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 sentinal 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 // sentinal 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 /// \brief 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 /// \brief 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::RoundUpToAlignment(NumBits, 8)); 125 } 126 /// \brief 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 /// \brief 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 /// \brief 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 const Type *Type = FD->getType()->getBaseElementTypeUnsafe(); 157 if (const MemberPointerType *MPT = Type->getAs<MemberPointerType>()) 158 return Types.getCXXABI().isZeroInitializable(MPT); 159 if (const RecordType *RT = Type->getAs<RecordType>()) 160 return isZeroInitializable(RT->getDecl()); 161 return true; 162 } 163 bool isZeroInitializable(const RecordDecl *RD) { 164 return Types.getCGRecordLayout(RD).isZeroInitializable(); 165 } 166 void appendPaddingBytes(CharUnits Size) { 167 if (!Size.isZero()) 168 FieldTypes.push_back(getByteArrayType(Size)); 169 } 170 uint64_t getFieldBitOffset(const FieldDecl *FD) { 171 return Layout.getFieldOffset(FD->getFieldIndex()); 172 } 173 // Layout routines. 174 void setBitFieldInfo(const FieldDecl *FD, CharUnits StartOffset, 175 llvm::Type *StorageType); 176 /// \brief Lowers an ASTRecordLayout to a llvm type. 177 void lower(bool NonVirtualBaseType); 178 void lowerUnion(); 179 void accumulateFields(); 180 void accumulateBitFields(RecordDecl::field_iterator Field, 181 RecordDecl::field_iterator FieldEnd); 182 void accumulateBases(); 183 void accumulateVPtrs(); 184 void accumulateVBases(); 185 /// \brief Recursively searches all of the bases to find out if a vbase is 186 /// not the primary vbase of some base class. 187 bool hasOwnStorage(const CXXRecordDecl *Decl, const CXXRecordDecl *Query); 188 void calculateZeroInit(); 189 /// \brief Lowers bitfield storage types to I8 arrays for bitfields with tail 190 /// padding that is or can potentially be used. 191 void clipTailPadding(); 192 /// \brief Determines if we need a packed llvm struct. 193 void determinePacked(bool NVBaseType); 194 /// \brief Inserts padding everwhere it's needed. 195 void insertPadding(); 196 /// \brief Fills out the structures that are ultimately consumed. 197 void fillOutputFields(); 198 // Input memoization fields. 199 CodeGenTypes &Types; 200 const ASTContext &Context; 201 const RecordDecl *D; 202 const CXXRecordDecl *RD; 203 const ASTRecordLayout &Layout; 204 const llvm::DataLayout &DataLayout; 205 // Helpful intermediate data-structures. 206 std::vector<MemberInfo> Members; 207 // Output fields, consumed by CodeGenTypes::ComputeRecordLayout. 208 SmallVector<llvm::Type *, 16> FieldTypes; 209 llvm::DenseMap<const FieldDecl *, unsigned> Fields; 210 llvm::DenseMap<const FieldDecl *, CGBitFieldInfo> BitFields; 211 llvm::DenseMap<const CXXRecordDecl *, unsigned> NonVirtualBases; 212 llvm::DenseMap<const CXXRecordDecl *, unsigned> VirtualBases; 213 bool IsZeroInitializable : 1; 214 bool IsZeroInitializableAsBase : 1; 215 bool Packed : 1; 216 private: 217 CGRecordLowering(const CGRecordLowering &) = delete; 218 void operator =(const CGRecordLowering &) = delete; 219 }; 220 } // namespace { 221 222 CGRecordLowering::CGRecordLowering(CodeGenTypes &Types, const RecordDecl *D, bool Packed) 223 : Types(Types), Context(Types.getContext()), D(D), 224 RD(dyn_cast<CXXRecordDecl>(D)), 225 Layout(Types.getContext().getASTRecordLayout(D)), 226 DataLayout(Types.getDataLayout()), IsZeroInitializable(true), 227 IsZeroInitializableAsBase(true), Packed(Packed) {} 228 229 void CGRecordLowering::setBitFieldInfo( 230 const FieldDecl *FD, CharUnits StartOffset, llvm::Type *StorageType) { 231 CGBitFieldInfo &Info = BitFields[FD->getCanonicalDecl()]; 232 Info.IsSigned = FD->getType()->isSignedIntegerOrEnumerationType(); 233 Info.Offset = (unsigned)(getFieldBitOffset(FD) - Context.toBits(StartOffset)); 234 Info.Size = FD->getBitWidthValue(Context); 235 Info.StorageSize = (unsigned)DataLayout.getTypeAllocSizeInBits(StorageType); 236 // Here we calculate the actual storage alignment of the bits. E.g if we've 237 // got an alignment >= 2 and the bitfield starts at offset 6 we've got an 238 // alignment of 2. 239 Info.StorageAlignment = 240 Layout.getAlignment().alignmentAtOffset(StartOffset).getQuantity(); 241 if (Info.Size > Info.StorageSize) 242 Info.Size = Info.StorageSize; 243 // Reverse the bit offsets for big endian machines. Because we represent 244 // a bitfield as a single large integer load, we can imagine the bits 245 // counting from the most-significant-bit instead of the 246 // least-significant-bit. 247 if (DataLayout.isBigEndian()) 248 Info.Offset = Info.StorageSize - (Info.Offset + Info.Size); 249 } 250 251 void CGRecordLowering::lower(bool NVBaseType) { 252 // The lowering process implemented in this function takes a variety of 253 // carefully ordered phases. 254 // 1) Store all members (fields and bases) in a list and sort them by offset. 255 // 2) Add a 1-byte capstone member at the Size of the structure. 256 // 3) Clip bitfield storages members if their tail padding is or might be 257 // used by another field or base. The clipping process uses the capstone 258 // by treating it as another object that occurs after the record. 259 // 4) Determine if the llvm-struct requires packing. It's important that this 260 // phase occur after clipping, because clipping changes the llvm type. 261 // This phase reads the offset of the capstone when determining packedness 262 // and updates the alignment of the capstone to be equal of the alignment 263 // of the record after doing so. 264 // 5) Insert padding everywhere it is needed. This phase requires 'Packed' to 265 // have been computed and needs to know the alignment of the record in 266 // order to understand if explicit tail padding is needed. 267 // 6) Remove the capstone, we don't need it anymore. 268 // 7) Determine if this record can be zero-initialized. This phase could have 269 // been placed anywhere after phase 1. 270 // 8) Format the complete list of members in a way that can be consumed by 271 // CodeGenTypes::ComputeRecordLayout. 272 CharUnits Size = NVBaseType ? Layout.getNonVirtualSize() : Layout.getSize(); 273 if (D->isUnion()) 274 return lowerUnion(); 275 accumulateFields(); 276 // RD implies C++. 277 if (RD) { 278 accumulateVPtrs(); 279 accumulateBases(); 280 if (Members.empty()) 281 return appendPaddingBytes(Size); 282 if (!NVBaseType) 283 accumulateVBases(); 284 } 285 std::stable_sort(Members.begin(), Members.end()); 286 Members.push_back(StorageInfo(Size, getIntNType(8))); 287 clipTailPadding(); 288 determinePacked(NVBaseType); 289 insertPadding(); 290 Members.pop_back(); 291 calculateZeroInit(); 292 fillOutputFields(); 293 } 294 295 void CGRecordLowering::lowerUnion() { 296 CharUnits LayoutSize = Layout.getSize(); 297 llvm::Type *StorageType = nullptr; 298 bool SeenNamedMember = false; 299 // Iterate through the fields setting bitFieldInfo and the Fields array. Also 300 // locate the "most appropriate" storage type. The heuristic for finding the 301 // storage type isn't necessary, the first (non-0-length-bitfield) field's 302 // type would work fine and be simpler but would be different than what we've 303 // been doing and cause lit tests to change. 304 for (const auto *Field : D->fields()) { 305 if (Field->isBitField()) { 306 // Skip 0 sized bitfields. 307 if (Field->getBitWidthValue(Context) == 0) 308 continue; 309 llvm::Type *FieldType = getStorageType(Field); 310 if (LayoutSize < getSize(FieldType)) 311 FieldType = getByteArrayType(LayoutSize); 312 setBitFieldInfo(Field, CharUnits::Zero(), FieldType); 313 } 314 Fields[Field->getCanonicalDecl()] = 0; 315 llvm::Type *FieldType = getStorageType(Field); 316 // Compute zero-initializable status. 317 // This union might not be zero initialized: it may contain a pointer to 318 // data member which might have some exotic initialization sequence. 319 // If this is the case, then we aught not to try and come up with a "better" 320 // type, it might not be very easy to come up with a Constant which 321 // correctly initializes it. 322 if (!SeenNamedMember && Field->getDeclName()) { 323 SeenNamedMember = true; 324 if (!isZeroInitializable(Field)) { 325 IsZeroInitializable = IsZeroInitializableAsBase = false; 326 StorageType = FieldType; 327 } 328 } 329 // Because our union isn't zero initializable, we won't be getting a better 330 // storage type. 331 if (!IsZeroInitializable) 332 continue; 333 // Conditionally update our storage type if we've got a new "better" one. 334 if (!StorageType || 335 getAlignment(FieldType) > getAlignment(StorageType) || 336 (getAlignment(FieldType) == getAlignment(StorageType) && 337 getSize(FieldType) > getSize(StorageType))) 338 StorageType = FieldType; 339 } 340 // If we have no storage type just pad to the appropriate size and return. 341 if (!StorageType) 342 return appendPaddingBytes(LayoutSize); 343 // If our storage size was bigger than our required size (can happen in the 344 // case of packed bitfields on Itanium) then just use an I8 array. 345 if (LayoutSize < getSize(StorageType)) 346 StorageType = getByteArrayType(LayoutSize); 347 FieldTypes.push_back(StorageType); 348 appendPaddingBytes(LayoutSize - getSize(StorageType)); 349 // Set packed if we need it. 350 if (LayoutSize % getAlignment(StorageType)) 351 Packed = true; 352 } 353 354 void CGRecordLowering::accumulateFields() { 355 for (RecordDecl::field_iterator Field = D->field_begin(), 356 FieldEnd = D->field_end(); 357 Field != FieldEnd;) 358 if (Field->isBitField()) { 359 RecordDecl::field_iterator Start = Field; 360 // Iterate to gather the list of bitfields. 361 for (++Field; Field != FieldEnd && Field->isBitField(); ++Field); 362 accumulateBitFields(Start, Field); 363 } else { 364 Members.push_back(MemberInfo( 365 bitsToCharUnits(getFieldBitOffset(*Field)), MemberInfo::Field, 366 getStorageType(*Field), *Field)); 367 ++Field; 368 } 369 } 370 371 void 372 CGRecordLowering::accumulateBitFields(RecordDecl::field_iterator Field, 373 RecordDecl::field_iterator FieldEnd) { 374 // Run stores the first element of the current run of bitfields. FieldEnd is 375 // used as a special value to note that we don't have a current run. A 376 // bitfield run is a contiguous collection of bitfields that can be stored in 377 // the same storage block. Zero-sized bitfields and bitfields that would 378 // cross an alignment boundary break a run and start a new one. 379 RecordDecl::field_iterator Run = FieldEnd; 380 // Tail is the offset of the first bit off the end of the current run. It's 381 // used to determine if the ASTRecordLayout is treating these two bitfields as 382 // contiguous. StartBitOffset is offset of the beginning of the Run. 383 uint64_t StartBitOffset, Tail = 0; 384 if (isDiscreteBitFieldABI()) { 385 for (; Field != FieldEnd; ++Field) { 386 uint64_t BitOffset = getFieldBitOffset(*Field); 387 // Zero-width bitfields end runs. 388 if (Field->getBitWidthValue(Context) == 0) { 389 Run = FieldEnd; 390 continue; 391 } 392 llvm::Type *Type = Types.ConvertTypeForMem(Field->getType()); 393 // If we don't have a run yet, or don't live within the previous run's 394 // allocated storage then we allocate some storage and start a new run. 395 if (Run == FieldEnd || BitOffset >= Tail) { 396 Run = Field; 397 StartBitOffset = BitOffset; 398 Tail = StartBitOffset + DataLayout.getTypeAllocSizeInBits(Type); 399 // Add the storage member to the record. This must be added to the 400 // record before the bitfield members so that it gets laid out before 401 // the bitfields it contains get laid out. 402 Members.push_back(StorageInfo(bitsToCharUnits(StartBitOffset), Type)); 403 } 404 // Bitfields get the offset of their storage but come afterward and remain 405 // there after a stable sort. 406 Members.push_back(MemberInfo(bitsToCharUnits(StartBitOffset), 407 MemberInfo::Field, nullptr, *Field)); 408 } 409 return; 410 } 411 for (;;) { 412 // Check to see if we need to start a new run. 413 if (Run == FieldEnd) { 414 // If we're out of fields, return. 415 if (Field == FieldEnd) 416 break; 417 // Any non-zero-length bitfield can start a new run. 418 if (Field->getBitWidthValue(Context) != 0) { 419 Run = Field; 420 StartBitOffset = getFieldBitOffset(*Field); 421 Tail = StartBitOffset + Field->getBitWidthValue(Context); 422 } 423 ++Field; 424 continue; 425 } 426 // Add bitfields to the run as long as they qualify. 427 if (Field != FieldEnd && Field->getBitWidthValue(Context) != 0 && 428 Tail == getFieldBitOffset(*Field)) { 429 Tail += Field->getBitWidthValue(Context); 430 ++Field; 431 continue; 432 } 433 // We've hit a break-point in the run and need to emit a storage field. 434 llvm::Type *Type = getIntNType(Tail - StartBitOffset); 435 // Add the storage member to the record and set the bitfield info for all of 436 // the bitfields in the run. Bitfields get the offset of their storage but 437 // come afterward and remain there after a stable sort. 438 Members.push_back(StorageInfo(bitsToCharUnits(StartBitOffset), Type)); 439 for (; Run != Field; ++Run) 440 Members.push_back(MemberInfo(bitsToCharUnits(StartBitOffset), 441 MemberInfo::Field, nullptr, *Run)); 442 Run = FieldEnd; 443 } 444 } 445 446 void CGRecordLowering::accumulateBases() { 447 // If we've got a primary virtual base, we need to add it with the bases. 448 if (Layout.isPrimaryBaseVirtual()) { 449 const CXXRecordDecl *BaseDecl = Layout.getPrimaryBase(); 450 Members.push_back(MemberInfo(CharUnits::Zero(), MemberInfo::Base, 451 getStorageType(BaseDecl), BaseDecl)); 452 } 453 // Accumulate the non-virtual bases. 454 for (const auto &Base : RD->bases()) { 455 if (Base.isVirtual()) 456 continue; 457 458 // Bases can be zero-sized even if not technically empty if they 459 // contain only a trailing array member. 460 const CXXRecordDecl *BaseDecl = Base.getType()->getAsCXXRecordDecl(); 461 if (!BaseDecl->isEmpty() && 462 !Context.getASTRecordLayout(BaseDecl).getSize().isZero()) 463 Members.push_back(MemberInfo(Layout.getBaseClassOffset(BaseDecl), 464 MemberInfo::Base, getStorageType(BaseDecl), BaseDecl)); 465 } 466 } 467 468 void CGRecordLowering::accumulateVPtrs() { 469 if (Layout.hasOwnVFPtr()) 470 Members.push_back(MemberInfo(CharUnits::Zero(), MemberInfo::VFPtr, 471 llvm::FunctionType::get(getIntNType(32), /*isVarArg=*/true)-> 472 getPointerTo()->getPointerTo())); 473 if (Layout.hasOwnVBPtr()) 474 Members.push_back(MemberInfo(Layout.getVBPtrOffset(), MemberInfo::VBPtr, 475 llvm::Type::getInt32PtrTy(Types.getLLVMContext()))); 476 } 477 478 void CGRecordLowering::accumulateVBases() { 479 CharUnits ScissorOffset = Layout.getNonVirtualSize(); 480 // In the itanium ABI, it's possible to place a vbase at a dsize that is 481 // smaller than the nvsize. Here we check to see if such a base is placed 482 // before the nvsize and set the scissor offset to that, instead of the 483 // nvsize. 484 if (isOverlappingVBaseABI()) 485 for (const auto &Base : RD->vbases()) { 486 const CXXRecordDecl *BaseDecl = Base.getType()->getAsCXXRecordDecl(); 487 if (BaseDecl->isEmpty()) 488 continue; 489 // If the vbase is a primary virtual base of some base, then it doesn't 490 // get its own storage location but instead lives inside of that base. 491 if (Context.isNearlyEmpty(BaseDecl) && !hasOwnStorage(RD, BaseDecl)) 492 continue; 493 ScissorOffset = std::min(ScissorOffset, 494 Layout.getVBaseClassOffset(BaseDecl)); 495 } 496 Members.push_back(MemberInfo(ScissorOffset, MemberInfo::Scissor, nullptr, 497 RD)); 498 for (const auto &Base : RD->vbases()) { 499 const CXXRecordDecl *BaseDecl = Base.getType()->getAsCXXRecordDecl(); 500 if (BaseDecl->isEmpty()) 501 continue; 502 CharUnits Offset = Layout.getVBaseClassOffset(BaseDecl); 503 // If the vbase is a primary virtual base of some base, then it doesn't 504 // get its own storage location but instead lives inside of that base. 505 if (isOverlappingVBaseABI() && 506 Context.isNearlyEmpty(BaseDecl) && 507 !hasOwnStorage(RD, BaseDecl)) { 508 Members.push_back(MemberInfo(Offset, MemberInfo::VBase, nullptr, 509 BaseDecl)); 510 continue; 511 } 512 // If we've got a vtordisp, add it as a storage type. 513 if (Layout.getVBaseOffsetsMap().find(BaseDecl)->second.hasVtorDisp()) 514 Members.push_back(StorageInfo(Offset - CharUnits::fromQuantity(4), 515 getIntNType(32))); 516 Members.push_back(MemberInfo(Offset, MemberInfo::VBase, 517 getStorageType(BaseDecl), BaseDecl)); 518 } 519 } 520 521 bool CGRecordLowering::hasOwnStorage(const CXXRecordDecl *Decl, 522 const CXXRecordDecl *Query) { 523 const ASTRecordLayout &DeclLayout = Context.getASTRecordLayout(Decl); 524 if (DeclLayout.isPrimaryBaseVirtual() && DeclLayout.getPrimaryBase() == Query) 525 return false; 526 for (const auto &Base : Decl->bases()) 527 if (!hasOwnStorage(Base.getType()->getAsCXXRecordDecl(), Query)) 528 return false; 529 return true; 530 } 531 532 void CGRecordLowering::calculateZeroInit() { 533 for (std::vector<MemberInfo>::const_iterator Member = Members.begin(), 534 MemberEnd = Members.end(); 535 IsZeroInitializableAsBase && Member != MemberEnd; ++Member) { 536 if (Member->Kind == MemberInfo::Field) { 537 if (!Member->FD || isZeroInitializable(Member->FD)) 538 continue; 539 IsZeroInitializable = IsZeroInitializableAsBase = false; 540 } else if (Member->Kind == MemberInfo::Base || 541 Member->Kind == MemberInfo::VBase) { 542 if (isZeroInitializable(Member->RD)) 543 continue; 544 IsZeroInitializable = false; 545 if (Member->Kind == MemberInfo::Base) 546 IsZeroInitializableAsBase = false; 547 } 548 } 549 } 550 551 void CGRecordLowering::clipTailPadding() { 552 std::vector<MemberInfo>::iterator Prior = Members.begin(); 553 CharUnits Tail = getSize(Prior->Data); 554 for (std::vector<MemberInfo>::iterator Member = Prior + 1, 555 MemberEnd = Members.end(); 556 Member != MemberEnd; ++Member) { 557 // Only members with data and the scissor can cut into tail padding. 558 if (!Member->Data && Member->Kind != MemberInfo::Scissor) 559 continue; 560 if (Member->Offset < Tail) { 561 assert(Prior->Kind == MemberInfo::Field && !Prior->FD && 562 "Only storage fields have tail padding!"); 563 Prior->Data = getByteArrayType(bitsToCharUnits(llvm::RoundUpToAlignment( 564 cast<llvm::IntegerType>(Prior->Data)->getIntegerBitWidth(), 8))); 565 } 566 if (Member->Data) 567 Prior = Member; 568 Tail = Prior->Offset + getSize(Prior->Data); 569 } 570 } 571 572 void CGRecordLowering::determinePacked(bool NVBaseType) { 573 if (Packed) 574 return; 575 CharUnits Alignment = CharUnits::One(); 576 CharUnits NVAlignment = CharUnits::One(); 577 CharUnits NVSize = 578 !NVBaseType && RD ? Layout.getNonVirtualSize() : CharUnits::Zero(); 579 for (std::vector<MemberInfo>::const_iterator Member = Members.begin(), 580 MemberEnd = Members.end(); 581 Member != MemberEnd; ++Member) { 582 if (!Member->Data) 583 continue; 584 // If any member falls at an offset that it not a multiple of its alignment, 585 // then the entire record must be packed. 586 if (Member->Offset % getAlignment(Member->Data)) 587 Packed = true; 588 if (Member->Offset < NVSize) 589 NVAlignment = std::max(NVAlignment, getAlignment(Member->Data)); 590 Alignment = std::max(Alignment, getAlignment(Member->Data)); 591 } 592 // If the size of the record (the capstone's offset) is not a multiple of the 593 // record's alignment, it must be packed. 594 if (Members.back().Offset % Alignment) 595 Packed = true; 596 // If the non-virtual sub-object is not a multiple of the non-virtual 597 // sub-object's alignment, it must be packed. We cannot have a packed 598 // non-virtual sub-object and an unpacked complete object or vise versa. 599 if (NVSize % NVAlignment) 600 Packed = true; 601 // Update the alignment of the sentinal. 602 if (!Packed) 603 Members.back().Data = getIntNType(Context.toBits(Alignment)); 604 } 605 606 void CGRecordLowering::insertPadding() { 607 std::vector<std::pair<CharUnits, CharUnits> > Padding; 608 CharUnits Size = CharUnits::Zero(); 609 for (std::vector<MemberInfo>::const_iterator Member = Members.begin(), 610 MemberEnd = Members.end(); 611 Member != MemberEnd; ++Member) { 612 if (!Member->Data) 613 continue; 614 CharUnits Offset = Member->Offset; 615 assert(Offset >= Size); 616 // Insert padding if we need to. 617 if (Offset != Size.RoundUpToAlignment(Packed ? CharUnits::One() : 618 getAlignment(Member->Data))) 619 Padding.push_back(std::make_pair(Size, Offset - Size)); 620 Size = Offset + getSize(Member->Data); 621 } 622 if (Padding.empty()) 623 return; 624 // Add the padding to the Members list and sort it. 625 for (std::vector<std::pair<CharUnits, CharUnits> >::const_iterator 626 Pad = Padding.begin(), PadEnd = Padding.end(); 627 Pad != PadEnd; ++Pad) 628 Members.push_back(StorageInfo(Pad->first, getByteArrayType(Pad->second))); 629 std::stable_sort(Members.begin(), Members.end()); 630 } 631 632 void CGRecordLowering::fillOutputFields() { 633 for (std::vector<MemberInfo>::const_iterator Member = Members.begin(), 634 MemberEnd = Members.end(); 635 Member != MemberEnd; ++Member) { 636 if (Member->Data) 637 FieldTypes.push_back(Member->Data); 638 if (Member->Kind == MemberInfo::Field) { 639 if (Member->FD) 640 Fields[Member->FD->getCanonicalDecl()] = FieldTypes.size() - 1; 641 // A field without storage must be a bitfield. 642 if (!Member->Data) 643 setBitFieldInfo(Member->FD, Member->Offset, FieldTypes.back()); 644 } else if (Member->Kind == MemberInfo::Base) 645 NonVirtualBases[Member->RD] = FieldTypes.size() - 1; 646 else if (Member->Kind == MemberInfo::VBase) 647 VirtualBases[Member->RD] = FieldTypes.size() - 1; 648 } 649 } 650 651 CGBitFieldInfo CGBitFieldInfo::MakeInfo(CodeGenTypes &Types, 652 const FieldDecl *FD, 653 uint64_t Offset, uint64_t Size, 654 uint64_t StorageSize, 655 uint64_t StorageAlignment) { 656 // This function is vestigial from CGRecordLayoutBuilder days but is still 657 // used in GCObjCRuntime.cpp. That usage has a "fixme" attached to it that 658 // when addressed will allow for the removal of this function. 659 llvm::Type *Ty = Types.ConvertTypeForMem(FD->getType()); 660 CharUnits TypeSizeInBytes = 661 CharUnits::fromQuantity(Types.getDataLayout().getTypeAllocSize(Ty)); 662 uint64_t TypeSizeInBits = Types.getContext().toBits(TypeSizeInBytes); 663 664 bool IsSigned = FD->getType()->isSignedIntegerOrEnumerationType(); 665 666 if (Size > TypeSizeInBits) { 667 // We have a wide bit-field. The extra bits are only used for padding, so 668 // if we have a bitfield of type T, with size N: 669 // 670 // T t : N; 671 // 672 // We can just assume that it's: 673 // 674 // T t : sizeof(T); 675 // 676 Size = TypeSizeInBits; 677 } 678 679 // Reverse the bit offsets for big endian machines. Because we represent 680 // a bitfield as a single large integer load, we can imagine the bits 681 // counting from the most-significant-bit instead of the 682 // least-significant-bit. 683 if (Types.getDataLayout().isBigEndian()) { 684 Offset = StorageSize - (Offset + Size); 685 } 686 687 return CGBitFieldInfo(Offset, Size, IsSigned, StorageSize, StorageAlignment); 688 } 689 690 CGRecordLayout *CodeGenTypes::ComputeRecordLayout(const RecordDecl *D, 691 llvm::StructType *Ty) { 692 CGRecordLowering Builder(*this, D, /*Packed=*/false); 693 694 Builder.lower(/*NonVirtualBaseType=*/false); 695 696 // If we're in C++, compute the base subobject type. 697 llvm::StructType *BaseTy = nullptr; 698 if (isa<CXXRecordDecl>(D) && !D->isUnion() && !D->hasAttr<FinalAttr>()) { 699 BaseTy = Ty; 700 if (Builder.Layout.getNonVirtualSize() != Builder.Layout.getSize()) { 701 CGRecordLowering BaseBuilder(*this, D, /*Packed=*/Builder.Packed); 702 BaseBuilder.lower(/*NonVirtualBaseType=*/true); 703 BaseTy = llvm::StructType::create( 704 getLLVMContext(), BaseBuilder.FieldTypes, "", BaseBuilder.Packed); 705 addRecordTypeName(D, BaseTy, ".base"); 706 // BaseTy and Ty must agree on their packedness for getLLVMFieldNo to work 707 // on both of them with the same index. 708 assert(Builder.Packed == BaseBuilder.Packed && 709 "Non-virtual and complete types must agree on packedness"); 710 } 711 } 712 713 // Fill in the struct *after* computing the base type. Filling in the body 714 // signifies that the type is no longer opaque and record layout is complete, 715 // but we may need to recursively layout D while laying D out as a base type. 716 Ty->setBody(Builder.FieldTypes, Builder.Packed); 717 718 CGRecordLayout *RL = 719 new CGRecordLayout(Ty, BaseTy, Builder.IsZeroInitializable, 720 Builder.IsZeroInitializableAsBase); 721 722 RL->NonVirtualBases.swap(Builder.NonVirtualBases); 723 RL->CompleteObjectVirtualBases.swap(Builder.VirtualBases); 724 725 // Add all the field numbers. 726 RL->FieldInfo.swap(Builder.Fields); 727 728 // Add bitfield info. 729 RL->BitFields.swap(Builder.BitFields); 730 731 // Dump the layout, if requested. 732 if (getContext().getLangOpts().DumpRecordLayouts) { 733 llvm::outs() << "\n*** Dumping IRgen Record Layout\n"; 734 llvm::outs() << "Record: "; 735 D->dump(llvm::outs()); 736 llvm::outs() << "\nLayout: "; 737 RL->print(llvm::outs()); 738 } 739 740 #ifndef NDEBUG 741 // Verify that the computed LLVM struct size matches the AST layout size. 742 const ASTRecordLayout &Layout = getContext().getASTRecordLayout(D); 743 744 uint64_t TypeSizeInBits = getContext().toBits(Layout.getSize()); 745 assert(TypeSizeInBits == getDataLayout().getTypeAllocSizeInBits(Ty) && 746 "Type size mismatch!"); 747 748 if (BaseTy) { 749 CharUnits NonVirtualSize = Layout.getNonVirtualSize(); 750 751 uint64_t AlignedNonVirtualTypeSizeInBits = 752 getContext().toBits(NonVirtualSize); 753 754 assert(AlignedNonVirtualTypeSizeInBits == 755 getDataLayout().getTypeAllocSizeInBits(BaseTy) && 756 "Type size mismatch!"); 757 } 758 759 // Verify that the LLVM and AST field offsets agree. 760 llvm::StructType *ST = 761 dyn_cast<llvm::StructType>(RL->getLLVMType()); 762 const llvm::StructLayout *SL = getDataLayout().getStructLayout(ST); 763 764 const ASTRecordLayout &AST_RL = getContext().getASTRecordLayout(D); 765 RecordDecl::field_iterator it = D->field_begin(); 766 for (unsigned i = 0, e = AST_RL.getFieldCount(); i != e; ++i, ++it) { 767 const FieldDecl *FD = *it; 768 769 // For non-bit-fields, just check that the LLVM struct offset matches the 770 // AST offset. 771 if (!FD->isBitField()) { 772 unsigned FieldNo = RL->getLLVMFieldNo(FD); 773 assert(AST_RL.getFieldOffset(i) == SL->getElementOffsetInBits(FieldNo) && 774 "Invalid field offset!"); 775 continue; 776 } 777 778 // Ignore unnamed bit-fields. 779 if (!FD->getDeclName()) 780 continue; 781 782 // Don't inspect zero-length bitfields. 783 if (FD->getBitWidthValue(getContext()) == 0) 784 continue; 785 786 const CGBitFieldInfo &Info = RL->getBitFieldInfo(FD); 787 llvm::Type *ElementTy = ST->getTypeAtIndex(RL->getLLVMFieldNo(FD)); 788 789 // Unions have overlapping elements dictating their layout, but for 790 // non-unions we can verify that this section of the layout is the exact 791 // expected size. 792 if (D->isUnion()) { 793 // For unions we verify that the start is zero and the size 794 // is in-bounds. However, on BE systems, the offset may be non-zero, but 795 // the size + offset should match the storage size in that case as it 796 // "starts" at the back. 797 if (getDataLayout().isBigEndian()) 798 assert(static_cast<unsigned>(Info.Offset + Info.Size) == 799 Info.StorageSize && 800 "Big endian union bitfield does not end at the back"); 801 else 802 assert(Info.Offset == 0 && 803 "Little endian union bitfield with a non-zero offset"); 804 assert(Info.StorageSize <= SL->getSizeInBits() && 805 "Union not large enough for bitfield storage"); 806 } else { 807 assert(Info.StorageSize == 808 getDataLayout().getTypeAllocSizeInBits(ElementTy) && 809 "Storage size does not match the element type size"); 810 } 811 assert(Info.Size > 0 && "Empty bitfield!"); 812 assert(static_cast<unsigned>(Info.Offset) + Info.Size <= Info.StorageSize && 813 "Bitfield outside of its allocated storage"); 814 } 815 #endif 816 817 return RL; 818 } 819 820 void CGRecordLayout::print(raw_ostream &OS) const { 821 OS << "<CGRecordLayout\n"; 822 OS << " LLVMType:" << *CompleteObjectType << "\n"; 823 if (BaseSubobjectType) 824 OS << " NonVirtualBaseLLVMType:" << *BaseSubobjectType << "\n"; 825 OS << " IsZeroInitializable:" << IsZeroInitializable << "\n"; 826 OS << " BitFields:[\n"; 827 828 // Print bit-field infos in declaration order. 829 std::vector<std::pair<unsigned, const CGBitFieldInfo*> > BFIs; 830 for (llvm::DenseMap<const FieldDecl*, CGBitFieldInfo>::const_iterator 831 it = BitFields.begin(), ie = BitFields.end(); 832 it != ie; ++it) { 833 const RecordDecl *RD = it->first->getParent(); 834 unsigned Index = 0; 835 for (RecordDecl::field_iterator 836 it2 = RD->field_begin(); *it2 != it->first; ++it2) 837 ++Index; 838 BFIs.push_back(std::make_pair(Index, &it->second)); 839 } 840 llvm::array_pod_sort(BFIs.begin(), BFIs.end()); 841 for (unsigned i = 0, e = BFIs.size(); i != e; ++i) { 842 OS.indent(4); 843 BFIs[i].second->print(OS); 844 OS << "\n"; 845 } 846 847 OS << "]>\n"; 848 } 849 850 void CGRecordLayout::dump() const { 851 print(llvm::errs()); 852 } 853 854 void CGBitFieldInfo::print(raw_ostream &OS) const { 855 OS << "<CGBitFieldInfo" 856 << " Offset:" << Offset 857 << " Size:" << Size 858 << " IsSigned:" << IsSigned 859 << " StorageSize:" << StorageSize 860 << " StorageAlignment:" << StorageAlignment << ">"; 861 } 862 863 void CGBitFieldInfo::dump() const { 864 print(llvm::errs()); 865 } 866