Skip to content

Commit

Permalink
[flang][OpenMP] Support reduction of POINTER variables (#95148)
Browse files Browse the repository at this point in the history
Just treat them the same as ALLOCATABLE. gfortran doesn't allow POINTER
objects in a REDUCTION clause, but so far as I can tell the standard
explicitly allows it (openmp5.2 section 5.5.5).
  • Loading branch information
tblah committed Jun 14, 2024
1 parent d62ff71 commit b6b0f97
Show file tree
Hide file tree
Showing 3 changed files with 255 additions and 8 deletions.
22 changes: 14 additions & 8 deletions flang/lib/Lower/OpenMP/ReductionProcessor.cpp
Original file line number Diff line number Diff line change
Expand Up @@ -332,15 +332,17 @@ static void genBoxCombiner(fir::FirOpBuilder &builder, mlir::Location loc,
fir::unwrapRefType(boxTy.getEleTy()));
fir::HeapType heapTy =
mlir::dyn_cast_or_null<fir::HeapType>(boxTy.getEleTy());
if ((!seqTy || seqTy.hasUnknownShape()) && !heapTy)
fir::PointerType ptrTy =
mlir::dyn_cast_or_null<fir::PointerType>(boxTy.getEleTy());
if ((!seqTy || seqTy.hasUnknownShape()) && !heapTy && !ptrTy)
TODO(loc, "Unsupported boxed type in OpenMP reduction");

// load fir.ref<fir.box<...>>
mlir::Value lhsAddr = lhs;
lhs = builder.create<fir::LoadOp>(loc, lhs);
rhs = builder.create<fir::LoadOp>(loc, rhs);

if (heapTy && !seqTy) {
if ((heapTy || ptrTy) && !seqTy) {
// get box contents (heap pointers)
lhs = builder.create<fir::BoxAddrOp>(loc, lhs);
rhs = builder.create<fir::BoxAddrOp>(loc, rhs);
Expand All @@ -350,8 +352,10 @@ static void genBoxCombiner(fir::FirOpBuilder &builder, mlir::Location loc,
lhs = builder.create<fir::LoadOp>(loc, lhs);
rhs = builder.create<fir::LoadOp>(loc, rhs);

mlir::Type eleTy = heapTy ? heapTy.getEleTy() : ptrTy.getEleTy();

mlir::Value result = ReductionProcessor::createScalarCombiner(
builder, loc, redId, heapTy.getEleTy(), lhs, rhs);
builder, loc, redId, eleTy, lhs, rhs);
builder.create<fir::StoreOp>(loc, result, lhsValAddr);
builder.create<mlir::omp::YieldOp>(loc, lhsAddr);
return;
Expand Down Expand Up @@ -439,7 +443,7 @@ createReductionCleanupRegion(fir::FirOpBuilder &builder, mlir::Location loc,

mlir::Type valTy = fir::unwrapRefType(redTy);
if (auto boxTy = mlir::dyn_cast_or_null<fir::BaseBoxType>(valTy)) {
if (!mlir::isa<fir::HeapType>(boxTy.getEleTy())) {
if (!mlir::isa<fir::HeapType, fir::PointerType>(boxTy.getEleTy())) {
mlir::Type innerTy = fir::extractSequenceType(boxTy);
if (!mlir::isa<fir::SequenceType>(innerTy))
typeError();
Expand Down Expand Up @@ -533,12 +537,13 @@ createReductionInitRegion(fir::FirOpBuilder &builder, mlir::Location loc,
// all arrays are boxed
if (auto boxTy = mlir::dyn_cast_or_null<fir::BaseBoxType>(ty)) {
assert(isByRef && "passing boxes by value is unsupported");
bool isAllocatable = mlir::isa<fir::HeapType>(boxTy.getEleTy());
bool isAllocatableOrPointer =
mlir::isa<fir::HeapType, fir::PointerType>(boxTy.getEleTy());
mlir::Value boxAlloca = builder.create<fir::AllocaOp>(loc, ty);
mlir::Type innerTy = fir::unwrapRefType(boxTy.getEleTy());
if (fir::isa_trivial(innerTy)) {
// boxed non-sequence value e.g. !fir.box<!fir.heap<i32>>
if (!isAllocatable)
if (!isAllocatableOrPointer)
TODO(loc, "Reduction of non-allocatable trivial typed box");

fir::IfOp ifUnallocated = handleNullAllocatable(boxAlloca);
Expand All @@ -560,7 +565,7 @@ createReductionInitRegion(fir::FirOpBuilder &builder, mlir::Location loc,
TODO(loc, "Unsupported boxed type for reduction");

fir::IfOp ifUnallocated{nullptr};
if (isAllocatable) {
if (isAllocatableOrPointer) {
ifUnallocated = handleNullAllocatable(boxAlloca);
builder.setInsertionPointToStart(&ifUnallocated.getElseRegion().front());
}
Expand All @@ -587,7 +592,8 @@ createReductionInitRegion(fir::FirOpBuilder &builder, mlir::Location loc,
mlir::OpBuilder::InsertionGuard guard(builder);
createReductionCleanupRegion(builder, loc, reductionDecl);
} else {
assert(!isAllocatable && "Allocatable arrays must be heap allocated");
assert(!isAllocatableOrPointer &&
"Pointer-like arrays must be heap allocated");
}

// Put the temporary inside of a box:
Expand Down
131 changes: 131 additions & 0 deletions flang/test/Lower/OpenMP/parallel-reduction-pointer-array.f90
Original file line number Diff line number Diff line change
@@ -0,0 +1,131 @@
! RUN: bbc -emit-hlfir -fopenmp -o - %s | FileCheck %s
! RUN: %flang_fc1 -emit-hlfir -fopenmp -o - %s | FileCheck %s

program reduce
integer :: i = 0
integer, dimension(:), pointer :: r

allocate(r(2))

!$omp parallel do reduction(+:r)
do i=0,10
r(1) = i
r(2) = -i
enddo
!$omp end parallel do

print *,r
deallocate(r)

end program

! CHECK-LABEL: omp.declare_reduction @add_reduction_byref_box_ptr_Uxi32 : !fir.ref<!fir.box<!fir.ptr<!fir.array<?xi32>>>> init {
! CHECK: ^bb0(%[[VAL_0:.*]]: !fir.ref<!fir.box<!fir.ptr<!fir.array<?xi32>>>>):
! CHECK: %[[VAL_1:.*]] = arith.constant 0 : i32
! CHECK: %[[VAL_2:.*]] = fir.load %[[VAL_0]] : !fir.ref<!fir.box<!fir.ptr<!fir.array<?xi32>>>>
! CHECK: %[[VAL_3:.*]] = fir.alloca !fir.box<!fir.ptr<!fir.array<?xi32>>>
! CHECK: %[[VAL_4:.*]] = fir.box_addr %[[VAL_2]] : (!fir.box<!fir.ptr<!fir.array<?xi32>>>) -> !fir.ptr<!fir.array<?xi32>>
! CHECK: %[[VAL_5:.*]] = fir.convert %[[VAL_4]] : (!fir.ptr<!fir.array<?xi32>>) -> i64
! CHECK: %[[VAL_6:.*]] = arith.constant 0 : i64
! CHECK: %[[VAL_7:.*]] = arith.cmpi eq, %[[VAL_5]], %[[VAL_6]] : i64
! CHECK: fir.if %[[VAL_7]] {
! CHECK: %[[VAL_8:.*]] = fir.embox %[[VAL_4]] : (!fir.ptr<!fir.array<?xi32>>) -> !fir.box<!fir.ptr<!fir.array<?xi32>>>
! CHECK: fir.store %[[VAL_8]] to %[[VAL_3]] : !fir.ref<!fir.box<!fir.ptr<!fir.array<?xi32>>>>
! CHECK: } else {
! CHECK: %[[VAL_9:.*]] = arith.constant 0 : index
! CHECK: %[[VAL_10:.*]]:3 = fir.box_dims %[[VAL_2]], %[[VAL_9]] : (!fir.box<!fir.ptr<!fir.array<?xi32>>>, index) -> (index, index, index)
! CHECK: %[[VAL_11:.*]] = fir.shape %[[VAL_10]]#1 : (index) -> !fir.shape<1>
! CHECK: %[[VAL_12:.*]] = fir.allocmem !fir.array<?xi32>, %[[VAL_10]]#1 {bindc_name = ".tmp", uniq_name = ""}
! CHECK: %[[VAL_13:.*]] = arith.constant true
! CHECK: %[[VAL_14:.*]]:2 = hlfir.declare %[[VAL_12]](%[[VAL_11]]) {uniq_name = ".tmp"} : (!fir.heap<!fir.array<?xi32>>, !fir.shape<1>) -> (!fir.box<!fir.array<?xi32>>, !fir.heap<!fir.array<?xi32>>)
! CHECK: %[[VAL_15:.*]] = arith.constant 0 : index
! CHECK: %[[VAL_16:.*]]:3 = fir.box_dims %[[VAL_2]], %[[VAL_15]] : (!fir.box<!fir.ptr<!fir.array<?xi32>>>, index) -> (index, index, index)
! CHECK: %[[VAL_17:.*]] = fir.shape_shift %[[VAL_16]]#0, %[[VAL_16]]#1 : (index, index) -> !fir.shapeshift<1>
! CHECK: %[[VAL_18:.*]] = fir.rebox %[[VAL_14]]#0(%[[VAL_17]]) : (!fir.box<!fir.array<?xi32>>, !fir.shapeshift<1>) -> !fir.box<!fir.ptr<!fir.array<?xi32>>>
! CHECK: hlfir.assign %[[VAL_1]] to %[[VAL_18]] : i32, !fir.box<!fir.ptr<!fir.array<?xi32>>>
! CHECK: fir.store %[[VAL_18]] to %[[VAL_3]] : !fir.ref<!fir.box<!fir.ptr<!fir.array<?xi32>>>>
! CHECK: }
! CHECK: omp.yield(%[[VAL_3]] : !fir.ref<!fir.box<!fir.ptr<!fir.array<?xi32>>>>)
! CHECK-LABEL: } combiner {
! CHECK: ^bb0(%[[VAL_0:.*]]: !fir.ref<!fir.box<!fir.ptr<!fir.array<?xi32>>>>, %[[VAL_1:.*]]: !fir.ref<!fir.box<!fir.ptr<!fir.array<?xi32>>>>):
! CHECK: %[[VAL_2:.*]] = fir.load %[[VAL_0]] : !fir.ref<!fir.box<!fir.ptr<!fir.array<?xi32>>>>
! CHECK: %[[VAL_3:.*]] = fir.load %[[VAL_1]] : !fir.ref<!fir.box<!fir.ptr<!fir.array<?xi32>>>>
! CHECK: %[[VAL_4:.*]] = arith.constant 0 : index
! CHECK: %[[VAL_5:.*]]:3 = fir.box_dims %[[VAL_2]], %[[VAL_4]] : (!fir.box<!fir.ptr<!fir.array<?xi32>>>, index) -> (index, index, index)
! CHECK: %[[VAL_6:.*]] = fir.shape_shift %[[VAL_5]]#0, %[[VAL_5]]#1 : (index, index) -> !fir.shapeshift<1>
! CHECK: %[[VAL_7:.*]] = arith.constant 1 : index
! CHECK: fir.do_loop %[[VAL_8:.*]] = %[[VAL_7]] to %[[VAL_5]]#1 step %[[VAL_7]] unordered {
! CHECK: %[[VAL_9:.*]] = fir.array_coor %[[VAL_2]](%[[VAL_6]]) %[[VAL_8]] : (!fir.box<!fir.ptr<!fir.array<?xi32>>>, !fir.shapeshift<1>, index) -> !fir.ref<i32>
! CHECK: %[[VAL_10:.*]] = fir.array_coor %[[VAL_3]](%[[VAL_6]]) %[[VAL_8]] : (!fir.box<!fir.ptr<!fir.array<?xi32>>>, !fir.shapeshift<1>, index) -> !fir.ref<i32>
! CHECK: %[[VAL_11:.*]] = fir.load %[[VAL_9]] : !fir.ref<i32>
! CHECK: %[[VAL_12:.*]] = fir.load %[[VAL_10]] : !fir.ref<i32>
! CHECK: %[[VAL_13:.*]] = arith.addi %[[VAL_11]], %[[VAL_12]] : i32
! CHECK: fir.store %[[VAL_13]] to %[[VAL_9]] : !fir.ref<i32>
! CHECK: }
! CHECK: omp.yield(%[[VAL_0]] : !fir.ref<!fir.box<!fir.ptr<!fir.array<?xi32>>>>)
! CHECK-LABEL: } cleanup {
! CHECK: ^bb0(%[[VAL_0:.*]]: !fir.ref<!fir.box<!fir.ptr<!fir.array<?xi32>>>>):
! CHECK: %[[VAL_1:.*]] = fir.load %[[VAL_0]] : !fir.ref<!fir.box<!fir.ptr<!fir.array<?xi32>>>>
! CHECK: %[[VAL_2:.*]] = fir.box_addr %[[VAL_1]] : (!fir.box<!fir.ptr<!fir.array<?xi32>>>) -> !fir.ptr<!fir.array<?xi32>>
! CHECK: %[[VAL_3:.*]] = fir.convert %[[VAL_2]] : (!fir.ptr<!fir.array<?xi32>>) -> i64
! CHECK: %[[VAL_4:.*]] = arith.constant 0 : i64
! CHECK: %[[VAL_5:.*]] = arith.cmpi ne, %[[VAL_3]], %[[VAL_4]] : i64
! CHECK: fir.if %[[VAL_5]] {
! CHECK: %[[VAL_6:.*]] = fir.convert %[[VAL_2]] : (!fir.ptr<!fir.array<?xi32>>) -> !fir.heap<!fir.array<?xi32>>
! CHECK: fir.freemem %[[VAL_6]] : !fir.heap<!fir.array<?xi32>>
! CHECK: }
! CHECK: omp.yield
! CHECK: }

! CHECK-LABEL: func.func @_QQmain() attributes {fir.bindc_name = "reduce"} {
! CHECK: %[[VAL_0:.*]] = fir.address_of(@_QFEi) : !fir.ref<i32>
! CHECK: %[[VAL_1:.*]]:2 = hlfir.declare %[[VAL_0]] {uniq_name = "_QFEi"} : (!fir.ref<i32>) -> (!fir.ref<i32>, !fir.ref<i32>)
! CHECK: %[[VAL_2:.*]] = fir.address_of(@_QFEr) : !fir.ref<!fir.box<!fir.ptr<!fir.array<?xi32>>>>
! CHECK: %[[VAL_3:.*]]:2 = hlfir.declare %[[VAL_2]] {fortran_attrs = {{.*}}<pointer>, uniq_name = "_QFEr"} : (!fir.ref<!fir.box<!fir.ptr<!fir.array<?xi32>>>>) -> (!fir.ref<!fir.box<!fir.ptr<!fir.array<?xi32>>>>, !fir.ref<!fir.box<!fir.ptr<!fir.array<?xi32>>>>)
! CHECK: %[[VAL_4:.*]] = arith.constant false
! CHECK: %[[VAL_5:.*]] = fir.absent !fir.box<none>
! CHECK: %[[VAL_6:.*]] = fir.address_of(
! CHECK: %[[VAL_7:.*]] = arith.constant 8 : i32
! CHECK: %[[VAL_8:.*]] = fir.zero_bits !fir.ptr<!fir.array<?xi32>>
! CHECK: %[[VAL_9:.*]] = arith.constant 0 : index
! CHECK: %[[VAL_10:.*]] = fir.shape %[[VAL_9]] : (index) -> !fir.shape<1>
! CHECK: %[[VAL_11:.*]] = fir.embox %[[VAL_8]](%[[VAL_10]]) : (!fir.ptr<!fir.array<?xi32>>, !fir.shape<1>) -> !fir.box<!fir.ptr<!fir.array<?xi32>>>
! CHECK: fir.store %[[VAL_11]] to %[[VAL_3]]#1 : !fir.ref<!fir.box<!fir.ptr<!fir.array<?xi32>>>>
! CHECK: %[[VAL_12:.*]] = arith.constant 1 : index
! CHECK: %[[VAL_13:.*]] = arith.constant 2 : i32
! CHECK: %[[VAL_14:.*]] = arith.constant 0 : i32
! CHECK: %[[VAL_15:.*]] = fir.convert %[[VAL_3]]#1 : (!fir.ref<!fir.box<!fir.ptr<!fir.array<?xi32>>>>) -> !fir.ref<!fir.box<none>>
! CHECK: %[[VAL_16:.*]] = fir.convert %[[VAL_12]] : (index) -> i64
! CHECK: %[[VAL_17:.*]] = fir.convert %[[VAL_13]] : (i32) -> i64
! CHECK: %[[VAL_18:.*]] = fir.call @_FortranAPointerSetBounds(%[[VAL_15]], %[[VAL_14]], %[[VAL_16]], %[[VAL_17]]) fastmath<contract> : (!fir.ref<!fir.box<none>>, i32, i64, i64) -> none
! CHECK: %[[VAL_19:.*]] = fir.convert %[[VAL_3]]#1 : (!fir.ref<!fir.box<!fir.ptr<!fir.array<?xi32>>>>) -> !fir.ref<!fir.box<none>>
! CHECK: %[[VAL_20:.*]] = fir.convert %[[VAL_6]] : (!fir.ref<!fir.char<{{.*}}>>) -> !fir.ref<i8>
! CHECK: %[[VAL_21:.*]] = fir.call @_FortranAPointerAllocate(%[[VAL_19]], %[[VAL_4]], %[[VAL_5]], %[[VAL_20]], %[[VAL_7]]) fastmath<contract> : (!fir.ref<!fir.box<none>>, i1, !fir.box<none>, !fir.ref<i8>, i32) -> i32
! CHECK: omp.parallel {
! CHECK: %[[VAL_22:.*]] = fir.alloca i32 {bindc_name = "i", pinned, uniq_name = "_QFEi"}
! CHECK: %[[VAL_23:.*]]:2 = hlfir.declare %[[VAL_22]] {uniq_name = "_QFEi"} : (!fir.ref<i32>) -> (!fir.ref<i32>, !fir.ref<i32>)
! CHECK: %[[VAL_24:.*]] = arith.constant 0 : i32
! CHECK: %[[VAL_25:.*]] = arith.constant 10 : i32
! CHECK: %[[VAL_26:.*]] = arith.constant 1 : i32
! CHECK: omp.wsloop reduction(byref @add_reduction_byref_box_ptr_Uxi32 %[[VAL_3]]#0 -> %[[VAL_27:.*]] : !fir.ref<!fir.box<!fir.ptr<!fir.array<?xi32>>>>) {
! CHECK: omp.loop_nest (%[[VAL_28:.*]]) : i32 = (%[[VAL_24]]) to (%[[VAL_25]]) inclusive step (%[[VAL_26]]) {
! CHECK: %[[VAL_29:.*]]:2 = hlfir.declare %[[VAL_27]] {fortran_attrs = {{.*}}<pointer>, uniq_name = "_QFEr"} : (!fir.ref<!fir.box<!fir.ptr<!fir.array<?xi32>>>>) -> (!fir.ref<!fir.box<!fir.ptr<!fir.array<?xi32>>>>, !fir.ref<!fir.box<!fir.ptr<!fir.array<?xi32>>>>)
! CHECK: fir.store %[[VAL_28]] to %[[VAL_23]]#1 : !fir.ref<i32>
! CHECK: %[[VAL_30:.*]] = fir.load %[[VAL_23]]#0 : !fir.ref<i32>
! CHECK: %[[VAL_31:.*]] = fir.load %[[VAL_29]]#0 : !fir.ref<!fir.box<!fir.ptr<!fir.array<?xi32>>>>
! CHECK: %[[VAL_32:.*]] = arith.constant 1 : index
! CHECK: %[[VAL_33:.*]] = hlfir.designate %[[VAL_31]] (%[[VAL_32]]) : (!fir.box<!fir.ptr<!fir.array<?xi32>>>, index) -> !fir.ref<i32>
! CHECK: hlfir.assign %[[VAL_30]] to %[[VAL_33]] : i32, !fir.ref<i32>
! CHECK: %[[VAL_34:.*]] = fir.load %[[VAL_23]]#0 : !fir.ref<i32>
! CHECK: %[[VAL_35:.*]] = arith.constant 0 : i32
! CHECK: %[[VAL_36:.*]] = arith.subi %[[VAL_35]], %[[VAL_34]] : i32
! CHECK: %[[VAL_37:.*]] = fir.load %[[VAL_29]]#0 : !fir.ref<!fir.box<!fir.ptr<!fir.array<?xi32>>>>
! CHECK: %[[VAL_38:.*]] = arith.constant 2 : index
! CHECK: %[[VAL_39:.*]] = hlfir.designate %[[VAL_37]] (%[[VAL_38]]) : (!fir.box<!fir.ptr<!fir.array<?xi32>>>, index) -> !fir.ref<i32>
! CHECK: hlfir.assign %[[VAL_36]] to %[[VAL_39]] : i32, !fir.ref<i32>
! CHECK: omp.yield
! CHECK: }
! CHECK: omp.terminator
! CHECK: }
! CHECK: omp.terminator
! CHECK: }
Loading

0 comments on commit b6b0f97

Please sign in to comment.