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"
29using namespace clang;
30using namespace CodeGen;
31
32CodeGenTBAA::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
39CodeGenTBAA::~CodeGenTBAA() {
40}
41
42llvm::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    Root = MDHelper.createTBAARoot("Simple C/C++ TBAA");
49
50  return Root;
51}
52
53// For both scalar TBAA and struct-path aware TBAA, the scalar type has the
54// same format: name, parent node, and offset.
55llvm::MDNode *CodeGenTBAA::createTBAAScalarType(StringRef Name,
56                                                llvm::MDNode *Parent) {
57  return MDHelper.createTBAAScalarTypeNode(Name, Parent);
58}
59
60llvm::MDNode *CodeGenTBAA::getChar() {
61  // Define the root of the tree for user-accessible memory. C and C++
62  // give special powers to char and certain similar types. However,
63  // these special powers only cover user-accessible memory, and doesn't
64  // include things like vtables.
65  if (!Char)
66    Char = createTBAAScalarType("omnipotent char", getRoot());
67
68  return Char;
69}
70
71static bool TypeHasMayAlias(QualType QTy) {
72  // Tagged types have declarations, and therefore may have attributes.
73  if (const TagType *TTy = dyn_cast<TagType>(QTy))
74    return TTy->getDecl()->hasAttr<MayAliasAttr>();
75
76  // Typedef types have declarations, and therefore may have attributes.
77  if (const TypedefType *TTy = dyn_cast<TypedefType>(QTy)) {
78    if (TTy->getDecl()->hasAttr<MayAliasAttr>())
79      return true;
80    // Also, their underlying types may have relevant attributes.
81    return TypeHasMayAlias(TTy->desugar());
82  }
83
84  return false;
85}
86
87llvm::MDNode *
88CodeGenTBAA::getTBAAInfo(QualType QTy) {
89  // At -O0 or relaxed aliasing, TBAA is not emitted for regular types.
90  if (CodeGenOpts.OptimizationLevel == 0 || CodeGenOpts.RelaxedAliasing)
91    return nullptr;
92
93  // If the type has the may_alias attribute (even on a typedef), it is
94  // effectively in the general char alias class.
95  if (TypeHasMayAlias(QTy))
96    return getChar();
97
98  const Type *Ty = Context.getCanonicalType(QTy).getTypePtr();
99
100  if (llvm::MDNode *N = MetadataCache[Ty])
101    return N;
102
103  // Handle builtin types.
104  if (const BuiltinType *BTy = dyn_cast<BuiltinType>(Ty)) {
105    switch (BTy->getKind()) {
106    // Character types are special and can alias anything.
107    // In C++, this technically only includes "char" and "unsigned char",
108    // and not "signed char". In C, it includes all three. For now,
109    // the risk of exploiting this detail in C++ seems likely to outweigh
110    // the benefit.
111    case BuiltinType::Char_U:
112    case BuiltinType::Char_S:
113    case BuiltinType::UChar:
114    case BuiltinType::SChar:
115      return getChar();
116
117    // Unsigned types can alias their corresponding signed types.
118    case BuiltinType::UShort:
119      return getTBAAInfo(Context.ShortTy);
120    case BuiltinType::UInt:
121      return getTBAAInfo(Context.IntTy);
122    case BuiltinType::ULong:
123      return getTBAAInfo(Context.LongTy);
124    case BuiltinType::ULongLong:
125      return getTBAAInfo(Context.LongLongTy);
126    case BuiltinType::UInt128:
127      return getTBAAInfo(Context.Int128Ty);
128
129    // Treat all other builtin types as distinct types. This includes
130    // treating wchar_t, char16_t, and char32_t as distinct from their
131    // "underlying types".
132    default:
133      return MetadataCache[Ty] =
134        createTBAAScalarType(BTy->getName(Features), getChar());
135    }
136  }
137
138  // Handle pointers.
139  // TODO: Implement C++'s type "similarity" and consider dis-"similar"
140  // pointers distinct.
141  if (Ty->isPointerType())
142    return MetadataCache[Ty] = createTBAAScalarType("any pointer",
143                                                    getChar());
144
145  // Enum types are distinct types. In C++ they have "underlying types",
146  // however they aren't related for TBAA.
147  if (const EnumType *ETy = dyn_cast<EnumType>(Ty)) {
148    // In C++ mode, types have linkage, so we can rely on the ODR and
149    // on their mangled names, if they're external.
150    // TODO: Is there a way to get a program-wide unique name for a
151    // decl with local linkage or no linkage?
152    if (!Features.CPlusPlus || !ETy->getDecl()->isExternallyVisible())
153      return MetadataCache[Ty] = getChar();
154
155    SmallString<256> OutName;
156    llvm::raw_svector_ostream Out(OutName);
157    MContext.mangleTypeName(QualType(ETy, 0), Out);
158    Out.flush();
159    return MetadataCache[Ty] = createTBAAScalarType(OutName, getChar());
160  }
161
162  // For now, handle any other kind of type conservatively.
163  return MetadataCache[Ty] = getChar();
164}
165
166llvm::MDNode *CodeGenTBAA::getTBAAInfoForVTablePtr() {
167  return createTBAAScalarType("vtable pointer", getRoot());
168}
169
170bool
171CodeGenTBAA::CollectFields(uint64_t BaseOffset,
172                           QualType QTy,
173                           SmallVectorImpl<llvm::MDBuilder::TBAAStructField> &
174                             Fields,
175                           bool MayAlias) {
176  /* Things not handled yet include: C++ base classes, bitfields, */
177
178  if (const RecordType *TTy = QTy->getAs<RecordType>()) {
179    const RecordDecl *RD = TTy->getDecl()->getDefinition();
180    if (RD->hasFlexibleArrayMember())
181      return false;
182
183    // TODO: Handle C++ base classes.
184    if (const CXXRecordDecl *Decl = dyn_cast<CXXRecordDecl>(RD))
185      if (Decl->bases_begin() != Decl->bases_end())
186        return false;
187
188    const ASTRecordLayout &Layout = Context.getASTRecordLayout(RD);
189
190    unsigned idx = 0;
191    for (RecordDecl::field_iterator i = RD->field_begin(),
192         e = RD->field_end(); i != e; ++i, ++idx) {
193      uint64_t Offset = BaseOffset +
194                        Layout.getFieldOffset(idx) / Context.getCharWidth();
195      QualType FieldQTy = i->getType();
196      if (!CollectFields(Offset, FieldQTy, Fields,
197                         MayAlias || TypeHasMayAlias(FieldQTy)))
198        return false;
199    }
200    return true;
201  }
202
203  /* Otherwise, treat whatever it is as a field. */
204  uint64_t Offset = BaseOffset;
205  uint64_t Size = Context.getTypeSizeInChars(QTy).getQuantity();
206  llvm::MDNode *TBAAInfo = MayAlias ? getChar() : getTBAAInfo(QTy);
207  llvm::MDNode *TBAATag = getTBAAScalarTagInfo(TBAAInfo);
208  Fields.push_back(llvm::MDBuilder::TBAAStructField(Offset, Size, TBAATag));
209  return true;
210}
211
212llvm::MDNode *
213CodeGenTBAA::getTBAAStructInfo(QualType QTy) {
214  const Type *Ty = Context.getCanonicalType(QTy).getTypePtr();
215
216  if (llvm::MDNode *N = StructMetadataCache[Ty])
217    return N;
218
219  SmallVector<llvm::MDBuilder::TBAAStructField, 4> Fields;
220  if (CollectFields(0, QTy, Fields, TypeHasMayAlias(QTy)))
221    return MDHelper.createTBAAStructNode(Fields);
222
223  // For now, handle any other kind of type conservatively.
224  return StructMetadataCache[Ty] = nullptr;
225}
226
227/// Check if the given type can be handled by path-aware TBAA.
228static bool isTBAAPathStruct(QualType QTy) {
229  if (const RecordType *TTy = QTy->getAs<RecordType>()) {
230    const RecordDecl *RD = TTy->getDecl()->getDefinition();
231    if (RD->hasFlexibleArrayMember())
232      return false;
233    // RD can be struct, union, class, interface or enum.
234    // For now, we only handle struct and class.
235    if (RD->isStruct() || RD->isClass())
236      return true;
237  }
238  return false;
239}
240
241llvm::MDNode *
242CodeGenTBAA::getTBAAStructTypeInfo(QualType QTy) {
243  const Type *Ty = Context.getCanonicalType(QTy).getTypePtr();
244  assert(isTBAAPathStruct(QTy));
245
246  if (llvm::MDNode *N = StructTypeMetadataCache[Ty])
247    return N;
248
249  if (const RecordType *TTy = QTy->getAs<RecordType>()) {
250    const RecordDecl *RD = TTy->getDecl()->getDefinition();
251
252    const ASTRecordLayout &Layout = Context.getASTRecordLayout(RD);
253    SmallVector <std::pair<llvm::MDNode*, uint64_t>, 4> Fields;
254    unsigned idx = 0;
255    for (RecordDecl::field_iterator i = RD->field_begin(),
256         e = RD->field_end(); i != e; ++i, ++idx) {
257      QualType FieldQTy = i->getType();
258      llvm::MDNode *FieldNode;
259      if (isTBAAPathStruct(FieldQTy))
260        FieldNode = getTBAAStructTypeInfo(FieldQTy);
261      else
262        FieldNode = getTBAAInfo(FieldQTy);
263      if (!FieldNode)
264        return StructTypeMetadataCache[Ty] = nullptr;
265      Fields.push_back(std::make_pair(
266          FieldNode, Layout.getFieldOffset(idx) / Context.getCharWidth()));
267    }
268
269    SmallString<256> OutName;
270    if (Features.CPlusPlus) {
271      // Don't use the mangler for C code.
272      llvm::raw_svector_ostream Out(OutName);
273      MContext.mangleTypeName(QualType(Ty, 0), Out);
274      Out.flush();
275    } else {
276      OutName = RD->getName();
277    }
278    // Create the struct type node with a vector of pairs (offset, type).
279    return StructTypeMetadataCache[Ty] =
280      MDHelper.createTBAAStructTypeNode(OutName, Fields);
281  }
282
283  return StructMetadataCache[Ty] = nullptr;
284}
285
286/// Return a TBAA tag node for both scalar TBAA and struct-path aware TBAA.
287llvm::MDNode *
288CodeGenTBAA::getTBAAStructTagInfo(QualType BaseQTy, llvm::MDNode *AccessNode,
289                                  uint64_t Offset) {
290  if (!AccessNode)
291    return nullptr;
292
293  if (!CodeGenOpts.StructPathTBAA)
294    return getTBAAScalarTagInfo(AccessNode);
295
296  const Type *BTy = Context.getCanonicalType(BaseQTy).getTypePtr();
297  TBAAPathTag PathTag = TBAAPathTag(BTy, AccessNode, Offset);
298  if (llvm::MDNode *N = StructTagMetadataCache[PathTag])
299    return N;
300
301  llvm::MDNode *BNode = nullptr;
302  if (isTBAAPathStruct(BaseQTy))
303    BNode  = getTBAAStructTypeInfo(BaseQTy);
304  if (!BNode)
305    return StructTagMetadataCache[PathTag] =
306       MDHelper.createTBAAStructTagNode(AccessNode, AccessNode, 0);
307
308  return StructTagMetadataCache[PathTag] =
309    MDHelper.createTBAAStructTagNode(BNode, AccessNode, Offset);
310}
311
312llvm::MDNode *
313CodeGenTBAA::getTBAAScalarTagInfo(llvm::MDNode *AccessNode) {
314  if (!AccessNode)
315    return nullptr;
316  if (llvm::MDNode *N = ScalarTagMetadataCache[AccessNode])
317    return N;
318
319  return ScalarTagMetadataCache[AccessNode] =
320    MDHelper.createTBAAStructTagNode(AccessNode, AccessNode, 0);
321}
322