dc1d025 OS X: Add test for rpaths on Mac.
8576b3f OS X: Add support for @rpath in export files.
00d71bd Xcode: Add rpath support in Xcode generator.
94e7fef OS X: Add RPATH support for Mac.
RPATH support is activated on targets that have the MACOSX_RPATH
property turned on.
For install time, it is also useful to set INSTALL_RPATH to help
find dependent libraries with an @rpath in their install name.
Also adding detection of rpath conflicts when using frameworks.
Initialize the CMAKE_CONFIGURATION_TYPES cache entry early during
EnableLanguage like the Xcode generator does. Avoid depending on
the MSVC compiler information module to do it. Otherwise code like
project(MyProj NONE)
sets CMAKE_CONFIGURATION_TYPES late (in GenerateConfigurations), and
to only "Debug" and "Release" instead of the standard set of 4.
Reported-by: Paul Smith <paul@mad-scientist.net>
Commit dac78148 (...makes the mingw cross compiler work out of the
box..., 2007-08-02) added to CMAKE_SYSTEM_PROGRAM_PATH and
CMAKE_SYSTEM_LIBRARY_PATH paths like "/bin" and "/lib" with no Windows
drive letter so that cross-compiling to Windows from Linux would search
these paths under CMAKE_FIND_ROOT_PATH. Later commit 2a782880 (...use
CMAKE_SYSTEM_PREFIX_PATH when possible, 2008-01-16) generalized this
approach by instead adding "/" to CMAKE_SYSTEM_PREFIX_PATH.
Both commits assumed that the paths would never match anything on
Windows hosts without a drive letter. However, Windows evaluates these
paths relative to the current working drive letter so find_* commands
may report paths like "/lib/..." when paths like "c:/lib/..." exist on
what happens to be current drive. Such drive-less paths are not
reliable when the working drive changes, so we should not use them.
Fix WindowsPaths.cmake to add '/' to CMAKE_SYSTEM_PREFIX_PATH only when
cross-compiling to Windows from a non-Windows host. This will avoid
searching and finding local paths without a drive letter on Windows.
Update the CMAKE_<LANG>_LINK_EXECUTABLE rule variable to invoke the
linker directly instead of through the compiler. We already do this
for DLL linking with CMAKE_<LANG>_CREATE_SHARED_LIBRARY.
This also works around a VS 6 cl bug. While invoking the link tool
internally it fails to correctly quote flags like /pdb:... with spaces
in the value.
The MS tools create two types of PDB files as explained here:
http://msdn.microsoft.com/en-us/library/yd4f8bd1%28v=vs.71%29.aspxhttp://msdn.microsoft.com/en-us/library/yd4f8bd1%28v=vs.80%29.aspxhttp://msdn.microsoft.com/en-us/library/yd4f8bd1%28v=vs.90%29.aspxhttp://msdn.microsoft.com/en-us/library/yd4f8bd1%28v=vs.100%29.aspx
One is created by the compiler (/Fd) and the other by the linker (/pdb).
The two options should not specify the same file. Split them up.
In the VS IDE generators, simply drop ProgramDataBaseFileName to
take the VS default "/Fd$(IntDir)vc$(PlatformToolsetVersion).pdb".
In the Makefile generators, set "/Fd" on the compile line to be
the directory containing object files (with a trailing slash the
compiler will add the "vc$(PlatformToolsetVersion).pdb" filename
automatically). Drop the /Fd option from the exe link command
line and add "/pdb" instead (already done for dll linking).
Update these rules for both MSVC and Intel tools.
Drop support for PDB_OUTPUT_DIRECTORY and PDB_NAME in STATIC
libraries because the generated .pdb files are only from /Fd
and not real linker-generated .pdb files. Update documentation to
clarify that the PDB_* properties are only for linker .pdb files.
This regresses the PDBDirectoryAndName test for STATIC libraries.
Since it is not clear at this time what should be done for STATIC
library .pdb files, comment out the relevant portion of the test
and leave a TODO comment.
The flag was added incorrectly by commit 9c3a6eb4 (Need -brtl when creating
shared libraries, 2003-05-16). According to "man ld" the -G option implies
"-brtl -bnortllib ...", -brtl implies "-brtllib", and -brtllib should only be
used for executables, not shared libraries. Therefore it is incorrect and
unnecessary to specify -brtl explicitly after -G.
Reported-by: Kevin Burge <kcburge@gmail.com>
The GNU compiler front-ends on AIX invoke the linker with flags of the
form "-L/path/to/gnu/runtime/lib" to tell ld where to find the language
runtime libraries. They depend on the default libpath behavior
documented in "man ld" to add the -L paths also to the runtime libpath
so the dynamic loader can find the language runtime libraries. This
differs from platforms whose linkers have distinct -rpath flags that
non-system compilers can use to tell the dynamic loader where to find
their language runtime libraries.
Since commit 96fd5909 (Implement linking with paths to library files,
2008-01-22) CMake always passes "-Wl,-blibpath:" followed by any
project-defined RPATH plus "/usr/lib:/lib" in order to explicitly set
the runtime libpath and avoid getting all the project -L paths in the
runtime libpath. The explicit libpath prevents the GNU compiler runtime
library -L paths from being placed in the libpath and then the dynamic
loader fails to find the language runtime libraries.
CMake already detects the implicit link directories for each language
since commit 07ea19ad (Implicit link info for C, CXX, and Fortran,
2009-07-23). Add the implicit link directories to the explicit runtime
libpath for GNU compilers on AIX to fix this use case.
Since commit c70beb4b (change the default borland stack size, 2003-05-05),
commit 1b572eb9 (remove -H flags, 2003-05-08), and commit 2d411398 (Stack size
in generated programs should be 10 meg, 2003-06-12) CMake adds link flags to
select a 10MB stack. At the time this was for consistency with our behavior on
MS, but that was recently removed by commit 51af1da3 (Remove "/STACK:10000000"
from default linker flags, 2012-11-23).
Change our Embarcadero link flags to select the default stack and heap settings
according to the compiler documentation. This is more reliable than leaving
the flags out completely as it has been reported that the linker does not
always use its documented defaults.
Suggested-by: Mathäus Mendel <contato@mathausmendel.com>
Previously we hard-coded a list of implicit framework directories but
did not account for CMAKE_OSX_SYSROOT or for changes to the list across
OS X versions. Instead we should automatically detect the framework
directories for the active toolchain.
The parent commit added the "-Wl,-v" option to ask "ld" to print its
implicit directories. It displays a block such as:
Framework search paths:
/...
Parse this block to extract the list of framework directories.
Detection may fail on toolchains that do not list their framework
directories, such as older OS X linkers. Always treat the paths
<sdk>/Library/Frameworks
<sdk>/System/Library/Frameworks
<sdk>/Network/Library/Frameworks # Older OS X only
/System/Library/Frameworks
as implicit. Note that /System/Library/Frameworks should always be
considered implicit so that frameworks CMake finds there will not
override the SDK copies.
We detect the implicit link directories for the toolchain by adding a
flag to get verbose output from the compiler front-end while linking the
ABI detection binary. Newer OS X toolchains based on Clang do not add
the implicit link directories with -L options to their internal
invocation of "ld". Instead they use a linker that comes with the
toolchain and is already configured with the proper directories.
Add the "-Wl,-v" option to ask "ld" to print its implicit directories.
It displays them in a block such as:
Library search paths:
/...
Parse this block to extract the implicit link directories.
While at it, remove the checks introduced by commit efaf335b (Skip
implicit link information on Xcode, 2009-07-23) and commit 5195a664
(Skip implicit link info for multiple OS X archs, 2009-09-22). Discard
the non-system link directories added by Xcode. Discard all detected
implicit libraries in the multi-architecture case but keep the
directories. The directories are still useful without the libraries
just to suppress addition of explicit -L options for them.
Modern apps that use multiple threads do NOT want 10 Megabytes of RAM
per thread being used for each thread's stack... Just leave off the
/STACK: argument, and let the compiler use a reasonable default value
for the stack size.
If existing single-threaded apps require the /STACK: argument because
they do need a very large stack size, they can add the flag in their
own CMakeLists files.
From the option documentation of VS >= 7.1:
"In earlier versions of Visual C++, the compiler used several discrete
heaps, and each had a finite limit. Currently, the compiler dynamically
grows the heaps as necessary up to a total heap size limit, and requires
a fixed-size buffer only to construct precompiled headers. Consequently,
the /Zm compiler option is rarely necessary."
http://msdn.microsoft.com/en-us/library/bdscwf1c.aspx
Suggested-by: Adam Moss <adam@broadcom.com>
Xcode 3.2.6 is known to break the SDK Library/Frameworks layout.
Detect and warn about this case to tell users to fix their system.
Reported-by: Matthew Brett <matthew.brett@gmail.com>
Since commit 43b74793 (OS X: Further improve default CMAKE_OSX_SYSROOT
selection, 2012-09-21) we choose a default CMAKE_OSX_SYSROOT only when
one is needed. However, the change forgot that we require a sysroot
when a deployment target is requested. Teach Darwin.cmake to choose a
default CMAKE_OSX_SYSROOT when CMAKE_OSX_DEPLOYMENT_TARGET is set.
Reported-by: Matthew Brett <matthew.brett@gmail.com>
Reported-by: Bradley Giesbrecht <pixilla@macports.org>
Remove ancient checks left from commit f5d95fb0 (Complete rework of
makefile generators expect trouble, 2002-11-08). Modern FreeBSD and
NetBSD platforms support shared libraries. When cross-compiling the
/usr/include/dlfcn.h may not exist on the host but the toolchain still
supports shared libraries.
In the response file also linker options could be passed,
and because <OBJECTS> is replaced by a response file, it
is necessary that no compiler option follows <OBJECTS>.
This reverts commit 5598d9b2a0.
Since commit f1670ab1 (Ninja: don't confuse ninja's rsp files with
nmake's, 2012-09-26) Ninja generator response files are placed in
CMakeFiles/ so the previously existing check already avoids expanding
them.
Since commit 1786b121 (OS X: Allow CMAKE_OSX_SYSROOT to be a logical SDK
name, 2012-09-21) we support names like "macosx" or "macosx10.7" as the
specified value of CMAKE_OSX_SYSROOT. Extend the SDK name->path
conversion to save the original value and also convert into a temporary
variable for the Xcode generator. Re-implement the deployment target
sanity check to detect the version from the transformed path.
Since commit 230ea218 (OS X: Improve default CMAKE_OSX_SYSROOT
selection, 2012-09-21) we always set CMAKE_OSX_SYSROOT if any SDK is
found in order to support Makefile generator builds with Xcode >= 4.3
without the command-line tools installed. However, in the basic
POSIX-only case of the Makefile generator with command-line tools and no
CMAKE_OSX_ARCHITECTURES we should not select any SDK by default.
Xcode supports SDKROOT values that just name an SDK rather than
specifying the full path to it. Recognize these values and handle them.
For Xcode we just put the value directly in the generated project file.
For Makefile generators we ask xcodebuild to provide the full path to
the named SDK.
Suggested-by: Jason DiCioccio <jd@ods.org>
Simplify the search for OSX_DEVELOPER_ROOT and allow it to fail if no
"/Developer" exists. When it does exist, always find a MacOSX SDK
inside it to use as the default CMAKE_OSX_SYSROOT. Otherwise set
CMAKE_OSX_SYSROOT to empty.
Drop the last use of CMAKE_OSX_SYSROOT_DEFAULT. Replace internal
platform variable CMAKE_${lang}_HAS_ISYSROOT with a more general
CMAKE_${lang}_SYSROOT_FLAG variable. If the -isysroot flag exists and
CMAKE_OSX_SYSROOT points to an SDK (not "/") then always add it to
compiler command lines. This is already done in the Xcode IDE.
Clang has the same interface as GNU except that we do not need to test
for the deployment target and sysroot flags. Simply set variables
CMAKE_${lang}_HAS_ISYSROOT
CMAKE_${lang}_OSX_DEPLOYMENT_TARGET_FLAG
to true because every version of Clang available on OS X supports these
flags.
This enables CMake to create Makefiles targeting Windows CE devices.
CMake needs to be run within a cross compile command prompt and requires
a toolchain file which sets CMAKE_SYSTEM_NAME to "WindowsCE" and
optionally CMAKE_SYSTEM_VERSION.
In commit 485a940e (VS: Simplify MSVC version reporting, 2012-08-23) we
accidentally flipped the 0/1 values of MSVC_IDE. Flip them back and
teach the CheckCompilerRelatedVariables test to check the variable.
Remove the old-style "Windows-cl.cmake" and its helper "cl.cmake". Load
the information through new-style "Platform/Windows-MSVC-<lang>.cmake"
files. Factor information common to C and CXX into a helper file
"Platform/Windows-MSVC.cmake" loaded from the per-language files.
Teach Windows-cl.cmake to use CMAKE_(C|CXX)_COMPILER_VERSION to set the
"MSVC##" and MSVC_VERSION variables. It no longer needs the IDE generator
to dictate the version or to detect the version by running the
command-line tool for NMake and Ninja generators. Drop configuration of
CMakeCPlatform.cmake and CMakeCXXPlatform.cmake from Windows-cl.cmake.in
because all the results it saved are now cheap to compute every time.
At the top of a build tree we configure inside the CMakeFiles directory
files such as "CMakeSystem.cmake" and "CMake<lang>Compiler.cmake" to
save information detected about the system and compilers in use. The
method of detection and the exact results store varies across CMake
versions as things improve. This leads to problems when loading files
configured by a different version of CMake. Previously we ignored such
existing files only if the major.minor part of the CMake version
component changed, and depended on the CMakeCache.txt to tell us the
last version of CMake that wrote the files. This led to problems if the
user deletes the CMakeCache.txt or we add required information to the
files in a patch-level release of CMake (still a "feature point" release
by modern CMake versioning convention).
Ensure that we always have version-consistent platform information files
by storing them in a subdirectory named with the CMake version. Every
version of CMake will do its own system and compiler identification
checks even when a build tree has already been configured by another
version of CMake. Stored results will not clobber those from other
versions of CMake which may be run again on the same tree in the future.
Loaded results will match what the system and language modules expect.
Rename the undocumented variable CMAKE_PLATFORM_ROOT_BIN to
CMAKE_PLATFORM_INFO_DIR to clarify its purpose. The new variable points
at the version-specific directory while the old variable did not.
Several more recent Visual Studio Express editions are now available and
they support debug builds. Simplify our VS platform files by removing
support for these old tools. If anyone still uses them we can restore
support with a more modern way to test for them.
Ancient versions of CMake required else(), endif(), and similar block
termination commands to have arguments matching the command starting the
block. This is no longer the preferred style.
Run the following shell code:
for c in else endif endforeach endfunction endmacro endwhile; do
echo 's/\b'"$c"'\(\s*\)(.\+)/'"$c"'\1()/'
done >convert.sed &&
git ls-files -z -- bootstrap '*.cmake' '*.cmake.in' '*CMakeLists.txt' |
egrep -z -v '^(Utilities/cm|Source/kwsys/)' |
egrep -z -v 'Tests/CMakeTests/While-Endwhile-' |
xargs -0 sed -i -f convert.sed &&
rm convert.sed
Since commit c198730b (Detect Watcom compiler version with its id,
2011-12-07) the CMAKE_(C|CXX)_COMPILER_VERSION variables are set for the
Watcom compiler. Use these in Windows-wcl386.cmake to set the old
WATCOM1* version variables. This avoids using the old EXECUTE_PROCESS
command which failed due to extra quotes anyway.
Teach CMake to prefer the system default compiler automatically when no
compiler is specified. By default use "cc" for C, "CC" for C++, and
"f95" for Fortran. Load a new Platform/<os>-<lang>.cmake module to
allow each platform to specify for each language its system compiler
name(s) and/or exclude certain names.
Create Platform/(CYGWIN|Darwin|Linux|Windows)-CXX.cmake modules to
specify "c++" as the system C++ compiler name for these platforms. On
systems that use case-insensitive filesystems exclude C++ compiler names
that are distinguished from C compiler names only by case.
This will change the default compiler selection for existing build
scripts that do not specify a compiler when run on machines with
separate system and GNU compilers both installed in the PATH. We do not
make this change in default behavior lightly. However:
(1) If a given build really needs specific compilers one should specify
them explicitly e.g. by setting CC, CXX, and FC in the environment.
(2) The motivating case is to prefer the system Clang on newer OS X
systems over the older GNU compilers typically also installed. On
such systems the names "cc" and "c++" link to Clang. This is the
first platform known to CMake on which "c++" is not a GNU compiler.
The old behavior selected "gcc" for C and "c++" C++ and therefore
chooses GNU for C and Clang for C++ by default. The new behavior
selects GNU or Clang consistently for both languages on older or
newer OS X systems, respectively.
(3) Other than the motivating OS X case the conditions under which the
behavior changes do not tend to exist in default OS installations.
They typically occur only on non-GNU systems with manually-installed
GNU compilers.
(4) The consequences of the new behavior are not dire. At worst the
project fails to compile with the system compiler when it previously
worked with the non-system GNU compiler. Such failure is easy to
work around (see #1).
In short this change creates a more sensible default behavior everywhere
and fixes poor default behavior on a widely-used platform at the cost of
a modest change in behavior in less-common conditions.
4bb94c9 Ninja: sysconf() is declared in unistd.h
bb36759 Ninja: enable response file support on Mac (length 262144)
3a2c8e8 Ninja: disable work around when linking with mingw
3856e66 Ninja: error on missing rspfile_content
8c1e35c Ninja: remove some unused default arguments
7f647cf Ninja: also write link libraries to rsp file
The work around is only needed by older GCCs (only testet 4.4/4.7)
Ninja is very new so chances are high that there is also a new mingw.
Use slashes in link rsp file, because ar.exe can't handle \.
Among other flags this sets RPATH flags correctly so that CMake knows
how to treat CMAKE_PLATFORM_REQUIRED_RUNTIME_PATH for the ASM language.
This is the GNU compiler equivalent to commit a0bab7ae (Add ASM platform
information for XL compiler on AIX, 2011-03-02), made for XL.
bd34963 Refactor generation of shared library flags
55d7aa4 Add platform variable for flags specific to shared libraries
31d7a0f Add platform variables for position independent code flags
Store in new platform variables
CMAKE_${lang}_COMPILE_OPTIONS_PIC
CMAKE_${lang}_COMPILE_OPTIONS_PIE
flags for position independent code generation.
In almost all cases, this means duplication of the
CMAKE_SHARED_LIBRARY_${lang}_FLAGS for the _PIC case and using the
assumed pie equivalent for the _PIE case. Note that the GNU compiler
has supported -fPIE since 3.4 and that there is no -fPIC on GNU for
Windows or Cygwin.
There is a possibility that the _PIE variables are not correct.
However, as there is no backwards compatibility to be concerned about
(as the POSITION_INDEPENDENT_CODE property is not used anywhere yet),
the current state suffices.
The default for `CMAKE_FIND_FRAMEWORK`, defined in `Darwin.cmake` and
`Darwin-icc.cmake`, is now guarded so that it will not override command line
arguments passed by users.
Similarly for `CMAKE_FIND_APPBUNDLE`
Add a boolean target property NO_SONAME which may be used to disable
soname for the specified shared library or module even if the platform
supports it. This property should be useful for private shared
libraries or various plugins which live in private directories and have
not been designed to be found or loaded globally.
Replace references to <CMAKE_SHARED_LIBRARY_SONAME_${LANG}_FLAG> and
hard-coded -install_name flags with a conditional <SONAME_FLAG> which is
expanded to the value of the CMAKE_SHARED_LIBRARY_SONAME_${LANG}_FLAG
definition as long as soname supports is enabled for the target in
question. Keep expanding CMAKE_SHARED_LIBRARY_SONAME_${LANG}_FLAG in
rules in case third party projects still use it. Such projects would
not yet use NO_SONAME so the adjacent <TARGET_SONAME> will always be
expanded. Make <TARGET_INSTALLNAME_DIR> NO_SONAME aware as well. Since
-install_name is soname on OS X, this should not be a problem if this
variable is expanded only if soname is enabled.
The Ninja generator performs rule variable substitution only once
globally per rule to put its own placeholders. Final substitution is
performed by ninja at build time. Therefore we cannot conditionally
replace the soname placeholders on a per-target basis. Rather than
omitting $SONAME from rules.ninja, simply do not write its contents for
targets which have NO_SONAME. Since 3 variables are affected by
NO_SONAME ($SONAME, $SONAME_FLAG, $INSTALLNAME_DIR), set them only if
soname is enabled.
Use of the deprecated option with Intel 2011 produces
icl: command line remark #10010: option '/GX' is deprecated and will
be removed in a future release. See '/help deprecated'
so use its replacement option which has been supported for several
older versions anyway.
The parent commit added a search path relative to OSX_DEVELOPER_ROOT.
But with Xcode 4.3 the nested Applications folder is in a different
relative location compared to that root. This commit makes the intent
of the previous commit work with older and newer Xcode directory layouts.
Furthermore, it only adds paths that exist to the search path.
Since commit 4693cf84 (Xcode: Detect new default locations of Xcode 4.3
bits and pieces) Darwin.cmake detects the developer application
directory instead of hard-coding /Developer. Replace the hard-coded
path in CMAKE_SYSTEM_APPBUNDLE_PATH using the computed result.
0f4dfa6 CPack: Use real path to PackageMaker to find its version file (#12621)
4693cf8 Xcode: Detect new default locations of Xcode 4.3 bits and pieces (#12621)
8485208 Ninja: shell escape $(CMAKE_SOURCE_DIR) and $(CMAKE_BINARY_DIR)
df84767 Ninja: add support for OBJECT_OUTPUTS, fix PrecompiledHeader test case
48eb7fc Ninja: Avoid using 'this' in member initializers
bba37dd Ninja: Fix for PDB files with spaces in the path.
ac800f4 Ninja: Constify use of cmCustomCommand
9a0d5a8 Ninja: add /DEF: flag to linker call
d40eebd Ninja: Add a cache option CMAKE_ENABLE_NINJA to enable the ninja generator.
8c63433 Ninja: Add friend struct so it can access the private ConvertToNinjaPath.
dbe3dce Ninja: add .def file support
f1bb08f Ninja: ensure the output dir exists at compile time
7a6b5f4 Ninja: Remove an unnecessary variable
80ff210 Ninja: Use cmSystemTools::ExpandListArgument to split compile/link commands
d2731a3 Ninja: Add a missed license header
eabc9b0 Ninja: CMake: Adapt Ninja generator for per-target include dirs
bada88e Merge branch 'target-include-directories' into ninja-generator
54bd175 Ninja: windows msvc: create for each target a .pdb file
...
Xcode 4.3 installs into "/Applications" by default, from the Mac App Store.
Also, the paths to the available SDKs changed: they are now within the
Xcode.app bundle.
PackageMaker is installed as a separate program, and may be installed
anywhere. It is not installed with Xcode 4.3 by default anymore.
Download the "Auxiliary Tools for Xcode" to get PackageMaker.
Put PackageMaker inside the Xcode.app bundle, in its nested Applications
folder, or put it alongside Xcode in "/Applications" and CMake will find
it.
Update references to "find" paths: add new possible locations for finding
Xcode.app and PackageMaker.app. Prefer the most recent version's locations
first, but keep the old locations as fallback search paths, too.
Thanks to all the contributors who provided and tested out various patches
for fixing this issue. Especially, but by no means limited to:
Francisco Requena Espí, Jamie Kirkpatrick and drfrogsplat.
The Borland compiler was re-branded as CodeGear during 2007-2009 and
since 2009 is the Embarcadero compiler. They offer predefined macros:
http://docwiki.embarcadero.com/RADStudio/en/Predefined_Macros
and distinguish themselves by __CODEGEARC__ and __CODEGEARC_VERSION__.
Version 6.30 (C++Builder XE) changed the meaning of some flags:
http://docwiki.embarcadero.com/RADStudio/en/C%2B%2B_Compiler_Option_Changes_for_XE
Teach Embarcadero compiler information files to generate build rules
with flags matching the compiler version. Leave the flags unchanged
for old Borland versions. Always set the BORLAND toolchain indicator
for compatibility with existing projects that test it. Also set the
EMBARCADERO indicator for newer toolchains.
The Borland compiler is now the Embarcadero compiler. Rename the shared
platform information file to reflect this. This does not change the
interface, as old versions are still "Borland", but will allow new
versions released by Embarcadero to be supported cleanly.
Fix typo introduced in commit 66a08c10 (more uniform approach to enable
language, 2004-08-26). The optimization option should be /O2 for
Release configurations and /O1 for MinSizeRel.
Suggested-by: He Yuqi <yuqi.he@gmail.com>
The default OS X 10.4 linker incorrectly searches for dependencies of
linked shared libraries only under the -isysroot location. It fails to
find dependencies of linked shared libraries in cases such as the
ExportImport test. It produces errors like:
/usr/libexec/gcc/i686-apple-darwin8/4.0.1/ld: warning can't open dynamic library:
libtestLib3Imp.dylib
referenced from: /.../ExportImport/Root/lib/libtestLib3lib.1.2.dylib
(checking for undefined symbols may be affected) (No such file or directory, errno = 2)
/usr/libexec/gcc/i686-apple-darwin8/4.0.1/ld: Undefined symbols: _testLib3Imp
referenced from libtestLib3lib expected to be defined in
libtestLib3Imp.dylib
or with CMAKE_SKIP_RPATH off to enable install_name in the Export side:
/usr/libexec/gcc/i686-apple-darwin8/4.0.1/ld: warning can't open dynamic library:
/Developer/SDKs/MacOSX10.4u.sdk/.../ExportImport/Export/impl/libtestLib3Imp.dylib
referenced from: /.../ExportImport/Export/libtestLib3lib.1.2.dylib
(checking for undefined symbols may be affected) (No such file or directory, errno = 2)
/usr/libexec/gcc/i686-apple-darwin8/4.0.1/ld: Undefined symbols:_testLib3Imp
referenced from libtestLib3lib expected to be defined in
/.../ExportImport/Export/impl/libtestLib3Imp.dylib
Note how "/Developer/SDKs/MacOSX10.4u.sdk" is prepended to the dependent
library path.
Commit 2cff26fa (Support linking to shared libs with dependent libs,
2008-01-31) and commit 82fcaebe (Pass dependent library search path to
linker on some platforms, 2008-02-01) worked around the problem by
defining platform variable CMAKE_LINK_DEPENDENT_LIBRARY_FILES. It tells
CMake to link to dependent libraries explicitly by their path thus
telling the linker where to find them.
Unfortunately the workaround had the side effect of linking dependent
libraries and defeats most benefits of LINK_INTERFACE_LIBRARIES.
Fortunately OS X 10.5 and above do not need to find transitive
dependencies at all so we can avoid the workaround on Modern OS X.
Previously we linked C, Fortran, and ASM shared libraries compiled with
the HP compiler using a direct invocation of the linker (ld). This
behavior was left historically from support for an ancient HP C compiler
that did not know how to create shared libraries. Fortran shared
libraries need to be linked with the compiler to get the language
runtime library dependencies as is already done for C++.
Update the HP-UX-HP* platform information to use the compiler front end
when linking shared libraries. This works on modern HP tools and
produces correct behavior. If there is a need to support older tools
again we can add a special case for them.
ae62a1c Test CMAKE_GNUtoMS option in ExportImport on MinGW and MSys
afb00fe Add CMAKE_GNUtoMS option to convert GNU .dll.a to MS .lib
61e8629 Factor makefile generator link rule lookup into helper function
a603250 Load platform files that need to know the ABI when possible
ecd8414 Fortran: Detect pointer size in gfortran on MinGW
Teach the Windows-GNU.cmake platform file to look for Visual Studio
tools matching the target ABI. Add an extra step to the link command
for shared libraries and executables that export symbols and on which a
new GNUtoMS property is set (initialized by the CMAKE_GNUtoMS option).
Tell the GNU linker to output a module definition (.def) file listing
exported symbols in addition to the GNU-format import library (.dll.a).
Pass the .def file to the MS "lib" tool to construct a MS-format DLL
import library (.lib).
Teach the install(TARGETS) command to install the MS import library next
to the GNU one. Teach the install(EXPORT) and export() command to set
the IMPORTED_IMPLIB property pointing at the import library to use the
import library matching the tools in the importing project.
Use __SIZEOF_POINTER__ which the GNU Fortran compiler defines at least
on 64-bit MinGW. Assume default size 4 on MinGW if gfortran does not
define the size.
Use the "-shared" option to link shared libraries. The compiler does
not support "-Wl," or "-rpath" but does know how to pass "-soname"
through to the linker.
- Build wasn't properly using -soname linker args, so installed libraries
could depend on relative paths from the build directory.
- Consolidated GNU linker args to one place in the BlueGeneP-base platform
file, since ld is used by both XL and GNU toolchains on BlueGene.
Commit 6d434ee6 (Split XL compiler information files, 2009-09-30)
added Platform/AIX-(XL|VisualAge)-(C|CXX|Fortran).cmake modules
to support the old and new compiler branding for all languages.
Add the "AIX-VisualAge-Fortran" combination that was left out
accidentally.
Allows wlib to generate proper exports if two routines have
the same spelling, but different case (like Scale and scale).
Thanks to J Decker for the patch.