[flang][acc] remap symbol appearing in reduction clause #168876
Merged
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
llvmbot
added
flang
Member
|
@llvm/pr-subscribers-flang-fir-hlfir @llvm/pr-subscribers-openacc Author: None (jeanPerier) ChangesThis patch is a follow-up of #162306 for the reduction clause. Inside the compute region that carries the reduction clause, a new hlfir.declare is generated for symbol appearing in the reduction clause. The input of this hlfir.declare is the acc.reduction result. The related semantics::Symbol is remapped to the hlfir.declare result so that any reference to the symbol inside the compute region will use this SSA value as the starting point instead of the SSA value for the host address. Full diff: https://github.com/llvm/llvm-project/pull/168876.diff 2 Files Affected:
diff --git a/flang/lib/Lower/OpenACC.cpp b/flang/lib/Lower/OpenACC.cpp
index 98a3aced3f528..d10169e236471 100644
--- a/flang/lib/Lower/OpenACC.cpp
+++ b/flang/lib/Lower/OpenACC.cpp
@@ -1622,16 +1622,18 @@ static bool isSupportedReductionType(mlir::Type ty) {
return fir::isa_trivial(ty);
}
-static void
-genReductions(const Fortran::parser::AccObjectListWithReduction &objectList,
- Fortran::lower::AbstractConverter &converter,
- Fortran::semantics::SemanticsContext &semanticsContext,
- Fortran::lower::StatementContext &stmtCtx,
- llvm::SmallVectorImpl<mlir::Value> &reductionOperands,
- llvm::SmallVector<mlir::Attribute> &reductionRecipes,
- llvm::ArrayRef<mlir::Value> async,
- llvm::ArrayRef<mlir::Attribute> asyncDeviceTypes,
- llvm::ArrayRef<mlir::Attribute> asyncOnlyDeviceTypes) {
+static void genReductions(
+ const Fortran::parser::AccObjectListWithReduction &objectList,
+ Fortran::lower::AbstractConverter &converter,
+ Fortran::semantics::SemanticsContext &semanticsContext,
+ Fortran::lower::StatementContext &stmtCtx,
+ llvm::SmallVectorImpl<mlir::Value> &reductionOperands,
+ llvm::SmallVector<mlir::Attribute> &reductionRecipes,
+ llvm::ArrayRef<mlir::Value> async,
+ llvm::ArrayRef<mlir::Attribute> asyncDeviceTypes,
+ llvm::ArrayRef<mlir::Attribute> asyncOnlyDeviceTypes,
+ llvm::SmallVectorImpl<std::pair<mlir::Value, Fortran::semantics::SymbolRef>>
+ *symbolPairs = nullptr) {
fir::FirOpBuilder &builder = converter.getFirOpBuilder();
const auto &objects = std::get<Fortran::parser::AccObjectList>(objectList.t);
const auto &op = std::get<Fortran::parser::ReductionOperator>(objectList.t);
@@ -1644,6 +1646,8 @@ genReductions(const Fortran::parser::AccObjectListWithReduction &objectList,
Fortran::semantics::Symbol &symbol = getSymbolFromAccObject(accObject);
Fortran::semantics::MaybeExpr designator = Fortran::common::visit(
[&](auto &&s) { return ea.Analyze(s); }, accObject.u);
+ bool isWholeSymbol =
+ !designator || Fortran::evaluate::UnwrapWholeSymbolDataRef(*designator);
fir::factory::AddrAndBoundsInfo info =
Fortran::lower::gatherDataOperandAddrAndBounds<
mlir::acc::DataBoundsOp, mlir::acc::DataBoundsType>(
@@ -1680,6 +1684,11 @@ genReductions(const Fortran::parser::AccObjectListWithReduction &objectList,
reductionRecipes.push_back(mlir::SymbolRefAttr::get(
builder.getContext(), recipe.getSymName().str()));
reductionOperands.push_back(op.getAccVar());
+ // Track the symbol and its corresponding mlir::Value if requested so that
+ // accesses inside the compute/loop regions use the acc.reduction variable.
+ if (symbolPairs && isWholeSymbol)
+ symbolPairs->emplace_back(op.getAccVar(),
+ Fortran::semantics::SymbolRef(symbol));
}
}
@@ -2545,7 +2554,8 @@ static mlir::acc::LoopOp createLoopOp(
&clause.u)) {
genReductions(reductionClause->v, converter, semanticsContext, stmtCtx,
reductionOperands, reductionRecipes, /*async=*/{},
- /*asyncDeviceTypes=*/{}, /*asyncOnlyDeviceTypes=*/{});
+ /*asyncDeviceTypes=*/{}, /*asyncOnlyDeviceTypes=*/{},
+ &dataOperandSymbolPairs);
} else if (std::get_if<Fortran::parser::AccClause::Seq>(&clause.u)) {
for (auto crtDeviceTypeAttr : crtDeviceTypes)
seqDeviceTypes.push_back(crtDeviceTypeAttr);
@@ -2995,7 +3005,8 @@ static Op createComputeOp(
if (!combinedConstructs) {
genReductions(reductionClause->v, converter, semanticsContext, stmtCtx,
reductionOperands, reductionRecipes, async,
- asyncDeviceTypes, asyncOnlyDeviceTypes);
+ asyncDeviceTypes, asyncOnlyDeviceTypes,
+ &dataOperandSymbolPairs);
} else {
auto crtDataStart = dataClauseOperands.size();
genDataOperandOperations<mlir::acc::CopyinOp>(
diff --git a/flang/test/Lower/OpenACC/acc-reduction-remapping.f90 b/flang/test/Lower/OpenACC/acc-reduction-remapping.f90
new file mode 100644
index 0000000000000..6ee365f278678
--- /dev/null
+++ b/flang/test/Lower/OpenACC/acc-reduction-remapping.f90
@@ -0,0 +1,160 @@
+! Test remapping of variables appearing in OpenACC reduction clause
+! to the related acc dialect data operation result.
+
+! This tests checks how the hlfir.declare is recreated and used inside
+! the acc compute region.
+
+! RUN: bbc -fopenacc -emit-hlfir %s -o - | FileCheck %s
+
+subroutine scalar_combined(x, y)
+ real :: y, x(100)
+ !$acc parallel copyin(x) reduction(+:y)
+ do i=1,100
+ y = y + x(i)
+ end do
+ !$acc end parallel
+end subroutine
+
+subroutine scalar_split(x, y)
+ real :: y, x(100)
+ !$acc parallel copyin(x) copyout(y)
+ !$acc loop reduction(+:y)
+ do i=1,100
+ y = y + x(i)
+ end do
+ !$acc end parallel
+end subroutine
+
+subroutine array_combined(x, y, n)
+ integer(8) :: n
+ real :: y(n), x(100, n)
+ !$acc parallel copyin(x) reduction(+:y)
+ do j=1,n
+ do i=1,100
+ y(j) = y(j) + x(i, j)
+ end do
+ end do
+ !$acc end parallel
+end subroutine
+
+subroutine array_split(x, y, n)
+ integer(8) :: n
+ real :: y(n), x(100, n)
+ !$acc parallel copyin(x) copyout(y)
+ !$acc loop reduction(+:y)
+ do j=1,n
+ do i=1,100
+ y(j) = y(j) + x(i, j)
+ end do
+ end do
+ !$acc end parallel
+end subroutine
+
+! CHECK-LABEL: func.func @_QPscalar_combined(
+! CHECK: %[[DUMMY_SCOPE_0:.*]] = fir.dummy_scope : !fir.dscope
+! CHECK: %[[DECLARE_Y:.*]]:2 = hlfir.declare %{{.*}} dummy_scope %[[DUMMY_SCOPE_0]] arg 2 {uniq_name = "_QFscalar_combinedEy"} : (!fir.ref<f32>, !fir.dscope) -> (!fir.ref<f32>, !fir.ref<f32>)
+! CHECK: %[[REDUCTION_Y:.*]] = acc.reduction varPtr(%[[DECLARE_Y]]#0 : !fir.ref<f32>) -> !fir.ref<f32> {name = "y"}
+! CHECK: acc.parallel {{.*}} reduction(@reduction_add_ref_f32 -> %[[REDUCTION_Y]] : !fir.ref<f32>) {
+! CHECK: %[[DUMMY_SCOPE_1:.*]] = fir.dummy_scope : !fir.dscope
+! CHECK: %[[DECLARE_RED_PAR:.*]]:2 = hlfir.declare %[[REDUCTION_Y]] dummy_scope %[[DUMMY_SCOPE_1]] arg 2 {uniq_name = "_QFscalar_combinedEy"} : (!fir.ref<f32>, !fir.dscope) -> (!fir.ref<f32>, !fir.ref<f32>)
+! CHECK: %[[PRIVATE_I:.*]] = acc.private varPtr({{.*}}) -> !fir.ref<i32> {implicit = true, name = "i"}
+! CHECK: acc.loop private(@privatization_ref_i32 -> %[[PRIVATE_I]] : !fir.ref<i32>) {{.*}} {
+! CHECK: %[[DECLARE_I:.*]]:2 = hlfir.declare %[[PRIVATE_I]] {uniq_name = "_QFscalar_combinedEi"} : (!fir.ref<i32>) -> (!fir.ref<i32>, !fir.ref<i32>)
+! CHECK: fir.store %[[VAL_0:.*]] to %[[DECLARE_I]]#0 : !fir.ref<i32>
+! CHECK: %[[LOAD_RED:.*]] = fir.load %[[DECLARE_RED_PAR]]#0 : !fir.ref<f32>
+! CHECK: {{.*}} = hlfir.designate {{.*}} : (!fir.ref<!fir.array<100xf32>>, i64) -> !fir.ref<f32>
+! CHECK: {{.*}} = fir.load {{.*}} : !fir.ref<f32>
+! CHECK: %[[ADDF:.*]] = arith.addf %[[LOAD_RED]], {{.*}} {{.*}}: f32
+! CHECK: hlfir.assign %[[ADDF]] to %[[DECLARE_RED_PAR]]#0 : f32, !fir.ref<f32>
+! CHECK: acc.yield
+! CHECK: }
+! CHECK: acc.yield
+! CHECK: }
+! CHECK: return
+! CHECK: }
+!
+!
+! CHECK-LABEL: func.func @_QPscalar_split(
+! CHECK: %[[DUMMY_SCOPE_0:.*]] = fir.dummy_scope : !fir.dscope
+! CHECK: %[[DECLARE_Y:.*]]:2 = hlfir.declare %{{.*}} dummy_scope %[[DUMMY_SCOPE_0]] arg 2 {uniq_name = "_QFscalar_splitEy"} : (!fir.ref<f32>, !fir.dscope) -> (!fir.ref<f32>, !fir.ref<f32>)
+! CHECK: %[[CREATE_Y:.*]] = acc.create varPtr(%[[DECLARE_Y]]#0 : !fir.ref<f32>) -> !fir.ref<f32> {dataClause = #acc<data_clause acc_copyout>, name = "y"}
+! CHECK: acc.parallel dataOperands({{.*}}%[[CREATE_Y]] : {{.*}}) {
+! CHECK: %[[DUMMY_SCOPE_1:.*]] = fir.dummy_scope : !fir.dscope
+! CHECK: %[[DECLARE_Y_PAR:.*]]:2 = hlfir.declare %[[CREATE_Y]] dummy_scope %[[DUMMY_SCOPE_1]] arg 2 {uniq_name = "_QFscalar_splitEy"} : (!fir.ref<f32>, !fir.dscope) -> (!fir.ref<f32>, !fir.ref<f32>)
+! CHECK: %[[REDUCTION_Y:.*]] = acc.reduction varPtr(%[[DECLARE_Y_PAR]]#0 : !fir.ref<f32>) -> !fir.ref<f32> {name = "y"}
+! CHECK: %[[PRIVATE_I:.*]] = acc.private varPtr({{.*}}) -> !fir.ref<i32> {implicit = true, name = "i"}
+! CHECK: acc.loop private(@privatization_ref_i32 -> %[[PRIVATE_I]] : !fir.ref<i32>) reduction(@reduction_add_ref_f32 -> %[[REDUCTION_Y]] : !fir.ref<f32>) {{.*}} {
+! CHECK: %[[DUMMY_SCOPE_2:.*]] = fir.dummy_scope : !fir.dscope
+! CHECK: %[[DECLARE_RED:.*]]:2 = hlfir.declare %[[REDUCTION_Y]] dummy_scope %[[DUMMY_SCOPE_2]] arg 2 {uniq_name = "_QFscalar_splitEy"} : (!fir.ref<f32>, !fir.dscope) -> (!fir.ref<f32>, !fir.ref<f32>)
+! CHECK: %[[LOAD_RED:.*]] = fir.load %[[DECLARE_RED]]#0 : !fir.ref<f32>
+! CHECK: %[[ADDF:.*]] = arith.addf %[[LOAD_RED]], {{.*}} {{.*}}: f32
+! CHECK: hlfir.assign %[[ADDF]] to %[[DECLARE_RED]]#0 : f32, !fir.ref<f32>
+! CHECK: acc.yield
+! CHECK: }
+! CHECK: acc.yield
+! CHECK: }
+! CHECK: return
+! CHECK: }
+
+
+! CHECK-LABEL: func.func @_QParray_combined(
+! CHECK: %[[DUMMY_SCOPE_0:.*]] = fir.dummy_scope : !fir.dscope
+! CHECK: %[[DECLARE_N:.*]]:2 = hlfir.declare %{{.*}} dummy_scope %[[DUMMY_SCOPE_0]] arg 3 {uniq_name = "_QFarray_combinedEn"} : (!fir.ref<i64>, !fir.dscope) -> (!fir.ref<i64>, !fir.ref<i64>)
+! CHECK: %[[DECLARE_Y:.*]]:2 = hlfir.declare %{{.*}}({{.*}}) dummy_scope %[[DUMMY_SCOPE_0]] arg 2 {uniq_name = "_QFarray_combinedEy"} : (!fir.ref<!fir.array<?xf32>>, {{.*}}, !fir.dscope) -> (!fir.box<!fir.array<?xf32>>, !fir.ref<!fir.array<?xf32>>)
+! CHECK: %[[REDUCTION_Y:.*]] = acc.reduction var(%[[DECLARE_Y]]#0 : !fir.box<!fir.array<?xf32>>) -> !fir.box<!fir.array<?xf32>> {name = "y"}
+! CHECK: acc.parallel dataOperands({{.*}}) reduction(@reduction_add_box_Uxf32 -> %[[REDUCTION_Y]] : !fir.box<!fir.array<?xf32>>) {
+! CHECK: %[[DUMMY_SCOPE_1:.*]] = fir.dummy_scope : !fir.dscope
+! CHECK: %[[BOX_ADDR_RED:.*]] = fir.box_addr %[[REDUCTION_Y]] : (!fir.box<!fir.array<?xf32>>) -> !fir.ref<!fir.array<?xf32>>
+! CHECK: %[[DECLARE_Y_PAR:.*]]:2 = hlfir.declare %[[BOX_ADDR_RED]]({{.*}}) dummy_scope %[[DUMMY_SCOPE_1]] arg 2 {uniq_name = "_QFarray_combinedEy"} : (!fir.ref<!fir.array<?xf32>>, {{.*}}, !fir.dscope) -> (!fir.box<!fir.array<?xf32>>, !fir.ref<!fir.array<?xf32>>)
+! CHECK: %[[PRIVATE_J:.*]] = acc.private varPtr({{.*}}) -> !fir.ref<i32> {implicit = true, name = "j"}
+! CHECK: acc.loop private(@privatization_ref_i32 -> %[[PRIVATE_J]] : !fir.ref<i32>) {{.*}} {
+! CHECK: %[[PRIVATE_I:.*]] = acc.private varPtr({{.*}}) -> !fir.ref<i32> {implicit = true, name = "i"}
+! CHECK: acc.loop private(@privatization_ref_i32 -> %[[PRIVATE_I]] : !fir.ref<i32>) {{.*}} {
+! CHECK: %[[DESIGNATE_RED:.*]] = hlfir.designate %[[DECLARE_Y_PAR]]#0 ({{.*}}) : (!fir.box<!fir.array<?xf32>>, i64) -> !fir.ref<f32>
+! CHECK: %[[LOAD_OLD:.*]] = fir.load %[[DESIGNATE_RED]] : !fir.ref<f32>
+! CHECK: {{.*}} = hlfir.designate {{.*}} : (!fir.box<!fir.array<100x?xf32>>, i64, i64) -> !fir.ref<f32>
+! CHECK: {{.*}} = fir.load {{.*}} : !fir.ref<f32>
+! CHECK: %[[ADDF:.*]] = arith.addf %[[LOAD_OLD]], {{.*}} {{.*}}: f32
+! CHECK: %[[DESIGNATE_RED2:.*]] = hlfir.designate %[[DECLARE_Y_PAR]]#0 ({{.*}}) : (!fir.box<!fir.array<?xf32>>, i64) -> !fir.ref<f32>
+! CHECK: hlfir.assign %[[ADDF]] to %[[DESIGNATE_RED2]] : f32, !fir.ref<f32>
+! CHECK: acc.yield
+! CHECK: }
+! CHECK: acc.yield
+! CHECK: }
+! CHECK: acc.yield
+! CHECK: }
+! CHECK: return
+! CHECK: }
+
+
+! CHECK-LABEL: func.func @_QParray_split(
+! CHECK: %[[DUMMY_SCOPE_0:.*]] = fir.dummy_scope : !fir.dscope
+! CHECK: %[[DECLARE_Y:.*]]:2 = hlfir.declare %{{.*}}({{.*}}) dummy_scope %[[DUMMY_SCOPE_0]] arg 2 {uniq_name = "_QFarray_splitEy"} : (!fir.ref<!fir.array<?xf32>>, {{.*}}, !fir.dscope) -> (!fir.box<!fir.array<?xf32>>, !fir.ref<!fir.array<?xf32>>)
+! CHECK: %[[CREATE_Y:.*]] = acc.create var(%[[DECLARE_Y]]#0 : !fir.box<!fir.array<?xf32>>) -> !fir.box<!fir.array<?xf32>> {dataClause = #acc<data_clause acc_copyout>, name = "y"}
+! CHECK: acc.parallel dataOperands({{.*}}%[[CREATE_Y]] : {{.*}}) {
+! CHECK: %[[DUMMY_SCOPE_1:.*]] = fir.dummy_scope : !fir.dscope
+! CHECK: %[[BOX_ADDR_Y:.*]] = fir.box_addr %[[CREATE_Y]] : (!fir.box<!fir.array<?xf32>>) -> !fir.ref<!fir.array<?xf32>>
+! CHECK: %[[DECLARE_Y_PAR:.*]]:2 = hlfir.declare %[[BOX_ADDR_Y]]({{.*}}) dummy_scope %[[DUMMY_SCOPE_1]] arg 2 {uniq_name = "_QFarray_splitEy"} : (!fir.ref<!fir.array<?xf32>>, {{.*}}, !fir.dscope) -> (!fir.box<!fir.array<?xf32>>, !fir.ref<!fir.array<?xf32>>)
+! CHECK: %[[REDUCTION_Y:.*]] = acc.reduction var(%[[DECLARE_Y_PAR]]#0 : !fir.box<!fir.array<?xf32>>) -> !fir.box<!fir.array<?xf32>> {name = "y"}
+! CHECK: %[[PRIVATE_J:.*]] = acc.private varPtr({{.*}}) -> !fir.ref<i32> {implicit = true, name = "j"}
+! CHECK: acc.loop private(@privatization_ref_i32 -> %[[PRIVATE_J]] : !fir.ref<i32>) reduction(@reduction_add_box_Uxf32 -> %[[REDUCTION_Y]] : !fir.box<!fir.array<?xf32>>) {{.*}} {
+! CHECK: %[[BOX_ADDR_RED:.*]] = fir.box_addr %[[REDUCTION_Y]] : (!fir.box<!fir.array<?xf32>>) -> !fir.ref<!fir.array<?xf32>>
+! CHECK: %[[DUMMY_SCOPE_2:.*]] = fir.dummy_scope : !fir.dscope
+! CHECK: %[[DECLARE_Y_LOOP_PAR:.*]]:2 = hlfir.declare %[[BOX_ADDR_RED]]({{.*}}) dummy_scope %[[DUMMY_SCOPE_2]] arg 2 {uniq_name = "_QFarray_splitEy"} : (!fir.ref<!fir.array<?xf32>>, {{.*}}, !fir.dscope) -> (!fir.box<!fir.array<?xf32>>, !fir.ref<!fir.array<?xf32>>)
+! CHECK: %[[PRIVATE_I:.*]] = acc.private varPtr({{.*}}) -> !fir.ref<i32> {implicit = true, name = "i"}
+! CHECK: acc.loop private(@privatization_ref_i32 -> %[[PRIVATE_I]] : !fir.ref<i32>) {{.*}} {
+! CHECK: %[[DESIGNATE_RED:.*]] = hlfir.designate %[[DECLARE_Y_LOOP_PAR]]#0 ({{.*}}) : (!fir.box<!fir.array<?xf32>>, i64) -> !fir.ref<f32>
+! CHECK: %[[LOAD_OLD:.*]] = fir.load %[[DESIGNATE_RED]] : !fir.ref<f32>
+! CHECK: {{.*}} = hlfir.designate {{.*}} : (!fir.box<!fir.array<100x?xf32>>, i64, i64) -> !fir.ref<f32>
+! CHECK: {{.*}} = fir.load {{.*}} : !fir.ref<f32>
+! CHECK: %[[ADDF:.*]] = arith.addf %[[LOAD_OLD]], {{.*}} {{.*}}: f32
+! CHECK: %[[DESIGNATE_RED2:.*]] = hlfir.designate %[[DECLARE_Y_LOOP_PAR]]#0 ({{.*}}) : (!fir.box<!fir.array<?xf32>>, i64) -> !fir.ref<f32>
+! CHECK: hlfir.assign %[[ADDF]] to %[[DESIGNATE_RED2]] : f32, !fir.ref<f32>
+! CHECK: acc.yield
+! CHECK: }
+! CHECK: acc.yield
+! CHECK: }
+! CHECK: acc.yield
+! CHECK: }
+! CHECK: return
+! CHECK: }
|
🐧 Linux x64 Test Results
|
vzakhari
approved these changes
Nov 25, 2025
augusto2112
pushed a commit
to augusto2112/llvm-project
that referenced
this pull request
Dec 3, 2025
This patch is a follow-up of llvm#162306 for the reduction clause. Inside the compute region that carries the reduction clause, a new hlfir.declare is generated for symbol appearing in the reduction clause. The input of this hlfir.declare is the acc.reduction result. The related semantics::Symbol is remapped to the hlfir.declare result so that any reference to the symbol inside the compute region will use this SSA value as the starting point instead of the SSA value for the host address.
kcloudy0717
pushed a commit
to kcloudy0717/llvm-project
that referenced
this pull request
Dec 4, 2025
This patch is a follow-up of llvm#162306 for the reduction clause. Inside the compute region that carries the reduction clause, a new hlfir.declare is generated for symbol appearing in the reduction clause. The input of this hlfir.declare is the acc.reduction result. The related semantics::Symbol is remapped to the hlfir.declare result so that any reference to the symbol inside the compute region will use this SSA value as the starting point instead of the SSA value for the host address.
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
4 participants
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
This patch is a follow-up of #162306 for the reduction clause.
Inside the compute region that carries the reduction clause, a new hlfir.declare is generated for symbol appearing in the reduction clause. The input of this hlfir.declare is the acc.reduction result. The related semantics::Symbol is remapped to the hlfir.declare result so that any reference to the symbol inside the compute region will use this SSA value as the starting point instead of the SSA value for the host address.