1 //===-- CodeGenTBAA.cpp - TBAA information for LLVM CodeGen ---------------===//
2 //
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6 //
7 //===----------------------------------------------------------------------===//
8 //
9 // This is the code that manages TBAA information and defines the TBAA policy
10 // for the optimizer to use. Relevant standards text includes:
11 //
12 //   C99 6.5p7
13 //   C++ [basic.lval] (p10 in n3126, p15 in some earlier versions)
14 //
15 //===----------------------------------------------------------------------===//
16 
17 #include "CodeGenTBAA.h"
18 #include "clang/AST/ASTContext.h"
19 #include "clang/AST/Attr.h"
20 #include "clang/AST/Mangle.h"
21 #include "clang/AST/RecordLayout.h"
22 #include "clang/Basic/CodeGenOptions.h"
23 #include "llvm/ADT/SmallSet.h"
24 #include "llvm/IR/Constants.h"
25 #include "llvm/IR/LLVMContext.h"
26 #include "llvm/IR/Metadata.h"
27 #include "llvm/IR/Module.h"
28 #include "llvm/IR/Type.h"
29 using namespace clang;
30 using namespace CodeGen;
31 
32 CodeGenTBAA::CodeGenTBAA(ASTContext &Ctx, llvm::Module &M,
33                          const CodeGenOptions &CGO,
34                          const LangOptions &Features, MangleContext &MContext)
35   : Context(Ctx), Module(M), CodeGenOpts(CGO),
36     Features(Features), MContext(MContext), MDHelper(M.getContext()),
37     Root(nullptr), Char(nullptr)
38 {}
39 
40 CodeGenTBAA::~CodeGenTBAA() {
41 }
42 
43 llvm::MDNode *CodeGenTBAA::getRoot() {
44   // Define the root of the tree. This identifies the tree, so that
45   // if our LLVM IR is linked with LLVM IR from a different front-end
46   // (or a different version of this front-end), their TBAA trees will
47   // remain distinct, and the optimizer will treat them conservatively.
48   if (!Root) {
49     if (Features.CPlusPlus)
50       Root = MDHelper.createTBAARoot("Simple C++ TBAA");
51     else
52       Root = MDHelper.createTBAARoot("Simple C/C++ TBAA");
53   }
54 
55   return Root;
56 }
57 
58 llvm::MDNode *CodeGenTBAA::createScalarTypeNode(StringRef Name,
59                                                 llvm::MDNode *Parent,
60                                                 uint64_t Size) {
61   if (CodeGenOpts.NewStructPathTBAA) {
62     llvm::Metadata *Id = MDHelper.createString(Name);
63     return MDHelper.createTBAATypeNode(Parent, Size, Id);
64   }
65   return MDHelper.createTBAAScalarTypeNode(Name, Parent);
66 }
67 
68 llvm::MDNode *CodeGenTBAA::getChar() {
69   // Define the root of the tree for user-accessible memory. C and C++
70   // give special powers to char and certain similar types. However,
71   // these special powers only cover user-accessible memory, and doesn't
72   // include things like vtables.
73   if (!Char)
74     Char = createScalarTypeNode("omnipotent char", getRoot(), /* Size= */ 1);
75 
76   return Char;
77 }
78 
79 static bool TypeHasMayAlias(QualType QTy) {
80   // Tagged types have declarations, and therefore may have attributes.
81   if (auto *TD = QTy->getAsTagDecl())
82     if (TD->hasAttr<MayAliasAttr>())
83       return true;
84 
85   // Also look for may_alias as a declaration attribute on a typedef.
86   // FIXME: We should follow GCC and model may_alias as a type attribute
87   // rather than as a declaration attribute.
88   while (auto *TT = QTy->getAs<TypedefType>()) {
89     if (TT->getDecl()->hasAttr<MayAliasAttr>())
90       return true;
91     QTy = TT->desugar();
92   }
93   return false;
94 }
95 
96 /// Check if the given type is a valid base type to be used in access tags.
97 static bool isValidBaseType(QualType QTy) {
98   if (QTy->isReferenceType())
99     return false;
100   if (const RecordType *TTy = QTy->getAs<RecordType>()) {
101     const RecordDecl *RD = TTy->getDecl()->getDefinition();
102     // Incomplete types are not valid base access types.
103     if (!RD)
104       return false;
105     if (RD->hasFlexibleArrayMember())
106       return false;
107     // RD can be struct, union, class, interface or enum.
108     // For now, we only handle struct and class.
109     if (RD->isStruct() || RD->isClass())
110       return true;
111   }
112   return false;
113 }
114 
115 llvm::MDNode *CodeGenTBAA::getTypeInfoHelper(const Type *Ty) {
116   uint64_t Size = Context.getTypeSizeInChars(Ty).getQuantity();
117 
118   // Handle builtin types.
119   if (const BuiltinType *BTy = dyn_cast<BuiltinType>(Ty)) {
120     switch (BTy->getKind()) {
121     // Character types are special and can alias anything.
122     // In C++, this technically only includes "char" and "unsigned char",
123     // and not "signed char". In C, it includes all three. For now,
124     // the risk of exploiting this detail in C++ seems likely to outweigh
125     // the benefit.
126     case BuiltinType::Char_U:
127     case BuiltinType::Char_S:
128     case BuiltinType::UChar:
129     case BuiltinType::SChar:
130       return getChar();
131 
132     // Unsigned types can alias their corresponding signed types.
133     case BuiltinType::UShort:
134       return getTypeInfo(Context.ShortTy);
135     case BuiltinType::UInt:
136       return getTypeInfo(Context.IntTy);
137     case BuiltinType::ULong:
138       return getTypeInfo(Context.LongTy);
139     case BuiltinType::ULongLong:
140       return getTypeInfo(Context.LongLongTy);
141     case BuiltinType::UInt128:
142       return getTypeInfo(Context.Int128Ty);
143 
144     // Treat all other builtin types as distinct types. This includes
145     // treating wchar_t, char16_t, and char32_t as distinct from their
146     // "underlying types".
147     default:
148       return createScalarTypeNode(BTy->getName(Features), getChar(), Size);
149     }
150   }
151 
152   // C++1z [basic.lval]p10: "If a program attempts to access the stored value of
153   // an object through a glvalue of other than one of the following types the
154   // behavior is undefined: [...] a char, unsigned char, or std::byte type."
155   if (Ty->isStdByteType())
156     return getChar();
157 
158   // Handle pointers and references.
159   // TODO: Implement C++'s type "similarity" and consider dis-"similar"
160   // pointers distinct.
161   if (Ty->isPointerType() || Ty->isReferenceType())
162     return createScalarTypeNode("any pointer", getChar(), Size);
163 
164   // Accesses to arrays are accesses to objects of their element types.
165   if (CodeGenOpts.NewStructPathTBAA && Ty->isArrayType())
166     return getTypeInfo(cast<ArrayType>(Ty)->getElementType());
167 
168   // Enum types are distinct types. In C++ they have "underlying types",
169   // however they aren't related for TBAA.
170   if (const EnumType *ETy = dyn_cast<EnumType>(Ty)) {
171     // In C++ mode, types have linkage, so we can rely on the ODR and
172     // on their mangled names, if they're external.
173     // TODO: Is there a way to get a program-wide unique name for a
174     // decl with local linkage or no linkage?
175     if (!Features.CPlusPlus || !ETy->getDecl()->isExternallyVisible())
176       return getChar();
177 
178     SmallString<256> OutName;
179     llvm::raw_svector_ostream Out(OutName);
180     MContext.mangleTypeName(QualType(ETy, 0), Out);
181     return createScalarTypeNode(OutName, getChar(), Size);
182   }
183 
184   // For now, handle any other kind of type conservatively.
185   return getChar();
186 }
187 
188 llvm::MDNode *CodeGenTBAA::getTypeInfo(QualType QTy) {
189   // At -O0 or relaxed aliasing, TBAA is not emitted for regular types.
190   if (CodeGenOpts.OptimizationLevel == 0 || CodeGenOpts.RelaxedAliasing)
191     return nullptr;
192 
193   // If the type has the may_alias attribute (even on a typedef), it is
194   // effectively in the general char alias class.
195   if (TypeHasMayAlias(QTy))
196     return getChar();
197 
198   // We need this function to not fall back to returning the "omnipotent char"
199   // type node for aggregate and union types. Otherwise, any dereference of an
200   // aggregate will result into the may-alias access descriptor, meaning all
201   // subsequent accesses to direct and indirect members of that aggregate will
202   // be considered may-alias too.
203   // TODO: Combine getTypeInfo() and getBaseTypeInfo() into a single function.
204   if (isValidBaseType(QTy))
205     return getBaseTypeInfo(QTy);
206 
207   const Type *Ty = Context.getCanonicalType(QTy).getTypePtr();
208   if (llvm::MDNode *N = MetadataCache[Ty])
209     return N;
210 
211   // Note that the following helper call is allowed to add new nodes to the
212   // cache, which invalidates all its previously obtained iterators. So we
213   // first generate the node for the type and then add that node to the cache.
214   llvm::MDNode *TypeNode = getTypeInfoHelper(Ty);
215   return MetadataCache[Ty] = TypeNode;
216 }
217 
218 TBAAAccessInfo CodeGenTBAA::getAccessInfo(QualType AccessType) {
219   // Pointee values may have incomplete types, but they shall never be
220   // dereferenced.
221   if (AccessType->isIncompleteType())
222     return TBAAAccessInfo::getIncompleteInfo();
223 
224   if (TypeHasMayAlias(AccessType))
225     return TBAAAccessInfo::getMayAliasInfo();
226 
227   uint64_t Size = Context.getTypeSizeInChars(AccessType).getQuantity();
228   return TBAAAccessInfo(getTypeInfo(AccessType), Size);
229 }
230 
231 TBAAAccessInfo CodeGenTBAA::getVTablePtrAccessInfo(llvm::Type *VTablePtrType) {
232   llvm::DataLayout DL(&Module);
233   unsigned Size = DL.getPointerTypeSize(VTablePtrType);
234   return TBAAAccessInfo(createScalarTypeNode("vtable pointer", getRoot(), Size),
235                         Size);
236 }
237 
238 bool
239 CodeGenTBAA::CollectFields(uint64_t BaseOffset,
240                            QualType QTy,
241                            SmallVectorImpl<llvm::MDBuilder::TBAAStructField> &
242                              Fields,
243                            bool MayAlias) {
244   /* Things not handled yet include: C++ base classes, bitfields, */
245 
246   if (const RecordType *TTy = QTy->getAs<RecordType>()) {
247     const RecordDecl *RD = TTy->getDecl()->getDefinition();
248     if (RD->hasFlexibleArrayMember())
249       return false;
250 
251     // TODO: Handle C++ base classes.
252     if (const CXXRecordDecl *Decl = dyn_cast<CXXRecordDecl>(RD))
253       if (Decl->bases_begin() != Decl->bases_end())
254         return false;
255 
256     const ASTRecordLayout &Layout = Context.getASTRecordLayout(RD);
257 
258     unsigned idx = 0;
259     for (RecordDecl::field_iterator i = RD->field_begin(),
260          e = RD->field_end(); i != e; ++i, ++idx) {
261       if ((*i)->isZeroSize(Context) || (*i)->isUnnamedBitfield())
262         continue;
263       uint64_t Offset = BaseOffset +
264                         Layout.getFieldOffset(idx) / Context.getCharWidth();
265       QualType FieldQTy = i->getType();
266       if (!CollectFields(Offset, FieldQTy, Fields,
267                          MayAlias || TypeHasMayAlias(FieldQTy)))
268         return false;
269     }
270     return true;
271   }
272 
273   /* Otherwise, treat whatever it is as a field. */
274   uint64_t Offset = BaseOffset;
275   uint64_t Size = Context.getTypeSizeInChars(QTy).getQuantity();
276   llvm::MDNode *TBAAType = MayAlias ? getChar() : getTypeInfo(QTy);
277   llvm::MDNode *TBAATag = getAccessTagInfo(TBAAAccessInfo(TBAAType, Size));
278   Fields.push_back(llvm::MDBuilder::TBAAStructField(Offset, Size, TBAATag));
279   return true;
280 }
281 
282 llvm::MDNode *
283 CodeGenTBAA::getTBAAStructInfo(QualType QTy) {
284   const Type *Ty = Context.getCanonicalType(QTy).getTypePtr();
285 
286   if (llvm::MDNode *N = StructMetadataCache[Ty])
287     return N;
288 
289   SmallVector<llvm::MDBuilder::TBAAStructField, 4> Fields;
290   if (CollectFields(0, QTy, Fields, TypeHasMayAlias(QTy)))
291     return MDHelper.createTBAAStructNode(Fields);
292 
293   // For now, handle any other kind of type conservatively.
294   return StructMetadataCache[Ty] = nullptr;
295 }
296 
297 llvm::MDNode *CodeGenTBAA::getBaseTypeInfoHelper(const Type *Ty) {
298   if (auto *TTy = dyn_cast<RecordType>(Ty)) {
299     const RecordDecl *RD = TTy->getDecl()->getDefinition();
300     const ASTRecordLayout &Layout = Context.getASTRecordLayout(RD);
301     SmallVector<llvm::MDBuilder::TBAAStructField, 4> Fields;
302     for (FieldDecl *Field : RD->fields()) {
303       if (Field->isZeroSize(Context) || Field->isUnnamedBitfield())
304         continue;
305       QualType FieldQTy = Field->getType();
306       llvm::MDNode *TypeNode = isValidBaseType(FieldQTy) ?
307           getBaseTypeInfo(FieldQTy) : getTypeInfo(FieldQTy);
308       if (!TypeNode)
309         return BaseTypeMetadataCache[Ty] = nullptr;
310 
311       uint64_t BitOffset = Layout.getFieldOffset(Field->getFieldIndex());
312       uint64_t Offset = Context.toCharUnitsFromBits(BitOffset).getQuantity();
313       uint64_t Size = Context.getTypeSizeInChars(FieldQTy).getQuantity();
314       Fields.push_back(llvm::MDBuilder::TBAAStructField(Offset, Size,
315                                                         TypeNode));
316     }
317 
318     SmallString<256> OutName;
319     if (Features.CPlusPlus) {
320       // Don't use the mangler for C code.
321       llvm::raw_svector_ostream Out(OutName);
322       MContext.mangleTypeName(QualType(Ty, 0), Out);
323     } else {
324       OutName = RD->getName();
325     }
326 
327     if (CodeGenOpts.NewStructPathTBAA) {
328       llvm::MDNode *Parent = getChar();
329       uint64_t Size = Context.getTypeSizeInChars(Ty).getQuantity();
330       llvm::Metadata *Id = MDHelper.createString(OutName);
331       return MDHelper.createTBAATypeNode(Parent, Size, Id, Fields);
332     }
333 
334     // Create the struct type node with a vector of pairs (offset, type).
335     SmallVector<std::pair<llvm::MDNode*, uint64_t>, 4> OffsetsAndTypes;
336     for (const auto &Field : Fields)
337         OffsetsAndTypes.push_back(std::make_pair(Field.Type, Field.Offset));
338     return MDHelper.createTBAAStructTypeNode(OutName, OffsetsAndTypes);
339   }
340 
341   return nullptr;
342 }
343 
344 llvm::MDNode *CodeGenTBAA::getBaseTypeInfo(QualType QTy) {
345   if (!isValidBaseType(QTy))
346     return nullptr;
347 
348   const Type *Ty = Context.getCanonicalType(QTy).getTypePtr();
349   if (llvm::MDNode *N = BaseTypeMetadataCache[Ty])
350     return N;
351 
352   // Note that the following helper call is allowed to add new nodes to the
353   // cache, which invalidates all its previously obtained iterators. So we
354   // first generate the node for the type and then add that node to the cache.
355   llvm::MDNode *TypeNode = getBaseTypeInfoHelper(Ty);
356   return BaseTypeMetadataCache[Ty] = TypeNode;
357 }
358 
359 llvm::MDNode *CodeGenTBAA::getAccessTagInfo(TBAAAccessInfo Info) {
360   assert(!Info.isIncomplete() && "Access to an object of an incomplete type!");
361 
362   if (Info.isMayAlias())
363     Info = TBAAAccessInfo(getChar(), Info.Size);
364 
365   if (!Info.AccessType)
366     return nullptr;
367 
368   if (!CodeGenOpts.StructPathTBAA)
369     Info = TBAAAccessInfo(Info.AccessType, Info.Size);
370 
371   llvm::MDNode *&N = AccessTagMetadataCache[Info];
372   if (N)
373     return N;
374 
375   if (!Info.BaseType) {
376     Info.BaseType = Info.AccessType;
377     assert(!Info.Offset && "Nonzero offset for an access with no base type!");
378   }
379   if (CodeGenOpts.NewStructPathTBAA) {
380     return N = MDHelper.createTBAAAccessTag(Info.BaseType, Info.AccessType,
381                                             Info.Offset, Info.Size);
382   }
383   return N = MDHelper.createTBAAStructTagNode(Info.BaseType, Info.AccessType,
384                                               Info.Offset);
385 }
386 
387 TBAAAccessInfo CodeGenTBAA::mergeTBAAInfoForCast(TBAAAccessInfo SourceInfo,
388                                                  TBAAAccessInfo TargetInfo) {
389   if (SourceInfo.isMayAlias() || TargetInfo.isMayAlias())
390     return TBAAAccessInfo::getMayAliasInfo();
391   return TargetInfo;
392 }
393 
394 TBAAAccessInfo
395 CodeGenTBAA::mergeTBAAInfoForConditionalOperator(TBAAAccessInfo InfoA,
396                                                  TBAAAccessInfo InfoB) {
397   if (InfoA == InfoB)
398     return InfoA;
399 
400   if (!InfoA || !InfoB)
401     return TBAAAccessInfo();
402 
403   if (InfoA.isMayAlias() || InfoB.isMayAlias())
404     return TBAAAccessInfo::getMayAliasInfo();
405 
406   // TODO: Implement the rest of the logic here. For example, two accesses
407   // with same final access types result in an access to an object of that final
408   // access type regardless of their base types.
409   return TBAAAccessInfo::getMayAliasInfo();
410 }
411 
412 TBAAAccessInfo
413 CodeGenTBAA::mergeTBAAInfoForMemoryTransfer(TBAAAccessInfo DestInfo,
414                                             TBAAAccessInfo SrcInfo) {
415   if (DestInfo == SrcInfo)
416     return DestInfo;
417 
418   if (!DestInfo || !SrcInfo)
419     return TBAAAccessInfo();
420 
421   if (DestInfo.isMayAlias() || SrcInfo.isMayAlias())
422     return TBAAAccessInfo::getMayAliasInfo();
423 
424   // TODO: Implement the rest of the logic here. For example, two accesses
425   // with same final access types result in an access to an object of that final
426   // access type regardless of their base types.
427   return TBAAAccessInfo::getMayAliasInfo();
428 }
429