1 //=======- PaddingChecker.cpp ------------------------------------*- C++ -*-==// 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 file defines a checker that checks for padding that could be 11 // removed by re-ordering members. 12 // 13 //===----------------------------------------------------------------------===// 14 15 #include "ClangSACheckers.h" 16 #include "clang/AST/CharUnits.h" 17 #include "clang/AST/DeclTemplate.h" 18 #include "clang/AST/RecordLayout.h" 19 #include "clang/AST/RecursiveASTVisitor.h" 20 #include "clang/StaticAnalyzer/Core/BugReporter/BugReporter.h" 21 #include "clang/StaticAnalyzer/Core/BugReporter/BugType.h" 22 #include "clang/StaticAnalyzer/Core/Checker.h" 23 #include "clang/StaticAnalyzer/Core/PathSensitive/AnalysisManager.h" 24 #include "llvm/ADT/SmallString.h" 25 #include "llvm/Support/MathExtras.h" 26 #include "llvm/Support/raw_ostream.h" 27 #include <numeric> 28 29 using namespace clang; 30 using namespace ento; 31 32 namespace { 33 class PaddingChecker : public Checker<check::ASTDecl<TranslationUnitDecl>> { 34 private: 35 mutable std::unique_ptr<BugType> PaddingBug; 36 mutable int64_t AllowedPad; 37 mutable BugReporter *BR; 38 39 public: 40 void checkASTDecl(const TranslationUnitDecl *TUD, AnalysisManager &MGR, 41 BugReporter &BRArg) const { 42 BR = &BRArg; 43 AllowedPad = 44 MGR.getAnalyzerOptions() 45 .getCheckerIntegerOption("AllowedPad", 24, this); 46 assert(AllowedPad >= 0 && "AllowedPad option should be non-negative"); 47 48 // The calls to checkAST* from AnalysisConsumer don't 49 // visit template instantiations or lambda classes. We 50 // want to visit those, so we make our own RecursiveASTVisitor. 51 struct LocalVisitor : public RecursiveASTVisitor<LocalVisitor> { 52 const PaddingChecker *Checker; 53 bool shouldVisitTemplateInstantiations() const { return true; } 54 bool shouldVisitImplicitCode() const { return true; } 55 explicit LocalVisitor(const PaddingChecker *Checker) : Checker(Checker) {} 56 bool VisitRecordDecl(const RecordDecl *RD) { 57 Checker->visitRecord(RD); 58 return true; 59 } 60 bool VisitVarDecl(const VarDecl *VD) { 61 Checker->visitVariable(VD); 62 return true; 63 } 64 // TODO: Visit array new and mallocs for arrays. 65 }; 66 67 LocalVisitor visitor(this); 68 visitor.TraverseDecl(const_cast<TranslationUnitDecl *>(TUD)); 69 } 70 71 /// Look for records of overly padded types. If padding * 72 /// PadMultiplier exceeds AllowedPad, then generate a report. 73 /// PadMultiplier is used to share code with the array padding 74 /// checker. 75 void visitRecord(const RecordDecl *RD, uint64_t PadMultiplier = 1) const { 76 if (shouldSkipDecl(RD)) 77 return; 78 79 // TODO: Figure out why we are going through declarations and not only 80 // definitions. 81 if (!(RD = RD->getDefinition())) 82 return; 83 84 // This is the simplest correct case: a class with no fields and one base 85 // class. Other cases are more complicated because of how the base classes 86 // & fields might interact, so we don't bother dealing with them. 87 // TODO: Support other combinations of base classes and fields. 88 if (auto *CXXRD = dyn_cast<CXXRecordDecl>(RD)) 89 if (CXXRD->field_empty() && CXXRD->getNumBases() == 1) 90 return visitRecord(CXXRD->bases().begin()->getType()->getAsRecordDecl(), 91 PadMultiplier); 92 93 auto &ASTContext = RD->getASTContext(); 94 const ASTRecordLayout &RL = ASTContext.getASTRecordLayout(RD); 95 assert(llvm::isPowerOf2_64(RL.getAlignment().getQuantity())); 96 97 CharUnits BaselinePad = calculateBaselinePad(RD, ASTContext, RL); 98 if (BaselinePad.isZero()) 99 return; 100 101 CharUnits OptimalPad; 102 SmallVector<const FieldDecl *, 20> OptimalFieldsOrder; 103 std::tie(OptimalPad, OptimalFieldsOrder) = 104 calculateOptimalPad(RD, ASTContext, RL); 105 106 CharUnits DiffPad = PadMultiplier * (BaselinePad - OptimalPad); 107 if (DiffPad.getQuantity() <= AllowedPad) { 108 assert(!DiffPad.isNegative() && "DiffPad should not be negative"); 109 // There is not enough excess padding to trigger a warning. 110 return; 111 } 112 reportRecord(RD, BaselinePad, OptimalPad, OptimalFieldsOrder); 113 } 114 115 /// Look for arrays of overly padded types. If the padding of the 116 /// array type exceeds AllowedPad, then generate a report. 117 void visitVariable(const VarDecl *VD) const { 118 const ArrayType *ArrTy = VD->getType()->getAsArrayTypeUnsafe(); 119 if (ArrTy == nullptr) 120 return; 121 uint64_t Elts = 0; 122 if (const ConstantArrayType *CArrTy = dyn_cast<ConstantArrayType>(ArrTy)) 123 Elts = CArrTy->getSize().getZExtValue(); 124 if (Elts == 0) 125 return; 126 const RecordType *RT = ArrTy->getElementType()->getAs<RecordType>(); 127 if (RT == nullptr) 128 return; 129 130 // TODO: Recurse into the fields to see if they have excess padding. 131 visitRecord(RT->getDecl(), Elts); 132 } 133 134 bool shouldSkipDecl(const RecordDecl *RD) const { 135 // TODO: Figure out why we are going through declarations and not only 136 // definitions. 137 if (!(RD = RD->getDefinition())) 138 return true; 139 auto Location = RD->getLocation(); 140 // If the construct doesn't have a source file, then it's not something 141 // we want to diagnose. 142 if (!Location.isValid()) 143 return true; 144 SrcMgr::CharacteristicKind Kind = 145 BR->getSourceManager().getFileCharacteristic(Location); 146 // Throw out all records that come from system headers. 147 if (Kind != SrcMgr::C_User) 148 return true; 149 150 // Not going to attempt to optimize unions. 151 if (RD->isUnion()) 152 return true; 153 if (auto *CXXRD = dyn_cast<CXXRecordDecl>(RD)) { 154 // Tail padding with base classes ends up being very complicated. 155 // We will skip objects with base classes for now, unless they do not 156 // have fields. 157 // TODO: Handle more base class scenarios. 158 if (!CXXRD->field_empty() && CXXRD->getNumBases() != 0) 159 return true; 160 if (CXXRD->field_empty() && CXXRD->getNumBases() != 1) 161 return true; 162 // Virtual bases are complicated, skipping those for now. 163 if (CXXRD->getNumVBases() != 0) 164 return true; 165 // Can't layout a template, so skip it. We do still layout the 166 // instantiations though. 167 if (CXXRD->getTypeForDecl()->isDependentType()) 168 return true; 169 if (CXXRD->getTypeForDecl()->isInstantiationDependentType()) 170 return true; 171 } 172 // How do you reorder fields if you haven't got any? 173 else if (RD->field_empty()) 174 return true; 175 176 auto IsTrickyField = [](const FieldDecl *FD) -> bool { 177 // Bitfield layout is hard. 178 if (FD->isBitField()) 179 return true; 180 181 // Variable length arrays are tricky too. 182 QualType Ty = FD->getType(); 183 if (Ty->isIncompleteArrayType()) 184 return true; 185 return false; 186 }; 187 188 if (std::any_of(RD->field_begin(), RD->field_end(), IsTrickyField)) 189 return true; 190 return false; 191 } 192 193 static CharUnits calculateBaselinePad(const RecordDecl *RD, 194 const ASTContext &ASTContext, 195 const ASTRecordLayout &RL) { 196 CharUnits PaddingSum; 197 CharUnits Offset = ASTContext.toCharUnitsFromBits(RL.getFieldOffset(0)); 198 for (const FieldDecl *FD : RD->fields()) { 199 // This checker only cares about the padded size of the 200 // field, and not the data size. If the field is a record 201 // with tail padding, then we won't put that number in our 202 // total because reordering fields won't fix that problem. 203 CharUnits FieldSize = ASTContext.getTypeSizeInChars(FD->getType()); 204 auto FieldOffsetBits = RL.getFieldOffset(FD->getFieldIndex()); 205 CharUnits FieldOffset = ASTContext.toCharUnitsFromBits(FieldOffsetBits); 206 PaddingSum += (FieldOffset - Offset); 207 Offset = FieldOffset + FieldSize; 208 } 209 PaddingSum += RL.getSize() - Offset; 210 return PaddingSum; 211 } 212 213 /// Optimal padding overview: 214 /// 1. Find a close approximation to where we can place our first field. 215 /// This will usually be at offset 0. 216 /// 2. Try to find the best field that can legally be placed at the current 217 /// offset. 218 /// a. "Best" is the largest alignment that is legal, but smallest size. 219 /// This is to account for overly aligned types. 220 /// 3. If no fields can fit, pad by rounding the current offset up to the 221 /// smallest alignment requirement of our fields. Measure and track the 222 // amount of padding added. Go back to 2. 223 /// 4. Increment the current offset by the size of the chosen field. 224 /// 5. Remove the chosen field from the set of future possibilities. 225 /// 6. Go back to 2 if there are still unplaced fields. 226 /// 7. Add tail padding by rounding the current offset up to the structure 227 /// alignment. Track the amount of padding added. 228 229 static std::pair<CharUnits, SmallVector<const FieldDecl *, 20>> 230 calculateOptimalPad(const RecordDecl *RD, const ASTContext &ASTContext, 231 const ASTRecordLayout &RL) { 232 struct FieldInfo { 233 CharUnits Align; 234 CharUnits Size; 235 const FieldDecl *Field; 236 bool operator<(const FieldInfo &RHS) const { 237 // Order from small alignments to large alignments, 238 // then large sizes to small sizes. 239 // then large field indices to small field indices 240 return std::make_tuple(Align, -Size, 241 Field ? -static_cast<int>(Field->getFieldIndex()) 242 : 0) < 243 std::make_tuple( 244 RHS.Align, -RHS.Size, 245 RHS.Field ? -static_cast<int>(RHS.Field->getFieldIndex()) 246 : 0); 247 } 248 }; 249 SmallVector<FieldInfo, 20> Fields; 250 auto GatherSizesAndAlignments = [](const FieldDecl *FD) { 251 FieldInfo RetVal; 252 RetVal.Field = FD; 253 auto &Ctx = FD->getASTContext(); 254 std::tie(RetVal.Size, RetVal.Align) = 255 Ctx.getTypeInfoInChars(FD->getType()); 256 assert(llvm::isPowerOf2_64(RetVal.Align.getQuantity())); 257 if (auto Max = FD->getMaxAlignment()) 258 RetVal.Align = std::max(Ctx.toCharUnitsFromBits(Max), RetVal.Align); 259 return RetVal; 260 }; 261 std::transform(RD->field_begin(), RD->field_end(), 262 std::back_inserter(Fields), GatherSizesAndAlignments); 263 llvm::sort(Fields); 264 // This lets us skip over vptrs and non-virtual bases, 265 // so that we can just worry about the fields in our object. 266 // Note that this does cause us to miss some cases where we 267 // could pack more bytes in to a base class's tail padding. 268 CharUnits NewOffset = ASTContext.toCharUnitsFromBits(RL.getFieldOffset(0)); 269 CharUnits NewPad; 270 SmallVector<const FieldDecl *, 20> OptimalFieldsOrder; 271 while (!Fields.empty()) { 272 unsigned TrailingZeros = 273 llvm::countTrailingZeros((unsigned long long)NewOffset.getQuantity()); 274 // If NewOffset is zero, then countTrailingZeros will be 64. Shifting 275 // 64 will overflow our unsigned long long. Shifting 63 will turn 276 // our long long (and CharUnits internal type) negative. So shift 62. 277 long long CurAlignmentBits = 1ull << (std::min)(TrailingZeros, 62u); 278 CharUnits CurAlignment = CharUnits::fromQuantity(CurAlignmentBits); 279 FieldInfo InsertPoint = {CurAlignment, CharUnits::Zero(), nullptr}; 280 auto CurBegin = Fields.begin(); 281 auto CurEnd = Fields.end(); 282 283 // In the typical case, this will find the last element 284 // of the vector. We won't find a middle element unless 285 // we started on a poorly aligned address or have an overly 286 // aligned field. 287 auto Iter = std::upper_bound(CurBegin, CurEnd, InsertPoint); 288 if (Iter != CurBegin) { 289 // We found a field that we can layout with the current alignment. 290 --Iter; 291 NewOffset += Iter->Size; 292 OptimalFieldsOrder.push_back(Iter->Field); 293 Fields.erase(Iter); 294 } else { 295 // We are poorly aligned, and we need to pad in order to layout another 296 // field. Round up to at least the smallest field alignment that we 297 // currently have. 298 CharUnits NextOffset = NewOffset.alignTo(Fields[0].Align); 299 NewPad += NextOffset - NewOffset; 300 NewOffset = NextOffset; 301 } 302 } 303 // Calculate tail padding. 304 CharUnits NewSize = NewOffset.alignTo(RL.getAlignment()); 305 NewPad += NewSize - NewOffset; 306 return {NewPad, std::move(OptimalFieldsOrder)}; 307 } 308 309 void reportRecord( 310 const RecordDecl *RD, CharUnits BaselinePad, CharUnits OptimalPad, 311 const SmallVector<const FieldDecl *, 20> &OptimalFieldsOrder) const { 312 if (!PaddingBug) 313 PaddingBug = 314 llvm::make_unique<BugType>(this, "Excessive Padding", "Performance"); 315 316 SmallString<100> Buf; 317 llvm::raw_svector_ostream Os(Buf); 318 Os << "Excessive padding in '"; 319 Os << QualType::getAsString(RD->getTypeForDecl(), Qualifiers(), 320 LangOptions()) 321 << "'"; 322 323 if (auto *TSD = dyn_cast<ClassTemplateSpecializationDecl>(RD)) { 324 // TODO: make this show up better in the console output and in 325 // the HTML. Maybe just make it show up in HTML like the path 326 // diagnostics show. 327 SourceLocation ILoc = TSD->getPointOfInstantiation(); 328 if (ILoc.isValid()) 329 Os << " instantiated here: " 330 << ILoc.printToString(BR->getSourceManager()); 331 } 332 333 Os << " (" << BaselinePad.getQuantity() << " padding bytes, where " 334 << OptimalPad.getQuantity() << " is optimal). \n" 335 << "Optimal fields order: \n"; 336 for (const auto *FD : OptimalFieldsOrder) 337 Os << FD->getName() << ", \n"; 338 Os << "consider reordering the fields or adding explicit padding " 339 "members."; 340 341 PathDiagnosticLocation CELoc = 342 PathDiagnosticLocation::create(RD, BR->getSourceManager()); 343 auto Report = llvm::make_unique<BugReport>(*PaddingBug, Os.str(), CELoc); 344 Report->setDeclWithIssue(RD); 345 Report->addRange(RD->getSourceRange()); 346 BR->emitReport(std::move(Report)); 347 } 348 }; 349 } // namespace 350 351 void ento::registerPaddingChecker(CheckerManager &Mgr) { 352 Mgr.registerChecker<PaddingChecker>(); 353 } 354