CMake 2.6 Notes: Difference between revisions

From KitwarePublic
Jump to navigationJump to search
No edit summary
(Replace content with link to new CMake community wiki)
 
(2 intermediate revisions by one other user not shown)
Line 1: Line 1:
__TOC__
{{CMake/Template/Moved}}


This page documents some of the changes and new features available in CMake 2.6.
This page has moved [https://gitlab.kitware.com/cmake/community/wikis/doc/cmake/notes/2.6 here].
 
=Exporting and Importing Targets=
 
Please see our tutorial on [[CMake/Tutorials/Exporting_and_Importing_Targets|Exporting and Importing Targets]].
 
=Packages=
 
Please see our tutorial on [[CMake/Tutorials/Packaging|Packaging]].
 
=Preprocessor Definitions=
 
Preprocessor definitions may now be added to builds with much finer granularity than in previous versions of CMake.  There is a new property called <code>COMPILE_DEFINITIONS</code> that is defined directories, targets, and source files.  For example, the code
 
  add_library(mylib src1.c src2.c)
  add_executable(myexe main1.c)
 
  set_property(
    DIRECTORY
    PROPERTY COMPILE_DEFINITIONS A AV=1
    )
  set_property(
    TARGET mylib
    PROPERTY COMPILE_DEFINITIONS B BV=2
    )
  set_property(
    SOURCE src1.c
    PROPERTY COMPILE_DEFINITIONS C CV=3
    )
 
will build the source files with these definitions:
 
  src1.c:  -DA -DAV=1 -DB -DBV=2 -DC -DCV=3
  src2.c:  -DA -DAV=1 -DB -DBV=2
  main2.c:  -DA -DAV=1
 
When the <code>add_definitions</code> command is called with flags like "<code>-DX</code>" the definitions are extracted and added to the current directory's <code>COMPILE_DEFINITIONS</code> property.  When a new subdirectory is created with <code>add_subdirectory</code> the current state of the directory-level property is used to initialize the same property in the subdirectory.
 
Note in the above example that the <code>set_property</code> command will actually '''set''' the property and replace any existing value.  The command provides the <code>APPEND</code> option to help add more definitions without removing existing ones.  For example, the code
 
  set_property(
    SOURCE src1.c
    APPEND PROPERTY COMPILE_DEFINITIONS D DV=4
    )
 
will add the definitions "<code>-DD -DDV=4</code>" when building <code>src1.c</code>.
 
Definitions may also be added on a per-configuration basis using the <code>COMPILE_DEFINITIONS_<CONFIG></code> property.  For example, the code
 
  set_property(
    TARGET mylib
    PROPERTY COMPILE_DEFINITIONS_DEBUG MYLIB_DEBUG_MODE
    )
 
will build sources in mylib with <code>-DMYLIB_DEBUG_MODE</code> only when compiling in a <code>Debug</code> configuration.
 
=Link Line Generation=
 
CMake 2.6 implements a new approach to generating link lines for targets.
 
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 OS X) do not
support the <code>-Bstatic</code> 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 side-effects that affect compatibility with existing projects (documented in the next two subsections).
 
==Missing Linker Search Directories==
 
Projects used to be able to write this (wrong) code and it would work by accident:
 
  add_executable(myexe myexe.c)
  target_link_libraries(myexe /path/to/libA.so B)
 
where "<code>B</code>" is meant to link "<code>/path/to/libB.so</code>".  This code is incorrect
because it asks CMake to link to <code>B</code> but does not provide the proper
linker search path for it.  It used to work by accident because the
<code>-L/path/to</code> would get added as part of the implementation of linking to
A.  The correct code would be
 
  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)
 
In order to support projects that have this bug, we've added a
compatibility feature that adds the "<code>-L/path/to</code>" paths for all libraries
linked with full paths even though the linker will not need those paths
to find the main libraries.  See policy CMP0003 for details.
 
==Linking to System Libraries==
 
System libraries on UNIX-like systems are typically provided in <code>/usr/lib</code> or <code>/lib</code>.  These directories are considered implicit linker search paths because linkers automatically search these locations even without a flag like <code>-L/usr/lib</code>.  Consider the code
 
  find_library(M_LIB m)
  target_link_libraries(myexe ${M_LIB})
 
Typically the <code>find_library</code> command would find the math library
 
  /usr/lib/libm.so
 
Some platforms provide multiple versions of libraries corresponding to different architectures.  For example, on an IRIX machine 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 <code>find_library</code> may not know about all the architecture-specific system search paths used by the linker.  In fact when it finds <code>/usr/lib/libm.so</code> it may be finding a library of incorrect architecture.  If the link computation were to produce the line
 
  ... /usr/lib/libm.so ...
 
the linker might complain if <code>/usr/lib/libm.so</code> does not match the architecture it wants.
 
One solution to this problem is for the link computation to 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 library.  Unfortunately this solution suffers from the original problem of distinguishing between static and shared versions:
 
  /usr/lib/libm.a
  /usr/lib/libm.so
 
In order to ask the linker to find the static system library of the correct architecture it must produce the link line
 
  ... -Wl,-Bstatic -lm ... -Wl,-Bshared ...
 
This solution directly contradicts the original motivation to give the linker paths to libraries instead of <code>-l</code> options: not all platforms have an option like <code>-Bstatic</code>.
Fortunately the platforms that do not provide such flags also tend to not have architecture-specific implicit link directories.
 
The solution used by CMake is:
 
* Libraries not in implicit system locations are linked by passing the file path to the linker
* Libraries in implicit system locations are linked by
** passing the <code>-l</code> option if a flag like <code>-Bstatic</code> is available
** passing the file path to the linker otherwise
 
Users can override this behavior by using the IMPORTED targets feature:
 
  add_library(math STATIC IMPORTED)
  set_property(TARGET math PROPERTY IMPORTED_LOCATION /usr/lib/libm.a)
  add_executable(foo foo.c)
  target_link_libraries(foo math) # will link using full path
 
=CMake Policy Mechanism=
 
CMake 2.6 introduces a new mechanism for backwards compatibility support.  See [[CMake/Policies]] for more information.
 
{{CMake/Template/Footer}}

Latest revision as of 15:41, 30 April 2018


The CMake community Wiki has moved to the Kitware GitLab Instance.

This page has moved here.