1 //===--- CGRecordLayoutBuilder.cpp - CGRecordLayout builder ----*- C++ -*-===// 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 // Builder implementation for CGRecordLayout objects. 10 // 11 //===----------------------------------------------------------------------===// 12 13 #include "CGRecordLayout.h" 14 #include "CGCXXABI.h" 15 #include "CodeGenTypes.h" 16 #include "clang/AST/ASTContext.h" 17 #include "clang/AST/Attr.h" 18 #include "clang/AST/CXXInheritance.h" 19 #include "clang/AST/DeclCXX.h" 20 #include "clang/AST/Expr.h" 21 #include "clang/AST/RecordLayout.h" 22 #include "clang/Basic/CodeGenOptions.h" 23 #include "llvm/IR/DataLayout.h" 24 #include "llvm/IR/DerivedTypes.h" 25 #include "llvm/IR/Type.h" 26 #include "llvm/Support/Debug.h" 27 #include "llvm/Support/MathExtras.h" 28 #include "llvm/Support/raw_ostream.h" 29 using namespace clang; 30 using namespace CodeGen; 31 32 namespace { 33 /// The CGRecordLowering is responsible for lowering an ASTRecordLayout to an 34 /// llvm::Type. Some of the lowering is straightforward, some is not. Here we 35 /// detail some of the complexities and weirdnesses here. 36 /// * LLVM does not have unions - Unions can, in theory be represented by any 37 /// llvm::Type with correct size. We choose a field via a specific heuristic 38 /// and add padding if necessary. 39 /// * LLVM does not have bitfields - Bitfields are collected into contiguous 40 /// runs and allocated as a single storage type for the run. ASTRecordLayout 41 /// contains enough information to determine where the runs break. Microsoft 42 /// and Itanium follow different rules and use different codepaths. 43 /// * It is desired that, when possible, bitfields use the appropriate iN type 44 /// when lowered to llvm types. For example unsigned x : 24 gets lowered to 45 /// i24. This isn't always possible because i24 has storage size of 32 bit 46 /// and if it is possible to use that extra byte of padding we must use 47 /// [i8 x 3] instead of i24. The function clipTailPadding does this. 48 /// C++ examples that require clipping: 49 /// struct { int a : 24; char b; }; // a must be clipped, b goes at offset 3 50 /// struct A { int a : 24; }; // a must be clipped because a struct like B 51 // could exist: struct B : A { char b; }; // b goes at offset 3 52 /// * Clang ignores 0 sized bitfields and 0 sized bases but *not* zero sized 53 /// fields. The existing asserts suggest that LLVM assumes that *every* field 54 /// has an underlying storage type. Therefore empty structures containing 55 /// zero sized subobjects such as empty records or zero sized arrays still get 56 /// a zero sized (empty struct) storage type. 57 /// * Clang reads the complete type rather than the base type when generating 58 /// code to access fields. Bitfields in tail position with tail padding may 59 /// be clipped in the base class but not the complete class (we may discover 60 /// that the tail padding is not used in the complete class.) However, 61 /// because LLVM reads from the complete type it can generate incorrect code 62 /// if we do not clip the tail padding off of the bitfield in the complete 63 /// layout. This introduces a somewhat awkward extra unnecessary clip stage. 64 /// The location of the clip is stored internally as a sentinel of type 65 /// SCISSOR. If LLVM were updated to read base types (which it probably 66 /// should because locations of things such as VBases are bogus in the llvm 67 /// type anyway) then we could eliminate the SCISSOR. 68 /// * Itanium allows nearly empty primary virtual bases. These bases don't get 69 /// get their own storage because they're laid out as part of another base 70 /// or at the beginning of the structure. Determining if a VBase actually 71 /// gets storage awkwardly involves a walk of all bases. 72 /// * VFPtrs and VBPtrs do *not* make a record NotZeroInitializable. 73 struct CGRecordLowering { 74 // MemberInfo is a helper structure that contains information about a record 75 // member. In additional to the standard member types, there exists a 76 // sentinel member type that ensures correct rounding. 77 struct MemberInfo { 78 CharUnits Offset; 79 enum InfoKind { VFPtr, VBPtr, Field, Base, VBase, Scissor } Kind; 80 llvm::Type *Data; 81 union { 82 const FieldDecl *FD; 83 const CXXRecordDecl *RD; 84 }; 85 MemberInfo(CharUnits Offset, InfoKind Kind, llvm::Type *Data, 86 const FieldDecl *FD = nullptr) 87 : Offset(Offset), Kind(Kind), Data(Data), FD(FD) {} 88 MemberInfo(CharUnits Offset, InfoKind Kind, llvm::Type *Data, 89 const CXXRecordDecl *RD) 90 : Offset(Offset), Kind(Kind), Data(Data), RD(RD) {} 91 // MemberInfos are sorted so we define a < operator. 92 bool operator <(const MemberInfo& a) const { return Offset < a.Offset; } 93 }; 94 // The constructor. 95 CGRecordLowering(CodeGenTypes &Types, const RecordDecl *D, bool Packed); 96 // Short helper routines. 97 /// Constructs a MemberInfo instance from an offset and llvm::Type *. 98 MemberInfo StorageInfo(CharUnits Offset, llvm::Type *Data) { 99 return MemberInfo(Offset, MemberInfo::Field, Data); 100 } 101 102 /// The Microsoft bitfield layout rule allocates discrete storage 103 /// units of the field's formal type and only combines adjacent 104 /// fields of the same formal type. We want to emit a layout with 105 /// these discrete storage units instead of combining them into a 106 /// continuous run. 107 bool isDiscreteBitFieldABI() { 108 return Context.getTargetInfo().getCXXABI().isMicrosoft() || 109 D->isMsStruct(Context); 110 } 111 112 /// The Itanium base layout rule allows virtual bases to overlap 113 /// other bases, which complicates layout in specific ways. 114 /// 115 /// Note specifically that the ms_struct attribute doesn't change this. 116 bool isOverlappingVBaseABI() { 117 return !Context.getTargetInfo().getCXXABI().isMicrosoft(); 118 } 119 120 /// Wraps llvm::Type::getIntNTy with some implicit arguments. 121 llvm::Type *getIntNType(uint64_t NumBits) { 122 return llvm::Type::getIntNTy(Types.getLLVMContext(), 123 (unsigned)llvm::alignTo(NumBits, 8)); 124 } 125 /// Gets an llvm type of size NumBytes and alignment 1. 126 llvm::Type *getByteArrayType(CharUnits NumBytes) { 127 assert(!NumBytes.isZero() && "Empty byte arrays aren't allowed."); 128 llvm::Type *Type = llvm::Type::getInt8Ty(Types.getLLVMContext()); 129 return NumBytes == CharUnits::One() ? Type : 130 (llvm::Type *)llvm::ArrayType::get(Type, NumBytes.getQuantity()); 131 } 132 /// Gets the storage type for a field decl and handles storage 133 /// for itanium bitfields that are smaller than their declared type. 134 llvm::Type *getStorageType(const FieldDecl *FD) { 135 llvm::Type *Type = Types.ConvertTypeForMem(FD->getType()); 136 if (!FD->isBitField()) return Type; 137 if (isDiscreteBitFieldABI()) return Type; 138 return getIntNType(std::min(FD->getBitWidthValue(Context), 139 (unsigned)Context.toBits(getSize(Type)))); 140 } 141 /// Gets the llvm Basesubobject type from a CXXRecordDecl. 142 llvm::Type *getStorageType(const CXXRecordDecl *RD) { 143 return Types.getCGRecordLayout(RD).getBaseSubobjectLLVMType(); 144 } 145 CharUnits bitsToCharUnits(uint64_t BitOffset) { 146 return Context.toCharUnitsFromBits(BitOffset); 147 } 148 CharUnits getSize(llvm::Type *Type) { 149 return CharUnits::fromQuantity(DataLayout.getTypeAllocSize(Type)); 150 } 151 CharUnits getAlignment(llvm::Type *Type) { 152 return CharUnits::fromQuantity(DataLayout.getABITypeAlignment(Type)); 153 } 154 bool isZeroInitializable(const FieldDecl *FD) { 155 return Types.isZeroInitializable(FD->getType()); 156 } 157 bool isZeroInitializable(const RecordDecl *RD) { 158 return Types.isZeroInitializable(RD); 159 } 160 void appendPaddingBytes(CharUnits Size) { 161 if (!Size.isZero()) 162 FieldTypes.push_back(getByteArrayType(Size)); 163 } 164 uint64_t getFieldBitOffset(const FieldDecl *FD) { 165 return Layout.getFieldOffset(FD->getFieldIndex()); 166 } 167 // Layout routines. 168 void setBitFieldInfo(const FieldDecl *FD, CharUnits StartOffset, 169 llvm::Type *StorageType); 170 /// Lowers an ASTRecordLayout to a llvm type. 171 void lower(bool NonVirtualBaseType); 172 void lowerUnion(); 173 void accumulateFields(); 174 void accumulateBitFields(RecordDecl::field_iterator Field, 175 RecordDecl::field_iterator FieldEnd); 176 void accumulateBases(); 177 void accumulateVPtrs(); 178 void accumulateVBases(); 179 /// Recursively searches all of the bases to find out if a vbase is 180 /// not the primary vbase of some base class. 181 bool hasOwnStorage(const CXXRecordDecl *Decl, const CXXRecordDecl *Query); 182 void calculateZeroInit(); 183 /// Lowers bitfield storage types to I8 arrays for bitfields with tail 184 /// padding that is or can potentially be used. 185 void clipTailPadding(); 186 /// Determines if we need a packed llvm struct. 187 void determinePacked(bool NVBaseType); 188 /// Inserts padding everywhere it's needed. 189 void insertPadding(); 190 /// Fills out the structures that are ultimately consumed. 191 void fillOutputFields(); 192 // Input memoization fields. 193 CodeGenTypes &Types; 194 const ASTContext &Context; 195 const RecordDecl *D; 196 const CXXRecordDecl *RD; 197 const ASTRecordLayout &Layout; 198 const llvm::DataLayout &DataLayout; 199 // Helpful intermediate data-structures. 200 std::vector<MemberInfo> Members; 201 // Output fields, consumed by CodeGenTypes::ComputeRecordLayout. 202 SmallVector<llvm::Type *, 16> FieldTypes; 203 llvm::DenseMap<const FieldDecl *, unsigned> Fields; 204 llvm::DenseMap<const FieldDecl *, CGBitFieldInfo> BitFields; 205 llvm::DenseMap<const CXXRecordDecl *, unsigned> NonVirtualBases; 206 llvm::DenseMap<const CXXRecordDecl *, unsigned> VirtualBases; 207 bool IsZeroInitializable : 1; 208 bool IsZeroInitializableAsBase : 1; 209 bool Packed : 1; 210 private: 211 CGRecordLowering(const CGRecordLowering &) = delete; 212 void operator =(const CGRecordLowering &) = delete; 213 }; 214 } // namespace { 215 216 CGRecordLowering::CGRecordLowering(CodeGenTypes &Types, const RecordDecl *D, 217 bool Packed) 218 : Types(Types), Context(Types.getContext()), D(D), 219 RD(dyn_cast<CXXRecordDecl>(D)), 220 Layout(Types.getContext().getASTRecordLayout(D)), 221 DataLayout(Types.getDataLayout()), IsZeroInitializable(true), 222 IsZeroInitializableAsBase(true), Packed(Packed) {} 223 224 void CGRecordLowering::setBitFieldInfo( 225 const FieldDecl *FD, CharUnits StartOffset, llvm::Type *StorageType) { 226 CGBitFieldInfo &Info = BitFields[FD->getCanonicalDecl()]; 227 Info.IsSigned = FD->getType()->isSignedIntegerOrEnumerationType(); 228 Info.Offset = (unsigned)(getFieldBitOffset(FD) - Context.toBits(StartOffset)); 229 Info.Size = FD->getBitWidthValue(Context); 230 Info.StorageSize = (unsigned)DataLayout.getTypeAllocSizeInBits(StorageType); 231 Info.StorageOffset = StartOffset; 232 if (Info.Size > Info.StorageSize) 233 Info.Size = Info.StorageSize; 234 // Reverse the bit offsets for big endian machines. Because we represent 235 // a bitfield as a single large integer load, we can imagine the bits 236 // counting from the most-significant-bit instead of the 237 // least-significant-bit. 238 if (DataLayout.isBigEndian()) 239 Info.Offset = Info.StorageSize - (Info.Offset + Info.Size); 240 } 241 242 void CGRecordLowering::lower(bool NVBaseType) { 243 // The lowering process implemented in this function takes a variety of 244 // carefully ordered phases. 245 // 1) Store all members (fields and bases) in a list and sort them by offset. 246 // 2) Add a 1-byte capstone member at the Size of the structure. 247 // 3) Clip bitfield storages members if their tail padding is or might be 248 // used by another field or base. The clipping process uses the capstone 249 // by treating it as another object that occurs after the record. 250 // 4) Determine if the llvm-struct requires packing. It's important that this 251 // phase occur after clipping, because clipping changes the llvm type. 252 // This phase reads the offset of the capstone when determining packedness 253 // and updates the alignment of the capstone to be equal of the alignment 254 // of the record after doing so. 255 // 5) Insert padding everywhere it is needed. This phase requires 'Packed' to 256 // have been computed and needs to know the alignment of the record in 257 // order to understand if explicit tail padding is needed. 258 // 6) Remove the capstone, we don't need it anymore. 259 // 7) Determine if this record can be zero-initialized. This phase could have 260 // been placed anywhere after phase 1. 261 // 8) Format the complete list of members in a way that can be consumed by 262 // CodeGenTypes::ComputeRecordLayout. 263 CharUnits Size = NVBaseType ? Layout.getNonVirtualSize() : Layout.getSize(); 264 if (D->isUnion()) 265 return lowerUnion(); 266 accumulateFields(); 267 // RD implies C++. 268 if (RD) { 269 accumulateVPtrs(); 270 accumulateBases(); 271 if (Members.empty()) 272 return appendPaddingBytes(Size); 273 if (!NVBaseType) 274 accumulateVBases(); 275 } 276 llvm::stable_sort(Members); 277 Members.push_back(StorageInfo(Size, getIntNType(8))); 278 clipTailPadding(); 279 determinePacked(NVBaseType); 280 insertPadding(); 281 Members.pop_back(); 282 calculateZeroInit(); 283 fillOutputFields(); 284 } 285 286 void CGRecordLowering::lowerUnion() { 287 CharUnits LayoutSize = Layout.getSize(); 288 llvm::Type *StorageType = nullptr; 289 bool SeenNamedMember = false; 290 // Iterate through the fields setting bitFieldInfo and the Fields array. Also 291 // locate the "most appropriate" storage type. The heuristic for finding the 292 // storage type isn't necessary, the first (non-0-length-bitfield) field's 293 // type would work fine and be simpler but would be different than what we've 294 // been doing and cause lit tests to change. 295 for (const auto *Field : D->fields()) { 296 if (Field->isBitField()) { 297 if (Field->isZeroLengthBitField(Context)) 298 continue; 299 llvm::Type *FieldType = getStorageType(Field); 300 if (LayoutSize < getSize(FieldType)) 301 FieldType = getByteArrayType(LayoutSize); 302 setBitFieldInfo(Field, CharUnits::Zero(), FieldType); 303 } 304 Fields[Field->getCanonicalDecl()] = 0; 305 llvm::Type *FieldType = getStorageType(Field); 306 // Compute zero-initializable status. 307 // This union might not be zero initialized: it may contain a pointer to 308 // data member which might have some exotic initialization sequence. 309 // If this is the case, then we aught not to try and come up with a "better" 310 // type, it might not be very easy to come up with a Constant which 311 // correctly initializes it. 312 if (!SeenNamedMember) { 313 SeenNamedMember = Field->getIdentifier(); 314 if (!SeenNamedMember) 315 if (const auto *FieldRD = Field->getType()->getAsRecordDecl()) 316 SeenNamedMember = FieldRD->findFirstNamedDataMember(); 317 if (SeenNamedMember && !isZeroInitializable(Field)) { 318 IsZeroInitializable = IsZeroInitializableAsBase = false; 319 StorageType = FieldType; 320 } 321 } 322 // Because our union isn't zero initializable, we won't be getting a better 323 // storage type. 324 if (!IsZeroInitializable) 325 continue; 326 // Conditionally update our storage type if we've got a new "better" one. 327 if (!StorageType || 328 getAlignment(FieldType) > getAlignment(StorageType) || 329 (getAlignment(FieldType) == getAlignment(StorageType) && 330 getSize(FieldType) > getSize(StorageType))) 331 StorageType = FieldType; 332 } 333 // If we have no storage type just pad to the appropriate size and return. 334 if (!StorageType) 335 return appendPaddingBytes(LayoutSize); 336 // If our storage size was bigger than our required size (can happen in the 337 // case of packed bitfields on Itanium) then just use an I8 array. 338 if (LayoutSize < getSize(StorageType)) 339 StorageType = getByteArrayType(LayoutSize); 340 FieldTypes.push_back(StorageType); 341 appendPaddingBytes(LayoutSize - getSize(StorageType)); 342 // Set packed if we need it. 343 if (LayoutSize % getAlignment(StorageType)) 344 Packed = true; 345 } 346 347 void CGRecordLowering::accumulateFields() { 348 for (RecordDecl::field_iterator Field = D->field_begin(), 349 FieldEnd = D->field_end(); 350 Field != FieldEnd;) { 351 if (Field->isBitField()) { 352 RecordDecl::field_iterator Start = Field; 353 // Iterate to gather the list of bitfields. 354 for (++Field; Field != FieldEnd && Field->isBitField(); ++Field); 355 accumulateBitFields(Start, Field); 356 } else if (!Field->isZeroSize(Context)) { 357 Members.push_back(MemberInfo( 358 bitsToCharUnits(getFieldBitOffset(*Field)), MemberInfo::Field, 359 getStorageType(*Field), *Field)); 360 ++Field; 361 } else { 362 ++Field; 363 } 364 } 365 } 366 367 void 368 CGRecordLowering::accumulateBitFields(RecordDecl::field_iterator Field, 369 RecordDecl::field_iterator FieldEnd) { 370 // Run stores the first element of the current run of bitfields. FieldEnd is 371 // used as a special value to note that we don't have a current run. A 372 // bitfield run is a contiguous collection of bitfields that can be stored in 373 // the same storage block. Zero-sized bitfields and bitfields that would 374 // cross an alignment boundary break a run and start a new one. 375 RecordDecl::field_iterator Run = FieldEnd; 376 // Tail is the offset of the first bit off the end of the current run. It's 377 // used to determine if the ASTRecordLayout is treating these two bitfields as 378 // contiguous. StartBitOffset is offset of the beginning of the Run. 379 uint64_t StartBitOffset, Tail = 0; 380 if (isDiscreteBitFieldABI()) { 381 for (; Field != FieldEnd; ++Field) { 382 uint64_t BitOffset = getFieldBitOffset(*Field); 383 // Zero-width bitfields end runs. 384 if (Field->isZeroLengthBitField(Context)) { 385 Run = FieldEnd; 386 continue; 387 } 388 llvm::Type *Type = 389 Types.ConvertTypeForMem(Field->getType(), /*ForBitFields=*/true); 390 // If we don't have a run yet, or don't live within the previous run's 391 // allocated storage then we allocate some storage and start a new run. 392 if (Run == FieldEnd || BitOffset >= Tail) { 393 Run = Field; 394 StartBitOffset = BitOffset; 395 Tail = StartBitOffset + DataLayout.getTypeAllocSizeInBits(Type); 396 // Add the storage member to the record. This must be added to the 397 // record before the bitfield members so that it gets laid out before 398 // the bitfields it contains get laid out. 399 Members.push_back(StorageInfo(bitsToCharUnits(StartBitOffset), Type)); 400 } 401 // Bitfields get the offset of their storage but come afterward and remain 402 // there after a stable sort. 403 Members.push_back(MemberInfo(bitsToCharUnits(StartBitOffset), 404 MemberInfo::Field, nullptr, *Field)); 405 } 406 return; 407 } 408 409 // Check if OffsetInRecord (the size in bits of the current run) is better 410 // as a single field run. When OffsetInRecord has legal integer width, and 411 // its bitfield offset is naturally aligned, it is better to make the 412 // bitfield a separate storage component so as it can be accessed directly 413 // with lower cost. 414 auto IsBetterAsSingleFieldRun = [&](uint64_t OffsetInRecord, 415 uint64_t StartBitOffset) { 416 if (!Types.getCodeGenOpts().FineGrainedBitfieldAccesses) 417 return false; 418 if (OffsetInRecord < 8 || !llvm::isPowerOf2_64(OffsetInRecord) || 419 !DataLayout.fitsInLegalInteger(OffsetInRecord)) 420 return false; 421 // Make sure StartBitOffset is natually aligned if it is treated as an 422 // IType integer. 423 if (StartBitOffset % 424 Context.toBits(getAlignment(getIntNType(OffsetInRecord))) != 425 0) 426 return false; 427 return true; 428 }; 429 430 // The start field is better as a single field run. 431 bool StartFieldAsSingleRun = false; 432 for (;;) { 433 // Check to see if we need to start a new run. 434 if (Run == FieldEnd) { 435 // If we're out of fields, return. 436 if (Field == FieldEnd) 437 break; 438 // Any non-zero-length bitfield can start a new run. 439 if (!Field->isZeroLengthBitField(Context)) { 440 Run = Field; 441 StartBitOffset = getFieldBitOffset(*Field); 442 Tail = StartBitOffset + Field->getBitWidthValue(Context); 443 StartFieldAsSingleRun = IsBetterAsSingleFieldRun(Tail - StartBitOffset, 444 StartBitOffset); 445 } 446 ++Field; 447 continue; 448 } 449 450 // If the start field of a new run is better as a single run, or 451 // if current field (or consecutive fields) is better as a single run, or 452 // if current field has zero width bitfield and either 453 // UseZeroLengthBitfieldAlignment or UseBitFieldTypeAlignment is set to 454 // true, or 455 // if the offset of current field is inconsistent with the offset of 456 // previous field plus its offset, 457 // skip the block below and go ahead to emit the storage. 458 // Otherwise, try to add bitfields to the run. 459 if (!StartFieldAsSingleRun && Field != FieldEnd && 460 !IsBetterAsSingleFieldRun(Tail - StartBitOffset, StartBitOffset) && 461 (!Field->isZeroLengthBitField(Context) || 462 (!Context.getTargetInfo().useZeroLengthBitfieldAlignment() && 463 !Context.getTargetInfo().useBitFieldTypeAlignment())) && 464 Tail == getFieldBitOffset(*Field)) { 465 Tail += Field->getBitWidthValue(Context); 466 ++Field; 467 continue; 468 } 469 470 // We've hit a break-point in the run and need to emit a storage field. 471 llvm::Type *Type = getIntNType(Tail - StartBitOffset); 472 // Add the storage member to the record and set the bitfield info for all of 473 // the bitfields in the run. Bitfields get the offset of their storage but 474 // come afterward and remain there after a stable sort. 475 Members.push_back(StorageInfo(bitsToCharUnits(StartBitOffset), Type)); 476 for (; Run != Field; ++Run) 477 Members.push_back(MemberInfo(bitsToCharUnits(StartBitOffset), 478 MemberInfo::Field, nullptr, *Run)); 479 Run = FieldEnd; 480 StartFieldAsSingleRun = false; 481 } 482 } 483 484 void CGRecordLowering::accumulateBases() { 485 // If we've got a primary virtual base, we need to add it with the bases. 486 if (Layout.isPrimaryBaseVirtual()) { 487 const CXXRecordDecl *BaseDecl = Layout.getPrimaryBase(); 488 Members.push_back(MemberInfo(CharUnits::Zero(), MemberInfo::Base, 489 getStorageType(BaseDecl), BaseDecl)); 490 } 491 // Accumulate the non-virtual bases. 492 for (const auto &Base : RD->bases()) { 493 if (Base.isVirtual()) 494 continue; 495 496 // Bases can be zero-sized even if not technically empty if they 497 // contain only a trailing array member. 498 const CXXRecordDecl *BaseDecl = Base.getType()->getAsCXXRecordDecl(); 499 if (!BaseDecl->isEmpty() && 500 !Context.getASTRecordLayout(BaseDecl).getNonVirtualSize().isZero()) 501 Members.push_back(MemberInfo(Layout.getBaseClassOffset(BaseDecl), 502 MemberInfo::Base, getStorageType(BaseDecl), BaseDecl)); 503 } 504 } 505 506 void CGRecordLowering::accumulateVPtrs() { 507 if (Layout.hasOwnVFPtr()) 508 Members.push_back(MemberInfo(CharUnits::Zero(), MemberInfo::VFPtr, 509 llvm::FunctionType::get(getIntNType(32), /*isVarArg=*/true)-> 510 getPointerTo()->getPointerTo())); 511 if (Layout.hasOwnVBPtr()) 512 Members.push_back(MemberInfo(Layout.getVBPtrOffset(), MemberInfo::VBPtr, 513 llvm::Type::getInt32PtrTy(Types.getLLVMContext()))); 514 } 515 516 void CGRecordLowering::accumulateVBases() { 517 CharUnits ScissorOffset = Layout.getNonVirtualSize(); 518 // In the itanium ABI, it's possible to place a vbase at a dsize that is 519 // smaller than the nvsize. Here we check to see if such a base is placed 520 // before the nvsize and set the scissor offset to that, instead of the 521 // nvsize. 522 if (isOverlappingVBaseABI()) 523 for (const auto &Base : RD->vbases()) { 524 const CXXRecordDecl *BaseDecl = Base.getType()->getAsCXXRecordDecl(); 525 if (BaseDecl->isEmpty()) 526 continue; 527 // If the vbase is a primary virtual base of some base, then it doesn't 528 // get its own storage location but instead lives inside of that base. 529 if (Context.isNearlyEmpty(BaseDecl) && !hasOwnStorage(RD, BaseDecl)) 530 continue; 531 ScissorOffset = std::min(ScissorOffset, 532 Layout.getVBaseClassOffset(BaseDecl)); 533 } 534 Members.push_back(MemberInfo(ScissorOffset, MemberInfo::Scissor, nullptr, 535 RD)); 536 for (const auto &Base : RD->vbases()) { 537 const CXXRecordDecl *BaseDecl = Base.getType()->getAsCXXRecordDecl(); 538 if (BaseDecl->isEmpty()) 539 continue; 540 CharUnits Offset = Layout.getVBaseClassOffset(BaseDecl); 541 // If the vbase is a primary virtual base of some base, then it doesn't 542 // get its own storage location but instead lives inside of that base. 543 if (isOverlappingVBaseABI() && 544 Context.isNearlyEmpty(BaseDecl) && 545 !hasOwnStorage(RD, BaseDecl)) { 546 Members.push_back(MemberInfo(Offset, MemberInfo::VBase, nullptr, 547 BaseDecl)); 548 continue; 549 } 550 // If we've got a vtordisp, add it as a storage type. 551 if (Layout.getVBaseOffsetsMap().find(BaseDecl)->second.hasVtorDisp()) 552 Members.push_back(StorageInfo(Offset - CharUnits::fromQuantity(4), 553 getIntNType(32))); 554 Members.push_back(MemberInfo(Offset, MemberInfo::VBase, 555 getStorageType(BaseDecl), BaseDecl)); 556 } 557 } 558 559 bool CGRecordLowering::hasOwnStorage(const CXXRecordDecl *Decl, 560 const CXXRecordDecl *Query) { 561 const ASTRecordLayout &DeclLayout = Context.getASTRecordLayout(Decl); 562 if (DeclLayout.isPrimaryBaseVirtual() && DeclLayout.getPrimaryBase() == Query) 563 return false; 564 for (const auto &Base : Decl->bases()) 565 if (!hasOwnStorage(Base.getType()->getAsCXXRecordDecl(), Query)) 566 return false; 567 return true; 568 } 569 570 void CGRecordLowering::calculateZeroInit() { 571 for (std::vector<MemberInfo>::const_iterator Member = Members.begin(), 572 MemberEnd = Members.end(); 573 IsZeroInitializableAsBase && Member != MemberEnd; ++Member) { 574 if (Member->Kind == MemberInfo::Field) { 575 if (!Member->FD || isZeroInitializable(Member->FD)) 576 continue; 577 IsZeroInitializable = IsZeroInitializableAsBase = false; 578 } else if (Member->Kind == MemberInfo::Base || 579 Member->Kind == MemberInfo::VBase) { 580 if (isZeroInitializable(Member->RD)) 581 continue; 582 IsZeroInitializable = false; 583 if (Member->Kind == MemberInfo::Base) 584 IsZeroInitializableAsBase = false; 585 } 586 } 587 } 588 589 void CGRecordLowering::clipTailPadding() { 590 std::vector<MemberInfo>::iterator Prior = Members.begin(); 591 CharUnits Tail = getSize(Prior->Data); 592 for (std::vector<MemberInfo>::iterator Member = Prior + 1, 593 MemberEnd = Members.end(); 594 Member != MemberEnd; ++Member) { 595 // Only members with data and the scissor can cut into tail padding. 596 if (!Member->Data && Member->Kind != MemberInfo::Scissor) 597 continue; 598 if (Member->Offset < Tail) { 599 assert(Prior->Kind == MemberInfo::Field && 600 "Only storage fields have tail padding!"); 601 if (!Prior->FD || Prior->FD->isBitField()) 602 Prior->Data = getByteArrayType(bitsToCharUnits(llvm::alignTo( 603 cast<llvm::IntegerType>(Prior->Data)->getIntegerBitWidth(), 8))); 604 else { 605 assert(Prior->FD->hasAttr<NoUniqueAddressAttr>() && 606 "should not have reused this field's tail padding"); 607 Prior->Data = getByteArrayType( 608 Context.getTypeInfoDataSizeInChars(Prior->FD->getType()).first); 609 } 610 } 611 if (Member->Data) 612 Prior = Member; 613 Tail = Prior->Offset + getSize(Prior->Data); 614 } 615 } 616 617 void CGRecordLowering::determinePacked(bool NVBaseType) { 618 if (Packed) 619 return; 620 CharUnits Alignment = CharUnits::One(); 621 CharUnits NVAlignment = CharUnits::One(); 622 CharUnits NVSize = 623 !NVBaseType && RD ? Layout.getNonVirtualSize() : CharUnits::Zero(); 624 for (std::vector<MemberInfo>::const_iterator Member = Members.begin(), 625 MemberEnd = Members.end(); 626 Member != MemberEnd; ++Member) { 627 if (!Member->Data) 628 continue; 629 // If any member falls at an offset that it not a multiple of its alignment, 630 // then the entire record must be packed. 631 if (Member->Offset % getAlignment(Member->Data)) 632 Packed = true; 633 if (Member->Offset < NVSize) 634 NVAlignment = std::max(NVAlignment, getAlignment(Member->Data)); 635 Alignment = std::max(Alignment, getAlignment(Member->Data)); 636 } 637 // If the size of the record (the capstone's offset) is not a multiple of the 638 // record's alignment, it must be packed. 639 if (Members.back().Offset % Alignment) 640 Packed = true; 641 // If the non-virtual sub-object is not a multiple of the non-virtual 642 // sub-object's alignment, it must be packed. We cannot have a packed 643 // non-virtual sub-object and an unpacked complete object or vise versa. 644 if (NVSize % NVAlignment) 645 Packed = true; 646 // Update the alignment of the sentinel. 647 if (!Packed) 648 Members.back().Data = getIntNType(Context.toBits(Alignment)); 649 } 650 651 void CGRecordLowering::insertPadding() { 652 std::vector<std::pair<CharUnits, CharUnits> > Padding; 653 CharUnits Size = CharUnits::Zero(); 654 for (std::vector<MemberInfo>::const_iterator Member = Members.begin(), 655 MemberEnd = Members.end(); 656 Member != MemberEnd; ++Member) { 657 if (!Member->Data) 658 continue; 659 CharUnits Offset = Member->Offset; 660 assert(Offset >= Size); 661 // Insert padding if we need to. 662 if (Offset != 663 Size.alignTo(Packed ? CharUnits::One() : getAlignment(Member->Data))) 664 Padding.push_back(std::make_pair(Size, Offset - Size)); 665 Size = Offset + getSize(Member->Data); 666 } 667 if (Padding.empty()) 668 return; 669 // Add the padding to the Members list and sort it. 670 for (std::vector<std::pair<CharUnits, CharUnits> >::const_iterator 671 Pad = Padding.begin(), PadEnd = Padding.end(); 672 Pad != PadEnd; ++Pad) 673 Members.push_back(StorageInfo(Pad->first, getByteArrayType(Pad->second))); 674 llvm::stable_sort(Members); 675 } 676 677 void CGRecordLowering::fillOutputFields() { 678 for (std::vector<MemberInfo>::const_iterator Member = Members.begin(), 679 MemberEnd = Members.end(); 680 Member != MemberEnd; ++Member) { 681 if (Member->Data) 682 FieldTypes.push_back(Member->Data); 683 if (Member->Kind == MemberInfo::Field) { 684 if (Member->FD) 685 Fields[Member->FD->getCanonicalDecl()] = FieldTypes.size() - 1; 686 // A field without storage must be a bitfield. 687 if (!Member->Data) 688 setBitFieldInfo(Member->FD, Member->Offset, FieldTypes.back()); 689 } else if (Member->Kind == MemberInfo::Base) 690 NonVirtualBases[Member->RD] = FieldTypes.size() - 1; 691 else if (Member->Kind == MemberInfo::VBase) 692 VirtualBases[Member->RD] = FieldTypes.size() - 1; 693 } 694 } 695 696 CGBitFieldInfo CGBitFieldInfo::MakeInfo(CodeGenTypes &Types, 697 const FieldDecl *FD, 698 uint64_t Offset, uint64_t Size, 699 uint64_t StorageSize, 700 CharUnits StorageOffset) { 701 // This function is vestigial from CGRecordLayoutBuilder days but is still 702 // used in GCObjCRuntime.cpp. That usage has a "fixme" attached to it that 703 // when addressed will allow for the removal of this function. 704 llvm::Type *Ty = Types.ConvertTypeForMem(FD->getType()); 705 CharUnits TypeSizeInBytes = 706 CharUnits::fromQuantity(Types.getDataLayout().getTypeAllocSize(Ty)); 707 uint64_t TypeSizeInBits = Types.getContext().toBits(TypeSizeInBytes); 708 709 bool IsSigned = FD->getType()->isSignedIntegerOrEnumerationType(); 710 711 if (Size > TypeSizeInBits) { 712 // We have a wide bit-field. The extra bits are only used for padding, so 713 // if we have a bitfield of type T, with size N: 714 // 715 // T t : N; 716 // 717 // We can just assume that it's: 718 // 719 // T t : sizeof(T); 720 // 721 Size = TypeSizeInBits; 722 } 723 724 // Reverse the bit offsets for big endian machines. Because we represent 725 // a bitfield as a single large integer load, we can imagine the bits 726 // counting from the most-significant-bit instead of the 727 // least-significant-bit. 728 if (Types.getDataLayout().isBigEndian()) { 729 Offset = StorageSize - (Offset + Size); 730 } 731 732 return CGBitFieldInfo(Offset, Size, IsSigned, StorageSize, StorageOffset); 733 } 734 735 std::unique_ptr<CGRecordLayout> 736 CodeGenTypes::ComputeRecordLayout(const RecordDecl *D, llvm::StructType *Ty) { 737 CGRecordLowering Builder(*this, D, /*Packed=*/false); 738 739 Builder.lower(/*NonVirtualBaseType=*/false); 740 741 // If we're in C++, compute the base subobject type. 742 llvm::StructType *BaseTy = nullptr; 743 if (isa<CXXRecordDecl>(D) && !D->isUnion() && !D->hasAttr<FinalAttr>()) { 744 BaseTy = Ty; 745 if (Builder.Layout.getNonVirtualSize() != Builder.Layout.getSize()) { 746 CGRecordLowering BaseBuilder(*this, D, /*Packed=*/Builder.Packed); 747 BaseBuilder.lower(/*NonVirtualBaseType=*/true); 748 BaseTy = llvm::StructType::create( 749 getLLVMContext(), BaseBuilder.FieldTypes, "", BaseBuilder.Packed); 750 addRecordTypeName(D, BaseTy, ".base"); 751 // BaseTy and Ty must agree on their packedness for getLLVMFieldNo to work 752 // on both of them with the same index. 753 assert(Builder.Packed == BaseBuilder.Packed && 754 "Non-virtual and complete types must agree on packedness"); 755 } 756 } 757 758 // Fill in the struct *after* computing the base type. Filling in the body 759 // signifies that the type is no longer opaque and record layout is complete, 760 // but we may need to recursively layout D while laying D out as a base type. 761 Ty->setBody(Builder.FieldTypes, Builder.Packed); 762 763 auto RL = std::make_unique<CGRecordLayout>( 764 Ty, BaseTy, (bool)Builder.IsZeroInitializable, 765 (bool)Builder.IsZeroInitializableAsBase); 766 767 RL->NonVirtualBases.swap(Builder.NonVirtualBases); 768 RL->CompleteObjectVirtualBases.swap(Builder.VirtualBases); 769 770 // Add all the field numbers. 771 RL->FieldInfo.swap(Builder.Fields); 772 773 // Add bitfield info. 774 RL->BitFields.swap(Builder.BitFields); 775 776 // Dump the layout, if requested. 777 if (getContext().getLangOpts().DumpRecordLayouts) { 778 llvm::outs() << "\n*** Dumping IRgen Record Layout\n"; 779 llvm::outs() << "Record: "; 780 D->dump(llvm::outs()); 781 llvm::outs() << "\nLayout: "; 782 RL->print(llvm::outs()); 783 } 784 785 #ifndef NDEBUG 786 // Verify that the computed LLVM struct size matches the AST layout size. 787 const ASTRecordLayout &Layout = getContext().getASTRecordLayout(D); 788 789 uint64_t TypeSizeInBits = getContext().toBits(Layout.getSize()); 790 assert(TypeSizeInBits == getDataLayout().getTypeAllocSizeInBits(Ty) && 791 "Type size mismatch!"); 792 793 if (BaseTy) { 794 CharUnits NonVirtualSize = Layout.getNonVirtualSize(); 795 796 uint64_t AlignedNonVirtualTypeSizeInBits = 797 getContext().toBits(NonVirtualSize); 798 799 assert(AlignedNonVirtualTypeSizeInBits == 800 getDataLayout().getTypeAllocSizeInBits(BaseTy) && 801 "Type size mismatch!"); 802 } 803 804 // Verify that the LLVM and AST field offsets agree. 805 llvm::StructType *ST = RL->getLLVMType(); 806 const llvm::StructLayout *SL = getDataLayout().getStructLayout(ST); 807 808 const ASTRecordLayout &AST_RL = getContext().getASTRecordLayout(D); 809 RecordDecl::field_iterator it = D->field_begin(); 810 for (unsigned i = 0, e = AST_RL.getFieldCount(); i != e; ++i, ++it) { 811 const FieldDecl *FD = *it; 812 813 // Ignore zero-sized fields. 814 if (FD->isZeroSize(getContext())) 815 continue; 816 817 // For non-bit-fields, just check that the LLVM struct offset matches the 818 // AST offset. 819 if (!FD->isBitField()) { 820 unsigned FieldNo = RL->getLLVMFieldNo(FD); 821 assert(AST_RL.getFieldOffset(i) == SL->getElementOffsetInBits(FieldNo) && 822 "Invalid field offset!"); 823 continue; 824 } 825 826 // Ignore unnamed bit-fields. 827 if (!FD->getDeclName()) 828 continue; 829 830 const CGBitFieldInfo &Info = RL->getBitFieldInfo(FD); 831 llvm::Type *ElementTy = ST->getTypeAtIndex(RL->getLLVMFieldNo(FD)); 832 833 // Unions have overlapping elements dictating their layout, but for 834 // non-unions we can verify that this section of the layout is the exact 835 // expected size. 836 if (D->isUnion()) { 837 // For unions we verify that the start is zero and the size 838 // is in-bounds. However, on BE systems, the offset may be non-zero, but 839 // the size + offset should match the storage size in that case as it 840 // "starts" at the back. 841 if (getDataLayout().isBigEndian()) 842 assert(static_cast<unsigned>(Info.Offset + Info.Size) == 843 Info.StorageSize && 844 "Big endian union bitfield does not end at the back"); 845 else 846 assert(Info.Offset == 0 && 847 "Little endian union bitfield with a non-zero offset"); 848 assert(Info.StorageSize <= SL->getSizeInBits() && 849 "Union not large enough for bitfield storage"); 850 } else { 851 assert(Info.StorageSize == 852 getDataLayout().getTypeAllocSizeInBits(ElementTy) && 853 "Storage size does not match the element type size"); 854 } 855 assert(Info.Size > 0 && "Empty bitfield!"); 856 assert(static_cast<unsigned>(Info.Offset) + Info.Size <= Info.StorageSize && 857 "Bitfield outside of its allocated storage"); 858 } 859 #endif 860 861 return RL; 862 } 863 864 void CGRecordLayout::print(raw_ostream &OS) const { 865 OS << "<CGRecordLayout\n"; 866 OS << " LLVMType:" << *CompleteObjectType << "\n"; 867 if (BaseSubobjectType) 868 OS << " NonVirtualBaseLLVMType:" << *BaseSubobjectType << "\n"; 869 OS << " IsZeroInitializable:" << IsZeroInitializable << "\n"; 870 OS << " BitFields:[\n"; 871 872 // Print bit-field infos in declaration order. 873 std::vector<std::pair<unsigned, const CGBitFieldInfo*> > BFIs; 874 for (llvm::DenseMap<const FieldDecl*, CGBitFieldInfo>::const_iterator 875 it = BitFields.begin(), ie = BitFields.end(); 876 it != ie; ++it) { 877 const RecordDecl *RD = it->first->getParent(); 878 unsigned Index = 0; 879 for (RecordDecl::field_iterator 880 it2 = RD->field_begin(); *it2 != it->first; ++it2) 881 ++Index; 882 BFIs.push_back(std::make_pair(Index, &it->second)); 883 } 884 llvm::array_pod_sort(BFIs.begin(), BFIs.end()); 885 for (unsigned i = 0, e = BFIs.size(); i != e; ++i) { 886 OS.indent(4); 887 BFIs[i].second->print(OS); 888 OS << "\n"; 889 } 890 891 OS << "]>\n"; 892 } 893 894 LLVM_DUMP_METHOD void CGRecordLayout::dump() const { 895 print(llvm::errs()); 896 } 897 898 void CGBitFieldInfo::print(raw_ostream &OS) const { 899 OS << "<CGBitFieldInfo" 900 << " Offset:" << Offset 901 << " Size:" << Size 902 << " IsSigned:" << IsSigned 903 << " StorageSize:" << StorageSize 904 << " StorageOffset:" << StorageOffset.getQuantity() << ">"; 905 } 906 907 LLVM_DUMP_METHOD void CGBitFieldInfo::dump() const { 908 print(llvm::errs()); 909 } 910