[llvm-branch-commits] [flang] [flang][OpenMP] Use OmpDirectiveSpecification in ALLOCATE (PR #165865)

[Krzysztof Parzyszek via llvm-branch-commits]

================
@@ -2558,11 +2574,24 @@ bool OmpAttributeVisitor::Pre(const parser::OpenMPThreadprivate &x) {
   return true;
 }
 
-bool OmpAttributeVisitor::Pre(const parser::OpenMPDeclarativeAllocate &x) {
+bool OmpAttributeVisitor::Pre(const parser::OmpAllocateDirective &x) {
   PushContext(x.source, llvm::omp::Directive::OMPD_allocate);
-  if (const auto &list{std::get<std::optional<parser::OmpObjectList>>(x.t)}) {
-    ResolveOmpObjectList(*list, Symbol::Flag::OmpDeclarativeAllocateDirective);
+  assert(!partStack_.empty() && "Misplaced directive");
+
+  auto ompFlag{partStack_.back() == PartKind::SpecificationPart
+          ? Symbol::Flag::OmpDeclarativeAllocateDirective
+          : Symbol::Flag::OmpExecutableAllocateDirective};
+
+  parser::omp::OmpAllocateInfo info{parser::omp::SplitOmpAllocate(x)};
+  for (const parser::OmpAllocateDirective *ad : info.dirs) {
+    for (const parser::OmpArgument &arg : ad->BeginDir().Arguments().v) {
+      if (auto *object{omp::GetArgumentObject(arg)}) {
+        ResolveOmpObject(*object, ompFlag);
+      }
+    }
   }
+
+  PopContext();
----------------
kparzysz wrote:

Short answer: because we do all handling for this directive in this function, and don't want the traversal to descend into the subcomponents (we return "false" from Pre).  In that case Post will never be called.

We don't want to traverse any potentially nested directives because each one would push its own context, and the executable-allocate should really only push a single context for the entire construct.

https://github.com/llvm/llvm-project/pull/165865