[flang-commits] [flang] [flang][OpenMP] Support user-defined declare reduction with derived types (PR #184897)

[Tom Eccles via flang-commits]

================
@@ -1,14 +1,33 @@
-! RUN: not %flang_fc1 -emit-mlir -fopenmp %s -o - 2>&1 | FileCheck %s
+! RUN: %flang_fc1 -emit-hlfir -fopenmp %s -o - | FileCheck %s
 
 program test
   type t
      integer :: x
   end type t
-  ! CHECK: not yet implemented: Reduction of some types is not supported
   !$omp declare reduction(+:t: omp_out%x = omp_out%x + omp_in%x) initializer(omp_priv = t(0))
   type(t) :: a
   a = t(0)
   !$omp parallel reduction(+:a)
   a%x = a%x + 1
   !$omp end parallel
 end program test
+
+! CHECK: omp.declare_reduction @add_reduction_byref_rec__QFTt :
+! CHECK:   %[[ALLOCA:.*]] = fir.alloca [[TY:.*]]
+! CHECK:   omp.yield(%[[ALLOCA]] : !fir.ref<[[TY]]>)
+! CHECK: } init {
+! CHECK: ^bb0(%[[INIT_ARG0:.*]]: !fir.ref<[[TY]]>, %[[INIT_ARG1:.*]]: !fir.ref<[[TY]]>):
+! CHECK:   %{{.*}} = hlfir.declare %[[INIT_ARG0]] {uniq_name = "omp_orig"}
+! CHECK:   %{{.*}} = hlfir.declare %[[INIT_ARG0]] {uniq_name = "omp_priv"}
+! CHECK:   omp.yield(%[[INIT_ARG1]] : !fir.ref<[[TY]]>)
+! CHECK: } combiner {
+! CHECK: ^bb0(%[[ARG0:.*]]: !fir.ref<[[TY]]>, %[[ARG1:.*]]: !fir.ref<[[TY]]>):
+! CHECK:   %[[OMP_IN:.*]]:2 = hlfir.declare %[[ARG1]] {uniq_name = "omp_in"}
+! CHECK:   %[[OMP_OUT:.*]]:2 = hlfir.declare %[[ARG0]] {uniq_name = "omp_out"}
+! CHECK:   %[[OUT_X:.*]] = hlfir.designate %[[OMP_OUT]]#0{"x"} : (!fir.ref<[[TY]]>) -> !fir.ref<i32>
+! CHECK:   %[[OUT_X_VAL:.*]] = fir.load %[[OUT_X]] : !fir.ref<i32>
+! CHECK:   %[[IN_X:.*]] = hlfir.designate %[[OMP_IN]]#0{"x"} : (!fir.ref<[[TY]]>) -> !fir.ref<i32>
+! CHECK:   %[[IN_X_VAL:.*]] = fir.load %[[IN_X]] : !fir.ref<i32>
+! CHECK:   %{{.*}} = arith.addi %[[OUT_X_VAL]], %[[IN_X_VAL]] : i32
----------------
tblah wrote:

I prefer option 2. That way we should be able to re-use however assignment statements are usually lowered instead of rolling a different implementation inside of OpenMP, which will inevitably miss updates, and be less well tested than the main assignment lowering for all of Flang. You are right to highlight that manually lowering the LHS is not the right way to go. There are edge cases around allocatable assignment etc which it would be better not to try to get right in multiple places.

https://github.com/llvm/llvm-project/pull/184897