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/Mangle.h" 21 #include "llvm/LLVMContext.h" 22 #include "llvm/Metadata.h" 23 #include "llvm/Constants.h" 24 #include "llvm/Type.h" 25 using namespace clang; 26 using namespace CodeGen; 27 28 CodeGenTBAA::CodeGenTBAA(ASTContext &Ctx, llvm::LLVMContext& VMContext, 29 const LangOptions &Features, MangleContext &MContext) 30 : Context(Ctx), VMContext(VMContext), Features(Features), MContext(MContext), 31 MDHelper(VMContext), Root(0), Char(0) { 32 } 33 34 CodeGenTBAA::~CodeGenTBAA() { 35 } 36 37 llvm::MDNode *CodeGenTBAA::getRoot() { 38 // Define the root of the tree. This identifies the tree, so that 39 // if our LLVM IR is linked with LLVM IR from a different front-end 40 // (or a different version of this front-end), their TBAA trees will 41 // remain distinct, and the optimizer will treat them conservatively. 42 if (!Root) 43 Root = MDHelper.createTBAARoot("Simple C/C++ TBAA"); 44 45 return Root; 46 } 47 48 llvm::MDNode *CodeGenTBAA::getChar() { 49 // Define the root of the tree for user-accessible memory. C and C++ 50 // give special powers to char and certain similar types. However, 51 // these special powers only cover user-accessible memory, and doesn't 52 // include things like vtables. 53 if (!Char) 54 Char = MDHelper.createTBAANode("omnipotent char", getRoot()); 55 56 return Char; 57 } 58 59 static bool TypeHasMayAlias(QualType QTy) { 60 // Tagged types have declarations, and therefore may have attributes. 61 if (const TagType *TTy = dyn_cast<TagType>(QTy)) 62 return TTy->getDecl()->hasAttr<MayAliasAttr>(); 63 64 // Typedef types have declarations, and therefore may have attributes. 65 if (const TypedefType *TTy = dyn_cast<TypedefType>(QTy)) { 66 if (TTy->getDecl()->hasAttr<MayAliasAttr>()) 67 return true; 68 // Also, their underlying types may have relevant attributes. 69 return TypeHasMayAlias(TTy->desugar()); 70 } 71 72 return false; 73 } 74 75 llvm::MDNode * 76 CodeGenTBAA::getTBAAInfo(QualType QTy) { 77 // If the type has the may_alias attribute (even on a typedef), it is 78 // effectively in the general char alias class. 79 if (TypeHasMayAlias(QTy)) 80 return getChar(); 81 82 const Type *Ty = Context.getCanonicalType(QTy).getTypePtr(); 83 84 if (llvm::MDNode *N = MetadataCache[Ty]) 85 return N; 86 87 // Handle builtin types. 88 if (const BuiltinType *BTy = dyn_cast<BuiltinType>(Ty)) { 89 switch (BTy->getKind()) { 90 // Character types are special and can alias anything. 91 // In C++, this technically only includes "char" and "unsigned char", 92 // and not "signed char". In C, it includes all three. For now, 93 // the risk of exploiting this detail in C++ seems likely to outweigh 94 // the benefit. 95 case BuiltinType::Char_U: 96 case BuiltinType::Char_S: 97 case BuiltinType::UChar: 98 case BuiltinType::SChar: 99 return getChar(); 100 101 // Unsigned types can alias their corresponding signed types. 102 case BuiltinType::UShort: 103 return getTBAAInfo(Context.ShortTy); 104 case BuiltinType::UInt: 105 return getTBAAInfo(Context.IntTy); 106 case BuiltinType::ULong: 107 return getTBAAInfo(Context.LongTy); 108 case BuiltinType::ULongLong: 109 return getTBAAInfo(Context.LongLongTy); 110 case BuiltinType::UInt128: 111 return getTBAAInfo(Context.Int128Ty); 112 113 // Treat all other builtin types as distinct types. This includes 114 // treating wchar_t, char16_t, and char32_t as distinct from their 115 // "underlying types". 116 default: 117 return MetadataCache[Ty] = 118 MDHelper.createTBAANode(BTy->getName(Features), getChar()); 119 } 120 } 121 122 // Handle pointers. 123 // TODO: Implement C++'s type "similarity" and consider dis-"similar" 124 // pointers distinct. 125 if (Ty->isPointerType()) 126 return MetadataCache[Ty] = MDHelper.createTBAANode("any pointer", 127 getChar()); 128 129 // Enum types are distinct types. In C++ they have "underlying types", 130 // however they aren't related for TBAA. 131 if (const EnumType *ETy = dyn_cast<EnumType>(Ty)) { 132 // In C mode, two anonymous enums are compatible iff their members 133 // are the same -- see C99 6.2.7p1. For now, be conservative. We could 134 // theoretically implement this by combining information about all the 135 // members into a single identifying MDNode. 136 if (!Features.CPlusPlus && 137 ETy->getDecl()->getTypedefNameForAnonDecl()) 138 return MetadataCache[Ty] = getChar(); 139 140 // In C++ mode, types have linkage, so we can rely on the ODR and 141 // on their mangled names, if they're external. 142 // TODO: Is there a way to get a program-wide unique name for a 143 // decl with local linkage or no linkage? 144 if (Features.CPlusPlus && 145 ETy->getDecl()->getLinkage() != ExternalLinkage) 146 return MetadataCache[Ty] = getChar(); 147 148 // TODO: This is using the RTTI name. Is there a better way to get 149 // a unique string for a type? 150 SmallString<256> OutName; 151 llvm::raw_svector_ostream Out(OutName); 152 MContext.mangleCXXRTTIName(QualType(ETy, 0), Out); 153 Out.flush(); 154 return MetadataCache[Ty] = MDHelper.createTBAANode(OutName, getChar()); 155 } 156 157 // For now, handle any other kind of type conservatively. 158 return MetadataCache[Ty] = getChar(); 159 } 160 161 llvm::MDNode *CodeGenTBAA::getTBAAInfoForVTablePtr() { 162 return MDHelper.createTBAANode("vtable pointer", getRoot()); 163 } 164
