Note the distinction between the version (1.0) and the release label
(2013.03). The latter is not really usable to check for API
compatibility. (See <http://www.semver.org>).
Although they don't use Boost::UnitTest, they can at least pass, so we
can use them to detect simple compilation and runtime errors, although
we miss the semantic check.
(If you have time, please make them proper unit tests)
If we write a blank value into the cache, we tell CMake that ERT isn't
found and it won't check again. If we delete it from the cache in case
we didn't find the directories, then CMake will run the find module
again the next time we configure. Just probing for directories is cheap
so this is no problem. This allows us to install ERT, do a reconfigure
and it will then be picked up by the build system.
If we find ERT and it doesn't compile properly, a blank value will still
be written, and it won't try again the next time, so you must wipe the
cache if you *upgrade* ERT into a newer and more compatible version.
It gets tiresome to rebuild the project everytime one makes changes to
any of the unit tests. We don't likely publish results based on the
debug code anyway, so tracking the version number there is not such a
pressing issue.
Otherwise, the library may be declared "done" before the version number
process have had any change to touch files it depends upon. This ensures
that the version number is correct before we start building the library.
Akin to config.h, we write project-version.h whenever the VCS sha hash
of the project changes. This file can then be included to embed this
into the project. Since this changes more frequently, we choose another
file than config.h. Care is also taken to not rewrite the header if the
information doesn't change, but the project is rebuilt.
The difference between the version number (1.0) and the label (2013.03)
is: The version is time-independent but shows the level of compatibility,
see e.g. <http://www.semver.org>, whereas the label is a marketing tool
which describes the project progress.
that one was due to the fact that the constructor arguments were no
longer used to initialize (unused) private member variables. These
warnings did not appear in CLang for some reason. Again, thanks to
Bård Skaflestad for the review.
'work' is a small, constant size array for temporary data so it can be
allocated on the stack and the 'work' parameter can be eliminated entirely.
Thanks to Bård Skaflestad for the suggestion!
most of them quite insignificant, but still annoying. The only
exception is the warning about the changed alignment for the 'work'
argument of spu_implicit_assemble(). AFAICT, the only reason why it
worked was that the pointer produced by malloc() was passed
directly. (malloc() seems to fulfill all alignment criteria.) To fix
this, I've changed that argument's type from char* to double*.
