[PATCH] D83998: [flang][openacc] Basic name resolution infrastructure for OpenACC construct

[Tim Keith via Phabricator via llvm-commits]

tskeith added inline comments.


================
Comment at: flang/lib/Semantics/resolve-names.cpp:1191
+  }
+  return nullptr;
+}
----------------
This function could be implement as:
`return parser::Unwrap<parser::DoConstruct>(x);`


================
Comment at: flang/lib/Semantics/resolve-names.cpp:6488
   inExecutionPart_ = false;
+  ResolveAccParts(x);
   ResolveOmpParts(x);
----------------
Is it worthwhile skipping this unless `-fopenacc` is present? (Same with OpenMP?)


================
Comment at: flang/lib/Semantics/resolve-names.cpp:6838
+  }
+  return nullptr;
+}
----------------
This return is not reachable.

A better way to write this function might be:
```
  if (!name) {
    return nullptr;
  } else if (auto *prev{GetContext().scope.parent().FindCommonBlock(name->source)}) {
    name->symbol  = prev;
    return prev;
  } else {
    return nullptr;
  }
```


================
Comment at: flang/lib/Semantics/resolve-names.cpp:7416
+  if (!context().AnyFatalError()) {
+    AccAttributeVisitor{context(), *this}.Walk(node);
+  }
----------------
Do you need to walk the tree twice for the same reason as OpenMP below? If so, can it be a single function with the visitor as a template parameter?


================
Comment at: flang/lib/Semantics/resolve-names.cpp:1085
 
+template <typename T> class DirectiveAttributeVisitor {
+public:
----------------
clementval wrote:
> clementval wrote:
> > clementval wrote:
> > > klausler wrote:
> > > > clementval wrote:
> > > > > klausler wrote:
> > > > > > clementval wrote:
> > > > > > > klausler wrote:
> > > > > > > > clementval wrote:
> > > > > > > > > clementval wrote:
> > > > > > > > > > klausler wrote:
> > > > > > > > > > > resolve-names.cpp is already a large source file.  Does the directive name resolution code have to be in there in order to use things that are local to that file?  If not, maybe all the directive name resolution code could reside elsewhere.
> > > > > > > > > > I think we can move that into a separate file. I'll update the patch. Thanks for the suggestion.
> > > > > > > > > @klausler I would have to move some class declaration from `resolve-name.cpp` to `resolve-name.h`. If this is fine I will split the files. 
> > > > > > > > Which classes?
> > > > > > > > 
> > > > > > > > Can they go into resolve-names-utils.{h,cpp} instead?
> > > > > > > In fact most of the class declarations would have to be moved to the `resolve-name.h` file in order to be able to split the file and have its own for directive resolution.
> > > > > > That may be the case for `AccVisitor`, but can `AccAttributeVisitor` be extracted into its own header and C++ source file?
> > > > > After some work on this, this would need to move all the declarations in to a header file (`resolve-name.h`) because the `Acc/OmpVisitor` and `Acc/OmpAttributeVisitor` use classes declared inside `resolve-names.cpp` and those classes derived from other internal classes as well. 
> > > > > 
> > > > > It would be much easier to keep those declarations together. One solution to reduce the size of the file might be to extract the declarations in the `.h` file and keep only the implementation in the `.cpp` file but still together. Introducing new files for the directive part makes it harder since they are interdependent. 
> > > > > 
> > > > > What do you think is the best to move forward @klausler? 
> > > > Please don't embark on a wholesale restructuring of resolve-names.cpp without consulting with @tskeith.  If there isn't an obvious way to extract the name resolution for directives, just leave it in there.
> > > The problem is that `AccAttributeVisitor`, `OmpAttributeVisitor` and `DirectiveAttributeVisitor` are using `ResolveNamesVisitor` that is not declared in a `.h` file but in `resolve-names.cpp`. This `class` needs then lots of other classes declared in `resolve-names.cpp`. So there is no easy way to extract them without a refactoring. 
> > Sure will wait on his inputs. 
> Basically almost all the classes declared in `resolve-names.cpp` because one inherits from another or uses other classes and we end up moving the whole things. 
I think it's a good idea for these to be in a separate file, but it will require some restructuring. The main problem is ResolveDesignator. It is used both to resolve a Designator and to get the Name from an already-resolve designator. It looks like you only need the latter function, and that could be moved to resolve-names-utils.{h,cpp}.

I suggest you land this change without worrying about that restructuring and then I'll take a look at it in a separate change (unless you want to).


Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D83998/new/

https://reviews.llvm.org/D83998