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"
29using namespace clang;
30using namespace CodeGen;
31
32namespace {
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.
73struct 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 /// Helper function to check if we are targeting AAPCS.
113 bool isAAPCS() const {
114 return Context.getTargetInfo().getABI().startswith("aapcs");
115 }
116
117 /// Helper function to check if the target machine is BigEndian.
118 bool isBE() const { return Context.getTargetInfo().isBigEndian(); }
119
120 /// The Itanium base layout rule allows virtual bases to overlap
121 /// other bases, which complicates layout in specific ways.
122 ///
123 /// Note specifically that the ms_struct attribute doesn't change this.
124 bool isOverlappingVBaseABI() {
125 return !Context.getTargetInfo().getCXXABI().isMicrosoft();
126 }
127
128 /// Wraps llvm::Type::getIntNTy with some implicit arguments.
129 llvm::Type *getIntNType(uint64_t NumBits) {
130 unsigned AlignedBits = llvm::alignTo(NumBits, Context.getCharWidth());
131 return llvm::Type::getIntNTy(Types.getLLVMContext(), AlignedBits);
132 }
133 /// Get the LLVM type sized as one character unit.
134 llvm::Type *getCharType() {
135 return llvm::Type::getIntNTy(Types.getLLVMContext(),
136 Context.getCharWidth());
137 }
138 /// Gets an llvm type of size NumChars and alignment 1.
139 llvm::Type *getByteArrayType(CharUnits NumChars) {
140 assert(!NumChars.isZero() && "Empty byte arrays aren't allowed.");
141 llvm::Type *Type = getCharType();
142 return NumChars == CharUnits::One() ? Type :
143 (llvm::Type *)llvm::ArrayType::get(Type, NumChars.getQuantity());
144 }
145 /// Gets the storage type for a field decl and handles storage
146 /// for itanium bitfields that are smaller than their declared type.
147 llvm::Type *getStorageType(const FieldDecl *FD) {
148 llvm::Type *Type = Types.ConvertTypeForMem(FD->getType());
149 if (!FD->isBitField()) return Type;
150 if (isDiscreteBitFieldABI()) return Type;
151 return getIntNType(std::min(FD->getBitWidthValue(Context),
152 (unsigned)Context.toBits(getSize(Type))));
153 }
154 /// Gets the llvm Basesubobject type from a CXXRecordDecl.
155 llvm::Type *getStorageType(const CXXRecordDecl *RD) {
156 return Types.getCGRecordLayout(RD).getBaseSubobjectLLVMType();
157 }
158 CharUnits bitsToCharUnits(uint64_t BitOffset) {
159 return Context.toCharUnitsFromBits(BitOffset);
160 }
161 CharUnits getSize(llvm::Type *Type) {
162 return CharUnits::fromQuantity(DataLayout.getTypeAllocSize(Type));
163 }
164 CharUnits getAlignment(llvm::Type *Type) {
165 return CharUnits::fromQuantity(DataLayout.getABITypeAlignment(Type));
166 }
167 bool isZeroInitializable(const FieldDecl *FD) {
168 return Types.isZeroInitializable(FD->getType());
169 }
170 bool isZeroInitializable(const RecordDecl *RD) {
171 return Types.isZeroInitializable(RD);
172 }
173 void appendPaddingBytes(CharUnits Size) {
174 if (!Size.isZero())
175 FieldTypes.push_back(getByteArrayType(Size));
176 }
177 uint64_t getFieldBitOffset(const FieldDecl *FD) {
178 return Layout.getFieldOffset(FD->getFieldIndex());
179 }
180 // Layout routines.
181 void setBitFieldInfo(const FieldDecl *FD, CharUnits StartOffset,
182 llvm::Type *StorageType);
183 /// Lowers an ASTRecordLayout to a llvm type.
184 void lower(bool NonVirtualBaseType);
185 void lowerUnion();
186 void accumulateFields();
187 void accumulateBitFields(RecordDecl::field_iterator Field,
188 RecordDecl::field_iterator FieldEnd);
189 void computeVolatileBitfields();
190 void accumulateBases();
191 void accumulateVPtrs();
192 void accumulateVBases();
193 /// Recursively searches all of the bases to find out if a vbase is
194 /// not the primary vbase of some base class.
195 bool hasOwnStorage(const CXXRecordDecl *Decl, const CXXRecordDecl *Query);
196 void calculateZeroInit();
197 /// Lowers bitfield storage types to I8 arrays for bitfields with tail
198 /// padding that is or can potentially be used.
199 void clipTailPadding();
200 /// Determines if we need a packed llvm struct.
201 void determinePacked(bool NVBaseType);
202 /// Inserts padding everywhere it's needed.
203 void insertPadding();
204 /// Fills out the structures that are ultimately consumed.
205 void fillOutputFields();
206 // Input memoization fields.
207 CodeGenTypes &Types;
208 const ASTContext &Context;
209 const RecordDecl *D;
210 const CXXRecordDecl *RD;
211 const ASTRecordLayout &Layout;
212 const llvm::DataLayout &DataLayout;
213 // Helpful intermediate data-structures.
214 std::vector<MemberInfo> Members;
215 // Output fields, consumed by CodeGenTypes::ComputeRecordLayout.
216 SmallVector<llvm::Type *, 16> FieldTypes;
217 llvm::DenseMap<const FieldDecl *, unsigned> Fields;
218 llvm::DenseMap<const FieldDecl *, CGBitFieldInfo> BitFields;
219 llvm::DenseMap<const CXXRecordDecl *, unsigned> NonVirtualBases;
220 llvm::DenseMap<const CXXRecordDecl *, unsigned> VirtualBases;
221 bool IsZeroInitializable : 1;
222 bool IsZeroInitializableAsBase : 1;
223 bool Packed : 1;
224private:
225 CGRecordLowering(const CGRecordLowering &) = delete;
226 void operator =(const CGRecordLowering &) = delete;
227};
228} // namespace {
229
230CGRecordLowering::CGRecordLowering(CodeGenTypes &Types, const RecordDecl *D,
231 bool Packed)
232 : Types(Types), Context(Types.getContext()), D(D),
233 RD(dyn_cast<CXXRecordDecl>(D)),
234 Layout(Types.getContext().getASTRecordLayout(D)),
235 DataLayout(Types.getDataLayout()), IsZeroInitializable(true),
236 IsZeroInitializableAsBase(true), Packed(Packed) {}
237
238void CGRecordLowering::setBitFieldInfo(
239 const FieldDecl *FD, CharUnits StartOffset, llvm::Type *StorageType) {
240 CGBitFieldInfo &Info = BitFields[FD->getCanonicalDecl()];
241 Info.IsSigned = FD->getType()->isSignedIntegerOrEnumerationType();
242 Info.Offset = (unsigned)(getFieldBitOffset(FD) - Context.toBits(StartOffset));
243 Info.Size = FD->getBitWidthValue(Context);
244 Info.StorageSize = (unsigned)DataLayout.getTypeAllocSizeInBits(StorageType);
245 Info.StorageOffset = StartOffset;
246 if (Info.Size > Info.StorageSize)
247 Info.Size = Info.StorageSize;
248 // Reverse the bit offsets for big endian machines. Because we represent
249 // a bitfield as a single large integer load, we can imagine the bits
250 // counting from the most-significant-bit instead of the
251 // least-significant-bit.
252 if (DataLayout.isBigEndian())
253 Info.Offset = Info.StorageSize - (Info.Offset + Info.Size);
254
255 Info.VolatileStorageSize = 0;
256 Info.VolatileOffset = 0;
257 Info.VolatileStorageOffset = CharUnits::Zero();
258}
259
260void CGRecordLowering::lower(bool NVBaseType) {
261 // The lowering process implemented in this function takes a variety of
262 // carefully ordered phases.
263 // 1) Store all members (fields and bases) in a list and sort them by offset.
264 // 2) Add a 1-byte capstone member at the Size of the structure.
265 // 3) Clip bitfield storages members if their tail padding is or might be
266 // used by another field or base. The clipping process uses the capstone
267 // by treating it as another object that occurs after the record.
268 // 4) Determine if the llvm-struct requires packing. It's important that this
269 // phase occur after clipping, because clipping changes the llvm type.
270 // This phase reads the offset of the capstone when determining packedness
271 // and updates the alignment of the capstone to be equal of the alignment
272 // of the record after doing so.
273 // 5) Insert padding everywhere it is needed. This phase requires 'Packed' to
274 // have been computed and needs to know the alignment of the record in
275 // order to understand if explicit tail padding is needed.
276 // 6) Remove the capstone, we don't need it anymore.
277 // 7) Determine if this record can be zero-initialized. This phase could have
278 // been placed anywhere after phase 1.
279 // 8) Format the complete list of members in a way that can be consumed by
280 // CodeGenTypes::ComputeRecordLayout.
281 CharUnits Size = NVBaseType ? Layout.getNonVirtualSize() : Layout.getSize();
282 if (D->isUnion()) {
283 lowerUnion();
284 computeVolatileBitfields();
285 return;
286 }
287 accumulateFields();
288 // RD implies C++.
289 if (RD) {
290 accumulateVPtrs();
291 accumulateBases();
292 if (Members.empty()) {
293 appendPaddingBytes(Size);
294 computeVolatileBitfields();
295 return;
296 }
297 if (!NVBaseType)
298 accumulateVBases();
299 }
300 llvm::stable_sort(Members);
301 Members.push_back(StorageInfo(Size, getIntNType(8)));
302 clipTailPadding();
303 determinePacked(NVBaseType);
304 insertPadding();
305 Members.pop_back();
306 calculateZeroInit();
307 fillOutputFields();
308 computeVolatileBitfields();
309}
310
311void CGRecordLowering::lowerUnion() {
312 CharUnits LayoutSize = Layout.getSize();
313 llvm::Type *StorageType = nullptr;
314 bool SeenNamedMember = false;
315 // Iterate through the fields setting bitFieldInfo and the Fields array. Also
316 // locate the "most appropriate" storage type. The heuristic for finding the
317 // storage type isn't necessary, the first (non-0-length-bitfield) field's
318 // type would work fine and be simpler but would be different than what we've
319 // been doing and cause lit tests to change.
320 for (const auto *Field : D->fields()) {
321 if (Field->isBitField()) {
322 if (Field->isZeroLengthBitField(Context))
323 continue;
324 llvm::Type *FieldType = getStorageType(Field);
325 if (LayoutSize < getSize(FieldType))
326 FieldType = getByteArrayType(LayoutSize);
327 setBitFieldInfo(Field, CharUnits::Zero(), FieldType);
328 }
329 Fields[Field->getCanonicalDecl()] = 0;
330 llvm::Type *FieldType = getStorageType(Field);
331 // Compute zero-initializable status.
332 // This union might not be zero initialized: it may contain a pointer to
333 // data member which might have some exotic initialization sequence.
334 // If this is the case, then we aught not to try and come up with a "better"
335 // type, it might not be very easy to come up with a Constant which
336 // correctly initializes it.
337 if (!SeenNamedMember) {
338 SeenNamedMember = Field->getIdentifier();
339 if (!SeenNamedMember)
340 if (const auto *FieldRD = Field->getType()->getAsRecordDecl())
341 SeenNamedMember = FieldRD->findFirstNamedDataMember();
342 if (SeenNamedMember && !isZeroInitializable(Field)) {
343 IsZeroInitializable = IsZeroInitializableAsBase = false;
344 StorageType = FieldType;
345 }
346 }
347 // Because our union isn't zero initializable, we won't be getting a better
348 // storage type.
349 if (!IsZeroInitializable)
350 continue;
351 // Conditionally update our storage type if we've got a new "better" one.
352 if (!StorageType ||
353 getAlignment(FieldType) > getAlignment(StorageType) ||
354 (getAlignment(FieldType) == getAlignment(StorageType) &&
355 getSize(FieldType) > getSize(StorageType)))
356 StorageType = FieldType;
357 }
358 // If we have no storage type just pad to the appropriate size and return.
359 if (!StorageType)
360 return appendPaddingBytes(LayoutSize);
361 // If our storage size was bigger than our required size (can happen in the
362 // case of packed bitfields on Itanium) then just use an I8 array.
363 if (LayoutSize < getSize(StorageType))
364 StorageType = getByteArrayType(LayoutSize);
365 FieldTypes.push_back(StorageType);
366 appendPaddingBytes(LayoutSize - getSize(StorageType));
367 // Set packed if we need it.
368 if (LayoutSize % getAlignment(StorageType))
369 Packed = true;
370}
371
372void CGRecordLowering::accumulateFields() {
373 for (RecordDecl::field_iterator Field = D->field_begin(),
374 FieldEnd = D->field_end();
375 Field != FieldEnd;) {
376 if (Field->isBitField()) {
377 RecordDecl::field_iterator Start = Field;
378 // Iterate to gather the list of bitfields.
379 for (++Field; Field != FieldEnd && Field->isBitField(); ++Field);
380 accumulateBitFields(Start, Field);
381 } else if (!Field->isZeroSize(Context)) {
382 Members.push_back(MemberInfo(
383 bitsToCharUnits(getFieldBitOffset(*Field)), MemberInfo::Field,
384 getStorageType(*Field), *Field));
385 ++Field;
386 } else {
387 ++Field;
388 }
389 }
390}
391
392void
393CGRecordLowering::accumulateBitFields(RecordDecl::field_iterator Field,
394 RecordDecl::field_iterator FieldEnd) {
395 // Run stores the first element of the current run of bitfields. FieldEnd is
396 // used as a special value to note that we don't have a current run. A
397 // bitfield run is a contiguous collection of bitfields that can be stored in
398 // the same storage block. Zero-sized bitfields and bitfields that would
399 // cross an alignment boundary break a run and start a new one.
400 RecordDecl::field_iterator Run = FieldEnd;
401 // Tail is the offset of the first bit off the end of the current run. It's
402 // used to determine if the ASTRecordLayout is treating these two bitfields as
403 // contiguous. StartBitOffset is offset of the beginning of the Run.
404 uint64_t StartBitOffset, Tail = 0;
405 if (isDiscreteBitFieldABI()) {
406 for (; Field != FieldEnd; ++Field) {
407 uint64_t BitOffset = getFieldBitOffset(*Field);
408 // Zero-width bitfields end runs.
409 if (Field->isZeroLengthBitField(Context)) {
410 Run = FieldEnd;
411 continue;
412 }
413 llvm::Type *Type =
414 Types.ConvertTypeForMem(Field->getType(), /*ForBitFields=*/true);
415 // If we don't have a run yet, or don't live within the previous run's
416 // allocated storage then we allocate some storage and start a new run.
417 if (Run == FieldEnd || BitOffset >= Tail) {
418 Run = Field;
419 StartBitOffset = BitOffset;
420 Tail = StartBitOffset + DataLayout.getTypeAllocSizeInBits(Type);
421 // Add the storage member to the record. This must be added to the
422 // record before the bitfield members so that it gets laid out before
423 // the bitfields it contains get laid out.
424 Members.push_back(StorageInfo(bitsToCharUnits(StartBitOffset), Type));
425 }
426 // Bitfields get the offset of their storage but come afterward and remain
427 // there after a stable sort.
428 Members.push_back(MemberInfo(bitsToCharUnits(StartBitOffset),
429 MemberInfo::Field, nullptr, *Field));
430 }
431 return;
432 }
433
434 // Check if OffsetInRecord (the size in bits of the current run) is better
435 // as a single field run. When OffsetInRecord has legal integer width, and
436 // its bitfield offset is naturally aligned, it is better to make the
437 // bitfield a separate storage component so as it can be accessed directly
438 // with lower cost.
439 auto IsBetterAsSingleFieldRun = [&](uint64_t OffsetInRecord,
440 uint64_t StartBitOffset) {
441 if (!Types.getCodeGenOpts().FineGrainedBitfieldAccesses)
442 return false;
443 if (OffsetInRecord < 8 || !llvm::isPowerOf2_64(OffsetInRecord) ||
444 !DataLayout.fitsInLegalInteger(OffsetInRecord))
445 return false;
446 // Make sure StartBitOffset is naturally aligned if it is treated as an
447 // IType integer.
448 if (StartBitOffset %
449 Context.toBits(getAlignment(getIntNType(OffsetInRecord))) !=
450 0)
451 return false;
452 return true;
453 };
454
455 // The start field is better as a single field run.
456 bool StartFieldAsSingleRun = false;
457 for (;;) {
458 // Check to see if we need to start a new run.
459 if (Run == FieldEnd) {
460 // If we're out of fields, return.
461 if (Field == FieldEnd)
462 break;
463 // Any non-zero-length bitfield can start a new run.
464 if (!Field->isZeroLengthBitField(Context)) {
465 Run = Field;
466 StartBitOffset = getFieldBitOffset(*Field);
467 Tail = StartBitOffset + Field->getBitWidthValue(Context);
468 StartFieldAsSingleRun = IsBetterAsSingleFieldRun(Tail - StartBitOffset,
469 StartBitOffset);
470 }
471 ++Field;
472 continue;
473 }
474
475 // If the start field of a new run is better as a single run, or
476 // if current field (or consecutive fields) is better as a single run, or
477 // if current field has zero width bitfield and either
478 // UseZeroLengthBitfieldAlignment or UseBitFieldTypeAlignment is set to
479 // true, or
480 // if the offset of current field is inconsistent with the offset of
481 // previous field plus its offset,
482 // skip the block below and go ahead to emit the storage.
483 // Otherwise, try to add bitfields to the run.
484 if (!StartFieldAsSingleRun && Field != FieldEnd &&
485 !IsBetterAsSingleFieldRun(Tail - StartBitOffset, StartBitOffset) &&
486 (!Field->isZeroLengthBitField(Context) ||
487 (!Context.getTargetInfo().useZeroLengthBitfieldAlignment() &&
488 !Context.getTargetInfo().useBitFieldTypeAlignment())) &&
489 Tail == getFieldBitOffset(*Field)) {
490 Tail += Field->getBitWidthValue(Context);
491 ++Field;
492 continue;
493 }
494
495 // We've hit a break-point in the run and need to emit a storage field.
496 llvm::Type *Type = getIntNType(Tail - StartBitOffset);
497 // Add the storage member to the record and set the bitfield info for all of
498 // the bitfields in the run. Bitfields get the offset of their storage but
499 // come afterward and remain there after a stable sort.
500 Members.push_back(StorageInfo(bitsToCharUnits(StartBitOffset), Type));
501 for (; Run != Field; ++Run)
502 Members.push_back(MemberInfo(bitsToCharUnits(StartBitOffset),
503 MemberInfo::Field, nullptr, *Run));
504 Run = FieldEnd;
505 StartFieldAsSingleRun = false;
506 }
507}
508
509void CGRecordLowering::accumulateBases() {
510 // If we've got a primary virtual base, we need to add it with the bases.
511 if (Layout.isPrimaryBaseVirtual()) {
512 const CXXRecordDecl *BaseDecl = Layout.getPrimaryBase();
513 Members.push_back(MemberInfo(CharUnits::Zero(), MemberInfo::Base,
514 getStorageType(BaseDecl), BaseDecl));
515 }
516 // Accumulate the non-virtual bases.
517 for (const auto &Base : RD->bases()) {
518 if (Base.isVirtual())
519 continue;
520
521 // Bases can be zero-sized even if not technically empty if they
522 // contain only a trailing array member.
523 const CXXRecordDecl *BaseDecl = Base.getType()->getAsCXXRecordDecl();
524 if (!BaseDecl->isEmpty() &&
525 !Context.getASTRecordLayout(BaseDecl).getNonVirtualSize().isZero())
526 Members.push_back(MemberInfo(Layout.getBaseClassOffset(BaseDecl),
527 MemberInfo::Base, getStorageType(BaseDecl), BaseDecl));
528 }
529}
530
531/// The AAPCS that defines that, when possible, bit-fields should
532/// be accessed using containers of the declared type width:
533/// When a volatile bit-field is read, and its container does not overlap with
534/// any non-bit-field member or any zero length bit-field member, its container
535/// must be read exactly once using the access width appropriate to the type of
536/// the container. When a volatile bit-field is written, and its container does
537/// not overlap with any non-bit-field member or any zero-length bit-field
538/// member, its container must be read exactly once and written exactly once
539/// using the access width appropriate to the type of the container. The two
540/// accesses are not atomic.
541///
542/// Enforcing the width restriction can be disabled using
543/// -fno-aapcs-bitfield-width.
544void CGRecordLowering::computeVolatileBitfields() {
545 if (!isAAPCS() || !Types.getCodeGenOpts().AAPCSBitfieldWidth)
546 return;
547
548 for (auto &I : BitFields) {
549 const FieldDecl *Field = I.first;
550 CGBitFieldInfo &Info = I.second;
551 llvm::Type *ResLTy = Types.ConvertTypeForMem(Field->getType());
552 // If the record alignment is less than the type width, we can't enforce a
553 // aligned load, bail out.
554 if ((uint64_t)(Context.toBits(Layout.getAlignment())) <
555 ResLTy->getPrimitiveSizeInBits())
556 continue;
557 // CGRecordLowering::setBitFieldInfo() pre-adjusts the bit-field offsets
558 // for big-endian targets, but it assumes a container of width
559 // Info.StorageSize. Since AAPCS uses a different container size (width
560 // of the type), we first undo that calculation here and redo it once
561 // the bit-field offset within the new container is calculated.
562 const unsigned OldOffset =
563 isBE() ? Info.StorageSize - (Info.Offset + Info.Size) : Info.Offset;
564 // Offset to the bit-field from the beginning of the struct.
565 const unsigned AbsoluteOffset =
566 Context.toBits(Info.StorageOffset) + OldOffset;
567
568 // Container size is the width of the bit-field type.
569 const unsigned StorageSize = ResLTy->getPrimitiveSizeInBits();
570 // Nothing to do if the access uses the desired
571 // container width and is naturally aligned.
572 if (Info.StorageSize == StorageSize && (OldOffset % StorageSize == 0))
573 continue;
574
575 // Offset within the container.
576 unsigned Offset = AbsoluteOffset & (StorageSize - 1);
577 // Bail out if an aligned load of the container cannot cover the entire
578 // bit-field. This can happen for example, if the bit-field is part of a
579 // packed struct. AAPCS does not define access rules for such cases, we let
580 // clang to follow its own rules.
581 if (Offset + Info.Size > StorageSize)
582 continue;
583
584 // Re-adjust offsets for big-endian targets.
585 if (isBE())
586 Offset = StorageSize - (Offset + Info.Size);
587
588 const CharUnits StorageOffset =
589 Context.toCharUnitsFromBits(AbsoluteOffset & ~(StorageSize - 1));
590 const CharUnits End = StorageOffset +
591 Context.toCharUnitsFromBits(StorageSize) -
592 CharUnits::One();
593
594 const ASTRecordLayout &Layout =
595 Context.getASTRecordLayout(Field->getParent());
596 // If we access outside memory outside the record, than bail out.
597 const CharUnits RecordSize = Layout.getSize();
598 if (End >= RecordSize)
599 continue;
600
601 // Bail out if performing this load would access non-bit-fields members.
602 bool Conflict = false;
603 for (const auto *F : D->fields()) {
604 // Allow sized bit-fields overlaps.
605 if (F->isBitField() && !F->isZeroLengthBitField(Context))
606 continue;
607
608 const CharUnits FOffset = Context.toCharUnitsFromBits(
609 Layout.getFieldOffset(F->getFieldIndex()));
610
611 // As C11 defines, a zero sized bit-field defines a barrier, so
612 // fields after and before it should be race condition free.
613 // The AAPCS acknowledges it and imposes no restritions when the
614 // natural container overlaps a zero-length bit-field.
615 if (F->isZeroLengthBitField(Context)) {
616 if (End > FOffset && StorageOffset < FOffset) {
617 Conflict = true;
618 break;
619 }
620 }
621
622 const CharUnits FEnd =
623 FOffset +
624 Context.toCharUnitsFromBits(
625 Types.ConvertTypeForMem(F->getType())->getPrimitiveSizeInBits()) -
626 CharUnits::One();
627 // If no overlap, continue.
628 if (End < FOffset || FEnd < StorageOffset)
629 continue;
630
631 // The desired load overlaps a non-bit-field member, bail out.
632 Conflict = true;
633 break;
634 }
635
636 if (Conflict)
637 continue;
638 // Write the new bit-field access parameters.
639 // As the storage offset now is defined as the number of elements from the
640 // start of the structure, we should divide the Offset by the element size.
641 Info.VolatileStorageOffset =
642 StorageOffset / Context.toCharUnitsFromBits(StorageSize).getQuantity();
643 Info.VolatileStorageSize = StorageSize;
644 Info.VolatileOffset = Offset;
645 }
646}
647
648void CGRecordLowering::accumulateVPtrs() {
649 if (Layout.hasOwnVFPtr())
650 Members.push_back(MemberInfo(CharUnits::Zero(), MemberInfo::VFPtr,
651 llvm::FunctionType::get(getIntNType(32), /*isVarArg=*/true)->
652 getPointerTo()->getPointerTo()));
653 if (Layout.hasOwnVBPtr())
654 Members.push_back(MemberInfo(Layout.getVBPtrOffset(), MemberInfo::VBPtr,
655 llvm::Type::getInt32PtrTy(Types.getLLVMContext())));
656}
657
658void CGRecordLowering::accumulateVBases() {
659 CharUnits ScissorOffset = Layout.getNonVirtualSize();
660 // In the itanium ABI, it's possible to place a vbase at a dsize that is
661 // smaller than the nvsize. Here we check to see if such a base is placed
662 // before the nvsize and set the scissor offset to that, instead of the
663 // nvsize.
664 if (isOverlappingVBaseABI())
665 for (const auto &Base : RD->vbases()) {
666 const CXXRecordDecl *BaseDecl = Base.getType()->getAsCXXRecordDecl();
667 if (BaseDecl->isEmpty())
668 continue;
669 // If the vbase is a primary virtual base of some base, then it doesn't
670 // get its own storage location but instead lives inside of that base.
671 if (Context.isNearlyEmpty(BaseDecl) && !hasOwnStorage(RD, BaseDecl))
672 continue;
673 ScissorOffset = std::min(ScissorOffset,
674 Layout.getVBaseClassOffset(BaseDecl));
675 }
676 Members.push_back(MemberInfo(ScissorOffset, MemberInfo::Scissor, nullptr,
677 RD));
678 for (const auto &Base : RD->vbases()) {
679 const CXXRecordDecl *BaseDecl = Base.getType()->getAsCXXRecordDecl();
680 if (BaseDecl->isEmpty())
681 continue;
682 CharUnits Offset = Layout.getVBaseClassOffset(BaseDecl);
683 // If the vbase is a primary virtual base of some base, then it doesn't
684 // get its own storage location but instead lives inside of that base.
685 if (isOverlappingVBaseABI() &&
686 Context.isNearlyEmpty(BaseDecl) &&
687 !hasOwnStorage(RD, BaseDecl)) {
688 Members.push_back(MemberInfo(Offset, MemberInfo::VBase, nullptr,
689 BaseDecl));
690 continue;
691 }
692 // If we've got a vtordisp, add it as a storage type.
693 if (Layout.getVBaseOffsetsMap().find(BaseDecl)->second.hasVtorDisp())
694 Members.push_back(StorageInfo(Offset - CharUnits::fromQuantity(4),
695 getIntNType(32)));
696 Members.push_back(MemberInfo(Offset, MemberInfo::VBase,
697 getStorageType(BaseDecl), BaseDecl));
698 }
699}
700
701bool CGRecordLowering::hasOwnStorage(const CXXRecordDecl *Decl,
702 const CXXRecordDecl *Query) {
703 const ASTRecordLayout &DeclLayout = Context.getASTRecordLayout(Decl);
704 if (DeclLayout.isPrimaryBaseVirtual() && DeclLayout.getPrimaryBase() == Query)
705 return false;
706 for (const auto &Base : Decl->bases())
707 if (!hasOwnStorage(Base.getType()->getAsCXXRecordDecl(), Query))
708 return false;
709 return true;
710}
711
712void CGRecordLowering::calculateZeroInit() {
713 for (std::vector<MemberInfo>::const_iterator Member = Members.begin(),
714 MemberEnd = Members.end();
715 IsZeroInitializableAsBase && Member != MemberEnd; ++Member) {
716 if (Member->Kind == MemberInfo::Field) {
717 if (!Member->FD || isZeroInitializable(Member->FD))
718 continue;
719 IsZeroInitializable = IsZeroInitializableAsBase = false;
720 } else if (Member->Kind == MemberInfo::Base ||
721 Member->Kind == MemberInfo::VBase) {
722 if (isZeroInitializable(Member->RD))
723 continue;
724 IsZeroInitializable = false;
725 if (Member->Kind == MemberInfo::Base)
726 IsZeroInitializableAsBase = false;
727 }
728 }
729}
730
731void CGRecordLowering::clipTailPadding() {
732 std::vector<MemberInfo>::iterator Prior = Members.begin();
733 CharUnits Tail = getSize(Prior->Data);
734 for (std::vector<MemberInfo>::iterator Member = Prior + 1,
735 MemberEnd = Members.end();
736 Member != MemberEnd; ++Member) {
737 // Only members with data and the scissor can cut into tail padding.
738 if (!Member->Data && Member->Kind != MemberInfo::Scissor)
739 continue;
740 if (Member->Offset < Tail) {
741 assert(Prior->Kind == MemberInfo::Field &&
742 "Only storage fields have tail padding!");
743 if (!Prior->FD || Prior->FD->isBitField())
744 Prior->Data = getByteArrayType(bitsToCharUnits(llvm::alignTo(
745 cast<llvm::IntegerType>(Prior->Data)->getIntegerBitWidth(), 8)));
746 else {
747 assert(Prior->FD->hasAttr<NoUniqueAddressAttr>() &&
748 "should not have reused this field's tail padding");
749 Prior->Data = getByteArrayType(
750 Context.getTypeInfoDataSizeInChars(Prior->FD->getType()).Width);
751 }
752 }
753 if (Member->Data)
754 Prior = Member;
755 Tail = Prior->Offset + getSize(Prior->Data);
756 }
757}
758
759void CGRecordLowering::determinePacked(bool NVBaseType) {
760 if (Packed)
761 return;
762 CharUnits Alignment = CharUnits::One();
763 CharUnits NVAlignment = CharUnits::One();
764 CharUnits NVSize =
765 !NVBaseType && RD ? Layout.getNonVirtualSize() : CharUnits::Zero();
766 for (std::vector<MemberInfo>::const_iterator Member = Members.begin(),
767 MemberEnd = Members.end();
768 Member != MemberEnd; ++Member) {
769 if (!Member->Data)
770 continue;
771 // If any member falls at an offset that it not a multiple of its alignment,
772 // then the entire record must be packed.
773 if (Member->Offset % getAlignment(Member->Data))
774 Packed = true;
775 if (Member->Offset < NVSize)
776 NVAlignment = std::max(NVAlignment, getAlignment(Member->Data));
777 Alignment = std::max(Alignment, getAlignment(Member->Data));
778 }
779 // If the size of the record (the capstone's offset) is not a multiple of the
780 // record's alignment, it must be packed.
781 if (Members.back().Offset % Alignment)
782 Packed = true;
783 // If the non-virtual sub-object is not a multiple of the non-virtual
784 // sub-object's alignment, it must be packed. We cannot have a packed
785 // non-virtual sub-object and an unpacked complete object or vise versa.
786 if (NVSize % NVAlignment)
787 Packed = true;
788 // Update the alignment of the sentinel.
789 if (!Packed)
790 Members.back().Data = getIntNType(Context.toBits(Alignment));
791}
792
793void CGRecordLowering::insertPadding() {
794 std::vector<std::pair<CharUnits, CharUnits> > Padding;
795 CharUnits Size = CharUnits::Zero();
796 for (std::vector<MemberInfo>::const_iterator Member = Members.begin(),
797 MemberEnd = Members.end();
798 Member != MemberEnd; ++Member) {
799 if (!Member->Data)
800 continue;
801 CharUnits Offset = Member->Offset;
802 assert(Offset >= Size);
803 // Insert padding if we need to.
804 if (Offset !=
805 Size.alignTo(Packed ? CharUnits::One() : getAlignment(Member->Data)))
806 Padding.push_back(std::make_pair(Size, Offset - Size));
807 Size = Offset + getSize(Member->Data);
808 }
809 if (Padding.empty())
810 return;
811 // Add the padding to the Members list and sort it.
812 for (std::vector<std::pair<CharUnits, CharUnits> >::const_iterator
813 Pad = Padding.begin(), PadEnd = Padding.end();
814 Pad != PadEnd; ++Pad)
815 Members.push_back(StorageInfo(Pad->first, getByteArrayType(Pad->second)));
816 llvm::stable_sort(Members);
817}
818
819void CGRecordLowering::fillOutputFields() {
820 for (std::vector<MemberInfo>::const_iterator Member = Members.begin(),
821 MemberEnd = Members.end();
822 Member != MemberEnd; ++Member) {
823 if (Member->Data)
824 FieldTypes.push_back(Member->Data);
825 if (Member->Kind == MemberInfo::Field) {
826 if (Member->FD)
827 Fields[Member->FD->getCanonicalDecl()] = FieldTypes.size() - 1;
828 // A field without storage must be a bitfield.
829 if (!Member->Data)
830 setBitFieldInfo(Member->FD, Member->Offset, FieldTypes.back());
831 } else if (Member->Kind == MemberInfo::Base)
832 NonVirtualBases[Member->RD] = FieldTypes.size() - 1;
833 else if (Member->Kind == MemberInfo::VBase)
834 VirtualBases[Member->RD] = FieldTypes.size() - 1;
835 }
836}
837
838CGBitFieldInfo CGBitFieldInfo::MakeInfo(CodeGenTypes &Types,
839 const FieldDecl *FD,
840 uint64_t Offset, uint64_t Size,
841 uint64_t StorageSize,
842 CharUnits StorageOffset) {
843 // This function is vestigial from CGRecordLayoutBuilder days but is still
844 // used in GCObjCRuntime.cpp. That usage has a "fixme" attached to it that
845 // when addressed will allow for the removal of this function.
846 llvm::Type *Ty = Types.ConvertTypeForMem(FD->getType());
847 CharUnits TypeSizeInBytes =
848 CharUnits::fromQuantity(Types.getDataLayout().getTypeAllocSize(Ty));
849 uint64_t TypeSizeInBits = Types.getContext().toBits(TypeSizeInBytes);
850
851 bool IsSigned = FD->getType()->isSignedIntegerOrEnumerationType();
852
853 if (Size > TypeSizeInBits) {
854 // We have a wide bit-field. The extra bits are only used for padding, so
855 // if we have a bitfield of type T, with size N:
856 //
857 // T t : N;
858 //
859 // We can just assume that it's:
860 //
861 // T t : sizeof(T);
862 //
863 Size = TypeSizeInBits;
864 }
865
866 // Reverse the bit offsets for big endian machines. Because we represent
867 // a bitfield as a single large integer load, we can imagine the bits
868 // counting from the most-significant-bit instead of the
869 // least-significant-bit.
870 if (Types.getDataLayout().isBigEndian()) {
871 Offset = StorageSize - (Offset + Size);
872 }
873
874 return CGBitFieldInfo(Offset, Size, IsSigned, StorageSize, StorageOffset);
875}
876
877std::unique_ptr<CGRecordLayout>
878CodeGenTypes::ComputeRecordLayout(const RecordDecl *D, llvm::StructType *Ty) {
879 CGRecordLowering Builder(*this, D, /*Packed=*/false);
880
881 Builder.lower(/*NonVirtualBaseType=*/false);
882
883 // If we're in C++, compute the base subobject type.
884 llvm::StructType *BaseTy = nullptr;
885 if (isa<CXXRecordDecl>(D) && !D->isUnion() && !D->hasAttr<FinalAttr>()) {
886 BaseTy = Ty;
887 if (Builder.Layout.getNonVirtualSize() != Builder.Layout.getSize()) {
888 CGRecordLowering BaseBuilder(*this, D, /*Packed=*/Builder.Packed);
889 BaseBuilder.lower(/*NonVirtualBaseType=*/true);
890 BaseTy = llvm::StructType::create(
891 getLLVMContext(), BaseBuilder.FieldTypes, "", BaseBuilder.Packed);
892 addRecordTypeName(D, BaseTy, ".base");
893 // BaseTy and Ty must agree on their packedness for getLLVMFieldNo to work
894 // on both of them with the same index.
895 assert(Builder.Packed == BaseBuilder.Packed &&
896 "Non-virtual and complete types must agree on packedness");
897 }
898 }
899
900 // Fill in the struct *after* computing the base type. Filling in the body
901 // signifies that the type is no longer opaque and record layout is complete,
902 // but we may need to recursively layout D while laying D out as a base type.
903 Ty->setBody(Builder.FieldTypes, Builder.Packed);
904
905 auto RL = std::make_unique<CGRecordLayout>(
906 Ty, BaseTy, (bool)Builder.IsZeroInitializable,
907 (bool)Builder.IsZeroInitializableAsBase);
908
909 RL->NonVirtualBases.swap(Builder.NonVirtualBases);
910 RL->CompleteObjectVirtualBases.swap(Builder.VirtualBases);
911
912 // Add all the field numbers.
913 RL->FieldInfo.swap(Builder.Fields);
914
915 // Add bitfield info.
916 RL->BitFields.swap(Builder.BitFields);
917
918 // Dump the layout, if requested.
919 if (getContext().getLangOpts().DumpRecordLayouts) {
920 llvm::outs() << "\n*** Dumping IRgen Record Layout\n";
921 llvm::outs() << "Record: ";
922 D->dump(llvm::outs());
923 llvm::outs() << "\nLayout: ";
924 RL->print(llvm::outs());
925 }
926
927#ifndef NDEBUG
928 // Verify that the computed LLVM struct size matches the AST layout size.
929 const ASTRecordLayout &Layout = getContext().getASTRecordLayout(D);
930
931 uint64_t TypeSizeInBits = getContext().toBits(Layout.getSize());
932 assert(TypeSizeInBits == getDataLayout().getTypeAllocSizeInBits(Ty) &&
933 "Type size mismatch!");
934
935 if (BaseTy) {
936 CharUnits NonVirtualSize = Layout.getNonVirtualSize();
937
938 uint64_t AlignedNonVirtualTypeSizeInBits =
939 getContext().toBits(NonVirtualSize);
940
941 assert(AlignedNonVirtualTypeSizeInBits ==
942 getDataLayout().getTypeAllocSizeInBits(BaseTy) &&
943 "Type size mismatch!");
944 }
945
946 // Verify that the LLVM and AST field offsets agree.
947 llvm::StructType *ST = RL->getLLVMType();
948 const llvm::StructLayout *SL = getDataLayout().getStructLayout(ST);
949
950 const ASTRecordLayout &AST_RL = getContext().getASTRecordLayout(D);
951 RecordDecl::field_iterator it = D->field_begin();
952 for (unsigned i = 0, e = AST_RL.getFieldCount(); i != e; ++i, ++it) {
953 const FieldDecl *FD = *it;
954
955 // Ignore zero-sized fields.
956 if (FD->isZeroSize(getContext()))
957 continue;
958
959 // For non-bit-fields, just check that the LLVM struct offset matches the
960 // AST offset.
961 if (!FD->isBitField()) {
962 unsigned FieldNo = RL->getLLVMFieldNo(FD);
963 assert(AST_RL.getFieldOffset(i) == SL->getElementOffsetInBits(FieldNo) &&
964 "Invalid field offset!");
965 continue;
966 }
967
968 // Ignore unnamed bit-fields.
969 if (!FD->getDeclName())
970 continue;
971
972 const CGBitFieldInfo &Info = RL->getBitFieldInfo(FD);
973 llvm::Type *ElementTy = ST->getTypeAtIndex(RL->getLLVMFieldNo(FD));
974
975 // Unions have overlapping elements dictating their layout, but for
976 // non-unions we can verify that this section of the layout is the exact
977 // expected size.
978 if (D->isUnion()) {
979 // For unions we verify that the start is zero and the size
980 // is in-bounds. However, on BE systems, the offset may be non-zero, but
981 // the size + offset should match the storage size in that case as it
982 // "starts" at the back.
983 if (getDataLayout().isBigEndian())
984 assert(static_cast<unsigned>(Info.Offset + Info.Size) ==
985 Info.StorageSize &&
986 "Big endian union bitfield does not end at the back");
987 else
988 assert(Info.Offset == 0 &&
989 "Little endian union bitfield with a non-zero offset");
990 assert(Info.StorageSize <= SL->getSizeInBits() &&
991 "Union not large enough for bitfield storage");
992 } else {
993 assert((Info.StorageSize ==
994 getDataLayout().getTypeAllocSizeInBits(ElementTy) ||
995 Info.VolatileStorageSize ==
996 getDataLayout().getTypeAllocSizeInBits(ElementTy)) &&
997 "Storage size does not match the element type size");
998 }
999 assert(Info.Size > 0 && "Empty bitfield!");
1000 assert(static_cast<unsigned>(Info.Offset) + Info.Size <= Info.StorageSize &&
1001 "Bitfield outside of its allocated storage");
1002 }
1003#endif
1004
1005 return RL;
1006}
1007
1008void CGRecordLayout::print(raw_ostream &OS) const {
1009 OS << "<CGRecordLayout\n";
1010 OS << " LLVMType:" << *CompleteObjectType << "\n";
1011 if (BaseSubobjectType)
1012 OS << " NonVirtualBaseLLVMType:" << *BaseSubobjectType << "\n";
1013 OS << " IsZeroInitializable:" << IsZeroInitializable << "\n";
1014 OS << " BitFields:[\n";
1015
1016 // Print bit-field infos in declaration order.
1017 std::vector<std::pair<unsigned, const CGBitFieldInfo*> > BFIs;
1018 for (llvm::DenseMap<const FieldDecl*, CGBitFieldInfo>::const_iterator
1019 it = BitFields.begin(), ie = BitFields.end();
1020 it != ie; ++it) {
1021 const RecordDecl *RD = it->first->getParent();
1022 unsigned Index = 0;
1023 for (RecordDecl::field_iterator
1024 it2 = RD->field_begin(); *it2 != it->first; ++it2)
1025 ++Index;
1026 BFIs.push_back(std::make_pair(Index, &it->second));
1027 }
1028 llvm::array_pod_sort(BFIs.begin(), BFIs.end());
1029 for (unsigned i = 0, e = BFIs.size(); i != e; ++i) {
1030 OS.indent(4);
1031 BFIs[i].second->print(OS);
1032 OS << "\n";
1033 }
1034
1035 OS << "]>\n";
1036}
1037
1038LLVM_DUMP_METHOD void CGRecordLayout::dump() const {
1039 print(llvm::errs());
1040}
1041
1042void CGBitFieldInfo::print(raw_ostream &OS) const {
1043 OS << "<CGBitFieldInfo"
1044 << " Offset:" << Offset << " Size:" << Size << " IsSigned:" << IsSigned
1045 << " StorageSize:" << StorageSize
1046 << " StorageOffset:" << StorageOffset.getQuantity()
1047 << " VolatileOffset:" << VolatileOffset
1048 << " VolatileStorageSize:" << VolatileStorageSize
1049 << " VolatileStorageOffset:" << VolatileStorageOffset.getQuantity() << ">";
1050}
1051
1052LLVM_DUMP_METHOD void CGBitFieldInfo::dump() const {
1053 print(llvm::errs());
1054}
1055