[flang][Lower][OpenMP] Fix reduction on array sections aborting in lowering

[blazie2004]

Summary

This regression was introduced by #196094, which added a special lowering path for reductions on a single array element, such as a(2).

The problem is that Flang also treated an array section like a(2:96) as if it were a single element. Because of this, the section was sent to a code path that only supports scalar elements.

That path produced an array type that the reduction initialization code could not handle, so Flang reached a TODO and aborted with a “not yet implemented” error.

Fix

The fix is to use the special element path only when the expression has rank 0, which means it represents one single value. Array sections have rank greater than 0, so they should continue through the existing boxed-array path.

Fixes : 209462

[llvmorg-github-actions]

@llvm/pr-subscribers-flang-openmp

@llvm/pr-subscribers-flang-fir-hlfir

Author: jay0x (blazie2004)

Changes

Summary

This regression was introduced by #196094, which added a special lowering path for reductions on a single array element, such as a(2).

The problem is that Flang also treated an array section like a(2:96) as if it were a single element. Because of this, the section was sent to a code path that only supports scalar elements.

That path produced an array type that the reduction initialization code could not handle, so Flang reached a TODO and aborted with a “not yet implemented” error.

Fix

The fix is to use the special element path only when the expression has rank 0, which means it represents one single value. Array sections have rank greater than 0, so they should continue through the existing boxed-array path.

Fixes : 209462

---

Full diff: https://github.com/llvm/llvm-project/pull/209701.diff

2 Files Affected:

• (modified) flang/lib/Lower/Support/ReductionProcessor.cpp (+5-1)
• (added) flang/test/Lower/OpenMP/reduction-array-section.f90 (+39)

diff --git a/flang/lib/Lower/Support/ReductionProcessor.cpp b/flang/lib/Lower/Support/ReductionProcessor.cpp
index b0a2fe63ef81f..80d34285101f1 100644
--- a/flang/lib/Lower/Support/ReductionProcessor.cpp
+++ b/flang/lib/Lower/Support/ReductionProcessor.cpp
@@ -387,7 +387,11 @@ bool ReductionProcessor::isExpressionLoweredAsReductionObject(
   if (!object || !object->ref())
     return false;
   const SomeExpr &expr = *object->ref();
-  return evaluate::IsArrayElement(expr);
+  // Only genuine single array elements (rank 0) are lowered via the element
+  // path. Array sections such as a(2:96) and vector subscripts have rank > 0;
+  // lowering them here produces an unsupported sequence type and aborts in
+  // PrivateReductionUtils. Let them fall back to the boxed whole-array path.
+  return evaluate::IsArrayElement(expr) && expr.Rank() == 0;
 }
 
 template <typename ParentDeclOpType>
diff --git a/flang/test/Lower/OpenMP/reduction-array-section.f90 b/flang/test/Lower/OpenMP/reduction-array-section.f90
new file mode 100644
index 0000000000000..65e8b2c8fa018
--- /dev/null
+++ b/flang/test/Lower/OpenMP/reduction-array-section.f90
@@ -0,0 +1,39 @@
+! Regression test for reductions on array *sections* (e.g. a(2:96)).
+!
+! An array section has rank > 0, so it must be lowered using the boxed
+! whole-array reduction (@add_reduction_byref_box_*) and the section applied
+! via hlfir.designate inside the region. It must NOT be routed through the
+! single-element path (uniq_name = "omp.reduction.element"), which only supports
+! rank-0 references and otherwise aborts in PrivateReductionUtils with
+! "creating reduction/privatization init region for unsupported type".
+
+! RUN: %flang_fc1 -emit-hlfir -fopenmp -fopenmp-version=50 %s -o - | FileCheck %s --implicit-check-not=omp.reduction.element
+
+subroutine reduction_array_section(a, n)
+  integer :: a(100), n
+!$omp parallel do reduction(+: a(2:96))
+  do i = 1, n
+    a(2:96) = a(2:96) + i
+  end do
+end subroutine
+
+! CHECK: omp.declare_reduction @[[RED:add_reduction_byref_box_100xi32]] : !fir.ref<!fir.box<!fir.array<100xi32>>>
+
+! CHECK-LABEL: func.func @_QPreduction_array_section
+! CHECK: omp.wsloop {{.*}} reduction(byref @[[RED]] %{{[0-9]+}} -> %[[ARG:.*]] : !fir.ref<!fir.box<!fir.array<100xi32>>>) {
+! CHECK: %[[DECL:.*]]:2 = hlfir.declare %[[ARG]] {uniq_name = "_QFreduction_array_sectionEa"}
+! CHECK: %[[BOX:.*]] = fir.load %[[DECL]]#0 : !fir.ref<!fir.box<!fir.array<100xi32>>>
+! CHECK: hlfir.designate %[[BOX]] (%c2:%c96:%c1) {{.*}} -> !fir.ref<!fir.array<95xi32>>
+
+subroutine reduction_array_section_simd(a, n)
+  integer :: a(100), n
+!$omp parallel do simd reduction(+: a(2:96))
+  do i = 1, n
+    a(2:96) = a(2:96) + i
+  end do
+end subroutine
+
+! CHECK-LABEL: func.func @_QPreduction_array_section_simd
+! CHECK: omp.wsloop reduction(byref @[[RED]] %{{[0-9]+}} -> %[[WSARG:.*]] : !fir.ref<!fir.box<!fir.array<100xi32>>>) {
+! CHECK: omp.simd {{.*}} reduction(byref @[[RED]] %[[WSARG]] -> %[[SIMDARG:.*]] : !fir.ref<!fir.box<!fir.array<100xi32>>>) {
+! CHECK: hlfir.declare %[[SIMDARG]] {uniq_name = "_QFreduction_array_section_simdEa"}

[Stylie777]

Change LGTM, just some nits that will need fixing before merging. No need for re-review from me once this is fixed.

Just to check, does this need a Assisted-by: notation in the description for AI? https://llvm.org/docs/AIToolPolicy.html

Thanks for catching this, the regression did not flag in any tests when I created the original patch.

[Stylie777]

nit: Usual practise is to place the RUN line at the top of the test file.

[tblah]

LGTM. Thanks

[tblah]

Please make sure you use a commit message which uses the normal format for other commits in flang. In this case something like

[flang][Lower][OpenMP] Fix abort lowering array section reductions

[tblah]

/cherry-pick 6c560cd

[llvmbot]

/pull-request #211237