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     case BuiltinType::UShortFract:
145       return getTypeInfo(Context.ShortFractTy);
146     case BuiltinType::UFract:
147       return getTypeInfo(Context.FractTy);
148     case BuiltinType::ULongFract:
149       return getTypeInfo(Context.LongFractTy);
150 
151     case BuiltinType::SatUShortFract:
152       return getTypeInfo(Context.SatShortFractTy);
153     case BuiltinType::SatUFract:
154       return getTypeInfo(Context.SatFractTy);
155     case BuiltinType::SatULongFract:
156       return getTypeInfo(Context.SatLongFractTy);
157 
158     case BuiltinType::UShortAccum:
159       return getTypeInfo(Context.ShortAccumTy);
160     case BuiltinType::UAccum:
161       return getTypeInfo(Context.AccumTy);
162     case BuiltinType::ULongAccum:
163       return getTypeInfo(Context.LongAccumTy);
164 
165     case BuiltinType::SatUShortAccum:
166       return getTypeInfo(Context.SatShortAccumTy);
167     case BuiltinType::SatUAccum:
168       return getTypeInfo(Context.SatAccumTy);
169     case BuiltinType::SatULongAccum:
170       return getTypeInfo(Context.SatLongAccumTy);
171 
172     // Treat all other builtin types as distinct types. This includes
173     // treating wchar_t, char16_t, and char32_t as distinct from their
174     // "underlying types".
175     default:
176       return createScalarTypeNode(BTy->getName(Features), getChar(), Size);
177     }
178   }
179 
180   // C++1z [basic.lval]p10: "If a program attempts to access the stored value of
181   // an object through a glvalue of other than one of the following types the
182   // behavior is undefined: [...] a char, unsigned char, or std::byte type."
183   if (Ty->isStdByteType())
184     return getChar();
185 
186   // Handle pointers and references.
187   // TODO: Implement C++'s type "similarity" and consider dis-"similar"
188   // pointers distinct.
189   if (Ty->isPointerType() || Ty->isReferenceType())
190     return createScalarTypeNode("any pointer", getChar(), Size);
191 
192   // Accesses to arrays are accesses to objects of their element types.
193   if (CodeGenOpts.NewStructPathTBAA && Ty->isArrayType())
194     return getTypeInfo(cast<ArrayType>(Ty)->getElementType());
195 
196   // Enum types are distinct types. In C++ they have "underlying types",
197   // however they aren't related for TBAA.
198   if (const EnumType *ETy = dyn_cast<EnumType>(Ty)) {
199     // In C++ mode, types have linkage, so we can rely on the ODR and
200     // on their mangled names, if they're external.
201     // TODO: Is there a way to get a program-wide unique name for a
202     // decl with local linkage or no linkage?
203     if (!Features.CPlusPlus || !ETy->getDecl()->isExternallyVisible())
204       return getChar();
205 
206     SmallString<256> OutName;
207     llvm::raw_svector_ostream Out(OutName);
208     MContext.mangleTypeName(QualType(ETy, 0), Out);
209     return createScalarTypeNode(OutName, getChar(), Size);
210   }
211 
212   if (const auto *EIT = dyn_cast<BitIntType>(Ty)) {
213     SmallString<256> OutName;
214     llvm::raw_svector_ostream Out(OutName);
215     // Don't specify signed/unsigned since integer types can alias despite sign
216     // differences.
217     Out << "_BitInt(" << EIT->getNumBits() << ')';
218     return createScalarTypeNode(OutName, getChar(), Size);
219   }
220 
221   // For now, handle any other kind of type conservatively.
222   return getChar();
223 }
224 
225 llvm::MDNode *CodeGenTBAA::getTypeInfo(QualType QTy) {
226   // At -O0 or relaxed aliasing, TBAA is not emitted for regular types.
227   if (CodeGenOpts.OptimizationLevel == 0 || CodeGenOpts.RelaxedAliasing)
228     return nullptr;
229 
230   // If the type has the may_alias attribute (even on a typedef), it is
231   // effectively in the general char alias class.
232   if (TypeHasMayAlias(QTy))
233     return getChar();
234 
235   // We need this function to not fall back to returning the "omnipotent char"
236   // type node for aggregate and union types. Otherwise, any dereference of an
237   // aggregate will result into the may-alias access descriptor, meaning all
238   // subsequent accesses to direct and indirect members of that aggregate will
239   // be considered may-alias too.
240   // TODO: Combine getTypeInfo() and getBaseTypeInfo() into a single function.
241   if (isValidBaseType(QTy))
242     return getBaseTypeInfo(QTy);
243 
244   const Type *Ty = Context.getCanonicalType(QTy).getTypePtr();
245   if (llvm::MDNode *N = MetadataCache[Ty])
246     return N;
247 
248   // Note that the following helper call is allowed to add new nodes to the
249   // cache, which invalidates all its previously obtained iterators. So we
250   // first generate the node for the type and then add that node to the cache.
251   llvm::MDNode *TypeNode = getTypeInfoHelper(Ty);
252   return MetadataCache[Ty] = TypeNode;
253 }
254 
255 TBAAAccessInfo CodeGenTBAA::getAccessInfo(QualType AccessType) {
256   // Pointee values may have incomplete types, but they shall never be
257   // dereferenced.
258   if (AccessType->isIncompleteType())
259     return TBAAAccessInfo::getIncompleteInfo();
260 
261   if (TypeHasMayAlias(AccessType))
262     return TBAAAccessInfo::getMayAliasInfo();
263 
264   uint64_t Size = Context.getTypeSizeInChars(AccessType).getQuantity();
265   return TBAAAccessInfo(getTypeInfo(AccessType), Size);
266 }
267 
268 TBAAAccessInfo CodeGenTBAA::getVTablePtrAccessInfo(llvm::Type *VTablePtrType) {
269   llvm::DataLayout DL(&Module);
270   unsigned Size = DL.getPointerTypeSize(VTablePtrType);
271   return TBAAAccessInfo(createScalarTypeNode("vtable pointer", getRoot(), Size),
272                         Size);
273 }
274 
275 bool
276 CodeGenTBAA::CollectFields(uint64_t BaseOffset,
277                            QualType QTy,
278                            SmallVectorImpl<llvm::MDBuilder::TBAAStructField> &
279                              Fields,
280                            bool MayAlias) {
281   /* Things not handled yet include: C++ base classes, bitfields, */
282 
283   if (const RecordType *TTy = QTy->getAs<RecordType>()) {
284     const RecordDecl *RD = TTy->getDecl()->getDefinition();
285     if (RD->hasFlexibleArrayMember())
286       return false;
287 
288     // TODO: Handle C++ base classes.
289     if (const CXXRecordDecl *Decl = dyn_cast<CXXRecordDecl>(RD))
290       if (Decl->bases_begin() != Decl->bases_end())
291         return false;
292 
293     const ASTRecordLayout &Layout = Context.getASTRecordLayout(RD);
294 
295     unsigned idx = 0;
296     for (RecordDecl::field_iterator i = RD->field_begin(),
297          e = RD->field_end(); i != e; ++i, ++idx) {
298       if ((*i)->isZeroSize(Context) || (*i)->isUnnamedBitfield())
299         continue;
300       uint64_t Offset = BaseOffset +
301                         Layout.getFieldOffset(idx) / Context.getCharWidth();
302       QualType FieldQTy = i->getType();
303       if (!CollectFields(Offset, FieldQTy, Fields,
304                          MayAlias || TypeHasMayAlias(FieldQTy)))
305         return false;
306     }
307     return true;
308   }
309 
310   /* Otherwise, treat whatever it is as a field. */
311   uint64_t Offset = BaseOffset;
312   uint64_t Size = Context.getTypeSizeInChars(QTy).getQuantity();
313   llvm::MDNode *TBAAType = MayAlias ? getChar() : getTypeInfo(QTy);
314   llvm::MDNode *TBAATag = getAccessTagInfo(TBAAAccessInfo(TBAAType, Size));
315   Fields.push_back(llvm::MDBuilder::TBAAStructField(Offset, Size, TBAATag));
316   return true;
317 }
318 
319 llvm::MDNode *
320 CodeGenTBAA::getTBAAStructInfo(QualType QTy) {
321   const Type *Ty = Context.getCanonicalType(QTy).getTypePtr();
322 
323   if (llvm::MDNode *N = StructMetadataCache[Ty])
324     return N;
325 
326   SmallVector<llvm::MDBuilder::TBAAStructField, 4> Fields;
327   if (CollectFields(0, QTy, Fields, TypeHasMayAlias(QTy)))
328     return MDHelper.createTBAAStructNode(Fields);
329 
330   // For now, handle any other kind of type conservatively.
331   return StructMetadataCache[Ty] = nullptr;
332 }
333 
334 llvm::MDNode *CodeGenTBAA::getBaseTypeInfoHelper(const Type *Ty) {
335   if (auto *TTy = dyn_cast<RecordType>(Ty)) {
336     const RecordDecl *RD = TTy->getDecl()->getDefinition();
337     const ASTRecordLayout &Layout = Context.getASTRecordLayout(RD);
338     using TBAAStructField = llvm::MDBuilder::TBAAStructField;
339     SmallVector<TBAAStructField, 4> Fields;
340     if (const CXXRecordDecl *CXXRD = dyn_cast<CXXRecordDecl>(RD)) {
341       // Handle C++ base classes. Non-virtual bases can treated a a kind of
342       // field. Virtual bases are more complex and omitted, but avoid an
343       // incomplete view for NewStructPathTBAA.
344       if (CodeGenOpts.NewStructPathTBAA && CXXRD->getNumVBases() != 0)
345         return BaseTypeMetadataCache[Ty] = nullptr;
346       for (const CXXBaseSpecifier &B : CXXRD->bases()) {
347         if (B.isVirtual())
348           continue;
349         QualType BaseQTy = B.getType();
350         const CXXRecordDecl *BaseRD = BaseQTy->getAsCXXRecordDecl();
351         if (BaseRD->isEmpty())
352           continue;
353         llvm::MDNode *TypeNode = isValidBaseType(BaseQTy)
354                                      ? getBaseTypeInfo(BaseQTy)
355                                      : getTypeInfo(BaseQTy);
356         if (!TypeNode)
357           return BaseTypeMetadataCache[Ty] = nullptr;
358         uint64_t Offset = Layout.getBaseClassOffset(BaseRD).getQuantity();
359         uint64_t Size =
360             Context.getASTRecordLayout(BaseRD).getDataSize().getQuantity();
361         Fields.push_back(
362             llvm::MDBuilder::TBAAStructField(Offset, Size, TypeNode));
363       }
364       // The order in which base class subobjects are allocated is unspecified,
365       // so may differ from declaration order. In particular, Itanium ABI will
366       // allocate a primary base first.
367       // Since we exclude empty subobjects, the objects are not overlapping and
368       // their offsets are unique.
369       llvm::sort(Fields,
370                  [](const TBAAStructField &A, const TBAAStructField &B) {
371                    return A.Offset < B.Offset;
372                  });
373     }
374     for (FieldDecl *Field : RD->fields()) {
375       if (Field->isZeroSize(Context) || Field->isUnnamedBitfield())
376         continue;
377       QualType FieldQTy = Field->getType();
378       llvm::MDNode *TypeNode = isValidBaseType(FieldQTy) ?
379           getBaseTypeInfo(FieldQTy) : getTypeInfo(FieldQTy);
380       if (!TypeNode)
381         return BaseTypeMetadataCache[Ty] = nullptr;
382 
383       uint64_t BitOffset = Layout.getFieldOffset(Field->getFieldIndex());
384       uint64_t Offset = Context.toCharUnitsFromBits(BitOffset).getQuantity();
385       uint64_t Size = Context.getTypeSizeInChars(FieldQTy).getQuantity();
386       Fields.push_back(llvm::MDBuilder::TBAAStructField(Offset, Size,
387                                                         TypeNode));
388     }
389 
390     SmallString<256> OutName;
391     if (Features.CPlusPlus) {
392       // Don't use the mangler for C code.
393       llvm::raw_svector_ostream Out(OutName);
394       MContext.mangleTypeName(QualType(Ty, 0), Out);
395     } else {
396       OutName = RD->getName();
397     }
398 
399     if (CodeGenOpts.NewStructPathTBAA) {
400       llvm::MDNode *Parent = getChar();
401       uint64_t Size = Context.getTypeSizeInChars(Ty).getQuantity();
402       llvm::Metadata *Id = MDHelper.createString(OutName);
403       return MDHelper.createTBAATypeNode(Parent, Size, Id, Fields);
404     }
405 
406     // Create the struct type node with a vector of pairs (offset, type).
407     SmallVector<std::pair<llvm::MDNode*, uint64_t>, 4> OffsetsAndTypes;
408     for (const auto &Field : Fields)
409         OffsetsAndTypes.push_back(std::make_pair(Field.Type, Field.Offset));
410     return MDHelper.createTBAAStructTypeNode(OutName, OffsetsAndTypes);
411   }
412 
413   return nullptr;
414 }
415 
416 llvm::MDNode *CodeGenTBAA::getBaseTypeInfo(QualType QTy) {
417   if (!isValidBaseType(QTy))
418     return nullptr;
419 
420   const Type *Ty = Context.getCanonicalType(QTy).getTypePtr();
421   if (llvm::MDNode *N = BaseTypeMetadataCache[Ty])
422     return N;
423 
424   // Note that the following helper call is allowed to add new nodes to the
425   // cache, which invalidates all its previously obtained iterators. So we
426   // first generate the node for the type and then add that node to the cache.
427   llvm::MDNode *TypeNode = getBaseTypeInfoHelper(Ty);
428   return BaseTypeMetadataCache[Ty] = TypeNode;
429 }
430 
431 llvm::MDNode *CodeGenTBAA::getAccessTagInfo(TBAAAccessInfo Info) {
432   assert(!Info.isIncomplete() && "Access to an object of an incomplete type!");
433 
434   if (Info.isMayAlias())
435     Info = TBAAAccessInfo(getChar(), Info.Size);
436 
437   if (!Info.AccessType)
438     return nullptr;
439 
440   if (!CodeGenOpts.StructPathTBAA)
441     Info = TBAAAccessInfo(Info.AccessType, Info.Size);
442 
443   llvm::MDNode *&N = AccessTagMetadataCache[Info];
444   if (N)
445     return N;
446 
447   if (!Info.BaseType) {
448     Info.BaseType = Info.AccessType;
449     assert(!Info.Offset && "Nonzero offset for an access with no base type!");
450   }
451   if (CodeGenOpts.NewStructPathTBAA) {
452     return N = MDHelper.createTBAAAccessTag(Info.BaseType, Info.AccessType,
453                                             Info.Offset, Info.Size);
454   }
455   return N = MDHelper.createTBAAStructTagNode(Info.BaseType, Info.AccessType,
456                                               Info.Offset);
457 }
458 
459 TBAAAccessInfo CodeGenTBAA::mergeTBAAInfoForCast(TBAAAccessInfo SourceInfo,
460                                                  TBAAAccessInfo TargetInfo) {
461   if (SourceInfo.isMayAlias() || TargetInfo.isMayAlias())
462     return TBAAAccessInfo::getMayAliasInfo();
463   return TargetInfo;
464 }
465 
466 TBAAAccessInfo
467 CodeGenTBAA::mergeTBAAInfoForConditionalOperator(TBAAAccessInfo InfoA,
468                                                  TBAAAccessInfo InfoB) {
469   if (InfoA == InfoB)
470     return InfoA;
471 
472   if (!InfoA || !InfoB)
473     return TBAAAccessInfo();
474 
475   if (InfoA.isMayAlias() || InfoB.isMayAlias())
476     return TBAAAccessInfo::getMayAliasInfo();
477 
478   // TODO: Implement the rest of the logic here. For example, two accesses
479   // with same final access types result in an access to an object of that final
480   // access type regardless of their base types.
481   return TBAAAccessInfo::getMayAliasInfo();
482 }
483 
484 TBAAAccessInfo
485 CodeGenTBAA::mergeTBAAInfoForMemoryTransfer(TBAAAccessInfo DestInfo,
486                                             TBAAAccessInfo SrcInfo) {
487   if (DestInfo == SrcInfo)
488     return DestInfo;
489 
490   if (!DestInfo || !SrcInfo)
491     return TBAAAccessInfo();
492 
493   if (DestInfo.isMayAlias() || SrcInfo.isMayAlias())
494     return TBAAAccessInfo::getMayAliasInfo();
495 
496   // TODO: Implement the rest of the logic here. For example, two accesses
497   // with same final access types result in an access to an object of that final
498   // access type regardless of their base types.
499   return TBAAAccessInfo::getMayAliasInfo();
500 }
501