Commit 089fe1c1 (Optimize genex evaluation for includes and
defines., 2013-02-01) introduced an optimization on DAG processing
to not reprocess properties on targets which have already been seen.
This was refactored slightly in commit 8dfdf1c7 (Fix the tests for
evaluating includes and defines., 2013-02-18), but was not extended
to cover COMPILE_OPTIONS in commit 80ca9c4b (Add COMPILE_OPTIONS target
property., 2013-05-16).
This omission causes the same performance regression in running
cmake on LLVM which 089fe1c1 fixed before, but this time for the
transitive evaluation of the COMPILE_OPTIONS property.
We should also check whether the INTERFACE_ variant of a property
is being read, and in the case of the compile definitions, we should
test the _<CONFIG> suffixed variants. That is already available
through the use of the methods.
This way, we use the ALREADY_SEEN optimization when evaluating
the includes of a target in 'external' generator expressions, ie, those
used in a add_custom_command invokation, as opposed to evaluating the
INCLUDE_DIRECTORIES of a target itself via GetIncludeDirectories.
Before this patch, the following is reported falsely as a self-reference:
target_link_libraries(empty2 LINK_PUBLIC empty3)
target_link_libraries(empty3 LINK_PUBLIC empty2)
add_custom_target(...
-DINCLUDES=$<TARGET_PROPERTY:empty2,INTERFACE_INCLUDE_DIRECTORIES>
)
The reason is that the existing code assumed that all reading of
include directories would be done through cmTarget::GetIncludeDirectories()
and would therefore be initialized with a DagChecker. That is not the case
if reading the property with an 'external' generator expression.
While porting boost to use these features, the generation step took
too long (several minutes before I stopped it). The reason was that
the boost libraries form a large interdependent mesh. The libraries
list their dependencies in their INTERFACE such as:
$<LINKED:boost::core>;$<LINKED:boost::config>;$<LINKED:boost::mpl>
As boost::core already depends on the boost::config libraries, that
expression has no impact on the end-content, as it is removed after
the generation step. There is no DAG issue though, so the generator
expression evaluation would fully evaluate them. In the case of the
config library, it also depends on the core library, so all depends
are followed through that again, despite the fact that they've just
been evaluated. After this patch, the evaluation skips libraries if
they have already been seen via depends or directly in the content.
This patch keeps track of targets whose INTERFACE has been consumed
already. The INCLUDE_DIRECTORIES and COMPILE_DEFINITIONS properties
are whitelisted because repeated content will be stripped out later
during generation. For other properties now and in the future, that
may not be the case.
These interface-related link-libraries properties are used to determine
the value of the other INTERFACE properties, so we were getting infinite
recursion and segfaults otherwise.
Constructs such as these are an error as they are direct self-references:
set_property(TARGET foo APPEND PROPERTY
INCLUDE_DIRECTORIES $<TARGET_PROPERTY:foo,INCLUDE_DIRECTORIES>)
set_property(TARGET foo APPEND PROPERTY
INCLUDE_DIRECTORIES $<TARGET_PROPERTY:INCLUDE_DIRECTORIES>)
However, this is an indirect self-reference in a cycle, and not an error:
set_property(TARGET foo APPEND PROPERTY
INCLUDE_DIRECTORIES $<TARGET_PROPERTY:bar,INCLUDE_DIRECTORIES>)
set_property(TARGET bar APPEND PROPERTY
INCLUDE_DIRECTORIES $<TARGET_PROPERTY:foo,INCLUDE_DIRECTORIES>)
Stephen Kelly