Skip to content

Commit 1e4e2b3

Browse files
clementvalclementval
authored
[flang][cuda] Import type descriptor in the gpu module when needed (#149157)
1 parent 7caa0c9 commit 1e4e2b3

File tree

2 files changed

+53
-12
lines changed

2 files changed

+53
-12
lines changed

‎flang/lib/Optimizer/Transforms/CUFDeviceGlobal.cpp

Lines changed: 30 additions & 12 deletions
Original file line numberDiff line numberDiff line change
@@ -53,21 +53,26 @@ static void processAddrOfOp(fir::AddrOfOp addrOfOp,
5353
}
5454
}
5555

56+
static void processTypeDescriptor(fir::RecordType recTy,
57+
mlir::SymbolTable &symbolTable,
58+
llvm::DenseSet<fir::GlobalOp> &candidates) {
59+
if (auto globalOp = symbolTable.lookup<fir::GlobalOp>(
60+
fir::NameUniquer::getTypeDescriptorName(recTy.getName()))) {
61+
if (!candidates.contains(globalOp)) {
62+
globalOp.walk([&](fir::AddrOfOp op) {
63+
processAddrOfOp(op, symbolTable, candidates,
64+
/*recurseInGlobal=*/true);
65+
});
66+
candidates.insert(globalOp);
67+
}
68+
}
69+
}
70+
5671
static void processEmboxOp(fir::EmboxOp emboxOp, mlir::SymbolTable &symbolTable,
5772
llvm::DenseSet<fir::GlobalOp> &candidates) {
5873
if (auto recTy = mlir::dyn_cast<fir::RecordType>(
59-
fir::unwrapRefType(emboxOp.getMemref().getType()))) {
60-
if (auto globalOp = symbolTable.lookup<fir::GlobalOp>(
61-
fir::NameUniquer::getTypeDescriptorName(recTy.getName()))) {
62-
if (!candidates.contains(globalOp)) {
63-
globalOp.walk([&](fir::AddrOfOp op) {
64-
processAddrOfOp(op, symbolTable, candidates,
65-
/*recurseInGlobal=*/true);
66-
});
67-
candidates.insert(globalOp);
68-
}
69-
}
70-
}
74+
fir::unwrapRefType(emboxOp.getMemref().getType())))
75+
processTypeDescriptor(recTy, symbolTable, candidates);
7176
}
7277

7378
static void
@@ -85,6 +90,17 @@ prepareImplicitDeviceGlobals(mlir::func::FuncOp funcOp,
8590
}
8691
}
8792

93+
static void
94+
processPotentialTypeDescriptor(mlir::Type candidateType,
95+
mlir::SymbolTable &symbolTable,
96+
llvm::DenseSet<fir::GlobalOp> &candidates) {
97+
if (auto boxTy = mlir::dyn_cast<fir::BaseBoxType>(candidateType))
98+
candidateType = boxTy.getEleTy();
99+
candidateType = fir::unwrapSequenceType(fir::unwrapRefType(candidateType));
100+
if (auto recTy = mlir::dyn_cast<fir::RecordType>(candidateType))
101+
processTypeDescriptor(recTy, symbolTable, candidates);
102+
}
103+
88104
class CUFDeviceGlobal : public fir::impl::CUFDeviceGlobalBase<CUFDeviceGlobal> {
89105
public:
90106
void runOnOperation() override {
@@ -115,6 +131,8 @@ class CUFDeviceGlobal : public fir::impl::CUFDeviceGlobalBase<CUFDeviceGlobal> {
115131
for (auto globalOp : mod.getOps<fir::GlobalOp>()) {
116132
if (cuf::isRegisteredDeviceGlobal(globalOp)) {
117133
candidates.insert(globalOp);
134+
processPotentialTypeDescriptor(globalOp.getType(), parentSymTable,
135+
candidates);
118136
} else if (globalOp.getConstant() &&
119137
mlir::isa<fir::SequenceType>(
120138
fir::unwrapRefType(globalOp.resultType()))) {

‎flang/test/Fir/CUDA/cuda-device-global.f90

Lines changed: 23 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -24,3 +24,26 @@ module attributes {fir.defaultkind = "a1c4d8i4l4r4", fir.kindmap = "", gpu.conta
2424
// CHECK: gpu.module @cuda_device_mod
2525
// CHECK-DAG: fir.global @_QMm2ECc
2626
// CHECK-DAG: fir.global @_QMm1ECb
27+
28+
// -----
29+
30+
module attributes {fir.defaultkind = "a1c4d8i4l4r4", fir.kindmap = "", gpu.container_module} {
31+
fir.global @_QMmEddarrays {data_attr = #cuf.cuda<managed>} : !fir.box<!fir.heap<!fir.array<?x!fir.type<_QMmTdevicearrays{phi_r:!fir.box<!fir.heap<!fir.array<?xf64>>>,phi_i:!fir.box<!fir.heap<!fir.array<?xf64>>>,phi0_r:!fir.box<!fir.heap<!fir.array<?xf64>>>,phi0_i:!fir.box<!fir.heap<!fir.array<?xf64>>>,buf_r:!fir.box<!fir.heap<!fir.array<?xf64>>>,buf_i:!fir.box<!fir.heap<!fir.array<?xf64>>>}>>>> {
32+
%c0 = arith.constant 0 : index
33+
%0 = fir.zero_bits !fir.heap<!fir.array<?x!fir.type<_QMmTdevicearrays{phi_r:!fir.box<!fir.heap<!fir.array<?xf64>>>,phi_i:!fir.box<!fir.heap<!fir.array<?xf64>>>,phi0_r:!fir.box<!fir.heap<!fir.array<?xf64>>>,phi0_i:!fir.box<!fir.heap<!fir.array<?xf64>>>,buf_r:!fir.box<!fir.heap<!fir.array<?xf64>>>,buf_i:!fir.box<!fir.heap<!fir.array<?xf64>>>}>>>
34+
%1 = fir.shape %c0 : (index) -> !fir.shape<1>
35+
%2 = fir.embox %0(%1) {allocator_idx = 3 : i32} : (!fir.heap<!fir.array<?x!fir.type<_QMmTdevicearrays{phi_r:!fir.box<!fir.heap<!fir.array<?xf64>>>,phi_i:!fir.box<!fir.heap<!fir.array<?xf64>>>,phi0_r:!fir.box<!fir.heap<!fir.array<?xf64>>>,phi0_i:!fir.box<!fir.heap<!fir.array<?xf64>>>,buf_r:!fir.box<!fir.heap<!fir.array<?xf64>>>,buf_i:!fir.box<!fir.heap<!fir.array<?xf64>>>}>>>, !fir.shape<1>) -> !fir.box<!fir.heap<!fir.array<?x!fir.type<_QMmTdevicearrays{phi_r:!fir.box<!fir.heap<!fir.array<?xf64>>>,phi_i:!fir.box<!fir.heap<!fir.array<?xf64>>>,phi0_r:!fir.box<!fir.heap<!fir.array<?xf64>>>,phi0_i:!fir.box<!fir.heap<!fir.array<?xf64>>>,buf_r:!fir.box<!fir.heap<!fir.array<?xf64>>>,buf_i:!fir.box<!fir.heap<!fir.array<?xf64>>>}>>>>
36+
fir.has_value %2 : !fir.box<!fir.heap<!fir.array<?x!fir.type<_QMmTdevicearrays{phi_r:!fir.box<!fir.heap<!fir.array<?xf64>>>,phi_i:!fir.box<!fir.heap<!fir.array<?xf64>>>,phi0_r:!fir.box<!fir.heap<!fir.array<?xf64>>>,phi0_i:!fir.box<!fir.heap<!fir.array<?xf64>>>,buf_r:!fir.box<!fir.heap<!fir.array<?xf64>>>,buf_i:!fir.box<!fir.heap<!fir.array<?xf64>>>}>>>>
37+
}
38+
fir.global linkonce_odr @_QMmE.dt.devicearrays constant target : !fir.type<_QM__fortran_type_infoTderivedtype{binding:!fir.box<!fir.ptr<!fir.array<?x!fir.type<_QM__fortran_type_infoTbinding{proc:!fir.type<_QM__fortran_builtinsT__builtin_c_funptr{__address:i64}>,name:!fir.box<!fir.ptr<!fir.char<1,?>>>}>>>>,name:!fir.box<!fir.ptr<!fir.char<1,?>>>,sizeinbytes:i64,uninstantiated:!fir.box<!fir.ptr<!fir.type<_QM__fortran_type_infoTderivedtype>>>,kindparameter:!fir.box<!fir.ptr<!fir.array<?xi64>>>,lenparameterkind:!fir.box<!fir.ptr<!fir.array<?xi8>>>,component:!fir.box<!fir.ptr<!fir.array<?x!fir.type<_QM__fortran_type_infoTcomponent{name:!fir.box<!fir.ptr<!fir.char<1,?>>>,genre:i8,category:i8,kind:i8,rank:i8,__padding0:!fir.array<4xi8>,offset:i64,characterlen:!fir.type<_QM__fortran_type_infoTvalue{genre:i8,__padding0:!fir.array<7xi8>,value:i64}>,derived:!fir.box<!fir.ptr<!fir.type<_QM__fortran_type_infoTderivedtype>>>,lenvalue:!fir.box<!fir.ptr<!fir.array<?x!fir.type<_QM__fortran_type_infoTvalue{genre:i8,__padding0:!fir.array<7xi8>,value:i64}>>>>,bounds:!fir.box<!fir.ptr<!fir.array<?x?x!fir.type<_QM__fortran_type_infoTvalue{genre:i8,__padding0:!fir.array<7xi8>,value:i64}>>>>,initialization:!fir.type<_QM__fortran_builtinsT__builtin_c_ptr{__address:i64}>}>>>>,procptr:!fir.box<!fir.ptr<!fir.array<?x!fir.type<_QM__fortran_type_infoTprocptrcomponent{name:!fir.box<!fir.ptr<!fir.char<1,?>>>,offset:i64,initialization:!fir.type<_QM__fortran_builtinsT__builtin_c_funptr{__address:i64}>}>>>>,special:!fir.box<!fir.ptr<!fir.array<?x!fir.type<_QM__fortran_type_infoTspecialbinding{which:i8,isargdescriptorset:i8,istypebound:i8,isargcontiguousset:i8,__padding0:!fir.array<4xi8>,proc:!fir.type<_QM__fortran_builtinsT__builtin_c_funptr{__address:i64}>}>>>>,specialbitset:i32,hasparent:i8,noinitializationneeded:i8,nodestructionneeded:i8,nofinalizationneeded:i8,nodefinedassignment:i8,__padding0:!fir.array<3xi8>}> {
39+
%0 = fir.undefined !fir.type<_QM__fortran_type_infoTderivedtype{binding:!fir.box<!fir.ptr<!fir.array<?x!fir.type<_QM__fortran_type_infoTbinding{proc:!fir.type<_QM__fortran_builtinsT__builtin_c_funptr{__address:i64}>,name:!fir.box<!fir.ptr<!fir.char<1,?>>>}>>>>,name:!fir.box<!fir.ptr<!fir.char<1,?>>>,sizeinbytes:i64,uninstantiated:!fir.box<!fir.ptr<!fir.type<_QM__fortran_type_infoTderivedtype>>>,kindparameter:!fir.box<!fir.ptr<!fir.array<?xi64>>>,lenparameterkind:!fir.box<!fir.ptr<!fir.array<?xi8>>>,component:!fir.box<!fir.ptr<!fir.array<?x!fir.type<_QM__fortran_type_infoTcomponent{name:!fir.box<!fir.ptr<!fir.char<1,?>>>,genre:i8,category:i8,kind:i8,rank:i8,__padding0:!fir.array<4xi8>,offset:i64,characterlen:!fir.type<_QM__fortran_type_infoTvalue{genre:i8,__padding0:!fir.array<7xi8>,value:i64}>,derived:!fir.box<!fir.ptr<!fir.type<_QM__fortran_type_infoTderivedtype>>>,lenvalue:!fir.box<!fir.ptr<!fir.array<?x!fir.type<_QM__fortran_type_infoTvalue{genre:i8,__padding0:!fir.array<7xi8>,value:i64}>>>>,bounds:!fir.box<!fir.ptr<!fir.array<?x?x!fir.type<_QM__fortran_type_infoTvalue{genre:i8,__padding0:!fir.array<7xi8>,value:i64}>>>>,initialization:!fir.type<_QM__fortran_builtinsT__builtin_c_ptr{__address:i64}>}>>>>,procptr:!fir.box<!fir.ptr<!fir.array<?x!fir.type<_QM__fortran_type_infoTprocptrcomponent{name:!fir.box<!fir.ptr<!fir.char<1,?>>>,offset:i64,initialization:!fir.type<_QM__fortran_builtinsT__builtin_c_funptr{__address:i64}>}>>>>,special:!fir.box<!fir.ptr<!fir.array<?x!fir.type<_QM__fortran_type_infoTspecialbinding{which:i8,isargdescriptorset:i8,istypebound:i8,isargcontiguousset:i8,__padding0:!fir.array<4xi8>,proc:!fir.type<_QM__fortran_builtinsT__builtin_c_funptr{__address:i64}>}>>>>,specialbitset:i32,hasparent:i8,noinitializationneeded:i8,nodestructionneeded:i8,nofinalizationneeded:i8,nodefinedassignment:i8,__padding0:!fir.array<3xi8>}>
40+
fir.has_value %0 : !fir.type<_QM__fortran_type_infoTderivedtype{binding:!fir.box<!fir.ptr<!fir.array<?x!fir.type<_QM__fortran_type_infoTbinding{proc:!fir.type<_QM__fortran_builtinsT__builtin_c_funptr{__address:i64}>,name:!fir.box<!fir.ptr<!fir.char<1,?>>>}>>>>,name:!fir.box<!fir.ptr<!fir.char<1,?>>>,sizeinbytes:i64,uninstantiated:!fir.box<!fir.ptr<!fir.type<_QM__fortran_type_infoTderivedtype>>>,kindparameter:!fir.box<!fir.ptr<!fir.array<?xi64>>>,lenparameterkind:!fir.box<!fir.ptr<!fir.array<?xi8>>>,component:!fir.box<!fir.ptr<!fir.array<?x!fir.type<_QM__fortran_type_infoTcomponent{name:!fir.box<!fir.ptr<!fir.char<1,?>>>,genre:i8,category:i8,kind:i8,rank:i8,__padding0:!fir.array<4xi8>,offset:i64,characterlen:!fir.type<_QM__fortran_type_infoTvalue{genre:i8,__padding0:!fir.array<7xi8>,value:i64}>,derived:!fir.box<!fir.ptr<!fir.type<_QM__fortran_type_infoTderivedtype>>>,lenvalue:!fir.box<!fir.ptr<!fir.array<?x!fir.type<_QM__fortran_type_infoTvalue{genre:i8,__padding0:!fir.array<7xi8>,value:i64}>>>>,bounds:!fir.box<!fir.ptr<!fir.array<?x?x!fir.type<_QM__fortran_type_infoTvalue{genre:i8,__padding0:!fir.array<7xi8>,value:i64}>>>>,initialization:!fir.type<_QM__fortran_builtinsT__builtin_c_ptr{__address:i64}>}>>>>,procptr:!fir.box<!fir.ptr<!fir.array<?x!fir.type<_QM__fortran_type_infoTprocptrcomponent{name:!fir.box<!fir.ptr<!fir.char<1,?>>>,offset:i64,initialization:!fir.type<_QM__fortran_builtinsT__builtin_c_funptr{__address:i64}>}>>>>,special:!fir.box<!fir.ptr<!fir.array<?x!fir.type<_QM__fortran_type_infoTspecialbinding{which:i8,isargdescriptorset:i8,istypebound:i8,isargcontiguousset:i8,__padding0:!fir.array<4xi8>,proc:!fir.type<_QM__fortran_builtinsT__builtin_c_funptr{__address:i64}>}>>>>,specialbitset:i32,hasparent:i8,noinitializationneeded:i8,nodestructionneeded:i8,nofinalizationneeded:i8,nodefinedassignment:i8,__padding0:!fir.array<3xi8>}>
41+
}
42+
}
43+
44+
45+
// CHECK-NAG: fir.global @_QMmEddarrays
46+
// CHECK-NAG: fir.global linkonce_odr @_QMmE.dt.devicearrays
47+
// CHECK: gpu.module @cuda_device_mod
48+
// CHECK-NAG: fir.global @_QMmEddarrays
49+
// CHECK-NAG: fir.global linkonce_odr @_QMmE.dt.devicearrays

0 commit comments

Comments
 (0)