On Tue, Mar 31, 2009 at 3:27 PM, BRM <bm_witness at yahoo.com> wrote:
Notice my original API suggestion - the project controls its own header - just not the list of available items.
So essentially:
1) Cmake runs, finds packages, builds list
2) user add extra items to list
3) header generated
Noting from my original API example:
cmake_autoheader(C, path/to/cmake/autoheader_output1.h some_other_package )
cmake_autoheader(C, path/to/cmake/autoheader_output2.h some_other_package )
cmake_autoheader(C, path/to/cmake/autoheader_output3.h some_other_package )
Adding in the filtering as I suggested:
cmake_autoheader(C, path/to/cmake/autoheader_output1.h some_other_package [filter 1])
cmake_autoheader(C, path/to/cmake/autoheader_output2.h some_other_package [filter 2])
cmake_autoheader(C, path/to/cmake/autoheader_output3.h some_other_package [filter 3])
Each project could easily define its own header (or headers) - there would be no limit. And it should probably generate the header immediately based on when the cmake_autoheader() is called, based on the state of the list at that point. (I think it would be too complex to push it until the end of the file; but that might be possible too.)
As to resolving between sub-projects, it would likely be good to internally register a scope for the variables - the scope being the path from the root CMakeLists.txt to the present one. If a parent, sibling, or child node wanted to access a different node's variables, it would have to do so using the filters and explicitly name the path. This would be easy to do in the filters as the project designer would know the paths between the projects. The output in the header would be no different - just the #define HAS_VARIABLE no matter what node it was in.
So take the following:
src/CMakeLists.txt
src/subdir1/CMakeLists.txt
src/subdir1/subdir1a/CMakeLists.txt
src/subdir2/CMakeLists.txt
src/subdir1/CMakeLists.txt would be able to access its variables with no scoping whatsoever. But to access a variable in src/CMakeLists.txt, it would have to provide explicit scoping to it via a filter when generating the autoheader via cmake_autoheader(). It would need to do the same to access variables in either src/subdir1/subdir1a/CMakeLists.txt and src/subdir2/CMakeLists.txt.
However, I would only guarantee that the cmake_autoheader() would be able to get what is in the global list at the time it is run. So src/subdir1/CMakeLists.txt would likely not be able to see src/subdir2/CMakeLists.txt's variables, or all of the variables in src/CMakeLists.txt (as some may be defined after src/subdir1/CMakeLists.txt is run). It could control what it or its child nodes know though - based on when it makes the call, but that can be left to the project designer.
Quote Philip Lowman <philip at yhbt.com>:
===============================
Why not simply make cmake_autoheader() reentrant (and I assume a command) and have it process last? This would solve one of your critiques of not having to pass lists around. You could also eliminate the need to specify the filename everywhere and a logical name could be used. It would presumably call the functionality provided by configure_file() internally per logical name.
cmake_autoheader(INIT myconfig ${CMAKE_CURRENT_BINARY_DIR}/config.h)
cmake_autoheader(CONFIG myconfig
... stuff like we already have ...)
add_subdirectory(foo)
foo/ => cmake_autoheader(CONFIG myconfig
... add additional stuff)
Ben's Reply:
=========
Only issue I can see with it being re-entrant (good idea overall) is that you would be limited to having one file. Otherwise love it.
Quote Philip Lowman <philip at yhbt.com>:
===============================
I hear you on trying to automatically add #cmakedefine variables to a list. Ultimately I think Bill and co. are right that trying to retrofit find_package() and any other place where you may (or may not) want to add a #cmakedefine like check_function_exists, etc. is simply going to be too complex. Also as for find_package() someone out there will certainly want HAVE_OPENSCENEGRAPH to work instead of HAVE_OSG or HAVE_OSG2 or whatever. You seem to like HAS_FOO, others probably prefer USE_FOO or HAVE_FOO_H. Clearly we're never all going to agree.
Even if it was easy to do this there is another risk. Users won't care to really define which variables they want #cmakedefined or not so they'll simply define everything. Then you'll end up with what GNU autotools seems to do now which is to create a ginormous config.h.in file for pretty much anything you type in configure.ac. Trying to wade through which of these #defines are actually necessary or not is an annoying exercise in the use of grep. Also many of these #defines are superfluous on modern compilers/platforms.
Ben's Reply:
=========
If you noticed my original API example, I had the Find function use its own variable for defining the variable. I honestly could care less what it is actually called. There is already precedence (I think) for FOO_FOUND. I quoted 'HAS_FOO' only to keep the discussion simple as someone quoted that from the Autotools equivalent functionality. But what it is called in the end makes no difference to me. All I care about is that there is consistency between projects so developer's don't have to either (a) figure out what something is called, or (b) figure out what to call something. Simplicity.
I proposed having it done by the Find functions as that would provide a unified area for any given library, and thus minimize overlapping defines. It would be easier to move between projects, and re-use would be better. Simplicity.
As to whether a user wants to call something else by the same name - well, I'd have to say "tough luck", and the precedences seems to already be set by the Find functions - they already define a series of variables that the user cannot change and if they conflict with variables the user wants, then its "touch luck, use something else". So I don't think that either is or will be an issue - it certainly doesn't seem to have been already. Precedence is already there to what I can tell. (http://www.cmake.org/Wiki/CMake_HowToFindInstalledSoftware, http://www.cmake.org/cmake/help/cmake2.6docs.html#command:find_package)
I also don't think that would be overly complex - it's just updating the various Find functions to add a call to the correct API.
So would you rather that a project designer have to bloat their CMakeFiles.txt with something like:
if (FOUND_FOO)
cmake_autoheader(CONFIG myconfig HAS_FOO)
for every package they want found? (So if you have 30 packages, you'd have 60 lines added)
Or would it be better to only have to define it once (like a project), and then have everything add to that and only have one other line required to generate the header?
Simplicity is best, and I think the simplest solution means not having user's bloat their CMakeFiles.txt - it should be part of the system provided by CMake.
As another POV, suppose you are working on 10 projects. You started 5 of them, and at different times joined 5 others. If you relied on user's to define the variables for the header, then you could theoretically have 10 different variables. Or if you were smart about your own projects, you'd have 4 - 1 for 1 project + your 5, and 3 others. If were absolutely lucky against all odds, you'd have 1 - but that would likely only occur if you had a common participant on all projects OR if someone reviewed all the projects and said "Let's make these all the same". Imagine having to move between those projects? If they had different variables (which is more likely than not), then you'd have to remember which variable was for which project.
Now consider the POV where the system predefines what the variable format would be, and what the variable names are for non-project-specific variables - e.g. stuff that CMake finds FOR your project via the Find Functions. In that same scenario, you'd have 1 variable for the same package on all projects.
Ultimately there are three choices to make:
1) Do we want to have a reliable naming convention for variables in the API?
2) Do we want to have reliable, cross-project names for the same project in the API?
3) Do we want to make it easy for users to use the API?
I think all the answer to all the above should be "YES". And I think the user-visible syntax/work to support using it should be absolutely minimal.
Perhaps there is a middle ground, somewhat, but it would require more work on the part of the CMake developers I think:
1) Have the scope limited to a project/module. When the user creates a project, they can optionally have it register itself for Autoheader support.
2) Whenever a variable is added, the supporting API function checks with the internals and registers the variables with all registered modules that are in scope. (I think the scoping here is probably already well defined.)
This would really eliminate much of any user-visible API - other than to do two things:
1) When the project/module is defined:
PROJECT(name C)
CMAKE_AUTOHEADER(name file [optional filter])
2) A generic API to add/remove variables - the user would not necessarily have to use it; but it would be used for the CMake provided back-end (e.g. in the Find functions) for the same purpose
CMAKE_AUTOHEADER_ADD_VAR(myvar)
CMAKE_AUTOHEADER_REMOVE_VAR(myvar)
When CMake generates the project, it would also generate the autoheader - e.g. it could be among the last steps.
Ben