1 //===- InputFiles.cpp -----------------------------------------------------===//
2 //
3 //                             The LLVM Linker
4 //
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
7 //
8 //===----------------------------------------------------------------------===//
9 
10 #include "InputFiles.h"
11 #include "InputSection.h"
12 #include "LinkerScript.h"
13 #include "SymbolTable.h"
14 #include "Symbols.h"
15 #include "SyntheticSections.h"
16 #include "lld/Common/ErrorHandler.h"
17 #include "lld/Common/Memory.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/ARMAttributeParser.h"
27 #include "llvm/Support/ARMBuildAttributes.h"
28 #include "llvm/Support/Path.h"
29 #include "llvm/Support/TarWriter.h"
30 #include "llvm/Support/raw_ostream.h"
31 
32 using namespace llvm;
33 using namespace llvm::ELF;
34 using namespace llvm::object;
35 using namespace llvm::sys;
36 using namespace llvm::sys::fs;
37 
38 using namespace lld;
39 using namespace lld::elf;
40 
41 std::vector<BinaryFile *> elf::BinaryFiles;
42 std::vector<BitcodeFile *> elf::BitcodeFiles;
43 std::vector<InputFile *> elf::ObjectFiles;
44 std::vector<InputFile *> elf::SharedFiles;
45 
46 TarWriter *elf::Tar;
47 
48 InputFile::InputFile(Kind K, MemoryBufferRef M) : MB(M), FileKind(K) {}
49 
50 Optional<MemoryBufferRef> elf::readFile(StringRef Path) {
51   // The --chroot option changes our virtual root directory.
52   // This is useful when you are dealing with files created by --reproduce.
53   if (!Config->Chroot.empty() && Path.startswith("/"))
54     Path = Saver.save(Config->Chroot + Path);
55 
56   log(Path);
57 
58   auto MBOrErr = MemoryBuffer::getFile(Path);
59   if (auto EC = MBOrErr.getError()) {
60     error("cannot open " + Path + ": " + EC.message());
61     return None;
62   }
63 
64   std::unique_ptr<MemoryBuffer> &MB = *MBOrErr;
65   MemoryBufferRef MBRef = MB->getMemBufferRef();
66   make<std::unique_ptr<MemoryBuffer>>(std::move(MB)); // take MB ownership
67 
68   if (Tar)
69     Tar->append(relativeToRoot(Path), MBRef.getBuffer());
70   return MBRef;
71 }
72 
73 // Concatenates arguments to construct a string representing an error location.
74 static std::string createFileLineMsg(StringRef Path, unsigned Line) {
75   std::string Filename = path::filename(Path);
76   std::string Lineno = ":" + std::to_string(Line);
77   if (Filename == Path)
78     return Filename + Lineno;
79   return Filename + Lineno + " (" + Path.str() + Lineno + ")";
80 }
81 
82 template <class ELFT>
83 static std::string getSrcMsgAux(ObjFile<ELFT> &File, const Symbol &Sym,
84                                 InputSectionBase &Sec, uint64_t Offset) {
85   // In DWARF, functions and variables are stored to different places.
86   // First, lookup a function for a given offset.
87   if (Optional<DILineInfo> Info = File.getDILineInfo(&Sec, Offset))
88     return createFileLineMsg(Info->FileName, Info->Line);
89 
90   // If it failed, lookup again as a variable.
91   if (Optional<std::pair<std::string, unsigned>> FileLine =
92           File.getVariableLoc(Sym.getName()))
93     return createFileLineMsg(FileLine->first, FileLine->second);
94 
95   // File.SourceFile contains STT_FILE symbol, and that is a last resort.
96   return File.SourceFile;
97 }
98 
99 std::string InputFile::getSrcMsg(const Symbol &Sym, InputSectionBase &Sec,
100                                  uint64_t Offset) {
101   if (kind() != ObjKind)
102     return "";
103   switch (Config->EKind) {
104   default:
105     llvm_unreachable("Invalid kind");
106   case ELF32LEKind:
107     return getSrcMsgAux(cast<ObjFile<ELF32LE>>(*this), Sym, Sec, Offset);
108   case ELF32BEKind:
109     return getSrcMsgAux(cast<ObjFile<ELF32BE>>(*this), Sym, Sec, Offset);
110   case ELF64LEKind:
111     return getSrcMsgAux(cast<ObjFile<ELF64LE>>(*this), Sym, Sec, Offset);
112   case ELF64BEKind:
113     return getSrcMsgAux(cast<ObjFile<ELF64BE>>(*this), Sym, Sec, Offset);
114   }
115 }
116 
117 template <class ELFT> void ObjFile<ELFT>::initializeDwarf() {
118   Dwarf = llvm::make_unique<DWARFContext>(make_unique<LLDDwarfObj<ELFT>>(this));
119   const DWARFObject &Obj = Dwarf->getDWARFObj();
120   DwarfLine.reset(new DWARFDebugLine);
121   DWARFDataExtractor LineData(Obj, Obj.getLineSection(), Config->IsLE,
122                               Config->Wordsize);
123 
124   for (std::unique_ptr<DWARFCompileUnit> &CU : Dwarf->compile_units()) {
125     const DWARFDebugLine::LineTable *LT = Dwarf->getLineTableForUnit(CU.get());
126     if (!LT)
127       continue;
128     LineTables.push_back(LT);
129 
130     // Loop over variable records and insert them to VariableLoc.
131     for (const auto &Entry : CU->dies()) {
132       DWARFDie Die(CU.get(), &Entry);
133       // Skip all tags that are not variables.
134       if (Die.getTag() != dwarf::DW_TAG_variable)
135         continue;
136 
137       // Skip if a local variable because we don't need them for generating
138       // error messages. In general, only non-local symbols can fail to be
139       // linked.
140       if (!dwarf::toUnsigned(Die.find(dwarf::DW_AT_external), 0))
141         continue;
142 
143       // Get the source filename index for the variable.
144       unsigned File = dwarf::toUnsigned(Die.find(dwarf::DW_AT_decl_file), 0);
145       if (!LT->hasFileAtIndex(File))
146         continue;
147 
148       // Get the line number on which the variable is declared.
149       unsigned Line = dwarf::toUnsigned(Die.find(dwarf::DW_AT_decl_line), 0);
150 
151       // Get the name of the variable and add the collected information to
152       // VariableLoc. Usually Name is non-empty, but it can be empty if the
153       // input object file lacks some debug info.
154       StringRef Name = dwarf::toString(Die.find(dwarf::DW_AT_name), "");
155       if (!Name.empty())
156         VariableLoc.insert({Name, {LT, File, Line}});
157     }
158   }
159 }
160 
161 // Returns the pair of file name and line number describing location of data
162 // object (variable, array, etc) definition.
163 template <class ELFT>
164 Optional<std::pair<std::string, unsigned>>
165 ObjFile<ELFT>::getVariableLoc(StringRef Name) {
166   llvm::call_once(InitDwarfLine, [this]() { initializeDwarf(); });
167 
168   // Return if we have no debug information about data object.
169   auto It = VariableLoc.find(Name);
170   if (It == VariableLoc.end())
171     return None;
172 
173   // Take file name string from line table.
174   std::string FileName;
175   if (!It->second.LT->getFileNameByIndex(
176           It->second.File, nullptr,
177           DILineInfoSpecifier::FileLineInfoKind::AbsoluteFilePath, FileName))
178     return None;
179 
180   return std::make_pair(FileName, It->second.Line);
181 }
182 
183 // Returns source line information for a given offset
184 // using DWARF debug info.
185 template <class ELFT>
186 Optional<DILineInfo> ObjFile<ELFT>::getDILineInfo(InputSectionBase *S,
187                                                   uint64_t Offset) {
188   llvm::call_once(InitDwarfLine, [this]() { initializeDwarf(); });
189 
190   // Use fake address calcuated by adding section file offset and offset in
191   // section. See comments for ObjectInfo class.
192   DILineInfo Info;
193   for (const llvm::DWARFDebugLine::LineTable *LT : LineTables)
194     if (LT->getFileLineInfoForAddress(
195             S->getOffsetInFile() + Offset, nullptr,
196             DILineInfoSpecifier::FileLineInfoKind::AbsoluteFilePath, Info))
197       return Info;
198   return None;
199 }
200 
201 // Returns source line information for a given offset using DWARF debug info.
202 template <class ELFT>
203 std::string ObjFile<ELFT>::getLineInfo(InputSectionBase *S, uint64_t Offset) {
204   if (Optional<DILineInfo> Info = getDILineInfo(S, Offset))
205     return Info->FileName + ":" + std::to_string(Info->Line);
206   return "";
207 }
208 
209 // Returns "<internal>", "foo.a(bar.o)" or "baz.o".
210 std::string lld::toString(const InputFile *F) {
211   if (!F)
212     return "<internal>";
213 
214   if (F->ToStringCache.empty()) {
215     if (F->ArchiveName.empty())
216       F->ToStringCache = F->getName();
217     else
218       F->ToStringCache = (F->ArchiveName + "(" + F->getName() + ")").str();
219   }
220   return F->ToStringCache;
221 }
222 
223 template <class ELFT>
224 ELFFileBase<ELFT>::ELFFileBase(Kind K, MemoryBufferRef MB) : InputFile(K, MB) {
225   if (ELFT::TargetEndianness == support::little)
226     EKind = ELFT::Is64Bits ? ELF64LEKind : ELF32LEKind;
227   else
228     EKind = ELFT::Is64Bits ? ELF64BEKind : ELF32BEKind;
229 
230   EMachine = getObj().getHeader()->e_machine;
231   OSABI = getObj().getHeader()->e_ident[llvm::ELF::EI_OSABI];
232 }
233 
234 template <class ELFT>
235 typename ELFT::SymRange ELFFileBase<ELFT>::getGlobalELFSyms() {
236   return makeArrayRef(ELFSyms.begin() + FirstGlobal, ELFSyms.end());
237 }
238 
239 template <class ELFT>
240 uint32_t ELFFileBase<ELFT>::getSectionIndex(const Elf_Sym &Sym) const {
241   return CHECK(getObj().getSectionIndex(&Sym, ELFSyms, SymtabSHNDX), this);
242 }
243 
244 template <class ELFT>
245 void ELFFileBase<ELFT>::initSymtab(ArrayRef<Elf_Shdr> Sections,
246                                    const Elf_Shdr *Symtab) {
247   FirstGlobal = Symtab->sh_info;
248   ELFSyms = CHECK(getObj().symbols(Symtab), this);
249   if (FirstGlobal == 0 || FirstGlobal > ELFSyms.size())
250     fatal(toString(this) + ": invalid sh_info in symbol table");
251 
252   StringTable =
253       CHECK(getObj().getStringTableForSymtab(*Symtab, Sections), this);
254 }
255 
256 template <class ELFT>
257 ObjFile<ELFT>::ObjFile(MemoryBufferRef M, StringRef ArchiveName)
258     : ELFFileBase<ELFT>(Base::ObjKind, M) {
259   this->ArchiveName = ArchiveName;
260 }
261 
262 template <class ELFT> ArrayRef<Symbol *> ObjFile<ELFT>::getLocalSymbols() {
263   if (this->Symbols.empty())
264     return {};
265   return makeArrayRef(this->Symbols).slice(1, this->FirstGlobal - 1);
266 }
267 
268 template <class ELFT> ArrayRef<Symbol *> ObjFile<ELFT>::getGlobalSymbols() {
269   return makeArrayRef(this->Symbols).slice(this->FirstGlobal);
270 }
271 
272 template <class ELFT>
273 void ObjFile<ELFT>::parse(DenseSet<CachedHashStringRef> &ComdatGroups) {
274   // Read a section table. JustSymbols is usually false.
275   if (this->JustSymbols)
276     initializeJustSymbols();
277   else
278     initializeSections(ComdatGroups);
279 
280   // Read a symbol table.
281   initializeSymbols();
282 }
283 
284 // Sections with SHT_GROUP and comdat bits define comdat section groups.
285 // They are identified and deduplicated by group name. This function
286 // returns a group name.
287 template <class ELFT>
288 StringRef ObjFile<ELFT>::getShtGroupSignature(ArrayRef<Elf_Shdr> Sections,
289                                               const Elf_Shdr &Sec) {
290   // Group signatures are stored as symbol names in object files.
291   // sh_info contains a symbol index, so we fetch a symbol and read its name.
292   if (this->ELFSyms.empty())
293     this->initSymtab(
294         Sections, CHECK(object::getSection<ELFT>(Sections, Sec.sh_link), this));
295 
296   const Elf_Sym *Sym =
297       CHECK(object::getSymbol<ELFT>(this->ELFSyms, Sec.sh_info), this);
298   StringRef Signature = CHECK(Sym->getName(this->StringTable), this);
299 
300   // As a special case, if a symbol is a section symbol and has no name,
301   // we use a section name as a signature.
302   //
303   // Such SHT_GROUP sections are invalid from the perspective of the ELF
304   // standard, but GNU gold 1.14 (the newest version as of July 2017) or
305   // older produce such sections as outputs for the -r option, so we need
306   // a bug-compatibility.
307   if (Signature.empty() && Sym->getType() == STT_SECTION)
308     return getSectionName(Sec);
309   return Signature;
310 }
311 
312 template <class ELFT>
313 ArrayRef<typename ObjFile<ELFT>::Elf_Word>
314 ObjFile<ELFT>::getShtGroupEntries(const Elf_Shdr &Sec) {
315   const ELFFile<ELFT> &Obj = this->getObj();
316   ArrayRef<Elf_Word> Entries =
317       CHECK(Obj.template getSectionContentsAsArray<Elf_Word>(&Sec), this);
318   if (Entries.empty() || Entries[0] != GRP_COMDAT)
319     fatal(toString(this) + ": unsupported SHT_GROUP format");
320   return Entries.slice(1);
321 }
322 
323 template <class ELFT> bool ObjFile<ELFT>::shouldMerge(const Elf_Shdr &Sec) {
324   // On a regular link we don't merge sections if -O0 (default is -O1). This
325   // sometimes makes the linker significantly faster, although the output will
326   // be bigger.
327   //
328   // Doing the same for -r would create a problem as it would combine sections
329   // with different sh_entsize. One option would be to just copy every SHF_MERGE
330   // section as is to the output. While this would produce a valid ELF file with
331   // usable SHF_MERGE sections, tools like (llvm-)?dwarfdump get confused when
332   // they see two .debug_str. We could have separate logic for combining
333   // SHF_MERGE sections based both on their name and sh_entsize, but that seems
334   // to be more trouble than it is worth. Instead, we just use the regular (-O1)
335   // logic for -r.
336   if (Config->Optimize == 0 && !Config->Relocatable)
337     return false;
338 
339   // A mergeable section with size 0 is useless because they don't have
340   // any data to merge. A mergeable string section with size 0 can be
341   // argued as invalid because it doesn't end with a null character.
342   // We'll avoid a mess by handling them as if they were non-mergeable.
343   if (Sec.sh_size == 0)
344     return false;
345 
346   // Check for sh_entsize. The ELF spec is not clear about the zero
347   // sh_entsize. It says that "the member [sh_entsize] contains 0 if
348   // the section does not hold a table of fixed-size entries". We know
349   // that Rust 1.13 produces a string mergeable section with a zero
350   // sh_entsize. Here we just accept it rather than being picky about it.
351   uint64_t EntSize = Sec.sh_entsize;
352   if (EntSize == 0)
353     return false;
354   if (Sec.sh_size % EntSize)
355     fatal(toString(this) +
356           ": SHF_MERGE section size must be a multiple of sh_entsize");
357 
358   uint64_t Flags = Sec.sh_flags;
359   if (!(Flags & SHF_MERGE))
360     return false;
361   if (Flags & SHF_WRITE)
362     fatal(toString(this) + ": writable SHF_MERGE section is not supported");
363 
364   return true;
365 }
366 
367 // This is for --just-symbols.
368 //
369 // --just-symbols is a very minor feature that allows you to link your
370 // output against other existing program, so that if you load both your
371 // program and the other program into memory, your output can refer the
372 // other program's symbols.
373 //
374 // When the option is given, we link "just symbols". The section table is
375 // initialized with null pointers.
376 template <class ELFT> void ObjFile<ELFT>::initializeJustSymbols() {
377   ArrayRef<Elf_Shdr> ObjSections = CHECK(this->getObj().sections(), this);
378   this->Sections.resize(ObjSections.size());
379 
380   for (const Elf_Shdr &Sec : ObjSections) {
381     if (Sec.sh_type != SHT_SYMTAB)
382       continue;
383     this->initSymtab(ObjSections, &Sec);
384     return;
385   }
386 }
387 
388 template <class ELFT>
389 void ObjFile<ELFT>::initializeSections(
390     DenseSet<CachedHashStringRef> &ComdatGroups) {
391   const ELFFile<ELFT> &Obj = this->getObj();
392 
393   ArrayRef<Elf_Shdr> ObjSections = CHECK(this->getObj().sections(), this);
394   uint64_t Size = ObjSections.size();
395   this->Sections.resize(Size);
396   this->SectionStringTable =
397       CHECK(Obj.getSectionStringTable(ObjSections), this);
398 
399   for (size_t I = 0, E = ObjSections.size(); I < E; I++) {
400     if (this->Sections[I] == &InputSection::Discarded)
401       continue;
402     const Elf_Shdr &Sec = ObjSections[I];
403 
404     // SHF_EXCLUDE'ed sections are discarded by the linker. However,
405     // if -r is given, we'll let the final link discard such sections.
406     // This is compatible with GNU.
407     if ((Sec.sh_flags & SHF_EXCLUDE) && !Config->Relocatable) {
408       this->Sections[I] = &InputSection::Discarded;
409       continue;
410     }
411 
412     switch (Sec.sh_type) {
413     case SHT_GROUP: {
414       // De-duplicate section groups by their signatures.
415       StringRef Signature = getShtGroupSignature(ObjSections, Sec);
416       bool IsNew = ComdatGroups.insert(CachedHashStringRef(Signature)).second;
417       this->Sections[I] = &InputSection::Discarded;
418 
419       // If it is a new section group, we want to keep group members.
420       // Group leader sections, which contain indices of group members, are
421       // discarded because they are useless beyond this point. The only
422       // exception is the -r option because in order to produce re-linkable
423       // object files, we want to pass through basically everything.
424       if (IsNew) {
425         if (Config->Relocatable)
426           this->Sections[I] = createInputSection(Sec);
427         continue;
428       }
429 
430       // Otherwise, discard group members.
431       for (uint32_t SecIndex : getShtGroupEntries(Sec)) {
432         if (SecIndex >= Size)
433           fatal(toString(this) +
434                 ": invalid section index in group: " + Twine(SecIndex));
435         this->Sections[SecIndex] = &InputSection::Discarded;
436       }
437       break;
438     }
439     case SHT_SYMTAB:
440       this->initSymtab(ObjSections, &Sec);
441       break;
442     case SHT_SYMTAB_SHNDX:
443       this->SymtabSHNDX = CHECK(Obj.getSHNDXTable(Sec, ObjSections), this);
444       break;
445     case SHT_STRTAB:
446     case SHT_NULL:
447       break;
448     default:
449       this->Sections[I] = createInputSection(Sec);
450     }
451 
452     // .ARM.exidx sections have a reverse dependency on the InputSection they
453     // have a SHF_LINK_ORDER dependency, this is identified by the sh_link.
454     if (Sec.sh_flags & SHF_LINK_ORDER) {
455       if (Sec.sh_link >= this->Sections.size())
456         fatal(toString(this) +
457               ": invalid sh_link index: " + Twine(Sec.sh_link));
458 
459       InputSectionBase *LinkSec = this->Sections[Sec.sh_link];
460       InputSection *IS = cast<InputSection>(this->Sections[I]);
461       LinkSec->DependentSections.push_back(IS);
462       if (!isa<InputSection>(LinkSec))
463         error("a section " + IS->Name +
464               " with SHF_LINK_ORDER should not refer a non-regular "
465               "section: " +
466               toString(LinkSec));
467     }
468   }
469 }
470 
471 // The ARM support in lld makes some use of instructions that are not available
472 // on all ARM architectures. Namely:
473 // - Use of BLX instruction for interworking between ARM and Thumb state.
474 // - Use of the extended Thumb branch encoding in relocation.
475 // - Use of the MOVT/MOVW instructions in Thumb Thunks.
476 // The ARM Attributes section contains information about the architecture chosen
477 // at compile time. We follow the convention that if at least one input object
478 // is compiled with an architecture that supports these features then lld is
479 // permitted to use them.
480 static void updateSupportedARMFeatures(const ARMAttributeParser &Attributes) {
481   if (!Attributes.hasAttribute(ARMBuildAttrs::CPU_arch))
482     return;
483   auto Arch = Attributes.getAttributeValue(ARMBuildAttrs::CPU_arch);
484   switch (Arch) {
485   case ARMBuildAttrs::Pre_v4:
486   case ARMBuildAttrs::v4:
487   case ARMBuildAttrs::v4T:
488     // Architectures prior to v5 do not support BLX instruction
489     break;
490   case ARMBuildAttrs::v5T:
491   case ARMBuildAttrs::v5TE:
492   case ARMBuildAttrs::v5TEJ:
493   case ARMBuildAttrs::v6:
494   case ARMBuildAttrs::v6KZ:
495   case ARMBuildAttrs::v6K:
496     Config->ARMHasBlx = true;
497     // Architectures used in pre-Cortex processors do not support
498     // The J1 = 1 J2 = 1 Thumb branch range extension, with the exception
499     // of Architecture v6T2 (arm1156t2-s and arm1156t2f-s) that do.
500     break;
501   default:
502     // All other Architectures have BLX and extended branch encoding
503     Config->ARMHasBlx = true;
504     Config->ARMJ1J2BranchEncoding = true;
505     if (Arch != ARMBuildAttrs::v6_M && Arch != ARMBuildAttrs::v6S_M)
506       // All Architectures used in Cortex processors with the exception
507       // of v6-M and v6S-M have the MOVT and MOVW instructions.
508       Config->ARMHasMovtMovw = true;
509     break;
510   }
511 }
512 
513 template <class ELFT>
514 InputSectionBase *ObjFile<ELFT>::getRelocTarget(const Elf_Shdr &Sec) {
515   uint32_t Idx = Sec.sh_info;
516   if (Idx >= this->Sections.size())
517     fatal(toString(this) + ": invalid relocated section index: " + Twine(Idx));
518   InputSectionBase *Target = this->Sections[Idx];
519 
520   // Strictly speaking, a relocation section must be included in the
521   // group of the section it relocates. However, LLVM 3.3 and earlier
522   // would fail to do so, so we gracefully handle that case.
523   if (Target == &InputSection::Discarded)
524     return nullptr;
525 
526   if (!Target)
527     fatal(toString(this) + ": unsupported relocation reference");
528   return Target;
529 }
530 
531 // Create a regular InputSection class that has the same contents
532 // as a given section.
533 static InputSection *toRegularSection(MergeInputSection *Sec) {
534   return make<InputSection>(Sec->File, Sec->Flags, Sec->Type, Sec->Alignment,
535                             Sec->Data, Sec->Name);
536 }
537 
538 template <class ELFT>
539 InputSectionBase *ObjFile<ELFT>::createInputSection(const Elf_Shdr &Sec) {
540   StringRef Name = getSectionName(Sec);
541 
542   switch (Sec.sh_type) {
543   case SHT_ARM_ATTRIBUTES: {
544     if (Config->EMachine != EM_ARM)
545       break;
546     ARMAttributeParser Attributes;
547     ArrayRef<uint8_t> Contents = check(this->getObj().getSectionContents(&Sec));
548     Attributes.Parse(Contents, /*isLittle*/ Config->EKind == ELF32LEKind);
549     updateSupportedARMFeatures(Attributes);
550     // FIXME: Retain the first attribute section we see. The eglibc ARM
551     // dynamic loaders require the presence of an attribute section for dlopen
552     // to work. In a full implementation we would merge all attribute sections.
553     if (InX::ARMAttributes == nullptr) {
554       InX::ARMAttributes = make<InputSection>(*this, Sec, Name);
555       return InX::ARMAttributes;
556     }
557     return &InputSection::Discarded;
558   }
559   case SHT_RELA:
560   case SHT_REL: {
561     // Find a relocation target section and associate this section with that.
562     // Target may have been discarded if it is in a different section group
563     // and the group is discarded, even though it's a violation of the
564     // spec. We handle that situation gracefully by discarding dangling
565     // relocation sections.
566     InputSectionBase *Target = getRelocTarget(Sec);
567     if (!Target)
568       return nullptr;
569 
570     // This section contains relocation information.
571     // If -r is given, we do not interpret or apply relocation
572     // but just copy relocation sections to output.
573     if (Config->Relocatable)
574       return make<InputSection>(*this, Sec, Name);
575 
576     if (Target->FirstRelocation)
577       fatal(toString(this) +
578             ": multiple relocation sections to one section are not supported");
579 
580     // ELF spec allows mergeable sections with relocations, but they are
581     // rare, and it is in practice hard to merge such sections by contents,
582     // because applying relocations at end of linking changes section
583     // contents. So, we simply handle such sections as non-mergeable ones.
584     // Degrading like this is acceptable because section merging is optional.
585     if (auto *MS = dyn_cast<MergeInputSection>(Target)) {
586       Target = toRegularSection(MS);
587       this->Sections[Sec.sh_info] = Target;
588     }
589 
590     if (Sec.sh_type == SHT_RELA) {
591       ArrayRef<Elf_Rela> Rels = CHECK(this->getObj().relas(&Sec), this);
592       Target->FirstRelocation = Rels.begin();
593       Target->NumRelocations = Rels.size();
594       Target->AreRelocsRela = true;
595     } else {
596       ArrayRef<Elf_Rel> Rels = CHECK(this->getObj().rels(&Sec), this);
597       Target->FirstRelocation = Rels.begin();
598       Target->NumRelocations = Rels.size();
599       Target->AreRelocsRela = false;
600     }
601     assert(isUInt<31>(Target->NumRelocations));
602 
603     // Relocation sections processed by the linker are usually removed
604     // from the output, so returning `nullptr` for the normal case.
605     // However, if -emit-relocs is given, we need to leave them in the output.
606     // (Some post link analysis tools need this information.)
607     if (Config->EmitRelocs) {
608       InputSection *RelocSec = make<InputSection>(*this, Sec, Name);
609       // We will not emit relocation section if target was discarded.
610       Target->DependentSections.push_back(RelocSec);
611       return RelocSec;
612     }
613     return nullptr;
614   }
615   }
616 
617   // The GNU linker uses .note.GNU-stack section as a marker indicating
618   // that the code in the object file does not expect that the stack is
619   // executable (in terms of NX bit). If all input files have the marker,
620   // the GNU linker adds a PT_GNU_STACK segment to tells the loader to
621   // make the stack non-executable. Most object files have this section as
622   // of 2017.
623   //
624   // But making the stack non-executable is a norm today for security
625   // reasons. Failure to do so may result in a serious security issue.
626   // Therefore, we make LLD always add PT_GNU_STACK unless it is
627   // explicitly told to do otherwise (by -z execstack). Because the stack
628   // executable-ness is controlled solely by command line options,
629   // .note.GNU-stack sections are simply ignored.
630   if (Name == ".note.GNU-stack")
631     return &InputSection::Discarded;
632 
633   // Split stacks is a feature to support a discontiguous stack. At least
634   // as of 2017, it seems that the feature is not being used widely.
635   // Only GNU gold supports that. We don't. For the details about that,
636   // see https://gcc.gnu.org/wiki/SplitStacks
637   if (Name == ".note.GNU-split-stack") {
638     error(toString(this) +
639           ": object file compiled with -fsplit-stack is not supported");
640     return &InputSection::Discarded;
641   }
642 
643   // The linkonce feature is a sort of proto-comdat. Some glibc i386 object
644   // files contain definitions of symbol "__x86.get_pc_thunk.bx" in linkonce
645   // sections. Drop those sections to avoid duplicate symbol errors.
646   // FIXME: This is glibc PR20543, we should remove this hack once that has been
647   // fixed for a while.
648   if (Name.startswith(".gnu.linkonce."))
649     return &InputSection::Discarded;
650 
651   // If we are creating a new .build-id section, strip existing .build-id
652   // sections so that the output won't have more than one .build-id.
653   // This is not usually a problem because input object files normally don't
654   // have .build-id sections, but you can create such files by
655   // "ld.{bfd,gold,lld} -r --build-id", and we want to guard against it.
656   if (Name == ".note.gnu.build-id" && Config->BuildId != BuildIdKind::None)
657     return &InputSection::Discarded;
658 
659   // The linker merges EH (exception handling) frames and creates a
660   // .eh_frame_hdr section for runtime. So we handle them with a special
661   // class. For relocatable outputs, they are just passed through.
662   if (Name == ".eh_frame" && !Config->Relocatable)
663     return make<EhInputSection>(*this, Sec, Name);
664 
665   if (shouldMerge(Sec))
666     return make<MergeInputSection>(*this, Sec, Name);
667   return make<InputSection>(*this, Sec, Name);
668 }
669 
670 template <class ELFT>
671 StringRef ObjFile<ELFT>::getSectionName(const Elf_Shdr &Sec) {
672   return CHECK(this->getObj().getSectionName(&Sec, SectionStringTable), this);
673 }
674 
675 template <class ELFT> void ObjFile<ELFT>::initializeSymbols() {
676   this->Symbols.reserve(this->ELFSyms.size());
677   for (const Elf_Sym &Sym : this->ELFSyms)
678     this->Symbols.push_back(createSymbol(&Sym));
679 }
680 
681 template <class ELFT> Symbol *ObjFile<ELFT>::createSymbol(const Elf_Sym *Sym) {
682   int Binding = Sym->getBinding();
683 
684   uint32_t SecIdx = this->getSectionIndex(*Sym);
685   if (SecIdx >= this->Sections.size())
686     fatal(toString(this) + ": invalid section index: " + Twine(SecIdx));
687 
688   InputSectionBase *Sec = this->Sections[SecIdx];
689   uint8_t StOther = Sym->st_other;
690   uint8_t Type = Sym->getType();
691   uint64_t Value = Sym->st_value;
692   uint64_t Size = Sym->st_size;
693 
694   if (Binding == STB_LOCAL) {
695     if (Sym->getType() == STT_FILE)
696       SourceFile = CHECK(Sym->getName(this->StringTable), this);
697 
698     if (this->StringTable.size() <= Sym->st_name)
699       fatal(toString(this) + ": invalid symbol name offset");
700 
701     StringRefZ Name = this->StringTable.data() + Sym->st_name;
702     if (Sym->st_shndx == SHN_UNDEF)
703       return make<Undefined>(this, Name, Binding, StOther, Type);
704 
705     return make<Defined>(this, Name, Binding, StOther, Type, Value, Size, Sec);
706   }
707 
708   StringRef Name = CHECK(Sym->getName(this->StringTable), this);
709 
710   switch (Sym->st_shndx) {
711   case SHN_UNDEF:
712     return Symtab->addUndefined<ELFT>(Name, Binding, StOther, Type,
713                                       /*CanOmitFromDynSym=*/false, this);
714   case SHN_COMMON:
715     if (Value == 0 || Value >= UINT32_MAX)
716       fatal(toString(this) + ": common symbol '" + Name +
717             "' has invalid alignment: " + Twine(Value));
718     return Symtab->addCommon(Name, Size, Value, Binding, StOther, Type, *this);
719   }
720 
721   switch (Binding) {
722   default:
723     fatal(toString(this) + ": unexpected binding: " + Twine(Binding));
724   case STB_GLOBAL:
725   case STB_WEAK:
726   case STB_GNU_UNIQUE:
727     if (Sec == &InputSection::Discarded)
728       return Symtab->addUndefined<ELFT>(Name, Binding, StOther, Type,
729                                         /*CanOmitFromDynSym=*/false, this);
730     return Symtab->addRegular(Name, StOther, Type, Value, Size, Binding, Sec,
731                               this);
732   }
733 }
734 
735 ArchiveFile::ArchiveFile(std::unique_ptr<Archive> &&File)
736     : InputFile(ArchiveKind, File->getMemoryBufferRef()),
737       File(std::move(File)) {}
738 
739 template <class ELFT> void ArchiveFile::parse() {
740   for (const Archive::Symbol &Sym : File->symbols())
741     Symtab->addLazyArchive<ELFT>(Sym.getName(), *this, Sym);
742 }
743 
744 // Returns a buffer pointing to a member file containing a given symbol.
745 InputFile *ArchiveFile::fetch(const Archive::Symbol &Sym) {
746   Archive::Child C =
747       CHECK(Sym.getMember(), toString(this) +
748                                  ": could not get the member for symbol " +
749                                  Sym.getName());
750 
751   if (!Seen.insert(C.getChildOffset()).second)
752     return nullptr;
753 
754   MemoryBufferRef MB =
755       CHECK(C.getMemoryBufferRef(),
756             toString(this) +
757                 ": could not get the buffer for the member defining symbol " +
758                 Sym.getName());
759 
760   if (Tar && C.getParent()->isThin())
761     Tar->append(relativeToRoot(CHECK(C.getFullName(), this)), MB.getBuffer());
762 
763   return createObjectFile(MB, getName(),
764                           C.getParent()->isThin() ? 0 : C.getChildOffset());
765 }
766 
767 template <class ELFT>
768 SharedFile<ELFT>::SharedFile(MemoryBufferRef M, StringRef DefaultSoName)
769     : ELFFileBase<ELFT>(Base::SharedKind, M), SoName(DefaultSoName),
770       IsNeeded(!Config->AsNeeded) {}
771 
772 // Partially parse the shared object file so that we can call
773 // getSoName on this object.
774 template <class ELFT> void SharedFile<ELFT>::parseSoName() {
775   const Elf_Shdr *DynamicSec = nullptr;
776   const ELFFile<ELFT> Obj = this->getObj();
777   ArrayRef<Elf_Shdr> Sections = CHECK(Obj.sections(), this);
778 
779   // Search for .dynsym, .dynamic, .symtab, .gnu.version and .gnu.version_d.
780   for (const Elf_Shdr &Sec : Sections) {
781     switch (Sec.sh_type) {
782     default:
783       continue;
784     case SHT_DYNSYM:
785       this->initSymtab(Sections, &Sec);
786       break;
787     case SHT_DYNAMIC:
788       DynamicSec = &Sec;
789       break;
790     case SHT_SYMTAB_SHNDX:
791       this->SymtabSHNDX = CHECK(Obj.getSHNDXTable(Sec, Sections), this);
792       break;
793     case SHT_GNU_versym:
794       this->VersymSec = &Sec;
795       break;
796     case SHT_GNU_verdef:
797       this->VerdefSec = &Sec;
798       break;
799     }
800   }
801 
802   if (this->VersymSec && this->ELFSyms.empty())
803     error("SHT_GNU_versym should be associated with symbol table");
804 
805   // Search for a DT_SONAME tag to initialize this->SoName.
806   if (!DynamicSec)
807     return;
808   ArrayRef<Elf_Dyn> Arr =
809       CHECK(Obj.template getSectionContentsAsArray<Elf_Dyn>(DynamicSec), this);
810   for (const Elf_Dyn &Dyn : Arr) {
811     if (Dyn.d_tag == DT_SONAME) {
812       uint64_t Val = Dyn.getVal();
813       if (Val >= this->StringTable.size())
814         fatal(toString(this) + ": invalid DT_SONAME entry");
815       SoName = this->StringTable.data() + Val;
816       return;
817     }
818   }
819 }
820 
821 // Parses ".gnu.version" section which is a parallel array for the symbol table.
822 // If a given file doesn't have ".gnu.version" section, returns VER_NDX_GLOBAL.
823 template <class ELFT> std::vector<uint32_t> SharedFile<ELFT>::parseVersyms() {
824   size_t Size = this->ELFSyms.size() - this->FirstGlobal;
825   if (!VersymSec)
826     return std::vector<uint32_t>(Size, VER_NDX_GLOBAL);
827 
828   const char *Base = this->MB.getBuffer().data();
829   const Elf_Versym *Versym =
830       reinterpret_cast<const Elf_Versym *>(Base + VersymSec->sh_offset) +
831       this->FirstGlobal;
832 
833   std::vector<uint32_t> Ret(Size);
834   for (size_t I = 0; I < Size; ++I)
835     Ret[I] = Versym[I].vs_index;
836   return Ret;
837 }
838 
839 // Parse the version definitions in the object file if present. Returns a vector
840 // whose nth element contains a pointer to the Elf_Verdef for version identifier
841 // n. Version identifiers that are not definitions map to nullptr.
842 template <class ELFT>
843 std::vector<const typename ELFT::Verdef *> SharedFile<ELFT>::parseVerdefs() {
844   if (!VerdefSec)
845     return {};
846 
847   // We cannot determine the largest verdef identifier without inspecting
848   // every Elf_Verdef, but both bfd and gold assign verdef identifiers
849   // sequentially starting from 1, so we predict that the largest identifier
850   // will be VerdefCount.
851   unsigned VerdefCount = VerdefSec->sh_info;
852   std::vector<const Elf_Verdef *> Verdefs(VerdefCount + 1);
853 
854   // Build the Verdefs array by following the chain of Elf_Verdef objects
855   // from the start of the .gnu.version_d section.
856   const char *Base = this->MB.getBuffer().data();
857   const char *Verdef = Base + VerdefSec->sh_offset;
858   for (unsigned I = 0; I != VerdefCount; ++I) {
859     auto *CurVerdef = reinterpret_cast<const Elf_Verdef *>(Verdef);
860     Verdef += CurVerdef->vd_next;
861     unsigned VerdefIndex = CurVerdef->vd_ndx;
862     if (Verdefs.size() <= VerdefIndex)
863       Verdefs.resize(VerdefIndex + 1);
864     Verdefs[VerdefIndex] = CurVerdef;
865   }
866 
867   return Verdefs;
868 }
869 
870 // We do not usually care about alignments of data in shared object
871 // files because the loader takes care of it. However, if we promote a
872 // DSO symbol to point to .bss due to copy relocation, we need to keep
873 // the original alignment requirements. We infer it in this function.
874 template <class ELFT>
875 uint32_t SharedFile<ELFT>::getAlignment(ArrayRef<Elf_Shdr> Sections,
876                                         const Elf_Sym &Sym) {
877   uint64_t Ret = 1;
878   if (Sym.st_value)
879     Ret = 1ULL << countTrailingZeros((uint64_t)Sym.st_value);
880   if (0 < Sym.st_shndx && Sym.st_shndx < Sections.size())
881     Ret = std::min<uint64_t>(Ret, Sections[Sym.st_shndx].sh_addralign);
882 
883   if (Ret > UINT32_MAX)
884     error(toString(this) + ": alignment too large: " +
885           CHECK(Sym.getName(this->StringTable), this));
886   return Ret;
887 }
888 
889 // Fully parse the shared object file. This must be called after parseSoName().
890 //
891 // This function parses symbol versions. If a DSO has version information,
892 // the file has a ".gnu.version_d" section which contains symbol version
893 // definitions. Each symbol is associated to one version through a table in
894 // ".gnu.version" section. That table is a parallel array for the symbol
895 // table, and each table entry contains an index in ".gnu.version_d".
896 //
897 // The special index 0 is reserved for VERF_NDX_LOCAL and 1 is for
898 // VER_NDX_GLOBAL. There's no table entry for these special versions in
899 // ".gnu.version_d".
900 //
901 // The file format for symbol versioning is perhaps a bit more complicated
902 // than necessary, but you can easily understand the code if you wrap your
903 // head around the data structure described above.
904 template <class ELFT> void SharedFile<ELFT>::parseRest() {
905   Verdefs = parseVerdefs();                       // parse .gnu.version_d
906   std::vector<uint32_t> Versyms = parseVersyms(); // parse .gnu.version
907   ArrayRef<Elf_Shdr> Sections = CHECK(this->getObj().sections(), this);
908 
909   // Add symbols to the symbol table.
910   ArrayRef<Elf_Sym> Syms = this->getGlobalELFSyms();
911   for (size_t I = 0; I < Syms.size(); ++I) {
912     const Elf_Sym &Sym = Syms[I];
913 
914     StringRef Name = CHECK(Sym.getName(this->StringTable), this);
915     if (Sym.isUndefined()) {
916       Symbol *S = Symtab->addUndefined<ELFT>(Name, Sym.getBinding(),
917                                              Sym.st_other, Sym.getType(),
918                                              /*CanOmitFromDynSym=*/false, this);
919       S->ExportDynamic = true;
920       continue;
921     }
922 
923     // ELF spec requires that all local symbols precede weak or global
924     // symbols in each symbol table, and the index of first non-local symbol
925     // is stored to sh_info. If a local symbol appears after some non-local
926     // symbol, that's a violation of the spec.
927     if (Sym.getBinding() == STB_LOCAL) {
928       warn("found local symbol '" + Name +
929            "' in global part of symbol table in file " + toString(this));
930       continue;
931     }
932 
933     // MIPS BFD linker puts _gp_disp symbol into DSO files and incorrectly
934     // assigns VER_NDX_LOCAL to this section global symbol. Here is a
935     // workaround for this bug.
936     uint32_t Idx = Versyms[I] & ~VERSYM_HIDDEN;
937     if (Config->EMachine == EM_MIPS && Idx == VER_NDX_LOCAL &&
938         Name == "_gp_disp")
939       continue;
940 
941     uint64_t Alignment = getAlignment(Sections, Sym);
942     if (!(Versyms[I] & VERSYM_HIDDEN))
943       Symtab->addShared(Name, *this, Sym, Alignment, Idx);
944 
945     // Also add the symbol with the versioned name to handle undefined symbols
946     // with explicit versions.
947     if (Idx == VER_NDX_GLOBAL)
948       continue;
949 
950     if (Idx >= Verdefs.size() || Idx == VER_NDX_LOCAL) {
951       error("corrupt input file: version definition index " + Twine(Idx) +
952             " for symbol " + Name + " is out of bounds\n>>> defined in " +
953             toString(this));
954       continue;
955     }
956 
957     StringRef VerName =
958         this->StringTable.data() + Verdefs[Idx]->getAux()->vda_name;
959     Name = Saver.save(Name + "@" + VerName);
960     Symtab->addShared(Name, *this, Sym, Alignment, Idx);
961   }
962 }
963 
964 static ELFKind getBitcodeELFKind(const Triple &T) {
965   if (T.isLittleEndian())
966     return T.isArch64Bit() ? ELF64LEKind : ELF32LEKind;
967   return T.isArch64Bit() ? ELF64BEKind : ELF32BEKind;
968 }
969 
970 static uint8_t getBitcodeMachineKind(StringRef Path, const Triple &T) {
971   switch (T.getArch()) {
972   case Triple::aarch64:
973     return EM_AARCH64;
974   case Triple::arm:
975   case Triple::thumb:
976     return EM_ARM;
977   case Triple::avr:
978     return EM_AVR;
979   case Triple::mips:
980   case Triple::mipsel:
981   case Triple::mips64:
982   case Triple::mips64el:
983     return EM_MIPS;
984   case Triple::ppc:
985     return EM_PPC;
986   case Triple::ppc64:
987     return EM_PPC64;
988   case Triple::x86:
989     return T.isOSIAMCU() ? EM_IAMCU : EM_386;
990   case Triple::x86_64:
991     return EM_X86_64;
992   default:
993     fatal(Path + ": could not infer e_machine from bitcode target triple " +
994           T.str());
995   }
996 }
997 
998 BitcodeFile::BitcodeFile(MemoryBufferRef MB, StringRef ArchiveName,
999                          uint64_t OffsetInArchive)
1000     : InputFile(BitcodeKind, MB) {
1001   this->ArchiveName = ArchiveName;
1002 
1003   // Here we pass a new MemoryBufferRef which is identified by ArchiveName
1004   // (the fully resolved path of the archive) + member name + offset of the
1005   // member in the archive.
1006   // ThinLTO uses the MemoryBufferRef identifier to access its internal
1007   // data structures and if two archives define two members with the same name,
1008   // this causes a collision which result in only one of the objects being
1009   // taken into consideration at LTO time (which very likely causes undefined
1010   // symbols later in the link stage).
1011   MemoryBufferRef MBRef(MB.getBuffer(),
1012                         Saver.save(ArchiveName + MB.getBufferIdentifier() +
1013                                    utostr(OffsetInArchive)));
1014   Obj = CHECK(lto::InputFile::create(MBRef), this);
1015 
1016   Triple T(Obj->getTargetTriple());
1017   EKind = getBitcodeELFKind(T);
1018   EMachine = getBitcodeMachineKind(MB.getBufferIdentifier(), T);
1019 }
1020 
1021 static uint8_t mapVisibility(GlobalValue::VisibilityTypes GvVisibility) {
1022   switch (GvVisibility) {
1023   case GlobalValue::DefaultVisibility:
1024     return STV_DEFAULT;
1025   case GlobalValue::HiddenVisibility:
1026     return STV_HIDDEN;
1027   case GlobalValue::ProtectedVisibility:
1028     return STV_PROTECTED;
1029   }
1030   llvm_unreachable("unknown visibility");
1031 }
1032 
1033 template <class ELFT>
1034 static Symbol *createBitcodeSymbol(const std::vector<bool> &KeptComdats,
1035                                    const lto::InputFile::Symbol &ObjSym,
1036                                    BitcodeFile &F) {
1037   StringRef NameRef = Saver.save(ObjSym.getName());
1038   uint32_t Binding = ObjSym.isWeak() ? STB_WEAK : STB_GLOBAL;
1039 
1040   uint8_t Type = ObjSym.isTLS() ? STT_TLS : STT_NOTYPE;
1041   uint8_t Visibility = mapVisibility(ObjSym.getVisibility());
1042   bool CanOmitFromDynSym = ObjSym.canBeOmittedFromSymbolTable();
1043 
1044   int C = ObjSym.getComdatIndex();
1045   if (C != -1 && !KeptComdats[C])
1046     return Symtab->addUndefined<ELFT>(NameRef, Binding, Visibility, Type,
1047                                       CanOmitFromDynSym, &F);
1048 
1049   if (ObjSym.isUndefined())
1050     return Symtab->addUndefined<ELFT>(NameRef, Binding, Visibility, Type,
1051                                       CanOmitFromDynSym, &F);
1052 
1053   if (ObjSym.isCommon())
1054     return Symtab->addCommon(NameRef, ObjSym.getCommonSize(),
1055                              ObjSym.getCommonAlignment(), Binding, Visibility,
1056                              STT_OBJECT, F);
1057 
1058   return Symtab->addBitcode(NameRef, Binding, Visibility, Type,
1059                             CanOmitFromDynSym, F);
1060 }
1061 
1062 template <class ELFT>
1063 void BitcodeFile::parse(DenseSet<CachedHashStringRef> &ComdatGroups) {
1064   std::vector<bool> KeptComdats;
1065   for (StringRef S : Obj->getComdatTable())
1066     KeptComdats.push_back(ComdatGroups.insert(CachedHashStringRef(S)).second);
1067 
1068   for (const lto::InputFile::Symbol &ObjSym : Obj->symbols())
1069     Symbols.push_back(createBitcodeSymbol<ELFT>(KeptComdats, ObjSym, *this));
1070 }
1071 
1072 static ELFKind getELFKind(MemoryBufferRef MB) {
1073   unsigned char Size;
1074   unsigned char Endian;
1075   std::tie(Size, Endian) = getElfArchType(MB.getBuffer());
1076 
1077   if (Endian != ELFDATA2LSB && Endian != ELFDATA2MSB)
1078     fatal(MB.getBufferIdentifier() + ": invalid data encoding");
1079   if (Size != ELFCLASS32 && Size != ELFCLASS64)
1080     fatal(MB.getBufferIdentifier() + ": invalid file class");
1081 
1082   size_t BufSize = MB.getBuffer().size();
1083   if ((Size == ELFCLASS32 && BufSize < sizeof(Elf32_Ehdr)) ||
1084       (Size == ELFCLASS64 && BufSize < sizeof(Elf64_Ehdr)))
1085     fatal(MB.getBufferIdentifier() + ": file is too short");
1086 
1087   if (Size == ELFCLASS32)
1088     return (Endian == ELFDATA2LSB) ? ELF32LEKind : ELF32BEKind;
1089   return (Endian == ELFDATA2LSB) ? ELF64LEKind : ELF64BEKind;
1090 }
1091 
1092 void BinaryFile::parse() {
1093   ArrayRef<uint8_t> Data = toArrayRef(MB.getBuffer());
1094   auto *Section = make<InputSection>(this, SHF_ALLOC | SHF_WRITE, SHT_PROGBITS,
1095                                      8, Data, ".data");
1096   Sections.push_back(Section);
1097 
1098   // For each input file foo that is embedded to a result as a binary
1099   // blob, we define _binary_foo_{start,end,size} symbols, so that
1100   // user programs can access blobs by name. Non-alphanumeric
1101   // characters in a filename are replaced with underscore.
1102   std::string S = "_binary_" + MB.getBufferIdentifier().str();
1103   for (size_t I = 0; I < S.size(); ++I)
1104     if (!isAlnum(S[I]))
1105       S[I] = '_';
1106 
1107   Symtab->addRegular(Saver.save(S + "_start"), STV_DEFAULT, STT_OBJECT, 0, 0,
1108                      STB_GLOBAL, Section, nullptr);
1109   Symtab->addRegular(Saver.save(S + "_end"), STV_DEFAULT, STT_OBJECT,
1110                      Data.size(), 0, STB_GLOBAL, Section, nullptr);
1111   Symtab->addRegular(Saver.save(S + "_size"), STV_DEFAULT, STT_OBJECT,
1112                      Data.size(), 0, STB_GLOBAL, nullptr, nullptr);
1113 }
1114 
1115 static bool isBitcode(MemoryBufferRef MB) {
1116   using namespace sys::fs;
1117   return identify_magic(MB.getBuffer()) == file_magic::bitcode;
1118 }
1119 
1120 InputFile *elf::createObjectFile(MemoryBufferRef MB, StringRef ArchiveName,
1121                                  uint64_t OffsetInArchive) {
1122   if (isBitcode(MB))
1123     return make<BitcodeFile>(MB, ArchiveName, OffsetInArchive);
1124 
1125   switch (getELFKind(MB)) {
1126   case ELF32LEKind:
1127     return make<ObjFile<ELF32LE>>(MB, ArchiveName);
1128   case ELF32BEKind:
1129     return make<ObjFile<ELF32BE>>(MB, ArchiveName);
1130   case ELF64LEKind:
1131     return make<ObjFile<ELF64LE>>(MB, ArchiveName);
1132   case ELF64BEKind:
1133     return make<ObjFile<ELF64BE>>(MB, ArchiveName);
1134   default:
1135     llvm_unreachable("getELFKind");
1136   }
1137 }
1138 
1139 InputFile *elf::createSharedFile(MemoryBufferRef MB, StringRef DefaultSoName) {
1140   switch (getELFKind(MB)) {
1141   case ELF32LEKind:
1142     return make<SharedFile<ELF32LE>>(MB, DefaultSoName);
1143   case ELF32BEKind:
1144     return make<SharedFile<ELF32BE>>(MB, DefaultSoName);
1145   case ELF64LEKind:
1146     return make<SharedFile<ELF64LE>>(MB, DefaultSoName);
1147   case ELF64BEKind:
1148     return make<SharedFile<ELF64BE>>(MB, DefaultSoName);
1149   default:
1150     llvm_unreachable("getELFKind");
1151   }
1152 }
1153 
1154 MemoryBufferRef LazyObjFile::getBuffer() {
1155   if (Seen)
1156     return MemoryBufferRef();
1157   Seen = true;
1158   return MB;
1159 }
1160 
1161 InputFile *LazyObjFile::fetch() {
1162   MemoryBufferRef MBRef = getBuffer();
1163   if (MBRef.getBuffer().empty())
1164     return nullptr;
1165   return createObjectFile(MBRef, ArchiveName, OffsetInArchive);
1166 }
1167 
1168 template <class ELFT> void LazyObjFile::parse() {
1169   // A lazy object file wraps either a bitcode file or an ELF file.
1170   if (isBitcode(this->MB)) {
1171     std::unique_ptr<lto::InputFile> Obj =
1172         CHECK(lto::InputFile::create(this->MB), this);
1173     for (const lto::InputFile::Symbol &Sym : Obj->symbols())
1174       if (!Sym.isUndefined())
1175         Symtab->addLazyObject<ELFT>(Saver.save(Sym.getName()), *this);
1176     return;
1177   }
1178 
1179   switch (getELFKind(this->MB)) {
1180   case ELF32LEKind:
1181     addElfSymbols<ELF32LE>();
1182     return;
1183   case ELF32BEKind:
1184     addElfSymbols<ELF32BE>();
1185     return;
1186   case ELF64LEKind:
1187     addElfSymbols<ELF64LE>();
1188     return;
1189   case ELF64BEKind:
1190     addElfSymbols<ELF64BE>();
1191     return;
1192   default:
1193     llvm_unreachable("getELFKind");
1194   }
1195 }
1196 
1197 template <class ELFT> void LazyObjFile::addElfSymbols() {
1198   ELFFile<ELFT> Obj = check(ELFFile<ELFT>::create(MB.getBuffer()));
1199   ArrayRef<typename ELFT::Shdr> Sections = CHECK(Obj.sections(), this);
1200 
1201   for (const typename ELFT::Shdr &Sec : Sections) {
1202     if (Sec.sh_type != SHT_SYMTAB)
1203       continue;
1204 
1205     typename ELFT::SymRange Syms = CHECK(Obj.symbols(&Sec), this);
1206     uint32_t FirstGlobal = Sec.sh_info;
1207     StringRef StringTable =
1208         CHECK(Obj.getStringTableForSymtab(Sec, Sections), this);
1209 
1210     for (const typename ELFT::Sym &Sym : Syms.slice(FirstGlobal))
1211       if (Sym.st_shndx != SHN_UNDEF)
1212         Symtab->addLazyObject<ELFT>(CHECK(Sym.getName(StringTable), this),
1213                                     *this);
1214     return;
1215   }
1216 }
1217 
1218 template void ArchiveFile::parse<ELF32LE>();
1219 template void ArchiveFile::parse<ELF32BE>();
1220 template void ArchiveFile::parse<ELF64LE>();
1221 template void ArchiveFile::parse<ELF64BE>();
1222 
1223 template void BitcodeFile::parse<ELF32LE>(DenseSet<CachedHashStringRef> &);
1224 template void BitcodeFile::parse<ELF32BE>(DenseSet<CachedHashStringRef> &);
1225 template void BitcodeFile::parse<ELF64LE>(DenseSet<CachedHashStringRef> &);
1226 template void BitcodeFile::parse<ELF64BE>(DenseSet<CachedHashStringRef> &);
1227 
1228 template void LazyObjFile::parse<ELF32LE>();
1229 template void LazyObjFile::parse<ELF32BE>();
1230 template void LazyObjFile::parse<ELF64LE>();
1231 template void LazyObjFile::parse<ELF64BE>();
1232 
1233 template class elf::ELFFileBase<ELF32LE>;
1234 template class elf::ELFFileBase<ELF32BE>;
1235 template class elf::ELFFileBase<ELF64LE>;
1236 template class elf::ELFFileBase<ELF64BE>;
1237 
1238 template class elf::ObjFile<ELF32LE>;
1239 template class elf::ObjFile<ELF32BE>;
1240 template class elf::ObjFile<ELF64LE>;
1241 template class elf::ObjFile<ELF64BE>;
1242 
1243 template class elf::SharedFile<ELF32LE>;
1244 template class elf::SharedFile<ELF32BE>;
1245 template class elf::SharedFile<ELF64LE>;
1246 template class elf::SharedFile<ELF64BE>;
1247