The HAVE_UMFPACK_WITHOUT_CHOLMOD test would generate a false success.
New test contains program with more calls to the umfpack API, reflecting
our actual usage.
This will hopefully guide some (buggy) compiler variants to not start
interpreting the header as a full program, or at least give a more
sensible error message.
Templates can now be reused across projects. The output files are still
put in the root of the output tree because that is where the client
programs will look for them.
The source code of ERT is located in a devel/ subdirectory, but it is
not usual (?) to have this structure in a separate build directory as
well. If you however build everything in the same directory as the
source code, then this sub-directory needs to be accounted for.
Using umfpack_dl_solve as a test function increases the chance that the
probe will not be able to link successfully if it has not found all
libraries necessary (to link other than trivial test programs).
These files are text files which specify which platform specific
libraries we should be looking for. We need libtool installed to
use them, but they can be generated pretty inexpensively from the
CMake script itself.
There is a convention that one can do `pkg-config --variable=includedir`
and have it return the path to the source files (even if that should
have been taken care of by options in --cflags).
If more than one match is written, then DUNE_CHECK_MODULES will return
several matches, which becomes a syntax error on the resulting command-
line. Thus, check if there is a match before writing anything.
Dunecontrol scans the module directory for certain files and phrases
which must be present if the project can be used as a dependency for
other DUNE modules.
The old implementation erraneously used only the first argument that was
passed to the testing procedure due to the way lists are passed in CMake
when unquoted. Thus, the test failed for SuiteSparse >= 4.0 which needs
more than one dependency.
Also add dependency to BLAS and LAPACK which was hidden because the
caller already included these itself.
Previous implementation had a separate search path for headers and for
the library, which in the case of "RedHat-multiarch" made it miss the
library (it was looking for /usr/lib/lib64 instead of /usr/lib64); the
new version starts looking from /usr.
UMFPACK is by default dependent on CHOLMOD, but on some systems it is
compiled with -DNCHOLMOD; in those cases we should not try to link in
the library -- it may not be available and in any cases it is not used,
so it is just a superfluous dependency.
The way the test was written earlier, it would accept if it found
components that were transitively pulled in even if the originally
requested component was not!
At this time no probe is added to check if an UMFPACK without dependency
to CHOLMOD is installed; thus if you need/want UMFPACK, you'll have to
go for the entire SparseSuite.
dunecontrol will read the dune.module file first, before calling any
configuration files; thus it cannot be initialized from the script
but the version numbers must be hardcoded here and updated manually.
Note that using pkg-config is NOT recommended because of:
1. It doesn't know which language you are using, so language-specific
options such as -std=c++11 or -std=c99 cannot be passed along. Yet,
the build will corrupt if you don't get these correct.
2. It has no means of communicating which defines that should be set
in config.h, effectively disabling them all, possibly leading to
inconsistencies between the library and the executable.
