[flang][cuda] Use declare op results instead of memref (#106287)

#106120 Simplify the data transfer when possible by using the reference
and a shape. This bypass the declare op. In order to keep the declare op
around, use the second results of the declare op which achieve the same.

Author: clementval

flang/lib/Lower/Bridge.cpp

@@ -4262,7 +4262,7 @@ class FirConverter : public Fortran::lower::AbstractConverter {
         if (!declOp.getShape())
           return val;
         if (mlir::isa<fir::ReferenceType>(declOp.getMemref().getType()))
-          return declOp.getMemref();
+          return declOp.getResults()[1];
       }
       return val;
     };

flang/test/Lower/CUDA/cuda-data-transfer.cuf

@@ -199,8 +199,8 @@ end subroutine
 ! CHECK-SAME: %[[ARG0:.*]]: !fir.ref<!fir.array<?xi32>> {cuf.data_attr = #cuf.cuda<device>, fir.bindc_name = "a"}, %[[ARG1:.*]]: !fir.ref<!fir.array<10xi32>> {fir.bindc_name = "b"}, %[[ARG2:.*]]: !fir.ref<i32> {fir.bindc_name = "n"})
 ! CHECK: %[[B:.*]]:2 = hlfir.declare %[[ARG1]](%{{.*}}) dummy_scope %{{.*}} {uniq_name = "_QFsub8Eb"} : (!fir.ref<!fir.array<10xi32>>, !fir.shape<1>, !fir.dscope) -> (!fir.ref<!fir.array<10xi32>>, !fir.ref<!fir.array<10xi32>>)
 ! CHECK: %[[A:.*]]:2 = hlfir.declare %[[ARG0]](%{{.*}}) dummy_scope %{{.*}} {data_attr = #cuf.cuda<device>, uniq_name = "_QFsub8Ea"} : (!fir.ref<!fir.array<?xi32>>, !fir.shape<1>, !fir.dscope) -> (!fir.box<!fir.array<?xi32>>, !fir.ref<!fir.array<?xi32>>)
-! CHECK: cuf.data_transfer %[[ARG0]] to %[[B]]#0, %{{.*}} : !fir.shape<1> {transfer_kind = #cuf.cuda_transfer<device_host>} : !fir.ref<!fir.array<?xi32>>, !fir.ref<!fir.array<10xi32>>
-! CHECK: cuf.data_transfer %[[B]]#0 to %[[ARG0]], %{{.*}} : !fir.shape<1> {transfer_kind = #cuf.cuda_transfer<host_device>} : !fir.ref<!fir.array<10xi32>>, !fir.ref<!fir.array<?xi32>>
+! CHECK: cuf.data_transfer %[[A]]#1 to %[[B]]#0, %{{.*}} : !fir.shape<1> {transfer_kind = #cuf.cuda_transfer<device_host>} : !fir.ref<!fir.array<?xi32>>, !fir.ref<!fir.array<10xi32>>
+! CHECK: cuf.data_transfer %[[B]]#0 to %[[A]]#1, %{{.*}} : !fir.shape<1> {transfer_kind = #cuf.cuda_transfer<host_device>} : !fir.ref<!fir.array<10xi32>>, !fir.ref<!fir.array<?xi32>>
 
 subroutine sub9(a)
   integer, pinned, allocatable :: a(:)
@@ -286,6 +286,7 @@ end subroutine
 
 ! CHECK-LABEL: func.func @_QPsub15(
 ! CHECK-SAME: %[[ARG0:.*]]: !fir.ref<!fir.array<?xf32>> {cuf.data_attr = #cuf.cuda<device>, fir.bindc_name = "a_dev"}, %[[ARG1:.*]]: !fir.ref<!fir.array<?xf32>> {fir.bindc_name = "a_host"}
-! CHECK: %{{.*}} = fir.shape %{{.*}} : (index) -> !fir.shape<1>
+! CHECK: %[[ADEV:.*]]:2 = hlfir.declare %[[ARG0]](%{{.*}}) dummy_scope %{{.*}} {data_attr = #cuf.cuda<device>, uniq_name = "_QFsub15Ea_dev"} : (!fir.ref<!fir.array<?xf32>>, !fir.shape<1>, !fir.dscope) -> (!fir.box<!fir.array<?xf32>>, !fir.ref<!fir.array<?xf32>>)
 ! CHECK: %[[SHAPE:.*]] = fir.shape %{{.*}} : (index) -> !fir.shape<1>
-! CHECK: cuf.data_transfer %[[ARG1]] to %[[ARG0]], %[[SHAPE]] : !fir.shape<1> {transfer_kind = #cuf.cuda_transfer<host_device>} : !fir.ref<!fir.array<?xf32>>, !fir.ref<!fir.array<?xf32>>
+! CHECK: %[[AHOST:.*]]:2 = hlfir.declare %[[ARG1]](%{{.*}}) dummy_scope %{{.*}} {uniq_name = "_QFsub15Ea_host"} : (!fir.ref<!fir.array<?xf32>>, !fir.shape<1>, !fir.dscope) -> (!fir.box<!fir.array<?xf32>>, !fir.ref<!fir.array<?xf32>>)
+! CHECK: cuf.data_transfer %[[AHOST]]#1 to %[[ADEV]]#1, %[[SHAPE]] : !fir.shape<1> {transfer_kind = #cuf.cuda_transfer<host_device>} : !fir.ref<!fir.array<?xf32>>, !fir.ref<!fir.array<?xf32>>