1 //===- SyntheticSections.cpp ----------------------------------------------===//
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file contains linker-synthesized sections. Currently,
11 // synthetic sections are created either output sections or input sections,
12 // but we are rewriting code so that all synthetic sections are created as
15 //===----------------------------------------------------------------------===//
17 #include "SyntheticSections.h"
20 #include "InputFiles.h"
21 #include "LinkerScript.h"
22 #include "OutputSections.h"
24 #include "SymbolTable.h"
28 #include "lld/Common/ErrorHandler.h"
29 #include "lld/Common/Memory.h"
30 #include "lld/Common/Threads.h"
31 #include "lld/Common/Version.h"
32 #include "llvm/BinaryFormat/Dwarf.h"
33 #include "llvm/DebugInfo/DWARF/DWARFDebugPubTable.h"
34 #include "llvm/Object/Decompressor.h"
35 #include "llvm/Object/ELFObjectFile.h"
36 #include "llvm/Support/Endian.h"
37 #include "llvm/Support/LEB128.h"
38 #include "llvm/Support/MD5.h"
39 #include "llvm/Support/RandomNumberGenerator.h"
40 #include "llvm/Support/SHA1.h"
41 #include "llvm/Support/xxhash.h"
46 using namespace llvm::dwarf;
47 using namespace llvm::ELF;
48 using namespace llvm::object;
49 using namespace llvm::support;
50 using namespace llvm::support::endian;
53 using namespace lld::elf;
55 constexpr size_t MergeNoTailSection::NumShards;
57 static void write32(void *Buf, uint32_t Val) {
58 endian::write32(Buf, Val, Config->Endianness);
61 uint64_t SyntheticSection::getVA() const {
62 if (OutputSection *Sec = getParent())
63 return Sec->Addr + OutSecOff;
67 // Returns an LLD version string.
68 static ArrayRef<uint8_t> getVersion() {
69 // Check LLD_VERSION first for ease of testing.
70 // You can get consitent output by using the environment variable.
71 // This is only for testing.
72 StringRef S = getenv("LLD_VERSION");
74 S = Saver.save(Twine("Linker: ") + getLLDVersion());
76 // +1 to include the terminating '\0'.
77 return {(const uint8_t *)S.data(), S.size() + 1};
80 // Creates a .comment section containing LLD version info.
81 // With this feature, you can identify LLD-generated binaries easily
82 // by "readelf --string-dump .comment <file>".
83 // The returned object is a mergeable string section.
84 MergeInputSection *elf::createCommentSection() {
85 return make<MergeInputSection>(SHF_MERGE | SHF_STRINGS, SHT_PROGBITS, 1,
86 getVersion(), ".comment");
89 // .MIPS.abiflags section.
91 MipsAbiFlagsSection<ELFT>::MipsAbiFlagsSection(Elf_Mips_ABIFlags Flags)
92 : SyntheticSection(SHF_ALLOC, SHT_MIPS_ABIFLAGS, 8, ".MIPS.abiflags"),
94 this->Entsize = sizeof(Elf_Mips_ABIFlags);
97 template <class ELFT> void MipsAbiFlagsSection<ELFT>::writeTo(uint8_t *Buf) {
98 memcpy(Buf, &Flags, sizeof(Flags));
101 template <class ELFT>
102 MipsAbiFlagsSection<ELFT> *MipsAbiFlagsSection<ELFT>::create() {
103 Elf_Mips_ABIFlags Flags = {};
106 for (InputSectionBase *Sec : InputSections) {
107 if (Sec->Type != SHT_MIPS_ABIFLAGS)
112 std::string Filename = toString(Sec->File);
113 const size_t Size = Sec->Data.size();
114 // Older version of BFD (such as the default FreeBSD linker) concatenate
115 // .MIPS.abiflags instead of merging. To allow for this case (or potential
116 // zero padding) we ignore everything after the first Elf_Mips_ABIFlags
117 if (Size < sizeof(Elf_Mips_ABIFlags)) {
118 error(Filename + ": invalid size of .MIPS.abiflags section: got " +
119 Twine(Size) + " instead of " + Twine(sizeof(Elf_Mips_ABIFlags)));
122 auto *S = reinterpret_cast<const Elf_Mips_ABIFlags *>(Sec->Data.data());
123 if (S->version != 0) {
124 error(Filename + ": unexpected .MIPS.abiflags version " +
129 // LLD checks ISA compatibility in calcMipsEFlags(). Here we just
130 // select the highest number of ISA/Rev/Ext.
131 Flags.isa_level = std::max(Flags.isa_level, S->isa_level);
132 Flags.isa_rev = std::max(Flags.isa_rev, S->isa_rev);
133 Flags.isa_ext = std::max(Flags.isa_ext, S->isa_ext);
134 Flags.gpr_size = std::max(Flags.gpr_size, S->gpr_size);
135 Flags.cpr1_size = std::max(Flags.cpr1_size, S->cpr1_size);
136 Flags.cpr2_size = std::max(Flags.cpr2_size, S->cpr2_size);
137 Flags.ases |= S->ases;
138 Flags.flags1 |= S->flags1;
139 Flags.flags2 |= S->flags2;
140 Flags.fp_abi = elf::getMipsFpAbiFlag(Flags.fp_abi, S->fp_abi, Filename);
144 return make<MipsAbiFlagsSection<ELFT>>(Flags);
148 // .MIPS.options section.
149 template <class ELFT>
150 MipsOptionsSection<ELFT>::MipsOptionsSection(Elf_Mips_RegInfo Reginfo)
151 : SyntheticSection(SHF_ALLOC, SHT_MIPS_OPTIONS, 8, ".MIPS.options"),
153 this->Entsize = sizeof(Elf_Mips_Options) + sizeof(Elf_Mips_RegInfo);
156 template <class ELFT> void MipsOptionsSection<ELFT>::writeTo(uint8_t *Buf) {
157 auto *Options = reinterpret_cast<Elf_Mips_Options *>(Buf);
158 Options->kind = ODK_REGINFO;
159 Options->size = getSize();
161 if (!Config->Relocatable)
162 Reginfo.ri_gp_value = InX::MipsGot->getGp();
163 memcpy(Buf + sizeof(Elf_Mips_Options), &Reginfo, sizeof(Reginfo));
166 template <class ELFT>
167 MipsOptionsSection<ELFT> *MipsOptionsSection<ELFT>::create() {
172 std::vector<InputSectionBase *> Sections;
173 for (InputSectionBase *Sec : InputSections)
174 if (Sec->Type == SHT_MIPS_OPTIONS)
175 Sections.push_back(Sec);
177 if (Sections.empty())
180 Elf_Mips_RegInfo Reginfo = {};
181 for (InputSectionBase *Sec : Sections) {
184 std::string Filename = toString(Sec->File);
185 ArrayRef<uint8_t> D = Sec->Data;
188 if (D.size() < sizeof(Elf_Mips_Options)) {
189 error(Filename + ": invalid size of .MIPS.options section");
193 auto *Opt = reinterpret_cast<const Elf_Mips_Options *>(D.data());
194 if (Opt->kind == ODK_REGINFO) {
195 if (Config->Relocatable && Opt->getRegInfo().ri_gp_value)
196 error(Filename + ": unsupported non-zero ri_gp_value");
197 Reginfo.ri_gprmask |= Opt->getRegInfo().ri_gprmask;
198 Sec->getFile<ELFT>()->MipsGp0 = Opt->getRegInfo().ri_gp_value;
203 fatal(Filename + ": zero option descriptor size");
204 D = D.slice(Opt->size);
208 return make<MipsOptionsSection<ELFT>>(Reginfo);
211 // MIPS .reginfo section.
212 template <class ELFT>
213 MipsReginfoSection<ELFT>::MipsReginfoSection(Elf_Mips_RegInfo Reginfo)
214 : SyntheticSection(SHF_ALLOC, SHT_MIPS_REGINFO, 4, ".reginfo"),
216 this->Entsize = sizeof(Elf_Mips_RegInfo);
219 template <class ELFT> void MipsReginfoSection<ELFT>::writeTo(uint8_t *Buf) {
220 if (!Config->Relocatable)
221 Reginfo.ri_gp_value = InX::MipsGot->getGp();
222 memcpy(Buf, &Reginfo, sizeof(Reginfo));
225 template <class ELFT>
226 MipsReginfoSection<ELFT> *MipsReginfoSection<ELFT>::create() {
227 // Section should be alive for O32 and N32 ABIs only.
231 std::vector<InputSectionBase *> Sections;
232 for (InputSectionBase *Sec : InputSections)
233 if (Sec->Type == SHT_MIPS_REGINFO)
234 Sections.push_back(Sec);
236 if (Sections.empty())
239 Elf_Mips_RegInfo Reginfo = {};
240 for (InputSectionBase *Sec : Sections) {
243 if (Sec->Data.size() != sizeof(Elf_Mips_RegInfo)) {
244 error(toString(Sec->File) + ": invalid size of .reginfo section");
247 auto *R = reinterpret_cast<const Elf_Mips_RegInfo *>(Sec->Data.data());
248 if (Config->Relocatable && R->ri_gp_value)
249 error(toString(Sec->File) + ": unsupported non-zero ri_gp_value");
251 Reginfo.ri_gprmask |= R->ri_gprmask;
252 Sec->getFile<ELFT>()->MipsGp0 = R->ri_gp_value;
255 return make<MipsReginfoSection<ELFT>>(Reginfo);
258 InputSection *elf::createInterpSection() {
259 // StringSaver guarantees that the returned string ends with '\0'.
260 StringRef S = Saver.save(Config->DynamicLinker);
261 ArrayRef<uint8_t> Contents = {(const uint8_t *)S.data(), S.size() + 1};
263 auto *Sec = make<InputSection>(nullptr, SHF_ALLOC, SHT_PROGBITS, 1, Contents,
269 Symbol *elf::addSyntheticLocal(StringRef Name, uint8_t Type, uint64_t Value,
270 uint64_t Size, InputSectionBase &Section) {
271 auto *S = make<Defined>(Section.File, Name, STB_LOCAL, STV_DEFAULT, Type,
272 Value, Size, &Section);
274 InX::SymTab->addSymbol(S);
278 static size_t getHashSize() {
279 switch (Config->BuildId) {
280 case BuildIdKind::Fast:
282 case BuildIdKind::Md5:
283 case BuildIdKind::Uuid:
285 case BuildIdKind::Sha1:
287 case BuildIdKind::Hexstring:
288 return Config->BuildIdVector.size();
290 llvm_unreachable("unknown BuildIdKind");
294 BuildIdSection::BuildIdSection()
295 : SyntheticSection(SHF_ALLOC, SHT_NOTE, 4, ".note.gnu.build-id"),
296 HashSize(getHashSize()) {}
298 void BuildIdSection::writeTo(uint8_t *Buf) {
299 write32(Buf, 4); // Name size
300 write32(Buf + 4, HashSize); // Content size
301 write32(Buf + 8, NT_GNU_BUILD_ID); // Type
302 memcpy(Buf + 12, "GNU", 4); // Name string
306 // Split one uint8 array into small pieces of uint8 arrays.
307 static std::vector<ArrayRef<uint8_t>> split(ArrayRef<uint8_t> Arr,
309 std::vector<ArrayRef<uint8_t>> Ret;
310 while (Arr.size() > ChunkSize) {
311 Ret.push_back(Arr.take_front(ChunkSize));
312 Arr = Arr.drop_front(ChunkSize);
319 // Computes a hash value of Data using a given hash function.
320 // In order to utilize multiple cores, we first split data into 1MB
321 // chunks, compute a hash for each chunk, and then compute a hash value
322 // of the hash values.
323 void BuildIdSection::computeHash(
324 llvm::ArrayRef<uint8_t> Data,
325 std::function<void(uint8_t *Dest, ArrayRef<uint8_t> Arr)> HashFn) {
326 std::vector<ArrayRef<uint8_t>> Chunks = split(Data, 1024 * 1024);
327 std::vector<uint8_t> Hashes(Chunks.size() * HashSize);
329 // Compute hash values.
330 parallelForEachN(0, Chunks.size(), [&](size_t I) {
331 HashFn(Hashes.data() + I * HashSize, Chunks[I]);
334 // Write to the final output buffer.
335 HashFn(HashBuf, Hashes);
338 BssSection::BssSection(StringRef Name, uint64_t Size, uint32_t Alignment)
339 : SyntheticSection(SHF_ALLOC | SHF_WRITE, SHT_NOBITS, Alignment, Name) {
341 if (OutputSection *Sec = getParent())
342 Sec->Alignment = std::max(Sec->Alignment, Alignment);
346 void BuildIdSection::writeBuildId(ArrayRef<uint8_t> Buf) {
347 switch (Config->BuildId) {
348 case BuildIdKind::Fast:
349 computeHash(Buf, [](uint8_t *Dest, ArrayRef<uint8_t> Arr) {
350 write64le(Dest, xxHash64(toStringRef(Arr)));
353 case BuildIdKind::Md5:
354 computeHash(Buf, [](uint8_t *Dest, ArrayRef<uint8_t> Arr) {
355 memcpy(Dest, MD5::hash(Arr).data(), 16);
358 case BuildIdKind::Sha1:
359 computeHash(Buf, [](uint8_t *Dest, ArrayRef<uint8_t> Arr) {
360 memcpy(Dest, SHA1::hash(Arr).data(), 20);
363 case BuildIdKind::Uuid:
364 if (auto EC = getRandomBytes(HashBuf, HashSize))
365 error("entropy source failure: " + EC.message());
367 case BuildIdKind::Hexstring:
368 memcpy(HashBuf, Config->BuildIdVector.data(), Config->BuildIdVector.size());
371 llvm_unreachable("unknown BuildIdKind");
375 EhFrameSection::EhFrameSection()
376 : SyntheticSection(SHF_ALLOC, SHT_PROGBITS, 1, ".eh_frame") {}
378 // Search for an existing CIE record or create a new one.
379 // CIE records from input object files are uniquified by their contents
380 // and where their relocations point to.
381 template <class ELFT, class RelTy>
382 CieRecord *EhFrameSection::addCie(EhSectionPiece &Cie, ArrayRef<RelTy> Rels) {
383 auto *Sec = cast<EhInputSection>(Cie.Sec);
384 if (read32(Cie.data().data() + 4, Config->Endianness) != 0)
385 fatal(toString(Sec) + ": CIE expected at beginning of .eh_frame");
387 Symbol *Personality = nullptr;
388 unsigned FirstRelI = Cie.FirstRelocation;
389 if (FirstRelI != (unsigned)-1)
391 &Sec->template getFile<ELFT>()->getRelocTargetSym(Rels[FirstRelI]);
393 // Search for an existing CIE by CIE contents/relocation target pair.
394 CieRecord *&Rec = CieMap[{Cie.data(), Personality}];
396 // If not found, create a new one.
398 Rec = make<CieRecord>();
400 CieRecords.push_back(Rec);
405 // There is one FDE per function. Returns true if a given FDE
406 // points to a live function.
407 template <class ELFT, class RelTy>
408 bool EhFrameSection::isFdeLive(EhSectionPiece &Fde, ArrayRef<RelTy> Rels) {
409 auto *Sec = cast<EhInputSection>(Fde.Sec);
410 unsigned FirstRelI = Fde.FirstRelocation;
412 // An FDE should point to some function because FDEs are to describe
413 // functions. That's however not always the case due to an issue of
414 // ld.gold with -r. ld.gold may discard only functions and leave their
415 // corresponding FDEs, which results in creating bad .eh_frame sections.
416 // To deal with that, we ignore such FDEs.
417 if (FirstRelI == (unsigned)-1)
420 const RelTy &Rel = Rels[FirstRelI];
421 Symbol &B = Sec->template getFile<ELFT>()->getRelocTargetSym(Rel);
423 // FDEs for garbage-collected or merged-by-ICF sections are dead.
424 if (auto *D = dyn_cast<Defined>(&B))
425 if (SectionBase *Sec = D->Section)
430 // .eh_frame is a sequence of CIE or FDE records. In general, there
431 // is one CIE record per input object file which is followed by
432 // a list of FDEs. This function searches an existing CIE or create a new
433 // one and associates FDEs to the CIE.
434 template <class ELFT, class RelTy>
435 void EhFrameSection::addSectionAux(EhInputSection *Sec, ArrayRef<RelTy> Rels) {
436 DenseMap<size_t, CieRecord *> OffsetToCie;
437 for (EhSectionPiece &Piece : Sec->Pieces) {
438 // The empty record is the end marker.
442 size_t Offset = Piece.InputOff;
443 uint32_t ID = read32(Piece.data().data() + 4, Config->Endianness);
445 OffsetToCie[Offset] = addCie<ELFT>(Piece, Rels);
449 uint32_t CieOffset = Offset + 4 - ID;
450 CieRecord *Rec = OffsetToCie[CieOffset];
452 fatal(toString(Sec) + ": invalid CIE reference");
454 if (!isFdeLive<ELFT>(Piece, Rels))
456 Rec->Fdes.push_back(&Piece);
461 template <class ELFT> void EhFrameSection::addSection(InputSectionBase *C) {
462 auto *Sec = cast<EhInputSection>(C);
465 Alignment = std::max(Alignment, Sec->Alignment);
466 Sections.push_back(Sec);
468 for (auto *DS : Sec->DependentSections)
469 DependentSections.push_back(DS);
471 // .eh_frame is a sequence of CIE or FDE records. This function
472 // splits it into pieces so that we can call
473 // SplitInputSection::getSectionPiece on the section.
475 if (Sec->Pieces.empty())
478 if (Sec->AreRelocsRela)
479 addSectionAux<ELFT>(Sec, Sec->template relas<ELFT>());
481 addSectionAux<ELFT>(Sec, Sec->template rels<ELFT>());
484 static void writeCieFde(uint8_t *Buf, ArrayRef<uint8_t> D) {
485 memcpy(Buf, D.data(), D.size());
487 size_t Aligned = alignTo(D.size(), Config->Wordsize);
489 // Zero-clear trailing padding if it exists.
490 memset(Buf + D.size(), 0, Aligned - D.size());
492 // Fix the size field. -4 since size does not include the size field itself.
493 write32(Buf, Aligned - 4);
496 void EhFrameSection::finalizeContents() {
498 return; // Already finalized.
501 for (CieRecord *Rec : CieRecords) {
502 Rec->Cie->OutputOff = Off;
503 Off += alignTo(Rec->Cie->Size, Config->Wordsize);
505 for (EhSectionPiece *Fde : Rec->Fdes) {
506 Fde->OutputOff = Off;
507 Off += alignTo(Fde->Size, Config->Wordsize);
511 // The LSB standard does not allow a .eh_frame section with zero
512 // Call Frame Information records. Therefore add a CIE record length
513 // 0 as a terminator if this .eh_frame section is empty.
520 // Returns data for .eh_frame_hdr. .eh_frame_hdr is a binary search table
521 // to get an FDE from an address to which FDE is applied. This function
522 // returns a list of such pairs.
523 std::vector<EhFrameSection::FdeData> EhFrameSection::getFdeData() const {
524 uint8_t *Buf = getParent()->Loc + OutSecOff;
525 std::vector<FdeData> Ret;
527 for (CieRecord *Rec : CieRecords) {
528 uint8_t Enc = getFdeEncoding(Rec->Cie);
529 for (EhSectionPiece *Fde : Rec->Fdes) {
530 uint32_t Pc = getFdePc(Buf, Fde->OutputOff, Enc);
531 uint32_t FdeVA = getParent()->Addr + Fde->OutputOff;
532 Ret.push_back({Pc, FdeVA});
538 static uint64_t readFdeAddr(uint8_t *Buf, int Size) {
540 case DW_EH_PE_udata2:
541 return read16(Buf, Config->Endianness);
542 case DW_EH_PE_udata4:
543 return read32(Buf, Config->Endianness);
544 case DW_EH_PE_udata8:
545 return read64(Buf, Config->Endianness);
546 case DW_EH_PE_absptr:
547 return readUint(Buf);
549 fatal("unknown FDE size encoding");
552 // Returns the VA to which a given FDE (on a mmap'ed buffer) is applied to.
553 // We need it to create .eh_frame_hdr section.
554 uint64_t EhFrameSection::getFdePc(uint8_t *Buf, size_t FdeOff,
556 // The starting address to which this FDE applies is
557 // stored at FDE + 8 byte.
558 size_t Off = FdeOff + 8;
559 uint64_t Addr = readFdeAddr(Buf + Off, Enc & 0x7);
560 if ((Enc & 0x70) == DW_EH_PE_absptr)
562 if ((Enc & 0x70) == DW_EH_PE_pcrel)
563 return Addr + getParent()->Addr + Off;
564 fatal("unknown FDE size relative encoding");
567 void EhFrameSection::writeTo(uint8_t *Buf) {
568 // Write CIE and FDE records.
569 for (CieRecord *Rec : CieRecords) {
570 size_t CieOffset = Rec->Cie->OutputOff;
571 writeCieFde(Buf + CieOffset, Rec->Cie->data());
573 for (EhSectionPiece *Fde : Rec->Fdes) {
574 size_t Off = Fde->OutputOff;
575 writeCieFde(Buf + Off, Fde->data());
577 // FDE's second word should have the offset to an associated CIE.
579 write32(Buf + Off + 4, Off + 4 - CieOffset);
583 // Apply relocations. .eh_frame section contents are not contiguous
584 // in the output buffer, but relocateAlloc() still works because
585 // getOffset() takes care of discontiguous section pieces.
586 for (EhInputSection *S : Sections)
587 S->relocateAlloc(Buf, nullptr);
590 GotSection::GotSection()
591 : SyntheticSection(SHF_ALLOC | SHF_WRITE, SHT_PROGBITS,
592 Target->GotEntrySize, ".got") {}
594 void GotSection::addEntry(Symbol &Sym) {
595 Sym.GotIndex = NumEntries;
599 bool GotSection::addDynTlsEntry(Symbol &Sym) {
600 if (Sym.GlobalDynIndex != -1U)
602 Sym.GlobalDynIndex = NumEntries;
603 // Global Dynamic TLS entries take two GOT slots.
608 // Reserves TLS entries for a TLS module ID and a TLS block offset.
609 // In total it takes two GOT slots.
610 bool GotSection::addTlsIndex() {
611 if (TlsIndexOff != uint32_t(-1))
613 TlsIndexOff = NumEntries * Config->Wordsize;
618 uint64_t GotSection::getGlobalDynAddr(const Symbol &B) const {
619 return this->getVA() + B.GlobalDynIndex * Config->Wordsize;
622 uint64_t GotSection::getGlobalDynOffset(const Symbol &B) const {
623 return B.GlobalDynIndex * Config->Wordsize;
626 void GotSection::finalizeContents() { Size = NumEntries * Config->Wordsize; }
628 bool GotSection::empty() const {
629 // We need to emit a GOT even if it's empty if there's a relocation that is
630 // relative to GOT(such as GOTOFFREL) or there's a symbol that points to a GOT
631 // (i.e. _GLOBAL_OFFSET_TABLE_).
632 return NumEntries == 0 && !HasGotOffRel && !ElfSym::GlobalOffsetTable;
635 void GotSection::writeTo(uint8_t *Buf) {
636 // Buf points to the start of this section's buffer,
637 // whereas InputSectionBase::relocateAlloc() expects its argument
638 // to point to the start of the output section.
639 relocateAlloc(Buf - OutSecOff, Buf - OutSecOff + Size);
642 MipsGotSection::MipsGotSection()
643 : SyntheticSection(SHF_ALLOC | SHF_WRITE | SHF_MIPS_GPREL, SHT_PROGBITS, 16,
646 void MipsGotSection::addEntry(Symbol &Sym, int64_t Addend, RelExpr Expr) {
647 // For "true" local symbols which can be referenced from the same module
648 // only compiler creates two instructions for address loading:
650 // lw $8, 0($gp) # R_MIPS_GOT16
651 // addi $8, $8, 0 # R_MIPS_LO16
653 // The first instruction loads high 16 bits of the symbol address while
654 // the second adds an offset. That allows to reduce number of required
655 // GOT entries because only one global offset table entry is necessary
656 // for every 64 KBytes of local data. So for local symbols we need to
657 // allocate number of GOT entries to hold all required "page" addresses.
659 // All global symbols (hidden and regular) considered by compiler uniformly.
660 // It always generates a single `lw` instruction and R_MIPS_GOT16 relocation
661 // to load address of the symbol. So for each such symbol we need to
662 // allocate dedicated GOT entry to store its address.
664 // If a symbol is preemptible we need help of dynamic linker to get its
665 // final address. The corresponding GOT entries are allocated in the
666 // "global" part of GOT. Entries for non preemptible global symbol allocated
667 // in the "local" part of GOT.
669 // See "Global Offset Table" in Chapter 5:
670 // ftp://www.linux-mips.org/pub/linux/mips/doc/ABI/mipsabi.pdf
671 if (Expr == R_MIPS_GOT_LOCAL_PAGE) {
672 // At this point we do not know final symbol value so to reduce number
673 // of allocated GOT entries do the following trick. Save all output
674 // sections referenced by GOT relocations. Then later in the `finalize`
675 // method calculate number of "pages" required to cover all saved output
676 // section and allocate appropriate number of GOT entries.
677 PageIndexMap.insert({Sym.getOutputSection(), 0});
681 // GOT entries created for MIPS TLS relocations behave like
682 // almost GOT entries from other ABIs. They go to the end
683 // of the global offset table.
684 Sym.GotIndex = TlsEntries.size();
685 TlsEntries.push_back(&Sym);
688 auto AddEntry = [&](Symbol &S, uint64_t A, GotEntries &Items) {
689 if (S.isInGot() && !A)
691 size_t NewIndex = Items.size();
692 if (!EntryIndexMap.insert({{&S, A}, NewIndex}).second)
694 Items.emplace_back(&S, A);
696 S.GotIndex = NewIndex;
698 if (Sym.IsPreemptible) {
699 // Ignore addends for preemptible symbols. They got single GOT entry anyway.
700 AddEntry(Sym, 0, GlobalEntries);
701 Sym.IsInGlobalMipsGot = true;
702 } else if (Expr == R_MIPS_GOT_OFF32) {
703 AddEntry(Sym, Addend, LocalEntries32);
704 Sym.Is32BitMipsGot = true;
706 // Hold local GOT entries accessed via a 16-bit index separately.
707 // That allows to write them in the beginning of the GOT and keep
708 // their indexes as less as possible to escape relocation's overflow.
709 AddEntry(Sym, Addend, LocalEntries);
713 bool MipsGotSection::addDynTlsEntry(Symbol &Sym) {
714 if (Sym.GlobalDynIndex != -1U)
716 Sym.GlobalDynIndex = TlsEntries.size();
717 // Global Dynamic TLS entries take two GOT slots.
718 TlsEntries.push_back(nullptr);
719 TlsEntries.push_back(&Sym);
723 // Reserves TLS entries for a TLS module ID and a TLS block offset.
724 // In total it takes two GOT slots.
725 bool MipsGotSection::addTlsIndex() {
726 if (TlsIndexOff != uint32_t(-1))
728 TlsIndexOff = TlsEntries.size() * Config->Wordsize;
729 TlsEntries.push_back(nullptr);
730 TlsEntries.push_back(nullptr);
734 static uint64_t getMipsPageAddr(uint64_t Addr) {
735 return (Addr + 0x8000) & ~0xffff;
738 static uint64_t getMipsPageCount(uint64_t Size) {
739 return (Size + 0xfffe) / 0xffff + 1;
742 uint64_t MipsGotSection::getPageEntryOffset(const Symbol &B,
743 int64_t Addend) const {
744 const OutputSection *OutSec = B.getOutputSection();
745 uint64_t SecAddr = getMipsPageAddr(OutSec->Addr);
746 uint64_t SymAddr = getMipsPageAddr(B.getVA(Addend));
747 uint64_t Index = PageIndexMap.lookup(OutSec) + (SymAddr - SecAddr) / 0xffff;
748 assert(Index < PageEntriesNum);
749 return (HeaderEntriesNum + Index) * Config->Wordsize;
752 uint64_t MipsGotSection::getSymEntryOffset(const Symbol &B,
753 int64_t Addend) const {
754 // Calculate offset of the GOT entries block: TLS, global, local.
755 uint64_t Index = HeaderEntriesNum + PageEntriesNum;
757 Index += LocalEntries.size() + LocalEntries32.size() + GlobalEntries.size();
758 else if (B.IsInGlobalMipsGot)
759 Index += LocalEntries.size() + LocalEntries32.size();
760 else if (B.Is32BitMipsGot)
761 Index += LocalEntries.size();
762 // Calculate offset of the GOT entry in the block.
766 auto It = EntryIndexMap.find({&B, Addend});
767 assert(It != EntryIndexMap.end());
770 return Index * Config->Wordsize;
773 uint64_t MipsGotSection::getTlsOffset() const {
774 return (getLocalEntriesNum() + GlobalEntries.size()) * Config->Wordsize;
777 uint64_t MipsGotSection::getGlobalDynOffset(const Symbol &B) const {
778 return B.GlobalDynIndex * Config->Wordsize;
781 const Symbol *MipsGotSection::getFirstGlobalEntry() const {
782 return GlobalEntries.empty() ? nullptr : GlobalEntries.front().first;
785 unsigned MipsGotSection::getLocalEntriesNum() const {
786 return HeaderEntriesNum + PageEntriesNum + LocalEntries.size() +
787 LocalEntries32.size();
790 void MipsGotSection::finalizeContents() { updateAllocSize(); }
792 bool MipsGotSection::updateAllocSize() {
794 for (std::pair<const OutputSection *, size_t> &P : PageIndexMap) {
795 // For each output section referenced by GOT page relocations calculate
796 // and save into PageIndexMap an upper bound of MIPS GOT entries required
797 // to store page addresses of local symbols. We assume the worst case -
798 // each 64kb page of the output section has at least one GOT relocation
799 // against it. And take in account the case when the section intersects
801 P.second = PageEntriesNum;
802 PageEntriesNum += getMipsPageCount(P.first->Size);
804 Size = (getLocalEntriesNum() + GlobalEntries.size() + TlsEntries.size()) *
809 bool MipsGotSection::empty() const {
810 // We add the .got section to the result for dynamic MIPS target because
811 // its address and properties are mentioned in the .dynamic section.
812 return Config->Relocatable;
815 uint64_t MipsGotSection::getGp() const { return ElfSym::MipsGp->getVA(0); }
817 void MipsGotSection::writeTo(uint8_t *Buf) {
818 // Set the MSB of the second GOT slot. This is not required by any
819 // MIPS ABI documentation, though.
821 // There is a comment in glibc saying that "The MSB of got[1] of a
822 // gnu object is set to identify gnu objects," and in GNU gold it
823 // says "the second entry will be used by some runtime loaders".
824 // But how this field is being used is unclear.
826 // We are not really willing to mimic other linkers behaviors
827 // without understanding why they do that, but because all files
828 // generated by GNU tools have this special GOT value, and because
829 // we've been doing this for years, it is probably a safe bet to
830 // keep doing this for now. We really need to revisit this to see
831 // if we had to do this.
832 writeUint(Buf + Config->Wordsize, (uint64_t)1 << (Config->Wordsize * 8 - 1));
833 Buf += HeaderEntriesNum * Config->Wordsize;
834 // Write 'page address' entries to the local part of the GOT.
835 for (std::pair<const OutputSection *, size_t> &L : PageIndexMap) {
836 size_t PageCount = getMipsPageCount(L.first->Size);
837 uint64_t FirstPageAddr = getMipsPageAddr(L.first->Addr);
838 for (size_t PI = 0; PI < PageCount; ++PI) {
839 uint8_t *Entry = Buf + (L.second + PI) * Config->Wordsize;
840 writeUint(Entry, FirstPageAddr + PI * 0x10000);
843 Buf += PageEntriesNum * Config->Wordsize;
844 auto AddEntry = [&](const GotEntry &SA) {
845 uint8_t *Entry = Buf;
846 Buf += Config->Wordsize;
847 const Symbol *Sym = SA.first;
848 uint64_t VA = Sym->getVA(SA.second);
849 if (Sym->StOther & STO_MIPS_MICROMIPS)
851 writeUint(Entry, VA);
853 std::for_each(std::begin(LocalEntries), std::end(LocalEntries), AddEntry);
854 std::for_each(std::begin(LocalEntries32), std::end(LocalEntries32), AddEntry);
855 std::for_each(std::begin(GlobalEntries), std::end(GlobalEntries), AddEntry);
856 // Initialize TLS-related GOT entries. If the entry has a corresponding
857 // dynamic relocations, leave it initialized by zero. Write down adjusted
858 // TLS symbol's values otherwise. To calculate the adjustments use offsets
859 // for thread-local storage.
860 // https://www.linux-mips.org/wiki/NPTL
861 if (TlsIndexOff != -1U && !Config->Pic)
862 writeUint(Buf + TlsIndexOff, 1);
863 for (const Symbol *B : TlsEntries) {
864 if (!B || B->IsPreemptible)
866 uint64_t VA = B->getVA();
867 if (B->GotIndex != -1U) {
868 uint8_t *Entry = Buf + B->GotIndex * Config->Wordsize;
869 writeUint(Entry, VA - 0x7000);
871 if (B->GlobalDynIndex != -1U) {
872 uint8_t *Entry = Buf + B->GlobalDynIndex * Config->Wordsize;
874 Entry += Config->Wordsize;
875 writeUint(Entry, VA - 0x8000);
880 GotPltSection::GotPltSection()
881 : SyntheticSection(SHF_ALLOC | SHF_WRITE, SHT_PROGBITS,
882 Target->GotPltEntrySize, ".got.plt") {}
884 void GotPltSection::addEntry(Symbol &Sym) {
885 Sym.GotPltIndex = Target->GotPltHeaderEntriesNum + Entries.size();
886 Entries.push_back(&Sym);
889 size_t GotPltSection::getSize() const {
890 return (Target->GotPltHeaderEntriesNum + Entries.size()) *
891 Target->GotPltEntrySize;
894 void GotPltSection::writeTo(uint8_t *Buf) {
895 Target->writeGotPltHeader(Buf);
896 Buf += Target->GotPltHeaderEntriesNum * Target->GotPltEntrySize;
897 for (const Symbol *B : Entries) {
898 Target->writeGotPlt(Buf, *B);
899 Buf += Config->Wordsize;
903 // On ARM the IgotPltSection is part of the GotSection, on other Targets it is
904 // part of the .got.plt
905 IgotPltSection::IgotPltSection()
906 : SyntheticSection(SHF_ALLOC | SHF_WRITE, SHT_PROGBITS,
907 Target->GotPltEntrySize,
908 Config->EMachine == EM_ARM ? ".got" : ".got.plt") {}
910 void IgotPltSection::addEntry(Symbol &Sym) {
912 Sym.GotPltIndex = Entries.size();
913 Entries.push_back(&Sym);
916 size_t IgotPltSection::getSize() const {
917 return Entries.size() * Target->GotPltEntrySize;
920 void IgotPltSection::writeTo(uint8_t *Buf) {
921 for (const Symbol *B : Entries) {
922 Target->writeIgotPlt(Buf, *B);
923 Buf += Config->Wordsize;
927 StringTableSection::StringTableSection(StringRef Name, bool Dynamic)
928 : SyntheticSection(Dynamic ? (uint64_t)SHF_ALLOC : 0, SHT_STRTAB, 1, Name),
930 // ELF string tables start with a NUL byte.
934 // Adds a string to the string table. If HashIt is true we hash and check for
935 // duplicates. It is optional because the name of global symbols are already
936 // uniqued and hashing them again has a big cost for a small value: uniquing
937 // them with some other string that happens to be the same.
938 unsigned StringTableSection::addString(StringRef S, bool HashIt) {
940 auto R = StringMap.insert(std::make_pair(S, this->Size));
942 return R.first->second;
944 unsigned Ret = this->Size;
945 this->Size = this->Size + S.size() + 1;
946 Strings.push_back(S);
950 void StringTableSection::writeTo(uint8_t *Buf) {
951 for (StringRef S : Strings) {
952 memcpy(Buf, S.data(), S.size());
953 Buf[S.size()] = '\0';
958 // Returns the number of version definition entries. Because the first entry
959 // is for the version definition itself, it is the number of versioned symbols
960 // plus one. Note that we don't support multiple versions yet.
961 static unsigned getVerDefNum() { return Config->VersionDefinitions.size() + 1; }
963 template <class ELFT>
964 DynamicSection<ELFT>::DynamicSection()
965 : SyntheticSection(SHF_ALLOC | SHF_WRITE, SHT_DYNAMIC, Config->Wordsize,
967 this->Entsize = ELFT::Is64Bits ? 16 : 8;
969 // .dynamic section is not writable on MIPS and on Fuchsia OS
970 // which passes -z rodynamic.
971 // See "Special Section" in Chapter 4 in the following document:
972 // ftp://www.linux-mips.org/pub/linux/mips/doc/ABI/mipsabi.pdf
973 if (Config->EMachine == EM_MIPS || Config->ZRodynamic)
974 this->Flags = SHF_ALLOC;
976 // Add strings to .dynstr early so that .dynstr's size will be
978 for (StringRef S : Config->FilterList)
979 addInt(DT_FILTER, InX::DynStrTab->addString(S));
980 for (StringRef S : Config->AuxiliaryList)
981 addInt(DT_AUXILIARY, InX::DynStrTab->addString(S));
983 if (!Config->Rpath.empty())
984 addInt(Config->EnableNewDtags ? DT_RUNPATH : DT_RPATH,
985 InX::DynStrTab->addString(Config->Rpath));
987 for (InputFile *File : SharedFiles) {
988 SharedFile<ELFT> *F = cast<SharedFile<ELFT>>(File);
990 addInt(DT_NEEDED, InX::DynStrTab->addString(F->SoName));
992 if (!Config->SoName.empty())
993 addInt(DT_SONAME, InX::DynStrTab->addString(Config->SoName));
996 template <class ELFT>
997 void DynamicSection<ELFT>::add(int32_t Tag, std::function<uint64_t()> Fn) {
998 Entries.push_back({Tag, Fn});
1001 template <class ELFT>
1002 void DynamicSection<ELFT>::addInt(int32_t Tag, uint64_t Val) {
1003 Entries.push_back({Tag, [=] { return Val; }});
1006 template <class ELFT>
1007 void DynamicSection<ELFT>::addInSec(int32_t Tag, InputSection *Sec) {
1009 {Tag, [=] { return Sec->getParent()->Addr + Sec->OutSecOff; }});
1012 template <class ELFT>
1013 void DynamicSection<ELFT>::addOutSec(int32_t Tag, OutputSection *Sec) {
1014 Entries.push_back({Tag, [=] { return Sec->Addr; }});
1017 template <class ELFT>
1018 void DynamicSection<ELFT>::addSize(int32_t Tag, OutputSection *Sec) {
1019 Entries.push_back({Tag, [=] { return Sec->Size; }});
1022 template <class ELFT>
1023 void DynamicSection<ELFT>::addSym(int32_t Tag, Symbol *Sym) {
1024 Entries.push_back({Tag, [=] { return Sym->getVA(); }});
1027 // Add remaining entries to complete .dynamic contents.
1028 template <class ELFT> void DynamicSection<ELFT>::finalizeContents() {
1030 return; // Already finalized.
1032 // Set DT_FLAGS and DT_FLAGS_1.
1033 uint32_t DtFlags = 0;
1034 uint32_t DtFlags1 = 0;
1035 if (Config->Bsymbolic)
1036 DtFlags |= DF_SYMBOLIC;
1037 if (Config->ZNodelete)
1038 DtFlags1 |= DF_1_NODELETE;
1039 if (Config->ZNodlopen)
1040 DtFlags1 |= DF_1_NOOPEN;
1042 DtFlags |= DF_BIND_NOW;
1043 DtFlags1 |= DF_1_NOW;
1045 if (Config->ZOrigin) {
1046 DtFlags |= DF_ORIGIN;
1047 DtFlags1 |= DF_1_ORIGIN;
1051 addInt(DT_FLAGS, DtFlags);
1053 addInt(DT_FLAGS_1, DtFlags1);
1055 // DT_DEBUG is a pointer to debug informaion used by debuggers at runtime. We
1056 // need it for each process, so we don't write it for DSOs. The loader writes
1057 // the pointer into this entry.
1059 // DT_DEBUG is the only .dynamic entry that needs to be written to. Some
1060 // systems (currently only Fuchsia OS) provide other means to give the
1061 // debugger this information. Such systems may choose make .dynamic read-only.
1062 // If the target is such a system (used -z rodynamic) don't write DT_DEBUG.
1063 if (!Config->Shared && !Config->Relocatable && !Config->ZRodynamic)
1064 addInt(DT_DEBUG, 0);
1066 this->Link = InX::DynStrTab->getParent()->SectionIndex;
1067 if (!InX::RelaDyn->empty()) {
1068 addInSec(InX::RelaDyn->DynamicTag, InX::RelaDyn);
1069 addSize(InX::RelaDyn->SizeDynamicTag, InX::RelaDyn->getParent());
1071 bool IsRela = Config->IsRela;
1072 addInt(IsRela ? DT_RELAENT : DT_RELENT,
1073 IsRela ? sizeof(Elf_Rela) : sizeof(Elf_Rel));
1075 // MIPS dynamic loader does not support RELCOUNT tag.
1076 // The problem is in the tight relation between dynamic
1077 // relocations and GOT. So do not emit this tag on MIPS.
1078 if (Config->EMachine != EM_MIPS) {
1079 size_t NumRelativeRels = InX::RelaDyn->getRelativeRelocCount();
1080 if (Config->ZCombreloc && NumRelativeRels)
1081 addInt(IsRela ? DT_RELACOUNT : DT_RELCOUNT, NumRelativeRels);
1084 // .rel[a].plt section usually consists of two parts, containing plt and
1085 // iplt relocations. It is possible to have only iplt relocations in the
1086 // output. In that case RelaPlt is empty and have zero offset, the same offset
1087 // as RelaIplt have. And we still want to emit proper dynamic tags for that
1088 // case, so here we always use RelaPlt as marker for the begining of
1089 // .rel[a].plt section.
1090 if (InX::RelaPlt->getParent()->Live) {
1091 addInSec(DT_JMPREL, InX::RelaPlt);
1092 addSize(DT_PLTRELSZ, InX::RelaPlt->getParent());
1093 switch (Config->EMachine) {
1095 addInSec(DT_MIPS_PLTGOT, InX::GotPlt);
1098 addInSec(DT_PLTGOT, InX::Plt);
1101 addInSec(DT_PLTGOT, InX::GotPlt);
1104 addInt(DT_PLTREL, Config->IsRela ? DT_RELA : DT_REL);
1107 addInSec(DT_SYMTAB, InX::DynSymTab);
1108 addInt(DT_SYMENT, sizeof(Elf_Sym));
1109 addInSec(DT_STRTAB, InX::DynStrTab);
1110 addInt(DT_STRSZ, InX::DynStrTab->getSize());
1112 addInt(DT_TEXTREL, 0);
1113 if (InX::GnuHashTab)
1114 addInSec(DT_GNU_HASH, InX::GnuHashTab);
1116 addInSec(DT_HASH, InX::HashTab);
1118 if (Out::PreinitArray) {
1119 addOutSec(DT_PREINIT_ARRAY, Out::PreinitArray);
1120 addSize(DT_PREINIT_ARRAYSZ, Out::PreinitArray);
1122 if (Out::InitArray) {
1123 addOutSec(DT_INIT_ARRAY, Out::InitArray);
1124 addSize(DT_INIT_ARRAYSZ, Out::InitArray);
1126 if (Out::FiniArray) {
1127 addOutSec(DT_FINI_ARRAY, Out::FiniArray);
1128 addSize(DT_FINI_ARRAYSZ, Out::FiniArray);
1131 if (Symbol *B = Symtab->find(Config->Init))
1134 if (Symbol *B = Symtab->find(Config->Fini))
1138 bool HasVerNeed = In<ELFT>::VerNeed->getNeedNum() != 0;
1139 if (HasVerNeed || In<ELFT>::VerDef)
1140 addInSec(DT_VERSYM, In<ELFT>::VerSym);
1141 if (In<ELFT>::VerDef) {
1142 addInSec(DT_VERDEF, In<ELFT>::VerDef);
1143 addInt(DT_VERDEFNUM, getVerDefNum());
1146 addInSec(DT_VERNEED, In<ELFT>::VerNeed);
1147 addInt(DT_VERNEEDNUM, In<ELFT>::VerNeed->getNeedNum());
1150 if (Config->EMachine == EM_MIPS) {
1151 addInt(DT_MIPS_RLD_VERSION, 1);
1152 addInt(DT_MIPS_FLAGS, RHF_NOTPOT);
1153 addInt(DT_MIPS_BASE_ADDRESS, Target->getImageBase());
1154 addInt(DT_MIPS_SYMTABNO, InX::DynSymTab->getNumSymbols());
1156 add(DT_MIPS_LOCAL_GOTNO, [] { return InX::MipsGot->getLocalEntriesNum(); });
1158 if (const Symbol *B = InX::MipsGot->getFirstGlobalEntry())
1159 addInt(DT_MIPS_GOTSYM, B->DynsymIndex);
1161 addInt(DT_MIPS_GOTSYM, InX::DynSymTab->getNumSymbols());
1162 addInSec(DT_PLTGOT, InX::MipsGot);
1163 if (InX::MipsRldMap)
1164 addInSec(DT_MIPS_RLD_MAP, InX::MipsRldMap);
1169 getParent()->Link = this->Link;
1170 this->Size = Entries.size() * this->Entsize;
1173 template <class ELFT> void DynamicSection<ELFT>::writeTo(uint8_t *Buf) {
1174 auto *P = reinterpret_cast<Elf_Dyn *>(Buf);
1176 for (std::pair<int32_t, std::function<uint64_t()>> &KV : Entries) {
1177 P->d_tag = KV.first;
1178 P->d_un.d_val = KV.second();
1183 uint64_t DynamicReloc::getOffset() const {
1184 return InputSec->getOutputSection()->Addr + InputSec->getOffset(OffsetInSec);
1187 int64_t DynamicReloc::getAddend() const {
1189 return Sym->getVA(Addend);
1193 uint32_t DynamicReloc::getSymIndex() const {
1194 if (Sym && !UseSymVA)
1195 return Sym->DynsymIndex;
1199 RelocationBaseSection::RelocationBaseSection(StringRef Name, uint32_t Type,
1201 int32_t SizeDynamicTag)
1202 : SyntheticSection(SHF_ALLOC, Type, Config->Wordsize, Name),
1203 DynamicTag(DynamicTag), SizeDynamicTag(SizeDynamicTag) {}
1205 void RelocationBaseSection::addReloc(const DynamicReloc &Reloc) {
1206 if (Reloc.Type == Target->RelativeRel)
1207 ++NumRelativeRelocs;
1208 Relocs.push_back(Reloc);
1211 void RelocationBaseSection::finalizeContents() {
1212 // If all relocations are R_*_RELATIVE they don't refer to any
1213 // dynamic symbol and we don't need a dynamic symbol table. If that
1214 // is the case, just use 0 as the link.
1215 Link = InX::DynSymTab ? InX::DynSymTab->getParent()->SectionIndex : 0;
1217 // Set required output section properties.
1218 getParent()->Link = Link;
1221 template <class ELFT>
1222 static void encodeDynamicReloc(typename ELFT::Rela *P,
1223 const DynamicReloc &Rel) {
1225 P->r_addend = Rel.getAddend();
1226 P->r_offset = Rel.getOffset();
1227 if (Config->EMachine == EM_MIPS && Rel.getInputSec() == InX::MipsGot)
1228 // The MIPS GOT section contains dynamic relocations that correspond to TLS
1229 // entries. These entries are placed after the global and local sections of
1230 // the GOT. At the point when we create these relocations, the size of the
1231 // global and local sections is unknown, so the offset that we store in the
1232 // TLS entry's DynamicReloc is relative to the start of the TLS section of
1233 // the GOT, rather than being relative to the start of the GOT. This line of
1234 // code adds the size of the global and local sections to the virtual
1235 // address computed by getOffset() in order to adjust it into the TLS
1237 P->r_offset += InX::MipsGot->getTlsOffset();
1238 P->setSymbolAndType(Rel.getSymIndex(), Rel.Type, Config->IsMips64EL);
1241 template <class ELFT>
1242 RelocationSection<ELFT>::RelocationSection(StringRef Name, bool Sort)
1243 : RelocationBaseSection(Name, Config->IsRela ? SHT_RELA : SHT_REL,
1244 Config->IsRela ? DT_RELA : DT_REL,
1245 Config->IsRela ? DT_RELASZ : DT_RELSZ),
1247 this->Entsize = Config->IsRela ? sizeof(Elf_Rela) : sizeof(Elf_Rel);
1250 template <class ELFT, class RelTy>
1251 static bool compRelocations(const RelTy &A, const RelTy &B) {
1252 bool AIsRel = A.getType(Config->IsMips64EL) == Target->RelativeRel;
1253 bool BIsRel = B.getType(Config->IsMips64EL) == Target->RelativeRel;
1254 if (AIsRel != BIsRel)
1257 return A.getSymbol(Config->IsMips64EL) < B.getSymbol(Config->IsMips64EL);
1260 template <class ELFT> void RelocationSection<ELFT>::writeTo(uint8_t *Buf) {
1261 uint8_t *BufBegin = Buf;
1262 for (const DynamicReloc &Rel : Relocs) {
1263 encodeDynamicReloc<ELFT>(reinterpret_cast<Elf_Rela *>(Buf), Rel);
1264 Buf += Config->IsRela ? sizeof(Elf_Rela) : sizeof(Elf_Rel);
1269 std::stable_sort((Elf_Rela *)BufBegin,
1270 (Elf_Rela *)BufBegin + Relocs.size(),
1271 compRelocations<ELFT, Elf_Rela>);
1273 std::stable_sort((Elf_Rel *)BufBegin, (Elf_Rel *)BufBegin + Relocs.size(),
1274 compRelocations<ELFT, Elf_Rel>);
1278 template <class ELFT> unsigned RelocationSection<ELFT>::getRelocOffset() {
1279 return this->Entsize * Relocs.size();
1282 template <class ELFT>
1283 AndroidPackedRelocationSection<ELFT>::AndroidPackedRelocationSection(
1285 : RelocationBaseSection(
1286 Name, Config->IsRela ? SHT_ANDROID_RELA : SHT_ANDROID_REL,
1287 Config->IsRela ? DT_ANDROID_RELA : DT_ANDROID_REL,
1288 Config->IsRela ? DT_ANDROID_RELASZ : DT_ANDROID_RELSZ) {
1292 template <class ELFT>
1293 bool AndroidPackedRelocationSection<ELFT>::updateAllocSize() {
1294 // This function computes the contents of an Android-format packed relocation
1297 // This format compresses relocations by using relocation groups to factor out
1298 // fields that are common between relocations and storing deltas from previous
1299 // relocations in SLEB128 format (which has a short representation for small
1300 // numbers). A good example of a relocation type with common fields is
1301 // R_*_RELATIVE, which is normally used to represent function pointers in
1302 // vtables. In the REL format, each relative relocation has the same r_info
1303 // field, and is only different from other relative relocations in terms of
1304 // the r_offset field. By sorting relocations by offset, grouping them by
1305 // r_info and representing each relocation with only the delta from the
1306 // previous offset, each 8-byte relocation can be compressed to as little as 1
1307 // byte (or less with run-length encoding). This relocation packer was able to
1308 // reduce the size of the relocation section in an Android Chromium DSO from
1309 // 2,911,184 bytes to 174,693 bytes, or 6% of the original size.
1311 // A relocation section consists of a header containing the literal bytes
1312 // 'APS2' followed by a sequence of SLEB128-encoded integers. The first two
1313 // elements are the total number of relocations in the section and an initial
1314 // r_offset value. The remaining elements define a sequence of relocation
1315 // groups. Each relocation group starts with a header consisting of the
1316 // following elements:
1318 // - the number of relocations in the relocation group
1319 // - flags for the relocation group
1320 // - (if RELOCATION_GROUPED_BY_OFFSET_DELTA_FLAG is set) the r_offset delta
1321 // for each relocation in the group.
1322 // - (if RELOCATION_GROUPED_BY_INFO_FLAG is set) the value of the r_info
1323 // field for each relocation in the group.
1324 // - (if RELOCATION_GROUP_HAS_ADDEND_FLAG and
1325 // RELOCATION_GROUPED_BY_ADDEND_FLAG are set) the r_addend delta for
1326 // each relocation in the group.
1328 // Following the relocation group header are descriptions of each of the
1329 // relocations in the group. They consist of the following elements:
1331 // - (if RELOCATION_GROUPED_BY_OFFSET_DELTA_FLAG is not set) the r_offset
1332 // delta for this relocation.
1333 // - (if RELOCATION_GROUPED_BY_INFO_FLAG is not set) the value of the r_info
1334 // field for this relocation.
1335 // - (if RELOCATION_GROUP_HAS_ADDEND_FLAG is set and
1336 // RELOCATION_GROUPED_BY_ADDEND_FLAG is not set) the r_addend delta for
1339 size_t OldSize = RelocData.size();
1341 RelocData = {'A', 'P', 'S', '2'};
1342 raw_svector_ostream OS(RelocData);
1343 auto Add = [&](int64_t V) { encodeSLEB128(V, OS); };
1345 // The format header includes the number of relocations and the initial
1346 // offset (we set this to zero because the first relocation group will
1347 // perform the initial adjustment).
1351 std::vector<Elf_Rela> Relatives, NonRelatives;
1353 for (const DynamicReloc &Rel : Relocs) {
1355 encodeDynamicReloc<ELFT>(&R, Rel);
1357 if (R.getType(Config->IsMips64EL) == Target->RelativeRel)
1358 Relatives.push_back(R);
1360 NonRelatives.push_back(R);
1363 std::sort(Relatives.begin(), Relatives.end(),
1364 [](const Elf_Rel &A, const Elf_Rel &B) {
1365 return A.r_offset < B.r_offset;
1368 // Try to find groups of relative relocations which are spaced one word
1369 // apart from one another. These generally correspond to vtable entries. The
1370 // format allows these groups to be encoded using a sort of run-length
1371 // encoding, but each group will cost 7 bytes in addition to the offset from
1372 // the previous group, so it is only profitable to do this for groups of
1373 // size 8 or larger.
1374 std::vector<Elf_Rela> UngroupedRelatives;
1375 std::vector<std::vector<Elf_Rela>> RelativeGroups;
1376 for (auto I = Relatives.begin(), E = Relatives.end(); I != E;) {
1377 std::vector<Elf_Rela> Group;
1379 Group.push_back(*I++);
1380 } while (I != E && (I - 1)->r_offset + Config->Wordsize == I->r_offset);
1382 if (Group.size() < 8)
1383 UngroupedRelatives.insert(UngroupedRelatives.end(), Group.begin(),
1386 RelativeGroups.emplace_back(std::move(Group));
1389 unsigned HasAddendIfRela =
1390 Config->IsRela ? RELOCATION_GROUP_HAS_ADDEND_FLAG : 0;
1392 uint64_t Offset = 0;
1393 uint64_t Addend = 0;
1395 // Emit the run-length encoding for the groups of adjacent relative
1396 // relocations. Each group is represented using two groups in the packed
1397 // format. The first is used to set the current offset to the start of the
1398 // group (and also encodes the first relocation), and the second encodes the
1399 // remaining relocations.
1400 for (std::vector<Elf_Rela> &G : RelativeGroups) {
1401 // The first relocation in the group.
1403 Add(RELOCATION_GROUPED_BY_OFFSET_DELTA_FLAG |
1404 RELOCATION_GROUPED_BY_INFO_FLAG | HasAddendIfRela);
1405 Add(G[0].r_offset - Offset);
1406 Add(Target->RelativeRel);
1407 if (Config->IsRela) {
1408 Add(G[0].r_addend - Addend);
1409 Addend = G[0].r_addend;
1412 // The remaining relocations.
1414 Add(RELOCATION_GROUPED_BY_OFFSET_DELTA_FLAG |
1415 RELOCATION_GROUPED_BY_INFO_FLAG | HasAddendIfRela);
1416 Add(Config->Wordsize);
1417 Add(Target->RelativeRel);
1418 if (Config->IsRela) {
1419 for (auto I = G.begin() + 1, E = G.end(); I != E; ++I) {
1420 Add(I->r_addend - Addend);
1421 Addend = I->r_addend;
1425 Offset = G.back().r_offset;
1428 // Now the ungrouped relatives.
1429 if (!UngroupedRelatives.empty()) {
1430 Add(UngroupedRelatives.size());
1431 Add(RELOCATION_GROUPED_BY_INFO_FLAG | HasAddendIfRela);
1432 Add(Target->RelativeRel);
1433 for (Elf_Rela &R : UngroupedRelatives) {
1434 Add(R.r_offset - Offset);
1435 Offset = R.r_offset;
1436 if (Config->IsRela) {
1437 Add(R.r_addend - Addend);
1438 Addend = R.r_addend;
1443 // Finally the non-relative relocations.
1444 std::sort(NonRelatives.begin(), NonRelatives.end(),
1445 [](const Elf_Rela &A, const Elf_Rela &B) {
1446 return A.r_offset < B.r_offset;
1448 if (!NonRelatives.empty()) {
1449 Add(NonRelatives.size());
1450 Add(HasAddendIfRela);
1451 for (Elf_Rela &R : NonRelatives) {
1452 Add(R.r_offset - Offset);
1453 Offset = R.r_offset;
1455 if (Config->IsRela) {
1456 Add(R.r_addend - Addend);
1457 Addend = R.r_addend;
1462 // Returns whether the section size changed. We need to keep recomputing both
1463 // section layout and the contents of this section until the size converges
1464 // because changing this section's size can affect section layout, which in
1465 // turn can affect the sizes of the LEB-encoded integers stored in this
1467 return RelocData.size() != OldSize;
1470 SymbolTableBaseSection::SymbolTableBaseSection(StringTableSection &StrTabSec)
1471 : SyntheticSection(StrTabSec.isDynamic() ? (uint64_t)SHF_ALLOC : 0,
1472 StrTabSec.isDynamic() ? SHT_DYNSYM : SHT_SYMTAB,
1474 StrTabSec.isDynamic() ? ".dynsym" : ".symtab"),
1475 StrTabSec(StrTabSec) {}
1477 // Orders symbols according to their positions in the GOT,
1478 // in compliance with MIPS ABI rules.
1479 // See "Global Offset Table" in Chapter 5 in the following document
1480 // for detailed description:
1481 // ftp://www.linux-mips.org/pub/linux/mips/doc/ABI/mipsabi.pdf
1482 static bool sortMipsSymbols(const SymbolTableEntry &L,
1483 const SymbolTableEntry &R) {
1484 // Sort entries related to non-local preemptible symbols by GOT indexes.
1485 // All other entries go to the first part of GOT in arbitrary order.
1486 bool LIsInLocalGot = !L.Sym->IsInGlobalMipsGot;
1487 bool RIsInLocalGot = !R.Sym->IsInGlobalMipsGot;
1488 if (LIsInLocalGot || RIsInLocalGot)
1489 return !RIsInLocalGot;
1490 return L.Sym->GotIndex < R.Sym->GotIndex;
1493 void SymbolTableBaseSection::finalizeContents() {
1494 getParent()->Link = StrTabSec.getParent()->SectionIndex;
1496 // If it is a .dynsym, there should be no local symbols, but we need
1497 // to do a few things for the dynamic linker.
1498 if (this->Type == SHT_DYNSYM) {
1499 // Section's Info field has the index of the first non-local symbol.
1500 // Because the first symbol entry is a null entry, 1 is the first.
1501 getParent()->Info = 1;
1503 if (InX::GnuHashTab) {
1504 // NB: It also sorts Symbols to meet the GNU hash table requirements.
1505 InX::GnuHashTab->addSymbols(Symbols);
1506 } else if (Config->EMachine == EM_MIPS) {
1507 std::stable_sort(Symbols.begin(), Symbols.end(), sortMipsSymbols);
1511 for (const SymbolTableEntry &S : Symbols) S.Sym->DynsymIndex = ++I;
1516 // The ELF spec requires that all local symbols precede global symbols, so we
1517 // sort symbol entries in this function. (For .dynsym, we don't do that because
1518 // symbols for dynamic linking are inherently all globals.)
1519 void SymbolTableBaseSection::postThunkContents() {
1520 if (this->Type == SHT_DYNSYM)
1522 // move all local symbols before global symbols.
1523 auto It = std::stable_partition(
1524 Symbols.begin(), Symbols.end(), [](const SymbolTableEntry &S) {
1525 return S.Sym->isLocal() || S.Sym->computeBinding() == STB_LOCAL;
1527 size_t NumLocals = It - Symbols.begin();
1528 getParent()->Info = NumLocals + 1;
1531 void SymbolTableBaseSection::addSymbol(Symbol *B) {
1532 // Adding a local symbol to a .dynsym is a bug.
1533 assert(this->Type != SHT_DYNSYM || !B->isLocal());
1535 bool HashIt = B->isLocal();
1536 Symbols.push_back({B, StrTabSec.addString(B->getName(), HashIt)});
1539 size_t SymbolTableBaseSection::getSymbolIndex(Symbol *Sym) {
1540 // Initializes symbol lookup tables lazily. This is used only
1541 // for -r or -emit-relocs.
1542 llvm::call_once(OnceFlag, [&] {
1543 SymbolIndexMap.reserve(Symbols.size());
1545 for (const SymbolTableEntry &E : Symbols) {
1546 if (E.Sym->Type == STT_SECTION)
1547 SectionIndexMap[E.Sym->getOutputSection()] = ++I;
1549 SymbolIndexMap[E.Sym] = ++I;
1553 // Section symbols are mapped based on their output sections
1554 // to maintain their semantics.
1555 if (Sym->Type == STT_SECTION)
1556 return SectionIndexMap.lookup(Sym->getOutputSection());
1557 return SymbolIndexMap.lookup(Sym);
1560 template <class ELFT>
1561 SymbolTableSection<ELFT>::SymbolTableSection(StringTableSection &StrTabSec)
1562 : SymbolTableBaseSection(StrTabSec) {
1563 this->Entsize = sizeof(Elf_Sym);
1566 // Write the internal symbol table contents to the output symbol table.
1567 template <class ELFT> void SymbolTableSection<ELFT>::writeTo(uint8_t *Buf) {
1568 // The first entry is a null entry as per the ELF spec.
1569 memset(Buf, 0, sizeof(Elf_Sym));
1570 Buf += sizeof(Elf_Sym);
1572 auto *ESym = reinterpret_cast<Elf_Sym *>(Buf);
1574 for (SymbolTableEntry &Ent : Symbols) {
1575 Symbol *Sym = Ent.Sym;
1577 // Set st_info and st_other.
1579 if (Sym->isLocal()) {
1580 ESym->setBindingAndType(STB_LOCAL, Sym->Type);
1582 ESym->setBindingAndType(Sym->computeBinding(), Sym->Type);
1583 ESym->setVisibility(Sym->Visibility);
1586 ESym->st_name = Ent.StrTabOffset;
1588 // Set a section index.
1589 BssSection *CommonSec = nullptr;
1590 if (!Config->DefineCommon)
1591 if (auto *D = dyn_cast<Defined>(Sym))
1592 CommonSec = dyn_cast_or_null<BssSection>(D->Section);
1594 ESym->st_shndx = SHN_COMMON;
1595 else if (const OutputSection *OutSec = Sym->getOutputSection())
1596 ESym->st_shndx = OutSec->SectionIndex;
1597 else if (isa<Defined>(Sym))
1598 ESym->st_shndx = SHN_ABS;
1600 ESym->st_shndx = SHN_UNDEF;
1602 // Copy symbol size if it is a defined symbol. st_size is not significant
1603 // for undefined symbols, so whether copying it or not is up to us if that's
1604 // the case. We'll leave it as zero because by not setting a value, we can
1605 // get the exact same outputs for two sets of input files that differ only
1606 // in undefined symbol size in DSOs.
1607 if (ESym->st_shndx == SHN_UNDEF)
1610 ESym->st_size = Sym->getSize();
1612 // st_value is usually an address of a symbol, but that has a
1613 // special meaining for uninstantiated common symbols (this can
1614 // occur if -r is given).
1616 ESym->st_value = CommonSec->Alignment;
1618 ESym->st_value = Sym->getVA();
1623 // On MIPS we need to mark symbol which has a PLT entry and requires
1624 // pointer equality by STO_MIPS_PLT flag. That is necessary to help
1625 // dynamic linker distinguish such symbols and MIPS lazy-binding stubs.
1626 // https://sourceware.org/ml/binutils/2008-07/txt00000.txt
1627 if (Config->EMachine == EM_MIPS) {
1628 auto *ESym = reinterpret_cast<Elf_Sym *>(Buf);
1630 for (SymbolTableEntry &Ent : Symbols) {
1631 Symbol *Sym = Ent.Sym;
1632 if (Sym->isInPlt() && Sym->NeedsPltAddr)
1633 ESym->st_other |= STO_MIPS_PLT;
1634 if (isMicroMips()) {
1635 // Set STO_MIPS_MICROMIPS flag and less-significant bit for
1636 // defined microMIPS symbols and shared symbols with PLT record.
1637 if ((Sym->isDefined() && (Sym->StOther & STO_MIPS_MICROMIPS)) ||
1638 (Sym->isShared() && Sym->NeedsPltAddr)) {
1639 if (StrTabSec.isDynamic())
1640 ESym->st_value |= 1;
1641 ESym->st_other |= STO_MIPS_MICROMIPS;
1644 if (Config->Relocatable)
1645 if (auto *D = dyn_cast<Defined>(Sym))
1646 if (isMipsPIC<ELFT>(D))
1647 ESym->st_other |= STO_MIPS_PIC;
1653 // .hash and .gnu.hash sections contain on-disk hash tables that map
1654 // symbol names to their dynamic symbol table indices. Their purpose
1655 // is to help the dynamic linker resolve symbols quickly. If ELF files
1656 // don't have them, the dynamic linker has to do linear search on all
1657 // dynamic symbols, which makes programs slower. Therefore, a .hash
1658 // section is added to a DSO by default. A .gnu.hash is added if you
1659 // give the -hash-style=gnu or -hash-style=both option.
1661 // The Unix semantics of resolving dynamic symbols is somewhat expensive.
1662 // Each ELF file has a list of DSOs that the ELF file depends on and a
1663 // list of dynamic symbols that need to be resolved from any of the
1664 // DSOs. That means resolving all dynamic symbols takes O(m)*O(n)
1665 // where m is the number of DSOs and n is the number of dynamic
1666 // symbols. For modern large programs, both m and n are large. So
1667 // making each step faster by using hash tables substiantially
1668 // improves time to load programs.
1670 // (Note that this is not the only way to design the shared library.
1671 // For instance, the Windows DLL takes a different approach. On
1672 // Windows, each dynamic symbol has a name of DLL from which the symbol
1673 // has to be resolved. That makes the cost of symbol resolution O(n).
1674 // This disables some hacky techniques you can use on Unix such as
1675 // LD_PRELOAD, but this is arguably better semantics than the Unix ones.)
1677 // Due to historical reasons, we have two different hash tables, .hash
1678 // and .gnu.hash. They are for the same purpose, and .gnu.hash is a new
1679 // and better version of .hash. .hash is just an on-disk hash table, but
1680 // .gnu.hash has a bloom filter in addition to a hash table to skip
1681 // DSOs very quickly. If you are sure that your dynamic linker knows
1682 // about .gnu.hash, you want to specify -hash-style=gnu. Otherwise, a
1683 // safe bet is to specify -hash-style=both for backward compatibilty.
1684 GnuHashTableSection::GnuHashTableSection()
1685 : SyntheticSection(SHF_ALLOC, SHT_GNU_HASH, Config->Wordsize, ".gnu.hash") {
1688 void GnuHashTableSection::finalizeContents() {
1689 getParent()->Link = InX::DynSymTab->getParent()->SectionIndex;
1691 // Computes bloom filter size in word size. We want to allocate 8
1692 // bits for each symbol. It must be a power of two.
1693 if (Symbols.empty())
1696 MaskWords = NextPowerOf2((Symbols.size() - 1) / Config->Wordsize);
1698 Size = 16; // Header
1699 Size += Config->Wordsize * MaskWords; // Bloom filter
1700 Size += NBuckets * 4; // Hash buckets
1701 Size += Symbols.size() * 4; // Hash values
1704 void GnuHashTableSection::writeTo(uint8_t *Buf) {
1705 // The output buffer is not guaranteed to be zero-cleared because we pre-
1706 // fill executable sections with trap instructions. This is a precaution
1707 // for that case, which happens only when -no-rosegment is given.
1708 memset(Buf, 0, Size);
1711 write32(Buf, NBuckets);
1712 write32(Buf + 4, InX::DynSymTab->getNumSymbols() - Symbols.size());
1713 write32(Buf + 8, MaskWords);
1714 write32(Buf + 12, getShift2());
1717 // Write a bloom filter and a hash table.
1718 writeBloomFilter(Buf);
1719 Buf += Config->Wordsize * MaskWords;
1720 writeHashTable(Buf);
1723 // This function writes a 2-bit bloom filter. This bloom filter alone
1724 // usually filters out 80% or more of all symbol lookups [1].
1725 // The dynamic linker uses the hash table only when a symbol is not
1726 // filtered out by a bloom filter.
1728 // [1] Ulrich Drepper (2011), "How To Write Shared Libraries" (Ver. 4.1.2),
1729 // p.9, https://www.akkadia.org/drepper/dsohowto.pdf
1730 void GnuHashTableSection::writeBloomFilter(uint8_t *Buf) {
1731 const unsigned C = Config->Wordsize * 8;
1732 for (const Entry &Sym : Symbols) {
1733 size_t I = (Sym.Hash / C) & (MaskWords - 1);
1734 uint64_t Val = readUint(Buf + I * Config->Wordsize);
1735 Val |= uint64_t(1) << (Sym.Hash % C);
1736 Val |= uint64_t(1) << ((Sym.Hash >> getShift2()) % C);
1737 writeUint(Buf + I * Config->Wordsize, Val);
1741 void GnuHashTableSection::writeHashTable(uint8_t *Buf) {
1742 uint32_t *Buckets = reinterpret_cast<uint32_t *>(Buf);
1743 uint32_t OldBucket = -1;
1744 uint32_t *Values = Buckets + NBuckets;
1745 for (auto I = Symbols.begin(), E = Symbols.end(); I != E; ++I) {
1746 // Write a hash value. It represents a sequence of chains that share the
1747 // same hash modulo value. The last element of each chain is terminated by
1749 uint32_t Hash = I->Hash;
1750 bool IsLastInChain = (I + 1) == E || I->BucketIdx != (I + 1)->BucketIdx;
1751 Hash = IsLastInChain ? Hash | 1 : Hash & ~1;
1752 write32(Values++, Hash);
1754 if (I->BucketIdx == OldBucket)
1756 // Write a hash bucket. Hash buckets contain indices in the following hash
1758 write32(Buckets + I->BucketIdx, I->Sym->DynsymIndex);
1759 OldBucket = I->BucketIdx;
1763 static uint32_t hashGnu(StringRef Name) {
1765 for (uint8_t C : Name)
1766 H = (H << 5) + H + C;
1770 // Add symbols to this symbol hash table. Note that this function
1771 // destructively sort a given vector -- which is needed because
1772 // GNU-style hash table places some sorting requirements.
1773 void GnuHashTableSection::addSymbols(std::vector<SymbolTableEntry> &V) {
1774 // We cannot use 'auto' for Mid because GCC 6.1 cannot deduce
1775 // its type correctly.
1776 std::vector<SymbolTableEntry>::iterator Mid =
1777 std::stable_partition(V.begin(), V.end(), [](const SymbolTableEntry &S) {
1778 // Shared symbols that this executable preempts are special. The dynamic
1779 // linker has to look them up, so they have to be in the hash table.
1780 if (auto *SS = dyn_cast<SharedSymbol>(S.Sym))
1781 return SS->CopyRelSec == nullptr && !SS->NeedsPltAddr;
1782 return !S.Sym->isDefined();
1787 // We chose load factor 4 for the on-disk hash table. For each hash
1788 // collision, the dynamic linker will compare a uint32_t hash value.
1789 // Since the integer comparison is quite fast, we believe we can make
1790 // the load factor even larger. 4 is just a conservative choice.
1791 NBuckets = std::max<size_t>((V.end() - Mid) / 4, 1);
1793 for (SymbolTableEntry &Ent : llvm::make_range(Mid, V.end())) {
1794 Symbol *B = Ent.Sym;
1795 uint32_t Hash = hashGnu(B->getName());
1796 uint32_t BucketIdx = Hash % NBuckets;
1797 Symbols.push_back({B, Ent.StrTabOffset, Hash, BucketIdx});
1801 Symbols.begin(), Symbols.end(),
1802 [](const Entry &L, const Entry &R) { return L.BucketIdx < R.BucketIdx; });
1804 V.erase(Mid, V.end());
1805 for (const Entry &Ent : Symbols)
1806 V.push_back({Ent.Sym, Ent.StrTabOffset});
1809 HashTableSection::HashTableSection()
1810 : SyntheticSection(SHF_ALLOC, SHT_HASH, 4, ".hash") {
1814 void HashTableSection::finalizeContents() {
1815 getParent()->Link = InX::DynSymTab->getParent()->SectionIndex;
1817 unsigned NumEntries = 2; // nbucket and nchain.
1818 NumEntries += InX::DynSymTab->getNumSymbols(); // The chain entries.
1820 // Create as many buckets as there are symbols.
1821 NumEntries += InX::DynSymTab->getNumSymbols();
1822 this->Size = NumEntries * 4;
1825 void HashTableSection::writeTo(uint8_t *Buf) {
1826 // See comment in GnuHashTableSection::writeTo.
1827 memset(Buf, 0, Size);
1829 unsigned NumSymbols = InX::DynSymTab->getNumSymbols();
1831 uint32_t *P = reinterpret_cast<uint32_t *>(Buf);
1832 write32(P++, NumSymbols); // nbucket
1833 write32(P++, NumSymbols); // nchain
1835 uint32_t *Buckets = P;
1836 uint32_t *Chains = P + NumSymbols;
1838 for (const SymbolTableEntry &S : InX::DynSymTab->getSymbols()) {
1839 Symbol *Sym = S.Sym;
1840 StringRef Name = Sym->getName();
1841 unsigned I = Sym->DynsymIndex;
1842 uint32_t Hash = hashSysV(Name) % NumSymbols;
1843 Chains[I] = Buckets[Hash];
1844 write32(Buckets + Hash, I);
1848 PltSection::PltSection(size_t S)
1849 : SyntheticSection(SHF_ALLOC | SHF_EXECINSTR, SHT_PROGBITS, 16, ".plt"),
1851 // The PLT needs to be writable on SPARC as the dynamic linker will
1852 // modify the instructions in the PLT entries.
1853 if (Config->EMachine == EM_SPARCV9)
1854 this->Flags |= SHF_WRITE;
1857 void PltSection::writeTo(uint8_t *Buf) {
1858 // At beginning of PLT but not the IPLT, we have code to call the dynamic
1859 // linker to resolve dynsyms at runtime. Write such code.
1860 if (HeaderSize != 0)
1861 Target->writePltHeader(Buf);
1862 size_t Off = HeaderSize;
1863 // The IPlt is immediately after the Plt, account for this in RelOff
1864 unsigned PltOff = getPltRelocOff();
1866 for (auto &I : Entries) {
1867 const Symbol *B = I.first;
1868 unsigned RelOff = I.second + PltOff;
1869 uint64_t Got = B->getGotPltVA();
1870 uint64_t Plt = this->getVA() + Off;
1871 Target->writePlt(Buf + Off, Got, Plt, B->PltIndex, RelOff);
1872 Off += Target->PltEntrySize;
1876 template <class ELFT> void PltSection::addEntry(Symbol &Sym) {
1877 Sym.PltIndex = Entries.size();
1878 RelocationBaseSection *PltRelocSection = InX::RelaPlt;
1879 if (HeaderSize == 0) {
1880 PltRelocSection = InX::RelaIplt;
1881 Sym.IsInIplt = true;
1884 static_cast<RelocationSection<ELFT> *>(PltRelocSection)->getRelocOffset();
1885 Entries.push_back(std::make_pair(&Sym, RelOff));
1888 size_t PltSection::getSize() const {
1889 return HeaderSize + Entries.size() * Target->PltEntrySize;
1892 // Some architectures such as additional symbols in the PLT section. For
1893 // example ARM uses mapping symbols to aid disassembly
1894 void PltSection::addSymbols() {
1895 // The PLT may have symbols defined for the Header, the IPLT has no header
1896 if (HeaderSize != 0)
1897 Target->addPltHeaderSymbols(*this);
1898 size_t Off = HeaderSize;
1899 for (size_t I = 0; I < Entries.size(); ++I) {
1900 Target->addPltSymbols(*this, Off);
1901 Off += Target->PltEntrySize;
1905 unsigned PltSection::getPltRelocOff() const {
1906 return (HeaderSize == 0) ? InX::Plt->getSize() : 0;
1909 // The string hash function for .gdb_index.
1910 static uint32_t computeGdbHash(StringRef S) {
1913 H = H * 67 + tolower(C) - 113;
1917 static std::vector<GdbIndexChunk::CuEntry> readCuList(DWARFContext &Dwarf) {
1918 std::vector<GdbIndexChunk::CuEntry> Ret;
1919 for (std::unique_ptr<DWARFCompileUnit> &Cu : Dwarf.compile_units())
1920 Ret.push_back({Cu->getOffset(), Cu->getLength() + 4});
1924 static std::vector<GdbIndexChunk::AddressEntry>
1925 readAddressAreas(DWARFContext &Dwarf, InputSection *Sec) {
1926 std::vector<GdbIndexChunk::AddressEntry> Ret;
1929 for (std::unique_ptr<DWARFCompileUnit> &Cu : Dwarf.compile_units()) {
1930 DWARFAddressRangesVector Ranges;
1931 Cu->collectAddressRanges(Ranges);
1933 ArrayRef<InputSectionBase *> Sections = Sec->File->getSections();
1934 for (DWARFAddressRange &R : Ranges) {
1935 InputSectionBase *S = Sections[R.SectionIndex];
1936 if (!S || S == &InputSection::Discarded || !S->Live)
1938 // Range list with zero size has no effect.
1939 if (R.LowPC == R.HighPC)
1941 auto *IS = cast<InputSection>(S);
1942 uint64_t Offset = IS->getOffsetInFile();
1943 Ret.push_back({IS, R.LowPC - Offset, R.HighPC - Offset, CuIdx});
1950 static std::vector<GdbIndexChunk::NameTypeEntry>
1951 readPubNamesAndTypes(DWARFContext &Dwarf) {
1952 StringRef Sec1 = Dwarf.getDWARFObj().getGnuPubNamesSection();
1953 StringRef Sec2 = Dwarf.getDWARFObj().getGnuPubTypesSection();
1955 std::vector<GdbIndexChunk::NameTypeEntry> Ret;
1956 for (StringRef Sec : {Sec1, Sec2}) {
1957 DWARFDebugPubTable Table(Sec, Config->IsLE, true);
1958 for (const DWARFDebugPubTable::Set &Set : Table.getData()) {
1959 for (const DWARFDebugPubTable::Entry &Ent : Set.Entries) {
1960 CachedHashStringRef S(Ent.Name, computeGdbHash(Ent.Name));
1961 Ret.push_back({S, Ent.Descriptor.toBits()});
1968 static std::vector<InputSection *> getDebugInfoSections() {
1969 std::vector<InputSection *> Ret;
1970 for (InputSectionBase *S : InputSections)
1971 if (InputSection *IS = dyn_cast<InputSection>(S))
1972 if (IS->Name == ".debug_info")
1977 void GdbIndexSection::fixCuIndex() {
1979 for (GdbIndexChunk &Chunk : Chunks) {
1980 for (GdbIndexChunk::AddressEntry &Ent : Chunk.AddressAreas)
1982 Idx += Chunk.CompilationUnits.size();
1986 std::vector<std::vector<uint32_t>> GdbIndexSection::createCuVectors() {
1987 std::vector<std::vector<uint32_t>> Ret;
1991 for (GdbIndexChunk &Chunk : Chunks) {
1992 for (GdbIndexChunk::NameTypeEntry &Ent : Chunk.NamesAndTypes) {
1993 GdbSymbol *&Sym = Symbols[Ent.Name];
1995 Sym = make<GdbSymbol>(GdbSymbol{Ent.Name.hash(), Off, Ret.size()});
1996 Off += Ent.Name.size() + 1;
2000 // gcc 5.4.1 produces a buggy .debug_gnu_pubnames that contains
2001 // duplicate entries, so we want to dedup them.
2002 std::vector<uint32_t> &Vec = Ret[Sym->CuVectorIndex];
2003 uint32_t Val = (Ent.Type << 24) | Idx;
2004 if (Vec.empty() || Vec.back() != Val)
2007 Idx += Chunk.CompilationUnits.size();
2010 StringPoolSize = Off;
2014 template <class ELFT> GdbIndexSection *elf::createGdbIndex() {
2015 // Gather debug info to create a .gdb_index section.
2016 std::vector<InputSection *> Sections = getDebugInfoSections();
2017 std::vector<GdbIndexChunk> Chunks(Sections.size());
2019 parallelForEachN(0, Chunks.size(), [&](size_t I) {
2020 ObjFile<ELFT> *File = Sections[I]->getFile<ELFT>();
2021 DWARFContext Dwarf(make_unique<LLDDwarfObj<ELFT>>(File));
2023 Chunks[I].DebugInfoSec = Sections[I];
2024 Chunks[I].CompilationUnits = readCuList(Dwarf);
2025 Chunks[I].AddressAreas = readAddressAreas(Dwarf, Sections[I]);
2026 Chunks[I].NamesAndTypes = readPubNamesAndTypes(Dwarf);
2029 // .debug_gnu_pub{names,types} are useless in executables.
2030 // They are present in input object files solely for creating
2031 // a .gdb_index. So we can remove it from the output.
2032 for (InputSectionBase *S : InputSections)
2033 if (S->Name == ".debug_gnu_pubnames" || S->Name == ".debug_gnu_pubtypes")
2036 // Create a .gdb_index and returns it.
2037 return make<GdbIndexSection>(std::move(Chunks));
2040 static size_t getCuSize(ArrayRef<GdbIndexChunk> Arr) {
2042 for (const GdbIndexChunk &D : Arr)
2043 Ret += D.CompilationUnits.size();
2047 static size_t getAddressAreaSize(ArrayRef<GdbIndexChunk> Arr) {
2049 for (const GdbIndexChunk &D : Arr)
2050 Ret += D.AddressAreas.size();
2054 std::vector<GdbSymbol *> GdbIndexSection::createGdbSymtab() {
2055 uint32_t Size = NextPowerOf2(Symbols.size() * 4 / 3);
2059 uint32_t Mask = Size - 1;
2060 std::vector<GdbSymbol *> Ret(Size);
2062 for (auto &KV : Symbols) {
2063 GdbSymbol *Sym = KV.second;
2064 uint32_t I = Sym->NameHash & Mask;
2065 uint32_t Step = ((Sym->NameHash * 17) & Mask) | 1;
2068 I = (I + Step) & Mask;
2074 GdbIndexSection::GdbIndexSection(std::vector<GdbIndexChunk> &&C)
2075 : SyntheticSection(0, SHT_PROGBITS, 1, ".gdb_index"), Chunks(std::move(C)) {
2077 CuVectors = createCuVectors();
2078 GdbSymtab = createGdbSymtab();
2080 // Compute offsets early to know the section size.
2081 // Each chunk size needs to be in sync with what we write in writeTo.
2082 CuTypesOffset = CuListOffset + getCuSize(Chunks) * 16;
2083 SymtabOffset = CuTypesOffset + getAddressAreaSize(Chunks) * 20;
2084 ConstantPoolOffset = SymtabOffset + GdbSymtab.size() * 8;
2087 for (ArrayRef<uint32_t> Vec : CuVectors) {
2088 CuVectorOffsets.push_back(Off);
2089 Off += (Vec.size() + 1) * 4;
2091 StringPoolOffset = ConstantPoolOffset + Off;
2094 size_t GdbIndexSection::getSize() const {
2095 return StringPoolOffset + StringPoolSize;
2098 void GdbIndexSection::writeTo(uint8_t *Buf) {
2099 // Write the section header.
2101 write32le(Buf + 4, CuListOffset);
2102 write32le(Buf + 8, CuTypesOffset);
2103 write32le(Buf + 12, CuTypesOffset);
2104 write32le(Buf + 16, SymtabOffset);
2105 write32le(Buf + 20, ConstantPoolOffset);
2108 // Write the CU list.
2109 for (GdbIndexChunk &D : Chunks) {
2110 for (GdbIndexChunk::CuEntry &Cu : D.CompilationUnits) {
2111 write64le(Buf, D.DebugInfoSec->OutSecOff + Cu.CuOffset);
2112 write64le(Buf + 8, Cu.CuLength);
2117 // Write the address area.
2118 for (GdbIndexChunk &D : Chunks) {
2119 for (GdbIndexChunk::AddressEntry &E : D.AddressAreas) {
2121 E.Section->getParent()->Addr + E.Section->getOffset(0);
2122 write64le(Buf, BaseAddr + E.LowAddress);
2123 write64le(Buf + 8, BaseAddr + E.HighAddress);
2124 write32le(Buf + 16, E.CuIndex);
2129 // Write the symbol table.
2130 for (GdbSymbol *Sym : GdbSymtab) {
2132 write32le(Buf, Sym->NameOffset + StringPoolOffset - ConstantPoolOffset);
2133 write32le(Buf + 4, CuVectorOffsets[Sym->CuVectorIndex]);
2138 // Write the CU vectors.
2139 for (ArrayRef<uint32_t> Vec : CuVectors) {
2140 write32le(Buf, Vec.size());
2142 for (uint32_t Val : Vec) {
2143 write32le(Buf, Val);
2148 // Write the string pool.
2149 for (auto &KV : Symbols) {
2150 CachedHashStringRef S = KV.first;
2151 GdbSymbol *Sym = KV.second;
2152 size_t Off = Sym->NameOffset;
2153 memcpy(Buf + Off, S.val().data(), S.size());
2154 Buf[Off + S.size()] = '\0';
2158 bool GdbIndexSection::empty() const { return !Out::DebugInfo; }
2160 EhFrameHeader::EhFrameHeader()
2161 : SyntheticSection(SHF_ALLOC, SHT_PROGBITS, 1, ".eh_frame_hdr") {}
2163 // .eh_frame_hdr contains a binary search table of pointers to FDEs.
2164 // Each entry of the search table consists of two values,
2165 // the starting PC from where FDEs covers, and the FDE's address.
2166 // It is sorted by PC.
2167 void EhFrameHeader::writeTo(uint8_t *Buf) {
2168 typedef EhFrameSection::FdeData FdeData;
2170 std::vector<FdeData> Fdes = InX::EhFrame->getFdeData();
2172 // Sort the FDE list by their PC and uniqueify. Usually there is only
2173 // one FDE for a PC (i.e. function), but if ICF merges two functions
2174 // into one, there can be more than one FDEs pointing to the address.
2175 auto Less = [](const FdeData &A, const FdeData &B) { return A.Pc < B.Pc; };
2176 std::stable_sort(Fdes.begin(), Fdes.end(), Less);
2177 auto Eq = [](const FdeData &A, const FdeData &B) { return A.Pc == B.Pc; };
2178 Fdes.erase(std::unique(Fdes.begin(), Fdes.end(), Eq), Fdes.end());
2181 Buf[1] = DW_EH_PE_pcrel | DW_EH_PE_sdata4;
2182 Buf[2] = DW_EH_PE_udata4;
2183 Buf[3] = DW_EH_PE_datarel | DW_EH_PE_sdata4;
2184 write32(Buf + 4, InX::EhFrame->getParent()->Addr - this->getVA() - 4);
2185 write32(Buf + 8, Fdes.size());
2188 uint64_t VA = this->getVA();
2189 for (FdeData &Fde : Fdes) {
2190 write32(Buf, Fde.Pc - VA);
2191 write32(Buf + 4, Fde.FdeVA - VA);
2196 size_t EhFrameHeader::getSize() const {
2197 // .eh_frame_hdr has a 12 bytes header followed by an array of FDEs.
2198 return 12 + InX::EhFrame->NumFdes * 8;
2201 bool EhFrameHeader::empty() const { return InX::EhFrame->empty(); }
2203 template <class ELFT>
2204 VersionDefinitionSection<ELFT>::VersionDefinitionSection()
2205 : SyntheticSection(SHF_ALLOC, SHT_GNU_verdef, sizeof(uint32_t),
2206 ".gnu.version_d") {}
2208 static StringRef getFileDefName() {
2209 if (!Config->SoName.empty())
2210 return Config->SoName;
2211 return Config->OutputFile;
2214 template <class ELFT> void VersionDefinitionSection<ELFT>::finalizeContents() {
2215 FileDefNameOff = InX::DynStrTab->addString(getFileDefName());
2216 for (VersionDefinition &V : Config->VersionDefinitions)
2217 V.NameOff = InX::DynStrTab->addString(V.Name);
2219 getParent()->Link = InX::DynStrTab->getParent()->SectionIndex;
2221 // sh_info should be set to the number of definitions. This fact is missed in
2222 // documentation, but confirmed by binutils community:
2223 // https://sourceware.org/ml/binutils/2014-11/msg00355.html
2224 getParent()->Info = getVerDefNum();
2227 template <class ELFT>
2228 void VersionDefinitionSection<ELFT>::writeOne(uint8_t *Buf, uint32_t Index,
2229 StringRef Name, size_t NameOff) {
2230 auto *Verdef = reinterpret_cast<Elf_Verdef *>(Buf);
2231 Verdef->vd_version = 1;
2233 Verdef->vd_aux = sizeof(Elf_Verdef);
2234 Verdef->vd_next = sizeof(Elf_Verdef) + sizeof(Elf_Verdaux);
2235 Verdef->vd_flags = (Index == 1 ? VER_FLG_BASE : 0);
2236 Verdef->vd_ndx = Index;
2237 Verdef->vd_hash = hashSysV(Name);
2239 auto *Verdaux = reinterpret_cast<Elf_Verdaux *>(Buf + sizeof(Elf_Verdef));
2240 Verdaux->vda_name = NameOff;
2241 Verdaux->vda_next = 0;
2244 template <class ELFT>
2245 void VersionDefinitionSection<ELFT>::writeTo(uint8_t *Buf) {
2246 writeOne(Buf, 1, getFileDefName(), FileDefNameOff);
2248 for (VersionDefinition &V : Config->VersionDefinitions) {
2249 Buf += sizeof(Elf_Verdef) + sizeof(Elf_Verdaux);
2250 writeOne(Buf, V.Id, V.Name, V.NameOff);
2253 // Need to terminate the last version definition.
2254 Elf_Verdef *Verdef = reinterpret_cast<Elf_Verdef *>(Buf);
2255 Verdef->vd_next = 0;
2258 template <class ELFT> size_t VersionDefinitionSection<ELFT>::getSize() const {
2259 return (sizeof(Elf_Verdef) + sizeof(Elf_Verdaux)) * getVerDefNum();
2262 template <class ELFT>
2263 VersionTableSection<ELFT>::VersionTableSection()
2264 : SyntheticSection(SHF_ALLOC, SHT_GNU_versym, sizeof(uint16_t),
2266 this->Entsize = sizeof(Elf_Versym);
2269 template <class ELFT> void VersionTableSection<ELFT>::finalizeContents() {
2270 // At the moment of june 2016 GNU docs does not mention that sh_link field
2271 // should be set, but Sun docs do. Also readelf relies on this field.
2272 getParent()->Link = InX::DynSymTab->getParent()->SectionIndex;
2275 template <class ELFT> size_t VersionTableSection<ELFT>::getSize() const {
2276 return sizeof(Elf_Versym) * (InX::DynSymTab->getSymbols().size() + 1);
2279 template <class ELFT> void VersionTableSection<ELFT>::writeTo(uint8_t *Buf) {
2280 auto *OutVersym = reinterpret_cast<Elf_Versym *>(Buf) + 1;
2281 for (const SymbolTableEntry &S : InX::DynSymTab->getSymbols()) {
2282 OutVersym->vs_index = S.Sym->VersionId;
2287 template <class ELFT> bool VersionTableSection<ELFT>::empty() const {
2288 return !In<ELFT>::VerDef && In<ELFT>::VerNeed->empty();
2291 template <class ELFT>
2292 VersionNeedSection<ELFT>::VersionNeedSection()
2293 : SyntheticSection(SHF_ALLOC, SHT_GNU_verneed, sizeof(uint32_t),
2295 // Identifiers in verneed section start at 2 because 0 and 1 are reserved
2296 // for VER_NDX_LOCAL and VER_NDX_GLOBAL.
2297 // First identifiers are reserved by verdef section if it exist.
2298 NextIndex = getVerDefNum() + 1;
2301 template <class ELFT>
2302 void VersionNeedSection<ELFT>::addSymbol(SharedSymbol *SS) {
2303 SharedFile<ELFT> &File = SS->getFile<ELFT>();
2304 const typename ELFT::Verdef *Ver = File.Verdefs[SS->VerdefIndex];
2306 SS->VersionId = VER_NDX_GLOBAL;
2310 // If we don't already know that we need an Elf_Verneed for this DSO, prepare
2311 // to create one by adding it to our needed list and creating a dynstr entry
2313 if (File.VerdefMap.empty())
2314 Needed.push_back({&File, InX::DynStrTab->addString(File.SoName)});
2315 typename SharedFile<ELFT>::NeededVer &NV = File.VerdefMap[Ver];
2316 // If we don't already know that we need an Elf_Vernaux for this Elf_Verdef,
2317 // prepare to create one by allocating a version identifier and creating a
2318 // dynstr entry for the version name.
2319 if (NV.Index == 0) {
2320 NV.StrTab = InX::DynStrTab->addString(File.getStringTable().data() +
2321 Ver->getAux()->vda_name);
2322 NV.Index = NextIndex++;
2324 SS->VersionId = NV.Index;
2327 template <class ELFT> void VersionNeedSection<ELFT>::writeTo(uint8_t *Buf) {
2328 // The Elf_Verneeds need to appear first, followed by the Elf_Vernauxs.
2329 auto *Verneed = reinterpret_cast<Elf_Verneed *>(Buf);
2330 auto *Vernaux = reinterpret_cast<Elf_Vernaux *>(Verneed + Needed.size());
2332 for (std::pair<SharedFile<ELFT> *, size_t> &P : Needed) {
2333 // Create an Elf_Verneed for this DSO.
2334 Verneed->vn_version = 1;
2335 Verneed->vn_cnt = P.first->VerdefMap.size();
2336 Verneed->vn_file = P.second;
2338 reinterpret_cast<char *>(Vernaux) - reinterpret_cast<char *>(Verneed);
2339 Verneed->vn_next = sizeof(Elf_Verneed);
2342 // Create the Elf_Vernauxs for this Elf_Verneed. The loop iterates over
2343 // VerdefMap, which will only contain references to needed version
2344 // definitions. Each Elf_Vernaux is based on the information contained in
2345 // the Elf_Verdef in the source DSO. This loop iterates over a std::map of
2346 // pointers, but is deterministic because the pointers refer to Elf_Verdef
2347 // data structures within a single input file.
2348 for (auto &NV : P.first->VerdefMap) {
2349 Vernaux->vna_hash = NV.first->vd_hash;
2350 Vernaux->vna_flags = 0;
2351 Vernaux->vna_other = NV.second.Index;
2352 Vernaux->vna_name = NV.second.StrTab;
2353 Vernaux->vna_next = sizeof(Elf_Vernaux);
2357 Vernaux[-1].vna_next = 0;
2359 Verneed[-1].vn_next = 0;
2362 template <class ELFT> void VersionNeedSection<ELFT>::finalizeContents() {
2363 getParent()->Link = InX::DynStrTab->getParent()->SectionIndex;
2364 getParent()->Info = Needed.size();
2367 template <class ELFT> size_t VersionNeedSection<ELFT>::getSize() const {
2368 unsigned Size = Needed.size() * sizeof(Elf_Verneed);
2369 for (const std::pair<SharedFile<ELFT> *, size_t> &P : Needed)
2370 Size += P.first->VerdefMap.size() * sizeof(Elf_Vernaux);
2374 template <class ELFT> bool VersionNeedSection<ELFT>::empty() const {
2375 return getNeedNum() == 0;
2378 void MergeSyntheticSection::addSection(MergeInputSection *MS) {
2380 Sections.push_back(MS);
2383 MergeTailSection::MergeTailSection(StringRef Name, uint32_t Type,
2384 uint64_t Flags, uint32_t Alignment)
2385 : MergeSyntheticSection(Name, Type, Flags, Alignment),
2386 Builder(StringTableBuilder::RAW, Alignment) {}
2388 size_t MergeTailSection::getSize() const { return Builder.getSize(); }
2390 void MergeTailSection::writeTo(uint8_t *Buf) { Builder.write(Buf); }
2392 void MergeTailSection::finalizeContents() {
2393 // Add all string pieces to the string table builder to create section
2395 for (MergeInputSection *Sec : Sections)
2396 for (size_t I = 0, E = Sec->Pieces.size(); I != E; ++I)
2397 if (Sec->Pieces[I].Live)
2398 Builder.add(Sec->getData(I));
2400 // Fix the string table content. After this, the contents will never change.
2403 // finalize() fixed tail-optimized strings, so we can now get
2404 // offsets of strings. Get an offset for each string and save it
2405 // to a corresponding StringPiece for easy access.
2406 for (MergeInputSection *Sec : Sections)
2407 for (size_t I = 0, E = Sec->Pieces.size(); I != E; ++I)
2408 if (Sec->Pieces[I].Live)
2409 Sec->Pieces[I].OutputOff = Builder.getOffset(Sec->getData(I));
2412 void MergeNoTailSection::writeTo(uint8_t *Buf) {
2413 for (size_t I = 0; I < NumShards; ++I)
2414 Shards[I].write(Buf + ShardOffsets[I]);
2417 // This function is very hot (i.e. it can take several seconds to finish)
2418 // because sometimes the number of inputs is in an order of magnitude of
2419 // millions. So, we use multi-threading.
2421 // For any strings S and T, we know S is not mergeable with T if S's hash
2422 // value is different from T's. If that's the case, we can safely put S and
2423 // T into different string builders without worrying about merge misses.
2424 // We do it in parallel.
2425 void MergeNoTailSection::finalizeContents() {
2426 // Initializes string table builders.
2427 for (size_t I = 0; I < NumShards; ++I)
2428 Shards.emplace_back(StringTableBuilder::RAW, Alignment);
2430 // Concurrency level. Must be a power of 2 to avoid expensive modulo
2431 // operations in the following tight loop.
2432 size_t Concurrency = 1;
2435 std::min<size_t>(PowerOf2Floor(hardware_concurrency()), NumShards);
2437 // Add section pieces to the builders.
2438 parallelForEachN(0, Concurrency, [&](size_t ThreadId) {
2439 for (MergeInputSection *Sec : Sections) {
2440 for (size_t I = 0, E = Sec->Pieces.size(); I != E; ++I) {
2441 size_t ShardId = getShardId(Sec->Pieces[I].Hash);
2442 if ((ShardId & (Concurrency - 1)) == ThreadId && Sec->Pieces[I].Live)
2443 Sec->Pieces[I].OutputOff = Shards[ShardId].add(Sec->getData(I));
2448 // Compute an in-section offset for each shard.
2450 for (size_t I = 0; I < NumShards; ++I) {
2451 Shards[I].finalizeInOrder();
2452 if (Shards[I].getSize() > 0)
2453 Off = alignTo(Off, Alignment);
2454 ShardOffsets[I] = Off;
2455 Off += Shards[I].getSize();
2459 // So far, section pieces have offsets from beginning of shards, but
2460 // we want offsets from beginning of the whole section. Fix them.
2461 parallelForEach(Sections, [&](MergeInputSection *Sec) {
2462 for (size_t I = 0, E = Sec->Pieces.size(); I != E; ++I)
2463 if (Sec->Pieces[I].Live)
2464 Sec->Pieces[I].OutputOff +=
2465 ShardOffsets[getShardId(Sec->Pieces[I].Hash)];
2469 static MergeSyntheticSection *createMergeSynthetic(StringRef Name,
2472 uint32_t Alignment) {
2473 bool ShouldTailMerge = (Flags & SHF_STRINGS) && Config->Optimize >= 2;
2474 if (ShouldTailMerge)
2475 return make<MergeTailSection>(Name, Type, Flags, Alignment);
2476 return make<MergeNoTailSection>(Name, Type, Flags, Alignment);
2479 // Debug sections may be compressed by zlib. Uncompress if exists.
2480 void elf::decompressSections() {
2481 parallelForEach(InputSections, [](InputSectionBase *Sec) {
2483 Sec->maybeUncompress();
2487 // This function scans over the inputsections to create mergeable
2488 // synthetic sections.
2490 // It removes MergeInputSections from the input section array and adds
2491 // new synthetic sections at the location of the first input section
2492 // that it replaces. It then finalizes each synthetic section in order
2493 // to compute an output offset for each piece of each input section.
2494 void elf::mergeSections() {
2495 // splitIntoPieces needs to be called on each MergeInputSection
2496 // before calling finalizeContents(). Do that first.
2497 parallelForEach(InputSections, [](InputSectionBase *Sec) {
2499 if (auto *S = dyn_cast<MergeInputSection>(Sec))
2500 S->splitIntoPieces();
2503 std::vector<MergeSyntheticSection *> MergeSections;
2504 for (InputSectionBase *&S : InputSections) {
2505 MergeInputSection *MS = dyn_cast<MergeInputSection>(S);
2509 // We do not want to handle sections that are not alive, so just remove
2510 // them instead of trying to merge.
2514 StringRef OutsecName = getOutputSectionName(MS);
2515 uint32_t Alignment = std::max<uint32_t>(MS->Alignment, MS->Entsize);
2517 auto I = llvm::find_if(MergeSections, [=](MergeSyntheticSection *Sec) {
2518 // While we could create a single synthetic section for two different
2519 // values of Entsize, it is better to take Entsize into consideration.
2521 // With a single synthetic section no two pieces with different Entsize
2522 // could be equal, so we may as well have two sections.
2524 // Using Entsize in here also allows us to propagate it to the synthetic
2526 return Sec->Name == OutsecName && Sec->Flags == MS->Flags &&
2527 Sec->Entsize == MS->Entsize && Sec->Alignment == Alignment;
2529 if (I == MergeSections.end()) {
2530 MergeSyntheticSection *Syn =
2531 createMergeSynthetic(OutsecName, MS->Type, MS->Flags, Alignment);
2532 MergeSections.push_back(Syn);
2533 I = std::prev(MergeSections.end());
2535 Syn->Entsize = MS->Entsize;
2539 (*I)->addSection(MS);
2541 for (auto *MS : MergeSections)
2542 MS->finalizeContents();
2544 std::vector<InputSectionBase *> &V = InputSections;
2545 V.erase(std::remove(V.begin(), V.end(), nullptr), V.end());
2548 MipsRldMapSection::MipsRldMapSection()
2549 : SyntheticSection(SHF_ALLOC | SHF_WRITE, SHT_PROGBITS, Config->Wordsize,
2552 ARMExidxSentinelSection::ARMExidxSentinelSection()
2553 : SyntheticSection(SHF_ALLOC | SHF_LINK_ORDER, SHT_ARM_EXIDX,
2554 Config->Wordsize, ".ARM.exidx") {}
2556 // Write a terminating sentinel entry to the end of the .ARM.exidx table.
2557 // This section will have been sorted last in the .ARM.exidx table.
2558 // This table entry will have the form:
2559 // | PREL31 upper bound of code that has exception tables | EXIDX_CANTUNWIND |
2560 // The sentinel must have the PREL31 value of an address higher than any
2561 // address described by any other table entry.
2562 void ARMExidxSentinelSection::writeTo(uint8_t *Buf) {
2565 Highest->getParent()->Addr + Highest->getOffset(Highest->getSize());
2566 uint64_t P = getVA();
2567 Target->relocateOne(Buf, R_ARM_PREL31, S - P);
2568 write32le(Buf + 4, 1);
2571 // The sentinel has to be removed if there are no other .ARM.exidx entries.
2572 bool ARMExidxSentinelSection::empty() const {
2573 OutputSection *OS = getParent();
2574 for (auto *B : OS->SectionCommands)
2575 if (auto *ISD = dyn_cast<InputSectionDescription>(B))
2576 for (auto *S : ISD->Sections)
2577 if (!isa<ARMExidxSentinelSection>(S))
2582 ThunkSection::ThunkSection(OutputSection *OS, uint64_t Off)
2583 : SyntheticSection(SHF_ALLOC | SHF_EXECINSTR, SHT_PROGBITS,
2584 Config->Wordsize, ".text.thunk") {
2586 this->OutSecOff = Off;
2589 void ThunkSection::addThunk(Thunk *T) {
2590 uint64_t Off = alignTo(Size, T->Alignment);
2592 Thunks.push_back(T);
2593 T->addSymbols(*this);
2594 Size = Off + T->size();
2597 void ThunkSection::writeTo(uint8_t *Buf) {
2598 for (const Thunk *T : Thunks)
2599 T->writeTo(Buf + T->Offset, *this);
2602 InputSection *ThunkSection::getTargetInputSection() const {
2605 const Thunk *T = Thunks.front();
2606 return T->getTargetInputSection();
2609 InputSection *InX::ARMAttributes;
2610 BssSection *InX::Bss;
2611 BssSection *InX::BssRelRo;
2612 BuildIdSection *InX::BuildId;
2613 EhFrameHeader *InX::EhFrameHdr;
2614 EhFrameSection *InX::EhFrame;
2615 SyntheticSection *InX::Dynamic;
2616 StringTableSection *InX::DynStrTab;
2617 SymbolTableBaseSection *InX::DynSymTab;
2618 InputSection *InX::Interp;
2619 GdbIndexSection *InX::GdbIndex;
2620 GotSection *InX::Got;
2621 GotPltSection *InX::GotPlt;
2622 GnuHashTableSection *InX::GnuHashTab;
2623 HashTableSection *InX::HashTab;
2624 IgotPltSection *InX::IgotPlt;
2625 MipsGotSection *InX::MipsGot;
2626 MipsRldMapSection *InX::MipsRldMap;
2627 PltSection *InX::Plt;
2628 PltSection *InX::Iplt;
2629 RelocationBaseSection *InX::RelaDyn;
2630 RelocationBaseSection *InX::RelaPlt;
2631 RelocationBaseSection *InX::RelaIplt;
2632 StringTableSection *InX::ShStrTab;
2633 StringTableSection *InX::StrTab;
2634 SymbolTableBaseSection *InX::SymTab;
2636 template GdbIndexSection *elf::createGdbIndex<ELF32LE>();
2637 template GdbIndexSection *elf::createGdbIndex<ELF32BE>();
2638 template GdbIndexSection *elf::createGdbIndex<ELF64LE>();
2639 template GdbIndexSection *elf::createGdbIndex<ELF64BE>();
2641 template void EhFrameSection::addSection<ELF32LE>(InputSectionBase *);
2642 template void EhFrameSection::addSection<ELF32BE>(InputSectionBase *);
2643 template void EhFrameSection::addSection<ELF64LE>(InputSectionBase *);
2644 template void EhFrameSection::addSection<ELF64BE>(InputSectionBase *);
2646 template void PltSection::addEntry<ELF32LE>(Symbol &Sym);
2647 template void PltSection::addEntry<ELF32BE>(Symbol &Sym);
2648 template void PltSection::addEntry<ELF64LE>(Symbol &Sym);
2649 template void PltSection::addEntry<ELF64BE>(Symbol &Sym);
2651 template class elf::MipsAbiFlagsSection<ELF32LE>;
2652 template class elf::MipsAbiFlagsSection<ELF32BE>;
2653 template class elf::MipsAbiFlagsSection<ELF64LE>;
2654 template class elf::MipsAbiFlagsSection<ELF64BE>;
2656 template class elf::MipsOptionsSection<ELF32LE>;
2657 template class elf::MipsOptionsSection<ELF32BE>;
2658 template class elf::MipsOptionsSection<ELF64LE>;
2659 template class elf::MipsOptionsSection<ELF64BE>;
2661 template class elf::MipsReginfoSection<ELF32LE>;
2662 template class elf::MipsReginfoSection<ELF32BE>;
2663 template class elf::MipsReginfoSection<ELF64LE>;
2664 template class elf::MipsReginfoSection<ELF64BE>;
2666 template class elf::DynamicSection<ELF32LE>;
2667 template class elf::DynamicSection<ELF32BE>;
2668 template class elf::DynamicSection<ELF64LE>;
2669 template class elf::DynamicSection<ELF64BE>;
2671 template class elf::RelocationSection<ELF32LE>;
2672 template class elf::RelocationSection<ELF32BE>;
2673 template class elf::RelocationSection<ELF64LE>;
2674 template class elf::RelocationSection<ELF64BE>;
2676 template class elf::AndroidPackedRelocationSection<ELF32LE>;
2677 template class elf::AndroidPackedRelocationSection<ELF32BE>;
2678 template class elf::AndroidPackedRelocationSection<ELF64LE>;
2679 template class elf::AndroidPackedRelocationSection<ELF64BE>;
2681 template class elf::SymbolTableSection<ELF32LE>;
2682 template class elf::SymbolTableSection<ELF32BE>;
2683 template class elf::SymbolTableSection<ELF64LE>;
2684 template class elf::SymbolTableSection<ELF64BE>;
2686 template class elf::VersionTableSection<ELF32LE>;
2687 template class elf::VersionTableSection<ELF32BE>;
2688 template class elf::VersionTableSection<ELF64LE>;
2689 template class elf::VersionTableSection<ELF64BE>;
2691 template class elf::VersionNeedSection<ELF32LE>;
2692 template class elf::VersionNeedSection<ELF32BE>;
2693 template class elf::VersionNeedSection<ELF64LE>;
2694 template class elf::VersionNeedSection<ELF64BE>;
2696 template class elf::VersionDefinitionSection<ELF32LE>;
2697 template class elf::VersionDefinitionSection<ELF32BE>;
2698 template class elf::VersionDefinitionSection<ELF64LE>;
2699 template class elf::VersionDefinitionSection<ELF64BE>;