38c762c Merge 'remove-CTestTest3' into ctest-file-checksum
46df0b4 Activate retry code on any curl submit failure.
8705497 Checksum test should use CMAKE_TESTS_CDASH_SERVER
d0d1cdd Mock checksum failure output for old CDash versions
af5ef0c Testing for CTest checksum
86e81b5 CTest should resubmit in the checksum failed case
d6b7107 Fix subscript out of range crash
082c87e Cross-platform fixes for checksum/retry code
e525649 Checksums on CTest submit files, and retry timed out submissions.
Previously, we were only going into the retry block
for time out conditions. But a "could not connect"
response, or really any sort of curl failure, is
also a condition where we should retry the submit
if the user has requested a retry.
Fix niggly in ExternalProject that would inadvertently
create a log file of output when a command was explicitly
set to the empty string.
Also, increase the default value for showing context
around build errors to 10 lines, up from 6. Also add
the ability to customize this setting so that users
can put it back to 6 or even lower if they want, or
show more if they want.
Clang is smart enough to recognize that "*(int*)0=0" invokes undefined
behavior, warns, and produces an "undefined instruction". The message
in commit "Avoid Clang optimizer bug in testProcess-[45]" (2010-07-02)
is incorrect; this is not a Clang bug. It really is undefined
behavior. Use "*(int*)1=0" to produce a crash instead.
Clang's optimizer, as of clang version 2.8 (trunk 107463), produces the
undefined instruction 'ud2' for the code "*(int*)0=0" on OS X x86_64.
It causes our crash tests to fail because the child process exits with
an invalid instruction instead of a segmentation fault. Work around the
bug by using "*(int*)1=0" in this case.
Add option KWSYS_TEST_BOGUS_FAILURES that can be set by a containing
project or in the CMake cache to list tests known to fail consistently
on a buggy system.
Commit ff1f8d0b (Fix or cast more integer conversions in cmake) changed
a member type from int to size_t. Update the types of variables
compared to these values to be unsigned also.
Both possible result values need to be convertible to the same type.
Some compilers fail to recognize that they can construct std::string
from the empty string literal, so state it explicitly.
Since commit "Provide unix-sytle command line parsing" (2009-07-13) the
reference to KWSYS_C_HAS_PTRDIFF_T in System.c has been meaningless
because the macro was never passed to the compiler!
Separation by ',' only works in VS 2008 and below and does not work in
the PlayStation3 VS plugin. Separation by ';' works in VS 10 and all
prior versions.
Change types of local variables, or casting, or re-arrange
expressions to get rid of "conversion may alter value" warnings
as seen on recent dashboard submissions from londinium.kitware.
I naively assumed in my previous commit that the Convert call
would correctly convert a relative path file name correctly
relative to the makefile's current output directory. It actually
converts it relative to the process's current working directory.
So it would be different depending on how you launched cmake-gui.
This commit ensures that the generated files are always the same
by starting with a full path to begin with, based on the makefile
GetCurrentOutputDirectory method.
Git's diff-tree format has no '\n'-terminated blank line at the end of
its commit message body block if there are no diff lines. Instead the
message body is terminated by '\0' and there is no diff section. Teach
CTest to parse the format in this case.
This is especially important for the Visual Studio 10
generator and its quirky current working directory
behavior.
Also, emit more information about exactly what files are
out of date when cmakeCheckStampFile returns false.
Commit 02f7cfbc (Need to remove the MAKEFLAGS when cmake starts,
2003-05-13) removed the MAKEFLAGS environment variable when CMake starts
to prevent try_compile() from inheriting "make -i". This is unnecessary
in script mode (cmake -P). Instead remove the variable only when
configuring a project.
Tru64's make(1) resolves relative paths in "include" directives with
respect to the includer. This is inconsistent with all other known make
tools. Note that this make tool treats the path literally so we cannot
use our standard FULL path code which escapes spaces. Instead qualify
the paths with $(CMAKE_BINARY_DIR) to avoid the problem.
GCC places the vtable in the object implementing the first non-pure,
non-inline virtual method. Since the symbol is not weak on Tru64, make
the location unique by putting the destructor in a single object file.
The DynamicLoader::LibPrefix and DynamicLoader::LibExtension methods
previously hard-coded the module name components for each platform. Set
them from the CMAKE_SHARED_MODULE_PREFIX and CMAKE_SHARED_MODULE_SUFFIX
CMake variables instead. This ensures consistency in a program that
uses these methods to construct the file names for its own modules.
Use 'git fetch' followed by 'git reset' to update the source tree. This
is better than 'git pull' because it can handle a rewritten upstream
branch and does not leave local modifications. After fetch, parse
FETCH_HEAD to find the merge head that 'git pull' would choose to track
the upstream branch. Then reset to the selected head.
In the normal fast-forward case the behavior remains unchanged.
However, now local modifications and commits will be erased, and
upstream rewrites are handled smoothly. This ensures that the upstream
branch is tested as expected.
Commit 67277bac (Teach ctest_update about Git submodules, 2010-05-04)
accidentally logged "git submodule update" with the prefixes "pull-out"
and "pull-err". Fix it to use "submodule-out" and "submodule-err"
instead.
The CopyFileIfDifferent, CopyFileAlways, CopyAFile and CopyADirectory
methods should always copy permissions. The special cases in which a
caller would pass copyPermissions=false should be handled at the call
site. The parameter needlessly complicates the interface and semantics
of these methods.
If there is a .bat or .cmd file used as a custom command
then the Borland Makefiles generator (specifically) requires
using the "call " syntax before the name of the .bat or .cmd
file. This fix applies to all Makefile based generators where
WindowsShell is true.
On Windows 7 the file size reported by 'stat' on a new file sometimes
reports zero even though the real size is correct. This causes our
CopyFileAlways method to falsely detect copy failure. Work around the
problem by trusting the state of ofstream after writing the file.
Commit 0fafdb7e (Do not copy permissions of files when making the copy
in an install rule, 2008-12-18) added special behavior to KWSys file
copy methods for this special case. Use a local solution to avoid use
of the special behavior so it can be removed later.
Commit 8d0161c8 (Trust umask for file permissions, 2010-01-12) taught
these commands to set permissions to 0666 explicitly. The intention was
to let the open() call inside ofstream handle permsisions so that umask
would be honored. Now we set permissions only when we need to preserve
those on an existing file. New files will be created with umask-based
permissions.
The LINK_FLAGS property is defined only for targets that really link.
These include executables and shared libraries. For static libraries we
define the STATIC_LIBRARY_FLAGS property. Teach the Xcode generator to
make this distinction.
Improve FILE(DOWNLOAD ...):
- Add percent complete progress output to the FILE DOWNLOAD
command. This progress output is off by default to
preserve existing behavior. To turn it on, pass
SHOW_PROGRESS as an argument.
- Add EXPECTED_MD5 argument. Verify that the downloaded
file has the expected md5 sum after download is complete.
- Add documentation for SHOW_PROGRESS and EXPECTED_MD5.
When the destination file exists already and has the
expected md5 sum, then do not bother re-downloading
the file. ("Short circuit" return.)
Also, add a test that checks for the status output
indicating that the short circuit behavior is actually
occurring. Use a binary file for the test so that the
md5 sum is guaranteed to be the same on all platforms
regardless of "shifting text file line ending" issues.
Improve ExternalProject:
- Add argument URL_MD5.
- Add verify step that compares md5 sum of .tar.gz file
before extracting it.
- Add md5 check to download step, too, to prevent
unnecessary downloads.
- Emit a warning message when a file is not verified.
Indicate that the file may be corrupt or that no
checksum was specified.
Only generate .filters files if they are different than the last time
they were generated. This should prevent the unnecessary reloads
being triggered with Visual Studio 2010 builds.
kFreeBSD and Hurd have the same userland as Linux. This change is
necessary to enable kwsysProcessKill() to kill child processes on
kFreeBSD. The bug was detected by CTestTestTimeout test.
Patch from "Modestas Vainius <modestas@vainius.eu>".
See issue #10432.
This is #10466. It didn't crash for me, but also didn't work.
The issue was that a line ending with something else than a # was considered as the
end of the documentation block. If there was simply the end of the file we didn't
get into the correct branch.
Alex
Add the <Email>...</Email> element in Update.xml for each commit
reported. This field was defined by Dart but never really used.
Distributed version control systems use author name and email
instead of a user id, so now it makes sense to use this field.
The expectation of users of the MSVC60, MSVC70, MSVC71, MSVC80, MSVC90
and the new MSVC10 variables is that at most one of them will be set
for any given build tree. This change enforces that expectation for
build trees using Makefile generators. It also fixes the one mismatch
in that expectation to be found in the Visual Studio generator world:
previously, the VS 7.1 generator would set *both* MSVC70 and MSVC71;
now, it only sets MSVC71.
With these changes, user expectations are now met, and the recently
introduced CheckCompilerRelatedVariables test should pass everywhere.
Since commit "Switch to using libarchive from libtar" (2009-10-30) the
the implementation of "tar xz" has printed all paths from the tarball as
a single line with no separator. This fixes the logic to extract
silently as expected.
Git does not automatically checkout the matching version of a submodule
when it checks out a new version of the parent project in the work tree.
If the submodule reference changed in the parent project then we were
reporting the submodule path as a local modification. Work around the
problem in ctest_update using "git submodule update" after "git pull".
For projects with no submodules this is a no-op. See issue #10662.
Also add a submodule to the test project for CTest.UpdateGIT to test the
work-around.
The PDB file name for VCLinkerTool is specified by the xml attribute
"ProgramDatabaseFile", not "ProgramDataBaseFile" (note the lower-case
character 'b'). VS seems to cope with the incorrect capitalization but
the combination of VS 7.1 and Incredibuild does not. See issue #10614.
Report in <UpdateReturnStatus> when the update command line returns
non-zero. Otherwise the failure may be silently ignored since the error
was previously recorded only in the local log file.
These were implementation details of the unused methods
cmMakefile::RegisterData
cmMakefile::LookupData
We simply remove the methods, members, and class cmData.
For builds from Git repositories, add "-g<commit>" to the end of the
version number. If the source tree is modified, append "-dirty".
For builds from CVS checkouts, add "-cvs-<branch>".
Add the [.tweak] version component throughout the policy implementation.
Document all components for the cmake_policy(VERSION) command. Record
the tweak level in which each policy was introduced (0 for all current
policies). In generated documentation we report the tweak level only if
it is not zero. This preserves existing documentation.
The command now accepts four version components in the format
major[.minor[.patch[.tweak]]]
This corresponds to the new versioning scheme introduced recently.
Prepare to switch to the workflow described by "git help workflows". In
this workflow, the "master" branch is always used to integrate topics
ready for release. Brand new work merges into a "next" branch instead.
We need a new versioning scheme to work this way because the version on
"master" must always increase.
We no longer use an even/odd minor number to distinguish releases from
development versions. Since we still support cvs checkout of our source
tree we cannot depend on "git describe" to compute a version number
based on the history graph. We can use the CCYYMMDD nightly date stamp
to get a monotonically increasing version component.
The new version format is "major.minor.patch.(tweak|date)". Releases
use a tweak level in the half-open range [0,20000000), which is smaller
than any current or future date. For tweak=0 we do not show the tweak
component, leaving the format "major.minor.patch" for most releases.
Development versions use date=CCYYMMDD for the tweak level. The
major.minor.patch part of development versions on "master" always
matches the most recent release.
For example, a first-parent traversal of "master" might see
v2.8.1 2.8.1.20100422 v2.8.2
| | |
----o----o----o----o----o----o----o----o----
Since the date appears in the tweak component, the next release can
increment the patch level (or any more significant component) to be
greater than any version leading to it. Topic branches not ready for
release are published only on "next" so we know that all versions on
master lead between two releases.
Commit "Support more special characters in file(STRINGS)" (2009-10-06)
attempted to support parsing strings from binaries produced by the
Portland Group Fortran compiler. The compiler seems to put an extra
byte just at the end of its string literals. Previously we dealt with
this by explicitly enumerating bytes known to occur, but it seems that
many such possibilities exist. Now we support extraction of strings
that end in any non-ASCII character.
The SaveRestoreEnvironment helper object makes sure that the
original environment is restored immediately after the
StartProcess call returns rather than waiting for the end
of the test. When tests are executed in parallel, there is
no guarantee about the ordering of EndTest calls relative
to StartTest calls. In fact, it would be odd for them to
be nested nicely. Therefore, to avoid the corruption of
the calling ctest's environment, the original environment
must be restored before ForkProcess returns.
Detect the runtime linker's search path and add to the compile time
linker's search path. This is needed because OpenBSD's static linker
does not search for shared library dependencies in the same places as
the runtime linker.
Teach kwsysProcessKill to identify processes on this platform using the "ps"
command just as on Linux. Patch from Modestas Vainius <modax@debian.org>.
See issue #10432.
The compiler id is checked for C++ and C, if there is not one
of those available, then just default to gcc. This makes it
work with Fortran, or None projects.
This makes the behavior of the build with the Visual Studio generators
equivalent to the behavior of makefile based builds. After an error
in a custom command sequence, the build stops and reports an error
rather than executing the remaining commands in the sequence.
Create platform variable "CMAKE_<LANG>_RESPONSE_FILE_LINK_FLAG" to
specify an alternative to "@" for referencing response files. It
applies specifically to response files with linker options.
See issue #10401.
Response files are parsed by tools, not by shells. We teach
cmLocalGenerator::Convert() a new "RESPONSE" output format and use it
for objects listed in response files. It does not do special slash or
MSYS root translation like the "SHELL" format does. This is necessary
for GNU tools on Windows to understand response file content.
See issue #10401.
Allow the user to set the CMake variable CTEST_COST_DATA_FILE, which will be used to store the cost data from test runs. If not set, defaults to the original location in the build tree Testing/Temporary dir.
Ensure that the HTML documentation generated by CMake complies with
"XHTML 1.0 Strict":
- All tags are properly closed and DOCTYPE is specified
- Useful for downstream XML-processors (e.g. for extracting section
titles)
See issue #10338.
Signed-off-by: Simon Harvey <simon.harvey@cambridgeflowsolutions.com>
http://public.kitware.com/Bug/view.php?id=9978
Now instead of one linked resource for each project() just one linked
resource to the top level source directory is created.
This should really avoid this type of name clashes. And to me it looks also
much less confusing.
Hopefully the name "[Source directory]" containing a space and square
brackets doesn't lead to problems somewhere. Here it works.
Alex
In commit 'Create KWSYS_PLATFORM_INFO_TEST macro' (2009-11-20) we
implemented the macro to use a cache entry to avoid re-running the
try_compile(). However, the output copied from the try_compile is
needed on every configure. If the user wipes out the build tree but not
the cache file then the try_compile() will not re-run to recreate the
needed file. We address the problem by teaching the macro to run the
try_compile() whenever its output file does not exist.
We store custom command rule hashes in CMakeFiles/CMakeRuleHashes.txt
persistently across CMake runs. When the rule hash changes we delete
the custom command output file and write a new hash into the persistence
file.
This functionality was first added by the commit 'Introduce "rule
hashes" to help rebuild files when rules change.' (2008-06-02).
However, the implementation in cmGlobalGenerator::CheckRuleHashes kept
the file open for read when attempting to rewrite a new file. On
Windows filesystems this prevented the new version of the file from
being written! This caused the first set of rule hashes to be used
forever within a build tree, meaning that all custom commands whose
rules changed would be rebuilt every time CMake regenerated the build
tree.
In this commit we address the problem by splitting the read and write
operations into separate methods. This ensures that the input stream is
closed before the output stream opens the file.
We use 'git diff-index' to detect local modifications after pull. On
some filesystems the work tree timestamps of a few files may be dated
after the index, making them appear as locally modified. We address the
problem by using 'git update-index --refresh' to refresh the index and
avoid false local modifications.
Our internal path processing methods assume no trailing slashes, but bzr
adds trailing slashes to updated directories. This can lead to empty
entries in Update.xml files. We address the problem by stripping the
slashes as soon as they are parsed.
Cygwin versions .dll files by putting the version number in the file
name. Our fix to issue #3571 taught CMake to do this, but it used the
VERSION target property. It is better to use the SOVERSION property
since that is the interface (rather than implementation) version.
Change based on patch from issue #10122.
When building on Cygwin without -mwin32, the _WIN32 macro may not be
defined. SharedForward must still set the PATH environment variable to
ensure runtime dependencies are found.
The 'ldd' wrapping feature uses 'cygcheck' for now since a real ldd tool
is not available in Cygwin 1.5. We can change to use the real ldd when
we choose to stop supporting legacy Cygwin and require 1.7.
When kwsys is built using GCC visibility support can be used. This is similar
to the way that Windows exports symbols in DLLs, and requires projects that
build kwsys to change the default visibility using some compiler flags. See
http://gcc.gnu.org/wiki/Visibility for more details about GCC visibility.
Makefile dependencies must be escaped using cmLocalGenerator::Convert
with the cmLocalGenerator::MAKEFILE option. This fixes Fortran module
dependencies with spaces in the path. We test the fix by adding a space
to one of the module paths in the Fortran test.
The improved text gives very clear information when either the CMakeLists.txt or CTestConfig.cmake file is missing. Hopefully, it makes it easier for those poor future souls who encounter these messages to solve their problems quickly.
If APPEND is given to ctest_start, it will read the tag from the current existing Testing/TAG file rather than creating a new one based on the current time stamp. This allows a developer to run several dashboard scripts in a row, all of which will share the same tag/stamp/buildid when they finally get submitted to CDash. Now you can split the running of build phases and test phases for the same dashboard row into multiple scripts.
Before this commit, the value of PATH at cmake time was put into the eclipse
project file. The problem with this is that this will be lost the first time
cmake is rerun from an build inside eclipse which was started without the
environment externally already set.
This patch now:
-adds the env.var to the cache if it is not already in the cache
-reuses the variable from the cache if it is in the cache, but not in the env.
-uses the variable from the cache if it contains the whole content of the
current env.var (e.g. if it is the full PATH plus the MSVC dirs)
Also store INTEL_LICENSE_FILE in the project file if an Intel compiler is used.
Alex
The Eclipse and KDevelop generators set the VERBOSE environment variable to
TRUE in the project files, because they must be able to "see" the full
command lines and errors, otherwise they can't parse the errors. But the
VERBOSE env.var. also enables cmakes own verbose output, which can be quite
long. This commit introduces an environment variable CMAKE_NO_VERBOSE, which
when set disables cmake's verbose output also when VERBOSE is set. This
env.var is now set by both the Eclipse and the KDevelop generators.
Alex
Previously we produced commit times formatted like
1261403774 -0500
which is what the Git plumbing prints. Now we use a human-readable
format like
2009-12-21 15:28:06 -0500
which is still easy to machine-parse.
Previously we escaped quotes in <UpdateCommand>...</UpdateCommand>
values using '"'. This is not necessary because the value is in
xml CDATA and not an xml attribute.
A Subversion revision is unique across the entire repository, but work
trees typically correspond only to a subdirectory below the root path.
In order to specify the version of the source code that was tested,
CTest now submits a <SVNPath> element in Update.xml that specifies the
directory of the repository that corresponds to the work tree. In
combination with the revision number this uniquely specifies the tested
source. See issue #7541.
We teach CTest to report in a <Revision> element the revision of the
source tree that was tested. This makes sense for all modern VCS tools
because they version the whole tree. We simply omit this element for
CVS because it only versions files. See issue #7541.
The SharedForward header contains a preprocessor table mapping from
platform to equivalents for ldd and LD_LIBRARY_PATH. This commit fixes
the table preprocessor directives to guarantee at most one platform.
This generalizes the commit "Fix compilation of VTK on debian/sparc".
The TestSharedForward executable and TestDynload module do not actually
link to a KWSys library, but it is nice to build them after the
libraries just like all other test binaries.
This also works around a universal binary bug in Xcode 2.x. It forgets
to create the output directory for the executable before linking it. We
avoid the problem by putting the library in the directory first.
CMake has a special case for the "make install" target when building
CMake itself. We use the just-built CMake to install itself since an
existing CMake installation cannot replace itself (at least on Windows).
We simplify the code that computes the location of the CMake binary by
taking advantage of existing generator support for target lookup. This
will make it robust to any changes in CMake's own CMakeLists.txt files
in the future.
Some fixes for including Qt frameworks.
Remove extra "QtGui.framework" so its not Contents/Frameworks/QtGui.framework/QtGui.framwork/... anymore.
Also include QtGui Resource folder, so a Cocoa/Qt based cmake-gui app works.
CTest filters the output from tools and tests to ensure that the XML
build/test result documents it generates have valid characters.
Previously we just converted all non-ASCII bytes into XML-escaped
Unicode characters of the corresponding index. This does not preserve
tool output encoded in UTF-8.
We now assume UTF-8 output from tools and implement decoding as
specified in RFC 3629. Valid characters are preserved, possibly with
XML escaping. Invalid byte sequences and characters are converted to
human-readable hex values with distinguishing tags. See issue #10003.
We re-arrange EXECUTABLE_OUTPUT_PATH settings to avoid putting utility
and test executables in the 'bin' directory of the build tree. This
makes the directory look like that in the installation tree, except that
on multi-configuration generators we still use a per-config
subdirectory.
KWSys should not set variables outside its namespace. It can honor the
EXECUTABLE_OUTPUT_PATH set by a host project, but tere is no need for it
to set a default in the host project cache.
The DumpDocumentation executable and some supporting code and tests were
completely unused by CMake. Generation of documentation is done by the
individual executables with --help* options. In this commit we simply
remove the unused code, executable, and test.
The Intel Fortran plugin to VS defines VFFortranCompilerTool as the
compiler tool. This commit fixes generated projects to use that tool
for per-source settings instead of VCCLCompilerTool. We were already
using it for target-wide compiler settings.
In order to kill process trees we need to list all processes to find
those whose parent we are killing. We implement process listing on
OpenSolaris by using "ps -ef" and parsing the resulting format:
UID PID PPID C STIME TTY TIME CMD
%*s %d %d %*[^\n]\n
In order to kill process trees we need to list all processes to find
those whose parent we are killing. We implement process listing on QNX
using "ps -Af" and parsing the resulting format:
UID PID PPID C STIME TTY TIME CMD
%*d %d %d %*[^\n]\n
We enumerate processes to identify those whose parent is being killed so
that we can recursively kill the children. Enumeration uses the
Process32(First|Next) windows API functions, which accept PROCESSENTRY32
objects to be filled. This commit corrects the declaration of the entry
structure to account for its size on 64-bit Windows.
On UNIX systems we kill a tree of processes by performing a DFS walk of
the tree. We send SIGSTOP to each process encountered, recursively
handle its children, and then send SIGKILL.
We once used the above approach in the past, but it was removed by the
commit "Do not send both SIGSTOP and SIGKILL when killing a process".
The commit was meant to work-around an OS X 10.3 bug in which the child
would not always honor SIGKILL after SIGSTOP. At the time we wrongly
assumed that the process tree remains intact after SIGKILL and before
the child is reaped. In fact the grandchildren may be re-parented to
ppid=1 even before the child is reaped, which causes the DFS walk to
miss them.
The Watcom tools do their own command-line parsing and do not accept
double-quotes. Instead we single-quote the target output name when
invoking wlink and other Watcom tools. This fixes support for spaces in
the target output directory path when it is not under the build tree.
We create CMake Policy CMP0015 to make link_directories() treat relative
paths with respect to the source tree while retaining compatibility.
This makes it consistent with include_directories() and other commands.
Changes based on patch from Alex. See issue #9697.
In CTest command-driven script mode we support starting without a source
tree. Previously the ctest_start() command would do some initialization
but could not do anything that required CTestConfig.cmake from the input
source tree. Later, ctest_update() would run CTEST_CHECKOUT_COMMAND to
create the source tree, and then re-initialize everything. This
delayed-initialization approach led to many complicated cases of which
only some worked. For example, the second initialization only worked
correctly in Nightly mode and simply failed for Experimental and
Continuous builds.
A simpler solution is to run CTEST_CHECKOUT_COMMAND during ctest_start()
and then have a single initialization path. In principle this change in
behavior could break scripts that set the checkout command after
ctest_start() but before ctest_update(). However, the convention we've
always followed has been to set all variables before ctest_start().
See issue #9450.
We make the cmCTest::Initialize method private since it is only called
from inside the class implementation. We also combine the two boolean
arguments into one since they both meant the same thing.
We suppress Intel warning 1572 because the cases where we do equality
tests are valid. Since this project does not do numerical computations
we need not worry about real instances against which this warning
protects.
-use CMAKE_EXECUTABLE_FORMAT and CMAKE_SYSTEM_NAME to decide which binary
parsers to load (ELF/Mach O/PE)
-use CMAKE_(C|CXX)_COMPILER_ID to load the respective compiler error parser
-remove EclipseToolchainType, which was a mixture between compiler and operating system
Alex
The commit "Enable loose loop constructs in KWSys" set the minimum
required CMake version to 2.4.5. This regressed the setting of CMP0003,
so we restore it in this commit.
The FundamentalType header needs to know type sizes at preprocessing
time. This commit teaches it to avoid using CHECK_TYPE_SIZE because the
macro does not work for types whose size varies across architectuers in
Mac OS X universal binaries. Fortunately the Mac compilers provide just
enough information to detect the needed type sizes during preprocessing.
We now use preprocessor macros instead of configuration tests whenever
they are available. As a side effect this reduces the number of
try-compiles needed with GCC.
See issue #9913.
This macro helps KWSys perform try-compile tests that extract 'INFO'
strings out of compiled binaries. It works for CMake 2.6 and above.
On CMake 2.4 it always returns an empty list of information values,
so this should be used only as an optimization until 2.6 is required.
In KWSys IOStream we need to detect whether 'long long' exists but we do
not need its size. We avoid using CHECK_TYPE_SIZE because it does not
work for types whose size varies across architectuers in Mac OS X
universal binaries. See issue #9913.
The list of libraries to be linked into the current target must be
specified using windows slashes to that UNC paths such as
\\server\share\somelibrary.lib
work correctly. See issue #9917.
VS 10 provides $(Configuration) and $(ConfigurationName) but only the
former is documented so we prefer it. This also makes CMAKE_CFG_INTDIR
consistent with its documentation. See issue #9916.
The commit "Fix get_filename_component ABSOLUTE mode" broke REALPATH
treatment of relative paths because it stopped storing the absolute path
in local variable 'filename'. This commit fixes the call to GetRealPath
to use the proper local variable and adds a test.
The ctest subdirs command now checks the relative path first, and if that does not exist, also checks if the given path was absolute. Thanks vodall for the patch.
Previously we silently ignored such calls and set nothing. The commit
"Initialize directory scope with closure of parent" inroduced a bad test
for the top scope. This commit fixes the test to avoid dereferencing a
null pointer, and adds a warning when the case is encountered.
This allows for a built in bzip and zip capability, so external tools
will not be needed for these packagers. The cmake -E tar xf should be
able to handle all compression types now as well.
We create per-configuration target properties to specify ARCHIVE,
LIBRARY, and RUNTIME output directories. The properties override the
generic properties for the <CONFIG> configuration:
ARCHIVE_OUTPUT_DIRECTORY -> ARCHIVE_OUTPUT_DIRECTORY_<CONFIG>
LIBRARY_OUTPUT_DIRECTORY -> LIBRARY_OUTPUT_DIRECTORY_<CONFIG>
RUNTIME_OUTPUT_DIRECTORY -> RUNTIME_OUTPUT_DIRECTORY_<CONFIG>
For multi-configuration generators, the per-configuration subdirectory
normally appended to the generic output directory is not added to the
configuration-specific property values. This allows projects to set the
exact location at which binaries will be placed for each configuration.
See issue #9163.
VS 6 forgets to create the output directory for a static library if it
differs from the intermediate files directory. We work around this VS
bug by creating a pre-link event on the library target to make the
directory.
This commit removes use of configuration-less cmTarget::GetDirectory()
by the VS 6 generator (except for compatibility with user templates).
We replace OUTPUT_DIRECTORY_<CONFIG> tokens in the templates using the
per-configuration result of cmTarget::GetDirectory(config).
The try_compile() command's COPY_FILE option should not try to actually
copy the file if it cannot be found. Some C runtime library's fopen
cannot handle an empty file name.
Previously the Xcode generator set SYMROOT to be the target output
directory. This told Xcode to put the "<proj>.build" directory in the
output path too.
This commit sets SYMROOT, CONFIGURATION_BUILD_DIR, and OBJROOT to put
intermediate files in the build directory corresponding to the source
directory that created each target. This is more consistent with the VS
IDE generators. Now only the build output files (actual targets) go to
the target output directory.
The commit "modified the if command to address bug 9123 some" changed
the if() command behavior with respect to named boolean constants. It
introduced policy CMP0012 to provide compatibility. However, it also
changed behavior with respect to numbers (like '2') but did not cover
the change with the policy. Also, the behavior it created for numbers
is confusing ('2' is false).
This commit teaches if() to recognize numbers again, and treats them
like the C language does in terms of boolean conversion. We also fix
the CMP0012 check to trigger in all cases where the result of boolean
coersion differs from that produced by CMake 2.6.4.
The if() command reports its arguments at the beginning of some error
messages. Originally it reported the un-expanded form of the arguments
because in ancient CMake versions no context information was available.
Now it is more useful to see the real arguments, which may be mentioned
in the main error message. Since full context information is now
available, users can refer back to the source if they need to see the
unexpanded form of the arguments.
For example, the code
set(regex "++")
if("x" MATCHES "${regex}")
endif()
now produces the message
if given arguments:
"x" "MATCHES" "++"
Regular expression "++" cannot compile
instead of
if given arguments
"x" MATCHES "${regex}"
Regular expression "++" cannot compile
The commit "Set version info for shared libs on OSX" taught the Xcode
generator to honor VERSION and SOVERSION properties. However, it also
set version '1.0.0' as the default when no version property is set,
which is inconsistent with the Makefiles generator. This commit fixes
the default to '0.0.0' for consistency.
See issue #9773.
The reverted commit attempted to preserve the "../" PREFIX work-around
for avoiding per-config build directories in the VS IDE generators.
However, the original reporter has concluded that a "../" PREFIX no
longer works everywhere in VS 10 project files anyway. Rather than set
OutputPath, this commit restores the $(OutDir)$(TargetName)$(TargetExt)
default.
See issue #9768.
This commit fixes the cmparseMSBuildXML.py script to generate correct
flag table entries for booleans with optional value. These flags use
two entries: the first should ignore the value and enable the option,
and the second should use the value if given. Previously the first
entry did not recognize flags with values.
In particular this fixes flags like /MP4, but the change corrects
matching of some other flags too. See issue #9771.
This commit teaches the VS 10 generator to detect the -D_UNICODE option
in preprocessor definitions and set the CharacterSet attribute to the
value 'Unicode'. This was already done for other VS IDE versions.
See issue #9769
The default $(OutDir)$(TargetName)$(TargetExt) for this value works in
most cases because we set the three properties. However, if the target
property PREFIX contains a path component (not documented but happens to
work in other VS generators) we drop it from TargetName and do not put
it in OutDir either. This commit corrects the resulting path by setting
the OutputPath property explicitly with the full path.
See issue #9768.
Intel Fortran on Mac OS X enables Fortran support in Xcode. This commit
teaches CMake to associate Fortran sources properly in Xcode projects.
See issue #9739.
The VS 10 flag table generation script did not produce correct entries
for precompiled header flags. Since precompiled header flag translation
requires multiple entries cooperating in a certain order, it is not
worth the time to make the generation script work automatically. This
commit manually adds the proper entries.
See issue #9753.
MS changed the location of the Microsoft.Cpp.$(Platform).user.props
file. This commit teaches the VS 10 generator about the new location.
See issue #9759.
Versioned UNIX libraries and executables produce multiple names for a
single target using one of
cmake -E cmake_symlink_library
cmake -E cmake_symlink_executable
to create symlinks to the real file for the extra names. However, when
cross-compiling from Windows to Linux we cannot create symlinks. This
commit teaches CMake to make copies instead of symbolic links when
running on windows. While this approach does not produce exactly what
Linux wants to see, at least the build will complete and the binary will
run on the target system. See issue #9171.
During installation of a target we generate "tweak" rules to update the
installed file (RPATH, strip, ranlib, etc.). However, some targets
install multiple files, such as the versioned names of a shared library.
Previously the extra files for a target have always been symbolic links,
but for cross-compiling from Windows to UNIX they may need to be copies.
This commit teaches the generated install scripts to loop over all files
installed for the target to apply tweaks to those that are not symlinks.
See issue #9171.
In cmInstallTargetGenerator::GenerateScriptForConfig we were computing
the full 'from' paths for all target files to be installed, but only
computing a 'to' path for the "main" target file. This commit teaches
the method to compute both 'from' and 'to' paths for every target file
to be installed. The result is cleaner, easier to follow, and will
allow installation tweaks to be added later on all target files.
We factor the implementation of
cmake -E cmake_symlink_library
cmake -E cmake_symlink_executable
out of cmake::ExecuteCMakeCommand into methods
cmake::SymlinkLibrary
cmake::SymlinkExecutable
plus a helper method cmake::SymlinkInternal.
The cmInstallTargetGenerator methods AddStripRule and AddRanlibRule do
not need the target type argument. They can simply use the type of the
target for which the generator instance was created.
The CMAKE_OSX_ARCHITECTURES variable works only as a global setting.
This commit defines target properties
OSX_ARCHITECTURES
OSX_ARCHITECTURES_<CONFIG>
to specify OS X architectures on a per-target and per-configuration
basis. See issue #8725.
Both generators use the CMAKE_EDIT_COMMAND variable to determine whether
they should add the edit_cache target, i.e. they don't add it if it's
ccmake, since this does not work inside the output log view of
Eclipse/Codeblocks. But instead of requiring the variable to be set they now
check it for 0 and handle this appropriately. This should help Dave getting
some testing for them :-)
Alex
In VS 8 and greater this commit implements
add_dependencies(myexe mylib) # depend without linking
by adding the
LinkLibraryDependencies="false"
option to project files. Previously the above code would cause myexe to
link to mylib in VS 8 and greater. This option prevents dependencies
specified only in the solution from being linked. We already specify
the real link library dependencies in the project files, and any project
depending on this to link would not have worked in Makefile generators.
We were already avoiding this problem in VS 7.1 and below by inserting
intermediate mylib_UTILITY targets. It was more important for those
versions because if a static library depended on another library the
librarian would copy the dependees into the depender! This is no longer
the case with VS 8 and above so we do not need that workaround.
See issue #9732.
In Visual Studio project files we pass compiler flags to the whole
target based on the linker language, which works for MS tools and
combinations of C and C++. For the Intel Fortran plugin though the
generated .vfproj files should never contain C or C++ options.
We generate .vfproj files only for targets consisting only of Fortran
code. Now that the linker language is computed transitively through
linking it is possible that the linker language is C++ for an otherwise
Fortran-only project. This commit forces Fortran as the linker language
for the purpose of specifying target-wide flags in .vfproj files.
See issue #9719.
The commit "Avoid non-root copies of root-only targets" moved the check
for root-only targets into cmGlobalGenerator::GetTargetSets to avoid
adding multiple ALL_BUILD targets to the "original" target set. This
approach did not work for ZERO_CHECK targets though because those are
pulled in by dependency analysis.
Instead we eliminate duplicate ZERO_CHECK targets altogether and refer
to a single one from all solution files. This cleans up VS 10 project
file references to ZERO_CHECK targets anyway.
We revert commit "Create CMake Policy CMP0015 to fix set(CACHE)" because
the NEW behavior of the policy breaks a valid use case:
# CMakeLists.txt
option(BUILD_SHARED_LIBS "..." ON)
add_library(mylib ...)
set(BUILD_SHARED_LIBS OFF) # we want only mylib to be shared
add_subdirectory(ThirdParty)
# ThirdParty/CMakeLists.txt
option(BUILD_SHARED_LIBS "..." ON)
# uh, oh, with NEW behavior this dir uses shared libs!!!
We'll re-introduce the policy later with a different change in behavior
to resolve the motivating case, which was more subtle but less common.
See issue #9008.
The commit "Really fix color check for dependency scanning" disabled
color unless "make COLOR=ON" is specified. This restores the previous
default behavior when CMAKE_COLOR_MAKEFILE is ON while retaining the
previous commit's fix. See issue #9680.
The commit "Fix color check for dependency scanning" was meant to
disable color if CMAKE_COLOR_MAKEFILE was off. It did remove use of the
activation option '--color' but it failed to make the default false when
the option was missing. This commit corrects that. See issue #9680.
The commit "add support for borland run time flag for shared builds"
started using the value of CMAKE_SHARED_LIBRARY_CXX_FLAGS to link
executables because Borland requires some flags both at compile and link
time. This change ended up propagating all the way to the current
Makefile generators and the behavior applies on all platforms.
In general it is incorrect to use these flags to link executables. The
commit "Split Borland compiler information files" re-wrote Borland
support to work without this behavior, so we remove it in this commit.
Patch from Ben Hutchings. See issue #9659.
This commit re-writes Borland compiler build rules. We split the rules
into modern <os>-<id>-<lang> information modules but share a common
macro between languages to avoid duplication.
We also address a bug in the previous rules that would build some target
types against the static Borland runtime and others against the shared
Borland runtime in one build tree. Now we always use the shared runtime
as is the default in the rules for MS tools.
When <pkg>_DIR is set to an incorrect version we search again and store
the result in the variable, even if it is <pkg>_DIR-NOTFOUND.
There was a bug in the case when the new search does not find anything
and the old value came from a cache entry with UNINITALIZED type. The
command used to try to load a package configuration file from the last
place searched, and would leave the old wrong value in the entry. This
commit fixes the behavior to avoid trying to load a missing file and to
set the value to <pkg>_DIR-NOTFOUND as expected.
In cmMakefile::AddCacheDefinition we collapse paths specified in PATH or
FILEPATH cache entries originally specified on the command line with
UNINITALIZED type. This commit fixes the logic to avoid collapsing
<var>-NOTFOUND and other false values. The change allows other CMake
code to force a NOTFOUND value on an entry with UNINITALIZED type.
The commit "Support more special characters in file(STRINGS)" added code
using a 'char' literal as an array index. Some compilers warn about
this because char might be a signed type, leading to negative indices.
We replace the literal with an integer to avoid the warning.
The commits "Teach Fortran compiler identification about the Portland
Group compiler" and "Fix previous change to file(STRINGS) command"
taught file(STRINGS) to recognize the form-feed '\f' character as part
of string literals. The Portland Group Fortran compiler also puts 0x14
bytes at the end of string literals in some cases. We generalize the
previous solution and add the new character in this commit.
Scanning the dependencies for kdelibs/khtml/ (when all files are scanned) is
now down to 4.6 s from 6.3 s before this change (without the
headerLocationCache it takes about 14 s here).
It doesn't really make sense to include the complete include path as part
of the key for the map, since the include path will be the same for
all files in a project, so it doesn't add anything.
Alex
In cmGlobalGenerator::GetTargetSets we collect targets from all local
generators in a tree or subtree corresponding to a project() command.
Some targets, such as ALL_BUILD, are duplicated in each subdirectory
with a project() command. For such targets we should keep only the copy
for the top-most (root) local generator.
Previously this filtering was done in each VS IDE generator, but it is
easier to do it in one place when the targets are first encountered.
This also fixes bad ALL_BUILD dependencies generated for VS 7.0 because
the cmGlobalVisualStudio7Generator::WriteTargetDepends method was not
filtering out duplicates. Now we avoid duplicates from the start.
The commit "Target copy ctor should copy internal state" created a new
cmTargetInternals constructor but failed to initialize a POD member that
the original constructor initializes. This commit fixes it.
Ideally we should never copy cmTarget instances, but it is a pain to
remove current uses of it. The pimplized portion of cmTarget has mostly
members that cache results, but some are part of the object state.
These should be copied in the copy ctor instead of re-initialized.
The Encoding key is now deprecated by the FreeDesktop standard and all
strings are required to be encoded in UTF-8. This desktop entry
explicitly specified an Encoding of UTF-8, which is harmless but no
longer necessary. See
http://standards.freedesktop.org/desktop-entry-spec/1.0/apc.html
for details. Patch from Modestas Vainius. See issue #9659.
In cmTarget we compute the link implementation, link interface, and link
closure structures on-demand and cache the results. This commit teaches
cmTarget to invalidate results after a LINK_INTERFACE_* property changes
or a new link library is added. We also clear the results at the end of
the Configure step to ensure the Generate step uses up-to-date results.
In cmTarget::SetProperty and cmTarget::AppendProperty we check whether
changing the property invalidates cached information. The check was
duplicated in the two methods, so this commit moves the check into a
helper method called from both.
This method is called during ConfigureFinalPass on every target. It
gives each target a chance to do some final processing after it is known
that no more commands will affect it. Currently we just call the old
AnalyzeLibDependencies that used to be called directly.
This commit re-words the warning message produced for CMP0012 to avoid
the word 'you' since often the person reading the message is not the
author of the code. We also add an example of the bad OLD behavior to
the policy documentation.
These policies were originally developed during the 2.7.x series and
intended for 2.6.5. There was never a 2.6.5 release, so we should refer
to 2.8.0 instead.
This commit creates target and directory properties to enable the Intel
interprocedural optimization support on Linux. Enabling it adds the
compiler option '-ipo' and uses 'xiar' to create archives.
See issue #9615.
This creates cmTarget::GetFeature and cmMakefile::GetFeature methods to
query "build feature" properties. These methods handle local-to-global
scope and per-configuration property lookup. Specific build features
will be defined later.
We create cmMakefileTargetGenerator::AddFeatureFlags to consolidate
addition of language flags. Currently it just adds the flags from
generic per-language flag variables (AddLanguageFlags).
There is confusion whether the file "currently being processed" inside a
function or macro is the file containing the definition or not. This
commit explicitly describes the behavior. See issue #9646.
In VS 6, 7, and 7.1 solutions we implement
add_dependencies(myexe mylib) # depend without linking
by creating an intermediate mylib_UTILITY target with dependencies
myexe -> mylib_UTILITY -> mylib
to avoid linking myexe to mylib. Previously these extra targets were
only added to the solution files in an ancestor directory of that
defining mylib. For example, in the project:
# CMakeLists.txt
project(TOP)
add_subdirectory(A)
add_subdirectory(b)
# A/CMakeLists.txt
add_library(mylib ...)
# B/CMakeLists.txt
project(B)
add_executable(myexe ...)
add_dependencies(myexe mylib)
the solution for TOP would have mylib_UTILITY but the solution for B
would not even though it pulls in mylib through the dependency. This
commit fixes solutions generated in other directories so that any
solution that has mylib will get mylib_UTILITY also.
See issue #9568.
This commit cleans up the declaration, definition, and invocations of
the GetTargetSets method and related code. There is no change in
function except to make the method virtual.
This teaches the VS 6 IDE generator to use the transitive closure of
target dependencies. This re-uses much of the implementation already
done for the other VS>6 generators. See issue #8223 and issue #9568.
We move cmGlobalVisualStudio7Generator::OrderedTargetDependSet up to
cmGlobalVisualStudioGenerator so it can be re-used for other VS
versions. See issue #9568.
Some find-modules use get_filename_component() to expand registry
values. We need to look in both the 32-bit and 64-bit registry views
when expanding values. We prefer the one that the target application
would see. See issue #8792.
When building through NMake with VS 6, the module definition file must
be passed without spaces in the path. This is because 'cl -link' does
not escape the spaces when passing the value on to the linker.
Since Haiku does not have /usr (and therefore /usr/local), this commit
changes the default install prefix to the equivalent directory of
/boot/common.
See issue #9607.
We recognize .def source files and map them to the /DEF:<file> option in
the MSVC tools. Previously this worked only for shared libraries. This
commit cleans up the implementation and makes it work for executables
too. See issue #9613.
When constructing cmVisualStudioGeneratorOptions to parse options for
tools 'link' and 'lib' the tool type is now Linker, not Compiler. This
tells it not to recognize flags starting in '/D' as preprocessor macros,
such as the '/DEF:<file>' linker option. See issue #9613.
The commit "cmparseMSBuildXML should output StringProperty values too"
fixed the script that generated
Source/cmVS10CLFlagTable.h
Source/cmVS10LibFlagTable.h
Source/cmVS10LinkFlagTable.h
This commit updates the files with the new output.
Some older STL implementations invoke the comparison functor as a const
object, so the function call operator must be 'const' qualified. This
fixes the commit "Fix support for OLD behavior of policy CMP0002" to
compile on older STLs.
The commit "Cleanup make progress rule generation code" introduced a map
from target name to the progress.make file location. Policy CMP0002's
OLD behavior allows duplicate target names in different directories, so
only one ends up with a progress.make file. This commit fixes the map
to order by target name first and build directory second, restoring
support for duplicate target names.
This converts the CMake license to a pure 3-clause OSI-approved BSD
License. We drop the previous license clause requiring modified
versions to be plainly marked. We also update the CMake copyright to
cover the full development time range.
This converts the KWSys license to a pure 3-clause OSI-approved BSD
License. We drop the previous license clause requiring modified
versions to be plainly marked. We also update the KWSys copyright to
cover the full development time range.
This commit adds KWSys configuration option KWSYS_INSTALL_DOC_DIR to
specify the directory for installation of documentation. We use it to
put the KWSys Copyright.txt file at the location
${KWSYS_INSTALL_DOC_DIR}/${KWSYS_NAMESPACE}/Copyright.txt
in the project installation tree. This helps containing projects meet
the license requirement to distribute the copyright and license with
binary forms.
The commit "Generate proper Intel Fortran project version" replaced the
hard-coded 9.10 value with a runtime registry lookup of the real
version. Version 10.1 actually uses project file format 9.10, so this
commit switches it back for that version. See issue #9169.
Applications on Haiku are discouraged from storing their data in $HOME.
This teaches export(PACKAGE) and find_package() to use the BeAPI on
Haiku to store the package registry instead of using ~/.cmake/packages.
See issue #9603.
Set the working_dir entry in the codeblocks project file of executable
targets to the directory where the executable is created. Then when running
CB, executing the target (not building), will run it from that directory.
Alex
CMake policies CMP0014 and CMP0015 were implemented in the development
series version 2.7.x but will be first released in 2.8.0. Now that the
development version number is higher than that (2.9.x) we can update
their version of introduction to the actual release number.
In CMake Policy documentation we specify the default behavior for the
current version of CMake. This commit fixes that version by reporting
the full version number instead of just major and minor.
KWSys Robot
Brad King
Kai Wasserbäch
Bill Hoffman
David Cole
Zach Mullen
David Genest
Christoph Watzl
Clinton Stimpson
Alex Neundorf
Ruben Van Boxem
Dave Partyka
Chuck Atkins
Pat Marion
Simon Harvey
Marcus Hanwell
Luis Ibanez
Zack Galbreath