Module System#
VTK 9.0 introduces a new build system compared to previous versions. This version uses CMake’s built-in functionality for behaviors that were performed manually in the previous iteration of the build system.
Terminology#
module: A unit of API provided by a project. This is the core of the system and there are lots of features available through this mechanism that are not provided by CMake’s library or other usage requirements.
group: A configure-time collection of modules. These may be used to control whether member modules will be built or not with a single flag.
kit: A collection of modules for which all the compiled code is placed in a single library.
property: An attribute of a module. Only of real interest to developers of the module system and its extensions.
autoinit: A mechanism for triggering registration to global registries based on the complete set of linked-to libraries.
third party: A module representing an external dependency.
enable status: A 4-way state to allow for “weak” and “strong” selection or deselection of a module or group for building.
Principles#
The module system was designed with a number of principles in mind. These should be followed as much as possible when developing extensions as well.
The minimum CMake version required by the module system should be as low as possible to get the required features. For example, if a new feature is available in 3.15 that improves core module functionality, that’d be a reasonable reason to require it. But a bugfix in 3.10 that can be worked around should not bump the minimum version. Currently CMake 3.8 is expected to work, though various features (such as kits) are only available with newer CMake versions.
Build tree looks like the install tree. The layout of the build tree is set up to mirror the layout of the install tree. This allows more code content to be shared between build and install time.
Convention over configuration. CMake conventions should be followed. Of note, projects are assumed to be “well-behaved” including, but not limited to:
use of
BUILD_SHARED_LIBS
to control shared vs. static library compilation;use of
GNUInstallDirs
; andsensible defaults based on things like
CMAKE_PROJECT_NAME
as set by theproject()
function.
Configuration through API. Where configuration is provided, instead of using global state or “magic” variables, configuration should be provided through parameters to the API functions provided. Concessions are made for rarely-used functionality or where the API would be complicated to plumb through the required information. These variables (which are typically parameterized) are documented at the end of this document. Such variables should be named so that it is unambiguous that they are for the module system.
Don’t pollute the environment. Variables should be cleaned up at the end of macros and functions should use variable names that don’t conflict with the caller environment (usually by prefixing with
_function_name_
or the like).Relocatable installs. Install trees should not bake-in paths from the build tree or build machine (at least by default). This makes it easier to create packages from install trees instead of having to run a post-processing step over it before it may be used for distributable packages.
Build process#
Building modules involves two phases. The first phase is called “scanning” and
involves collecting all the information necessary for the second phase,
“building”. Scanning uses the vtk_module_scan()
function to search the
vtk.module files for metadata, gathers the set of modules
to build and returns them to the caller. That list of modules is eventually passed to
vtk_module_build()
which sorts the modules for their build order and then
builds each module in turn. This separation allows for scanning and building
modules in different groups. For example, the main set of modules may be scanned
to determine which of some internal set of modules are required by those which
is then scanned separately with different options.
Scanning should occur from the leaf-most module set and work its way inward to the lower levels. This is done so that modules in the lower level that are required higher up can be enabled gracefully. Builds should start at the lower level and move up the tree so that targets required by the higher groups exist when they are built.
Modules#
Modules are described by vtk.module files. These files are “scanned” using
the vtk_module_scan()
function. They provide all the information necessary for
the module system to:
provide cache variables for selecting the module (e.g.,
VTK_MODULE_ENABLE_ModuleName
);construct the dependency tree to automatically enable or disable modules based on whether it is built or not;
provide module-level metadata (such as exclusion from any wrapping and marking modules as third party)
The vtk.module files are read and “parsed”, but not executed directly. This
ensures that the module files do not contain any procedural CMake code. The
files may contain comments starting with #
like CMake code. They may either
be passed manually to vtk_module_scan()
or discovered by using the
vtk_module_find_modules()
convenience function.
The most important (and only required) parameter is the NAME
of a module.
This is used as the target name in CMake and is how the module’s target should
be referred to in all CMake code, inside the build and from the
find_package
which provides the module. To change the
name of the compiled artifact (library or executable), the LIBRARY_NAME
argument may be used.
It is highly recommended to provide a DESCRIPTION
for the module. This is
added to the documentation for the cache variable so that the user has more than
just the module name to know what the module’s purpose is.
Modules may also belong to groups which are created implicitly by adding
modules to the same-named group. Groups are listed under the GROUPS
argument
and are checked in order for a non-default setting to use.
A module may be hidden by using the CONDITION
argument. The values passed to
this field is added into a CMake if
statement and checked for validity (all
quoting is passed along verbatim). If the condition evaluates to FALSE
, the
module is treated as if it did not exist at all.
Module metadata#
A number of pieces of metadata are considered important enough to indicate them at the module level. These are used for managing slightly different workflows for modules which have these properties.
EXCLUDE_WRAP
: This marks the module with a flag that all language wrapping facilities should use to know that this module is not meant for wrapping in any language. Usually this is for modules containing user interface classes, low-level functionality, or logic that is language specific.IMPLEMENTABLE
andIMPLEMENTS
: These are used by the autoinit functionality to trigger the static factory registration calls. A module which is listed under anIMPLEMENTS
list must be marked asIMPLEMENTABLE
itself.THIRD_PARTY
: Indicates that the module represents a third party dependency. It may be internal or external to the source tree, but may be used as an additional configuration point if necessary. These modules are implicitlyEXCLUDE_WRAP
, notIMPLEMENTABLE
and do notIMPLEMENTS
any module.
Enabling modules for build#
Modules are enabled in a number of ways. These ways allow for project control
and user control of which modules should be built or not. There are 4 states for
controlling a module’s enable status
as well as a DEFAULT
setting which is used to allow for other mechanisms to select the enable status:
YES
: The module must be built.NO
: The module must not be built. If aYES
module has aNO
module in its dependency tree, an error is raised.WANT
: The module should be built. It will not be built, however, if it depends on aNO
module.DONT_WANT
: The module doesn’t need to be built. It will be built if aYES
orWANT
module depends on it.DEFAULT
: Look at other metadata to determine the status.
The first check for modules are via the REQUEST_MODULES
and REJECT_MODULES
arguments to the vtk_module_scan
function. Modules passed to
REQUEST_MODULES
are treated as if they use YES
and REJECT_MODULES
as if
they use NO
. A module may not be passed to both arguments. Modules selected
in this way do not have CMake cache variables exposed for them (since it is
assumed they are selected via some other mechanism outside the module system).
The next selector is the VTK_MODULE_ENABLE_
variable for the module. This is
added to the cache and defaults to DEFAULT
. Assuming HIDE_MODULES_FROM_CACHE
is not set to ON
, this setting is exposed in the cache and allows users to
change the status of modules not handled via the REQUEST_MODULES
and
REJECT_MODULES
mechanism.
If a module is still selected as DEFAULT
, the list of GROUPS
it is a member
of is used. In order, each group is looked at for a non-DEFAULT
value. If so,
its value is used for the module. Groups also default to using DEFAULT
for
their setting, but a project may set the _vtk_module_group_default_${group}
variable to change this default value.
After all of the above logic, if a module is still marked as DEFAULT
, the
WANT_BY_DEFAULT
argument to vtk_module_scan()
is used to determine whether
it is treated as a WANT
or DONT_WANT
request.
Now that all modules have a non-DEFAULT
enable setting, the set of modules and
kits that are available may be determined by traversing the dependency tree of
the modules.
Dependencies#
Modules have three types of dependencies:
DEPENDS
: These are dependencies which must be available and are transitively provided to modules depending on this module. The API of the module may be affected by changes in these modules. This includes, but is not limited to, classes in this module inherit or expose classes from the dependent modules.PRIVATE_DEPENDS
: Dependencies which are only used in the implementation details of the module. The API of the module is not affected by changes in these modules.OPTIONAL_DEPENDS
: Dependencies which will be used if available, but the implementation can cope with their absence. These are always treated asPRIVATE_DEPENDS
if they are available.
Modules which are listed in DEPENDS
or PRIVATE_DEPENDS
are always available
to the module and can be assumed to exist if the module is being built. Modules
listed in OPTIONAL_DEPENDS
cannot be assumed to exist. In CMake code, a
TARGET optional_depend
condition may be used to detect whether it is available
or not. The module system will add a VTK_MODULE_ENABLE_${module}
compilation
definition set to either 0
or 1
if it is available for use in the module’s
code. This flag is made preprocessor-safe by replacing any ::
in the module
name with _
. So an optional dependency on Namespace::Target
will use a flag
named VTK_MODULE_ENABLE_Namespace_Target
.
At this stage, the dependency tree for all scanned modules is traversed, marking
dependencies of YES
modules as those that should be built, marking modules
depending on NO
modules as not to be built (and triggering an error if a
conflict is found). Any WANT
modules that have not been found in the trees of
YES
or NO
modules are then enabled with their dependencies.
There is a script to help figuring out dependencies when building your own modules
or VTK-dependant code (*.cxx, *.h) in order to generate a find_package
command. The required json argument is only available in a build tree though.
Utilities/Maintenance/FindNeededModules.py -s /path/to/sources -j path/to/vtk_build/modules.json
Testing#
There is some support for testing in the module system, but it is not as comprehensive as the build side. This is because testing infrastructure and strategies vary wildly between projects. Rather than trying to handle the minimum baseline of any plausible testing infrastructure or framework, the module system merely handles dependency management for testing and entering a subdirectory with the tests.
Modules may have TEST_DEPENDS
and TEST_OPTIONAL_DEPENDS
lists provided as
well. These modules are required or optionally used by the testing code for the
module.
When scanning, the ENABLE_TESTS
argument may be set to ON
, OFF
, WANT
(the default), or DEFAULT
. Modules which appear in TEST_DEPENDS
for the
module are affected by this setting.
ON
: Modules required for testing are treated as required. Tests will be enabled.OFF
: Tests will not be enabled.WANT
: If possible,TEST_DEPENDS
modules will also be enabled if they are not disabled in some other way.DEFAULT
: Check when tests are checked whether all ofTEST_DEPENDS
are available. If they are, enable testing for the module, otherwise skip it.
The only guarantee for testing provided is that all modules in the
TEST_DEPENDS
will be available before the testing is added and
TEST_OPTIONAL_DEPENDS
are available if they’d be available at all (i.e., they
won’t be made available later).
Modules may also have TEST_LABELS
set to ease labeling all tests for the
module. The module system itself does nothing with this other than set a global
property with the value. It is up to any test infrastructure used within the
module’s CMake code to make use of the value.
The tests for a module are expected to live in a subdirectory of the module code
itself. The name of this directory is given by the TEST_DIRECTORY_NAME
argument to the vtk_module_build()
function. If the directory is available and
the module’s testing is enabled, the module system will
add_subdirectory
this directory at the appropriate
time. This is decoupled so that testing code can depend on modules that depend
on the module that is being tested and the same TARGET ${dependency}
check can
be used for optional module dependencies.
Building modules#
After scanning is complete, vtk_module_scan()
returns a list of modules and
kits to build in the variables given by the PROVIDES_MODULES
and
PROVIDES_KITS
arguments to it. It also provides lists of modules that were
found during scanning that were not scanned by that call. These are given back
in the variables passed to the UNRECOGNIZED_MODULES
and REQUIRES_MODULES
variables.
The UNRECOGNIZED_MODULES
list contains modules passed to REQUIRES_MODULES
and REJECT_MODULES
that were not found during the scan. This typically
indicates that the values passed to those arguments were not constructed
properly. However, it may also mean that they should be passed on to further
scans if they may be found elsewhere. Callers should handle the variable as
necessary for their use case.
The REQUIRES_MODULES
are modules that were named as dependencies of the
scanned modules and need to be provided in some way before building the provided
modules (the build step will require that they exist when it tries to build the
modules which required them). These can be passed on to future
REQUIRES_MODULES
arguments in future scans or used to error out depending on
the use case of the caller.
When using vtk_module_build()
, the PROVIDES_MODULES
and PROVIDES_KITS
from
a single scan should be passed together. Multiple scans may be built together as
well if they all use the same build parameters as each other.
Build-time parameters#
The vtk_module_build()
function is where the decision to build with or without
kits is decided through the BUILD_WITH_KITS
option. Only if this is set will
kits be built for this set of modules.
The decision to default third party modules to using an external or internal
copy (where such a decision is possible) is done using the USE_EXTERNAL
argument.
Where build artifacts end up in the build tree are left to CMake’s typical variables for controlling these locations:
The defaults for these place outputs into the binary directory where the targets
were added. The module system will set these to be sensible for itself if they
are not already set, but it is recommended to set these at the top-level so that
targets not built under vtk_module_build()
also end up at a sensible location.
Library parameters#
When building libraries, it is sometimes useful to have top-level control of library metadata. For example, VTK suffixes its library filenames with a version number. The variables that control this include:
LIBRARY_NAME_SUFFIX
: If non-empty, all libraries and executable names will be suffixed with this value prefixed with a hyphen (e.g., a suffix offoo
will makeNamespace::Target
’s library be namedTarget-foo
or, if the module sets itsLIBRARY_NAME
tonsTarget
,nsTarget-foo
).VERSION
: Controls theVERSION
property for all library modules.SOVERSION
: Controls theSOVERSION
property for all library modules.
Installation support#
vtk_module_build()
also offers arguments to aid in installing module
artifacts. These include destinations for pieces that are installed, CMake
packaging controls, and components to use for the installations.
A number of destinations control arguments are provided:
ARCHIVE_DESTINATION
HEADERS_DESTINATION
LIBRARY_DESTINATION
RUNTIME_DESTINATION
CMAKE_DESTINATION
LICENSE_DESTINATION
HIERARCHY_DESTINATION
See the API documentation for default values for each which are based on
GNUInstallDirs
variables. Note that all installation
destinations are expected to be relative paths. This is because the conveniences
provided by the module system are all assumed to be installed to a single prefix
(CMAKE_INSTALL_PREFIX
) and placed underneath it.
Suppression of header installation is provided via the INSTALL_HEADERS
argument to vtk_module_build()
. Setting this to OFF
will suppress the
installation of:
headers
CMake package files
hierarchy files (since their use requires headers)
Basically, suppression of headers means that SDK components for the built modules are not available in the install tree.
Components for the installation are provided via the HEADERS_COMPONENT
and
TARGETS_COMPONENT
arguments. The former is used for SDK bits and the latter
for runtime bits (libraries, executables, etc.).
For CMake package installation, the PACKAGE
and INSTALL_EXPORT
arguments are
available. The former controls the names used by the CMake files created by the
module system while the former is the export set to use for the member modules
when creating those CMake files. Non-module targets may also exist in this
export set when vtk_module_build()
is called, but the export set is considered
“closed” afterwards since it has already been exported (if INSTALL_HEADERS
is
true).
Test data information#
The directory that is looked for in each module is specified by using the
TEST_DIRECTORY_NAME
argument. If it is set to the value of NONE
, no testing
directories will be searched for. It defaults to Testing
due to VTK’s
conventions.
The module system, due to VTK’s usage of it, has convenience parameters for
controlling the ExternalData
module that is available to
testing infrastructure. These include:
TEST_DATA_TARGET
: The data target to use for tests.TEST_INPUT_DATA_DIRECTORY
: WhereExternalData
should look for data files.TEST_OUTPUT_DATA_DIRECTORY
: WhereExternalData
should place the downloaded data files.TEST_OUTPUT_DIRECTORY
: Where tests should place output files.
Each is provided in the testing subdirectory as _vtk_build_${name}
, so the
TEST_DATA_TARGET
argument is available as _vtk_build_TEST_DATA_TARGET
.
Building a module#
Building a module is basically the same as a normal CMake library or executable, but is wrapped to use arguments to facilitate wrapping, exporting, and installation of the tools as well.
There are two main functions provided for this:
The former creates a library for the module being built while the latter can
create an executable for the module itself or create utility executable
associated with the module. The module system requires that the CMakeLists.txt
for a module create a target with the name of the module. In the case of
INTERFACE
modules, it suffices to create the module manually in many cases.
Libraries#
Most modules end up being libraries that can be linked against by other
libraries. Due to cross-platform support generally being a good thing, the
EXPORT_MACRO_PREFIX
argument is provided to specify the prefix for macro names
to be used by GenerateExportHeader
. By default, the
LIBRARY_NAME
for the module is transformed to uppercase to make the prefix.
Some modules may need to add additional information to the library name that
will be used that is not statically know and depends on other environmental
settings. The LIBRARY_NAME_SUFFIX
may be specified to add an additional suffix
to the LIBRARY_NAME
for the module. The vtk_module_build()
LIBRARY_NAME_SUFFIX
argument value will be appended to this name as well.
Normally, libraries are built according to the
BUILD_SHARED_LIBS
variable, however, some modules may
need to be built statically all the time. The FORCE_STATIC
parameter exists
for this purpose. This is generally only necessary if the module is in some
other must-be-static library’s dependency tree (which may happen for a number of
reasons). It is not an escape hatch for general usage; it is there because use
cases which only support static libraries (even in a shared build) exist.
If a library module is part of a kit and it is being built via the
vtk_module_build()
BUILD_WITH_KITS
argument, it will be built as an
OBJECT
library and the kit machinery in vtk_module_build()
will create the resulting kit library artifact.
Header-only modules must pass HEADER_ONLY
to create an INTERFACE
library
instead of expecting a linkable artifact.
Note
HEADER_ONLY
modules which are part of kits is currently untested. This
should be supported, but might not work at the moment.
Source listing#
Instead of using CMake’s “all sources in a single list” pattern for
add_library
, vtk_module_add_module()
classifies its source files explicitly:
SOURCES
HEADERS
TEMPLATES
The HEADERS
and TEMPLATES
are installed into the HEADERS_DESTINATION
specified to vtk_module_build()
and may be added to a subdirectory of this
destination by using the HEADERS_SUBDIR
argument. Note that the structure of
the header paths passed is ignored. If more structure is required from the
installed header layout, vtk_module_install_headers()
should be used.
Files passed via HEADERS
are treated as the API interface to the code of the
module and are added to properties so that language wrappers can
discover the API of the module.
Note
Only headers passed via HEADERS
are eligible for wrapping; those
installed via vtk_module_install_headers()
are not. This is a known limitation
at the moment.
There are also private variations for HEADERS
and TEMPLATES
named
PRIVATE_HEADERS
and PRIVATE_TEMPLATES
respectively. These are never
installed nor exposed to wrapping mechanisms.
There are also a couple of convenience parameters that use VTK’s file naming conventions to ease usage. These include:
CLASSES
: For each value<class>
, adds<class>.cxx
toSOURCES
and<class>.h
toHEADERS
.TEMPLATE_CLASSES
: For each value<class>
, adds<class>.txx
toTEMPLATES
and<class>.h
toHEADERS
.PRIVATE_CLASSES
: For each value<class>
, adds<class>.cxx
toSOURCES
and<class>.h
toPRIVATE_HEADERS
.PRIVATE_TEMPLATE_CLASSES
: For each value<class>
, adds<class>.txx
toPRIVATE_TEMPLATES
and<class>.h
toPRIVATE_HEADERS
.
Executables#
Executables may be created using vtk_module_add_executable()
. The first
argument is the name of the executable to build. Since the scanning phase does
not know what kind of target will be created for each module (and it may change
based on other configuration values), an executable module which claims it is
part of a kit raises an error since this is not possible to do.
For modules that are executables using this function, the metadata from the module information is used to set the relevant properties. The module dependencies are also automatically linked in the same way as a library module would do so.
For utility executables, NO_INSTALL
may be passed to keep it within the build
tree. It will not be available to consumers of the project. If the name of the
executable is different from the target name, BASENAME
may be used to change
the executable’s name.
Module APIs#
All of CMake’s target_
function calls have analogues for
modules. This is primarily due to the kits feature which causes the target name
created by the module system that is required to use the target_
functions
dependent on whether the module is a member of a kit and kits are being built.
The CMake version of the function and the module API analogue (as well as
differences, if any) is:
add_dependencies becomes
vtk_module_depend()
target_compile_features becomes
vtk_module_compile_features()
target_link_libraries becomes
vtk_module_link()
: When kits are enabled, anyPRIVATE
links are forwarded to the kit itself. This necessitates making all of these targets globally scoped rather than locally scoped.
Packaging support#
Getting installed packages to work for CMake is, unfortunately, not trivial. The module system provides some support for helping with this, but it does place some extra constraints on the project so that some assumptions that vastly simplify the process can be made.
Assumptions#
The main assumption is that all modules passed to a single vtk_module_build()
have the same CMake namespace (the part up to and including the ::
, if any,
in a module name. For exporting dependencies, that namespace matches the
PACKAGE
argument for vtk_module_build()
. These are done so that the
generated code can use
CMAKE_FIND_PACKAGE_NAME
variable can be
used to discover information about the package that is being found.
The package support also assumes that all modules may be queried using
COMPONENTS
and OPTIONAL_COMPONENTS
and that the component name for a module
corresponds to the name of a module without the namespace.
These rules basically mean that a module named Namespace::Target
will be found
using find_package(Namespace)
, that COMPONENTS Target
may be passed to
ensure that that module exists, and OPTIONAL_COMPONENTS Target
may be passed
to allow the component to not exist while not failing the main
find_package
call.
Creating a full package#
The module system provides no support for the top-level file that is used by
find_package
. This is because this logic is highly
project-specific and hard to generalize in a useful way. Instead, files are
generated which should be included from the main file.
Here, the list of files generated are based on the PACKAGE
argument passed to
vtk_module_build()
:
<PACKAGE>-targets.cmake
: The CMake-generated export file for the targets in theINSTALL_EXPORT
.<PACKAGE>-vtk-module-properties.cmake
: Properties for the targets exported into the build.
The module properties file must be included after the targets file so that they exist when it tries to add properties to the imported targets.
External dependencies#
Since the module system is heavily skewed towards using imported targets, these
targets show up by name in the find_package
of the
project as well. This means that these external projects need to be found to
recreate their imported targets at that time. To this end, there is the
vtk_module_export_find_packages()
function. This function writes a file named
according to its FILE_NAME
argument and place it in the build and install
trees according to its CMAKE_DESTINATION
argument.
This file will be populated with logic to determine whether third party packages
found using vtk_module_find_package()
are required during the
find_package
of the package or not. It will forward
REQUIRED
and QUIET
parameters to other find_package
calls as necessary based on the REQUIRED
and QUIET
flags for the package
and whether that call is involved in a non-optional COMPONENT
(a
component-less find_package
call is assumed to mean
“all components”).
This file should be included after the <PACKAGE>-vtk-module-properties.cmake
file generated by the vtk_module_build()
call so that it can use the module
dependency information set via that file.
After this file is included, for each component that it checks, it will set
${CMAKE_FIND_PACKAGE_NAME}_<component>_FOUND
to 0 if it is not valid and
append a reason to ${CMAKE_FIND_PACKAGE_NAME}_<component>_NOT_FOUND_MESSAGE
so that the package can collate the reason why things are not available.
Setting the _FOUND
variable#
The module system does not currently help in determining the top-level
${CMAKE_FIND_PACKAGE_NAME}_FOUND
variable based on the results of the
components that were requested and the status of dependent packages. This may be
provided at some point, but there has not currently been enough experience to
determine what patterns are available for factoring it out as a utility
function.
The general pattern should be to go through the list of components requested,
determine whether targets for those components exist. Then for each found
component, use the module dependency information to ensure that all targets in
the dependency trees are found (propagating not-found statuses through the
dependency tree). The ${CMAKE_FIND_PACKAGE_NAME}_NOT_FOUND_MESSAGE
should be
built up based on the reasons the find_package
call did
not work based on these discoveries.
This is the process for modules in a package, but packages may contain non-module components, and it is hard for the module system to provide support for them, so they are not attempted. See the CMake documentation for more details about creating a package configuration.
Advanced topics#
There are a number of advanced features provided by the module system that are not normally required in a simple project.
Kits#
Kits are described in vtk.kit files which act much like vtk.module
files. However, they only have NAME
, LIBRARY_NAME
, and DESCRIPTION
fields.
These all act just like they do in the vtk.module
context. These files may
either be passed manually to vtk_module_scan()
or discovered by using the
vtk_module_find_kits()
convenience function.
Before a module may be a member of a kit, a vtk.kit must declare it and be scanned at the same time. This means that kits may only contain modules that are scanned with them and cannot be extended later nor may kits be made of modules that they do not know about.
Requirements#
In order to actually use kits, CMake 3.12 is necessary in order to do the
OBJECT
library manipulations done behind the scenes to make it
Just Work. 3.8 is still the minimum version for using a project that is built
with kits however. This is only checked when kits are actually in use, so
projects requiring older CMake versions as their minimum version may still
provide kits so that users with newer CMake versions can use them.
Kits create a single library on disk, but the usage requirements of the modules should still be the same (except for that which is inherently required to be different by combining libraries). So include directories, compile definitions, and other usage requirements should not leak from other modules that are members of the same kit.
Autoinit#
The module system supports a mechanism for triggering static code construction
for modules which require it. This cannot be done through normal CMake usage
requirements because the requirements are intersectional. For example, a module
F
having a factory where module I
provides an implementation for it means
that a target linking to both F
and I
needs to ensure that I
registers its
implementation to the factory code. There is no such support in CMake and due to
the complexities and code generation involved with this support, it is unlikely
to exist.
Code which uses modules may call the vtk_module_autoinit()
function to use
this functionality. The list of modules passed to the function are used to
compute the defines necessary to trigger the registration to factories when
necessary.
For details on the implementation of the autoinit system, please see the relevant section in the API documentation.
Wrapping#
VTK comes with support for wrapping its classes into other languages. Currently, VTK supports wrapping its classes for use in the Python and Java languages. In order to wrap a set of modules for a language, a separate function is used for each language.
All languages read the headers of classes with a __VTK_WRAP__
preprocessor
definition defined. This may be used to hide methods or other details from the
wrapping code if wanted.
Python#
For Python, the vtk_module_wrap_python()
function must be used. This function
takes a list of modules in its MODULES
argument and creates Python modules
for use under the PYTHON_PACKAGE
package. No __init__.py
for this package
is created automatically and must be provided in some other way.
A target named by the TARGET
argument is created and installed. This target
may be linked to in order to be able to import static Python modules. In this
case, a header and function named according to the basename of TARGET
(e.g.,
VTK::PythonWrapped
has a basename of PythonWrapped
) must be used. The
header is named <TARGET_BASENAME>.h
and the function which adds the wrapped
modules to the static import table is <void TARGET_BASENAME>_load()
. This
function is also created in shared builds, but does nothing so that it may
always be called in static or shared builds.
The modules will be installed under the MODULE_DESTINATION
given to the
function into the PYTHON_PACKAGE
directory needed for it. The
vtk_module_python_default_destination()
function is used to determine a
default if one is not passed.
The Python wrappers define a __VTK_WRAP_PYTHON__
preprocessor definition when
reading code which may be used to hide methods or other details from the Python
wrapping code.
Java#
For Java, the vtk_module_wrap_java()
function must be used. This function
creates Java sources for classes in the modules passed in its MODULES
argument. The sources are written to a JAVA_OUTPUT
directory. These then can
be compiled by CMake normally.
For this purpose, there are <MODULE>Java
targets which contain a
_vtk_module_java_files
properties containing a list of .java
sources
generated for the given module. There is also a <MODULE>Java-java-sources
target which may be depended upon if just the source generation needs to used
in an add_dependencies
call.
The Java wrappers define a __VTK_WRAP_JAVA__
preprocessor definition when
reading code which may be used to hide methods or other details from the Java
wrapping code.
Hierarchy files#
Hierarchy files are used by the language wrapper tools to know the class
inheritance for classes within a module. Each module has a hierarchy file
associated with it. The path to a module’s hierarchy file is stored in its
hierarchy
module property.
Third party#
The module system has support for representing third party modules in its
build. These may be built as part of the project or represented using other
mechanisms (usually find_package
and a set of imported
targets from it).
The primary API is vtk_module_third_party()
which creates a
VTK_MODULE_USE_EXTERNAL_Namespace_Target
option for the module to switch
between an internal and external source for the third party code. This value
defaults to the setting of the USE_EXTERNAL
argument for the calling
vtk_module_build()
function. Arguments passed under the INTERNAL
and
EXTERNAL
arguments to this command are then passed on to
vtk_module_third_party_internal()
or vtk_module_third_party_external()
,
respectively, depending on the VTK_MODULE_USE_EXTERNAL_Namespace_Target
option.
Note that third party modules (marked as such by adding the THIRD_PARTY
keyword to a vtk.module
file) may not be part of a kit, be wrapped, or
participate in autoinit.
External third party modules#
External modules are found using CMake’s find_package
mechanism. In addition to the arguments supported by
vtk_module_find_package()
(except PRIVATE
and
PRIVATE_IF_SHARED
), information about the found package is used to construct
a module target which represents the third party package. The preferred
mechanism is to give a list of imported targets to the LIBRARIES
argument.
These will be added to the INTERFACE
of the module and provide the third
party package for use within the module system.
If imported targets are not available (they really should be created if not),
variable names may be passed to INCLUDE_DIRS
, LIBRARIES
, and DEFINITIONS
to create the module interface.
In addition, any variables which should be forwarded from the package to the
rest of the build may be specified using the USE_VARIABLES
argument.
The STANDARD_INCLUDE_DIRS
argument creates an include interface for the
module target which includes the “standard” module include directories to.
Basically, the source and binary directories of the module.
Internal third party modules#
Internal modules are those that may be built as part of the build. These should
ideally specify a set of LICENSE_FILES
indicating the license status of the
third party code. These files will be installed along with the third party
package to aid in any licensing requirements of the code. It is also
recommended to set the VERSION
argument so that it is known what version of
the code is provided at a glance.
By default, the LIBRARY_NAME
of the module is used as the name of the
subdirectory to include, but this may be changed by using the SUBDIRECTORY
argument.
Header-only third party modules may be indicated by using the HEADER_ONLY
argument. Modules which represent multiple libraries at once from a project may
use the INTERFACE
argument.
The STANDARD_INCLUDE_DIRS
argument creates an include interface for the
module target which includes the “standard” module include directories to.
Basically, the source and binary directories of the module. A subdirectory may
be used by setting the HEADERS_SUBDIR
option. It is implied for
HEADERS_ONLY
third party modules.
After the subdirectory is added a target with the module’s name must exist.
However, a target is automatically created if it is HEADERS_ONLY
.
Properly shipping internal third party code#
There are many things that really should be done to ship internal third party code (also known as vendoring) properly. The issue is mainly that the internal code may conflict with other code bringing in another copy of the same package into a process. Most platforms do not behave well in this situation.
In order to avoid conflicts at every level possible, a process called “name mangling” should be performed. A non-exhaustive list of name manglings that must be done to fully handle this case includes:
moving headers to a subdirectory (to avoid compilations from finding incompatible headers);
changing the library name (to avoid DLL lookups from finding incompatible copies); and
mangling symbols (to avoid symbol lookup from confusing two copies in the same process).
Some projects may need further work like editing CMake APIs or the like to be mangled as well.
Moving headers and changing library names is fairly straightforward by editing
CMake code. Mangling symbols usually involves creating a header which has a
#define
for each public symbol to change its name at runtime to be distinct
from another copy that may end up existing in the same process from another
project.
Typically, a header needs to be created at the module level which hides the differences between third party code which may or may not be provided by an external package. In this case, it is recommended that code using the third party module use unmangled names and let the module interface and mangling headers handle the mangling at that level.
Debugging#
The module system can output debugging information about its inner workings by
using the _vtk_module_log
variable. This variable is a list of “domains” to
log about, or the special ALL
value causes all domains to log output. The
following domains are used in the internals of the module system:
kit
: discovery and membership of kitsmodule
: discovery andCONDITION
results of modulesenable
: resolution of the enable status of modulesprovide
: determination of module provisionbuilding
: when building a module occurstesting
: missing test dependencies
It is encouraged that projects expose user-friendly flags to control logging
rather than exposing _vtk_module_log
directly.
Control variables#
These variables do not follow the API convention and are used if set:
_vtk_module_warnings
: If enabled, “strict” warnings are generated. These are not strictly problems, but may be used as linting for improving usage of the module system._vtk_module_log
: A list of “domains” to output debugging information._vtk_module_group_default_${group}
: used to set a non-DEFAULT
default for group settings.
Some mechanisms use global properties instead:
_vtk_module_autoinit_include
: The file that needs to be included in order to make theVTK_MODULE_AUTOINIT
symbol available for use in the autoinit support.
SPDX files generation#
The generation of VTK module SPDX files relies on three components:
SPDX arguments in
vtk_module_build()
SPDX arguments in each vtk.module
SPDX Tags in the sources files
SPDX files are named after <ModuleName>.spdx
and are generated for
all VTK modules.
Generated SPDX files are based on the SPDX 2.2 specification.
If some information is missing, VTK will warn during configuration or during build
but the SPDX file will still be generated with unknown fields being attributed a
NOASSERTION
or other default value.
The collected license identifiers are joined together using AND
keyword.
Similarly all collected copyright texts are joined using a new line.
SPDX arguments in vtk_module_build
#
Support for SPDX file generation requires to specify the following
vtk_module_build()
arguments:
GENERATE_SPDX
SPDX_DOCUMENT_NAMESPACE
SPDX_DOWNLOAD_LOCATION
GENERATE_SPDX
is used to enable the generation and install of SPDX file for
each modules. Set this to ON
to enable it.
SPDX_DOCUMENT_NAMESPACE
is used as a basename for the DocumentNamespace
SPDX field. The name of the module will simply be appended to the basename.
If not provided, https://vtk.org/spdx
will be used. This is the value VTK
project uses as well. Note that the namespace does not need to be an actual
website URL, but just a unique Uniform Resource Identifier (URI).
Caution
If VTK decide to host SPDX files in the future, the namespace in use for the VTK SPDX files may change accordingly.
SPDX_DOWNLOAD_LOCATION
is used as a basename for the PackageDownloadLocation
when not provided at module level. The relative path to the module will simply
be appended in order to generate the actual PackageDownloadLocation
SPDX field.
If not provided at module or in vtk_module_build()
, NOASSERTION
will be used.
SPDX arguments in vtk.module
#
Defining these three arguments in vtk.module is required:
SPDX_LICENSE_IDENTIFIER
SPDX_COPYRIGHT_TEXT
SPDX_DOWNLOAD_LOCATION
SPDX_LICENSE_IDENTIFIER
is an expected field corresponding to the PackageLicenseDeclared
SPDX field that is considered as the global license for all files of the
module that are not parsed during generation. This field is used to set
the PackageLicenseConcluded
SPDX field.
Note
The SPDX generation system do not and cannot replace the LICENSE_FILES
mechanism.
Indeed, certains license (e.g Apache 2.0) requires additional files (e.g NOTICE
) to
also be distributed.
SPDX_COPYRIGHT_TEXT
is an expected field that correspond to the copyright applying
to all files that are not parsed during generation, it is used to generate PackageCopyrightText
.
SPDX_DOWNLOAD_LOCATION
is a optional field for modules (see above for setting
this in vtk_module_build
) and expected field for third parties.
If provided, it is used as is for the PackageDownloadLocation
SPDX field.
SPDX arguments in vtk_module_add_module
#
It is possible to specify a SPDX_SKIP_REGEX
when adding a module in order
to skip specific file during SPDX tags parsing.
It is a python regex which is used to match with the filename of the source files.
Custom license support#
If the VTK module contains a custom license that is not part of the SPDX license list then adding a custom license may be needed.
The SPDX generation system support to specify exactly one custom license by module, supplemental to standard licenses.
The text of this license should be made available in a file and added to the module definition using SPDX_CUSTOM_LICENSE_FILE
, the name of the license should be specified using SPDX_CUSTOM_LICENSE_NAME
(eg: LicenseName
and the SPDX_LICENSE_IDENTIFIER
for this license
should be LicenseRef-
followed by the name (eg: LicenseRef-licenseName
). See this entry for more info.
Note
If this custom license is to be added to VTK proper, it must be compatible with the BSD-3-Clause license of VTK and not add more restriction to the code.
Limitations#
Correctness of the
SPDX-FileCopyrightText
andSPDX-License-Identifier
tags is not ensured. The value will be used as is.The generated SPDX files only include the Package information section. This means that there are no File information sections describing source files or build artifacts.
Third party source files are not parsed for SPDX tags.
Adding empty lines between
// SPDX-FileCopyrightText
and// SPDX-License-Identifier
tags is not supported.Certain files are not parsed at all, eg: cmake files, python files, test files, …