As opm-common is search in the toplevel CMakeLists.txt of each module in
config mode and the dependencies and defines are only processed in
opm_find_package(opm-common), we allow a second search using module mode
to trigger opm_find_package.
Otherwise some defines and macros will not be set as this is done
in opm_find_package. This might also make sense for dunecontrol
which always will set *_DIR.
Theses variables are already set in opm-output-config.cmake. Unfortunately,
theres is still a bug there which will be fixed by a PR in opm-output. After
that everything should work and have less guess work in it.
Previously we assumed it to be ecl (like the project name).
That is not correct. With this commit we now use the correct
default clone directory, libecl, in the sibling search.
Even though people are telling that ecl is already found using
sibling search it turned out that this statement is false. It
was always found using the CMake package cache. When installing
this lead to the installed package using ecl from the build tree.
With this commit we first try to find ecl without the package cache
but maybe using sibling search. Building installed packages no works
by setting -DSIBLING_SEARCH=Off -DCMAKE_INSTALL_PREFIX=/install/path
We use ${module}_DIR to set the correct path when sibling search is activated.
The package configuration files set all the necessary variable and we save
us a lot of CMake magic.
make this default to off.
additionally, check that the found superlu version
is compatible with the dune-istl version in use.
in particular superlu5 is not compatible with dune-istl 2.4
Previously, we checked for header CpGrid.hpp which finally
tried to include the non-existent generic geometry headers
as the checks for dune-geometry versions do not work during
a CMake run.
With this patch we search for the header GlobalIdMapper.hpp which
does not use any version checks and therefore works for every DUNE
version.
These take precendence over the package registry where we have no control
on which entry will be used. The registry is only the last resort.
On some Linux version the most recently created package registry key will
be used. Please not that an entry is only created for a build try that is
not already in the registry. If there is already key, then not even the
modification date of the key will be updated. This default behaviour might
lead to strange mixes of build configurations.
CONFIG mode is usually a fallback when Find* modules are not found, but
opm-parser is now mainly meant to be used with config mode, so it's the
first thing being tried. If that fails, continue with the old logic.
Some legacy requires HAVE_ERT being set somewhere, so this logic is
injected here.
Previously, this failed if one dependency was a static library compiled
without -fPIC. This is is the case for library libsuitesparseconfig on
Debian which has no dynamic alternative. The error messages read e.g.
```
/usr/bin/ld: /usr/lib/gcc/x86_64-linux-gnu/4.9/../../../x86_64-linux-gnu/libsuitesparseconfig.a(SuiteSparse_config.o): relocation R_X86_64_PC32 against undefined symbol `malloc@@GLIBC_2.2.5' can not be used when making a shared object; recompile with -fPIC
/usr/bin/ld: final link failed: Bad value
```
With this patch we change the target_link_libraries call when using
BUILD_SHARED_LIBS=ON. We only pass the dynamic libraries as PUBLIC such
that they will get linked to the library. The static ones are passed as
INTERFACE libraries. This means that they will be automatically linked
to executables which link to the main library. Currently this transitive
linking will only work in the same module as we do not correctly export
libraries. But our build system explicitly links the others ones anyway.
