1 //===-- CodeGenTBAA.cpp - TBAA information for LLVM CodeGen ---------------===//
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 // This is the code that manages TBAA information and defines the TBAA policy
10 // for the optimizer to use. Relevant standards text includes:
13 // C++ [basic.lval] (p10 in n3126, p15 in some earlier versions)
15 //===----------------------------------------------------------------------===//
17 #include "CodeGenTBAA.h"
18 #include "clang/AST/ASTContext.h"
19 #include "clang/AST/Attr.h"
20 #include "clang/AST/Mangle.h"
21 #include "clang/AST/RecordLayout.h"
22 #include "clang/Basic/CodeGenOptions.h"
23 #include "llvm/ADT/SmallSet.h"
24 #include "llvm/IR/Constants.h"
25 #include "llvm/IR/LLVMContext.h"
26 #include "llvm/IR/Metadata.h"
27 #include "llvm/IR/Module.h"
28 #include "llvm/IR/Type.h"
29 using namespace clang;
30 using namespace CodeGen;
32 CodeGenTBAA::CodeGenTBAA(ASTContext &Ctx, llvm::Module &M,
33 const CodeGenOptions &CGO,
34 const LangOptions &Features, MangleContext &MContext)
35 : Context(Ctx), Module(M), CodeGenOpts(CGO),
36 Features(Features), MContext(MContext), MDHelper(M.getContext()),
37 Root(nullptr), Char(nullptr)
40 CodeGenTBAA::~CodeGenTBAA() {
43 llvm::MDNode *CodeGenTBAA::getRoot() {
44 // Define the root of the tree. This identifies the tree, so that
45 // if our LLVM IR is linked with LLVM IR from a different front-end
46 // (or a different version of this front-end), their TBAA trees will
47 // remain distinct, and the optimizer will treat them conservatively.
49 if (Features.CPlusPlus)
50 Root = MDHelper.createTBAARoot("Simple C++ TBAA");
52 Root = MDHelper.createTBAARoot("Simple C/C++ TBAA");
58 llvm::MDNode *CodeGenTBAA::createScalarTypeNode(StringRef Name,
61 if (CodeGenOpts.NewStructPathTBAA) {
62 llvm::Metadata *Id = MDHelper.createString(Name);
63 return MDHelper.createTBAATypeNode(Parent, Size, Id);
65 return MDHelper.createTBAAScalarTypeNode(Name, Parent);
68 llvm::MDNode *CodeGenTBAA::getChar() {
69 // Define the root of the tree for user-accessible memory. C and C++
70 // give special powers to char and certain similar types. However,
71 // these special powers only cover user-accessible memory, and doesn't
72 // include things like vtables.
74 Char = createScalarTypeNode("omnipotent char", getRoot(), /* Size= */ 1);
79 static bool TypeHasMayAlias(QualType QTy) {
80 // Tagged types have declarations, and therefore may have attributes.
81 if (const TagType *TTy = dyn_cast<TagType>(QTy))
82 return TTy->getDecl()->hasAttr<MayAliasAttr>();
84 // Typedef types have declarations, and therefore may have attributes.
85 if (const TypedefType *TTy = dyn_cast<TypedefType>(QTy)) {
86 if (TTy->getDecl()->hasAttr<MayAliasAttr>())
88 // Also, their underlying types may have relevant attributes.
89 return TypeHasMayAlias(TTy->desugar());
95 /// Check if the given type is a valid base type to be used in access tags.
96 static bool isValidBaseType(QualType QTy) {
97 if (QTy->isReferenceType())
99 if (const RecordType *TTy = QTy->getAs<RecordType>()) {
100 const RecordDecl *RD = TTy->getDecl()->getDefinition();
101 // Incomplete types are not valid base access types.
104 if (RD->hasFlexibleArrayMember())
106 // RD can be struct, union, class, interface or enum.
107 // For now, we only handle struct and class.
108 if (RD->isStruct() || RD->isClass())
114 llvm::MDNode *CodeGenTBAA::getTypeInfoHelper(const Type *Ty) {
115 uint64_t Size = Context.getTypeSizeInChars(Ty).getQuantity();
117 // Handle builtin types.
118 if (const BuiltinType *BTy = dyn_cast<BuiltinType>(Ty)) {
119 switch (BTy->getKind()) {
120 // Character types are special and can alias anything.
121 // In C++, this technically only includes "char" and "unsigned char",
122 // and not "signed char". In C, it includes all three. For now,
123 // the risk of exploiting this detail in C++ seems likely to outweigh
125 case BuiltinType::Char_U:
126 case BuiltinType::Char_S:
127 case BuiltinType::UChar:
128 case BuiltinType::SChar:
131 // Unsigned types can alias their corresponding signed types.
132 case BuiltinType::UShort:
133 return getTypeInfo(Context.ShortTy);
134 case BuiltinType::UInt:
135 return getTypeInfo(Context.IntTy);
136 case BuiltinType::ULong:
137 return getTypeInfo(Context.LongTy);
138 case BuiltinType::ULongLong:
139 return getTypeInfo(Context.LongLongTy);
140 case BuiltinType::UInt128:
141 return getTypeInfo(Context.Int128Ty);
143 // Treat all other builtin types as distinct types. This includes
144 // treating wchar_t, char16_t, and char32_t as distinct from their
145 // "underlying types".
147 return createScalarTypeNode(BTy->getName(Features), getChar(), Size);
151 // C++1z [basic.lval]p10: "If a program attempts to access the stored value of
152 // an object through a glvalue of other than one of the following types the
153 // behavior is undefined: [...] a char, unsigned char, or std::byte type."
154 if (Ty->isStdByteType())
157 // Handle pointers and references.
158 // TODO: Implement C++'s type "similarity" and consider dis-"similar"
159 // pointers distinct.
160 if (Ty->isPointerType() || Ty->isReferenceType())
161 return createScalarTypeNode("any pointer", getChar(), Size);
163 // Accesses to arrays are accesses to objects of their element types.
164 if (CodeGenOpts.NewStructPathTBAA && Ty->isArrayType())
165 return getTypeInfo(cast<ArrayType>(Ty)->getElementType());
167 // Enum types are distinct types. In C++ they have "underlying types",
168 // however they aren't related for TBAA.
169 if (const EnumType *ETy = dyn_cast<EnumType>(Ty)) {
170 // In C++ mode, types have linkage, so we can rely on the ODR and
171 // on their mangled names, if they're external.
172 // TODO: Is there a way to get a program-wide unique name for a
173 // decl with local linkage or no linkage?
174 if (!Features.CPlusPlus || !ETy->getDecl()->isExternallyVisible())
177 SmallString<256> OutName;
178 llvm::raw_svector_ostream Out(OutName);
179 MContext.mangleTypeName(QualType(ETy, 0), Out);
180 return createScalarTypeNode(OutName, getChar(), Size);
183 // For now, handle any other kind of type conservatively.
187 llvm::MDNode *CodeGenTBAA::getTypeInfo(QualType QTy) {
188 // At -O0 or relaxed aliasing, TBAA is not emitted for regular types.
189 if (CodeGenOpts.OptimizationLevel == 0 || CodeGenOpts.RelaxedAliasing)
192 // If the type has the may_alias attribute (even on a typedef), it is
193 // effectively in the general char alias class.
194 if (TypeHasMayAlias(QTy))
197 // We need this function to not fall back to returning the "omnipotent char"
198 // type node for aggregate and union types. Otherwise, any dereference of an
199 // aggregate will result into the may-alias access descriptor, meaning all
200 // subsequent accesses to direct and indirect members of that aggregate will
201 // be considered may-alias too.
202 // TODO: Combine getTypeInfo() and getBaseTypeInfo() into a single function.
203 if (isValidBaseType(QTy))
204 return getBaseTypeInfo(QTy);
206 const Type *Ty = Context.getCanonicalType(QTy).getTypePtr();
207 if (llvm::MDNode *N = MetadataCache[Ty])
210 // Note that the following helper call is allowed to add new nodes to the
211 // cache, which invalidates all its previously obtained iterators. So we
212 // first generate the node for the type and then add that node to the cache.
213 llvm::MDNode *TypeNode = getTypeInfoHelper(Ty);
214 return MetadataCache[Ty] = TypeNode;
217 TBAAAccessInfo CodeGenTBAA::getAccessInfo(QualType AccessType) {
218 // Pointee values may have incomplete types, but they shall never be
220 if (AccessType->isIncompleteType())
221 return TBAAAccessInfo::getIncompleteInfo();
223 if (TypeHasMayAlias(AccessType))
224 return TBAAAccessInfo::getMayAliasInfo();
226 uint64_t Size = Context.getTypeSizeInChars(AccessType).getQuantity();
227 return TBAAAccessInfo(getTypeInfo(AccessType), Size);
230 TBAAAccessInfo CodeGenTBAA::getVTablePtrAccessInfo(llvm::Type *VTablePtrType) {
231 llvm::DataLayout DL(&Module);
232 unsigned Size = DL.getPointerTypeSize(VTablePtrType);
233 return TBAAAccessInfo(createScalarTypeNode("vtable pointer", getRoot(), Size),
238 CodeGenTBAA::CollectFields(uint64_t BaseOffset,
240 SmallVectorImpl<llvm::MDBuilder::TBAAStructField> &
243 /* Things not handled yet include: C++ base classes, bitfields, */
245 if (const RecordType *TTy = QTy->getAs<RecordType>()) {
246 const RecordDecl *RD = TTy->getDecl()->getDefinition();
247 if (RD->hasFlexibleArrayMember())
250 // TODO: Handle C++ base classes.
251 if (const CXXRecordDecl *Decl = dyn_cast<CXXRecordDecl>(RD))
252 if (Decl->bases_begin() != Decl->bases_end())
255 const ASTRecordLayout &Layout = Context.getASTRecordLayout(RD);
258 for (RecordDecl::field_iterator i = RD->field_begin(),
259 e = RD->field_end(); i != e; ++i, ++idx) {
260 if ((*i)->isZeroSize(Context) || (*i)->isUnnamedBitfield())
262 uint64_t Offset = BaseOffset +
263 Layout.getFieldOffset(idx) / Context.getCharWidth();
264 QualType FieldQTy = i->getType();
265 if (!CollectFields(Offset, FieldQTy, Fields,
266 MayAlias || TypeHasMayAlias(FieldQTy)))
272 /* Otherwise, treat whatever it is as a field. */
273 uint64_t Offset = BaseOffset;
274 uint64_t Size = Context.getTypeSizeInChars(QTy).getQuantity();
275 llvm::MDNode *TBAAType = MayAlias ? getChar() : getTypeInfo(QTy);
276 llvm::MDNode *TBAATag = getAccessTagInfo(TBAAAccessInfo(TBAAType, Size));
277 Fields.push_back(llvm::MDBuilder::TBAAStructField(Offset, Size, TBAATag));
282 CodeGenTBAA::getTBAAStructInfo(QualType QTy) {
283 const Type *Ty = Context.getCanonicalType(QTy).getTypePtr();
285 if (llvm::MDNode *N = StructMetadataCache[Ty])
288 SmallVector<llvm::MDBuilder::TBAAStructField, 4> Fields;
289 if (CollectFields(0, QTy, Fields, TypeHasMayAlias(QTy)))
290 return MDHelper.createTBAAStructNode(Fields);
292 // For now, handle any other kind of type conservatively.
293 return StructMetadataCache[Ty] = nullptr;
296 llvm::MDNode *CodeGenTBAA::getBaseTypeInfoHelper(const Type *Ty) {
297 if (auto *TTy = dyn_cast<RecordType>(Ty)) {
298 const RecordDecl *RD = TTy->getDecl()->getDefinition();
299 const ASTRecordLayout &Layout = Context.getASTRecordLayout(RD);
300 SmallVector<llvm::MDBuilder::TBAAStructField, 4> Fields;
301 for (FieldDecl *Field : RD->fields()) {
302 if (Field->isZeroSize(Context) || Field->isUnnamedBitfield())
304 QualType FieldQTy = Field->getType();
305 llvm::MDNode *TypeNode = isValidBaseType(FieldQTy) ?
306 getBaseTypeInfo(FieldQTy) : getTypeInfo(FieldQTy);
308 return BaseTypeMetadataCache[Ty] = nullptr;
310 uint64_t BitOffset = Layout.getFieldOffset(Field->getFieldIndex());
311 uint64_t Offset = Context.toCharUnitsFromBits(BitOffset).getQuantity();
312 uint64_t Size = Context.getTypeSizeInChars(FieldQTy).getQuantity();
313 Fields.push_back(llvm::MDBuilder::TBAAStructField(Offset, Size,
317 SmallString<256> OutName;
318 if (Features.CPlusPlus) {
319 // Don't use the mangler for C code.
320 llvm::raw_svector_ostream Out(OutName);
321 MContext.mangleTypeName(QualType(Ty, 0), Out);
323 OutName = RD->getName();
326 if (CodeGenOpts.NewStructPathTBAA) {
327 llvm::MDNode *Parent = getChar();
328 uint64_t Size = Context.getTypeSizeInChars(Ty).getQuantity();
329 llvm::Metadata *Id = MDHelper.createString(OutName);
330 return MDHelper.createTBAATypeNode(Parent, Size, Id, Fields);
333 // Create the struct type node with a vector of pairs (offset, type).
334 SmallVector<std::pair<llvm::MDNode*, uint64_t>, 4> OffsetsAndTypes;
335 for (const auto &Field : Fields)
336 OffsetsAndTypes.push_back(std::make_pair(Field.Type, Field.Offset));
337 return MDHelper.createTBAAStructTypeNode(OutName, OffsetsAndTypes);
343 llvm::MDNode *CodeGenTBAA::getBaseTypeInfo(QualType QTy) {
344 if (!isValidBaseType(QTy))
347 const Type *Ty = Context.getCanonicalType(QTy).getTypePtr();
348 if (llvm::MDNode *N = BaseTypeMetadataCache[Ty])
351 // Note that the following helper call is allowed to add new nodes to the
352 // cache, which invalidates all its previously obtained iterators. So we
353 // first generate the node for the type and then add that node to the cache.
354 llvm::MDNode *TypeNode = getBaseTypeInfoHelper(Ty);
355 return BaseTypeMetadataCache[Ty] = TypeNode;
358 llvm::MDNode *CodeGenTBAA::getAccessTagInfo(TBAAAccessInfo Info) {
359 assert(!Info.isIncomplete() && "Access to an object of an incomplete type!");
361 if (Info.isMayAlias())
362 Info = TBAAAccessInfo(getChar(), Info.Size);
364 if (!Info.AccessType)
367 if (!CodeGenOpts.StructPathTBAA)
368 Info = TBAAAccessInfo(Info.AccessType, Info.Size);
370 llvm::MDNode *&N = AccessTagMetadataCache[Info];
374 if (!Info.BaseType) {
375 Info.BaseType = Info.AccessType;
376 assert(!Info.Offset && "Nonzero offset for an access with no base type!");
378 if (CodeGenOpts.NewStructPathTBAA) {
379 return N = MDHelper.createTBAAAccessTag(Info.BaseType, Info.AccessType,
380 Info.Offset, Info.Size);
382 return N = MDHelper.createTBAAStructTagNode(Info.BaseType, Info.AccessType,
386 TBAAAccessInfo CodeGenTBAA::mergeTBAAInfoForCast(TBAAAccessInfo SourceInfo,
387 TBAAAccessInfo TargetInfo) {
388 if (SourceInfo.isMayAlias() || TargetInfo.isMayAlias())
389 return TBAAAccessInfo::getMayAliasInfo();
394 CodeGenTBAA::mergeTBAAInfoForConditionalOperator(TBAAAccessInfo InfoA,
395 TBAAAccessInfo InfoB) {
399 if (!InfoA || !InfoB)
400 return TBAAAccessInfo();
402 if (InfoA.isMayAlias() || InfoB.isMayAlias())
403 return TBAAAccessInfo::getMayAliasInfo();
405 // TODO: Implement the rest of the logic here. For example, two accesses
406 // with same final access types result in an access to an object of that final
407 // access type regardless of their base types.
408 return TBAAAccessInfo::getMayAliasInfo();
412 CodeGenTBAA::mergeTBAAInfoForMemoryTransfer(TBAAAccessInfo DestInfo,
413 TBAAAccessInfo SrcInfo) {
414 if (DestInfo == SrcInfo)
417 if (!DestInfo || !SrcInfo)
418 return TBAAAccessInfo();
420 if (DestInfo.isMayAlias() || SrcInfo.isMayAlias())
421 return TBAAAccessInfo::getMayAliasInfo();
423 // TODO: Implement the rest of the logic here. For example, two accesses
424 // with same final access types result in an access to an object of that final
425 // access type regardless of their base types.
426 return TBAAAccessInfo::getMayAliasInfo();