Skip to content

Commit 083808d

Browse files
committed
[flang][cuda] Support device component in a pointer or allocatable derived-type
1 parent fd5fc76 commit 083808d

File tree

2 files changed

+55
-3
lines changed

2 files changed

+55
-3
lines changed

flang/lib/Lower/ConvertVariable.cpp

Lines changed: 23 additions & 3 deletions
Original file line numberDiff line numberDiff line change
@@ -814,6 +814,10 @@ initializeDeviceComponentAllocator(Fortran::lower::AbstractConverter &converter,
814814
if (auto boxTy = mlir::dyn_cast<fir::BaseBoxType>(baseTy))
815815
baseTy = boxTy.getEleTy();
816816
baseTy = fir::unwrapRefType(baseTy);
817+
818+
if (mlir::isa<fir::SequenceType>(baseTy))
819+
TODO(loc, "array of derived-type with device component");
820+
817821
auto recTy =
818822
mlir::dyn_cast<fir::RecordType>(fir::unwrapSequenceType(baseTy));
819823
assert(recTy && "expected fir::RecordType");
@@ -824,7 +828,7 @@ initializeDeviceComponentAllocator(Fortran::lower::AbstractConverter &converter,
824828
if (Fortran::semantics::IsDeviceAllocatable(sym)) {
825829
unsigned fieldIdx = recTy.getFieldIndex(sym.name().ToString());
826830
mlir::Type fieldTy;
827-
std::vector<mlir::Value> coordinates;
831+
llvm::SmallVector<mlir::Value> coordinates;
828832

829833
if (fieldIdx != std::numeric_limits<unsigned>::max()) {
830834
// Field found in the base record type.
@@ -867,8 +871,24 @@ initializeDeviceComponentAllocator(Fortran::lower::AbstractConverter &converter,
867871
TODO(loc, "device resident component in complex derived-type "
868872
"hierarchy");
869873

870-
mlir::Value comp = builder.create<fir::CoordinateOp>(
871-
loc, builder.getRefType(fieldTy), fir::getBase(exv), coordinates);
874+
mlir::Value base = fir::getBase(exv);
875+
mlir::Value comp;
876+
if (mlir::isa<fir::BaseBoxType>(fir::unwrapRefType(base.getType()))) {
877+
mlir::Value box = builder.create<fir::LoadOp>(loc, base);
878+
mlir::Value addr = builder.create<fir::BoxAddrOp>(loc, box);
879+
llvm::SmallVector<mlir::Value> lenParams;
880+
assert(coordinates.size() == 1 && "expect one coordinate");
881+
auto field = mlir::dyn_cast<fir::FieldIndexOp>(
882+
coordinates[0].getDefiningOp());
883+
comp = builder.create<hlfir::DesignateOp>(
884+
loc, builder.getRefType(fieldTy), addr,
885+
/*component=*/field.getFieldName(),
886+
/*componentShape=*/mlir::Value{},
887+
hlfir::DesignateOp::Subscripts{});
888+
} else {
889+
comp = builder.create<fir::CoordinateOp>(
890+
loc, builder.getRefType(fieldTy), base, coordinates);
891+
}
872892
cuf::DataAttributeAttr dataAttr =
873893
Fortran::lower::translateSymbolCUFDataAttribute(
874894
builder.getContext(), sym);

flang/test/Lower/CUDA/cuda-set-allocator.cuf

Lines changed: 32 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -21,4 +21,36 @@ contains
2121
! CHECK: %[[Z:.*]] = fir.coordinate_of %[[DT]]#0, z : (!fir.ref<!fir.type<_QMm1Tty_device{x:!fir.box<!fir.heap<!fir.array<?xi32>>>,y:i32,z:!fir.box<!fir.heap<!fir.array<?xi32>>>}>>) -> !fir.ref<!fir.box<!fir.heap<!fir.array<?xi32>>>>
2222
! CHECK: cuf.set_allocator_idx %[[Z]] : !fir.ref<!fir.box<!fir.heap<!fir.array<?xi32>>>> {data_attr = #cuf.cuda<device>}
2323

24+
subroutine sub2()
25+
type(ty_device), pointer :: d1
26+
end subroutine
27+
28+
! CHECK-LABEL: func.func @_QMm1Psub2()
29+
! CHECK: %[[ALLOC:.*]] = cuf.alloc !fir.box<!fir.ptr<!fir.type<_QMm1Tty_device{x:!fir.box<!fir.heap<!fir.array<?xi32>>>,y:i32,z:!fir.box<!fir.heap<!fir.array<?xi32>>>}>>> {bindc_name = "d1", data_attr = #cuf.cuda<managed>, uniq_name = "_QMm1Fsub2Ed1"} -> !fir.ref<!fir.box<!fir.ptr<!fir.type<_QMm1Tty_device{x:!fir.box<!fir.heap<!fir.array<?xi32>>>,y:i32,z:!fir.box<!fir.heap<!fir.array<?xi32>>>}>>>>
30+
! CHECK: %[[DECL:.*]]:2 = hlfir.declare %[[ALLOC]] {data_attr = #cuf.cuda<managed>, fortran_attrs = #fir.var_attrs<pointer>, uniq_name = "_QMm1Fsub2Ed1"} : (!fir.ref<!fir.box<!fir.ptr<!fir.type<_QMm1Tty_device{x:!fir.box<!fir.heap<!fir.array<?xi32>>>,y:i32,z:!fir.box<!fir.heap<!fir.array<?xi32>>>}>>>>) -> (!fir.ref<!fir.box<!fir.ptr<!fir.type<_QMm1Tty_device{x:!fir.box<!fir.heap<!fir.array<?xi32>>>,y:i32,z:!fir.box<!fir.heap<!fir.array<?xi32>>>}>>>>, !fir.ref<!fir.box<!fir.ptr<!fir.type<_QMm1Tty_device{x:!fir.box<!fir.heap<!fir.array<?xi32>>>,y:i32,z:!fir.box<!fir.heap<!fir.array<?xi32>>>}>>>>)
31+
! CHECK: %[[LOAD1:.*]] = fir.load %[[DECL]]#0 : !fir.ref<!fir.box<!fir.ptr<!fir.type<_QMm1Tty_device{x:!fir.box<!fir.heap<!fir.array<?xi32>>>,y:i32,z:!fir.box<!fir.heap<!fir.array<?xi32>>>}>>>>
32+
! CHECK: %[[ADDR1:.*]] = fir.box_addr %[[LOAD1]] : (!fir.box<!fir.ptr<!fir.type<_QMm1Tty_device{x:!fir.box<!fir.heap<!fir.array<?xi32>>>,y:i32,z:!fir.box<!fir.heap<!fir.array<?xi32>>>}>>>) -> !fir.ptr<!fir.type<_QMm1Tty_device{x:!fir.box<!fir.heap<!fir.array<?xi32>>>,y:i32,z:!fir.box<!fir.heap<!fir.array<?xi32>>>}>>
33+
! CHECK: %[[DESIGNATE1:.*]] = hlfir.designate %[[ADDR1]]{"x"} : (!fir.ptr<!fir.type<_QMm1Tty_device{x:!fir.box<!fir.heap<!fir.array<?xi32>>>,y:i32,z:!fir.box<!fir.heap<!fir.array<?xi32>>>}>>) -> !fir.ref<!fir.box<!fir.heap<!fir.array<?xi32>>>>
34+
! CHECK: cuf.set_allocator_idx %[[DESIGNATE1]] : !fir.ref<!fir.box<!fir.heap<!fir.array<?xi32>>>> {data_attr = #cuf.cuda<device>}
35+
! CHECK: %[[LOAD2:.*]] = fir.load %[[DECL]]#0 : !fir.ref<!fir.box<!fir.ptr<!fir.type<_QMm1Tty_device{x:!fir.box<!fir.heap<!fir.array<?xi32>>>,y:i32,z:!fir.box<!fir.heap<!fir.array<?xi32>>>}>>>>
36+
! CHECK: %[[ADDR2:.*]] = fir.box_addr %[[LOAD2]] : (!fir.box<!fir.ptr<!fir.type<_QMm1Tty_device{x:!fir.box<!fir.heap<!fir.array<?xi32>>>,y:i32,z:!fir.box<!fir.heap<!fir.array<?xi32>>>}>>>) -> !fir.ptr<!fir.type<_QMm1Tty_device{x:!fir.box<!fir.heap<!fir.array<?xi32>>>,y:i32,z:!fir.box<!fir.heap<!fir.array<?xi32>>>}>>
37+
! CHECK: %[[DESIGNATE2:.*]] = hlfir.designate %[[ADDR2]]{"z"} : (!fir.ptr<!fir.type<_QMm1Tty_device{x:!fir.box<!fir.heap<!fir.array<?xi32>>>,y:i32,z:!fir.box<!fir.heap<!fir.array<?xi32>>>}>>) -> !fir.ref<!fir.box<!fir.heap<!fir.array<?xi32>>>>
38+
! CHECK: cuf.set_allocator_idx %[[DESIGNATE2]] : !fir.ref<!fir.box<!fir.heap<!fir.array<?xi32>>>> {data_attr = #cuf.cuda<device>}
39+
40+
subroutine sub3()
41+
type(ty_device), allocatable :: d1
42+
end subroutine
43+
44+
! CHECK-LABEL: func.func @_QMm1Psub3()
45+
! CHECK: %[[ALLOC:.*]] = cuf.alloc !fir.box<!fir.heap<!fir.type<_QMm1Tty_device{x:!fir.box<!fir.heap<!fir.array<?xi32>>>,y:i32,z:!fir.box<!fir.heap<!fir.array<?xi32>>>}>>> {bindc_name = "d1", data_attr = #cuf.cuda<managed>, uniq_name = "_QMm1Fsub3Ed1"} -> !fir.ref<!fir.box<!fir.heap<!fir.type<_QMm1Tty_device{x:!fir.box<!fir.heap<!fir.array<?xi32>>>,y:i32,z:!fir.box<!fir.heap<!fir.array<?xi32>>>}>>>>
46+
! CHECK: %[[DECL:.*]]:2 = hlfir.declare %[[ALLOC]] {data_attr = #cuf.cuda<managed>, fortran_attrs = #fir.var_attrs<allocatable>, uniq_name = "_QMm1Fsub3Ed1"} : (!fir.ref<!fir.box<!fir.heap<!fir.type<_QMm1Tty_device{x:!fir.box<!fir.heap<!fir.array<?xi32>>>,y:i32,z:!fir.box<!fir.heap<!fir.array<?xi32>>>}>>>>) -> (!fir.ref<!fir.box<!fir.heap<!fir.type<_QMm1Tty_device{x:!fir.box<!fir.heap<!fir.array<?xi32>>>,y:i32,z:!fir.box<!fir.heap<!fir.array<?xi32>>>}>>>>, !fir.ref<!fir.box<!fir.heap<!fir.type<_QMm1Tty_device{x:!fir.box<!fir.heap<!fir.array<?xi32>>>,y:i32,z:!fir.box<!fir.heap<!fir.array<?xi32>>>}>>>>)
47+
! CHECK: %[[LOAD1:.*]] = fir.load %[[DECL]]#0 : !fir.ref<!fir.box<!fir.heap<!fir.type<_QMm1Tty_device{x:!fir.box<!fir.heap<!fir.array<?xi32>>>,y:i32,z:!fir.box<!fir.heap<!fir.array<?xi32>>>}>>>>
48+
! CHECK: %[[ADDR1:.*]] = fir.box_addr %[[LOAD1]] : (!fir.box<!fir.heap<!fir.type<_QMm1Tty_device{x:!fir.box<!fir.heap<!fir.array<?xi32>>>,y:i32,z:!fir.box<!fir.heap<!fir.array<?xi32>>>}>>>) -> !fir.heap<!fir.type<_QMm1Tty_device{x:!fir.box<!fir.heap<!fir.array<?xi32>>>,y:i32,z:!fir.box<!fir.heap<!fir.array<?xi32>>>}>>
49+
! CHECK: %[[DESIGNATE1:.*]] = hlfir.designate %[[ADDR1]]{"x"} : (!fir.heap<!fir.type<_QMm1Tty_device{x:!fir.box<!fir.heap<!fir.array<?xi32>>>,y:i32,z:!fir.box<!fir.heap<!fir.array<?xi32>>>}>>) -> !fir.ref<!fir.box<!fir.heap<!fir.array<?xi32>>>>
50+
! CHECK: cuf.set_allocator_idx %[[DESIGNATE1]] : !fir.ref<!fir.box<!fir.heap<!fir.array<?xi32>>>> {data_attr = #cuf.cuda<device>}
51+
! CHECK: %[[LOAD2:.*]] = fir.load %[[DECL]]#0 : !fir.ref<!fir.box<!fir.heap<!fir.type<_QMm1Tty_device{x:!fir.box<!fir.heap<!fir.array<?xi32>>>,y:i32,z:!fir.box<!fir.heap<!fir.array<?xi32>>>}>>>>
52+
! CHECK: %[[ADDR2:.*]] = fir.box_addr %[[LOAD2]] : (!fir.box<!fir.heap<!fir.type<_QMm1Tty_device{x:!fir.box<!fir.heap<!fir.array<?xi32>>>,y:i32,z:!fir.box<!fir.heap<!fir.array<?xi32>>>}>>>) -> !fir.heap<!fir.type<_QMm1Tty_device{x:!fir.box<!fir.heap<!fir.array<?xi32>>>,y:i32,z:!fir.box<!fir.heap<!fir.array<?xi32>>>}>>
53+
! CHECK: %[[DESIGNATE2:.*]] = hlfir.designate %[[ADDR2]]{"z"} : (!fir.heap<!fir.type<_QMm1Tty_device{x:!fir.box<!fir.heap<!fir.array<?xi32>>>,y:i32,z:!fir.box<!fir.heap<!fir.array<?xi32>>>}>>) -> !fir.ref<!fir.box<!fir.heap<!fir.array<?xi32>>>>
54+
! CHECK: cuf.set_allocator_idx %[[DESIGNATE2]] : !fir.ref<!fir.box<!fir.heap<!fir.array<?xi32>>>> {data_attr = #cuf.cuda<device>}
55+
2456
end module

0 commit comments

Comments
 (0)