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 "ABIInfoImpl.h"
14#include "CGCXXABI.h"
15#include "CGRecordLayout.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/Basic/CodeGenOptions.h"
24#include "clang/CodeGenUtils/RecordLayoutUtils.h"
25#include "clang/CodeGenUtils/TargetUtils.h"
26#include "llvm/IR/DataLayout.h"
27#include "llvm/IR/DerivedTypes.h"
28#include "llvm/IR/Type.h"
29#include "llvm/Support/Debug.h"
30#include "llvm/Support/MathExtras.h"
31#include "llvm/Support/raw_ostream.h"
32using namespace clang;
33using namespace CodeGen;
34
35namespace {
36/// The CGRecordLowering is responsible for lowering an ASTRecordLayout to an
37/// llvm::Type. Some of the lowering is straightforward, some is not. Here we
38/// detail some of the complexities and weirdnesses here.
39/// * LLVM does not have unions - Unions can, in theory be represented by any
40/// llvm::Type with correct size. We choose a field via a specific heuristic
41/// and add padding if necessary.
42/// * LLVM does not have bitfields - Bitfields are collected into contiguous
43/// runs and allocated as a single storage type for the run. ASTRecordLayout
44/// contains enough information to determine where the runs break. Microsoft
45/// and Itanium follow different rules and use different codepaths.
46/// * It is desired that, when possible, bitfields use the appropriate iN type
47/// when lowered to llvm types. For example unsigned x : 24 gets lowered to
48/// i24. This isn't always possible because i24 has storage size of 32 bit
49/// and if it is possible to use that extra byte of padding we must use [i8 x
50/// 3] instead of i24. This is computed when accumulating bitfields in
51/// accumulateBitfields.
52/// C++ examples that require clipping:
53/// struct { int a : 24; char b; }; // a must be clipped, b goes at offset 3
54/// struct A { int a : 24; ~A(); }; // a must be clipped because:
55/// struct B : A { char b; }; // b goes at offset 3
56/// * The allocation of bitfield access units is described in more detail in
57/// CGRecordLowering::accumulateBitFields.
58/// * Clang ignores 0 sized bitfields and 0 sized bases but *not* zero sized
59/// fields. The existing asserts suggest that LLVM assumes that *every* field
60/// has an underlying storage type. Therefore empty structures containing
61/// zero sized subobjects such as empty records or zero sized arrays still get
62/// a zero sized (empty struct) storage type.
63/// * Clang reads the complete type rather than the base type when generating
64/// code to access fields. Bitfields in tail position with tail padding may
65/// be clipped in the base class but not the complete class (we may discover
66/// that the tail padding is not used in the complete class.) However,
67/// because LLVM reads from the complete type it can generate incorrect code
68/// if we do not clip the tail padding off of the bitfield in the complete
69/// layout.
70/// * Itanium allows nearly empty primary virtual bases. These bases don't get
71/// get their own storage because they're laid out as part of another base
72/// or at the beginning of the structure. Determining if a VBase actually
73/// gets storage awkwardly involves a walk of all bases.
74/// * VFPtrs and VBPtrs do *not* make a record NotZeroInitializable.
75struct CGRecordLowering {
76 // MemberInfo is a helper structure that contains information about a record
77 // member. In additional to the standard member types, there exists a
78 // sentinel member type that ensures correct rounding.
79 struct MemberInfo {
80 CharUnits Offset;
81 enum InfoKind { VFPtr, VBPtr, Field, Base, VBase } Kind;
82 llvm::Type *Data;
83 union {
84 const FieldDecl *FD;
85 const CXXRecordDecl *RD;
86 };
87 MemberInfo(CharUnits Offset, InfoKind Kind, llvm::Type *Data,
88 const FieldDecl *FD = nullptr)
89 : Offset(Offset), Kind(Kind), Data(Data), FD(FD) {}
90 MemberInfo(CharUnits Offset, InfoKind Kind, llvm::Type *Data,
91 const CXXRecordDecl *RD)
92 : Offset(Offset), Kind(Kind), Data(Data), RD(RD) {}
93 // MemberInfos are sorted so we define a < operator.
94 bool operator <(const MemberInfo& a) const { return Offset < a.Offset; }
95 };
96 // The constructor.
97 CGRecordLowering(CodeGenTypes &Types, const RecordDecl *D, bool Packed);
98 // Short helper routines.
99 /// Constructs a MemberInfo instance from an offset and llvm::Type *.
100 static MemberInfo StorageInfo(CharUnits Offset, llvm::Type *Data) {
101 return MemberInfo(Offset, MemberInfo::Field, Data);
102 }
103
104 /// Helper function to check if the target machine is BigEndian.
105 bool isBE() const { return Context.getTargetInfo().isBigEndian(); }
106
107 /// Wraps llvm::Type::getIntNTy with some implicit arguments.
108 llvm::Type *getIntNType(uint64_t NumBits) const {
109 unsigned AlignedBits = llvm::alignTo(Value: NumBits, Align: Context.getCharWidth());
110 return llvm::Type::getIntNTy(C&: Types.getLLVMContext(), N: AlignedBits);
111 }
112 /// Get the LLVM type sized as one character unit.
113 llvm::Type *getCharType() const {
114 return llvm::Type::getIntNTy(C&: Types.getLLVMContext(),
115 N: Context.getCharWidth());
116 }
117 /// Gets an llvm type of size NumChars and alignment 1.
118 llvm::Type *getByteArrayType(CharUnits NumChars) const {
119 assert(!NumChars.isZero() && "Empty byte arrays aren't allowed.");
120 llvm::Type *Type = getCharType();
121 return NumChars == CharUnits::One() ? Type :
122 (llvm::Type *)llvm::ArrayType::get(ElementType: Type, NumElements: NumChars.getQuantity());
123 }
124 /// Gets the storage type for a field decl and handles storage
125 /// for itanium bitfields that are smaller than their declared type.
126 llvm::Type *getStorageType(const FieldDecl *FD) const {
127 llvm::Type *Type = Types.ConvertTypeForMem(T: FD->getType());
128 if (!FD->isBitField()) return Type;
129 if (CodeGenUtils::isDiscreteBitFieldABI(Ctx: Context, RD: D))
130 return Type;
131 return getIntNType(NumBits: std::min(a: FD->getBitWidthValue(),
132 b: (unsigned)Context.toBits(CharSize: getSize(Type))));
133 }
134 /// Gets the llvm Basesubobject type from a CXXRecordDecl.
135 llvm::Type *getStorageType(const CXXRecordDecl *RD) const {
136 return Types.getCGRecordLayout(RD).getBaseSubobjectLLVMType();
137 }
138 CharUnits bitsToCharUnits(uint64_t BitOffset) const {
139 return Context.toCharUnitsFromBits(BitSize: BitOffset);
140 }
141 CharUnits getSize(llvm::Type *Type) const {
142 return CharUnits::fromQuantity(Quantity: DataLayout.getTypeAllocSize(Ty: Type));
143 }
144 CharUnits getAlignment(llvm::Type *Type) const {
145 return CharUnits::fromQuantity(Quantity: DataLayout.getABITypeAlign(Ty: Type));
146 }
147 bool isZeroInitializable(const FieldDecl *FD) const {
148 return Types.isZeroInitializable(T: FD->getType());
149 }
150 bool isZeroInitializable(const RecordDecl *RD) const {
151 return Types.isZeroInitializable(RD);
152 }
153 void appendPaddingBytes(CharUnits Size) {
154 if (!Size.isZero())
155 FieldTypes.push_back(Elt: getByteArrayType(NumChars: Size));
156 }
157 uint64_t getFieldBitOffset(const FieldDecl *FD) const {
158 return Layout.getFieldOffset(FieldNo: FD->getFieldIndex());
159 }
160 // Layout routines.
161 void setBitFieldInfo(const FieldDecl *FD, CharUnits StartOffset,
162 llvm::Type *StorageType);
163 /// Lowers an ASTRecordLayout to a llvm type.
164 void lower(bool NonVirtualBaseType);
165 void lowerUnion(bool isNonVirtualBaseType);
166 void accumulateFields(bool isNonVirtualBaseType);
167 RecordDecl::field_iterator
168 accumulateBitFields(bool isNonVirtualBaseType,
169 RecordDecl::field_iterator Field,
170 RecordDecl::field_iterator FieldEnd);
171 void computeVolatileBitfields();
172 void accumulateBases();
173 void accumulateVPtrs();
174 void accumulateVBases();
175 void calculateZeroInit();
176 CharUnits calculateTailClippingOffset(bool isNonVirtualBaseType) const;
177 void checkBitfieldClipping(bool isNonVirtualBaseType) const;
178 /// Determines if we need a packed llvm struct.
179 void determinePacked(bool NVBaseType);
180 /// Inserts padding everywhere it's needed.
181 void insertPadding();
182 /// Fills out the structures that are ultimately consumed.
183 void fillOutputFields();
184 // Input memoization fields.
185 CodeGenTypes &Types;
186 const ASTContext &Context;
187 const RecordDecl *D;
188 const CXXRecordDecl *RD;
189 const ASTRecordLayout &Layout;
190 const llvm::DataLayout &DataLayout;
191 // Helpful intermediate data-structures.
192 std::vector<MemberInfo> Members;
193 // Output fields, consumed by CodeGenTypes::ComputeRecordLayout.
194 SmallVector<llvm::Type *, 16> FieldTypes;
195 llvm::DenseMap<const FieldDecl *, unsigned> Fields;
196 llvm::DenseMap<const FieldDecl *, CGBitFieldInfo> BitFields;
197 llvm::DenseMap<const CXXRecordDecl *, unsigned> NonVirtualBases;
198 llvm::DenseMap<const CXXRecordDecl *, unsigned> VirtualBases;
199 bool IsZeroInitializable : 1;
200 bool IsZeroInitializableAsBase : 1;
201 bool Packed : 1;
202private:
203 CGRecordLowering(const CGRecordLowering &) = delete;
204 void operator =(const CGRecordLowering &) = delete;
205};
206} // namespace {
207
208CGRecordLowering::CGRecordLowering(CodeGenTypes &Types, const RecordDecl *D,
209 bool Packed)
210 : Types(Types), Context(Types.getContext()), D(D),
211 RD(dyn_cast<CXXRecordDecl>(Val: D)),
212 Layout(Types.getContext().getASTRecordLayout(D)),
213 DataLayout(Types.getDataLayout()), IsZeroInitializable(true),
214 IsZeroInitializableAsBase(true), Packed(Packed) {}
215
216void CGRecordLowering::setBitFieldInfo(
217 const FieldDecl *FD, CharUnits StartOffset, llvm::Type *StorageType) {
218 uint64_t Offset = getFieldBitOffset(FD) - Context.toBits(CharSize: StartOffset);
219 uint64_t StorageSize =
220 DataLayout.getTypeAllocSizeInBits(Ty: StorageType).getFixedValue();
221 // CGBitFieldInfo::MakeInfo does the same clamp
222 llvm::Type *FieldType = Types.ConvertTypeForMem(T: FD->getType());
223 uint64_t Size =
224 std::min(l: {uint64_t(FD->getBitWidthValue()), StorageSize,
225 DataLayout.getTypeAllocSizeInBits(Ty: FieldType).getFixedValue()});
226 // Reverse the bit offsets for big endian machines. Because we represent
227 // a bitfield as a single large integer load, we can imagine the bits
228 // counting from the most-significant-bit instead of the
229 // least-significant-bit.
230 if (DataLayout.isBigEndian())
231 Offset = StorageSize - (Offset + Size);
232
233 BitFields[FD->getCanonicalDecl()] = CGBitFieldInfo(
234 Offset, Size, FD->getType()->isSignedIntegerOrEnumerationType(),
235 StorageSize, StartOffset);
236}
237
238void CGRecordLowering::lower(bool NVBaseType) {
239 // The lowering process implemented in this function takes a variety of
240 // carefully ordered phases.
241 // 1) Store all members (fields and bases) in a list and sort them by offset.
242 // 2) Add a 1-byte capstone member at the Size of the structure.
243 // 3) Clip bitfield storages members if their tail padding is or might be
244 // used by another field or base. The clipping process uses the capstone
245 // by treating it as another object that occurs after the record.
246 // 4) Determine if the llvm-struct requires packing. It's important that this
247 // phase occur after clipping, because clipping changes the llvm type.
248 // This phase reads the offset of the capstone when determining packedness
249 // and updates the alignment of the capstone to be equal of the alignment
250 // of the record after doing so.
251 // 5) Insert padding everywhere it is needed. This phase requires 'Packed' to
252 // have been computed and needs to know the alignment of the record in
253 // order to understand if explicit tail padding is needed.
254 // 6) Remove the capstone, we don't need it anymore.
255 // 7) Determine if this record can be zero-initialized. This phase could have
256 // been placed anywhere after phase 1.
257 // 8) Format the complete list of members in a way that can be consumed by
258 // CodeGenTypes::ComputeRecordLayout.
259 CharUnits Size = NVBaseType ? Layout.getNonVirtualSize() : Layout.getSize();
260 if (D->isUnion()) {
261 lowerUnion(isNonVirtualBaseType: NVBaseType);
262 computeVolatileBitfields();
263 return;
264 }
265 accumulateFields(isNonVirtualBaseType: NVBaseType);
266 // RD implies C++.
267 if (RD) {
268 accumulateVPtrs();
269 accumulateBases();
270 if (Members.empty()) {
271 appendPaddingBytes(Size);
272 computeVolatileBitfields();
273 return;
274 }
275 if (!NVBaseType)
276 accumulateVBases();
277 }
278 llvm::stable_sort(Range&: Members);
279 checkBitfieldClipping(isNonVirtualBaseType: NVBaseType);
280 Members.push_back(x: StorageInfo(Offset: Size, Data: getIntNType(NumBits: 8)));
281 determinePacked(NVBaseType);
282 insertPadding();
283 Members.pop_back();
284 calculateZeroInit();
285 fillOutputFields();
286 computeVolatileBitfields();
287}
288
289void CGRecordLowering::lowerUnion(bool isNonVirtualBaseType) {
290 CharUnits LayoutSize =
291 isNonVirtualBaseType ? Layout.getDataSize() : Layout.getSize();
292 llvm::Type *StorageType = nullptr;
293 bool SeenNamedMember = false;
294 // Iterate through the fields setting bitFieldInfo and the Fields array. Also
295 // locate the "most appropriate" storage type. The heuristic for finding the
296 // storage type isn't necessary, the first (non-0-length-bitfield) field's
297 // type would work fine and be simpler but would be different than what we've
298 // been doing and cause lit tests to change.
299 for (const auto *Field : D->fields()) {
300 if (Field->isBitField()) {
301 if (Field->isZeroLengthBitField())
302 continue;
303 llvm::Type *FieldType = getStorageType(FD: Field);
304 if (LayoutSize < getSize(Type: FieldType))
305 FieldType = getByteArrayType(NumChars: LayoutSize);
306 setBitFieldInfo(FD: Field, StartOffset: CharUnits::Zero(), StorageType: FieldType);
307 }
308 Fields[Field->getCanonicalDecl()] = 0;
309 llvm::Type *FieldType = getStorageType(FD: Field);
310 // Compute zero-initializable status.
311 // This union might not be zero initialized: it may contain a pointer to
312 // data member which might have some exotic initialization sequence.
313 // If this is the case, then we aught not to try and come up with a "better"
314 // type, it might not be very easy to come up with a Constant which
315 // correctly initializes it.
316 if (!SeenNamedMember) {
317 SeenNamedMember = Field->getIdentifier();
318 if (!SeenNamedMember)
319 if (const auto *FieldRD = Field->getType()->getAsRecordDecl())
320 SeenNamedMember = FieldRD->findFirstNamedDataMember();
321 if (SeenNamedMember && !isZeroInitializable(FD: Field)) {
322 IsZeroInitializable = IsZeroInitializableAsBase = false;
323 StorageType = FieldType;
324 }
325 }
326 // Because our union isn't zero initializable, we won't be getting a better
327 // storage type.
328 if (!IsZeroInitializable)
329 continue;
330 // Conditionally update our storage type if we've got a new "better" one.
331 if (!StorageType ||
332 getAlignment(Type: FieldType) > getAlignment(Type: StorageType) ||
333 (getAlignment(Type: FieldType) == getAlignment(Type: StorageType) &&
334 getSize(Type: FieldType) > getSize(Type: StorageType)))
335 StorageType = FieldType;
336 }
337 // If we have no storage type just pad to the appropriate size and return.
338 if (!StorageType)
339 return appendPaddingBytes(Size: LayoutSize);
340 // If our storage size was bigger than our required size (can happen in the
341 // case of packed bitfields on Itanium) then just use an I8 array.
342 if (LayoutSize < getSize(Type: StorageType))
343 StorageType = getByteArrayType(NumChars: LayoutSize);
344 FieldTypes.push_back(Elt: StorageType);
345 appendPaddingBytes(Size: LayoutSize - getSize(Type: StorageType));
346 // Set packed if we need it.
347 const auto StorageAlignment = getAlignment(Type: StorageType);
348 assert((Layout.getSize().isMultipleOf(StorageAlignment) ||
349 !Layout.getDataSize().isMultipleOf(StorageAlignment)) &&
350 "Union's standard layout and no_unique_address layout must agree on "
351 "packedness");
352 if (!Layout.getDataSize().isMultipleOf(N: StorageAlignment))
353 Packed = true;
354}
355
356void CGRecordLowering::accumulateFields(bool isNonVirtualBaseType) {
357 for (RecordDecl::field_iterator Field = D->field_begin(),
358 FieldEnd = D->field_end();
359 Field != FieldEnd;) {
360 if (Field->isBitField()) {
361 Field = accumulateBitFields(isNonVirtualBaseType, Field, FieldEnd);
362 assert((Field == FieldEnd || !Field->isBitField()) &&
363 "Failed to accumulate all the bitfields");
364 } else if (CodeGenUtils::isEmptyFieldForLayout(Ctx: Context, FD: *Field)) {
365 // Empty fields have no storage.
366 ++Field;
367 } else {
368 // Use base subobject layout for the potentially-overlapping field,
369 // as it is done in RecordLayoutBuilder
370 Members.push_back(x: MemberInfo(
371 bitsToCharUnits(BitOffset: getFieldBitOffset(FD: *Field)), MemberInfo::Field,
372 Field->isPotentiallyOverlapping()
373 ? getStorageType(RD: Field->getType()->getAsCXXRecordDecl())
374 : getStorageType(FD: *Field),
375 *Field));
376 ++Field;
377 }
378 }
379}
380
381// Create members for bitfields. Field is a bitfield, and FieldEnd is the end
382// iterator of the record. Return the first non-bitfield encountered. We need
383// to know whether this is the base or complete layout, as virtual bases could
384// affect the upper bound of bitfield access unit allocation.
385RecordDecl::field_iterator
386CGRecordLowering::accumulateBitFields(bool isNonVirtualBaseType,
387 RecordDecl::field_iterator Field,
388 RecordDecl::field_iterator FieldEnd) {
389 if (CodeGenUtils::isDiscreteBitFieldABI(Ctx: Context, RD: D)) {
390 // Run stores the first element of the current run of bitfields. FieldEnd is
391 // used as a special value to note that we don't have a current run. A
392 // bitfield run is a contiguous collection of bitfields that can be stored
393 // in the same storage block. Zero-sized bitfields and bitfields that would
394 // cross an alignment boundary break a run and start a new one.
395 RecordDecl::field_iterator Run = FieldEnd;
396 // Tail is the offset of the first bit off the end of the current run. It's
397 // used to determine if the ASTRecordLayout is treating these two bitfields
398 // as contiguous. StartBitOffset is offset of the beginning of the Run.
399 uint64_t StartBitOffset, Tail = 0;
400 for (; Field != FieldEnd && Field->isBitField(); ++Field) {
401 // Zero-width bitfields end runs.
402 if (Field->isZeroLengthBitField()) {
403 Run = FieldEnd;
404 continue;
405 }
406 uint64_t BitOffset = getFieldBitOffset(FD: *Field);
407 llvm::Type *Type = Types.ConvertTypeForMem(T: Field->getType());
408 // If we don't have a run yet, or don't live within the previous run's
409 // allocated storage then we allocate some storage and start a new run.
410 if (Run == FieldEnd || BitOffset >= Tail) {
411 Run = Field;
412 StartBitOffset = BitOffset;
413 Tail = StartBitOffset + DataLayout.getTypeAllocSizeInBits(Ty: Type);
414 // Add the storage member to the record. This must be added to the
415 // record before the bitfield members so that it gets laid out before
416 // the bitfields it contains get laid out.
417 Members.push_back(x: StorageInfo(Offset: bitsToCharUnits(BitOffset: StartBitOffset), Data: Type));
418 }
419 // Bitfields get the offset of their storage but come afterward and remain
420 // there after a stable sort.
421 Members.push_back(x: MemberInfo(bitsToCharUnits(BitOffset: StartBitOffset),
422 MemberInfo::Field, nullptr, *Field));
423 }
424 return Field;
425 }
426
427 // The SysV ABI can overlap bitfield storage units with both other bitfield
428 // storage units /and/ other non-bitfield data members. Accessing a sequence
429 // of bitfields mustn't interfere with adjacent non-bitfields -- they're
430 // permitted to be accessed in separate threads for instance.
431
432 // We split runs of bit-fields into a sequence of "access units". When we emit
433 // a load or store of a bit-field, we'll load/store the entire containing
434 // access unit. As mentioned, the standard requires that these loads and
435 // stores must not interfere with accesses to other memory locations, and it
436 // defines the bit-field's memory location as the current run of
437 // non-zero-width bit-fields. So an access unit must never overlap with
438 // non-bit-field storage or cross a zero-width bit-field. Otherwise, we're
439 // free to draw the lines as we see fit.
440
441 // Drawing these lines well can be complicated. LLVM generally can't modify a
442 // program to access memory that it didn't before, so using very narrow access
443 // units can prevent the compiler from using optimal access patterns. For
444 // example, suppose a run of bit-fields occupies four bytes in a struct. If we
445 // split that into four 1-byte access units, then a sequence of assignments
446 // that doesn't touch all four bytes may have to be emitted with multiple
447 // 8-bit stores instead of a single 32-bit store. On the other hand, if we use
448 // very wide access units, we may find ourselves emitting accesses to
449 // bit-fields we didn't really need to touch, just because LLVM was unable to
450 // clean up after us.
451
452 // It is desirable to have access units be aligned powers of 2 no larger than
453 // a register. (On non-strict alignment ISAs, the alignment requirement can be
454 // dropped.) A three byte access unit will be accessed using 2-byte and 1-byte
455 // accesses and bit manipulation. If no bitfield straddles across the two
456 // separate accesses, it is better to have separate 2-byte and 1-byte access
457 // units, as then LLVM will not generate unnecessary memory accesses, or bit
458 // manipulation. Similarly, on a strict-alignment architecture, it is better
459 // to keep access-units naturally aligned, to avoid similar bit
460 // manipulation synthesizing larger unaligned accesses.
461
462 // Bitfields that share parts of a single byte are, of necessity, placed in
463 // the same access unit. That unit will encompass a consecutive run where
464 // adjacent bitfields share parts of a byte. (The first bitfield of such an
465 // access unit will start at the beginning of a byte.)
466
467 // We then try and accumulate adjacent access units when the combined unit is
468 // naturally sized, no larger than a register, and (on a strict alignment
469 // ISA), naturally aligned. Note that this requires lookahead to one or more
470 // subsequent access units. For instance, consider a 2-byte access-unit
471 // followed by 2 1-byte units. We can merge that into a 4-byte access-unit,
472 // but we would not want to merge a 2-byte followed by a single 1-byte (and no
473 // available tail padding). We keep track of the best access unit seen so far,
474 // and use that when we determine we cannot accumulate any more. Then we start
475 // again at the bitfield following that best one.
476
477 // The accumulation is also prevented when:
478 // *) it would cross a character-aigned zero-width bitfield, or
479 // *) fine-grained bitfield access option is in effect.
480
481 CharUnits RegSize =
482 bitsToCharUnits(BitOffset: Context.getTargetInfo().getRegisterWidth());
483 unsigned CharBits = Context.getCharWidth();
484
485 // Limit of useable tail padding at end of the record. Computed lazily and
486 // cached here.
487 CharUnits ScissorOffset = CharUnits::Zero();
488
489 // Data about the start of the span we're accumulating to create an access
490 // unit from. Begin is the first bitfield of the span. If Begin is FieldEnd,
491 // we've not got a current span. The span starts at the BeginOffset character
492 // boundary. BitSizeSinceBegin is the size (in bits) of the span -- this might
493 // include padding when we've advanced to a subsequent bitfield run.
494 RecordDecl::field_iterator Begin = FieldEnd;
495 CharUnits BeginOffset;
496 uint64_t BitSizeSinceBegin;
497
498 // The (non-inclusive) end of the largest acceptable access unit we've found
499 // since Begin. If this is Begin, we're gathering the initial set of bitfields
500 // of a new span. BestEndOffset is the end of that acceptable access unit --
501 // it might extend beyond the last character of the bitfield run, using
502 // available padding characters.
503 RecordDecl::field_iterator BestEnd = Begin;
504 CharUnits BestEndOffset;
505 bool BestClipped; // Whether the representation must be in a byte array.
506
507 for (;;) {
508 // AtAlignedBoundary is true iff Field is the (potential) start of a new
509 // span (or the end of the bitfields). When true, LimitOffset is the
510 // character offset of that span and Barrier indicates whether the new
511 // span cannot be merged into the current one.
512 bool AtAlignedBoundary = false;
513 bool Barrier = false;
514
515 if (Field != FieldEnd && Field->isBitField()) {
516 uint64_t BitOffset = getFieldBitOffset(FD: *Field);
517 if (Begin == FieldEnd) {
518 // Beginning a new span.
519 Begin = Field;
520 BestEnd = Begin;
521
522 assert((BitOffset % CharBits) == 0 && "Not at start of char");
523 BeginOffset = bitsToCharUnits(BitOffset);
524 BitSizeSinceBegin = 0;
525 } else if ((BitOffset % CharBits) != 0) {
526 // Bitfield occupies the same character as previous bitfield, it must be
527 // part of the same span. This can include zero-length bitfields, should
528 // the target not align them to character boundaries. Such non-alignment
529 // is at variance with the standards, which require zero-length
530 // bitfields be a barrier between access units. But of course we can't
531 // achieve that in the middle of a character.
532 assert(BitOffset == Context.toBits(BeginOffset) + BitSizeSinceBegin &&
533 "Concatenating non-contiguous bitfields");
534 } else {
535 // Bitfield potentially begins a new span. This includes zero-length
536 // bitfields on non-aligning targets that lie at character boundaries
537 // (those are barriers to merging).
538 if (Field->isZeroLengthBitField())
539 Barrier = true;
540 AtAlignedBoundary = true;
541 }
542 } else {
543 // We've reached the end of the bitfield run. Either we're done, or this
544 // is a barrier for the current span.
545 if (Begin == FieldEnd)
546 break;
547
548 Barrier = true;
549 AtAlignedBoundary = true;
550 }
551
552 // InstallBest indicates whether we should create an access unit for the
553 // current best span: fields [Begin, BestEnd) occupying characters
554 // [BeginOffset, BestEndOffset).
555 bool InstallBest = false;
556 if (AtAlignedBoundary) {
557 // Field is the start of a new span or the end of the bitfields. The
558 // just-seen span now extends to BitSizeSinceBegin.
559
560 // Determine if we can accumulate that just-seen span into the current
561 // accumulation.
562 CharUnits AccessSize = bitsToCharUnits(BitOffset: BitSizeSinceBegin + CharBits - 1);
563 if (BestEnd == Begin) {
564 // This is the initial run at the start of a new span. By definition,
565 // this is the best seen so far.
566 BestEnd = Field;
567 BestEndOffset = BeginOffset + AccessSize;
568 // Assume clipped until proven not below.
569 BestClipped = true;
570 if (!BitSizeSinceBegin)
571 // A zero-sized initial span -- this will install nothing and reset
572 // for another.
573 InstallBest = true;
574 } else if (AccessSize > RegSize)
575 // Accumulating the just-seen span would create a multi-register access
576 // unit, which would increase register pressure.
577 InstallBest = true;
578
579 if (!InstallBest) {
580 // Determine if accumulating the just-seen span will create an expensive
581 // access unit or not.
582 llvm::Type *Type = getIntNType(NumBits: Context.toBits(CharSize: AccessSize));
583 if (!Context.getTargetInfo().hasCheapUnalignedBitFieldAccess()) {
584 // Unaligned accesses are expensive. Only accumulate if the new unit
585 // is naturally aligned. Otherwise install the best we have, which is
586 // either the initial access unit (can't do better), or a naturally
587 // aligned accumulation (since we would have already installed it if
588 // it wasn't naturally aligned).
589 CharUnits Align = getAlignment(Type);
590 if (Align > Layout.getAlignment())
591 // The alignment required is greater than the containing structure
592 // itself.
593 InstallBest = true;
594 else if (!BeginOffset.isMultipleOf(N: Align))
595 // The access unit is not at a naturally aligned offset within the
596 // structure.
597 InstallBest = true;
598
599 if (InstallBest && BestEnd == Field)
600 // We're installing the first span, whose clipping was presumed
601 // above. Compute it correctly.
602 if (getSize(Type) == AccessSize)
603 BestClipped = false;
604 }
605
606 if (!InstallBest) {
607 // Find the next used storage offset to determine what the limit of
608 // the current span is. That's either the offset of the next field
609 // with storage (which might be Field itself) or the end of the
610 // non-reusable tail padding.
611 CharUnits LimitOffset;
612 for (auto Probe = Field; Probe != FieldEnd; ++Probe)
613 if (!CodeGenUtils::isEmptyFieldForLayout(Ctx: Context, FD: *Probe)) {
614 // A member with storage sets the limit.
615 assert((getFieldBitOffset(*Probe) % CharBits) == 0 &&
616 "Next storage is not byte-aligned");
617 LimitOffset = bitsToCharUnits(BitOffset: getFieldBitOffset(FD: *Probe));
618 goto FoundLimit;
619 }
620 // We reached the end of the fields, determine the bounds of useable
621 // tail padding. As this can be complex for C++, we cache the result.
622 if (ScissorOffset.isZero()) {
623 ScissorOffset = calculateTailClippingOffset(isNonVirtualBaseType);
624 assert(!ScissorOffset.isZero() && "Tail clipping at zero");
625 }
626
627 LimitOffset = ScissorOffset;
628 FoundLimit:;
629
630 CharUnits TypeSize = getSize(Type);
631 if (BeginOffset + TypeSize <= LimitOffset) {
632 // There is space before LimitOffset to create a naturally-sized
633 // access unit.
634 BestEndOffset = BeginOffset + TypeSize;
635 BestEnd = Field;
636 BestClipped = false;
637 }
638
639 if (Barrier)
640 // The next field is a barrier that we cannot merge across.
641 InstallBest = true;
642 else if (Types.getCodeGenOpts().FineGrainedBitfieldAccesses)
643 // Fine-grained access, so no merging of spans.
644 InstallBest = true;
645 else
646 // Otherwise, we're not installing. Update the bit size
647 // of the current span to go all the way to LimitOffset, which is
648 // the (aligned) offset of next bitfield to consider.
649 BitSizeSinceBegin = Context.toBits(CharSize: LimitOffset - BeginOffset);
650 }
651 }
652 }
653
654 if (InstallBest) {
655 assert((Field == FieldEnd || !Field->isBitField() ||
656 (getFieldBitOffset(*Field) % CharBits) == 0) &&
657 "Installing but not at an aligned bitfield or limit");
658 CharUnits AccessSize = BestEndOffset - BeginOffset;
659 if (!AccessSize.isZero()) {
660 // Add the storage member for the access unit to the record. The
661 // bitfields get the offset of their storage but come afterward and
662 // remain there after a stable sort.
663 llvm::Type *Type;
664 if (BestClipped) {
665 assert(getSize(getIntNType(Context.toBits(AccessSize))) >
666 AccessSize &&
667 "Clipped access need not be clipped");
668 Type = getByteArrayType(NumChars: AccessSize);
669 } else {
670 Type = getIntNType(NumBits: Context.toBits(CharSize: AccessSize));
671 assert(getSize(Type) == AccessSize &&
672 "Unclipped access must be clipped");
673 }
674 Members.push_back(x: StorageInfo(Offset: BeginOffset, Data: Type));
675 for (; Begin != BestEnd; ++Begin)
676 if (!Begin->isZeroLengthBitField())
677 Members.push_back(
678 x: MemberInfo(BeginOffset, MemberInfo::Field, nullptr, *Begin));
679 }
680 // Reset to start a new span.
681 Field = BestEnd;
682 Begin = FieldEnd;
683 } else {
684 assert(Field != FieldEnd && Field->isBitField() &&
685 "Accumulating past end of bitfields");
686 assert(!Barrier && "Accumulating across barrier");
687 // Accumulate this bitfield into the current (potential) span.
688 BitSizeSinceBegin += Field->getBitWidthValue();
689 ++Field;
690 }
691 }
692
693 return Field;
694}
695
696void CGRecordLowering::accumulateBases() {
697 // If we've got a primary virtual base, we need to add it with the bases.
698 if (Layout.isPrimaryBaseVirtual()) {
699 const CXXRecordDecl *BaseDecl = Layout.getPrimaryBase();
700 Members.push_back(x: MemberInfo(CharUnits::Zero(), MemberInfo::Base,
701 getStorageType(RD: BaseDecl), BaseDecl));
702 }
703 // Accumulate the non-virtual bases.
704 for (const auto &Base : RD->bases()) {
705 if (Base.isVirtual())
706 continue;
707
708 // Bases can be zero-sized even if not technically empty if they
709 // contain only a trailing array member.
710 const CXXRecordDecl *BaseDecl = Base.getType()->getAsCXXRecordDecl();
711 if (!CodeGenUtils::isEmptyRecordForLayout(Ctx: Context, T: Base.getType()) &&
712 !Context.getASTRecordLayout(D: BaseDecl).getNonVirtualSize().isZero())
713 Members.push_back(x: MemberInfo(Layout.getBaseClassOffset(Base: BaseDecl),
714 MemberInfo::Base, getStorageType(RD: BaseDecl), BaseDecl));
715 }
716}
717
718/// The AAPCS that defines that, when possible, bit-fields should
719/// be accessed using containers of the declared type width:
720/// When a volatile bit-field is read, and its container does not overlap with
721/// any non-bit-field member or any zero length bit-field member, its container
722/// must be read exactly once using the access width appropriate to the type of
723/// the container. When a volatile bit-field is written, and its container does
724/// not overlap with any non-bit-field member or any zero-length bit-field
725/// member, its container must be read exactly once and written exactly once
726/// using the access width appropriate to the type of the container. The two
727/// accesses are not atomic.
728///
729/// Enforcing the width restriction can be disabled using
730/// -fno-aapcs-bitfield-width.
731void CGRecordLowering::computeVolatileBitfields() {
732 if (!CodeGenUtils::isAAPCS(TargetInfo: Context.getTargetInfo()) ||
733 !Types.getCodeGenOpts().AAPCSBitfieldWidth)
734 return;
735
736 for (auto &I : BitFields) {
737 const FieldDecl *Field = I.first;
738 CGBitFieldInfo &Info = I.second;
739 llvm::Type *ResLTy = Types.ConvertTypeForMem(T: Field->getType());
740 // If the record alignment is less than the type width, we can't enforce a
741 // aligned load, bail out.
742 if ((uint64_t)(Context.toBits(CharSize: Layout.getAlignment())) <
743 ResLTy->getPrimitiveSizeInBits())
744 continue;
745 // CGRecordLowering::setBitFieldInfo() pre-adjusts the bit-field offsets
746 // for big-endian targets, but it assumes a container of width
747 // Info.StorageSize. Since AAPCS uses a different container size (width
748 // of the type), we first undo that calculation here and redo it once
749 // the bit-field offset within the new container is calculated.
750 const unsigned OldOffset =
751 isBE() ? Info.StorageSize - (Info.Offset + Info.Size) : Info.Offset;
752 // Offset to the bit-field from the beginning of the struct.
753 const unsigned AbsoluteOffset =
754 Context.toBits(CharSize: Info.StorageOffset) + OldOffset;
755
756 // Container size is the width of the bit-field type.
757 const unsigned StorageSize = ResLTy->getPrimitiveSizeInBits();
758 // Nothing to do if the access uses the desired
759 // container width and is naturally aligned.
760 if (Info.StorageSize == StorageSize && (OldOffset % StorageSize == 0))
761 continue;
762
763 // Offset within the container.
764 unsigned Offset = AbsoluteOffset & (StorageSize - 1);
765 // Bail out if an aligned load of the container cannot cover the entire
766 // bit-field. This can happen for example, if the bit-field is part of a
767 // packed struct. AAPCS does not define access rules for such cases, we let
768 // clang to follow its own rules.
769 if (Offset + Info.Size > StorageSize)
770 continue;
771
772 // Re-adjust offsets for big-endian targets.
773 if (isBE())
774 Offset = StorageSize - (Offset + Info.Size);
775
776 const CharUnits StorageOffset =
777 Context.toCharUnitsFromBits(BitSize: AbsoluteOffset & ~(StorageSize - 1));
778 const CharUnits End = StorageOffset +
779 Context.toCharUnitsFromBits(BitSize: StorageSize) -
780 CharUnits::One();
781
782 const ASTRecordLayout &Layout =
783 Context.getASTRecordLayout(D: Field->getParent());
784 // If we access outside memory outside the record, than bail out.
785 const CharUnits RecordSize = Layout.getSize();
786 if (End >= RecordSize)
787 continue;
788
789 // Bail out if performing this load would access non-bit-fields members.
790 bool Conflict = false;
791 for (const auto *F : D->fields()) {
792 // Allow sized bit-fields overlaps.
793 if (F->isBitField() && !F->isZeroLengthBitField())
794 continue;
795
796 const CharUnits FOffset = Context.toCharUnitsFromBits(
797 BitSize: Layout.getFieldOffset(FieldNo: F->getFieldIndex()));
798
799 // As C11 defines, a zero sized bit-field defines a barrier, so
800 // fields after and before it should be race condition free.
801 // The AAPCS acknowledges it and imposes no restritions when the
802 // natural container overlaps a zero-length bit-field.
803 if (F->isZeroLengthBitField()) {
804 if (End > FOffset && StorageOffset < FOffset) {
805 Conflict = true;
806 break;
807 }
808 }
809
810 const CharUnits FEnd =
811 FOffset +
812 Context.toCharUnitsFromBits(
813 BitSize: Types.ConvertTypeForMem(T: F->getType())->getPrimitiveSizeInBits()) -
814 CharUnits::One();
815 // If no overlap, continue.
816 if (End < FOffset || FEnd < StorageOffset)
817 continue;
818
819 // The desired load overlaps a non-bit-field member, bail out.
820 Conflict = true;
821 break;
822 }
823
824 if (Conflict)
825 continue;
826 // Write the new bit-field access parameters.
827 // As the storage offset now is defined as the number of elements from the
828 // start of the structure, we should divide the Offset by the element size.
829 Info.VolatileStorageOffset =
830 StorageOffset / Context.toCharUnitsFromBits(BitSize: StorageSize).getQuantity();
831 Info.VolatileStorageSize = StorageSize;
832 Info.VolatileOffset = Offset;
833 }
834}
835
836void CGRecordLowering::accumulateVPtrs() {
837 llvm::Type *VPtrTy = Types.getCGM().GlobalsInt8PtrTy;
838 if (Layout.hasOwnVFPtr())
839 Members.push_back(x: MemberInfo(CharUnits::Zero(), MemberInfo::VFPtr, VPtrTy));
840 if (Layout.hasOwnVBPtr())
841 Members.push_back(
842 x: MemberInfo(Layout.getVBPtrOffset(), MemberInfo::VBPtr, VPtrTy));
843}
844
845CharUnits
846CGRecordLowering::calculateTailClippingOffset(bool isNonVirtualBaseType) const {
847 if (!RD)
848 return Layout.getDataSize();
849
850 CharUnits ScissorOffset = Layout.getNonVirtualSize();
851 // In the itanium ABI, it's possible to place a vbase at a dsize that is
852 // smaller than the nvsize. Here we check to see if such a base is placed
853 // before the nvsize and set the scissor offset to that, instead of the
854 // nvsize.
855 if (!isNonVirtualBaseType && CodeGenUtils::isOverlappingVBaseABI(Ctx: Context))
856 for (const auto &Base : RD->vbases()) {
857 const CXXRecordDecl *BaseDecl = Base.getType()->getAsCXXRecordDecl();
858 if (CodeGenUtils::isEmptyRecordForLayout(Ctx: Context, T: Base.getType()))
859 continue;
860 // If the vbase is a primary virtual base of some base, then it doesn't
861 // get its own storage location but instead lives inside of that base.
862 if (Context.isNearlyEmpty(RD: BaseDecl) &&
863 !CodeGenUtils::hasOwnStorage(Ctx: Context, Decl: RD, Query: BaseDecl))
864 continue;
865 ScissorOffset = std::min(a: ScissorOffset,
866 b: Layout.getVBaseClassOffset(VBase: BaseDecl));
867 }
868
869 return ScissorOffset;
870}
871
872void CGRecordLowering::accumulateVBases() {
873 for (const auto &Base : RD->vbases()) {
874 const CXXRecordDecl *BaseDecl = Base.getType()->getAsCXXRecordDecl();
875 if (CodeGenUtils::isEmptyRecordForLayout(Ctx: Context, T: Base.getType()))
876 continue;
877 CharUnits Offset = Layout.getVBaseClassOffset(VBase: BaseDecl);
878 // If the vbase is a primary virtual base of some base, then it doesn't
879 // get its own storage location but instead lives inside of that base.
880 if (CodeGenUtils::isOverlappingVBaseABI(Ctx: Context) &&
881 Context.isNearlyEmpty(RD: BaseDecl) &&
882 !CodeGenUtils::hasOwnStorage(Ctx: Context, Decl: RD, Query: BaseDecl)) {
883 Members.push_back(x: MemberInfo(Offset, MemberInfo::VBase, nullptr,
884 BaseDecl));
885 continue;
886 }
887 // If we've got a vtordisp, add it as a storage type.
888 if (Layout.getVBaseOffsetsMap().find(Val: BaseDecl)->second.hasVtorDisp())
889 Members.push_back(x: StorageInfo(Offset: Offset - CharUnits::fromQuantity(Quantity: 4),
890 Data: getIntNType(NumBits: 32)));
891 Members.push_back(x: MemberInfo(Offset, MemberInfo::VBase,
892 getStorageType(RD: BaseDecl), BaseDecl));
893 }
894}
895
896void CGRecordLowering::calculateZeroInit() {
897 for (std::vector<MemberInfo>::const_iterator Member = Members.begin(),
898 MemberEnd = Members.end();
899 IsZeroInitializableAsBase && Member != MemberEnd; ++Member) {
900 if (Member->Kind == MemberInfo::Field) {
901 if (!Member->FD || isZeroInitializable(FD: Member->FD))
902 continue;
903 IsZeroInitializable = IsZeroInitializableAsBase = false;
904 } else if (Member->Kind == MemberInfo::Base ||
905 Member->Kind == MemberInfo::VBase) {
906 if (isZeroInitializable(RD: Member->RD))
907 continue;
908 IsZeroInitializable = false;
909 if (Member->Kind == MemberInfo::Base)
910 IsZeroInitializableAsBase = false;
911 }
912 }
913}
914
915// Verify accumulateBitfields computed the correct storage representations.
916void CGRecordLowering::checkBitfieldClipping(bool IsNonVirtualBaseType) const {
917#ifndef NDEBUG
918 auto ScissorOffset = calculateTailClippingOffset(IsNonVirtualBaseType);
919 auto Tail = CharUnits::Zero();
920 for (const auto &M : Members) {
921 // Only members with data could possibly overlap.
922 if (!M.Data)
923 continue;
924
925 assert(M.Offset >= Tail && "Bitfield access unit is not clipped");
926 Tail = M.Offset + getSize(M.Data);
927 assert((Tail <= ScissorOffset || M.Offset >= ScissorOffset) &&
928 "Bitfield straddles scissor offset");
929 }
930#endif
931}
932
933void CGRecordLowering::determinePacked(bool NVBaseType) {
934 if (Packed)
935 return;
936 CharUnits Alignment = CharUnits::One();
937 CharUnits NVAlignment = CharUnits::One();
938 CharUnits NVSize =
939 !NVBaseType && RD ? Layout.getNonVirtualSize() : CharUnits::Zero();
940 for (const MemberInfo &Member : Members) {
941 if (!Member.Data)
942 continue;
943 // If any member falls at an offset that it not a multiple of its alignment,
944 // then the entire record must be packed.
945 if (!Member.Offset.isMultipleOf(N: getAlignment(Type: Member.Data)))
946 Packed = true;
947 if (Member.Offset < NVSize)
948 NVAlignment = std::max(a: NVAlignment, b: getAlignment(Type: Member.Data));
949 Alignment = std::max(a: Alignment, b: getAlignment(Type: Member.Data));
950 }
951 // If the size of the record (the capstone's offset) is not a multiple of the
952 // record's alignment, it must be packed.
953 if (!Members.back().Offset.isMultipleOf(N: Alignment))
954 Packed = true;
955 // If the non-virtual sub-object is not a multiple of the non-virtual
956 // sub-object's alignment, it must be packed. We cannot have a packed
957 // non-virtual sub-object and an unpacked complete object or vise versa.
958 if (!NVSize.isMultipleOf(N: NVAlignment))
959 Packed = true;
960 // Update the alignment of the sentinel.
961 if (!Packed)
962 Members.back().Data = getIntNType(NumBits: Context.toBits(CharSize: Alignment));
963}
964
965void CGRecordLowering::insertPadding() {
966 std::vector<std::pair<CharUnits, CharUnits> > Padding;
967 CharUnits Size = CharUnits::Zero();
968 for (const MemberInfo &Member : Members) {
969 if (!Member.Data)
970 continue;
971 CharUnits Offset = Member.Offset;
972 assert(Offset >= Size);
973 // Insert padding if we need to.
974 if (Offset !=
975 Size.alignTo(Align: Packed ? CharUnits::One() : getAlignment(Type: Member.Data)))
976 Padding.push_back(x: std::make_pair(x&: Size, y: Offset - Size));
977 Size = Offset + getSize(Type: Member.Data);
978 }
979 if (Padding.empty())
980 return;
981 // Add the padding to the Members list and sort it.
982 for (const auto &Pad : Padding)
983 Members.push_back(x: StorageInfo(Offset: Pad.first, Data: getByteArrayType(NumChars: Pad.second)));
984 llvm::stable_sort(Range&: Members);
985}
986
987void CGRecordLowering::fillOutputFields() {
988 for (const MemberInfo &Member : Members) {
989 if (Member.Data)
990 FieldTypes.push_back(Elt: Member.Data);
991 if (Member.Kind == MemberInfo::Field) {
992 if (Member.FD)
993 Fields[Member.FD->getCanonicalDecl()] = FieldTypes.size() - 1;
994 // A field without storage must be a bitfield.
995 if (!Member.Data) {
996 assert(Member.FD &&
997 "Member.Data is a nullptr so Member.FD should not be");
998 setBitFieldInfo(FD: Member.FD, StartOffset: Member.Offset, StorageType: FieldTypes.back());
999 }
1000 } else if (Member.Kind == MemberInfo::Base)
1001 NonVirtualBases[Member.RD] = FieldTypes.size() - 1;
1002 else if (Member.Kind == MemberInfo::VBase)
1003 VirtualBases[Member.RD] = FieldTypes.size() - 1;
1004 }
1005}
1006
1007CGBitFieldInfo CGBitFieldInfo::MakeInfo(CodeGenTypes &Types,
1008 const FieldDecl *FD,
1009 uint64_t Offset, uint64_t Size,
1010 uint64_t StorageSize,
1011 CharUnits StorageOffset) {
1012 // This function is vestigial from CGRecordLayoutBuilder days but is still
1013 // used in GCObjCRuntime.cpp. That usage has a "fixme" attached to it that
1014 // when addressed will allow for the removal of this function.
1015 llvm::Type *Ty = Types.ConvertTypeForMem(T: FD->getType());
1016 CharUnits TypeSizeInBytes =
1017 CharUnits::fromQuantity(Quantity: Types.getDataLayout().getTypeAllocSize(Ty));
1018 uint64_t TypeSizeInBits = Types.getContext().toBits(CharSize: TypeSizeInBytes);
1019
1020 bool IsSigned = FD->getType()->isSignedIntegerOrEnumerationType();
1021
1022 if (Size > TypeSizeInBits) {
1023 // We have a wide bit-field. The extra bits are only used for padding, so
1024 // if we have a bitfield of type T, with size N:
1025 //
1026 // T t : N;
1027 //
1028 // We can just assume that it's:
1029 //
1030 // T t : sizeof(T);
1031 //
1032 Size = TypeSizeInBits;
1033 }
1034
1035 // Reverse the bit offsets for big endian machines. Because we represent
1036 // a bitfield as a single large integer load, we can imagine the bits
1037 // counting from the most-significant-bit instead of the
1038 // least-significant-bit.
1039 if (Types.getDataLayout().isBigEndian()) {
1040 Offset = StorageSize - (Offset + Size);
1041 }
1042
1043 return CGBitFieldInfo(Offset, Size, IsSigned, StorageSize, StorageOffset);
1044}
1045
1046std::unique_ptr<CGRecordLayout>
1047CodeGenTypes::ComputeRecordLayout(const RecordDecl *D, llvm::StructType *Ty) {
1048 CGRecordLowering Builder(*this, D, /*Packed=*/false);
1049
1050 Builder.lower(/*NonVirtualBaseType=*/NVBaseType: false);
1051
1052 // If we're in C++, compute the base subobject type.
1053 llvm::StructType *BaseTy = nullptr;
1054 if (isa<CXXRecordDecl>(Val: D)) {
1055 BaseTy = Ty;
1056 if (Builder.Layout.getNonVirtualSize() != Builder.Layout.getSize()) {
1057 CGRecordLowering BaseBuilder(*this, D, /*Packed=*/Builder.Packed);
1058 BaseBuilder.lower(/*NonVirtualBaseType=*/NVBaseType: true);
1059 BaseTy = llvm::StructType::create(
1060 Context&: getLLVMContext(), Elements: BaseBuilder.FieldTypes, Name: "", isPacked: BaseBuilder.Packed);
1061 addRecordTypeName(RD: D, Ty: BaseTy, suffix: ".base");
1062 // BaseTy and Ty must agree on their packedness for getLLVMFieldNo to work
1063 // on both of them with the same index.
1064 assert(Builder.Packed == BaseBuilder.Packed &&
1065 "Non-virtual and complete types must agree on packedness");
1066 }
1067 }
1068
1069 // Fill in the struct *after* computing the base type. Filling in the body
1070 // signifies that the type is no longer opaque and record layout is complete,
1071 // but we may need to recursively layout D while laying D out as a base type.
1072 Ty->setBody(Elements: Builder.FieldTypes, isPacked: Builder.Packed);
1073
1074 auto RL = std::make_unique<CGRecordLayout>(
1075 args&: Ty, args&: BaseTy, args: (bool)Builder.IsZeroInitializable,
1076 args: (bool)Builder.IsZeroInitializableAsBase);
1077
1078 RL->NonVirtualBases.swap(RHS&: Builder.NonVirtualBases);
1079 RL->CompleteObjectVirtualBases.swap(RHS&: Builder.VirtualBases);
1080
1081 // Add all the field numbers.
1082 RL->FieldInfo.swap(RHS&: Builder.Fields);
1083
1084 // Add bitfield info.
1085 RL->BitFields.swap(RHS&: Builder.BitFields);
1086
1087 // Dump the layout, if requested.
1088 if (getContext().getLangOpts().DumpRecordLayouts) {
1089 llvm::outs() << "\n*** Dumping IRgen Record Layout\n";
1090 llvm::outs() << "Record: ";
1091 D->dump(Out&: llvm::outs());
1092 llvm::outs() << "\nLayout: ";
1093 RL->print(OS&: llvm::outs());
1094 }
1095
1096#ifndef NDEBUG
1097 // Verify that the computed LLVM struct size matches the AST layout size.
1098 const ASTRecordLayout &Layout = getContext().getASTRecordLayout(D);
1099
1100 uint64_t TypeSizeInBits = getContext().toBits(Layout.getSize());
1101 assert(TypeSizeInBits == getDataLayout().getTypeAllocSizeInBits(Ty) &&
1102 "Type size mismatch!");
1103
1104 if (BaseTy) {
1105 CharUnits NonVirtualSize = Layout.getNonVirtualSize();
1106
1107 uint64_t AlignedNonVirtualTypeSizeInBits =
1108 getContext().toBits(NonVirtualSize);
1109
1110 assert(AlignedNonVirtualTypeSizeInBits ==
1111 getDataLayout().getTypeAllocSizeInBits(BaseTy) &&
1112 "Type size mismatch!");
1113 }
1114
1115 // Verify that the LLVM and AST field offsets agree.
1116 llvm::StructType *ST = RL->getLLVMType();
1117 const llvm::StructLayout *SL = getDataLayout().getStructLayout(ST);
1118
1119 const ASTRecordLayout &AST_RL = getContext().getASTRecordLayout(D);
1120 RecordDecl::field_iterator it = D->field_begin();
1121 for (unsigned i = 0, e = AST_RL.getFieldCount(); i != e; ++i, ++it) {
1122 const FieldDecl *FD = *it;
1123
1124 // Ignore zero-sized fields.
1125 if (CodeGenUtils::isEmptyFieldForLayout(getContext(), FD))
1126 continue;
1127
1128 // For non-bit-fields, just check that the LLVM struct offset matches the
1129 // AST offset.
1130 if (!FD->isBitField()) {
1131 unsigned FieldNo = RL->getLLVMFieldNo(FD);
1132 assert(AST_RL.getFieldOffset(i) == SL->getElementOffsetInBits(FieldNo) &&
1133 "Invalid field offset!");
1134 continue;
1135 }
1136
1137 // Ignore unnamed bit-fields.
1138 if (!FD->getDeclName())
1139 continue;
1140
1141 const CGBitFieldInfo &Info = RL->getBitFieldInfo(FD);
1142 llvm::Type *ElementTy = ST->getTypeAtIndex(RL->getLLVMFieldNo(FD));
1143
1144 // Unions have overlapping elements dictating their layout, but for
1145 // non-unions we can verify that this section of the layout is the exact
1146 // expected size.
1147 if (D->isUnion()) {
1148 // For unions we verify that the start is zero and the size
1149 // is in-bounds. However, on BE systems, the offset may be non-zero, but
1150 // the size + offset should match the storage size in that case as it
1151 // "starts" at the back.
1152 if (getDataLayout().isBigEndian())
1153 assert(static_cast<unsigned>(Info.Offset + Info.Size) ==
1154 Info.StorageSize &&
1155 "Big endian union bitfield does not end at the back");
1156 else
1157 assert(Info.Offset == 0 &&
1158 "Little endian union bitfield with a non-zero offset");
1159 assert(Info.StorageSize <= SL->getSizeInBits() &&
1160 "Union not large enough for bitfield storage");
1161 } else {
1162 assert((Info.StorageSize ==
1163 getDataLayout().getTypeAllocSizeInBits(ElementTy) ||
1164 Info.VolatileStorageSize ==
1165 getDataLayout().getTypeAllocSizeInBits(ElementTy)) &&
1166 "Storage size does not match the element type size");
1167 }
1168 assert(Info.Size > 0 && "Empty bitfield!");
1169 assert(static_cast<unsigned>(Info.Offset) + Info.Size <= Info.StorageSize &&
1170 "Bitfield outside of its allocated storage");
1171 }
1172#endif
1173
1174 return RL;
1175}
1176
1177void CGRecordLayout::print(raw_ostream &OS) const {
1178 OS << "<CGRecordLayout\n";
1179 OS << " LLVMType:" << *CompleteObjectType << "\n";
1180 if (BaseSubobjectType)
1181 OS << " NonVirtualBaseLLVMType:" << *BaseSubobjectType << "\n";
1182 OS << " IsZeroInitializable:" << IsZeroInitializable << "\n";
1183 OS << " BitFields:[\n";
1184
1185 // Print bit-field infos in declaration order.
1186 std::vector<std::pair<unsigned, const CGBitFieldInfo*> > BFIs;
1187 for (const auto &BitField : BitFields) {
1188 const RecordDecl *RD = BitField.first->getParent();
1189 unsigned Index = 0;
1190 for (RecordDecl::field_iterator it2 = RD->field_begin();
1191 *it2 != BitField.first; ++it2)
1192 ++Index;
1193 BFIs.push_back(x: std::make_pair(x&: Index, y: &BitField.second));
1194 }
1195 llvm::array_pod_sort(Start: BFIs.begin(), End: BFIs.end());
1196 for (auto &BFI : BFIs) {
1197 OS.indent(NumSpaces: 4);
1198 BFI.second->print(OS);
1199 OS << "\n";
1200 }
1201
1202 OS << "]>\n";
1203}
1204
1205LLVM_DUMP_METHOD void CGRecordLayout::dump() const {
1206 print(OS&: llvm::errs());
1207}
1208
1209void CGBitFieldInfo::print(raw_ostream &OS) const {
1210 OS << "<CGBitFieldInfo"
1211 << " Offset:" << Offset << " Size:" << Size << " IsSigned:" << IsSigned
1212 << " StorageSize:" << StorageSize
1213 << " StorageOffset:" << StorageOffset.getQuantity()
1214 << " VolatileOffset:" << VolatileOffset
1215 << " VolatileStorageSize:" << VolatileStorageSize
1216 << " VolatileStorageOffset:" << VolatileStorageOffset.getQuantity() << ">";
1217}
1218
1219LLVM_DUMP_METHOD void CGBitFieldInfo::dump() const {
1220 print(OS&: llvm::errs());
1221}
1222