In SharedForward, the call to execvp warned on MinGW because the
signature declared in process.h has an extra const. We use an explicit
cast to convert the pointer type.
To detect when the launcher is running from the build tree we now test
if the directory containing it is the same as the build-tree directory
using an inode test instead of string comparison. This makes it more
robust on case-insensitive filesystems and other quirky situations.
If the arguments to a command fail to parse correctly due to a syntax
error, the command should not be invoked. This avoids problems created
by processing of commands with bad arguments. Even though the build
system will not be generated, the command may affect files on disk that
persist across CMake runs.
We now properly report the source location of command arguments inside
macros instead of using the macro invocation site. No information is
lost because full call-stack information is already reported.
Previously error messages produced by parsing of command argument
variable references, such as bad $KEY{VAR} syntax or a bad escape
sequence, did not provide good context information. Errors parsing
arguments inside macro invocations gave no context at all. Furthermore,
some errors such as a missing close curly "${VAR" would be reported but
build files would still be generated.
These changes teach CMake to report errors with good context information
for all command argument parsing problems. Policy CMP0010 is introduced
so that existing projects that built despite such errors will continue
to work.
Automatic generation of 64-bit library search paths must preserve
trailing slashes. This fixes a failure case exposed by the recent
rewrite of find_library, which assumes trailing slashes occur on all
search paths.
OpenBSD shared libraries use a ".so.<major>.<minor>" extension and do
not have a symlink with just a ".so" extension. Its "ld" is capable of
finding the library with the best version. This change adds support for
finding such libraries. See issue #3470.
Previously we searched for library files by enumerating every possible
combination of prefix and suffix. Now we load (and cache) directory
content from disk and search for matching file names. This should
reduce disk access. It will also allow more advanced matching rules in
the future. See issue #3470.
Previously the cmGlobalGenerator::GetDirectoryContent method would work
safely only during build system generation. These changes make it safe
to use during each configure step by flushing it at the beginning.
When looking for NOTFOUND libraries, use the direct dependencies of a
target instead of all dependencies. At least one target will trigger
the NOTFOUND error anyway because at least one must directly link it.
This removes another use of the old-style link line computation.
In some cases it may be useful to compute a "link" line for a static
library even though it will not be put in the generated build system.
This removes the assertion which previously diallowed the case.
In cmComputeLinkInformation items in the final link line returned by
GetItems now contain a pointer to their corresponding cmTarget if they
were produced by a target. This makes available the set of all targets
linked.
In cmGlobalGenerator we use cmComputeTargetDepends to construct a safe,
non-circular set of inter-target dependencies. This change enables use
of the results by the Xcode generator. It also removes a lot of old
code and another use of the old-style linking logic. See issue #7652.
Provide VERSION_LESS, VERSION_EQUAL, and VERSION_GREATER operators in
the if() command. This simplifies component-wise comparison of version
numbers in the form "major[.minor[.patch[.tweak]]]".
Make the number of version components specified explicitly available.
Set variables for unspecified version components to "0" instead of
leaving them unset. This simplifies version number handling for find-
and config-modules. Also support a fourth "tweak" version component
since some packages use them.
Use the new-style error reporting mechanism to provide more context
information for a find_package call with a bad package name. When the
package is not required, issue a warning instead of an error.
The set_property command unsets a property if it is given no value. In
the case of GLOBAL properties, the cmake::SetProperty method would
replace a NULL value with "NOTFOUND". Instead it should be left as NULL
so that the property is unset as expected. Once it is unset the
get_cmake_property command will still report NOTFOUND while the
get_property command will return the empty string as documented.
When the -C or --build-config option is used to specify the
configuration to be tested by CTest, do not override it with the
configuration in which CTest is built.
A Mac OS X Framework should provide a Resources/Info.plist file
containing meta-data about the framework. This change generates a
default Info.plist for frameworks and provides an interface for users to
customize it.
This change cleans up the implementation of cmXCodeObject to avoid
un-escaping and re-escaping string values. There is no need to store
the string in escaped form. It can be escaped once when it is printed
out to the generated project file.
The recent fix to avoid including flags in dependency inferral also
dropped them from chaining of dependencies through targets. This fix
restores chaining of flags through known dependency lists while still
leaving them out of inferred dependency lists.
This change introduces a new algorithm for link line construction. The
order it computes always begins with the exact link line specified by
the user. Dependencies of items specified by the user are tracked, and
those that are not already satisified by the line are appended to it at
the end with minimal repeats. This restores the behavior of CMake 2.4
and below while still fixing some of its bugs. See issue #7546.
In cmComputeLinkDepends link items that look like flags (starting in
'-') should not be included in dependency inferral. They are not
libraries and therefore have no dependencies. They should just be
passed through to the final link line unchanged. See issue #7546.
This introduces the unset() command to make it easy to unset CMake
variables, environment variables, and CMake cache variables. Previously
it was not even possible to unset ENV or CACHE variables (as in
completely remove them). Changes based on patch from Philip Lowman.
See issue #7507.
Some native build tools, particularly those for cross compiling, may
have a limit on the length of the full path to an object file name that
is lower than the platform otherwise supports. This change allows the
limit to be set by the project toolchain file through the variable
CMAKE_OBJECT_PATH_MAX.
It is useful to be able to test if a target has been created. Often
targets are created only inside conditions. Rather than storing the
result of the condition manually for testing by other parts of the
project, it is much easier for the other parts to just test for the
target's existence. This will also be useful when find-modules start
reporting results with IMPORTED targets and projects want to test if a
certain target is available.
According to "man select" on Linux it is possible that select() lies
about data being ready on a pipe in some subtle cases. We deal with
this by switching to non-blocking i/o and checking for EAGAIN. See
issue #7180.
The LINK_INTERFACE_LIBRARIES target property may not contain the
"debug", "optimized", or "general" keywords. These keywords are
supported only by the target_link_libraries (and link_libraries) command
and are not a generic library list feature in CMake. When a user
attempts to add one of these keywords to the property value, we now
produce an error message that refers users to alternative means.
The LINK_INTERFACE_LIBRARIES property does not apply for STATIC
libraries. The IMPORTED_LINK_INTERFACE_LIBRARIES property does apply
for STATIC libraries. State both explicitly in the documentation.
Also, clarify that the per-configuration version of these properties
completely overrids the generic version.
The compatibility check to allow linking to modules should test for
CMake 2.2, not the unreleased 2.3. See issue #7500. Furthermore, the
message should be more clear about fixing the code instead of setting
CMAKE_BACKWARDS_COMPATIBILITY unless one is just trying to build an
existing project.
In the future some policies may be set to REQUIRED_IF_USED or
REQUIRED_ALWAYS. This change clarifies the error messages users receive
when violating the requirements.
When creating an IMPORTED target for a library that has been found on
disk, it may not be known whether the library is STATIC or SHARED.
However, the library may still be linked using the file found from disk.
Use of an IMPORTED target is still important to allow per-configuration
files to be specified for the library.
This change creates an UNKNOWN type for IMPORTED library targets. The
IMPORTED_LOCATION property (and its per-config equivalents) specifies
the location of the library. CMake makes no assumptions about the
library that cannot be inferred from the file on disk. This will help
projects and find-modules import targets found on disk or specified by
the user.
In switch statements that deal with only a few target types, use a
'default' case for the remaining target types instead of listing them
explicitly. This will make it easier to add more types in the future.
Rename the recently added INTERFACE mode of the target_link_libraries()
command to LINK_INTERFACE_LIBRARIES. This makes it much more distinct
from a normal call to the command, and clearly states its connection to
the property of the same name. Also require the option to appear
immediately after the target name to make it a mode rather than an
option.
It is likely that projects or CMake modules in the future will need to
check the value of a policy setting. For example, if we add a policy
that affects the results of FindXYZ.cmake modules, the module code will
need to be able to check the policy.
During installation the RPATH and RUNPATH entries of ELF binaries are
edited to match the user specification. Usually either one entry is
present or both entries refer to the same string literal. In the case
that they are both present and refer to separate string literals we need
to update both. I have never seen this case in practice, but we should
do this just in case.
Removal of the RPATH and RUNPATH from ELF binaries must work when both
entries are present. Both entries should be removed. Previously only
one would be removed and the other would be blanked because it pointed
at the same string which was zeroed. This fixes gentoo bug number
224901.
Create an INTERFACE option to the target_link_libraries command to help
set the LINK_INTERFACE_LIBRARIES and LINK_INTERFACE_LIBRARIES_DEBUG
properties. This will help users specify link interfaces using
variables from Find*.cmake modules that include the 'debug' and
'optimized' keywords.
The "debug", "optimized", and "general" link library type specifier
arguments to the target_link_library commands are sometimes repeated in
user code due to variable expansion and other complications. Instead of
silently accepting the duplicates and trying to link to a bogus library
like "optimized.lib", warn and ignore the earlier specifiers.
The add_subdirectory() command's EXCLUDE_FROM_ALL option does not
override inter-target dependencies. This change clarifies the
documentation accordingly.
When an executable target within the project is named in
target_link_libraries for another target, but the executable does not
have the ENABLE_EXPORTS property set, then the executable cannot really
be linked. This is probably a case where the user intends to link to a
third-party library that happens to have the same name as an executable
target in the project (or else will get an error at build time). We
need to avoid making the other target depend on the executable target
incorrectly, since the executable may actually want to link to that
target and this is not a circular depenency.
After reporting an error about circular target dependencies do not try
to continue generation because the dependency computation object is not
in a useful state.
Custom command dependencies that are not full paths or targets may also
match source files. When one does, the full information about the
source file's location and name may be used. This fixes the case when a
custom commands depends by relative path on a source file generated by
another custom command specifying its output by relative path.
A name with an ambiguous extension may only match an unambiguous name
that is extended by one of the fixed set of extensions tried when
finding the source file on disk. This rule makes matching of source
files with ambiguous extensions much less aggressive but still
sufficient.
When generating RPATH entries on the link line using a repeated linker
flag (-R ... -R ... style) do not convert individual entries to a full
path. We need to preserve what the user requested.
When generating escape sequences for the native build tool do not put in
Makefile escapes for paths generated into link scripts. This fixes
putting "$ORIGIN" into the RPATH, and probably some other subtle
problems.
Creation of archive libraries with the unix 'ar' tool should be done
incrementally when the number of object files is large. This avoids
problems with the command line getting too many arguments.
When attempting to load the RPATH out of a non-ELF file cmELF would
crash because the check for a valid file was done with in correct
operator precedence. See bug#7392.
For historical reasons we still support naming of source files without
their extension. Sources without known extensions are located on disk
by iterating through a fixed set of possible extensions. We now want
users to always specify the extension, so the fixed set will not be
expanded and is preserved for compatibility with older projects.
This change adds recognition of extensions of all enabled languages to
avoid checking the disk for files whose extensions are unambiguous but
not in the original fixed set.
In CMake 2.4 the generated link line for a target always preserved the
originally specified libraries in their original order. Dependencies
were satisfied by inserting extra libraries into the line, though it had
some bugs. In CMake 2.6.0 we preserved only the items on the link line
that are not known to be shared libraries. This reduced excess
libraries on the link line. However, since we link to system libraries
(such as /usr/lib/libm.so) by asking the linker to search (-lm), some
linkers secretly replace the library with a static library in another
implicit search directory (developers can override this by using an
imported target to force linking by full path). When this happens the
order still matters.
To avoid this and other potential subtle issues this commit restores
preservation of all non-target items and static library targets. This
will create cases of unnecessary, duplicate shared libraries on the link
line if the user specifies them, but at least it will work. In the
future we can attempt a more advanced analysis to safely remove
duplicate shared libraries from the link line.
We preserve the order and multiplicity of libraries directly linked by a
target as specified by the user. Items known to be shared libraries may
be safely skipped because order preservation is only needed for static
libraries. However, CMake 2.4 did not skip shared libs, so we do the
same when in 2.4 compatibility mode.
We never explicitly specify system library directories in linker or
runtime search paths. Furthermore, libraries in these directories are
always linked by asking the linker to search for them. We need to
generate a warning when explicitly specified search directories contain
files that may hide the system libraries during the search.
This change introduces policy CMP0008 to decide how to treat full path
libraries that do not appear to be valid library file names. Such
libraries worked by accident in the VS IDE and Xcode generators with
CMake 2.4 and below. We support them in CMake 2.6 by introducing this
policy. See policy documentation added by this change for details.
Sometimes we ask the linker to search for a library for which the path
is known but for some reason cannot be specified by full path. In these
cases do not include the library in CMP0003 warnings because we know the
extra paths are not needed for it.
- This case worked accidentally in CMake 2.4, though not in Makefiles.
- Some projects build only with the VS IDE on windows and have this
mistake.
- Support them when 2.4 compatibility is enabled by adding the extension.
CMAKE[_SYSTEM]_(LIBRARY|PROGRAM|INCLUDE|PREFIX)_PATH variables
-moved CMAKE_CROSSCOMPILING from "Variables that modify behaviour" to
"variables that Provide Information", since it should be used only for
testing whether we are currently in cross compiling mode, not for switching
between the modes.
Alex
- The source-file signature of try_compile looks up the language
of the source file using the extension-to-language map so that
it knows what language to enable in the generated project.
- This map needs to be filled before loading a file specified by
CMAKE_USER_MAKE_RULES_OVERRIDE
CMAKE_USER_MAKE_RULES_OVERRIDE_<LANG>
so that the user file may call the try_compile() source-file
signature.
- It must still be re-filled after loading CMake<LANG>Information.cmake
in case the compiler- or platform-specific files added anything.
- See bug #7340.
Details:
==========
- New cpack_add_component, cpack_add_component_group, and
cpack_add_install_type "commands" defined as macros in the CPack
module.
- Documentation for all of the variables and commands in the CPack module.
- Added get_cmake_property(... COMPONENTS) to CMake to ask for the
names of all components. Used in the CPack module to automatically
build component-based installers. (Set CPACK_MONOLITHIC_INSTALL to
turn off component-based installation).
- A group can declare its PARENT_GROUP, to build an arbitrary
hierarchy of groups.
- New CPack command cpack_configure_downloads, which creates an
installer that downloads only the selected components on-the-fly.
Those components marked DOWNLOADED will be separate packages
downloaded on-the-fly (or, all packages can be marked as such with the
ALL option to cpack_configure_downloads). Individual components are
compressed with ZIP at installer-creation time and
downloaded/uncompressed by the installer as needed. This feature is
only available on Windows with NSIS at the moment.
- NSIS installers can install themselves and enable the "Change"
button in Add/Remove programs, allowing users to go back and install
or remove components. This can be disabled through
cpack_configure_downloads, because it's only really useful is most of
the application's functionality is in downloaded components.
- Bug fix: automatically install everything whose COMPONENT was not
specified (it's a hidden, required group)
- Bug fix: fixed removal of components when re-running the NSIS
installer and unchecking components
- Bug fix: NSIS installers now only install/remove the minimal
number of files when re-run to update the installation (or by clicking
"Change" in Add/Remove programs)
- The Info.plist file in app bundles should not be built.
- User-specified files such as foo.txt should not be built.
- Only files with a recognized language should be built,
just as in the Makefiles generators.
- See bug #7277.
- We used to always put LIBRARY_OUTPUT_PATH and EXECUTABLE_OUTPUT_PATH
in the cache if the project did not.
- In CMake 2.6 these variables should no longer be used.
- Now add them only if CMAKE_BACKWARDS_COMPATIBILITY is also cached.
- This happens only when CMP0001 is set to OLD or WARN or if
the user or project sets it. In any case compatibility is needed.
- Reported by Miguel A. Figueroa-Villanueva and Philip Lowman.
- Generated Xcode projects for application bundles list the
CMake-generated Info.plist input file as a resource.
- The location of the input file was moved by a previous commit,
but the reference to it as a resource file was not updated.
- This change moves the file to CMakeFiles/<tgt>.dir/Info.plist
to give it a more intuitive name in the Xcode project.
- We also update the reference to point at the correct location.
- See bug #7277.
- The Xcode generator creates one Info.plist input file which is
converted at build time by Xcode and placed in the final bundle.
- The <CONFIG>_OUTPUT_NAME target property can place different content
for the exe name in Info.plist on a per-configuration basis.
- Instead of generating a per-config Info.plist input file just let
Xcode put the name in at build time using the $(EXECUTABLE_NAME) var.
- The property tracks the value formed by add_definitions
and remove_definitions command invocations.
- The string should be maintained for use in returning for the
DEFINITIONS property value.
- It is no longer used for any other purpose.
- The DEFINITIONS property was recently documented as deprecated.
- See bug #7239.
- Fix documentation of get_directory_property command.
- Convert its list of computed directory properties to be
defined/documented directory properties.
ENH: -if no compiler has been found, don't test it, and also remove the compiler
information file again. This makes optionally enabling a language work
better.
Alex
- The short-hand forms do not document the NO_* options.
- CMake 2.4 and 2.6.0 accepted them accidentally, but also
treated the options as paths.
- Now the options are accepted but do not become paths.
- The CMAKE_PREFIX_PATH and similar variables have both
environment and CMake cache versions.
- Previously the environment value was checked before the
cache value.
- Now the cache value is favored because it is more specific.
- Hints are searched after user locations but before system locations
- The HINTS option should have paths provided by system introspection
- The PATHS option should have paths that are hard-coded guesses
- Locating a header inside a framework often requires globbing
- Previously the glob was <dir>/*/Headers/<name>
- Now the glob is <dir>/*.framework/Headers/<name>
- This is much faster when <dir> is not really a framework dir
- CMAKE_FIND_FRAMEWORK and CMAKE_FIND_APPBUNDLE are supposed to specify
whether to find frameworks/appbundles FIRST, LAST, ONLY, or NEVER.
- Previously this affected only the placement of CMAKE_FRAMEWORK_PATH
and CMAKE_APPBUNDLE_PATH with respect to the other path specifiers.
- Now it behaves as documented. The entire search path is inspected for
each kind of program, library, or header before trying the next kind.
- Additionally the ONLY mode is now honored for headers so that users
do not end up with a library in framework and a header from elsewhere.
- In cmFindBase when CheckCommonArgument returns true, set newStyle
- Otherwise if there are no PATHS then the ancient-style compatibility
mode is enabled and the common argument is treated as a path.
- Add each part of the search order in a separate method.
- Collect added paths in an ivar in cmFindCommon.
- Move user path storage up to cmFindCommon and share
between cmFindBase and cmFindPackageCommand.
- Expand user path registry values up in cmFindCommon
- Enables 32-/64-bit registry view for find_package
- Disables registry expansion for paths not specified
with the PATHS argument, which is not expected.
- cmFindBase should search both 32-bit and 64-bit registry views
for FIND_PROGRAM even if CMAKE_SIZEOF_VOID_P is not set.
- Needed because the variable is not available when CMAKE_MAKE_PROGRAM
is to be found.
- The rule hash should use only commands specified by the user.
- No make output (echo and progress) rules should be included.
- No outputs or dependencies need be included. The native build tool
will take care of them.
- Option was recently added but never released.
- Custom commands no longer depend on build.make so we do
not need the option.
- Rule hashes now take care of rebuilding when rules change
so the dependency is not needed.
- In CMake 2.4 custom commands would not rebuild when rules changed.
- In CMake 2.6.0 custom commands have a dependency on build.make
which causes them to rebuild when changed, but also when any
source is added or removed. This is too often.
- We cannot have a per-rule file because Windows filesystems
do not deal well with lots of small files.
- Instead we add a persistent CMakeFiles/CMakeRuleHashes.txt file
at the top of the build tree that is updated during each
CMake Generate step. It records a hash of the build rule for
each file to be built. When the hash changes the file is
removed so that it will be rebuilt.
- The target platform does not matter for finding executables
so find_program should expand to both 32-bit and 64-bit registry
values.
- See issue #7095.
- Add an argument to registry read/write/delete methods to specify
a 32-bit or 64-bit view.
- Default is the bit-ness of the running program.
- See issue #7095.
- Revert previous change to trust user-provided full paths.
- Instead trust them only far enough to determine the source language
but still check for existence for non-generated sources.
- Create MACOSX_BUNDLE_INFO_PLIST target property to specify template.
- Look for MacOSXBundleInfo.plist.in in CMAKE_MODULE_PATH by default.
- See issue #6983.
the generator on the first configure. It either ask for a toolchain file
or asks for all the information a toolchain file might contain.
Also added option for setting non-default compilers if not cross compiling.
Fixes#6849.
Also a bit of code cleanup and re-organizing.
- Tells CMake about the KWSYS_HEADER macro.
- Enables implicit dependencies of private source files.
- When a CMake new enough to support the property is required
the "#if 0" hack can be removed from the source files.
- Allows make rules to be created with no dependencies.
- Such rules will not re-run even if the commands themselves change.
- Useful to create rules that run only if the output is missing.
-generate and install the policy documentation files
-generate and install the docbook files for cmake, ctest, cpack, ccmake (cmake-gui not yet ?)
Alex
- Previously this was done implicitly by the check for a target
link language which checked all source full paths.
- The recent change to support computing a link language without
finding all the source files skipped the implicit check.
- This change adds an explicit check to find all source files.
the argument after -Wno-dev was skipped, which happened to be the source
directory, and so the current working directory was assumed as source
directory, although it was the build directory (maybe this didn't have an
effect if there was already a CMakeCache.txt in the build dir)
Alex
- Previously all links started in 'command_' which led to conflicts
and was confusing for non-command items.
- Use a per-section name that is meaningful to humans.
- Fix link id names to be valid HTML.
- CMake 1.8 and below did not do the check but could get in
infinite loops due to the local generate step.
- CMake 2.0 added the check but failed to perform it in directories
with no targets (see bug #678).
- CMake 2.2 removed the local generate which fixed the problem but
did not remove the check.
- Between CMake 2.4 and 2.6.0rc6 the check was fixed to work even
when no targets appear in a directory (see bug #6923).
- Bottom line: the check is no longer needed.
to help with debugging later on. The same should be done in the import
target files (but I didn't have time to do it yet).
STYLE: fix line length in cmListFileCache.cxx
Alex
- If new RPATH is empty then remove the entry completely
- Preserve file modification time so installation is not repeated
- If installed file already exists remove it if its RPATH
does not match that expected
- Place the built library in foo.framework/Versions/A/foo
- Do not create unused content symlinks (like PrivateHeaders)
- Do not use VERSION/SOVERSION properties for frameworks
- Make cmTarget::GetDirectory return by value
- Remove the foo.framework part from cmTarget::GetDirectory
- Correct install_name construction and conversion on install
- Fix MACOSX_PACKAGE_LOCATION under Xcode to use the
Versions/<version> directory for frameworks
- Update the Framework test to try these things
- Message for missing cmake_minimum_required is not issued
until the end of processing the top CMakeLists.txt file
- During processing a cmake_policy command may set behavior
- OLD behavior is to silently ignore the problem
- NEW behavior is to issue an error instead of a warning
- CMake 2.4 added link directories for targets linked
in the optimized configuration even when building debug
- Old behavior for policy CMP0003 must account for this
RUNTIME, ARCHIVE, LIBRARY, etc.) arguments are given.
If the component of a part of a target is queried, return the specific one,
if a specific one hasn't been set, return the generic one, if that hasn't
been set, return "Unspecified".
Alex
- In cmake_minimum_required do not set policy version if current
CMake is too old
- In cmPolicies::ApplyPolicyVersion report error if version is too
new or cannot be parsed
- Give example code to avoid the warning
- Make explanation more consise
- Explicitly state this is for compatibility
- Issue the warning for at most one target
- This will help projects support multiple CMake versions.
- In order to set a policy when using a newer CMake but still
working with an older CMake one may write
if(POLICY CMP1234)
cmake_policy(SET CMP1234 NEW)
endif(POLICY CMP1234)
- Note that since CMake 2.4 does not have if(POLICY) supporting
it will also require using "if(COMMAND cmake_policy)"
- Update policy CMP0000 to require use of the command
cmake_minimum_required and not cmake_policy
so there is only one way to avoid it.
- Explicitly specify the line users should add.
- Reference policy CMP0000 only at the end.
- Fix policy CMP0000 documentation to not suggest
use of the cmake_policy command.
- Add a paragraph introducing the policy mechanism
- Explicitly introduce the CMP<NNNN>, OLD, and NEW notation
- Note that setting policies by CMake version is preferred
- Fix SET signature to use CMP<NNNN> notation
- Add more details about the policy stack
Bill Hoffman
Brad King
David Cole
Alexander Neundorf
Clinton Stimpson
Ken Martin
Francois Bertel
Mathieu Malaterre
Bill Lorensen
Jeffrey Baumes
Sebastien Barre