1 //===- SyntheticSections.cpp ----------------------------------------------===//
2 //
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
6 //
7 //===----------------------------------------------------------------------===//
8 //
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
12 // input sections.
13 //
14 //===----------------------------------------------------------------------===//
15
16 #include "SyntheticSections.h"
17 #include "Config.h"
18 #include "DWARF.h"
19 #include "EhFrame.h"
20 #include "InputFiles.h"
21 #include "LinkerScript.h"
22 #include "OutputSections.h"
23 #include "SymbolTable.h"
24 #include "Symbols.h"
25 #include "Target.h"
26 #include "Thunks.h"
27 #include "Writer.h"
28 #include "lld/Common/CommonLinkerContext.h"
29 #include "lld/Common/DWARF.h"
30 #include "lld/Common/Strings.h"
31 #include "lld/Common/Version.h"
32 #include "llvm/ADT/STLExtras.h"
33 #include "llvm/ADT/SetOperations.h"
34 #include "llvm/ADT/StringExtras.h"
35 #include "llvm/BinaryFormat/Dwarf.h"
36 #include "llvm/BinaryFormat/ELF.h"
37 #include "llvm/DebugInfo/DWARF/DWARFDebugPubTable.h"
38 #include "llvm/Support/Endian.h"
39 #include "llvm/Support/LEB128.h"
40 #include "llvm/Support/Parallel.h"
41 #include "llvm/Support/TimeProfiler.h"
42 #include <cstdlib>
43
44 using namespace llvm;
45 using namespace llvm::dwarf;
46 using namespace llvm::ELF;
47 using namespace llvm::object;
48 using namespace llvm::support;
49 using namespace lld;
50 using namespace lld::elf;
51
52 using llvm::support::endian::read32le;
53 using llvm::support::endian::write32le;
54 using llvm::support::endian::write64le;
55
56 constexpr size_t MergeNoTailSection::numShards;
57
readUint(uint8_t * buf)58 static uint64_t readUint(uint8_t *buf) {
59 return config->is64 ? read64(buf) : read32(buf);
60 }
61
writeUint(uint8_t * buf,uint64_t val)62 static void writeUint(uint8_t *buf, uint64_t val) {
63 if (config->is64)
64 write64(buf, val);
65 else
66 write32(buf, val);
67 }
68
69 // Returns an LLD version string.
getVersion()70 static ArrayRef<uint8_t> getVersion() {
71 // Check LLD_VERSION first for ease of testing.
72 // You can get consistent output by using the environment variable.
73 // This is only for testing.
74 StringRef s = getenv("LLD_VERSION");
75 if (s.empty())
76 s = saver().save(Twine("Linker: ") + getLLDVersion());
77
78 // +1 to include the terminating '\0'.
79 return {(const uint8_t *)s.data(), s.size() + 1};
80 }
81
82 // Creates a .comment section containing LLD version info.
83 // With this feature, you can identify LLD-generated binaries easily
84 // by "readelf --string-dump .comment <file>".
85 // The returned object is a mergeable string section.
createCommentSection()86 MergeInputSection *elf::createCommentSection() {
87 auto *sec = make<MergeInputSection>(SHF_MERGE | SHF_STRINGS, SHT_PROGBITS, 1,
88 getVersion(), ".comment");
89 sec->splitIntoPieces();
90 return sec;
91 }
92
93 // .MIPS.abiflags section.
94 template <class ELFT>
MipsAbiFlagsSection(Elf_Mips_ABIFlags flags)95 MipsAbiFlagsSection<ELFT>::MipsAbiFlagsSection(Elf_Mips_ABIFlags flags)
96 : SyntheticSection(SHF_ALLOC, SHT_MIPS_ABIFLAGS, 8, ".MIPS.abiflags"),
97 flags(flags) {
98 this->entsize = sizeof(Elf_Mips_ABIFlags);
99 }
100
writeTo(uint8_t * buf)101 template <class ELFT> void MipsAbiFlagsSection<ELFT>::writeTo(uint8_t *buf) {
102 memcpy(buf, &flags, sizeof(flags));
103 }
104
105 template <class ELFT>
create()106 std::unique_ptr<MipsAbiFlagsSection<ELFT>> MipsAbiFlagsSection<ELFT>::create() {
107 Elf_Mips_ABIFlags flags = {};
108 bool create = false;
109
110 for (InputSectionBase *sec : ctx.inputSections) {
111 if (sec->type != SHT_MIPS_ABIFLAGS)
112 continue;
113 sec->markDead();
114 create = true;
115
116 std::string filename = toString(sec->file);
117 const size_t size = sec->content().size();
118 // Older version of BFD (such as the default FreeBSD linker) concatenate
119 // .MIPS.abiflags instead of merging. To allow for this case (or potential
120 // zero padding) we ignore everything after the first Elf_Mips_ABIFlags
121 if (size < sizeof(Elf_Mips_ABIFlags)) {
122 error(filename + ": invalid size of .MIPS.abiflags section: got " +
123 Twine(size) + " instead of " + Twine(sizeof(Elf_Mips_ABIFlags)));
124 return nullptr;
125 }
126 auto *s =
127 reinterpret_cast<const Elf_Mips_ABIFlags *>(sec->content().data());
128 if (s->version != 0) {
129 error(filename + ": unexpected .MIPS.abiflags version " +
130 Twine(s->version));
131 return nullptr;
132 }
133
134 // LLD checks ISA compatibility in calcMipsEFlags(). Here we just
135 // select the highest number of ISA/Rev/Ext.
136 flags.isa_level = std::max(flags.isa_level, s->isa_level);
137 flags.isa_rev = std::max(flags.isa_rev, s->isa_rev);
138 flags.isa_ext = std::max(flags.isa_ext, s->isa_ext);
139 flags.gpr_size = std::max(flags.gpr_size, s->gpr_size);
140 flags.cpr1_size = std::max(flags.cpr1_size, s->cpr1_size);
141 flags.cpr2_size = std::max(flags.cpr2_size, s->cpr2_size);
142 flags.ases |= s->ases;
143 flags.flags1 |= s->flags1;
144 flags.flags2 |= s->flags2;
145 flags.fp_abi = elf::getMipsFpAbiFlag(flags.fp_abi, s->fp_abi, filename);
146 };
147
148 if (create)
149 return std::make_unique<MipsAbiFlagsSection<ELFT>>(flags);
150 return nullptr;
151 }
152
153 // .MIPS.options section.
154 template <class ELFT>
MipsOptionsSection(Elf_Mips_RegInfo reginfo)155 MipsOptionsSection<ELFT>::MipsOptionsSection(Elf_Mips_RegInfo reginfo)
156 : SyntheticSection(SHF_ALLOC, SHT_MIPS_OPTIONS, 8, ".MIPS.options"),
157 reginfo(reginfo) {
158 this->entsize = sizeof(Elf_Mips_Options) + sizeof(Elf_Mips_RegInfo);
159 }
160
writeTo(uint8_t * buf)161 template <class ELFT> void MipsOptionsSection<ELFT>::writeTo(uint8_t *buf) {
162 auto *options = reinterpret_cast<Elf_Mips_Options *>(buf);
163 options->kind = ODK_REGINFO;
164 options->size = getSize();
165
166 if (!config->relocatable)
167 reginfo.ri_gp_value = in.mipsGot->getGp();
168 memcpy(buf + sizeof(Elf_Mips_Options), ®info, sizeof(reginfo));
169 }
170
171 template <class ELFT>
create()172 std::unique_ptr<MipsOptionsSection<ELFT>> MipsOptionsSection<ELFT>::create() {
173 // N64 ABI only.
174 if (!ELFT::Is64Bits)
175 return nullptr;
176
177 SmallVector<InputSectionBase *, 0> sections;
178 for (InputSectionBase *sec : ctx.inputSections)
179 if (sec->type == SHT_MIPS_OPTIONS)
180 sections.push_back(sec);
181
182 if (sections.empty())
183 return nullptr;
184
185 Elf_Mips_RegInfo reginfo = {};
186 for (InputSectionBase *sec : sections) {
187 sec->markDead();
188
189 std::string filename = toString(sec->file);
190 ArrayRef<uint8_t> d = sec->content();
191
192 while (!d.empty()) {
193 if (d.size() < sizeof(Elf_Mips_Options)) {
194 error(filename + ": invalid size of .MIPS.options section");
195 break;
196 }
197
198 auto *opt = reinterpret_cast<const Elf_Mips_Options *>(d.data());
199 if (opt->kind == ODK_REGINFO) {
200 reginfo.ri_gprmask |= opt->getRegInfo().ri_gprmask;
201 sec->getFile<ELFT>()->mipsGp0 = opt->getRegInfo().ri_gp_value;
202 break;
203 }
204
205 if (!opt->size)
206 fatal(filename + ": zero option descriptor size");
207 d = d.slice(opt->size);
208 }
209 };
210
211 return std::make_unique<MipsOptionsSection<ELFT>>(reginfo);
212 }
213
214 // MIPS .reginfo section.
215 template <class ELFT>
MipsReginfoSection(Elf_Mips_RegInfo reginfo)216 MipsReginfoSection<ELFT>::MipsReginfoSection(Elf_Mips_RegInfo reginfo)
217 : SyntheticSection(SHF_ALLOC, SHT_MIPS_REGINFO, 4, ".reginfo"),
218 reginfo(reginfo) {
219 this->entsize = sizeof(Elf_Mips_RegInfo);
220 }
221
writeTo(uint8_t * buf)222 template <class ELFT> void MipsReginfoSection<ELFT>::writeTo(uint8_t *buf) {
223 if (!config->relocatable)
224 reginfo.ri_gp_value = in.mipsGot->getGp();
225 memcpy(buf, ®info, sizeof(reginfo));
226 }
227
228 template <class ELFT>
create()229 std::unique_ptr<MipsReginfoSection<ELFT>> MipsReginfoSection<ELFT>::create() {
230 // Section should be alive for O32 and N32 ABIs only.
231 if (ELFT::Is64Bits)
232 return nullptr;
233
234 SmallVector<InputSectionBase *, 0> sections;
235 for (InputSectionBase *sec : ctx.inputSections)
236 if (sec->type == SHT_MIPS_REGINFO)
237 sections.push_back(sec);
238
239 if (sections.empty())
240 return nullptr;
241
242 Elf_Mips_RegInfo reginfo = {};
243 for (InputSectionBase *sec : sections) {
244 sec->markDead();
245
246 if (sec->content().size() != sizeof(Elf_Mips_RegInfo)) {
247 error(toString(sec->file) + ": invalid size of .reginfo section");
248 return nullptr;
249 }
250
251 auto *r = reinterpret_cast<const Elf_Mips_RegInfo *>(sec->content().data());
252 reginfo.ri_gprmask |= r->ri_gprmask;
253 sec->getFile<ELFT>()->mipsGp0 = r->ri_gp_value;
254 };
255
256 return std::make_unique<MipsReginfoSection<ELFT>>(reginfo);
257 }
258
createInterpSection()259 InputSection *elf::createInterpSection() {
260 // StringSaver guarantees that the returned string ends with '\0'.
261 StringRef s = saver().save(config->dynamicLinker);
262 ArrayRef<uint8_t> contents = {(const uint8_t *)s.data(), s.size() + 1};
263
264 return make<InputSection>(ctx.internalFile, SHF_ALLOC, SHT_PROGBITS, 1,
265 contents, ".interp");
266 }
267
addSyntheticLocal(StringRef name,uint8_t type,uint64_t value,uint64_t size,InputSectionBase & section)268 Defined *elf::addSyntheticLocal(StringRef name, uint8_t type, uint64_t value,
269 uint64_t size, InputSectionBase §ion) {
270 Defined *s = makeDefined(section.file, name, STB_LOCAL, STV_DEFAULT, type,
271 value, size, §ion);
272 if (in.symTab)
273 in.symTab->addSymbol(s);
274
275 if (config->emachine == EM_ARM && !config->isLE && config->armBe8 &&
276 (section.flags & SHF_EXECINSTR))
277 // Adding Linker generated mapping symbols to the arm specific mapping
278 // symbols list.
279 addArmSyntheticSectionMappingSymbol(s);
280
281 return s;
282 }
283
getHashSize()284 static size_t getHashSize() {
285 switch (config->buildId) {
286 case BuildIdKind::Fast:
287 return 8;
288 case BuildIdKind::Md5:
289 case BuildIdKind::Uuid:
290 return 16;
291 case BuildIdKind::Sha1:
292 return 20;
293 case BuildIdKind::Hexstring:
294 return config->buildIdVector.size();
295 default:
296 llvm_unreachable("unknown BuildIdKind");
297 }
298 }
299
300 // This class represents a linker-synthesized .note.gnu.property section.
301 //
302 // In x86 and AArch64, object files may contain feature flags indicating the
303 // features that they have used. The flags are stored in a .note.gnu.property
304 // section.
305 //
306 // lld reads the sections from input files and merges them by computing AND of
307 // the flags. The result is written as a new .note.gnu.property section.
308 //
309 // If the flag is zero (which indicates that the intersection of the feature
310 // sets is empty, or some input files didn't have .note.gnu.property sections),
311 // we don't create this section.
GnuPropertySection()312 GnuPropertySection::GnuPropertySection()
313 : SyntheticSection(llvm::ELF::SHF_ALLOC, llvm::ELF::SHT_NOTE,
314 config->wordsize, ".note.gnu.property") {}
315
writeTo(uint8_t * buf)316 void GnuPropertySection::writeTo(uint8_t *buf) {
317 uint32_t featureAndType = config->emachine == EM_AARCH64
318 ? GNU_PROPERTY_AARCH64_FEATURE_1_AND
319 : GNU_PROPERTY_X86_FEATURE_1_AND;
320
321 write32(buf, 4); // Name size
322 write32(buf + 4, config->is64 ? 16 : 12); // Content size
323 write32(buf + 8, NT_GNU_PROPERTY_TYPE_0); // Type
324 memcpy(buf + 12, "GNU", 4); // Name string
325 write32(buf + 16, featureAndType); // Feature type
326 write32(buf + 20, 4); // Feature size
327 write32(buf + 24, config->andFeatures); // Feature flags
328 if (config->is64)
329 write32(buf + 28, 0); // Padding
330 }
331
getSize() const332 size_t GnuPropertySection::getSize() const { return config->is64 ? 32 : 28; }
333
BuildIdSection()334 BuildIdSection::BuildIdSection()
335 : SyntheticSection(SHF_ALLOC, SHT_NOTE, 4, ".note.gnu.build-id"),
336 hashSize(getHashSize()) {}
337
writeTo(uint8_t * buf)338 void BuildIdSection::writeTo(uint8_t *buf) {
339 write32(buf, 4); // Name size
340 write32(buf + 4, hashSize); // Content size
341 write32(buf + 8, NT_GNU_BUILD_ID); // Type
342 memcpy(buf + 12, "GNU", 4); // Name string
343 hashBuf = buf + 16;
344 }
345
writeBuildId(ArrayRef<uint8_t> buf)346 void BuildIdSection::writeBuildId(ArrayRef<uint8_t> buf) {
347 assert(buf.size() == hashSize);
348 memcpy(hashBuf, buf.data(), hashSize);
349 }
350
BssSection(StringRef name,uint64_t size,uint32_t alignment)351 BssSection::BssSection(StringRef name, uint64_t size, uint32_t alignment)
352 : SyntheticSection(SHF_ALLOC | SHF_WRITE, SHT_NOBITS, alignment, name) {
353 this->bss = true;
354 this->size = size;
355 }
356
EhFrameSection()357 EhFrameSection::EhFrameSection()
358 : SyntheticSection(SHF_ALLOC, SHT_PROGBITS, 1, ".eh_frame") {}
359
360 // Search for an existing CIE record or create a new one.
361 // CIE records from input object files are uniquified by their contents
362 // and where their relocations point to.
363 template <class ELFT, class RelTy>
addCie(EhSectionPiece & cie,ArrayRef<RelTy> rels)364 CieRecord *EhFrameSection::addCie(EhSectionPiece &cie, ArrayRef<RelTy> rels) {
365 Symbol *personality = nullptr;
366 unsigned firstRelI = cie.firstRelocation;
367 if (firstRelI != (unsigned)-1)
368 personality =
369 &cie.sec->template getFile<ELFT>()->getRelocTargetSym(rels[firstRelI]);
370
371 // Search for an existing CIE by CIE contents/relocation target pair.
372 CieRecord *&rec = cieMap[{cie.data(), personality}];
373
374 // If not found, create a new one.
375 if (!rec) {
376 rec = make<CieRecord>();
377 rec->cie = &cie;
378 cieRecords.push_back(rec);
379 }
380 return rec;
381 }
382
383 // There is one FDE per function. Returns a non-null pointer to the function
384 // symbol if the given FDE points to a live function.
385 template <class ELFT, class RelTy>
isFdeLive(EhSectionPiece & fde,ArrayRef<RelTy> rels)386 Defined *EhFrameSection::isFdeLive(EhSectionPiece &fde, ArrayRef<RelTy> rels) {
387 auto *sec = cast<EhInputSection>(fde.sec);
388 unsigned firstRelI = fde.firstRelocation;
389
390 // An FDE should point to some function because FDEs are to describe
391 // functions. That's however not always the case due to an issue of
392 // ld.gold with -r. ld.gold may discard only functions and leave their
393 // corresponding FDEs, which results in creating bad .eh_frame sections.
394 // To deal with that, we ignore such FDEs.
395 if (firstRelI == (unsigned)-1)
396 return nullptr;
397
398 const RelTy &rel = rels[firstRelI];
399 Symbol &b = sec->template getFile<ELFT>()->getRelocTargetSym(rel);
400
401 // FDEs for garbage-collected or merged-by-ICF sections, or sections in
402 // another partition, are dead.
403 if (auto *d = dyn_cast<Defined>(&b))
404 if (!d->folded && d->section && d->section->partition == partition)
405 return d;
406 return nullptr;
407 }
408
409 // .eh_frame is a sequence of CIE or FDE records. In general, there
410 // is one CIE record per input object file which is followed by
411 // a list of FDEs. This function searches an existing CIE or create a new
412 // one and associates FDEs to the CIE.
413 template <class ELFT, class RelTy>
addRecords(EhInputSection * sec,ArrayRef<RelTy> rels)414 void EhFrameSection::addRecords(EhInputSection *sec, ArrayRef<RelTy> rels) {
415 offsetToCie.clear();
416 for (EhSectionPiece &cie : sec->cies)
417 offsetToCie[cie.inputOff] = addCie<ELFT>(cie, rels);
418 for (EhSectionPiece &fde : sec->fdes) {
419 uint32_t id = endian::read32<ELFT::TargetEndianness>(fde.data().data() + 4);
420 CieRecord *rec = offsetToCie[fde.inputOff + 4 - id];
421 if (!rec)
422 fatal(toString(sec) + ": invalid CIE reference");
423
424 if (!isFdeLive<ELFT>(fde, rels))
425 continue;
426 rec->fdes.push_back(&fde);
427 numFdes++;
428 }
429 }
430
431 template <class ELFT>
addSectionAux(EhInputSection * sec)432 void EhFrameSection::addSectionAux(EhInputSection *sec) {
433 if (!sec->isLive())
434 return;
435 const RelsOrRelas<ELFT> rels = sec->template relsOrRelas<ELFT>();
436 if (rels.areRelocsRel())
437 addRecords<ELFT>(sec, rels.rels);
438 else
439 addRecords<ELFT>(sec, rels.relas);
440 }
441
442 // Used by ICF<ELFT>::handleLSDA(). This function is very similar to
443 // EhFrameSection::addRecords().
444 template <class ELFT, class RelTy>
iterateFDEWithLSDAAux(EhInputSection & sec,ArrayRef<RelTy> rels,DenseSet<size_t> & ciesWithLSDA,llvm::function_ref<void (InputSection &)> fn)445 void EhFrameSection::iterateFDEWithLSDAAux(
446 EhInputSection &sec, ArrayRef<RelTy> rels, DenseSet<size_t> &ciesWithLSDA,
447 llvm::function_ref<void(InputSection &)> fn) {
448 for (EhSectionPiece &cie : sec.cies)
449 if (hasLSDA(cie))
450 ciesWithLSDA.insert(cie.inputOff);
451 for (EhSectionPiece &fde : sec.fdes) {
452 uint32_t id = endian::read32<ELFT::TargetEndianness>(fde.data().data() + 4);
453 if (!ciesWithLSDA.contains(fde.inputOff + 4 - id))
454 continue;
455
456 // The CIE has a LSDA argument. Call fn with d's section.
457 if (Defined *d = isFdeLive<ELFT>(fde, rels))
458 if (auto *s = dyn_cast_or_null<InputSection>(d->section))
459 fn(*s);
460 }
461 }
462
463 template <class ELFT>
iterateFDEWithLSDA(llvm::function_ref<void (InputSection &)> fn)464 void EhFrameSection::iterateFDEWithLSDA(
465 llvm::function_ref<void(InputSection &)> fn) {
466 DenseSet<size_t> ciesWithLSDA;
467 for (EhInputSection *sec : sections) {
468 ciesWithLSDA.clear();
469 const RelsOrRelas<ELFT> rels = sec->template relsOrRelas<ELFT>();
470 if (rels.areRelocsRel())
471 iterateFDEWithLSDAAux<ELFT>(*sec, rels.rels, ciesWithLSDA, fn);
472 else
473 iterateFDEWithLSDAAux<ELFT>(*sec, rels.relas, ciesWithLSDA, fn);
474 }
475 }
476
writeCieFde(uint8_t * buf,ArrayRef<uint8_t> d)477 static void writeCieFde(uint8_t *buf, ArrayRef<uint8_t> d) {
478 memcpy(buf, d.data(), d.size());
479 // Fix the size field. -4 since size does not include the size field itself.
480 write32(buf, d.size() - 4);
481 }
482
finalizeContents()483 void EhFrameSection::finalizeContents() {
484 assert(!this->size); // Not finalized.
485
486 switch (config->ekind) {
487 case ELFNoneKind:
488 llvm_unreachable("invalid ekind");
489 case ELF32LEKind:
490 for (EhInputSection *sec : sections)
491 addSectionAux<ELF32LE>(sec);
492 break;
493 case ELF32BEKind:
494 for (EhInputSection *sec : sections)
495 addSectionAux<ELF32BE>(sec);
496 break;
497 case ELF64LEKind:
498 for (EhInputSection *sec : sections)
499 addSectionAux<ELF64LE>(sec);
500 break;
501 case ELF64BEKind:
502 for (EhInputSection *sec : sections)
503 addSectionAux<ELF64BE>(sec);
504 break;
505 }
506
507 size_t off = 0;
508 for (CieRecord *rec : cieRecords) {
509 rec->cie->outputOff = off;
510 off += rec->cie->size;
511
512 for (EhSectionPiece *fde : rec->fdes) {
513 fde->outputOff = off;
514 off += fde->size;
515 }
516 }
517
518 // The LSB standard does not allow a .eh_frame section with zero
519 // Call Frame Information records. glibc unwind-dw2-fde.c
520 // classify_object_over_fdes expects there is a CIE record length 0 as a
521 // terminator. Thus we add one unconditionally.
522 off += 4;
523
524 this->size = off;
525 }
526
527 // Returns data for .eh_frame_hdr. .eh_frame_hdr is a binary search table
528 // to get an FDE from an address to which FDE is applied. This function
529 // returns a list of such pairs.
getFdeData() const530 SmallVector<EhFrameSection::FdeData, 0> EhFrameSection::getFdeData() const {
531 uint8_t *buf = Out::bufferStart + getParent()->offset + outSecOff;
532 SmallVector<FdeData, 0> ret;
533
534 uint64_t va = getPartition().ehFrameHdr->getVA();
535 for (CieRecord *rec : cieRecords) {
536 uint8_t enc = getFdeEncoding(rec->cie);
537 for (EhSectionPiece *fde : rec->fdes) {
538 uint64_t pc = getFdePc(buf, fde->outputOff, enc);
539 uint64_t fdeVA = getParent()->addr + fde->outputOff;
540 if (!isInt<32>(pc - va))
541 fatal(toString(fde->sec) + ": PC offset is too large: 0x" +
542 Twine::utohexstr(pc - va));
543 ret.push_back({uint32_t(pc - va), uint32_t(fdeVA - va)});
544 }
545 }
546
547 // Sort the FDE list by their PC and uniqueify. Usually there is only
548 // one FDE for a PC (i.e. function), but if ICF merges two functions
549 // into one, there can be more than one FDEs pointing to the address.
550 auto less = [](const FdeData &a, const FdeData &b) {
551 return a.pcRel < b.pcRel;
552 };
553 llvm::stable_sort(ret, less);
554 auto eq = [](const FdeData &a, const FdeData &b) {
555 return a.pcRel == b.pcRel;
556 };
557 ret.erase(std::unique(ret.begin(), ret.end(), eq), ret.end());
558
559 return ret;
560 }
561
readFdeAddr(uint8_t * buf,int size)562 static uint64_t readFdeAddr(uint8_t *buf, int size) {
563 switch (size) {
564 case DW_EH_PE_udata2:
565 return read16(buf);
566 case DW_EH_PE_sdata2:
567 return (int16_t)read16(buf);
568 case DW_EH_PE_udata4:
569 return read32(buf);
570 case DW_EH_PE_sdata4:
571 return (int32_t)read32(buf);
572 case DW_EH_PE_udata8:
573 case DW_EH_PE_sdata8:
574 return read64(buf);
575 case DW_EH_PE_absptr:
576 return readUint(buf);
577 }
578 fatal("unknown FDE size encoding");
579 }
580
581 // Returns the VA to which a given FDE (on a mmap'ed buffer) is applied to.
582 // We need it to create .eh_frame_hdr section.
getFdePc(uint8_t * buf,size_t fdeOff,uint8_t enc) const583 uint64_t EhFrameSection::getFdePc(uint8_t *buf, size_t fdeOff,
584 uint8_t enc) const {
585 // The starting address to which this FDE applies is
586 // stored at FDE + 8 byte. And this offset is within
587 // the .eh_frame section.
588 size_t off = fdeOff + 8;
589 uint64_t addr = readFdeAddr(buf + off, enc & 0xf);
590 if ((enc & 0x70) == DW_EH_PE_absptr)
591 return addr;
592 if ((enc & 0x70) == DW_EH_PE_pcrel)
593 return addr + getParent()->addr + off + outSecOff;
594 fatal("unknown FDE size relative encoding");
595 }
596
writeTo(uint8_t * buf)597 void EhFrameSection::writeTo(uint8_t *buf) {
598 // Write CIE and FDE records.
599 for (CieRecord *rec : cieRecords) {
600 size_t cieOffset = rec->cie->outputOff;
601 writeCieFde(buf + cieOffset, rec->cie->data());
602
603 for (EhSectionPiece *fde : rec->fdes) {
604 size_t off = fde->outputOff;
605 writeCieFde(buf + off, fde->data());
606
607 // FDE's second word should have the offset to an associated CIE.
608 // Write it.
609 write32(buf + off + 4, off + 4 - cieOffset);
610 }
611 }
612
613 // Apply relocations. .eh_frame section contents are not contiguous
614 // in the output buffer, but relocateAlloc() still works because
615 // getOffset() takes care of discontiguous section pieces.
616 for (EhInputSection *s : sections)
617 target->relocateAlloc(*s, buf);
618
619 if (getPartition().ehFrameHdr && getPartition().ehFrameHdr->getParent())
620 getPartition().ehFrameHdr->write();
621 }
622
GotSection()623 GotSection::GotSection()
624 : SyntheticSection(SHF_ALLOC | SHF_WRITE, SHT_PROGBITS,
625 target->gotEntrySize, ".got") {
626 numEntries = target->gotHeaderEntriesNum;
627 }
628
addConstant(const Relocation & r)629 void GotSection::addConstant(const Relocation &r) { relocations.push_back(r); }
addEntry(Symbol & sym)630 void GotSection::addEntry(Symbol &sym) {
631 assert(sym.auxIdx == symAux.size() - 1);
632 symAux.back().gotIdx = numEntries++;
633 }
634
addTlsDescEntry(Symbol & sym)635 bool GotSection::addTlsDescEntry(Symbol &sym) {
636 assert(sym.auxIdx == symAux.size() - 1);
637 symAux.back().tlsDescIdx = numEntries;
638 numEntries += 2;
639 return true;
640 }
641
addDynTlsEntry(Symbol & sym)642 bool GotSection::addDynTlsEntry(Symbol &sym) {
643 assert(sym.auxIdx == symAux.size() - 1);
644 symAux.back().tlsGdIdx = numEntries;
645 // Global Dynamic TLS entries take two GOT slots.
646 numEntries += 2;
647 return true;
648 }
649
650 // Reserves TLS entries for a TLS module ID and a TLS block offset.
651 // In total it takes two GOT slots.
addTlsIndex()652 bool GotSection::addTlsIndex() {
653 if (tlsIndexOff != uint32_t(-1))
654 return false;
655 tlsIndexOff = numEntries * config->wordsize;
656 numEntries += 2;
657 return true;
658 }
659
getTlsDescOffset(const Symbol & sym) const660 uint32_t GotSection::getTlsDescOffset(const Symbol &sym) const {
661 return sym.getTlsDescIdx() * config->wordsize;
662 }
663
getTlsDescAddr(const Symbol & sym) const664 uint64_t GotSection::getTlsDescAddr(const Symbol &sym) const {
665 return getVA() + getTlsDescOffset(sym);
666 }
667
getGlobalDynAddr(const Symbol & b) const668 uint64_t GotSection::getGlobalDynAddr(const Symbol &b) const {
669 return this->getVA() + b.getTlsGdIdx() * config->wordsize;
670 }
671
getGlobalDynOffset(const Symbol & b) const672 uint64_t GotSection::getGlobalDynOffset(const Symbol &b) const {
673 return b.getTlsGdIdx() * config->wordsize;
674 }
675
finalizeContents()676 void GotSection::finalizeContents() {
677 if (config->emachine == EM_PPC64 &&
678 numEntries <= target->gotHeaderEntriesNum && !ElfSym::globalOffsetTable)
679 size = 0;
680 else
681 size = numEntries * config->wordsize;
682 }
683
isNeeded() const684 bool GotSection::isNeeded() const {
685 // Needed if the GOT symbol is used or the number of entries is more than just
686 // the header. A GOT with just the header may not be needed.
687 return hasGotOffRel || numEntries > target->gotHeaderEntriesNum;
688 }
689
writeTo(uint8_t * buf)690 void GotSection::writeTo(uint8_t *buf) {
691 // On PPC64 .got may be needed but empty. Skip the write.
692 if (size == 0)
693 return;
694 target->writeGotHeader(buf);
695 target->relocateAlloc(*this, buf);
696 }
697
getMipsPageAddr(uint64_t addr)698 static uint64_t getMipsPageAddr(uint64_t addr) {
699 return (addr + 0x8000) & ~0xffff;
700 }
701
getMipsPageCount(uint64_t size)702 static uint64_t getMipsPageCount(uint64_t size) {
703 return (size + 0xfffe) / 0xffff + 1;
704 }
705
MipsGotSection()706 MipsGotSection::MipsGotSection()
707 : SyntheticSection(SHF_ALLOC | SHF_WRITE | SHF_MIPS_GPREL, SHT_PROGBITS, 16,
708 ".got") {}
709
addEntry(InputFile & file,Symbol & sym,int64_t addend,RelExpr expr)710 void MipsGotSection::addEntry(InputFile &file, Symbol &sym, int64_t addend,
711 RelExpr expr) {
712 FileGot &g = getGot(file);
713 if (expr == R_MIPS_GOT_LOCAL_PAGE) {
714 if (const OutputSection *os = sym.getOutputSection())
715 g.pagesMap.insert({os, {}});
716 else
717 g.local16.insert({{nullptr, getMipsPageAddr(sym.getVA(addend))}, 0});
718 } else if (sym.isTls())
719 g.tls.insert({&sym, 0});
720 else if (sym.isPreemptible && expr == R_ABS)
721 g.relocs.insert({&sym, 0});
722 else if (sym.isPreemptible)
723 g.global.insert({&sym, 0});
724 else if (expr == R_MIPS_GOT_OFF32)
725 g.local32.insert({{&sym, addend}, 0});
726 else
727 g.local16.insert({{&sym, addend}, 0});
728 }
729
addDynTlsEntry(InputFile & file,Symbol & sym)730 void MipsGotSection::addDynTlsEntry(InputFile &file, Symbol &sym) {
731 getGot(file).dynTlsSymbols.insert({&sym, 0});
732 }
733
addTlsIndex(InputFile & file)734 void MipsGotSection::addTlsIndex(InputFile &file) {
735 getGot(file).dynTlsSymbols.insert({nullptr, 0});
736 }
737
getEntriesNum() const738 size_t MipsGotSection::FileGot::getEntriesNum() const {
739 return getPageEntriesNum() + local16.size() + global.size() + relocs.size() +
740 tls.size() + dynTlsSymbols.size() * 2;
741 }
742
getPageEntriesNum() const743 size_t MipsGotSection::FileGot::getPageEntriesNum() const {
744 size_t num = 0;
745 for (const std::pair<const OutputSection *, FileGot::PageBlock> &p : pagesMap)
746 num += p.second.count;
747 return num;
748 }
749
getIndexedEntriesNum() const750 size_t MipsGotSection::FileGot::getIndexedEntriesNum() const {
751 size_t count = getPageEntriesNum() + local16.size() + global.size();
752 // If there are relocation-only entries in the GOT, TLS entries
753 // are allocated after them. TLS entries should be addressable
754 // by 16-bit index so count both reloc-only and TLS entries.
755 if (!tls.empty() || !dynTlsSymbols.empty())
756 count += relocs.size() + tls.size() + dynTlsSymbols.size() * 2;
757 return count;
758 }
759
getGot(InputFile & f)760 MipsGotSection::FileGot &MipsGotSection::getGot(InputFile &f) {
761 if (f.mipsGotIndex == uint32_t(-1)) {
762 gots.emplace_back();
763 gots.back().file = &f;
764 f.mipsGotIndex = gots.size() - 1;
765 }
766 return gots[f.mipsGotIndex];
767 }
768
getPageEntryOffset(const InputFile * f,const Symbol & sym,int64_t addend) const769 uint64_t MipsGotSection::getPageEntryOffset(const InputFile *f,
770 const Symbol &sym,
771 int64_t addend) const {
772 const FileGot &g = gots[f->mipsGotIndex];
773 uint64_t index = 0;
774 if (const OutputSection *outSec = sym.getOutputSection()) {
775 uint64_t secAddr = getMipsPageAddr(outSec->addr);
776 uint64_t symAddr = getMipsPageAddr(sym.getVA(addend));
777 index = g.pagesMap.lookup(outSec).firstIndex + (symAddr - secAddr) / 0xffff;
778 } else {
779 index = g.local16.lookup({nullptr, getMipsPageAddr(sym.getVA(addend))});
780 }
781 return index * config->wordsize;
782 }
783
getSymEntryOffset(const InputFile * f,const Symbol & s,int64_t addend) const784 uint64_t MipsGotSection::getSymEntryOffset(const InputFile *f, const Symbol &s,
785 int64_t addend) const {
786 const FileGot &g = gots[f->mipsGotIndex];
787 Symbol *sym = const_cast<Symbol *>(&s);
788 if (sym->isTls())
789 return g.tls.lookup(sym) * config->wordsize;
790 if (sym->isPreemptible)
791 return g.global.lookup(sym) * config->wordsize;
792 return g.local16.lookup({sym, addend}) * config->wordsize;
793 }
794
getTlsIndexOffset(const InputFile * f) const795 uint64_t MipsGotSection::getTlsIndexOffset(const InputFile *f) const {
796 const FileGot &g = gots[f->mipsGotIndex];
797 return g.dynTlsSymbols.lookup(nullptr) * config->wordsize;
798 }
799
getGlobalDynOffset(const InputFile * f,const Symbol & s) const800 uint64_t MipsGotSection::getGlobalDynOffset(const InputFile *f,
801 const Symbol &s) const {
802 const FileGot &g = gots[f->mipsGotIndex];
803 Symbol *sym = const_cast<Symbol *>(&s);
804 return g.dynTlsSymbols.lookup(sym) * config->wordsize;
805 }
806
getFirstGlobalEntry() const807 const Symbol *MipsGotSection::getFirstGlobalEntry() const {
808 if (gots.empty())
809 return nullptr;
810 const FileGot &primGot = gots.front();
811 if (!primGot.global.empty())
812 return primGot.global.front().first;
813 if (!primGot.relocs.empty())
814 return primGot.relocs.front().first;
815 return nullptr;
816 }
817
getLocalEntriesNum() const818 unsigned MipsGotSection::getLocalEntriesNum() const {
819 if (gots.empty())
820 return headerEntriesNum;
821 return headerEntriesNum + gots.front().getPageEntriesNum() +
822 gots.front().local16.size();
823 }
824
tryMergeGots(FileGot & dst,FileGot & src,bool isPrimary)825 bool MipsGotSection::tryMergeGots(FileGot &dst, FileGot &src, bool isPrimary) {
826 FileGot tmp = dst;
827 set_union(tmp.pagesMap, src.pagesMap);
828 set_union(tmp.local16, src.local16);
829 set_union(tmp.global, src.global);
830 set_union(tmp.relocs, src.relocs);
831 set_union(tmp.tls, src.tls);
832 set_union(tmp.dynTlsSymbols, src.dynTlsSymbols);
833
834 size_t count = isPrimary ? headerEntriesNum : 0;
835 count += tmp.getIndexedEntriesNum();
836
837 if (count * config->wordsize > config->mipsGotSize)
838 return false;
839
840 std::swap(tmp, dst);
841 return true;
842 }
843
finalizeContents()844 void MipsGotSection::finalizeContents() { updateAllocSize(); }
845
updateAllocSize()846 bool MipsGotSection::updateAllocSize() {
847 size = headerEntriesNum * config->wordsize;
848 for (const FileGot &g : gots)
849 size += g.getEntriesNum() * config->wordsize;
850 return false;
851 }
852
build()853 void MipsGotSection::build() {
854 if (gots.empty())
855 return;
856
857 std::vector<FileGot> mergedGots(1);
858
859 // For each GOT move non-preemptible symbols from the `Global`
860 // to `Local16` list. Preemptible symbol might become non-preemptible
861 // one if, for example, it gets a related copy relocation.
862 for (FileGot &got : gots) {
863 for (auto &p: got.global)
864 if (!p.first->isPreemptible)
865 got.local16.insert({{p.first, 0}, 0});
866 got.global.remove_if([&](const std::pair<Symbol *, size_t> &p) {
867 return !p.first->isPreemptible;
868 });
869 }
870
871 // For each GOT remove "reloc-only" entry if there is "global"
872 // entry for the same symbol. And add local entries which indexed
873 // using 32-bit value at the end of 16-bit entries.
874 for (FileGot &got : gots) {
875 got.relocs.remove_if([&](const std::pair<Symbol *, size_t> &p) {
876 return got.global.count(p.first);
877 });
878 set_union(got.local16, got.local32);
879 got.local32.clear();
880 }
881
882 // Evaluate number of "reloc-only" entries in the resulting GOT.
883 // To do that put all unique "reloc-only" and "global" entries
884 // from all GOTs to the future primary GOT.
885 FileGot *primGot = &mergedGots.front();
886 for (FileGot &got : gots) {
887 set_union(primGot->relocs, got.global);
888 set_union(primGot->relocs, got.relocs);
889 got.relocs.clear();
890 }
891
892 // Evaluate number of "page" entries in each GOT.
893 for (FileGot &got : gots) {
894 for (std::pair<const OutputSection *, FileGot::PageBlock> &p :
895 got.pagesMap) {
896 const OutputSection *os = p.first;
897 uint64_t secSize = 0;
898 for (SectionCommand *cmd : os->commands) {
899 if (auto *isd = dyn_cast<InputSectionDescription>(cmd))
900 for (InputSection *isec : isd->sections) {
901 uint64_t off = alignToPowerOf2(secSize, isec->addralign);
902 secSize = off + isec->getSize();
903 }
904 }
905 p.second.count = getMipsPageCount(secSize);
906 }
907 }
908
909 // Merge GOTs. Try to join as much as possible GOTs but do not exceed
910 // maximum GOT size. At first, try to fill the primary GOT because
911 // the primary GOT can be accessed in the most effective way. If it
912 // is not possible, try to fill the last GOT in the list, and finally
913 // create a new GOT if both attempts failed.
914 for (FileGot &srcGot : gots) {
915 InputFile *file = srcGot.file;
916 if (tryMergeGots(mergedGots.front(), srcGot, true)) {
917 file->mipsGotIndex = 0;
918 } else {
919 // If this is the first time we failed to merge with the primary GOT,
920 // MergedGots.back() will also be the primary GOT. We must make sure not
921 // to try to merge again with isPrimary=false, as otherwise, if the
922 // inputs are just right, we could allow the primary GOT to become 1 or 2
923 // words bigger due to ignoring the header size.
924 if (mergedGots.size() == 1 ||
925 !tryMergeGots(mergedGots.back(), srcGot, false)) {
926 mergedGots.emplace_back();
927 std::swap(mergedGots.back(), srcGot);
928 }
929 file->mipsGotIndex = mergedGots.size() - 1;
930 }
931 }
932 std::swap(gots, mergedGots);
933
934 // Reduce number of "reloc-only" entries in the primary GOT
935 // by subtracting "global" entries in the primary GOT.
936 primGot = &gots.front();
937 primGot->relocs.remove_if([&](const std::pair<Symbol *, size_t> &p) {
938 return primGot->global.count(p.first);
939 });
940
941 // Calculate indexes for each GOT entry.
942 size_t index = headerEntriesNum;
943 for (FileGot &got : gots) {
944 got.startIndex = &got == primGot ? 0 : index;
945 for (std::pair<const OutputSection *, FileGot::PageBlock> &p :
946 got.pagesMap) {
947 // For each output section referenced by GOT page relocations calculate
948 // and save into pagesMap an upper bound of MIPS GOT entries required
949 // to store page addresses of local symbols. We assume the worst case -
950 // each 64kb page of the output section has at least one GOT relocation
951 // against it. And take in account the case when the section intersects
952 // page boundaries.
953 p.second.firstIndex = index;
954 index += p.second.count;
955 }
956 for (auto &p: got.local16)
957 p.second = index++;
958 for (auto &p: got.global)
959 p.second = index++;
960 for (auto &p: got.relocs)
961 p.second = index++;
962 for (auto &p: got.tls)
963 p.second = index++;
964 for (auto &p: got.dynTlsSymbols) {
965 p.second = index;
966 index += 2;
967 }
968 }
969
970 // Update SymbolAux::gotIdx field to use this
971 // value later in the `sortMipsSymbols` function.
972 for (auto &p : primGot->global) {
973 if (p.first->auxIdx == 0)
974 p.first->allocateAux();
975 symAux.back().gotIdx = p.second;
976 }
977 for (auto &p : primGot->relocs) {
978 if (p.first->auxIdx == 0)
979 p.first->allocateAux();
980 symAux.back().gotIdx = p.second;
981 }
982
983 // Create dynamic relocations.
984 for (FileGot &got : gots) {
985 // Create dynamic relocations for TLS entries.
986 for (std::pair<Symbol *, size_t> &p : got.tls) {
987 Symbol *s = p.first;
988 uint64_t offset = p.second * config->wordsize;
989 // When building a shared library we still need a dynamic relocation
990 // for the TP-relative offset as we don't know how much other data will
991 // be allocated before us in the static TLS block.
992 if (s->isPreemptible || config->shared)
993 mainPart->relaDyn->addReloc({target->tlsGotRel, this, offset,
994 DynamicReloc::AgainstSymbolWithTargetVA,
995 *s, 0, R_ABS});
996 }
997 for (std::pair<Symbol *, size_t> &p : got.dynTlsSymbols) {
998 Symbol *s = p.first;
999 uint64_t offset = p.second * config->wordsize;
1000 if (s == nullptr) {
1001 if (!config->shared)
1002 continue;
1003 mainPart->relaDyn->addReloc({target->tlsModuleIndexRel, this, offset});
1004 } else {
1005 // When building a shared library we still need a dynamic relocation
1006 // for the module index. Therefore only checking for
1007 // S->isPreemptible is not sufficient (this happens e.g. for
1008 // thread-locals that have been marked as local through a linker script)
1009 if (!s->isPreemptible && !config->shared)
1010 continue;
1011 mainPart->relaDyn->addSymbolReloc(target->tlsModuleIndexRel, *this,
1012 offset, *s);
1013 // However, we can skip writing the TLS offset reloc for non-preemptible
1014 // symbols since it is known even in shared libraries
1015 if (!s->isPreemptible)
1016 continue;
1017 offset += config->wordsize;
1018 mainPart->relaDyn->addSymbolReloc(target->tlsOffsetRel, *this, offset,
1019 *s);
1020 }
1021 }
1022
1023 // Do not create dynamic relocations for non-TLS
1024 // entries in the primary GOT.
1025 if (&got == primGot)
1026 continue;
1027
1028 // Dynamic relocations for "global" entries.
1029 for (const std::pair<Symbol *, size_t> &p : got.global) {
1030 uint64_t offset = p.second * config->wordsize;
1031 mainPart->relaDyn->addSymbolReloc(target->relativeRel, *this, offset,
1032 *p.first);
1033 }
1034 if (!config->isPic)
1035 continue;
1036 // Dynamic relocations for "local" entries in case of PIC.
1037 for (const std::pair<const OutputSection *, FileGot::PageBlock> &l :
1038 got.pagesMap) {
1039 size_t pageCount = l.second.count;
1040 for (size_t pi = 0; pi < pageCount; ++pi) {
1041 uint64_t offset = (l.second.firstIndex + pi) * config->wordsize;
1042 mainPart->relaDyn->addReloc({target->relativeRel, this, offset, l.first,
1043 int64_t(pi * 0x10000)});
1044 }
1045 }
1046 for (const std::pair<GotEntry, size_t> &p : got.local16) {
1047 uint64_t offset = p.second * config->wordsize;
1048 mainPart->relaDyn->addReloc({target->relativeRel, this, offset,
1049 DynamicReloc::AddendOnlyWithTargetVA,
1050 *p.first.first, p.first.second, R_ABS});
1051 }
1052 }
1053 }
1054
isNeeded() const1055 bool MipsGotSection::isNeeded() const {
1056 // We add the .got section to the result for dynamic MIPS target because
1057 // its address and properties are mentioned in the .dynamic section.
1058 return !config->relocatable;
1059 }
1060
getGp(const InputFile * f) const1061 uint64_t MipsGotSection::getGp(const InputFile *f) const {
1062 // For files without related GOT or files refer a primary GOT
1063 // returns "common" _gp value. For secondary GOTs calculate
1064 // individual _gp values.
1065 if (!f || f->mipsGotIndex == uint32_t(-1) || f->mipsGotIndex == 0)
1066 return ElfSym::mipsGp->getVA(0);
1067 return getVA() + gots[f->mipsGotIndex].startIndex * config->wordsize + 0x7ff0;
1068 }
1069
writeTo(uint8_t * buf)1070 void MipsGotSection::writeTo(uint8_t *buf) {
1071 // Set the MSB of the second GOT slot. This is not required by any
1072 // MIPS ABI documentation, though.
1073 //
1074 // There is a comment in glibc saying that "The MSB of got[1] of a
1075 // gnu object is set to identify gnu objects," and in GNU gold it
1076 // says "the second entry will be used by some runtime loaders".
1077 // But how this field is being used is unclear.
1078 //
1079 // We are not really willing to mimic other linkers behaviors
1080 // without understanding why they do that, but because all files
1081 // generated by GNU tools have this special GOT value, and because
1082 // we've been doing this for years, it is probably a safe bet to
1083 // keep doing this for now. We really need to revisit this to see
1084 // if we had to do this.
1085 writeUint(buf + config->wordsize, (uint64_t)1 << (config->wordsize * 8 - 1));
1086 for (const FileGot &g : gots) {
1087 auto write = [&](size_t i, const Symbol *s, int64_t a) {
1088 uint64_t va = a;
1089 if (s)
1090 va = s->getVA(a);
1091 writeUint(buf + i * config->wordsize, va);
1092 };
1093 // Write 'page address' entries to the local part of the GOT.
1094 for (const std::pair<const OutputSection *, FileGot::PageBlock> &l :
1095 g.pagesMap) {
1096 size_t pageCount = l.second.count;
1097 uint64_t firstPageAddr = getMipsPageAddr(l.first->addr);
1098 for (size_t pi = 0; pi < pageCount; ++pi)
1099 write(l.second.firstIndex + pi, nullptr, firstPageAddr + pi * 0x10000);
1100 }
1101 // Local, global, TLS, reloc-only entries.
1102 // If TLS entry has a corresponding dynamic relocations, leave it
1103 // initialized by zero. Write down adjusted TLS symbol's values otherwise.
1104 // To calculate the adjustments use offsets for thread-local storage.
1105 // http://web.archive.org/web/20190324223224/https://www.linux-mips.org/wiki/NPTL
1106 for (const std::pair<GotEntry, size_t> &p : g.local16)
1107 write(p.second, p.first.first, p.first.second);
1108 // Write VA to the primary GOT only. For secondary GOTs that
1109 // will be done by REL32 dynamic relocations.
1110 if (&g == &gots.front())
1111 for (const std::pair<Symbol *, size_t> &p : g.global)
1112 write(p.second, p.first, 0);
1113 for (const std::pair<Symbol *, size_t> &p : g.relocs)
1114 write(p.second, p.first, 0);
1115 for (const std::pair<Symbol *, size_t> &p : g.tls)
1116 write(p.second, p.first,
1117 p.first->isPreemptible || config->shared ? 0 : -0x7000);
1118 for (const std::pair<Symbol *, size_t> &p : g.dynTlsSymbols) {
1119 if (p.first == nullptr && !config->shared)
1120 write(p.second, nullptr, 1);
1121 else if (p.first && !p.first->isPreemptible) {
1122 // If we are emitting a shared library with relocations we mustn't write
1123 // anything to the GOT here. When using Elf_Rel relocations the value
1124 // one will be treated as an addend and will cause crashes at runtime
1125 if (!config->shared)
1126 write(p.second, nullptr, 1);
1127 write(p.second + 1, p.first, -0x8000);
1128 }
1129 }
1130 }
1131 }
1132
1133 // On PowerPC the .plt section is used to hold the table of function addresses
1134 // instead of the .got.plt, and the type is SHT_NOBITS similar to a .bss
1135 // section. I don't know why we have a BSS style type for the section but it is
1136 // consistent across both 64-bit PowerPC ABIs as well as the 32-bit PowerPC ABI.
GotPltSection()1137 GotPltSection::GotPltSection()
1138 : SyntheticSection(SHF_ALLOC | SHF_WRITE, SHT_PROGBITS, config->wordsize,
1139 ".got.plt") {
1140 if (config->emachine == EM_PPC) {
1141 name = ".plt";
1142 } else if (config->emachine == EM_PPC64) {
1143 type = SHT_NOBITS;
1144 name = ".plt";
1145 }
1146 }
1147
addEntry(Symbol & sym)1148 void GotPltSection::addEntry(Symbol &sym) {
1149 assert(sym.auxIdx == symAux.size() - 1 &&
1150 symAux.back().pltIdx == entries.size());
1151 entries.push_back(&sym);
1152 }
1153
getSize() const1154 size_t GotPltSection::getSize() const {
1155 return (target->gotPltHeaderEntriesNum + entries.size()) *
1156 target->gotEntrySize;
1157 }
1158
writeTo(uint8_t * buf)1159 void GotPltSection::writeTo(uint8_t *buf) {
1160 target->writeGotPltHeader(buf);
1161 buf += target->gotPltHeaderEntriesNum * target->gotEntrySize;
1162 for (const Symbol *b : entries) {
1163 target->writeGotPlt(buf, *b);
1164 buf += target->gotEntrySize;
1165 }
1166 }
1167
isNeeded() const1168 bool GotPltSection::isNeeded() const {
1169 // We need to emit GOTPLT even if it's empty if there's a relocation relative
1170 // to it.
1171 return !entries.empty() || hasGotPltOffRel;
1172 }
1173
getIgotPltName()1174 static StringRef getIgotPltName() {
1175 // On ARM the IgotPltSection is part of the GotSection.
1176 if (config->emachine == EM_ARM)
1177 return ".got";
1178
1179 // On PowerPC64 the GotPltSection is renamed to '.plt' so the IgotPltSection
1180 // needs to be named the same.
1181 if (config->emachine == EM_PPC64)
1182 return ".plt";
1183
1184 return ".got.plt";
1185 }
1186
1187 // On PowerPC64 the GotPltSection type is SHT_NOBITS so we have to follow suit
1188 // with the IgotPltSection.
IgotPltSection()1189 IgotPltSection::IgotPltSection()
1190 : SyntheticSection(SHF_ALLOC | SHF_WRITE,
1191 config->emachine == EM_PPC64 ? SHT_NOBITS : SHT_PROGBITS,
1192 target->gotEntrySize, getIgotPltName()) {}
1193
addEntry(Symbol & sym)1194 void IgotPltSection::addEntry(Symbol &sym) {
1195 assert(symAux.back().pltIdx == entries.size());
1196 entries.push_back(&sym);
1197 }
1198
getSize() const1199 size_t IgotPltSection::getSize() const {
1200 return entries.size() * target->gotEntrySize;
1201 }
1202
writeTo(uint8_t * buf)1203 void IgotPltSection::writeTo(uint8_t *buf) {
1204 for (const Symbol *b : entries) {
1205 target->writeIgotPlt(buf, *b);
1206 buf += target->gotEntrySize;
1207 }
1208 }
1209
StringTableSection(StringRef name,bool dynamic)1210 StringTableSection::StringTableSection(StringRef name, bool dynamic)
1211 : SyntheticSection(dynamic ? (uint64_t)SHF_ALLOC : 0, SHT_STRTAB, 1, name),
1212 dynamic(dynamic) {
1213 // ELF string tables start with a NUL byte.
1214 strings.push_back("");
1215 stringMap.try_emplace(CachedHashStringRef(""), 0);
1216 size = 1;
1217 }
1218
1219 // Adds a string to the string table. If `hashIt` is true we hash and check for
1220 // duplicates. It is optional because the name of global symbols are already
1221 // uniqued and hashing them again has a big cost for a small value: uniquing
1222 // them with some other string that happens to be the same.
addString(StringRef s,bool hashIt)1223 unsigned StringTableSection::addString(StringRef s, bool hashIt) {
1224 if (hashIt) {
1225 auto r = stringMap.try_emplace(CachedHashStringRef(s), size);
1226 if (!r.second)
1227 return r.first->second;
1228 }
1229 if (s.empty())
1230 return 0;
1231 unsigned ret = this->size;
1232 this->size = this->size + s.size() + 1;
1233 strings.push_back(s);
1234 return ret;
1235 }
1236
writeTo(uint8_t * buf)1237 void StringTableSection::writeTo(uint8_t *buf) {
1238 for (StringRef s : strings) {
1239 memcpy(buf, s.data(), s.size());
1240 buf[s.size()] = '\0';
1241 buf += s.size() + 1;
1242 }
1243 }
1244
1245 // Returns the number of entries in .gnu.version_d: the number of
1246 // non-VER_NDX_LOCAL-non-VER_NDX_GLOBAL definitions, plus 1.
1247 // Note that we don't support vd_cnt > 1 yet.
getVerDefNum()1248 static unsigned getVerDefNum() {
1249 return namedVersionDefs().size() + 1;
1250 }
1251
1252 template <class ELFT>
DynamicSection()1253 DynamicSection<ELFT>::DynamicSection()
1254 : SyntheticSection(SHF_ALLOC | SHF_WRITE, SHT_DYNAMIC, config->wordsize,
1255 ".dynamic") {
1256 this->entsize = ELFT::Is64Bits ? 16 : 8;
1257
1258 // .dynamic section is not writable on MIPS and on Fuchsia OS
1259 // which passes -z rodynamic.
1260 // See "Special Section" in Chapter 4 in the following document:
1261 // ftp://www.linux-mips.org/pub/linux/mips/doc/ABI/mipsabi.pdf
1262 if (config->emachine == EM_MIPS || config->zRodynamic)
1263 this->flags = SHF_ALLOC;
1264 }
1265
1266 // The output section .rela.dyn may include these synthetic sections:
1267 //
1268 // - part.relaDyn
1269 // - in.relaIplt: this is included if in.relaIplt is named .rela.dyn
1270 // - in.relaPlt: this is included if a linker script places .rela.plt inside
1271 // .rela.dyn
1272 //
1273 // DT_RELASZ is the total size of the included sections.
addRelaSz(const RelocationBaseSection & relaDyn)1274 static uint64_t addRelaSz(const RelocationBaseSection &relaDyn) {
1275 size_t size = relaDyn.getSize();
1276 if (in.relaIplt->getParent() == relaDyn.getParent())
1277 size += in.relaIplt->getSize();
1278 if (in.relaPlt->getParent() == relaDyn.getParent())
1279 size += in.relaPlt->getSize();
1280 return size;
1281 }
1282
1283 // A Linker script may assign the RELA relocation sections to the same
1284 // output section. When this occurs we cannot just use the OutputSection
1285 // Size. Moreover the [DT_JMPREL, DT_JMPREL + DT_PLTRELSZ) is permitted to
1286 // overlap with the [DT_RELA, DT_RELA + DT_RELASZ).
addPltRelSz()1287 static uint64_t addPltRelSz() {
1288 size_t size = in.relaPlt->getSize();
1289 if (in.relaIplt->getParent() == in.relaPlt->getParent() &&
1290 in.relaIplt->name == in.relaPlt->name)
1291 size += in.relaIplt->getSize();
1292 return size;
1293 }
1294
1295 // Add remaining entries to complete .dynamic contents.
1296 template <class ELFT>
1297 std::vector<std::pair<int32_t, uint64_t>>
computeContents()1298 DynamicSection<ELFT>::computeContents() {
1299 elf::Partition &part = getPartition();
1300 bool isMain = part.name.empty();
1301 std::vector<std::pair<int32_t, uint64_t>> entries;
1302
1303 auto addInt = [&](int32_t tag, uint64_t val) {
1304 entries.emplace_back(tag, val);
1305 };
1306 auto addInSec = [&](int32_t tag, const InputSection &sec) {
1307 entries.emplace_back(tag, sec.getVA());
1308 };
1309
1310 for (StringRef s : config->filterList)
1311 addInt(DT_FILTER, part.dynStrTab->addString(s));
1312 for (StringRef s : config->auxiliaryList)
1313 addInt(DT_AUXILIARY, part.dynStrTab->addString(s));
1314
1315 if (!config->rpath.empty())
1316 addInt(config->enableNewDtags ? DT_RUNPATH : DT_RPATH,
1317 part.dynStrTab->addString(config->rpath));
1318
1319 for (SharedFile *file : ctx.sharedFiles)
1320 if (file->isNeeded)
1321 addInt(DT_NEEDED, part.dynStrTab->addString(file->soName));
1322
1323 if (isMain) {
1324 if (!config->soName.empty())
1325 addInt(DT_SONAME, part.dynStrTab->addString(config->soName));
1326 } else {
1327 if (!config->soName.empty())
1328 addInt(DT_NEEDED, part.dynStrTab->addString(config->soName));
1329 addInt(DT_SONAME, part.dynStrTab->addString(part.name));
1330 }
1331
1332 // Set DT_FLAGS and DT_FLAGS_1.
1333 uint32_t dtFlags = 0;
1334 uint32_t dtFlags1 = 0;
1335 if (config->bsymbolic == BsymbolicKind::All)
1336 dtFlags |= DF_SYMBOLIC;
1337 if (config->zGlobal)
1338 dtFlags1 |= DF_1_GLOBAL;
1339 if (config->zInitfirst)
1340 dtFlags1 |= DF_1_INITFIRST;
1341 if (config->zInterpose)
1342 dtFlags1 |= DF_1_INTERPOSE;
1343 if (config->zNodefaultlib)
1344 dtFlags1 |= DF_1_NODEFLIB;
1345 if (config->zNodelete)
1346 dtFlags1 |= DF_1_NODELETE;
1347 if (config->zNodlopen)
1348 dtFlags1 |= DF_1_NOOPEN;
1349 if (config->pie)
1350 dtFlags1 |= DF_1_PIE;
1351 if (config->zNow) {
1352 dtFlags |= DF_BIND_NOW;
1353 dtFlags1 |= DF_1_NOW;
1354 }
1355 if (config->zOrigin) {
1356 dtFlags |= DF_ORIGIN;
1357 dtFlags1 |= DF_1_ORIGIN;
1358 }
1359 if (!config->zText)
1360 dtFlags |= DF_TEXTREL;
1361 if (ctx.hasTlsIe && config->shared)
1362 dtFlags |= DF_STATIC_TLS;
1363
1364 if (dtFlags)
1365 addInt(DT_FLAGS, dtFlags);
1366 if (dtFlags1)
1367 addInt(DT_FLAGS_1, dtFlags1);
1368
1369 // DT_DEBUG is a pointer to debug information used by debuggers at runtime. We
1370 // need it for each process, so we don't write it for DSOs. The loader writes
1371 // the pointer into this entry.
1372 //
1373 // DT_DEBUG is the only .dynamic entry that needs to be written to. Some
1374 // systems (currently only Fuchsia OS) provide other means to give the
1375 // debugger this information. Such systems may choose make .dynamic read-only.
1376 // If the target is such a system (used -z rodynamic) don't write DT_DEBUG.
1377 if (!config->shared && !config->relocatable && !config->zRodynamic)
1378 addInt(DT_DEBUG, 0);
1379
1380 if (part.relaDyn->isNeeded() ||
1381 (in.relaIplt->isNeeded() &&
1382 part.relaDyn->getParent() == in.relaIplt->getParent())) {
1383 addInSec(part.relaDyn->dynamicTag, *part.relaDyn);
1384 entries.emplace_back(part.relaDyn->sizeDynamicTag,
1385 addRelaSz(*part.relaDyn));
1386
1387 bool isRela = config->isRela;
1388 addInt(isRela ? DT_RELAENT : DT_RELENT,
1389 isRela ? sizeof(Elf_Rela) : sizeof(Elf_Rel));
1390
1391 // MIPS dynamic loader does not support RELCOUNT tag.
1392 // The problem is in the tight relation between dynamic
1393 // relocations and GOT. So do not emit this tag on MIPS.
1394 if (config->emachine != EM_MIPS) {
1395 size_t numRelativeRels = part.relaDyn->getRelativeRelocCount();
1396 if (config->zCombreloc && numRelativeRels)
1397 addInt(isRela ? DT_RELACOUNT : DT_RELCOUNT, numRelativeRels);
1398 }
1399 }
1400 if (part.relrDyn && part.relrDyn->getParent() &&
1401 !part.relrDyn->relocs.empty()) {
1402 addInSec(config->useAndroidRelrTags ? DT_ANDROID_RELR : DT_RELR,
1403 *part.relrDyn);
1404 addInt(config->useAndroidRelrTags ? DT_ANDROID_RELRSZ : DT_RELRSZ,
1405 part.relrDyn->getParent()->size);
1406 addInt(config->useAndroidRelrTags ? DT_ANDROID_RELRENT : DT_RELRENT,
1407 sizeof(Elf_Relr));
1408 }
1409 // .rel[a].plt section usually consists of two parts, containing plt and
1410 // iplt relocations. It is possible to have only iplt relocations in the
1411 // output. In that case relaPlt is empty and have zero offset, the same offset
1412 // as relaIplt has. And we still want to emit proper dynamic tags for that
1413 // case, so here we always use relaPlt as marker for the beginning of
1414 // .rel[a].plt section.
1415 if (isMain && (in.relaPlt->isNeeded() || in.relaIplt->isNeeded())) {
1416 addInSec(DT_JMPREL, *in.relaPlt);
1417 entries.emplace_back(DT_PLTRELSZ, addPltRelSz());
1418 switch (config->emachine) {
1419 case EM_MIPS:
1420 addInSec(DT_MIPS_PLTGOT, *in.gotPlt);
1421 break;
1422 case EM_S390:
1423 addInSec(DT_PLTGOT, *in.got);
1424 break;
1425 case EM_SPARCV9:
1426 addInSec(DT_PLTGOT, *in.plt);
1427 break;
1428 case EM_AARCH64:
1429 if (llvm::find_if(in.relaPlt->relocs, [](const DynamicReloc &r) {
1430 return r.type == target->pltRel &&
1431 r.sym->stOther & STO_AARCH64_VARIANT_PCS;
1432 }) != in.relaPlt->relocs.end())
1433 addInt(DT_AARCH64_VARIANT_PCS, 0);
1434 addInSec(DT_PLTGOT, *in.gotPlt);
1435 break;
1436 case EM_RISCV:
1437 if (llvm::any_of(in.relaPlt->relocs, [](const DynamicReloc &r) {
1438 return r.type == target->pltRel &&
1439 (r.sym->stOther & STO_RISCV_VARIANT_CC);
1440 }))
1441 addInt(DT_RISCV_VARIANT_CC, 0);
1442 [[fallthrough]];
1443 default:
1444 addInSec(DT_PLTGOT, *in.gotPlt);
1445 break;
1446 }
1447 addInt(DT_PLTREL, config->isRela ? DT_RELA : DT_REL);
1448 }
1449
1450 if (config->emachine == EM_AARCH64) {
1451 if (config->andFeatures & GNU_PROPERTY_AARCH64_FEATURE_1_BTI)
1452 addInt(DT_AARCH64_BTI_PLT, 0);
1453 if (config->zPacPlt)
1454 addInt(DT_AARCH64_PAC_PLT, 0);
1455
1456 if (hasMemtag()) {
1457 addInt(DT_AARCH64_MEMTAG_MODE, config->androidMemtagMode == NT_MEMTAG_LEVEL_ASYNC);
1458 addInt(DT_AARCH64_MEMTAG_HEAP, config->androidMemtagHeap);
1459 addInt(DT_AARCH64_MEMTAG_STACK, config->androidMemtagStack);
1460 if (mainPart->memtagGlobalDescriptors->isNeeded()) {
1461 addInSec(DT_AARCH64_MEMTAG_GLOBALS, *mainPart->memtagGlobalDescriptors);
1462 addInt(DT_AARCH64_MEMTAG_GLOBALSSZ,
1463 mainPart->memtagGlobalDescriptors->getSize());
1464 }
1465 }
1466 }
1467
1468 addInSec(DT_SYMTAB, *part.dynSymTab);
1469 addInt(DT_SYMENT, sizeof(Elf_Sym));
1470 addInSec(DT_STRTAB, *part.dynStrTab);
1471 addInt(DT_STRSZ, part.dynStrTab->getSize());
1472 if (!config->zText)
1473 addInt(DT_TEXTREL, 0);
1474 if (part.gnuHashTab && part.gnuHashTab->getParent())
1475 addInSec(DT_GNU_HASH, *part.gnuHashTab);
1476 if (part.hashTab && part.hashTab->getParent())
1477 addInSec(DT_HASH, *part.hashTab);
1478
1479 if (isMain) {
1480 if (Out::preinitArray) {
1481 addInt(DT_PREINIT_ARRAY, Out::preinitArray->addr);
1482 addInt(DT_PREINIT_ARRAYSZ, Out::preinitArray->size);
1483 }
1484 if (Out::initArray) {
1485 addInt(DT_INIT_ARRAY, Out::initArray->addr);
1486 addInt(DT_INIT_ARRAYSZ, Out::initArray->size);
1487 }
1488 if (Out::finiArray) {
1489 addInt(DT_FINI_ARRAY, Out::finiArray->addr);
1490 addInt(DT_FINI_ARRAYSZ, Out::finiArray->size);
1491 }
1492
1493 if (Symbol *b = symtab.find(config->init))
1494 if (b->isDefined())
1495 addInt(DT_INIT, b->getVA());
1496 if (Symbol *b = symtab.find(config->fini))
1497 if (b->isDefined())
1498 addInt(DT_FINI, b->getVA());
1499 }
1500
1501 if (part.verSym && part.verSym->isNeeded())
1502 addInSec(DT_VERSYM, *part.verSym);
1503 if (part.verDef && part.verDef->isLive()) {
1504 addInSec(DT_VERDEF, *part.verDef);
1505 addInt(DT_VERDEFNUM, getVerDefNum());
1506 }
1507 if (part.verNeed && part.verNeed->isNeeded()) {
1508 addInSec(DT_VERNEED, *part.verNeed);
1509 unsigned needNum = 0;
1510 for (SharedFile *f : ctx.sharedFiles)
1511 if (!f->vernauxs.empty())
1512 ++needNum;
1513 addInt(DT_VERNEEDNUM, needNum);
1514 }
1515
1516 if (config->emachine == EM_MIPS) {
1517 addInt(DT_MIPS_RLD_VERSION, 1);
1518 addInt(DT_MIPS_FLAGS, RHF_NOTPOT);
1519 addInt(DT_MIPS_BASE_ADDRESS, target->getImageBase());
1520 addInt(DT_MIPS_SYMTABNO, part.dynSymTab->getNumSymbols());
1521 addInt(DT_MIPS_LOCAL_GOTNO, in.mipsGot->getLocalEntriesNum());
1522
1523 if (const Symbol *b = in.mipsGot->getFirstGlobalEntry())
1524 addInt(DT_MIPS_GOTSYM, b->dynsymIndex);
1525 else
1526 addInt(DT_MIPS_GOTSYM, part.dynSymTab->getNumSymbols());
1527 addInSec(DT_PLTGOT, *in.mipsGot);
1528 if (in.mipsRldMap) {
1529 if (!config->pie)
1530 addInSec(DT_MIPS_RLD_MAP, *in.mipsRldMap);
1531 // Store the offset to the .rld_map section
1532 // relative to the address of the tag.
1533 addInt(DT_MIPS_RLD_MAP_REL,
1534 in.mipsRldMap->getVA() - (getVA() + entries.size() * entsize));
1535 }
1536 }
1537
1538 // DT_PPC_GOT indicates to glibc Secure PLT is used. If DT_PPC_GOT is absent,
1539 // glibc assumes the old-style BSS PLT layout which we don't support.
1540 if (config->emachine == EM_PPC)
1541 addInSec(DT_PPC_GOT, *in.got);
1542
1543 // Glink dynamic tag is required by the V2 abi if the plt section isn't empty.
1544 if (config->emachine == EM_PPC64 && in.plt->isNeeded()) {
1545 // The Glink tag points to 32 bytes before the first lazy symbol resolution
1546 // stub, which starts directly after the header.
1547 addInt(DT_PPC64_GLINK, in.plt->getVA() + target->pltHeaderSize - 32);
1548 }
1549
1550 if (config->emachine == EM_PPC64)
1551 addInt(DT_PPC64_OPT, getPPC64TargetInfo()->ppc64DynamicSectionOpt);
1552
1553 addInt(DT_NULL, 0);
1554 return entries;
1555 }
1556
finalizeContents()1557 template <class ELFT> void DynamicSection<ELFT>::finalizeContents() {
1558 if (OutputSection *sec = getPartition().dynStrTab->getParent())
1559 getParent()->link = sec->sectionIndex;
1560 this->size = computeContents().size() * this->entsize;
1561 }
1562
writeTo(uint8_t * buf)1563 template <class ELFT> void DynamicSection<ELFT>::writeTo(uint8_t *buf) {
1564 auto *p = reinterpret_cast<Elf_Dyn *>(buf);
1565
1566 for (std::pair<int32_t, uint64_t> kv : computeContents()) {
1567 p->d_tag = kv.first;
1568 p->d_un.d_val = kv.second;
1569 ++p;
1570 }
1571 }
1572
getOffset() const1573 uint64_t DynamicReloc::getOffset() const {
1574 return inputSec->getVA(offsetInSec);
1575 }
1576
computeAddend() const1577 int64_t DynamicReloc::computeAddend() const {
1578 switch (kind) {
1579 case AddendOnly:
1580 assert(sym == nullptr);
1581 return addend;
1582 case AgainstSymbol:
1583 assert(sym != nullptr);
1584 return addend;
1585 case AddendOnlyWithTargetVA:
1586 case AgainstSymbolWithTargetVA: {
1587 uint64_t ca = InputSection::getRelocTargetVA(inputSec->file, type, addend,
1588 getOffset(), *sym, expr);
1589 return config->is64 ? ca : SignExtend64<32>(ca);
1590 }
1591 case MipsMultiGotPage:
1592 assert(sym == nullptr);
1593 return getMipsPageAddr(outputSec->addr) + addend;
1594 }
1595 llvm_unreachable("Unknown DynamicReloc::Kind enum");
1596 }
1597
getSymIndex(SymbolTableBaseSection * symTab) const1598 uint32_t DynamicReloc::getSymIndex(SymbolTableBaseSection *symTab) const {
1599 if (!needsDynSymIndex())
1600 return 0;
1601
1602 size_t index = symTab->getSymbolIndex(sym);
1603 assert((index != 0 || (type != target->gotRel && type != target->pltRel) ||
1604 !mainPart->dynSymTab->getParent()) &&
1605 "GOT or PLT relocation must refer to symbol in dynamic symbol table");
1606 return index;
1607 }
1608
RelocationBaseSection(StringRef name,uint32_t type,int32_t dynamicTag,int32_t sizeDynamicTag,bool combreloc,unsigned concurrency)1609 RelocationBaseSection::RelocationBaseSection(StringRef name, uint32_t type,
1610 int32_t dynamicTag,
1611 int32_t sizeDynamicTag,
1612 bool combreloc,
1613 unsigned concurrency)
1614 : SyntheticSection(SHF_ALLOC, type, config->wordsize, name),
1615 dynamicTag(dynamicTag), sizeDynamicTag(sizeDynamicTag),
1616 relocsVec(concurrency), combreloc(combreloc) {}
1617
addSymbolReloc(RelType dynType,InputSectionBase & isec,uint64_t offsetInSec,Symbol & sym,int64_t addend,std::optional<RelType> addendRelType)1618 void RelocationBaseSection::addSymbolReloc(
1619 RelType dynType, InputSectionBase &isec, uint64_t offsetInSec, Symbol &sym,
1620 int64_t addend, std::optional<RelType> addendRelType) {
1621 addReloc(DynamicReloc::AgainstSymbol, dynType, isec, offsetInSec, sym, addend,
1622 R_ADDEND, addendRelType ? *addendRelType : target->noneRel);
1623 }
1624
addAddendOnlyRelocIfNonPreemptible(RelType dynType,GotSection & sec,uint64_t offsetInSec,Symbol & sym,RelType addendRelType)1625 void RelocationBaseSection::addAddendOnlyRelocIfNonPreemptible(
1626 RelType dynType, GotSection &sec, uint64_t offsetInSec, Symbol &sym,
1627 RelType addendRelType) {
1628 // No need to write an addend to the section for preemptible symbols.
1629 if (sym.isPreemptible)
1630 addReloc({dynType, &sec, offsetInSec, DynamicReloc::AgainstSymbol, sym, 0,
1631 R_ABS});
1632 else
1633 addReloc(DynamicReloc::AddendOnlyWithTargetVA, dynType, sec, offsetInSec,
1634 sym, 0, R_ABS, addendRelType);
1635 }
1636
mergeRels()1637 void RelocationBaseSection::mergeRels() {
1638 size_t newSize = relocs.size();
1639 for (const auto &v : relocsVec)
1640 newSize += v.size();
1641 relocs.reserve(newSize);
1642 for (const auto &v : relocsVec)
1643 llvm::append_range(relocs, v);
1644 relocsVec.clear();
1645 }
1646
partitionRels()1647 void RelocationBaseSection::partitionRels() {
1648 if (!combreloc)
1649 return;
1650 const RelType relativeRel = target->relativeRel;
1651 numRelativeRelocs =
1652 llvm::partition(relocs, [=](auto &r) { return r.type == relativeRel; }) -
1653 relocs.begin();
1654 }
1655
finalizeContents()1656 void RelocationBaseSection::finalizeContents() {
1657 SymbolTableBaseSection *symTab = getPartition().dynSymTab.get();
1658
1659 // When linking glibc statically, .rel{,a}.plt contains R_*_IRELATIVE
1660 // relocations due to IFUNC (e.g. strcpy). sh_link will be set to 0 in that
1661 // case.
1662 if (symTab && symTab->getParent())
1663 getParent()->link = symTab->getParent()->sectionIndex;
1664 else
1665 getParent()->link = 0;
1666
1667 if (in.relaPlt.get() == this && in.gotPlt->getParent()) {
1668 getParent()->flags |= ELF::SHF_INFO_LINK;
1669 getParent()->info = in.gotPlt->getParent()->sectionIndex;
1670 }
1671 if (in.relaIplt.get() == this && in.igotPlt->getParent()) {
1672 getParent()->flags |= ELF::SHF_INFO_LINK;
1673 getParent()->info = in.igotPlt->getParent()->sectionIndex;
1674 }
1675 }
1676
computeRaw(SymbolTableBaseSection * symtab)1677 void DynamicReloc::computeRaw(SymbolTableBaseSection *symtab) {
1678 r_offset = getOffset();
1679 r_sym = getSymIndex(symtab);
1680 addend = computeAddend();
1681 kind = AddendOnly; // Catch errors
1682 }
1683
computeRels()1684 void RelocationBaseSection::computeRels() {
1685 SymbolTableBaseSection *symTab = getPartition().dynSymTab.get();
1686 parallelForEach(relocs,
1687 [symTab](DynamicReloc &rel) { rel.computeRaw(symTab); });
1688 // Sort by (!IsRelative,SymIndex,r_offset). DT_REL[A]COUNT requires us to
1689 // place R_*_RELATIVE first. SymIndex is to improve locality, while r_offset
1690 // is to make results easier to read.
1691 if (combreloc) {
1692 auto nonRelative = relocs.begin() + numRelativeRelocs;
1693 parallelSort(relocs.begin(), nonRelative,
1694 [&](auto &a, auto &b) { return a.r_offset < b.r_offset; });
1695 // Non-relative relocations are few, so don't bother with parallelSort.
1696 llvm::sort(nonRelative, relocs.end(), [&](auto &a, auto &b) {
1697 return std::tie(a.r_sym, a.r_offset) < std::tie(b.r_sym, b.r_offset);
1698 });
1699 }
1700 }
1701
1702 template <class ELFT>
RelocationSection(StringRef name,bool combreloc,unsigned concurrency)1703 RelocationSection<ELFT>::RelocationSection(StringRef name, bool combreloc,
1704 unsigned concurrency)
1705 : RelocationBaseSection(name, config->isRela ? SHT_RELA : SHT_REL,
1706 config->isRela ? DT_RELA : DT_REL,
1707 config->isRela ? DT_RELASZ : DT_RELSZ, combreloc,
1708 concurrency) {
1709 this->entsize = config->isRela ? sizeof(Elf_Rela) : sizeof(Elf_Rel);
1710 }
1711
writeTo(uint8_t * buf)1712 template <class ELFT> void RelocationSection<ELFT>::writeTo(uint8_t *buf) {
1713 computeRels();
1714 for (const DynamicReloc &rel : relocs) {
1715 auto *p = reinterpret_cast<Elf_Rela *>(buf);
1716 p->r_offset = rel.r_offset;
1717 p->setSymbolAndType(rel.r_sym, rel.type, config->isMips64EL);
1718 if (config->isRela)
1719 p->r_addend = rel.addend;
1720 buf += config->isRela ? sizeof(Elf_Rela) : sizeof(Elf_Rel);
1721 }
1722 }
1723
RelrBaseSection(unsigned concurrency)1724 RelrBaseSection::RelrBaseSection(unsigned concurrency)
1725 : SyntheticSection(SHF_ALLOC,
1726 config->useAndroidRelrTags ? SHT_ANDROID_RELR : SHT_RELR,
1727 config->wordsize, ".relr.dyn"),
1728 relocsVec(concurrency) {}
1729
mergeRels()1730 void RelrBaseSection::mergeRels() {
1731 size_t newSize = relocs.size();
1732 for (const auto &v : relocsVec)
1733 newSize += v.size();
1734 relocs.reserve(newSize);
1735 for (const auto &v : relocsVec)
1736 llvm::append_range(relocs, v);
1737 relocsVec.clear();
1738 }
1739
1740 template <class ELFT>
AndroidPackedRelocationSection(StringRef name,unsigned concurrency)1741 AndroidPackedRelocationSection<ELFT>::AndroidPackedRelocationSection(
1742 StringRef name, unsigned concurrency)
1743 : RelocationBaseSection(
1744 name, config->isRela ? SHT_ANDROID_RELA : SHT_ANDROID_REL,
1745 config->isRela ? DT_ANDROID_RELA : DT_ANDROID_REL,
1746 config->isRela ? DT_ANDROID_RELASZ : DT_ANDROID_RELSZ,
1747 /*combreloc=*/false, concurrency) {
1748 this->entsize = 1;
1749 }
1750
1751 template <class ELFT>
updateAllocSize()1752 bool AndroidPackedRelocationSection<ELFT>::updateAllocSize() {
1753 // This function computes the contents of an Android-format packed relocation
1754 // section.
1755 //
1756 // This format compresses relocations by using relocation groups to factor out
1757 // fields that are common between relocations and storing deltas from previous
1758 // relocations in SLEB128 format (which has a short representation for small
1759 // numbers). A good example of a relocation type with common fields is
1760 // R_*_RELATIVE, which is normally used to represent function pointers in
1761 // vtables. In the REL format, each relative relocation has the same r_info
1762 // field, and is only different from other relative relocations in terms of
1763 // the r_offset field. By sorting relocations by offset, grouping them by
1764 // r_info and representing each relocation with only the delta from the
1765 // previous offset, each 8-byte relocation can be compressed to as little as 1
1766 // byte (or less with run-length encoding). This relocation packer was able to
1767 // reduce the size of the relocation section in an Android Chromium DSO from
1768 // 2,911,184 bytes to 174,693 bytes, or 6% of the original size.
1769 //
1770 // A relocation section consists of a header containing the literal bytes
1771 // 'APS2' followed by a sequence of SLEB128-encoded integers. The first two
1772 // elements are the total number of relocations in the section and an initial
1773 // r_offset value. The remaining elements define a sequence of relocation
1774 // groups. Each relocation group starts with a header consisting of the
1775 // following elements:
1776 //
1777 // - the number of relocations in the relocation group
1778 // - flags for the relocation group
1779 // - (if RELOCATION_GROUPED_BY_OFFSET_DELTA_FLAG is set) the r_offset delta
1780 // for each relocation in the group.
1781 // - (if RELOCATION_GROUPED_BY_INFO_FLAG is set) the value of the r_info
1782 // field for each relocation in the group.
1783 // - (if RELOCATION_GROUP_HAS_ADDEND_FLAG and
1784 // RELOCATION_GROUPED_BY_ADDEND_FLAG are set) the r_addend delta for
1785 // each relocation in the group.
1786 //
1787 // Following the relocation group header are descriptions of each of the
1788 // relocations in the group. They consist of the following elements:
1789 //
1790 // - (if RELOCATION_GROUPED_BY_OFFSET_DELTA_FLAG is not set) the r_offset
1791 // delta for this relocation.
1792 // - (if RELOCATION_GROUPED_BY_INFO_FLAG is not set) the value of the r_info
1793 // field for this relocation.
1794 // - (if RELOCATION_GROUP_HAS_ADDEND_FLAG is set and
1795 // RELOCATION_GROUPED_BY_ADDEND_FLAG is not set) the r_addend delta for
1796 // this relocation.
1797
1798 size_t oldSize = relocData.size();
1799
1800 relocData = {'A', 'P', 'S', '2'};
1801 raw_svector_ostream os(relocData);
1802 auto add = [&](int64_t v) { encodeSLEB128(v, os); };
1803
1804 // The format header includes the number of relocations and the initial
1805 // offset (we set this to zero because the first relocation group will
1806 // perform the initial adjustment).
1807 add(relocs.size());
1808 add(0);
1809
1810 std::vector<Elf_Rela> relatives, nonRelatives;
1811
1812 for (const DynamicReloc &rel : relocs) {
1813 Elf_Rela r;
1814 r.r_offset = rel.getOffset();
1815 r.setSymbolAndType(rel.getSymIndex(getPartition().dynSymTab.get()),
1816 rel.type, false);
1817 r.r_addend = config->isRela ? rel.computeAddend() : 0;
1818
1819 if (r.getType(config->isMips64EL) == target->relativeRel)
1820 relatives.push_back(r);
1821 else
1822 nonRelatives.push_back(r);
1823 }
1824
1825 llvm::sort(relatives, [](const Elf_Rel &a, const Elf_Rel &b) {
1826 return a.r_offset < b.r_offset;
1827 });
1828
1829 // Try to find groups of relative relocations which are spaced one word
1830 // apart from one another. These generally correspond to vtable entries. The
1831 // format allows these groups to be encoded using a sort of run-length
1832 // encoding, but each group will cost 7 bytes in addition to the offset from
1833 // the previous group, so it is only profitable to do this for groups of
1834 // size 8 or larger.
1835 std::vector<Elf_Rela> ungroupedRelatives;
1836 std::vector<std::vector<Elf_Rela>> relativeGroups;
1837 for (auto i = relatives.begin(), e = relatives.end(); i != e;) {
1838 std::vector<Elf_Rela> group;
1839 do {
1840 group.push_back(*i++);
1841 } while (i != e && (i - 1)->r_offset + config->wordsize == i->r_offset);
1842
1843 if (group.size() < 8)
1844 ungroupedRelatives.insert(ungroupedRelatives.end(), group.begin(),
1845 group.end());
1846 else
1847 relativeGroups.emplace_back(std::move(group));
1848 }
1849
1850 // For non-relative relocations, we would like to:
1851 // 1. Have relocations with the same symbol offset to be consecutive, so
1852 // that the runtime linker can speed-up symbol lookup by implementing an
1853 // 1-entry cache.
1854 // 2. Group relocations by r_info to reduce the size of the relocation
1855 // section.
1856 // Since the symbol offset is the high bits in r_info, sorting by r_info
1857 // allows us to do both.
1858 //
1859 // For Rela, we also want to sort by r_addend when r_info is the same. This
1860 // enables us to group by r_addend as well.
1861 llvm::sort(nonRelatives, [](const Elf_Rela &a, const Elf_Rela &b) {
1862 if (a.r_info != b.r_info)
1863 return a.r_info < b.r_info;
1864 if (a.r_addend != b.r_addend)
1865 return a.r_addend < b.r_addend;
1866 return a.r_offset < b.r_offset;
1867 });
1868
1869 // Group relocations with the same r_info. Note that each group emits a group
1870 // header and that may make the relocation section larger. It is hard to
1871 // estimate the size of a group header as the encoded size of that varies
1872 // based on r_info. However, we can approximate this trade-off by the number
1873 // of values encoded. Each group header contains 3 values, and each relocation
1874 // in a group encodes one less value, as compared to when it is not grouped.
1875 // Therefore, we only group relocations if there are 3 or more of them with
1876 // the same r_info.
1877 //
1878 // For Rela, the addend for most non-relative relocations is zero, and thus we
1879 // can usually get a smaller relocation section if we group relocations with 0
1880 // addend as well.
1881 std::vector<Elf_Rela> ungroupedNonRelatives;
1882 std::vector<std::vector<Elf_Rela>> nonRelativeGroups;
1883 for (auto i = nonRelatives.begin(), e = nonRelatives.end(); i != e;) {
1884 auto j = i + 1;
1885 while (j != e && i->r_info == j->r_info &&
1886 (!config->isRela || i->r_addend == j->r_addend))
1887 ++j;
1888 if (j - i < 3 || (config->isRela && i->r_addend != 0))
1889 ungroupedNonRelatives.insert(ungroupedNonRelatives.end(), i, j);
1890 else
1891 nonRelativeGroups.emplace_back(i, j);
1892 i = j;
1893 }
1894
1895 // Sort ungrouped relocations by offset to minimize the encoded length.
1896 llvm::sort(ungroupedNonRelatives, [](const Elf_Rela &a, const Elf_Rela &b) {
1897 return a.r_offset < b.r_offset;
1898 });
1899
1900 unsigned hasAddendIfRela =
1901 config->isRela ? RELOCATION_GROUP_HAS_ADDEND_FLAG : 0;
1902
1903 uint64_t offset = 0;
1904 uint64_t addend = 0;
1905
1906 // Emit the run-length encoding for the groups of adjacent relative
1907 // relocations. Each group is represented using two groups in the packed
1908 // format. The first is used to set the current offset to the start of the
1909 // group (and also encodes the first relocation), and the second encodes the
1910 // remaining relocations.
1911 for (std::vector<Elf_Rela> &g : relativeGroups) {
1912 // The first relocation in the group.
1913 add(1);
1914 add(RELOCATION_GROUPED_BY_OFFSET_DELTA_FLAG |
1915 RELOCATION_GROUPED_BY_INFO_FLAG | hasAddendIfRela);
1916 add(g[0].r_offset - offset);
1917 add(target->relativeRel);
1918 if (config->isRela) {
1919 add(g[0].r_addend - addend);
1920 addend = g[0].r_addend;
1921 }
1922
1923 // The remaining relocations.
1924 add(g.size() - 1);
1925 add(RELOCATION_GROUPED_BY_OFFSET_DELTA_FLAG |
1926 RELOCATION_GROUPED_BY_INFO_FLAG | hasAddendIfRela);
1927 add(config->wordsize);
1928 add(target->relativeRel);
1929 if (config->isRela) {
1930 for (const auto &i : llvm::drop_begin(g)) {
1931 add(i.r_addend - addend);
1932 addend = i.r_addend;
1933 }
1934 }
1935
1936 offset = g.back().r_offset;
1937 }
1938
1939 // Now the ungrouped relatives.
1940 if (!ungroupedRelatives.empty()) {
1941 add(ungroupedRelatives.size());
1942 add(RELOCATION_GROUPED_BY_INFO_FLAG | hasAddendIfRela);
1943 add(target->relativeRel);
1944 for (Elf_Rela &r : ungroupedRelatives) {
1945 add(r.r_offset - offset);
1946 offset = r.r_offset;
1947 if (config->isRela) {
1948 add(r.r_addend - addend);
1949 addend = r.r_addend;
1950 }
1951 }
1952 }
1953
1954 // Grouped non-relatives.
1955 for (ArrayRef<Elf_Rela> g : nonRelativeGroups) {
1956 add(g.size());
1957 add(RELOCATION_GROUPED_BY_INFO_FLAG);
1958 add(g[0].r_info);
1959 for (const Elf_Rela &r : g) {
1960 add(r.r_offset - offset);
1961 offset = r.r_offset;
1962 }
1963 addend = 0;
1964 }
1965
1966 // Finally the ungrouped non-relative relocations.
1967 if (!ungroupedNonRelatives.empty()) {
1968 add(ungroupedNonRelatives.size());
1969 add(hasAddendIfRela);
1970 for (Elf_Rela &r : ungroupedNonRelatives) {
1971 add(r.r_offset - offset);
1972 offset = r.r_offset;
1973 add(r.r_info);
1974 if (config->isRela) {
1975 add(r.r_addend - addend);
1976 addend = r.r_addend;
1977 }
1978 }
1979 }
1980
1981 // Don't allow the section to shrink; otherwise the size of the section can
1982 // oscillate infinitely.
1983 if (relocData.size() < oldSize)
1984 relocData.append(oldSize - relocData.size(), 0);
1985
1986 // Returns whether the section size changed. We need to keep recomputing both
1987 // section layout and the contents of this section until the size converges
1988 // because changing this section's size can affect section layout, which in
1989 // turn can affect the sizes of the LEB-encoded integers stored in this
1990 // section.
1991 return relocData.size() != oldSize;
1992 }
1993
1994 template <class ELFT>
RelrSection(unsigned concurrency)1995 RelrSection<ELFT>::RelrSection(unsigned concurrency)
1996 : RelrBaseSection(concurrency) {
1997 this->entsize = config->wordsize;
1998 }
1999
updateAllocSize()2000 template <class ELFT> bool RelrSection<ELFT>::updateAllocSize() {
2001 // This function computes the contents of an SHT_RELR packed relocation
2002 // section.
2003 //
2004 // Proposal for adding SHT_RELR sections to generic-abi is here:
2005 // https://groups.google.com/forum/#!topic/generic-abi/bX460iggiKg
2006 //
2007 // The encoded sequence of Elf64_Relr entries in a SHT_RELR section looks
2008 // like [ AAAAAAAA BBBBBBB1 BBBBBBB1 ... AAAAAAAA BBBBBB1 ... ]
2009 //
2010 // i.e. start with an address, followed by any number of bitmaps. The address
2011 // entry encodes 1 relocation. The subsequent bitmap entries encode up to 63
2012 // relocations each, at subsequent offsets following the last address entry.
2013 //
2014 // The bitmap entries must have 1 in the least significant bit. The assumption
2015 // here is that an address cannot have 1 in lsb. Odd addresses are not
2016 // supported.
2017 //
2018 // Excluding the least significant bit in the bitmap, each non-zero bit in
2019 // the bitmap represents a relocation to be applied to a corresponding machine
2020 // word that follows the base address word. The second least significant bit
2021 // represents the machine word immediately following the initial address, and
2022 // each bit that follows represents the next word, in linear order. As such,
2023 // a single bitmap can encode up to 31 relocations in a 32-bit object, and
2024 // 63 relocations in a 64-bit object.
2025 //
2026 // This encoding has a couple of interesting properties:
2027 // 1. Looking at any entry, it is clear whether it's an address or a bitmap:
2028 // even means address, odd means bitmap.
2029 // 2. Just a simple list of addresses is a valid encoding.
2030
2031 size_t oldSize = relrRelocs.size();
2032 relrRelocs.clear();
2033
2034 // Same as Config->Wordsize but faster because this is a compile-time
2035 // constant.
2036 const size_t wordsize = sizeof(typename ELFT::uint);
2037
2038 // Number of bits to use for the relocation offsets bitmap.
2039 // Must be either 63 or 31.
2040 const size_t nBits = wordsize * 8 - 1;
2041
2042 // Get offsets for all relative relocations and sort them.
2043 std::unique_ptr<uint64_t[]> offsets(new uint64_t[relocs.size()]);
2044 for (auto [i, r] : llvm::enumerate(relocs))
2045 offsets[i] = r.getOffset();
2046 llvm::sort(offsets.get(), offsets.get() + relocs.size());
2047
2048 // For each leading relocation, find following ones that can be folded
2049 // as a bitmap and fold them.
2050 for (size_t i = 0, e = relocs.size(); i != e;) {
2051 // Add a leading relocation.
2052 relrRelocs.push_back(Elf_Relr(offsets[i]));
2053 uint64_t base = offsets[i] + wordsize;
2054 ++i;
2055
2056 // Find foldable relocations to construct bitmaps.
2057 for (;;) {
2058 uint64_t bitmap = 0;
2059 for (; i != e; ++i) {
2060 uint64_t d = offsets[i] - base;
2061 if (d >= nBits * wordsize || d % wordsize)
2062 break;
2063 bitmap |= uint64_t(1) << (d / wordsize);
2064 }
2065 if (!bitmap)
2066 break;
2067 relrRelocs.push_back(Elf_Relr((bitmap << 1) | 1));
2068 base += nBits * wordsize;
2069 }
2070 }
2071
2072 // Don't allow the section to shrink; otherwise the size of the section can
2073 // oscillate infinitely. Trailing 1s do not decode to more relocations.
2074 if (relrRelocs.size() < oldSize) {
2075 log(".relr.dyn needs " + Twine(oldSize - relrRelocs.size()) +
2076 " padding word(s)");
2077 relrRelocs.resize(oldSize, Elf_Relr(1));
2078 }
2079
2080 return relrRelocs.size() != oldSize;
2081 }
2082
SymbolTableBaseSection(StringTableSection & strTabSec)2083 SymbolTableBaseSection::SymbolTableBaseSection(StringTableSection &strTabSec)
2084 : SyntheticSection(strTabSec.isDynamic() ? (uint64_t)SHF_ALLOC : 0,
2085 strTabSec.isDynamic() ? SHT_DYNSYM : SHT_SYMTAB,
2086 config->wordsize,
2087 strTabSec.isDynamic() ? ".dynsym" : ".symtab"),
2088 strTabSec(strTabSec) {}
2089
2090 // Orders symbols according to their positions in the GOT,
2091 // in compliance with MIPS ABI rules.
2092 // See "Global Offset Table" in Chapter 5 in the following document
2093 // for detailed description:
2094 // ftp://www.linux-mips.org/pub/linux/mips/doc/ABI/mipsabi.pdf
sortMipsSymbols(const SymbolTableEntry & l,const SymbolTableEntry & r)2095 static bool sortMipsSymbols(const SymbolTableEntry &l,
2096 const SymbolTableEntry &r) {
2097 // Sort entries related to non-local preemptible symbols by GOT indexes.
2098 // All other entries go to the beginning of a dynsym in arbitrary order.
2099 if (l.sym->isInGot() && r.sym->isInGot())
2100 return l.sym->getGotIdx() < r.sym->getGotIdx();
2101 if (!l.sym->isInGot() && !r.sym->isInGot())
2102 return false;
2103 return !l.sym->isInGot();
2104 }
2105
finalizeContents()2106 void SymbolTableBaseSection::finalizeContents() {
2107 if (OutputSection *sec = strTabSec.getParent())
2108 getParent()->link = sec->sectionIndex;
2109
2110 if (this->type != SHT_DYNSYM) {
2111 sortSymTabSymbols();
2112 return;
2113 }
2114
2115 // If it is a .dynsym, there should be no local symbols, but we need
2116 // to do a few things for the dynamic linker.
2117
2118 // Section's Info field has the index of the first non-local symbol.
2119 // Because the first symbol entry is a null entry, 1 is the first.
2120 getParent()->info = 1;
2121
2122 if (getPartition().gnuHashTab) {
2123 // NB: It also sorts Symbols to meet the GNU hash table requirements.
2124 getPartition().gnuHashTab->addSymbols(symbols);
2125 } else if (config->emachine == EM_MIPS) {
2126 llvm::stable_sort(symbols, sortMipsSymbols);
2127 }
2128
2129 // Only the main partition's dynsym indexes are stored in the symbols
2130 // themselves. All other partitions use a lookup table.
2131 if (this == mainPart->dynSymTab.get()) {
2132 size_t i = 0;
2133 for (const SymbolTableEntry &s : symbols)
2134 s.sym->dynsymIndex = ++i;
2135 }
2136 }
2137
2138 // The ELF spec requires that all local symbols precede global symbols, so we
2139 // sort symbol entries in this function. (For .dynsym, we don't do that because
2140 // symbols for dynamic linking are inherently all globals.)
2141 //
2142 // Aside from above, we put local symbols in groups starting with the STT_FILE
2143 // symbol. That is convenient for purpose of identifying where are local symbols
2144 // coming from.
sortSymTabSymbols()2145 void SymbolTableBaseSection::sortSymTabSymbols() {
2146 // Move all local symbols before global symbols.
2147 auto e = std::stable_partition(
2148 symbols.begin(), symbols.end(),
2149 [](const SymbolTableEntry &s) { return s.sym->isLocal(); });
2150 size_t numLocals = e - symbols.begin();
2151 getParent()->info = numLocals + 1;
2152
2153 // We want to group the local symbols by file. For that we rebuild the local
2154 // part of the symbols vector. We do not need to care about the STT_FILE
2155 // symbols, they are already naturally placed first in each group. That
2156 // happens because STT_FILE is always the first symbol in the object and hence
2157 // precede all other local symbols we add for a file.
2158 MapVector<InputFile *, SmallVector<SymbolTableEntry, 0>> arr;
2159 for (const SymbolTableEntry &s : llvm::make_range(symbols.begin(), e))
2160 arr[s.sym->file].push_back(s);
2161
2162 auto i = symbols.begin();
2163 for (auto &p : arr)
2164 for (SymbolTableEntry &entry : p.second)
2165 *i++ = entry;
2166 }
2167
addSymbol(Symbol * b)2168 void SymbolTableBaseSection::addSymbol(Symbol *b) {
2169 // Adding a local symbol to a .dynsym is a bug.
2170 assert(this->type != SHT_DYNSYM || !b->isLocal());
2171 symbols.push_back({b, strTabSec.addString(b->getName(), false)});
2172 }
2173
getSymbolIndex(Symbol * sym)2174 size_t SymbolTableBaseSection::getSymbolIndex(Symbol *sym) {
2175 if (this == mainPart->dynSymTab.get())
2176 return sym->dynsymIndex;
2177
2178 // Initializes symbol lookup tables lazily. This is used only for -r,
2179 // --emit-relocs and dynsyms in partitions other than the main one.
2180 llvm::call_once(onceFlag, [&] {
2181 symbolIndexMap.reserve(symbols.size());
2182 size_t i = 0;
2183 for (const SymbolTableEntry &e : symbols) {
2184 if (e.sym->type == STT_SECTION)
2185 sectionIndexMap[e.sym->getOutputSection()] = ++i;
2186 else
2187 symbolIndexMap[e.sym] = ++i;
2188 }
2189 });
2190
2191 // Section symbols are mapped based on their output sections
2192 // to maintain their semantics.
2193 if (sym->type == STT_SECTION)
2194 return sectionIndexMap.lookup(sym->getOutputSection());
2195 return symbolIndexMap.lookup(sym);
2196 }
2197
2198 template <class ELFT>
SymbolTableSection(StringTableSection & strTabSec)2199 SymbolTableSection<ELFT>::SymbolTableSection(StringTableSection &strTabSec)
2200 : SymbolTableBaseSection(strTabSec) {
2201 this->entsize = sizeof(Elf_Sym);
2202 }
2203
getCommonSec(Symbol * sym)2204 static BssSection *getCommonSec(Symbol *sym) {
2205 if (config->relocatable)
2206 if (auto *d = dyn_cast<Defined>(sym))
2207 return dyn_cast_or_null<BssSection>(d->section);
2208 return nullptr;
2209 }
2210
getSymSectionIndex(Symbol * sym)2211 static uint32_t getSymSectionIndex(Symbol *sym) {
2212 assert(!(sym->hasFlag(NEEDS_COPY) && sym->isObject()));
2213 if (!isa<Defined>(sym) || sym->hasFlag(NEEDS_COPY))
2214 return SHN_UNDEF;
2215 if (const OutputSection *os = sym->getOutputSection())
2216 return os->sectionIndex >= SHN_LORESERVE ? (uint32_t)SHN_XINDEX
2217 : os->sectionIndex;
2218 return SHN_ABS;
2219 }
2220
2221 // Write the internal symbol table contents to the output symbol table.
writeTo(uint8_t * buf)2222 template <class ELFT> void SymbolTableSection<ELFT>::writeTo(uint8_t *buf) {
2223 // The first entry is a null entry as per the ELF spec.
2224 buf += sizeof(Elf_Sym);
2225
2226 auto *eSym = reinterpret_cast<Elf_Sym *>(buf);
2227
2228 for (SymbolTableEntry &ent : symbols) {
2229 Symbol *sym = ent.sym;
2230 bool isDefinedHere = type == SHT_SYMTAB || sym->partition == partition;
2231
2232 // Set st_name, st_info and st_other.
2233 eSym->st_name = ent.strTabOffset;
2234 eSym->setBindingAndType(sym->binding, sym->type);
2235 eSym->st_other = sym->stOther;
2236
2237 if (BssSection *commonSec = getCommonSec(sym)) {
2238 // When -r is specified, a COMMON symbol is not allocated. Its st_shndx
2239 // holds SHN_COMMON and st_value holds the alignment.
2240 eSym->st_shndx = SHN_COMMON;
2241 eSym->st_value = commonSec->addralign;
2242 eSym->st_size = cast<Defined>(sym)->size;
2243 } else {
2244 const uint32_t shndx = getSymSectionIndex(sym);
2245 if (isDefinedHere) {
2246 eSym->st_shndx = shndx;
2247 eSym->st_value = sym->getVA();
2248 // Copy symbol size if it is a defined symbol. st_size is not
2249 // significant for undefined symbols, so whether copying it or not is up
2250 // to us if that's the case. We'll leave it as zero because by not
2251 // setting a value, we can get the exact same outputs for two sets of
2252 // input files that differ only in undefined symbol size in DSOs.
2253 eSym->st_size = shndx != SHN_UNDEF ? cast<Defined>(sym)->size : 0;
2254 } else {
2255 eSym->st_shndx = 0;
2256 eSym->st_value = 0;
2257 eSym->st_size = 0;
2258 }
2259 }
2260
2261 ++eSym;
2262 }
2263
2264 // On MIPS we need to mark symbol which has a PLT entry and requires
2265 // pointer equality by STO_MIPS_PLT flag. That is necessary to help
2266 // dynamic linker distinguish such symbols and MIPS lazy-binding stubs.
2267 // https://sourceware.org/ml/binutils/2008-07/txt00000.txt
2268 if (config->emachine == EM_MIPS) {
2269 auto *eSym = reinterpret_cast<Elf_Sym *>(buf);
2270
2271 for (SymbolTableEntry &ent : symbols) {
2272 Symbol *sym = ent.sym;
2273 if (sym->isInPlt() && sym->hasFlag(NEEDS_COPY))
2274 eSym->st_other |= STO_MIPS_PLT;
2275 if (isMicroMips()) {
2276 // We already set the less-significant bit for symbols
2277 // marked by the `STO_MIPS_MICROMIPS` flag and for microMIPS PLT
2278 // records. That allows us to distinguish such symbols in
2279 // the `MIPS<ELFT>::relocate()` routine. Now we should
2280 // clear that bit for non-dynamic symbol table, so tools
2281 // like `objdump` will be able to deal with a correct
2282 // symbol position.
2283 if (sym->isDefined() &&
2284 ((sym->stOther & STO_MIPS_MICROMIPS) || sym->hasFlag(NEEDS_COPY))) {
2285 if (!strTabSec.isDynamic())
2286 eSym->st_value &= ~1;
2287 eSym->st_other |= STO_MIPS_MICROMIPS;
2288 }
2289 }
2290 if (config->relocatable)
2291 if (auto *d = dyn_cast<Defined>(sym))
2292 if (isMipsPIC<ELFT>(d))
2293 eSym->st_other |= STO_MIPS_PIC;
2294 ++eSym;
2295 }
2296 }
2297 }
2298
SymtabShndxSection()2299 SymtabShndxSection::SymtabShndxSection()
2300 : SyntheticSection(0, SHT_SYMTAB_SHNDX, 4, ".symtab_shndx") {
2301 this->entsize = 4;
2302 }
2303
writeTo(uint8_t * buf)2304 void SymtabShndxSection::writeTo(uint8_t *buf) {
2305 // We write an array of 32 bit values, where each value has 1:1 association
2306 // with an entry in .symtab. If the corresponding entry contains SHN_XINDEX,
2307 // we need to write actual index, otherwise, we must write SHN_UNDEF(0).
2308 buf += 4; // Ignore .symtab[0] entry.
2309 for (const SymbolTableEntry &entry : in.symTab->getSymbols()) {
2310 if (!getCommonSec(entry.sym) && getSymSectionIndex(entry.sym) == SHN_XINDEX)
2311 write32(buf, entry.sym->getOutputSection()->sectionIndex);
2312 buf += 4;
2313 }
2314 }
2315
isNeeded() const2316 bool SymtabShndxSection::isNeeded() const {
2317 // SHT_SYMTAB can hold symbols with section indices values up to
2318 // SHN_LORESERVE. If we need more, we want to use extension SHT_SYMTAB_SHNDX
2319 // section. Problem is that we reveal the final section indices a bit too
2320 // late, and we do not know them here. For simplicity, we just always create
2321 // a .symtab_shndx section when the amount of output sections is huge.
2322 size_t size = 0;
2323 for (SectionCommand *cmd : script->sectionCommands)
2324 if (isa<OutputDesc>(cmd))
2325 ++size;
2326 return size >= SHN_LORESERVE;
2327 }
2328
finalizeContents()2329 void SymtabShndxSection::finalizeContents() {
2330 getParent()->link = in.symTab->getParent()->sectionIndex;
2331 }
2332
getSize() const2333 size_t SymtabShndxSection::getSize() const {
2334 return in.symTab->getNumSymbols() * 4;
2335 }
2336
2337 // .hash and .gnu.hash sections contain on-disk hash tables that map
2338 // symbol names to their dynamic symbol table indices. Their purpose
2339 // is to help the dynamic linker resolve symbols quickly. If ELF files
2340 // don't have them, the dynamic linker has to do linear search on all
2341 // dynamic symbols, which makes programs slower. Therefore, a .hash
2342 // section is added to a DSO by default.
2343 //
2344 // The Unix semantics of resolving dynamic symbols is somewhat expensive.
2345 // Each ELF file has a list of DSOs that the ELF file depends on and a
2346 // list of dynamic symbols that need to be resolved from any of the
2347 // DSOs. That means resolving all dynamic symbols takes O(m)*O(n)
2348 // where m is the number of DSOs and n is the number of dynamic
2349 // symbols. For modern large programs, both m and n are large. So
2350 // making each step faster by using hash tables substantially
2351 // improves time to load programs.
2352 //
2353 // (Note that this is not the only way to design the shared library.
2354 // For instance, the Windows DLL takes a different approach. On
2355 // Windows, each dynamic symbol has a name of DLL from which the symbol
2356 // has to be resolved. That makes the cost of symbol resolution O(n).
2357 // This disables some hacky techniques you can use on Unix such as
2358 // LD_PRELOAD, but this is arguably better semantics than the Unix ones.)
2359 //
2360 // Due to historical reasons, we have two different hash tables, .hash
2361 // and .gnu.hash. They are for the same purpose, and .gnu.hash is a new
2362 // and better version of .hash. .hash is just an on-disk hash table, but
2363 // .gnu.hash has a bloom filter in addition to a hash table to skip
2364 // DSOs very quickly. If you are sure that your dynamic linker knows
2365 // about .gnu.hash, you want to specify --hash-style=gnu. Otherwise, a
2366 // safe bet is to specify --hash-style=both for backward compatibility.
GnuHashTableSection()2367 GnuHashTableSection::GnuHashTableSection()
2368 : SyntheticSection(SHF_ALLOC, SHT_GNU_HASH, config->wordsize, ".gnu.hash") {
2369 }
2370
finalizeContents()2371 void GnuHashTableSection::finalizeContents() {
2372 if (OutputSection *sec = getPartition().dynSymTab->getParent())
2373 getParent()->link = sec->sectionIndex;
2374
2375 // Computes bloom filter size in word size. We want to allocate 12
2376 // bits for each symbol. It must be a power of two.
2377 if (symbols.empty()) {
2378 maskWords = 1;
2379 } else {
2380 uint64_t numBits = symbols.size() * 12;
2381 maskWords = NextPowerOf2(numBits / (config->wordsize * 8));
2382 }
2383
2384 size = 16; // Header
2385 size += config->wordsize * maskWords; // Bloom filter
2386 size += nBuckets * 4; // Hash buckets
2387 size += symbols.size() * 4; // Hash values
2388 }
2389
writeTo(uint8_t * buf)2390 void GnuHashTableSection::writeTo(uint8_t *buf) {
2391 // Write a header.
2392 write32(buf, nBuckets);
2393 write32(buf + 4, getPartition().dynSymTab->getNumSymbols() - symbols.size());
2394 write32(buf + 8, maskWords);
2395 write32(buf + 12, Shift2);
2396 buf += 16;
2397
2398 // Write the 2-bit bloom filter.
2399 const unsigned c = config->is64 ? 64 : 32;
2400 for (const Entry &sym : symbols) {
2401 // When C = 64, we choose a word with bits [6:...] and set 1 to two bits in
2402 // the word using bits [0:5] and [26:31].
2403 size_t i = (sym.hash / c) & (maskWords - 1);
2404 uint64_t val = readUint(buf + i * config->wordsize);
2405 val |= uint64_t(1) << (sym.hash % c);
2406 val |= uint64_t(1) << ((sym.hash >> Shift2) % c);
2407 writeUint(buf + i * config->wordsize, val);
2408 }
2409 buf += config->wordsize * maskWords;
2410
2411 // Write the hash table.
2412 uint32_t *buckets = reinterpret_cast<uint32_t *>(buf);
2413 uint32_t oldBucket = -1;
2414 uint32_t *values = buckets + nBuckets;
2415 for (auto i = symbols.begin(), e = symbols.end(); i != e; ++i) {
2416 // Write a hash value. It represents a sequence of chains that share the
2417 // same hash modulo value. The last element of each chain is terminated by
2418 // LSB 1.
2419 uint32_t hash = i->hash;
2420 bool isLastInChain = (i + 1) == e || i->bucketIdx != (i + 1)->bucketIdx;
2421 hash = isLastInChain ? hash | 1 : hash & ~1;
2422 write32(values++, hash);
2423
2424 if (i->bucketIdx == oldBucket)
2425 continue;
2426 // Write a hash bucket. Hash buckets contain indices in the following hash
2427 // value table.
2428 write32(buckets + i->bucketIdx,
2429 getPartition().dynSymTab->getSymbolIndex(i->sym));
2430 oldBucket = i->bucketIdx;
2431 }
2432 }
2433
2434 // Add symbols to this symbol hash table. Note that this function
2435 // destructively sort a given vector -- which is needed because
2436 // GNU-style hash table places some sorting requirements.
addSymbols(SmallVectorImpl<SymbolTableEntry> & v)2437 void GnuHashTableSection::addSymbols(SmallVectorImpl<SymbolTableEntry> &v) {
2438 // We cannot use 'auto' for Mid because GCC 6.1 cannot deduce
2439 // its type correctly.
2440 auto mid =
2441 std::stable_partition(v.begin(), v.end(), [&](const SymbolTableEntry &s) {
2442 return !s.sym->isDefined() || s.sym->partition != partition;
2443 });
2444
2445 // We chose load factor 4 for the on-disk hash table. For each hash
2446 // collision, the dynamic linker will compare a uint32_t hash value.
2447 // Since the integer comparison is quite fast, we believe we can
2448 // make the load factor even larger. 4 is just a conservative choice.
2449 //
2450 // Note that we don't want to create a zero-sized hash table because
2451 // Android loader as of 2018 doesn't like a .gnu.hash containing such
2452 // table. If that's the case, we create a hash table with one unused
2453 // dummy slot.
2454 nBuckets = std::max<size_t>((v.end() - mid) / 4, 1);
2455
2456 if (mid == v.end())
2457 return;
2458
2459 for (SymbolTableEntry &ent : llvm::make_range(mid, v.end())) {
2460 Symbol *b = ent.sym;
2461 uint32_t hash = hashGnu(b->getName());
2462 uint32_t bucketIdx = hash % nBuckets;
2463 symbols.push_back({b, ent.strTabOffset, hash, bucketIdx});
2464 }
2465
2466 llvm::sort(symbols, [](const Entry &l, const Entry &r) {
2467 return std::tie(l.bucketIdx, l.strTabOffset) <
2468 std::tie(r.bucketIdx, r.strTabOffset);
2469 });
2470
2471 v.erase(mid, v.end());
2472 for (const Entry &ent : symbols)
2473 v.push_back({ent.sym, ent.strTabOffset});
2474 }
2475
HashTableSection()2476 HashTableSection::HashTableSection()
2477 : SyntheticSection(SHF_ALLOC, SHT_HASH, 4, ".hash") {
2478 this->entsize = 4;
2479 }
2480
finalizeContents()2481 void HashTableSection::finalizeContents() {
2482 SymbolTableBaseSection *symTab = getPartition().dynSymTab.get();
2483
2484 if (OutputSection *sec = symTab->getParent())
2485 getParent()->link = sec->sectionIndex;
2486
2487 unsigned numEntries = 2; // nbucket and nchain.
2488 numEntries += symTab->getNumSymbols(); // The chain entries.
2489
2490 // Create as many buckets as there are symbols.
2491 numEntries += symTab->getNumSymbols();
2492 this->size = numEntries * 4;
2493 }
2494
writeTo(uint8_t * buf)2495 void HashTableSection::writeTo(uint8_t *buf) {
2496 SymbolTableBaseSection *symTab = getPartition().dynSymTab.get();
2497 unsigned numSymbols = symTab->getNumSymbols();
2498
2499 uint32_t *p = reinterpret_cast<uint32_t *>(buf);
2500 write32(p++, numSymbols); // nbucket
2501 write32(p++, numSymbols); // nchain
2502
2503 uint32_t *buckets = p;
2504 uint32_t *chains = p + numSymbols;
2505
2506 for (const SymbolTableEntry &s : symTab->getSymbols()) {
2507 Symbol *sym = s.sym;
2508 StringRef name = sym->getName();
2509 unsigned i = sym->dynsymIndex;
2510 uint32_t hash = hashSysV(name) % numSymbols;
2511 chains[i] = buckets[hash];
2512 write32(buckets + hash, i);
2513 }
2514 }
2515
PltSection()2516 PltSection::PltSection()
2517 : SyntheticSection(SHF_ALLOC | SHF_EXECINSTR, SHT_PROGBITS, 16, ".plt"),
2518 headerSize(target->pltHeaderSize) {
2519 // On PowerPC, this section contains lazy symbol resolvers.
2520 if (config->emachine == EM_PPC64) {
2521 name = ".glink";
2522 addralign = 4;
2523 }
2524
2525 // On x86 when IBT is enabled, this section contains the second PLT (lazy
2526 // symbol resolvers).
2527 if ((config->emachine == EM_386 || config->emachine == EM_X86_64) &&
2528 (config->andFeatures & GNU_PROPERTY_X86_FEATURE_1_IBT))
2529 name = ".plt.sec";
2530
2531 // The PLT needs to be writable on SPARC as the dynamic linker will
2532 // modify the instructions in the PLT entries.
2533 if (config->emachine == EM_SPARCV9)
2534 this->flags |= SHF_WRITE;
2535 }
2536
writeTo(uint8_t * buf)2537 void PltSection::writeTo(uint8_t *buf) {
2538 // At beginning of PLT, we have code to call the dynamic
2539 // linker to resolve dynsyms at runtime. Write such code.
2540 target->writePltHeader(buf);
2541 size_t off = headerSize;
2542
2543 for (const Symbol *sym : entries) {
2544 target->writePlt(buf + off, *sym, getVA() + off);
2545 off += target->pltEntrySize;
2546 }
2547 }
2548
addEntry(Symbol & sym)2549 void PltSection::addEntry(Symbol &sym) {
2550 assert(sym.auxIdx == symAux.size() - 1);
2551 symAux.back().pltIdx = entries.size();
2552 entries.push_back(&sym);
2553 }
2554
getSize() const2555 size_t PltSection::getSize() const {
2556 return headerSize + entries.size() * target->pltEntrySize;
2557 }
2558
isNeeded() const2559 bool PltSection::isNeeded() const {
2560 // For -z retpolineplt, .iplt needs the .plt header.
2561 return !entries.empty() || (config->zRetpolineplt && in.iplt->isNeeded());
2562 }
2563
2564 // Used by ARM to add mapping symbols in the PLT section, which aid
2565 // disassembly.
addSymbols()2566 void PltSection::addSymbols() {
2567 target->addPltHeaderSymbols(*this);
2568
2569 size_t off = headerSize;
2570 for (size_t i = 0; i < entries.size(); ++i) {
2571 target->addPltSymbols(*this, off);
2572 off += target->pltEntrySize;
2573 }
2574 }
2575
IpltSection()2576 IpltSection::IpltSection()
2577 : SyntheticSection(SHF_ALLOC | SHF_EXECINSTR, SHT_PROGBITS, 16, ".iplt") {
2578 if (config->emachine == EM_PPC || config->emachine == EM_PPC64) {
2579 name = ".glink";
2580 addralign = 4;
2581 }
2582 }
2583
writeTo(uint8_t * buf)2584 void IpltSection::writeTo(uint8_t *buf) {
2585 uint32_t off = 0;
2586 for (const Symbol *sym : entries) {
2587 target->writeIplt(buf + off, *sym, getVA() + off);
2588 off += target->ipltEntrySize;
2589 }
2590 }
2591
getSize() const2592 size_t IpltSection::getSize() const {
2593 return entries.size() * target->ipltEntrySize;
2594 }
2595
addEntry(Symbol & sym)2596 void IpltSection::addEntry(Symbol &sym) {
2597 assert(sym.auxIdx == symAux.size() - 1);
2598 symAux.back().pltIdx = entries.size();
2599 entries.push_back(&sym);
2600 }
2601
2602 // ARM uses mapping symbols to aid disassembly.
addSymbols()2603 void IpltSection::addSymbols() {
2604 size_t off = 0;
2605 for (size_t i = 0, e = entries.size(); i != e; ++i) {
2606 target->addPltSymbols(*this, off);
2607 off += target->pltEntrySize;
2608 }
2609 }
2610
PPC32GlinkSection()2611 PPC32GlinkSection::PPC32GlinkSection() {
2612 name = ".glink";
2613 addralign = 4;
2614 }
2615
writeTo(uint8_t * buf)2616 void PPC32GlinkSection::writeTo(uint8_t *buf) {
2617 writePPC32GlinkSection(buf, entries.size());
2618 }
2619
getSize() const2620 size_t PPC32GlinkSection::getSize() const {
2621 return headerSize + entries.size() * target->pltEntrySize + footerSize;
2622 }
2623
2624 // This is an x86-only extra PLT section and used only when a security
2625 // enhancement feature called CET is enabled. In this comment, I'll explain what
2626 // the feature is and why we have two PLT sections if CET is enabled.
2627 //
2628 // So, what does CET do? CET introduces a new restriction to indirect jump
2629 // instructions. CET works this way. Assume that CET is enabled. Then, if you
2630 // execute an indirect jump instruction, the processor verifies that a special
2631 // "landing pad" instruction (which is actually a repurposed NOP instruction and
2632 // now called "endbr32" or "endbr64") is at the jump target. If the jump target
2633 // does not start with that instruction, the processor raises an exception
2634 // instead of continuing executing code.
2635 //
2636 // If CET is enabled, the compiler emits endbr to all locations where indirect
2637 // jumps may jump to.
2638 //
2639 // This mechanism makes it extremely hard to transfer the control to a middle of
2640 // a function that is not supporsed to be a indirect jump target, preventing
2641 // certain types of attacks such as ROP or JOP.
2642 //
2643 // Note that the processors in the market as of 2019 don't actually support the
2644 // feature. Only the spec is available at the moment.
2645 //
2646 // Now, I'll explain why we have this extra PLT section for CET.
2647 //
2648 // Since you can indirectly jump to a PLT entry, we have to make PLT entries
2649 // start with endbr. The problem is there's no extra space for endbr (which is 4
2650 // bytes long), as the PLT entry is only 16 bytes long and all bytes are already
2651 // used.
2652 //
2653 // In order to deal with the issue, we split a PLT entry into two PLT entries.
2654 // Remember that each PLT entry contains code to jump to an address read from
2655 // .got.plt AND code to resolve a dynamic symbol lazily. With the 2-PLT scheme,
2656 // the former code is written to .plt.sec, and the latter code is written to
2657 // .plt.
2658 //
2659 // Lazy symbol resolution in the 2-PLT scheme works in the usual way, except
2660 // that the regular .plt is now called .plt.sec and .plt is repurposed to
2661 // contain only code for lazy symbol resolution.
2662 //
2663 // In other words, this is how the 2-PLT scheme works. Application code is
2664 // supposed to jump to .plt.sec to call an external function. Each .plt.sec
2665 // entry contains code to read an address from a corresponding .got.plt entry
2666 // and jump to that address. Addresses in .got.plt initially point to .plt, so
2667 // when an application calls an external function for the first time, the
2668 // control is transferred to a function that resolves a symbol name from
2669 // external shared object files. That function then rewrites a .got.plt entry
2670 // with a resolved address, so that the subsequent function calls directly jump
2671 // to a desired location from .plt.sec.
2672 //
2673 // There is an open question as to whether the 2-PLT scheme was desirable or
2674 // not. We could have simply extended the PLT entry size to 32-bytes to
2675 // accommodate endbr, and that scheme would have been much simpler than the
2676 // 2-PLT scheme. One reason to split PLT was, by doing that, we could keep hot
2677 // code (.plt.sec) from cold code (.plt). But as far as I know no one proved
2678 // that the optimization actually makes a difference.
2679 //
2680 // That said, the 2-PLT scheme is a part of the ABI, debuggers and other tools
2681 // depend on it, so we implement the ABI.
IBTPltSection()2682 IBTPltSection::IBTPltSection()
2683 : SyntheticSection(SHF_ALLOC | SHF_EXECINSTR, SHT_PROGBITS, 16, ".plt") {}
2684
writeTo(uint8_t * buf)2685 void IBTPltSection::writeTo(uint8_t *buf) {
2686 target->writeIBTPlt(buf, in.plt->getNumEntries());
2687 }
2688
getSize() const2689 size_t IBTPltSection::getSize() const {
2690 // 16 is the header size of .plt.
2691 return 16 + in.plt->getNumEntries() * target->pltEntrySize;
2692 }
2693
isNeeded() const2694 bool IBTPltSection::isNeeded() const { return in.plt->getNumEntries() > 0; }
2695
RelroPaddingSection()2696 RelroPaddingSection::RelroPaddingSection()
2697 : SyntheticSection(SHF_ALLOC | SHF_WRITE, SHT_NOBITS, 1, ".relro_padding") {
2698 }
2699
2700 // The string hash function for .gdb_index.
computeGdbHash(StringRef s)2701 static uint32_t computeGdbHash(StringRef s) {
2702 uint32_t h = 0;
2703 for (uint8_t c : s)
2704 h = h * 67 + toLower(c) - 113;
2705 return h;
2706 }
2707
GdbIndexSection()2708 GdbIndexSection::GdbIndexSection()
2709 : SyntheticSection(0, SHT_PROGBITS, 1, ".gdb_index") {}
2710
2711 // Returns the desired size of an on-disk hash table for a .gdb_index section.
2712 // There's a tradeoff between size and collision rate. We aim 75% utilization.
computeSymtabSize() const2713 size_t GdbIndexSection::computeSymtabSize() const {
2714 return std::max<size_t>(NextPowerOf2(symbols.size() * 4 / 3), 1024);
2715 }
2716
2717 static SmallVector<GdbIndexSection::CuEntry, 0>
readCuList(DWARFContext & dwarf)2718 readCuList(DWARFContext &dwarf) {
2719 SmallVector<GdbIndexSection::CuEntry, 0> ret;
2720 for (std::unique_ptr<DWARFUnit> &cu : dwarf.compile_units())
2721 ret.push_back({cu->getOffset(), cu->getLength() + 4});
2722 return ret;
2723 }
2724
2725 static SmallVector<GdbIndexSection::AddressEntry, 0>
readAddressAreas(DWARFContext & dwarf,InputSection * sec)2726 readAddressAreas(DWARFContext &dwarf, InputSection *sec) {
2727 SmallVector<GdbIndexSection::AddressEntry, 0> ret;
2728
2729 uint32_t cuIdx = 0;
2730 for (std::unique_ptr<DWARFUnit> &cu : dwarf.compile_units()) {
2731 if (Error e = cu->tryExtractDIEsIfNeeded(false)) {
2732 warn(toString(sec) + ": " + toString(std::move(e)));
2733 return {};
2734 }
2735 Expected<DWARFAddressRangesVector> ranges = cu->collectAddressRanges();
2736 if (!ranges) {
2737 warn(toString(sec) + ": " + toString(ranges.takeError()));
2738 return {};
2739 }
2740
2741 ArrayRef<InputSectionBase *> sections = sec->file->getSections();
2742 for (DWARFAddressRange &r : *ranges) {
2743 if (r.SectionIndex == -1ULL)
2744 continue;
2745 // Range list with zero size has no effect.
2746 InputSectionBase *s = sections[r.SectionIndex];
2747 if (s && s != &InputSection::discarded && s->isLive())
2748 if (r.LowPC != r.HighPC)
2749 ret.push_back({cast<InputSection>(s), r.LowPC, r.HighPC, cuIdx});
2750 }
2751 ++cuIdx;
2752 }
2753
2754 return ret;
2755 }
2756
2757 template <class ELFT>
2758 static SmallVector<GdbIndexSection::NameAttrEntry, 0>
readPubNamesAndTypes(const LLDDwarfObj<ELFT> & obj,const SmallVectorImpl<GdbIndexSection::CuEntry> & cus)2759 readPubNamesAndTypes(const LLDDwarfObj<ELFT> &obj,
2760 const SmallVectorImpl<GdbIndexSection::CuEntry> &cus) {
2761 const LLDDWARFSection &pubNames = obj.getGnuPubnamesSection();
2762 const LLDDWARFSection &pubTypes = obj.getGnuPubtypesSection();
2763
2764 SmallVector<GdbIndexSection::NameAttrEntry, 0> ret;
2765 for (const LLDDWARFSection *pub : {&pubNames, &pubTypes}) {
2766 DWARFDataExtractor data(obj, *pub, config->isLE, config->wordsize);
2767 DWARFDebugPubTable table;
2768 table.extract(data, /*GnuStyle=*/true, [&](Error e) {
2769 warn(toString(pub->sec) + ": " + toString(std::move(e)));
2770 });
2771 for (const DWARFDebugPubTable::Set &set : table.getData()) {
2772 // The value written into the constant pool is kind << 24 | cuIndex. As we
2773 // don't know how many compilation units precede this object to compute
2774 // cuIndex, we compute (kind << 24 | cuIndexInThisObject) instead, and add
2775 // the number of preceding compilation units later.
2776 uint32_t i = llvm::partition_point(cus,
2777 [&](GdbIndexSection::CuEntry cu) {
2778 return cu.cuOffset < set.Offset;
2779 }) -
2780 cus.begin();
2781 for (const DWARFDebugPubTable::Entry &ent : set.Entries)
2782 ret.push_back({{ent.Name, computeGdbHash(ent.Name)},
2783 (ent.Descriptor.toBits() << 24) | i});
2784 }
2785 }
2786 return ret;
2787 }
2788
2789 // Create a list of symbols from a given list of symbol names and types
2790 // by uniquifying them by name.
2791 static std::pair<SmallVector<GdbIndexSection::GdbSymbol, 0>, size_t>
createSymbols(ArrayRef<SmallVector<GdbIndexSection::NameAttrEntry,0>> nameAttrs,const SmallVector<GdbIndexSection::GdbChunk,0> & chunks)2792 createSymbols(
2793 ArrayRef<SmallVector<GdbIndexSection::NameAttrEntry, 0>> nameAttrs,
2794 const SmallVector<GdbIndexSection::GdbChunk, 0> &chunks) {
2795 using GdbSymbol = GdbIndexSection::GdbSymbol;
2796 using NameAttrEntry = GdbIndexSection::NameAttrEntry;
2797
2798 // For each chunk, compute the number of compilation units preceding it.
2799 uint32_t cuIdx = 0;
2800 std::unique_ptr<uint32_t[]> cuIdxs(new uint32_t[chunks.size()]);
2801 for (uint32_t i = 0, e = chunks.size(); i != e; ++i) {
2802 cuIdxs[i] = cuIdx;
2803 cuIdx += chunks[i].compilationUnits.size();
2804 }
2805
2806 // The number of symbols we will handle in this function is of the order
2807 // of millions for very large executables, so we use multi-threading to
2808 // speed it up.
2809 constexpr size_t numShards = 32;
2810 const size_t concurrency =
2811 llvm::bit_floor(std::min<size_t>(config->threadCount, numShards));
2812
2813 // A sharded map to uniquify symbols by name.
2814 auto map =
2815 std::make_unique<DenseMap<CachedHashStringRef, size_t>[]>(numShards);
2816 size_t shift = 32 - llvm::countr_zero(numShards);
2817
2818 // Instantiate GdbSymbols while uniqufying them by name.
2819 auto symbols = std::make_unique<SmallVector<GdbSymbol, 0>[]>(numShards);
2820
2821 parallelFor(0, concurrency, [&](size_t threadId) {
2822 uint32_t i = 0;
2823 for (ArrayRef<NameAttrEntry> entries : nameAttrs) {
2824 for (const NameAttrEntry &ent : entries) {
2825 size_t shardId = ent.name.hash() >> shift;
2826 if ((shardId & (concurrency - 1)) != threadId)
2827 continue;
2828
2829 uint32_t v = ent.cuIndexAndAttrs + cuIdxs[i];
2830 size_t &idx = map[shardId][ent.name];
2831 if (idx) {
2832 symbols[shardId][idx - 1].cuVector.push_back(v);
2833 continue;
2834 }
2835
2836 idx = symbols[shardId].size() + 1;
2837 symbols[shardId].push_back({ent.name, {v}, 0, 0});
2838 }
2839 ++i;
2840 }
2841 });
2842
2843 size_t numSymbols = 0;
2844 for (ArrayRef<GdbSymbol> v : ArrayRef(symbols.get(), numShards))
2845 numSymbols += v.size();
2846
2847 // The return type is a flattened vector, so we'll copy each vector
2848 // contents to Ret.
2849 SmallVector<GdbSymbol, 0> ret;
2850 ret.reserve(numSymbols);
2851 for (SmallVector<GdbSymbol, 0> &vec :
2852 MutableArrayRef(symbols.get(), numShards))
2853 for (GdbSymbol &sym : vec)
2854 ret.push_back(std::move(sym));
2855
2856 // CU vectors and symbol names are adjacent in the output file.
2857 // We can compute their offsets in the output file now.
2858 size_t off = 0;
2859 for (GdbSymbol &sym : ret) {
2860 sym.cuVectorOff = off;
2861 off += (sym.cuVector.size() + 1) * 4;
2862 }
2863 for (GdbSymbol &sym : ret) {
2864 sym.nameOff = off;
2865 off += sym.name.size() + 1;
2866 }
2867 // If off overflows, the last symbol's nameOff likely overflows.
2868 if (!isUInt<32>(off))
2869 errorOrWarn("--gdb-index: constant pool size (" + Twine(off) +
2870 ") exceeds UINT32_MAX");
2871
2872 return {ret, off};
2873 }
2874
2875 // Returns a newly-created .gdb_index section.
create()2876 template <class ELFT> GdbIndexSection *GdbIndexSection::create() {
2877 llvm::TimeTraceScope timeScope("Create gdb index");
2878
2879 // Collect InputFiles with .debug_info. See the comment in
2880 // LLDDwarfObj<ELFT>::LLDDwarfObj. If we do lightweight parsing in the future,
2881 // note that isec->data() may uncompress the full content, which should be
2882 // parallelized.
2883 SetVector<InputFile *> files;
2884 for (InputSectionBase *s : ctx.inputSections) {
2885 InputSection *isec = dyn_cast<InputSection>(s);
2886 if (!isec)
2887 continue;
2888 // .debug_gnu_pub{names,types} are useless in executables.
2889 // They are present in input object files solely for creating
2890 // a .gdb_index. So we can remove them from the output.
2891 if (s->name == ".debug_gnu_pubnames" || s->name == ".debug_gnu_pubtypes")
2892 s->markDead();
2893 else if (isec->name == ".debug_info")
2894 files.insert(isec->file);
2895 }
2896 // Drop .rel[a].debug_gnu_pub{names,types} for --emit-relocs.
2897 llvm::erase_if(ctx.inputSections, [](InputSectionBase *s) {
2898 if (auto *isec = dyn_cast<InputSection>(s))
2899 if (InputSectionBase *rel = isec->getRelocatedSection())
2900 return !rel->isLive();
2901 return !s->isLive();
2902 });
2903
2904 SmallVector<GdbChunk, 0> chunks(files.size());
2905 SmallVector<SmallVector<NameAttrEntry, 0>, 0> nameAttrs(files.size());
2906
2907 parallelFor(0, files.size(), [&](size_t i) {
2908 // To keep memory usage low, we don't want to keep cached DWARFContext, so
2909 // avoid getDwarf() here.
2910 ObjFile<ELFT> *file = cast<ObjFile<ELFT>>(files[i]);
2911 DWARFContext dwarf(std::make_unique<LLDDwarfObj<ELFT>>(file));
2912 auto &dobj = static_cast<const LLDDwarfObj<ELFT> &>(dwarf.getDWARFObj());
2913
2914 // If the are multiple compile units .debug_info (very rare ld -r --unique),
2915 // this only picks the last one. Other address ranges are lost.
2916 chunks[i].sec = dobj.getInfoSection();
2917 chunks[i].compilationUnits = readCuList(dwarf);
2918 chunks[i].addressAreas = readAddressAreas(dwarf, chunks[i].sec);
2919 nameAttrs[i] = readPubNamesAndTypes<ELFT>(dobj, chunks[i].compilationUnits);
2920 });
2921
2922 auto *ret = make<GdbIndexSection>();
2923 ret->chunks = std::move(chunks);
2924 std::tie(ret->symbols, ret->size) = createSymbols(nameAttrs, ret->chunks);
2925
2926 // Count the areas other than the constant pool.
2927 ret->size += sizeof(GdbIndexHeader) + ret->computeSymtabSize() * 8;
2928 for (GdbChunk &chunk : ret->chunks)
2929 ret->size +=
2930 chunk.compilationUnits.size() * 16 + chunk.addressAreas.size() * 20;
2931
2932 return ret;
2933 }
2934
writeTo(uint8_t * buf)2935 void GdbIndexSection::writeTo(uint8_t *buf) {
2936 // Write the header.
2937 auto *hdr = reinterpret_cast<GdbIndexHeader *>(buf);
2938 uint8_t *start = buf;
2939 hdr->version = 7;
2940 buf += sizeof(*hdr);
2941
2942 // Write the CU list.
2943 hdr->cuListOff = buf - start;
2944 for (GdbChunk &chunk : chunks) {
2945 for (CuEntry &cu : chunk.compilationUnits) {
2946 write64le(buf, chunk.sec->outSecOff + cu.cuOffset);
2947 write64le(buf + 8, cu.cuLength);
2948 buf += 16;
2949 }
2950 }
2951
2952 // Write the address area.
2953 hdr->cuTypesOff = buf - start;
2954 hdr->addressAreaOff = buf - start;
2955 uint32_t cuOff = 0;
2956 for (GdbChunk &chunk : chunks) {
2957 for (AddressEntry &e : chunk.addressAreas) {
2958 // In the case of ICF there may be duplicate address range entries.
2959 const uint64_t baseAddr = e.section->repl->getVA(0);
2960 write64le(buf, baseAddr + e.lowAddress);
2961 write64le(buf + 8, baseAddr + e.highAddress);
2962 write32le(buf + 16, e.cuIndex + cuOff);
2963 buf += 20;
2964 }
2965 cuOff += chunk.compilationUnits.size();
2966 }
2967
2968 // Write the on-disk open-addressing hash table containing symbols.
2969 hdr->symtabOff = buf - start;
2970 size_t symtabSize = computeSymtabSize();
2971 uint32_t mask = symtabSize - 1;
2972
2973 for (GdbSymbol &sym : symbols) {
2974 uint32_t h = sym.name.hash();
2975 uint32_t i = h & mask;
2976 uint32_t step = ((h * 17) & mask) | 1;
2977
2978 while (read32le(buf + i * 8))
2979 i = (i + step) & mask;
2980
2981 write32le(buf + i * 8, sym.nameOff);
2982 write32le(buf + i * 8 + 4, sym.cuVectorOff);
2983 }
2984
2985 buf += symtabSize * 8;
2986
2987 // Write the string pool.
2988 hdr->constantPoolOff = buf - start;
2989 parallelForEach(symbols, [&](GdbSymbol &sym) {
2990 memcpy(buf + sym.nameOff, sym.name.data(), sym.name.size());
2991 });
2992
2993 // Write the CU vectors.
2994 for (GdbSymbol &sym : symbols) {
2995 write32le(buf, sym.cuVector.size());
2996 buf += 4;
2997 for (uint32_t val : sym.cuVector) {
2998 write32le(buf, val);
2999 buf += 4;
3000 }
3001 }
3002 }
3003
isNeeded() const3004 bool GdbIndexSection::isNeeded() const { return !chunks.empty(); }
3005
EhFrameHeader()3006 EhFrameHeader::EhFrameHeader()
3007 : SyntheticSection(SHF_ALLOC, SHT_PROGBITS, 4, ".eh_frame_hdr") {}
3008
writeTo(uint8_t * buf)3009 void EhFrameHeader::writeTo(uint8_t *buf) {
3010 // Unlike most sections, the EhFrameHeader section is written while writing
3011 // another section, namely EhFrameSection, which calls the write() function
3012 // below from its writeTo() function. This is necessary because the contents
3013 // of EhFrameHeader depend on the relocated contents of EhFrameSection and we
3014 // don't know which order the sections will be written in.
3015 }
3016
3017 // .eh_frame_hdr contains a binary search table of pointers to FDEs.
3018 // Each entry of the search table consists of two values,
3019 // the starting PC from where FDEs covers, and the FDE's address.
3020 // It is sorted by PC.
write()3021 void EhFrameHeader::write() {
3022 uint8_t *buf = Out::bufferStart + getParent()->offset + outSecOff;
3023 using FdeData = EhFrameSection::FdeData;
3024 SmallVector<FdeData, 0> fdes = getPartition().ehFrame->getFdeData();
3025
3026 buf[0] = 1;
3027 buf[1] = DW_EH_PE_pcrel | DW_EH_PE_sdata4;
3028 buf[2] = DW_EH_PE_udata4;
3029 buf[3] = DW_EH_PE_datarel | DW_EH_PE_sdata4;
3030 write32(buf + 4,
3031 getPartition().ehFrame->getParent()->addr - this->getVA() - 4);
3032 write32(buf + 8, fdes.size());
3033 buf += 12;
3034
3035 for (FdeData &fde : fdes) {
3036 write32(buf, fde.pcRel);
3037 write32(buf + 4, fde.fdeVARel);
3038 buf += 8;
3039 }
3040 }
3041
getSize() const3042 size_t EhFrameHeader::getSize() const {
3043 // .eh_frame_hdr has a 12 bytes header followed by an array of FDEs.
3044 return 12 + getPartition().ehFrame->numFdes * 8;
3045 }
3046
isNeeded() const3047 bool EhFrameHeader::isNeeded() const {
3048 return isLive() && getPartition().ehFrame->isNeeded();
3049 }
3050
VersionDefinitionSection()3051 VersionDefinitionSection::VersionDefinitionSection()
3052 : SyntheticSection(SHF_ALLOC, SHT_GNU_verdef, sizeof(uint32_t),
3053 ".gnu.version_d") {}
3054
getFileDefName()3055 StringRef VersionDefinitionSection::getFileDefName() {
3056 if (!getPartition().name.empty())
3057 return getPartition().name;
3058 if (!config->soName.empty())
3059 return config->soName;
3060 return config->outputFile;
3061 }
3062
finalizeContents()3063 void VersionDefinitionSection::finalizeContents() {
3064 fileDefNameOff = getPartition().dynStrTab->addString(getFileDefName());
3065 for (const VersionDefinition &v : namedVersionDefs())
3066 verDefNameOffs.push_back(getPartition().dynStrTab->addString(v.name));
3067
3068 if (OutputSection *sec = getPartition().dynStrTab->getParent())
3069 getParent()->link = sec->sectionIndex;
3070
3071 // sh_info should be set to the number of definitions. This fact is missed in
3072 // documentation, but confirmed by binutils community:
3073 // https://sourceware.org/ml/binutils/2014-11/msg00355.html
3074 getParent()->info = getVerDefNum();
3075 }
3076
writeOne(uint8_t * buf,uint32_t index,StringRef name,size_t nameOff)3077 void VersionDefinitionSection::writeOne(uint8_t *buf, uint32_t index,
3078 StringRef name, size_t nameOff) {
3079 uint16_t flags = index == 1 ? VER_FLG_BASE : 0;
3080
3081 // Write a verdef.
3082 write16(buf, 1); // vd_version
3083 write16(buf + 2, flags); // vd_flags
3084 write16(buf + 4, index); // vd_ndx
3085 write16(buf + 6, 1); // vd_cnt
3086 write32(buf + 8, hashSysV(name)); // vd_hash
3087 write32(buf + 12, 20); // vd_aux
3088 write32(buf + 16, 28); // vd_next
3089
3090 // Write a veraux.
3091 write32(buf + 20, nameOff); // vda_name
3092 write32(buf + 24, 0); // vda_next
3093 }
3094
writeTo(uint8_t * buf)3095 void VersionDefinitionSection::writeTo(uint8_t *buf) {
3096 writeOne(buf, 1, getFileDefName(), fileDefNameOff);
3097
3098 auto nameOffIt = verDefNameOffs.begin();
3099 for (const VersionDefinition &v : namedVersionDefs()) {
3100 buf += EntrySize;
3101 writeOne(buf, v.id, v.name, *nameOffIt++);
3102 }
3103
3104 // Need to terminate the last version definition.
3105 write32(buf + 16, 0); // vd_next
3106 }
3107
getSize() const3108 size_t VersionDefinitionSection::getSize() const {
3109 return EntrySize * getVerDefNum();
3110 }
3111
3112 // .gnu.version is a table where each entry is 2 byte long.
VersionTableSection()3113 VersionTableSection::VersionTableSection()
3114 : SyntheticSection(SHF_ALLOC, SHT_GNU_versym, sizeof(uint16_t),
3115 ".gnu.version") {
3116 this->entsize = 2;
3117 }
3118
finalizeContents()3119 void VersionTableSection::finalizeContents() {
3120 // At the moment of june 2016 GNU docs does not mention that sh_link field
3121 // should be set, but Sun docs do. Also readelf relies on this field.
3122 getParent()->link = getPartition().dynSymTab->getParent()->sectionIndex;
3123 }
3124
getSize() const3125 size_t VersionTableSection::getSize() const {
3126 return (getPartition().dynSymTab->getSymbols().size() + 1) * 2;
3127 }
3128
writeTo(uint8_t * buf)3129 void VersionTableSection::writeTo(uint8_t *buf) {
3130 buf += 2;
3131 for (const SymbolTableEntry &s : getPartition().dynSymTab->getSymbols()) {
3132 // For an unextracted lazy symbol (undefined weak), it must have been
3133 // converted to Undefined and have VER_NDX_GLOBAL version here.
3134 assert(!s.sym->isLazy());
3135 write16(buf, s.sym->versionId);
3136 buf += 2;
3137 }
3138 }
3139
isNeeded() const3140 bool VersionTableSection::isNeeded() const {
3141 return isLive() &&
3142 (getPartition().verDef || getPartition().verNeed->isNeeded());
3143 }
3144
addVerneed(Symbol * ss)3145 void elf::addVerneed(Symbol *ss) {
3146 auto &file = cast<SharedFile>(*ss->file);
3147 if (ss->versionId == VER_NDX_GLOBAL)
3148 return;
3149
3150 if (file.vernauxs.empty())
3151 file.vernauxs.resize(file.verdefs.size());
3152
3153 // Select a version identifier for the vernaux data structure, if we haven't
3154 // already allocated one. The verdef identifiers cover the range
3155 // [1..getVerDefNum()]; this causes the vernaux identifiers to start from
3156 // getVerDefNum()+1.
3157 if (file.vernauxs[ss->versionId] == 0)
3158 file.vernauxs[ss->versionId] = ++SharedFile::vernauxNum + getVerDefNum();
3159
3160 ss->versionId = file.vernauxs[ss->versionId];
3161 }
3162
3163 template <class ELFT>
VersionNeedSection()3164 VersionNeedSection<ELFT>::VersionNeedSection()
3165 : SyntheticSection(SHF_ALLOC, SHT_GNU_verneed, sizeof(uint32_t),
3166 ".gnu.version_r") {}
3167
finalizeContents()3168 template <class ELFT> void VersionNeedSection<ELFT>::finalizeContents() {
3169 for (SharedFile *f : ctx.sharedFiles) {
3170 if (f->vernauxs.empty())
3171 continue;
3172 verneeds.emplace_back();
3173 Verneed &vn = verneeds.back();
3174 vn.nameStrTab = getPartition().dynStrTab->addString(f->soName);
3175 bool isLibc = config->relrGlibc && f->soName.starts_with("libc.so.");
3176 bool isGlibc2 = false;
3177 for (unsigned i = 0; i != f->vernauxs.size(); ++i) {
3178 if (f->vernauxs[i] == 0)
3179 continue;
3180 auto *verdef =
3181 reinterpret_cast<const typename ELFT::Verdef *>(f->verdefs[i]);
3182 StringRef ver(f->getStringTable().data() + verdef->getAux()->vda_name);
3183 if (isLibc && ver.starts_with("GLIBC_2."))
3184 isGlibc2 = true;
3185 vn.vernauxs.push_back({verdef->vd_hash, f->vernauxs[i],
3186 getPartition().dynStrTab->addString(ver)});
3187 }
3188 if (isGlibc2) {
3189 const char *ver = "GLIBC_ABI_DT_RELR";
3190 vn.vernauxs.push_back({hashSysV(ver),
3191 ++SharedFile::vernauxNum + getVerDefNum(),
3192 getPartition().dynStrTab->addString(ver)});
3193 }
3194 }
3195
3196 if (OutputSection *sec = getPartition().dynStrTab->getParent())
3197 getParent()->link = sec->sectionIndex;
3198 getParent()->info = verneeds.size();
3199 }
3200
writeTo(uint8_t * buf)3201 template <class ELFT> void VersionNeedSection<ELFT>::writeTo(uint8_t *buf) {
3202 // The Elf_Verneeds need to appear first, followed by the Elf_Vernauxs.
3203 auto *verneed = reinterpret_cast<Elf_Verneed *>(buf);
3204 auto *vernaux = reinterpret_cast<Elf_Vernaux *>(verneed + verneeds.size());
3205
3206 for (auto &vn : verneeds) {
3207 // Create an Elf_Verneed for this DSO.
3208 verneed->vn_version = 1;
3209 verneed->vn_cnt = vn.vernauxs.size();
3210 verneed->vn_file = vn.nameStrTab;
3211 verneed->vn_aux =
3212 reinterpret_cast<char *>(vernaux) - reinterpret_cast<char *>(verneed);
3213 verneed->vn_next = sizeof(Elf_Verneed);
3214 ++verneed;
3215
3216 // Create the Elf_Vernauxs for this Elf_Verneed.
3217 for (auto &vna : vn.vernauxs) {
3218 vernaux->vna_hash = vna.hash;
3219 vernaux->vna_flags = 0;
3220 vernaux->vna_other = vna.verneedIndex;
3221 vernaux->vna_name = vna.nameStrTab;
3222 vernaux->vna_next = sizeof(Elf_Vernaux);
3223 ++vernaux;
3224 }
3225
3226 vernaux[-1].vna_next = 0;
3227 }
3228 verneed[-1].vn_next = 0;
3229 }
3230
getSize() const3231 template <class ELFT> size_t VersionNeedSection<ELFT>::getSize() const {
3232 return verneeds.size() * sizeof(Elf_Verneed) +
3233 SharedFile::vernauxNum * sizeof(Elf_Vernaux);
3234 }
3235
isNeeded() const3236 template <class ELFT> bool VersionNeedSection<ELFT>::isNeeded() const {
3237 return isLive() && SharedFile::vernauxNum != 0;
3238 }
3239
addSection(MergeInputSection * ms)3240 void MergeSyntheticSection::addSection(MergeInputSection *ms) {
3241 ms->parent = this;
3242 sections.push_back(ms);
3243 assert(addralign == ms->addralign || !(ms->flags & SHF_STRINGS));
3244 addralign = std::max(addralign, ms->addralign);
3245 }
3246
MergeTailSection(StringRef name,uint32_t type,uint64_t flags,uint32_t alignment)3247 MergeTailSection::MergeTailSection(StringRef name, uint32_t type,
3248 uint64_t flags, uint32_t alignment)
3249 : MergeSyntheticSection(name, type, flags, alignment),
3250 builder(StringTableBuilder::RAW, llvm::Align(alignment)) {}
3251
getSize() const3252 size_t MergeTailSection::getSize() const { return builder.getSize(); }
3253
writeTo(uint8_t * buf)3254 void MergeTailSection::writeTo(uint8_t *buf) { builder.write(buf); }
3255
finalizeContents()3256 void MergeTailSection::finalizeContents() {
3257 // Add all string pieces to the string table builder to create section
3258 // contents.
3259 for (MergeInputSection *sec : sections)
3260 for (size_t i = 0, e = sec->pieces.size(); i != e; ++i)
3261 if (sec->pieces[i].live)
3262 builder.add(sec->getData(i));
3263
3264 // Fix the string table content. After this, the contents will never change.
3265 builder.finalize();
3266
3267 // finalize() fixed tail-optimized strings, so we can now get
3268 // offsets of strings. Get an offset for each string and save it
3269 // to a corresponding SectionPiece for easy access.
3270 for (MergeInputSection *sec : sections)
3271 for (size_t i = 0, e = sec->pieces.size(); i != e; ++i)
3272 if (sec->pieces[i].live)
3273 sec->pieces[i].outputOff = builder.getOffset(sec->getData(i));
3274 }
3275
writeTo(uint8_t * buf)3276 void MergeNoTailSection::writeTo(uint8_t *buf) {
3277 parallelFor(0, numShards,
3278 [&](size_t i) { shards[i].write(buf + shardOffsets[i]); });
3279 }
3280
3281 // This function is very hot (i.e. it can take several seconds to finish)
3282 // because sometimes the number of inputs is in an order of magnitude of
3283 // millions. So, we use multi-threading.
3284 //
3285 // For any strings S and T, we know S is not mergeable with T if S's hash
3286 // value is different from T's. If that's the case, we can safely put S and
3287 // T into different string builders without worrying about merge misses.
3288 // We do it in parallel.
finalizeContents()3289 void MergeNoTailSection::finalizeContents() {
3290 // Initializes string table builders.
3291 for (size_t i = 0; i < numShards; ++i)
3292 shards.emplace_back(StringTableBuilder::RAW, llvm::Align(addralign));
3293
3294 // Concurrency level. Must be a power of 2 to avoid expensive modulo
3295 // operations in the following tight loop.
3296 const size_t concurrency =
3297 llvm::bit_floor(std::min<size_t>(config->threadCount, numShards));
3298
3299 // Add section pieces to the builders.
3300 parallelFor(0, concurrency, [&](size_t threadId) {
3301 for (MergeInputSection *sec : sections) {
3302 for (size_t i = 0, e = sec->pieces.size(); i != e; ++i) {
3303 if (!sec->pieces[i].live)
3304 continue;
3305 size_t shardId = getShardId(sec->pieces[i].hash);
3306 if ((shardId & (concurrency - 1)) == threadId)
3307 sec->pieces[i].outputOff = shards[shardId].add(sec->getData(i));
3308 }
3309 }
3310 });
3311
3312 // Compute an in-section offset for each shard.
3313 size_t off = 0;
3314 for (size_t i = 0; i < numShards; ++i) {
3315 shards[i].finalizeInOrder();
3316 if (shards[i].getSize() > 0)
3317 off = alignToPowerOf2(off, addralign);
3318 shardOffsets[i] = off;
3319 off += shards[i].getSize();
3320 }
3321 size = off;
3322
3323 // So far, section pieces have offsets from beginning of shards, but
3324 // we want offsets from beginning of the whole section. Fix them.
3325 parallelForEach(sections, [&](MergeInputSection *sec) {
3326 for (size_t i = 0, e = sec->pieces.size(); i != e; ++i)
3327 if (sec->pieces[i].live)
3328 sec->pieces[i].outputOff +=
3329 shardOffsets[getShardId(sec->pieces[i].hash)];
3330 });
3331 }
3332
splitSections()3333 template <class ELFT> void elf::splitSections() {
3334 llvm::TimeTraceScope timeScope("Split sections");
3335 // splitIntoPieces needs to be called on each MergeInputSection
3336 // before calling finalizeContents().
3337 parallelForEach(ctx.objectFiles, [](ELFFileBase *file) {
3338 for (InputSectionBase *sec : file->getSections()) {
3339 if (!sec)
3340 continue;
3341 if (auto *s = dyn_cast<MergeInputSection>(sec))
3342 s->splitIntoPieces();
3343 else if (auto *eh = dyn_cast<EhInputSection>(sec))
3344 eh->split<ELFT>();
3345 }
3346 });
3347 }
3348
combineEhSections()3349 void elf::combineEhSections() {
3350 llvm::TimeTraceScope timeScope("Combine EH sections");
3351 for (EhInputSection *sec : ctx.ehInputSections) {
3352 EhFrameSection &eh = *sec->getPartition().ehFrame;
3353 sec->parent = &eh;
3354 eh.addralign = std::max(eh.addralign, sec->addralign);
3355 eh.sections.push_back(sec);
3356 llvm::append_range(eh.dependentSections, sec->dependentSections);
3357 }
3358
3359 if (!mainPart->armExidx)
3360 return;
3361 llvm::erase_if(ctx.inputSections, [](InputSectionBase *s) {
3362 // Ignore dead sections and the partition end marker (.part.end),
3363 // whose partition number is out of bounds.
3364 if (!s->isLive() || s->partition == 255)
3365 return false;
3366 Partition &part = s->getPartition();
3367 return s->kind() == SectionBase::Regular && part.armExidx &&
3368 part.armExidx->addSection(cast<InputSection>(s));
3369 });
3370 }
3371
MipsRldMapSection()3372 MipsRldMapSection::MipsRldMapSection()
3373 : SyntheticSection(SHF_ALLOC | SHF_WRITE, SHT_PROGBITS, config->wordsize,
3374 ".rld_map") {}
3375
ARMExidxSyntheticSection()3376 ARMExidxSyntheticSection::ARMExidxSyntheticSection()
3377 : SyntheticSection(SHF_ALLOC | SHF_LINK_ORDER, SHT_ARM_EXIDX,
3378 config->wordsize, ".ARM.exidx") {}
3379
findExidxSection(InputSection * isec)3380 static InputSection *findExidxSection(InputSection *isec) {
3381 for (InputSection *d : isec->dependentSections)
3382 if (d->type == SHT_ARM_EXIDX && d->isLive())
3383 return d;
3384 return nullptr;
3385 }
3386
isValidExidxSectionDep(InputSection * isec)3387 static bool isValidExidxSectionDep(InputSection *isec) {
3388 return (isec->flags & SHF_ALLOC) && (isec->flags & SHF_EXECINSTR) &&
3389 isec->getSize() > 0;
3390 }
3391
addSection(InputSection * isec)3392 bool ARMExidxSyntheticSection::addSection(InputSection *isec) {
3393 if (isec->type == SHT_ARM_EXIDX) {
3394 if (InputSection *dep = isec->getLinkOrderDep())
3395 if (isValidExidxSectionDep(dep)) {
3396 exidxSections.push_back(isec);
3397 // Every exidxSection is 8 bytes, we need an estimate of
3398 // size before assignAddresses can be called. Final size
3399 // will only be known after finalize is called.
3400 size += 8;
3401 }
3402 return true;
3403 }
3404
3405 if (isValidExidxSectionDep(isec)) {
3406 executableSections.push_back(isec);
3407 return false;
3408 }
3409
3410 // FIXME: we do not output a relocation section when --emit-relocs is used
3411 // as we do not have relocation sections for linker generated table entries
3412 // and we would have to erase at a late stage relocations from merged entries.
3413 // Given that exception tables are already position independent and a binary
3414 // analyzer could derive the relocations we choose to erase the relocations.
3415 if (config->emitRelocs && isec->type == SHT_REL)
3416 if (InputSectionBase *ex = isec->getRelocatedSection())
3417 if (isa<InputSection>(ex) && ex->type == SHT_ARM_EXIDX)
3418 return true;
3419
3420 return false;
3421 }
3422
3423 // References to .ARM.Extab Sections have bit 31 clear and are not the
3424 // special EXIDX_CANTUNWIND bit-pattern.
isExtabRef(uint32_t unwind)3425 static bool isExtabRef(uint32_t unwind) {
3426 return (unwind & 0x80000000) == 0 && unwind != 0x1;
3427 }
3428
3429 // Return true if the .ARM.exidx section Cur can be merged into the .ARM.exidx
3430 // section Prev, where Cur follows Prev in the table. This can be done if the
3431 // unwinding instructions in Cur are identical to Prev. Linker generated
3432 // EXIDX_CANTUNWIND entries are represented by nullptr as they do not have an
3433 // InputSection.
isDuplicateArmExidxSec(InputSection * prev,InputSection * cur)3434 static bool isDuplicateArmExidxSec(InputSection *prev, InputSection *cur) {
3435 // Get the last table Entry from the previous .ARM.exidx section. If Prev is
3436 // nullptr then it will be a synthesized EXIDX_CANTUNWIND entry.
3437 uint32_t prevUnwind = 1;
3438 if (prev)
3439 prevUnwind = read32(prev->content().data() + prev->content().size() - 4);
3440 if (isExtabRef(prevUnwind))
3441 return false;
3442
3443 // We consider the unwind instructions of an .ARM.exidx table entry
3444 // a duplicate if the previous unwind instructions if:
3445 // - Both are the special EXIDX_CANTUNWIND.
3446 // - Both are the same inline unwind instructions.
3447 // We do not attempt to follow and check links into .ARM.extab tables as
3448 // consecutive identical entries are rare and the effort to check that they
3449 // are identical is high.
3450
3451 // If Cur is nullptr then this is synthesized EXIDX_CANTUNWIND entry.
3452 if (cur == nullptr)
3453 return prevUnwind == 1;
3454
3455 for (uint32_t offset = 4; offset < (uint32_t)cur->content().size(); offset +=8) {
3456 uint32_t curUnwind = read32(cur->content().data() + offset);
3457 if (isExtabRef(curUnwind) || curUnwind != prevUnwind)
3458 return false;
3459 }
3460 // All table entries in this .ARM.exidx Section can be merged into the
3461 // previous Section.
3462 return true;
3463 }
3464
3465 // The .ARM.exidx table must be sorted in ascending order of the address of the
3466 // functions the table describes. std::optionally duplicate adjacent table
3467 // entries can be removed. At the end of the function the executableSections
3468 // must be sorted in ascending order of address, Sentinel is set to the
3469 // InputSection with the highest address and any InputSections that have
3470 // mergeable .ARM.exidx table entries are removed from it.
finalizeContents()3471 void ARMExidxSyntheticSection::finalizeContents() {
3472 // The executableSections and exidxSections that we use to derive the final
3473 // contents of this SyntheticSection are populated before
3474 // processSectionCommands() and ICF. A /DISCARD/ entry in SECTIONS command or
3475 // ICF may remove executable InputSections and their dependent .ARM.exidx
3476 // section that we recorded earlier.
3477 auto isDiscarded = [](const InputSection *isec) { return !isec->isLive(); };
3478 llvm::erase_if(exidxSections, isDiscarded);
3479 // We need to remove discarded InputSections and InputSections without
3480 // .ARM.exidx sections that if we generated the .ARM.exidx it would be out
3481 // of range.
3482 auto isDiscardedOrOutOfRange = [this](InputSection *isec) {
3483 if (!isec->isLive())
3484 return true;
3485 if (findExidxSection(isec))
3486 return false;
3487 int64_t off = static_cast<int64_t>(isec->getVA() - getVA());
3488 return off != llvm::SignExtend64(off, 31);
3489 };
3490 llvm::erase_if(executableSections, isDiscardedOrOutOfRange);
3491
3492 // Sort the executable sections that may or may not have associated
3493 // .ARM.exidx sections by order of ascending address. This requires the
3494 // relative positions of InputSections and OutputSections to be known.
3495 auto compareByFilePosition = [](const InputSection *a,
3496 const InputSection *b) {
3497 OutputSection *aOut = a->getParent();
3498 OutputSection *bOut = b->getParent();
3499
3500 if (aOut != bOut)
3501 return aOut->addr < bOut->addr;
3502 return a->outSecOff < b->outSecOff;
3503 };
3504 llvm::stable_sort(executableSections, compareByFilePosition);
3505 sentinel = executableSections.back();
3506 // std::optionally merge adjacent duplicate entries.
3507 if (config->mergeArmExidx) {
3508 SmallVector<InputSection *, 0> selectedSections;
3509 selectedSections.reserve(executableSections.size());
3510 selectedSections.push_back(executableSections[0]);
3511 size_t prev = 0;
3512 for (size_t i = 1; i < executableSections.size(); ++i) {
3513 InputSection *ex1 = findExidxSection(executableSections[prev]);
3514 InputSection *ex2 = findExidxSection(executableSections[i]);
3515 if (!isDuplicateArmExidxSec(ex1, ex2)) {
3516 selectedSections.push_back(executableSections[i]);
3517 prev = i;
3518 }
3519 }
3520 executableSections = std::move(selectedSections);
3521 }
3522 // offset is within the SyntheticSection.
3523 size_t offset = 0;
3524 size = 0;
3525 for (InputSection *isec : executableSections) {
3526 if (InputSection *d = findExidxSection(isec)) {
3527 d->outSecOff = offset;
3528 d->parent = getParent();
3529 offset += d->getSize();
3530 } else {
3531 offset += 8;
3532 }
3533 }
3534 // Size includes Sentinel.
3535 size = offset + 8;
3536 }
3537
getLinkOrderDep() const3538 InputSection *ARMExidxSyntheticSection::getLinkOrderDep() const {
3539 return executableSections.front();
3540 }
3541
3542 // To write the .ARM.exidx table from the ExecutableSections we have three cases
3543 // 1.) The InputSection has a .ARM.exidx InputSection in its dependent sections.
3544 // We write the .ARM.exidx section contents and apply its relocations.
3545 // 2.) The InputSection does not have a dependent .ARM.exidx InputSection. We
3546 // must write the contents of an EXIDX_CANTUNWIND directly. We use the
3547 // start of the InputSection as the purpose of the linker generated
3548 // section is to terminate the address range of the previous entry.
3549 // 3.) A trailing EXIDX_CANTUNWIND sentinel section is required at the end of
3550 // the table to terminate the address range of the final entry.
writeTo(uint8_t * buf)3551 void ARMExidxSyntheticSection::writeTo(uint8_t *buf) {
3552
3553 // A linker generated CANTUNWIND entry is made up of two words:
3554 // 0x0 with R_ARM_PREL31 relocation to target.
3555 // 0x1 with EXIDX_CANTUNWIND.
3556 uint64_t offset = 0;
3557 for (InputSection *isec : executableSections) {
3558 assert(isec->getParent() != nullptr);
3559 if (InputSection *d = findExidxSection(isec)) {
3560 for (int dataOffset = 0; dataOffset != (int)d->content().size();
3561 dataOffset += 4)
3562 write32(buf + offset + dataOffset,
3563 read32(d->content().data() + dataOffset));
3564 // Recalculate outSecOff as finalizeAddressDependentContent()
3565 // may have altered syntheticSection outSecOff.
3566 d->outSecOff = offset + outSecOff;
3567 target->relocateAlloc(*d, buf + offset);
3568 offset += d->getSize();
3569 } else {
3570 // A Linker generated CANTUNWIND section.
3571 write32(buf + offset + 0, 0x0);
3572 write32(buf + offset + 4, 0x1);
3573 uint64_t s = isec->getVA();
3574 uint64_t p = getVA() + offset;
3575 target->relocateNoSym(buf + offset, R_ARM_PREL31, s - p);
3576 offset += 8;
3577 }
3578 }
3579 // Write Sentinel CANTUNWIND entry.
3580 write32(buf + offset + 0, 0x0);
3581 write32(buf + offset + 4, 0x1);
3582 uint64_t s = sentinel->getVA(sentinel->getSize());
3583 uint64_t p = getVA() + offset;
3584 target->relocateNoSym(buf + offset, R_ARM_PREL31, s - p);
3585 assert(size == offset + 8);
3586 }
3587
isNeeded() const3588 bool ARMExidxSyntheticSection::isNeeded() const {
3589 return llvm::any_of(exidxSections,
3590 [](InputSection *isec) { return isec->isLive(); });
3591 }
3592
ThunkSection(OutputSection * os,uint64_t off)3593 ThunkSection::ThunkSection(OutputSection *os, uint64_t off)
3594 : SyntheticSection(SHF_ALLOC | SHF_EXECINSTR, SHT_PROGBITS,
3595 config->emachine == EM_PPC64 ? 16 : 4, ".text.thunk") {
3596 this->parent = os;
3597 this->outSecOff = off;
3598 }
3599
getSize() const3600 size_t ThunkSection::getSize() const {
3601 if (roundUpSizeForErrata)
3602 return alignTo(size, 4096);
3603 return size;
3604 }
3605
addThunk(Thunk * t)3606 void ThunkSection::addThunk(Thunk *t) {
3607 thunks.push_back(t);
3608 t->addSymbols(*this);
3609 }
3610
writeTo(uint8_t * buf)3611 void ThunkSection::writeTo(uint8_t *buf) {
3612 for (Thunk *t : thunks)
3613 t->writeTo(buf + t->offset);
3614 }
3615
getTargetInputSection() const3616 InputSection *ThunkSection::getTargetInputSection() const {
3617 if (thunks.empty())
3618 return nullptr;
3619 const Thunk *t = thunks.front();
3620 return t->getTargetInputSection();
3621 }
3622
assignOffsets()3623 bool ThunkSection::assignOffsets() {
3624 uint64_t off = 0;
3625 for (Thunk *t : thunks) {
3626 off = alignToPowerOf2(off, t->alignment);
3627 t->setOffset(off);
3628 uint32_t size = t->size();
3629 t->getThunkTargetSym()->size = size;
3630 off += size;
3631 }
3632 bool changed = off != size;
3633 size = off;
3634 return changed;
3635 }
3636
PPC32Got2Section()3637 PPC32Got2Section::PPC32Got2Section()
3638 : SyntheticSection(SHF_ALLOC | SHF_WRITE, SHT_PROGBITS, 4, ".got2") {}
3639
isNeeded() const3640 bool PPC32Got2Section::isNeeded() const {
3641 // See the comment below. This is not needed if there is no other
3642 // InputSection.
3643 for (SectionCommand *cmd : getParent()->commands)
3644 if (auto *isd = dyn_cast<InputSectionDescription>(cmd))
3645 for (InputSection *isec : isd->sections)
3646 if (isec != this)
3647 return true;
3648 return false;
3649 }
3650
finalizeContents()3651 void PPC32Got2Section::finalizeContents() {
3652 // PPC32 may create multiple GOT sections for -fPIC/-fPIE, one per file in
3653 // .got2 . This function computes outSecOff of each .got2 to be used in
3654 // PPC32PltCallStub::writeTo(). The purpose of this empty synthetic section is
3655 // to collect input sections named ".got2".
3656 for (SectionCommand *cmd : getParent()->commands)
3657 if (auto *isd = dyn_cast<InputSectionDescription>(cmd)) {
3658 for (InputSection *isec : isd->sections) {
3659 // isec->file may be nullptr for MergeSyntheticSection.
3660 if (isec != this && isec->file)
3661 isec->file->ppc32Got2 = isec;
3662 }
3663 }
3664 }
3665
3666 // If linking position-dependent code then the table will store the addresses
3667 // directly in the binary so the section has type SHT_PROGBITS. If linking
3668 // position-independent code the section has type SHT_NOBITS since it will be
3669 // allocated and filled in by the dynamic linker.
PPC64LongBranchTargetSection()3670 PPC64LongBranchTargetSection::PPC64LongBranchTargetSection()
3671 : SyntheticSection(SHF_ALLOC | SHF_WRITE,
3672 config->isPic ? SHT_NOBITS : SHT_PROGBITS, 8,
3673 ".branch_lt") {}
3674
getEntryVA(const Symbol * sym,int64_t addend)3675 uint64_t PPC64LongBranchTargetSection::getEntryVA(const Symbol *sym,
3676 int64_t addend) {
3677 return getVA() + entry_index.find({sym, addend})->second * 8;
3678 }
3679
3680 std::optional<uint32_t>
addEntry(const Symbol * sym,int64_t addend)3681 PPC64LongBranchTargetSection::addEntry(const Symbol *sym, int64_t addend) {
3682 auto res =
3683 entry_index.try_emplace(std::make_pair(sym, addend), entries.size());
3684 if (!res.second)
3685 return std::nullopt;
3686 entries.emplace_back(sym, addend);
3687 return res.first->second;
3688 }
3689
getSize() const3690 size_t PPC64LongBranchTargetSection::getSize() const {
3691 return entries.size() * 8;
3692 }
3693
writeTo(uint8_t * buf)3694 void PPC64LongBranchTargetSection::writeTo(uint8_t *buf) {
3695 // If linking non-pic we have the final addresses of the targets and they get
3696 // written to the table directly. For pic the dynamic linker will allocate
3697 // the section and fill it.
3698 if (config->isPic)
3699 return;
3700
3701 for (auto entry : entries) {
3702 const Symbol *sym = entry.first;
3703 int64_t addend = entry.second;
3704 assert(sym->getVA());
3705 // Need calls to branch to the local entry-point since a long-branch
3706 // must be a local-call.
3707 write64(buf, sym->getVA(addend) +
3708 getPPC64GlobalEntryToLocalEntryOffset(sym->stOther));
3709 buf += 8;
3710 }
3711 }
3712
isNeeded() const3713 bool PPC64LongBranchTargetSection::isNeeded() const {
3714 // `removeUnusedSyntheticSections()` is called before thunk allocation which
3715 // is too early to determine if this section will be empty or not. We need
3716 // Finalized to keep the section alive until after thunk creation. Finalized
3717 // only gets set to true once `finalizeSections()` is called after thunk
3718 // creation. Because of this, if we don't create any long-branch thunks we end
3719 // up with an empty .branch_lt section in the binary.
3720 return !finalized || !entries.empty();
3721 }
3722
getAbiVersion()3723 static uint8_t getAbiVersion() {
3724 // MIPS non-PIC executable gets ABI version 1.
3725 if (config->emachine == EM_MIPS) {
3726 if (!config->isPic && !config->relocatable &&
3727 (config->eflags & (EF_MIPS_PIC | EF_MIPS_CPIC)) == EF_MIPS_CPIC)
3728 return 1;
3729 return 0;
3730 }
3731
3732 if (config->emachine == EM_AMDGPU && !ctx.objectFiles.empty()) {
3733 uint8_t ver = ctx.objectFiles[0]->abiVersion;
3734 for (InputFile *file : ArrayRef(ctx.objectFiles).slice(1))
3735 if (file->abiVersion != ver)
3736 error("incompatible ABI version: " + toString(file));
3737 return ver;
3738 }
3739
3740 return 0;
3741 }
3742
writeEhdr(uint8_t * buf,Partition & part)3743 template <typename ELFT> void elf::writeEhdr(uint8_t *buf, Partition &part) {
3744 memcpy(buf, "\177ELF", 4);
3745
3746 auto *eHdr = reinterpret_cast<typename ELFT::Ehdr *>(buf);
3747 eHdr->e_ident[EI_CLASS] = config->is64 ? ELFCLASS64 : ELFCLASS32;
3748 eHdr->e_ident[EI_DATA] = config->isLE ? ELFDATA2LSB : ELFDATA2MSB;
3749 eHdr->e_ident[EI_VERSION] = EV_CURRENT;
3750 eHdr->e_ident[EI_OSABI] = config->osabi;
3751 eHdr->e_ident[EI_ABIVERSION] = getAbiVersion();
3752 eHdr->e_machine = config->emachine;
3753 eHdr->e_version = EV_CURRENT;
3754 eHdr->e_flags = config->eflags;
3755 eHdr->e_ehsize = sizeof(typename ELFT::Ehdr);
3756 eHdr->e_phnum = part.phdrs.size();
3757 eHdr->e_shentsize = sizeof(typename ELFT::Shdr);
3758
3759 if (!config->relocatable) {
3760 eHdr->e_phoff = sizeof(typename ELFT::Ehdr);
3761 eHdr->e_phentsize = sizeof(typename ELFT::Phdr);
3762 }
3763 }
3764
writePhdrs(uint8_t * buf,Partition & part)3765 template <typename ELFT> void elf::writePhdrs(uint8_t *buf, Partition &part) {
3766 // Write the program header table.
3767 auto *hBuf = reinterpret_cast<typename ELFT::Phdr *>(buf);
3768 for (PhdrEntry *p : part.phdrs) {
3769 hBuf->p_type = p->p_type;
3770 hBuf->p_flags = p->p_flags;
3771 hBuf->p_offset = p->p_offset;
3772 hBuf->p_vaddr = p->p_vaddr;
3773 hBuf->p_paddr = p->p_paddr;
3774 hBuf->p_filesz = p->p_filesz;
3775 hBuf->p_memsz = p->p_memsz;
3776 hBuf->p_align = p->p_align;
3777 ++hBuf;
3778 }
3779 }
3780
3781 template <typename ELFT>
PartitionElfHeaderSection()3782 PartitionElfHeaderSection<ELFT>::PartitionElfHeaderSection()
3783 : SyntheticSection(SHF_ALLOC, SHT_LLVM_PART_EHDR, 1, "") {}
3784
3785 template <typename ELFT>
getSize() const3786 size_t PartitionElfHeaderSection<ELFT>::getSize() const {
3787 return sizeof(typename ELFT::Ehdr);
3788 }
3789
3790 template <typename ELFT>
writeTo(uint8_t * buf)3791 void PartitionElfHeaderSection<ELFT>::writeTo(uint8_t *buf) {
3792 writeEhdr<ELFT>(buf, getPartition());
3793
3794 // Loadable partitions are always ET_DYN.
3795 auto *eHdr = reinterpret_cast<typename ELFT::Ehdr *>(buf);
3796 eHdr->e_type = ET_DYN;
3797 }
3798
3799 template <typename ELFT>
PartitionProgramHeadersSection()3800 PartitionProgramHeadersSection<ELFT>::PartitionProgramHeadersSection()
3801 : SyntheticSection(SHF_ALLOC, SHT_LLVM_PART_PHDR, 1, ".phdrs") {}
3802
3803 template <typename ELFT>
getSize() const3804 size_t PartitionProgramHeadersSection<ELFT>::getSize() const {
3805 return sizeof(typename ELFT::Phdr) * getPartition().phdrs.size();
3806 }
3807
3808 template <typename ELFT>
writeTo(uint8_t * buf)3809 void PartitionProgramHeadersSection<ELFT>::writeTo(uint8_t *buf) {
3810 writePhdrs<ELFT>(buf, getPartition());
3811 }
3812
PartitionIndexSection()3813 PartitionIndexSection::PartitionIndexSection()
3814 : SyntheticSection(SHF_ALLOC, SHT_PROGBITS, 4, ".rodata") {}
3815
getSize() const3816 size_t PartitionIndexSection::getSize() const {
3817 return 12 * (partitions.size() - 1);
3818 }
3819
finalizeContents()3820 void PartitionIndexSection::finalizeContents() {
3821 for (size_t i = 1; i != partitions.size(); ++i)
3822 partitions[i].nameStrTab = mainPart->dynStrTab->addString(partitions[i].name);
3823 }
3824
writeTo(uint8_t * buf)3825 void PartitionIndexSection::writeTo(uint8_t *buf) {
3826 uint64_t va = getVA();
3827 for (size_t i = 1; i != partitions.size(); ++i) {
3828 write32(buf, mainPart->dynStrTab->getVA() + partitions[i].nameStrTab - va);
3829 write32(buf + 4, partitions[i].elfHeader->getVA() - (va + 4));
3830
3831 SyntheticSection *next = i == partitions.size() - 1
3832 ? in.partEnd.get()
3833 : partitions[i + 1].elfHeader.get();
3834 write32(buf + 8, next->getVA() - partitions[i].elfHeader->getVA());
3835
3836 va += 12;
3837 buf += 12;
3838 }
3839 }
3840
reset()3841 void InStruct::reset() {
3842 attributes.reset();
3843 riscvAttributes.reset();
3844 bss.reset();
3845 bssRelRo.reset();
3846 got.reset();
3847 gotPlt.reset();
3848 igotPlt.reset();
3849 relroPadding.reset();
3850 armCmseSGSection.reset();
3851 ppc64LongBranchTarget.reset();
3852 mipsAbiFlags.reset();
3853 mipsGot.reset();
3854 mipsOptions.reset();
3855 mipsReginfo.reset();
3856 mipsRldMap.reset();
3857 partEnd.reset();
3858 partIndex.reset();
3859 plt.reset();
3860 iplt.reset();
3861 ppc32Got2.reset();
3862 ibtPlt.reset();
3863 relaPlt.reset();
3864 relaIplt.reset();
3865 shStrTab.reset();
3866 strTab.reset();
3867 symTab.reset();
3868 symTabShndx.reset();
3869 }
3870
3871 constexpr char kMemtagAndroidNoteName[] = "Android";
writeTo(uint8_t * buf)3872 void MemtagAndroidNote::writeTo(uint8_t *buf) {
3873 static_assert(
3874 sizeof(kMemtagAndroidNoteName) == 8,
3875 "Android 11 & 12 have an ABI that the note name is 8 bytes long. Keep it "
3876 "that way for backwards compatibility.");
3877
3878 write32(buf, sizeof(kMemtagAndroidNoteName));
3879 write32(buf + 4, sizeof(uint32_t));
3880 write32(buf + 8, ELF::NT_ANDROID_TYPE_MEMTAG);
3881 memcpy(buf + 12, kMemtagAndroidNoteName, sizeof(kMemtagAndroidNoteName));
3882 buf += 12 + alignTo(sizeof(kMemtagAndroidNoteName), 4);
3883
3884 uint32_t value = 0;
3885 value |= config->androidMemtagMode;
3886 if (config->androidMemtagHeap)
3887 value |= ELF::NT_MEMTAG_HEAP;
3888 // Note, MTE stack is an ABI break. Attempting to run an MTE stack-enabled
3889 // binary on Android 11 or 12 will result in a checkfail in the loader.
3890 if (config->androidMemtagStack)
3891 value |= ELF::NT_MEMTAG_STACK;
3892 write32(buf, value); // note value
3893 }
3894
getSize() const3895 size_t MemtagAndroidNote::getSize() const {
3896 return sizeof(llvm::ELF::Elf64_Nhdr) +
3897 /*namesz=*/alignTo(sizeof(kMemtagAndroidNoteName), 4) +
3898 /*descsz=*/sizeof(uint32_t);
3899 }
3900
writeTo(uint8_t * buf)3901 void PackageMetadataNote::writeTo(uint8_t *buf) {
3902 write32(buf, 4);
3903 write32(buf + 4, config->packageMetadata.size() + 1);
3904 write32(buf + 8, FDO_PACKAGING_METADATA);
3905 memcpy(buf + 12, "FDO", 4);
3906 memcpy(buf + 16, config->packageMetadata.data(),
3907 config->packageMetadata.size());
3908 }
3909
getSize() const3910 size_t PackageMetadataNote::getSize() const {
3911 return sizeof(llvm::ELF::Elf64_Nhdr) + 4 +
3912 alignTo(config->packageMetadata.size() + 1, 4);
3913 }
3914
3915 // Helper function, return the size of the ULEB128 for 'v', optionally writing
3916 // it to `*(buf + offset)` if `buf` is non-null.
computeOrWriteULEB128(uint64_t v,uint8_t * buf,size_t offset)3917 static size_t computeOrWriteULEB128(uint64_t v, uint8_t *buf, size_t offset) {
3918 if (buf)
3919 return encodeULEB128(v, buf + offset);
3920 return getULEB128Size(v);
3921 }
3922
3923 // https://github.com/ARM-software/abi-aa/blob/main/memtagabielf64/memtagabielf64.rst#83encoding-of-sht_aarch64_memtag_globals_dynamic
3924 constexpr uint64_t kMemtagStepSizeBits = 3;
3925 constexpr uint64_t kMemtagGranuleSize = 16;
3926 static size_t
createMemtagGlobalDescriptors(const SmallVector<const Symbol *,0> & symbols,uint8_t * buf=nullptr)3927 createMemtagGlobalDescriptors(const SmallVector<const Symbol *, 0> &symbols,
3928 uint8_t *buf = nullptr) {
3929 size_t sectionSize = 0;
3930 uint64_t lastGlobalEnd = 0;
3931
3932 for (const Symbol *sym : symbols) {
3933 if (!includeInSymtab(*sym))
3934 continue;
3935 const uint64_t addr = sym->getVA();
3936 const uint64_t size = sym->getSize();
3937
3938 if (addr <= kMemtagGranuleSize && buf != nullptr)
3939 errorOrWarn("address of the tagged symbol \"" + sym->getName() +
3940 "\" falls in the ELF header. This is indicative of a "
3941 "compiler/linker bug");
3942 if (addr % kMemtagGranuleSize != 0)
3943 errorOrWarn("address of the tagged symbol \"" + sym->getName() +
3944 "\" at 0x" + Twine::utohexstr(addr) +
3945 "\" is not granule (16-byte) aligned");
3946 if (size == 0)
3947 errorOrWarn("size of the tagged symbol \"" + sym->getName() +
3948 "\" is not allowed to be zero");
3949 if (size % kMemtagGranuleSize != 0)
3950 errorOrWarn("size of the tagged symbol \"" + sym->getName() +
3951 "\" (size 0x" + Twine::utohexstr(size) +
3952 ") is not granule (16-byte) aligned");
3953
3954 const uint64_t sizeToEncode = size / kMemtagGranuleSize;
3955 const uint64_t stepToEncode = ((addr - lastGlobalEnd) / kMemtagGranuleSize)
3956 << kMemtagStepSizeBits;
3957 if (sizeToEncode < (1 << kMemtagStepSizeBits)) {
3958 sectionSize += computeOrWriteULEB128(stepToEncode | sizeToEncode, buf, sectionSize);
3959 } else {
3960 sectionSize += computeOrWriteULEB128(stepToEncode, buf, sectionSize);
3961 sectionSize += computeOrWriteULEB128(sizeToEncode - 1, buf, sectionSize);
3962 }
3963 lastGlobalEnd = addr + size;
3964 }
3965
3966 return sectionSize;
3967 }
3968
updateAllocSize()3969 bool MemtagGlobalDescriptors::updateAllocSize() {
3970 size_t oldSize = getSize();
3971 std::stable_sort(symbols.begin(), symbols.end(),
3972 [](const Symbol *s1, const Symbol *s2) {
3973 return s1->getVA() < s2->getVA();
3974 });
3975 return oldSize != getSize();
3976 }
3977
writeTo(uint8_t * buf)3978 void MemtagGlobalDescriptors::writeTo(uint8_t *buf) {
3979 createMemtagGlobalDescriptors(symbols, buf);
3980 }
3981
getSize() const3982 size_t MemtagGlobalDescriptors::getSize() const {
3983 return createMemtagGlobalDescriptors(symbols);
3984 }
3985
3986 InStruct elf::in;
3987
3988 std::vector<Partition> elf::partitions;
3989 Partition *elf::mainPart;
3990
3991 template GdbIndexSection *GdbIndexSection::create<ELF32LE>();
3992 template GdbIndexSection *GdbIndexSection::create<ELF32BE>();
3993 template GdbIndexSection *GdbIndexSection::create<ELF64LE>();
3994 template GdbIndexSection *GdbIndexSection::create<ELF64BE>();
3995
3996 template void elf::splitSections<ELF32LE>();
3997 template void elf::splitSections<ELF32BE>();
3998 template void elf::splitSections<ELF64LE>();
3999 template void elf::splitSections<ELF64BE>();
4000
4001 template class elf::MipsAbiFlagsSection<ELF32LE>;
4002 template class elf::MipsAbiFlagsSection<ELF32BE>;
4003 template class elf::MipsAbiFlagsSection<ELF64LE>;
4004 template class elf::MipsAbiFlagsSection<ELF64BE>;
4005
4006 template class elf::MipsOptionsSection<ELF32LE>;
4007 template class elf::MipsOptionsSection<ELF32BE>;
4008 template class elf::MipsOptionsSection<ELF64LE>;
4009 template class elf::MipsOptionsSection<ELF64BE>;
4010
4011 template void EhFrameSection::iterateFDEWithLSDA<ELF32LE>(
4012 function_ref<void(InputSection &)>);
4013 template void EhFrameSection::iterateFDEWithLSDA<ELF32BE>(
4014 function_ref<void(InputSection &)>);
4015 template void EhFrameSection::iterateFDEWithLSDA<ELF64LE>(
4016 function_ref<void(InputSection &)>);
4017 template void EhFrameSection::iterateFDEWithLSDA<ELF64BE>(
4018 function_ref<void(InputSection &)>);
4019
4020 template class elf::MipsReginfoSection<ELF32LE>;
4021 template class elf::MipsReginfoSection<ELF32BE>;
4022 template class elf::MipsReginfoSection<ELF64LE>;
4023 template class elf::MipsReginfoSection<ELF64BE>;
4024
4025 template class elf::DynamicSection<ELF32LE>;
4026 template class elf::DynamicSection<ELF32BE>;
4027 template class elf::DynamicSection<ELF64LE>;
4028 template class elf::DynamicSection<ELF64BE>;
4029
4030 template class elf::RelocationSection<ELF32LE>;
4031 template class elf::RelocationSection<ELF32BE>;
4032 template class elf::RelocationSection<ELF64LE>;
4033 template class elf::RelocationSection<ELF64BE>;
4034
4035 template class elf::AndroidPackedRelocationSection<ELF32LE>;
4036 template class elf::AndroidPackedRelocationSection<ELF32BE>;
4037 template class elf::AndroidPackedRelocationSection<ELF64LE>;
4038 template class elf::AndroidPackedRelocationSection<ELF64BE>;
4039
4040 template class elf::RelrSection<ELF32LE>;
4041 template class elf::RelrSection<ELF32BE>;
4042 template class elf::RelrSection<ELF64LE>;
4043 template class elf::RelrSection<ELF64BE>;
4044
4045 template class elf::SymbolTableSection<ELF32LE>;
4046 template class elf::SymbolTableSection<ELF32BE>;
4047 template class elf::SymbolTableSection<ELF64LE>;
4048 template class elf::SymbolTableSection<ELF64BE>;
4049
4050 template class elf::VersionNeedSection<ELF32LE>;
4051 template class elf::VersionNeedSection<ELF32BE>;
4052 template class elf::VersionNeedSection<ELF64LE>;
4053 template class elf::VersionNeedSection<ELF64BE>;
4054
4055 template void elf::writeEhdr<ELF32LE>(uint8_t *Buf, Partition &Part);
4056 template void elf::writeEhdr<ELF32BE>(uint8_t *Buf, Partition &Part);
4057 template void elf::writeEhdr<ELF64LE>(uint8_t *Buf, Partition &Part);
4058 template void elf::writeEhdr<ELF64BE>(uint8_t *Buf, Partition &Part);
4059
4060 template void elf::writePhdrs<ELF32LE>(uint8_t *Buf, Partition &Part);
4061 template void elf::writePhdrs<ELF32BE>(uint8_t *Buf, Partition &Part);
4062 template void elf::writePhdrs<ELF64LE>(uint8_t *Buf, Partition &Part);
4063 template void elf::writePhdrs<ELF64BE>(uint8_t *Buf, Partition &Part);
4064
4065 template class elf::PartitionElfHeaderSection<ELF32LE>;
4066 template class elf::PartitionElfHeaderSection<ELF32BE>;
4067 template class elf::PartitionElfHeaderSection<ELF64LE>;
4068 template class elf::PartitionElfHeaderSection<ELF64BE>;
4069
4070 template class elf::PartitionProgramHeadersSection<ELF32LE>;
4071 template class elf::PartitionProgramHeadersSection<ELF32BE>;
4072 template class elf::PartitionProgramHeadersSection<ELF64LE>;
4073 template class elf::PartitionProgramHeadersSection<ELF64BE>;
4074