1 //===--- CodeGenTypes.cpp - TBAA information for LLVM CodeGen -------------===// 2 // 3 // The LLVM Compiler Infrastructure 4 // 5 // This file is distributed under the University of Illinois Open Source 6 // License. See LICENSE.TXT for details. 7 // 8 //===----------------------------------------------------------------------===// 9 // 10 // This is the code that manages TBAA information and defines the TBAA policy 11 // for the optimizer to use. Relevant standards text includes: 12 // 13 // C99 6.5p7 14 // C++ [basic.lval] (p10 in n3126, p15 in some earlier versions) 15 // 16 //===----------------------------------------------------------------------===// 17 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; 31 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) { 37 } 38 39 CodeGenTBAA::~CodeGenTBAA() { 40 } 41 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. 47 if (!Root) { 48 if (Features.CPlusPlus) 49 Root = MDHelper.createTBAARoot("Simple C++ TBAA"); 50 else 51 Root = MDHelper.createTBAARoot("Simple C/C++ TBAA"); 52 } 53 54 return Root; 55 } 56 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); 62 } 63 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. 69 if (!Char) 70 Char = createTBAAScalarType("omnipotent char", getRoot()); 71 72 return Char; 73 } 74 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>(); 79 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>()) 83 return true; 84 // Also, their underlying types may have relevant attributes. 85 return TypeHasMayAlias(TTy->desugar()); 86 } 87 88 return false; 89 } 90 91 llvm::MDNode * 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) 95 return nullptr; 96 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)) 100 return getChar(); 101 102 const Type *Ty = Context.getCanonicalType(QTy).getTypePtr(); 103 104 if (llvm::MDNode *N = MetadataCache[Ty]) 105 return N; 106 107 // 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 114 // the benefit. 115 case BuiltinType::Char_U: 116 case BuiltinType::Char_S: 117 case BuiltinType::UChar: 118 case BuiltinType::SChar: 119 return getChar(); 120 121 // 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); 132 133 // 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". 136 default: 137 return MetadataCache[Ty] = 138 createTBAAScalarType(BTy->getName(Features), getChar()); 139 } 140 } 141 142 // Handle pointers. 143 // TODO: Implement C++'s type "similarity" and consider dis-"similar" 144 // pointers distinct. 145 if (Ty->isPointerType()) 146 return MetadataCache[Ty] = createTBAAScalarType("any pointer", 147 getChar()); 148 149 // Enum types are distinct types. In C++ they have "underlying types", 150 // however they aren't related for TBAA. 151 if (const EnumType *ETy = dyn_cast<EnumType>(Ty)) { 152 // In C++ mode, types have linkage, so we can rely on the ODR and 153 // on their mangled names, if they're external. 154 // TODO: Is there a way to get a program-wide unique name for a 155 // decl with local linkage or no linkage? 156 if (!Features.CPlusPlus || !ETy->getDecl()->isExternallyVisible()) 157 return MetadataCache[Ty] = getChar(); 158 159 SmallString<256> OutName; 160 llvm::raw_svector_ostream Out(OutName); 161 MContext.mangleTypeName(QualType(ETy, 0), Out); 162 return MetadataCache[Ty] = createTBAAScalarType(OutName, getChar()); 163 } 164 165 // For now, handle any other kind of type conservatively. 166 return MetadataCache[Ty] = getChar(); 167 } 168 169 llvm::MDNode *CodeGenTBAA::getTBAAInfoForVTablePtr() { 170 return createTBAAScalarType("vtable pointer", getRoot()); 171 } 172 173 bool 174 CodeGenTBAA::CollectFields(uint64_t BaseOffset, 175 QualType QTy, 176 SmallVectorImpl<llvm::MDBuilder::TBAAStructField> & 177 Fields, 178 bool MayAlias) { 179 /* Things not handled yet include: C++ base classes, bitfields, */ 180 181 if (const RecordType *TTy = QTy->getAs<RecordType>()) { 182 const RecordDecl *RD = TTy->getDecl()->getDefinition(); 183 if (RD->hasFlexibleArrayMember()) 184 return false; 185 186 // TODO: Handle C++ base classes. 187 if (const CXXRecordDecl *Decl = dyn_cast<CXXRecordDecl>(RD)) 188 if (Decl->bases_begin() != Decl->bases_end()) 189 return false; 190 191 const ASTRecordLayout &Layout = Context.getASTRecordLayout(RD); 192 193 unsigned idx = 0; 194 for (RecordDecl::field_iterator i = RD->field_begin(), 195 e = RD->field_end(); i != e; ++i, ++idx) { 196 uint64_t Offset = BaseOffset + 197 Layout.getFieldOffset(idx) / Context.getCharWidth(); 198 QualType FieldQTy = i->getType(); 199 if (!CollectFields(Offset, FieldQTy, Fields, 200 MayAlias || TypeHasMayAlias(FieldQTy))) 201 return false; 202 } 203 return true; 204 } 205 206 /* Otherwise, treat whatever it is as a field. */ 207 uint64_t Offset = BaseOffset; 208 uint64_t Size = Context.getTypeSizeInChars(QTy).getQuantity(); 209 llvm::MDNode *TBAAInfo = MayAlias ? getChar() : getTBAAInfo(QTy); 210 llvm::MDNode *TBAATag = getTBAAScalarTagInfo(TBAAInfo); 211 Fields.push_back(llvm::MDBuilder::TBAAStructField(Offset, Size, TBAATag)); 212 return true; 213 } 214 215 llvm::MDNode * 216 CodeGenTBAA::getTBAAStructInfo(QualType QTy) { 217 const Type *Ty = Context.getCanonicalType(QTy).getTypePtr(); 218 219 if (llvm::MDNode *N = StructMetadataCache[Ty]) 220 return N; 221 222 SmallVector<llvm::MDBuilder::TBAAStructField, 4> Fields; 223 if (CollectFields(0, QTy, Fields, TypeHasMayAlias(QTy))) 224 return MDHelper.createTBAAStructNode(Fields); 225 226 // For now, handle any other kind of type conservatively. 227 return StructMetadataCache[Ty] = nullptr; 228 } 229 230 /// Check if the given type can be handled by path-aware TBAA. 231 static bool isTBAAPathStruct(QualType QTy) { 232 if (const RecordType *TTy = QTy->getAs<RecordType>()) { 233 const RecordDecl *RD = TTy->getDecl()->getDefinition(); 234 if (RD->hasFlexibleArrayMember()) 235 return false; 236 // RD can be struct, union, class, interface or enum. 237 // For now, we only handle struct and class. 238 if (RD->isStruct() || RD->isClass()) 239 return true; 240 } 241 return false; 242 } 243 244 llvm::MDNode * 245 CodeGenTBAA::getTBAAStructTypeInfo(QualType QTy) { 246 const Type *Ty = Context.getCanonicalType(QTy).getTypePtr(); 247 assert(isTBAAPathStruct(QTy)); 248 249 if (llvm::MDNode *N = StructTypeMetadataCache[Ty]) 250 return N; 251 252 if (const RecordType *TTy = QTy->getAs<RecordType>()) { 253 const RecordDecl *RD = TTy->getDecl()->getDefinition(); 254 255 const ASTRecordLayout &Layout = Context.getASTRecordLayout(RD); 256 SmallVector <std::pair<llvm::MDNode*, uint64_t>, 4> Fields; 257 unsigned idx = 0; 258 for (RecordDecl::field_iterator i = RD->field_begin(), 259 e = RD->field_end(); i != e; ++i, ++idx) { 260 QualType FieldQTy = i->getType(); 261 llvm::MDNode *FieldNode; 262 if (isTBAAPathStruct(FieldQTy)) 263 FieldNode = getTBAAStructTypeInfo(FieldQTy); 264 else 265 FieldNode = getTBAAInfo(FieldQTy); 266 if (!FieldNode) 267 return StructTypeMetadataCache[Ty] = nullptr; 268 Fields.push_back(std::make_pair( 269 FieldNode, Layout.getFieldOffset(idx) / Context.getCharWidth())); 270 } 271 272 SmallString<256> OutName; 273 if (Features.CPlusPlus) { 274 // Don't use the mangler for C code. 275 llvm::raw_svector_ostream Out(OutName); 276 MContext.mangleTypeName(QualType(Ty, 0), Out); 277 } else { 278 OutName = RD->getName(); 279 } 280 // Create the struct type node with a vector of pairs (offset, type). 281 return StructTypeMetadataCache[Ty] = 282 MDHelper.createTBAAStructTypeNode(OutName, Fields); 283 } 284 285 return StructMetadataCache[Ty] = nullptr; 286 } 287 288 /// Return a TBAA tag node for both scalar TBAA and struct-path aware TBAA. 289 llvm::MDNode * 290 CodeGenTBAA::getTBAAStructTagInfo(QualType BaseQTy, llvm::MDNode *AccessNode, 291 uint64_t Offset) { 292 if (!AccessNode) 293 return nullptr; 294 295 if (!CodeGenOpts.StructPathTBAA) 296 return getTBAAScalarTagInfo(AccessNode); 297 298 const Type *BTy = Context.getCanonicalType(BaseQTy).getTypePtr(); 299 TBAAPathTag PathTag = TBAAPathTag(BTy, AccessNode, Offset); 300 if (llvm::MDNode *N = StructTagMetadataCache[PathTag]) 301 return N; 302 303 llvm::MDNode *BNode = nullptr; 304 if (isTBAAPathStruct(BaseQTy)) 305 BNode = getTBAAStructTypeInfo(BaseQTy); 306 if (!BNode) 307 return StructTagMetadataCache[PathTag] = 308 MDHelper.createTBAAStructTagNode(AccessNode, AccessNode, 0); 309 310 return StructTagMetadataCache[PathTag] = 311 MDHelper.createTBAAStructTagNode(BNode, AccessNode, Offset); 312 } 313 314 llvm::MDNode * 315 CodeGenTBAA::getTBAAScalarTagInfo(llvm::MDNode *AccessNode) { 316 if (!AccessNode) 317 return nullptr; 318 if (llvm::MDNode *N = ScalarTagMetadataCache[AccessNode]) 319 return N; 320 321 return ScalarTagMetadataCache[AccessNode] = 322 MDHelper.createTBAAStructTagNode(AccessNode, AccessNode, 0); 323 } 324