[flang-commits] [flang] dbd6eb6 - [flang][OpenMP] lower reductions of assumed shape arrays (#86982)
via flang-commits
flang-commits at lists.llvm.org
Thu Apr 4 03:14:31 PDT 2024
Author: Tom Eccles
Date: 2024-04-04T11:14:28+01:00
New Revision: dbd6eb6779bc7073c5466e9689eb5a69736d1120
URL: https://github.com/llvm/llvm-project/commit/dbd6eb6779bc7073c5466e9689eb5a69736d1120
DIFF: https://github.com/llvm/llvm-project/commit/dbd6eb6779bc7073c5466e9689eb5a69736d1120.diff
LOG: [flang][OpenMP] lower reductions of assumed shape arrays (#86982)
Patch 1: https://github.com/llvm/llvm-project/pull/86978
Patch 2: https://github.com/llvm/llvm-project/pull/86979
Added:
flang/test/Lower/OpenMP/wsloop-reduction-array-assumed-shape.f90
Modified:
flang/lib/Lower/OpenMP/ReductionProcessor.cpp
Removed:
################################################################################
diff --git a/flang/lib/Lower/OpenMP/ReductionProcessor.cpp b/flang/lib/Lower/OpenMP/ReductionProcessor.cpp
index 6a8447a00e4c6d..c1c94119fd9083 100644
--- a/flang/lib/Lower/OpenMP/ReductionProcessor.cpp
+++ b/flang/lib/Lower/OpenMP/ReductionProcessor.cpp
@@ -523,11 +523,16 @@ void ReductionProcessor::addDeclareReduction(
if (reductionSymbols)
reductionSymbols->push_back(symbol);
mlir::Value symVal = converter.getSymbolAddress(*symbol);
- auto redType = mlir::cast<fir::ReferenceType>(symVal.getType());
+ mlir::Type eleType;
+ auto refType = mlir::dyn_cast_or_null<fir::ReferenceType>(symVal.getType());
+ if (refType)
+ eleType = refType.getEleTy();
+ else
+ eleType = symVal.getType();
// all arrays must be boxed so that we have convenient access to all the
// information needed to iterate over the array
- if (mlir::isa<fir::SequenceType>(redType.getEleTy())) {
+ if (mlir::isa<fir::SequenceType>(eleType)) {
// For Host associated symbols, use `SymbolBox` instead
Fortran::lower::SymbolBox symBox =
converter.lookupOneLevelUpSymbol(*symbol);
@@ -542,11 +547,25 @@ void ReductionProcessor::addDeclareReduction(
builder.create<fir::StoreOp>(currentLocation, box, alloca);
symVal = alloca;
- redType = mlir::cast<fir::ReferenceType>(symVal.getType());
+ } else if (mlir::isa<fir::BaseBoxType>(symVal.getType())) {
+ // boxed arrays are passed as values not by reference. Unfortunately,
+ // we can't pass a box by value to omp.redution_declare, so turn it
+ // into a reference
+
+ auto alloca =
+ builder.create<fir::AllocaOp>(currentLocation, symVal.getType());
+ builder.create<fir::StoreOp>(currentLocation, symVal, alloca);
+ symVal = alloca;
} else if (auto declOp = symVal.getDefiningOp<hlfir::DeclareOp>()) {
symVal = declOp.getBase();
}
+ // this isn't the same as the by-val and by-ref passing later in the
+ // pipeline. Both styles assume that the variable is a reference at
+ // this point
+ assert(mlir::isa<fir::ReferenceType>(symVal.getType()) &&
+ "reduction input var is a reference");
+
reductionVars.push_back(symVal);
}
const bool isByRef = doReductionByRef(reductionVars);
diff --git a/flang/test/Lower/OpenMP/wsloop-reduction-array-assumed-shape.f90 b/flang/test/Lower/OpenMP/wsloop-reduction-array-assumed-shape.f90
new file mode 100644
index 00000000000000..a1f339faea5cd5
--- /dev/null
+++ b/flang/test/Lower/OpenMP/wsloop-reduction-array-assumed-shape.f90
@@ -0,0 +1,90 @@
+! RUN: bbc -emit-hlfir -fopenmp -o - %s | FileCheck %s
+! RUN: %flang_fc1 -emit-hlfir -fopenmp -o - %s | FileCheck %s
+
+program reduce_assumed_shape
+real(8), dimension(2) :: r
+r = 0
+call reduce(r)
+print *, r
+
+contains
+subroutine reduce(r)
+ implicit none
+ real(8),intent(inout) :: r(:)
+ integer :: i = 0
+
+ !$omp parallel do reduction(+:r)
+ do i=0,10
+ r(1) = i
+ r(2) = 1
+ enddo
+ !$omp end parallel do
+end subroutine
+end program
+
+! CHECK-LABEL: omp.declare_reduction @add_reduction_byref_box_Uxf64 : !fir.ref<!fir.box<!fir.array<?xf64>>> init {
+! CHECK: ^bb0(%[[VAL_0:.*]]: !fir.ref<!fir.box<!fir.array<?xf64>>>):
+! CHECK: %[[VAL_1:.*]] = arith.constant 0.000000e+00 : f64
+! CHECK: %[[VAL_2:.*]] = fir.load %[[VAL_0]] : !fir.ref<!fir.box<!fir.array<?xf64>>>
+! CHECK: %[[VAL_3:.*]] = arith.constant 0 : index
+! CHECK: %[[VAL_4:.*]]:3 = fir.box_dims %[[VAL_2]], %[[VAL_3]] : (!fir.box<!fir.array<?xf64>>, index) -> (index, index, index)
+! CHECK: %[[VAL_5:.*]] = fir.shape %[[VAL_4]]#1 : (index) -> !fir.shape<1>
+! CHECK: %[[VAL_6:.*]] = fir.alloca !fir.array<?xf64>, %[[VAL_4]]#1 {bindc_name = ".tmp"}
+! CHECK: %[[VAL_7:.*]]:2 = hlfir.declare %[[VAL_6]](%[[VAL_5]]) {uniq_name = ".tmp"} : (!fir.ref<!fir.array<?xf64>>, !fir.shape<1>) -> (!fir.box<!fir.array<?xf64>>, !fir.ref<!fir.array<?xf64>>)
+! CHECK: hlfir.assign %[[VAL_1]] to %[[VAL_7]]#0 : f64, !fir.box<!fir.array<?xf64>>
+! CHECK: %[[VAL_8:.*]] = fir.alloca !fir.box<!fir.array<?xf64>>
+! CHECK: fir.store %[[VAL_7]]#0 to %[[VAL_8]] : !fir.ref<!fir.box<!fir.array<?xf64>>>
+! CHECK: omp.yield(%[[VAL_8]] : !fir.ref<!fir.box<!fir.array<?xf64>>>)
+
+! CHECK-LABEL: } combiner {
+! CHECK: ^bb0(%[[VAL_0:.*]]: !fir.ref<!fir.box<!fir.array<?xf64>>>, %[[VAL_1:.*]]: !fir.ref<!fir.box<!fir.array<?xf64>>>):
+! CHECK: %[[VAL_2:.*]] = fir.load %[[VAL_0]] : !fir.ref<!fir.box<!fir.array<?xf64>>>
+! CHECK: %[[VAL_3:.*]] = fir.load %[[VAL_1]] : !fir.ref<!fir.box<!fir.array<?xf64>>>
+! CHECK: %[[VAL_4:.*]] = arith.constant 0 : index
+! CHECK: %[[VAL_5:.*]]:3 = fir.box_dims %[[VAL_2]], %[[VAL_4]] : (!fir.box<!fir.array<?xf64>>, 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.array<?xf64>>, !fir.shapeshift<1>, index) -> !fir.ref<f64>
+! CHECK: %[[VAL_10:.*]] = fir.array_coor %[[VAL_3]](%[[VAL_6]]) %[[VAL_8]] : (!fir.box<!fir.array<?xf64>>, !fir.shapeshift<1>, index) -> !fir.ref<f64>
+! CHECK: %[[VAL_11:.*]] = fir.load %[[VAL_9]] : !fir.ref<f64>
+! CHECK: %[[VAL_12:.*]] = fir.load %[[VAL_10]] : !fir.ref<f64>
+! CHECK: %[[VAL_13:.*]] = arith.addf %[[VAL_11]], %[[VAL_12]] fastmath<contract> : f64
+! CHECK: fir.store %[[VAL_13]] to %[[VAL_9]] : !fir.ref<f64>
+! CHECK: }
+! CHECK: omp.yield(%[[VAL_0]] : !fir.ref<!fir.box<!fir.array<?xf64>>>)
+! CHECK: }
+
+! CHECK-LABEL: func.func private @_QFPreduce(
+! CHECK-SAME: %[[VAL_0:.*]]: !fir.box<!fir.array<?xf64>> {fir.bindc_name = "r"}) attributes {{.*}} {
+! CHECK: %[[VAL_1:.*]] = fir.address_of(@_QFFreduceEi) : !fir.ref<i32>
+! CHECK: %[[VAL_2:.*]]:2 = hlfir.declare %[[VAL_1]] {uniq_name = "_QFFreduceEi"} : (!fir.ref<i32>) -> (!fir.ref<i32>, !fir.ref<i32>)
+! CHECK: %[[VAL_3:.*]]:2 = hlfir.declare %[[VAL_0]] {fortran_attrs = {{.*}}, uniq_name = "_QFFreduceEr"} : (!fir.box<!fir.array<?xf64>>) -> (!fir.box<!fir.array<?xf64>>, !fir.box<!fir.array<?xf64>>)
+! CHECK: omp.parallel {
+! CHECK: %[[VAL_4:.*]] = fir.alloca i32 {adapt.valuebyref, pinned}
+! CHECK: %[[VAL_5:.*]]:2 = hlfir.declare %[[VAL_4]] {uniq_name = "_QFFreduceEi"} : (!fir.ref<i32>) -> (!fir.ref<i32>, !fir.ref<i32>)
+! CHECK: %[[VAL_6:.*]] = arith.constant 0 : i32
+! CHECK: %[[VAL_7:.*]] = arith.constant 10 : i32
+! CHECK: %[[VAL_8:.*]] = arith.constant 1 : i32
+! CHECK: %[[VAL_9:.*]] = fir.alloca !fir.box<!fir.array<?xf64>>
+! CHECK: fir.store %[[VAL_3]]#1 to %[[VAL_9]] : !fir.ref<!fir.box<!fir.array<?xf64>>>
+! CHECK: omp.wsloop byref reduction(@add_reduction_byref_box_Uxf64 %[[VAL_9]] -> %[[VAL_10:.*]] : !fir.ref<!fir.box<!fir.array<?xf64>>>) for (%[[VAL_11:.*]]) : i32 = (%[[VAL_6]]) to (%[[VAL_7]]) inclusive step (%[[VAL_8]]) {
+! CHECK: fir.store %[[VAL_11]] to %[[VAL_5]]#1 : !fir.ref<i32>
+! CHECK: %[[VAL_12:.*]]:2 = hlfir.declare %[[VAL_10]] {fortran_attrs = {{.*}}, uniq_name = "_QFFreduceEr"} : (!fir.ref<!fir.box<!fir.array<?xf64>>>) -> (!fir.ref<!fir.box<!fir.array<?xf64>>>, !fir.ref<!fir.box<!fir.array<?xf64>>>)
+! CHECK: %[[VAL_13:.*]] = fir.load %[[VAL_5]]#0 : !fir.ref<i32>
+! CHECK: %[[VAL_14:.*]] = fir.convert %[[VAL_13]] : (i32) -> f64
+! CHECK: %[[VAL_15:.*]] = fir.load %[[VAL_12]]#0 : !fir.ref<!fir.box<!fir.array<?xf64>>>
+! CHECK: %[[VAL_16:.*]] = arith.constant 1 : index
+! CHECK: %[[VAL_17:.*]] = hlfir.designate %[[VAL_15]] (%[[VAL_16]]) : (!fir.box<!fir.array<?xf64>>, index) -> !fir.ref<f64>
+! CHECK: hlfir.assign %[[VAL_14]] to %[[VAL_17]] : f64, !fir.ref<f64>
+! CHECK: %[[VAL_18:.*]] = arith.constant 1.000000e+00 : f64
+! CHECK: %[[VAL_19:.*]] = fir.load %[[VAL_12]]#0 : !fir.ref<!fir.box<!fir.array<?xf64>>>
+! CHECK: %[[VAL_20:.*]] = arith.constant 2 : index
+! CHECK: %[[VAL_21:.*]] = hlfir.designate %[[VAL_19]] (%[[VAL_20]]) : (!fir.box<!fir.array<?xf64>>, index) -> !fir.ref<f64>
+! CHECK: hlfir.assign %[[VAL_18]] to %[[VAL_21]] : f64, !fir.ref<f64>
+! CHECK: omp.yield
+! CHECK: }
+! CHECK: omp.terminator
+! CHECK: }
+! CHECK: return
+! CHECK: }
More information about the flang-commits
mailing list