The evaluation files must be known before cmTargetTraceDependencies
attempts to find them, but we must actually generate the files after
cmTargetTraceDependencies, as that can add to target SOURCES. The
limitation is that the generated output name must not depend on the
SOURCES of a target if the generated file is used by that target.
Mark the output files as GENERATED so that trace dependencies does
not expect them to already exist in the filesystem.
Move the invokation of ForceLinkerLanguage in the Generate logic
to after the generated file names are known. ForceLinkerLanguage
tries to determine the sources of a target (in order to determine
an already-known language) and otherwise fails to get information
about the generated file.
Test that the output of file(GENERATE) can be used as a target source
file and that accessing the target SOURCES in the name of the output
file is an error. Accessing the TARGET_OBJECTS would be a similar
error if it was legal to use that generator expression in this
context. That is not currently possible and is a different error
condition, so test the current error output as a reminder to change
the expected output if that becomes possible in the future. Test
that generated rule files resulting from cmTargetTraceDependencies
appear in the SOURCES generated in the output file.
Allow setting build properties based on the features available
for a target. The availability of features is determined at
generate-time by evaluating the link implementation.
Ensure that the <LANG>_STANDARD determined while evaluating
COMPILE_FEATURES in the link implementation is not lower than that
provided by the INTERFACE of the link implementation. This is
similar to handling of transitive properties such as
POSITION_INDEPENDENT_CODE.
The output of this expression may contain macros for IDEs to replace
such as $(Configuration), $(CURRENT_ARCH) etc. To avoid generating
content which is not usable in other contexts, report an error if
there is an attempt to use it in other contexts.
This commit may be reverted in the future if a solution to the
above difference is implemented.
Use the clang RemoveCStrCalls tool to automatically migrate the
code. This was only run on linux, so does not have any positive or
negative effect on other platforms.
Casts from std::string -> cmStdString were high on the list of things
taking up time. Avoid such implicit casts across function calls by just
using std::string everywhere.
The comment that the symbol name is too long is no longer relevant since
modern debuggers alias the templates anyways and the size is a
non-issue since the underlying methods are generated since it's
inherited.
The char-by-char parsing causes lots of reallocations which shouldn't be
necessary. To improve this, fast-path strings without a semicolon,
reserve space in the result, and insert into the result in chunks.
Code such as
target_include_directories(foo INTERFACE
$<INSTALL_INTERFACE:include$<FOO>>
)
should be treated as a relative directory, despite the genex, after
the INSTALL_INTERFACE is stripped away.
Previously, this would generate a relative directory on export, which
would be an error on import, so no policy is needed.
This has follow-on effects for other methods and classes. Further
work on making the use of const cmTarget pointers common can be
done, particularly with a view to generate-time methods.
Re-insert the semicolon which was removed during splitting.
Commit d777b8e7 (Genex: Allow relative paths in INSTALL_INTERFACE.,
2013-07-25) introduced the prefixItems method to allow relative paths
in the argument of the INSTALL_INTERFACE expression. That method was
buggy in that it did not re-introduce the semicolon separator in
the result.
This bug also affects paths which are already absolute in user code.
Similar incomplete generator expressions are already tested
in the GeneratorExpression unit test, but those are executed
with add_custom_target. The generator expressions in the include
directories are run through the preprocessor, whereas the ones
run through add_custom_target are not.
We can't find both preprocessing expressions at once, because then
the BUILD_INTERFACE will always be favored if both are present, even
if INSTALL_INTERFACE appears first.
This was affecting the behavior of install(EXPORT) because the
INTERFACE_INCLUDE_DIRECTORIES contained entries like
/foo/include;$<INSTALL_INTERFACE:/bar/include>
As the INSTALL_INTERFACE always evaluates to '0', it always needs
to be preprocessed properly.
Extend the range of valid target names with the + sign. This character
can commonly be used for target names, such as those containing 'c++'.
Add a test but skip it for Borland and Watcom tools which do not support
the character.
Suggested-By: Benjamin Kloster
In the case of input like
foo$<1:bar>
the preGenex should be 'foo'. In that case, the search for a ';'
will not find one, and there is no preceding input to process as a
non-genex list.
Previously, the result of 'splitting' such a string would instead
be a vector containing the same string two times.
8dfdf1c Fix the tests for evaluating includes and defines.
98a6725 Fix constness of accessors.
7e70744 Expand includes and defines transitively in 'external' genexes.
d1a2729 Fix DAG checker finding cycling dependencies.
e72eaad Workaround broken code where a target has itself in its link iface.
ec2c67b Strip stray semicolons when evaluating generator expressions.
This tracking was added during the development of commit 042ecf04
(Add API to calculate link-interface-dependent bool properties
or error., 2013-01-06), but was never used.
It was not necessary to use the content because what is really
useful in that logic is to determine if a property has been implied
to be null by appearing in a LINK_LIBRARIES genex.
I think the motivating usecase for developing the feature of
keeping track of the targets relevant to a property was that I
thought it would make it possible to allow requiring granular
compatibility of interface properties only for targets which
depended on the interface property. Eg:
add_library(foo ...)
add_library(bar ...)
add_executable(user ...)
# Read the INTERFACE_POSITION_INDEPENDENT_CODE from bar, but not
# from foo:
target_link_libraries(user foo $<$<TARGET_PROPERTY:POSTITION_INDEPENDENT_CODE>:bar>)
This obviously doesn't make sense. We require that INTERFACE
properties are consistent across all linked targets instead.
Generator expressions whose output depends on the configuration
now record that fact. The GetIncludeDirectories method can use
that result to cache the include directories for later calls.
GetIncludeDirectories is called multiple times for a target
for each configuration, so this should restore performance for
multi-config generators.
This is for specifying INCLUDE_DIRECTORIES relevant to the build-location
or the install location for example:
set_property(TARGET foo PROPERTY
INTERFACE_INCLUDE_DIRECTORIES
"$<BUILD_INTERFACE:${CMAKE_CURRENT_BINARY_DIR};${CMAKE_CURRENT_SOURCE_DIR}>"
"$<INSTALL_INTERFACE:${CMAKE_INSTALL_PREFIX}/include>"
)
A 'bar' target can then use:
set_property(TARGET bar PROPERTY
INCLUDE_DIRECTORIES
"$<TARGET_PROPERTY:foo,INTERFACE_INCLUDE_DIRECTORIES>"
)
and it will work whether foo is in the same project, or an imported target
from an installation location, or an imported target from a build location
generated by the export() command.
Because the generator expressions are only evaluated at build-time, these
new expressions are equivalent to the ZeroNode and OneNode.
The GeneratorExpression test is split into parts. Some shells can't run
the custom command as it is getting too long.
Following from the discussion here:
http://thread.gmane.org/gmane.comp.programming.tools.cmake.devel/3615/focus=5170
(Re: Generator expressisons in target properties, 26 Oct 12:10)
we can't split cmTarget API for linking into cmGeneratorTarget. In
the future we will probably also need to move the include and compile
definitions API back to cmTarget so that it can be used by export().
This fixes a regression introduced in commit 290e92ad (Move
GetIncludeDirectories to cmGeneratorTarget, 2012-09-16) which loops over
cmGeneratorTargets before they get created, so the container is empty.
There are two overloads, so that it can use the operational
target when a target property is being evaluated, and a target
can alternatively be specified by name.
At this point, the generators don't chain. That comes later.
Removing the Process() API and removing the parameters from the
constructor will allow cmGeneratorExpressions to be cached and evaluated
with multiple configs for example, such as when evaluating target
properties. This requires the creation of a new compiled representation
of cmGeneratorExpression. The cmListFileBacktrace remains in the
constructor so that we can record where a particular generator
expression appeared in the CMakeLists file.
The expressions may be parsed and then cached and evaluated multiple
times. They are evaluated lazily so that literals such as ',' can be
treated as universal parameter separators, and can be processed from
results without appearing literally, and without interfering with the
parsing/evaluation of the entire expression.
This expression evaluates to '1' or '0' to indicate whether the build
configuration for which the expression is evaluated matches tha named
configuration. In combination with the "$<0:...>" and "$<1:...>"
expressions this allows per-configuration content to be generated.
Stephen Kelly
Sean McBride
Brad King
Ben Boeckel