1 //===- SyntheticSections.cpp ----------------------------------------------===//
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
7 //===----------------------------------------------------------------------===//
9 // This file contains linker-synthesized sections. Currently,
10 // synthetic sections are created either output sections or input sections,
11 // but we are rewriting code so that all synthetic sections are created as
14 //===----------------------------------------------------------------------===//
16 #include "SyntheticSections.h"
18 #include "InputFiles.h"
19 #include "LinkerScript.h"
20 #include "OutputSections.h"
21 #include "SymbolTable.h"
25 #include "lld/Common/ErrorHandler.h"
26 #include "lld/Common/Memory.h"
27 #include "lld/Common/Strings.h"
28 #include "lld/Common/Threads.h"
29 #include "lld/Common/Version.h"
30 #include "llvm/ADT/SetOperations.h"
31 #include "llvm/ADT/StringExtras.h"
32 #include "llvm/BinaryFormat/Dwarf.h"
33 #include "llvm/DebugInfo/DWARF/DWARFDebugPubTable.h"
34 #include "llvm/Object/ELFObjectFile.h"
35 #include "llvm/Support/Compression.h"
36 #include "llvm/Support/Endian.h"
37 #include "llvm/Support/LEB128.h"
38 #include "llvm/Support/MD5.h"
43 using namespace llvm::dwarf;
44 using namespace llvm::ELF;
45 using namespace llvm::object;
46 using namespace llvm::support;
48 using llvm::support::endian::read32le;
49 using llvm::support::endian::write32le;
50 using llvm::support::endian::write64le;
54 constexpr size_t MergeNoTailSection::numShards;
56 static uint64_t readUint(uint8_t *buf) {
57 return config->is64 ? read64(buf) : read32(buf);
60 static void writeUint(uint8_t *buf, uint64_t val) {
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 consistent 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 *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 = 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 = in.mipsGot->getGp();
163 memcpy(buf + sizeof(Elf_Mips_Options), ®info, 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 reginfo.ri_gprmask |= opt->getRegInfo().ri_gprmask;
196 sec->getFile<ELFT>()->mipsGp0 = opt->getRegInfo().ri_gp_value;
201 fatal(filename + ": zero option descriptor size");
202 d = d.slice(opt->size);
206 return make<MipsOptionsSection<ELFT>>(reginfo);
209 // MIPS .reginfo section.
210 template <class ELFT>
211 MipsReginfoSection<ELFT>::MipsReginfoSection(Elf_Mips_RegInfo reginfo)
212 : SyntheticSection(SHF_ALLOC, SHT_MIPS_REGINFO, 4, ".reginfo"),
214 this->entsize = sizeof(Elf_Mips_RegInfo);
217 template <class ELFT> void MipsReginfoSection<ELFT>::writeTo(uint8_t *buf) {
218 if (!config->relocatable)
219 reginfo.ri_gp_value = in.mipsGot->getGp();
220 memcpy(buf, ®info, sizeof(reginfo));
223 template <class ELFT>
224 MipsReginfoSection<ELFT> *MipsReginfoSection<ELFT>::create() {
225 // Section should be alive for O32 and N32 ABIs only.
229 std::vector<InputSectionBase *> sections;
230 for (InputSectionBase *sec : inputSections)
231 if (sec->type == SHT_MIPS_REGINFO)
232 sections.push_back(sec);
234 if (sections.empty())
237 Elf_Mips_RegInfo reginfo = {};
238 for (InputSectionBase *sec : sections) {
241 if (sec->data().size() != sizeof(Elf_Mips_RegInfo)) {
242 error(toString(sec->file) + ": invalid size of .reginfo section");
246 auto *r = reinterpret_cast<const Elf_Mips_RegInfo *>(sec->data().data());
247 reginfo.ri_gprmask |= r->ri_gprmask;
248 sec->getFile<ELFT>()->mipsGp0 = r->ri_gp_value;
251 return make<MipsReginfoSection<ELFT>>(reginfo);
254 InputSection *createInterpSection() {
255 // StringSaver guarantees that the returned string ends with '\0'.
256 StringRef s = saver.save(config->dynamicLinker);
257 ArrayRef<uint8_t> contents = {(const uint8_t *)s.data(), s.size() + 1};
259 return make<InputSection>(nullptr, SHF_ALLOC, SHT_PROGBITS, 1, contents,
263 Defined *addSyntheticLocal(StringRef name, uint8_t type, uint64_t value,
264 uint64_t size, InputSectionBase §ion) {
265 auto *s = make<Defined>(section.file, name, STB_LOCAL, STV_DEFAULT, type,
266 value, size, §ion);
268 in.symTab->addSymbol(s);
272 static size_t getHashSize() {
273 switch (config->buildId) {
274 case BuildIdKind::Fast:
276 case BuildIdKind::Md5:
277 case BuildIdKind::Uuid:
279 case BuildIdKind::Sha1:
281 case BuildIdKind::Hexstring:
282 return config->buildIdVector.size();
284 llvm_unreachable("unknown BuildIdKind");
288 // This class represents a linker-synthesized .note.gnu.property section.
290 // In x86 and AArch64, object files may contain feature flags indicating the
291 // features that they have used. The flags are stored in a .note.gnu.property
294 // lld reads the sections from input files and merges them by computing AND of
295 // the flags. The result is written as a new .note.gnu.property section.
297 // If the flag is zero (which indicates that the intersection of the feature
298 // sets is empty, or some input files didn't have .note.gnu.property sections),
299 // we don't create this section.
300 GnuPropertySection::GnuPropertySection()
301 : SyntheticSection(llvm::ELF::SHF_ALLOC, llvm::ELF::SHT_NOTE,
302 config->wordsize, ".note.gnu.property") {}
304 void GnuPropertySection::writeTo(uint8_t *buf) {
305 uint32_t featureAndType = config->emachine == EM_AARCH64
306 ? GNU_PROPERTY_AARCH64_FEATURE_1_AND
307 : GNU_PROPERTY_X86_FEATURE_1_AND;
309 write32(buf, 4); // Name size
310 write32(buf + 4, config->is64 ? 16 : 12); // Content size
311 write32(buf + 8, NT_GNU_PROPERTY_TYPE_0); // Type
312 memcpy(buf + 12, "GNU", 4); // Name string
313 write32(buf + 16, featureAndType); // Feature type
314 write32(buf + 20, 4); // Feature size
315 write32(buf + 24, config->andFeatures); // Feature flags
317 write32(buf + 28, 0); // Padding
320 size_t GnuPropertySection::getSize() const { return config->is64 ? 32 : 28; }
322 BuildIdSection::BuildIdSection()
323 : SyntheticSection(SHF_ALLOC, SHT_NOTE, 4, ".note.gnu.build-id"),
324 hashSize(getHashSize()) {}
326 void BuildIdSection::writeTo(uint8_t *buf) {
327 write32(buf, 4); // Name size
328 write32(buf + 4, hashSize); // Content size
329 write32(buf + 8, NT_GNU_BUILD_ID); // Type
330 memcpy(buf + 12, "GNU", 4); // Name string
334 void BuildIdSection::writeBuildId(ArrayRef<uint8_t> buf) {
335 assert(buf.size() == hashSize);
336 memcpy(hashBuf, buf.data(), hashSize);
339 BssSection::BssSection(StringRef name, uint64_t size, uint32_t alignment)
340 : SyntheticSection(SHF_ALLOC | SHF_WRITE, SHT_NOBITS, alignment, name) {
345 EhFrameSection::EhFrameSection()
346 : SyntheticSection(SHF_ALLOC, SHT_PROGBITS, 1, ".eh_frame") {}
348 // Search for an existing CIE record or create a new one.
349 // CIE records from input object files are uniquified by their contents
350 // and where their relocations point to.
351 template <class ELFT, class RelTy>
352 CieRecord *EhFrameSection::addCie(EhSectionPiece &cie, ArrayRef<RelTy> rels) {
353 Symbol *personality = nullptr;
354 unsigned firstRelI = cie.firstRelocation;
355 if (firstRelI != (unsigned)-1)
357 &cie.sec->template getFile<ELFT>()->getRelocTargetSym(rels[firstRelI]);
359 // Search for an existing CIE by CIE contents/relocation target pair.
360 CieRecord *&rec = cieMap[{cie.data(), personality}];
362 // If not found, create a new one.
364 rec = make<CieRecord>();
366 cieRecords.push_back(rec);
371 // There is one FDE per function. Returns true if a given FDE
372 // points to a live function.
373 template <class ELFT, class RelTy>
374 bool EhFrameSection::isFdeLive(EhSectionPiece &fde, ArrayRef<RelTy> rels) {
375 auto *sec = cast<EhInputSection>(fde.sec);
376 unsigned firstRelI = fde.firstRelocation;
378 // An FDE should point to some function because FDEs are to describe
379 // functions. That's however not always the case due to an issue of
380 // ld.gold with -r. ld.gold may discard only functions and leave their
381 // corresponding FDEs, which results in creating bad .eh_frame sections.
382 // To deal with that, we ignore such FDEs.
383 if (firstRelI == (unsigned)-1)
386 const RelTy &rel = rels[firstRelI];
387 Symbol &b = sec->template getFile<ELFT>()->getRelocTargetSym(rel);
389 // FDEs for garbage-collected or merged-by-ICF sections, or sections in
390 // another partition, are dead.
391 if (auto *d = dyn_cast<Defined>(&b))
392 if (SectionBase *sec = d->section)
393 return sec->partition == partition;
397 // .eh_frame is a sequence of CIE or FDE records. In general, there
398 // is one CIE record per input object file which is followed by
399 // a list of FDEs. This function searches an existing CIE or create a new
400 // one and associates FDEs to the CIE.
401 template <class ELFT, class RelTy>
402 void EhFrameSection::addRecords(EhInputSection *sec, ArrayRef<RelTy> rels) {
404 for (EhSectionPiece &piece : sec->pieces) {
405 // The empty record is the end marker.
409 size_t offset = piece.inputOff;
410 uint32_t id = read32(piece.data().data() + 4);
412 offsetToCie[offset] = addCie<ELFT>(piece, rels);
416 uint32_t cieOffset = offset + 4 - id;
417 CieRecord *rec = offsetToCie[cieOffset];
419 fatal(toString(sec) + ": invalid CIE reference");
421 if (!isFdeLive<ELFT>(piece, rels))
423 rec->fdes.push_back(&piece);
428 template <class ELFT>
429 void EhFrameSection::addSectionAux(EhInputSection *sec) {
432 if (sec->areRelocsRela)
433 addRecords<ELFT>(sec, sec->template relas<ELFT>());
435 addRecords<ELFT>(sec, sec->template rels<ELFT>());
438 void EhFrameSection::addSection(EhInputSection *sec) {
441 alignment = std::max(alignment, sec->alignment);
442 sections.push_back(sec);
444 for (auto *ds : sec->dependentSections)
445 dependentSections.push_back(ds);
448 static void writeCieFde(uint8_t *buf, ArrayRef<uint8_t> d) {
449 memcpy(buf, d.data(), d.size());
451 size_t aligned = alignTo(d.size(), config->wordsize);
453 // Zero-clear trailing padding if it exists.
454 memset(buf + d.size(), 0, aligned - d.size());
456 // Fix the size field. -4 since size does not include the size field itself.
457 write32(buf, aligned - 4);
460 void EhFrameSection::finalizeContents() {
461 assert(!this->size); // Not finalized.
463 switch (config->ekind) {
465 llvm_unreachable("invalid ekind");
467 for (EhInputSection *sec : sections)
468 addSectionAux<ELF32LE>(sec);
471 for (EhInputSection *sec : sections)
472 addSectionAux<ELF32BE>(sec);
475 for (EhInputSection *sec : sections)
476 addSectionAux<ELF64LE>(sec);
479 for (EhInputSection *sec : sections)
480 addSectionAux<ELF64BE>(sec);
485 for (CieRecord *rec : cieRecords) {
486 rec->cie->outputOff = off;
487 off += alignTo(rec->cie->size, config->wordsize);
489 for (EhSectionPiece *fde : rec->fdes) {
490 fde->outputOff = off;
491 off += alignTo(fde->size, config->wordsize);
495 // The LSB standard does not allow a .eh_frame section with zero
496 // Call Frame Information records. glibc unwind-dw2-fde.c
497 // classify_object_over_fdes expects there is a CIE record length 0 as a
498 // terminator. Thus we add one unconditionally.
504 // Returns data for .eh_frame_hdr. .eh_frame_hdr is a binary search table
505 // to get an FDE from an address to which FDE is applied. This function
506 // returns a list of such pairs.
507 std::vector<EhFrameSection::FdeData> EhFrameSection::getFdeData() const {
508 uint8_t *buf = Out::bufferStart + getParent()->offset + outSecOff;
509 std::vector<FdeData> ret;
511 uint64_t va = getPartition().ehFrameHdr->getVA();
512 for (CieRecord *rec : cieRecords) {
513 uint8_t enc = getFdeEncoding(rec->cie);
514 for (EhSectionPiece *fde : rec->fdes) {
515 uint64_t pc = getFdePc(buf, fde->outputOff, enc);
516 uint64_t fdeVA = getParent()->addr + fde->outputOff;
517 if (!isInt<32>(pc - va))
518 fatal(toString(fde->sec) + ": PC offset is too large: 0x" +
519 Twine::utohexstr(pc - va));
520 ret.push_back({uint32_t(pc - va), uint32_t(fdeVA - va)});
524 // Sort the FDE list by their PC and uniqueify. Usually there is only
525 // one FDE for a PC (i.e. function), but if ICF merges two functions
526 // into one, there can be more than one FDEs pointing to the address.
527 auto less = [](const FdeData &a, const FdeData &b) {
528 return a.pcRel < b.pcRel;
530 llvm::stable_sort(ret, less);
531 auto eq = [](const FdeData &a, const FdeData &b) {
532 return a.pcRel == b.pcRel;
534 ret.erase(std::unique(ret.begin(), ret.end(), eq), ret.end());
539 static uint64_t readFdeAddr(uint8_t *buf, int size) {
541 case DW_EH_PE_udata2:
543 case DW_EH_PE_sdata2:
544 return (int16_t)read16(buf);
545 case DW_EH_PE_udata4:
547 case DW_EH_PE_sdata4:
548 return (int32_t)read32(buf);
549 case DW_EH_PE_udata8:
550 case DW_EH_PE_sdata8:
552 case DW_EH_PE_absptr:
553 return readUint(buf);
555 fatal("unknown FDE size encoding");
558 // Returns the VA to which a given FDE (on a mmap'ed buffer) is applied to.
559 // We need it to create .eh_frame_hdr section.
560 uint64_t EhFrameSection::getFdePc(uint8_t *buf, size_t fdeOff,
562 // The starting address to which this FDE applies is
563 // stored at FDE + 8 byte.
564 size_t off = fdeOff + 8;
565 uint64_t addr = readFdeAddr(buf + off, enc & 0xf);
566 if ((enc & 0x70) == DW_EH_PE_absptr)
568 if ((enc & 0x70) == DW_EH_PE_pcrel)
569 return addr + getParent()->addr + off;
570 fatal("unknown FDE size relative encoding");
573 void EhFrameSection::writeTo(uint8_t *buf) {
574 // Write CIE and FDE records.
575 for (CieRecord *rec : cieRecords) {
576 size_t cieOffset = rec->cie->outputOff;
577 writeCieFde(buf + cieOffset, rec->cie->data());
579 for (EhSectionPiece *fde : rec->fdes) {
580 size_t off = fde->outputOff;
581 writeCieFde(buf + off, fde->data());
583 // FDE's second word should have the offset to an associated CIE.
585 write32(buf + off + 4, off + 4 - cieOffset);
589 // Apply relocations. .eh_frame section contents are not contiguous
590 // in the output buffer, but relocateAlloc() still works because
591 // getOffset() takes care of discontiguous section pieces.
592 for (EhInputSection *s : sections)
593 s->relocateAlloc(buf, nullptr);
595 if (getPartition().ehFrameHdr && getPartition().ehFrameHdr->getParent())
596 getPartition().ehFrameHdr->write();
599 GotSection::GotSection()
600 : SyntheticSection(SHF_ALLOC | SHF_WRITE, SHT_PROGBITS, config->wordsize,
602 // If ElfSym::globalOffsetTable is relative to .got and is referenced,
603 // increase numEntries by the number of entries used to emit
604 // ElfSym::globalOffsetTable.
605 if (ElfSym::globalOffsetTable && !target->gotBaseSymInGotPlt)
606 numEntries += target->gotHeaderEntriesNum;
609 void GotSection::addEntry(Symbol &sym) {
610 sym.gotIndex = numEntries;
614 bool GotSection::addDynTlsEntry(Symbol &sym) {
615 if (sym.globalDynIndex != -1U)
617 sym.globalDynIndex = numEntries;
618 // Global Dynamic TLS entries take two GOT slots.
623 // Reserves TLS entries for a TLS module ID and a TLS block offset.
624 // In total it takes two GOT slots.
625 bool GotSection::addTlsIndex() {
626 if (tlsIndexOff != uint32_t(-1))
628 tlsIndexOff = numEntries * config->wordsize;
633 uint64_t GotSection::getGlobalDynAddr(const Symbol &b) const {
634 return this->getVA() + b.globalDynIndex * config->wordsize;
637 uint64_t GotSection::getGlobalDynOffset(const Symbol &b) const {
638 return b.globalDynIndex * config->wordsize;
641 void GotSection::finalizeContents() {
642 size = numEntries * config->wordsize;
645 bool GotSection::isNeeded() const {
646 // We need to emit a GOT even if it's empty if there's a relocation that is
647 // relative to GOT(such as GOTOFFREL).
648 return numEntries || hasGotOffRel;
651 void GotSection::writeTo(uint8_t *buf) {
652 // Buf points to the start of this section's buffer,
653 // whereas InputSectionBase::relocateAlloc() expects its argument
654 // to point to the start of the output section.
655 target->writeGotHeader(buf);
656 relocateAlloc(buf - outSecOff, buf - outSecOff + size);
659 static uint64_t getMipsPageAddr(uint64_t addr) {
660 return (addr + 0x8000) & ~0xffff;
663 static uint64_t getMipsPageCount(uint64_t size) {
664 return (size + 0xfffe) / 0xffff + 1;
667 MipsGotSection::MipsGotSection()
668 : SyntheticSection(SHF_ALLOC | SHF_WRITE | SHF_MIPS_GPREL, SHT_PROGBITS, 16,
671 void MipsGotSection::addEntry(InputFile &file, Symbol &sym, int64_t addend,
673 FileGot &g = getGot(file);
674 if (expr == R_MIPS_GOT_LOCAL_PAGE) {
675 if (const OutputSection *os = sym.getOutputSection())
676 g.pagesMap.insert({os, {}});
678 g.local16.insert({{nullptr, getMipsPageAddr(sym.getVA(addend))}, 0});
679 } else if (sym.isTls())
680 g.tls.insert({&sym, 0});
681 else if (sym.isPreemptible && expr == R_ABS)
682 g.relocs.insert({&sym, 0});
683 else if (sym.isPreemptible)
684 g.global.insert({&sym, 0});
685 else if (expr == R_MIPS_GOT_OFF32)
686 g.local32.insert({{&sym, addend}, 0});
688 g.local16.insert({{&sym, addend}, 0});
691 void MipsGotSection::addDynTlsEntry(InputFile &file, Symbol &sym) {
692 getGot(file).dynTlsSymbols.insert({&sym, 0});
695 void MipsGotSection::addTlsIndex(InputFile &file) {
696 getGot(file).dynTlsSymbols.insert({nullptr, 0});
699 size_t MipsGotSection::FileGot::getEntriesNum() const {
700 return getPageEntriesNum() + local16.size() + global.size() + relocs.size() +
701 tls.size() + dynTlsSymbols.size() * 2;
704 size_t MipsGotSection::FileGot::getPageEntriesNum() const {
706 for (const std::pair<const OutputSection *, FileGot::PageBlock> &p : pagesMap)
707 num += p.second.count;
711 size_t MipsGotSection::FileGot::getIndexedEntriesNum() const {
712 size_t count = getPageEntriesNum() + local16.size() + global.size();
713 // If there are relocation-only entries in the GOT, TLS entries
714 // are allocated after them. TLS entries should be addressable
715 // by 16-bit index so count both reloc-only and TLS entries.
716 if (!tls.empty() || !dynTlsSymbols.empty())
717 count += relocs.size() + tls.size() + dynTlsSymbols.size() * 2;
721 MipsGotSection::FileGot &MipsGotSection::getGot(InputFile &f) {
722 if (!f.mipsGotIndex.hasValue()) {
724 gots.back().file = &f;
725 f.mipsGotIndex = gots.size() - 1;
727 return gots[*f.mipsGotIndex];
730 uint64_t MipsGotSection::getPageEntryOffset(const InputFile *f,
732 int64_t addend) const {
733 const FileGot &g = gots[*f->mipsGotIndex];
735 if (const OutputSection *outSec = sym.getOutputSection()) {
736 uint64_t secAddr = getMipsPageAddr(outSec->addr);
737 uint64_t symAddr = getMipsPageAddr(sym.getVA(addend));
738 index = g.pagesMap.lookup(outSec).firstIndex + (symAddr - secAddr) / 0xffff;
740 index = g.local16.lookup({nullptr, getMipsPageAddr(sym.getVA(addend))});
742 return index * config->wordsize;
745 uint64_t MipsGotSection::getSymEntryOffset(const InputFile *f, const Symbol &s,
746 int64_t addend) const {
747 const FileGot &g = gots[*f->mipsGotIndex];
748 Symbol *sym = const_cast<Symbol *>(&s);
750 return g.tls.lookup(sym) * config->wordsize;
751 if (sym->isPreemptible)
752 return g.global.lookup(sym) * config->wordsize;
753 return g.local16.lookup({sym, addend}) * config->wordsize;
756 uint64_t MipsGotSection::getTlsIndexOffset(const InputFile *f) const {
757 const FileGot &g = gots[*f->mipsGotIndex];
758 return g.dynTlsSymbols.lookup(nullptr) * config->wordsize;
761 uint64_t MipsGotSection::getGlobalDynOffset(const InputFile *f,
762 const Symbol &s) const {
763 const FileGot &g = gots[*f->mipsGotIndex];
764 Symbol *sym = const_cast<Symbol *>(&s);
765 return g.dynTlsSymbols.lookup(sym) * config->wordsize;
768 const Symbol *MipsGotSection::getFirstGlobalEntry() const {
771 const FileGot &primGot = gots.front();
772 if (!primGot.global.empty())
773 return primGot.global.front().first;
774 if (!primGot.relocs.empty())
775 return primGot.relocs.front().first;
779 unsigned MipsGotSection::getLocalEntriesNum() const {
781 return headerEntriesNum;
782 return headerEntriesNum + gots.front().getPageEntriesNum() +
783 gots.front().local16.size();
786 bool MipsGotSection::tryMergeGots(FileGot &dst, FileGot &src, bool isPrimary) {
788 set_union(tmp.pagesMap, src.pagesMap);
789 set_union(tmp.local16, src.local16);
790 set_union(tmp.global, src.global);
791 set_union(tmp.relocs, src.relocs);
792 set_union(tmp.tls, src.tls);
793 set_union(tmp.dynTlsSymbols, src.dynTlsSymbols);
795 size_t count = isPrimary ? headerEntriesNum : 0;
796 count += tmp.getIndexedEntriesNum();
798 if (count * config->wordsize > config->mipsGotSize)
805 void MipsGotSection::finalizeContents() { updateAllocSize(); }
807 bool MipsGotSection::updateAllocSize() {
808 size = headerEntriesNum * config->wordsize;
809 for (const FileGot &g : gots)
810 size += g.getEntriesNum() * config->wordsize;
814 void MipsGotSection::build() {
818 std::vector<FileGot> mergedGots(1);
820 // For each GOT move non-preemptible symbols from the `Global`
821 // to `Local16` list. Preemptible symbol might become non-preemptible
822 // one if, for example, it gets a related copy relocation.
823 for (FileGot &got : gots) {
824 for (auto &p: got.global)
825 if (!p.first->isPreemptible)
826 got.local16.insert({{p.first, 0}, 0});
827 got.global.remove_if([&](const std::pair<Symbol *, size_t> &p) {
828 return !p.first->isPreemptible;
832 // For each GOT remove "reloc-only" entry if there is "global"
833 // entry for the same symbol. And add local entries which indexed
834 // using 32-bit value at the end of 16-bit entries.
835 for (FileGot &got : gots) {
836 got.relocs.remove_if([&](const std::pair<Symbol *, size_t> &p) {
837 return got.global.count(p.first);
839 set_union(got.local16, got.local32);
843 // Evaluate number of "reloc-only" entries in the resulting GOT.
844 // To do that put all unique "reloc-only" and "global" entries
845 // from all GOTs to the future primary GOT.
846 FileGot *primGot = &mergedGots.front();
847 for (FileGot &got : gots) {
848 set_union(primGot->relocs, got.global);
849 set_union(primGot->relocs, got.relocs);
853 // Evaluate number of "page" entries in each GOT.
854 for (FileGot &got : gots) {
855 for (std::pair<const OutputSection *, FileGot::PageBlock> &p :
857 const OutputSection *os = p.first;
858 uint64_t secSize = 0;
859 for (BaseCommand *cmd : os->sectionCommands) {
860 if (auto *isd = dyn_cast<InputSectionDescription>(cmd))
861 for (InputSection *isec : isd->sections) {
862 uint64_t off = alignTo(secSize, isec->alignment);
863 secSize = off + isec->getSize();
866 p.second.count = getMipsPageCount(secSize);
870 // Merge GOTs. Try to join as much as possible GOTs but do not exceed
871 // maximum GOT size. At first, try to fill the primary GOT because
872 // the primary GOT can be accessed in the most effective way. If it
873 // is not possible, try to fill the last GOT in the list, and finally
874 // create a new GOT if both attempts failed.
875 for (FileGot &srcGot : gots) {
876 InputFile *file = srcGot.file;
877 if (tryMergeGots(mergedGots.front(), srcGot, true)) {
878 file->mipsGotIndex = 0;
880 // If this is the first time we failed to merge with the primary GOT,
881 // MergedGots.back() will also be the primary GOT. We must make sure not
882 // to try to merge again with isPrimary=false, as otherwise, if the
883 // inputs are just right, we could allow the primary GOT to become 1 or 2
884 // words bigger due to ignoring the header size.
885 if (mergedGots.size() == 1 ||
886 !tryMergeGots(mergedGots.back(), srcGot, false)) {
887 mergedGots.emplace_back();
888 std::swap(mergedGots.back(), srcGot);
890 file->mipsGotIndex = mergedGots.size() - 1;
893 std::swap(gots, mergedGots);
895 // Reduce number of "reloc-only" entries in the primary GOT
896 // by subtracting "global" entries in the primary GOT.
897 primGot = &gots.front();
898 primGot->relocs.remove_if([&](const std::pair<Symbol *, size_t> &p) {
899 return primGot->global.count(p.first);
902 // Calculate indexes for each GOT entry.
903 size_t index = headerEntriesNum;
904 for (FileGot &got : gots) {
905 got.startIndex = &got == primGot ? 0 : index;
906 for (std::pair<const OutputSection *, FileGot::PageBlock> &p :
908 // For each output section referenced by GOT page relocations calculate
909 // and save into pagesMap an upper bound of MIPS GOT entries required
910 // to store page addresses of local symbols. We assume the worst case -
911 // each 64kb page of the output section has at least one GOT relocation
912 // against it. And take in account the case when the section intersects
914 p.second.firstIndex = index;
915 index += p.second.count;
917 for (auto &p: got.local16)
919 for (auto &p: got.global)
921 for (auto &p: got.relocs)
923 for (auto &p: got.tls)
925 for (auto &p: got.dynTlsSymbols) {
931 // Update Symbol::gotIndex field to use this
932 // value later in the `sortMipsSymbols` function.
933 for (auto &p : primGot->global)
934 p.first->gotIndex = p.second;
935 for (auto &p : primGot->relocs)
936 p.first->gotIndex = p.second;
938 // Create dynamic relocations.
939 for (FileGot &got : gots) {
940 // Create dynamic relocations for TLS entries.
941 for (std::pair<Symbol *, size_t> &p : got.tls) {
943 uint64_t offset = p.second * config->wordsize;
944 if (s->isPreemptible)
945 mainPart->relaDyn->addReloc(target->tlsGotRel, this, offset, s);
947 for (std::pair<Symbol *, size_t> &p : got.dynTlsSymbols) {
949 uint64_t offset = p.second * config->wordsize;
953 mainPart->relaDyn->addReloc(target->tlsModuleIndexRel, this, offset, s);
955 // When building a shared library we still need a dynamic relocation
956 // for the module index. Therefore only checking for
957 // S->isPreemptible is not sufficient (this happens e.g. for
958 // thread-locals that have been marked as local through a linker script)
959 if (!s->isPreemptible && !config->isPic)
961 mainPart->relaDyn->addReloc(target->tlsModuleIndexRel, this, offset, s);
962 // However, we can skip writing the TLS offset reloc for non-preemptible
963 // symbols since it is known even in shared libraries
964 if (!s->isPreemptible)
966 offset += config->wordsize;
967 mainPart->relaDyn->addReloc(target->tlsOffsetRel, this, offset, s);
971 // Do not create dynamic relocations for non-TLS
972 // entries in the primary GOT.
976 // Dynamic relocations for "global" entries.
977 for (const std::pair<Symbol *, size_t> &p : got.global) {
978 uint64_t offset = p.second * config->wordsize;
979 mainPart->relaDyn->addReloc(target->relativeRel, this, offset, p.first);
983 // Dynamic relocations for "local" entries in case of PIC.
984 for (const std::pair<const OutputSection *, FileGot::PageBlock> &l :
986 size_t pageCount = l.second.count;
987 for (size_t pi = 0; pi < pageCount; ++pi) {
988 uint64_t offset = (l.second.firstIndex + pi) * config->wordsize;
989 mainPart->relaDyn->addReloc({target->relativeRel, this, offset, l.first,
990 int64_t(pi * 0x10000)});
993 for (const std::pair<GotEntry, size_t> &p : got.local16) {
994 uint64_t offset = p.second * config->wordsize;
995 mainPart->relaDyn->addReloc({target->relativeRel, this, offset, true,
996 p.first.first, p.first.second});
1001 bool MipsGotSection::isNeeded() const {
1002 // We add the .got section to the result for dynamic MIPS target because
1003 // its address and properties are mentioned in the .dynamic section.
1004 return !config->relocatable;
1007 uint64_t MipsGotSection::getGp(const InputFile *f) const {
1008 // For files without related GOT or files refer a primary GOT
1009 // returns "common" _gp value. For secondary GOTs calculate
1010 // individual _gp values.
1011 if (!f || !f->mipsGotIndex.hasValue() || *f->mipsGotIndex == 0)
1012 return ElfSym::mipsGp->getVA(0);
1013 return getVA() + gots[*f->mipsGotIndex].startIndex * config->wordsize +
1017 void MipsGotSection::writeTo(uint8_t *buf) {
1018 // Set the MSB of the second GOT slot. This is not required by any
1019 // MIPS ABI documentation, though.
1021 // There is a comment in glibc saying that "The MSB of got[1] of a
1022 // gnu object is set to identify gnu objects," and in GNU gold it
1023 // says "the second entry will be used by some runtime loaders".
1024 // But how this field is being used is unclear.
1026 // We are not really willing to mimic other linkers behaviors
1027 // without understanding why they do that, but because all files
1028 // generated by GNU tools have this special GOT value, and because
1029 // we've been doing this for years, it is probably a safe bet to
1030 // keep doing this for now. We really need to revisit this to see
1031 // if we had to do this.
1032 writeUint(buf + config->wordsize, (uint64_t)1 << (config->wordsize * 8 - 1));
1033 for (const FileGot &g : gots) {
1034 auto write = [&](size_t i, const Symbol *s, int64_t a) {
1038 writeUint(buf + i * config->wordsize, va);
1040 // Write 'page address' entries to the local part of the GOT.
1041 for (const std::pair<const OutputSection *, FileGot::PageBlock> &l :
1043 size_t pageCount = l.second.count;
1044 uint64_t firstPageAddr = getMipsPageAddr(l.first->addr);
1045 for (size_t pi = 0; pi < pageCount; ++pi)
1046 write(l.second.firstIndex + pi, nullptr, firstPageAddr + pi * 0x10000);
1048 // Local, global, TLS, reloc-only entries.
1049 // If TLS entry has a corresponding dynamic relocations, leave it
1050 // initialized by zero. Write down adjusted TLS symbol's values otherwise.
1051 // To calculate the adjustments use offsets for thread-local storage.
1052 // https://www.linux-mips.org/wiki/NPTL
1053 for (const std::pair<GotEntry, size_t> &p : g.local16)
1054 write(p.second, p.first.first, p.first.second);
1055 // Write VA to the primary GOT only. For secondary GOTs that
1056 // will be done by REL32 dynamic relocations.
1057 if (&g == &gots.front())
1058 for (const std::pair<Symbol *, size_t> &p : g.global)
1059 write(p.second, p.first, 0);
1060 for (const std::pair<Symbol *, size_t> &p : g.relocs)
1061 write(p.second, p.first, 0);
1062 for (const std::pair<Symbol *, size_t> &p : g.tls)
1063 write(p.second, p.first, p.first->isPreemptible ? 0 : -0x7000);
1064 for (const std::pair<Symbol *, size_t> &p : g.dynTlsSymbols) {
1065 if (p.first == nullptr && !config->isPic)
1066 write(p.second, nullptr, 1);
1067 else if (p.first && !p.first->isPreemptible) {
1068 // If we are emitting PIC code with relocations we mustn't write
1069 // anything to the GOT here. When using Elf_Rel relocations the value
1070 // one will be treated as an addend and will cause crashes at runtime
1072 write(p.second, nullptr, 1);
1073 write(p.second + 1, p.first, -0x8000);
1079 // On PowerPC the .plt section is used to hold the table of function addresses
1080 // instead of the .got.plt, and the type is SHT_NOBITS similar to a .bss
1081 // section. I don't know why we have a BSS style type for the section but it is
1082 // consistent across both 64-bit PowerPC ABIs as well as the 32-bit PowerPC ABI.
1083 GotPltSection::GotPltSection()
1084 : SyntheticSection(SHF_ALLOC | SHF_WRITE, SHT_PROGBITS, config->wordsize,
1086 if (config->emachine == EM_PPC) {
1088 } else if (config->emachine == EM_PPC64) {
1094 void GotPltSection::addEntry(Symbol &sym) {
1095 assert(sym.pltIndex == entries.size());
1096 entries.push_back(&sym);
1099 size_t GotPltSection::getSize() const {
1100 return (target->gotPltHeaderEntriesNum + entries.size()) * config->wordsize;
1103 void GotPltSection::writeTo(uint8_t *buf) {
1104 target->writeGotPltHeader(buf);
1105 buf += target->gotPltHeaderEntriesNum * config->wordsize;
1106 for (const Symbol *b : entries) {
1107 target->writeGotPlt(buf, *b);
1108 buf += config->wordsize;
1112 bool GotPltSection::isNeeded() const {
1113 // We need to emit GOTPLT even if it's empty if there's a relocation relative
1115 return !entries.empty() || hasGotPltOffRel;
1118 static StringRef getIgotPltName() {
1119 // On ARM the IgotPltSection is part of the GotSection.
1120 if (config->emachine == EM_ARM)
1123 // On PowerPC64 the GotPltSection is renamed to '.plt' so the IgotPltSection
1124 // needs to be named the same.
1125 if (config->emachine == EM_PPC64)
1131 // On PowerPC64 the GotPltSection type is SHT_NOBITS so we have to follow suit
1132 // with the IgotPltSection.
1133 IgotPltSection::IgotPltSection()
1134 : SyntheticSection(SHF_ALLOC | SHF_WRITE,
1135 config->emachine == EM_PPC64 ? SHT_NOBITS : SHT_PROGBITS,
1136 config->wordsize, getIgotPltName()) {}
1138 void IgotPltSection::addEntry(Symbol &sym) {
1139 assert(sym.pltIndex == entries.size());
1140 entries.push_back(&sym);
1143 size_t IgotPltSection::getSize() const {
1144 return entries.size() * config->wordsize;
1147 void IgotPltSection::writeTo(uint8_t *buf) {
1148 for (const Symbol *b : entries) {
1149 target->writeIgotPlt(buf, *b);
1150 buf += config->wordsize;
1154 StringTableSection::StringTableSection(StringRef name, bool dynamic)
1155 : SyntheticSection(dynamic ? (uint64_t)SHF_ALLOC : 0, SHT_STRTAB, 1, name),
1157 // ELF string tables start with a NUL byte.
1161 // Adds a string to the string table. If `hashIt` is true we hash and check for
1162 // duplicates. It is optional because the name of global symbols are already
1163 // uniqued and hashing them again has a big cost for a small value: uniquing
1164 // them with some other string that happens to be the same.
1165 unsigned StringTableSection::addString(StringRef s, bool hashIt) {
1167 auto r = stringMap.insert(std::make_pair(s, this->size));
1169 return r.first->second;
1171 unsigned ret = this->size;
1172 this->size = this->size + s.size() + 1;
1173 strings.push_back(s);
1177 void StringTableSection::writeTo(uint8_t *buf) {
1178 for (StringRef s : strings) {
1179 memcpy(buf, s.data(), s.size());
1180 buf[s.size()] = '\0';
1181 buf += s.size() + 1;
1185 // Returns the number of entries in .gnu.version_d: the number of
1186 // non-VER_NDX_LOCAL-non-VER_NDX_GLOBAL definitions, plus 1.
1187 // Note that we don't support vd_cnt > 1 yet.
1188 static unsigned getVerDefNum() {
1189 return namedVersionDefs().size() + 1;
1192 template <class ELFT>
1193 DynamicSection<ELFT>::DynamicSection()
1194 : SyntheticSection(SHF_ALLOC | SHF_WRITE, SHT_DYNAMIC, config->wordsize,
1196 this->entsize = ELFT::Is64Bits ? 16 : 8;
1198 // .dynamic section is not writable on MIPS and on Fuchsia OS
1199 // which passes -z rodynamic.
1200 // See "Special Section" in Chapter 4 in the following document:
1201 // ftp://www.linux-mips.org/pub/linux/mips/doc/ABI/mipsabi.pdf
1202 if (config->emachine == EM_MIPS || config->zRodynamic)
1203 this->flags = SHF_ALLOC;
1206 template <class ELFT>
1207 void DynamicSection<ELFT>::add(int32_t tag, std::function<uint64_t()> fn) {
1208 entries.push_back({tag, fn});
1211 template <class ELFT>
1212 void DynamicSection<ELFT>::addInt(int32_t tag, uint64_t val) {
1213 entries.push_back({tag, [=] { return val; }});
1216 template <class ELFT>
1217 void DynamicSection<ELFT>::addInSec(int32_t tag, InputSection *sec) {
1218 entries.push_back({tag, [=] { return sec->getVA(0); }});
1221 template <class ELFT>
1222 void DynamicSection<ELFT>::addInSecRelative(int32_t tag, InputSection *sec) {
1223 size_t tagOffset = entries.size() * entsize;
1225 {tag, [=] { return sec->getVA(0) - (getVA() + tagOffset); }});
1228 template <class ELFT>
1229 void DynamicSection<ELFT>::addOutSec(int32_t tag, OutputSection *sec) {
1230 entries.push_back({tag, [=] { return sec->addr; }});
1233 template <class ELFT>
1234 void DynamicSection<ELFT>::addSize(int32_t tag, OutputSection *sec) {
1235 entries.push_back({tag, [=] { return sec->size; }});
1238 template <class ELFT>
1239 void DynamicSection<ELFT>::addSym(int32_t tag, Symbol *sym) {
1240 entries.push_back({tag, [=] { return sym->getVA(); }});
1243 // The output section .rela.dyn may include these synthetic sections:
1246 // - in.relaIplt: this is included if in.relaIplt is named .rela.dyn
1247 // - in.relaPlt: this is included if a linker script places .rela.plt inside
1250 // DT_RELASZ is the total size of the included sections.
1251 static std::function<uint64_t()> addRelaSz(RelocationBaseSection *relaDyn) {
1253 size_t size = relaDyn->getSize();
1254 if (in.relaIplt->getParent() == relaDyn->getParent())
1255 size += in.relaIplt->getSize();
1256 if (in.relaPlt->getParent() == relaDyn->getParent())
1257 size += in.relaPlt->getSize();
1262 // A Linker script may assign the RELA relocation sections to the same
1263 // output section. When this occurs we cannot just use the OutputSection
1264 // Size. Moreover the [DT_JMPREL, DT_JMPREL + DT_PLTRELSZ) is permitted to
1265 // overlap with the [DT_RELA, DT_RELA + DT_RELASZ).
1266 static uint64_t addPltRelSz() {
1267 size_t size = in.relaPlt->getSize();
1268 if (in.relaIplt->getParent() == in.relaPlt->getParent() &&
1269 in.relaIplt->name == in.relaPlt->name)
1270 size += in.relaIplt->getSize();
1274 // Add remaining entries to complete .dynamic contents.
1275 template <class ELFT> void DynamicSection<ELFT>::finalizeContents() {
1276 Partition &part = getPartition();
1277 bool isMain = part.name.empty();
1279 for (StringRef s : config->filterList)
1280 addInt(DT_FILTER, part.dynStrTab->addString(s));
1281 for (StringRef s : config->auxiliaryList)
1282 addInt(DT_AUXILIARY, part.dynStrTab->addString(s));
1284 if (!config->rpath.empty())
1285 addInt(config->enableNewDtags ? DT_RUNPATH : DT_RPATH,
1286 part.dynStrTab->addString(config->rpath));
1288 for (SharedFile *file : sharedFiles)
1290 addInt(DT_NEEDED, part.dynStrTab->addString(file->soName));
1293 if (!config->soName.empty())
1294 addInt(DT_SONAME, part.dynStrTab->addString(config->soName));
1296 if (!config->soName.empty())
1297 addInt(DT_NEEDED, part.dynStrTab->addString(config->soName));
1298 addInt(DT_SONAME, part.dynStrTab->addString(part.name));
1301 // Set DT_FLAGS and DT_FLAGS_1.
1302 uint32_t dtFlags = 0;
1303 uint32_t dtFlags1 = 0;
1304 if (config->bsymbolic)
1305 dtFlags |= DF_SYMBOLIC;
1306 if (config->zGlobal)
1307 dtFlags1 |= DF_1_GLOBAL;
1308 if (config->zInitfirst)
1309 dtFlags1 |= DF_1_INITFIRST;
1310 if (config->zInterpose)
1311 dtFlags1 |= DF_1_INTERPOSE;
1312 if (config->zNodefaultlib)
1313 dtFlags1 |= DF_1_NODEFLIB;
1314 if (config->zNodelete)
1315 dtFlags1 |= DF_1_NODELETE;
1316 if (config->zNodlopen)
1317 dtFlags1 |= DF_1_NOOPEN;
1319 dtFlags |= DF_BIND_NOW;
1320 dtFlags1 |= DF_1_NOW;
1322 if (config->zOrigin) {
1323 dtFlags |= DF_ORIGIN;
1324 dtFlags1 |= DF_1_ORIGIN;
1327 dtFlags |= DF_TEXTREL;
1328 if (config->hasStaticTlsModel)
1329 dtFlags |= DF_STATIC_TLS;
1332 addInt(DT_FLAGS, dtFlags);
1334 addInt(DT_FLAGS_1, dtFlags1);
1336 // DT_DEBUG is a pointer to debug information used by debuggers at runtime. We
1337 // need it for each process, so we don't write it for DSOs. The loader writes
1338 // the pointer into this entry.
1340 // DT_DEBUG is the only .dynamic entry that needs to be written to. Some
1341 // systems (currently only Fuchsia OS) provide other means to give the
1342 // debugger this information. Such systems may choose make .dynamic read-only.
1343 // If the target is such a system (used -z rodynamic) don't write DT_DEBUG.
1344 if (!config->shared && !config->relocatable && !config->zRodynamic)
1345 addInt(DT_DEBUG, 0);
1347 if (OutputSection *sec = part.dynStrTab->getParent())
1348 this->link = sec->sectionIndex;
1350 if (part.relaDyn->isNeeded() ||
1351 (in.relaIplt->isNeeded() &&
1352 part.relaDyn->getParent() == in.relaIplt->getParent())) {
1353 addInSec(part.relaDyn->dynamicTag, part.relaDyn);
1354 entries.push_back({part.relaDyn->sizeDynamicTag, addRelaSz(part.relaDyn)});
1356 bool isRela = config->isRela;
1357 addInt(isRela ? DT_RELAENT : DT_RELENT,
1358 isRela ? sizeof(Elf_Rela) : sizeof(Elf_Rel));
1360 // MIPS dynamic loader does not support RELCOUNT tag.
1361 // The problem is in the tight relation between dynamic
1362 // relocations and GOT. So do not emit this tag on MIPS.
1363 if (config->emachine != EM_MIPS) {
1364 size_t numRelativeRels = part.relaDyn->getRelativeRelocCount();
1365 if (config->zCombreloc && numRelativeRels)
1366 addInt(isRela ? DT_RELACOUNT : DT_RELCOUNT, numRelativeRels);
1369 if (part.relrDyn && !part.relrDyn->relocs.empty()) {
1370 addInSec(config->useAndroidRelrTags ? DT_ANDROID_RELR : DT_RELR,
1372 addSize(config->useAndroidRelrTags ? DT_ANDROID_RELRSZ : DT_RELRSZ,
1373 part.relrDyn->getParent());
1374 addInt(config->useAndroidRelrTags ? DT_ANDROID_RELRENT : DT_RELRENT,
1377 // .rel[a].plt section usually consists of two parts, containing plt and
1378 // iplt relocations. It is possible to have only iplt relocations in the
1379 // output. In that case relaPlt is empty and have zero offset, the same offset
1380 // as relaIplt has. And we still want to emit proper dynamic tags for that
1381 // case, so here we always use relaPlt as marker for the beginning of
1382 // .rel[a].plt section.
1383 if (isMain && (in.relaPlt->isNeeded() || in.relaIplt->isNeeded())) {
1384 addInSec(DT_JMPREL, in.relaPlt);
1385 entries.push_back({DT_PLTRELSZ, addPltRelSz});
1386 switch (config->emachine) {
1388 addInSec(DT_MIPS_PLTGOT, in.gotPlt);
1391 addInSec(DT_PLTGOT, in.plt);
1394 addInSec(DT_PLTGOT, in.gotPlt);
1397 addInt(DT_PLTREL, config->isRela ? DT_RELA : DT_REL);
1400 if (config->emachine == EM_AARCH64) {
1401 if (config->andFeatures & GNU_PROPERTY_AARCH64_FEATURE_1_BTI)
1402 addInt(DT_AARCH64_BTI_PLT, 0);
1403 if (config->andFeatures & GNU_PROPERTY_AARCH64_FEATURE_1_PAC)
1404 addInt(DT_AARCH64_PAC_PLT, 0);
1407 addInSec(DT_SYMTAB, part.dynSymTab);
1408 addInt(DT_SYMENT, sizeof(Elf_Sym));
1409 addInSec(DT_STRTAB, part.dynStrTab);
1410 addInt(DT_STRSZ, part.dynStrTab->getSize());
1412 addInt(DT_TEXTREL, 0);
1413 if (part.gnuHashTab)
1414 addInSec(DT_GNU_HASH, part.gnuHashTab);
1416 addInSec(DT_HASH, part.hashTab);
1419 if (Out::preinitArray) {
1420 addOutSec(DT_PREINIT_ARRAY, Out::preinitArray);
1421 addSize(DT_PREINIT_ARRAYSZ, Out::preinitArray);
1423 if (Out::initArray) {
1424 addOutSec(DT_INIT_ARRAY, Out::initArray);
1425 addSize(DT_INIT_ARRAYSZ, Out::initArray);
1427 if (Out::finiArray) {
1428 addOutSec(DT_FINI_ARRAY, Out::finiArray);
1429 addSize(DT_FINI_ARRAYSZ, Out::finiArray);
1432 if (Symbol *b = symtab->find(config->init))
1435 if (Symbol *b = symtab->find(config->fini))
1440 if (part.verSym && part.verSym->isNeeded())
1441 addInSec(DT_VERSYM, part.verSym);
1442 if (part.verDef && part.verDef->isLive()) {
1443 addInSec(DT_VERDEF, part.verDef);
1444 addInt(DT_VERDEFNUM, getVerDefNum());
1446 if (part.verNeed && part.verNeed->isNeeded()) {
1447 addInSec(DT_VERNEED, part.verNeed);
1448 unsigned needNum = 0;
1449 for (SharedFile *f : sharedFiles)
1450 if (!f->vernauxs.empty())
1452 addInt(DT_VERNEEDNUM, needNum);
1455 if (config->emachine == EM_MIPS) {
1456 addInt(DT_MIPS_RLD_VERSION, 1);
1457 addInt(DT_MIPS_FLAGS, RHF_NOTPOT);
1458 addInt(DT_MIPS_BASE_ADDRESS, target->getImageBase());
1459 addInt(DT_MIPS_SYMTABNO, part.dynSymTab->getNumSymbols());
1461 add(DT_MIPS_LOCAL_GOTNO, [] { return in.mipsGot->getLocalEntriesNum(); });
1463 if (const Symbol *b = in.mipsGot->getFirstGlobalEntry())
1464 addInt(DT_MIPS_GOTSYM, b->dynsymIndex);
1466 addInt(DT_MIPS_GOTSYM, part.dynSymTab->getNumSymbols());
1467 addInSec(DT_PLTGOT, in.mipsGot);
1468 if (in.mipsRldMap) {
1470 addInSec(DT_MIPS_RLD_MAP, in.mipsRldMap);
1471 // Store the offset to the .rld_map section
1472 // relative to the address of the tag.
1473 addInSecRelative(DT_MIPS_RLD_MAP_REL, in.mipsRldMap);
1477 // DT_PPC_GOT indicates to glibc Secure PLT is used. If DT_PPC_GOT is absent,
1478 // glibc assumes the old-style BSS PLT layout which we don't support.
1479 if (config->emachine == EM_PPC)
1480 add(DT_PPC_GOT, [] { return in.got->getVA(); });
1482 // Glink dynamic tag is required by the V2 abi if the plt section isn't empty.
1483 if (config->emachine == EM_PPC64 && in.plt->isNeeded()) {
1484 // The Glink tag points to 32 bytes before the first lazy symbol resolution
1485 // stub, which starts directly after the header.
1486 entries.push_back({DT_PPC64_GLINK, [=] {
1487 unsigned offset = target->pltHeaderSize - 32;
1488 return in.plt->getVA(0) + offset;
1494 getParent()->link = this->link;
1495 this->size = entries.size() * this->entsize;
1498 template <class ELFT> void DynamicSection<ELFT>::writeTo(uint8_t *buf) {
1499 auto *p = reinterpret_cast<Elf_Dyn *>(buf);
1501 for (std::pair<int32_t, std::function<uint64_t()>> &kv : entries) {
1502 p->d_tag = kv.first;
1503 p->d_un.d_val = kv.second();
1508 uint64_t DynamicReloc::getOffset() const {
1509 return inputSec->getVA(offsetInSec);
1512 int64_t DynamicReloc::computeAddend() const {
1514 return sym->getVA(addend);
1517 // See the comment in the DynamicReloc ctor.
1518 return getMipsPageAddr(outputSec->addr) + addend;
1521 uint32_t DynamicReloc::getSymIndex(SymbolTableBaseSection *symTab) const {
1522 if (sym && !useSymVA)
1523 return symTab->getSymbolIndex(sym);
1527 RelocationBaseSection::RelocationBaseSection(StringRef name, uint32_t type,
1529 int32_t sizeDynamicTag)
1530 : SyntheticSection(SHF_ALLOC, type, config->wordsize, name),
1531 dynamicTag(dynamicTag), sizeDynamicTag(sizeDynamicTag) {}
1533 void RelocationBaseSection::addReloc(RelType dynType, InputSectionBase *isec,
1534 uint64_t offsetInSec, Symbol *sym) {
1535 addReloc({dynType, isec, offsetInSec, false, sym, 0});
1538 void RelocationBaseSection::addReloc(RelType dynType,
1539 InputSectionBase *inputSec,
1540 uint64_t offsetInSec, Symbol *sym,
1541 int64_t addend, RelExpr expr,
1543 // Write the addends to the relocated address if required. We skip
1544 // it if the written value would be zero.
1545 if (config->writeAddends && (expr != R_ADDEND || addend != 0))
1546 inputSec->relocations.push_back({expr, type, offsetInSec, addend, sym});
1547 addReloc({dynType, inputSec, offsetInSec, expr != R_ADDEND, sym, addend});
1550 void RelocationBaseSection::addReloc(const DynamicReloc &reloc) {
1551 if (reloc.type == target->relativeRel)
1552 ++numRelativeRelocs;
1553 relocs.push_back(reloc);
1556 void RelocationBaseSection::finalizeContents() {
1557 SymbolTableBaseSection *symTab = getPartition().dynSymTab;
1559 // When linking glibc statically, .rel{,a}.plt contains R_*_IRELATIVE
1560 // relocations due to IFUNC (e.g. strcpy). sh_link will be set to 0 in that
1562 if (symTab && symTab->getParent())
1563 getParent()->link = symTab->getParent()->sectionIndex;
1565 getParent()->link = 0;
1567 if (in.relaPlt == this)
1568 getParent()->info = in.gotPlt->getParent()->sectionIndex;
1569 if (in.relaIplt == this)
1570 getParent()->info = in.igotPlt->getParent()->sectionIndex;
1573 RelrBaseSection::RelrBaseSection()
1574 : SyntheticSection(SHF_ALLOC,
1575 config->useAndroidRelrTags ? SHT_ANDROID_RELR : SHT_RELR,
1576 config->wordsize, ".relr.dyn") {}
1578 template <class ELFT>
1579 static void encodeDynamicReloc(SymbolTableBaseSection *symTab,
1580 typename ELFT::Rela *p,
1581 const DynamicReloc &rel) {
1583 p->r_addend = rel.computeAddend();
1584 p->r_offset = rel.getOffset();
1585 p->setSymbolAndType(rel.getSymIndex(symTab), rel.type, config->isMips64EL);
1588 template <class ELFT>
1589 RelocationSection<ELFT>::RelocationSection(StringRef name, bool sort)
1590 : RelocationBaseSection(name, config->isRela ? SHT_RELA : SHT_REL,
1591 config->isRela ? DT_RELA : DT_REL,
1592 config->isRela ? DT_RELASZ : DT_RELSZ),
1594 this->entsize = config->isRela ? sizeof(Elf_Rela) : sizeof(Elf_Rel);
1597 template <class ELFT> void RelocationSection<ELFT>::writeTo(uint8_t *buf) {
1598 SymbolTableBaseSection *symTab = getPartition().dynSymTab;
1600 // Sort by (!IsRelative,SymIndex,r_offset). DT_REL[A]COUNT requires us to
1601 // place R_*_RELATIVE first. SymIndex is to improve locality, while r_offset
1602 // is to make results easier to read.
1605 relocs, [&](const DynamicReloc &a, const DynamicReloc &b) {
1606 return std::make_tuple(a.type != target->relativeRel,
1607 a.getSymIndex(symTab), a.getOffset()) <
1608 std::make_tuple(b.type != target->relativeRel,
1609 b.getSymIndex(symTab), b.getOffset());
1612 for (const DynamicReloc &rel : relocs) {
1613 encodeDynamicReloc<ELFT>(symTab, reinterpret_cast<Elf_Rela *>(buf), rel);
1614 buf += config->isRela ? sizeof(Elf_Rela) : sizeof(Elf_Rel);
1618 template <class ELFT>
1619 AndroidPackedRelocationSection<ELFT>::AndroidPackedRelocationSection(
1621 : RelocationBaseSection(
1622 name, config->isRela ? SHT_ANDROID_RELA : SHT_ANDROID_REL,
1623 config->isRela ? DT_ANDROID_RELA : DT_ANDROID_REL,
1624 config->isRela ? DT_ANDROID_RELASZ : DT_ANDROID_RELSZ) {
1628 template <class ELFT>
1629 bool AndroidPackedRelocationSection<ELFT>::updateAllocSize() {
1630 // This function computes the contents of an Android-format packed relocation
1633 // This format compresses relocations by using relocation groups to factor out
1634 // fields that are common between relocations and storing deltas from previous
1635 // relocations in SLEB128 format (which has a short representation for small
1636 // numbers). A good example of a relocation type with common fields is
1637 // R_*_RELATIVE, which is normally used to represent function pointers in
1638 // vtables. In the REL format, each relative relocation has the same r_info
1639 // field, and is only different from other relative relocations in terms of
1640 // the r_offset field. By sorting relocations by offset, grouping them by
1641 // r_info and representing each relocation with only the delta from the
1642 // previous offset, each 8-byte relocation can be compressed to as little as 1
1643 // byte (or less with run-length encoding). This relocation packer was able to
1644 // reduce the size of the relocation section in an Android Chromium DSO from
1645 // 2,911,184 bytes to 174,693 bytes, or 6% of the original size.
1647 // A relocation section consists of a header containing the literal bytes
1648 // 'APS2' followed by a sequence of SLEB128-encoded integers. The first two
1649 // elements are the total number of relocations in the section and an initial
1650 // r_offset value. The remaining elements define a sequence of relocation
1651 // groups. Each relocation group starts with a header consisting of the
1652 // following elements:
1654 // - the number of relocations in the relocation group
1655 // - flags for the relocation group
1656 // - (if RELOCATION_GROUPED_BY_OFFSET_DELTA_FLAG is set) the r_offset delta
1657 // for each relocation in the group.
1658 // - (if RELOCATION_GROUPED_BY_INFO_FLAG is set) the value of the r_info
1659 // field for each relocation in the group.
1660 // - (if RELOCATION_GROUP_HAS_ADDEND_FLAG and
1661 // RELOCATION_GROUPED_BY_ADDEND_FLAG are set) the r_addend delta for
1662 // each relocation in the group.
1664 // Following the relocation group header are descriptions of each of the
1665 // relocations in the group. They consist of the following elements:
1667 // - (if RELOCATION_GROUPED_BY_OFFSET_DELTA_FLAG is not set) the r_offset
1668 // delta for this relocation.
1669 // - (if RELOCATION_GROUPED_BY_INFO_FLAG is not set) the value of the r_info
1670 // field for this relocation.
1671 // - (if RELOCATION_GROUP_HAS_ADDEND_FLAG is set and
1672 // RELOCATION_GROUPED_BY_ADDEND_FLAG is not set) the r_addend delta for
1675 size_t oldSize = relocData.size();
1677 relocData = {'A', 'P', 'S', '2'};
1678 raw_svector_ostream os(relocData);
1679 auto add = [&](int64_t v) { encodeSLEB128(v, os); };
1681 // The format header includes the number of relocations and the initial
1682 // offset (we set this to zero because the first relocation group will
1683 // perform the initial adjustment).
1687 std::vector<Elf_Rela> relatives, nonRelatives;
1689 for (const DynamicReloc &rel : relocs) {
1691 encodeDynamicReloc<ELFT>(getPartition().dynSymTab, &r, rel);
1693 if (r.getType(config->isMips64EL) == target->relativeRel)
1694 relatives.push_back(r);
1696 nonRelatives.push_back(r);
1699 llvm::sort(relatives, [](const Elf_Rel &a, const Elf_Rel &b) {
1700 return a.r_offset < b.r_offset;
1703 // Try to find groups of relative relocations which are spaced one word
1704 // apart from one another. These generally correspond to vtable entries. The
1705 // format allows these groups to be encoded using a sort of run-length
1706 // encoding, but each group will cost 7 bytes in addition to the offset from
1707 // the previous group, so it is only profitable to do this for groups of
1708 // size 8 or larger.
1709 std::vector<Elf_Rela> ungroupedRelatives;
1710 std::vector<std::vector<Elf_Rela>> relativeGroups;
1711 for (auto i = relatives.begin(), e = relatives.end(); i != e;) {
1712 std::vector<Elf_Rela> group;
1714 group.push_back(*i++);
1715 } while (i != e && (i - 1)->r_offset + config->wordsize == i->r_offset);
1717 if (group.size() < 8)
1718 ungroupedRelatives.insert(ungroupedRelatives.end(), group.begin(),
1721 relativeGroups.emplace_back(std::move(group));
1724 // For non-relative relocations, we would like to:
1725 // 1. Have relocations with the same symbol offset to be consecutive, so
1726 // that the runtime linker can speed-up symbol lookup by implementing an
1728 // 2. Group relocations by r_info to reduce the size of the relocation
1730 // Since the symbol offset is the high bits in r_info, sorting by r_info
1731 // allows us to do both.
1733 // For Rela, we also want to sort by r_addend when r_info is the same. This
1734 // enables us to group by r_addend as well.
1735 llvm::stable_sort(nonRelatives, [](const Elf_Rela &a, const Elf_Rela &b) {
1736 if (a.r_info != b.r_info)
1737 return a.r_info < b.r_info;
1739 return a.r_addend < b.r_addend;
1743 // Group relocations with the same r_info. Note that each group emits a group
1744 // header and that may make the relocation section larger. It is hard to
1745 // estimate the size of a group header as the encoded size of that varies
1746 // based on r_info. However, we can approximate this trade-off by the number
1747 // of values encoded. Each group header contains 3 values, and each relocation
1748 // in a group encodes one less value, as compared to when it is not grouped.
1749 // Therefore, we only group relocations if there are 3 or more of them with
1752 // For Rela, the addend for most non-relative relocations is zero, and thus we
1753 // can usually get a smaller relocation section if we group relocations with 0
1755 std::vector<Elf_Rela> ungroupedNonRelatives;
1756 std::vector<std::vector<Elf_Rela>> nonRelativeGroups;
1757 for (auto i = nonRelatives.begin(), e = nonRelatives.end(); i != e;) {
1759 while (j != e && i->r_info == j->r_info &&
1760 (!config->isRela || i->r_addend == j->r_addend))
1762 if (j - i < 3 || (config->isRela && i->r_addend != 0))
1763 ungroupedNonRelatives.insert(ungroupedNonRelatives.end(), i, j);
1765 nonRelativeGroups.emplace_back(i, j);
1769 // Sort ungrouped relocations by offset to minimize the encoded length.
1770 llvm::sort(ungroupedNonRelatives, [](const Elf_Rela &a, const Elf_Rela &b) {
1771 return a.r_offset < b.r_offset;
1774 unsigned hasAddendIfRela =
1775 config->isRela ? RELOCATION_GROUP_HAS_ADDEND_FLAG : 0;
1777 uint64_t offset = 0;
1778 uint64_t addend = 0;
1780 // Emit the run-length encoding for the groups of adjacent relative
1781 // relocations. Each group is represented using two groups in the packed
1782 // format. The first is used to set the current offset to the start of the
1783 // group (and also encodes the first relocation), and the second encodes the
1784 // remaining relocations.
1785 for (std::vector<Elf_Rela> &g : relativeGroups) {
1786 // The first relocation in the group.
1788 add(RELOCATION_GROUPED_BY_OFFSET_DELTA_FLAG |
1789 RELOCATION_GROUPED_BY_INFO_FLAG | hasAddendIfRela);
1790 add(g[0].r_offset - offset);
1791 add(target->relativeRel);
1792 if (config->isRela) {
1793 add(g[0].r_addend - addend);
1794 addend = g[0].r_addend;
1797 // The remaining relocations.
1799 add(RELOCATION_GROUPED_BY_OFFSET_DELTA_FLAG |
1800 RELOCATION_GROUPED_BY_INFO_FLAG | hasAddendIfRela);
1801 add(config->wordsize);
1802 add(target->relativeRel);
1803 if (config->isRela) {
1804 for (auto i = g.begin() + 1, e = g.end(); i != e; ++i) {
1805 add(i->r_addend - addend);
1806 addend = i->r_addend;
1810 offset = g.back().r_offset;
1813 // Now the ungrouped relatives.
1814 if (!ungroupedRelatives.empty()) {
1815 add(ungroupedRelatives.size());
1816 add(RELOCATION_GROUPED_BY_INFO_FLAG | hasAddendIfRela);
1817 add(target->relativeRel);
1818 for (Elf_Rela &r : ungroupedRelatives) {
1819 add(r.r_offset - offset);
1820 offset = r.r_offset;
1821 if (config->isRela) {
1822 add(r.r_addend - addend);
1823 addend = r.r_addend;
1828 // Grouped non-relatives.
1829 for (ArrayRef<Elf_Rela> g : nonRelativeGroups) {
1831 add(RELOCATION_GROUPED_BY_INFO_FLAG);
1833 for (const Elf_Rela &r : g) {
1834 add(r.r_offset - offset);
1835 offset = r.r_offset;
1840 // Finally the ungrouped non-relative relocations.
1841 if (!ungroupedNonRelatives.empty()) {
1842 add(ungroupedNonRelatives.size());
1843 add(hasAddendIfRela);
1844 for (Elf_Rela &r : ungroupedNonRelatives) {
1845 add(r.r_offset - offset);
1846 offset = r.r_offset;
1848 if (config->isRela) {
1849 add(r.r_addend - addend);
1850 addend = r.r_addend;
1855 // Don't allow the section to shrink; otherwise the size of the section can
1856 // oscillate infinitely.
1857 if (relocData.size() < oldSize)
1858 relocData.append(oldSize - relocData.size(), 0);
1860 // Returns whether the section size changed. We need to keep recomputing both
1861 // section layout and the contents of this section until the size converges
1862 // because changing this section's size can affect section layout, which in
1863 // turn can affect the sizes of the LEB-encoded integers stored in this
1865 return relocData.size() != oldSize;
1868 template <class ELFT> RelrSection<ELFT>::RelrSection() {
1869 this->entsize = config->wordsize;
1872 template <class ELFT> bool RelrSection<ELFT>::updateAllocSize() {
1873 // This function computes the contents of an SHT_RELR packed relocation
1876 // Proposal for adding SHT_RELR sections to generic-abi is here:
1877 // https://groups.google.com/forum/#!topic/generic-abi/bX460iggiKg
1879 // The encoded sequence of Elf64_Relr entries in a SHT_RELR section looks
1880 // like [ AAAAAAAA BBBBBBB1 BBBBBBB1 ... AAAAAAAA BBBBBB1 ... ]
1882 // i.e. start with an address, followed by any number of bitmaps. The address
1883 // entry encodes 1 relocation. The subsequent bitmap entries encode up to 63
1884 // relocations each, at subsequent offsets following the last address entry.
1886 // The bitmap entries must have 1 in the least significant bit. The assumption
1887 // here is that an address cannot have 1 in lsb. Odd addresses are not
1890 // Excluding the least significant bit in the bitmap, each non-zero bit in
1891 // the bitmap represents a relocation to be applied to a corresponding machine
1892 // word that follows the base address word. The second least significant bit
1893 // represents the machine word immediately following the initial address, and
1894 // each bit that follows represents the next word, in linear order. As such,
1895 // a single bitmap can encode up to 31 relocations in a 32-bit object, and
1896 // 63 relocations in a 64-bit object.
1898 // This encoding has a couple of interesting properties:
1899 // 1. Looking at any entry, it is clear whether it's an address or a bitmap:
1900 // even means address, odd means bitmap.
1901 // 2. Just a simple list of addresses is a valid encoding.
1903 size_t oldSize = relrRelocs.size();
1906 // Same as Config->Wordsize but faster because this is a compile-time
1908 const size_t wordsize = sizeof(typename ELFT::uint);
1910 // Number of bits to use for the relocation offsets bitmap.
1911 // Must be either 63 or 31.
1912 const size_t nBits = wordsize * 8 - 1;
1914 // Get offsets for all relative relocations and sort them.
1915 std::vector<uint64_t> offsets;
1916 for (const RelativeReloc &rel : relocs)
1917 offsets.push_back(rel.getOffset());
1918 llvm::sort(offsets);
1920 // For each leading relocation, find following ones that can be folded
1921 // as a bitmap and fold them.
1922 for (size_t i = 0, e = offsets.size(); i < e;) {
1923 // Add a leading relocation.
1924 relrRelocs.push_back(Elf_Relr(offsets[i]));
1925 uint64_t base = offsets[i] + wordsize;
1928 // Find foldable relocations to construct bitmaps.
1930 uint64_t bitmap = 0;
1933 uint64_t delta = offsets[i] - base;
1935 // If it is too far, it cannot be folded.
1936 if (delta >= nBits * wordsize)
1939 // If it is not a multiple of wordsize away, it cannot be folded.
1940 if (delta % wordsize)
1944 bitmap |= 1ULL << (delta / wordsize);
1951 relrRelocs.push_back(Elf_Relr((bitmap << 1) | 1));
1952 base += nBits * wordsize;
1956 // Don't allow the section to shrink; otherwise the size of the section can
1957 // oscillate infinitely. Trailing 1s do not decode to more relocations.
1958 if (relrRelocs.size() < oldSize) {
1959 log(".relr.dyn needs " + Twine(oldSize - relrRelocs.size()) +
1960 " padding word(s)");
1961 relrRelocs.resize(oldSize, Elf_Relr(1));
1964 return relrRelocs.size() != oldSize;
1967 SymbolTableBaseSection::SymbolTableBaseSection(StringTableSection &strTabSec)
1968 : SyntheticSection(strTabSec.isDynamic() ? (uint64_t)SHF_ALLOC : 0,
1969 strTabSec.isDynamic() ? SHT_DYNSYM : SHT_SYMTAB,
1971 strTabSec.isDynamic() ? ".dynsym" : ".symtab"),
1972 strTabSec(strTabSec) {}
1974 // Orders symbols according to their positions in the GOT,
1975 // in compliance with MIPS ABI rules.
1976 // See "Global Offset Table" in Chapter 5 in the following document
1977 // for detailed description:
1978 // ftp://www.linux-mips.org/pub/linux/mips/doc/ABI/mipsabi.pdf
1979 static bool sortMipsSymbols(const SymbolTableEntry &l,
1980 const SymbolTableEntry &r) {
1981 // Sort entries related to non-local preemptible symbols by GOT indexes.
1982 // All other entries go to the beginning of a dynsym in arbitrary order.
1983 if (l.sym->isInGot() && r.sym->isInGot())
1984 return l.sym->gotIndex < r.sym->gotIndex;
1985 if (!l.sym->isInGot() && !r.sym->isInGot())
1987 return !l.sym->isInGot();
1990 void SymbolTableBaseSection::finalizeContents() {
1991 if (OutputSection *sec = strTabSec.getParent())
1992 getParent()->link = sec->sectionIndex;
1994 if (this->type != SHT_DYNSYM) {
1995 sortSymTabSymbols();
1999 // If it is a .dynsym, there should be no local symbols, but we need
2000 // to do a few things for the dynamic linker.
2002 // Section's Info field has the index of the first non-local symbol.
2003 // Because the first symbol entry is a null entry, 1 is the first.
2004 getParent()->info = 1;
2006 if (getPartition().gnuHashTab) {
2007 // NB: It also sorts Symbols to meet the GNU hash table requirements.
2008 getPartition().gnuHashTab->addSymbols(symbols);
2009 } else if (config->emachine == EM_MIPS) {
2010 llvm::stable_sort(symbols, sortMipsSymbols);
2013 // Only the main partition's dynsym indexes are stored in the symbols
2014 // themselves. All other partitions use a lookup table.
2015 if (this == mainPart->dynSymTab) {
2017 for (const SymbolTableEntry &s : symbols)
2018 s.sym->dynsymIndex = ++i;
2022 // The ELF spec requires that all local symbols precede global symbols, so we
2023 // sort symbol entries in this function. (For .dynsym, we don't do that because
2024 // symbols for dynamic linking are inherently all globals.)
2026 // Aside from above, we put local symbols in groups starting with the STT_FILE
2027 // symbol. That is convenient for purpose of identifying where are local symbols
2029 void SymbolTableBaseSection::sortSymTabSymbols() {
2030 // Move all local symbols before global symbols.
2031 auto e = std::stable_partition(
2032 symbols.begin(), symbols.end(), [](const SymbolTableEntry &s) {
2033 return s.sym->isLocal() || s.sym->computeBinding() == STB_LOCAL;
2035 size_t numLocals = e - symbols.begin();
2036 getParent()->info = numLocals + 1;
2038 // We want to group the local symbols by file. For that we rebuild the local
2039 // part of the symbols vector. We do not need to care about the STT_FILE
2040 // symbols, they are already naturally placed first in each group. That
2041 // happens because STT_FILE is always the first symbol in the object and hence
2042 // precede all other local symbols we add for a file.
2043 MapVector<InputFile *, std::vector<SymbolTableEntry>> arr;
2044 for (const SymbolTableEntry &s : llvm::make_range(symbols.begin(), e))
2045 arr[s.sym->file].push_back(s);
2047 auto i = symbols.begin();
2048 for (std::pair<InputFile *, std::vector<SymbolTableEntry>> &p : arr)
2049 for (SymbolTableEntry &entry : p.second)
2053 void SymbolTableBaseSection::addSymbol(Symbol *b) {
2054 // Adding a local symbol to a .dynsym is a bug.
2055 assert(this->type != SHT_DYNSYM || !b->isLocal());
2057 bool hashIt = b->isLocal();
2058 symbols.push_back({b, strTabSec.addString(b->getName(), hashIt)});
2061 size_t SymbolTableBaseSection::getSymbolIndex(Symbol *sym) {
2062 if (this == mainPart->dynSymTab)
2063 return sym->dynsymIndex;
2065 // Initializes symbol lookup tables lazily. This is used only for -r,
2066 // -emit-relocs and dynsyms in partitions other than the main one.
2067 llvm::call_once(onceFlag, [&] {
2068 symbolIndexMap.reserve(symbols.size());
2070 for (const SymbolTableEntry &e : symbols) {
2071 if (e.sym->type == STT_SECTION)
2072 sectionIndexMap[e.sym->getOutputSection()] = ++i;
2074 symbolIndexMap[e.sym] = ++i;
2078 // Section symbols are mapped based on their output sections
2079 // to maintain their semantics.
2080 if (sym->type == STT_SECTION)
2081 return sectionIndexMap.lookup(sym->getOutputSection());
2082 return symbolIndexMap.lookup(sym);
2085 template <class ELFT>
2086 SymbolTableSection<ELFT>::SymbolTableSection(StringTableSection &strTabSec)
2087 : SymbolTableBaseSection(strTabSec) {
2088 this->entsize = sizeof(Elf_Sym);
2091 static BssSection *getCommonSec(Symbol *sym) {
2092 if (!config->defineCommon)
2093 if (auto *d = dyn_cast<Defined>(sym))
2094 return dyn_cast_or_null<BssSection>(d->section);
2098 static uint32_t getSymSectionIndex(Symbol *sym) {
2099 if (getCommonSec(sym))
2101 if (!isa<Defined>(sym) || sym->needsPltAddr)
2103 if (const OutputSection *os = sym->getOutputSection())
2104 return os->sectionIndex >= SHN_LORESERVE ? (uint32_t)SHN_XINDEX
2109 // Write the internal symbol table contents to the output symbol table.
2110 template <class ELFT> void SymbolTableSection<ELFT>::writeTo(uint8_t *buf) {
2111 // The first entry is a null entry as per the ELF spec.
2112 memset(buf, 0, sizeof(Elf_Sym));
2113 buf += sizeof(Elf_Sym);
2115 auto *eSym = reinterpret_cast<Elf_Sym *>(buf);
2117 for (SymbolTableEntry &ent : symbols) {
2118 Symbol *sym = ent.sym;
2119 bool isDefinedHere = type == SHT_SYMTAB || sym->partition == partition;
2121 // Set st_info and st_other.
2123 if (sym->isLocal()) {
2124 eSym->setBindingAndType(STB_LOCAL, sym->type);
2126 eSym->setBindingAndType(sym->computeBinding(), sym->type);
2127 eSym->setVisibility(sym->visibility);
2130 // The 3 most significant bits of st_other are used by OpenPOWER ABI.
2131 // See getPPC64GlobalEntryToLocalEntryOffset() for more details.
2132 if (config->emachine == EM_PPC64)
2133 eSym->st_other |= sym->stOther & 0xe0;
2135 eSym->st_name = ent.strTabOffset;
2137 eSym->st_shndx = getSymSectionIndex(ent.sym);
2141 // Copy symbol size if it is a defined symbol. st_size is not significant
2142 // for undefined symbols, so whether copying it or not is up to us if that's
2143 // the case. We'll leave it as zero because by not setting a value, we can
2144 // get the exact same outputs for two sets of input files that differ only
2145 // in undefined symbol size in DSOs.
2146 if (eSym->st_shndx == SHN_UNDEF || !isDefinedHere)
2149 eSym->st_size = sym->getSize();
2151 // st_value is usually an address of a symbol, but that has a
2152 // special meaining for uninstantiated common symbols (this can
2153 // occur if -r is given).
2154 if (BssSection *commonSec = getCommonSec(ent.sym))
2155 eSym->st_value = commonSec->alignment;
2156 else if (isDefinedHere)
2157 eSym->st_value = sym->getVA();
2164 // On MIPS we need to mark symbol which has a PLT entry and requires
2165 // pointer equality by STO_MIPS_PLT flag. That is necessary to help
2166 // dynamic linker distinguish such symbols and MIPS lazy-binding stubs.
2167 // https://sourceware.org/ml/binutils/2008-07/txt00000.txt
2168 if (config->emachine == EM_MIPS) {
2169 auto *eSym = reinterpret_cast<Elf_Sym *>(buf);
2171 for (SymbolTableEntry &ent : symbols) {
2172 Symbol *sym = ent.sym;
2173 if (sym->isInPlt() && sym->needsPltAddr)
2174 eSym->st_other |= STO_MIPS_PLT;
2175 if (isMicroMips()) {
2176 // We already set the less-significant bit for symbols
2177 // marked by the `STO_MIPS_MICROMIPS` flag and for microMIPS PLT
2178 // records. That allows us to distinguish such symbols in
2179 // the `MIPS<ELFT>::relocateOne()` routine. Now we should
2180 // clear that bit for non-dynamic symbol table, so tools
2181 // like `objdump` will be able to deal with a correct
2183 if (sym->isDefined() &&
2184 ((sym->stOther & STO_MIPS_MICROMIPS) || sym->needsPltAddr)) {
2185 if (!strTabSec.isDynamic())
2186 eSym->st_value &= ~1;
2187 eSym->st_other |= STO_MIPS_MICROMIPS;
2190 if (config->relocatable)
2191 if (auto *d = dyn_cast<Defined>(sym))
2192 if (isMipsPIC<ELFT>(d))
2193 eSym->st_other |= STO_MIPS_PIC;
2199 SymtabShndxSection::SymtabShndxSection()
2200 : SyntheticSection(0, SHT_SYMTAB_SHNDX, 4, ".symtab_shndx") {
2204 void SymtabShndxSection::writeTo(uint8_t *buf) {
2205 // We write an array of 32 bit values, where each value has 1:1 association
2206 // with an entry in .symtab. If the corresponding entry contains SHN_XINDEX,
2207 // we need to write actual index, otherwise, we must write SHN_UNDEF(0).
2208 buf += 4; // Ignore .symtab[0] entry.
2209 for (const SymbolTableEntry &entry : in.symTab->getSymbols()) {
2210 if (getSymSectionIndex(entry.sym) == SHN_XINDEX)
2211 write32(buf, entry.sym->getOutputSection()->sectionIndex);
2216 bool SymtabShndxSection::isNeeded() const {
2217 // SHT_SYMTAB can hold symbols with section indices values up to
2218 // SHN_LORESERVE. If we need more, we want to use extension SHT_SYMTAB_SHNDX
2219 // section. Problem is that we reveal the final section indices a bit too
2220 // late, and we do not know them here. For simplicity, we just always create
2221 // a .symtab_shndx section when the amount of output sections is huge.
2223 for (BaseCommand *base : script->sectionCommands)
2224 if (isa<OutputSection>(base))
2226 return size >= SHN_LORESERVE;
2229 void SymtabShndxSection::finalizeContents() {
2230 getParent()->link = in.symTab->getParent()->sectionIndex;
2233 size_t SymtabShndxSection::getSize() const {
2234 return in.symTab->getNumSymbols() * 4;
2237 // .hash and .gnu.hash sections contain on-disk hash tables that map
2238 // symbol names to their dynamic symbol table indices. Their purpose
2239 // is to help the dynamic linker resolve symbols quickly. If ELF files
2240 // don't have them, the dynamic linker has to do linear search on all
2241 // dynamic symbols, which makes programs slower. Therefore, a .hash
2242 // section is added to a DSO by default. A .gnu.hash is added if you
2243 // give the -hash-style=gnu or -hash-style=both option.
2245 // The Unix semantics of resolving dynamic symbols is somewhat expensive.
2246 // Each ELF file has a list of DSOs that the ELF file depends on and a
2247 // list of dynamic symbols that need to be resolved from any of the
2248 // DSOs. That means resolving all dynamic symbols takes O(m)*O(n)
2249 // where m is the number of DSOs and n is the number of dynamic
2250 // symbols. For modern large programs, both m and n are large. So
2251 // making each step faster by using hash tables substiantially
2252 // improves time to load programs.
2254 // (Note that this is not the only way to design the shared library.
2255 // For instance, the Windows DLL takes a different approach. On
2256 // Windows, each dynamic symbol has a name of DLL from which the symbol
2257 // has to be resolved. That makes the cost of symbol resolution O(n).
2258 // This disables some hacky techniques you can use on Unix such as
2259 // LD_PRELOAD, but this is arguably better semantics than the Unix ones.)
2261 // Due to historical reasons, we have two different hash tables, .hash
2262 // and .gnu.hash. They are for the same purpose, and .gnu.hash is a new
2263 // and better version of .hash. .hash is just an on-disk hash table, but
2264 // .gnu.hash has a bloom filter in addition to a hash table to skip
2265 // DSOs very quickly. If you are sure that your dynamic linker knows
2266 // about .gnu.hash, you want to specify -hash-style=gnu. Otherwise, a
2267 // safe bet is to specify -hash-style=both for backward compatibility.
2268 GnuHashTableSection::GnuHashTableSection()
2269 : SyntheticSection(SHF_ALLOC, SHT_GNU_HASH, config->wordsize, ".gnu.hash") {
2272 void GnuHashTableSection::finalizeContents() {
2273 if (OutputSection *sec = getPartition().dynSymTab->getParent())
2274 getParent()->link = sec->sectionIndex;
2276 // Computes bloom filter size in word size. We want to allocate 12
2277 // bits for each symbol. It must be a power of two.
2278 if (symbols.empty()) {
2281 uint64_t numBits = symbols.size() * 12;
2282 maskWords = NextPowerOf2(numBits / (config->wordsize * 8));
2285 size = 16; // Header
2286 size += config->wordsize * maskWords; // Bloom filter
2287 size += nBuckets * 4; // Hash buckets
2288 size += symbols.size() * 4; // Hash values
2291 void GnuHashTableSection::writeTo(uint8_t *buf) {
2292 // The output buffer is not guaranteed to be zero-cleared because we pre-
2293 // fill executable sections with trap instructions. This is a precaution
2294 // for that case, which happens only when -no-rosegment is given.
2295 memset(buf, 0, size);
2298 write32(buf, nBuckets);
2299 write32(buf + 4, getPartition().dynSymTab->getNumSymbols() - symbols.size());
2300 write32(buf + 8, maskWords);
2301 write32(buf + 12, Shift2);
2304 // Write a bloom filter and a hash table.
2305 writeBloomFilter(buf);
2306 buf += config->wordsize * maskWords;
2307 writeHashTable(buf);
2310 // This function writes a 2-bit bloom filter. This bloom filter alone
2311 // usually filters out 80% or more of all symbol lookups [1].
2312 // The dynamic linker uses the hash table only when a symbol is not
2313 // filtered out by a bloom filter.
2315 // [1] Ulrich Drepper (2011), "How To Write Shared Libraries" (Ver. 4.1.2),
2316 // p.9, https://www.akkadia.org/drepper/dsohowto.pdf
2317 void GnuHashTableSection::writeBloomFilter(uint8_t *buf) {
2318 unsigned c = config->is64 ? 64 : 32;
2319 for (const Entry &sym : symbols) {
2320 // When C = 64, we choose a word with bits [6:...] and set 1 to two bits in
2321 // the word using bits [0:5] and [26:31].
2322 size_t i = (sym.hash / c) & (maskWords - 1);
2323 uint64_t val = readUint(buf + i * config->wordsize);
2324 val |= uint64_t(1) << (sym.hash % c);
2325 val |= uint64_t(1) << ((sym.hash >> Shift2) % c);
2326 writeUint(buf + i * config->wordsize, val);
2330 void GnuHashTableSection::writeHashTable(uint8_t *buf) {
2331 uint32_t *buckets = reinterpret_cast<uint32_t *>(buf);
2332 uint32_t oldBucket = -1;
2333 uint32_t *values = buckets + nBuckets;
2334 for (auto i = symbols.begin(), e = symbols.end(); i != e; ++i) {
2335 // Write a hash value. It represents a sequence of chains that share the
2336 // same hash modulo value. The last element of each chain is terminated by
2338 uint32_t hash = i->hash;
2339 bool isLastInChain = (i + 1) == e || i->bucketIdx != (i + 1)->bucketIdx;
2340 hash = isLastInChain ? hash | 1 : hash & ~1;
2341 write32(values++, hash);
2343 if (i->bucketIdx == oldBucket)
2345 // Write a hash bucket. Hash buckets contain indices in the following hash
2347 write32(buckets + i->bucketIdx,
2348 getPartition().dynSymTab->getSymbolIndex(i->sym));
2349 oldBucket = i->bucketIdx;
2353 static uint32_t hashGnu(StringRef name) {
2355 for (uint8_t c : name)
2356 h = (h << 5) + h + c;
2360 // Add symbols to this symbol hash table. Note that this function
2361 // destructively sort a given vector -- which is needed because
2362 // GNU-style hash table places some sorting requirements.
2363 void GnuHashTableSection::addSymbols(std::vector<SymbolTableEntry> &v) {
2364 // We cannot use 'auto' for Mid because GCC 6.1 cannot deduce
2365 // its type correctly.
2366 std::vector<SymbolTableEntry>::iterator mid =
2367 std::stable_partition(v.begin(), v.end(), [&](const SymbolTableEntry &s) {
2368 return !s.sym->isDefined() || s.sym->partition != partition;
2371 // We chose load factor 4 for the on-disk hash table. For each hash
2372 // collision, the dynamic linker will compare a uint32_t hash value.
2373 // Since the integer comparison is quite fast, we believe we can
2374 // make the load factor even larger. 4 is just a conservative choice.
2376 // Note that we don't want to create a zero-sized hash table because
2377 // Android loader as of 2018 doesn't like a .gnu.hash containing such
2378 // table. If that's the case, we create a hash table with one unused
2380 nBuckets = std::max<size_t>((v.end() - mid) / 4, 1);
2385 for (SymbolTableEntry &ent : llvm::make_range(mid, v.end())) {
2386 Symbol *b = ent.sym;
2387 uint32_t hash = hashGnu(b->getName());
2388 uint32_t bucketIdx = hash % nBuckets;
2389 symbols.push_back({b, ent.strTabOffset, hash, bucketIdx});
2392 llvm::stable_sort(symbols, [](const Entry &l, const Entry &r) {
2393 return l.bucketIdx < r.bucketIdx;
2396 v.erase(mid, v.end());
2397 for (const Entry &ent : symbols)
2398 v.push_back({ent.sym, ent.strTabOffset});
2401 HashTableSection::HashTableSection()
2402 : SyntheticSection(SHF_ALLOC, SHT_HASH, 4, ".hash") {
2406 void HashTableSection::finalizeContents() {
2407 SymbolTableBaseSection *symTab = getPartition().dynSymTab;
2409 if (OutputSection *sec = symTab->getParent())
2410 getParent()->link = sec->sectionIndex;
2412 unsigned numEntries = 2; // nbucket and nchain.
2413 numEntries += symTab->getNumSymbols(); // The chain entries.
2415 // Create as many buckets as there are symbols.
2416 numEntries += symTab->getNumSymbols();
2417 this->size = numEntries * 4;
2420 void HashTableSection::writeTo(uint8_t *buf) {
2421 SymbolTableBaseSection *symTab = getPartition().dynSymTab;
2423 // See comment in GnuHashTableSection::writeTo.
2424 memset(buf, 0, size);
2426 unsigned numSymbols = symTab->getNumSymbols();
2428 uint32_t *p = reinterpret_cast<uint32_t *>(buf);
2429 write32(p++, numSymbols); // nbucket
2430 write32(p++, numSymbols); // nchain
2432 uint32_t *buckets = p;
2433 uint32_t *chains = p + numSymbols;
2435 for (const SymbolTableEntry &s : symTab->getSymbols()) {
2436 Symbol *sym = s.sym;
2437 StringRef name = sym->getName();
2438 unsigned i = sym->dynsymIndex;
2439 uint32_t hash = hashSysV(name) % numSymbols;
2440 chains[i] = buckets[hash];
2441 write32(buckets + hash, i);
2445 PltSection::PltSection()
2446 : SyntheticSection(SHF_ALLOC | SHF_EXECINSTR, SHT_PROGBITS, 16, ".plt"),
2447 headerSize(target->pltHeaderSize) {
2448 // On PowerPC, this section contains lazy symbol resolvers.
2449 if (config->emachine == EM_PPC || config->emachine == EM_PPC64) {
2452 // PLTresolve is at the end.
2453 if (config->emachine == EM_PPC)
2457 // On x86 when IBT is enabled, this section contains the second PLT (lazy
2458 // symbol resolvers).
2459 if ((config->emachine == EM_386 || config->emachine == EM_X86_64) &&
2460 (config->andFeatures & GNU_PROPERTY_X86_FEATURE_1_IBT))
2463 // The PLT needs to be writable on SPARC as the dynamic linker will
2464 // modify the instructions in the PLT entries.
2465 if (config->emachine == EM_SPARCV9)
2466 this->flags |= SHF_WRITE;
2469 void PltSection::writeTo(uint8_t *buf) {
2470 if (config->emachine == EM_PPC) {
2471 writePPC32GlinkSection(buf, entries.size());
2475 // At beginning of PLT, we have code to call the dynamic
2476 // linker to resolve dynsyms at runtime. Write such code.
2477 target->writePltHeader(buf);
2478 size_t off = headerSize;
2480 for (const Symbol *sym : entries) {
2481 target->writePlt(buf + off, *sym, getVA() + off);
2482 off += target->pltEntrySize;
2486 void PltSection::addEntry(Symbol &sym) {
2487 sym.pltIndex = entries.size();
2488 entries.push_back(&sym);
2491 size_t PltSection::getSize() const {
2492 return headerSize + entries.size() * target->pltEntrySize + footerSize;
2495 bool PltSection::isNeeded() const {
2496 // For -z retpolineplt, .iplt needs the .plt header.
2497 return !entries.empty() || (config->zRetpolineplt && in.iplt->isNeeded());
2500 // Used by ARM to add mapping symbols in the PLT section, which aid
2502 void PltSection::addSymbols() {
2503 target->addPltHeaderSymbols(*this);
2505 size_t off = headerSize;
2506 for (size_t i = 0; i < entries.size(); ++i) {
2507 target->addPltSymbols(*this, off);
2508 off += target->pltEntrySize;
2512 IpltSection::IpltSection()
2513 : SyntheticSection(SHF_ALLOC | SHF_EXECINSTR, SHT_PROGBITS, 16, ".iplt") {
2514 if (config->emachine == EM_PPC || config->emachine == EM_PPC64) {
2520 void IpltSection::writeTo(uint8_t *buf) {
2522 for (const Symbol *sym : entries) {
2523 target->writeIplt(buf + off, *sym, getVA() + off);
2524 off += target->ipltEntrySize;
2528 size_t IpltSection::getSize() const {
2529 return entries.size() * target->ipltEntrySize;
2532 void IpltSection::addEntry(Symbol &sym) {
2533 sym.pltIndex = entries.size();
2534 entries.push_back(&sym);
2537 // ARM uses mapping symbols to aid disassembly.
2538 void IpltSection::addSymbols() {
2540 for (size_t i = 0, e = entries.size(); i != e; ++i) {
2541 target->addPltSymbols(*this, off);
2542 off += target->pltEntrySize;
2546 // This is an x86-only extra PLT section and used only when a security
2547 // enhancement feature called CET is enabled. In this comment, I'll explain what
2548 // the feature is and why we have two PLT sections if CET is enabled.
2550 // So, what does CET do? CET introduces a new restriction to indirect jump
2551 // instructions. CET works this way. Assume that CET is enabled. Then, if you
2552 // execute an indirect jump instruction, the processor verifies that a special
2553 // "landing pad" instruction (which is actually a repurposed NOP instruction and
2554 // now called "endbr32" or "endbr64") is at the jump target. If the jump target
2555 // does not start with that instruction, the processor raises an exception
2556 // instead of continuing executing code.
2558 // If CET is enabled, the compiler emits endbr to all locations where indirect
2559 // jumps may jump to.
2561 // This mechanism makes it extremely hard to transfer the control to a middle of
2562 // a function that is not supporsed to be a indirect jump target, preventing
2563 // certain types of attacks such as ROP or JOP.
2565 // Note that the processors in the market as of 2019 don't actually support the
2566 // feature. Only the spec is available at the moment.
2568 // Now, I'll explain why we have this extra PLT section for CET.
2570 // Since you can indirectly jump to a PLT entry, we have to make PLT entries
2571 // start with endbr. The problem is there's no extra space for endbr (which is 4
2572 // bytes long), as the PLT entry is only 16 bytes long and all bytes are already
2575 // In order to deal with the issue, we split a PLT entry into two PLT entries.
2576 // Remember that each PLT entry contains code to jump to an address read from
2577 // .got.plt AND code to resolve a dynamic symbol lazily. With the 2-PLT scheme,
2578 // the former code is written to .plt.sec, and the latter code is written to
2581 // Lazy symbol resolution in the 2-PLT scheme works in the usual way, except
2582 // that the regular .plt is now called .plt.sec and .plt is repurposed to
2583 // contain only code for lazy symbol resolution.
2585 // In other words, this is how the 2-PLT scheme works. Application code is
2586 // supposed to jump to .plt.sec to call an external function. Each .plt.sec
2587 // entry contains code to read an address from a corresponding .got.plt entry
2588 // and jump to that address. Addresses in .got.plt initially point to .plt, so
2589 // when an application calls an external function for the first time, the
2590 // control is transferred to a function that resolves a symbol name from
2591 // external shared object files. That function then rewrites a .got.plt entry
2592 // with a resolved address, so that the subsequent function calls directly jump
2593 // to a desired location from .plt.sec.
2595 // There is an open question as to whether the 2-PLT scheme was desirable or
2596 // not. We could have simply extended the PLT entry size to 32-bytes to
2597 // accommodate endbr, and that scheme would have been much simpler than the
2598 // 2-PLT scheme. One reason to split PLT was, by doing that, we could keep hot
2599 // code (.plt.sec) from cold code (.plt). But as far as I know no one proved
2600 // that the optimization actually makes a difference.
2602 // That said, the 2-PLT scheme is a part of the ABI, debuggers and other tools
2603 // depend on it, so we implement the ABI.
2604 IBTPltSection::IBTPltSection()
2605 : SyntheticSection(SHF_ALLOC | SHF_EXECINSTR, SHT_PROGBITS, 16, ".plt") {}
2607 void IBTPltSection::writeTo(uint8_t *buf) {
2608 target->writeIBTPlt(buf, in.plt->getNumEntries());
2611 size_t IBTPltSection::getSize() const {
2612 // 16 is the header size of .plt.
2613 return 16 + in.plt->getNumEntries() * target->pltEntrySize;
2616 // The string hash function for .gdb_index.
2617 static uint32_t computeGdbHash(StringRef s) {
2620 h = h * 67 + toLower(c) - 113;
2624 GdbIndexSection::GdbIndexSection()
2625 : SyntheticSection(0, SHT_PROGBITS, 1, ".gdb_index") {}
2627 // Returns the desired size of an on-disk hash table for a .gdb_index section.
2628 // There's a tradeoff between size and collision rate. We aim 75% utilization.
2629 size_t GdbIndexSection::computeSymtabSize() const {
2630 return std::max<size_t>(NextPowerOf2(symbols.size() * 4 / 3), 1024);
2633 // Compute the output section size.
2634 void GdbIndexSection::initOutputSize() {
2635 size = sizeof(GdbIndexHeader) + computeSymtabSize() * 8;
2637 for (GdbChunk &chunk : chunks)
2638 size += chunk.compilationUnits.size() * 16 + chunk.addressAreas.size() * 20;
2640 // Add the constant pool size if exists.
2641 if (!symbols.empty()) {
2642 GdbSymbol &sym = symbols.back();
2643 size += sym.nameOff + sym.name.size() + 1;
2647 static std::vector<InputSection *> getDebugInfoSections() {
2648 std::vector<InputSection *> ret;
2649 for (InputSectionBase *s : inputSections)
2650 if (InputSection *isec = dyn_cast<InputSection>(s))
2651 if (isec->name == ".debug_info")
2652 ret.push_back(isec);
2656 static std::vector<GdbIndexSection::CuEntry> readCuList(DWARFContext &dwarf) {
2657 std::vector<GdbIndexSection::CuEntry> ret;
2658 for (std::unique_ptr<DWARFUnit> &cu : dwarf.compile_units())
2659 ret.push_back({cu->getOffset(), cu->getLength() + 4});
2663 static std::vector<GdbIndexSection::AddressEntry>
2664 readAddressAreas(DWARFContext &dwarf, InputSection *sec) {
2665 std::vector<GdbIndexSection::AddressEntry> ret;
2668 for (std::unique_ptr<DWARFUnit> &cu : dwarf.compile_units()) {
2669 if (Error e = cu->tryExtractDIEsIfNeeded(false)) {
2670 error(toString(sec) + ": " + toString(std::move(e)));
2673 Expected<DWARFAddressRangesVector> ranges = cu->collectAddressRanges();
2675 error(toString(sec) + ": " + toString(ranges.takeError()));
2679 ArrayRef<InputSectionBase *> sections = sec->file->getSections();
2680 for (DWARFAddressRange &r : *ranges) {
2681 if (r.SectionIndex == -1ULL)
2683 InputSectionBase *s = sections[r.SectionIndex];
2684 if (!s || s == &InputSection::discarded || !s->isLive())
2686 // Range list with zero size has no effect.
2687 if (r.LowPC == r.HighPC)
2689 auto *isec = cast<InputSection>(s);
2690 uint64_t offset = isec->getOffsetInFile();
2691 ret.push_back({isec, r.LowPC - offset, r.HighPC - offset, cuIdx});
2699 template <class ELFT>
2700 static std::vector<GdbIndexSection::NameAttrEntry>
2701 readPubNamesAndTypes(const LLDDwarfObj<ELFT> &obj,
2702 const std::vector<GdbIndexSection::CuEntry> &cus) {
2703 const DWARFSection &pubNames = obj.getGnuPubnamesSection();
2704 const DWARFSection &pubTypes = obj.getGnuPubtypesSection();
2706 std::vector<GdbIndexSection::NameAttrEntry> ret;
2707 for (const DWARFSection *pub : {&pubNames, &pubTypes}) {
2708 DWARFDebugPubTable table(obj, *pub, config->isLE, true);
2709 for (const DWARFDebugPubTable::Set &set : table.getData()) {
2710 // The value written into the constant pool is kind << 24 | cuIndex. As we
2711 // don't know how many compilation units precede this object to compute
2712 // cuIndex, we compute (kind << 24 | cuIndexInThisObject) instead, and add
2713 // the number of preceding compilation units later.
2714 uint32_t i = llvm::partition_point(cus,
2715 [&](GdbIndexSection::CuEntry cu) {
2716 return cu.cuOffset < set.Offset;
2719 for (const DWARFDebugPubTable::Entry &ent : set.Entries)
2720 ret.push_back({{ent.Name, computeGdbHash(ent.Name)},
2721 (ent.Descriptor.toBits() << 24) | i});
2727 // Create a list of symbols from a given list of symbol names and types
2728 // by uniquifying them by name.
2729 static std::vector<GdbIndexSection::GdbSymbol>
2730 createSymbols(ArrayRef<std::vector<GdbIndexSection::NameAttrEntry>> nameAttrs,
2731 const std::vector<GdbIndexSection::GdbChunk> &chunks) {
2732 using GdbSymbol = GdbIndexSection::GdbSymbol;
2733 using NameAttrEntry = GdbIndexSection::NameAttrEntry;
2735 // For each chunk, compute the number of compilation units preceding it.
2737 std::vector<uint32_t> cuIdxs(chunks.size());
2738 for (uint32_t i = 0, e = chunks.size(); i != e; ++i) {
2740 cuIdx += chunks[i].compilationUnits.size();
2743 // The number of symbols we will handle in this function is of the order
2744 // of millions for very large executables, so we use multi-threading to
2746 size_t numShards = 32;
2747 size_t concurrency = 1;
2750 std::min<size_t>(PowerOf2Floor(hardware_concurrency()), numShards);
2752 // A sharded map to uniquify symbols by name.
2753 std::vector<DenseMap<CachedHashStringRef, size_t>> map(numShards);
2754 size_t shift = 32 - countTrailingZeros(numShards);
2756 // Instantiate GdbSymbols while uniqufying them by name.
2757 std::vector<std::vector<GdbSymbol>> symbols(numShards);
2758 parallelForEachN(0, concurrency, [&](size_t threadId) {
2760 for (ArrayRef<NameAttrEntry> entries : nameAttrs) {
2761 for (const NameAttrEntry &ent : entries) {
2762 size_t shardId = ent.name.hash() >> shift;
2763 if ((shardId & (concurrency - 1)) != threadId)
2766 uint32_t v = ent.cuIndexAndAttrs + cuIdxs[i];
2767 size_t &idx = map[shardId][ent.name];
2769 symbols[shardId][idx - 1].cuVector.push_back(v);
2773 idx = symbols[shardId].size() + 1;
2774 symbols[shardId].push_back({ent.name, {v}, 0, 0});
2780 size_t numSymbols = 0;
2781 for (ArrayRef<GdbSymbol> v : symbols)
2782 numSymbols += v.size();
2784 // The return type is a flattened vector, so we'll copy each vector
2786 std::vector<GdbSymbol> ret;
2787 ret.reserve(numSymbols);
2788 for (std::vector<GdbSymbol> &vec : symbols)
2789 for (GdbSymbol &sym : vec)
2790 ret.push_back(std::move(sym));
2792 // CU vectors and symbol names are adjacent in the output file.
2793 // We can compute their offsets in the output file now.
2795 for (GdbSymbol &sym : ret) {
2796 sym.cuVectorOff = off;
2797 off += (sym.cuVector.size() + 1) * 4;
2799 for (GdbSymbol &sym : ret) {
2801 off += sym.name.size() + 1;
2807 // Returns a newly-created .gdb_index section.
2808 template <class ELFT> GdbIndexSection *GdbIndexSection::create() {
2809 std::vector<InputSection *> sections = getDebugInfoSections();
2811 // .debug_gnu_pub{names,types} are useless in executables.
2812 // They are present in input object files solely for creating
2813 // a .gdb_index. So we can remove them from the output.
2814 for (InputSectionBase *s : inputSections)
2815 if (s->name == ".debug_gnu_pubnames" || s->name == ".debug_gnu_pubtypes")
2818 std::vector<GdbChunk> chunks(sections.size());
2819 std::vector<std::vector<NameAttrEntry>> nameAttrs(sections.size());
2821 parallelForEachN(0, sections.size(), [&](size_t i) {
2822 ObjFile<ELFT> *file = sections[i]->getFile<ELFT>();
2823 DWARFContext dwarf(std::make_unique<LLDDwarfObj<ELFT>>(file));
2825 chunks[i].sec = sections[i];
2826 chunks[i].compilationUnits = readCuList(dwarf);
2827 chunks[i].addressAreas = readAddressAreas(dwarf, sections[i]);
2828 nameAttrs[i] = readPubNamesAndTypes<ELFT>(
2829 static_cast<const LLDDwarfObj<ELFT> &>(dwarf.getDWARFObj()),
2830 chunks[i].compilationUnits);
2833 auto *ret = make<GdbIndexSection>();
2834 ret->chunks = std::move(chunks);
2835 ret->symbols = createSymbols(nameAttrs, ret->chunks);
2836 ret->initOutputSize();
2840 void GdbIndexSection::writeTo(uint8_t *buf) {
2841 // Write the header.
2842 auto *hdr = reinterpret_cast<GdbIndexHeader *>(buf);
2843 uint8_t *start = buf;
2845 buf += sizeof(*hdr);
2847 // Write the CU list.
2848 hdr->cuListOff = buf - start;
2849 for (GdbChunk &chunk : chunks) {
2850 for (CuEntry &cu : chunk.compilationUnits) {
2851 write64le(buf, chunk.sec->outSecOff + cu.cuOffset);
2852 write64le(buf + 8, cu.cuLength);
2857 // Write the address area.
2858 hdr->cuTypesOff = buf - start;
2859 hdr->addressAreaOff = buf - start;
2861 for (GdbChunk &chunk : chunks) {
2862 for (AddressEntry &e : chunk.addressAreas) {
2863 uint64_t baseAddr = e.section->getVA(0);
2864 write64le(buf, baseAddr + e.lowAddress);
2865 write64le(buf + 8, baseAddr + e.highAddress);
2866 write32le(buf + 16, e.cuIndex + cuOff);
2869 cuOff += chunk.compilationUnits.size();
2872 // Write the on-disk open-addressing hash table containing symbols.
2873 hdr->symtabOff = buf - start;
2874 size_t symtabSize = computeSymtabSize();
2875 uint32_t mask = symtabSize - 1;
2877 for (GdbSymbol &sym : symbols) {
2878 uint32_t h = sym.name.hash();
2879 uint32_t i = h & mask;
2880 uint32_t step = ((h * 17) & mask) | 1;
2882 while (read32le(buf + i * 8))
2883 i = (i + step) & mask;
2885 write32le(buf + i * 8, sym.nameOff);
2886 write32le(buf + i * 8 + 4, sym.cuVectorOff);
2889 buf += symtabSize * 8;
2891 // Write the string pool.
2892 hdr->constantPoolOff = buf - start;
2893 parallelForEach(symbols, [&](GdbSymbol &sym) {
2894 memcpy(buf + sym.nameOff, sym.name.data(), sym.name.size());
2897 // Write the CU vectors.
2898 for (GdbSymbol &sym : symbols) {
2899 write32le(buf, sym.cuVector.size());
2901 for (uint32_t val : sym.cuVector) {
2902 write32le(buf, val);
2908 bool GdbIndexSection::isNeeded() const { return !chunks.empty(); }
2910 EhFrameHeader::EhFrameHeader()
2911 : SyntheticSection(SHF_ALLOC, SHT_PROGBITS, 4, ".eh_frame_hdr") {}
2913 void EhFrameHeader::writeTo(uint8_t *buf) {
2914 // Unlike most sections, the EhFrameHeader section is written while writing
2915 // another section, namely EhFrameSection, which calls the write() function
2916 // below from its writeTo() function. This is necessary because the contents
2917 // of EhFrameHeader depend on the relocated contents of EhFrameSection and we
2918 // don't know which order the sections will be written in.
2921 // .eh_frame_hdr contains a binary search table of pointers to FDEs.
2922 // Each entry of the search table consists of two values,
2923 // the starting PC from where FDEs covers, and the FDE's address.
2924 // It is sorted by PC.
2925 void EhFrameHeader::write() {
2926 uint8_t *buf = Out::bufferStart + getParent()->offset + outSecOff;
2927 using FdeData = EhFrameSection::FdeData;
2929 std::vector<FdeData> fdes = getPartition().ehFrame->getFdeData();
2932 buf[1] = DW_EH_PE_pcrel | DW_EH_PE_sdata4;
2933 buf[2] = DW_EH_PE_udata4;
2934 buf[3] = DW_EH_PE_datarel | DW_EH_PE_sdata4;
2936 getPartition().ehFrame->getParent()->addr - this->getVA() - 4);
2937 write32(buf + 8, fdes.size());
2940 for (FdeData &fde : fdes) {
2941 write32(buf, fde.pcRel);
2942 write32(buf + 4, fde.fdeVARel);
2947 size_t EhFrameHeader::getSize() const {
2948 // .eh_frame_hdr has a 12 bytes header followed by an array of FDEs.
2949 return 12 + getPartition().ehFrame->numFdes * 8;
2952 bool EhFrameHeader::isNeeded() const {
2953 return isLive() && getPartition().ehFrame->isNeeded();
2956 VersionDefinitionSection::VersionDefinitionSection()
2957 : SyntheticSection(SHF_ALLOC, SHT_GNU_verdef, sizeof(uint32_t),
2958 ".gnu.version_d") {}
2960 StringRef VersionDefinitionSection::getFileDefName() {
2961 if (!getPartition().name.empty())
2962 return getPartition().name;
2963 if (!config->soName.empty())
2964 return config->soName;
2965 return config->outputFile;
2968 void VersionDefinitionSection::finalizeContents() {
2969 fileDefNameOff = getPartition().dynStrTab->addString(getFileDefName());
2970 for (const VersionDefinition &v : namedVersionDefs())
2971 verDefNameOffs.push_back(getPartition().dynStrTab->addString(v.name));
2973 if (OutputSection *sec = getPartition().dynStrTab->getParent())
2974 getParent()->link = sec->sectionIndex;
2976 // sh_info should be set to the number of definitions. This fact is missed in
2977 // documentation, but confirmed by binutils community:
2978 // https://sourceware.org/ml/binutils/2014-11/msg00355.html
2979 getParent()->info = getVerDefNum();
2982 void VersionDefinitionSection::writeOne(uint8_t *buf, uint32_t index,
2983 StringRef name, size_t nameOff) {
2984 uint16_t flags = index == 1 ? VER_FLG_BASE : 0;
2987 write16(buf, 1); // vd_version
2988 write16(buf + 2, flags); // vd_flags
2989 write16(buf + 4, index); // vd_ndx
2990 write16(buf + 6, 1); // vd_cnt
2991 write32(buf + 8, hashSysV(name)); // vd_hash
2992 write32(buf + 12, 20); // vd_aux
2993 write32(buf + 16, 28); // vd_next
2996 write32(buf + 20, nameOff); // vda_name
2997 write32(buf + 24, 0); // vda_next
3000 void VersionDefinitionSection::writeTo(uint8_t *buf) {
3001 writeOne(buf, 1, getFileDefName(), fileDefNameOff);
3003 auto nameOffIt = verDefNameOffs.begin();
3004 for (const VersionDefinition &v : namedVersionDefs()) {
3006 writeOne(buf, v.id, v.name, *nameOffIt++);
3009 // Need to terminate the last version definition.
3010 write32(buf + 16, 0); // vd_next
3013 size_t VersionDefinitionSection::getSize() const {
3014 return EntrySize * getVerDefNum();
3017 // .gnu.version is a table where each entry is 2 byte long.
3018 VersionTableSection::VersionTableSection()
3019 : SyntheticSection(SHF_ALLOC, SHT_GNU_versym, sizeof(uint16_t),
3024 void VersionTableSection::finalizeContents() {
3025 // At the moment of june 2016 GNU docs does not mention that sh_link field
3026 // should be set, but Sun docs do. Also readelf relies on this field.
3027 getParent()->link = getPartition().dynSymTab->getParent()->sectionIndex;
3030 size_t VersionTableSection::getSize() const {
3031 return (getPartition().dynSymTab->getSymbols().size() + 1) * 2;
3034 void VersionTableSection::writeTo(uint8_t *buf) {
3036 for (const SymbolTableEntry &s : getPartition().dynSymTab->getSymbols()) {
3037 write16(buf, s.sym->versionId);
3042 bool VersionTableSection::isNeeded() const {
3044 (getPartition().verDef || getPartition().verNeed->isNeeded());
3047 void addVerneed(Symbol *ss) {
3048 auto &file = cast<SharedFile>(*ss->file);
3049 if (ss->verdefIndex == VER_NDX_GLOBAL) {
3050 ss->versionId = VER_NDX_GLOBAL;
3054 if (file.vernauxs.empty())
3055 file.vernauxs.resize(file.verdefs.size());
3057 // Select a version identifier for the vernaux data structure, if we haven't
3058 // already allocated one. The verdef identifiers cover the range
3059 // [1..getVerDefNum()]; this causes the vernaux identifiers to start from
3060 // getVerDefNum()+1.
3061 if (file.vernauxs[ss->verdefIndex] == 0)
3062 file.vernauxs[ss->verdefIndex] = ++SharedFile::vernauxNum + getVerDefNum();
3064 ss->versionId = file.vernauxs[ss->verdefIndex];
3067 template <class ELFT>
3068 VersionNeedSection<ELFT>::VersionNeedSection()
3069 : SyntheticSection(SHF_ALLOC, SHT_GNU_verneed, sizeof(uint32_t),
3070 ".gnu.version_r") {}
3072 template <class ELFT> void VersionNeedSection<ELFT>::finalizeContents() {
3073 for (SharedFile *f : sharedFiles) {
3074 if (f->vernauxs.empty())
3076 verneeds.emplace_back();
3077 Verneed &vn = verneeds.back();
3078 vn.nameStrTab = getPartition().dynStrTab->addString(f->soName);
3079 for (unsigned i = 0; i != f->vernauxs.size(); ++i) {
3080 if (f->vernauxs[i] == 0)
3083 reinterpret_cast<const typename ELFT::Verdef *>(f->verdefs[i]);
3084 vn.vernauxs.push_back(
3085 {verdef->vd_hash, f->vernauxs[i],
3086 getPartition().dynStrTab->addString(f->getStringTable().data() +
3087 verdef->getAux()->vda_name)});
3091 if (OutputSection *sec = getPartition().dynStrTab->getParent())
3092 getParent()->link = sec->sectionIndex;
3093 getParent()->info = verneeds.size();
3096 template <class ELFT> void VersionNeedSection<ELFT>::writeTo(uint8_t *buf) {
3097 // The Elf_Verneeds need to appear first, followed by the Elf_Vernauxs.
3098 auto *verneed = reinterpret_cast<Elf_Verneed *>(buf);
3099 auto *vernaux = reinterpret_cast<Elf_Vernaux *>(verneed + verneeds.size());
3101 for (auto &vn : verneeds) {
3102 // Create an Elf_Verneed for this DSO.
3103 verneed->vn_version = 1;
3104 verneed->vn_cnt = vn.vernauxs.size();
3105 verneed->vn_file = vn.nameStrTab;
3107 reinterpret_cast<char *>(vernaux) - reinterpret_cast<char *>(verneed);
3108 verneed->vn_next = sizeof(Elf_Verneed);
3111 // Create the Elf_Vernauxs for this Elf_Verneed.
3112 for (auto &vna : vn.vernauxs) {
3113 vernaux->vna_hash = vna.hash;
3114 vernaux->vna_flags = 0;
3115 vernaux->vna_other = vna.verneedIndex;
3116 vernaux->vna_name = vna.nameStrTab;
3117 vernaux->vna_next = sizeof(Elf_Vernaux);
3121 vernaux[-1].vna_next = 0;
3123 verneed[-1].vn_next = 0;
3126 template <class ELFT> size_t VersionNeedSection<ELFT>::getSize() const {
3127 return verneeds.size() * sizeof(Elf_Verneed) +
3128 SharedFile::vernauxNum * sizeof(Elf_Vernaux);
3131 template <class ELFT> bool VersionNeedSection<ELFT>::isNeeded() const {
3132 return isLive() && SharedFile::vernauxNum != 0;
3135 void MergeSyntheticSection::addSection(MergeInputSection *ms) {
3137 sections.push_back(ms);
3138 assert(alignment == ms->alignment || !(ms->flags & SHF_STRINGS));
3139 alignment = std::max(alignment, ms->alignment);
3142 MergeTailSection::MergeTailSection(StringRef name, uint32_t type,
3143 uint64_t flags, uint32_t alignment)
3144 : MergeSyntheticSection(name, type, flags, alignment),
3145 builder(StringTableBuilder::RAW, alignment) {}
3147 size_t MergeTailSection::getSize() const { return builder.getSize(); }
3149 void MergeTailSection::writeTo(uint8_t *buf) { builder.write(buf); }
3151 void MergeTailSection::finalizeContents() {
3152 // Add all string pieces to the string table builder to create section
3154 for (MergeInputSection *sec : sections)
3155 for (size_t i = 0, e = sec->pieces.size(); i != e; ++i)
3156 if (sec->pieces[i].live)
3157 builder.add(sec->getData(i));
3159 // Fix the string table content. After this, the contents will never change.
3162 // finalize() fixed tail-optimized strings, so we can now get
3163 // offsets of strings. Get an offset for each string and save it
3164 // to a corresponding SectionPiece for easy access.
3165 for (MergeInputSection *sec : sections)
3166 for (size_t i = 0, e = sec->pieces.size(); i != e; ++i)
3167 if (sec->pieces[i].live)
3168 sec->pieces[i].outputOff = builder.getOffset(sec->getData(i));
3171 void MergeNoTailSection::writeTo(uint8_t *buf) {
3172 for (size_t i = 0; i < numShards; ++i)
3173 shards[i].write(buf + shardOffsets[i]);
3176 // This function is very hot (i.e. it can take several seconds to finish)
3177 // because sometimes the number of inputs is in an order of magnitude of
3178 // millions. So, we use multi-threading.
3180 // For any strings S and T, we know S is not mergeable with T if S's hash
3181 // value is different from T's. If that's the case, we can safely put S and
3182 // T into different string builders without worrying about merge misses.
3183 // We do it in parallel.
3184 void MergeNoTailSection::finalizeContents() {
3185 // Initializes string table builders.
3186 for (size_t i = 0; i < numShards; ++i)
3187 shards.emplace_back(StringTableBuilder::RAW, alignment);
3189 // Concurrency level. Must be a power of 2 to avoid expensive modulo
3190 // operations in the following tight loop.
3191 size_t concurrency = 1;
3194 std::min<size_t>(PowerOf2Floor(hardware_concurrency()), numShards);
3196 // Add section pieces to the builders.
3197 parallelForEachN(0, concurrency, [&](size_t threadId) {
3198 for (MergeInputSection *sec : sections) {
3199 for (size_t i = 0, e = sec->pieces.size(); i != e; ++i) {
3200 if (!sec->pieces[i].live)
3202 size_t shardId = getShardId(sec->pieces[i].hash);
3203 if ((shardId & (concurrency - 1)) == threadId)
3204 sec->pieces[i].outputOff = shards[shardId].add(sec->getData(i));
3209 // Compute an in-section offset for each shard.
3211 for (size_t i = 0; i < numShards; ++i) {
3212 shards[i].finalizeInOrder();
3213 if (shards[i].getSize() > 0)
3214 off = alignTo(off, alignment);
3215 shardOffsets[i] = off;
3216 off += shards[i].getSize();
3220 // So far, section pieces have offsets from beginning of shards, but
3221 // we want offsets from beginning of the whole section. Fix them.
3222 parallelForEach(sections, [&](MergeInputSection *sec) {
3223 for (size_t i = 0, e = sec->pieces.size(); i != e; ++i)
3224 if (sec->pieces[i].live)
3225 sec->pieces[i].outputOff +=
3226 shardOffsets[getShardId(sec->pieces[i].hash)];
3230 MergeSyntheticSection *createMergeSynthetic(StringRef name, uint32_t type,
3232 uint32_t alignment) {
3233 bool shouldTailMerge = (flags & SHF_STRINGS) && config->optimize >= 2;
3234 if (shouldTailMerge)
3235 return make<MergeTailSection>(name, type, flags, alignment);
3236 return make<MergeNoTailSection>(name, type, flags, alignment);
3239 template <class ELFT> void splitSections() {
3240 // splitIntoPieces needs to be called on each MergeInputSection
3241 // before calling finalizeContents().
3242 parallelForEach(inputSections, [](InputSectionBase *sec) {
3243 if (auto *s = dyn_cast<MergeInputSection>(sec))
3244 s->splitIntoPieces();
3245 else if (auto *eh = dyn_cast<EhInputSection>(sec))
3250 MipsRldMapSection::MipsRldMapSection()
3251 : SyntheticSection(SHF_ALLOC | SHF_WRITE, SHT_PROGBITS, config->wordsize,
3254 ARMExidxSyntheticSection::ARMExidxSyntheticSection()
3255 : SyntheticSection(SHF_ALLOC | SHF_LINK_ORDER, SHT_ARM_EXIDX,
3256 config->wordsize, ".ARM.exidx") {}
3258 static InputSection *findExidxSection(InputSection *isec) {
3259 for (InputSection *d : isec->dependentSections)
3260 if (d->type == SHT_ARM_EXIDX)
3265 static bool isValidExidxSectionDep(InputSection *isec) {
3266 return (isec->flags & SHF_ALLOC) && (isec->flags & SHF_EXECINSTR) &&
3267 isec->getSize() > 0;
3270 bool ARMExidxSyntheticSection::addSection(InputSection *isec) {
3271 if (isec->type == SHT_ARM_EXIDX) {
3272 if (InputSection *dep = isec->getLinkOrderDep())
3273 if (isValidExidxSectionDep(dep))
3274 exidxSections.push_back(isec);
3278 if (isValidExidxSectionDep(isec)) {
3279 executableSections.push_back(isec);
3283 // FIXME: we do not output a relocation section when --emit-relocs is used
3284 // as we do not have relocation sections for linker generated table entries
3285 // and we would have to erase at a late stage relocations from merged entries.
3286 // Given that exception tables are already position independent and a binary
3287 // analyzer could derive the relocations we choose to erase the relocations.
3288 if (config->emitRelocs && isec->type == SHT_REL)
3289 if (InputSectionBase *ex = isec->getRelocatedSection())
3290 if (isa<InputSection>(ex) && ex->type == SHT_ARM_EXIDX)
3296 // References to .ARM.Extab Sections have bit 31 clear and are not the
3297 // special EXIDX_CANTUNWIND bit-pattern.
3298 static bool isExtabRef(uint32_t unwind) {
3299 return (unwind & 0x80000000) == 0 && unwind != 0x1;
3302 // Return true if the .ARM.exidx section Cur can be merged into the .ARM.exidx
3303 // section Prev, where Cur follows Prev in the table. This can be done if the
3304 // unwinding instructions in Cur are identical to Prev. Linker generated
3305 // EXIDX_CANTUNWIND entries are represented by nullptr as they do not have an
3307 static bool isDuplicateArmExidxSec(InputSection *prev, InputSection *cur) {
3313 // Get the last table Entry from the previous .ARM.exidx section. If Prev is
3314 // nullptr then it will be a synthesized EXIDX_CANTUNWIND entry.
3315 ExidxEntry prevEntry = {ulittle32_t(0), ulittle32_t(1)};
3317 prevEntry = prev->getDataAs<ExidxEntry>().back();
3318 if (isExtabRef(prevEntry.unwind))
3321 // We consider the unwind instructions of an .ARM.exidx table entry
3322 // a duplicate if the previous unwind instructions if:
3323 // - Both are the special EXIDX_CANTUNWIND.
3324 // - Both are the same inline unwind instructions.
3325 // We do not attempt to follow and check links into .ARM.extab tables as
3326 // consecutive identical entries are rare and the effort to check that they
3327 // are identical is high.
3329 // If Cur is nullptr then this is synthesized EXIDX_CANTUNWIND entry.
3331 return prevEntry.unwind == 1;
3333 for (const ExidxEntry entry : cur->getDataAs<ExidxEntry>())
3334 if (isExtabRef(entry.unwind) || entry.unwind != prevEntry.unwind)
3337 // All table entries in this .ARM.exidx Section can be merged into the
3338 // previous Section.
3342 // The .ARM.exidx table must be sorted in ascending order of the address of the
3343 // functions the table describes. Optionally duplicate adjacent table entries
3344 // can be removed. At the end of the function the executableSections must be
3345 // sorted in ascending order of address, Sentinel is set to the InputSection
3346 // with the highest address and any InputSections that have mergeable
3347 // .ARM.exidx table entries are removed from it.
3348 void ARMExidxSyntheticSection::finalizeContents() {
3349 // The executableSections and exidxSections that we use to derive the final
3350 // contents of this SyntheticSection are populated before
3351 // processSectionCommands() and ICF. A /DISCARD/ entry in SECTIONS command or
3352 // ICF may remove executable InputSections and their dependent .ARM.exidx
3353 // section that we recorded earlier.
3354 auto isDiscarded = [](const InputSection *isec) { return !isec->isLive(); };
3355 llvm::erase_if(executableSections, isDiscarded);
3356 llvm::erase_if(exidxSections, isDiscarded);
3358 // Sort the executable sections that may or may not have associated
3359 // .ARM.exidx sections by order of ascending address. This requires the
3360 // relative positions of InputSections to be known.
3361 auto compareByFilePosition = [](const InputSection *a,
3362 const InputSection *b) {
3363 OutputSection *aOut = a->getParent();
3364 OutputSection *bOut = b->getParent();
3367 return aOut->sectionIndex < bOut->sectionIndex;
3368 return a->outSecOff < b->outSecOff;
3370 llvm::stable_sort(executableSections, compareByFilePosition);
3371 sentinel = executableSections.back();
3372 // Optionally merge adjacent duplicate entries.
3373 if (config->mergeArmExidx) {
3374 std::vector<InputSection *> selectedSections;
3375 selectedSections.reserve(executableSections.size());
3376 selectedSections.push_back(executableSections[0]);
3378 for (size_t i = 1; i < executableSections.size(); ++i) {
3379 InputSection *ex1 = findExidxSection(executableSections[prev]);
3380 InputSection *ex2 = findExidxSection(executableSections[i]);
3381 if (!isDuplicateArmExidxSec(ex1, ex2)) {
3382 selectedSections.push_back(executableSections[i]);
3386 executableSections = std::move(selectedSections);
3391 for (InputSection *isec : executableSections) {
3392 if (InputSection *d = findExidxSection(isec)) {
3393 d->outSecOff = offset;
3394 d->parent = getParent();
3395 offset += d->getSize();
3400 // Size includes Sentinel.
3404 InputSection *ARMExidxSyntheticSection::getLinkOrderDep() const {
3405 return executableSections.front();
3408 // To write the .ARM.exidx table from the ExecutableSections we have three cases
3409 // 1.) The InputSection has a .ARM.exidx InputSection in its dependent sections.
3410 // We write the .ARM.exidx section contents and apply its relocations.
3411 // 2.) The InputSection does not have a dependent .ARM.exidx InputSection. We
3412 // must write the contents of an EXIDX_CANTUNWIND directly. We use the
3413 // start of the InputSection as the purpose of the linker generated
3414 // section is to terminate the address range of the previous entry.
3415 // 3.) A trailing EXIDX_CANTUNWIND sentinel section is required at the end of
3416 // the table to terminate the address range of the final entry.
3417 void ARMExidxSyntheticSection::writeTo(uint8_t *buf) {
3419 const uint8_t cantUnwindData[8] = {0, 0, 0, 0, // PREL31 to target
3420 1, 0, 0, 0}; // EXIDX_CANTUNWIND
3422 uint64_t offset = 0;
3423 for (InputSection *isec : executableSections) {
3424 assert(isec->getParent() != nullptr);
3425 if (InputSection *d = findExidxSection(isec)) {
3426 memcpy(buf + offset, d->data().data(), d->data().size());
3427 d->relocateAlloc(buf, buf + d->getSize());
3428 offset += d->getSize();
3430 // A Linker generated CANTUNWIND section.
3431 memcpy(buf + offset, cantUnwindData, sizeof(cantUnwindData));
3432 uint64_t s = isec->getVA();
3433 uint64_t p = getVA() + offset;
3434 target->relocateOne(buf + offset, R_ARM_PREL31, s - p);
3439 memcpy(buf + offset, cantUnwindData, sizeof(cantUnwindData));
3440 uint64_t s = sentinel->getVA(sentinel->getSize());
3441 uint64_t p = getVA() + offset;
3442 target->relocateOne(buf + offset, R_ARM_PREL31, s - p);
3443 assert(size == offset + 8);
3446 bool ARMExidxSyntheticSection::isNeeded() const {
3447 return llvm::find_if(exidxSections, [](InputSection *isec) {
3448 return isec->isLive();
3449 }) != exidxSections.end();
3452 bool ARMExidxSyntheticSection::classof(const SectionBase *d) {
3453 return d->kind() == InputSectionBase::Synthetic && d->type == SHT_ARM_EXIDX;
3456 ThunkSection::ThunkSection(OutputSection *os, uint64_t off)
3457 : SyntheticSection(SHF_ALLOC | SHF_EXECINSTR, SHT_PROGBITS, 4,
3460 this->outSecOff = off;
3463 size_t ThunkSection::getSize() const {
3464 if (roundUpSizeForErrata)
3465 return alignTo(size, 4096);
3469 void ThunkSection::addThunk(Thunk *t) {
3470 thunks.push_back(t);
3471 t->addSymbols(*this);
3474 void ThunkSection::writeTo(uint8_t *buf) {
3475 for (Thunk *t : thunks)
3476 t->writeTo(buf + t->offset);
3479 InputSection *ThunkSection::getTargetInputSection() const {
3482 const Thunk *t = thunks.front();
3483 return t->getTargetInputSection();
3486 bool ThunkSection::assignOffsets() {
3488 for (Thunk *t : thunks) {
3489 off = alignTo(off, t->alignment);
3491 uint32_t size = t->size();
3492 t->getThunkTargetSym()->size = size;
3495 bool changed = off != size;
3500 PPC32Got2Section::PPC32Got2Section()
3501 : SyntheticSection(SHF_ALLOC | SHF_WRITE, SHT_PROGBITS, 4, ".got2") {}
3503 bool PPC32Got2Section::isNeeded() const {
3504 // See the comment below. This is not needed if there is no other
3506 for (BaseCommand *base : getParent()->sectionCommands)
3507 if (auto *isd = dyn_cast<InputSectionDescription>(base))
3508 for (InputSection *isec : isd->sections)
3514 void PPC32Got2Section::finalizeContents() {
3515 // PPC32 may create multiple GOT sections for -fPIC/-fPIE, one per file in
3516 // .got2 . This function computes outSecOff of each .got2 to be used in
3517 // PPC32PltCallStub::writeTo(). The purpose of this empty synthetic section is
3518 // to collect input sections named ".got2".
3519 uint32_t offset = 0;
3520 for (BaseCommand *base : getParent()->sectionCommands)
3521 if (auto *isd = dyn_cast<InputSectionDescription>(base)) {
3522 for (InputSection *isec : isd->sections) {
3525 isec->file->ppc32Got2OutSecOff = offset;
3526 offset += (uint32_t)isec->getSize();
3531 // If linking position-dependent code then the table will store the addresses
3532 // directly in the binary so the section has type SHT_PROGBITS. If linking
3533 // position-independent code the section has type SHT_NOBITS since it will be
3534 // allocated and filled in by the dynamic linker.
3535 PPC64LongBranchTargetSection::PPC64LongBranchTargetSection()
3536 : SyntheticSection(SHF_ALLOC | SHF_WRITE,
3537 config->isPic ? SHT_NOBITS : SHT_PROGBITS, 8,
3540 uint64_t PPC64LongBranchTargetSection::getEntryVA(const Symbol *sym,
3542 return getVA() + entry_index.find({sym, addend})->second * 8;
3545 Optional<uint32_t> PPC64LongBranchTargetSection::addEntry(const Symbol *sym,
3548 entry_index.try_emplace(std::make_pair(sym, addend), entries.size());
3551 entries.emplace_back(sym, addend);
3552 return res.first->second;
3555 size_t PPC64LongBranchTargetSection::getSize() const {
3556 return entries.size() * 8;
3559 void PPC64LongBranchTargetSection::writeTo(uint8_t *buf) {
3560 // If linking non-pic we have the final addresses of the targets and they get
3561 // written to the table directly. For pic the dynamic linker will allocate
3562 // the section and fill it it.
3566 for (auto entry : entries) {
3567 const Symbol *sym = entry.first;
3568 int64_t addend = entry.second;
3569 assert(sym->getVA());
3570 // Need calls to branch to the local entry-point since a long-branch
3571 // must be a local-call.
3572 write64(buf, sym->getVA(addend) +
3573 getPPC64GlobalEntryToLocalEntryOffset(sym->stOther));
3578 bool PPC64LongBranchTargetSection::isNeeded() const {
3579 // `removeUnusedSyntheticSections()` is called before thunk allocation which
3580 // is too early to determine if this section will be empty or not. We need
3581 // Finalized to keep the section alive until after thunk creation. Finalized
3582 // only gets set to true once `finalizeSections()` is called after thunk
3583 // creation. Because of this, if we don't create any long-branch thunks we end
3584 // up with an empty .branch_lt section in the binary.
3585 return !finalized || !entries.empty();
3588 static uint8_t getAbiVersion() {
3589 // MIPS non-PIC executable gets ABI version 1.
3590 if (config->emachine == EM_MIPS) {
3591 if (!config->isPic && !config->relocatable &&
3592 (config->eflags & (EF_MIPS_PIC | EF_MIPS_CPIC)) == EF_MIPS_CPIC)
3597 if (config->emachine == EM_AMDGPU) {
3598 uint8_t ver = objectFiles[0]->abiVersion;
3599 for (InputFile *file : makeArrayRef(objectFiles).slice(1))
3600 if (file->abiVersion != ver)
3601 error("incompatible ABI version: " + toString(file));
3608 template <typename ELFT> void writeEhdr(uint8_t *buf, Partition &part) {
3609 // For executable segments, the trap instructions are written before writing
3610 // the header. Setting Elf header bytes to zero ensures that any unused bytes
3611 // in header are zero-cleared, instead of having trap instructions.
3612 memset(buf, 0, sizeof(typename ELFT::Ehdr));
3613 memcpy(buf, "\177ELF", 4);
3615 auto *eHdr = reinterpret_cast<typename ELFT::Ehdr *>(buf);
3616 eHdr->e_ident[EI_CLASS] = config->is64 ? ELFCLASS64 : ELFCLASS32;
3617 eHdr->e_ident[EI_DATA] = config->isLE ? ELFDATA2LSB : ELFDATA2MSB;
3618 eHdr->e_ident[EI_VERSION] = EV_CURRENT;
3619 eHdr->e_ident[EI_OSABI] = config->osabi;
3620 eHdr->e_ident[EI_ABIVERSION] = getAbiVersion();
3621 eHdr->e_machine = config->emachine;
3622 eHdr->e_version = EV_CURRENT;
3623 eHdr->e_flags = config->eflags;
3624 eHdr->e_ehsize = sizeof(typename ELFT::Ehdr);
3625 eHdr->e_phnum = part.phdrs.size();
3626 eHdr->e_shentsize = sizeof(typename ELFT::Shdr);
3628 if (!config->relocatable) {
3629 eHdr->e_phoff = sizeof(typename ELFT::Ehdr);
3630 eHdr->e_phentsize = sizeof(typename ELFT::Phdr);
3634 template <typename ELFT> void writePhdrs(uint8_t *buf, Partition &part) {
3635 // Write the program header table.
3636 auto *hBuf = reinterpret_cast<typename ELFT::Phdr *>(buf);
3637 for (PhdrEntry *p : part.phdrs) {
3638 hBuf->p_type = p->p_type;
3639 hBuf->p_flags = p->p_flags;
3640 hBuf->p_offset = p->p_offset;
3641 hBuf->p_vaddr = p->p_vaddr;
3642 hBuf->p_paddr = p->p_paddr;
3643 hBuf->p_filesz = p->p_filesz;
3644 hBuf->p_memsz = p->p_memsz;
3645 hBuf->p_align = p->p_align;
3650 template <typename ELFT>
3651 PartitionElfHeaderSection<ELFT>::PartitionElfHeaderSection()
3652 : SyntheticSection(SHF_ALLOC, SHT_LLVM_PART_EHDR, 1, "") {}
3654 template <typename ELFT>
3655 size_t PartitionElfHeaderSection<ELFT>::getSize() const {
3656 return sizeof(typename ELFT::Ehdr);
3659 template <typename ELFT>
3660 void PartitionElfHeaderSection<ELFT>::writeTo(uint8_t *buf) {
3661 writeEhdr<ELFT>(buf, getPartition());
3663 // Loadable partitions are always ET_DYN.
3664 auto *eHdr = reinterpret_cast<typename ELFT::Ehdr *>(buf);
3665 eHdr->e_type = ET_DYN;
3668 template <typename ELFT>
3669 PartitionProgramHeadersSection<ELFT>::PartitionProgramHeadersSection()
3670 : SyntheticSection(SHF_ALLOC, SHT_LLVM_PART_PHDR, 1, ".phdrs") {}
3672 template <typename ELFT>
3673 size_t PartitionProgramHeadersSection<ELFT>::getSize() const {
3674 return sizeof(typename ELFT::Phdr) * getPartition().phdrs.size();
3677 template <typename ELFT>
3678 void PartitionProgramHeadersSection<ELFT>::writeTo(uint8_t *buf) {
3679 writePhdrs<ELFT>(buf, getPartition());
3682 PartitionIndexSection::PartitionIndexSection()
3683 : SyntheticSection(SHF_ALLOC, SHT_PROGBITS, 4, ".rodata") {}
3685 size_t PartitionIndexSection::getSize() const {
3686 return 12 * (partitions.size() - 1);
3689 void PartitionIndexSection::finalizeContents() {
3690 for (size_t i = 1; i != partitions.size(); ++i)
3691 partitions[i].nameStrTab = mainPart->dynStrTab->addString(partitions[i].name);
3694 void PartitionIndexSection::writeTo(uint8_t *buf) {
3695 uint64_t va = getVA();
3696 for (size_t i = 1; i != partitions.size(); ++i) {
3697 write32(buf, mainPart->dynStrTab->getVA() + partitions[i].nameStrTab - va);
3698 write32(buf + 4, partitions[i].elfHeader->getVA() - (va + 4));
3700 SyntheticSection *next =
3701 i == partitions.size() - 1 ? in.partEnd : partitions[i + 1].elfHeader;
3702 write32(buf + 8, next->getVA() - partitions[i].elfHeader->getVA());
3711 std::vector<Partition> partitions;
3712 Partition *mainPart;
3714 template GdbIndexSection *GdbIndexSection::create<ELF32LE>();
3715 template GdbIndexSection *GdbIndexSection::create<ELF32BE>();
3716 template GdbIndexSection *GdbIndexSection::create<ELF64LE>();
3717 template GdbIndexSection *GdbIndexSection::create<ELF64BE>();
3719 template void splitSections<ELF32LE>();
3720 template void splitSections<ELF32BE>();
3721 template void splitSections<ELF64LE>();
3722 template void splitSections<ELF64BE>();
3724 template class MipsAbiFlagsSection<ELF32LE>;
3725 template class MipsAbiFlagsSection<ELF32BE>;
3726 template class MipsAbiFlagsSection<ELF64LE>;
3727 template class MipsAbiFlagsSection<ELF64BE>;
3729 template class MipsOptionsSection<ELF32LE>;
3730 template class MipsOptionsSection<ELF32BE>;
3731 template class MipsOptionsSection<ELF64LE>;
3732 template class MipsOptionsSection<ELF64BE>;
3734 template class MipsReginfoSection<ELF32LE>;
3735 template class MipsReginfoSection<ELF32BE>;
3736 template class MipsReginfoSection<ELF64LE>;
3737 template class MipsReginfoSection<ELF64BE>;
3739 template class DynamicSection<ELF32LE>;
3740 template class DynamicSection<ELF32BE>;
3741 template class DynamicSection<ELF64LE>;
3742 template class DynamicSection<ELF64BE>;
3744 template class RelocationSection<ELF32LE>;
3745 template class RelocationSection<ELF32BE>;
3746 template class RelocationSection<ELF64LE>;
3747 template class RelocationSection<ELF64BE>;
3749 template class AndroidPackedRelocationSection<ELF32LE>;
3750 template class AndroidPackedRelocationSection<ELF32BE>;
3751 template class AndroidPackedRelocationSection<ELF64LE>;
3752 template class AndroidPackedRelocationSection<ELF64BE>;
3754 template class RelrSection<ELF32LE>;
3755 template class RelrSection<ELF32BE>;
3756 template class RelrSection<ELF64LE>;
3757 template class RelrSection<ELF64BE>;
3759 template class SymbolTableSection<ELF32LE>;
3760 template class SymbolTableSection<ELF32BE>;
3761 template class SymbolTableSection<ELF64LE>;
3762 template class SymbolTableSection<ELF64BE>;
3764 template class VersionNeedSection<ELF32LE>;
3765 template class VersionNeedSection<ELF32BE>;
3766 template class VersionNeedSection<ELF64LE>;
3767 template class VersionNeedSection<ELF64BE>;
3769 template void writeEhdr<ELF32LE>(uint8_t *Buf, Partition &Part);
3770 template void writeEhdr<ELF32BE>(uint8_t *Buf, Partition &Part);
3771 template void writeEhdr<ELF64LE>(uint8_t *Buf, Partition &Part);
3772 template void writeEhdr<ELF64BE>(uint8_t *Buf, Partition &Part);
3774 template void writePhdrs<ELF32LE>(uint8_t *Buf, Partition &Part);
3775 template void writePhdrs<ELF32BE>(uint8_t *Buf, Partition &Part);
3776 template void writePhdrs<ELF64LE>(uint8_t *Buf, Partition &Part);
3777 template void writePhdrs<ELF64BE>(uint8_t *Buf, Partition &Part);
3779 template class PartitionElfHeaderSection<ELF32LE>;
3780 template class PartitionElfHeaderSection<ELF32BE>;
3781 template class PartitionElfHeaderSection<ELF64LE>;
3782 template class PartitionElfHeaderSection<ELF64BE>;
3784 template class PartitionProgramHeadersSection<ELF32LE>;
3785 template class PartitionProgramHeadersSection<ELF32BE>;
3786 template class PartitionProgramHeadersSection<ELF64LE>;
3787 template class PartitionProgramHeadersSection<ELF64BE>;