[flang][cuda] Copying device globals in the gpu module

Author: Renaud-K

flang/lib/Optimizer/Transforms/CUFDeviceGlobal.cpp

@@ -11,6 +11,7 @@
 #include "flang/Optimizer/Dialect/FIRDialect.h"
 #include "flang/Optimizer/Dialect/FIROps.h"
 #include "flang/Optimizer/HLFIR/HLFIROps.h"
+#include "flang/Optimizer/Transforms/CUFCommon.h"
 #include "flang/Runtime/CUDA/common.h"
 #include "flang/Runtime/allocatable.h"
 #include "mlir/IR/SymbolTable.h"
@@ -58,6 +59,31 @@ class CUFDeviceGlobal : public fir::impl::CUFDeviceGlobalBase<CUFDeviceGlobal> {
       prepareImplicitDeviceGlobals(funcOp, symTable);
       return mlir::WalkResult::advance();
     });
+
+    mlir::SymbolTable parentSymTable(mod);
+    auto gpuMod =
+        parentSymTable.lookup<mlir::gpu::GPUModuleOp>(cudaDeviceModuleName);
+    if (gpuMod) {
+      mlir::SymbolTable gpuSymTable(gpuMod);
+      for (auto globalOp : mod.getOps<fir::GlobalOp>()) {
+        auto attr = globalOp.getDataAttrAttr();
+        if (!attr)
+          continue;
+        switch (attr.getValue()) {
+        case cuf::DataAttribute::Device:
+        case cuf::DataAttribute::Constant:
+        case cuf::DataAttribute::Managed: {
+          auto globalName{globalOp.getSymbol().getValue()};
+          if (gpuSymTable.lookup<fir::GlobalOp>(globalName)) {
+            break;
+          }
+          gpuSymTable.insert(globalOp->clone());
+        } break;
+        default:
+          break;
+        }
+      }
+    }
   }
 };
 } // namespace

flang/test/Fir/CUDA/cuda-device-global.f90

@@ -0,0 +1,14 @@
+
+// RUN: fir-opt --split-input-file --cuf-device-global %s | FileCheck %s
+
+
+// -----// IR Dump After CUFLaunchToGPU (cuf-fir-launch-to-gpu) //----- //
+module attributes {fir.defaultkind = "a1c4d8i4l4r4", fir.kindmap = "", gpu.container_module} {
+  fir.global @_QMmtestsEn(dense<[3, 4, 5, 6, 7]> : tensor<5xi32>) {data_attr = #cuf.cuda<device>} : !fir.array<5xi32>
+
+  gpu.module @cuda_device_mod [#nvvm.target] {
+  }
+}
+
+// CHECK: gpu.module @cuda_device_mod [#nvvm.target] 
+// CHECK-NEXT: fir.global @_QMmtestsEn(dense<[3, 4, 5, 6, 7]> : tensor<5xi32>) {data_attr = #cuf.cuda<device>} : !fir.array<5xi32>