1 //===--- CGRecordLayoutBuilder.cpp - CGRecordLayout builder ----*- 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 // Builder implementation for CGRecordLayout objects. 11 // 12 //===----------------------------------------------------------------------===// 13 14 #include "CGRecordLayout.h" 15 #include "CGCXXABI.h" 16 #include "CodeGenTypes.h" 17 #include "clang/AST/ASTContext.h" 18 #include "clang/AST/Attr.h" 19 #include "clang/AST/CXXInheritance.h" 20 #include "clang/AST/DeclCXX.h" 21 #include "clang/AST/Expr.h" 22 #include "clang/AST/RecordLayout.h" 23 #include "clang/Frontend/CodeGenOptions.h" 24 #include "llvm/IR/DataLayout.h" 25 #include "llvm/IR/DerivedTypes.h" 26 #include "llvm/IR/Type.h" 27 #include "llvm/Support/Debug.h" 28 #include "llvm/Support/MathExtras.h" 29 #include "llvm/Support/raw_ostream.h" 30 using namespace clang; 31 using namespace CodeGen; 32 33 namespace { 34 /// The CGRecordLowering is responsible for lowering an ASTRecordLayout to an 35 /// llvm::Type. Some of the lowering is straightforward, some is not. Here we 36 /// detail some of the complexities and weirdnesses here. 37 /// * LLVM does not have unions - Unions can, in theory be represented by any 38 /// llvm::Type with correct size. We choose a field via a specific heuristic 39 /// and add padding if necessary. 40 /// * LLVM does not have bitfields - Bitfields are collected into contiguous 41 /// runs and allocated as a single storage type for the run. ASTRecordLayout 42 /// contains enough information to determine where the runs break. Microsoft 43 /// and Itanium follow different rules and use different codepaths. 44 /// * It is desired that, when possible, bitfields use the appropriate iN type 45 /// when lowered to llvm types. For example unsigned x : 24 gets lowered to 46 /// i24. This isn't always possible because i24 has storage size of 32 bit 47 /// and if it is possible to use that extra byte of padding we must use 48 /// [i8 x 3] instead of i24. The function clipTailPadding does this. 49 /// C++ examples that require clipping: 50 /// struct { int a : 24; char b; }; // a must be clipped, b goes at offset 3 51 /// struct A { int a : 24; }; // a must be clipped because a struct like B 52 // could exist: struct B : A { char b; }; // b goes at offset 3 53 /// * Clang ignores 0 sized bitfields and 0 sized bases but *not* zero sized 54 /// fields. The existing asserts suggest that LLVM assumes that *every* field 55 /// has an underlying storage type. Therefore empty structures containing 56 /// zero sized subobjects such as empty records or zero sized arrays still get 57 /// a zero sized (empty struct) storage type. 58 /// * Clang reads the complete type rather than the base type when generating 59 /// code to access fields. Bitfields in tail position with tail padding may 60 /// be clipped in the base class but not the complete class (we may discover 61 /// that the tail padding is not used in the complete class.) However, 62 /// because LLVM reads from the complete type it can generate incorrect code 63 /// if we do not clip the tail padding off of the bitfield in the complete 64 /// layout. This introduces a somewhat awkward extra unnecessary clip stage. 65 /// The location of the clip is stored internally as a sentinel of type 66 /// SCISSOR. If LLVM were updated to read base types (which it probably 67 /// should because locations of things such as VBases are bogus in the llvm 68 /// type anyway) then we could eliminate the SCISSOR. 69 /// * Itanium allows nearly empty primary virtual bases. These bases don't get 70 /// get their own storage because they're laid out as part of another base 71 /// or at the beginning of the structure. Determining if a VBase actually 72 /// gets storage awkwardly involves a walk of all bases. 73 /// * VFPtrs and VBPtrs do *not* make a record NotZeroInitializable. 74 struct CGRecordLowering { 75 // MemberInfo is a helper structure that contains information about a record 76 // member. In additional to the standard member types, there exists a 77 // sentinel member type that ensures correct rounding. 78 struct MemberInfo { 79 CharUnits Offset; 80 enum InfoKind { VFPtr, VBPtr, Field, Base, VBase, Scissor } Kind; 81 llvm::Type *Data; 82 union { 83 const FieldDecl *FD; 84 const CXXRecordDecl *RD; 85 }; 86 MemberInfo(CharUnits Offset, InfoKind Kind, llvm::Type *Data, 87 const FieldDecl *FD = nullptr) 88 : Offset(Offset), Kind(Kind), Data(Data), FD(FD) {} 89 MemberInfo(CharUnits Offset, InfoKind Kind, llvm::Type *Data, 90 const CXXRecordDecl *RD) 91 : Offset(Offset), Kind(Kind), Data(Data), RD(RD) {} 92 // MemberInfos are sorted so we define a < operator. 93 bool operator <(const MemberInfo& a) const { return Offset < a.Offset; } 94 }; 95 // The constructor. 96 CGRecordLowering(CodeGenTypes &Types, const RecordDecl *D, bool Packed); 97 // Short helper routines. 98 /// Constructs a MemberInfo instance from an offset and llvm::Type *. 99 MemberInfo StorageInfo(CharUnits Offset, llvm::Type *Data) { 100 return MemberInfo(Offset, MemberInfo::Field, Data); 101 } 102 103 /// The Microsoft bitfield layout rule allocates discrete storage 104 /// units of the field's formal type and only combines adjacent 105 /// fields of the same formal type. We want to emit a layout with 106 /// these discrete storage units instead of combining them into a 107 /// continuous run. 108 bool isDiscreteBitFieldABI() { 109 return Context.getTargetInfo().getCXXABI().isMicrosoft() || 110 D->isMsStruct(Context); 111 } 112 113 /// The Itanium base layout rule allows virtual bases to overlap 114 /// other bases, which complicates layout in specific ways. 115 /// 116 /// Note specifically that the ms_struct attribute doesn't change this. 117 bool isOverlappingVBaseABI() { 118 return !Context.getTargetInfo().getCXXABI().isMicrosoft(); 119 } 120 121 /// Wraps llvm::Type::getIntNTy with some implicit arguments. 122 llvm::Type *getIntNType(uint64_t NumBits) { 123 return llvm::Type::getIntNTy(Types.getLLVMContext(), 124 (unsigned)llvm::alignTo(NumBits, 8)); 125 } 126 /// Gets an llvm type of size NumBytes and alignment 1. 127 llvm::Type *getByteArrayType(CharUnits NumBytes) { 128 assert(!NumBytes.isZero() && "Empty byte arrays aren't allowed."); 129 llvm::Type *Type = llvm::Type::getInt8Ty(Types.getLLVMContext()); 130 return NumBytes == CharUnits::One() ? Type : 131 (llvm::Type *)llvm::ArrayType::get(Type, NumBytes.getQuantity()); 132 } 133 /// Gets the storage type for a field decl and handles storage 134 /// for itanium bitfields that are smaller than their declared type. 135 llvm::Type *getStorageType(const FieldDecl *FD) { 136 llvm::Type *Type = Types.ConvertTypeForMem(FD->getType()); 137 if (!FD->isBitField()) return Type; 138 if (isDiscreteBitFieldABI()) return Type; 139 return getIntNType(std::min(FD->getBitWidthValue(Context), 140 (unsigned)Context.toBits(getSize(Type)))); 141 } 142 /// Gets the llvm Basesubobject type from a CXXRecordDecl. 143 llvm::Type *getStorageType(const CXXRecordDecl *RD) { 144 return Types.getCGRecordLayout(RD).getBaseSubobjectLLVMType(); 145 } 146 CharUnits bitsToCharUnits(uint64_t BitOffset) { 147 return Context.toCharUnitsFromBits(BitOffset); 148 } 149 CharUnits getSize(llvm::Type *Type) { 150 return CharUnits::fromQuantity(DataLayout.getTypeAllocSize(Type)); 151 } 152 CharUnits getAlignment(llvm::Type *Type) { 153 return CharUnits::fromQuantity(DataLayout.getABITypeAlignment(Type)); 154 } 155 bool isZeroInitializable(const FieldDecl *FD) { 156 return Types.isZeroInitializable(FD->getType()); 157 } 158 bool isZeroInitializable(const RecordDecl *RD) { 159 return Types.isZeroInitializable(RD); 160 } 161 void appendPaddingBytes(CharUnits Size) { 162 if (!Size.isZero()) 163 FieldTypes.push_back(getByteArrayType(Size)); 164 } 165 uint64_t getFieldBitOffset(const FieldDecl *FD) { 166 return Layout.getFieldOffset(FD->getFieldIndex()); 167 } 168 // Layout routines. 169 void setBitFieldInfo(const FieldDecl *FD, CharUnits StartOffset, 170 llvm::Type *StorageType); 171 /// Lowers an ASTRecordLayout to a llvm type. 172 void lower(bool NonVirtualBaseType); 173 void lowerUnion(); 174 void accumulateFields(); 175 void accumulateBitFields(RecordDecl::field_iterator Field, 176 RecordDecl::field_iterator FieldEnd); 177 void accumulateBases(); 178 void accumulateVPtrs(); 179 void accumulateVBases(); 180 /// Recursively searches all of the bases to find out if a vbase is 181 /// not the primary vbase of some base class. 182 bool hasOwnStorage(const CXXRecordDecl *Decl, const CXXRecordDecl *Query); 183 void calculateZeroInit(); 184 /// Lowers bitfield storage types to I8 arrays for bitfields with tail 185 /// padding that is or can potentially be used. 186 void clipTailPadding(); 187 /// Determines if we need a packed llvm struct. 188 void determinePacked(bool NVBaseType); 189 /// Inserts padding everywhere it's needed. 190 void insertPadding(); 191 /// Fills out the structures that are ultimately consumed. 192 void fillOutputFields(); 193 // Input memoization fields. 194 CodeGenTypes &Types; 195 const ASTContext &Context; 196 const RecordDecl *D; 197 const CXXRecordDecl *RD; 198 const ASTRecordLayout &Layout; 199 const llvm::DataLayout &DataLayout; 200 // Helpful intermediate data-structures. 201 std::vector<MemberInfo> Members; 202 // Output fields, consumed by CodeGenTypes::ComputeRecordLayout. 203 SmallVector<llvm::Type *, 16> FieldTypes; 204 llvm::DenseMap<const FieldDecl *, unsigned> Fields; 205 llvm::DenseMap<const FieldDecl *, CGBitFieldInfo> BitFields; 206 llvm::DenseMap<const CXXRecordDecl *, unsigned> NonVirtualBases; 207 llvm::DenseMap<const CXXRecordDecl *, unsigned> VirtualBases; 208 bool IsZeroInitializable : 1; 209 bool IsZeroInitializableAsBase : 1; 210 bool Packed : 1; 211 private: 212 CGRecordLowering(const CGRecordLowering &) = delete; 213 void operator =(const CGRecordLowering &) = delete; 214 }; 215 } // namespace { 216 217 CGRecordLowering::CGRecordLowering(CodeGenTypes &Types, const RecordDecl *D, 218 bool Packed) 219 : Types(Types), Context(Types.getContext()), D(D), 220 RD(dyn_cast<CXXRecordDecl>(D)), 221 Layout(Types.getContext().getASTRecordLayout(D)), 222 DataLayout(Types.getDataLayout()), IsZeroInitializable(true), 223 IsZeroInitializableAsBase(true), Packed(Packed) {} 224 225 void CGRecordLowering::setBitFieldInfo( 226 const FieldDecl *FD, CharUnits StartOffset, llvm::Type *StorageType) { 227 CGBitFieldInfo &Info = BitFields[FD->getCanonicalDecl()]; 228 Info.IsSigned = FD->getType()->isSignedIntegerOrEnumerationType(); 229 Info.Offset = (unsigned)(getFieldBitOffset(FD) - Context.toBits(StartOffset)); 230 Info.Size = FD->getBitWidthValue(Context); 231 Info.StorageSize = (unsigned)DataLayout.getTypeAllocSizeInBits(StorageType); 232 Info.StorageOffset = StartOffset; 233 if (Info.Size > Info.StorageSize) 234 Info.Size = Info.StorageSize; 235 // Reverse the bit offsets for big endian machines. Because we represent 236 // a bitfield as a single large integer load, we can imagine the bits 237 // counting from the most-significant-bit instead of the 238 // least-significant-bit. 239 if (DataLayout.isBigEndian()) 240 Info.Offset = Info.StorageSize - (Info.Offset + Info.Size); 241 } 242 243 void CGRecordLowering::lower(bool NVBaseType) { 244 // The lowering process implemented in this function takes a variety of 245 // carefully ordered phases. 246 // 1) Store all members (fields and bases) in a list and sort them by offset. 247 // 2) Add a 1-byte capstone member at the Size of the structure. 248 // 3) Clip bitfield storages members if their tail padding is or might be 249 // used by another field or base. The clipping process uses the capstone 250 // by treating it as another object that occurs after the record. 251 // 4) Determine if the llvm-struct requires packing. It's important that this 252 // phase occur after clipping, because clipping changes the llvm type. 253 // This phase reads the offset of the capstone when determining packedness 254 // and updates the alignment of the capstone to be equal of the alignment 255 // of the record after doing so. 256 // 5) Insert padding everywhere it is needed. This phase requires 'Packed' to 257 // have been computed and needs to know the alignment of the record in 258 // order to understand if explicit tail padding is needed. 259 // 6) Remove the capstone, we don't need it anymore. 260 // 7) Determine if this record can be zero-initialized. This phase could have 261 // been placed anywhere after phase 1. 262 // 8) Format the complete list of members in a way that can be consumed by 263 // CodeGenTypes::ComputeRecordLayout. 264 CharUnits Size = NVBaseType ? Layout.getNonVirtualSize() : Layout.getSize(); 265 if (D->isUnion()) 266 return lowerUnion(); 267 accumulateFields(); 268 // RD implies C++. 269 if (RD) { 270 accumulateVPtrs(); 271 accumulateBases(); 272 if (Members.empty()) 273 return appendPaddingBytes(Size); 274 if (!NVBaseType) 275 accumulateVBases(); 276 } 277 std::stable_sort(Members.begin(), Members.end()); 278 Members.push_back(StorageInfo(Size, getIntNType(8))); 279 clipTailPadding(); 280 determinePacked(NVBaseType); 281 insertPadding(); 282 Members.pop_back(); 283 calculateZeroInit(); 284 fillOutputFields(); 285 } 286 287 void CGRecordLowering::lowerUnion() { 288 CharUnits LayoutSize = Layout.getSize(); 289 llvm::Type *StorageType = nullptr; 290 bool SeenNamedMember = false; 291 // Iterate through the fields setting bitFieldInfo and the Fields array. Also 292 // locate the "most appropriate" storage type. The heuristic for finding the 293 // storage type isn't necessary, the first (non-0-length-bitfield) field's 294 // type would work fine and be simpler but would be different than what we've 295 // been doing and cause lit tests to change. 296 for (const auto *Field : D->fields()) { 297 if (Field->isBitField()) { 298 if (Field->isZeroLengthBitField(Context)) 299 continue; 300 llvm::Type *FieldType = getStorageType(Field); 301 if (LayoutSize < getSize(FieldType)) 302 FieldType = getByteArrayType(LayoutSize); 303 setBitFieldInfo(Field, CharUnits::Zero(), FieldType); 304 } 305 Fields[Field->getCanonicalDecl()] = 0; 306 llvm::Type *FieldType = getStorageType(Field); 307 // Compute zero-initializable status. 308 // This union might not be zero initialized: it may contain a pointer to 309 // data member which might have some exotic initialization sequence. 310 // If this is the case, then we aught not to try and come up with a "better" 311 // type, it might not be very easy to come up with a Constant which 312 // correctly initializes it. 313 if (!SeenNamedMember) { 314 SeenNamedMember = Field->getIdentifier(); 315 if (!SeenNamedMember) 316 if (const auto *FieldRD = Field->getType()->getAsRecordDecl()) 317 SeenNamedMember = FieldRD->findFirstNamedDataMember(); 318 if (SeenNamedMember && !isZeroInitializable(Field)) { 319 IsZeroInitializable = IsZeroInitializableAsBase = false; 320 StorageType = FieldType; 321 } 322 } 323 // Because our union isn't zero initializable, we won't be getting a better 324 // storage type. 325 if (!IsZeroInitializable) 326 continue; 327 // Conditionally update our storage type if we've got a new "better" one. 328 if (!StorageType || 329 getAlignment(FieldType) > getAlignment(StorageType) || 330 (getAlignment(FieldType) == getAlignment(StorageType) && 331 getSize(FieldType) > getSize(StorageType))) 332 StorageType = FieldType; 333 } 334 // If we have no storage type just pad to the appropriate size and return. 335 if (!StorageType) 336 return appendPaddingBytes(LayoutSize); 337 // If our storage size was bigger than our required size (can happen in the 338 // case of packed bitfields on Itanium) then just use an I8 array. 339 if (LayoutSize < getSize(StorageType)) 340 StorageType = getByteArrayType(LayoutSize); 341 FieldTypes.push_back(StorageType); 342 appendPaddingBytes(LayoutSize - getSize(StorageType)); 343 // Set packed if we need it. 344 if (LayoutSize % getAlignment(StorageType)) 345 Packed = true; 346 } 347 348 void CGRecordLowering::accumulateFields() { 349 for (RecordDecl::field_iterator Field = D->field_begin(), 350 FieldEnd = D->field_end(); 351 Field != FieldEnd;) 352 if (Field->isBitField()) { 353 RecordDecl::field_iterator Start = Field; 354 // Iterate to gather the list of bitfields. 355 for (++Field; Field != FieldEnd && Field->isBitField(); ++Field); 356 accumulateBitFields(Start, Field); 357 } else { 358 Members.push_back(MemberInfo( 359 bitsToCharUnits(getFieldBitOffset(*Field)), MemberInfo::Field, 360 getStorageType(*Field), *Field)); 361 ++Field; 362 } 363 } 364 365 void 366 CGRecordLowering::accumulateBitFields(RecordDecl::field_iterator Field, 367 RecordDecl::field_iterator FieldEnd) { 368 // Run stores the first element of the current run of bitfields. FieldEnd is 369 // used as a special value to note that we don't have a current run. A 370 // bitfield run is a contiguous collection of bitfields that can be stored in 371 // the same storage block. Zero-sized bitfields and bitfields that would 372 // cross an alignment boundary break a run and start a new one. 373 RecordDecl::field_iterator Run = FieldEnd; 374 // Tail is the offset of the first bit off the end of the current run. It's 375 // used to determine if the ASTRecordLayout is treating these two bitfields as 376 // contiguous. StartBitOffset is offset of the beginning of the Run. 377 uint64_t StartBitOffset, Tail = 0; 378 if (isDiscreteBitFieldABI()) { 379 for (; Field != FieldEnd; ++Field) { 380 uint64_t BitOffset = getFieldBitOffset(*Field); 381 // Zero-width bitfields end runs. 382 if (Field->isZeroLengthBitField(Context)) { 383 Run = FieldEnd; 384 continue; 385 } 386 llvm::Type *Type = Types.ConvertTypeForMem(Field->getType()); 387 // If we don't have a run yet, or don't live within the previous run's 388 // allocated storage then we allocate some storage and start a new run. 389 if (Run == FieldEnd || BitOffset >= Tail) { 390 Run = Field; 391 StartBitOffset = BitOffset; 392 Tail = StartBitOffset + DataLayout.getTypeAllocSizeInBits(Type); 393 // Add the storage member to the record. This must be added to the 394 // record before the bitfield members so that it gets laid out before 395 // the bitfields it contains get laid out. 396 Members.push_back(StorageInfo(bitsToCharUnits(StartBitOffset), Type)); 397 } 398 // Bitfields get the offset of their storage but come afterward and remain 399 // there after a stable sort. 400 Members.push_back(MemberInfo(bitsToCharUnits(StartBitOffset), 401 MemberInfo::Field, nullptr, *Field)); 402 } 403 return; 404 } 405 406 // Check if OffsetInRecord is better as a single field run. When OffsetInRecord 407 // has legal integer width, and its bitfield offset is naturally aligned, it 408 // is better to make the bitfield a separate storage component so as it can be 409 // accessed directly with lower cost. 410 auto IsBetterAsSingleFieldRun = [&](uint64_t OffsetInRecord, 411 uint64_t StartBitOffset) { 412 if (!Types.getCodeGenOpts().FineGrainedBitfieldAccesses) 413 return false; 414 if (!DataLayout.isLegalInteger(OffsetInRecord)) 415 return false; 416 // Make sure StartBitOffset is natually aligned if it is treated as an 417 // IType integer. 418 if (StartBitOffset % 419 Context.toBits(getAlignment(getIntNType(OffsetInRecord))) != 420 0) 421 return false; 422 return true; 423 }; 424 425 // The start field is better as a single field run. 426 bool StartFieldAsSingleRun = false; 427 for (;;) { 428 // Check to see if we need to start a new run. 429 if (Run == FieldEnd) { 430 // If we're out of fields, return. 431 if (Field == FieldEnd) 432 break; 433 // Any non-zero-length bitfield can start a new run. 434 if (!Field->isZeroLengthBitField(Context)) { 435 Run = Field; 436 StartBitOffset = getFieldBitOffset(*Field); 437 Tail = StartBitOffset + Field->getBitWidthValue(Context); 438 StartFieldAsSingleRun = IsBetterAsSingleFieldRun(Tail - StartBitOffset, 439 StartBitOffset); 440 } 441 ++Field; 442 continue; 443 } 444 445 // If the start field of a new run is better as a single run, or 446 // if current field (or consecutive fields) is better as a single run, or 447 // if current field has zero width bitfield and either 448 // UseZeroLengthBitfieldAlignment or UseBitFieldTypeAlignment is set to 449 // true, or 450 // if the offset of current field is inconsistent with the offset of 451 // previous field plus its offset, 452 // skip the block below and go ahead to emit the storage. 453 // Otherwise, try to add bitfields to the run. 454 if (!StartFieldAsSingleRun && Field != FieldEnd && 455 !IsBetterAsSingleFieldRun(Tail - StartBitOffset, StartBitOffset) && 456 (!Field->isZeroLengthBitField(Context) || 457 (!Context.getTargetInfo().useZeroLengthBitfieldAlignment() && 458 !Context.getTargetInfo().useBitFieldTypeAlignment())) && 459 Tail == getFieldBitOffset(*Field)) { 460 Tail += Field->getBitWidthValue(Context); 461 ++Field; 462 continue; 463 } 464 465 // We've hit a break-point in the run and need to emit a storage field. 466 llvm::Type *Type = getIntNType(Tail - StartBitOffset); 467 // Add the storage member to the record and set the bitfield info for all of 468 // the bitfields in the run. Bitfields get the offset of their storage but 469 // come afterward and remain there after a stable sort. 470 Members.push_back(StorageInfo(bitsToCharUnits(StartBitOffset), Type)); 471 for (; Run != Field; ++Run) 472 Members.push_back(MemberInfo(bitsToCharUnits(StartBitOffset), 473 MemberInfo::Field, nullptr, *Run)); 474 Run = FieldEnd; 475 StartFieldAsSingleRun = false; 476 } 477 } 478 479 void CGRecordLowering::accumulateBases() { 480 // If we've got a primary virtual base, we need to add it with the bases. 481 if (Layout.isPrimaryBaseVirtual()) { 482 const CXXRecordDecl *BaseDecl = Layout.getPrimaryBase(); 483 Members.push_back(MemberInfo(CharUnits::Zero(), MemberInfo::Base, 484 getStorageType(BaseDecl), BaseDecl)); 485 } 486 // Accumulate the non-virtual bases. 487 for (const auto &Base : RD->bases()) { 488 if (Base.isVirtual()) 489 continue; 490 491 // Bases can be zero-sized even if not technically empty if they 492 // contain only a trailing array member. 493 const CXXRecordDecl *BaseDecl = Base.getType()->getAsCXXRecordDecl(); 494 if (!BaseDecl->isEmpty() && 495 !Context.getASTRecordLayout(BaseDecl).getNonVirtualSize().isZero()) 496 Members.push_back(MemberInfo(Layout.getBaseClassOffset(BaseDecl), 497 MemberInfo::Base, getStorageType(BaseDecl), BaseDecl)); 498 } 499 } 500 501 void CGRecordLowering::accumulateVPtrs() { 502 if (Layout.hasOwnVFPtr()) 503 Members.push_back(MemberInfo(CharUnits::Zero(), MemberInfo::VFPtr, 504 llvm::FunctionType::get(getIntNType(32), /*isVarArg=*/true)-> 505 getPointerTo()->getPointerTo())); 506 if (Layout.hasOwnVBPtr()) 507 Members.push_back(MemberInfo(Layout.getVBPtrOffset(), MemberInfo::VBPtr, 508 llvm::Type::getInt32PtrTy(Types.getLLVMContext()))); 509 } 510 511 void CGRecordLowering::accumulateVBases() { 512 CharUnits ScissorOffset = Layout.getNonVirtualSize(); 513 // In the itanium ABI, it's possible to place a vbase at a dsize that is 514 // smaller than the nvsize. Here we check to see if such a base is placed 515 // before the nvsize and set the scissor offset to that, instead of the 516 // nvsize. 517 if (isOverlappingVBaseABI()) 518 for (const auto &Base : RD->vbases()) { 519 const CXXRecordDecl *BaseDecl = Base.getType()->getAsCXXRecordDecl(); 520 if (BaseDecl->isEmpty()) 521 continue; 522 // If the vbase is a primary virtual base of some base, then it doesn't 523 // get its own storage location but instead lives inside of that base. 524 if (Context.isNearlyEmpty(BaseDecl) && !hasOwnStorage(RD, BaseDecl)) 525 continue; 526 ScissorOffset = std::min(ScissorOffset, 527 Layout.getVBaseClassOffset(BaseDecl)); 528 } 529 Members.push_back(MemberInfo(ScissorOffset, MemberInfo::Scissor, nullptr, 530 RD)); 531 for (const auto &Base : RD->vbases()) { 532 const CXXRecordDecl *BaseDecl = Base.getType()->getAsCXXRecordDecl(); 533 if (BaseDecl->isEmpty()) 534 continue; 535 CharUnits Offset = Layout.getVBaseClassOffset(BaseDecl); 536 // If the vbase is a primary virtual base of some base, then it doesn't 537 // get its own storage location but instead lives inside of that base. 538 if (isOverlappingVBaseABI() && 539 Context.isNearlyEmpty(BaseDecl) && 540 !hasOwnStorage(RD, BaseDecl)) { 541 Members.push_back(MemberInfo(Offset, MemberInfo::VBase, nullptr, 542 BaseDecl)); 543 continue; 544 } 545 // If we've got a vtordisp, add it as a storage type. 546 if (Layout.getVBaseOffsetsMap().find(BaseDecl)->second.hasVtorDisp()) 547 Members.push_back(StorageInfo(Offset - CharUnits::fromQuantity(4), 548 getIntNType(32))); 549 Members.push_back(MemberInfo(Offset, MemberInfo::VBase, 550 getStorageType(BaseDecl), BaseDecl)); 551 } 552 } 553 554 bool CGRecordLowering::hasOwnStorage(const CXXRecordDecl *Decl, 555 const CXXRecordDecl *Query) { 556 const ASTRecordLayout &DeclLayout = Context.getASTRecordLayout(Decl); 557 if (DeclLayout.isPrimaryBaseVirtual() && DeclLayout.getPrimaryBase() == Query) 558 return false; 559 for (const auto &Base : Decl->bases()) 560 if (!hasOwnStorage(Base.getType()->getAsCXXRecordDecl(), Query)) 561 return false; 562 return true; 563 } 564 565 void CGRecordLowering::calculateZeroInit() { 566 for (std::vector<MemberInfo>::const_iterator Member = Members.begin(), 567 MemberEnd = Members.end(); 568 IsZeroInitializableAsBase && Member != MemberEnd; ++Member) { 569 if (Member->Kind == MemberInfo::Field) { 570 if (!Member->FD || isZeroInitializable(Member->FD)) 571 continue; 572 IsZeroInitializable = IsZeroInitializableAsBase = false; 573 } else if (Member->Kind == MemberInfo::Base || 574 Member->Kind == MemberInfo::VBase) { 575 if (isZeroInitializable(Member->RD)) 576 continue; 577 IsZeroInitializable = false; 578 if (Member->Kind == MemberInfo::Base) 579 IsZeroInitializableAsBase = false; 580 } 581 } 582 } 583 584 void CGRecordLowering::clipTailPadding() { 585 std::vector<MemberInfo>::iterator Prior = Members.begin(); 586 CharUnits Tail = getSize(Prior->Data); 587 for (std::vector<MemberInfo>::iterator Member = Prior + 1, 588 MemberEnd = Members.end(); 589 Member != MemberEnd; ++Member) { 590 // Only members with data and the scissor can cut into tail padding. 591 if (!Member->Data && Member->Kind != MemberInfo::Scissor) 592 continue; 593 if (Member->Offset < Tail) { 594 assert(Prior->Kind == MemberInfo::Field && !Prior->FD && 595 "Only storage fields have tail padding!"); 596 Prior->Data = getByteArrayType(bitsToCharUnits(llvm::alignTo( 597 cast<llvm::IntegerType>(Prior->Data)->getIntegerBitWidth(), 8))); 598 } 599 if (Member->Data) 600 Prior = Member; 601 Tail = Prior->Offset + getSize(Prior->Data); 602 } 603 } 604 605 void CGRecordLowering::determinePacked(bool NVBaseType) { 606 if (Packed) 607 return; 608 CharUnits Alignment = CharUnits::One(); 609 CharUnits NVAlignment = CharUnits::One(); 610 CharUnits NVSize = 611 !NVBaseType && RD ? Layout.getNonVirtualSize() : CharUnits::Zero(); 612 for (std::vector<MemberInfo>::const_iterator Member = Members.begin(), 613 MemberEnd = Members.end(); 614 Member != MemberEnd; ++Member) { 615 if (!Member->Data) 616 continue; 617 // If any member falls at an offset that it not a multiple of its alignment, 618 // then the entire record must be packed. 619 if (Member->Offset % getAlignment(Member->Data)) 620 Packed = true; 621 if (Member->Offset < NVSize) 622 NVAlignment = std::max(NVAlignment, getAlignment(Member->Data)); 623 Alignment = std::max(Alignment, getAlignment(Member->Data)); 624 } 625 // If the size of the record (the capstone's offset) is not a multiple of the 626 // record's alignment, it must be packed. 627 if (Members.back().Offset % Alignment) 628 Packed = true; 629 // If the non-virtual sub-object is not a multiple of the non-virtual 630 // sub-object's alignment, it must be packed. We cannot have a packed 631 // non-virtual sub-object and an unpacked complete object or vise versa. 632 if (NVSize % NVAlignment) 633 Packed = true; 634 // Update the alignment of the sentinel. 635 if (!Packed) 636 Members.back().Data = getIntNType(Context.toBits(Alignment)); 637 } 638 639 void CGRecordLowering::insertPadding() { 640 std::vector<std::pair<CharUnits, CharUnits> > Padding; 641 CharUnits Size = CharUnits::Zero(); 642 for (std::vector<MemberInfo>::const_iterator Member = Members.begin(), 643 MemberEnd = Members.end(); 644 Member != MemberEnd; ++Member) { 645 if (!Member->Data) 646 continue; 647 CharUnits Offset = Member->Offset; 648 assert(Offset >= Size); 649 // Insert padding if we need to. 650 if (Offset != 651 Size.alignTo(Packed ? CharUnits::One() : getAlignment(Member->Data))) 652 Padding.push_back(std::make_pair(Size, Offset - Size)); 653 Size = Offset + getSize(Member->Data); 654 } 655 if (Padding.empty()) 656 return; 657 // Add the padding to the Members list and sort it. 658 for (std::vector<std::pair<CharUnits, CharUnits> >::const_iterator 659 Pad = Padding.begin(), PadEnd = Padding.end(); 660 Pad != PadEnd; ++Pad) 661 Members.push_back(StorageInfo(Pad->first, getByteArrayType(Pad->second))); 662 std::stable_sort(Members.begin(), Members.end()); 663 } 664 665 void CGRecordLowering::fillOutputFields() { 666 for (std::vector<MemberInfo>::const_iterator Member = Members.begin(), 667 MemberEnd = Members.end(); 668 Member != MemberEnd; ++Member) { 669 if (Member->Data) 670 FieldTypes.push_back(Member->Data); 671 if (Member->Kind == MemberInfo::Field) { 672 if (Member->FD) 673 Fields[Member->FD->getCanonicalDecl()] = FieldTypes.size() - 1; 674 // A field without storage must be a bitfield. 675 if (!Member->Data) 676 setBitFieldInfo(Member->FD, Member->Offset, FieldTypes.back()); 677 } else if (Member->Kind == MemberInfo::Base) 678 NonVirtualBases[Member->RD] = FieldTypes.size() - 1; 679 else if (Member->Kind == MemberInfo::VBase) 680 VirtualBases[Member->RD] = FieldTypes.size() - 1; 681 } 682 } 683 684 CGBitFieldInfo CGBitFieldInfo::MakeInfo(CodeGenTypes &Types, 685 const FieldDecl *FD, 686 uint64_t Offset, uint64_t Size, 687 uint64_t StorageSize, 688 CharUnits StorageOffset) { 689 // This function is vestigial from CGRecordLayoutBuilder days but is still 690 // used in GCObjCRuntime.cpp. That usage has a "fixme" attached to it that 691 // when addressed will allow for the removal of this function. 692 llvm::Type *Ty = Types.ConvertTypeForMem(FD->getType()); 693 CharUnits TypeSizeInBytes = 694 CharUnits::fromQuantity(Types.getDataLayout().getTypeAllocSize(Ty)); 695 uint64_t TypeSizeInBits = Types.getContext().toBits(TypeSizeInBytes); 696 697 bool IsSigned = FD->getType()->isSignedIntegerOrEnumerationType(); 698 699 if (Size > TypeSizeInBits) { 700 // We have a wide bit-field. The extra bits are only used for padding, so 701 // if we have a bitfield of type T, with size N: 702 // 703 // T t : N; 704 // 705 // We can just assume that it's: 706 // 707 // T t : sizeof(T); 708 // 709 Size = TypeSizeInBits; 710 } 711 712 // Reverse the bit offsets for big endian machines. Because we represent 713 // a bitfield as a single large integer load, we can imagine the bits 714 // counting from the most-significant-bit instead of the 715 // least-significant-bit. 716 if (Types.getDataLayout().isBigEndian()) { 717 Offset = StorageSize - (Offset + Size); 718 } 719 720 return CGBitFieldInfo(Offset, Size, IsSigned, StorageSize, StorageOffset); 721 } 722 723 CGRecordLayout *CodeGenTypes::ComputeRecordLayout(const RecordDecl *D, 724 llvm::StructType *Ty) { 725 CGRecordLowering Builder(*this, D, /*Packed=*/false); 726 727 Builder.lower(/*NonVirtualBaseType=*/false); 728 729 // If we're in C++, compute the base subobject type. 730 llvm::StructType *BaseTy = nullptr; 731 if (isa<CXXRecordDecl>(D) && !D->isUnion() && !D->hasAttr<FinalAttr>()) { 732 BaseTy = Ty; 733 if (Builder.Layout.getNonVirtualSize() != Builder.Layout.getSize()) { 734 CGRecordLowering BaseBuilder(*this, D, /*Packed=*/Builder.Packed); 735 BaseBuilder.lower(/*NonVirtualBaseType=*/true); 736 BaseTy = llvm::StructType::create( 737 getLLVMContext(), BaseBuilder.FieldTypes, "", BaseBuilder.Packed); 738 addRecordTypeName(D, BaseTy, ".base"); 739 // BaseTy and Ty must agree on their packedness for getLLVMFieldNo to work 740 // on both of them with the same index. 741 assert(Builder.Packed == BaseBuilder.Packed && 742 "Non-virtual and complete types must agree on packedness"); 743 } 744 } 745 746 // Fill in the struct *after* computing the base type. Filling in the body 747 // signifies that the type is no longer opaque and record layout is complete, 748 // but we may need to recursively layout D while laying D out as a base type. 749 Ty->setBody(Builder.FieldTypes, Builder.Packed); 750 751 CGRecordLayout *RL = 752 new CGRecordLayout(Ty, BaseTy, Builder.IsZeroInitializable, 753 Builder.IsZeroInitializableAsBase); 754 755 RL->NonVirtualBases.swap(Builder.NonVirtualBases); 756 RL->CompleteObjectVirtualBases.swap(Builder.VirtualBases); 757 758 // Add all the field numbers. 759 RL->FieldInfo.swap(Builder.Fields); 760 761 // Add bitfield info. 762 RL->BitFields.swap(Builder.BitFields); 763 764 // Dump the layout, if requested. 765 if (getContext().getLangOpts().DumpRecordLayouts) { 766 llvm::outs() << "\n*** Dumping IRgen Record Layout\n"; 767 llvm::outs() << "Record: "; 768 D->dump(llvm::outs()); 769 llvm::outs() << "\nLayout: "; 770 RL->print(llvm::outs()); 771 } 772 773 #ifndef NDEBUG 774 // Verify that the computed LLVM struct size matches the AST layout size. 775 const ASTRecordLayout &Layout = getContext().getASTRecordLayout(D); 776 777 uint64_t TypeSizeInBits = getContext().toBits(Layout.getSize()); 778 assert(TypeSizeInBits == getDataLayout().getTypeAllocSizeInBits(Ty) && 779 "Type size mismatch!"); 780 781 if (BaseTy) { 782 CharUnits NonVirtualSize = Layout.getNonVirtualSize(); 783 784 uint64_t AlignedNonVirtualTypeSizeInBits = 785 getContext().toBits(NonVirtualSize); 786 787 assert(AlignedNonVirtualTypeSizeInBits == 788 getDataLayout().getTypeAllocSizeInBits(BaseTy) && 789 "Type size mismatch!"); 790 } 791 792 // Verify that the LLVM and AST field offsets agree. 793 llvm::StructType *ST = RL->getLLVMType(); 794 const llvm::StructLayout *SL = getDataLayout().getStructLayout(ST); 795 796 const ASTRecordLayout &AST_RL = getContext().getASTRecordLayout(D); 797 RecordDecl::field_iterator it = D->field_begin(); 798 for (unsigned i = 0, e = AST_RL.getFieldCount(); i != e; ++i, ++it) { 799 const FieldDecl *FD = *it; 800 801 // For non-bit-fields, just check that the LLVM struct offset matches the 802 // AST offset. 803 if (!FD->isBitField()) { 804 unsigned FieldNo = RL->getLLVMFieldNo(FD); 805 assert(AST_RL.getFieldOffset(i) == SL->getElementOffsetInBits(FieldNo) && 806 "Invalid field offset!"); 807 continue; 808 } 809 810 // Ignore unnamed bit-fields. 811 if (!FD->getDeclName()) 812 continue; 813 814 // Don't inspect zero-length bitfields. 815 if (FD->isZeroLengthBitField(getContext())) 816 continue; 817 818 const CGBitFieldInfo &Info = RL->getBitFieldInfo(FD); 819 llvm::Type *ElementTy = ST->getTypeAtIndex(RL->getLLVMFieldNo(FD)); 820 821 // Unions have overlapping elements dictating their layout, but for 822 // non-unions we can verify that this section of the layout is the exact 823 // expected size. 824 if (D->isUnion()) { 825 // For unions we verify that the start is zero and the size 826 // is in-bounds. However, on BE systems, the offset may be non-zero, but 827 // the size + offset should match the storage size in that case as it 828 // "starts" at the back. 829 if (getDataLayout().isBigEndian()) 830 assert(static_cast<unsigned>(Info.Offset + Info.Size) == 831 Info.StorageSize && 832 "Big endian union bitfield does not end at the back"); 833 else 834 assert(Info.Offset == 0 && 835 "Little endian union bitfield with a non-zero offset"); 836 assert(Info.StorageSize <= SL->getSizeInBits() && 837 "Union not large enough for bitfield storage"); 838 } else { 839 assert(Info.StorageSize == 840 getDataLayout().getTypeAllocSizeInBits(ElementTy) && 841 "Storage size does not match the element type size"); 842 } 843 assert(Info.Size > 0 && "Empty bitfield!"); 844 assert(static_cast<unsigned>(Info.Offset) + Info.Size <= Info.StorageSize && 845 "Bitfield outside of its allocated storage"); 846 } 847 #endif 848 849 return RL; 850 } 851 852 void CGRecordLayout::print(raw_ostream &OS) const { 853 OS << "<CGRecordLayout\n"; 854 OS << " LLVMType:" << *CompleteObjectType << "\n"; 855 if (BaseSubobjectType) 856 OS << " NonVirtualBaseLLVMType:" << *BaseSubobjectType << "\n"; 857 OS << " IsZeroInitializable:" << IsZeroInitializable << "\n"; 858 OS << " BitFields:[\n"; 859 860 // Print bit-field infos in declaration order. 861 std::vector<std::pair<unsigned, const CGBitFieldInfo*> > BFIs; 862 for (llvm::DenseMap<const FieldDecl*, CGBitFieldInfo>::const_iterator 863 it = BitFields.begin(), ie = BitFields.end(); 864 it != ie; ++it) { 865 const RecordDecl *RD = it->first->getParent(); 866 unsigned Index = 0; 867 for (RecordDecl::field_iterator 868 it2 = RD->field_begin(); *it2 != it->first; ++it2) 869 ++Index; 870 BFIs.push_back(std::make_pair(Index, &it->second)); 871 } 872 llvm::array_pod_sort(BFIs.begin(), BFIs.end()); 873 for (unsigned i = 0, e = BFIs.size(); i != e; ++i) { 874 OS.indent(4); 875 BFIs[i].second->print(OS); 876 OS << "\n"; 877 } 878 879 OS << "]>\n"; 880 } 881 882 LLVM_DUMP_METHOD void CGRecordLayout::dump() const { 883 print(llvm::errs()); 884 } 885 886 void CGBitFieldInfo::print(raw_ostream &OS) const { 887 OS << "<CGBitFieldInfo" 888 << " Offset:" << Offset 889 << " Size:" << Size 890 << " IsSigned:" << IsSigned 891 << " StorageSize:" << StorageSize 892 << " StorageOffset:" << StorageOffset.getQuantity() << ">"; 893 } 894 895 LLVM_DUMP_METHOD void CGBitFieldInfo::dump() const { 896 print(llvm::errs()); 897 } 898