[cfe-dev] OpenMP in Flang and Clang

[Johannes Doerfert via cfe-dev]

Since we now merged F18, yay!, we should start reusing OpenMP
"infrastructure" between LLVM/Clang and LLVM/Flang. This will reduce
duplication (coding & maintenance), improve testing, and thereby should
get us to OpenMP 5.X support faster.

I started to look at `parse-tree.h` and `openmp-parsers.cpp` which
contain a lot of code dealing with the OpenMP "grammar". Arguably, we
don't want the duplication with
   `llvm/include/llvm/Frontend/OpenMP/OMPKinds.def`
(and the almost empty "old" clang file `OpenMPKinds.def`).

I think there are multiple steps we should take. Note that I have to
list them in some order, please don't interpret this as the order I
suggest we work on them.

   Move and merge the ENUM_CLASS definitions that can be shared into
   OMPKinds.def. This might require to extend the macros such that we
   have a flag for C/C++ or Fortran only but that should not be too hard.
   The existing Clang code (or better OpenMPIRBuilder by now) could
   benefit from new enum classes already in Flang, e.g. OmpCancelType,
   while things like the kind of an OmpScheduleClause should only be
   defined once.

   Determine how we can merge the definitions in `parse-tree.h` with the
   ones in `OpenMPClauses.h`. I was hoping to trim down the latter anyway
   as it seems there is quite a bit of duplication in there. I was also
   hoping we could generate them from the macro file (or a table-gen
   file) so we don't have to write all the boilerplate when new
   directives and clauses are added. As an example, in the new OpenMP
   context selector support for Clang new selector sets, selectors, and
   traits can be added via macros in the OMPKinds.def file. Boilerplate
   like parsing, pretty printing, error messages, will work right away.
   Only if the existing logic to determine if the selector matches a
   context is not sufficient you need to modify one more place.
   (This is not true for implementation defined extensions that change
    the behavior of the entire thing but anyway.)

   Rethink the use of `std::tuple` in and `std::variant` in (the OpenMP
   parts) of parse-tree.h. The latter is used for OmpLinearClause for
   example. As far as I can tell, the two members of the variant have two
   common fields and one has a third, an enum. In clang this is usually
   solved by having an "unknown" or "none" member in the enum which
   allows to remove the duplication and variant completely. I would like
   to replace tuples so we can give their constitutes proper names. I
   think a similar point was made before when we discussed tooling with
   (or extending) Flang.

I hope to get some feedback and thoughts on this.

Thanks,
   Johannes