1 //===--- CodeGenTypes.cpp - TBAA information for LLVM CodeGen -------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This is the code that manages TBAA information and defines the TBAA policy
11 // for the optimizer to use. Relevant standards text includes:
14 // C++ [basic.lval] (p10 in n3126, p15 in some earlier versions)
16 //===----------------------------------------------------------------------===//
18 #include "CodeGenTBAA.h"
19 #include "clang/AST/ASTContext.h"
20 #include "clang/AST/Attr.h"
21 #include "clang/AST/Mangle.h"
22 #include "clang/AST/RecordLayout.h"
23 #include "clang/Frontend/CodeGenOptions.h"
24 #include "llvm/ADT/SmallSet.h"
25 #include "llvm/IR/Constants.h"
26 #include "llvm/IR/LLVMContext.h"
27 #include "llvm/IR/Metadata.h"
28 #include "llvm/IR/Type.h"
29 using namespace clang;
30 using namespace CodeGen;
32 CodeGenTBAA::CodeGenTBAA(ASTContext &Ctx, llvm::LLVMContext& VMContext,
33 const CodeGenOptions &CGO,
34 const LangOptions &Features, MangleContext &MContext)
35 : Context(Ctx), CodeGenOpts(CGO), Features(Features), MContext(MContext),
36 MDHelper(VMContext), Root(nullptr), Char(nullptr) {
39 CodeGenTBAA::~CodeGenTBAA() {
42 llvm::MDNode *CodeGenTBAA::getRoot() {
43 // Define the root of the tree. This identifies the tree, so that
44 // if our LLVM IR is linked with LLVM IR from a different front-end
45 // (or a different version of this front-end), their TBAA trees will
46 // remain distinct, and the optimizer will treat them conservatively.
48 if (Features.CPlusPlus)
49 Root = MDHelper.createTBAARoot("Simple C++ TBAA");
51 Root = MDHelper.createTBAARoot("Simple C/C++ TBAA");
57 // For both scalar TBAA and struct-path aware TBAA, the scalar type has the
58 // same format: name, parent node, and offset.
59 llvm::MDNode *CodeGenTBAA::createTBAAScalarType(StringRef Name,
60 llvm::MDNode *Parent) {
61 return MDHelper.createTBAAScalarTypeNode(Name, Parent);
64 llvm::MDNode *CodeGenTBAA::getChar() {
65 // Define the root of the tree for user-accessible memory. C and C++
66 // give special powers to char and certain similar types. However,
67 // these special powers only cover user-accessible memory, and doesn't
68 // include things like vtables.
70 Char = createTBAAScalarType("omnipotent char", getRoot());
75 static bool TypeHasMayAlias(QualType QTy) {
76 // Tagged types have declarations, and therefore may have attributes.
77 if (const TagType *TTy = dyn_cast<TagType>(QTy))
78 return TTy->getDecl()->hasAttr<MayAliasAttr>();
80 // Typedef types have declarations, and therefore may have attributes.
81 if (const TypedefType *TTy = dyn_cast<TypedefType>(QTy)) {
82 if (TTy->getDecl()->hasAttr<MayAliasAttr>())
84 // Also, their underlying types may have relevant attributes.
85 return TypeHasMayAlias(TTy->desugar());
92 CodeGenTBAA::getTBAAInfo(QualType QTy) {
93 // At -O0 or relaxed aliasing, TBAA is not emitted for regular types.
94 if (CodeGenOpts.OptimizationLevel == 0 || CodeGenOpts.RelaxedAliasing)
97 // If the type has the may_alias attribute (even on a typedef), it is
98 // effectively in the general char alias class.
99 if (TypeHasMayAlias(QTy))
102 const Type *Ty = Context.getCanonicalType(QTy).getTypePtr();
104 if (llvm::MDNode *N = MetadataCache[Ty])
107 // Handle builtin types.
108 if (const BuiltinType *BTy = dyn_cast<BuiltinType>(Ty)) {
109 switch (BTy->getKind()) {
110 // Character types are special and can alias anything.
111 // In C++, this technically only includes "char" and "unsigned char",
112 // and not "signed char". In C, it includes all three. For now,
113 // the risk of exploiting this detail in C++ seems likely to outweigh
115 case BuiltinType::Char_U:
116 case BuiltinType::Char_S:
117 case BuiltinType::UChar:
118 case BuiltinType::SChar:
121 // Unsigned types can alias their corresponding signed types.
122 case BuiltinType::UShort:
123 return getTBAAInfo(Context.ShortTy);
124 case BuiltinType::UInt:
125 return getTBAAInfo(Context.IntTy);
126 case BuiltinType::ULong:
127 return getTBAAInfo(Context.LongTy);
128 case BuiltinType::ULongLong:
129 return getTBAAInfo(Context.LongLongTy);
130 case BuiltinType::UInt128:
131 return getTBAAInfo(Context.Int128Ty);
133 // Treat all other builtin types as distinct types. This includes
134 // treating wchar_t, char16_t, and char32_t as distinct from their
135 // "underlying types".
137 return MetadataCache[Ty] =
138 createTBAAScalarType(BTy->getName(Features), getChar());
142 // C++1z [basic.lval]p10: "If a program attempts to access the stored value of
143 // an object through a glvalue of other than one of the following types the
144 // behavior is undefined: [...] a char, unsigned char, or std::byte type."
145 if (Ty->isStdByteType())
146 return MetadataCache[Ty] = getChar();
149 // TODO: Implement C++'s type "similarity" and consider dis-"similar"
150 // pointers distinct.
151 if (Ty->isPointerType())
152 return MetadataCache[Ty] = createTBAAScalarType("any pointer",
155 // Enum types are distinct types. In C++ they have "underlying types",
156 // however they aren't related for TBAA.
157 if (const EnumType *ETy = dyn_cast<EnumType>(Ty)) {
158 // In C++ mode, types have linkage, so we can rely on the ODR and
159 // on their mangled names, if they're external.
160 // TODO: Is there a way to get a program-wide unique name for a
161 // decl with local linkage or no linkage?
162 if (!Features.CPlusPlus || !ETy->getDecl()->isExternallyVisible())
163 return MetadataCache[Ty] = getChar();
165 SmallString<256> OutName;
166 llvm::raw_svector_ostream Out(OutName);
167 MContext.mangleTypeName(QualType(ETy, 0), Out);
168 return MetadataCache[Ty] = createTBAAScalarType(OutName, getChar());
171 // For now, handle any other kind of type conservatively.
172 return MetadataCache[Ty] = getChar();
175 llvm::MDNode *CodeGenTBAA::getTBAAInfoForVTablePtr() {
176 return createTBAAScalarType("vtable pointer", getRoot());
180 CodeGenTBAA::CollectFields(uint64_t BaseOffset,
182 SmallVectorImpl<llvm::MDBuilder::TBAAStructField> &
185 /* Things not handled yet include: C++ base classes, bitfields, */
187 if (const RecordType *TTy = QTy->getAs<RecordType>()) {
188 const RecordDecl *RD = TTy->getDecl()->getDefinition();
189 if (RD->hasFlexibleArrayMember())
192 // TODO: Handle C++ base classes.
193 if (const CXXRecordDecl *Decl = dyn_cast<CXXRecordDecl>(RD))
194 if (Decl->bases_begin() != Decl->bases_end())
197 const ASTRecordLayout &Layout = Context.getASTRecordLayout(RD);
200 for (RecordDecl::field_iterator i = RD->field_begin(),
201 e = RD->field_end(); i != e; ++i, ++idx) {
202 uint64_t Offset = BaseOffset +
203 Layout.getFieldOffset(idx) / Context.getCharWidth();
204 QualType FieldQTy = i->getType();
205 if (!CollectFields(Offset, FieldQTy, Fields,
206 MayAlias || TypeHasMayAlias(FieldQTy)))
212 /* Otherwise, treat whatever it is as a field. */
213 uint64_t Offset = BaseOffset;
214 uint64_t Size = Context.getTypeSizeInChars(QTy).getQuantity();
215 llvm::MDNode *TBAAInfo = MayAlias ? getChar() : getTBAAInfo(QTy);
216 llvm::MDNode *TBAATag = getTBAAScalarTagInfo(TBAAInfo);
217 Fields.push_back(llvm::MDBuilder::TBAAStructField(Offset, Size, TBAATag));
222 CodeGenTBAA::getTBAAStructInfo(QualType QTy) {
223 const Type *Ty = Context.getCanonicalType(QTy).getTypePtr();
225 if (llvm::MDNode *N = StructMetadataCache[Ty])
228 SmallVector<llvm::MDBuilder::TBAAStructField, 4> Fields;
229 if (CollectFields(0, QTy, Fields, TypeHasMayAlias(QTy)))
230 return MDHelper.createTBAAStructNode(Fields);
232 // For now, handle any other kind of type conservatively.
233 return StructMetadataCache[Ty] = nullptr;
236 /// Check if the given type can be handled by path-aware TBAA.
237 static bool isTBAAPathStruct(QualType QTy) {
238 if (const RecordType *TTy = QTy->getAs<RecordType>()) {
239 const RecordDecl *RD = TTy->getDecl()->getDefinition();
240 if (RD->hasFlexibleArrayMember())
242 // RD can be struct, union, class, interface or enum.
243 // For now, we only handle struct and class.
244 if (RD->isStruct() || RD->isClass())
251 CodeGenTBAA::getTBAAStructTypeInfo(QualType QTy) {
252 const Type *Ty = Context.getCanonicalType(QTy).getTypePtr();
253 assert(isTBAAPathStruct(QTy));
255 if (llvm::MDNode *N = StructTypeMetadataCache[Ty])
258 if (const RecordType *TTy = QTy->getAs<RecordType>()) {
259 const RecordDecl *RD = TTy->getDecl()->getDefinition();
261 const ASTRecordLayout &Layout = Context.getASTRecordLayout(RD);
262 SmallVector <std::pair<llvm::MDNode*, uint64_t>, 4> Fields;
264 for (RecordDecl::field_iterator i = RD->field_begin(),
265 e = RD->field_end(); i != e; ++i, ++idx) {
266 QualType FieldQTy = i->getType();
267 llvm::MDNode *FieldNode;
268 if (isTBAAPathStruct(FieldQTy))
269 FieldNode = getTBAAStructTypeInfo(FieldQTy);
271 FieldNode = getTBAAInfo(FieldQTy);
273 return StructTypeMetadataCache[Ty] = nullptr;
274 Fields.push_back(std::make_pair(
275 FieldNode, Layout.getFieldOffset(idx) / Context.getCharWidth()));
278 SmallString<256> OutName;
279 if (Features.CPlusPlus) {
280 // Don't use the mangler for C code.
281 llvm::raw_svector_ostream Out(OutName);
282 MContext.mangleTypeName(QualType(Ty, 0), Out);
284 OutName = RD->getName();
286 // Create the struct type node with a vector of pairs (offset, type).
287 return StructTypeMetadataCache[Ty] =
288 MDHelper.createTBAAStructTypeNode(OutName, Fields);
291 return StructMetadataCache[Ty] = nullptr;
294 /// Return a TBAA tag node for both scalar TBAA and struct-path aware TBAA.
296 CodeGenTBAA::getTBAAStructTagInfo(QualType BaseQTy, llvm::MDNode *AccessNode,
301 if (!CodeGenOpts.StructPathTBAA)
302 return getTBAAScalarTagInfo(AccessNode);
304 const Type *BTy = Context.getCanonicalType(BaseQTy).getTypePtr();
305 TBAAPathTag PathTag = TBAAPathTag(BTy, AccessNode, Offset);
306 if (llvm::MDNode *N = StructTagMetadataCache[PathTag])
309 llvm::MDNode *BNode = nullptr;
310 if (isTBAAPathStruct(BaseQTy))
311 BNode = getTBAAStructTypeInfo(BaseQTy);
313 return StructTagMetadataCache[PathTag] =
314 MDHelper.createTBAAStructTagNode(AccessNode, AccessNode, 0);
316 return StructTagMetadataCache[PathTag] =
317 MDHelper.createTBAAStructTagNode(BNode, AccessNode, Offset);
321 CodeGenTBAA::getTBAAScalarTagInfo(llvm::MDNode *AccessNode) {
324 if (llvm::MDNode *N = ScalarTagMetadataCache[AccessNode])
327 return ScalarTagMetadataCache[AccessNode] =
328 MDHelper.createTBAAStructTagNode(AccessNode, AccessNode, 0);