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