuss a few of these commands and when they are useful. First, consider the add_dependencies() (page 273) command which creates a dependency between two targets. CMake automatically creates dependencies between targets when it can determine them. For example, CMake will automatically create a dependency for an executable target that depends on a library target. The add_dependencies command is typically used to specify inter target dependencies between targets where at least one of the targets is a custom target (see Chapter 6 for more information on custom targets).

The include_regular_expression() (page 316) command also relates to dependencies. This com­mand controls the regular expression that is used for tracing source code dependencies. By default, CMake will trace all the dependencies for a source file including system files such as stdio.h. If you specify a regular expression with the include_regular_expression command, that regular expression will be used to limit which include files are processed. For example; if your software project's include files all started with the prefix foo (e.g. fooMain.c fooStruct.h, etc), you could specify a regular expression of ^foo.*$ to limit the dependency checking to just the files of your project.

Occasionally you might want to get a listing of all the source files that another source file depends on. This is useful when you have a program that uses pieces of a large library, but are unsure which pieces it is using. The output_required_files() (page 349) command will take a source file and produce a list of all the other source files it depends on. You could then use this list to produce a reduced version of the library that only contains the necessary files for your program.

Some tools, such as Rational Purify on the Sun platform, are run by inserting an extra command before the final link step. So, instead of

CC foo.o -o foo

The link step would be

purify CC foo.o -o foo

It is possible to do this with CMake. To run an extra program in front of the link line, change the rule variables CMAKE_CXX_LINK_EXECUTABLE and CMAKE_C_LINK_EXECUTABLE. Rule variables are described in Chapter 11. The values for these variables are contained in the file Modules/CMakeDefaultMakeRuleVariables. cmake, and they are sometimes redefined in Modules/Platform/*.cmake. Make sure it is set after the PROJECT() (page 327) command in the CMakeLists file. Here is a small example of using purify to link a program called foo

project (foo)

set (CMAKE_CXX_LINK_EXECUTABLE
    "purify $(CMAKE_CXX_LINK_EXECUTABLE)")
add_executable (foo foo.cxx)

Of course, for a generic CMakeLists file, you should have some if checks for the correct platform. This will only work for the Makefile generators because the rule variables are not used by the IDE generators. Another option would be to use $(PURIFY) instead of plain purify. This would pass through CMake into the Makefile and be a make variable. The variable could be defined on the command line like this: make PURIFY=purify. If not specified then it would just use the regular rule for linking a C++ executable as PURIFY would be expanded by make to nothing.


Notes