[flang-commits] [flang] [Flang][OpenMP] Fix crash and IR errors for user-defined reduction on allocatable variables (PR #186765)
via flang-commits
flang-commits at lists.llvm.org
Thu Jul 2 00:53:28 PDT 2026
================
@@ -803,6 +804,185 @@ bool ReductionProcessor::processReductionArguments(
&redOperator.u)) {
if (!ReductionProcessor::supportedIntrinsicProcReduction(
*reductionIntrinsic)) {
+ if (isByRef) {
+ // create a new declare_reduction for the boxed type, reusing
+ // the existing init and combiner
+ semantics::Symbol *sym = reductionIntrinsic->v.sym();
+ llvm::StringRef baseName = {sym->name().begin(),
+ sym->name().size()};
+ mlir::ModuleOp module = builder.getModule();
+ auto existingDecl = module.lookupSymbol<OpType>(baseName);
+ if (!existingDecl) {
+ TODO(currentLocation,
+ "User-defined reductions on allocatable or pointer "
+ "variables: cannot find base reduction declaration");
+ }
+
+ std::string byrefName = getReductionName(
+ baseName, builder.getKindMap(), redType, isByRef);
+
+ mlir::Region &existingInitRegion =
+ existingDecl.getInitializerRegion();
+ auto genInitValueCB =
+ [&existingInitRegion](fir::FirOpBuilder &builder,
+ mlir::Location loc, mlir::Type elemTy,
+ mlir::Value) -> mlir::Value {
+ // unwrap box type to get the scalar element type
+ mlir::Type scalarTy = unwrapSeqOrBoxedType(elemTy);
+ // find a constant-producing op in the existing init region
+ mlir::Operation *constOp = nullptr;
+ existingInitRegion.walk([&](mlir::Operation *op) {
----------------
MattPD wrote:
This silently initializes the private copy to the wrong value when the initializer is not a literal constant. The value is reconstructed by scanning the base init region for the first literal constant and cloning it, so it reproduces the base declaration only for a single-literal initializer. When the initializer reads `omp_orig`, or is an expression or a function call, no constant matches and the fallback substitutes silently: integer becomes `0`, and character returns an empty value, so the private is never stored and stays uninitialized.
You can see it by saving this to `t.f90`:
```fortran
program p
integer, allocatable :: a
integer :: i
!$omp declare reduction (mymax : integer : omp_out = max(omp_out, omp_in)) &
!$omp& initializer (omp_priv = omp_orig)
allocate(a); a = 0
!$omp parallel do reduction(mymax : a)
do i = 1, 4
a = max(a, i)
end do
end program
```
and running `flang -fc1 -emit-hlfir -fopenmp -fopenmp-version=45 t.f90 -o -`. The boxed `@mymax_byref_box_heap_i32` init stores `arith.constant 0`, while the scalar `@mymax` init loads and yields `omp_orig`. The non-allocatable version reduces correctly, so only the allocatable path diverges, and it compiles with no diagnostic.
Would it be more robust to clone the base declaration's init-region body (remapping its `omp_orig` argument to the loaded original element), or to emit a TODO when the initializer is not a single literal, rather than substitute a constant, zero, or empty value?
https://github.com/llvm/llvm-project/pull/186765
More information about the flang-commits
mailing list