[flang][cuda] Fix __ldXX function signatures (#130628)

Function are taking array reference and not array directly.

Author: clementval

flang/lib/Optimizer/Builder/IntrinsicCall.cpp

@@ -3688,8 +3688,9 @@ IntrinsicLibrary::genCUDALDXXFunc(mlir::Type resultType,
   mlir::Value res = builder.create<fir::AllocaOp>(loc, resTy);
   if (mlir::isa<fir::BaseBoxType>(arg.getType()))
     arg = builder.create<fir::BoxAddrOp>(loc, arg);
+  mlir::Type refResTy = fir::ReferenceType::get(resTy);
   mlir::FunctionType ftype =
-      mlir::FunctionType::get(arg.getContext(), {resTy, resTy}, {});
+      mlir::FunctionType::get(arg.getContext(), {refResTy, refResTy}, {});
   auto funcOp = builder.createFunction(loc, fctName, ftype);
   llvm::SmallVector<mlir::Value> funcArgs;
   funcArgs.push_back(res);

flang/test/Lower/CUDA/cuda-device-proc.cuf

@@ -295,3 +295,30 @@ end
 ! CHECK: fir.call @__ldcs_r8x2_(%{{.*}}, %{{.*}}) fastmath<contract> : (!fir.ref<!fir.array<2xf64>>, !fir.ref<!fir.array<?xf64>>) -> ()
 ! CHECK: fir.call @__ldlu_r8x2_(%{{.*}}, %{{.*}}) fastmath<contract> : (!fir.ref<!fir.array<2xf64>>, !fir.ref<!fir.array<?xf64>>) -> ()
 ! CHECK: fir.call @__ldcv_r8x2_(%{{.*}}, %{{.*}}) fastmath<contract> : (!fir.ref<!fir.array<2xf64>>, !fir.ref<!fir.array<?xf64>>) -> ()
+
+
+! CHECK-DAG: func.func private @__ldca_i4x4_(!fir.ref<!fir.array<4xi32>>, !fir.ref<!fir.array<4xi32>>)
+! CHECK-DAG: func.func private @__ldcg_i4x4_(!fir.ref<!fir.array<4xi32>>, !fir.ref<!fir.array<4xi32>>)
+! CHECK-DAG: func.func private @__ldcs_i4x4_(!fir.ref<!fir.array<4xi32>>, !fir.ref<!fir.array<4xi32>>)
+! CHECK-DAG: func.func private @__ldlu_i4x4_(!fir.ref<!fir.array<4xi32>>, !fir.ref<!fir.array<4xi32>>)
+! CHECK-DAG: func.func private @__ldcv_i4x4_(!fir.ref<!fir.array<4xi32>>, !fir.ref<!fir.array<4xi32>>)
+! CHECK-DAG: func.func private @__ldca_i8x2_(!fir.ref<!fir.array<2xi64>>, !fir.ref<!fir.array<2xi64>>)
+! CHECK-DAG: func.func private @__ldcg_i8x2_(!fir.ref<!fir.array<2xi64>>, !fir.ref<!fir.array<2xi64>>)
+! CHECK-DAG: func.func private @__ldcs_i8x2_(!fir.ref<!fir.array<2xi64>>, !fir.ref<!fir.array<2xi64>>)
+! CHECK-DAG: func.func private @__ldlu_i8x2_(!fir.ref<!fir.array<2xi64>>, !fir.ref<!fir.array<2xi64>>)
+! CHECK-DAG: func.func private @__ldcv_i8x2_(!fir.ref<!fir.array<2xi64>>, !fir.ref<!fir.array<2xi64>>)
+! CHECK-DAG: func.func private @__ldca_r4x4_(!fir.ref<!fir.array<4xf32>>, !fir.ref<!fir.array<4xf32>>)
+! CHECK-DAG: func.func private @__ldcg_r4x4_(!fir.ref<!fir.array<4xf32>>, !fir.ref<!fir.array<4xf32>>)
+! CHECK-DAG: func.func private @__ldcs_r4x4_(!fir.ref<!fir.array<4xf32>>, !fir.ref<!fir.array<4xf32>>)
+! CHECK-DAG: func.func private @__ldlu_r4x4_(!fir.ref<!fir.array<4xf32>>, !fir.ref<!fir.array<4xf32>>)
+! CHECK-DAG: func.func private @__ldcv_r4x4_(!fir.ref<!fir.array<4xf32>>, !fir.ref<!fir.array<4xf32>>)
+! CHECK-DAG: func.func private @__ldca_r2x2_(!fir.ref<!fir.array<2xf16>>, !fir.ref<!fir.array<2xf16>>)
+! CHECK-DAG: func.func private @__ldcg_r2x2_(!fir.ref<!fir.array<2xf16>>, !fir.ref<!fir.array<2xf16>>)
+! CHECK-DAG: func.func private @__ldcs_r2x2_(!fir.ref<!fir.array<2xf16>>, !fir.ref<!fir.array<2xf16>>)
+! CHECK-DAG: func.func private @__ldlu_r2x2_(!fir.ref<!fir.array<2xf16>>, !fir.ref<!fir.array<2xf16>>)
+! CHECK-DAG: func.func private @__ldcv_r2x2_(!fir.ref<!fir.array<2xf16>>, !fir.ref<!fir.array<2xf16>>)
+! CHECK-DAG: func.func private @__ldca_r8x2_(!fir.ref<!fir.array<2xf64>>, !fir.ref<!fir.array<2xf64>>)
+! CHECK-DAG: func.func private @__ldcg_r8x2_(!fir.ref<!fir.array<2xf64>>, !fir.ref<!fir.array<2xf64>>)
+! CHECK-DAG: func.func private @__ldcs_r8x2_(!fir.ref<!fir.array<2xf64>>, !fir.ref<!fir.array<2xf64>>)
+! CHECK-DAG: func.func private @__ldlu_r8x2_(!fir.ref<!fir.array<2xf64>>, !fir.ref<!fir.array<2xf64>>)
+! CHECK-DAG: func.func private @__ldcv_r8x2_(!fir.ref<!fir.array<2xf64>>, !fir.ref<!fir.array<2xf64>>)