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     };
MemberInfo__anonaf82dbae0111::CGRecordLowering::MemberInfo86     MemberInfo(CharUnits Offset, InfoKind Kind, llvm::Type *Data,
87                const FieldDecl *FD = nullptr)
88       : Offset(Offset), Kind(Kind), Data(Data), FD(FD) {}
MemberInfo__anonaf82dbae0111::CGRecordLowering::MemberInfo89     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.
operator <__anonaf82dbae0111::CGRecordLowering::MemberInfo93     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 *.
StorageInfo__anonaf82dbae0111::CGRecordLowering99   MemberInfo StorageInfo(CharUnits Offset, llvm::Type *Data) {
100     return MemberInfo(Offset, MemberInfo::Field, Data);
101   }
useMSABI__anonaf82dbae0111::CGRecordLowering102   bool useMSABI() {
103     return Context.getTargetInfo().getCXXABI().isMicrosoft() ||
104            D->isMsStruct(Context);
105   }
106   /// \brief Wraps llvm::Type::getIntNTy with some implicit arguments.
getIntNType__anonaf82dbae0111::CGRecordLowering107   llvm::Type *getIntNType(uint64_t NumBits) {
108     return llvm::Type::getIntNTy(Types.getLLVMContext(),
109         (unsigned)llvm::RoundUpToAlignment(NumBits, 8));
110   }
111   /// \brief Gets an llvm type of size NumBytes and alignment 1.
getByteArrayType__anonaf82dbae0111::CGRecordLowering112   llvm::Type *getByteArrayType(CharUnits NumBytes) {
113     assert(!NumBytes.isZero() && "Empty byte arrays aren't allowed.");
114     llvm::Type *Type = llvm::Type::getInt8Ty(Types.getLLVMContext());
115     return NumBytes == CharUnits::One() ? Type :
116         (llvm::Type *)llvm::ArrayType::get(Type, NumBytes.getQuantity());
117   }
118   /// \brief Gets the storage type for a field decl and handles storage
119   /// for itanium bitfields that are smaller than their declared type.
getStorageType__anonaf82dbae0111::CGRecordLowering120   llvm::Type *getStorageType(const FieldDecl *FD) {
121     llvm::Type *Type = Types.ConvertTypeForMem(FD->getType());
122     return useMSABI() || !FD->isBitField() ? Type :
123         getIntNType(std::min(FD->getBitWidthValue(Context),
124                              (unsigned)Context.toBits(getSize(Type))));
125   }
126   /// \brief Gets the llvm Basesubobject type from a CXXRecordDecl.
getStorageType__anonaf82dbae0111::CGRecordLowering127   llvm::Type *getStorageType(const CXXRecordDecl *RD) {
128     return Types.getCGRecordLayout(RD).getBaseSubobjectLLVMType();
129   }
bitsToCharUnits__anonaf82dbae0111::CGRecordLowering130   CharUnits bitsToCharUnits(uint64_t BitOffset) {
131     return Context.toCharUnitsFromBits(BitOffset);
132   }
getSize__anonaf82dbae0111::CGRecordLowering133   CharUnits getSize(llvm::Type *Type) {
134     return CharUnits::fromQuantity(DataLayout.getTypeAllocSize(Type));
135   }
getAlignment__anonaf82dbae0111::CGRecordLowering136   CharUnits getAlignment(llvm::Type *Type) {
137     return CharUnits::fromQuantity(DataLayout.getABITypeAlignment(Type));
138   }
isZeroInitializable__anonaf82dbae0111::CGRecordLowering139   bool isZeroInitializable(const FieldDecl *FD) {
140     const Type *Type = FD->getType()->getBaseElementTypeUnsafe();
141     if (const MemberPointerType *MPT = Type->getAs<MemberPointerType>())
142       return Types.getCXXABI().isZeroInitializable(MPT);
143     if (const RecordType *RT = Type->getAs<RecordType>())
144       return isZeroInitializable(RT->getDecl());
145     return true;
146   }
isZeroInitializable__anonaf82dbae0111::CGRecordLowering147   bool isZeroInitializable(const RecordDecl *RD) {
148     return Types.getCGRecordLayout(RD).isZeroInitializable();
149   }
appendPaddingBytes__anonaf82dbae0111::CGRecordLowering150   void appendPaddingBytes(CharUnits Size) {
151     if (!Size.isZero())
152       FieldTypes.push_back(getByteArrayType(Size));
153   }
getFieldBitOffset__anonaf82dbae0111::CGRecordLowering154   uint64_t getFieldBitOffset(const FieldDecl *FD) {
155     return Layout.getFieldOffset(FD->getFieldIndex());
156   }
157   // Layout routines.
158   void setBitFieldInfo(const FieldDecl *FD, CharUnits StartOffset,
159                        llvm::Type *StorageType);
160   /// \brief Lowers an ASTRecordLayout to a llvm type.
161   void lower(bool NonVirtualBaseType);
162   void lowerUnion();
163   void accumulateFields();
164   void accumulateBitFields(RecordDecl::field_iterator Field,
165                         RecordDecl::field_iterator FieldEnd);
166   void accumulateBases();
167   void accumulateVPtrs();
168   void accumulateVBases();
169   /// \brief Recursively searches all of the bases to find out if a vbase is
170   /// not the primary vbase of some base class.
171   bool hasOwnStorage(const CXXRecordDecl *Decl, const CXXRecordDecl *Query);
172   void calculateZeroInit();
173   /// \brief Lowers bitfield storage types to I8 arrays for bitfields with tail
174   /// padding that is or can potentially be used.
175   void clipTailPadding();
176   /// \brief Determines if we need a packed llvm struct.
177   void determinePacked(bool NVBaseType);
178   /// \brief Inserts padding everwhere it's needed.
179   void insertPadding();
180   /// \brief Fills out the structures that are ultimately consumed.
181   void fillOutputFields();
182   // Input memoization fields.
183   CodeGenTypes &Types;
184   const ASTContext &Context;
185   const RecordDecl *D;
186   const CXXRecordDecl *RD;
187   const ASTRecordLayout &Layout;
188   const llvm::DataLayout &DataLayout;
189   // Helpful intermediate data-structures.
190   std::vector<MemberInfo> Members;
191   // Output fields, consumed by CodeGenTypes::ComputeRecordLayout.
192   SmallVector<llvm::Type *, 16> FieldTypes;
193   llvm::DenseMap<const FieldDecl *, unsigned> Fields;
194   llvm::DenseMap<const FieldDecl *, CGBitFieldInfo> BitFields;
195   llvm::DenseMap<const CXXRecordDecl *, unsigned> NonVirtualBases;
196   llvm::DenseMap<const CXXRecordDecl *, unsigned> VirtualBases;
197   bool IsZeroInitializable : 1;
198   bool IsZeroInitializableAsBase : 1;
199   bool Packed : 1;
200 private:
201   CGRecordLowering(const CGRecordLowering &) = delete;
202   void operator =(const CGRecordLowering &) = delete;
203 };
204 } // namespace {
205 
CGRecordLowering(CodeGenTypes & Types,const RecordDecl * D,bool Packed)206 CGRecordLowering::CGRecordLowering(CodeGenTypes &Types, const RecordDecl *D,                                 bool Packed)
207   : Types(Types), Context(Types.getContext()), D(D),
208     RD(dyn_cast<CXXRecordDecl>(D)),
209     Layout(Types.getContext().getASTRecordLayout(D)),
210     DataLayout(Types.getDataLayout()), IsZeroInitializable(true),
211     IsZeroInitializableAsBase(true), Packed(Packed) {}
212 
setBitFieldInfo(const FieldDecl * FD,CharUnits StartOffset,llvm::Type * StorageType)213 void CGRecordLowering::setBitFieldInfo(
214     const FieldDecl *FD, CharUnits StartOffset, llvm::Type *StorageType) {
215   CGBitFieldInfo &Info = BitFields[FD->getCanonicalDecl()];
216   Info.IsSigned = FD->getType()->isSignedIntegerOrEnumerationType();
217   Info.Offset = (unsigned)(getFieldBitOffset(FD) - Context.toBits(StartOffset));
218   Info.Size = FD->getBitWidthValue(Context);
219   Info.StorageSize = (unsigned)DataLayout.getTypeAllocSizeInBits(StorageType);
220   // Here we calculate the actual storage alignment of the bits.  E.g if we've
221   // got an alignment >= 2 and the bitfield starts at offset 6 we've got an
222   // alignment of 2.
223   Info.StorageAlignment =
224       Layout.getAlignment().alignmentAtOffset(StartOffset).getQuantity();
225   if (Info.Size > Info.StorageSize)
226     Info.Size = Info.StorageSize;
227   // Reverse the bit offsets for big endian machines. Because we represent
228   // a bitfield as a single large integer load, we can imagine the bits
229   // counting from the most-significant-bit instead of the
230   // least-significant-bit.
231   if (DataLayout.isBigEndian())
232     Info.Offset = Info.StorageSize - (Info.Offset + Info.Size);
233 }
234 
lower(bool NVBaseType)235 void CGRecordLowering::lower(bool NVBaseType) {
236   // The lowering process implemented in this function takes a variety of
237   // carefully ordered phases.
238   // 1) Store all members (fields and bases) in a list and sort them by offset.
239   // 2) Add a 1-byte capstone member at the Size of the structure.
240   // 3) Clip bitfield storages members if their tail padding is or might be
241   //    used by another field or base.  The clipping process uses the capstone
242   //    by treating it as another object that occurs after the record.
243   // 4) Determine if the llvm-struct requires packing.  It's important that this
244   //    phase occur after clipping, because clipping changes the llvm type.
245   //    This phase reads the offset of the capstone when determining packedness
246   //    and updates the alignment of the capstone to be equal of the alignment
247   //    of the record after doing so.
248   // 5) Insert padding everywhere it is needed.  This phase requires 'Packed' to
249   //    have been computed and needs to know the alignment of the record in
250   //    order to understand if explicit tail padding is needed.
251   // 6) Remove the capstone, we don't need it anymore.
252   // 7) Determine if this record can be zero-initialized.  This phase could have
253   //    been placed anywhere after phase 1.
254   // 8) Format the complete list of members in a way that can be consumed by
255   //    CodeGenTypes::ComputeRecordLayout.
256   CharUnits Size = NVBaseType ? Layout.getNonVirtualSize() : Layout.getSize();
257   if (D->isUnion())
258     return lowerUnion();
259   accumulateFields();
260   // RD implies C++.
261   if (RD) {
262     accumulateVPtrs();
263     accumulateBases();
264     if (Members.empty())
265       return appendPaddingBytes(Size);
266     if (!NVBaseType)
267       accumulateVBases();
268   }
269   std::stable_sort(Members.begin(), Members.end());
270   Members.push_back(StorageInfo(Size, getIntNType(8)));
271   clipTailPadding();
272   determinePacked(NVBaseType);
273   insertPadding();
274   Members.pop_back();
275   calculateZeroInit();
276   fillOutputFields();
277 }
278 
lowerUnion()279 void CGRecordLowering::lowerUnion() {
280   CharUnits LayoutSize = Layout.getSize();
281   llvm::Type *StorageType = nullptr;
282   bool SeenNamedMember = false;
283   // Iterate through the fields setting bitFieldInfo and the Fields array. Also
284   // locate the "most appropriate" storage type.  The heuristic for finding the
285   // storage type isn't necessary, the first (non-0-length-bitfield) field's
286   // type would work fine and be simpler but would be different than what we've
287   // been doing and cause lit tests to change.
288   for (const auto *Field : D->fields()) {
289     if (Field->isBitField()) {
290       // Skip 0 sized bitfields.
291       if (Field->getBitWidthValue(Context) == 0)
292         continue;
293       llvm::Type *FieldType = getStorageType(Field);
294       if (LayoutSize < getSize(FieldType))
295         FieldType = getByteArrayType(LayoutSize);
296       setBitFieldInfo(Field, CharUnits::Zero(), FieldType);
297     }
298     Fields[Field->getCanonicalDecl()] = 0;
299     llvm::Type *FieldType = getStorageType(Field);
300     // Compute zero-initializable status.
301     // This union might not be zero initialized: it may contain a pointer to
302     // data member which might have some exotic initialization sequence.
303     // If this is the case, then we aught not to try and come up with a "better"
304     // type, it might not be very easy to come up with a Constant which
305     // correctly initializes it.
306     if (!SeenNamedMember && Field->getDeclName()) {
307       SeenNamedMember = true;
308       if (!isZeroInitializable(Field)) {
309         IsZeroInitializable = IsZeroInitializableAsBase = false;
310         StorageType = FieldType;
311       }
312     }
313     // Because our union isn't zero initializable, we won't be getting a better
314     // storage type.
315     if (!IsZeroInitializable)
316       continue;
317     // Conditionally update our storage type if we've got a new "better" one.
318     if (!StorageType ||
319         getAlignment(FieldType) >  getAlignment(StorageType) ||
320         (getAlignment(FieldType) == getAlignment(StorageType) &&
321         getSize(FieldType) > getSize(StorageType)))
322       StorageType = FieldType;
323   }
324   // If we have no storage type just pad to the appropriate size and return.
325   if (!StorageType)
326     return appendPaddingBytes(LayoutSize);
327   // If our storage size was bigger than our required size (can happen in the
328   // case of packed bitfields on Itanium) then just use an I8 array.
329   if (LayoutSize < getSize(StorageType))
330     StorageType = getByteArrayType(LayoutSize);
331   FieldTypes.push_back(StorageType);
332   appendPaddingBytes(LayoutSize - getSize(StorageType));
333   // Set packed if we need it.
334   if (LayoutSize % getAlignment(StorageType))
335     Packed = true;
336 }
337 
accumulateFields()338 void CGRecordLowering::accumulateFields() {
339   for (RecordDecl::field_iterator Field = D->field_begin(),
340                                   FieldEnd = D->field_end();
341     Field != FieldEnd;)
342     if (Field->isBitField()) {
343       RecordDecl::field_iterator Start = Field;
344       // Iterate to gather the list of bitfields.
345       for (++Field; Field != FieldEnd && Field->isBitField(); ++Field);
346       accumulateBitFields(Start, Field);
347     } else {
348       Members.push_back(MemberInfo(
349           bitsToCharUnits(getFieldBitOffset(*Field)), MemberInfo::Field,
350           getStorageType(*Field), *Field));
351       ++Field;
352     }
353 }
354 
355 void
accumulateBitFields(RecordDecl::field_iterator Field,RecordDecl::field_iterator FieldEnd)356 CGRecordLowering::accumulateBitFields(RecordDecl::field_iterator Field,
357                                       RecordDecl::field_iterator FieldEnd) {
358   // Run stores the first element of the current run of bitfields.  FieldEnd is
359   // used as a special value to note that we don't have a current run.  A
360   // bitfield run is a contiguous collection of bitfields that can be stored in
361   // the same storage block.  Zero-sized bitfields and bitfields that would
362   // cross an alignment boundary break a run and start a new one.
363   RecordDecl::field_iterator Run = FieldEnd;
364   // Tail is the offset of the first bit off the end of the current run.  It's
365   // used to determine if the ASTRecordLayout is treating these two bitfields as
366   // contiguous.  StartBitOffset is offset of the beginning of the Run.
367   uint64_t StartBitOffset, Tail = 0;
368   if (useMSABI()) {
369     for (; Field != FieldEnd; ++Field) {
370       uint64_t BitOffset = getFieldBitOffset(*Field);
371       // Zero-width bitfields end runs.
372       if (Field->getBitWidthValue(Context) == 0) {
373         Run = FieldEnd;
374         continue;
375       }
376       llvm::Type *Type = Types.ConvertTypeForMem(Field->getType());
377       // If we don't have a run yet, or don't live within the previous run's
378       // allocated storage then we allocate some storage and start a new run.
379       if (Run == FieldEnd || BitOffset >= Tail) {
380         Run = Field;
381         StartBitOffset = BitOffset;
382         Tail = StartBitOffset + DataLayout.getTypeAllocSizeInBits(Type);
383         // Add the storage member to the record.  This must be added to the
384         // record before the bitfield members so that it gets laid out before
385         // the bitfields it contains get laid out.
386         Members.push_back(StorageInfo(bitsToCharUnits(StartBitOffset), Type));
387       }
388       // Bitfields get the offset of their storage but come afterward and remain
389       // there after a stable sort.
390       Members.push_back(MemberInfo(bitsToCharUnits(StartBitOffset),
391                                    MemberInfo::Field, nullptr, *Field));
392     }
393     return;
394   }
395   for (;;) {
396     // Check to see if we need to start a new run.
397     if (Run == FieldEnd) {
398       // If we're out of fields, return.
399       if (Field == FieldEnd)
400         break;
401       // Any non-zero-length bitfield can start a new run.
402       if (Field->getBitWidthValue(Context) != 0) {
403         Run = Field;
404         StartBitOffset = getFieldBitOffset(*Field);
405         Tail = StartBitOffset + Field->getBitWidthValue(Context);
406       }
407       ++Field;
408       continue;
409     }
410     // Add bitfields to the run as long as they qualify.
411     if (Field != FieldEnd && Field->getBitWidthValue(Context) != 0 &&
412         Tail == getFieldBitOffset(*Field)) {
413       Tail += Field->getBitWidthValue(Context);
414       ++Field;
415       continue;
416     }
417     // We've hit a break-point in the run and need to emit a storage field.
418     llvm::Type *Type = getIntNType(Tail - StartBitOffset);
419     // Add the storage member to the record and set the bitfield info for all of
420     // the bitfields in the run.  Bitfields get the offset of their storage but
421     // come afterward and remain there after a stable sort.
422     Members.push_back(StorageInfo(bitsToCharUnits(StartBitOffset), Type));
423     for (; Run != Field; ++Run)
424       Members.push_back(MemberInfo(bitsToCharUnits(StartBitOffset),
425                                    MemberInfo::Field, nullptr, *Run));
426     Run = FieldEnd;
427   }
428 }
429 
accumulateBases()430 void CGRecordLowering::accumulateBases() {
431   // If we've got a primary virtual base, we need to add it with the bases.
432   if (Layout.isPrimaryBaseVirtual()) {
433     const CXXRecordDecl *BaseDecl = Layout.getPrimaryBase();
434     Members.push_back(MemberInfo(CharUnits::Zero(), MemberInfo::Base,
435                                  getStorageType(BaseDecl), BaseDecl));
436   }
437   // Accumulate the non-virtual bases.
438   for (const auto &Base : RD->bases()) {
439     if (Base.isVirtual())
440       continue;
441     const CXXRecordDecl *BaseDecl = Base.getType()->getAsCXXRecordDecl();
442     if (!BaseDecl->isEmpty())
443       Members.push_back(MemberInfo(Layout.getBaseClassOffset(BaseDecl),
444           MemberInfo::Base, getStorageType(BaseDecl), BaseDecl));
445   }
446 }
447 
accumulateVPtrs()448 void CGRecordLowering::accumulateVPtrs() {
449   if (Layout.hasOwnVFPtr())
450     Members.push_back(MemberInfo(CharUnits::Zero(), MemberInfo::VFPtr,
451         llvm::FunctionType::get(getIntNType(32), /*isVarArg=*/true)->
452             getPointerTo()->getPointerTo()));
453   if (Layout.hasOwnVBPtr())
454     Members.push_back(MemberInfo(Layout.getVBPtrOffset(), MemberInfo::VBPtr,
455         llvm::Type::getInt32PtrTy(Types.getLLVMContext())));
456 }
457 
accumulateVBases()458 void CGRecordLowering::accumulateVBases() {
459   CharUnits ScissorOffset = Layout.getNonVirtualSize();
460   // In the itanium ABI, it's possible to place a vbase at a dsize that is
461   // smaller than the nvsize.  Here we check to see if such a base is placed
462   // before the nvsize and set the scissor offset to that, instead of the
463   // nvsize.
464   if (!useMSABI())
465     for (const auto &Base : RD->vbases()) {
466       const CXXRecordDecl *BaseDecl = Base.getType()->getAsCXXRecordDecl();
467       if (BaseDecl->isEmpty())
468         continue;
469       // If the vbase is a primary virtual base of some base, then it doesn't
470       // get its own storage location but instead lives inside of that base.
471       if (Context.isNearlyEmpty(BaseDecl) && !hasOwnStorage(RD, BaseDecl))
472         continue;
473       ScissorOffset = std::min(ScissorOffset,
474                                Layout.getVBaseClassOffset(BaseDecl));
475     }
476   Members.push_back(MemberInfo(ScissorOffset, MemberInfo::Scissor, nullptr,
477                                RD));
478   for (const auto &Base : RD->vbases()) {
479     const CXXRecordDecl *BaseDecl = Base.getType()->getAsCXXRecordDecl();
480     if (BaseDecl->isEmpty())
481       continue;
482     CharUnits Offset = Layout.getVBaseClassOffset(BaseDecl);
483     // If the vbase is a primary virtual base of some base, then it doesn't
484     // get its own storage location but instead lives inside of that base.
485     if (!useMSABI() && Context.isNearlyEmpty(BaseDecl) &&
486         !hasOwnStorage(RD, BaseDecl)) {
487       Members.push_back(MemberInfo(Offset, MemberInfo::VBase, nullptr,
488                                    BaseDecl));
489       continue;
490     }
491     // If we've got a vtordisp, add it as a storage type.
492     if (Layout.getVBaseOffsetsMap().find(BaseDecl)->second.hasVtorDisp())
493       Members.push_back(StorageInfo(Offset - CharUnits::fromQuantity(4),
494                                     getIntNType(32)));
495     Members.push_back(MemberInfo(Offset, MemberInfo::VBase,
496                                  getStorageType(BaseDecl), BaseDecl));
497   }
498 }
499 
hasOwnStorage(const CXXRecordDecl * Decl,const CXXRecordDecl * Query)500 bool CGRecordLowering::hasOwnStorage(const CXXRecordDecl *Decl,
501                                      const CXXRecordDecl *Query) {
502   const ASTRecordLayout &DeclLayout = Context.getASTRecordLayout(Decl);
503   if (DeclLayout.isPrimaryBaseVirtual() && DeclLayout.getPrimaryBase() == Query)
504     return false;
505   for (const auto &Base : Decl->bases())
506     if (!hasOwnStorage(Base.getType()->getAsCXXRecordDecl(), Query))
507       return false;
508   return true;
509 }
510 
calculateZeroInit()511 void CGRecordLowering::calculateZeroInit() {
512   for (std::vector<MemberInfo>::const_iterator Member = Members.begin(),
513                                                MemberEnd = Members.end();
514        IsZeroInitializableAsBase && Member != MemberEnd; ++Member) {
515     if (Member->Kind == MemberInfo::Field) {
516       if (!Member->FD || isZeroInitializable(Member->FD))
517         continue;
518       IsZeroInitializable = IsZeroInitializableAsBase = false;
519     } else if (Member->Kind == MemberInfo::Base ||
520                Member->Kind == MemberInfo::VBase) {
521       if (isZeroInitializable(Member->RD))
522         continue;
523       IsZeroInitializable = false;
524       if (Member->Kind == MemberInfo::Base)
525         IsZeroInitializableAsBase = false;
526     }
527   }
528 }
529 
clipTailPadding()530 void CGRecordLowering::clipTailPadding() {
531   std::vector<MemberInfo>::iterator Prior = Members.begin();
532   CharUnits Tail = getSize(Prior->Data);
533   for (std::vector<MemberInfo>::iterator Member = Prior + 1,
534                                          MemberEnd = Members.end();
535        Member != MemberEnd; ++Member) {
536     // Only members with data and the scissor can cut into tail padding.
537     if (!Member->Data && Member->Kind != MemberInfo::Scissor)
538       continue;
539     if (Member->Offset < Tail) {
540       assert(Prior->Kind == MemberInfo::Field && !Prior->FD &&
541              "Only storage fields have tail padding!");
542       Prior->Data = getByteArrayType(bitsToCharUnits(llvm::RoundUpToAlignment(
543           cast<llvm::IntegerType>(Prior->Data)->getIntegerBitWidth(), 8)));
544     }
545     if (Member->Data)
546       Prior = Member;
547     Tail = Prior->Offset + getSize(Prior->Data);
548   }
549 }
550 
determinePacked(bool NVBaseType)551 void CGRecordLowering::determinePacked(bool NVBaseType) {
552   if (Packed)
553     return;
554   CharUnits Alignment = CharUnits::One();
555   CharUnits NVAlignment = CharUnits::One();
556   CharUnits NVSize =
557       !NVBaseType && RD ? Layout.getNonVirtualSize() : CharUnits::Zero();
558   for (std::vector<MemberInfo>::const_iterator Member = Members.begin(),
559                                                MemberEnd = Members.end();
560        Member != MemberEnd; ++Member) {
561     if (!Member->Data)
562       continue;
563     // If any member falls at an offset that it not a multiple of its alignment,
564     // then the entire record must be packed.
565     if (Member->Offset % getAlignment(Member->Data))
566       Packed = true;
567     if (Member->Offset < NVSize)
568       NVAlignment = std::max(NVAlignment, getAlignment(Member->Data));
569     Alignment = std::max(Alignment, getAlignment(Member->Data));
570   }
571   // If the size of the record (the capstone's offset) is not a multiple of the
572   // record's alignment, it must be packed.
573   if (Members.back().Offset % Alignment)
574     Packed = true;
575   // If the non-virtual sub-object is not a multiple of the non-virtual
576   // sub-object's alignment, it must be packed.  We cannot have a packed
577   // non-virtual sub-object and an unpacked complete object or vise versa.
578   if (NVSize % NVAlignment)
579     Packed = true;
580   // Update the alignment of the sentinal.
581   if (!Packed)
582     Members.back().Data = getIntNType(Context.toBits(Alignment));
583 }
584 
insertPadding()585 void CGRecordLowering::insertPadding() {
586   std::vector<std::pair<CharUnits, CharUnits> > Padding;
587   CharUnits Size = CharUnits::Zero();
588   for (std::vector<MemberInfo>::const_iterator Member = Members.begin(),
589                                                MemberEnd = Members.end();
590        Member != MemberEnd; ++Member) {
591     if (!Member->Data)
592       continue;
593     CharUnits Offset = Member->Offset;
594     assert(Offset >= Size);
595     // Insert padding if we need to.
596     if (Offset != Size.RoundUpToAlignment(Packed ? CharUnits::One() :
597                                           getAlignment(Member->Data)))
598       Padding.push_back(std::make_pair(Size, Offset - Size));
599     Size = Offset + getSize(Member->Data);
600   }
601   if (Padding.empty())
602     return;
603   // Add the padding to the Members list and sort it.
604   for (std::vector<std::pair<CharUnits, CharUnits> >::const_iterator
605         Pad = Padding.begin(), PadEnd = Padding.end();
606         Pad != PadEnd; ++Pad)
607     Members.push_back(StorageInfo(Pad->first, getByteArrayType(Pad->second)));
608   std::stable_sort(Members.begin(), Members.end());
609 }
610 
fillOutputFields()611 void CGRecordLowering::fillOutputFields() {
612   for (std::vector<MemberInfo>::const_iterator Member = Members.begin(),
613                                                MemberEnd = Members.end();
614        Member != MemberEnd; ++Member) {
615     if (Member->Data)
616       FieldTypes.push_back(Member->Data);
617     if (Member->Kind == MemberInfo::Field) {
618       if (Member->FD)
619         Fields[Member->FD->getCanonicalDecl()] = FieldTypes.size() - 1;
620       // A field without storage must be a bitfield.
621       if (!Member->Data)
622         setBitFieldInfo(Member->FD, Member->Offset, FieldTypes.back());
623     } else if (Member->Kind == MemberInfo::Base)
624       NonVirtualBases[Member->RD] = FieldTypes.size() - 1;
625     else if (Member->Kind == MemberInfo::VBase)
626       VirtualBases[Member->RD] = FieldTypes.size() - 1;
627   }
628 }
629 
MakeInfo(CodeGenTypes & Types,const FieldDecl * FD,uint64_t Offset,uint64_t Size,uint64_t StorageSize,uint64_t StorageAlignment)630 CGBitFieldInfo CGBitFieldInfo::MakeInfo(CodeGenTypes &Types,
631                                         const FieldDecl *FD,
632                                         uint64_t Offset, uint64_t Size,
633                                         uint64_t StorageSize,
634                                         uint64_t StorageAlignment) {
635   // This function is vestigial from CGRecordLayoutBuilder days but is still
636   // used in GCObjCRuntime.cpp.  That usage has a "fixme" attached to it that
637   // when addressed will allow for the removal of this function.
638   llvm::Type *Ty = Types.ConvertTypeForMem(FD->getType());
639   CharUnits TypeSizeInBytes =
640     CharUnits::fromQuantity(Types.getDataLayout().getTypeAllocSize(Ty));
641   uint64_t TypeSizeInBits = Types.getContext().toBits(TypeSizeInBytes);
642 
643   bool IsSigned = FD->getType()->isSignedIntegerOrEnumerationType();
644 
645   if (Size > TypeSizeInBits) {
646     // We have a wide bit-field. The extra bits are only used for padding, so
647     // if we have a bitfield of type T, with size N:
648     //
649     // T t : N;
650     //
651     // We can just assume that it's:
652     //
653     // T t : sizeof(T);
654     //
655     Size = TypeSizeInBits;
656   }
657 
658   // Reverse the bit offsets for big endian machines. Because we represent
659   // a bitfield as a single large integer load, we can imagine the bits
660   // counting from the most-significant-bit instead of the
661   // least-significant-bit.
662   if (Types.getDataLayout().isBigEndian()) {
663     Offset = StorageSize - (Offset + Size);
664   }
665 
666   return CGBitFieldInfo(Offset, Size, IsSigned, StorageSize, StorageAlignment);
667 }
668 
ComputeRecordLayout(const RecordDecl * D,llvm::StructType * Ty)669 CGRecordLayout *CodeGenTypes::ComputeRecordLayout(const RecordDecl *D,
670                                                   llvm::StructType *Ty) {
671   CGRecordLowering Builder(*this, D, /*Packed=*/false);
672 
673   Builder.lower(/*NonVirtualBaseType=*/false);
674 
675   // If we're in C++, compute the base subobject type.
676   llvm::StructType *BaseTy = nullptr;
677   if (isa<CXXRecordDecl>(D) && !D->isUnion() && !D->hasAttr<FinalAttr>()) {
678     BaseTy = Ty;
679     if (Builder.Layout.getNonVirtualSize() != Builder.Layout.getSize()) {
680       CGRecordLowering BaseBuilder(*this, D, /*Packed=*/Builder.Packed);
681       BaseBuilder.lower(/*NonVirtualBaseType=*/true);
682       BaseTy = llvm::StructType::create(
683           getLLVMContext(), BaseBuilder.FieldTypes, "", BaseBuilder.Packed);
684       addRecordTypeName(D, BaseTy, ".base");
685       // BaseTy and Ty must agree on their packedness for getLLVMFieldNo to work
686       // on both of them with the same index.
687       assert(Builder.Packed == BaseBuilder.Packed &&
688              "Non-virtual and complete types must agree on packedness");
689     }
690   }
691 
692   // Fill in the struct *after* computing the base type.  Filling in the body
693   // signifies that the type is no longer opaque and record layout is complete,
694   // but we may need to recursively layout D while laying D out as a base type.
695   Ty->setBody(Builder.FieldTypes, Builder.Packed);
696 
697   CGRecordLayout *RL =
698     new CGRecordLayout(Ty, BaseTy, Builder.IsZeroInitializable,
699                         Builder.IsZeroInitializableAsBase);
700 
701   RL->NonVirtualBases.swap(Builder.NonVirtualBases);
702   RL->CompleteObjectVirtualBases.swap(Builder.VirtualBases);
703 
704   // Add all the field numbers.
705   RL->FieldInfo.swap(Builder.Fields);
706 
707   // Add bitfield info.
708   RL->BitFields.swap(Builder.BitFields);
709 
710   // Dump the layout, if requested.
711   if (getContext().getLangOpts().DumpRecordLayouts) {
712     llvm::outs() << "\n*** Dumping IRgen Record Layout\n";
713     llvm::outs() << "Record: ";
714     D->dump(llvm::outs());
715     llvm::outs() << "\nLayout: ";
716     RL->print(llvm::outs());
717   }
718 
719 #ifndef NDEBUG
720   // Verify that the computed LLVM struct size matches the AST layout size.
721   const ASTRecordLayout &Layout = getContext().getASTRecordLayout(D);
722 
723   uint64_t TypeSizeInBits = getContext().toBits(Layout.getSize());
724   assert(TypeSizeInBits == getDataLayout().getTypeAllocSizeInBits(Ty) &&
725          "Type size mismatch!");
726 
727   if (BaseTy) {
728     CharUnits NonVirtualSize  = Layout.getNonVirtualSize();
729 
730     uint64_t AlignedNonVirtualTypeSizeInBits =
731       getContext().toBits(NonVirtualSize);
732 
733     assert(AlignedNonVirtualTypeSizeInBits ==
734            getDataLayout().getTypeAllocSizeInBits(BaseTy) &&
735            "Type size mismatch!");
736   }
737 
738   // Verify that the LLVM and AST field offsets agree.
739   llvm::StructType *ST =
740     dyn_cast<llvm::StructType>(RL->getLLVMType());
741   const llvm::StructLayout *SL = getDataLayout().getStructLayout(ST);
742 
743   const ASTRecordLayout &AST_RL = getContext().getASTRecordLayout(D);
744   RecordDecl::field_iterator it = D->field_begin();
745   for (unsigned i = 0, e = AST_RL.getFieldCount(); i != e; ++i, ++it) {
746     const FieldDecl *FD = *it;
747 
748     // For non-bit-fields, just check that the LLVM struct offset matches the
749     // AST offset.
750     if (!FD->isBitField()) {
751       unsigned FieldNo = RL->getLLVMFieldNo(FD);
752       assert(AST_RL.getFieldOffset(i) == SL->getElementOffsetInBits(FieldNo) &&
753              "Invalid field offset!");
754       continue;
755     }
756 
757     // Ignore unnamed bit-fields.
758     if (!FD->getDeclName())
759       continue;
760 
761     // Don't inspect zero-length bitfields.
762     if (FD->getBitWidthValue(getContext()) == 0)
763       continue;
764 
765     const CGBitFieldInfo &Info = RL->getBitFieldInfo(FD);
766     llvm::Type *ElementTy = ST->getTypeAtIndex(RL->getLLVMFieldNo(FD));
767 
768     // Unions have overlapping elements dictating their layout, but for
769     // non-unions we can verify that this section of the layout is the exact
770     // expected size.
771     if (D->isUnion()) {
772       // For unions we verify that the start is zero and the size
773       // is in-bounds. However, on BE systems, the offset may be non-zero, but
774       // the size + offset should match the storage size in that case as it
775       // "starts" at the back.
776       if (getDataLayout().isBigEndian())
777         assert(static_cast<unsigned>(Info.Offset + Info.Size) ==
778                Info.StorageSize &&
779                "Big endian union bitfield does not end at the back");
780       else
781         assert(Info.Offset == 0 &&
782                "Little endian union bitfield with a non-zero offset");
783       assert(Info.StorageSize <= SL->getSizeInBits() &&
784              "Union not large enough for bitfield storage");
785     } else {
786       assert(Info.StorageSize ==
787              getDataLayout().getTypeAllocSizeInBits(ElementTy) &&
788              "Storage size does not match the element type size");
789     }
790     assert(Info.Size > 0 && "Empty bitfield!");
791     assert(static_cast<unsigned>(Info.Offset) + Info.Size <= Info.StorageSize &&
792            "Bitfield outside of its allocated storage");
793   }
794 #endif
795 
796   return RL;
797 }
798 
print(raw_ostream & OS) const799 void CGRecordLayout::print(raw_ostream &OS) const {
800   OS << "<CGRecordLayout\n";
801   OS << "  LLVMType:" << *CompleteObjectType << "\n";
802   if (BaseSubobjectType)
803     OS << "  NonVirtualBaseLLVMType:" << *BaseSubobjectType << "\n";
804   OS << "  IsZeroInitializable:" << IsZeroInitializable << "\n";
805   OS << "  BitFields:[\n";
806 
807   // Print bit-field infos in declaration order.
808   std::vector<std::pair<unsigned, const CGBitFieldInfo*> > BFIs;
809   for (llvm::DenseMap<const FieldDecl*, CGBitFieldInfo>::const_iterator
810          it = BitFields.begin(), ie = BitFields.end();
811        it != ie; ++it) {
812     const RecordDecl *RD = it->first->getParent();
813     unsigned Index = 0;
814     for (RecordDecl::field_iterator
815            it2 = RD->field_begin(); *it2 != it->first; ++it2)
816       ++Index;
817     BFIs.push_back(std::make_pair(Index, &it->second));
818   }
819   llvm::array_pod_sort(BFIs.begin(), BFIs.end());
820   for (unsigned i = 0, e = BFIs.size(); i != e; ++i) {
821     OS.indent(4);
822     BFIs[i].second->print(OS);
823     OS << "\n";
824   }
825 
826   OS << "]>\n";
827 }
828 
dump() const829 void CGRecordLayout::dump() const {
830   print(llvm::errs());
831 }
832 
print(raw_ostream & OS) const833 void CGBitFieldInfo::print(raw_ostream &OS) const {
834   OS << "<CGBitFieldInfo"
835      << " Offset:" << Offset
836      << " Size:" << Size
837      << " IsSigned:" << IsSigned
838      << " StorageSize:" << StorageSize
839      << " StorageAlignment:" << StorageAlignment << ">";
840 }
841 
dump() const842 void CGBitFieldInfo::dump() const {
843   print(llvm::errs());
844 }
845