The `CMAKE_LINK_LIBRARY_FILE_FLAG` variable is meant for linkers that
want library file paths to be preceded by a flag. This is used only
for OpenWatcom to add the `library` argument before library file paths.
Refactor the approach to treat `CMAKE_LINK_LIBRARY_FILE_FLAG` as a
command-line string fragment to add just before the library file path.
This has two advantages:
* `CMAKE_LINK_LIBRARY_FILE_FLAG` now works like `CMAKE_LINK_LIBRARY_FLAG`.
* `CMAKE_LINK_LIBRARY_FILE_FLAG` can now be an attached flag whose value
is the library file path.
Technically this is a change in behavior, but this setting was created
for internal use and should be rarely used outside of CMake itself.
Fixes: #19541
The NDK build system now uses only a single toolchain in
<ndk>/toolchains/llvm/prebuilt/<host>
Its compilers are always `bin/{clang,clang++}` and its binutils are
always `bin/<triple>-*`. It is a standalone toolchain:
* The Anrdoid API level is specified at the end of `--target=`.
* The standard library may be specified via `-stdlib=`.
* No need to pass system includes or libraries explicitly.
* No need to pass `--sysroot` or `-gcc-toolchain`.
Teach CMake to recognize NDK versions that have a unified
toolchain with its own sysroot and use the above approach.
Fixes: #18739
We've long created shared objects on AIX using the linker's `-G` option
(also offered by the XL front-end). The `-G` option implies `-brtl` and
enables runtime linking. This has been largely unnecessary because we
provide all dependencies on the link line and both XL and GNU compilers
offer builtin behavior to export symbols. Since commit 0f150b69d3 (AIX:
Explicitly compute shared object exports for both XL and GNU,
2019-07-11) we compute exports explicitly and consistently.
Therefore runtime linking is no longer necessary for shared objects.
We've also long created executables on AIX using the linker's `-brtl`
option to enable runtime linking in case they load plugins at runtime.
Since commit 9f5c2040bf (AIX: Explicitly compute executable exports for
both XL and GNU, 2019-07-12) and commit 2fa920c0cd (AIX: Create import
library for executables with exports, 2019-07-16) we now provide the
linker enough information to fully resolve symbols in plugins up front.
Therefore runtime linking is no longer necessary for executables.
Drop use of `-G` for creating shared objects and use the XL `-qmkshrobj`
and GCC `-shared` options instead. Both invoke the linker with the
`-bM:SRE -bnoentry` options to create a shared object without runtime
linking enabled. Also drop use of `-brtl` for creating executables.
Issue: #19163
On AIX, plugins meant to be loaded into executables via `dlopen` must be
linked with access to a list of symbols exported from the executable in
order to use them (when not using runtime linking). The AIX linker
supports specifying this list as an "import file" passed on the command
line either via the `-bI:...` option or (with a leading `#! .` line) as
a normal input file like any other library file.
The linker import file plays the same role on AIX as import libraries do
on Windows. Teach CMake to enable its import library abstraction on AIX
for executables with the `ENABLE_EXPORTS` target property set. Teach
our internal `ExportImportList` script to optionally generate a leading
`#! .` line at the top of the generated export/import list. Update our
rule for linking an executable with exports to generate a public-facing
"import library" implemented as an AIX linker import file.
With this approach, our existing infrastructure for handling import
libraries on Windows will now work for AIX linker import files too:
* Plugins that link to their executable's symbols will be automatically
linked using the import file on the command line.
* The executable's import file will be (optionally) installed and
exported for use in linking externally-built plugins.
This will allow executables and their plugins to build even if we later
turn off runtime linking.
Issue: #19163
On AIX, symbols in executables must be exported in order to be visible
to modules (plugins) they load via `dlopen`. Prior to policy `CMP0065`,
CMake linked all executables with flags to export symbols, but the NEW
behavior for that policy is to do so only for executables that have the
`ENABLE_EXPORTS` target property set. In both cases, CMake has always
used the AIX linker option `-bexpall` option to export symbols from
executables.
This has worked fairly well with the XL compiler, but with the GNU
compiler it works only for C ABI symbols. The reason is that `-bexpall`
does not export symbols starting in `_` but the GNU C++ ABI mangles all
symbols with a leading `_`. Therefore we have only supported C ABI
plugins with the GNU compiler on AIX. Some projects have tried to work
around this by replacing `-bexpall` with `-bexpfull`, but the latter
often exports symbols that we do not want exported.
Avoid using `-bexpall` for executables by instead using by our own
internal `ExportImportList` script to compute symbol export lists from
the object files to be linked into an executable. Pass the explicitly
computed export list to the AIX linker's `-bE:...` option. We already
do this for shared object exports.
Issue: #19163
On AIX, symbols in shared objects must be exported in order to be
visible to dependents (similar to Windows). The AIX linker provides a
`-bE:...` option to specify a file listing symbols to be exported.
Compilers offer some features to help:
* When the XL compiler is invoked with its `-qmkshrobj`/`-G` options for
creating shared objects (without/with runtime linking), it recognizes
when no explicit `-bE:...` linker option is specified and runs a
`CreateExportList` tool provided with the compiler to compute one from
the object files. Since commit d468a2c2cb (XL: Avoid copying archives
into shared libraries that link them, 2011-04-07, v2.8.5~153^2) CMake
runs `CreateExportList` explicitly to ensure it only looks at the object
files and not any library files.
* When the GNU compiler is invoked with its `-shared` option for creating
shared objects, its internal `collect2` tool recognizes when no explicit
`-bE:...` linker option is specified and computes one itself from the
object files. However, it sometimes includes extra symbols such as
`.__init_aix_libgcc_cxa_atexit`.
Introduce our own internal `ExportImportList` script to compute symbol
export lists from object files. Use a basic implementation for now: it
can be extended as needed later. Update our shared library creation
rules to run the script explicitly for both the XL and GNU compilers.
Issue: #19163
We removed `-brtl` in commit bce7a2a3a5 (AIX: Do not use -brtl to create
shared libraries, 2013-03-11, v2.8.11~103^2~1) but it was added again by
commit f254276fc1 (AIX,HP-UX: Fix RPATH handling when CMP0065 is set to
NEW, 2015-12-11, v3.4.2~4^2). Since the latter commit we initialize the
`CMAKE_{SHARED,MODULE}_LINKER_FLAGS` to use the `-brtl` linker flag.
This is unnecessary because we already use the `-G` linker flag which
implies `-brtl`.
The latter commit also moved `-brtl` to `CMAKE_EXE_LINKER_FLAGS` from
flags that were always included in executable link lines with CMP0065
OLD behavior and are not part of the change intended by CMP0065. Leave
this for now as we've always enabled runtime linking for executables
(and implicitly done so via -G for shared libraries and modules).
Issue: #13997
Issue: #19163
In commit fb3370b6a1 (MSVC: Add abstraction for runtime library
selection, 2019-04-10, v3.15.0-rc1~229^2) we overlooked updating flags
for CUDA on Windows, where nvcc uses MSVC as the host compiler. Add
them now and update the MSVCRuntimeLibrary test to cover CUDA.
Fixes: #19428
In commit c4b4d8b3a6 (POSITION_INDEPENDENT_CODE: Manage link flags for
executables, 2018-10-02, v3.14.0-rc1~395^2) we accidentally removed our
Android-specific logic for PIE under the CMP0083 OLD behavior. Restore
it and also implement Android-specific logic for CMP0083 NEW behavior.
Fixes: #19393
Both CMAKE_FIND_ROOT_PATH_MODE_INCLUDE and
CMAKE_FIND_ROOT_PATH_MODE_LIBRARY are set to "ONLY" when cross
building to iOS, but appears that CMAKE_FIND_ROOT_PATH_MODE_PACKAGE
was overlooked.
This causes packages to be searched for in the host system as well,
which is incorrect and can lead to linking issues.
Set CMAKE_FIND_ROOT_PATH_MODE_PACKAGE to "ONLY" as well.
CMAKE_FIND_ROOT_PATH_MODE_PROGRAM is not touched, because a user
might want to find programs / tools on the host system.
Currently the value is hardcoded to contain only the sysroot for
the respective darwin platform. This means that it can not be changed
in a custom toolchain file.
Instead of overriding the value, simply append it. This is similar
to how it is done in the Google provided Android toolchain file.
The usecase is to allow specifying addiitonal roots to look for
3rd party packages which are definitely not present in the default
sysroot.
Currently CMAKE_MACOSX_BUNDLE is always set to true when compiling
for iOS. This poses a problem when using the source file
variant of try_compile. Even if a custom value is passed via
the CMAKE_FLAGS option, it would still be overridden by the
Darwin.cmake file.
Only set the value in case no other value was provided before.
We do not add default warning flags on other compilers, and having
a warning flag in the default flags makes it hard for projects to
customize the warning level. They need to use string processing
to remove `/W3` from `CMAKE_{C,CXX}_FLAGS`. Therefore we should
drop it.
However, projects may be using string processing to replace `/W3`
with another flag, so we cannot simply drop it. Add a policy to
drop it in a compatible way.
Fixes: #18317
Replace our hard-coded defaults for `/MD` and `/MDd` with a first-class
abstraction to select the runtime library from an enumeration of logical
names. We've long hesitated to do this because the idea of "runtime
library selection" touches on related concepts on several platforms.
Avoid that scope creep by simply defining an abstraction that applies
only when targeting the MSVC ABI on Windows.
Removing the old default flags requires a policy because existing
projects may rely on string processing to edit them and choose a runtime
library under the old behavior. Add policy CMP0091 to provide
compatibility.
Fixes: #19108
This was forgotten in commit 626c51f47b (VS: Update for Visual Studio
2019 Preview 2, 2019-01-24, v3.14.0-rc1~74^2) when the toolset was
first renumbered to `v142`.
Issue: #19125
-- Rename platform script so it runs before initial try_compile() in
project() command.
-- Fix incorrect variable name GHS_OS_DIR_OPTION
-- Remove unnecessary ".*" from REGEX expression for GHS_CANDIDATE_OS_DIRS
-- Forward GHS_OS_DIR_OPTION to try_compile() and preserve trailing
whitespace of the variable.
Since commit 11da882a12 (Apple: Introduce separate system name for iOS,
tvOS, and watchOS, 2018-01-15, v3.14.0-rc1~14^2~1) we support setting
`CMAKE_SYSTEM_NAME` to `iOS`. Existing iOS toolchain files already
set `IOS` as a short-hand variable, so do the same here.
e53a968ed5 MSVC: Use -D instead of /D in RC_FLAGS
1a281a1acd RC: Pass output file in a way that llvm-rc 7 and below understand
fa339ced67 CMakeVersion.rc: Avoid preprocessor definitions to support llvm-rc
Acked-by: Kitware Robot <kwrobot@kitware.com>
Merge-request: !3007
CrayPrgEnv:
- add a new function __cmake_craype_linktype() that determines what
link mode the Cray compiler wrapper will use in a more sophisticated
way than just MATCHing for static/dynamic on the command line.
- add a new function __cmake_craype_setupenv() that does a
once-per-cmake-run setup that does the following:
1. does a basic check of the wrapper's configuration. Running
cmake and then changing module and/or linktype configuration
may cause build problems (since the data in the cmake cache
may no longer be correct after the change). We look for this
and warn the user about it.
2. uses the "module" provided PKG_CONFIG_PATH environment variable
to add additional prefixes to the system prefix path. This
function used to be done by CrayLinuxEnvironment using the
compiler implicit include/link paths but that is intended
only for cross-compiling on Cray front-end nodes. Since
CrayPrgEnv runs on both front-end and compute nodes, we
migrate this function here.
CrayLinuxEnvironment:
- No need to set variables like CMAKE_SHARED_LIBRARY_PREFIX to values
that have already been properly established by CMakeGenericSystem.cmake.
Remove redundant sets of CMAKE_SHARED_LIBRARY_PREFIX,
CMAKE_SHARED_LIBRARY_SUFFIX, CMAKE_STATIC_LIBRARY_PREFIX,
CMAKE_STATIC_LIBRARY_SUFFIX, CMAKE_FIND_LIBRARY_PREFIXES, and
CMAKE_DL_LIBS.
- No need to add $ENV{SYSROOT_DIR}/usr/include to CMAKE_SYSTEM_INCLUDE_PATH
when we already added $ENV{SYSROOT_DIR}/usr to CMAKE_SYSTEM_PREFIX_PATH.
- Remove __cray_list_intersect(), __list_clean_dupes(), and buggy
code that adds compiler implicit includes/libs to
CMAKE_SYSTEM_INCLUDE_PATH and CMAKE_SYSTEM_LIBRARY_PATH. This
function has migrated to CrayPrgEnv.cmake, as noted above.
See discussion in issue #17413 for additional details.
Since commit 7cd65c97fa (Add CMAKE_SYSROOT variable to set --sysroot
when cross compiling., 2013-04-13, v3.0.0-rc1~342^2) we have prefixed
the value of `CMAKE_SYSROOT` to implicit include directories. This was
done because we hard-coded `/usr/include` as an implicit include
directory without accounting for the sysroot. Instead we should prefix
the hard-coded paths when they are constructed. Update the
`Platform/UnixPaths` module to do this as `Platform/Darwin` already
does.
Since commit 5990ecb741 (Compute implicit include directories from
compiler output, 2018-12-07, v3.14.0-rc1~108^2) the values of the
`CMAKE_<LANG>_IMPLICIT_INCLUDE_DIRECTORIES` variables are computed from
a real compiler invocation so they already account for the sysroot
prefix. In commit 6fc3382944 (Update logic for sysroot in detected
implicit include directories, 2019-02-13, v3.14.0-rc2~6^2) we attempted
to apply the prefix conditionally, but that is incorrect because the
compiler's real implicit include directories are not all under the
sysroot. Instead assume that all implicit include directories in
`CMAKE_<LANG>_IMPLICIT_INCLUDE_DIRECTORIES` already have the sysroot
prefix if needed. Code that constructs the value must be responsible
for that because it is the only place that knows.
The change in commit 15ad830062 (Refactor exclusion of -I/usr/include to
avoid per-language values, 2019-01-21, v3.14.0-rc1~108^2~4) caused the
exclusion to apply to Fortran, but it was only meant for C, CXX, and
CUDA. The purpose of the change was to prepare for the value of
`CMAKE_<LANG>_IMPLICIT_INCLUDE_DIRECTORIES` to be computed from the
actual compiler instead of hard-coded. We need to preserve exclusion of
`-I/usr/include` if the compiler has any implicit include directory that
looks intended to replace it, e.g. `<sdk>/usr/include` on macOS.
Fixes: #18914
In commit 1293ed8507 (ParseImplicitIncludeInfo: keep implicit incl.
consistent when rerunning cmake, 2019-01-30, v3.14.0-rc1~26^2) the
`Platform/UnixPaths` module was updated to add `/usr/include` to
`CMAKE_{C,CXX,CUDA}_IMPLICIT_INCLUDE_DIRECTORIES` through an
initialization variable used by `CMakeDetermineCompilerABI` instead of
directly. This approach makes it only a default that can be overridden
by detection of the implicit include directories really used by the
compiler.
The addition of `<sdk>/usr/include` to default implicit include
directories by the `Platform/Darwin` module needs the same update but
was accidentally left out of the original commit.
e8ee8cab97 Xcode: Completely disable code signing for compiler id detection
11da882a12 Apple: Introduce separate system name for iOS, tvOS, and watchOS
36cf44a7a3 Tests: Isolate RunCMake.XcodeProject per-device cases from host arch
Acked-by: Kitware Robot <kwrobot@kitware.com>
Merge-request: !2392
- Remove code signing requirements for non-macOS
- Do not set deployment target for non-macOS
- Build static library for compiler feature detection for non-macOS
- Use framework to run CompilerId tests for watchOS
- Port tests to new SDK handling
- Add new Apple cross-compiling section to toolchain documentation
Closes: #17870
Ninja 1.9 supports the depfile format generated by this compiler.
Use `deps = gcc` when the version of Ninja is new enough.
Unfortunately the Intel Compiler for Windows does not properly
escape spaces in paths written to a depfile so if there is a
space in the path we must still fall back to `deps = msvc`.
Fixes: #18855
The first time you run cmake, it sets the implicit include path
to the value reported by the parser (and this value gets saved
in CMake${lang}Compiler.cmake). But if you re-run cmake,
UnixPaths.cmake blindly appends an extra /usr/include to the
value saved in CMake${lang}Compiler.cmake. That should not be
harmful in most cases, but we want later runs of cmake to be
consistent with the initial one. Resolve using a solution
suggested by Brad King:
- UnixPaths now sets the default implicit include path in a new
variable named _CMAKE_${lang}_IMPLICIT_INCLUDE_DIRECTORIES_INIT
This value is only used the first time cmake is run (by
CMakeDetermineCompilerABI.cmake when it calls the implicit
include parser).
- if CMakeDetermineCompilerABI.cmake successfully calls the
implicit include parser, it overwrites the value in
_CMAKE_${lang}_IMPLICIT_INCLUDE_DIRECTORIES_INIT with the
value returned by the parser
- CMakeDetermineCompilerABI.cmake always sets
CMAKE_${lang}_IMPLICIT_INCLUDE_DIRECTORIES to the above value
of _CMAKE_${lang}_IMPLICIT_INCLUDE_DIRECTORIES_INIT
- the final value of CMAKE_${lang}_IMPLICIT_INCLUDE_DIRECTORIES gets
saved to CMake${lang}Compiler.cmake when it is regenerated after
the compiler tests are done.
- CMakeDetermineCompilerABI.cmake is only executed the first time cmake
is run. Additional runs of cmake directly load the implicit include
path from the value saved in CMake${lang}Compiler.cmake (the parser
and _INIT variable are not used).
The above depends on UnixPaths.cmake being loaded to set the _INIT value
before CMakeDetermineCompilerABI.cmake runs the implicit include parser.
Add a `CMAKE_PLATFORM_IMPLICIT_INCLUDE_DIRECTORIES` to contain the
hard-coded list of paths to be excluded from `-I` arguments so that the
values remain excluded even if the per-language
`CMAKE_<LANG>_IMPLICIT_INCLUDE_DIRECTORIES` variants change.
This is needed to preserve our historical exclusion of `-I/usr/include`
even when it is not a real implicit include directory. A policy may be
needed to remove it later.
-- standalone platforms will not build if bsp/os is specified in project file
-- integrity platforms will always print these directives because they are required
-- cleanup -os_dir setting
allow customization of the actual setting because it is determined by tool-set customization files
remove variable that was set but never used
-- add message when using default values
