[flang-commits] [flang] [Flang][OpenMP] Optimize target updates of derived-type scalars (PR #219488)

[via flang-commits]

================
@@ -4300,6 +4302,173 @@ static mlir::omp::TargetDataOp genTargetDataOp(
   return targetDataOp;
 }
 
+struct TargetUpdateKernelEntry {
+  mlir::omp::MapInfoOp mapInfo;
+  mlir::Value hostPtr;
+  mlir::Type componentType;
+};
+
+static bool hasAMDGCNTarget(mlir::ModuleOp module) {
+  auto offloadModule = llvm::cast<mlir::omp::OffloadModuleInterface>(*module);
+  if (offloadModule.getIsTargetDevice())
+    return fir::getTargetTriple(module).isAMDGCN();
+  return llvm::any_of(
+      offloadModule.getTargetTriples(), [](mlir::Attribute attr) {
+        auto tripleAttr = llvm::dyn_cast<mlir::StringAttr>(attr);
+        return tripleAttr && llvm::Triple(tripleAttr.getValue()).isAMDGCN();
+      });
+}
+
+static std::optional<TargetUpdateKernelEntry>
+getTargetUpdateKernelEntry(mlir::Value mapVar) {
+  auto mapInfo = mapVar.getDefiningOp<mlir::omp::MapInfoOp>();
+  if (!mapInfo)
+    return std::nullopt;
+
+  // Keep the fast path to plain synchronous H2D motion. In particular, do not
+  // silently weaken `present` motion modifiers.
+  if (mapInfo.getMapType() != mlir::omp::ClauseMapFlags::to ||
+      mapInfo.getVarPtrPtr() || !mapInfo.getMembers().empty() ||
+      !mapInfo.getBounds().empty() || mapInfo.getMapperId())
+    return std::nullopt;
+
+  mlir::Value hostPtr = mapInfo.getVarPtr();
+  auto designate = hostPtr.getDefiningOp<hlfir::DesignateOp>();
+  if (!designate || !designate.getComponent() ||
+      designate.getComponentShape() || !designate.getIndices().empty() ||
+      !designate.getSubstring().empty() || designate.getComplexPart() ||
+      designate.getShape() || !designate.getTypeparams().empty())
+    return std::nullopt;
+
+  mlir::Type baseType = fir::unwrapRefType(designate.getMemref().getType());
+  auto recordType = mlir::dyn_cast<fir::RecordType>(baseType);
+  if (!recordType || recordType.getNumLenParams() != 0)
+    return std::nullopt;
+
+  llvm::StringRef component = designate.getComponent()->getValue();
+  mlir::Type componentType = recordType.getType(component);
+  if (!componentType || !fir::isa_trivial(componentType))
+    return std::nullopt;
+
+  return TargetUpdateKernelEntry{mapInfo, hostPtr, componentType};
+}
+
+static mlir::omp::TargetOp
+genTargetUpdateKernel(lower::AbstractConverter &converter, mlir::Location loc,
+                      llvm::ArrayRef<TargetUpdateKernelEntry> entries) {
+  fir::FirOpBuilder &builder = converter.getFirOpBuilder();
+  mlir::omp::TargetExtOperands targetClauseOps;
+  targetClauseOps.kernelType = mlir::omp::TargetExecModeAttr::get(
+      builder.getContext(), mlir::omp::TargetExecMode::generic);
+
+  llvm::SmallVector<mlir::Value> destinationMaps;
+  destinationMaps.reserve(entries.size());
+
+  llvm::SmallVector<mlir::Type> sourceTypes;
+  llvm::transform(
+      entries, std::back_inserter(sourceTypes),
+      [](const TargetUpdateKernelEntry &entry) { return entry.componentType; });
+  mlir::TupleType sourceType =
+      mlir::TupleType::get(builder.getContext(), sourceTypes);
+  mlir::Value sourcePack = builder.createTemporary(loc, sourceType);
+
+  for (auto indexedEntry : llvm::enumerate(entries)) {
----------------
agozillon wrote:

Nit: might be able to use the [i, &entry] syntax to save you the accesses in the loop

https://github.com/llvm/llvm-project/pull/219488