1 //===-- TypeConverter.cpp -- type conversion --------------------*- C++ -*-===//
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
7 //===----------------------------------------------------------------------===//
9 // Coding style: https://mlir.llvm.org/getting_started/DeveloperGuide/
11 //===----------------------------------------------------------------------===//
13 #define DEBUG_TYPE "flang-type-conversion"
15 #include "flang/Optimizer/CodeGen/TypeConverter.h"
16 #include "DescriptorModel.h"
17 #include "flang/Optimizer/Builder/Todo.h" // remove when TODO's are done
18 #include "flang/Optimizer/CodeGen/TBAABuilder.h"
19 #include "flang/Optimizer/CodeGen/Target.h"
20 #include "flang/Optimizer/Dialect/FIRType.h"
21 #include "flang/Optimizer/Dialect/Support/FIRContext.h"
22 #include "flang/Optimizer/Dialect/Support/KindMapping.h"
23 #include "mlir/Conversion/LLVMCommon/TypeConverter.h"
24 #include "llvm/ADT/ScopeExit.h"
25 #include "llvm/Support/Debug.h"
29 LLVMTypeConverter::LLVMTypeConverter(mlir::ModuleOp module
, bool applyTBAA
,
30 bool forceUnifiedTBAATree
,
31 const mlir::DataLayout
&dl
)
32 : mlir::LLVMTypeConverter(module
.getContext()),
33 kindMapping(getKindMapping(module
)),
34 specifics(CodeGenSpecifics::get(module
.getContext(),
35 getTargetTriple(module
),
36 getKindMapping(module
), dl
)),
37 tbaaBuilder(std::make_unique
<TBAABuilder
>(module
->getContext(), applyTBAA
,
38 forceUnifiedTBAATree
)) {
39 LLVM_DEBUG(llvm::dbgs() << "FIR type converter\n");
41 // Each conversion should return a value of type mlir::Type.
42 addConversion([&](BoxType box
) { return convertBoxType(box
); });
43 addConversion([&](BoxCharType boxchar
) {
44 LLVM_DEBUG(llvm::dbgs() << "type convert: " << boxchar
<< '\n');
45 return convertType(specifics
->boxcharMemoryType(boxchar
.getEleTy()));
47 addConversion([&](BoxProcType boxproc
) {
48 // TODO: Support for this type will be added later when the Fortran 2003
49 // procedure pointer feature is implemented.
53 [&](fir::ClassType classTy
) { return convertBoxType(classTy
); });
55 [&](fir::CharacterType charTy
) { return convertCharType(charTy
); });
57 [&](fir::ComplexType cmplx
) { return convertComplexType(cmplx
); });
58 addConversion([&](fir::FieldType field
) {
59 // Convert to i32 because of LLVM GEP indexing restriction.
60 return mlir::IntegerType::get(field
.getContext(), 32);
62 addConversion([&](HeapType heap
) { return convertPointerLike(heap
); });
63 addConversion([&](fir::IntegerType intTy
) {
64 return mlir::IntegerType::get(
65 &getContext(), kindMapping
.getIntegerBitsize(intTy
.getFKind()));
67 addConversion([&](fir::LenType field
) {
68 // Get size of len paramter from the descriptor.
69 return getModel
<Fortran::runtime::typeInfo::TypeParameterValue
>()(
72 addConversion([&](fir::LogicalType boolTy
) {
73 return mlir::IntegerType::get(
74 &getContext(), kindMapping
.getLogicalBitsize(boolTy
.getFKind()));
76 addConversion([&](fir::LLVMPointerType pointer
) {
77 return convertPointerLike(pointer
);
80 [&](fir::PointerType pointer
) { return convertPointerLike(pointer
); });
82 [&](fir::RecordType derived
, llvm::SmallVectorImpl
<mlir::Type
> &results
) {
83 return convertRecordType(derived
, results
);
86 [&](fir::RealType real
) { return convertRealType(real
.getFKind()); });
88 [&](fir::ReferenceType ref
) { return convertPointerLike(ref
); });
89 addConversion([&](fir::SequenceType sequence
) {
90 return convertSequenceType(sequence
);
92 addConversion([&](fir::TypeDescType tdesc
) {
93 return convertTypeDescType(tdesc
.getContext());
95 addConversion([&](fir::VectorType vecTy
) {
96 return mlir::VectorType::get(llvm::ArrayRef
<int64_t>(vecTy
.getLen()),
97 convertType(vecTy
.getEleTy()));
99 addConversion([&](mlir::TupleType tuple
) {
100 LLVM_DEBUG(llvm::dbgs() << "type convert: " << tuple
<< '\n');
101 llvm::SmallVector
<mlir::Type
> members
;
102 for (auto mem
: tuple
.getTypes()) {
103 // Prevent fir.box from degenerating to a pointer to a descriptor in the
104 // context of a tuple type.
105 if (auto box
= mem
.dyn_cast
<fir::BaseBoxType
>())
106 members
.push_back(convertBoxTypeAsStruct(box
));
108 members
.push_back(convertType(mem
).cast
<mlir::Type
>());
110 return mlir::LLVM::LLVMStructType::getLiteral(&getContext(), members
,
113 addConversion([&](mlir::NoneType none
) {
114 return mlir::LLVM::LLVMStructType::getLiteral(
115 none
.getContext(), std::nullopt
, /*isPacked=*/false);
117 // FIXME: https://reviews.llvm.org/D82831 introduced an automatic
118 // materialization of conversion around function calls that is not working
119 // well with fir lowering to llvm (incorrect llvm.mlir.cast are inserted).
120 // Workaround until better analysis: register a handler that does not insert
122 addSourceMaterialization(
123 [&](mlir::OpBuilder
&builder
, mlir::Type resultType
,
124 mlir::ValueRange inputs
,
125 mlir::Location loc
) -> std::optional
<mlir::Value
> {
126 if (inputs
.size() != 1)
130 // Similar FIXME workaround here (needed for compare.fir/select-type.fir
131 // as well as rebox-global.fir tests). This is needed to cope with the
132 // the fact that codegen does not lower some operation results to the LLVM
133 // type produced by this LLVMTypeConverter. For instance, inside FIR
134 // globals, fir.box are lowered to llvm.struct, while the fir.box type
135 // conversion translates it into an llvm.ptr<llvm.struct<>> because
136 // descriptors are manipulated in memory outside of global initializers
137 // where this is not possible. Hence, MLIR inserts
138 // builtin.unrealized_conversion_cast after the translation of operations
139 // producing fir.box in fir.global codegen. addSourceMaterialization and
140 // addTargetMaterialization allow ignoring these ops and removing them
141 // after codegen assuming the type discrepencies are intended (like for
142 // fir.box inside globals).
143 addTargetMaterialization(
144 [&](mlir::OpBuilder
&builder
, mlir::Type resultType
,
145 mlir::ValueRange inputs
,
146 mlir::Location loc
) -> std::optional
<mlir::Value
> {
147 if (inputs
.size() != 1)
153 // i32 is used here because LLVM wants i32 constants when indexing into struct
154 // types. Indexing into other aggregate types is more flexible.
155 mlir::Type
LLVMTypeConverter::offsetType() const {
156 return mlir::IntegerType::get(&getContext(), 32);
159 // i64 can be used to index into aggregates like arrays
160 mlir::Type
LLVMTypeConverter::indexType() const {
161 return mlir::IntegerType::get(&getContext(), 64);
164 // fir.type<name(p : TY'...){f : TY...}> --> llvm<"%name = { ty... }">
165 std::optional
<mlir::LogicalResult
> LLVMTypeConverter::convertRecordType(
166 fir::RecordType derived
, llvm::SmallVectorImpl
<mlir::Type
> &results
) {
167 auto name
= derived
.getName();
168 auto st
= mlir::LLVM::LLVMStructType::getIdentified(&getContext(), name
);
170 auto &callStack
= getCurrentThreadRecursiveStack();
171 if (llvm::count(callStack
, derived
)) {
172 results
.push_back(st
);
173 return mlir::success();
175 callStack
.push_back(derived
);
176 auto popConversionCallStack
=
177 llvm::make_scope_exit([&callStack
]() { callStack
.pop_back(); });
179 llvm::SmallVector
<mlir::Type
> members
;
180 for (auto mem
: derived
.getTypeList()) {
181 // Prevent fir.box from degenerating to a pointer to a descriptor in the
182 // context of a record type.
183 if (auto box
= mem
.second
.dyn_cast
<fir::BaseBoxType
>())
184 members
.push_back(convertBoxTypeAsStruct(box
));
186 members
.push_back(convertType(mem
.second
).cast
<mlir::Type
>());
188 if (mlir::failed(st
.setBody(members
, /*isPacked=*/false)))
189 return mlir::failure();
190 results
.push_back(st
);
191 return mlir::success();
194 // Is an extended descriptor needed given the element type of a fir.box type ?
195 // Extended descriptors are required for derived types.
196 bool LLVMTypeConverter::requiresExtendedDesc(mlir::Type boxElementType
) const {
197 auto eleTy
= fir::unwrapSequenceType(boxElementType
);
198 return eleTy
.isa
<fir::RecordType
>();
201 // This corresponds to the descriptor as defined in ISO_Fortran_binding.h and
202 // the addendum defined in descriptor.h.
203 mlir::Type
LLVMTypeConverter::convertBoxTypeAsStruct(BaseBoxType box
,
205 // (base_addr*, elem_len, version, rank, type, attribute, f18Addendum, [dim]
206 llvm::SmallVector
<mlir::Type
> dataDescFields
;
207 mlir::Type ele
= box
.getEleTy();
208 // remove fir.heap/fir.ref/fir.ptr
209 if (auto removeIndirection
= fir::dyn_cast_ptrEleTy(ele
))
210 ele
= removeIndirection
;
211 auto eleTy
= convertType(ele
);
213 if (ele
.isa
<SequenceType
>() && eleTy
.isa
<mlir::LLVM::LLVMPointerType
>())
214 dataDescFields
.push_back(eleTy
);
216 dataDescFields
.push_back(
217 mlir::LLVM::LLVMPointerType::get(eleTy
.getContext()));
219 dataDescFields
.push_back(
220 getDescFieldTypeModel
<kElemLenPosInBox
>()(&getContext()));
222 dataDescFields
.push_back(
223 getDescFieldTypeModel
<kVersionPosInBox
>()(&getContext()));
225 dataDescFields
.push_back(
226 getDescFieldTypeModel
<kRankPosInBox
>()(&getContext()));
228 dataDescFields
.push_back(
229 getDescFieldTypeModel
<kTypePosInBox
>()(&getContext()));
231 dataDescFields
.push_back(
232 getDescFieldTypeModel
<kAttributePosInBox
>()(&getContext()));
234 dataDescFields
.push_back(
235 getDescFieldTypeModel
<kF18AddendumPosInBox
>()(&getContext()));
237 if (rank
== unknownRank()) {
238 if (auto seqTy
= ele
.dyn_cast
<SequenceType
>())
239 rank
= seqTy
.getDimension();
244 auto rowTy
= getDescFieldTypeModel
<kDimsPosInBox
>()(&getContext());
245 dataDescFields
.push_back(mlir::LLVM::LLVMArrayType::get(rowTy
, rank
));
247 // opt-type-ptr: i8* (see fir.tdesc)
248 if (requiresExtendedDesc(ele
) || fir::isUnlimitedPolymorphicType(box
)) {
249 dataDescFields
.push_back(
250 getExtendedDescFieldTypeModel
<kOptTypePtrPosInBox
>()(&getContext()));
252 getExtendedDescFieldTypeModel
<kOptRowTypePosInBox
>()(&getContext());
253 dataDescFields
.push_back(mlir::LLVM::LLVMArrayType::get(rowTy
, 1));
254 if (auto recTy
= fir::unwrapSequenceType(ele
).dyn_cast
<fir::RecordType
>())
255 if (recTy
.getNumLenParams() > 0) {
256 // The descriptor design needs to be clarified regarding the number of
257 // length parameters in the addendum. Since it can change for
258 // polymorphic allocatables, it seems all length parameters cannot
259 // always possibly be placed in the addendum.
260 TODO_NOLOC("extended descriptor derived with length parameters");
261 unsigned numLenParams
= recTy
.getNumLenParams();
262 dataDescFields
.push_back(
263 mlir::LLVM::LLVMArrayType::get(rowTy
, numLenParams
));
266 return mlir::LLVM::LLVMStructType::getLiteral(&getContext(), dataDescFields
,
270 /// Convert fir.box type to the corresponding llvm struct type instead of a
271 /// pointer to this struct type.
272 mlir::Type
LLVMTypeConverter::convertBoxType(BaseBoxType box
, int rank
) const {
273 // TODO: send the box type and the converted LLVM structure layout
274 // to tbaaBuilder for proper creation of TBAATypeDescriptorOp.
275 return mlir::LLVM::LLVMPointerType::get(box
.getContext());
278 // fir.boxproc<any> --> llvm<"{ any*, i8* }">
279 mlir::Type
LLVMTypeConverter::convertBoxProcType(BoxProcType boxproc
) const {
280 auto funcTy
= convertType(boxproc
.getEleTy());
281 auto voidPtrTy
= mlir::LLVM::LLVMPointerType::get(boxproc
.getContext());
282 llvm::SmallVector
<mlir::Type
, 2> tuple
= {funcTy
, voidPtrTy
};
283 return mlir::LLVM::LLVMStructType::getLiteral(boxproc
.getContext(), tuple
,
287 unsigned LLVMTypeConverter::characterBitsize(fir::CharacterType charTy
) const {
288 return kindMapping
.getCharacterBitsize(charTy
.getFKind());
291 // fir.char<k,?> --> llvm<"ix"> where ix is scaled by kind mapping
292 // fir.char<k,n> --> llvm.array<n x "ix">
293 mlir::Type
LLVMTypeConverter::convertCharType(fir::CharacterType charTy
) const {
294 auto iTy
= mlir::IntegerType::get(&getContext(), characterBitsize(charTy
));
295 if (charTy
.getLen() == fir::CharacterType::unknownLen())
297 return mlir::LLVM::LLVMArrayType::get(iTy
, charTy
.getLen());
300 // convert a front-end kind value to either a std or LLVM IR dialect type
301 // fir.real<n> --> llvm.anyfloat where anyfloat is a kind mapping
302 mlir::Type
LLVMTypeConverter::convertRealType(fir::KindTy kind
) const {
303 return fir::fromRealTypeID(&getContext(), kindMapping
.getRealTypeID(kind
),
307 // fir.array<c ... :any> --> llvm<"[...[c x any]]">
308 mlir::Type
LLVMTypeConverter::convertSequenceType(SequenceType seq
) const {
309 auto baseTy
= convertType(seq
.getEleTy());
310 if (characterWithDynamicLen(seq
.getEleTy()))
312 auto shape
= seq
.getShape();
313 auto constRows
= seq
.getConstantRows();
315 decltype(constRows
) i
= constRows
;
316 for (auto e
: shape
) {
317 baseTy
= mlir::LLVM::LLVMArrayType::get(baseTy
, e
);
321 if (!seq
.hasDynamicExtents())
327 // fir.tdesc<any> --> llvm<"i8*">
328 // TODO: For now use a void*, however pointer identity is not sufficient for
329 // the f18 object v. class distinction (F2003).
331 LLVMTypeConverter::convertTypeDescType(mlir::MLIRContext
*ctx
) const {
332 return mlir::LLVM::LLVMPointerType::get(ctx
);
335 // Relay TBAA tag attachment to TBAABuilder.
336 void LLVMTypeConverter::attachTBAATag(mlir::LLVM::AliasAnalysisOpInterface op
,
337 mlir::Type baseFIRType
,
338 mlir::Type accessFIRType
,
339 mlir::LLVM::GEPOp gep
) const {
340 tbaaBuilder
->attachTBAATag(op
, baseFIRType
, accessFIRType
, gep
);