[flang-commits] [flang] [flang][acc] Implement PointerLikeType API gen[Allocate/Free/Copy] (PR #163660)

[Razvan Lupusoru via flang-commits]

================
@@ -751,4 +751,226 @@ template bool OpenACCMappableModel<fir::PointerType>::generatePrivateDestroy(
     mlir::Type type, mlir::OpBuilder &builder, mlir::Location loc,
     mlir::Value privatized) const;
 
+template <typename Ty>
+mlir::Value OpenACCPointerLikeModel<Ty>::genAllocate(
+    mlir::Type pointer, mlir::OpBuilder &builder, mlir::Location loc,
+    llvm::StringRef varName, mlir::Type varType, mlir::Value originalVar,
+    bool &needsFree) const {
+
+  // Get the element type from the pointer type
+  mlir::Type eleTy = mlir::cast<Ty>(pointer).getElementType();
+
+  // Unlimited polymorphic (class(*)) cannot be handled - size unknown
+  if (fir::isUnlimitedPolymorphicType(eleTy))
+    return {};
+
+  // Character types with dynamic length cannot be handled without a descriptor.
+  if (auto charTy = mlir::dyn_cast<fir::CharacterType>(eleTy))
+    if (charTy.hasDynamicLen())
+      return {};
+
+  // Return null for dynamic or unknown shape arrays because the size of the
+  // allocation cannot be determined simply from the type.
+  if (auto seqTy = mlir::dyn_cast<fir::SequenceType>(eleTy))
+    if (seqTy.hasDynamicExtents() || seqTy.hasUnknownShape())
+      return {};
----------------
razvanlupusoru wrote:

> fir::hasDynamicSize(eleTy) should be equivalent

Thank you for the suggestion to use hasDynamicSize - I will use that instead.

> Which leads me to another question, what is the expected behavior of fir.ref<fir.box/class<T>>? Is it to allocate/free/copy descriptors or to also allocate/free/copy the target data of the descriptor?

I think that `genAllocate` should allow for simple shallow allocations - so allocated descriptor storage is enough.  `generatePrivateInit` from MappableType (which is accessible through both `fir.ref<fir.box>` and `fir.box` should handle the complexity of allocating both descriptor and data.

However, I have decided in the current PR, I will reject boxes in this API and we can revisit if we need shallow allocations of it.

https://github.com/llvm/llvm-project/pull/163660