1 //===- lib/MC/ELFObjectWriter.cpp - ELF File Writer -----------------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file implements ELF object file writer information.
12 //===----------------------------------------------------------------------===//
14 #include "llvm/MC/MCELFObjectWriter.h"
15 #include "llvm/ADT/STLExtras.h"
16 #include "llvm/ADT/SmallPtrSet.h"
17 #include "llvm/ADT/SmallString.h"
18 #include "llvm/ADT/StringMap.h"
19 #include "llvm/MC/MCAsmBackend.h"
20 #include "llvm/MC/MCAsmInfo.h"
21 #include "llvm/MC/MCAsmLayout.h"
22 #include "llvm/MC/MCAssembler.h"
23 #include "llvm/MC/MCContext.h"
24 #include "llvm/MC/MCExpr.h"
25 #include "llvm/MC/MCFixupKindInfo.h"
26 #include "llvm/MC/MCObjectWriter.h"
27 #include "llvm/MC/MCSectionELF.h"
28 #include "llvm/MC/MCSymbolELF.h"
29 #include "llvm/MC/MCValue.h"
30 #include "llvm/MC/StringTableBuilder.h"
31 #include "llvm/Support/Compression.h"
32 #include "llvm/Support/Debug.h"
33 #include "llvm/Support/ELF.h"
34 #include "llvm/Support/Endian.h"
35 #include "llvm/Support/ErrorHandling.h"
40 #define DEBUG_TYPE "reloc-info"
44 typedef DenseMap<const MCSectionELF *, uint32_t> SectionIndexMapTy;
46 class ELFObjectWriter;
48 class SymbolTableWriter {
49 ELFObjectWriter &EWriter;
52 // indexes we are going to write to .symtab_shndx.
53 std::vector<uint32_t> ShndxIndexes;
55 // The numbel of symbols written so far.
58 void createSymtabShndx();
60 template <typename T> void write(T Value);
63 SymbolTableWriter(ELFObjectWriter &EWriter, bool Is64Bit);
65 void writeSymbol(uint32_t name, uint8_t info, uint64_t value, uint64_t size,
66 uint8_t other, uint32_t shndx, bool Reserved);
68 ArrayRef<uint32_t> getShndxIndexes() const { return ShndxIndexes; }
71 class ELFObjectWriter : public MCObjectWriter {
72 static bool isFixupKindPCRel(const MCAssembler &Asm, unsigned Kind);
73 static uint64_t SymbolValue(const MCSymbol &Sym, const MCAsmLayout &Layout);
74 static bool isInSymtab(const MCAsmLayout &Layout, const MCSymbolELF &Symbol,
75 bool Used, bool Renamed);
77 /// Helper struct for containing some precomputed information on symbols.
78 struct ELFSymbolData {
79 const MCSymbolELF *Symbol;
80 uint32_t SectionIndex;
83 // Support lexicographic sorting.
84 bool operator<(const ELFSymbolData &RHS) const {
85 unsigned LHSType = Symbol->getType();
86 unsigned RHSType = RHS.Symbol->getType();
87 if (LHSType == ELF::STT_SECTION && RHSType != ELF::STT_SECTION)
89 if (LHSType != ELF::STT_SECTION && RHSType == ELF::STT_SECTION)
91 if (LHSType == ELF::STT_SECTION && RHSType == ELF::STT_SECTION)
92 return SectionIndex < RHS.SectionIndex;
93 return Name < RHS.Name;
97 /// The target specific ELF writer instance.
98 std::unique_ptr<MCELFObjectTargetWriter> TargetObjectWriter;
100 DenseMap<const MCSymbolELF *, const MCSymbolELF *> Renames;
102 llvm::DenseMap<const MCSectionELF *, std::vector<ELFRelocationEntry>>
106 /// @name Symbol Table Data
109 StringTableBuilder StrTabBuilder;
113 // This holds the symbol table index of the last local symbol.
114 unsigned LastLocalSymbolIndex;
115 // This holds the .strtab section index.
116 unsigned StringTableIndex;
117 // This holds the .symtab section index.
118 unsigned SymbolTableIndex;
120 // Sections in the order they are to be output in the section table.
121 std::vector<const MCSectionELF *> SectionTable;
122 unsigned addToSectionTable(const MCSectionELF *Sec);
124 // TargetObjectWriter wrappers.
125 bool is64Bit() const { return TargetObjectWriter->is64Bit(); }
126 bool hasRelocationAddend() const {
127 return TargetObjectWriter->hasRelocationAddend();
129 unsigned GetRelocType(const MCValue &Target, const MCFixup &Fixup,
130 bool IsPCRel) const {
131 return TargetObjectWriter->GetRelocType(Target, Fixup, IsPCRel);
134 void align(unsigned Alignment);
137 ELFObjectWriter(MCELFObjectTargetWriter *MOTW, raw_pwrite_stream &OS,
139 : MCObjectWriter(OS, IsLittleEndian), TargetObjectWriter(MOTW) {}
141 void reset() override {
144 StrTabBuilder.clear();
145 SectionTable.clear();
146 MCObjectWriter::reset();
149 ~ELFObjectWriter() override;
151 void WriteWord(uint64_t W) {
158 template <typename T> void write(T Val) {
160 support::endian::Writer<support::little>(getStream()).write(Val);
162 support::endian::Writer<support::big>(getStream()).write(Val);
165 void writeHeader(const MCAssembler &Asm);
167 void writeSymbol(SymbolTableWriter &Writer, uint32_t StringIndex,
168 ELFSymbolData &MSD, const MCAsmLayout &Layout);
170 // Start and end offset of each section
171 typedef std::map<const MCSectionELF *, std::pair<uint64_t, uint64_t>>
174 bool shouldRelocateWithSymbol(const MCAssembler &Asm,
175 const MCSymbolRefExpr *RefA,
176 const MCSymbol *Sym, uint64_t C,
177 unsigned Type) const;
179 void recordRelocation(MCAssembler &Asm, const MCAsmLayout &Layout,
180 const MCFragment *Fragment, const MCFixup &Fixup,
181 MCValue Target, bool &IsPCRel,
182 uint64_t &FixedValue) override;
184 // Map from a signature symbol to the group section index
185 typedef DenseMap<const MCSymbol *, unsigned> RevGroupMapTy;
187 /// Compute the symbol table data
189 /// \param Asm - The assembler.
190 /// \param SectionIndexMap - Maps a section to its index.
191 /// \param RevGroupMap - Maps a signature symbol to the group section.
192 void computeSymbolTable(MCAssembler &Asm, const MCAsmLayout &Layout,
193 const SectionIndexMapTy &SectionIndexMap,
194 const RevGroupMapTy &RevGroupMap,
195 SectionOffsetsTy &SectionOffsets);
197 MCSectionELF *createRelocationSection(MCContext &Ctx,
198 const MCSectionELF &Sec);
200 const MCSectionELF *createStringTable(MCContext &Ctx);
202 void executePostLayoutBinding(MCAssembler &Asm,
203 const MCAsmLayout &Layout) override;
205 void writeSectionHeader(const MCAsmLayout &Layout,
206 const SectionIndexMapTy &SectionIndexMap,
207 const SectionOffsetsTy &SectionOffsets);
209 void writeSectionData(const MCAssembler &Asm, MCSection &Sec,
210 const MCAsmLayout &Layout);
212 void WriteSecHdrEntry(uint32_t Name, uint32_t Type, uint64_t Flags,
213 uint64_t Address, uint64_t Offset, uint64_t Size,
214 uint32_t Link, uint32_t Info, uint64_t Alignment,
217 void writeRelocations(const MCAssembler &Asm, const MCSectionELF &Sec);
219 bool isSymbolRefDifferenceFullyResolvedImpl(const MCAssembler &Asm,
220 const MCSymbol &SymA,
221 const MCFragment &FB,
223 bool IsPCRel) const override;
225 bool isWeak(const MCSymbol &Sym) const override;
227 void writeObject(MCAssembler &Asm, const MCAsmLayout &Layout) override;
228 void writeSection(const SectionIndexMapTy &SectionIndexMap,
229 uint32_t GroupSymbolIndex, uint64_t Offset, uint64_t Size,
230 const MCSectionELF &Section);
234 void ELFObjectWriter::align(unsigned Alignment) {
235 uint64_t Padding = OffsetToAlignment(getStream().tell(), Alignment);
239 unsigned ELFObjectWriter::addToSectionTable(const MCSectionELF *Sec) {
240 SectionTable.push_back(Sec);
241 StrTabBuilder.add(Sec->getSectionName());
242 return SectionTable.size();
245 void SymbolTableWriter::createSymtabShndx() {
246 if (!ShndxIndexes.empty())
249 ShndxIndexes.resize(NumWritten);
252 template <typename T> void SymbolTableWriter::write(T Value) {
253 EWriter.write(Value);
256 SymbolTableWriter::SymbolTableWriter(ELFObjectWriter &EWriter, bool Is64Bit)
257 : EWriter(EWriter), Is64Bit(Is64Bit), NumWritten(0) {}
259 void SymbolTableWriter::writeSymbol(uint32_t name, uint8_t info, uint64_t value,
260 uint64_t size, uint8_t other,
261 uint32_t shndx, bool Reserved) {
262 bool LargeIndex = shndx >= ELF::SHN_LORESERVE && !Reserved;
267 if (!ShndxIndexes.empty()) {
269 ShndxIndexes.push_back(shndx);
271 ShndxIndexes.push_back(0);
274 uint16_t Index = LargeIndex ? uint16_t(ELF::SHN_XINDEX) : shndx;
277 write(name); // st_name
278 write(info); // st_info
279 write(other); // st_other
280 write(Index); // st_shndx
281 write(value); // st_value
282 write(size); // st_size
284 write(name); // st_name
285 write(uint32_t(value)); // st_value
286 write(uint32_t(size)); // st_size
287 write(info); // st_info
288 write(other); // st_other
289 write(Index); // st_shndx
295 bool ELFObjectWriter::isFixupKindPCRel(const MCAssembler &Asm, unsigned Kind) {
296 const MCFixupKindInfo &FKI =
297 Asm.getBackend().getFixupKindInfo((MCFixupKind) Kind);
299 return FKI.Flags & MCFixupKindInfo::FKF_IsPCRel;
302 ELFObjectWriter::~ELFObjectWriter()
305 // Emit the ELF header.
306 void ELFObjectWriter::writeHeader(const MCAssembler &Asm) {
312 // emitWord method behaves differently for ELF32 and ELF64, writing
313 // 4 bytes in the former and 8 in the latter.
315 writeBytes(ELF::ElfMagic); // e_ident[EI_MAG0] to e_ident[EI_MAG3]
317 write8(is64Bit() ? ELF::ELFCLASS64 : ELF::ELFCLASS32); // e_ident[EI_CLASS]
320 write8(isLittleEndian() ? ELF::ELFDATA2LSB : ELF::ELFDATA2MSB);
322 write8(ELF::EV_CURRENT); // e_ident[EI_VERSION]
324 write8(TargetObjectWriter->getOSABI());
325 write8(0); // e_ident[EI_ABIVERSION]
327 WriteZeros(ELF::EI_NIDENT - ELF::EI_PAD);
329 write16(ELF::ET_REL); // e_type
331 write16(TargetObjectWriter->getEMachine()); // e_machine = target
333 write32(ELF::EV_CURRENT); // e_version
334 WriteWord(0); // e_entry, no entry point in .o file
335 WriteWord(0); // e_phoff, no program header for .o
336 WriteWord(0); // e_shoff = sec hdr table off in bytes
338 // e_flags = whatever the target wants
339 write32(Asm.getELFHeaderEFlags());
341 // e_ehsize = ELF header size
342 write16(is64Bit() ? sizeof(ELF::Elf64_Ehdr) : sizeof(ELF::Elf32_Ehdr));
344 write16(0); // e_phentsize = prog header entry size
345 write16(0); // e_phnum = # prog header entries = 0
347 // e_shentsize = Section header entry size
348 write16(is64Bit() ? sizeof(ELF::Elf64_Shdr) : sizeof(ELF::Elf32_Shdr));
350 // e_shnum = # of section header ents
353 // e_shstrndx = Section # of '.shstrtab'
354 assert(StringTableIndex < ELF::SHN_LORESERVE);
355 write16(StringTableIndex);
358 uint64_t ELFObjectWriter::SymbolValue(const MCSymbol &Sym,
359 const MCAsmLayout &Layout) {
360 if (Sym.isCommon() && Sym.isExternal())
361 return Sym.getCommonAlignment();
364 if (!Layout.getSymbolOffset(Sym, Res))
367 if (Layout.getAssembler().isThumbFunc(&Sym))
373 void ELFObjectWriter::executePostLayoutBinding(MCAssembler &Asm,
374 const MCAsmLayout &Layout) {
375 // The presence of symbol versions causes undefined symbols and
376 // versions declared with @@@ to be renamed.
378 for (const MCSymbol &A : Asm.symbols()) {
379 const auto &Alias = cast<MCSymbolELF>(A);
381 if (!Alias.isVariable())
383 auto *Ref = dyn_cast<MCSymbolRefExpr>(Alias.getVariableValue());
386 const auto &Symbol = cast<MCSymbolELF>(Ref->getSymbol());
388 StringRef AliasName = Alias.getName();
389 size_t Pos = AliasName.find('@');
390 if (Pos == StringRef::npos)
393 // Aliases defined with .symvar copy the binding from the symbol they alias.
394 // This is the first place we are able to copy this information.
395 Alias.setExternal(Symbol.isExternal());
396 Alias.setBinding(Symbol.getBinding());
398 StringRef Rest = AliasName.substr(Pos);
399 if (!Symbol.isUndefined() && !Rest.startswith("@@@"))
402 // FIXME: produce a better error message.
403 if (Symbol.isUndefined() && Rest.startswith("@@") &&
404 !Rest.startswith("@@@"))
405 report_fatal_error("A @@ version cannot be undefined");
407 Renames.insert(std::make_pair(&Symbol, &Alias));
411 static uint8_t mergeTypeForSet(uint8_t origType, uint8_t newType) {
412 uint8_t Type = newType;
414 // Propagation rules:
415 // IFUNC > FUNC > OBJECT > NOTYPE
416 // TLS_OBJECT > OBJECT > NOTYPE
418 // dont let the new type degrade the old type
422 case ELF::STT_GNU_IFUNC:
423 if (Type == ELF::STT_FUNC || Type == ELF::STT_OBJECT ||
424 Type == ELF::STT_NOTYPE || Type == ELF::STT_TLS)
425 Type = ELF::STT_GNU_IFUNC;
428 if (Type == ELF::STT_OBJECT || Type == ELF::STT_NOTYPE ||
429 Type == ELF::STT_TLS)
430 Type = ELF::STT_FUNC;
432 case ELF::STT_OBJECT:
433 if (Type == ELF::STT_NOTYPE)
434 Type = ELF::STT_OBJECT;
437 if (Type == ELF::STT_OBJECT || Type == ELF::STT_NOTYPE ||
438 Type == ELF::STT_GNU_IFUNC || Type == ELF::STT_FUNC)
446 void ELFObjectWriter::writeSymbol(SymbolTableWriter &Writer,
447 uint32_t StringIndex, ELFSymbolData &MSD,
448 const MCAsmLayout &Layout) {
449 const auto &Symbol = cast<MCSymbolELF>(*MSD.Symbol);
450 assert((!Symbol.getFragment() ||
451 (Symbol.getFragment()->getParent() == &Symbol.getSection())) &&
452 "The symbol's section doesn't match the fragment's symbol");
453 const MCSymbolELF *Base =
454 cast_or_null<MCSymbolELF>(Layout.getBaseSymbol(Symbol));
456 // This has to be in sync with when computeSymbolTable uses SHN_ABS or
458 bool IsReserved = !Base || Symbol.isCommon();
460 // Binding and Type share the same byte as upper and lower nibbles
461 uint8_t Binding = Symbol.getBinding();
462 uint8_t Type = Symbol.getType();
464 Type = mergeTypeForSet(Type, Base->getType());
466 uint8_t Info = (Binding << 4) | Type;
468 // Other and Visibility share the same byte with Visibility using the lower
470 uint8_t Visibility = Symbol.getVisibility();
471 uint8_t Other = Symbol.getOther() | Visibility;
473 uint64_t Value = SymbolValue(*MSD.Symbol, Layout);
476 const MCExpr *ESize = MSD.Symbol->getSize();
478 ESize = Base->getSize();
482 if (!ESize->evaluateKnownAbsolute(Res, Layout))
483 report_fatal_error("Size expression must be absolute.");
487 // Write out the symbol table entry
488 Writer.writeSymbol(StringIndex, Info, Value, Size, Other, MSD.SectionIndex,
492 // It is always valid to create a relocation with a symbol. It is preferable
493 // to use a relocation with a section if that is possible. Using the section
494 // allows us to omit some local symbols from the symbol table.
495 bool ELFObjectWriter::shouldRelocateWithSymbol(const MCAssembler &Asm,
496 const MCSymbolRefExpr *RefA,
497 const MCSymbol *S, uint64_t C,
498 unsigned Type) const {
499 const auto *Sym = cast_or_null<MCSymbolELF>(S);
500 // A PCRel relocation to an absolute value has no symbol (or section). We
501 // represent that with a relocation to a null section.
505 MCSymbolRefExpr::VariantKind Kind = RefA->getKind();
509 // The .odp creation emits a relocation against the symbol ".TOC." which
510 // create a R_PPC64_TOC relocation. However the relocation symbol name
511 // in final object creation should be NULL, since the symbol does not
512 // really exist, it is just the reference to TOC base for the current
513 // object file. Since the symbol is undefined, returning false results
514 // in a relocation with a null section which is the desired result.
515 case MCSymbolRefExpr::VK_PPC_TOCBASE:
518 // These VariantKind cause the relocation to refer to something other than
519 // the symbol itself, like a linker generated table. Since the address of
520 // symbol is not relevant, we cannot replace the symbol with the
521 // section and patch the difference in the addend.
522 case MCSymbolRefExpr::VK_GOT:
523 case MCSymbolRefExpr::VK_PLT:
524 case MCSymbolRefExpr::VK_GOTPCREL:
525 case MCSymbolRefExpr::VK_Mips_GOT:
526 case MCSymbolRefExpr::VK_PPC_GOT_LO:
527 case MCSymbolRefExpr::VK_PPC_GOT_HI:
528 case MCSymbolRefExpr::VK_PPC_GOT_HA:
532 // An undefined symbol is not in any section, so the relocation has to point
533 // to the symbol itself.
534 assert(Sym && "Expected a symbol");
535 if (Sym->isUndefined())
538 unsigned Binding = Sym->getBinding();
541 llvm_unreachable("Invalid Binding");
545 // If the symbol is weak, it might be overridden by a symbol in another
546 // file. The relocation has to point to the symbol so that the linker
549 case ELF::STB_GLOBAL:
550 // Global ELF symbols can be preempted by the dynamic linker. The relocation
551 // has to point to the symbol for a reason analogous to the STB_WEAK case.
555 // If a relocation points to a mergeable section, we have to be careful.
556 // If the offset is zero, a relocation with the section will encode the
557 // same information. With a non-zero offset, the situation is different.
558 // For example, a relocation can point 42 bytes past the end of a string.
559 // If we change such a relocation to use the section, the linker would think
560 // that it pointed to another string and subtracting 42 at runtime will
561 // produce the wrong value.
562 auto &Sec = cast<MCSectionELF>(Sym->getSection());
563 unsigned Flags = Sec.getFlags();
564 if (Flags & ELF::SHF_MERGE) {
568 // It looks like gold has a bug (http://sourceware.org/PR16794) and can
569 // only handle section relocations to mergeable sections if using RELA.
570 if (!hasRelocationAddend())
574 // Most TLS relocations use a got, so they need the symbol. Even those that
575 // are just an offset (@tpoff), require a symbol in gold versions before
576 // 5efeedf61e4fe720fd3e9a08e6c91c10abb66d42 (2014-09-26) which fixed
577 // http://sourceware.org/PR16773.
578 if (Flags & ELF::SHF_TLS)
581 // If the symbol is a thumb function the final relocation must set the lowest
582 // bit. With a symbol that is done by just having the symbol have that bit
583 // set, so we would lose the bit if we relocated with the section.
584 // FIXME: We could use the section but add the bit to the relocation value.
585 if (Asm.isThumbFunc(Sym))
588 if (TargetObjectWriter->needsRelocateWithSymbol(*Sym, Type))
593 // True if the assembler knows nothing about the final value of the symbol.
594 // This doesn't cover the comdat issues, since in those cases the assembler
595 // can at least know that all symbols in the section will move together.
596 static bool isWeak(const MCSymbolELF &Sym) {
597 if (Sym.getType() == ELF::STT_GNU_IFUNC)
600 switch (Sym.getBinding()) {
602 llvm_unreachable("Unknown binding");
605 case ELF::STB_GLOBAL:
608 case ELF::STB_GNU_UNIQUE:
613 void ELFObjectWriter::recordRelocation(MCAssembler &Asm,
614 const MCAsmLayout &Layout,
615 const MCFragment *Fragment,
616 const MCFixup &Fixup, MCValue Target,
617 bool &IsPCRel, uint64_t &FixedValue) {
618 const MCSectionELF &FixupSection = cast<MCSectionELF>(*Fragment->getParent());
619 uint64_t C = Target.getConstant();
620 uint64_t FixupOffset = Layout.getFragmentOffset(Fragment) + Fixup.getOffset();
622 if (const MCSymbolRefExpr *RefB = Target.getSymB()) {
623 assert(RefB->getKind() == MCSymbolRefExpr::VK_None &&
624 "Should not have constructed this");
626 // Let A, B and C being the components of Target and R be the location of
627 // the fixup. If the fixup is not pcrel, we want to compute (A - B + C).
628 // If it is pcrel, we want to compute (A - B + C - R).
630 // In general, ELF has no relocations for -B. It can only represent (A + C)
631 // or (A + C - R). If B = R + K and the relocation is not pcrel, we can
632 // replace B to implement it: (A - R - K + C)
634 Asm.getContext().reportFatalError(
636 "No relocation available to represent this relative expression");
638 const auto &SymB = cast<MCSymbolELF>(RefB->getSymbol());
640 if (SymB.isUndefined())
641 Asm.getContext().reportFatalError(
643 Twine("symbol '") + SymB.getName() +
644 "' can not be undefined in a subtraction expression");
646 assert(!SymB.isAbsolute() && "Should have been folded");
647 const MCSection &SecB = SymB.getSection();
648 if (&SecB != &FixupSection)
649 Asm.getContext().reportFatalError(
650 Fixup.getLoc(), "Cannot represent a difference across sections");
653 Asm.getContext().reportFatalError(
654 Fixup.getLoc(), "Cannot represent a subtraction with a weak symbol");
656 uint64_t SymBOffset = Layout.getSymbolOffset(SymB);
657 uint64_t K = SymBOffset - FixupOffset;
662 // We either rejected the fixup or folded B into C at this point.
663 const MCSymbolRefExpr *RefA = Target.getSymA();
664 const auto *SymA = RefA ? cast<MCSymbolELF>(&RefA->getSymbol()) : nullptr;
666 bool ViaWeakRef = false;
667 if (SymA && SymA->isVariable()) {
668 const MCExpr *Expr = SymA->getVariableValue();
669 if (const auto *Inner = dyn_cast<MCSymbolRefExpr>(Expr)) {
670 if (Inner->getKind() == MCSymbolRefExpr::VK_WEAKREF) {
671 SymA = cast<MCSymbolELF>(&Inner->getSymbol());
677 unsigned Type = GetRelocType(Target, Fixup, IsPCRel);
678 bool RelocateWithSymbol = shouldRelocateWithSymbol(Asm, RefA, SymA, C, Type);
679 if (!RelocateWithSymbol && SymA && !SymA->isUndefined())
680 C += Layout.getSymbolOffset(*SymA);
683 if (hasRelocationAddend()) {
690 if (!RelocateWithSymbol) {
691 const MCSection *SecA =
692 (SymA && !SymA->isUndefined()) ? &SymA->getSection() : nullptr;
693 auto *ELFSec = cast_or_null<MCSectionELF>(SecA);
694 const auto *SectionSymbol =
695 ELFSec ? cast<MCSymbolELF>(ELFSec->getBeginSymbol()) : nullptr;
697 SectionSymbol->setUsedInReloc();
698 ELFRelocationEntry Rec(FixupOffset, SectionSymbol, Type, Addend);
699 Relocations[&FixupSection].push_back(Rec);
704 if (const MCSymbolELF *R = Renames.lookup(SymA))
708 SymA->setIsWeakrefUsedInReloc();
710 SymA->setUsedInReloc();
712 ELFRelocationEntry Rec(FixupOffset, SymA, Type, Addend);
713 Relocations[&FixupSection].push_back(Rec);
717 bool ELFObjectWriter::isInSymtab(const MCAsmLayout &Layout,
718 const MCSymbolELF &Symbol, bool Used,
720 if (Symbol.isVariable()) {
721 const MCExpr *Expr = Symbol.getVariableValue();
722 if (const MCSymbolRefExpr *Ref = dyn_cast<MCSymbolRefExpr>(Expr)) {
723 if (Ref->getKind() == MCSymbolRefExpr::VK_WEAKREF)
734 if (Symbol.isVariable() && Symbol.isUndefined()) {
735 // FIXME: this is here just to diagnose the case of a var = commmon_sym.
736 Layout.getBaseSymbol(Symbol);
740 if (Symbol.isUndefined() && !Symbol.isBindingSet())
743 if (Symbol.isTemporary())
746 if (Symbol.getType() == ELF::STT_SECTION)
752 void ELFObjectWriter::computeSymbolTable(
753 MCAssembler &Asm, const MCAsmLayout &Layout,
754 const SectionIndexMapTy &SectionIndexMap, const RevGroupMapTy &RevGroupMap,
755 SectionOffsetsTy &SectionOffsets) {
756 MCContext &Ctx = Asm.getContext();
757 SymbolTableWriter Writer(*this, is64Bit());
760 unsigned EntrySize = is64Bit() ? ELF::SYMENTRY_SIZE64 : ELF::SYMENTRY_SIZE32;
761 MCSectionELF *SymtabSection =
762 Ctx.getELFSection(".symtab", ELF::SHT_SYMTAB, 0, EntrySize, "");
763 SymtabSection->setAlignment(is64Bit() ? 8 : 4);
764 SymbolTableIndex = addToSectionTable(SymtabSection);
766 align(SymtabSection->getAlignment());
767 uint64_t SecStart = getStream().tell();
769 // The first entry is the undefined symbol entry.
770 Writer.writeSymbol(0, 0, 0, 0, 0, 0, false);
772 std::vector<ELFSymbolData> LocalSymbolData;
773 std::vector<ELFSymbolData> ExternalSymbolData;
775 // Add the data for the symbols.
776 bool HasLargeSectionIndex = false;
777 for (const MCSymbol &S : Asm.symbols()) {
778 const auto &Symbol = cast<MCSymbolELF>(S);
779 bool Used = Symbol.isUsedInReloc();
780 bool WeakrefUsed = Symbol.isWeakrefUsedInReloc();
781 bool isSignature = Symbol.isSignature();
783 if (!isInSymtab(Layout, Symbol, Used || WeakrefUsed || isSignature,
784 Renames.count(&Symbol)))
787 if (Symbol.isTemporary() && Symbol.isUndefined())
788 Ctx.reportFatalError(SMLoc(), "Undefined temporary");
791 MSD.Symbol = cast<MCSymbolELF>(&Symbol);
793 bool Local = Symbol.getBinding() == ELF::STB_LOCAL;
794 assert(Local || !Symbol.isTemporary());
796 if (Symbol.isAbsolute()) {
797 MSD.SectionIndex = ELF::SHN_ABS;
798 } else if (Symbol.isCommon()) {
800 MSD.SectionIndex = ELF::SHN_COMMON;
801 } else if (Symbol.isUndefined()) {
802 if (isSignature && !Used) {
803 MSD.SectionIndex = RevGroupMap.lookup(&Symbol);
804 if (MSD.SectionIndex >= ELF::SHN_LORESERVE)
805 HasLargeSectionIndex = true;
807 MSD.SectionIndex = ELF::SHN_UNDEF;
810 const MCSectionELF &Section =
811 static_cast<const MCSectionELF &>(Symbol.getSection());
812 MSD.SectionIndex = SectionIndexMap.lookup(&Section);
813 assert(MSD.SectionIndex && "Invalid section index!");
814 if (MSD.SectionIndex >= ELF::SHN_LORESERVE)
815 HasLargeSectionIndex = true;
818 // The @@@ in symbol version is replaced with @ in undefined symbols and @@
821 // FIXME: All name handling should be done before we get to the writer,
822 // including dealing with GNU-style version suffixes. Fixing this isn't
825 // We thus have to be careful to not perform the symbol version replacement
828 // The ELF format is used on Windows by the MCJIT engine. Thus, on
829 // Windows, the ELFObjectWriter can encounter symbols mangled using the MS
830 // Visual Studio C++ name mangling scheme. Symbols mangled using the MSVC
831 // C++ name mangling can legally have "@@@" as a sub-string. In that case,
832 // the EFLObjectWriter should not interpret the "@@@" sub-string as
833 // specifying GNU-style symbol versioning. The ELFObjectWriter therefore
834 // checks for the MSVC C++ name mangling prefix which is either "?", "@?",
835 // "__imp_?" or "__imp_@?".
837 // It would have been interesting to perform the MS mangling prefix check
838 // only when the target triple is of the form *-pc-windows-elf. But, it
839 // seems that this information is not easily accessible from the
841 StringRef Name = Symbol.getName();
843 if (!Name.startswith("?") && !Name.startswith("@?") &&
844 !Name.startswith("__imp_?") && !Name.startswith("__imp_@?")) {
845 // This symbol isn't following the MSVC C++ name mangling convention. We
846 // can thus safely interpret the @@@ in symbol names as specifying symbol
848 size_t Pos = Name.find("@@@");
849 if (Pos != StringRef::npos) {
850 Buf += Name.substr(0, Pos);
851 unsigned Skip = MSD.SectionIndex == ELF::SHN_UNDEF ? 2 : 1;
852 Buf += Name.substr(Pos + Skip);
857 // Sections have their own string table
858 if (Symbol.getType() != ELF::STT_SECTION)
859 MSD.Name = StrTabBuilder.add(Name);
862 LocalSymbolData.push_back(MSD);
864 ExternalSymbolData.push_back(MSD);
867 // This holds the .symtab_shndx section index.
868 unsigned SymtabShndxSectionIndex = 0;
870 if (HasLargeSectionIndex) {
871 MCSectionELF *SymtabShndxSection =
872 Ctx.getELFSection(".symtab_shndxr", ELF::SHT_SYMTAB_SHNDX, 0, 4, "");
873 SymtabShndxSectionIndex = addToSectionTable(SymtabShndxSection);
874 SymtabShndxSection->setAlignment(4);
877 ArrayRef<std::string> FileNames = Asm.getFileNames();
878 for (const std::string &Name : FileNames)
879 StrTabBuilder.add(Name);
881 StrTabBuilder.finalize(StringTableBuilder::ELF);
883 for (const std::string &Name : FileNames)
884 Writer.writeSymbol(StrTabBuilder.getOffset(Name),
885 ELF::STT_FILE | ELF::STB_LOCAL, 0, 0, ELF::STV_DEFAULT,
888 // Symbols are required to be in lexicographic order.
889 array_pod_sort(LocalSymbolData.begin(), LocalSymbolData.end());
890 array_pod_sort(ExternalSymbolData.begin(), ExternalSymbolData.end());
892 // Set the symbol indices. Local symbols must come before all other
893 // symbols with non-local bindings.
894 unsigned Index = FileNames.size() + 1;
896 for (ELFSymbolData &MSD : LocalSymbolData) {
897 unsigned StringIndex = MSD.Symbol->getType() == ELF::STT_SECTION
899 : StrTabBuilder.getOffset(MSD.Name);
900 MSD.Symbol->setIndex(Index++);
901 writeSymbol(Writer, StringIndex, MSD, Layout);
904 // Write the symbol table entries.
905 LastLocalSymbolIndex = Index;
907 for (ELFSymbolData &MSD : ExternalSymbolData) {
908 unsigned StringIndex = StrTabBuilder.getOffset(MSD.Name);
909 MSD.Symbol->setIndex(Index++);
910 writeSymbol(Writer, StringIndex, MSD, Layout);
911 assert(MSD.Symbol->getBinding() != ELF::STB_LOCAL);
914 uint64_t SecEnd = getStream().tell();
915 SectionOffsets[SymtabSection] = std::make_pair(SecStart, SecEnd);
917 ArrayRef<uint32_t> ShndxIndexes = Writer.getShndxIndexes();
918 if (ShndxIndexes.empty()) {
919 assert(SymtabShndxSectionIndex == 0);
922 assert(SymtabShndxSectionIndex != 0);
924 SecStart = getStream().tell();
925 const MCSectionELF *SymtabShndxSection =
926 SectionTable[SymtabShndxSectionIndex - 1];
927 for (uint32_t Index : ShndxIndexes)
929 SecEnd = getStream().tell();
930 SectionOffsets[SymtabShndxSection] = std::make_pair(SecStart, SecEnd);
934 ELFObjectWriter::createRelocationSection(MCContext &Ctx,
935 const MCSectionELF &Sec) {
936 if (Relocations[&Sec].empty())
939 const StringRef SectionName = Sec.getSectionName();
940 std::string RelaSectionName = hasRelocationAddend() ? ".rela" : ".rel";
941 RelaSectionName += SectionName;
944 if (hasRelocationAddend())
945 EntrySize = is64Bit() ? sizeof(ELF::Elf64_Rela) : sizeof(ELF::Elf32_Rela);
947 EntrySize = is64Bit() ? sizeof(ELF::Elf64_Rel) : sizeof(ELF::Elf32_Rel);
950 if (Sec.getFlags() & ELF::SHF_GROUP)
951 Flags = ELF::SHF_GROUP;
953 MCSectionELF *RelaSection = Ctx.createELFRelSection(
954 RelaSectionName, hasRelocationAddend() ? ELF::SHT_RELA : ELF::SHT_REL,
955 Flags, EntrySize, Sec.getGroup(), &Sec);
956 RelaSection->setAlignment(is64Bit() ? 8 : 4);
960 // Include the debug info compression header:
961 // "ZLIB" followed by 8 bytes representing the uncompressed size of the section,
962 // useful for consumers to preallocate a buffer to decompress into.
964 prependCompressionHeader(uint64_t Size,
965 SmallVectorImpl<char> &CompressedContents) {
966 const StringRef Magic = "ZLIB";
967 if (Size <= Magic.size() + sizeof(Size) + CompressedContents.size())
969 if (sys::IsLittleEndianHost)
970 sys::swapByteOrder(Size);
971 CompressedContents.insert(CompressedContents.begin(),
972 Magic.size() + sizeof(Size), 0);
973 std::copy(Magic.begin(), Magic.end(), CompressedContents.begin());
974 std::copy(reinterpret_cast<char *>(&Size),
975 reinterpret_cast<char *>(&Size + 1),
976 CompressedContents.begin() + Magic.size());
980 void ELFObjectWriter::writeSectionData(const MCAssembler &Asm, MCSection &Sec,
981 const MCAsmLayout &Layout) {
982 MCSectionELF &Section = static_cast<MCSectionELF &>(Sec);
983 StringRef SectionName = Section.getSectionName();
985 // Compressing debug_frame requires handling alignment fragments which is
986 // more work (possibly generalizing MCAssembler.cpp:writeFragment to allow
987 // for writing to arbitrary buffers) for little benefit.
988 if (!Asm.getContext().getAsmInfo()->compressDebugSections() ||
989 !SectionName.startswith(".debug_") || SectionName == ".debug_frame") {
990 Asm.writeSectionData(&Section, Layout);
994 SmallVector<char, 128> UncompressedData;
995 raw_svector_ostream VecOS(UncompressedData);
996 raw_pwrite_stream &OldStream = getStream();
998 Asm.writeSectionData(&Section, Layout);
999 setStream(OldStream);
1001 SmallVector<char, 128> CompressedContents;
1002 zlib::Status Success = zlib::compress(
1003 StringRef(UncompressedData.data(), UncompressedData.size()),
1004 CompressedContents);
1005 if (Success != zlib::StatusOK) {
1006 getStream() << UncompressedData;
1010 if (!prependCompressionHeader(UncompressedData.size(), CompressedContents)) {
1011 getStream() << UncompressedData;
1014 Asm.getContext().renameELFSection(&Section,
1015 (".z" + SectionName.drop_front(1)).str());
1016 getStream() << CompressedContents;
1019 void ELFObjectWriter::WriteSecHdrEntry(uint32_t Name, uint32_t Type,
1020 uint64_t Flags, uint64_t Address,
1021 uint64_t Offset, uint64_t Size,
1022 uint32_t Link, uint32_t Info,
1024 uint64_t EntrySize) {
1025 write32(Name); // sh_name: index into string table
1026 write32(Type); // sh_type
1027 WriteWord(Flags); // sh_flags
1028 WriteWord(Address); // sh_addr
1029 WriteWord(Offset); // sh_offset
1030 WriteWord(Size); // sh_size
1031 write32(Link); // sh_link
1032 write32(Info); // sh_info
1033 WriteWord(Alignment); // sh_addralign
1034 WriteWord(EntrySize); // sh_entsize
1037 void ELFObjectWriter::writeRelocations(const MCAssembler &Asm,
1038 const MCSectionELF &Sec) {
1039 std::vector<ELFRelocationEntry> &Relocs = Relocations[&Sec];
1041 // Sort the relocation entries. Most targets just sort by Offset, but some
1042 // (e.g., MIPS) have additional constraints.
1043 TargetObjectWriter->sortRelocs(Asm, Relocs);
1045 for (unsigned i = 0, e = Relocs.size(); i != e; ++i) {
1046 const ELFRelocationEntry &Entry = Relocs[e - i - 1];
1047 unsigned Index = Entry.Symbol ? Entry.Symbol->getIndex() : 0;
1050 write(Entry.Offset);
1051 if (TargetObjectWriter->isN64()) {
1052 write(uint32_t(Index));
1054 write(TargetObjectWriter->getRSsym(Entry.Type));
1055 write(TargetObjectWriter->getRType3(Entry.Type));
1056 write(TargetObjectWriter->getRType2(Entry.Type));
1057 write(TargetObjectWriter->getRType(Entry.Type));
1059 struct ELF::Elf64_Rela ERE64;
1060 ERE64.setSymbolAndType(Index, Entry.Type);
1061 write(ERE64.r_info);
1063 if (hasRelocationAddend())
1064 write(Entry.Addend);
1066 write(uint32_t(Entry.Offset));
1068 struct ELF::Elf32_Rela ERE32;
1069 ERE32.setSymbolAndType(Index, Entry.Type);
1070 write(ERE32.r_info);
1072 if (hasRelocationAddend())
1073 write(uint32_t(Entry.Addend));
1078 const MCSectionELF *ELFObjectWriter::createStringTable(MCContext &Ctx) {
1079 const MCSectionELF *StrtabSection = SectionTable[StringTableIndex - 1];
1080 getStream() << StrTabBuilder.data();
1081 return StrtabSection;
1084 void ELFObjectWriter::writeSection(const SectionIndexMapTy &SectionIndexMap,
1085 uint32_t GroupSymbolIndex, uint64_t Offset,
1086 uint64_t Size, const MCSectionELF &Section) {
1087 uint64_t sh_link = 0;
1088 uint64_t sh_info = 0;
1090 switch(Section.getType()) {
1095 case ELF::SHT_DYNAMIC:
1096 llvm_unreachable("SHT_DYNAMIC in a relocatable object");
1099 case ELF::SHT_RELA: {
1100 sh_link = SymbolTableIndex;
1101 assert(sh_link && ".symtab not found");
1102 const MCSectionELF *InfoSection = Section.getAssociatedSection();
1103 sh_info = SectionIndexMap.lookup(InfoSection);
1107 case ELF::SHT_SYMTAB:
1108 case ELF::SHT_DYNSYM:
1109 sh_link = StringTableIndex;
1110 sh_info = LastLocalSymbolIndex;
1113 case ELF::SHT_SYMTAB_SHNDX:
1114 sh_link = SymbolTableIndex;
1117 case ELF::SHT_GROUP:
1118 sh_link = SymbolTableIndex;
1119 sh_info = GroupSymbolIndex;
1123 if (TargetObjectWriter->getEMachine() == ELF::EM_ARM &&
1124 Section.getType() == ELF::SHT_ARM_EXIDX)
1125 sh_link = SectionIndexMap.lookup(Section.getAssociatedSection());
1127 WriteSecHdrEntry(StrTabBuilder.getOffset(Section.getSectionName()),
1128 Section.getType(), Section.getFlags(), 0, Offset, Size,
1129 sh_link, sh_info, Section.getAlignment(),
1130 Section.getEntrySize());
1133 void ELFObjectWriter::writeSectionHeader(
1134 const MCAsmLayout &Layout, const SectionIndexMapTy &SectionIndexMap,
1135 const SectionOffsetsTy &SectionOffsets) {
1136 const unsigned NumSections = SectionTable.size();
1138 // Null section first.
1139 uint64_t FirstSectionSize =
1140 (NumSections + 1) >= ELF::SHN_LORESERVE ? NumSections + 1 : 0;
1141 WriteSecHdrEntry(0, 0, 0, 0, 0, FirstSectionSize, 0, 0, 0, 0);
1143 for (const MCSectionELF *Section : SectionTable) {
1144 uint32_t GroupSymbolIndex;
1145 unsigned Type = Section->getType();
1146 if (Type != ELF::SHT_GROUP)
1147 GroupSymbolIndex = 0;
1149 GroupSymbolIndex = Section->getGroup()->getIndex();
1151 const std::pair<uint64_t, uint64_t> &Offsets =
1152 SectionOffsets.find(Section)->second;
1154 if (Type == ELF::SHT_NOBITS)
1155 Size = Layout.getSectionAddressSize(Section);
1157 Size = Offsets.second - Offsets.first;
1159 writeSection(SectionIndexMap, GroupSymbolIndex, Offsets.first, Size,
1164 void ELFObjectWriter::writeObject(MCAssembler &Asm,
1165 const MCAsmLayout &Layout) {
1166 MCContext &Ctx = Asm.getContext();
1167 MCSectionELF *StrtabSection =
1168 Ctx.getELFSection(".strtab", ELF::SHT_STRTAB, 0);
1169 StringTableIndex = addToSectionTable(StrtabSection);
1171 RevGroupMapTy RevGroupMap;
1172 SectionIndexMapTy SectionIndexMap;
1174 std::map<const MCSymbol *, std::vector<const MCSectionELF *>> GroupMembers;
1176 // Write out the ELF header ...
1179 // ... then the sections ...
1180 SectionOffsetsTy SectionOffsets;
1181 std::vector<MCSectionELF *> Groups;
1182 std::vector<MCSectionELF *> Relocations;
1183 for (MCSection &Sec : Asm) {
1184 MCSectionELF &Section = static_cast<MCSectionELF &>(Sec);
1186 align(Section.getAlignment());
1188 // Remember the offset into the file for this section.
1189 uint64_t SecStart = getStream().tell();
1191 const MCSymbolELF *SignatureSymbol = Section.getGroup();
1192 writeSectionData(Asm, Section, Layout);
1194 uint64_t SecEnd = getStream().tell();
1195 SectionOffsets[&Section] = std::make_pair(SecStart, SecEnd);
1197 MCSectionELF *RelSection = createRelocationSection(Ctx, Section);
1199 if (SignatureSymbol) {
1200 Asm.registerSymbol(*SignatureSymbol);
1201 unsigned &GroupIdx = RevGroupMap[SignatureSymbol];
1203 MCSectionELF *Group = Ctx.createELFGroupSection(SignatureSymbol);
1204 GroupIdx = addToSectionTable(Group);
1205 Group->setAlignment(4);
1206 Groups.push_back(Group);
1208 std::vector<const MCSectionELF *> &Members =
1209 GroupMembers[SignatureSymbol];
1210 Members.push_back(&Section);
1212 Members.push_back(RelSection);
1215 SectionIndexMap[&Section] = addToSectionTable(&Section);
1217 SectionIndexMap[RelSection] = addToSectionTable(RelSection);
1218 Relocations.push_back(RelSection);
1222 for (MCSectionELF *Group : Groups) {
1223 align(Group->getAlignment());
1225 // Remember the offset into the file for this section.
1226 uint64_t SecStart = getStream().tell();
1228 const MCSymbol *SignatureSymbol = Group->getGroup();
1229 assert(SignatureSymbol);
1230 write(uint32_t(ELF::GRP_COMDAT));
1231 for (const MCSectionELF *Member : GroupMembers[SignatureSymbol]) {
1232 uint32_t SecIndex = SectionIndexMap.lookup(Member);
1236 uint64_t SecEnd = getStream().tell();
1237 SectionOffsets[Group] = std::make_pair(SecStart, SecEnd);
1240 // Compute symbol table information.
1241 computeSymbolTable(Asm, Layout, SectionIndexMap, RevGroupMap, SectionOffsets);
1243 for (MCSectionELF *RelSection : Relocations) {
1244 align(RelSection->getAlignment());
1246 // Remember the offset into the file for this section.
1247 uint64_t SecStart = getStream().tell();
1249 writeRelocations(Asm, *RelSection->getAssociatedSection());
1251 uint64_t SecEnd = getStream().tell();
1252 SectionOffsets[RelSection] = std::make_pair(SecStart, SecEnd);
1256 uint64_t SecStart = getStream().tell();
1257 const MCSectionELF *Sec = createStringTable(Ctx);
1258 uint64_t SecEnd = getStream().tell();
1259 SectionOffsets[Sec] = std::make_pair(SecStart, SecEnd);
1262 uint64_t NaturalAlignment = is64Bit() ? 8 : 4;
1263 align(NaturalAlignment);
1265 const unsigned SectionHeaderOffset = getStream().tell();
1267 // ... then the section header table ...
1268 writeSectionHeader(Layout, SectionIndexMap, SectionOffsets);
1270 uint16_t NumSections = (SectionTable.size() + 1 >= ELF::SHN_LORESERVE)
1271 ? (uint16_t)ELF::SHN_UNDEF
1272 : SectionTable.size() + 1;
1273 if (sys::IsLittleEndianHost != IsLittleEndian)
1274 sys::swapByteOrder(NumSections);
1275 unsigned NumSectionsOffset;
1278 uint64_t Val = SectionHeaderOffset;
1279 if (sys::IsLittleEndianHost != IsLittleEndian)
1280 sys::swapByteOrder(Val);
1281 getStream().pwrite(reinterpret_cast<char *>(&Val), sizeof(Val),
1282 offsetof(ELF::Elf64_Ehdr, e_shoff));
1283 NumSectionsOffset = offsetof(ELF::Elf64_Ehdr, e_shnum);
1285 uint32_t Val = SectionHeaderOffset;
1286 if (sys::IsLittleEndianHost != IsLittleEndian)
1287 sys::swapByteOrder(Val);
1288 getStream().pwrite(reinterpret_cast<char *>(&Val), sizeof(Val),
1289 offsetof(ELF::Elf32_Ehdr, e_shoff));
1290 NumSectionsOffset = offsetof(ELF::Elf32_Ehdr, e_shnum);
1292 getStream().pwrite(reinterpret_cast<char *>(&NumSections),
1293 sizeof(NumSections), NumSectionsOffset);
1296 bool ELFObjectWriter::isSymbolRefDifferenceFullyResolvedImpl(
1297 const MCAssembler &Asm, const MCSymbol &SA, const MCFragment &FB,
1298 bool InSet, bool IsPCRel) const {
1299 const auto &SymA = cast<MCSymbolELF>(SA);
1305 return MCObjectWriter::isSymbolRefDifferenceFullyResolvedImpl(Asm, SymA, FB,
1309 bool ELFObjectWriter::isWeak(const MCSymbol &S) const {
1310 const auto &Sym = cast<MCSymbolELF>(S);
1314 // It is invalid to replace a reference to a global in a comdat
1315 // with a reference to a local since out of comdat references
1316 // to a local are forbidden.
1317 // We could try to return false for more cases, like the reference
1318 // being in the same comdat or Sym being an alias to another global,
1319 // but it is not clear if it is worth the effort.
1320 if (Sym.getBinding() != ELF::STB_GLOBAL)
1323 if (!Sym.isInSection())
1326 const auto &Sec = cast<MCSectionELF>(Sym.getSection());
1327 return Sec.getGroup();
1330 MCObjectWriter *llvm::createELFObjectWriter(MCELFObjectTargetWriter *MOTW,
1331 raw_pwrite_stream &OS,
1332 bool IsLittleEndian) {
1333 return new ELFObjectWriter(MOTW, OS, IsLittleEndian);