Custom command dependencies are followed for each target's source files
and add their transitive closure to the corresponding target. This
means that when a custom command in one target has a dependency on a
custom command in another target, both will appear in the dependent
target's sources. For the Makefile, VS IDE, and Xcode generators this
is not a problem because each target gets its own independent build
system that is evaluated in target dependency order. By the time the
dependent target is built the custom command that belongs to one of its
dependencies will already have been brought up to date.
For the Ninja generator we need to generate a monolithic build system
covering all targets so we can have only one copy of a custom command.
This means that we need to reconcile the target-level ordering
dependencies from its appearance in multiple targets to include only the
least-dependent common set. This is done by computing the set
intersection of the dependencies of all the targets containing a custom
command. However, we previously included only the direct dependencies
so any target-level dependency not directly added to all targets into
which a custom command propagates was discarded.
Fix this by computing the transitive closure of dependencies for each
target and then intersecting those sets. That will get the common set
of dependencies. Also add a test to cover a case in which the
incorrectly dropped target ordering dependencies would fail.
A common idiom in CMake-based build systems is to have custom commands
that generate files not listed explicitly as outputs so that these
files do not have to be newer than the inputs. The file modification
times of such "byproducts" are updated only when their content changes.
Then other build rules can depend on the byproducts explicitly so that
their dependents rebuild when the content of the original byproducts
really does change.
This "undeclared byproduct" approach is necessary for Makefile, VS, and
Xcode build tools because if a byproduct were listed as an output of a
rule then the rule would always rerun when the input is newer than the
byproduct but the byproduct may never be updated.
Ninja solves this problem by offering a 'restat' feature to check
whether an output was really modified after running a rule and tracking
the fact that it is up to date separately from its timestamp. However,
Ninja also stats all dependencies up front and will only restat files
that are listed as outputs of rules with the 'restat' option enabled.
Therefore an undeclared byproduct that does not exist at the start of
the build will be considered missing and the build will fail even if
other dependencies would cause the byproduct to be available before its
dependents build.
CMake works around this limitation by adding 'phony' build rules for
custom command dependencies in the build tree that do not have any
explicit specification of what produces them. This is not optimal
because it prevents Ninja from reporting an error when an input to a
rule really is missing. A better approach is to allow projects to
explicitly specify the byproducts of their custom commands so that no
phony rules are needed for them. In order to work with the non-Ninja
generators, the byproducts must be known separately from the outputs.
Add a new "BYPRODUCTS" option to the add_custom_command and
add_custom_target commands to specify byproducts explicitly. Teach the
Ninja generator to specify byproducts as outputs of the custom commands.
In the case of POST_BUILD, PRE_LINK, and PRE_BUILD events on targets
that link, the byproducts must be specified as outputs of the link rule
that runs the commands. Activate 'restat' for such rules so that Ninja
knows it needs to check the byproducts, but not for link rules that have
no byproducts.
Teach the add_custom_command and add_custom_target commands a new
USES_TERMINAL option. Use it to tell the generator to give the command
direct access to the terminal if possible.
Brad King
Peter Collingbourne