[flang-commits] [flang] [llvm] [Flang][OpenMP] Improve use_device_addr code generation (PR #221265)

[Akash Banerjee via flang-commits]

================
@@ -1212,6 +1231,57 @@ class MapInfoFinalizationPass
                argIface.getUseDevicePtrBlockArgs());
   }
 
+  /// Check if we can optimize the descriptor mappings for use_device_addr
+  bool canOptimizeUseDeviceAddrMapping(fir::FirOpBuilder &builder,
+                                       mlir::Value descriptor,
+                                       mlir::omp::MapInfoOp op,
+                                       mlir::Operation *target) {
+    // optimize only temporary descriptors (i.e. allocated on function stack)
+    if (!isAllocaOp(descriptor))
+      return false;
+    // check if given descriptor is mapped as use_device_addr argument
+    if (getUseDeviceAddrBlockArg(op, *target) == nullptr)
+      return false;
+    auto module = builder.getModule();
+    // check if the OpenMP offload target device is specified
+    auto iface =
+        llvm::cast<mlir::omp::OffloadModuleInterface>(module.getOperation());
+    if (iface.getTargetTriples().empty())
+      return false;
+
+    // Only optimize descriptors whose element size is known at compile time.
+    // Excluded:
+    //   - polymorphic entities (!fir.class): elem_len is a runtime property,
+    //     since the dynamic type may extend the declared type.
+    //   - deferred-length characters (!fir.char<k,?>) and parameterized
+    //     derived types: LEN parameters are runtime values.
+    //   - assumed-rank arrays (!fir.array<*:T>): descriptor layout depends on
+    //     a rank that is not known here.
+    //
+    // TODO: The restrictions can be lifted if fir.create_box operation supports
+    // creation of box with dynamical element size.
+    bool isArray = false;
+    bool knownRanks = false;
+    fir::BaseBoxType baseBoxTy = mlir::dyn_cast<fir::BaseBoxType>(
+        fir::unwrapRefType(descriptor.getType()));
+    if (baseBoxTy) {
+      isArray = baseBoxTy.isArray();
+      knownRanks = !baseBoxTy.isAssumedRank();
+    }
+    if (!isArray)
+      return false;
+    if (!knownRanks)
+      return false;
+    auto eleTy = baseBoxTy.unwrapInnerType();
+    if (fir::hasDynamicSize(eleTy))
+      return false;
+    if (fir::isPolymorphicType(eleTy))
----------------
TIFitis wrote:

```suggestion
    if (fir::isPolymorphicType(baseBoxTy))
```

I think we should pass `baseBoxTy` to `fir::isPolymorphicType` here. `unwrapInnerType()` returns the element type, losing the `!fir.class` wrapper. Checking `eleTy` therefore misses `class(t)` arrays and still allows them into the `fir.create_box` path, which does not support polymorphic descriptors.

https://github.com/llvm/llvm-project/pull/221265