OpenMP Semantic checks for combined constructs

Hi all,

I’ve been looking at the OpenMP semantics checks for combined constructs and I think we need a slight redesign of the OpenMP semantics checks to accommodate these. Currently we’re copying the checks from sub-constructs into the combined constructs, but particularly in the target section of the OpenMP spec there are a huge number of combinations of combined constructs and duplicating all this code seems unwieldy. In addition, if a future version of the OpenMP spec adds for example a new allowed clause for the do directive, currently we will need to update the allowed clauses for do not only in do itself but in every combined construct in which do appears (of which there are a large number).

As such, I propose we restructure the OpenMP semantics checks to include a table of constructs and which clauses are allowed for each, so that this can be indexed not only for each construct itself but also the combined constructs of which it is a member. This will require rewriting most of the semantics checks for OpenMP constructs to use this mechanism, but I think the benefits outweigh the costs of doing this.

Does anyone have any input on this?

Thanks
David Truby

Fwiw, I’m in favor.

Generally, i think having a single location that specifies what causes/combinations are allowed is a good idea. I’d even go as far to say we should put this “table” into llvm/lib/frontend/OpenMP with the goal to reuse it in clang later as well. Anything that will help us keep the two in sync is going to be really helpful in the long run.

Currently, information like this lives in:
  clang/include/clang/Basic/OpenMPKinds.def
but I started to move stuff out of there into (something like*)
   llvm/include/llvm/Frontend/OpenMP/Kinds.def

* working title, see [0,1] and the discussion about lib/frontend on llvm-dev

[0] https://reviews.llvm.org/D69853
[1] https://reviews.llvm.org/D69785