1 //===- llvm/DataLayout.h - Data size & alignment info -----------*- C++ -*-===//
2 //
3 // The LLVM Compiler Infrastructure
4 //
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
7 //
8 //===----------------------------------------------------------------------===//
9 //
10 // This file defines layout properties related to datatype size/offset/alignment
11 // information. It uses lazy annotations to cache information about how
12 // structure types are laid out and used.
13 //
14 // This structure should be created once, filled in if the defaults are not
15 // correct and then passed around by const&. None of the members functions
16 // require modification to the object.
17 //
18 //===----------------------------------------------------------------------===//
19
20 #ifndef LLVM_IR_DATALAYOUT_H
21 #define LLVM_IR_DATALAYOUT_H
22
23 #include "llvm/ADT/ArrayRef.h"
24 #include "llvm/ADT/STLExtras.h"
25 #include "llvm/ADT/SmallVector.h"
26 #include "llvm/ADT/StringRef.h"
27 #include "llvm/IR/DerivedTypes.h"
28 #include "llvm/IR/Type.h"
29 #include "llvm/Pass.h"
30 #include "llvm/Support/Casting.h"
31 #include "llvm/Support/ErrorHandling.h"
32 #include "llvm/Support/MathExtras.h"
33 #include <cassert>
34 #include <cstdint>
35 #include <string>
36
37 // This needs to be outside of the namespace, to avoid conflict with llvm-c
38 // decl.
39 using LLVMTargetDataRef = struct LLVMOpaqueTargetData *;
40
41 namespace llvm {
42
43 class GlobalVariable;
44 class LLVMContext;
45 class Module;
46 class StructLayout;
47 class Triple;
48 class Value;
49
50 /// Enum used to categorize the alignment types stored by LayoutAlignElem
51 enum AlignTypeEnum {
52 INVALID_ALIGN = 0,
53 INTEGER_ALIGN = 'i',
54 VECTOR_ALIGN = 'v',
55 FLOAT_ALIGN = 'f',
56 AGGREGATE_ALIGN = 'a'
57 };
58
59 // FIXME: Currently the DataLayout string carries a "preferred alignment"
60 // for types. As the DataLayout is module/global, this should likely be
61 // sunk down to an FTTI element that is queried rather than a global
62 // preference.
63
64 /// Layout alignment element.
65 ///
66 /// Stores the alignment data associated with a given alignment type (integer,
67 /// vector, float) and type bit width.
68 ///
69 /// \note The unusual order of elements in the structure attempts to reduce
70 /// padding and make the structure slightly more cache friendly.
71 struct LayoutAlignElem {
72 /// Alignment type from \c AlignTypeEnum
73 unsigned AlignType : 8;
74 unsigned TypeBitWidth : 24;
75 unsigned ABIAlign : 16;
76 unsigned PrefAlign : 16;
77
78 static LayoutAlignElem get(AlignTypeEnum align_type, unsigned abi_align,
79 unsigned pref_align, uint32_t bit_width);
80
81 bool operator==(const LayoutAlignElem &rhs) const;
82 };
83
84 /// Layout pointer alignment element.
85 ///
86 /// Stores the alignment data associated with a given pointer and address space.
87 ///
88 /// \note The unusual order of elements in the structure attempts to reduce
89 /// padding and make the structure slightly more cache friendly.
90 struct PointerAlignElem {
91 unsigned ABIAlign;
92 unsigned PrefAlign;
93 uint32_t TypeByteWidth;
94 uint32_t AddressSpace;
95 uint32_t IndexWidth;
96
97 /// Initializer
98 static PointerAlignElem get(uint32_t AddressSpace, unsigned ABIAlign,
99 unsigned PrefAlign, uint32_t TypeByteWidth,
100 uint32_t IndexWidth);
101
102 bool operator==(const PointerAlignElem &rhs) const;
103 };
104
105 /// A parsed version of the target data layout string in and methods for
106 /// querying it.
107 ///
108 /// The target data layout string is specified *by the target* - a frontend
109 /// generating LLVM IR is required to generate the right target data for the
110 /// target being codegen'd to.
111 class DataLayout {
112 private:
113 /// Defaults to false.
114 bool BigEndian;
115
116 unsigned AllocaAddrSpace;
117 unsigned StackNaturalAlign;
118 unsigned ProgramAddrSpace;
119
120 enum ManglingModeT {
121 MM_None,
122 MM_ELF,
123 MM_MachO,
124 MM_WinCOFF,
125 MM_WinCOFFX86,
126 MM_Mips
127 };
128 ManglingModeT ManglingMode;
129
130 SmallVector<unsigned char, 8> LegalIntWidths;
131
132 /// Primitive type alignment data. This is sorted by type and bit
133 /// width during construction.
134 using AlignmentsTy = SmallVector<LayoutAlignElem, 16>;
135 AlignmentsTy Alignments;
136
137 AlignmentsTy::const_iterator
findAlignmentLowerBound(AlignTypeEnum AlignType,uint32_t BitWidth)138 findAlignmentLowerBound(AlignTypeEnum AlignType, uint32_t BitWidth) const {
139 return const_cast<DataLayout *>(this)->findAlignmentLowerBound(AlignType,
140 BitWidth);
141 }
142
143 AlignmentsTy::iterator
144 findAlignmentLowerBound(AlignTypeEnum AlignType, uint32_t BitWidth);
145
146 /// The string representation used to create this DataLayout
147 std::string StringRepresentation;
148
149 using PointersTy = SmallVector<PointerAlignElem, 8>;
150 PointersTy Pointers;
151
152 PointersTy::const_iterator
findPointerLowerBound(uint32_t AddressSpace)153 findPointerLowerBound(uint32_t AddressSpace) const {
154 return const_cast<DataLayout *>(this)->findPointerLowerBound(AddressSpace);
155 }
156
157 PointersTy::iterator findPointerLowerBound(uint32_t AddressSpace);
158
159 // The StructType -> StructLayout map.
160 mutable void *LayoutMap = nullptr;
161
162 /// Pointers in these address spaces are non-integral, and don't have a
163 /// well-defined bitwise representation.
164 SmallVector<unsigned, 8> NonIntegralAddressSpaces;
165
166 void setAlignment(AlignTypeEnum align_type, unsigned abi_align,
167 unsigned pref_align, uint32_t bit_width);
168 unsigned getAlignmentInfo(AlignTypeEnum align_type, uint32_t bit_width,
169 bool ABIAlign, Type *Ty) const;
170 void setPointerAlignment(uint32_t AddrSpace, unsigned ABIAlign,
171 unsigned PrefAlign, uint32_t TypeByteWidth,
172 uint32_t IndexWidth);
173
174 /// Internal helper method that returns requested alignment for type.
175 unsigned getAlignment(Type *Ty, bool abi_or_pref) const;
176
177 /// Parses a target data specification string. Assert if the string is
178 /// malformed.
179 void parseSpecifier(StringRef LayoutDescription);
180
181 // Free all internal data structures.
182 void clear();
183
184 public:
185 /// Constructs a DataLayout from a specification string. See reset().
DataLayout(StringRef LayoutDescription)186 explicit DataLayout(StringRef LayoutDescription) {
187 reset(LayoutDescription);
188 }
189
190 /// Initialize target data from properties stored in the module.
191 explicit DataLayout(const Module *M);
192
DataLayout(const DataLayout & DL)193 DataLayout(const DataLayout &DL) { *this = DL; }
194
195 ~DataLayout(); // Not virtual, do not subclass this class
196
197 DataLayout &operator=(const DataLayout &DL) {
198 clear();
199 StringRepresentation = DL.StringRepresentation;
200 BigEndian = DL.isBigEndian();
201 AllocaAddrSpace = DL.AllocaAddrSpace;
202 StackNaturalAlign = DL.StackNaturalAlign;
203 ProgramAddrSpace = DL.ProgramAddrSpace;
204 ManglingMode = DL.ManglingMode;
205 LegalIntWidths = DL.LegalIntWidths;
206 Alignments = DL.Alignments;
207 Pointers = DL.Pointers;
208 NonIntegralAddressSpaces = DL.NonIntegralAddressSpaces;
209 return *this;
210 }
211
212 bool operator==(const DataLayout &Other) const;
213 bool operator!=(const DataLayout &Other) const { return !(*this == Other); }
214
215 void init(const Module *M);
216
217 /// Parse a data layout string (with fallback to default values).
218 void reset(StringRef LayoutDescription);
219
220 /// Layout endianness...
isLittleEndian()221 bool isLittleEndian() const { return !BigEndian; }
isBigEndian()222 bool isBigEndian() const { return BigEndian; }
223
224 /// Returns the string representation of the DataLayout.
225 ///
226 /// This representation is in the same format accepted by the string
227 /// constructor above. This should not be used to compare two DataLayout as
228 /// different string can represent the same layout.
getStringRepresentation()229 const std::string &getStringRepresentation() const {
230 return StringRepresentation;
231 }
232
233 /// Test if the DataLayout was constructed from an empty string.
isDefault()234 bool isDefault() const { return StringRepresentation.empty(); }
235
236 /// Returns true if the specified type is known to be a native integer
237 /// type supported by the CPU.
238 ///
239 /// For example, i64 is not native on most 32-bit CPUs and i37 is not native
240 /// on any known one. This returns false if the integer width is not legal.
241 ///
242 /// The width is specified in bits.
isLegalInteger(uint64_t Width)243 bool isLegalInteger(uint64_t Width) const {
244 for (unsigned LegalIntWidth : LegalIntWidths)
245 if (LegalIntWidth == Width)
246 return true;
247 return false;
248 }
249
isIllegalInteger(uint64_t Width)250 bool isIllegalInteger(uint64_t Width) const { return !isLegalInteger(Width); }
251
252 /// Returns true if the given alignment exceeds the natural stack alignment.
exceedsNaturalStackAlignment(unsigned Align)253 bool exceedsNaturalStackAlignment(unsigned Align) const {
254 return (StackNaturalAlign != 0) && (Align > StackNaturalAlign);
255 }
256
getStackAlignment()257 unsigned getStackAlignment() const { return StackNaturalAlign; }
getAllocaAddrSpace()258 unsigned getAllocaAddrSpace() const { return AllocaAddrSpace; }
259
getProgramAddressSpace()260 unsigned getProgramAddressSpace() const { return ProgramAddrSpace; }
261
hasMicrosoftFastStdCallMangling()262 bool hasMicrosoftFastStdCallMangling() const {
263 return ManglingMode == MM_WinCOFFX86;
264 }
265
266 /// Returns true if symbols with leading question marks should not receive IR
267 /// mangling. True for Windows mangling modes.
doNotMangleLeadingQuestionMark()268 bool doNotMangleLeadingQuestionMark() const {
269 return ManglingMode == MM_WinCOFF || ManglingMode == MM_WinCOFFX86;
270 }
271
hasLinkerPrivateGlobalPrefix()272 bool hasLinkerPrivateGlobalPrefix() const { return ManglingMode == MM_MachO; }
273
getLinkerPrivateGlobalPrefix()274 StringRef getLinkerPrivateGlobalPrefix() const {
275 if (ManglingMode == MM_MachO)
276 return "l";
277 return "";
278 }
279
getGlobalPrefix()280 char getGlobalPrefix() const {
281 switch (ManglingMode) {
282 case MM_None:
283 case MM_ELF:
284 case MM_Mips:
285 case MM_WinCOFF:
286 return '\0';
287 case MM_MachO:
288 case MM_WinCOFFX86:
289 return '_';
290 }
291 llvm_unreachable("invalid mangling mode");
292 }
293
getPrivateGlobalPrefix()294 StringRef getPrivateGlobalPrefix() const {
295 switch (ManglingMode) {
296 case MM_None:
297 return "";
298 case MM_ELF:
299 case MM_WinCOFF:
300 return ".L";
301 case MM_Mips:
302 return "$";
303 case MM_MachO:
304 case MM_WinCOFFX86:
305 return "L";
306 }
307 llvm_unreachable("invalid mangling mode");
308 }
309
310 static const char *getManglingComponent(const Triple &T);
311
312 /// Returns true if the specified type fits in a native integer type
313 /// supported by the CPU.
314 ///
315 /// For example, if the CPU only supports i32 as a native integer type, then
316 /// i27 fits in a legal integer type but i45 does not.
fitsInLegalInteger(unsigned Width)317 bool fitsInLegalInteger(unsigned Width) const {
318 for (unsigned LegalIntWidth : LegalIntWidths)
319 if (Width <= LegalIntWidth)
320 return true;
321 return false;
322 }
323
324 /// Layout pointer alignment
325 unsigned getPointerABIAlignment(unsigned AS) const;
326
327 /// Return target's alignment for stack-based pointers
328 /// FIXME: The defaults need to be removed once all of
329 /// the backends/clients are updated.
330 unsigned getPointerPrefAlignment(unsigned AS = 0) const;
331
332 /// Layout pointer size
333 /// FIXME: The defaults need to be removed once all of
334 /// the backends/clients are updated.
335 unsigned getPointerSize(unsigned AS = 0) const;
336
337 // Index size used for address calculation.
338 unsigned getIndexSize(unsigned AS) const;
339
340 /// Return the address spaces containing non-integral pointers. Pointers in
341 /// this address space don't have a well-defined bitwise representation.
getNonIntegralAddressSpaces()342 ArrayRef<unsigned> getNonIntegralAddressSpaces() const {
343 return NonIntegralAddressSpaces;
344 }
345
isNonIntegralPointerType(PointerType * PT)346 bool isNonIntegralPointerType(PointerType *PT) const {
347 ArrayRef<unsigned> NonIntegralSpaces = getNonIntegralAddressSpaces();
348 return find(NonIntegralSpaces, PT->getAddressSpace()) !=
349 NonIntegralSpaces.end();
350 }
351
isNonIntegralPointerType(Type * Ty)352 bool isNonIntegralPointerType(Type *Ty) const {
353 auto *PTy = dyn_cast<PointerType>(Ty);
354 return PTy && isNonIntegralPointerType(PTy);
355 }
356
357 /// Layout pointer size, in bits
358 /// FIXME: The defaults need to be removed once all of
359 /// the backends/clients are updated.
360 unsigned getPointerSizeInBits(unsigned AS = 0) const {
361 return getPointerSize(AS) * 8;
362 }
363
364 /// Size in bits of index used for address calculation in getelementptr.
getIndexSizeInBits(unsigned AS)365 unsigned getIndexSizeInBits(unsigned AS) const {
366 return getIndexSize(AS) * 8;
367 }
368
369 /// Layout pointer size, in bits, based on the type. If this function is
370 /// called with a pointer type, then the type size of the pointer is returned.
371 /// If this function is called with a vector of pointers, then the type size
372 /// of the pointer is returned. This should only be called with a pointer or
373 /// vector of pointers.
374 unsigned getPointerTypeSizeInBits(Type *) const;
375
376 /// Layout size of the index used in GEP calculation.
377 /// The function should be called with pointer or vector of pointers type.
378 unsigned getIndexTypeSizeInBits(Type *Ty) const;
379
getPointerTypeSize(Type * Ty)380 unsigned getPointerTypeSize(Type *Ty) const {
381 return getPointerTypeSizeInBits(Ty) / 8;
382 }
383
384 /// Size examples:
385 ///
386 /// Type SizeInBits StoreSizeInBits AllocSizeInBits[*]
387 /// ---- ---------- --------------- ---------------
388 /// i1 1 8 8
389 /// i8 8 8 8
390 /// i19 19 24 32
391 /// i32 32 32 32
392 /// i100 100 104 128
393 /// i128 128 128 128
394 /// Float 32 32 32
395 /// Double 64 64 64
396 /// X86_FP80 80 80 96
397 ///
398 /// [*] The alloc size depends on the alignment, and thus on the target.
399 /// These values are for x86-32 linux.
400
401 /// Returns the number of bits necessary to hold the specified type.
402 ///
403 /// For example, returns 36 for i36 and 80 for x86_fp80. The type passed must
404 /// have a size (Type::isSized() must return true).
405 uint64_t getTypeSizeInBits(Type *Ty) const;
406
407 /// Returns the maximum number of bytes that may be overwritten by
408 /// storing the specified type.
409 ///
410 /// For example, returns 5 for i36 and 10 for x86_fp80.
getTypeStoreSize(Type * Ty)411 uint64_t getTypeStoreSize(Type *Ty) const {
412 return (getTypeSizeInBits(Ty) + 7) / 8;
413 }
414
415 /// Returns the maximum number of bits that may be overwritten by
416 /// storing the specified type; always a multiple of 8.
417 ///
418 /// For example, returns 40 for i36 and 80 for x86_fp80.
getTypeStoreSizeInBits(Type * Ty)419 uint64_t getTypeStoreSizeInBits(Type *Ty) const {
420 return 8 * getTypeStoreSize(Ty);
421 }
422
423 /// Returns the offset in bytes between successive objects of the
424 /// specified type, including alignment padding.
425 ///
426 /// This is the amount that alloca reserves for this type. For example,
427 /// returns 12 or 16 for x86_fp80, depending on alignment.
getTypeAllocSize(Type * Ty)428 uint64_t getTypeAllocSize(Type *Ty) const {
429 // Round up to the next alignment boundary.
430 return alignTo(getTypeStoreSize(Ty), getABITypeAlignment(Ty));
431 }
432
433 /// Returns the offset in bits between successive objects of the
434 /// specified type, including alignment padding; always a multiple of 8.
435 ///
436 /// This is the amount that alloca reserves for this type. For example,
437 /// returns 96 or 128 for x86_fp80, depending on alignment.
getTypeAllocSizeInBits(Type * Ty)438 uint64_t getTypeAllocSizeInBits(Type *Ty) const {
439 return 8 * getTypeAllocSize(Ty);
440 }
441
442 /// Returns the minimum ABI-required alignment for the specified type.
443 unsigned getABITypeAlignment(Type *Ty) const;
444
445 /// Returns the minimum ABI-required alignment for an integer type of
446 /// the specified bitwidth.
447 unsigned getABIIntegerTypeAlignment(unsigned BitWidth) const;
448
449 /// Returns the preferred stack/global alignment for the specified
450 /// type.
451 ///
452 /// This is always at least as good as the ABI alignment.
453 unsigned getPrefTypeAlignment(Type *Ty) const;
454
455 /// Returns the preferred alignment for the specified type, returned as
456 /// log2 of the value (a shift amount).
457 unsigned getPreferredTypeAlignmentShift(Type *Ty) const;
458
459 /// Returns an integer type with size at least as big as that of a
460 /// pointer in the given address space.
461 IntegerType *getIntPtrType(LLVMContext &C, unsigned AddressSpace = 0) const;
462
463 /// Returns an integer (vector of integer) type with size at least as
464 /// big as that of a pointer of the given pointer (vector of pointer) type.
465 Type *getIntPtrType(Type *) const;
466
467 /// Returns the smallest integer type with size at least as big as
468 /// Width bits.
469 Type *getSmallestLegalIntType(LLVMContext &C, unsigned Width = 0) const;
470
471 /// Returns the largest legal integer type, or null if none are set.
getLargestLegalIntType(LLVMContext & C)472 Type *getLargestLegalIntType(LLVMContext &C) const {
473 unsigned LargestSize = getLargestLegalIntTypeSizeInBits();
474 return (LargestSize == 0) ? nullptr : Type::getIntNTy(C, LargestSize);
475 }
476
477 /// Returns the size of largest legal integer type size, or 0 if none
478 /// are set.
479 unsigned getLargestLegalIntTypeSizeInBits() const;
480
481 /// Returns the type of a GEP index.
482 /// If it was not specified explicitly, it will be the integer type of the
483 /// pointer width - IntPtrType.
484 Type *getIndexType(Type *PtrTy) const;
485
486 /// Returns the offset from the beginning of the type for the specified
487 /// indices.
488 ///
489 /// Note that this takes the element type, not the pointer type.
490 /// This is used to implement getelementptr.
491 int64_t getIndexedOffsetInType(Type *ElemTy, ArrayRef<Value *> Indices) const;
492
493 /// Returns a StructLayout object, indicating the alignment of the
494 /// struct, its size, and the offsets of its fields.
495 ///
496 /// Note that this information is lazily cached.
497 const StructLayout *getStructLayout(StructType *Ty) const;
498
499 /// Returns the preferred alignment of the specified global.
500 ///
501 /// This includes an explicitly requested alignment (if the global has one).
502 unsigned getPreferredAlignment(const GlobalVariable *GV) const;
503
504 /// Returns the preferred alignment of the specified global, returned
505 /// in log form.
506 ///
507 /// This includes an explicitly requested alignment (if the global has one).
508 unsigned getPreferredAlignmentLog(const GlobalVariable *GV) const;
509 };
510
unwrap(LLVMTargetDataRef P)511 inline DataLayout *unwrap(LLVMTargetDataRef P) {
512 return reinterpret_cast<DataLayout *>(P);
513 }
514
wrap(const DataLayout * P)515 inline LLVMTargetDataRef wrap(const DataLayout *P) {
516 return reinterpret_cast<LLVMTargetDataRef>(const_cast<DataLayout *>(P));
517 }
518
519 /// Used to lazily calculate structure layout information for a target machine,
520 /// based on the DataLayout structure.
521 class StructLayout {
522 uint64_t StructSize;
523 unsigned StructAlignment;
524 unsigned IsPadded : 1;
525 unsigned NumElements : 31;
526 uint64_t MemberOffsets[1]; // variable sized array!
527
528 public:
getSizeInBytes()529 uint64_t getSizeInBytes() const { return StructSize; }
530
getSizeInBits()531 uint64_t getSizeInBits() const { return 8 * StructSize; }
532
getAlignment()533 unsigned getAlignment() const { return StructAlignment; }
534
535 /// Returns whether the struct has padding or not between its fields.
536 /// NB: Padding in nested element is not taken into account.
hasPadding()537 bool hasPadding() const { return IsPadded; }
538
539 /// Given a valid byte offset into the structure, returns the structure
540 /// index that contains it.
541 unsigned getElementContainingOffset(uint64_t Offset) const;
542
getElementOffset(unsigned Idx)543 uint64_t getElementOffset(unsigned Idx) const {
544 assert(Idx < NumElements && "Invalid element idx!");
545 return MemberOffsets[Idx];
546 }
547
getElementOffsetInBits(unsigned Idx)548 uint64_t getElementOffsetInBits(unsigned Idx) const {
549 return getElementOffset(Idx) * 8;
550 }
551
552 private:
553 friend class DataLayout; // Only DataLayout can create this class
554
555 StructLayout(StructType *ST, const DataLayout &DL);
556 };
557
558 // The implementation of this method is provided inline as it is particularly
559 // well suited to constant folding when called on a specific Type subclass.
getTypeSizeInBits(Type * Ty)560 inline uint64_t DataLayout::getTypeSizeInBits(Type *Ty) const {
561 assert(Ty->isSized() && "Cannot getTypeInfo() on a type that is unsized!");
562 switch (Ty->getTypeID()) {
563 case Type::LabelTyID:
564 return getPointerSizeInBits(0);
565 case Type::PointerTyID:
566 return getPointerSizeInBits(Ty->getPointerAddressSpace());
567 case Type::ArrayTyID: {
568 ArrayType *ATy = cast<ArrayType>(Ty);
569 return ATy->getNumElements() *
570 getTypeAllocSizeInBits(ATy->getElementType());
571 }
572 case Type::StructTyID:
573 // Get the layout annotation... which is lazily created on demand.
574 return getStructLayout(cast<StructType>(Ty))->getSizeInBits();
575 case Type::IntegerTyID:
576 return Ty->getIntegerBitWidth();
577 case Type::HalfTyID:
578 return 16;
579 case Type::FloatTyID:
580 return 32;
581 case Type::DoubleTyID:
582 case Type::X86_MMXTyID:
583 return 64;
584 case Type::PPC_FP128TyID:
585 case Type::FP128TyID:
586 return 128;
587 // In memory objects this is always aligned to a higher boundary, but
588 // only 80 bits contain information.
589 case Type::X86_FP80TyID:
590 return 80;
591 case Type::VectorTyID: {
592 VectorType *VTy = cast<VectorType>(Ty);
593 return VTy->getNumElements() * getTypeSizeInBits(VTy->getElementType());
594 }
595 default:
596 llvm_unreachable("DataLayout::getTypeSizeInBits(): Unsupported type");
597 }
598 }
599
600 } // end namespace llvm
601
602 #endif // LLVM_IR_DATALAYOUT_H
603