1 //===- InputFiles.cpp -----------------------------------------------------===//
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 #include "InputFiles.h"
12 #include "InputSection.h"
13 #include "LinkerScript.h"
15 #include "SymbolTable.h"
17 #include "SyntheticSections.h"
18 #include "llvm/ADT/STLExtras.h"
19 #include "llvm/CodeGen/Analysis.h"
20 #include "llvm/DebugInfo/DWARF/DWARFContext.h"
21 #include "llvm/IR/LLVMContext.h"
22 #include "llvm/IR/Module.h"
23 #include "llvm/LTO/LTO.h"
24 #include "llvm/MC/StringTableBuilder.h"
25 #include "llvm/Object/ELFObjectFile.h"
26 #include "llvm/Support/Path.h"
27 #include "llvm/Support/TarWriter.h"
28 #include "llvm/Support/raw_ostream.h"
31 using namespace llvm::ELF;
32 using namespace llvm::object;
33 using namespace llvm::sys::fs;
36 using namespace lld::elf;
40 InputFile::InputFile(Kind K, MemoryBufferRef M) : MB(M), FileKind(K) {}
43 // In ELF object file all section addresses are zero. If we have multiple
44 // .text sections (when using -ffunction-section or comdat group) then
45 // LLVM DWARF parser will not be able to parse .debug_line correctly, unless
46 // we assign each section some unique address. This callback method assigns
47 // each section an address equal to its offset in ELF object file.
48 class ObjectInfo : public LoadedObjectInfo {
50 uint64_t getSectionLoadAddress(const object::SectionRef &Sec) const override {
51 return static_cast<const ELFSectionRef &>(Sec).getOffset();
53 std::unique_ptr<LoadedObjectInfo> clone() const override {
54 return std::unique_ptr<LoadedObjectInfo>();
59 Optional<MemoryBufferRef> elf::readFile(StringRef Path) {
61 auto MBOrErr = MemoryBuffer::getFile(Path);
62 if (auto EC = MBOrErr.getError()) {
63 error("cannot open " + Path + ": " + EC.message());
67 std::unique_ptr<MemoryBuffer> &MB = *MBOrErr;
68 MemoryBufferRef MBRef = MB->getMemBufferRef();
69 make<std::unique_ptr<MemoryBuffer>>(std::move(MB)); // take MB ownership
72 Tar->append(relativeToRoot(Path), MBRef.getBuffer());
76 template <class ELFT> void elf::ObjectFile<ELFT>::initializeDwarfLine() {
77 std::unique_ptr<object::ObjectFile> Obj =
78 check(object::ObjectFile::createObjectFile(this->MB), toString(this));
81 DWARFContextInMemory Dwarf(*Obj, &ObjInfo);
82 DwarfLine.reset(new DWARFDebugLine(&Dwarf.getLineSection().Relocs));
83 DataExtractor LineData(Dwarf.getLineSection().Data, Config->IsLE,
86 // The second parameter is offset in .debug_line section
87 // for compilation unit (CU) of interest. We have only one
88 // CU (object file), so offset is always 0.
89 DwarfLine->getOrParseLineTable(LineData, 0);
92 // Returns source line information for a given offset
93 // using DWARF debug info.
95 Optional<DILineInfo> elf::ObjectFile<ELFT>::getDILineInfo(InputSectionBase *S,
98 initializeDwarfLine();
100 // The offset to CU is 0.
101 const DWARFDebugLine::LineTable *Tbl = DwarfLine->getLineTable(0);
105 // Use fake address calcuated by adding section file offset and offset in
106 // section. See comments for ObjectInfo class.
108 Tbl->getFileLineInfoForAddress(
109 S->getOffsetInFile() + Offset, nullptr,
110 DILineInfoSpecifier::FileLineInfoKind::AbsoluteFilePath, Info);
116 // Returns source line information for a given offset
117 // using DWARF debug info.
118 template <class ELFT>
119 std::string elf::ObjectFile<ELFT>::getLineInfo(InputSectionBase *S,
121 if (Optional<DILineInfo> Info = getDILineInfo(S, Offset))
122 return Info->FileName + ":" + std::to_string(Info->Line);
126 // Returns "<internal>", "foo.a(bar.o)" or "baz.o".
127 std::string lld::toString(const InputFile *F) {
131 if (F->ToStringCache.empty()) {
132 if (F->ArchiveName.empty())
133 F->ToStringCache = F->getName();
135 F->ToStringCache = (F->ArchiveName + "(" + F->getName() + ")").str();
137 return F->ToStringCache;
140 template <class ELFT>
141 ELFFileBase<ELFT>::ELFFileBase(Kind K, MemoryBufferRef MB) : InputFile(K, MB) {
142 if (ELFT::TargetEndianness == support::little)
143 EKind = ELFT::Is64Bits ? ELF64LEKind : ELF32LEKind;
145 EKind = ELFT::Is64Bits ? ELF64BEKind : ELF32BEKind;
147 EMachine = getObj().getHeader()->e_machine;
148 OSABI = getObj().getHeader()->e_ident[llvm::ELF::EI_OSABI];
151 template <class ELFT>
152 typename ELFT::SymRange ELFFileBase<ELFT>::getGlobalSymbols() {
153 return makeArrayRef(Symbols.begin() + FirstNonLocal, Symbols.end());
156 template <class ELFT>
157 uint32_t ELFFileBase<ELFT>::getSectionIndex(const Elf_Sym &Sym) const {
158 return check(getObj().getSectionIndex(&Sym, Symbols, SymtabSHNDX),
162 template <class ELFT>
163 void ELFFileBase<ELFT>::initSymtab(ArrayRef<Elf_Shdr> Sections,
164 const Elf_Shdr *Symtab) {
165 FirstNonLocal = Symtab->sh_info;
166 Symbols = check(getObj().symbols(Symtab), toString(this));
167 if (FirstNonLocal == 0 || FirstNonLocal > Symbols.size())
168 fatal(toString(this) + ": invalid sh_info in symbol table");
170 StringTable = check(getObj().getStringTableForSymtab(*Symtab, Sections),
174 template <class ELFT>
175 elf::ObjectFile<ELFT>::ObjectFile(MemoryBufferRef M, StringRef ArchiveName)
176 : ELFFileBase<ELFT>(Base::ObjectKind, M) {
177 this->ArchiveName = ArchiveName;
180 template <class ELFT>
181 ArrayRef<SymbolBody *> elf::ObjectFile<ELFT>::getLocalSymbols() {
182 if (this->SymbolBodies.empty())
183 return this->SymbolBodies;
184 return makeArrayRef(this->SymbolBodies).slice(1, this->FirstNonLocal - 1);
187 template <class ELFT>
188 ArrayRef<SymbolBody *> elf::ObjectFile<ELFT>::getSymbols() {
189 if (this->SymbolBodies.empty())
190 return this->SymbolBodies;
191 return makeArrayRef(this->SymbolBodies).slice(1);
194 template <class ELFT>
195 void elf::ObjectFile<ELFT>::parse(DenseSet<CachedHashStringRef> &ComdatGroups) {
196 // Read section and symbol tables.
197 initializeSections(ComdatGroups);
201 // Sections with SHT_GROUP and comdat bits define comdat section groups.
202 // They are identified and deduplicated by group name. This function
203 // returns a group name.
204 template <class ELFT>
206 elf::ObjectFile<ELFT>::getShtGroupSignature(ArrayRef<Elf_Shdr> Sections,
207 const Elf_Shdr &Sec) {
208 if (this->Symbols.empty())
211 check(object::getSection<ELFT>(Sections, Sec.sh_link), toString(this)));
212 const Elf_Sym *Sym = check(
213 object::getSymbol<ELFT>(this->Symbols, Sec.sh_info), toString(this));
214 return check(Sym->getName(this->StringTable), toString(this));
217 template <class ELFT>
218 ArrayRef<typename elf::ObjectFile<ELFT>::Elf_Word>
219 elf::ObjectFile<ELFT>::getShtGroupEntries(const Elf_Shdr &Sec) {
220 const ELFFile<ELFT> &Obj = this->getObj();
221 ArrayRef<Elf_Word> Entries = check(
222 Obj.template getSectionContentsAsArray<Elf_Word>(&Sec), toString(this));
223 if (Entries.empty() || Entries[0] != GRP_COMDAT)
224 fatal(toString(this) + ": unsupported SHT_GROUP format");
225 return Entries.slice(1);
228 template <class ELFT>
229 bool elf::ObjectFile<ELFT>::shouldMerge(const Elf_Shdr &Sec) {
230 // We don't merge sections if -O0 (default is -O1). This makes sometimes
231 // the linker significantly faster, although the output will be bigger.
232 if (Config->Optimize == 0)
235 // Do not merge sections if generating a relocatable object. It makes
236 // the code simpler because we do not need to update relocation addends
237 // to reflect changes introduced by merging. Instead of that we write
238 // such "merge" sections into separate OutputSections and keep SHF_MERGE
239 // / SHF_STRINGS flags and sh_entsize value to be able to perform merging
240 // later during a final linking.
241 if (Config->Relocatable)
244 // A mergeable section with size 0 is useless because they don't have
245 // any data to merge. A mergeable string section with size 0 can be
246 // argued as invalid because it doesn't end with a null character.
247 // We'll avoid a mess by handling them as if they were non-mergeable.
248 if (Sec.sh_size == 0)
251 // Check for sh_entsize. The ELF spec is not clear about the zero
252 // sh_entsize. It says that "the member [sh_entsize] contains 0 if
253 // the section does not hold a table of fixed-size entries". We know
254 // that Rust 1.13 produces a string mergeable section with a zero
255 // sh_entsize. Here we just accept it rather than being picky about it.
256 uint64_t EntSize = Sec.sh_entsize;
259 if (Sec.sh_size % EntSize)
260 fatal(toString(this) +
261 ": SHF_MERGE section size must be a multiple of sh_entsize");
263 uint64_t Flags = Sec.sh_flags;
264 if (!(Flags & SHF_MERGE))
266 if (Flags & SHF_WRITE)
267 fatal(toString(this) + ": writable SHF_MERGE section is not supported");
269 // Don't try to merge if the alignment is larger than the sh_entsize and this
270 // is not SHF_STRINGS.
272 // Since this is not a SHF_STRINGS, we would need to pad after every entity.
273 // It would be equivalent for the producer of the .o to just set a larger
275 if (Flags & SHF_STRINGS)
278 return Sec.sh_addralign <= EntSize;
281 template <class ELFT>
282 void elf::ObjectFile<ELFT>::initializeSections(
283 DenseSet<CachedHashStringRef> &ComdatGroups) {
284 const ELFFile<ELFT> &Obj = this->getObj();
286 ArrayRef<Elf_Shdr> ObjSections =
287 check(this->getObj().sections(), toString(this));
288 uint64_t Size = ObjSections.size();
289 this->Sections.resize(Size);
291 StringRef SectionStringTable =
292 check(Obj.getSectionStringTable(ObjSections), toString(this));
294 for (size_t I = 0, E = ObjSections.size(); I < E; I++) {
295 if (this->Sections[I] == &InputSection::Discarded)
297 const Elf_Shdr &Sec = ObjSections[I];
299 // SHF_EXCLUDE'ed sections are discarded by the linker. However,
300 // if -r is given, we'll let the final link discard such sections.
301 // This is compatible with GNU.
302 if ((Sec.sh_flags & SHF_EXCLUDE) && !Config->Relocatable) {
303 this->Sections[I] = &InputSection::Discarded;
307 switch (Sec.sh_type) {
309 // We discard comdat sections usually. When -r we should not do that. We
310 // still do deduplication in this case to simplify implementation, because
311 // otherwise merging group sections together would requre additional
312 // regeneration of its contents.
313 bool New = ComdatGroups
314 .insert(CachedHashStringRef(
315 getShtGroupSignature(ObjSections, Sec)))
317 if (New && Config->Relocatable)
318 this->Sections[I] = createInputSection(Sec, SectionStringTable);
320 this->Sections[I] = &InputSection::Discarded;
324 for (uint32_t SecIndex : getShtGroupEntries(Sec)) {
325 if (SecIndex >= Size)
326 fatal(toString(this) +
327 ": invalid section index in group: " + Twine(SecIndex));
328 this->Sections[SecIndex] = &InputSection::Discarded;
333 this->initSymtab(ObjSections, &Sec);
335 case SHT_SYMTAB_SHNDX:
337 check(Obj.getSHNDXTable(Sec, ObjSections), toString(this));
343 this->Sections[I] = createInputSection(Sec, SectionStringTable);
346 // .ARM.exidx sections have a reverse dependency on the InputSection they
347 // have a SHF_LINK_ORDER dependency, this is identified by the sh_link.
348 if (Sec.sh_flags & SHF_LINK_ORDER) {
349 if (Sec.sh_link >= this->Sections.size())
350 fatal(toString(this) + ": invalid sh_link index: " +
352 this->Sections[Sec.sh_link]->DependentSections.push_back(
358 template <class ELFT>
359 InputSectionBase *elf::ObjectFile<ELFT>::getRelocTarget(const Elf_Shdr &Sec) {
360 uint32_t Idx = Sec.sh_info;
361 if (Idx >= this->Sections.size())
362 fatal(toString(this) + ": invalid relocated section index: " + Twine(Idx));
363 InputSectionBase *Target = this->Sections[Idx];
365 // Strictly speaking, a relocation section must be included in the
366 // group of the section it relocates. However, LLVM 3.3 and earlier
367 // would fail to do so, so we gracefully handle that case.
368 if (Target == &InputSection::Discarded)
372 fatal(toString(this) + ": unsupported relocation reference");
376 // Create a regular InputSection class that has the same contents
377 // as a given section.
378 InputSectionBase *toRegularSection(MergeInputSection *Sec) {
379 auto *Ret = make<InputSection>(Sec->Flags, Sec->Type, Sec->Alignment,
380 Sec->Data, Sec->Name);
381 Ret->File = Sec->File;
385 template <class ELFT>
387 elf::ObjectFile<ELFT>::createInputSection(const Elf_Shdr &Sec,
388 StringRef SectionStringTable) {
389 StringRef Name = check(
390 this->getObj().getSectionName(&Sec, SectionStringTable), toString(this));
392 switch (Sec.sh_type) {
393 case SHT_ARM_ATTRIBUTES:
394 // FIXME: ARM meta-data section. Retain the first attribute section
395 // we see. The eglibc ARM dynamic loaders require the presence of an
396 // attribute section for dlopen to work.
397 // In a full implementation we would merge all attribute sections.
398 if (InX::ARMAttributes == nullptr) {
399 InX::ARMAttributes = make<InputSection>(this, &Sec, Name);
400 return InX::ARMAttributes;
402 return &InputSection::Discarded;
405 // Find the relocation target section and associate this
406 // section with it. Target can be discarded, for example
407 // if it is a duplicated member of SHT_GROUP section, we
408 // do not create or proccess relocatable sections then.
409 InputSectionBase *Target = getRelocTarget(Sec);
413 // This section contains relocation information.
414 // If -r is given, we do not interpret or apply relocation
415 // but just copy relocation sections to output.
416 if (Config->Relocatable)
417 return make<InputSection>(this, &Sec, Name);
419 if (Target->FirstRelocation)
420 fatal(toString(this) +
421 ": multiple relocation sections to one section are not supported");
423 // Mergeable sections with relocations are tricky because relocations
424 // need to be taken into account when comparing section contents for
425 // merging. It's not worth supporting such mergeable sections because
426 // they are rare and it'd complicates the internal design (we usually
427 // have to determine if two sections are mergeable early in the link
428 // process much before applying relocations). We simply handle mergeable
429 // sections with relocations as non-mergeable.
430 if (auto *MS = dyn_cast<MergeInputSection>(Target)) {
431 Target = toRegularSection(MS);
432 this->Sections[Sec.sh_info] = Target;
435 size_t NumRelocations;
436 if (Sec.sh_type == SHT_RELA) {
437 ArrayRef<Elf_Rela> Rels =
438 check(this->getObj().relas(&Sec), toString(this));
439 Target->FirstRelocation = Rels.begin();
440 NumRelocations = Rels.size();
441 Target->AreRelocsRela = true;
443 ArrayRef<Elf_Rel> Rels = check(this->getObj().rels(&Sec), toString(this));
444 Target->FirstRelocation = Rels.begin();
445 NumRelocations = Rels.size();
446 Target->AreRelocsRela = false;
448 assert(isUInt<31>(NumRelocations));
449 Target->NumRelocations = NumRelocations;
451 // Relocation sections processed by the linker are usually removed
452 // from the output, so returning `nullptr` for the normal case.
453 // However, if -emit-relocs is given, we need to leave them in the output.
454 // (Some post link analysis tools need this information.)
455 if (Config->EmitRelocs) {
456 InputSection *RelocSec = make<InputSection>(this, &Sec, Name);
457 // We will not emit relocation section if target was discarded.
458 Target->DependentSections.push_back(RelocSec);
465 // The GNU linker uses .note.GNU-stack section as a marker indicating
466 // that the code in the object file does not expect that the stack is
467 // executable (in terms of NX bit). If all input files have the marker,
468 // the GNU linker adds a PT_GNU_STACK segment to tells the loader to
469 // make the stack non-executable. Most object files have this section as
472 // But making the stack non-executable is a norm today for security
473 // reasons. Failure to do so may result in a serious security issue.
474 // Therefore, we make LLD always add PT_GNU_STACK unless it is
475 // explicitly told to do otherwise (by -z execstack). Because the stack
476 // executable-ness is controlled solely by command line options,
477 // .note.GNU-stack sections are simply ignored.
478 if (Name == ".note.GNU-stack")
479 return &InputSection::Discarded;
481 // Split stacks is a feature to support a discontiguous stack. At least
482 // as of 2017, it seems that the feature is not being used widely.
483 // Only GNU gold supports that. We don't. For the details about that,
484 // see https://gcc.gnu.org/wiki/SplitStacks
485 if (Name == ".note.GNU-split-stack") {
486 error(toString(this) +
487 ": object file compiled with -fsplit-stack is not supported");
488 return &InputSection::Discarded;
491 if (Config->Strip != StripPolicy::None && Name.startswith(".debug"))
492 return &InputSection::Discarded;
494 // If -gdb-index is given, LLD creates .gdb_index section, and that
495 // section serves the same purpose as .debug_gnu_pub{names,types} sections.
496 // If that's the case, we want to eliminate .debug_gnu_pub{names,types}
497 // because they are redundant and can waste large amount of disk space
498 // (for example, they are about 400 MiB in total for a clang debug build.)
499 if (Config->GdbIndex &&
500 (Name == ".debug_gnu_pubnames" || Name == ".debug_gnu_pubtypes"))
501 return &InputSection::Discarded;
503 // The linkonce feature is a sort of proto-comdat. Some glibc i386 object
504 // files contain definitions of symbol "__x86.get_pc_thunk.bx" in linkonce
505 // sections. Drop those sections to avoid duplicate symbol errors.
506 // FIXME: This is glibc PR20543, we should remove this hack once that has been
507 // fixed for a while.
508 if (Name.startswith(".gnu.linkonce."))
509 return &InputSection::Discarded;
511 // The linker merges EH (exception handling) frames and creates a
512 // .eh_frame_hdr section for runtime. So we handle them with a special
513 // class. For relocatable outputs, they are just passed through.
514 if (Name == ".eh_frame" && !Config->Relocatable)
515 return make<EhInputSection>(this, &Sec, Name);
517 if (shouldMerge(Sec))
518 return make<MergeInputSection>(this, &Sec, Name);
519 return make<InputSection>(this, &Sec, Name);
522 template <class ELFT> void elf::ObjectFile<ELFT>::initializeSymbols() {
523 SymbolBodies.reserve(this->Symbols.size());
524 for (const Elf_Sym &Sym : this->Symbols)
525 SymbolBodies.push_back(createSymbolBody(&Sym));
528 template <class ELFT>
529 InputSectionBase *elf::ObjectFile<ELFT>::getSection(const Elf_Sym &Sym) const {
530 uint32_t Index = this->getSectionIndex(Sym);
531 if (Index >= this->Sections.size())
532 fatal(toString(this) + ": invalid section index: " + Twine(Index));
533 InputSectionBase *S = this->Sections[Index];
535 // We found that GNU assembler 2.17.50 [FreeBSD] 2007-07-03 could
536 // generate broken objects. STT_SECTION/STT_NOTYPE symbols can be
537 // associated with SHT_REL[A]/SHT_SYMTAB/SHT_STRTAB sections.
538 // In this case it is fine for section to be null here as we do not
539 // allocate sections of these types.
541 if (Index == 0 || Sym.getType() == STT_SECTION ||
542 Sym.getType() == STT_NOTYPE)
544 fatal(toString(this) + ": invalid section index: " + Twine(Index));
547 if (S == &InputSection::Discarded)
552 template <class ELFT>
553 SymbolBody *elf::ObjectFile<ELFT>::createSymbolBody(const Elf_Sym *Sym) {
554 int Binding = Sym->getBinding();
555 InputSectionBase *Sec = getSection(*Sym);
557 uint8_t StOther = Sym->st_other;
558 uint8_t Type = Sym->getType();
559 uint64_t Value = Sym->st_value;
560 uint64_t Size = Sym->st_size;
562 if (Binding == STB_LOCAL) {
563 if (Sym->getType() == STT_FILE)
564 SourceFile = check(Sym->getName(this->StringTable), toString(this));
566 if (this->StringTable.size() <= Sym->st_name)
567 fatal(toString(this) + ": invalid symbol name offset");
569 StringRefZ Name = this->StringTable.data() + Sym->st_name;
570 if (Sym->st_shndx == SHN_UNDEF)
571 return make<Undefined>(Name, /*IsLocal=*/true, StOther, Type, this);
573 return make<DefinedRegular>(Name, /*IsLocal=*/true, StOther, Type, Value,
577 StringRef Name = check(Sym->getName(this->StringTable), toString(this));
579 switch (Sym->st_shndx) {
581 return elf::Symtab<ELFT>::X
582 ->addUndefined(Name, /*IsLocal=*/false, Binding, StOther, Type,
583 /*CanOmitFromDynSym=*/false, this)
586 if (Value == 0 || Value >= UINT32_MAX)
587 fatal(toString(this) + ": common symbol '" + Name +
588 "' has invalid alignment: " + Twine(Value));
589 return elf::Symtab<ELFT>::X
590 ->addCommon(Name, Size, Value, Binding, StOther, Type, this)
596 fatal(toString(this) + ": unexpected binding: " + Twine(Binding));
600 if (Sec == &InputSection::Discarded)
601 return elf::Symtab<ELFT>::X
602 ->addUndefined(Name, /*IsLocal=*/false, Binding, StOther, Type,
603 /*CanOmitFromDynSym=*/false, this)
605 return elf::Symtab<ELFT>::X
606 ->addRegular(Name, StOther, Type, Value, Size, Binding, Sec, this)
611 ArchiveFile::ArchiveFile(std::unique_ptr<Archive> &&File)
612 : InputFile(ArchiveKind, File->getMemoryBufferRef()),
613 File(std::move(File)) {}
615 template <class ELFT> void ArchiveFile::parse() {
616 for (const Archive::Symbol &Sym : File->symbols())
617 Symtab<ELFT>::X->addLazyArchive(this, Sym);
620 // Returns a buffer pointing to a member file containing a given symbol.
621 std::pair<MemoryBufferRef, uint64_t>
622 ArchiveFile::getMember(const Archive::Symbol *Sym) {
624 check(Sym->getMember(), toString(this) +
625 ": could not get the member for symbol " +
628 if (!Seen.insert(C.getChildOffset()).second)
629 return {MemoryBufferRef(), 0};
631 MemoryBufferRef Ret =
632 check(C.getMemoryBufferRef(),
634 ": could not get the buffer for the member defining symbol " +
637 if (C.getParent()->isThin() && Tar)
638 Tar->append(relativeToRoot(check(C.getFullName(), toString(this))),
640 if (C.getParent()->isThin())
642 return {Ret, C.getChildOffset()};
645 template <class ELFT>
646 SharedFile<ELFT>::SharedFile(MemoryBufferRef M, StringRef DefaultSoName)
647 : ELFFileBase<ELFT>(Base::SharedKind, M), SoName(DefaultSoName),
648 AsNeeded(Config->AsNeeded) {}
650 template <class ELFT>
651 const typename ELFT::Shdr *
652 SharedFile<ELFT>::getSection(const Elf_Sym &Sym) const {
654 this->getObj().getSection(&Sym, this->Symbols, this->SymtabSHNDX),
658 // Partially parse the shared object file so that we can call
659 // getSoName on this object.
660 template <class ELFT> void SharedFile<ELFT>::parseSoName() {
661 const Elf_Shdr *DynamicSec = nullptr;
662 const ELFFile<ELFT> Obj = this->getObj();
663 ArrayRef<Elf_Shdr> Sections = check(Obj.sections(), toString(this));
665 // Search for .dynsym, .dynamic, .symtab, .gnu.version and .gnu.version_d.
666 for (const Elf_Shdr &Sec : Sections) {
667 switch (Sec.sh_type) {
671 this->initSymtab(Sections, &Sec);
676 case SHT_SYMTAB_SHNDX:
678 check(Obj.getSHNDXTable(Sec, Sections), toString(this));
681 this->VersymSec = &Sec;
684 this->VerdefSec = &Sec;
689 if (this->VersymSec && this->Symbols.empty())
690 error("SHT_GNU_versym should be associated with symbol table");
692 // Search for a DT_SONAME tag to initialize this->SoName.
695 ArrayRef<Elf_Dyn> Arr =
696 check(Obj.template getSectionContentsAsArray<Elf_Dyn>(DynamicSec),
698 for (const Elf_Dyn &Dyn : Arr) {
699 if (Dyn.d_tag == DT_SONAME) {
700 uint64_t Val = Dyn.getVal();
701 if (Val >= this->StringTable.size())
702 fatal(toString(this) + ": invalid DT_SONAME entry");
703 SoName = this->StringTable.data() + Val;
709 // Parse the version definitions in the object file if present. Returns a vector
710 // whose nth element contains a pointer to the Elf_Verdef for version identifier
711 // n. Version identifiers that are not definitions map to nullptr. The array
712 // always has at least length 1.
713 template <class ELFT>
714 std::vector<const typename ELFT::Verdef *>
715 SharedFile<ELFT>::parseVerdefs(const Elf_Versym *&Versym) {
716 std::vector<const Elf_Verdef *> Verdefs(1);
717 // We only need to process symbol versions for this DSO if it has both a
718 // versym and a verdef section, which indicates that the DSO contains symbol
719 // version definitions.
720 if (!VersymSec || !VerdefSec)
723 // The location of the first global versym entry.
724 const char *Base = this->MB.getBuffer().data();
725 Versym = reinterpret_cast<const Elf_Versym *>(Base + VersymSec->sh_offset) +
728 // We cannot determine the largest verdef identifier without inspecting
729 // every Elf_Verdef, but both bfd and gold assign verdef identifiers
730 // sequentially starting from 1, so we predict that the largest identifier
731 // will be VerdefCount.
732 unsigned VerdefCount = VerdefSec->sh_info;
733 Verdefs.resize(VerdefCount + 1);
735 // Build the Verdefs array by following the chain of Elf_Verdef objects
736 // from the start of the .gnu.version_d section.
737 const char *Verdef = Base + VerdefSec->sh_offset;
738 for (unsigned I = 0; I != VerdefCount; ++I) {
739 auto *CurVerdef = reinterpret_cast<const Elf_Verdef *>(Verdef);
740 Verdef += CurVerdef->vd_next;
741 unsigned VerdefIndex = CurVerdef->vd_ndx;
742 if (Verdefs.size() <= VerdefIndex)
743 Verdefs.resize(VerdefIndex + 1);
744 Verdefs[VerdefIndex] = CurVerdef;
750 // Fully parse the shared object file. This must be called after parseSoName().
751 template <class ELFT> void SharedFile<ELFT>::parseRest() {
752 // Create mapping from version identifiers to Elf_Verdef entries.
753 const Elf_Versym *Versym = nullptr;
754 std::vector<const Elf_Verdef *> Verdefs = parseVerdefs(Versym);
756 Elf_Sym_Range Syms = this->getGlobalSymbols();
757 for (const Elf_Sym &Sym : Syms) {
758 unsigned VersymIndex = 0;
760 VersymIndex = Versym->vs_index;
763 bool Hidden = VersymIndex & VERSYM_HIDDEN;
764 VersymIndex = VersymIndex & ~VERSYM_HIDDEN;
766 StringRef Name = check(Sym.getName(this->StringTable), toString(this));
767 if (Sym.isUndefined()) {
768 Undefs.push_back(Name);
772 // Ignore local symbols.
773 if (Versym && VersymIndex == VER_NDX_LOCAL)
776 const Elf_Verdef *V =
777 VersymIndex == VER_NDX_GLOBAL ? nullptr : Verdefs[VersymIndex];
780 elf::Symtab<ELFT>::X->addShared(this, Name, Sym, V);
782 // Also add the symbol with the versioned name to handle undefined symbols
783 // with explicit versions.
785 StringRef VerName = this->StringTable.data() + V->getAux()->vda_name;
786 Name = Saver.save(Name + "@" + VerName);
787 elf::Symtab<ELFT>::X->addShared(this, Name, Sym, V);
792 static ELFKind getBitcodeELFKind(const Triple &T) {
793 if (T.isLittleEndian())
794 return T.isArch64Bit() ? ELF64LEKind : ELF32LEKind;
795 return T.isArch64Bit() ? ELF64BEKind : ELF32BEKind;
798 static uint8_t getBitcodeMachineKind(StringRef Path, const Triple &T) {
799 switch (T.getArch()) {
800 case Triple::aarch64:
808 case Triple::mips64el:
815 return T.isOSIAMCU() ? EM_IAMCU : EM_386;
819 fatal(Path + ": could not infer e_machine from bitcode target triple " +
824 BitcodeFile::BitcodeFile(MemoryBufferRef MB, StringRef ArchiveName,
825 uint64_t OffsetInArchive)
826 : InputFile(BitcodeKind, MB) {
827 this->ArchiveName = ArchiveName;
829 // Here we pass a new MemoryBufferRef which is identified by ArchiveName
830 // (the fully resolved path of the archive) + member name + offset of the
831 // member in the archive.
832 // ThinLTO uses the MemoryBufferRef identifier to access its internal
833 // data structures and if two archives define two members with the same name,
834 // this causes a collision which result in only one of the objects being
835 // taken into consideration at LTO time (which very likely causes undefined
836 // symbols later in the link stage).
837 MemoryBufferRef MBRef(MB.getBuffer(),
838 Saver.save(ArchiveName + MB.getBufferIdentifier() +
839 utostr(OffsetInArchive)));
840 Obj = check(lto::InputFile::create(MBRef), toString(this));
842 Triple T(Obj->getTargetTriple());
843 EKind = getBitcodeELFKind(T);
844 EMachine = getBitcodeMachineKind(MB.getBufferIdentifier(), T);
847 static uint8_t mapVisibility(GlobalValue::VisibilityTypes GvVisibility) {
848 switch (GvVisibility) {
849 case GlobalValue::DefaultVisibility:
851 case GlobalValue::HiddenVisibility:
853 case GlobalValue::ProtectedVisibility:
854 return STV_PROTECTED;
856 llvm_unreachable("unknown visibility");
859 template <class ELFT>
860 static Symbol *createBitcodeSymbol(const std::vector<bool> &KeptComdats,
861 const lto::InputFile::Symbol &ObjSym,
863 StringRef NameRef = Saver.save(ObjSym.getName());
864 uint32_t Binding = ObjSym.isWeak() ? STB_WEAK : STB_GLOBAL;
866 uint8_t Type = ObjSym.isTLS() ? STT_TLS : STT_NOTYPE;
867 uint8_t Visibility = mapVisibility(ObjSym.getVisibility());
868 bool CanOmitFromDynSym = ObjSym.canBeOmittedFromSymbolTable();
870 int C = ObjSym.getComdatIndex();
871 if (C != -1 && !KeptComdats[C])
872 return Symtab<ELFT>::X->addUndefined(NameRef, /*IsLocal=*/false, Binding,
873 Visibility, Type, CanOmitFromDynSym,
876 if (ObjSym.isUndefined())
877 return Symtab<ELFT>::X->addUndefined(NameRef, /*IsLocal=*/false, Binding,
878 Visibility, Type, CanOmitFromDynSym,
881 if (ObjSym.isCommon())
882 return Symtab<ELFT>::X->addCommon(NameRef, ObjSym.getCommonSize(),
883 ObjSym.getCommonAlignment(), Binding,
884 Visibility, STT_OBJECT, F);
886 return Symtab<ELFT>::X->addBitcode(NameRef, Binding, Visibility, Type,
887 CanOmitFromDynSym, F);
890 template <class ELFT>
891 void BitcodeFile::parse(DenseSet<CachedHashStringRef> &ComdatGroups) {
892 std::vector<bool> KeptComdats;
893 for (StringRef S : Obj->getComdatTable())
894 KeptComdats.push_back(ComdatGroups.insert(CachedHashStringRef(S)).second);
896 for (const lto::InputFile::Symbol &ObjSym : Obj->symbols())
897 Symbols.push_back(createBitcodeSymbol<ELFT>(KeptComdats, ObjSym, this));
900 static ELFKind getELFKind(MemoryBufferRef MB) {
902 unsigned char Endian;
903 std::tie(Size, Endian) = getElfArchType(MB.getBuffer());
905 if (Endian != ELFDATA2LSB && Endian != ELFDATA2MSB)
906 fatal(MB.getBufferIdentifier() + ": invalid data encoding");
907 if (Size != ELFCLASS32 && Size != ELFCLASS64)
908 fatal(MB.getBufferIdentifier() + ": invalid file class");
910 size_t BufSize = MB.getBuffer().size();
911 if ((Size == ELFCLASS32 && BufSize < sizeof(Elf32_Ehdr)) ||
912 (Size == ELFCLASS64 && BufSize < sizeof(Elf64_Ehdr)))
913 fatal(MB.getBufferIdentifier() + ": file is too short");
915 if (Size == ELFCLASS32)
916 return (Endian == ELFDATA2LSB) ? ELF32LEKind : ELF32BEKind;
917 return (Endian == ELFDATA2LSB) ? ELF64LEKind : ELF64BEKind;
920 template <class ELFT> void BinaryFile::parse() {
921 ArrayRef<uint8_t> Data = toArrayRef(MB.getBuffer());
923 make<InputSection>(SHF_ALLOC | SHF_WRITE, SHT_PROGBITS, 8, Data, ".data");
924 Sections.push_back(Section);
926 // For each input file foo that is embedded to a result as a binary
927 // blob, we define _binary_foo_{start,end,size} symbols, so that
928 // user programs can access blobs by name. Non-alphanumeric
929 // characters in a filename are replaced with underscore.
930 std::string S = "_binary_" + MB.getBufferIdentifier().str();
931 for (size_t I = 0; I < S.size(); ++I)
935 elf::Symtab<ELFT>::X->addRegular(Saver.save(S + "_start"), STV_DEFAULT,
936 STT_OBJECT, 0, 0, STB_GLOBAL, Section,
938 elf::Symtab<ELFT>::X->addRegular(Saver.save(S + "_end"), STV_DEFAULT,
939 STT_OBJECT, Data.size(), 0, STB_GLOBAL,
941 elf::Symtab<ELFT>::X->addRegular(Saver.save(S + "_size"), STV_DEFAULT,
942 STT_OBJECT, Data.size(), 0, STB_GLOBAL,
946 static bool isBitcode(MemoryBufferRef MB) {
947 using namespace sys::fs;
948 return identify_magic(MB.getBuffer()) == file_magic::bitcode;
951 InputFile *elf::createObjectFile(MemoryBufferRef MB, StringRef ArchiveName,
952 uint64_t OffsetInArchive) {
954 return make<BitcodeFile>(MB, ArchiveName, OffsetInArchive);
956 switch (getELFKind(MB)) {
958 return make<ObjectFile<ELF32LE>>(MB, ArchiveName);
960 return make<ObjectFile<ELF32BE>>(MB, ArchiveName);
962 return make<ObjectFile<ELF64LE>>(MB, ArchiveName);
964 return make<ObjectFile<ELF64BE>>(MB, ArchiveName);
966 llvm_unreachable("getELFKind");
970 InputFile *elf::createSharedFile(MemoryBufferRef MB, StringRef DefaultSoName) {
971 switch (getELFKind(MB)) {
973 return make<SharedFile<ELF32LE>>(MB, DefaultSoName);
975 return make<SharedFile<ELF32BE>>(MB, DefaultSoName);
977 return make<SharedFile<ELF64LE>>(MB, DefaultSoName);
979 return make<SharedFile<ELF64BE>>(MB, DefaultSoName);
981 llvm_unreachable("getELFKind");
985 MemoryBufferRef LazyObjectFile::getBuffer() {
987 return MemoryBufferRef();
992 InputFile *LazyObjectFile::fetch() {
993 MemoryBufferRef MBRef = getBuffer();
994 if (MBRef.getBuffer().empty())
996 return createObjectFile(MBRef, ArchiveName, OffsetInArchive);
999 template <class ELFT> void LazyObjectFile::parse() {
1000 for (StringRef Sym : getSymbols())
1001 Symtab<ELFT>::X->addLazyObject(Sym, *this);
1004 template <class ELFT> std::vector<StringRef> LazyObjectFile::getElfSymbols() {
1005 typedef typename ELFT::Shdr Elf_Shdr;
1006 typedef typename ELFT::Sym Elf_Sym;
1007 typedef typename ELFT::SymRange Elf_Sym_Range;
1009 const ELFFile<ELFT> Obj(this->MB.getBuffer());
1010 ArrayRef<Elf_Shdr> Sections = check(Obj.sections(), toString(this));
1011 for (const Elf_Shdr &Sec : Sections) {
1012 if (Sec.sh_type != SHT_SYMTAB)
1015 Elf_Sym_Range Syms = check(Obj.symbols(&Sec), toString(this));
1016 uint32_t FirstNonLocal = Sec.sh_info;
1017 StringRef StringTable =
1018 check(Obj.getStringTableForSymtab(Sec, Sections), toString(this));
1019 std::vector<StringRef> V;
1021 for (const Elf_Sym &Sym : Syms.slice(FirstNonLocal))
1022 if (Sym.st_shndx != SHN_UNDEF)
1023 V.push_back(check(Sym.getName(StringTable), toString(this)));
1029 std::vector<StringRef> LazyObjectFile::getBitcodeSymbols() {
1030 std::unique_ptr<lto::InputFile> Obj =
1031 check(lto::InputFile::create(this->MB), toString(this));
1032 std::vector<StringRef> V;
1033 for (const lto::InputFile::Symbol &Sym : Obj->symbols())
1034 if (!Sym.isUndefined())
1035 V.push_back(Saver.save(Sym.getName()));
1039 // Returns a vector of globally-visible defined symbol names.
1040 std::vector<StringRef> LazyObjectFile::getSymbols() {
1041 if (isBitcode(this->MB))
1042 return getBitcodeSymbols();
1044 switch (getELFKind(this->MB)) {
1046 return getElfSymbols<ELF32LE>();
1048 return getElfSymbols<ELF32BE>();
1050 return getElfSymbols<ELF64LE>();
1052 return getElfSymbols<ELF64BE>();
1054 llvm_unreachable("getELFKind");
1058 template void ArchiveFile::parse<ELF32LE>();
1059 template void ArchiveFile::parse<ELF32BE>();
1060 template void ArchiveFile::parse<ELF64LE>();
1061 template void ArchiveFile::parse<ELF64BE>();
1063 template void BitcodeFile::parse<ELF32LE>(DenseSet<CachedHashStringRef> &);
1064 template void BitcodeFile::parse<ELF32BE>(DenseSet<CachedHashStringRef> &);
1065 template void BitcodeFile::parse<ELF64LE>(DenseSet<CachedHashStringRef> &);
1066 template void BitcodeFile::parse<ELF64BE>(DenseSet<CachedHashStringRef> &);
1068 template void LazyObjectFile::parse<ELF32LE>();
1069 template void LazyObjectFile::parse<ELF32BE>();
1070 template void LazyObjectFile::parse<ELF64LE>();
1071 template void LazyObjectFile::parse<ELF64BE>();
1073 template class elf::ELFFileBase<ELF32LE>;
1074 template class elf::ELFFileBase<ELF32BE>;
1075 template class elf::ELFFileBase<ELF64LE>;
1076 template class elf::ELFFileBase<ELF64BE>;
1078 template class elf::ObjectFile<ELF32LE>;
1079 template class elf::ObjectFile<ELF32BE>;
1080 template class elf::ObjectFile<ELF64LE>;
1081 template class elf::ObjectFile<ELF64BE>;
1083 template class elf::SharedFile<ELF32LE>;
1084 template class elf::SharedFile<ELF32BE>;
1085 template class elf::SharedFile<ELF64LE>;
1086 template class elf::SharedFile<ELF64BE>;
1088 template void BinaryFile::parse<ELF32LE>();
1089 template void BinaryFile::parse<ELF32BE>();
1090 template void BinaryFile::parse<ELF64LE>();
1091 template void BinaryFile::parse<ELF64BE>();