Move deactivation of cxx bindings to config.h

deactivate cxx bindings similar to FindMPI

Add Changelog entry

Test for ParMETIS added

Move HAVE_PARMETIS and dune_register_package_flags to AddParMETISFlags

Add ENABLE_PTSCOTCH_PARMETIS for enabling ptscotch replacement for ParMETIS

Set the following variables always to 1:
 DUNE_HAVE_CXX_CLASS_TEMPLATE_ARGUMENT_DEDUCTION
 DUNE_HAVE_CXX_OPTIONAL
 DUNE_HAVE_CXX_VARIANT
 DUNE_SUPPORTS_CXX_THROW_IN_CONSTEXPR
 DUNE_HAVE_C_ALIGNED_ALLOC

Markus Blatt authored 5 years ago and Christoph Grüninger committed 5 years ago

FindLAPACK and FindBLAS check for the name mangling with added
underline as a fallback, anyway. Hence there is no need for using
a Fortran compiler. With this commit we now simply add another check
for a LAPACK function to see whether it has an underline added and
adapt the existing macro in config.h accordingly.

Closes #184

Santiago Ospina De Los Ríos authored 5 years ago and Christoph Grüninger committed 5 years ago

This is an approved struct for the C++20 standard. The implementation
comes from the suggested implementation in cppreference.com

Using `__has_include(<variant>)` unfortunately does not
work as expected, because for `g++-7 -std=c++14` the
`variant` header exists but does not define `std::variant`

We recently switched to built-in feature test macros for these, but as I had to
learn, libc++ does not have those macros. So back to the pedestrian way of
doing things...

We can't use a define from a header (__cpp_lib_experimental_detect in
this case) to detect whether we should include the header. Without
C++17 __has_include, CMake will have to do the work after all.

Inline variables are a C++17 feature that gives variables the same
type of linkage as inline functions, which avoids object file bloat
and linker problems.

For us, this is mostly interesting in the context of constexpr
variables like the partition sets provided in dune-grid: Right now,
a separate copy of those objects is emitted by each translation unit,
and those copies do not get merged by the linker. This has created
problems in the past and forced us to place the variables in an
anonymous namespace, which creates unique symbol names for each
translation unit.

With inline variables, the problem will eventually go away.

Using `#define` instead of `#cmakedefine` resulted in `HAVE_TBB` to be
defined as `TRUE`.  This change makes it defined to `1` instead.

Rene Milk authored 10 years ago and Dominic Kempf committed 7 years ago

usage: add_dune_tbb_flags(TARGETS)
also adds appropriately defined HAVE_TBB to config.h

(cherry picked from commit 36aec8e62786d436521220e6b86d43da061417c6)

Conflicts:
	cmake/modules/DuneCommonMacros.cmake
	config.h.cmake

This solves a problem that is introduced by unconditionally setting
`HAVE_VC=1` in config.h:

If a compilation unit is compiled without the `-fabi-version=6` or similar
then the compilers vector types (which are used by Vc under the hood if
available) will mangle to the same string, irrespective of the number of lanes
they actually use.  This makes certain overloaded function definitions in the
Vc headers clash, since their only difference is whether they apply to
e.g. AVX types (4 lanes) or SSE types (2 lanes).

This is normally not a problem, since any program that uses Vc needs to have a
call to `add_dune_vc_flags()` in `CMakeLists.txt`.  However, since I now want
to put support for vectorized types into the `fmatrix.hh`, any compilation
unit that uses a `FieldMatrix` will automatically include the Vc headers
because `config.h` unconditionally sets `HAVE_VC=1`.

Using the HAVE/ENABLE-trick makes sure that the value of `HAVE_VC` is in sync
with the compilation flags.

Note: I tried using `COMPILE_DEFINITION` for `ENABLE_VC=1` in
      `dune_register_package_flags()`, but that will always add
      `-DENABLE_VC=1` to the compiler command line, even when
      `add_dune_vc_flags()` hasn't been called for that target, and
      `dune_enable_all_packages()` isn't in use.  Using `COMPILE_OPTIONS`
      instead seems inappropriate, but does work.

Christoph Grüninger authored 9 years ago and Christoph Grüninger committed 9 years ago

Fixes dune-istl/#5

This adresses #26.

Revert "Merge remote-tracking branch 'refs/remotes/core/feature/vc' into patches"

This reverts commit 68d8b5a5, reversing
changes made to e8d207bb.

We require MPI 2.1 anyway and nobody has reported problems.

We require MPI 2.1 anyway and nobody has reported problems.

Tobias Malkmus authored 9 years ago and Christoph Grüninger committed 9 years ago

This patch adds the cmake findpackage script which locates Tim Davis' SuiteSparsae collection.

If a certain setting of the SuiteSparse library should be used it can be found using
find_package( SuiteSparse COMPONENETS <list-of-components> )
with a specific list of components.
Existence of components can be check using the SUITESPARSE_<comp>_FOUND/HAVE_SUITESPARSE_<comp> cmake variable.

The UMFPack search routines are redirected to the more general FindSuiteSparse script.

Old feature support macros which were tested until 2.4, kept around
for one more release. As these are now always supported due to the
new compiler requirements, they are directly defined without an
explicit test.

These commits broke compiling / linking if one package was found
but the includes / libraries were not added. We need a different
approach to hide the ENABLE_* trick from the user and keep full
testing capabilities.
This reverts commits 0f12caff
and d75c3a46.