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;
119 // This holds the .symtab_shndx section index.
120 unsigned SymtabShndxSectionIndex = 0;
122 // Sections in the order they are to be output in the section table.
123 std::vector<const MCSectionELF *> SectionTable;
124 unsigned addToSectionTable(const MCSectionELF *Sec);
126 // TargetObjectWriter wrappers.
127 bool is64Bit() const { return TargetObjectWriter->is64Bit(); }
128 bool hasRelocationAddend() const {
129 return TargetObjectWriter->hasRelocationAddend();
131 unsigned GetRelocType(const MCValue &Target, const MCFixup &Fixup,
132 bool IsPCRel) const {
133 return TargetObjectWriter->GetRelocType(Target, Fixup, IsPCRel);
136 void align(unsigned Alignment);
139 ELFObjectWriter(MCELFObjectTargetWriter *MOTW, raw_pwrite_stream &OS,
141 : MCObjectWriter(OS, IsLittleEndian), TargetObjectWriter(MOTW) {}
143 void reset() override {
146 StrTabBuilder.clear();
147 SectionTable.clear();
148 MCObjectWriter::reset();
151 ~ELFObjectWriter() override;
153 void WriteWord(uint64_t W) {
160 template <typename T> void write(T Val) {
162 support::endian::Writer<support::little>(OS).write(Val);
164 support::endian::Writer<support::big>(OS).write(Val);
167 void writeHeader(const MCAssembler &Asm);
169 void writeSymbol(SymbolTableWriter &Writer, uint32_t StringIndex,
170 ELFSymbolData &MSD, const MCAsmLayout &Layout);
172 // Start and end offset of each section
173 typedef std::map<const MCSectionELF *, std::pair<uint64_t, uint64_t>>
176 bool shouldRelocateWithSymbol(const MCAssembler &Asm,
177 const MCSymbolRefExpr *RefA,
178 const MCSymbol *Sym, uint64_t C,
179 unsigned Type) const;
181 void recordRelocation(MCAssembler &Asm, const MCAsmLayout &Layout,
182 const MCFragment *Fragment, const MCFixup &Fixup,
183 MCValue Target, bool &IsPCRel,
184 uint64_t &FixedValue) override;
186 // Map from a signature symbol to the group section index
187 typedef DenseMap<const MCSymbol *, unsigned> RevGroupMapTy;
189 /// Compute the symbol table data
191 /// \param Asm - The assembler.
192 /// \param SectionIndexMap - Maps a section to its index.
193 /// \param RevGroupMap - Maps a signature symbol to the group section.
194 void computeSymbolTable(MCAssembler &Asm, const MCAsmLayout &Layout,
195 const SectionIndexMapTy &SectionIndexMap,
196 const RevGroupMapTy &RevGroupMap,
197 SectionOffsetsTy &SectionOffsets);
199 MCSectionELF *createRelocationSection(MCContext &Ctx,
200 const MCSectionELF &Sec);
202 const MCSectionELF *createStringTable(MCContext &Ctx);
204 void executePostLayoutBinding(MCAssembler &Asm,
205 const MCAsmLayout &Layout) override;
207 void writeSectionHeader(const MCAsmLayout &Layout,
208 const SectionIndexMapTy &SectionIndexMap,
209 const SectionOffsetsTy &SectionOffsets);
211 void writeSectionData(const MCAssembler &Asm, MCSection &Sec,
212 const MCAsmLayout &Layout);
214 void WriteSecHdrEntry(uint32_t Name, uint32_t Type, uint64_t Flags,
215 uint64_t Address, uint64_t Offset, uint64_t Size,
216 uint32_t Link, uint32_t Info, uint64_t Alignment,
219 void writeRelocations(const MCAssembler &Asm, const MCSectionELF &Sec);
221 bool isSymbolRefDifferenceFullyResolvedImpl(const MCAssembler &Asm,
222 const MCSymbol &SymA,
223 const MCFragment &FB,
225 bool IsPCRel) const override;
227 bool isWeak(const MCSymbol &Sym) const override;
229 void writeObject(MCAssembler &Asm, const MCAsmLayout &Layout) override;
230 void writeSection(const SectionIndexMapTy &SectionIndexMap,
231 uint32_t GroupSymbolIndex, uint64_t Offset, uint64_t Size,
232 const MCSectionELF &Section);
236 void ELFObjectWriter::align(unsigned Alignment) {
237 uint64_t Padding = OffsetToAlignment(OS.tell(), Alignment);
241 unsigned ELFObjectWriter::addToSectionTable(const MCSectionELF *Sec) {
242 SectionTable.push_back(Sec);
243 StrTabBuilder.add(Sec->getSectionName());
244 return SectionTable.size();
247 void SymbolTableWriter::createSymtabShndx() {
248 if (!ShndxIndexes.empty())
251 ShndxIndexes.resize(NumWritten);
254 template <typename T> void SymbolTableWriter::write(T Value) {
255 EWriter.write(Value);
258 SymbolTableWriter::SymbolTableWriter(ELFObjectWriter &EWriter, bool Is64Bit)
259 : EWriter(EWriter), Is64Bit(Is64Bit), NumWritten(0) {}
261 void SymbolTableWriter::writeSymbol(uint32_t name, uint8_t info, uint64_t value,
262 uint64_t size, uint8_t other,
263 uint32_t shndx, bool Reserved) {
264 bool LargeIndex = shndx >= ELF::SHN_LORESERVE && !Reserved;
269 if (!ShndxIndexes.empty()) {
271 ShndxIndexes.push_back(shndx);
273 ShndxIndexes.push_back(0);
276 uint16_t Index = LargeIndex ? uint16_t(ELF::SHN_XINDEX) : shndx;
279 write(name); // st_name
280 write(info); // st_info
281 write(other); // st_other
282 write(Index); // st_shndx
283 write(value); // st_value
284 write(size); // st_size
286 write(name); // st_name
287 write(uint32_t(value)); // st_value
288 write(uint32_t(size)); // st_size
289 write(info); // st_info
290 write(other); // st_other
291 write(Index); // st_shndx
297 bool ELFObjectWriter::isFixupKindPCRel(const MCAssembler &Asm, unsigned Kind) {
298 const MCFixupKindInfo &FKI =
299 Asm.getBackend().getFixupKindInfo((MCFixupKind) Kind);
301 return FKI.Flags & MCFixupKindInfo::FKF_IsPCRel;
304 ELFObjectWriter::~ELFObjectWriter()
307 // Emit the ELF header.
308 void ELFObjectWriter::writeHeader(const MCAssembler &Asm) {
314 // emitWord method behaves differently for ELF32 and ELF64, writing
315 // 4 bytes in the former and 8 in the latter.
317 writeBytes(ELF::ElfMagic); // e_ident[EI_MAG0] to e_ident[EI_MAG3]
319 write8(is64Bit() ? ELF::ELFCLASS64 : ELF::ELFCLASS32); // e_ident[EI_CLASS]
322 write8(isLittleEndian() ? ELF::ELFDATA2LSB : ELF::ELFDATA2MSB);
324 write8(ELF::EV_CURRENT); // e_ident[EI_VERSION]
326 write8(TargetObjectWriter->getOSABI());
327 write8(0); // e_ident[EI_ABIVERSION]
329 WriteZeros(ELF::EI_NIDENT - ELF::EI_PAD);
331 write16(ELF::ET_REL); // e_type
333 write16(TargetObjectWriter->getEMachine()); // e_machine = target
335 write32(ELF::EV_CURRENT); // e_version
336 WriteWord(0); // e_entry, no entry point in .o file
337 WriteWord(0); // e_phoff, no program header for .o
338 WriteWord(0); // e_shoff = sec hdr table off in bytes
340 // e_flags = whatever the target wants
341 write32(Asm.getELFHeaderEFlags());
343 // e_ehsize = ELF header size
344 write16(is64Bit() ? sizeof(ELF::Elf64_Ehdr) : sizeof(ELF::Elf32_Ehdr));
346 write16(0); // e_phentsize = prog header entry size
347 write16(0); // e_phnum = # prog header entries = 0
349 // e_shentsize = Section header entry size
350 write16(is64Bit() ? sizeof(ELF::Elf64_Shdr) : sizeof(ELF::Elf32_Shdr));
352 // e_shnum = # of section header ents
355 // e_shstrndx = Section # of '.shstrtab'
356 assert(StringTableIndex < ELF::SHN_LORESERVE);
357 write16(StringTableIndex);
360 uint64_t ELFObjectWriter::SymbolValue(const MCSymbol &Sym,
361 const MCAsmLayout &Layout) {
362 if (Sym.isCommon() && Sym.isExternal())
363 return Sym.getCommonAlignment();
366 if (!Layout.getSymbolOffset(Sym, Res))
369 if (Layout.getAssembler().isThumbFunc(&Sym))
375 void ELFObjectWriter::executePostLayoutBinding(MCAssembler &Asm,
376 const MCAsmLayout &Layout) {
377 // The presence of symbol versions causes undefined symbols and
378 // versions declared with @@@ to be renamed.
380 for (const MCSymbol &A : Asm.symbols()) {
381 const auto &Alias = cast<MCSymbolELF>(A);
383 if (!Alias.isVariable())
385 auto *Ref = dyn_cast<MCSymbolRefExpr>(Alias.getVariableValue());
388 const auto &Symbol = cast<MCSymbolELF>(Ref->getSymbol());
390 StringRef AliasName = Alias.getName();
391 size_t Pos = AliasName.find('@');
392 if (Pos == StringRef::npos)
395 // Aliases defined with .symvar copy the binding from the symbol they alias.
396 // This is the first place we are able to copy this information.
397 Alias.setExternal(Symbol.isExternal());
398 Alias.setBinding(Symbol.getBinding());
400 StringRef Rest = AliasName.substr(Pos);
401 if (!Symbol.isUndefined() && !Rest.startswith("@@@"))
404 // FIXME: produce a better error message.
405 if (Symbol.isUndefined() && Rest.startswith("@@") &&
406 !Rest.startswith("@@@"))
407 report_fatal_error("A @@ version cannot be undefined");
409 Renames.insert(std::make_pair(&Symbol, &Alias));
413 static uint8_t mergeTypeForSet(uint8_t origType, uint8_t newType) {
414 uint8_t Type = newType;
416 // Propagation rules:
417 // IFUNC > FUNC > OBJECT > NOTYPE
418 // TLS_OBJECT > OBJECT > NOTYPE
420 // dont let the new type degrade the old type
424 case ELF::STT_GNU_IFUNC:
425 if (Type == ELF::STT_FUNC || Type == ELF::STT_OBJECT ||
426 Type == ELF::STT_NOTYPE || Type == ELF::STT_TLS)
427 Type = ELF::STT_GNU_IFUNC;
430 if (Type == ELF::STT_OBJECT || Type == ELF::STT_NOTYPE ||
431 Type == ELF::STT_TLS)
432 Type = ELF::STT_FUNC;
434 case ELF::STT_OBJECT:
435 if (Type == ELF::STT_NOTYPE)
436 Type = ELF::STT_OBJECT;
439 if (Type == ELF::STT_OBJECT || Type == ELF::STT_NOTYPE ||
440 Type == ELF::STT_GNU_IFUNC || Type == ELF::STT_FUNC)
448 void ELFObjectWriter::writeSymbol(SymbolTableWriter &Writer,
449 uint32_t StringIndex, ELFSymbolData &MSD,
450 const MCAsmLayout &Layout) {
451 const auto &Symbol = cast<MCSymbolELF>(*MSD.Symbol);
452 assert((!Symbol.getFragment() ||
453 (Symbol.getFragment()->getParent() == &Symbol.getSection())) &&
454 "The symbol's section doesn't match the fragment's symbol");
455 const MCSymbolELF *Base =
456 cast_or_null<MCSymbolELF>(Layout.getBaseSymbol(Symbol));
458 // This has to be in sync with when computeSymbolTable uses SHN_ABS or
460 bool IsReserved = !Base || Symbol.isCommon();
462 // Binding and Type share the same byte as upper and lower nibbles
463 uint8_t Binding = Symbol.getBinding();
464 uint8_t Type = Symbol.getType();
466 Type = mergeTypeForSet(Type, Base->getType());
468 uint8_t Info = (Binding << 4) | Type;
470 // Other and Visibility share the same byte with Visibility using the lower
472 uint8_t Visibility = Symbol.getVisibility();
473 uint8_t Other = Symbol.getOther() | Visibility;
475 uint64_t Value = SymbolValue(*MSD.Symbol, Layout);
478 const MCExpr *ESize = MSD.Symbol->getSize();
480 ESize = Base->getSize();
484 if (!ESize->evaluateKnownAbsolute(Res, Layout))
485 report_fatal_error("Size expression must be absolute.");
489 // Write out the symbol table entry
490 Writer.writeSymbol(StringIndex, Info, Value, Size, Other, MSD.SectionIndex,
494 // It is always valid to create a relocation with a symbol. It is preferable
495 // to use a relocation with a section if that is possible. Using the section
496 // allows us to omit some local symbols from the symbol table.
497 bool ELFObjectWriter::shouldRelocateWithSymbol(const MCAssembler &Asm,
498 const MCSymbolRefExpr *RefA,
499 const MCSymbol *S, uint64_t C,
500 unsigned Type) const {
501 const auto *Sym = cast_or_null<MCSymbolELF>(S);
502 // A PCRel relocation to an absolute value has no symbol (or section). We
503 // represent that with a relocation to a null section.
507 MCSymbolRefExpr::VariantKind Kind = RefA->getKind();
511 // The .odp creation emits a relocation against the symbol ".TOC." which
512 // create a R_PPC64_TOC relocation. However the relocation symbol name
513 // in final object creation should be NULL, since the symbol does not
514 // really exist, it is just the reference to TOC base for the current
515 // object file. Since the symbol is undefined, returning false results
516 // in a relocation with a null section which is the desired result.
517 case MCSymbolRefExpr::VK_PPC_TOCBASE:
520 // These VariantKind cause the relocation to refer to something other than
521 // the symbol itself, like a linker generated table. Since the address of
522 // symbol is not relevant, we cannot replace the symbol with the
523 // section and patch the difference in the addend.
524 case MCSymbolRefExpr::VK_GOT:
525 case MCSymbolRefExpr::VK_PLT:
526 case MCSymbolRefExpr::VK_GOTPCREL:
527 case MCSymbolRefExpr::VK_Mips_GOT:
528 case MCSymbolRefExpr::VK_PPC_GOT_LO:
529 case MCSymbolRefExpr::VK_PPC_GOT_HI:
530 case MCSymbolRefExpr::VK_PPC_GOT_HA:
534 // An undefined symbol is not in any section, so the relocation has to point
535 // to the symbol itself.
536 assert(Sym && "Expected a symbol");
537 if (Sym->isUndefined())
540 unsigned Binding = Sym->getBinding();
543 llvm_unreachable("Invalid Binding");
547 // If the symbol is weak, it might be overridden by a symbol in another
548 // file. The relocation has to point to the symbol so that the linker
551 case ELF::STB_GLOBAL:
552 // Global ELF symbols can be preempted by the dynamic linker. The relocation
553 // has to point to the symbol for a reason analogous to the STB_WEAK case.
557 // If a relocation points to a mergeable section, we have to be careful.
558 // If the offset is zero, a relocation with the section will encode the
559 // same information. With a non-zero offset, the situation is different.
560 // For example, a relocation can point 42 bytes past the end of a string.
561 // If we change such a relocation to use the section, the linker would think
562 // that it pointed to another string and subtracting 42 at runtime will
563 // produce the wrong value.
564 auto &Sec = cast<MCSectionELF>(Sym->getSection());
565 unsigned Flags = Sec.getFlags();
566 if (Flags & ELF::SHF_MERGE) {
570 // It looks like gold has a bug (http://sourceware.org/PR16794) and can
571 // only handle section relocations to mergeable sections if using RELA.
572 if (!hasRelocationAddend())
576 // Most TLS relocations use a got, so they need the symbol. Even those that
577 // are just an offset (@tpoff), require a symbol in gold versions before
578 // 5efeedf61e4fe720fd3e9a08e6c91c10abb66d42 (2014-09-26) which fixed
579 // http://sourceware.org/PR16773.
580 if (Flags & ELF::SHF_TLS)
583 // If the symbol is a thumb function the final relocation must set the lowest
584 // bit. With a symbol that is done by just having the symbol have that bit
585 // set, so we would lose the bit if we relocated with the section.
586 // FIXME: We could use the section but add the bit to the relocation value.
587 if (Asm.isThumbFunc(Sym))
590 if (TargetObjectWriter->needsRelocateWithSymbol(*Sym, Type))
595 // True if the assembler knows nothing about the final value of the symbol.
596 // This doesn't cover the comdat issues, since in those cases the assembler
597 // can at least know that all symbols in the section will move together.
598 static bool isWeak(const MCSymbolELF &Sym) {
599 if (Sym.getType() == ELF::STT_GNU_IFUNC)
602 switch (Sym.getBinding()) {
604 llvm_unreachable("Unknown binding");
607 case ELF::STB_GLOBAL:
610 case ELF::STB_GNU_UNIQUE:
615 void ELFObjectWriter::recordRelocation(MCAssembler &Asm,
616 const MCAsmLayout &Layout,
617 const MCFragment *Fragment,
618 const MCFixup &Fixup, MCValue Target,
619 bool &IsPCRel, uint64_t &FixedValue) {
620 const MCSectionELF &FixupSection = cast<MCSectionELF>(*Fragment->getParent());
621 uint64_t C = Target.getConstant();
622 uint64_t FixupOffset = Layout.getFragmentOffset(Fragment) + Fixup.getOffset();
624 if (const MCSymbolRefExpr *RefB = Target.getSymB()) {
625 assert(RefB->getKind() == MCSymbolRefExpr::VK_None &&
626 "Should not have constructed this");
628 // Let A, B and C being the components of Target and R be the location of
629 // the fixup. If the fixup is not pcrel, we want to compute (A - B + C).
630 // If it is pcrel, we want to compute (A - B + C - R).
632 // In general, ELF has no relocations for -B. It can only represent (A + C)
633 // or (A + C - R). If B = R + K and the relocation is not pcrel, we can
634 // replace B to implement it: (A - R - K + C)
636 Asm.getContext().reportFatalError(
638 "No relocation available to represent this relative expression");
640 const auto &SymB = cast<MCSymbolELF>(RefB->getSymbol());
642 if (SymB.isUndefined())
643 Asm.getContext().reportFatalError(
645 Twine("symbol '") + SymB.getName() +
646 "' can not be undefined in a subtraction expression");
648 assert(!SymB.isAbsolute() && "Should have been folded");
649 const MCSection &SecB = SymB.getSection();
650 if (&SecB != &FixupSection)
651 Asm.getContext().reportFatalError(
652 Fixup.getLoc(), "Cannot represent a difference across sections");
655 Asm.getContext().reportFatalError(
656 Fixup.getLoc(), "Cannot represent a subtraction with a weak symbol");
658 uint64_t SymBOffset = Layout.getSymbolOffset(SymB);
659 uint64_t K = SymBOffset - FixupOffset;
664 // We either rejected the fixup or folded B into C at this point.
665 const MCSymbolRefExpr *RefA = Target.getSymA();
666 const auto *SymA = RefA ? cast<MCSymbolELF>(&RefA->getSymbol()) : nullptr;
668 bool ViaWeakRef = false;
669 if (SymA && SymA->isVariable()) {
670 const MCExpr *Expr = SymA->getVariableValue();
671 if (const auto *Inner = dyn_cast<MCSymbolRefExpr>(Expr)) {
672 if (Inner->getKind() == MCSymbolRefExpr::VK_WEAKREF) {
673 SymA = cast<MCSymbolELF>(&Inner->getSymbol());
679 unsigned Type = GetRelocType(Target, Fixup, IsPCRel);
680 bool RelocateWithSymbol = shouldRelocateWithSymbol(Asm, RefA, SymA, C, Type);
681 if (!RelocateWithSymbol && SymA && !SymA->isUndefined())
682 C += Layout.getSymbolOffset(*SymA);
685 if (hasRelocationAddend()) {
692 if (!RelocateWithSymbol) {
693 const MCSection *SecA =
694 (SymA && !SymA->isUndefined()) ? &SymA->getSection() : nullptr;
695 auto *ELFSec = cast_or_null<MCSectionELF>(SecA);
696 const auto *SectionSymbol =
697 ELFSec ? cast<MCSymbolELF>(ELFSec->getBeginSymbol()) : nullptr;
699 SectionSymbol->setUsedInReloc();
700 ELFRelocationEntry Rec(FixupOffset, SectionSymbol, Type, Addend);
701 Relocations[&FixupSection].push_back(Rec);
706 if (const MCSymbolELF *R = Renames.lookup(SymA))
710 SymA->setIsWeakrefUsedInReloc();
712 SymA->setUsedInReloc();
714 ELFRelocationEntry Rec(FixupOffset, SymA, Type, Addend);
715 Relocations[&FixupSection].push_back(Rec);
719 bool ELFObjectWriter::isInSymtab(const MCAsmLayout &Layout,
720 const MCSymbolELF &Symbol, bool Used,
722 if (Symbol.isVariable()) {
723 const MCExpr *Expr = Symbol.getVariableValue();
724 if (const MCSymbolRefExpr *Ref = dyn_cast<MCSymbolRefExpr>(Expr)) {
725 if (Ref->getKind() == MCSymbolRefExpr::VK_WEAKREF)
736 if (Symbol.isVariable() && Symbol.isUndefined()) {
737 // FIXME: this is here just to diagnose the case of a var = commmon_sym.
738 Layout.getBaseSymbol(Symbol);
742 if (Symbol.isUndefined() && !Symbol.isBindingSet())
745 if (Symbol.isTemporary())
748 if (Symbol.getType() == ELF::STT_SECTION)
754 void ELFObjectWriter::computeSymbolTable(
755 MCAssembler &Asm, const MCAsmLayout &Layout,
756 const SectionIndexMapTy &SectionIndexMap, const RevGroupMapTy &RevGroupMap,
757 SectionOffsetsTy &SectionOffsets) {
758 MCContext &Ctx = Asm.getContext();
759 SymbolTableWriter Writer(*this, is64Bit());
762 unsigned EntrySize = is64Bit() ? ELF::SYMENTRY_SIZE64 : ELF::SYMENTRY_SIZE32;
763 MCSectionELF *SymtabSection =
764 Ctx.getELFSection(".symtab", ELF::SHT_SYMTAB, 0, EntrySize, "");
765 SymtabSection->setAlignment(is64Bit() ? 8 : 4);
766 SymbolTableIndex = addToSectionTable(SymtabSection);
768 align(SymtabSection->getAlignment());
769 uint64_t SecStart = OS.tell();
771 // The first entry is the undefined symbol entry.
772 Writer.writeSymbol(0, 0, 0, 0, 0, 0, false);
774 std::vector<ELFSymbolData> LocalSymbolData;
775 std::vector<ELFSymbolData> ExternalSymbolData;
777 // Add the data for the symbols.
778 bool HasLargeSectionIndex = false;
779 for (const MCSymbol &S : Asm.symbols()) {
780 const auto &Symbol = cast<MCSymbolELF>(S);
781 bool Used = Symbol.isUsedInReloc();
782 bool WeakrefUsed = Symbol.isWeakrefUsedInReloc();
783 bool isSignature = Symbol.isSignature();
785 if (!isInSymtab(Layout, Symbol, Used || WeakrefUsed || isSignature,
786 Renames.count(&Symbol)))
790 MSD.Symbol = cast<MCSymbolELF>(&Symbol);
792 bool Local = Symbol.getBinding() == ELF::STB_LOCAL;
793 if (Symbol.isAbsolute()) {
794 MSD.SectionIndex = ELF::SHN_ABS;
795 } else if (Symbol.isCommon()) {
797 MSD.SectionIndex = ELF::SHN_COMMON;
798 } else if (Symbol.isUndefined()) {
799 if (isSignature && !Used) {
800 MSD.SectionIndex = RevGroupMap.lookup(&Symbol);
801 if (MSD.SectionIndex >= ELF::SHN_LORESERVE)
802 HasLargeSectionIndex = true;
804 MSD.SectionIndex = ELF::SHN_UNDEF;
807 const MCSectionELF &Section =
808 static_cast<const MCSectionELF &>(Symbol.getSection());
809 MSD.SectionIndex = SectionIndexMap.lookup(&Section);
810 assert(MSD.SectionIndex && "Invalid section index!");
811 if (MSD.SectionIndex >= ELF::SHN_LORESERVE)
812 HasLargeSectionIndex = true;
815 // The @@@ in symbol version is replaced with @ in undefined symbols and @@
818 // FIXME: All name handling should be done before we get to the writer,
819 // including dealing with GNU-style version suffixes. Fixing this isn't
822 // We thus have to be careful to not perform the symbol version replacement
825 // The ELF format is used on Windows by the MCJIT engine. Thus, on
826 // Windows, the ELFObjectWriter can encounter symbols mangled using the MS
827 // Visual Studio C++ name mangling scheme. Symbols mangled using the MSVC
828 // C++ name mangling can legally have "@@@" as a sub-string. In that case,
829 // the EFLObjectWriter should not interpret the "@@@" sub-string as
830 // specifying GNU-style symbol versioning. The ELFObjectWriter therefore
831 // checks for the MSVC C++ name mangling prefix which is either "?", "@?",
832 // "__imp_?" or "__imp_@?".
834 // It would have been interesting to perform the MS mangling prefix check
835 // only when the target triple is of the form *-pc-windows-elf. But, it
836 // seems that this information is not easily accessible from the
838 StringRef Name = Symbol.getName();
839 if (!Name.startswith("?") && !Name.startswith("@?") &&
840 !Name.startswith("__imp_?") && !Name.startswith("__imp_@?")) {
841 // This symbol isn't following the MSVC C++ name mangling convention. We
842 // can thus safely interpret the @@@ in symbol names as specifying symbol
845 size_t Pos = Name.find("@@@");
846 if (Pos != StringRef::npos) {
847 Buf += Name.substr(0, Pos);
848 unsigned Skip = MSD.SectionIndex == ELF::SHN_UNDEF ? 2 : 1;
849 Buf += Name.substr(Pos + Skip);
854 // Sections have their own string table
855 if (Symbol.getType() != ELF::STT_SECTION)
856 MSD.Name = StrTabBuilder.add(Name);
859 LocalSymbolData.push_back(MSD);
861 ExternalSymbolData.push_back(MSD);
864 if (HasLargeSectionIndex) {
865 MCSectionELF *SymtabShndxSection =
866 Ctx.getELFSection(".symtab_shndxr", ELF::SHT_SYMTAB_SHNDX, 0, 4, "");
867 SymtabShndxSectionIndex = addToSectionTable(SymtabShndxSection);
868 SymtabShndxSection->setAlignment(4);
871 ArrayRef<std::string> FileNames = Asm.getFileNames();
872 for (const std::string &Name : FileNames)
873 StrTabBuilder.add(Name);
875 StrTabBuilder.finalize(StringTableBuilder::ELF);
877 for (const std::string &Name : FileNames)
878 Writer.writeSymbol(StrTabBuilder.getOffset(Name),
879 ELF::STT_FILE | ELF::STB_LOCAL, 0, 0, ELF::STV_DEFAULT,
882 // Symbols are required to be in lexicographic order.
883 array_pod_sort(LocalSymbolData.begin(), LocalSymbolData.end());
884 array_pod_sort(ExternalSymbolData.begin(), ExternalSymbolData.end());
886 // Set the symbol indices. Local symbols must come before all other
887 // symbols with non-local bindings.
888 unsigned Index = FileNames.size() + 1;
890 for (ELFSymbolData &MSD : LocalSymbolData) {
891 unsigned StringIndex = MSD.Symbol->getType() == ELF::STT_SECTION
893 : StrTabBuilder.getOffset(MSD.Name);
894 MSD.Symbol->setIndex(Index++);
895 writeSymbol(Writer, StringIndex, MSD, Layout);
898 // Write the symbol table entries.
899 LastLocalSymbolIndex = Index;
901 for (ELFSymbolData &MSD : ExternalSymbolData) {
902 unsigned StringIndex = StrTabBuilder.getOffset(MSD.Name);
903 MSD.Symbol->setIndex(Index++);
904 writeSymbol(Writer, StringIndex, MSD, Layout);
905 assert(MSD.Symbol->getBinding() != ELF::STB_LOCAL);
908 uint64_t SecEnd = OS.tell();
909 SectionOffsets[SymtabSection] = std::make_pair(SecStart, SecEnd);
911 ArrayRef<uint32_t> ShndxIndexes = Writer.getShndxIndexes();
912 if (ShndxIndexes.empty()) {
913 assert(SymtabShndxSectionIndex == 0);
916 assert(SymtabShndxSectionIndex != 0);
918 SecStart = OS.tell();
919 const MCSectionELF *SymtabShndxSection =
920 SectionTable[SymtabShndxSectionIndex - 1];
921 for (uint32_t Index : ShndxIndexes)
924 SectionOffsets[SymtabShndxSection] = std::make_pair(SecStart, SecEnd);
928 ELFObjectWriter::createRelocationSection(MCContext &Ctx,
929 const MCSectionELF &Sec) {
930 if (Relocations[&Sec].empty())
933 const StringRef SectionName = Sec.getSectionName();
934 std::string RelaSectionName = hasRelocationAddend() ? ".rela" : ".rel";
935 RelaSectionName += SectionName;
938 if (hasRelocationAddend())
939 EntrySize = is64Bit() ? sizeof(ELF::Elf64_Rela) : sizeof(ELF::Elf32_Rela);
941 EntrySize = is64Bit() ? sizeof(ELF::Elf64_Rel) : sizeof(ELF::Elf32_Rel);
944 if (Sec.getFlags() & ELF::SHF_GROUP)
945 Flags = ELF::SHF_GROUP;
947 MCSectionELF *RelaSection = Ctx.createELFRelSection(
948 RelaSectionName, hasRelocationAddend() ? ELF::SHT_RELA : ELF::SHT_REL,
949 Flags, EntrySize, Sec.getGroup(), &Sec);
950 RelaSection->setAlignment(is64Bit() ? 8 : 4);
954 static SmallVector<char, 128>
955 getUncompressedData(const MCAsmLayout &Layout,
956 const MCSection::FragmentListType &Fragments) {
957 SmallVector<char, 128> UncompressedData;
958 for (const MCFragment &F : Fragments) {
959 const SmallVectorImpl<char> *Contents;
960 switch (F.getKind()) {
961 case MCFragment::FT_Data:
962 Contents = &cast<MCDataFragment>(F).getContents();
964 case MCFragment::FT_Dwarf:
965 Contents = &cast<MCDwarfLineAddrFragment>(F).getContents();
967 case MCFragment::FT_DwarfFrame:
968 Contents = &cast<MCDwarfCallFrameFragment>(F).getContents();
972 "Not expecting any other fragment types in a debug_* section");
974 UncompressedData.append(Contents->begin(), Contents->end());
976 return UncompressedData;
979 // Include the debug info compression header:
980 // "ZLIB" followed by 8 bytes representing the uncompressed size of the section,
981 // useful for consumers to preallocate a buffer to decompress into.
983 prependCompressionHeader(uint64_t Size,
984 SmallVectorImpl<char> &CompressedContents) {
985 const StringRef Magic = "ZLIB";
986 if (Size <= Magic.size() + sizeof(Size) + CompressedContents.size())
988 if (sys::IsLittleEndianHost)
989 sys::swapByteOrder(Size);
990 CompressedContents.insert(CompressedContents.begin(),
991 Magic.size() + sizeof(Size), 0);
992 std::copy(Magic.begin(), Magic.end(), CompressedContents.begin());
993 std::copy(reinterpret_cast<char *>(&Size),
994 reinterpret_cast<char *>(&Size + 1),
995 CompressedContents.begin() + Magic.size());
999 void ELFObjectWriter::writeSectionData(const MCAssembler &Asm, MCSection &Sec,
1000 const MCAsmLayout &Layout) {
1001 MCSectionELF &Section = static_cast<MCSectionELF &>(Sec);
1002 StringRef SectionName = Section.getSectionName();
1004 // Compressing debug_frame requires handling alignment fragments which is
1005 // more work (possibly generalizing MCAssembler.cpp:writeFragment to allow
1006 // for writing to arbitrary buffers) for little benefit.
1007 if (!Asm.getContext().getAsmInfo()->compressDebugSections() ||
1008 !SectionName.startswith(".debug_") || SectionName == ".debug_frame") {
1009 Asm.writeSectionData(&Section, Layout);
1013 // Gather the uncompressed data from all the fragments.
1014 const MCSection::FragmentListType &Fragments = Section.getFragmentList();
1015 SmallVector<char, 128> UncompressedData =
1016 getUncompressedData(Layout, Fragments);
1018 SmallVector<char, 128> CompressedContents;
1019 zlib::Status Success = zlib::compress(
1020 StringRef(UncompressedData.data(), UncompressedData.size()),
1021 CompressedContents);
1022 if (Success != zlib::StatusOK) {
1023 Asm.writeSectionData(&Section, Layout);
1027 if (!prependCompressionHeader(UncompressedData.size(), CompressedContents)) {
1028 Asm.writeSectionData(&Section, Layout);
1031 Asm.getContext().renameELFSection(&Section,
1032 (".z" + SectionName.drop_front(1)).str());
1033 OS << CompressedContents;
1036 void ELFObjectWriter::WriteSecHdrEntry(uint32_t Name, uint32_t Type,
1037 uint64_t Flags, uint64_t Address,
1038 uint64_t Offset, uint64_t Size,
1039 uint32_t Link, uint32_t Info,
1041 uint64_t EntrySize) {
1042 write32(Name); // sh_name: index into string table
1043 write32(Type); // sh_type
1044 WriteWord(Flags); // sh_flags
1045 WriteWord(Address); // sh_addr
1046 WriteWord(Offset); // sh_offset
1047 WriteWord(Size); // sh_size
1048 write32(Link); // sh_link
1049 write32(Info); // sh_info
1050 WriteWord(Alignment); // sh_addralign
1051 WriteWord(EntrySize); // sh_entsize
1054 void ELFObjectWriter::writeRelocations(const MCAssembler &Asm,
1055 const MCSectionELF &Sec) {
1056 std::vector<ELFRelocationEntry> &Relocs = Relocations[&Sec];
1058 // Sort the relocation entries. Most targets just sort by Offset, but some
1059 // (e.g., MIPS) have additional constraints.
1060 TargetObjectWriter->sortRelocs(Asm, Relocs);
1062 for (unsigned i = 0, e = Relocs.size(); i != e; ++i) {
1063 const ELFRelocationEntry &Entry = Relocs[e - i - 1];
1064 unsigned Index = Entry.Symbol ? Entry.Symbol->getIndex() : 0;
1067 write(Entry.Offset);
1068 if (TargetObjectWriter->isN64()) {
1069 write(uint32_t(Index));
1071 write(TargetObjectWriter->getRSsym(Entry.Type));
1072 write(TargetObjectWriter->getRType3(Entry.Type));
1073 write(TargetObjectWriter->getRType2(Entry.Type));
1074 write(TargetObjectWriter->getRType(Entry.Type));
1076 struct ELF::Elf64_Rela ERE64;
1077 ERE64.setSymbolAndType(Index, Entry.Type);
1078 write(ERE64.r_info);
1080 if (hasRelocationAddend())
1081 write(Entry.Addend);
1083 write(uint32_t(Entry.Offset));
1085 struct ELF::Elf32_Rela ERE32;
1086 ERE32.setSymbolAndType(Index, Entry.Type);
1087 write(ERE32.r_info);
1089 if (hasRelocationAddend())
1090 write(uint32_t(Entry.Addend));
1095 const MCSectionELF *ELFObjectWriter::createStringTable(MCContext &Ctx) {
1096 const MCSectionELF *StrtabSection = SectionTable[StringTableIndex - 1];
1097 OS << StrTabBuilder.data();
1098 return StrtabSection;
1101 void ELFObjectWriter::writeSection(const SectionIndexMapTy &SectionIndexMap,
1102 uint32_t GroupSymbolIndex, uint64_t Offset,
1103 uint64_t Size, const MCSectionELF &Section) {
1104 uint64_t sh_link = 0;
1105 uint64_t sh_info = 0;
1107 switch(Section.getType()) {
1112 case ELF::SHT_DYNAMIC:
1113 llvm_unreachable("SHT_DYNAMIC in a relocatable object");
1116 case ELF::SHT_RELA: {
1117 sh_link = SymbolTableIndex;
1118 assert(sh_link && ".symtab not found");
1119 const MCSectionELF *InfoSection = Section.getAssociatedSection();
1120 sh_info = SectionIndexMap.lookup(InfoSection);
1124 case ELF::SHT_SYMTAB:
1125 case ELF::SHT_DYNSYM:
1126 sh_link = StringTableIndex;
1127 sh_info = LastLocalSymbolIndex;
1130 case ELF::SHT_SYMTAB_SHNDX:
1131 sh_link = SymbolTableIndex;
1134 case ELF::SHT_GROUP:
1135 sh_link = SymbolTableIndex;
1136 sh_info = GroupSymbolIndex;
1140 if (TargetObjectWriter->getEMachine() == ELF::EM_ARM &&
1141 Section.getType() == ELF::SHT_ARM_EXIDX)
1142 sh_link = SectionIndexMap.lookup(Section.getAssociatedSection());
1144 WriteSecHdrEntry(StrTabBuilder.getOffset(Section.getSectionName()),
1145 Section.getType(), Section.getFlags(), 0, Offset, Size,
1146 sh_link, sh_info, Section.getAlignment(),
1147 Section.getEntrySize());
1150 void ELFObjectWriter::writeSectionHeader(
1151 const MCAsmLayout &Layout, const SectionIndexMapTy &SectionIndexMap,
1152 const SectionOffsetsTy &SectionOffsets) {
1153 const unsigned NumSections = SectionTable.size();
1155 // Null section first.
1156 uint64_t FirstSectionSize =
1157 (NumSections + 1) >= ELF::SHN_LORESERVE ? NumSections + 1 : 0;
1158 WriteSecHdrEntry(0, 0, 0, 0, 0, FirstSectionSize, 0, 0, 0, 0);
1160 for (const MCSectionELF *Section : SectionTable) {
1161 uint32_t GroupSymbolIndex;
1162 unsigned Type = Section->getType();
1163 if (Type != ELF::SHT_GROUP)
1164 GroupSymbolIndex = 0;
1166 GroupSymbolIndex = Section->getGroup()->getIndex();
1168 const std::pair<uint64_t, uint64_t> &Offsets =
1169 SectionOffsets.find(Section)->second;
1171 if (Type == ELF::SHT_NOBITS)
1172 Size = Layout.getSectionAddressSize(Section);
1174 Size = Offsets.second - Offsets.first;
1176 writeSection(SectionIndexMap, GroupSymbolIndex, Offsets.first, Size,
1181 void ELFObjectWriter::writeObject(MCAssembler &Asm,
1182 const MCAsmLayout &Layout) {
1183 MCContext &Ctx = Asm.getContext();
1184 MCSectionELF *StrtabSection =
1185 Ctx.getELFSection(".strtab", ELF::SHT_STRTAB, 0);
1186 StringTableIndex = addToSectionTable(StrtabSection);
1188 RevGroupMapTy RevGroupMap;
1189 SectionIndexMapTy SectionIndexMap;
1191 std::map<const MCSymbol *, std::vector<const MCSectionELF *>> GroupMembers;
1193 // Write out the ELF header ...
1196 // ... then the sections ...
1197 SectionOffsetsTy SectionOffsets;
1198 std::vector<MCSectionELF *> Groups;
1199 std::vector<MCSectionELF *> Relocations;
1200 for (MCSection &Sec : Asm) {
1201 MCSectionELF &Section = static_cast<MCSectionELF &>(Sec);
1203 align(Section.getAlignment());
1205 // Remember the offset into the file for this section.
1206 uint64_t SecStart = OS.tell();
1208 const MCSymbolELF *SignatureSymbol = Section.getGroup();
1209 writeSectionData(Asm, Section, Layout);
1211 uint64_t SecEnd = OS.tell();
1212 SectionOffsets[&Section] = std::make_pair(SecStart, SecEnd);
1214 MCSectionELF *RelSection = createRelocationSection(Ctx, Section);
1216 if (SignatureSymbol) {
1217 Asm.registerSymbol(*SignatureSymbol);
1218 unsigned &GroupIdx = RevGroupMap[SignatureSymbol];
1220 MCSectionELF *Group = Ctx.createELFGroupSection(SignatureSymbol);
1221 GroupIdx = addToSectionTable(Group);
1222 Group->setAlignment(4);
1223 Groups.push_back(Group);
1225 std::vector<const MCSectionELF *> &Members =
1226 GroupMembers[SignatureSymbol];
1227 Members.push_back(&Section);
1229 Members.push_back(RelSection);
1232 SectionIndexMap[&Section] = addToSectionTable(&Section);
1234 SectionIndexMap[RelSection] = addToSectionTable(RelSection);
1235 Relocations.push_back(RelSection);
1239 for (MCSectionELF *Group : Groups) {
1240 align(Group->getAlignment());
1242 // Remember the offset into the file for this section.
1243 uint64_t SecStart = OS.tell();
1245 const MCSymbol *SignatureSymbol = Group->getGroup();
1246 assert(SignatureSymbol);
1247 write(uint32_t(ELF::GRP_COMDAT));
1248 for (const MCSectionELF *Member : GroupMembers[SignatureSymbol]) {
1249 uint32_t SecIndex = SectionIndexMap.lookup(Member);
1253 uint64_t SecEnd = OS.tell();
1254 SectionOffsets[Group] = std::make_pair(SecStart, SecEnd);
1257 // Compute symbol table information.
1258 computeSymbolTable(Asm, Layout, SectionIndexMap, RevGroupMap, SectionOffsets);
1260 for (MCSectionELF *RelSection : Relocations) {
1261 align(RelSection->getAlignment());
1263 // Remember the offset into the file for this section.
1264 uint64_t SecStart = OS.tell();
1266 writeRelocations(Asm, *RelSection->getAssociatedSection());
1268 uint64_t SecEnd = OS.tell();
1269 SectionOffsets[RelSection] = std::make_pair(SecStart, SecEnd);
1273 uint64_t SecStart = OS.tell();
1274 const MCSectionELF *Sec = createStringTable(Ctx);
1275 uint64_t SecEnd = OS.tell();
1276 SectionOffsets[Sec] = std::make_pair(SecStart, SecEnd);
1279 uint64_t NaturalAlignment = is64Bit() ? 8 : 4;
1280 align(NaturalAlignment);
1282 const unsigned SectionHeaderOffset = OS.tell();
1284 // ... then the section header table ...
1285 writeSectionHeader(Layout, SectionIndexMap, SectionOffsets);
1287 uint16_t NumSections = (SectionTable.size() + 1 >= ELF::SHN_LORESERVE)
1288 ? (uint16_t)ELF::SHN_UNDEF
1289 : SectionTable.size() + 1;
1290 if (sys::IsLittleEndianHost != IsLittleEndian)
1291 sys::swapByteOrder(NumSections);
1292 unsigned NumSectionsOffset;
1295 uint64_t Val = SectionHeaderOffset;
1296 if (sys::IsLittleEndianHost != IsLittleEndian)
1297 sys::swapByteOrder(Val);
1298 OS.pwrite(reinterpret_cast<char *>(&Val), sizeof(Val),
1299 offsetof(ELF::Elf64_Ehdr, e_shoff));
1300 NumSectionsOffset = offsetof(ELF::Elf64_Ehdr, e_shnum);
1302 uint32_t Val = SectionHeaderOffset;
1303 if (sys::IsLittleEndianHost != IsLittleEndian)
1304 sys::swapByteOrder(Val);
1305 OS.pwrite(reinterpret_cast<char *>(&Val), sizeof(Val),
1306 offsetof(ELF::Elf32_Ehdr, e_shoff));
1307 NumSectionsOffset = offsetof(ELF::Elf32_Ehdr, e_shnum);
1309 OS.pwrite(reinterpret_cast<char *>(&NumSections), sizeof(NumSections),
1313 bool ELFObjectWriter::isSymbolRefDifferenceFullyResolvedImpl(
1314 const MCAssembler &Asm, const MCSymbol &SA, const MCFragment &FB,
1315 bool InSet, bool IsPCRel) const {
1316 const auto &SymA = cast<MCSymbolELF>(SA);
1322 return MCObjectWriter::isSymbolRefDifferenceFullyResolvedImpl(Asm, SymA, FB,
1326 bool ELFObjectWriter::isWeak(const MCSymbol &S) const {
1327 const auto &Sym = cast<MCSymbolELF>(S);
1331 // It is invalid to replace a reference to a global in a comdat
1332 // with a reference to a local since out of comdat references
1333 // to a local are forbidden.
1334 // We could try to return false for more cases, like the reference
1335 // being in the same comdat or Sym being an alias to another global,
1336 // but it is not clear if it is worth the effort.
1337 if (Sym.getBinding() != ELF::STB_GLOBAL)
1340 if (!Sym.isInSection())
1343 const auto &Sec = cast<MCSectionELF>(Sym.getSection());
1344 return Sec.getGroup();
1347 MCObjectWriter *llvm::createELFObjectWriter(MCELFObjectTargetWriter *MOTW,
1348 raw_pwrite_stream &OS,
1349 bool IsLittleEndian) {
1350 return new ELFObjectWriter(MOTW, OS, IsLittleEndian);