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 "InputFiles.h"
19 #include "LinkerScript.h"
20 #include "OutputSections.h"
21 #include "SymbolTable.h"
22 #include "Symbols.h"
23 #include "Target.h"
24 #include "Writer.h"
25 #include "lld/Common/DWARF.h"
26 #include "lld/Common/ErrorHandler.h"
27 #include "lld/Common/Memory.h"
28 #include "lld/Common/Strings.h"
29 #include "lld/Common/Version.h"
30 #include "llvm/ADT/SetOperations.h"
31 #include "llvm/ADT/StringExtras.h"
32 #include "llvm/BinaryFormat/Dwarf.h"
33 #include "llvm/DebugInfo/DWARF/DWARFDebugPubTable.h"
34 #include "llvm/Object/ELFObjectFile.h"
35 #include "llvm/Support/Compression.h"
36 #include "llvm/Support/Endian.h"
37 #include "llvm/Support/LEB128.h"
38 #include "llvm/Support/MD5.h"
39 #include "llvm/Support/Parallel.h"
40 #include "llvm/Support/TimeProfiler.h"
41 #include <cstdlib>
42 #include <thread>
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 return make<MergeInputSection>(SHF_MERGE | SHF_STRINGS, SHT_PROGBITS, 1,
88 getVersion(), ".comment");
89 }
90
91 // .MIPS.abiflags section.
92 template <class ELFT>
MipsAbiFlagsSection(Elf_Mips_ABIFlags flags)93 MipsAbiFlagsSection<ELFT>::MipsAbiFlagsSection(Elf_Mips_ABIFlags flags)
94 : SyntheticSection(SHF_ALLOC, SHT_MIPS_ABIFLAGS, 8, ".MIPS.abiflags"),
95 flags(flags) {
96 this->entsize = sizeof(Elf_Mips_ABIFlags);
97 }
98
writeTo(uint8_t * buf)99 template <class ELFT> void MipsAbiFlagsSection<ELFT>::writeTo(uint8_t *buf) {
100 memcpy(buf, &flags, sizeof(flags));
101 }
102
103 template <class ELFT>
create()104 MipsAbiFlagsSection<ELFT> *MipsAbiFlagsSection<ELFT>::create() {
105 Elf_Mips_ABIFlags flags = {};
106 bool create = false;
107
108 for (InputSectionBase *sec : inputSections) {
109 if (sec->type != SHT_MIPS_ABIFLAGS)
110 continue;
111 sec->markDead();
112 create = true;
113
114 std::string filename = toString(sec->file);
115 const size_t size = sec->data().size();
116 // Older version of BFD (such as the default FreeBSD linker) concatenate
117 // .MIPS.abiflags instead of merging. To allow for this case (or potential
118 // zero padding) we ignore everything after the first Elf_Mips_ABIFlags
119 if (size < sizeof(Elf_Mips_ABIFlags)) {
120 error(filename + ": invalid size of .MIPS.abiflags section: got " +
121 Twine(size) + " instead of " + Twine(sizeof(Elf_Mips_ABIFlags)));
122 return nullptr;
123 }
124 auto *s = reinterpret_cast<const Elf_Mips_ABIFlags *>(sec->data().data());
125 if (s->version != 0) {
126 error(filename + ": unexpected .MIPS.abiflags version " +
127 Twine(s->version));
128 return nullptr;
129 }
130
131 // LLD checks ISA compatibility in calcMipsEFlags(). Here we just
132 // select the highest number of ISA/Rev/Ext.
133 flags.isa_level = std::max(flags.isa_level, s->isa_level);
134 flags.isa_rev = std::max(flags.isa_rev, s->isa_rev);
135 flags.isa_ext = std::max(flags.isa_ext, s->isa_ext);
136 flags.gpr_size = std::max(flags.gpr_size, s->gpr_size);
137 flags.cpr1_size = std::max(flags.cpr1_size, s->cpr1_size);
138 flags.cpr2_size = std::max(flags.cpr2_size, s->cpr2_size);
139 flags.ases |= s->ases;
140 flags.flags1 |= s->flags1;
141 flags.flags2 |= s->flags2;
142 flags.fp_abi = elf::getMipsFpAbiFlag(flags.fp_abi, s->fp_abi, filename);
143 };
144
145 if (create)
146 return make<MipsAbiFlagsSection<ELFT>>(flags);
147 return nullptr;
148 }
149
150 // .MIPS.options section.
151 template <class ELFT>
MipsOptionsSection(Elf_Mips_RegInfo reginfo)152 MipsOptionsSection<ELFT>::MipsOptionsSection(Elf_Mips_RegInfo reginfo)
153 : SyntheticSection(SHF_ALLOC, SHT_MIPS_OPTIONS, 8, ".MIPS.options"),
154 reginfo(reginfo) {
155 this->entsize = sizeof(Elf_Mips_Options) + sizeof(Elf_Mips_RegInfo);
156 }
157
writeTo(uint8_t * buf)158 template <class ELFT> void MipsOptionsSection<ELFT>::writeTo(uint8_t *buf) {
159 auto *options = reinterpret_cast<Elf_Mips_Options *>(buf);
160 options->kind = ODK_REGINFO;
161 options->size = getSize();
162
163 if (!config->relocatable)
164 reginfo.ri_gp_value = in.mipsGot->getGp();
165 memcpy(buf + sizeof(Elf_Mips_Options), ®info, sizeof(reginfo));
166 }
167
168 template <class ELFT>
create()169 MipsOptionsSection<ELFT> *MipsOptionsSection<ELFT>::create() {
170 // N64 ABI only.
171 if (!ELFT::Is64Bits)
172 return nullptr;
173
174 std::vector<InputSectionBase *> sections;
175 for (InputSectionBase *sec : inputSections)
176 if (sec->type == SHT_MIPS_OPTIONS)
177 sections.push_back(sec);
178
179 if (sections.empty())
180 return nullptr;
181
182 Elf_Mips_RegInfo reginfo = {};
183 for (InputSectionBase *sec : sections) {
184 sec->markDead();
185
186 std::string filename = toString(sec->file);
187 ArrayRef<uint8_t> d = sec->data();
188
189 while (!d.empty()) {
190 if (d.size() < sizeof(Elf_Mips_Options)) {
191 error(filename + ": invalid size of .MIPS.options section");
192 break;
193 }
194
195 auto *opt = reinterpret_cast<const Elf_Mips_Options *>(d.data());
196 if (opt->kind == ODK_REGINFO) {
197 reginfo.ri_gprmask |= opt->getRegInfo().ri_gprmask;
198 sec->getFile<ELFT>()->mipsGp0 = opt->getRegInfo().ri_gp_value;
199 break;
200 }
201
202 if (!opt->size)
203 fatal(filename + ": zero option descriptor size");
204 d = d.slice(opt->size);
205 }
206 };
207
208 return make<MipsOptionsSection<ELFT>>(reginfo);
209 }
210
211 // MIPS .reginfo section.
212 template <class ELFT>
MipsReginfoSection(Elf_Mips_RegInfo reginfo)213 MipsReginfoSection<ELFT>::MipsReginfoSection(Elf_Mips_RegInfo reginfo)
214 : SyntheticSection(SHF_ALLOC, SHT_MIPS_REGINFO, 4, ".reginfo"),
215 reginfo(reginfo) {
216 this->entsize = sizeof(Elf_Mips_RegInfo);
217 }
218
writeTo(uint8_t * buf)219 template <class ELFT> void MipsReginfoSection<ELFT>::writeTo(uint8_t *buf) {
220 if (!config->relocatable)
221 reginfo.ri_gp_value = in.mipsGot->getGp();
222 memcpy(buf, ®info, sizeof(reginfo));
223 }
224
225 template <class ELFT>
create()226 MipsReginfoSection<ELFT> *MipsReginfoSection<ELFT>::create() {
227 // Section should be alive for O32 and N32 ABIs only.
228 if (ELFT::Is64Bits)
229 return nullptr;
230
231 std::vector<InputSectionBase *> sections;
232 for (InputSectionBase *sec : inputSections)
233 if (sec->type == SHT_MIPS_REGINFO)
234 sections.push_back(sec);
235
236 if (sections.empty())
237 return nullptr;
238
239 Elf_Mips_RegInfo reginfo = {};
240 for (InputSectionBase *sec : sections) {
241 sec->markDead();
242
243 if (sec->data().size() != sizeof(Elf_Mips_RegInfo)) {
244 error(toString(sec->file) + ": invalid size of .reginfo section");
245 return nullptr;
246 }
247
248 auto *r = reinterpret_cast<const Elf_Mips_RegInfo *>(sec->data().data());
249 reginfo.ri_gprmask |= r->ri_gprmask;
250 sec->getFile<ELFT>()->mipsGp0 = r->ri_gp_value;
251 };
252
253 return make<MipsReginfoSection<ELFT>>(reginfo);
254 }
255
createInterpSection()256 InputSection *elf::createInterpSection() {
257 // StringSaver guarantees that the returned string ends with '\0'.
258 StringRef s = saver.save(config->dynamicLinker);
259 ArrayRef<uint8_t> contents = {(const uint8_t *)s.data(), s.size() + 1};
260
261 return make<InputSection>(nullptr, SHF_ALLOC, SHT_PROGBITS, 1, contents,
262 ".interp");
263 }
264
addSyntheticLocal(StringRef name,uint8_t type,uint64_t value,uint64_t size,InputSectionBase & section)265 Defined *elf::addSyntheticLocal(StringRef name, uint8_t type, uint64_t value,
266 uint64_t size, InputSectionBase §ion) {
267 auto *s = make<Defined>(section.file, name, STB_LOCAL, STV_DEFAULT, type,
268 value, size, §ion);
269 if (in.symTab)
270 in.symTab->addSymbol(s);
271 return s;
272 }
273
getHashSize()274 static size_t getHashSize() {
275 switch (config->buildId) {
276 case BuildIdKind::Fast:
277 return 8;
278 case BuildIdKind::Md5:
279 case BuildIdKind::Uuid:
280 return 16;
281 case BuildIdKind::Sha1:
282 return 20;
283 case BuildIdKind::Hexstring:
284 return config->buildIdVector.size();
285 default:
286 llvm_unreachable("unknown BuildIdKind");
287 }
288 }
289
290 // This class represents a linker-synthesized .note.gnu.property section.
291 //
292 // In x86 and AArch64, object files may contain feature flags indicating the
293 // features that they have used. The flags are stored in a .note.gnu.property
294 // section.
295 //
296 // lld reads the sections from input files and merges them by computing AND of
297 // the flags. The result is written as a new .note.gnu.property section.
298 //
299 // If the flag is zero (which indicates that the intersection of the feature
300 // sets is empty, or some input files didn't have .note.gnu.property sections),
301 // we don't create this section.
GnuPropertySection()302 GnuPropertySection::GnuPropertySection()
303 : SyntheticSection(llvm::ELF::SHF_ALLOC, llvm::ELF::SHT_NOTE,
304 config->wordsize, ".note.gnu.property") {}
305
writeTo(uint8_t * buf)306 void GnuPropertySection::writeTo(uint8_t *buf) {
307 uint32_t featureAndType = config->emachine == EM_AARCH64
308 ? GNU_PROPERTY_AARCH64_FEATURE_1_AND
309 : GNU_PROPERTY_X86_FEATURE_1_AND;
310
311 write32(buf, 4); // Name size
312 write32(buf + 4, config->is64 ? 16 : 12); // Content size
313 write32(buf + 8, NT_GNU_PROPERTY_TYPE_0); // Type
314 memcpy(buf + 12, "GNU", 4); // Name string
315 write32(buf + 16, featureAndType); // Feature type
316 write32(buf + 20, 4); // Feature size
317 write32(buf + 24, config->andFeatures); // Feature flags
318 if (config->is64)
319 write32(buf + 28, 0); // Padding
320 }
321
getSize() const322 size_t GnuPropertySection::getSize() const { return config->is64 ? 32 : 28; }
323
BuildIdSection()324 BuildIdSection::BuildIdSection()
325 : SyntheticSection(SHF_ALLOC, SHT_NOTE, 4, ".note.gnu.build-id"),
326 hashSize(getHashSize()) {}
327
writeTo(uint8_t * buf)328 void BuildIdSection::writeTo(uint8_t *buf) {
329 write32(buf, 4); // Name size
330 write32(buf + 4, hashSize); // Content size
331 write32(buf + 8, NT_GNU_BUILD_ID); // Type
332 memcpy(buf + 12, "GNU", 4); // Name string
333 hashBuf = buf + 16;
334 }
335
writeBuildId(ArrayRef<uint8_t> buf)336 void BuildIdSection::writeBuildId(ArrayRef<uint8_t> buf) {
337 assert(buf.size() == hashSize);
338 memcpy(hashBuf, buf.data(), hashSize);
339 }
340
BssSection(StringRef name,uint64_t size,uint32_t alignment)341 BssSection::BssSection(StringRef name, uint64_t size, uint32_t alignment)
342 : SyntheticSection(SHF_ALLOC | SHF_WRITE, SHT_NOBITS, alignment, name) {
343 this->bss = true;
344 this->size = size;
345 }
346
EhFrameSection()347 EhFrameSection::EhFrameSection()
348 : SyntheticSection(SHF_ALLOC, SHT_PROGBITS, 1, ".eh_frame") {}
349
350 // Search for an existing CIE record or create a new one.
351 // CIE records from input object files are uniquified by their contents
352 // and where their relocations point to.
353 template <class ELFT, class RelTy>
addCie(EhSectionPiece & cie,ArrayRef<RelTy> rels)354 CieRecord *EhFrameSection::addCie(EhSectionPiece &cie, ArrayRef<RelTy> rels) {
355 Symbol *personality = nullptr;
356 unsigned firstRelI = cie.firstRelocation;
357 if (firstRelI != (unsigned)-1)
358 personality =
359 &cie.sec->template getFile<ELFT>()->getRelocTargetSym(rels[firstRelI]);
360
361 // Search for an existing CIE by CIE contents/relocation target pair.
362 CieRecord *&rec = cieMap[{cie.data(), personality}];
363
364 // If not found, create a new one.
365 if (!rec) {
366 rec = make<CieRecord>();
367 rec->cie = &cie;
368 cieRecords.push_back(rec);
369 }
370 return rec;
371 }
372
373 // There is one FDE per function. Returns a non-null pointer to the function
374 // symbol if the given FDE points to a live function.
375 template <class ELFT, class RelTy>
isFdeLive(EhSectionPiece & fde,ArrayRef<RelTy> rels)376 Defined *EhFrameSection::isFdeLive(EhSectionPiece &fde, ArrayRef<RelTy> rels) {
377 auto *sec = cast<EhInputSection>(fde.sec);
378 unsigned firstRelI = fde.firstRelocation;
379
380 // An FDE should point to some function because FDEs are to describe
381 // functions. That's however not always the case due to an issue of
382 // ld.gold with -r. ld.gold may discard only functions and leave their
383 // corresponding FDEs, which results in creating bad .eh_frame sections.
384 // To deal with that, we ignore such FDEs.
385 if (firstRelI == (unsigned)-1)
386 return nullptr;
387
388 const RelTy &rel = rels[firstRelI];
389 Symbol &b = sec->template getFile<ELFT>()->getRelocTargetSym(rel);
390
391 // FDEs for garbage-collected or merged-by-ICF sections, or sections in
392 // another partition, are dead.
393 if (auto *d = dyn_cast<Defined>(&b))
394 if (d->section && d->section->partition == partition)
395 return d;
396 return nullptr;
397 }
398
399 // .eh_frame is a sequence of CIE or FDE records. In general, there
400 // is one CIE record per input object file which is followed by
401 // a list of FDEs. This function searches an existing CIE or create a new
402 // one and associates FDEs to the CIE.
403 template <class ELFT, class RelTy>
addRecords(EhInputSection * sec,ArrayRef<RelTy> rels)404 void EhFrameSection::addRecords(EhInputSection *sec, ArrayRef<RelTy> rels) {
405 offsetToCie.clear();
406 for (EhSectionPiece &piece : sec->pieces) {
407 // The empty record is the end marker.
408 if (piece.size == 4)
409 return;
410
411 size_t offset = piece.inputOff;
412 uint32_t id = read32(piece.data().data() + 4);
413 if (id == 0) {
414 offsetToCie[offset] = addCie<ELFT>(piece, rels);
415 continue;
416 }
417
418 uint32_t cieOffset = offset + 4 - id;
419 CieRecord *rec = offsetToCie[cieOffset];
420 if (!rec)
421 fatal(toString(sec) + ": invalid CIE reference");
422
423 if (!isFdeLive<ELFT>(piece, rels))
424 continue;
425 rec->fdes.push_back(&piece);
426 numFdes++;
427 }
428 }
429
430 template <class ELFT>
addSectionAux(EhInputSection * sec)431 void EhFrameSection::addSectionAux(EhInputSection *sec) {
432 if (!sec->isLive())
433 return;
434 if (sec->areRelocsRela)
435 addRecords<ELFT>(sec, sec->template relas<ELFT>());
436 else
437 addRecords<ELFT>(sec, sec->template rels<ELFT>());
438 }
439
addSection(EhInputSection * sec)440 void EhFrameSection::addSection(EhInputSection *sec) {
441 sec->parent = this;
442
443 alignment = std::max(alignment, sec->alignment);
444 sections.push_back(sec);
445
446 for (auto *ds : sec->dependentSections)
447 dependentSections.push_back(ds);
448 }
449
450 // Used by ICF<ELFT>::handleLSDA(). This function is very similar to
451 // EhFrameSection::addRecords().
452 template <class ELFT, class RelTy>
iterateFDEWithLSDAAux(EhInputSection & sec,ArrayRef<RelTy> rels,DenseSet<size_t> & ciesWithLSDA,llvm::function_ref<void (InputSection &)> fn)453 void EhFrameSection::iterateFDEWithLSDAAux(
454 EhInputSection &sec, ArrayRef<RelTy> rels, DenseSet<size_t> &ciesWithLSDA,
455 llvm::function_ref<void(InputSection &)> fn) {
456 for (EhSectionPiece &piece : sec.pieces) {
457 // Skip ZERO terminator.
458 if (piece.size == 4)
459 continue;
460
461 size_t offset = piece.inputOff;
462 uint32_t id =
463 endian::read32<ELFT::TargetEndianness>(piece.data().data() + 4);
464 if (id == 0) {
465 if (hasLSDA(piece))
466 ciesWithLSDA.insert(offset);
467 continue;
468 }
469 uint32_t cieOffset = offset + 4 - id;
470 if (ciesWithLSDA.count(cieOffset) == 0)
471 continue;
472
473 // The CIE has a LSDA argument. Call fn with d's section.
474 if (Defined *d = isFdeLive<ELFT>(piece, rels))
475 if (auto *s = dyn_cast_or_null<InputSection>(d->section))
476 fn(*s);
477 }
478 }
479
480 template <class ELFT>
iterateFDEWithLSDA(llvm::function_ref<void (InputSection &)> fn)481 void EhFrameSection::iterateFDEWithLSDA(
482 llvm::function_ref<void(InputSection &)> fn) {
483 DenseSet<size_t> ciesWithLSDA;
484 for (EhInputSection *sec : sections) {
485 ciesWithLSDA.clear();
486 if (sec->areRelocsRela)
487 iterateFDEWithLSDAAux<ELFT>(*sec, sec->template relas<ELFT>(),
488 ciesWithLSDA, fn);
489 else
490 iterateFDEWithLSDAAux<ELFT>(*sec, sec->template rels<ELFT>(),
491 ciesWithLSDA, fn);
492 }
493 }
494
writeCieFde(uint8_t * buf,ArrayRef<uint8_t> d)495 static void writeCieFde(uint8_t *buf, ArrayRef<uint8_t> d) {
496 memcpy(buf, d.data(), d.size());
497
498 size_t aligned = alignTo(d.size(), config->wordsize);
499
500 // Zero-clear trailing padding if it exists.
501 memset(buf + d.size(), 0, aligned - d.size());
502
503 // Fix the size field. -4 since size does not include the size field itself.
504 write32(buf, aligned - 4);
505 }
506
finalizeContents()507 void EhFrameSection::finalizeContents() {
508 assert(!this->size); // Not finalized.
509
510 switch (config->ekind) {
511 case ELFNoneKind:
512 llvm_unreachable("invalid ekind");
513 case ELF32LEKind:
514 for (EhInputSection *sec : sections)
515 addSectionAux<ELF32LE>(sec);
516 break;
517 case ELF32BEKind:
518 for (EhInputSection *sec : sections)
519 addSectionAux<ELF32BE>(sec);
520 break;
521 case ELF64LEKind:
522 for (EhInputSection *sec : sections)
523 addSectionAux<ELF64LE>(sec);
524 break;
525 case ELF64BEKind:
526 for (EhInputSection *sec : sections)
527 addSectionAux<ELF64BE>(sec);
528 break;
529 }
530
531 size_t off = 0;
532 for (CieRecord *rec : cieRecords) {
533 rec->cie->outputOff = off;
534 off += alignTo(rec->cie->size, config->wordsize);
535
536 for (EhSectionPiece *fde : rec->fdes) {
537 fde->outputOff = off;
538 off += alignTo(fde->size, config->wordsize);
539 }
540 }
541
542 // The LSB standard does not allow a .eh_frame section with zero
543 // Call Frame Information records. glibc unwind-dw2-fde.c
544 // classify_object_over_fdes expects there is a CIE record length 0 as a
545 // terminator. Thus we add one unconditionally.
546 off += 4;
547
548 this->size = off;
549 }
550
551 // Returns data for .eh_frame_hdr. .eh_frame_hdr is a binary search table
552 // to get an FDE from an address to which FDE is applied. This function
553 // returns a list of such pairs.
getFdeData() const554 std::vector<EhFrameSection::FdeData> EhFrameSection::getFdeData() const {
555 uint8_t *buf = Out::bufferStart + getParent()->offset + outSecOff;
556 std::vector<FdeData> ret;
557
558 uint64_t va = getPartition().ehFrameHdr->getVA();
559 for (CieRecord *rec : cieRecords) {
560 uint8_t enc = getFdeEncoding(rec->cie);
561 for (EhSectionPiece *fde : rec->fdes) {
562 uint64_t pc = getFdePc(buf, fde->outputOff, enc);
563 uint64_t fdeVA = getParent()->addr + fde->outputOff;
564 if (!isInt<32>(pc - va))
565 fatal(toString(fde->sec) + ": PC offset is too large: 0x" +
566 Twine::utohexstr(pc - va));
567 ret.push_back({uint32_t(pc - va), uint32_t(fdeVA - va)});
568 }
569 }
570
571 // Sort the FDE list by their PC and uniqueify. Usually there is only
572 // one FDE for a PC (i.e. function), but if ICF merges two functions
573 // into one, there can be more than one FDEs pointing to the address.
574 auto less = [](const FdeData &a, const FdeData &b) {
575 return a.pcRel < b.pcRel;
576 };
577 llvm::stable_sort(ret, less);
578 auto eq = [](const FdeData &a, const FdeData &b) {
579 return a.pcRel == b.pcRel;
580 };
581 ret.erase(std::unique(ret.begin(), ret.end(), eq), ret.end());
582
583 return ret;
584 }
585
readFdeAddr(uint8_t * buf,int size)586 static uint64_t readFdeAddr(uint8_t *buf, int size) {
587 switch (size) {
588 case DW_EH_PE_udata2:
589 return read16(buf);
590 case DW_EH_PE_sdata2:
591 return (int16_t)read16(buf);
592 case DW_EH_PE_udata4:
593 return read32(buf);
594 case DW_EH_PE_sdata4:
595 return (int32_t)read32(buf);
596 case DW_EH_PE_udata8:
597 case DW_EH_PE_sdata8:
598 return read64(buf);
599 case DW_EH_PE_absptr:
600 return readUint(buf);
601 }
602 fatal("unknown FDE size encoding");
603 }
604
605 // Returns the VA to which a given FDE (on a mmap'ed buffer) is applied to.
606 // We need it to create .eh_frame_hdr section.
getFdePc(uint8_t * buf,size_t fdeOff,uint8_t enc) const607 uint64_t EhFrameSection::getFdePc(uint8_t *buf, size_t fdeOff,
608 uint8_t enc) const {
609 // The starting address to which this FDE applies is
610 // stored at FDE + 8 byte.
611 size_t off = fdeOff + 8;
612 uint64_t addr = readFdeAddr(buf + off, enc & 0xf);
613 if ((enc & 0x70) == DW_EH_PE_absptr)
614 return addr;
615 if ((enc & 0x70) == DW_EH_PE_pcrel)
616 return addr + getParent()->addr + off;
617 fatal("unknown FDE size relative encoding");
618 }
619
writeTo(uint8_t * buf)620 void EhFrameSection::writeTo(uint8_t *buf) {
621 // Write CIE and FDE records.
622 for (CieRecord *rec : cieRecords) {
623 size_t cieOffset = rec->cie->outputOff;
624 writeCieFde(buf + cieOffset, rec->cie->data());
625
626 for (EhSectionPiece *fde : rec->fdes) {
627 size_t off = fde->outputOff;
628 writeCieFde(buf + off, fde->data());
629
630 // FDE's second word should have the offset to an associated CIE.
631 // Write it.
632 write32(buf + off + 4, off + 4 - cieOffset);
633 }
634 }
635
636 // Apply relocations. .eh_frame section contents are not contiguous
637 // in the output buffer, but relocateAlloc() still works because
638 // getOffset() takes care of discontiguous section pieces.
639 for (EhInputSection *s : sections)
640 s->relocateAlloc(buf, nullptr);
641
642 if (getPartition().ehFrameHdr && getPartition().ehFrameHdr->getParent())
643 getPartition().ehFrameHdr->write();
644 }
645
GotSection()646 GotSection::GotSection()
647 : SyntheticSection(SHF_ALLOC | SHF_WRITE, SHT_PROGBITS, config->wordsize,
648 ".got") {
649 // If ElfSym::globalOffsetTable is relative to .got and is referenced,
650 // increase numEntries by the number of entries used to emit
651 // ElfSym::globalOffsetTable.
652 if (ElfSym::globalOffsetTable && !target->gotBaseSymInGotPlt)
653 numEntries += target->gotHeaderEntriesNum;
654 }
655
addEntry(Symbol & sym)656 void GotSection::addEntry(Symbol &sym) {
657 sym.gotIndex = numEntries;
658 ++numEntries;
659 }
660
addDynTlsEntry(Symbol & sym)661 bool GotSection::addDynTlsEntry(Symbol &sym) {
662 if (sym.globalDynIndex != -1U)
663 return false;
664 sym.globalDynIndex = numEntries;
665 // Global Dynamic TLS entries take two GOT slots.
666 numEntries += 2;
667 return true;
668 }
669
670 // Reserves TLS entries for a TLS module ID and a TLS block offset.
671 // In total it takes two GOT slots.
addTlsIndex()672 bool GotSection::addTlsIndex() {
673 if (tlsIndexOff != uint32_t(-1))
674 return false;
675 tlsIndexOff = numEntries * config->wordsize;
676 numEntries += 2;
677 return true;
678 }
679
getGlobalDynAddr(const Symbol & b) const680 uint64_t GotSection::getGlobalDynAddr(const Symbol &b) const {
681 return this->getVA() + b.globalDynIndex * config->wordsize;
682 }
683
getGlobalDynOffset(const Symbol & b) const684 uint64_t GotSection::getGlobalDynOffset(const Symbol &b) const {
685 return b.globalDynIndex * config->wordsize;
686 }
687
finalizeContents()688 void GotSection::finalizeContents() {
689 size = numEntries * config->wordsize;
690 }
691
isNeeded() const692 bool GotSection::isNeeded() const {
693 // We need to emit a GOT even if it's empty if there's a relocation that is
694 // relative to GOT(such as GOTOFFREL).
695 return numEntries || hasGotOffRel;
696 }
697
writeTo(uint8_t * buf)698 void GotSection::writeTo(uint8_t *buf) {
699 target->writeGotHeader(buf);
700 relocateAlloc(buf, buf + size);
701 }
702
getMipsPageAddr(uint64_t addr)703 static uint64_t getMipsPageAddr(uint64_t addr) {
704 return (addr + 0x8000) & ~0xffff;
705 }
706
getMipsPageCount(uint64_t size)707 static uint64_t getMipsPageCount(uint64_t size) {
708 return (size + 0xfffe) / 0xffff + 1;
709 }
710
MipsGotSection()711 MipsGotSection::MipsGotSection()
712 : SyntheticSection(SHF_ALLOC | SHF_WRITE | SHF_MIPS_GPREL, SHT_PROGBITS, 16,
713 ".got") {}
714
addEntry(InputFile & file,Symbol & sym,int64_t addend,RelExpr expr)715 void MipsGotSection::addEntry(InputFile &file, Symbol &sym, int64_t addend,
716 RelExpr expr) {
717 FileGot &g = getGot(file);
718 if (expr == R_MIPS_GOT_LOCAL_PAGE) {
719 if (const OutputSection *os = sym.getOutputSection())
720 g.pagesMap.insert({os, {}});
721 else
722 g.local16.insert({{nullptr, getMipsPageAddr(sym.getVA(addend))}, 0});
723 } else if (sym.isTls())
724 g.tls.insert({&sym, 0});
725 else if (sym.isPreemptible && expr == R_ABS)
726 g.relocs.insert({&sym, 0});
727 else if (sym.isPreemptible)
728 g.global.insert({&sym, 0});
729 else if (expr == R_MIPS_GOT_OFF32)
730 g.local32.insert({{&sym, addend}, 0});
731 else
732 g.local16.insert({{&sym, addend}, 0});
733 }
734
addDynTlsEntry(InputFile & file,Symbol & sym)735 void MipsGotSection::addDynTlsEntry(InputFile &file, Symbol &sym) {
736 getGot(file).dynTlsSymbols.insert({&sym, 0});
737 }
738
addTlsIndex(InputFile & file)739 void MipsGotSection::addTlsIndex(InputFile &file) {
740 getGot(file).dynTlsSymbols.insert({nullptr, 0});
741 }
742
getEntriesNum() const743 size_t MipsGotSection::FileGot::getEntriesNum() const {
744 return getPageEntriesNum() + local16.size() + global.size() + relocs.size() +
745 tls.size() + dynTlsSymbols.size() * 2;
746 }
747
getPageEntriesNum() const748 size_t MipsGotSection::FileGot::getPageEntriesNum() const {
749 size_t num = 0;
750 for (const std::pair<const OutputSection *, FileGot::PageBlock> &p : pagesMap)
751 num += p.second.count;
752 return num;
753 }
754
getIndexedEntriesNum() const755 size_t MipsGotSection::FileGot::getIndexedEntriesNum() const {
756 size_t count = getPageEntriesNum() + local16.size() + global.size();
757 // If there are relocation-only entries in the GOT, TLS entries
758 // are allocated after them. TLS entries should be addressable
759 // by 16-bit index so count both reloc-only and TLS entries.
760 if (!tls.empty() || !dynTlsSymbols.empty())
761 count += relocs.size() + tls.size() + dynTlsSymbols.size() * 2;
762 return count;
763 }
764
getGot(InputFile & f)765 MipsGotSection::FileGot &MipsGotSection::getGot(InputFile &f) {
766 if (!f.mipsGotIndex.hasValue()) {
767 gots.emplace_back();
768 gots.back().file = &f;
769 f.mipsGotIndex = gots.size() - 1;
770 }
771 return gots[*f.mipsGotIndex];
772 }
773
getPageEntryOffset(const InputFile * f,const Symbol & sym,int64_t addend) const774 uint64_t MipsGotSection::getPageEntryOffset(const InputFile *f,
775 const Symbol &sym,
776 int64_t addend) const {
777 const FileGot &g = gots[*f->mipsGotIndex];
778 uint64_t index = 0;
779 if (const OutputSection *outSec = sym.getOutputSection()) {
780 uint64_t secAddr = getMipsPageAddr(outSec->addr);
781 uint64_t symAddr = getMipsPageAddr(sym.getVA(addend));
782 index = g.pagesMap.lookup(outSec).firstIndex + (symAddr - secAddr) / 0xffff;
783 } else {
784 index = g.local16.lookup({nullptr, getMipsPageAddr(sym.getVA(addend))});
785 }
786 return index * config->wordsize;
787 }
788
getSymEntryOffset(const InputFile * f,const Symbol & s,int64_t addend) const789 uint64_t MipsGotSection::getSymEntryOffset(const InputFile *f, const Symbol &s,
790 int64_t addend) const {
791 const FileGot &g = gots[*f->mipsGotIndex];
792 Symbol *sym = const_cast<Symbol *>(&s);
793 if (sym->isTls())
794 return g.tls.lookup(sym) * config->wordsize;
795 if (sym->isPreemptible)
796 return g.global.lookup(sym) * config->wordsize;
797 return g.local16.lookup({sym, addend}) * config->wordsize;
798 }
799
getTlsIndexOffset(const InputFile * f) const800 uint64_t MipsGotSection::getTlsIndexOffset(const InputFile *f) const {
801 const FileGot &g = gots[*f->mipsGotIndex];
802 return g.dynTlsSymbols.lookup(nullptr) * config->wordsize;
803 }
804
getGlobalDynOffset(const InputFile * f,const Symbol & s) const805 uint64_t MipsGotSection::getGlobalDynOffset(const InputFile *f,
806 const Symbol &s) const {
807 const FileGot &g = gots[*f->mipsGotIndex];
808 Symbol *sym = const_cast<Symbol *>(&s);
809 return g.dynTlsSymbols.lookup(sym) * config->wordsize;
810 }
811
getFirstGlobalEntry() const812 const Symbol *MipsGotSection::getFirstGlobalEntry() const {
813 if (gots.empty())
814 return nullptr;
815 const FileGot &primGot = gots.front();
816 if (!primGot.global.empty())
817 return primGot.global.front().first;
818 if (!primGot.relocs.empty())
819 return primGot.relocs.front().first;
820 return nullptr;
821 }
822
getLocalEntriesNum() const823 unsigned MipsGotSection::getLocalEntriesNum() const {
824 if (gots.empty())
825 return headerEntriesNum;
826 return headerEntriesNum + gots.front().getPageEntriesNum() +
827 gots.front().local16.size();
828 }
829
tryMergeGots(FileGot & dst,FileGot & src,bool isPrimary)830 bool MipsGotSection::tryMergeGots(FileGot &dst, FileGot &src, bool isPrimary) {
831 FileGot tmp = dst;
832 set_union(tmp.pagesMap, src.pagesMap);
833 set_union(tmp.local16, src.local16);
834 set_union(tmp.global, src.global);
835 set_union(tmp.relocs, src.relocs);
836 set_union(tmp.tls, src.tls);
837 set_union(tmp.dynTlsSymbols, src.dynTlsSymbols);
838
839 size_t count = isPrimary ? headerEntriesNum : 0;
840 count += tmp.getIndexedEntriesNum();
841
842 if (count * config->wordsize > config->mipsGotSize)
843 return false;
844
845 std::swap(tmp, dst);
846 return true;
847 }
848
finalizeContents()849 void MipsGotSection::finalizeContents() { updateAllocSize(); }
850
updateAllocSize()851 bool MipsGotSection::updateAllocSize() {
852 size = headerEntriesNum * config->wordsize;
853 for (const FileGot &g : gots)
854 size += g.getEntriesNum() * config->wordsize;
855 return false;
856 }
857
build()858 void MipsGotSection::build() {
859 if (gots.empty())
860 return;
861
862 std::vector<FileGot> mergedGots(1);
863
864 // For each GOT move non-preemptible symbols from the `Global`
865 // to `Local16` list. Preemptible symbol might become non-preemptible
866 // one if, for example, it gets a related copy relocation.
867 for (FileGot &got : gots) {
868 for (auto &p: got.global)
869 if (!p.first->isPreemptible)
870 got.local16.insert({{p.first, 0}, 0});
871 got.global.remove_if([&](const std::pair<Symbol *, size_t> &p) {
872 return !p.first->isPreemptible;
873 });
874 }
875
876 // For each GOT remove "reloc-only" entry if there is "global"
877 // entry for the same symbol. And add local entries which indexed
878 // using 32-bit value at the end of 16-bit entries.
879 for (FileGot &got : gots) {
880 got.relocs.remove_if([&](const std::pair<Symbol *, size_t> &p) {
881 return got.global.count(p.first);
882 });
883 set_union(got.local16, got.local32);
884 got.local32.clear();
885 }
886
887 // Evaluate number of "reloc-only" entries in the resulting GOT.
888 // To do that put all unique "reloc-only" and "global" entries
889 // from all GOTs to the future primary GOT.
890 FileGot *primGot = &mergedGots.front();
891 for (FileGot &got : gots) {
892 set_union(primGot->relocs, got.global);
893 set_union(primGot->relocs, got.relocs);
894 got.relocs.clear();
895 }
896
897 // Evaluate number of "page" entries in each GOT.
898 for (FileGot &got : gots) {
899 for (std::pair<const OutputSection *, FileGot::PageBlock> &p :
900 got.pagesMap) {
901 const OutputSection *os = p.first;
902 uint64_t secSize = 0;
903 for (BaseCommand *cmd : os->sectionCommands) {
904 if (auto *isd = dyn_cast<InputSectionDescription>(cmd))
905 for (InputSection *isec : isd->sections) {
906 uint64_t off = alignTo(secSize, isec->alignment);
907 secSize = off + isec->getSize();
908 }
909 }
910 p.second.count = getMipsPageCount(secSize);
911 }
912 }
913
914 // Merge GOTs. Try to join as much as possible GOTs but do not exceed
915 // maximum GOT size. At first, try to fill the primary GOT because
916 // the primary GOT can be accessed in the most effective way. If it
917 // is not possible, try to fill the last GOT in the list, and finally
918 // create a new GOT if both attempts failed.
919 for (FileGot &srcGot : gots) {
920 InputFile *file = srcGot.file;
921 if (tryMergeGots(mergedGots.front(), srcGot, true)) {
922 file->mipsGotIndex = 0;
923 } else {
924 // If this is the first time we failed to merge with the primary GOT,
925 // MergedGots.back() will also be the primary GOT. We must make sure not
926 // to try to merge again with isPrimary=false, as otherwise, if the
927 // inputs are just right, we could allow the primary GOT to become 1 or 2
928 // words bigger due to ignoring the header size.
929 if (mergedGots.size() == 1 ||
930 !tryMergeGots(mergedGots.back(), srcGot, false)) {
931 mergedGots.emplace_back();
932 std::swap(mergedGots.back(), srcGot);
933 }
934 file->mipsGotIndex = mergedGots.size() - 1;
935 }
936 }
937 std::swap(gots, mergedGots);
938
939 // Reduce number of "reloc-only" entries in the primary GOT
940 // by subtracting "global" entries in the primary GOT.
941 primGot = &gots.front();
942 primGot->relocs.remove_if([&](const std::pair<Symbol *, size_t> &p) {
943 return primGot->global.count(p.first);
944 });
945
946 // Calculate indexes for each GOT entry.
947 size_t index = headerEntriesNum;
948 for (FileGot &got : gots) {
949 got.startIndex = &got == primGot ? 0 : index;
950 for (std::pair<const OutputSection *, FileGot::PageBlock> &p :
951 got.pagesMap) {
952 // For each output section referenced by GOT page relocations calculate
953 // and save into pagesMap an upper bound of MIPS GOT entries required
954 // to store page addresses of local symbols. We assume the worst case -
955 // each 64kb page of the output section has at least one GOT relocation
956 // against it. And take in account the case when the section intersects
957 // page boundaries.
958 p.second.firstIndex = index;
959 index += p.second.count;
960 }
961 for (auto &p: got.local16)
962 p.second = index++;
963 for (auto &p: got.global)
964 p.second = index++;
965 for (auto &p: got.relocs)
966 p.second = index++;
967 for (auto &p: got.tls)
968 p.second = index++;
969 for (auto &p: got.dynTlsSymbols) {
970 p.second = index;
971 index += 2;
972 }
973 }
974
975 // Update Symbol::gotIndex field to use this
976 // value later in the `sortMipsSymbols` function.
977 for (auto &p : primGot->global)
978 p.first->gotIndex = p.second;
979 for (auto &p : primGot->relocs)
980 p.first->gotIndex = p.second;
981
982 // Create dynamic relocations.
983 for (FileGot &got : gots) {
984 // Create dynamic relocations for TLS entries.
985 for (std::pair<Symbol *, size_t> &p : got.tls) {
986 Symbol *s = p.first;
987 uint64_t offset = p.second * config->wordsize;
988 if (s->isPreemptible)
989 mainPart->relaDyn->addReloc(target->tlsGotRel, this, offset, s);
990 }
991 for (std::pair<Symbol *, size_t> &p : got.dynTlsSymbols) {
992 Symbol *s = p.first;
993 uint64_t offset = p.second * config->wordsize;
994 if (s == nullptr) {
995 if (!config->isPic)
996 continue;
997 mainPart->relaDyn->addReloc(target->tlsModuleIndexRel, this, offset, s);
998 } else {
999 // When building a shared library we still need a dynamic relocation
1000 // for the module index. Therefore only checking for
1001 // S->isPreemptible is not sufficient (this happens e.g. for
1002 // thread-locals that have been marked as local through a linker script)
1003 if (!s->isPreemptible && !config->isPic)
1004 continue;
1005 mainPart->relaDyn->addReloc(target->tlsModuleIndexRel, this, offset, s);
1006 // However, we can skip writing the TLS offset reloc for non-preemptible
1007 // symbols since it is known even in shared libraries
1008 if (!s->isPreemptible)
1009 continue;
1010 offset += config->wordsize;
1011 mainPart->relaDyn->addReloc(target->tlsOffsetRel, this, offset, s);
1012 }
1013 }
1014
1015 // Do not create dynamic relocations for non-TLS
1016 // entries in the primary GOT.
1017 if (&got == primGot)
1018 continue;
1019
1020 // Dynamic relocations for "global" entries.
1021 for (const std::pair<Symbol *, size_t> &p : got.global) {
1022 uint64_t offset = p.second * config->wordsize;
1023 mainPart->relaDyn->addReloc(target->relativeRel, this, offset, p.first);
1024 }
1025 if (!config->isPic)
1026 continue;
1027 // Dynamic relocations for "local" entries in case of PIC.
1028 for (const std::pair<const OutputSection *, FileGot::PageBlock> &l :
1029 got.pagesMap) {
1030 size_t pageCount = l.second.count;
1031 for (size_t pi = 0; pi < pageCount; ++pi) {
1032 uint64_t offset = (l.second.firstIndex + pi) * config->wordsize;
1033 mainPart->relaDyn->addReloc({target->relativeRel, this, offset, l.first,
1034 int64_t(pi * 0x10000)});
1035 }
1036 }
1037 for (const std::pair<GotEntry, size_t> &p : got.local16) {
1038 uint64_t offset = p.second * config->wordsize;
1039 mainPart->relaDyn->addReloc({target->relativeRel, this, offset, true,
1040 p.first.first, p.first.second});
1041 }
1042 }
1043 }
1044
isNeeded() const1045 bool MipsGotSection::isNeeded() const {
1046 // We add the .got section to the result for dynamic MIPS target because
1047 // its address and properties are mentioned in the .dynamic section.
1048 return !config->relocatable;
1049 }
1050
getGp(const InputFile * f) const1051 uint64_t MipsGotSection::getGp(const InputFile *f) const {
1052 // For files without related GOT or files refer a primary GOT
1053 // returns "common" _gp value. For secondary GOTs calculate
1054 // individual _gp values.
1055 if (!f || !f->mipsGotIndex.hasValue() || *f->mipsGotIndex == 0)
1056 return ElfSym::mipsGp->getVA(0);
1057 return getVA() + gots[*f->mipsGotIndex].startIndex * config->wordsize +
1058 0x7ff0;
1059 }
1060
writeTo(uint8_t * buf)1061 void MipsGotSection::writeTo(uint8_t *buf) {
1062 // Set the MSB of the second GOT slot. This is not required by any
1063 // MIPS ABI documentation, though.
1064 //
1065 // There is a comment in glibc saying that "The MSB of got[1] of a
1066 // gnu object is set to identify gnu objects," and in GNU gold it
1067 // says "the second entry will be used by some runtime loaders".
1068 // But how this field is being used is unclear.
1069 //
1070 // We are not really willing to mimic other linkers behaviors
1071 // without understanding why they do that, but because all files
1072 // generated by GNU tools have this special GOT value, and because
1073 // we've been doing this for years, it is probably a safe bet to
1074 // keep doing this for now. We really need to revisit this to see
1075 // if we had to do this.
1076 writeUint(buf + config->wordsize, (uint64_t)1 << (config->wordsize * 8 - 1));
1077 for (const FileGot &g : gots) {
1078 auto write = [&](size_t i, const Symbol *s, int64_t a) {
1079 uint64_t va = a;
1080 if (s)
1081 va = s->getVA(a);
1082 writeUint(buf + i * config->wordsize, va);
1083 };
1084 // Write 'page address' entries to the local part of the GOT.
1085 for (const std::pair<const OutputSection *, FileGot::PageBlock> &l :
1086 g.pagesMap) {
1087 size_t pageCount = l.second.count;
1088 uint64_t firstPageAddr = getMipsPageAddr(l.first->addr);
1089 for (size_t pi = 0; pi < pageCount; ++pi)
1090 write(l.second.firstIndex + pi, nullptr, firstPageAddr + pi * 0x10000);
1091 }
1092 // Local, global, TLS, reloc-only entries.
1093 // If TLS entry has a corresponding dynamic relocations, leave it
1094 // initialized by zero. Write down adjusted TLS symbol's values otherwise.
1095 // To calculate the adjustments use offsets for thread-local storage.
1096 // https://www.linux-mips.org/wiki/NPTL
1097 for (const std::pair<GotEntry, size_t> &p : g.local16)
1098 write(p.second, p.first.first, p.first.second);
1099 // Write VA to the primary GOT only. For secondary GOTs that
1100 // will be done by REL32 dynamic relocations.
1101 if (&g == &gots.front())
1102 for (const std::pair<Symbol *, size_t> &p : g.global)
1103 write(p.second, p.first, 0);
1104 for (const std::pair<Symbol *, size_t> &p : g.relocs)
1105 write(p.second, p.first, 0);
1106 for (const std::pair<Symbol *, size_t> &p : g.tls)
1107 write(p.second, p.first, p.first->isPreemptible ? 0 : -0x7000);
1108 for (const std::pair<Symbol *, size_t> &p : g.dynTlsSymbols) {
1109 if (p.first == nullptr && !config->isPic)
1110 write(p.second, nullptr, 1);
1111 else if (p.first && !p.first->isPreemptible) {
1112 // If we are emitting PIC code with relocations we mustn't write
1113 // anything to the GOT here. When using Elf_Rel relocations the value
1114 // one will be treated as an addend and will cause crashes at runtime
1115 if (!config->isPic)
1116 write(p.second, nullptr, 1);
1117 write(p.second + 1, p.first, -0x8000);
1118 }
1119 }
1120 }
1121 }
1122
1123 // On PowerPC the .plt section is used to hold the table of function addresses
1124 // instead of the .got.plt, and the type is SHT_NOBITS similar to a .bss
1125 // section. I don't know why we have a BSS style type for the section but it is
1126 // consistent across both 64-bit PowerPC ABIs as well as the 32-bit PowerPC ABI.
GotPltSection()1127 GotPltSection::GotPltSection()
1128 : SyntheticSection(SHF_ALLOC | SHF_WRITE, SHT_PROGBITS, config->wordsize,
1129 ".got.plt") {
1130 if (config->emachine == EM_PPC) {
1131 name = ".plt";
1132 } else if (config->emachine == EM_PPC64) {
1133 type = SHT_NOBITS;
1134 name = ".plt";
1135 }
1136 }
1137
addEntry(Symbol & sym)1138 void GotPltSection::addEntry(Symbol &sym) {
1139 assert(sym.pltIndex == entries.size());
1140 entries.push_back(&sym);
1141 }
1142
getSize() const1143 size_t GotPltSection::getSize() const {
1144 return (target->gotPltHeaderEntriesNum + entries.size()) * config->wordsize;
1145 }
1146
writeTo(uint8_t * buf)1147 void GotPltSection::writeTo(uint8_t *buf) {
1148 target->writeGotPltHeader(buf);
1149 buf += target->gotPltHeaderEntriesNum * config->wordsize;
1150 for (const Symbol *b : entries) {
1151 target->writeGotPlt(buf, *b);
1152 buf += config->wordsize;
1153 }
1154 }
1155
isNeeded() const1156 bool GotPltSection::isNeeded() const {
1157 // We need to emit GOTPLT even if it's empty if there's a relocation relative
1158 // to it.
1159 return !entries.empty() || hasGotPltOffRel;
1160 }
1161
getIgotPltName()1162 static StringRef getIgotPltName() {
1163 // On ARM the IgotPltSection is part of the GotSection.
1164 if (config->emachine == EM_ARM)
1165 return ".got";
1166
1167 // On PowerPC64 the GotPltSection is renamed to '.plt' so the IgotPltSection
1168 // needs to be named the same.
1169 if (config->emachine == EM_PPC64)
1170 return ".plt";
1171
1172 return ".got.plt";
1173 }
1174
1175 // On PowerPC64 the GotPltSection type is SHT_NOBITS so we have to follow suit
1176 // with the IgotPltSection.
IgotPltSection()1177 IgotPltSection::IgotPltSection()
1178 : SyntheticSection(SHF_ALLOC | SHF_WRITE,
1179 config->emachine == EM_PPC64 ? SHT_NOBITS : SHT_PROGBITS,
1180 config->wordsize, getIgotPltName()) {}
1181
addEntry(Symbol & sym)1182 void IgotPltSection::addEntry(Symbol &sym) {
1183 assert(sym.pltIndex == entries.size());
1184 entries.push_back(&sym);
1185 }
1186
getSize() const1187 size_t IgotPltSection::getSize() const {
1188 return entries.size() * config->wordsize;
1189 }
1190
writeTo(uint8_t * buf)1191 void IgotPltSection::writeTo(uint8_t *buf) {
1192 for (const Symbol *b : entries) {
1193 target->writeIgotPlt(buf, *b);
1194 buf += config->wordsize;
1195 }
1196 }
1197
StringTableSection(StringRef name,bool dynamic)1198 StringTableSection::StringTableSection(StringRef name, bool dynamic)
1199 : SyntheticSection(dynamic ? (uint64_t)SHF_ALLOC : 0, SHT_STRTAB, 1, name),
1200 dynamic(dynamic) {
1201 // ELF string tables start with a NUL byte.
1202 addString("");
1203 }
1204
1205 // Adds a string to the string table. If `hashIt` is true we hash and check for
1206 // duplicates. It is optional because the name of global symbols are already
1207 // uniqued and hashing them again has a big cost for a small value: uniquing
1208 // them with some other string that happens to be the same.
addString(StringRef s,bool hashIt)1209 unsigned StringTableSection::addString(StringRef s, bool hashIt) {
1210 if (hashIt) {
1211 auto r = stringMap.insert(std::make_pair(s, this->size));
1212 if (!r.second)
1213 return r.first->second;
1214 }
1215 unsigned ret = this->size;
1216 this->size = this->size + s.size() + 1;
1217 strings.push_back(s);
1218 return ret;
1219 }
1220
writeTo(uint8_t * buf)1221 void StringTableSection::writeTo(uint8_t *buf) {
1222 for (StringRef s : strings) {
1223 memcpy(buf, s.data(), s.size());
1224 buf[s.size()] = '\0';
1225 buf += s.size() + 1;
1226 }
1227 }
1228
1229 // Returns the number of entries in .gnu.version_d: the number of
1230 // non-VER_NDX_LOCAL-non-VER_NDX_GLOBAL definitions, plus 1.
1231 // Note that we don't support vd_cnt > 1 yet.
getVerDefNum()1232 static unsigned getVerDefNum() {
1233 return namedVersionDefs().size() + 1;
1234 }
1235
1236 template <class ELFT>
DynamicSection()1237 DynamicSection<ELFT>::DynamicSection()
1238 : SyntheticSection(SHF_ALLOC | SHF_WRITE, SHT_DYNAMIC, config->wordsize,
1239 ".dynamic") {
1240 this->entsize = ELFT::Is64Bits ? 16 : 8;
1241
1242 // .dynamic section is not writable on MIPS and on Fuchsia OS
1243 // which passes -z rodynamic.
1244 // See "Special Section" in Chapter 4 in the following document:
1245 // ftp://www.linux-mips.org/pub/linux/mips/doc/ABI/mipsabi.pdf
1246 if (config->emachine == EM_MIPS || config->zRodynamic)
1247 this->flags = SHF_ALLOC;
1248 }
1249
1250 template <class ELFT>
add(int32_t tag,std::function<uint64_t ()> fn)1251 void DynamicSection<ELFT>::add(int32_t tag, std::function<uint64_t()> fn) {
1252 entries.push_back({tag, fn});
1253 }
1254
1255 template <class ELFT>
addInt(int32_t tag,uint64_t val)1256 void DynamicSection<ELFT>::addInt(int32_t tag, uint64_t val) {
1257 entries.push_back({tag, [=] { return val; }});
1258 }
1259
1260 template <class ELFT>
addInSec(int32_t tag,InputSection * sec)1261 void DynamicSection<ELFT>::addInSec(int32_t tag, InputSection *sec) {
1262 entries.push_back({tag, [=] { return sec->getVA(0); }});
1263 }
1264
1265 template <class ELFT>
addInSecRelative(int32_t tag,InputSection * sec)1266 void DynamicSection<ELFT>::addInSecRelative(int32_t tag, InputSection *sec) {
1267 size_t tagOffset = entries.size() * entsize;
1268 entries.push_back(
1269 {tag, [=] { return sec->getVA(0) - (getVA() + tagOffset); }});
1270 }
1271
1272 template <class ELFT>
addOutSec(int32_t tag,OutputSection * sec)1273 void DynamicSection<ELFT>::addOutSec(int32_t tag, OutputSection *sec) {
1274 entries.push_back({tag, [=] { return sec->addr; }});
1275 }
1276
1277 template <class ELFT>
addSize(int32_t tag,OutputSection * sec)1278 void DynamicSection<ELFT>::addSize(int32_t tag, OutputSection *sec) {
1279 entries.push_back({tag, [=] { return sec->size; }});
1280 }
1281
1282 template <class ELFT>
addSym(int32_t tag,Symbol * sym)1283 void DynamicSection<ELFT>::addSym(int32_t tag, Symbol *sym) {
1284 entries.push_back({tag, [=] { return sym->getVA(); }});
1285 }
1286
1287 // The output section .rela.dyn may include these synthetic sections:
1288 //
1289 // - part.relaDyn
1290 // - in.relaIplt: this is included if in.relaIplt is named .rela.dyn
1291 // - in.relaPlt: this is included if a linker script places .rela.plt inside
1292 // .rela.dyn
1293 //
1294 // DT_RELASZ is the total size of the included sections.
addRelaSz(RelocationBaseSection * relaDyn)1295 static std::function<uint64_t()> addRelaSz(RelocationBaseSection *relaDyn) {
1296 return [=]() {
1297 size_t size = relaDyn->getSize();
1298 if (in.relaIplt->getParent() == relaDyn->getParent())
1299 size += in.relaIplt->getSize();
1300 if (in.relaPlt->getParent() == relaDyn->getParent())
1301 size += in.relaPlt->getSize();
1302 return size;
1303 };
1304 }
1305
1306 // A Linker script may assign the RELA relocation sections to the same
1307 // output section. When this occurs we cannot just use the OutputSection
1308 // Size. Moreover the [DT_JMPREL, DT_JMPREL + DT_PLTRELSZ) is permitted to
1309 // overlap with the [DT_RELA, DT_RELA + DT_RELASZ).
addPltRelSz()1310 static uint64_t addPltRelSz() {
1311 size_t size = in.relaPlt->getSize();
1312 if (in.relaIplt->getParent() == in.relaPlt->getParent() &&
1313 in.relaIplt->name == in.relaPlt->name)
1314 size += in.relaIplt->getSize();
1315 return size;
1316 }
1317
1318 // Add remaining entries to complete .dynamic contents.
finalizeContents()1319 template <class ELFT> void DynamicSection<ELFT>::finalizeContents() {
1320 elf::Partition &part = getPartition();
1321 bool isMain = part.name.empty();
1322
1323 for (StringRef s : config->filterList)
1324 addInt(DT_FILTER, part.dynStrTab->addString(s));
1325 for (StringRef s : config->auxiliaryList)
1326 addInt(DT_AUXILIARY, part.dynStrTab->addString(s));
1327
1328 if (!config->rpath.empty())
1329 addInt(config->enableNewDtags ? DT_RUNPATH : DT_RPATH,
1330 part.dynStrTab->addString(config->rpath));
1331
1332 for (SharedFile *file : sharedFiles)
1333 if (file->isNeeded)
1334 addInt(DT_NEEDED, part.dynStrTab->addString(file->soName));
1335
1336 if (isMain) {
1337 if (!config->soName.empty())
1338 addInt(DT_SONAME, part.dynStrTab->addString(config->soName));
1339 } else {
1340 if (!config->soName.empty())
1341 addInt(DT_NEEDED, part.dynStrTab->addString(config->soName));
1342 addInt(DT_SONAME, part.dynStrTab->addString(part.name));
1343 }
1344
1345 // Set DT_FLAGS and DT_FLAGS_1.
1346 uint32_t dtFlags = 0;
1347 uint32_t dtFlags1 = 0;
1348 if (config->bsymbolic)
1349 dtFlags |= DF_SYMBOLIC;
1350 if (config->zGlobal)
1351 dtFlags1 |= DF_1_GLOBAL;
1352 if (config->zInitfirst)
1353 dtFlags1 |= DF_1_INITFIRST;
1354 if (config->zInterpose)
1355 dtFlags1 |= DF_1_INTERPOSE;
1356 if (config->zNodefaultlib)
1357 dtFlags1 |= DF_1_NODEFLIB;
1358 if (config->zNodelete)
1359 dtFlags1 |= DF_1_NODELETE;
1360 if (config->zNodlopen)
1361 dtFlags1 |= DF_1_NOOPEN;
1362 if (config->pie)
1363 dtFlags1 |= DF_1_PIE;
1364 if (config->zNow) {
1365 dtFlags |= DF_BIND_NOW;
1366 dtFlags1 |= DF_1_NOW;
1367 }
1368 if (config->zOrigin) {
1369 dtFlags |= DF_ORIGIN;
1370 dtFlags1 |= DF_1_ORIGIN;
1371 }
1372 if (!config->zText)
1373 dtFlags |= DF_TEXTREL;
1374 if (config->hasStaticTlsModel)
1375 dtFlags |= DF_STATIC_TLS;
1376
1377 if (dtFlags)
1378 addInt(DT_FLAGS, dtFlags);
1379 if (dtFlags1)
1380 addInt(DT_FLAGS_1, dtFlags1);
1381
1382 // DT_DEBUG is a pointer to debug information used by debuggers at runtime. We
1383 // need it for each process, so we don't write it for DSOs. The loader writes
1384 // the pointer into this entry.
1385 //
1386 // DT_DEBUG is the only .dynamic entry that needs to be written to. Some
1387 // systems (currently only Fuchsia OS) provide other means to give the
1388 // debugger this information. Such systems may choose make .dynamic read-only.
1389 // If the target is such a system (used -z rodynamic) don't write DT_DEBUG.
1390 if (!config->shared && !config->relocatable && !config->zRodynamic)
1391 addInt(DT_DEBUG, 0);
1392
1393 if (OutputSection *sec = part.dynStrTab->getParent())
1394 this->link = sec->sectionIndex;
1395
1396 if (part.relaDyn->isNeeded() ||
1397 (in.relaIplt->isNeeded() &&
1398 part.relaDyn->getParent() == in.relaIplt->getParent())) {
1399 addInSec(part.relaDyn->dynamicTag, part.relaDyn);
1400 entries.push_back({part.relaDyn->sizeDynamicTag, addRelaSz(part.relaDyn)});
1401
1402 bool isRela = config->isRela;
1403 addInt(isRela ? DT_RELAENT : DT_RELENT,
1404 isRela ? sizeof(Elf_Rela) : sizeof(Elf_Rel));
1405
1406 // MIPS dynamic loader does not support RELCOUNT tag.
1407 // The problem is in the tight relation between dynamic
1408 // relocations and GOT. So do not emit this tag on MIPS.
1409 if (config->emachine != EM_MIPS) {
1410 size_t numRelativeRels = part.relaDyn->getRelativeRelocCount();
1411 if (config->zCombreloc && numRelativeRels)
1412 addInt(isRela ? DT_RELACOUNT : DT_RELCOUNT, numRelativeRels);
1413 }
1414 }
1415 if (part.relrDyn && !part.relrDyn->relocs.empty()) {
1416 addInSec(config->useAndroidRelrTags ? DT_ANDROID_RELR : DT_RELR,
1417 part.relrDyn);
1418 addSize(config->useAndroidRelrTags ? DT_ANDROID_RELRSZ : DT_RELRSZ,
1419 part.relrDyn->getParent());
1420 addInt(config->useAndroidRelrTags ? DT_ANDROID_RELRENT : DT_RELRENT,
1421 sizeof(Elf_Relr));
1422 }
1423 // .rel[a].plt section usually consists of two parts, containing plt and
1424 // iplt relocations. It is possible to have only iplt relocations in the
1425 // output. In that case relaPlt is empty and have zero offset, the same offset
1426 // as relaIplt has. And we still want to emit proper dynamic tags for that
1427 // case, so here we always use relaPlt as marker for the beginning of
1428 // .rel[a].plt section.
1429 if (isMain && (in.relaPlt->isNeeded() || in.relaIplt->isNeeded())) {
1430 addInSec(DT_JMPREL, in.relaPlt);
1431 entries.push_back({DT_PLTRELSZ, addPltRelSz});
1432 switch (config->emachine) {
1433 case EM_MIPS:
1434 addInSec(DT_MIPS_PLTGOT, in.gotPlt);
1435 break;
1436 case EM_SPARCV9:
1437 addInSec(DT_PLTGOT, in.plt);
1438 break;
1439 case EM_AARCH64:
1440 if (llvm::find_if(in.relaPlt->relocs, [](const DynamicReloc &r) {
1441 return r.type == target->pltRel &&
1442 r.sym->stOther & STO_AARCH64_VARIANT_PCS;
1443 }) != in.relaPlt->relocs.end())
1444 addInt(DT_AARCH64_VARIANT_PCS, 0);
1445 LLVM_FALLTHROUGH;
1446 default:
1447 addInSec(DT_PLTGOT, in.gotPlt);
1448 break;
1449 }
1450 addInt(DT_PLTREL, config->isRela ? DT_RELA : DT_REL);
1451 }
1452
1453 if (config->emachine == EM_AARCH64) {
1454 if (config->andFeatures & GNU_PROPERTY_AARCH64_FEATURE_1_BTI)
1455 addInt(DT_AARCH64_BTI_PLT, 0);
1456 if (config->zPacPlt)
1457 addInt(DT_AARCH64_PAC_PLT, 0);
1458 }
1459
1460 addInSec(DT_SYMTAB, part.dynSymTab);
1461 addInt(DT_SYMENT, sizeof(Elf_Sym));
1462 addInSec(DT_STRTAB, part.dynStrTab);
1463 addInt(DT_STRSZ, part.dynStrTab->getSize());
1464 if (!config->zText)
1465 addInt(DT_TEXTREL, 0);
1466 if (part.gnuHashTab)
1467 addInSec(DT_GNU_HASH, part.gnuHashTab);
1468 if (part.hashTab)
1469 addInSec(DT_HASH, part.hashTab);
1470
1471 if (isMain) {
1472 if (Out::preinitArray) {
1473 addOutSec(DT_PREINIT_ARRAY, Out::preinitArray);
1474 addSize(DT_PREINIT_ARRAYSZ, Out::preinitArray);
1475 }
1476 if (Out::initArray) {
1477 addOutSec(DT_INIT_ARRAY, Out::initArray);
1478 addSize(DT_INIT_ARRAYSZ, Out::initArray);
1479 }
1480 if (Out::finiArray) {
1481 addOutSec(DT_FINI_ARRAY, Out::finiArray);
1482 addSize(DT_FINI_ARRAYSZ, Out::finiArray);
1483 }
1484
1485 if (Symbol *b = symtab->find(config->init))
1486 if (b->isDefined())
1487 addSym(DT_INIT, b);
1488 if (Symbol *b = symtab->find(config->fini))
1489 if (b->isDefined())
1490 addSym(DT_FINI, b);
1491 }
1492
1493 if (part.verSym && part.verSym->isNeeded())
1494 addInSec(DT_VERSYM, part.verSym);
1495 if (part.verDef && part.verDef->isLive()) {
1496 addInSec(DT_VERDEF, part.verDef);
1497 addInt(DT_VERDEFNUM, getVerDefNum());
1498 }
1499 if (part.verNeed && part.verNeed->isNeeded()) {
1500 addInSec(DT_VERNEED, part.verNeed);
1501 unsigned needNum = 0;
1502 for (SharedFile *f : sharedFiles)
1503 if (!f->vernauxs.empty())
1504 ++needNum;
1505 addInt(DT_VERNEEDNUM, needNum);
1506 }
1507
1508 if (config->emachine == EM_MIPS) {
1509 addInt(DT_MIPS_RLD_VERSION, 1);
1510 addInt(DT_MIPS_FLAGS, RHF_NOTPOT);
1511 addInt(DT_MIPS_BASE_ADDRESS, target->getImageBase());
1512 addInt(DT_MIPS_SYMTABNO, part.dynSymTab->getNumSymbols());
1513
1514 add(DT_MIPS_LOCAL_GOTNO, [] { return in.mipsGot->getLocalEntriesNum(); });
1515
1516 if (const Symbol *b = in.mipsGot->getFirstGlobalEntry())
1517 addInt(DT_MIPS_GOTSYM, b->dynsymIndex);
1518 else
1519 addInt(DT_MIPS_GOTSYM, part.dynSymTab->getNumSymbols());
1520 addInSec(DT_PLTGOT, in.mipsGot);
1521 if (in.mipsRldMap) {
1522 if (!config->pie)
1523 addInSec(DT_MIPS_RLD_MAP, in.mipsRldMap);
1524 // Store the offset to the .rld_map section
1525 // relative to the address of the tag.
1526 addInSecRelative(DT_MIPS_RLD_MAP_REL, in.mipsRldMap);
1527 }
1528 }
1529
1530 // DT_PPC_GOT indicates to glibc Secure PLT is used. If DT_PPC_GOT is absent,
1531 // glibc assumes the old-style BSS PLT layout which we don't support.
1532 if (config->emachine == EM_PPC)
1533 add(DT_PPC_GOT, [] { return in.got->getVA(); });
1534
1535 // Glink dynamic tag is required by the V2 abi if the plt section isn't empty.
1536 if (config->emachine == EM_PPC64 && in.plt->isNeeded()) {
1537 // The Glink tag points to 32 bytes before the first lazy symbol resolution
1538 // stub, which starts directly after the header.
1539 entries.push_back({DT_PPC64_GLINK, [=] {
1540 unsigned offset = target->pltHeaderSize - 32;
1541 return in.plt->getVA(0) + offset;
1542 }});
1543 }
1544
1545 addInt(DT_NULL, 0);
1546
1547 getParent()->link = this->link;
1548 this->size = entries.size() * this->entsize;
1549 }
1550
writeTo(uint8_t * buf)1551 template <class ELFT> void DynamicSection<ELFT>::writeTo(uint8_t *buf) {
1552 auto *p = reinterpret_cast<Elf_Dyn *>(buf);
1553
1554 for (std::pair<int32_t, std::function<uint64_t()>> &kv : entries) {
1555 p->d_tag = kv.first;
1556 p->d_un.d_val = kv.second();
1557 ++p;
1558 }
1559 }
1560
getOffset() const1561 uint64_t DynamicReloc::getOffset() const {
1562 return inputSec->getVA(offsetInSec);
1563 }
1564
computeAddend() const1565 int64_t DynamicReloc::computeAddend() const {
1566 if (useSymVA)
1567 return sym->getVA(addend);
1568 if (!outputSec)
1569 return addend;
1570 // See the comment in the DynamicReloc ctor.
1571 return getMipsPageAddr(outputSec->addr) + addend;
1572 }
1573
getSymIndex(SymbolTableBaseSection * symTab) const1574 uint32_t DynamicReloc::getSymIndex(SymbolTableBaseSection *symTab) const {
1575 if (sym && !useSymVA)
1576 return symTab->getSymbolIndex(sym);
1577 return 0;
1578 }
1579
RelocationBaseSection(StringRef name,uint32_t type,int32_t dynamicTag,int32_t sizeDynamicTag)1580 RelocationBaseSection::RelocationBaseSection(StringRef name, uint32_t type,
1581 int32_t dynamicTag,
1582 int32_t sizeDynamicTag)
1583 : SyntheticSection(SHF_ALLOC, type, config->wordsize, name),
1584 dynamicTag(dynamicTag), sizeDynamicTag(sizeDynamicTag) {}
1585
addReloc(RelType dynType,InputSectionBase * isec,uint64_t offsetInSec,Symbol * sym)1586 void RelocationBaseSection::addReloc(RelType dynType, InputSectionBase *isec,
1587 uint64_t offsetInSec, Symbol *sym) {
1588 addReloc({dynType, isec, offsetInSec, false, sym, 0});
1589 }
1590
addReloc(RelType dynType,InputSectionBase * inputSec,uint64_t offsetInSec,Symbol * sym,int64_t addend,RelExpr expr,RelType type)1591 void RelocationBaseSection::addReloc(RelType dynType,
1592 InputSectionBase *inputSec,
1593 uint64_t offsetInSec, Symbol *sym,
1594 int64_t addend, RelExpr expr,
1595 RelType type) {
1596 // Write the addends to the relocated address if required. We skip
1597 // it if the written value would be zero.
1598 if (config->writeAddends && (expr != R_ADDEND || addend != 0))
1599 inputSec->relocations.push_back({expr, type, offsetInSec, addend, sym});
1600 addReloc({dynType, inputSec, offsetInSec, expr != R_ADDEND, sym, addend});
1601 }
1602
addReloc(const DynamicReloc & reloc)1603 void RelocationBaseSection::addReloc(const DynamicReloc &reloc) {
1604 if (reloc.type == target->relativeRel)
1605 ++numRelativeRelocs;
1606 relocs.push_back(reloc);
1607 }
1608
finalizeContents()1609 void RelocationBaseSection::finalizeContents() {
1610 SymbolTableBaseSection *symTab = getPartition().dynSymTab;
1611
1612 // When linking glibc statically, .rel{,a}.plt contains R_*_IRELATIVE
1613 // relocations due to IFUNC (e.g. strcpy). sh_link will be set to 0 in that
1614 // case.
1615 if (symTab && symTab->getParent())
1616 getParent()->link = symTab->getParent()->sectionIndex;
1617 else
1618 getParent()->link = 0;
1619
1620 if (in.relaPlt == this) {
1621 getParent()->flags |= ELF::SHF_INFO_LINK;
1622 getParent()->info = in.gotPlt->getParent()->sectionIndex;
1623 }
1624 if (in.relaIplt == this) {
1625 getParent()->flags |= ELF::SHF_INFO_LINK;
1626 getParent()->info = in.igotPlt->getParent()->sectionIndex;
1627 }
1628 }
1629
RelrBaseSection()1630 RelrBaseSection::RelrBaseSection()
1631 : SyntheticSection(SHF_ALLOC,
1632 config->useAndroidRelrTags ? SHT_ANDROID_RELR : SHT_RELR,
1633 config->wordsize, ".relr.dyn") {}
1634
1635 template <class ELFT>
encodeDynamicReloc(SymbolTableBaseSection * symTab,typename ELFT::Rela * p,const DynamicReloc & rel)1636 static void encodeDynamicReloc(SymbolTableBaseSection *symTab,
1637 typename ELFT::Rela *p,
1638 const DynamicReloc &rel) {
1639 if (config->isRela)
1640 p->r_addend = rel.computeAddend();
1641 p->r_offset = rel.getOffset();
1642 p->setSymbolAndType(rel.getSymIndex(symTab), rel.type, config->isMips64EL);
1643 }
1644
1645 template <class ELFT>
RelocationSection(StringRef name,bool sort)1646 RelocationSection<ELFT>::RelocationSection(StringRef name, bool sort)
1647 : RelocationBaseSection(name, config->isRela ? SHT_RELA : SHT_REL,
1648 config->isRela ? DT_RELA : DT_REL,
1649 config->isRela ? DT_RELASZ : DT_RELSZ),
1650 sort(sort) {
1651 this->entsize = config->isRela ? sizeof(Elf_Rela) : sizeof(Elf_Rel);
1652 }
1653
writeTo(uint8_t * buf)1654 template <class ELFT> void RelocationSection<ELFT>::writeTo(uint8_t *buf) {
1655 SymbolTableBaseSection *symTab = getPartition().dynSymTab;
1656
1657 // Sort by (!IsRelative,SymIndex,r_offset). DT_REL[A]COUNT requires us to
1658 // place R_*_RELATIVE first. SymIndex is to improve locality, while r_offset
1659 // is to make results easier to read.
1660 if (sort)
1661 llvm::stable_sort(
1662 relocs, [&](const DynamicReloc &a, const DynamicReloc &b) {
1663 return std::make_tuple(a.type != target->relativeRel,
1664 a.getSymIndex(symTab), a.getOffset()) <
1665 std::make_tuple(b.type != target->relativeRel,
1666 b.getSymIndex(symTab), b.getOffset());
1667 });
1668
1669 for (const DynamicReloc &rel : relocs) {
1670 encodeDynamicReloc<ELFT>(symTab, reinterpret_cast<Elf_Rela *>(buf), rel);
1671 buf += config->isRela ? sizeof(Elf_Rela) : sizeof(Elf_Rel);
1672 }
1673 }
1674
1675 template <class ELFT>
AndroidPackedRelocationSection(StringRef name)1676 AndroidPackedRelocationSection<ELFT>::AndroidPackedRelocationSection(
1677 StringRef name)
1678 : RelocationBaseSection(
1679 name, config->isRela ? SHT_ANDROID_RELA : SHT_ANDROID_REL,
1680 config->isRela ? DT_ANDROID_RELA : DT_ANDROID_REL,
1681 config->isRela ? DT_ANDROID_RELASZ : DT_ANDROID_RELSZ) {
1682 this->entsize = 1;
1683 }
1684
1685 template <class ELFT>
updateAllocSize()1686 bool AndroidPackedRelocationSection<ELFT>::updateAllocSize() {
1687 // This function computes the contents of an Android-format packed relocation
1688 // section.
1689 //
1690 // This format compresses relocations by using relocation groups to factor out
1691 // fields that are common between relocations and storing deltas from previous
1692 // relocations in SLEB128 format (which has a short representation for small
1693 // numbers). A good example of a relocation type with common fields is
1694 // R_*_RELATIVE, which is normally used to represent function pointers in
1695 // vtables. In the REL format, each relative relocation has the same r_info
1696 // field, and is only different from other relative relocations in terms of
1697 // the r_offset field. By sorting relocations by offset, grouping them by
1698 // r_info and representing each relocation with only the delta from the
1699 // previous offset, each 8-byte relocation can be compressed to as little as 1
1700 // byte (or less with run-length encoding). This relocation packer was able to
1701 // reduce the size of the relocation section in an Android Chromium DSO from
1702 // 2,911,184 bytes to 174,693 bytes, or 6% of the original size.
1703 //
1704 // A relocation section consists of a header containing the literal bytes
1705 // 'APS2' followed by a sequence of SLEB128-encoded integers. The first two
1706 // elements are the total number of relocations in the section and an initial
1707 // r_offset value. The remaining elements define a sequence of relocation
1708 // groups. Each relocation group starts with a header consisting of the
1709 // following elements:
1710 //
1711 // - the number of relocations in the relocation group
1712 // - flags for the relocation group
1713 // - (if RELOCATION_GROUPED_BY_OFFSET_DELTA_FLAG is set) the r_offset delta
1714 // for each relocation in the group.
1715 // - (if RELOCATION_GROUPED_BY_INFO_FLAG is set) the value of the r_info
1716 // field for each relocation in the group.
1717 // - (if RELOCATION_GROUP_HAS_ADDEND_FLAG and
1718 // RELOCATION_GROUPED_BY_ADDEND_FLAG are set) the r_addend delta for
1719 // each relocation in the group.
1720 //
1721 // Following the relocation group header are descriptions of each of the
1722 // relocations in the group. They consist of the following elements:
1723 //
1724 // - (if RELOCATION_GROUPED_BY_OFFSET_DELTA_FLAG is not set) the r_offset
1725 // delta for this relocation.
1726 // - (if RELOCATION_GROUPED_BY_INFO_FLAG is not set) the value of the r_info
1727 // field for this relocation.
1728 // - (if RELOCATION_GROUP_HAS_ADDEND_FLAG is set and
1729 // RELOCATION_GROUPED_BY_ADDEND_FLAG is not set) the r_addend delta for
1730 // this relocation.
1731
1732 size_t oldSize = relocData.size();
1733
1734 relocData = {'A', 'P', 'S', '2'};
1735 raw_svector_ostream os(relocData);
1736 auto add = [&](int64_t v) { encodeSLEB128(v, os); };
1737
1738 // The format header includes the number of relocations and the initial
1739 // offset (we set this to zero because the first relocation group will
1740 // perform the initial adjustment).
1741 add(relocs.size());
1742 add(0);
1743
1744 std::vector<Elf_Rela> relatives, nonRelatives;
1745
1746 for (const DynamicReloc &rel : relocs) {
1747 Elf_Rela r;
1748 encodeDynamicReloc<ELFT>(getPartition().dynSymTab, &r, rel);
1749
1750 if (r.getType(config->isMips64EL) == target->relativeRel)
1751 relatives.push_back(r);
1752 else
1753 nonRelatives.push_back(r);
1754 }
1755
1756 llvm::sort(relatives, [](const Elf_Rel &a, const Elf_Rel &b) {
1757 return a.r_offset < b.r_offset;
1758 });
1759
1760 // Try to find groups of relative relocations which are spaced one word
1761 // apart from one another. These generally correspond to vtable entries. The
1762 // format allows these groups to be encoded using a sort of run-length
1763 // encoding, but each group will cost 7 bytes in addition to the offset from
1764 // the previous group, so it is only profitable to do this for groups of
1765 // size 8 or larger.
1766 std::vector<Elf_Rela> ungroupedRelatives;
1767 std::vector<std::vector<Elf_Rela>> relativeGroups;
1768 for (auto i = relatives.begin(), e = relatives.end(); i != e;) {
1769 std::vector<Elf_Rela> group;
1770 do {
1771 group.push_back(*i++);
1772 } while (i != e && (i - 1)->r_offset + config->wordsize == i->r_offset);
1773
1774 if (group.size() < 8)
1775 ungroupedRelatives.insert(ungroupedRelatives.end(), group.begin(),
1776 group.end());
1777 else
1778 relativeGroups.emplace_back(std::move(group));
1779 }
1780
1781 // For non-relative relocations, we would like to:
1782 // 1. Have relocations with the same symbol offset to be consecutive, so
1783 // that the runtime linker can speed-up symbol lookup by implementing an
1784 // 1-entry cache.
1785 // 2. Group relocations by r_info to reduce the size of the relocation
1786 // section.
1787 // Since the symbol offset is the high bits in r_info, sorting by r_info
1788 // allows us to do both.
1789 //
1790 // For Rela, we also want to sort by r_addend when r_info is the same. This
1791 // enables us to group by r_addend as well.
1792 llvm::stable_sort(nonRelatives, [](const Elf_Rela &a, const Elf_Rela &b) {
1793 if (a.r_info != b.r_info)
1794 return a.r_info < b.r_info;
1795 if (config->isRela)
1796 return a.r_addend < b.r_addend;
1797 return false;
1798 });
1799
1800 // Group relocations with the same r_info. Note that each group emits a group
1801 // header and that may make the relocation section larger. It is hard to
1802 // estimate the size of a group header as the encoded size of that varies
1803 // based on r_info. However, we can approximate this trade-off by the number
1804 // of values encoded. Each group header contains 3 values, and each relocation
1805 // in a group encodes one less value, as compared to when it is not grouped.
1806 // Therefore, we only group relocations if there are 3 or more of them with
1807 // the same r_info.
1808 //
1809 // For Rela, the addend for most non-relative relocations is zero, and thus we
1810 // can usually get a smaller relocation section if we group relocations with 0
1811 // addend as well.
1812 std::vector<Elf_Rela> ungroupedNonRelatives;
1813 std::vector<std::vector<Elf_Rela>> nonRelativeGroups;
1814 for (auto i = nonRelatives.begin(), e = nonRelatives.end(); i != e;) {
1815 auto j = i + 1;
1816 while (j != e && i->r_info == j->r_info &&
1817 (!config->isRela || i->r_addend == j->r_addend))
1818 ++j;
1819 if (j - i < 3 || (config->isRela && i->r_addend != 0))
1820 ungroupedNonRelatives.insert(ungroupedNonRelatives.end(), i, j);
1821 else
1822 nonRelativeGroups.emplace_back(i, j);
1823 i = j;
1824 }
1825
1826 // Sort ungrouped relocations by offset to minimize the encoded length.
1827 llvm::sort(ungroupedNonRelatives, [](const Elf_Rela &a, const Elf_Rela &b) {
1828 return a.r_offset < b.r_offset;
1829 });
1830
1831 unsigned hasAddendIfRela =
1832 config->isRela ? RELOCATION_GROUP_HAS_ADDEND_FLAG : 0;
1833
1834 uint64_t offset = 0;
1835 uint64_t addend = 0;
1836
1837 // Emit the run-length encoding for the groups of adjacent relative
1838 // relocations. Each group is represented using two groups in the packed
1839 // format. The first is used to set the current offset to the start of the
1840 // group (and also encodes the first relocation), and the second encodes the
1841 // remaining relocations.
1842 for (std::vector<Elf_Rela> &g : relativeGroups) {
1843 // The first relocation in the group.
1844 add(1);
1845 add(RELOCATION_GROUPED_BY_OFFSET_DELTA_FLAG |
1846 RELOCATION_GROUPED_BY_INFO_FLAG | hasAddendIfRela);
1847 add(g[0].r_offset - offset);
1848 add(target->relativeRel);
1849 if (config->isRela) {
1850 add(g[0].r_addend - addend);
1851 addend = g[0].r_addend;
1852 }
1853
1854 // The remaining relocations.
1855 add(g.size() - 1);
1856 add(RELOCATION_GROUPED_BY_OFFSET_DELTA_FLAG |
1857 RELOCATION_GROUPED_BY_INFO_FLAG | hasAddendIfRela);
1858 add(config->wordsize);
1859 add(target->relativeRel);
1860 if (config->isRela) {
1861 for (auto i = g.begin() + 1, e = g.end(); i != e; ++i) {
1862 add(i->r_addend - addend);
1863 addend = i->r_addend;
1864 }
1865 }
1866
1867 offset = g.back().r_offset;
1868 }
1869
1870 // Now the ungrouped relatives.
1871 if (!ungroupedRelatives.empty()) {
1872 add(ungroupedRelatives.size());
1873 add(RELOCATION_GROUPED_BY_INFO_FLAG | hasAddendIfRela);
1874 add(target->relativeRel);
1875 for (Elf_Rela &r : ungroupedRelatives) {
1876 add(r.r_offset - offset);
1877 offset = r.r_offset;
1878 if (config->isRela) {
1879 add(r.r_addend - addend);
1880 addend = r.r_addend;
1881 }
1882 }
1883 }
1884
1885 // Grouped non-relatives.
1886 for (ArrayRef<Elf_Rela> g : nonRelativeGroups) {
1887 add(g.size());
1888 add(RELOCATION_GROUPED_BY_INFO_FLAG);
1889 add(g[0].r_info);
1890 for (const Elf_Rela &r : g) {
1891 add(r.r_offset - offset);
1892 offset = r.r_offset;
1893 }
1894 addend = 0;
1895 }
1896
1897 // Finally the ungrouped non-relative relocations.
1898 if (!ungroupedNonRelatives.empty()) {
1899 add(ungroupedNonRelatives.size());
1900 add(hasAddendIfRela);
1901 for (Elf_Rela &r : ungroupedNonRelatives) {
1902 add(r.r_offset - offset);
1903 offset = r.r_offset;
1904 add(r.r_info);
1905 if (config->isRela) {
1906 add(r.r_addend - addend);
1907 addend = r.r_addend;
1908 }
1909 }
1910 }
1911
1912 // Don't allow the section to shrink; otherwise the size of the section can
1913 // oscillate infinitely.
1914 if (relocData.size() < oldSize)
1915 relocData.append(oldSize - relocData.size(), 0);
1916
1917 // Returns whether the section size changed. We need to keep recomputing both
1918 // section layout and the contents of this section until the size converges
1919 // because changing this section's size can affect section layout, which in
1920 // turn can affect the sizes of the LEB-encoded integers stored in this
1921 // section.
1922 return relocData.size() != oldSize;
1923 }
1924
RelrSection()1925 template <class ELFT> RelrSection<ELFT>::RelrSection() {
1926 this->entsize = config->wordsize;
1927 }
1928
updateAllocSize()1929 template <class ELFT> bool RelrSection<ELFT>::updateAllocSize() {
1930 // This function computes the contents of an SHT_RELR packed relocation
1931 // section.
1932 //
1933 // Proposal for adding SHT_RELR sections to generic-abi is here:
1934 // https://groups.google.com/forum/#!topic/generic-abi/bX460iggiKg
1935 //
1936 // The encoded sequence of Elf64_Relr entries in a SHT_RELR section looks
1937 // like [ AAAAAAAA BBBBBBB1 BBBBBBB1 ... AAAAAAAA BBBBBB1 ... ]
1938 //
1939 // i.e. start with an address, followed by any number of bitmaps. The address
1940 // entry encodes 1 relocation. The subsequent bitmap entries encode up to 63
1941 // relocations each, at subsequent offsets following the last address entry.
1942 //
1943 // The bitmap entries must have 1 in the least significant bit. The assumption
1944 // here is that an address cannot have 1 in lsb. Odd addresses are not
1945 // supported.
1946 //
1947 // Excluding the least significant bit in the bitmap, each non-zero bit in
1948 // the bitmap represents a relocation to be applied to a corresponding machine
1949 // word that follows the base address word. The second least significant bit
1950 // represents the machine word immediately following the initial address, and
1951 // each bit that follows represents the next word, in linear order. As such,
1952 // a single bitmap can encode up to 31 relocations in a 32-bit object, and
1953 // 63 relocations in a 64-bit object.
1954 //
1955 // This encoding has a couple of interesting properties:
1956 // 1. Looking at any entry, it is clear whether it's an address or a bitmap:
1957 // even means address, odd means bitmap.
1958 // 2. Just a simple list of addresses is a valid encoding.
1959
1960 size_t oldSize = relrRelocs.size();
1961 relrRelocs.clear();
1962
1963 // Same as Config->Wordsize but faster because this is a compile-time
1964 // constant.
1965 const size_t wordsize = sizeof(typename ELFT::uint);
1966
1967 // Number of bits to use for the relocation offsets bitmap.
1968 // Must be either 63 or 31.
1969 const size_t nBits = wordsize * 8 - 1;
1970
1971 // Get offsets for all relative relocations and sort them.
1972 std::vector<uint64_t> offsets;
1973 for (const RelativeReloc &rel : relocs)
1974 offsets.push_back(rel.getOffset());
1975 llvm::sort(offsets);
1976
1977 // For each leading relocation, find following ones that can be folded
1978 // as a bitmap and fold them.
1979 for (size_t i = 0, e = offsets.size(); i < e;) {
1980 // Add a leading relocation.
1981 relrRelocs.push_back(Elf_Relr(offsets[i]));
1982 uint64_t base = offsets[i] + wordsize;
1983 ++i;
1984
1985 // Find foldable relocations to construct bitmaps.
1986 while (i < e) {
1987 uint64_t bitmap = 0;
1988
1989 while (i < e) {
1990 uint64_t delta = offsets[i] - base;
1991
1992 // If it is too far, it cannot be folded.
1993 if (delta >= nBits * wordsize)
1994 break;
1995
1996 // If it is not a multiple of wordsize away, it cannot be folded.
1997 if (delta % wordsize)
1998 break;
1999
2000 // Fold it.
2001 bitmap |= 1ULL << (delta / wordsize);
2002 ++i;
2003 }
2004
2005 if (!bitmap)
2006 break;
2007
2008 relrRelocs.push_back(Elf_Relr((bitmap << 1) | 1));
2009 base += nBits * wordsize;
2010 }
2011 }
2012
2013 // Don't allow the section to shrink; otherwise the size of the section can
2014 // oscillate infinitely. Trailing 1s do not decode to more relocations.
2015 if (relrRelocs.size() < oldSize) {
2016 log(".relr.dyn needs " + Twine(oldSize - relrRelocs.size()) +
2017 " padding word(s)");
2018 relrRelocs.resize(oldSize, Elf_Relr(1));
2019 }
2020
2021 return relrRelocs.size() != oldSize;
2022 }
2023
SymbolTableBaseSection(StringTableSection & strTabSec)2024 SymbolTableBaseSection::SymbolTableBaseSection(StringTableSection &strTabSec)
2025 : SyntheticSection(strTabSec.isDynamic() ? (uint64_t)SHF_ALLOC : 0,
2026 strTabSec.isDynamic() ? SHT_DYNSYM : SHT_SYMTAB,
2027 config->wordsize,
2028 strTabSec.isDynamic() ? ".dynsym" : ".symtab"),
2029 strTabSec(strTabSec) {}
2030
2031 // Orders symbols according to their positions in the GOT,
2032 // in compliance with MIPS ABI rules.
2033 // See "Global Offset Table" in Chapter 5 in the following document
2034 // for detailed description:
2035 // ftp://www.linux-mips.org/pub/linux/mips/doc/ABI/mipsabi.pdf
sortMipsSymbols(const SymbolTableEntry & l,const SymbolTableEntry & r)2036 static bool sortMipsSymbols(const SymbolTableEntry &l,
2037 const SymbolTableEntry &r) {
2038 // Sort entries related to non-local preemptible symbols by GOT indexes.
2039 // All other entries go to the beginning of a dynsym in arbitrary order.
2040 if (l.sym->isInGot() && r.sym->isInGot())
2041 return l.sym->gotIndex < r.sym->gotIndex;
2042 if (!l.sym->isInGot() && !r.sym->isInGot())
2043 return false;
2044 return !l.sym->isInGot();
2045 }
2046
finalizeContents()2047 void SymbolTableBaseSection::finalizeContents() {
2048 if (OutputSection *sec = strTabSec.getParent())
2049 getParent()->link = sec->sectionIndex;
2050
2051 if (this->type != SHT_DYNSYM) {
2052 sortSymTabSymbols();
2053 return;
2054 }
2055
2056 // If it is a .dynsym, there should be no local symbols, but we need
2057 // to do a few things for the dynamic linker.
2058
2059 // Section's Info field has the index of the first non-local symbol.
2060 // Because the first symbol entry is a null entry, 1 is the first.
2061 getParent()->info = 1;
2062
2063 if (getPartition().gnuHashTab) {
2064 // NB: It also sorts Symbols to meet the GNU hash table requirements.
2065 getPartition().gnuHashTab->addSymbols(symbols);
2066 } else if (config->emachine == EM_MIPS) {
2067 llvm::stable_sort(symbols, sortMipsSymbols);
2068 }
2069
2070 // Only the main partition's dynsym indexes are stored in the symbols
2071 // themselves. All other partitions use a lookup table.
2072 if (this == mainPart->dynSymTab) {
2073 size_t i = 0;
2074 for (const SymbolTableEntry &s : symbols)
2075 s.sym->dynsymIndex = ++i;
2076 }
2077 }
2078
2079 // The ELF spec requires that all local symbols precede global symbols, so we
2080 // sort symbol entries in this function. (For .dynsym, we don't do that because
2081 // symbols for dynamic linking are inherently all globals.)
2082 //
2083 // Aside from above, we put local symbols in groups starting with the STT_FILE
2084 // symbol. That is convenient for purpose of identifying where are local symbols
2085 // coming from.
sortSymTabSymbols()2086 void SymbolTableBaseSection::sortSymTabSymbols() {
2087 // Move all local symbols before global symbols.
2088 auto e = std::stable_partition(
2089 symbols.begin(), symbols.end(), [](const SymbolTableEntry &s) {
2090 return s.sym->isLocal() || s.sym->computeBinding() == STB_LOCAL;
2091 });
2092 size_t numLocals = e - symbols.begin();
2093 getParent()->info = numLocals + 1;
2094
2095 // We want to group the local symbols by file. For that we rebuild the local
2096 // part of the symbols vector. We do not need to care about the STT_FILE
2097 // symbols, they are already naturally placed first in each group. That
2098 // happens because STT_FILE is always the first symbol in the object and hence
2099 // precede all other local symbols we add for a file.
2100 MapVector<InputFile *, std::vector<SymbolTableEntry>> arr;
2101 for (const SymbolTableEntry &s : llvm::make_range(symbols.begin(), e))
2102 arr[s.sym->file].push_back(s);
2103
2104 auto i = symbols.begin();
2105 for (std::pair<InputFile *, std::vector<SymbolTableEntry>> &p : arr)
2106 for (SymbolTableEntry &entry : p.second)
2107 *i++ = entry;
2108 }
2109
addSymbol(Symbol * b)2110 void SymbolTableBaseSection::addSymbol(Symbol *b) {
2111 // Adding a local symbol to a .dynsym is a bug.
2112 assert(this->type != SHT_DYNSYM || !b->isLocal());
2113
2114 bool hashIt = b->isLocal();
2115 symbols.push_back({b, strTabSec.addString(b->getName(), hashIt)});
2116 }
2117
getSymbolIndex(Symbol * sym)2118 size_t SymbolTableBaseSection::getSymbolIndex(Symbol *sym) {
2119 if (this == mainPart->dynSymTab)
2120 return sym->dynsymIndex;
2121
2122 // Initializes symbol lookup tables lazily. This is used only for -r,
2123 // -emit-relocs and dynsyms in partitions other than the main one.
2124 llvm::call_once(onceFlag, [&] {
2125 symbolIndexMap.reserve(symbols.size());
2126 size_t i = 0;
2127 for (const SymbolTableEntry &e : symbols) {
2128 if (e.sym->type == STT_SECTION)
2129 sectionIndexMap[e.sym->getOutputSection()] = ++i;
2130 else
2131 symbolIndexMap[e.sym] = ++i;
2132 }
2133 });
2134
2135 // Section symbols are mapped based on their output sections
2136 // to maintain their semantics.
2137 if (sym->type == STT_SECTION)
2138 return sectionIndexMap.lookup(sym->getOutputSection());
2139 return symbolIndexMap.lookup(sym);
2140 }
2141
2142 template <class ELFT>
SymbolTableSection(StringTableSection & strTabSec)2143 SymbolTableSection<ELFT>::SymbolTableSection(StringTableSection &strTabSec)
2144 : SymbolTableBaseSection(strTabSec) {
2145 this->entsize = sizeof(Elf_Sym);
2146 }
2147
getCommonSec(Symbol * sym)2148 static BssSection *getCommonSec(Symbol *sym) {
2149 if (!config->defineCommon)
2150 if (auto *d = dyn_cast<Defined>(sym))
2151 return dyn_cast_or_null<BssSection>(d->section);
2152 return nullptr;
2153 }
2154
getSymSectionIndex(Symbol * sym)2155 static uint32_t getSymSectionIndex(Symbol *sym) {
2156 if (getCommonSec(sym))
2157 return SHN_COMMON;
2158 if (!isa<Defined>(sym) || sym->needsPltAddr)
2159 return SHN_UNDEF;
2160 if (const OutputSection *os = sym->getOutputSection())
2161 return os->sectionIndex >= SHN_LORESERVE ? (uint32_t)SHN_XINDEX
2162 : os->sectionIndex;
2163 return SHN_ABS;
2164 }
2165
2166 // Write the internal symbol table contents to the output symbol table.
writeTo(uint8_t * buf)2167 template <class ELFT> void SymbolTableSection<ELFT>::writeTo(uint8_t *buf) {
2168 // The first entry is a null entry as per the ELF spec.
2169 memset(buf, 0, sizeof(Elf_Sym));
2170 buf += sizeof(Elf_Sym);
2171
2172 auto *eSym = reinterpret_cast<Elf_Sym *>(buf);
2173
2174 for (SymbolTableEntry &ent : symbols) {
2175 Symbol *sym = ent.sym;
2176 bool isDefinedHere = type == SHT_SYMTAB || sym->partition == partition;
2177
2178 // Set st_info and st_other.
2179 eSym->st_other = 0;
2180 if (sym->isLocal()) {
2181 eSym->setBindingAndType(STB_LOCAL, sym->type);
2182 } else {
2183 eSym->setBindingAndType(sym->computeBinding(), sym->type);
2184 eSym->setVisibility(sym->visibility);
2185 }
2186
2187 // The 3 most significant bits of st_other are used by OpenPOWER ABI.
2188 // See getPPC64GlobalEntryToLocalEntryOffset() for more details.
2189 if (config->emachine == EM_PPC64)
2190 eSym->st_other |= sym->stOther & 0xe0;
2191 // The most significant bit of st_other is used by AArch64 ABI for the
2192 // variant PCS.
2193 else if (config->emachine == EM_AARCH64)
2194 eSym->st_other |= sym->stOther & STO_AARCH64_VARIANT_PCS;
2195
2196 eSym->st_name = ent.strTabOffset;
2197 if (isDefinedHere)
2198 eSym->st_shndx = getSymSectionIndex(ent.sym);
2199 else
2200 eSym->st_shndx = 0;
2201
2202 // Copy symbol size if it is a defined symbol. st_size is not significant
2203 // for undefined symbols, so whether copying it or not is up to us if that's
2204 // the case. We'll leave it as zero because by not setting a value, we can
2205 // get the exact same outputs for two sets of input files that differ only
2206 // in undefined symbol size in DSOs.
2207 if (eSym->st_shndx == SHN_UNDEF || !isDefinedHere)
2208 eSym->st_size = 0;
2209 else
2210 eSym->st_size = sym->getSize();
2211
2212 // st_value is usually an address of a symbol, but that has a special
2213 // meaning for uninstantiated common symbols (--no-define-common).
2214 if (BssSection *commonSec = getCommonSec(ent.sym))
2215 eSym->st_value = commonSec->alignment;
2216 else if (isDefinedHere)
2217 eSym->st_value = sym->getVA();
2218 else
2219 eSym->st_value = 0;
2220
2221 ++eSym;
2222 }
2223
2224 // On MIPS we need to mark symbol which has a PLT entry and requires
2225 // pointer equality by STO_MIPS_PLT flag. That is necessary to help
2226 // dynamic linker distinguish such symbols and MIPS lazy-binding stubs.
2227 // https://sourceware.org/ml/binutils/2008-07/txt00000.txt
2228 if (config->emachine == EM_MIPS) {
2229 auto *eSym = reinterpret_cast<Elf_Sym *>(buf);
2230
2231 for (SymbolTableEntry &ent : symbols) {
2232 Symbol *sym = ent.sym;
2233 if (sym->isInPlt() && sym->needsPltAddr)
2234 eSym->st_other |= STO_MIPS_PLT;
2235 if (isMicroMips()) {
2236 // We already set the less-significant bit for symbols
2237 // marked by the `STO_MIPS_MICROMIPS` flag and for microMIPS PLT
2238 // records. That allows us to distinguish such symbols in
2239 // the `MIPS<ELFT>::relocate()` routine. Now we should
2240 // clear that bit for non-dynamic symbol table, so tools
2241 // like `objdump` will be able to deal with a correct
2242 // symbol position.
2243 if (sym->isDefined() &&
2244 ((sym->stOther & STO_MIPS_MICROMIPS) || sym->needsPltAddr)) {
2245 if (!strTabSec.isDynamic())
2246 eSym->st_value &= ~1;
2247 eSym->st_other |= STO_MIPS_MICROMIPS;
2248 }
2249 }
2250 if (config->relocatable)
2251 if (auto *d = dyn_cast<Defined>(sym))
2252 if (isMipsPIC<ELFT>(d))
2253 eSym->st_other |= STO_MIPS_PIC;
2254 ++eSym;
2255 }
2256 }
2257 }
2258
SymtabShndxSection()2259 SymtabShndxSection::SymtabShndxSection()
2260 : SyntheticSection(0, SHT_SYMTAB_SHNDX, 4, ".symtab_shndx") {
2261 this->entsize = 4;
2262 }
2263
writeTo(uint8_t * buf)2264 void SymtabShndxSection::writeTo(uint8_t *buf) {
2265 // We write an array of 32 bit values, where each value has 1:1 association
2266 // with an entry in .symtab. If the corresponding entry contains SHN_XINDEX,
2267 // we need to write actual index, otherwise, we must write SHN_UNDEF(0).
2268 buf += 4; // Ignore .symtab[0] entry.
2269 for (const SymbolTableEntry &entry : in.symTab->getSymbols()) {
2270 if (getSymSectionIndex(entry.sym) == SHN_XINDEX)
2271 write32(buf, entry.sym->getOutputSection()->sectionIndex);
2272 buf += 4;
2273 }
2274 }
2275
isNeeded() const2276 bool SymtabShndxSection::isNeeded() const {
2277 // SHT_SYMTAB can hold symbols with section indices values up to
2278 // SHN_LORESERVE. If we need more, we want to use extension SHT_SYMTAB_SHNDX
2279 // section. Problem is that we reveal the final section indices a bit too
2280 // late, and we do not know them here. For simplicity, we just always create
2281 // a .symtab_shndx section when the amount of output sections is huge.
2282 size_t size = 0;
2283 for (BaseCommand *base : script->sectionCommands)
2284 if (isa<OutputSection>(base))
2285 ++size;
2286 return size >= SHN_LORESERVE;
2287 }
2288
finalizeContents()2289 void SymtabShndxSection::finalizeContents() {
2290 getParent()->link = in.symTab->getParent()->sectionIndex;
2291 }
2292
getSize() const2293 size_t SymtabShndxSection::getSize() const {
2294 return in.symTab->getNumSymbols() * 4;
2295 }
2296
2297 // .hash and .gnu.hash sections contain on-disk hash tables that map
2298 // symbol names to their dynamic symbol table indices. Their purpose
2299 // is to help the dynamic linker resolve symbols quickly. If ELF files
2300 // don't have them, the dynamic linker has to do linear search on all
2301 // dynamic symbols, which makes programs slower. Therefore, a .hash
2302 // section is added to a DSO by default. A .gnu.hash is added if you
2303 // give the -hash-style=gnu or -hash-style=both option.
2304 //
2305 // The Unix semantics of resolving dynamic symbols is somewhat expensive.
2306 // Each ELF file has a list of DSOs that the ELF file depends on and a
2307 // list of dynamic symbols that need to be resolved from any of the
2308 // DSOs. That means resolving all dynamic symbols takes O(m)*O(n)
2309 // where m is the number of DSOs and n is the number of dynamic
2310 // symbols. For modern large programs, both m and n are large. So
2311 // making each step faster by using hash tables substantially
2312 // improves time to load programs.
2313 //
2314 // (Note that this is not the only way to design the shared library.
2315 // For instance, the Windows DLL takes a different approach. On
2316 // Windows, each dynamic symbol has a name of DLL from which the symbol
2317 // has to be resolved. That makes the cost of symbol resolution O(n).
2318 // This disables some hacky techniques you can use on Unix such as
2319 // LD_PRELOAD, but this is arguably better semantics than the Unix ones.)
2320 //
2321 // Due to historical reasons, we have two different hash tables, .hash
2322 // and .gnu.hash. They are for the same purpose, and .gnu.hash is a new
2323 // and better version of .hash. .hash is just an on-disk hash table, but
2324 // .gnu.hash has a bloom filter in addition to a hash table to skip
2325 // DSOs very quickly. If you are sure that your dynamic linker knows
2326 // about .gnu.hash, you want to specify -hash-style=gnu. Otherwise, a
2327 // safe bet is to specify -hash-style=both for backward compatibility.
GnuHashTableSection()2328 GnuHashTableSection::GnuHashTableSection()
2329 : SyntheticSection(SHF_ALLOC, SHT_GNU_HASH, config->wordsize, ".gnu.hash") {
2330 }
2331
finalizeContents()2332 void GnuHashTableSection::finalizeContents() {
2333 if (OutputSection *sec = getPartition().dynSymTab->getParent())
2334 getParent()->link = sec->sectionIndex;
2335
2336 // Computes bloom filter size in word size. We want to allocate 12
2337 // bits for each symbol. It must be a power of two.
2338 if (symbols.empty()) {
2339 maskWords = 1;
2340 } else {
2341 uint64_t numBits = symbols.size() * 12;
2342 maskWords = NextPowerOf2(numBits / (config->wordsize * 8));
2343 }
2344
2345 size = 16; // Header
2346 size += config->wordsize * maskWords; // Bloom filter
2347 size += nBuckets * 4; // Hash buckets
2348 size += symbols.size() * 4; // Hash values
2349 }
2350
writeTo(uint8_t * buf)2351 void GnuHashTableSection::writeTo(uint8_t *buf) {
2352 // The output buffer is not guaranteed to be zero-cleared because we pre-
2353 // fill executable sections with trap instructions. This is a precaution
2354 // for that case, which happens only when -no-rosegment is given.
2355 memset(buf, 0, size);
2356
2357 // Write a header.
2358 write32(buf, nBuckets);
2359 write32(buf + 4, getPartition().dynSymTab->getNumSymbols() - symbols.size());
2360 write32(buf + 8, maskWords);
2361 write32(buf + 12, Shift2);
2362 buf += 16;
2363
2364 // Write a bloom filter and a hash table.
2365 writeBloomFilter(buf);
2366 buf += config->wordsize * maskWords;
2367 writeHashTable(buf);
2368 }
2369
2370 // This function writes a 2-bit bloom filter. This bloom filter alone
2371 // usually filters out 80% or more of all symbol lookups [1].
2372 // The dynamic linker uses the hash table only when a symbol is not
2373 // filtered out by a bloom filter.
2374 //
2375 // [1] Ulrich Drepper (2011), "How To Write Shared Libraries" (Ver. 4.1.2),
2376 // p.9, https://www.akkadia.org/drepper/dsohowto.pdf
writeBloomFilter(uint8_t * buf)2377 void GnuHashTableSection::writeBloomFilter(uint8_t *buf) {
2378 unsigned c = config->is64 ? 64 : 32;
2379 for (const Entry &sym : symbols) {
2380 // When C = 64, we choose a word with bits [6:...] and set 1 to two bits in
2381 // the word using bits [0:5] and [26:31].
2382 size_t i = (sym.hash / c) & (maskWords - 1);
2383 uint64_t val = readUint(buf + i * config->wordsize);
2384 val |= uint64_t(1) << (sym.hash % c);
2385 val |= uint64_t(1) << ((sym.hash >> Shift2) % c);
2386 writeUint(buf + i * config->wordsize, val);
2387 }
2388 }
2389
writeHashTable(uint8_t * buf)2390 void GnuHashTableSection::writeHashTable(uint8_t *buf) {
2391 uint32_t *buckets = reinterpret_cast<uint32_t *>(buf);
2392 uint32_t oldBucket = -1;
2393 uint32_t *values = buckets + nBuckets;
2394 for (auto i = symbols.begin(), e = symbols.end(); i != e; ++i) {
2395 // Write a hash value. It represents a sequence of chains that share the
2396 // same hash modulo value. The last element of each chain is terminated by
2397 // LSB 1.
2398 uint32_t hash = i->hash;
2399 bool isLastInChain = (i + 1) == e || i->bucketIdx != (i + 1)->bucketIdx;
2400 hash = isLastInChain ? hash | 1 : hash & ~1;
2401 write32(values++, hash);
2402
2403 if (i->bucketIdx == oldBucket)
2404 continue;
2405 // Write a hash bucket. Hash buckets contain indices in the following hash
2406 // value table.
2407 write32(buckets + i->bucketIdx,
2408 getPartition().dynSymTab->getSymbolIndex(i->sym));
2409 oldBucket = i->bucketIdx;
2410 }
2411 }
2412
hashGnu(StringRef name)2413 static uint32_t hashGnu(StringRef name) {
2414 uint32_t h = 5381;
2415 for (uint8_t c : name)
2416 h = (h << 5) + h + c;
2417 return h;
2418 }
2419
2420 // Add symbols to this symbol hash table. Note that this function
2421 // destructively sort a given vector -- which is needed because
2422 // GNU-style hash table places some sorting requirements.
addSymbols(std::vector<SymbolTableEntry> & v)2423 void GnuHashTableSection::addSymbols(std::vector<SymbolTableEntry> &v) {
2424 // We cannot use 'auto' for Mid because GCC 6.1 cannot deduce
2425 // its type correctly.
2426 std::vector<SymbolTableEntry>::iterator mid =
2427 std::stable_partition(v.begin(), v.end(), [&](const SymbolTableEntry &s) {
2428 return !s.sym->isDefined() || s.sym->partition != partition;
2429 });
2430
2431 // We chose load factor 4 for the on-disk hash table. For each hash
2432 // collision, the dynamic linker will compare a uint32_t hash value.
2433 // Since the integer comparison is quite fast, we believe we can
2434 // make the load factor even larger. 4 is just a conservative choice.
2435 //
2436 // Note that we don't want to create a zero-sized hash table because
2437 // Android loader as of 2018 doesn't like a .gnu.hash containing such
2438 // table. If that's the case, we create a hash table with one unused
2439 // dummy slot.
2440 nBuckets = std::max<size_t>((v.end() - mid) / 4, 1);
2441
2442 if (mid == v.end())
2443 return;
2444
2445 for (SymbolTableEntry &ent : llvm::make_range(mid, v.end())) {
2446 Symbol *b = ent.sym;
2447 uint32_t hash = hashGnu(b->getName());
2448 uint32_t bucketIdx = hash % nBuckets;
2449 symbols.push_back({b, ent.strTabOffset, hash, bucketIdx});
2450 }
2451
2452 llvm::stable_sort(symbols, [](const Entry &l, const Entry &r) {
2453 return l.bucketIdx < r.bucketIdx;
2454 });
2455
2456 v.erase(mid, v.end());
2457 for (const Entry &ent : symbols)
2458 v.push_back({ent.sym, ent.strTabOffset});
2459 }
2460
HashTableSection()2461 HashTableSection::HashTableSection()
2462 : SyntheticSection(SHF_ALLOC, SHT_HASH, 4, ".hash") {
2463 this->entsize = 4;
2464 }
2465
finalizeContents()2466 void HashTableSection::finalizeContents() {
2467 SymbolTableBaseSection *symTab = getPartition().dynSymTab;
2468
2469 if (OutputSection *sec = symTab->getParent())
2470 getParent()->link = sec->sectionIndex;
2471
2472 unsigned numEntries = 2; // nbucket and nchain.
2473 numEntries += symTab->getNumSymbols(); // The chain entries.
2474
2475 // Create as many buckets as there are symbols.
2476 numEntries += symTab->getNumSymbols();
2477 this->size = numEntries * 4;
2478 }
2479
writeTo(uint8_t * buf)2480 void HashTableSection::writeTo(uint8_t *buf) {
2481 SymbolTableBaseSection *symTab = getPartition().dynSymTab;
2482
2483 // See comment in GnuHashTableSection::writeTo.
2484 memset(buf, 0, size);
2485
2486 unsigned numSymbols = symTab->getNumSymbols();
2487
2488 uint32_t *p = reinterpret_cast<uint32_t *>(buf);
2489 write32(p++, numSymbols); // nbucket
2490 write32(p++, numSymbols); // nchain
2491
2492 uint32_t *buckets = p;
2493 uint32_t *chains = p + numSymbols;
2494
2495 for (const SymbolTableEntry &s : symTab->getSymbols()) {
2496 Symbol *sym = s.sym;
2497 StringRef name = sym->getName();
2498 unsigned i = sym->dynsymIndex;
2499 uint32_t hash = hashSysV(name) % numSymbols;
2500 chains[i] = buckets[hash];
2501 write32(buckets + hash, i);
2502 }
2503 }
2504
PltSection()2505 PltSection::PltSection()
2506 : SyntheticSection(SHF_ALLOC | SHF_EXECINSTR, SHT_PROGBITS, 16, ".plt"),
2507 headerSize(target->pltHeaderSize) {
2508 // On PowerPC, this section contains lazy symbol resolvers.
2509 if (config->emachine == EM_PPC64) {
2510 name = ".glink";
2511 alignment = 4;
2512 }
2513
2514 // On x86 when IBT is enabled, this section contains the second PLT (lazy
2515 // symbol resolvers).
2516 if ((config->emachine == EM_386 || config->emachine == EM_X86_64) &&
2517 (config->andFeatures & GNU_PROPERTY_X86_FEATURE_1_IBT))
2518 name = ".plt.sec";
2519
2520 // The PLT needs to be writable on SPARC as the dynamic linker will
2521 // modify the instructions in the PLT entries.
2522 if (config->emachine == EM_SPARCV9)
2523 this->flags |= SHF_WRITE;
2524 }
2525
writeTo(uint8_t * buf)2526 void PltSection::writeTo(uint8_t *buf) {
2527 // At beginning of PLT, we have code to call the dynamic
2528 // linker to resolve dynsyms at runtime. Write such code.
2529 target->writePltHeader(buf);
2530 size_t off = headerSize;
2531
2532 for (const Symbol *sym : entries) {
2533 target->writePlt(buf + off, *sym, getVA() + off);
2534 off += target->pltEntrySize;
2535 }
2536 }
2537
addEntry(Symbol & sym)2538 void PltSection::addEntry(Symbol &sym) {
2539 sym.pltIndex = entries.size();
2540 entries.push_back(&sym);
2541 }
2542
getSize() const2543 size_t PltSection::getSize() const {
2544 return headerSize + entries.size() * target->pltEntrySize;
2545 }
2546
isNeeded() const2547 bool PltSection::isNeeded() const {
2548 // For -z retpolineplt, .iplt needs the .plt header.
2549 return !entries.empty() || (config->zRetpolineplt && in.iplt->isNeeded());
2550 }
2551
2552 // Used by ARM to add mapping symbols in the PLT section, which aid
2553 // disassembly.
addSymbols()2554 void PltSection::addSymbols() {
2555 target->addPltHeaderSymbols(*this);
2556
2557 size_t off = headerSize;
2558 for (size_t i = 0; i < entries.size(); ++i) {
2559 target->addPltSymbols(*this, off);
2560 off += target->pltEntrySize;
2561 }
2562 }
2563
IpltSection()2564 IpltSection::IpltSection()
2565 : SyntheticSection(SHF_ALLOC | SHF_EXECINSTR, SHT_PROGBITS, 16, ".iplt") {
2566 if (config->emachine == EM_PPC || config->emachine == EM_PPC64) {
2567 name = ".glink";
2568 alignment = 4;
2569 }
2570 }
2571
writeTo(uint8_t * buf)2572 void IpltSection::writeTo(uint8_t *buf) {
2573 uint32_t off = 0;
2574 for (const Symbol *sym : entries) {
2575 target->writeIplt(buf + off, *sym, getVA() + off);
2576 off += target->ipltEntrySize;
2577 }
2578 }
2579
getSize() const2580 size_t IpltSection::getSize() const {
2581 return entries.size() * target->ipltEntrySize;
2582 }
2583
addEntry(Symbol & sym)2584 void IpltSection::addEntry(Symbol &sym) {
2585 sym.pltIndex = entries.size();
2586 entries.push_back(&sym);
2587 }
2588
2589 // ARM uses mapping symbols to aid disassembly.
addSymbols()2590 void IpltSection::addSymbols() {
2591 size_t off = 0;
2592 for (size_t i = 0, e = entries.size(); i != e; ++i) {
2593 target->addPltSymbols(*this, off);
2594 off += target->pltEntrySize;
2595 }
2596 }
2597
PPC32GlinkSection()2598 PPC32GlinkSection::PPC32GlinkSection() {
2599 name = ".glink";
2600 alignment = 4;
2601 }
2602
writeTo(uint8_t * buf)2603 void PPC32GlinkSection::writeTo(uint8_t *buf) {
2604 writePPC32GlinkSection(buf, entries.size());
2605 }
2606
getSize() const2607 size_t PPC32GlinkSection::getSize() const {
2608 return headerSize + entries.size() * target->pltEntrySize + footerSize;
2609 }
2610
2611 // This is an x86-only extra PLT section and used only when a security
2612 // enhancement feature called CET is enabled. In this comment, I'll explain what
2613 // the feature is and why we have two PLT sections if CET is enabled.
2614 //
2615 // So, what does CET do? CET introduces a new restriction to indirect jump
2616 // instructions. CET works this way. Assume that CET is enabled. Then, if you
2617 // execute an indirect jump instruction, the processor verifies that a special
2618 // "landing pad" instruction (which is actually a repurposed NOP instruction and
2619 // now called "endbr32" or "endbr64") is at the jump target. If the jump target
2620 // does not start with that instruction, the processor raises an exception
2621 // instead of continuing executing code.
2622 //
2623 // If CET is enabled, the compiler emits endbr to all locations where indirect
2624 // jumps may jump to.
2625 //
2626 // This mechanism makes it extremely hard to transfer the control to a middle of
2627 // a function that is not supporsed to be a indirect jump target, preventing
2628 // certain types of attacks such as ROP or JOP.
2629 //
2630 // Note that the processors in the market as of 2019 don't actually support the
2631 // feature. Only the spec is available at the moment.
2632 //
2633 // Now, I'll explain why we have this extra PLT section for CET.
2634 //
2635 // Since you can indirectly jump to a PLT entry, we have to make PLT entries
2636 // start with endbr. The problem is there's no extra space for endbr (which is 4
2637 // bytes long), as the PLT entry is only 16 bytes long and all bytes are already
2638 // used.
2639 //
2640 // In order to deal with the issue, we split a PLT entry into two PLT entries.
2641 // Remember that each PLT entry contains code to jump to an address read from
2642 // .got.plt AND code to resolve a dynamic symbol lazily. With the 2-PLT scheme,
2643 // the former code is written to .plt.sec, and the latter code is written to
2644 // .plt.
2645 //
2646 // Lazy symbol resolution in the 2-PLT scheme works in the usual way, except
2647 // that the regular .plt is now called .plt.sec and .plt is repurposed to
2648 // contain only code for lazy symbol resolution.
2649 //
2650 // In other words, this is how the 2-PLT scheme works. Application code is
2651 // supposed to jump to .plt.sec to call an external function. Each .plt.sec
2652 // entry contains code to read an address from a corresponding .got.plt entry
2653 // and jump to that address. Addresses in .got.plt initially point to .plt, so
2654 // when an application calls an external function for the first time, the
2655 // control is transferred to a function that resolves a symbol name from
2656 // external shared object files. That function then rewrites a .got.plt entry
2657 // with a resolved address, so that the subsequent function calls directly jump
2658 // to a desired location from .plt.sec.
2659 //
2660 // There is an open question as to whether the 2-PLT scheme was desirable or
2661 // not. We could have simply extended the PLT entry size to 32-bytes to
2662 // accommodate endbr, and that scheme would have been much simpler than the
2663 // 2-PLT scheme. One reason to split PLT was, by doing that, we could keep hot
2664 // code (.plt.sec) from cold code (.plt). But as far as I know no one proved
2665 // that the optimization actually makes a difference.
2666 //
2667 // That said, the 2-PLT scheme is a part of the ABI, debuggers and other tools
2668 // depend on it, so we implement the ABI.
IBTPltSection()2669 IBTPltSection::IBTPltSection()
2670 : SyntheticSection(SHF_ALLOC | SHF_EXECINSTR, SHT_PROGBITS, 16, ".plt") {}
2671
writeTo(uint8_t * buf)2672 void IBTPltSection::writeTo(uint8_t *buf) {
2673 target->writeIBTPlt(buf, in.plt->getNumEntries());
2674 }
2675
getSize() const2676 size_t IBTPltSection::getSize() const {
2677 // 16 is the header size of .plt.
2678 return 16 + in.plt->getNumEntries() * target->pltEntrySize;
2679 }
2680
2681 // The string hash function for .gdb_index.
computeGdbHash(StringRef s)2682 static uint32_t computeGdbHash(StringRef s) {
2683 uint32_t h = 0;
2684 for (uint8_t c : s)
2685 h = h * 67 + toLower(c) - 113;
2686 return h;
2687 }
2688
GdbIndexSection()2689 GdbIndexSection::GdbIndexSection()
2690 : SyntheticSection(0, SHT_PROGBITS, 1, ".gdb_index") {}
2691
2692 // Returns the desired size of an on-disk hash table for a .gdb_index section.
2693 // There's a tradeoff between size and collision rate. We aim 75% utilization.
computeSymtabSize() const2694 size_t GdbIndexSection::computeSymtabSize() const {
2695 return std::max<size_t>(NextPowerOf2(symbols.size() * 4 / 3), 1024);
2696 }
2697
2698 // Compute the output section size.
initOutputSize()2699 void GdbIndexSection::initOutputSize() {
2700 size = sizeof(GdbIndexHeader) + computeSymtabSize() * 8;
2701
2702 for (GdbChunk &chunk : chunks)
2703 size += chunk.compilationUnits.size() * 16 + chunk.addressAreas.size() * 20;
2704
2705 // Add the constant pool size if exists.
2706 if (!symbols.empty()) {
2707 GdbSymbol &sym = symbols.back();
2708 size += sym.nameOff + sym.name.size() + 1;
2709 }
2710 }
2711
readCuList(DWARFContext & dwarf)2712 static std::vector<GdbIndexSection::CuEntry> readCuList(DWARFContext &dwarf) {
2713 std::vector<GdbIndexSection::CuEntry> ret;
2714 for (std::unique_ptr<DWARFUnit> &cu : dwarf.compile_units())
2715 ret.push_back({cu->getOffset(), cu->getLength() + 4});
2716 return ret;
2717 }
2718
2719 static std::vector<GdbIndexSection::AddressEntry>
readAddressAreas(DWARFContext & dwarf,InputSection * sec)2720 readAddressAreas(DWARFContext &dwarf, InputSection *sec) {
2721 std::vector<GdbIndexSection::AddressEntry> ret;
2722
2723 uint32_t cuIdx = 0;
2724 for (std::unique_ptr<DWARFUnit> &cu : dwarf.compile_units()) {
2725 if (Error e = cu->tryExtractDIEsIfNeeded(false)) {
2726 warn(toString(sec) + ": " + toString(std::move(e)));
2727 return {};
2728 }
2729 Expected<DWARFAddressRangesVector> ranges = cu->collectAddressRanges();
2730 if (!ranges) {
2731 warn(toString(sec) + ": " + toString(ranges.takeError()));
2732 return {};
2733 }
2734
2735 ArrayRef<InputSectionBase *> sections = sec->file->getSections();
2736 for (DWARFAddressRange &r : *ranges) {
2737 if (r.SectionIndex == -1ULL)
2738 continue;
2739 // Range list with zero size has no effect.
2740 InputSectionBase *s = sections[r.SectionIndex];
2741 if (s && s != &InputSection::discarded && s->isLive())
2742 if (r.LowPC != r.HighPC)
2743 ret.push_back({cast<InputSection>(s), r.LowPC, r.HighPC, cuIdx});
2744 }
2745 ++cuIdx;
2746 }
2747
2748 return ret;
2749 }
2750
2751 template <class ELFT>
2752 static std::vector<GdbIndexSection::NameAttrEntry>
readPubNamesAndTypes(const LLDDwarfObj<ELFT> & obj,const std::vector<GdbIndexSection::CuEntry> & cus)2753 readPubNamesAndTypes(const LLDDwarfObj<ELFT> &obj,
2754 const std::vector<GdbIndexSection::CuEntry> &cus) {
2755 const LLDDWARFSection &pubNames = obj.getGnuPubnamesSection();
2756 const LLDDWARFSection &pubTypes = obj.getGnuPubtypesSection();
2757
2758 std::vector<GdbIndexSection::NameAttrEntry> ret;
2759 for (const LLDDWARFSection *pub : {&pubNames, &pubTypes}) {
2760 DWARFDataExtractor data(obj, *pub, config->isLE, config->wordsize);
2761 DWARFDebugPubTable table;
2762 table.extract(data, /*GnuStyle=*/true, [&](Error e) {
2763 warn(toString(pub->sec) + ": " + toString(std::move(e)));
2764 });
2765 for (const DWARFDebugPubTable::Set &set : table.getData()) {
2766 // The value written into the constant pool is kind << 24 | cuIndex. As we
2767 // don't know how many compilation units precede this object to compute
2768 // cuIndex, we compute (kind << 24 | cuIndexInThisObject) instead, and add
2769 // the number of preceding compilation units later.
2770 uint32_t i = llvm::partition_point(cus,
2771 [&](GdbIndexSection::CuEntry cu) {
2772 return cu.cuOffset < set.Offset;
2773 }) -
2774 cus.begin();
2775 for (const DWARFDebugPubTable::Entry &ent : set.Entries)
2776 ret.push_back({{ent.Name, computeGdbHash(ent.Name)},
2777 (ent.Descriptor.toBits() << 24) | i});
2778 }
2779 }
2780 return ret;
2781 }
2782
2783 // Create a list of symbols from a given list of symbol names and types
2784 // by uniquifying them by name.
2785 static std::vector<GdbIndexSection::GdbSymbol>
createSymbols(ArrayRef<std::vector<GdbIndexSection::NameAttrEntry>> nameAttrs,const std::vector<GdbIndexSection::GdbChunk> & chunks)2786 createSymbols(ArrayRef<std::vector<GdbIndexSection::NameAttrEntry>> nameAttrs,
2787 const std::vector<GdbIndexSection::GdbChunk> &chunks) {
2788 using GdbSymbol = GdbIndexSection::GdbSymbol;
2789 using NameAttrEntry = GdbIndexSection::NameAttrEntry;
2790
2791 // For each chunk, compute the number of compilation units preceding it.
2792 uint32_t cuIdx = 0;
2793 std::vector<uint32_t> cuIdxs(chunks.size());
2794 for (uint32_t i = 0, e = chunks.size(); i != e; ++i) {
2795 cuIdxs[i] = cuIdx;
2796 cuIdx += chunks[i].compilationUnits.size();
2797 }
2798
2799 // The number of symbols we will handle in this function is of the order
2800 // of millions for very large executables, so we use multi-threading to
2801 // speed it up.
2802 constexpr size_t numShards = 32;
2803 size_t concurrency = PowerOf2Floor(
2804 std::min<size_t>(hardware_concurrency(parallel::strategy.ThreadsRequested)
2805 .compute_thread_count(),
2806 numShards));
2807
2808 // A sharded map to uniquify symbols by name.
2809 std::vector<DenseMap<CachedHashStringRef, size_t>> map(numShards);
2810 size_t shift = 32 - countTrailingZeros(numShards);
2811
2812 // Instantiate GdbSymbols while uniqufying them by name.
2813 std::vector<std::vector<GdbSymbol>> symbols(numShards);
2814 parallelForEachN(0, concurrency, [&](size_t threadId) {
2815 uint32_t i = 0;
2816 for (ArrayRef<NameAttrEntry> entries : nameAttrs) {
2817 for (const NameAttrEntry &ent : entries) {
2818 size_t shardId = ent.name.hash() >> shift;
2819 if ((shardId & (concurrency - 1)) != threadId)
2820 continue;
2821
2822 uint32_t v = ent.cuIndexAndAttrs + cuIdxs[i];
2823 size_t &idx = map[shardId][ent.name];
2824 if (idx) {
2825 symbols[shardId][idx - 1].cuVector.push_back(v);
2826 continue;
2827 }
2828
2829 idx = symbols[shardId].size() + 1;
2830 symbols[shardId].push_back({ent.name, {v}, 0, 0});
2831 }
2832 ++i;
2833 }
2834 });
2835
2836 size_t numSymbols = 0;
2837 for (ArrayRef<GdbSymbol> v : symbols)
2838 numSymbols += v.size();
2839
2840 // The return type is a flattened vector, so we'll copy each vector
2841 // contents to Ret.
2842 std::vector<GdbSymbol> ret;
2843 ret.reserve(numSymbols);
2844 for (std::vector<GdbSymbol> &vec : symbols)
2845 for (GdbSymbol &sym : vec)
2846 ret.push_back(std::move(sym));
2847
2848 // CU vectors and symbol names are adjacent in the output file.
2849 // We can compute their offsets in the output file now.
2850 size_t off = 0;
2851 for (GdbSymbol &sym : ret) {
2852 sym.cuVectorOff = off;
2853 off += (sym.cuVector.size() + 1) * 4;
2854 }
2855 for (GdbSymbol &sym : ret) {
2856 sym.nameOff = off;
2857 off += sym.name.size() + 1;
2858 }
2859
2860 return ret;
2861 }
2862
2863 // Returns a newly-created .gdb_index section.
create()2864 template <class ELFT> GdbIndexSection *GdbIndexSection::create() {
2865 // Collect InputFiles with .debug_info. See the comment in
2866 // LLDDwarfObj<ELFT>::LLDDwarfObj. If we do lightweight parsing in the future,
2867 // note that isec->data() may uncompress the full content, which should be
2868 // parallelized.
2869 SetVector<InputFile *> files;
2870 for (InputSectionBase *s : inputSections) {
2871 InputSection *isec = dyn_cast<InputSection>(s);
2872 if (!isec)
2873 continue;
2874 // .debug_gnu_pub{names,types} are useless in executables.
2875 // They are present in input object files solely for creating
2876 // a .gdb_index. So we can remove them from the output.
2877 if (s->name == ".debug_gnu_pubnames" || s->name == ".debug_gnu_pubtypes")
2878 s->markDead();
2879 else if (isec->name == ".debug_info")
2880 files.insert(isec->file);
2881 }
2882 // Drop .rel[a].debug_gnu_pub{names,types} for --emit-relocs.
2883 llvm::erase_if(inputSections, [](InputSectionBase *s) {
2884 if (auto *isec = dyn_cast<InputSection>(s))
2885 if (InputSectionBase *rel = isec->getRelocatedSection())
2886 return !rel->isLive();
2887 return !s->isLive();
2888 });
2889
2890 std::vector<GdbChunk> chunks(files.size());
2891 std::vector<std::vector<NameAttrEntry>> nameAttrs(files.size());
2892
2893 parallelForEachN(0, files.size(), [&](size_t i) {
2894 // To keep memory usage low, we don't want to keep cached DWARFContext, so
2895 // avoid getDwarf() here.
2896 ObjFile<ELFT> *file = cast<ObjFile<ELFT>>(files[i]);
2897 DWARFContext dwarf(std::make_unique<LLDDwarfObj<ELFT>>(file));
2898 auto &dobj = static_cast<const LLDDwarfObj<ELFT> &>(dwarf.getDWARFObj());
2899
2900 // If the are multiple compile units .debug_info (very rare ld -r --unique),
2901 // this only picks the last one. Other address ranges are lost.
2902 chunks[i].sec = dobj.getInfoSection();
2903 chunks[i].compilationUnits = readCuList(dwarf);
2904 chunks[i].addressAreas = readAddressAreas(dwarf, chunks[i].sec);
2905 nameAttrs[i] = readPubNamesAndTypes<ELFT>(dobj, chunks[i].compilationUnits);
2906 });
2907
2908 auto *ret = make<GdbIndexSection>();
2909 ret->chunks = std::move(chunks);
2910 ret->symbols = createSymbols(nameAttrs, ret->chunks);
2911 ret->initOutputSize();
2912 return ret;
2913 }
2914
writeTo(uint8_t * buf)2915 void GdbIndexSection::writeTo(uint8_t *buf) {
2916 // Write the header.
2917 auto *hdr = reinterpret_cast<GdbIndexHeader *>(buf);
2918 uint8_t *start = buf;
2919 hdr->version = 7;
2920 buf += sizeof(*hdr);
2921
2922 // Write the CU list.
2923 hdr->cuListOff = buf - start;
2924 for (GdbChunk &chunk : chunks) {
2925 for (CuEntry &cu : chunk.compilationUnits) {
2926 write64le(buf, chunk.sec->outSecOff + cu.cuOffset);
2927 write64le(buf + 8, cu.cuLength);
2928 buf += 16;
2929 }
2930 }
2931
2932 // Write the address area.
2933 hdr->cuTypesOff = buf - start;
2934 hdr->addressAreaOff = buf - start;
2935 uint32_t cuOff = 0;
2936 for (GdbChunk &chunk : chunks) {
2937 for (AddressEntry &e : chunk.addressAreas) {
2938 // In the case of ICF there may be duplicate address range entries.
2939 const uint64_t baseAddr = e.section->repl->getVA(0);
2940 write64le(buf, baseAddr + e.lowAddress);
2941 write64le(buf + 8, baseAddr + e.highAddress);
2942 write32le(buf + 16, e.cuIndex + cuOff);
2943 buf += 20;
2944 }
2945 cuOff += chunk.compilationUnits.size();
2946 }
2947
2948 // Write the on-disk open-addressing hash table containing symbols.
2949 hdr->symtabOff = buf - start;
2950 size_t symtabSize = computeSymtabSize();
2951 uint32_t mask = symtabSize - 1;
2952
2953 for (GdbSymbol &sym : symbols) {
2954 uint32_t h = sym.name.hash();
2955 uint32_t i = h & mask;
2956 uint32_t step = ((h * 17) & mask) | 1;
2957
2958 while (read32le(buf + i * 8))
2959 i = (i + step) & mask;
2960
2961 write32le(buf + i * 8, sym.nameOff);
2962 write32le(buf + i * 8 + 4, sym.cuVectorOff);
2963 }
2964
2965 buf += symtabSize * 8;
2966
2967 // Write the string pool.
2968 hdr->constantPoolOff = buf - start;
2969 parallelForEach(symbols, [&](GdbSymbol &sym) {
2970 memcpy(buf + sym.nameOff, sym.name.data(), sym.name.size());
2971 });
2972
2973 // Write the CU vectors.
2974 for (GdbSymbol &sym : symbols) {
2975 write32le(buf, sym.cuVector.size());
2976 buf += 4;
2977 for (uint32_t val : sym.cuVector) {
2978 write32le(buf, val);
2979 buf += 4;
2980 }
2981 }
2982 }
2983
isNeeded() const2984 bool GdbIndexSection::isNeeded() const { return !chunks.empty(); }
2985
EhFrameHeader()2986 EhFrameHeader::EhFrameHeader()
2987 : SyntheticSection(SHF_ALLOC, SHT_PROGBITS, 4, ".eh_frame_hdr") {}
2988
writeTo(uint8_t * buf)2989 void EhFrameHeader::writeTo(uint8_t *buf) {
2990 // Unlike most sections, the EhFrameHeader section is written while writing
2991 // another section, namely EhFrameSection, which calls the write() function
2992 // below from its writeTo() function. This is necessary because the contents
2993 // of EhFrameHeader depend on the relocated contents of EhFrameSection and we
2994 // don't know which order the sections will be written in.
2995 }
2996
2997 // .eh_frame_hdr contains a binary search table of pointers to FDEs.
2998 // Each entry of the search table consists of two values,
2999 // the starting PC from where FDEs covers, and the FDE's address.
3000 // It is sorted by PC.
write()3001 void EhFrameHeader::write() {
3002 uint8_t *buf = Out::bufferStart + getParent()->offset + outSecOff;
3003 using FdeData = EhFrameSection::FdeData;
3004
3005 std::vector<FdeData> fdes = getPartition().ehFrame->getFdeData();
3006
3007 buf[0] = 1;
3008 buf[1] = DW_EH_PE_pcrel | DW_EH_PE_sdata4;
3009 buf[2] = DW_EH_PE_udata4;
3010 buf[3] = DW_EH_PE_datarel | DW_EH_PE_sdata4;
3011 write32(buf + 4,
3012 getPartition().ehFrame->getParent()->addr - this->getVA() - 4);
3013 write32(buf + 8, fdes.size());
3014 buf += 12;
3015
3016 for (FdeData &fde : fdes) {
3017 write32(buf, fde.pcRel);
3018 write32(buf + 4, fde.fdeVARel);
3019 buf += 8;
3020 }
3021 }
3022
getSize() const3023 size_t EhFrameHeader::getSize() const {
3024 // .eh_frame_hdr has a 12 bytes header followed by an array of FDEs.
3025 return 12 + getPartition().ehFrame->numFdes * 8;
3026 }
3027
isNeeded() const3028 bool EhFrameHeader::isNeeded() const {
3029 return isLive() && getPartition().ehFrame->isNeeded();
3030 }
3031
VersionDefinitionSection()3032 VersionDefinitionSection::VersionDefinitionSection()
3033 : SyntheticSection(SHF_ALLOC, SHT_GNU_verdef, sizeof(uint32_t),
3034 ".gnu.version_d") {}
3035
getFileDefName()3036 StringRef VersionDefinitionSection::getFileDefName() {
3037 if (!getPartition().name.empty())
3038 return getPartition().name;
3039 if (!config->soName.empty())
3040 return config->soName;
3041 return config->outputFile;
3042 }
3043
finalizeContents()3044 void VersionDefinitionSection::finalizeContents() {
3045 fileDefNameOff = getPartition().dynStrTab->addString(getFileDefName());
3046 for (const VersionDefinition &v : namedVersionDefs())
3047 verDefNameOffs.push_back(getPartition().dynStrTab->addString(v.name));
3048
3049 if (OutputSection *sec = getPartition().dynStrTab->getParent())
3050 getParent()->link = sec->sectionIndex;
3051
3052 // sh_info should be set to the number of definitions. This fact is missed in
3053 // documentation, but confirmed by binutils community:
3054 // https://sourceware.org/ml/binutils/2014-11/msg00355.html
3055 getParent()->info = getVerDefNum();
3056 }
3057
writeOne(uint8_t * buf,uint32_t index,StringRef name,size_t nameOff)3058 void VersionDefinitionSection::writeOne(uint8_t *buf, uint32_t index,
3059 StringRef name, size_t nameOff) {
3060 uint16_t flags = index == 1 ? VER_FLG_BASE : 0;
3061
3062 // Write a verdef.
3063 write16(buf, 1); // vd_version
3064 write16(buf + 2, flags); // vd_flags
3065 write16(buf + 4, index); // vd_ndx
3066 write16(buf + 6, 1); // vd_cnt
3067 write32(buf + 8, hashSysV(name)); // vd_hash
3068 write32(buf + 12, 20); // vd_aux
3069 write32(buf + 16, 28); // vd_next
3070
3071 // Write a veraux.
3072 write32(buf + 20, nameOff); // vda_name
3073 write32(buf + 24, 0); // vda_next
3074 }
3075
writeTo(uint8_t * buf)3076 void VersionDefinitionSection::writeTo(uint8_t *buf) {
3077 writeOne(buf, 1, getFileDefName(), fileDefNameOff);
3078
3079 auto nameOffIt = verDefNameOffs.begin();
3080 for (const VersionDefinition &v : namedVersionDefs()) {
3081 buf += EntrySize;
3082 writeOne(buf, v.id, v.name, *nameOffIt++);
3083 }
3084
3085 // Need to terminate the last version definition.
3086 write32(buf + 16, 0); // vd_next
3087 }
3088
getSize() const3089 size_t VersionDefinitionSection::getSize() const {
3090 return EntrySize * getVerDefNum();
3091 }
3092
3093 // .gnu.version is a table where each entry is 2 byte long.
VersionTableSection()3094 VersionTableSection::VersionTableSection()
3095 : SyntheticSection(SHF_ALLOC, SHT_GNU_versym, sizeof(uint16_t),
3096 ".gnu.version") {
3097 this->entsize = 2;
3098 }
3099
finalizeContents()3100 void VersionTableSection::finalizeContents() {
3101 // At the moment of june 2016 GNU docs does not mention that sh_link field
3102 // should be set, but Sun docs do. Also readelf relies on this field.
3103 getParent()->link = getPartition().dynSymTab->getParent()->sectionIndex;
3104 }
3105
getSize() const3106 size_t VersionTableSection::getSize() const {
3107 return (getPartition().dynSymTab->getSymbols().size() + 1) * 2;
3108 }
3109
writeTo(uint8_t * buf)3110 void VersionTableSection::writeTo(uint8_t *buf) {
3111 buf += 2;
3112 for (const SymbolTableEntry &s : getPartition().dynSymTab->getSymbols()) {
3113 // Use the original versionId for an unfetched lazy symbol (undefined weak),
3114 // which must be VER_NDX_GLOBAL (an undefined versioned symbol is an error).
3115 write16(buf, s.sym->isLazy() ? VER_NDX_GLOBAL : s.sym->versionId);
3116 buf += 2;
3117 }
3118 }
3119
isNeeded() const3120 bool VersionTableSection::isNeeded() const {
3121 return isLive() &&
3122 (getPartition().verDef || getPartition().verNeed->isNeeded());
3123 }
3124
addVerneed(Symbol * ss)3125 void elf::addVerneed(Symbol *ss) {
3126 auto &file = cast<SharedFile>(*ss->file);
3127 if (ss->verdefIndex == VER_NDX_GLOBAL) {
3128 ss->versionId = VER_NDX_GLOBAL;
3129 return;
3130 }
3131
3132 if (file.vernauxs.empty())
3133 file.vernauxs.resize(file.verdefs.size());
3134
3135 // Select a version identifier for the vernaux data structure, if we haven't
3136 // already allocated one. The verdef identifiers cover the range
3137 // [1..getVerDefNum()]; this causes the vernaux identifiers to start from
3138 // getVerDefNum()+1.
3139 if (file.vernauxs[ss->verdefIndex] == 0)
3140 file.vernauxs[ss->verdefIndex] = ++SharedFile::vernauxNum + getVerDefNum();
3141
3142 ss->versionId = file.vernauxs[ss->verdefIndex];
3143 }
3144
3145 template <class ELFT>
VersionNeedSection()3146 VersionNeedSection<ELFT>::VersionNeedSection()
3147 : SyntheticSection(SHF_ALLOC, SHT_GNU_verneed, sizeof(uint32_t),
3148 ".gnu.version_r") {}
3149
finalizeContents()3150 template <class ELFT> void VersionNeedSection<ELFT>::finalizeContents() {
3151 for (SharedFile *f : sharedFiles) {
3152 if (f->vernauxs.empty())
3153 continue;
3154 verneeds.emplace_back();
3155 Verneed &vn = verneeds.back();
3156 vn.nameStrTab = getPartition().dynStrTab->addString(f->soName);
3157 for (unsigned i = 0; i != f->vernauxs.size(); ++i) {
3158 if (f->vernauxs[i] == 0)
3159 continue;
3160 auto *verdef =
3161 reinterpret_cast<const typename ELFT::Verdef *>(f->verdefs[i]);
3162 vn.vernauxs.push_back(
3163 {verdef->vd_hash, f->vernauxs[i],
3164 getPartition().dynStrTab->addString(f->getStringTable().data() +
3165 verdef->getAux()->vda_name)});
3166 }
3167 }
3168
3169 if (OutputSection *sec = getPartition().dynStrTab->getParent())
3170 getParent()->link = sec->sectionIndex;
3171 getParent()->info = verneeds.size();
3172 }
3173
writeTo(uint8_t * buf)3174 template <class ELFT> void VersionNeedSection<ELFT>::writeTo(uint8_t *buf) {
3175 // The Elf_Verneeds need to appear first, followed by the Elf_Vernauxs.
3176 auto *verneed = reinterpret_cast<Elf_Verneed *>(buf);
3177 auto *vernaux = reinterpret_cast<Elf_Vernaux *>(verneed + verneeds.size());
3178
3179 for (auto &vn : verneeds) {
3180 // Create an Elf_Verneed for this DSO.
3181 verneed->vn_version = 1;
3182 verneed->vn_cnt = vn.vernauxs.size();
3183 verneed->vn_file = vn.nameStrTab;
3184 verneed->vn_aux =
3185 reinterpret_cast<char *>(vernaux) - reinterpret_cast<char *>(verneed);
3186 verneed->vn_next = sizeof(Elf_Verneed);
3187 ++verneed;
3188
3189 // Create the Elf_Vernauxs for this Elf_Verneed.
3190 for (auto &vna : vn.vernauxs) {
3191 vernaux->vna_hash = vna.hash;
3192 vernaux->vna_flags = 0;
3193 vernaux->vna_other = vna.verneedIndex;
3194 vernaux->vna_name = vna.nameStrTab;
3195 vernaux->vna_next = sizeof(Elf_Vernaux);
3196 ++vernaux;
3197 }
3198
3199 vernaux[-1].vna_next = 0;
3200 }
3201 verneed[-1].vn_next = 0;
3202 }
3203
getSize() const3204 template <class ELFT> size_t VersionNeedSection<ELFT>::getSize() const {
3205 return verneeds.size() * sizeof(Elf_Verneed) +
3206 SharedFile::vernauxNum * sizeof(Elf_Vernaux);
3207 }
3208
isNeeded() const3209 template <class ELFT> bool VersionNeedSection<ELFT>::isNeeded() const {
3210 return isLive() && SharedFile::vernauxNum != 0;
3211 }
3212
addSection(MergeInputSection * ms)3213 void MergeSyntheticSection::addSection(MergeInputSection *ms) {
3214 ms->parent = this;
3215 sections.push_back(ms);
3216 assert(alignment == ms->alignment || !(ms->flags & SHF_STRINGS));
3217 alignment = std::max(alignment, ms->alignment);
3218 }
3219
MergeTailSection(StringRef name,uint32_t type,uint64_t flags,uint32_t alignment)3220 MergeTailSection::MergeTailSection(StringRef name, uint32_t type,
3221 uint64_t flags, uint32_t alignment)
3222 : MergeSyntheticSection(name, type, flags, alignment),
3223 builder(StringTableBuilder::RAW, alignment) {}
3224
getSize() const3225 size_t MergeTailSection::getSize() const { return builder.getSize(); }
3226
writeTo(uint8_t * buf)3227 void MergeTailSection::writeTo(uint8_t *buf) { builder.write(buf); }
3228
finalizeContents()3229 void MergeTailSection::finalizeContents() {
3230 // Add all string pieces to the string table builder to create section
3231 // contents.
3232 for (MergeInputSection *sec : sections)
3233 for (size_t i = 0, e = sec->pieces.size(); i != e; ++i)
3234 if (sec->pieces[i].live)
3235 builder.add(sec->getData(i));
3236
3237 // Fix the string table content. After this, the contents will never change.
3238 builder.finalize();
3239
3240 // finalize() fixed tail-optimized strings, so we can now get
3241 // offsets of strings. Get an offset for each string and save it
3242 // to a corresponding SectionPiece for easy access.
3243 for (MergeInputSection *sec : sections)
3244 for (size_t i = 0, e = sec->pieces.size(); i != e; ++i)
3245 if (sec->pieces[i].live)
3246 sec->pieces[i].outputOff = builder.getOffset(sec->getData(i));
3247 }
3248
writeTo(uint8_t * buf)3249 void MergeNoTailSection::writeTo(uint8_t *buf) {
3250 for (size_t i = 0; i < numShards; ++i)
3251 shards[i].write(buf + shardOffsets[i]);
3252 }
3253
3254 // This function is very hot (i.e. it can take several seconds to finish)
3255 // because sometimes the number of inputs is in an order of magnitude of
3256 // millions. So, we use multi-threading.
3257 //
3258 // For any strings S and T, we know S is not mergeable with T if S's hash
3259 // value is different from T's. If that's the case, we can safely put S and
3260 // T into different string builders without worrying about merge misses.
3261 // We do it in parallel.
finalizeContents()3262 void MergeNoTailSection::finalizeContents() {
3263 // Initializes string table builders.
3264 for (size_t i = 0; i < numShards; ++i)
3265 shards.emplace_back(StringTableBuilder::RAW, alignment);
3266
3267 // Concurrency level. Must be a power of 2 to avoid expensive modulo
3268 // operations in the following tight loop.
3269 size_t concurrency = PowerOf2Floor(
3270 std::min<size_t>(hardware_concurrency(parallel::strategy.ThreadsRequested)
3271 .compute_thread_count(),
3272 numShards));
3273
3274 // Add section pieces to the builders.
3275 parallelForEachN(0, concurrency, [&](size_t threadId) {
3276 for (MergeInputSection *sec : sections) {
3277 for (size_t i = 0, e = sec->pieces.size(); i != e; ++i) {
3278 if (!sec->pieces[i].live)
3279 continue;
3280 size_t shardId = getShardId(sec->pieces[i].hash);
3281 if ((shardId & (concurrency - 1)) == threadId)
3282 sec->pieces[i].outputOff = shards[shardId].add(sec->getData(i));
3283 }
3284 }
3285 });
3286
3287 // Compute an in-section offset for each shard.
3288 size_t off = 0;
3289 for (size_t i = 0; i < numShards; ++i) {
3290 shards[i].finalizeInOrder();
3291 if (shards[i].getSize() > 0)
3292 off = alignTo(off, alignment);
3293 shardOffsets[i] = off;
3294 off += shards[i].getSize();
3295 }
3296 size = off;
3297
3298 // So far, section pieces have offsets from beginning of shards, but
3299 // we want offsets from beginning of the whole section. Fix them.
3300 parallelForEach(sections, [&](MergeInputSection *sec) {
3301 for (size_t i = 0, e = sec->pieces.size(); i != e; ++i)
3302 if (sec->pieces[i].live)
3303 sec->pieces[i].outputOff +=
3304 shardOffsets[getShardId(sec->pieces[i].hash)];
3305 });
3306 }
3307
createMergeSynthetic(StringRef name,uint32_t type,uint64_t flags,uint32_t alignment)3308 MergeSyntheticSection *elf::createMergeSynthetic(StringRef name, uint32_t type,
3309 uint64_t flags,
3310 uint32_t alignment) {
3311 bool shouldTailMerge = (flags & SHF_STRINGS) && config->optimize >= 2;
3312 if (shouldTailMerge)
3313 return make<MergeTailSection>(name, type, flags, alignment);
3314 return make<MergeNoTailSection>(name, type, flags, alignment);
3315 }
3316
splitSections()3317 template <class ELFT> void elf::splitSections() {
3318 llvm::TimeTraceScope timeScope("Split sections");
3319 // splitIntoPieces needs to be called on each MergeInputSection
3320 // before calling finalizeContents().
3321 parallelForEach(inputSections, [](InputSectionBase *sec) {
3322 if (auto *s = dyn_cast<MergeInputSection>(sec))
3323 s->splitIntoPieces();
3324 else if (auto *eh = dyn_cast<EhInputSection>(sec))
3325 eh->split<ELFT>();
3326 });
3327 }
3328
MipsRldMapSection()3329 MipsRldMapSection::MipsRldMapSection()
3330 : SyntheticSection(SHF_ALLOC | SHF_WRITE, SHT_PROGBITS, config->wordsize,
3331 ".rld_map") {}
3332
ARMExidxSyntheticSection()3333 ARMExidxSyntheticSection::ARMExidxSyntheticSection()
3334 : SyntheticSection(SHF_ALLOC | SHF_LINK_ORDER, SHT_ARM_EXIDX,
3335 config->wordsize, ".ARM.exidx") {}
3336
findExidxSection(InputSection * isec)3337 static InputSection *findExidxSection(InputSection *isec) {
3338 for (InputSection *d : isec->dependentSections)
3339 if (d->type == SHT_ARM_EXIDX && d->isLive())
3340 return d;
3341 return nullptr;
3342 }
3343
isValidExidxSectionDep(InputSection * isec)3344 static bool isValidExidxSectionDep(InputSection *isec) {
3345 return (isec->flags & SHF_ALLOC) && (isec->flags & SHF_EXECINSTR) &&
3346 isec->getSize() > 0;
3347 }
3348
addSection(InputSection * isec)3349 bool ARMExidxSyntheticSection::addSection(InputSection *isec) {
3350 if (isec->type == SHT_ARM_EXIDX) {
3351 if (InputSection *dep = isec->getLinkOrderDep())
3352 if (isValidExidxSectionDep(dep)) {
3353 exidxSections.push_back(isec);
3354 // Every exidxSection is 8 bytes, we need an estimate of
3355 // size before assignAddresses can be called. Final size
3356 // will only be known after finalize is called.
3357 size += 8;
3358 }
3359 return true;
3360 }
3361
3362 if (isValidExidxSectionDep(isec)) {
3363 executableSections.push_back(isec);
3364 return false;
3365 }
3366
3367 // FIXME: we do not output a relocation section when --emit-relocs is used
3368 // as we do not have relocation sections for linker generated table entries
3369 // and we would have to erase at a late stage relocations from merged entries.
3370 // Given that exception tables are already position independent and a binary
3371 // analyzer could derive the relocations we choose to erase the relocations.
3372 if (config->emitRelocs && isec->type == SHT_REL)
3373 if (InputSectionBase *ex = isec->getRelocatedSection())
3374 if (isa<InputSection>(ex) && ex->type == SHT_ARM_EXIDX)
3375 return true;
3376
3377 return false;
3378 }
3379
3380 // References to .ARM.Extab Sections have bit 31 clear and are not the
3381 // special EXIDX_CANTUNWIND bit-pattern.
isExtabRef(uint32_t unwind)3382 static bool isExtabRef(uint32_t unwind) {
3383 return (unwind & 0x80000000) == 0 && unwind != 0x1;
3384 }
3385
3386 // Return true if the .ARM.exidx section Cur can be merged into the .ARM.exidx
3387 // section Prev, where Cur follows Prev in the table. This can be done if the
3388 // unwinding instructions in Cur are identical to Prev. Linker generated
3389 // EXIDX_CANTUNWIND entries are represented by nullptr as they do not have an
3390 // InputSection.
isDuplicateArmExidxSec(InputSection * prev,InputSection * cur)3391 static bool isDuplicateArmExidxSec(InputSection *prev, InputSection *cur) {
3392
3393 struct ExidxEntry {
3394 ulittle32_t fn;
3395 ulittle32_t unwind;
3396 };
3397 // Get the last table Entry from the previous .ARM.exidx section. If Prev is
3398 // nullptr then it will be a synthesized EXIDX_CANTUNWIND entry.
3399 ExidxEntry prevEntry = {ulittle32_t(0), ulittle32_t(1)};
3400 if (prev)
3401 prevEntry = prev->getDataAs<ExidxEntry>().back();
3402 if (isExtabRef(prevEntry.unwind))
3403 return false;
3404
3405 // We consider the unwind instructions of an .ARM.exidx table entry
3406 // a duplicate if the previous unwind instructions if:
3407 // - Both are the special EXIDX_CANTUNWIND.
3408 // - Both are the same inline unwind instructions.
3409 // We do not attempt to follow and check links into .ARM.extab tables as
3410 // consecutive identical entries are rare and the effort to check that they
3411 // are identical is high.
3412
3413 // If Cur is nullptr then this is synthesized EXIDX_CANTUNWIND entry.
3414 if (cur == nullptr)
3415 return prevEntry.unwind == 1;
3416
3417 for (const ExidxEntry entry : cur->getDataAs<ExidxEntry>())
3418 if (isExtabRef(entry.unwind) || entry.unwind != prevEntry.unwind)
3419 return false;
3420
3421 // All table entries in this .ARM.exidx Section can be merged into the
3422 // previous Section.
3423 return true;
3424 }
3425
3426 // The .ARM.exidx table must be sorted in ascending order of the address of the
3427 // functions the table describes. Optionally duplicate adjacent table entries
3428 // can be removed. At the end of the function the executableSections must be
3429 // sorted in ascending order of address, Sentinel is set to the InputSection
3430 // with the highest address and any InputSections that have mergeable
3431 // .ARM.exidx table entries are removed from it.
finalizeContents()3432 void ARMExidxSyntheticSection::finalizeContents() {
3433 // The executableSections and exidxSections that we use to derive the final
3434 // contents of this SyntheticSection are populated before
3435 // processSectionCommands() and ICF. A /DISCARD/ entry in SECTIONS command or
3436 // ICF may remove executable InputSections and their dependent .ARM.exidx
3437 // section that we recorded earlier.
3438 auto isDiscarded = [](const InputSection *isec) { return !isec->isLive(); };
3439 llvm::erase_if(exidxSections, isDiscarded);
3440 // We need to remove discarded InputSections and InputSections without
3441 // .ARM.exidx sections that if we generated the .ARM.exidx it would be out
3442 // of range.
3443 auto isDiscardedOrOutOfRange = [this](InputSection *isec) {
3444 if (!isec->isLive())
3445 return true;
3446 if (findExidxSection(isec))
3447 return false;
3448 int64_t off = static_cast<int64_t>(isec->getVA() - getVA());
3449 return off != llvm::SignExtend64(off, 31);
3450 };
3451 llvm::erase_if(executableSections, isDiscardedOrOutOfRange);
3452
3453 // Sort the executable sections that may or may not have associated
3454 // .ARM.exidx sections by order of ascending address. This requires the
3455 // relative positions of InputSections and OutputSections to be known.
3456 auto compareByFilePosition = [](const InputSection *a,
3457 const InputSection *b) {
3458 OutputSection *aOut = a->getParent();
3459 OutputSection *bOut = b->getParent();
3460
3461 if (aOut != bOut)
3462 return aOut->addr < bOut->addr;
3463 return a->outSecOff < b->outSecOff;
3464 };
3465 llvm::stable_sort(executableSections, compareByFilePosition);
3466 sentinel = executableSections.back();
3467 // Optionally merge adjacent duplicate entries.
3468 if (config->mergeArmExidx) {
3469 std::vector<InputSection *> selectedSections;
3470 selectedSections.reserve(executableSections.size());
3471 selectedSections.push_back(executableSections[0]);
3472 size_t prev = 0;
3473 for (size_t i = 1; i < executableSections.size(); ++i) {
3474 InputSection *ex1 = findExidxSection(executableSections[prev]);
3475 InputSection *ex2 = findExidxSection(executableSections[i]);
3476 if (!isDuplicateArmExidxSec(ex1, ex2)) {
3477 selectedSections.push_back(executableSections[i]);
3478 prev = i;
3479 }
3480 }
3481 executableSections = std::move(selectedSections);
3482 }
3483
3484 size_t offset = 0;
3485 size = 0;
3486 for (InputSection *isec : executableSections) {
3487 if (InputSection *d = findExidxSection(isec)) {
3488 d->outSecOff = offset;
3489 d->parent = getParent();
3490 offset += d->getSize();
3491 } else {
3492 offset += 8;
3493 }
3494 }
3495 // Size includes Sentinel.
3496 size = offset + 8;
3497 }
3498
getLinkOrderDep() const3499 InputSection *ARMExidxSyntheticSection::getLinkOrderDep() const {
3500 return executableSections.front();
3501 }
3502
3503 // To write the .ARM.exidx table from the ExecutableSections we have three cases
3504 // 1.) The InputSection has a .ARM.exidx InputSection in its dependent sections.
3505 // We write the .ARM.exidx section contents and apply its relocations.
3506 // 2.) The InputSection does not have a dependent .ARM.exidx InputSection. We
3507 // must write the contents of an EXIDX_CANTUNWIND directly. We use the
3508 // start of the InputSection as the purpose of the linker generated
3509 // section is to terminate the address range of the previous entry.
3510 // 3.) A trailing EXIDX_CANTUNWIND sentinel section is required at the end of
3511 // the table to terminate the address range of the final entry.
writeTo(uint8_t * buf)3512 void ARMExidxSyntheticSection::writeTo(uint8_t *buf) {
3513
3514 const uint8_t cantUnwindData[8] = {0, 0, 0, 0, // PREL31 to target
3515 1, 0, 0, 0}; // EXIDX_CANTUNWIND
3516
3517 uint64_t offset = 0;
3518 for (InputSection *isec : executableSections) {
3519 assert(isec->getParent() != nullptr);
3520 if (InputSection *d = findExidxSection(isec)) {
3521 memcpy(buf + offset, d->data().data(), d->data().size());
3522 d->relocateAlloc(buf + d->outSecOff, buf + d->outSecOff + d->getSize());
3523 offset += d->getSize();
3524 } else {
3525 // A Linker generated CANTUNWIND section.
3526 memcpy(buf + offset, cantUnwindData, sizeof(cantUnwindData));
3527 uint64_t s = isec->getVA();
3528 uint64_t p = getVA() + offset;
3529 target->relocateNoSym(buf + offset, R_ARM_PREL31, s - p);
3530 offset += 8;
3531 }
3532 }
3533 // Write Sentinel.
3534 memcpy(buf + offset, cantUnwindData, sizeof(cantUnwindData));
3535 uint64_t s = sentinel->getVA(sentinel->getSize());
3536 uint64_t p = getVA() + offset;
3537 target->relocateNoSym(buf + offset, R_ARM_PREL31, s - p);
3538 assert(size == offset + 8);
3539 }
3540
isNeeded() const3541 bool ARMExidxSyntheticSection::isNeeded() const {
3542 return llvm::find_if(exidxSections, [](InputSection *isec) {
3543 return isec->isLive();
3544 }) != exidxSections.end();
3545 }
3546
classof(const SectionBase * d)3547 bool ARMExidxSyntheticSection::classof(const SectionBase *d) {
3548 return d->kind() == InputSectionBase::Synthetic && d->type == SHT_ARM_EXIDX;
3549 }
3550
ThunkSection(OutputSection * os,uint64_t off)3551 ThunkSection::ThunkSection(OutputSection *os, uint64_t off)
3552 : SyntheticSection(SHF_ALLOC | SHF_EXECINSTR, SHT_PROGBITS,
3553 config->emachine == EM_PPC64 ? 16 : 4, ".text.thunk") {
3554 this->parent = os;
3555 this->outSecOff = off;
3556 }
3557
getSize() const3558 size_t ThunkSection::getSize() const {
3559 if (roundUpSizeForErrata)
3560 return alignTo(size, 4096);
3561 return size;
3562 }
3563
addThunk(Thunk * t)3564 void ThunkSection::addThunk(Thunk *t) {
3565 thunks.push_back(t);
3566 t->addSymbols(*this);
3567 }
3568
writeTo(uint8_t * buf)3569 void ThunkSection::writeTo(uint8_t *buf) {
3570 for (Thunk *t : thunks)
3571 t->writeTo(buf + t->offset);
3572 }
3573
getTargetInputSection() const3574 InputSection *ThunkSection::getTargetInputSection() const {
3575 if (thunks.empty())
3576 return nullptr;
3577 const Thunk *t = thunks.front();
3578 return t->getTargetInputSection();
3579 }
3580
assignOffsets()3581 bool ThunkSection::assignOffsets() {
3582 uint64_t off = 0;
3583 for (Thunk *t : thunks) {
3584 off = alignTo(off, t->alignment);
3585 t->setOffset(off);
3586 uint32_t size = t->size();
3587 t->getThunkTargetSym()->size = size;
3588 off += size;
3589 }
3590 bool changed = off != size;
3591 size = off;
3592 return changed;
3593 }
3594
PPC32Got2Section()3595 PPC32Got2Section::PPC32Got2Section()
3596 : SyntheticSection(SHF_ALLOC | SHF_WRITE, SHT_PROGBITS, 4, ".got2") {}
3597
isNeeded() const3598 bool PPC32Got2Section::isNeeded() const {
3599 // See the comment below. This is not needed if there is no other
3600 // InputSection.
3601 for (BaseCommand *base : getParent()->sectionCommands)
3602 if (auto *isd = dyn_cast<InputSectionDescription>(base))
3603 for (InputSection *isec : isd->sections)
3604 if (isec != this)
3605 return true;
3606 return false;
3607 }
3608
finalizeContents()3609 void PPC32Got2Section::finalizeContents() {
3610 // PPC32 may create multiple GOT sections for -fPIC/-fPIE, one per file in
3611 // .got2 . This function computes outSecOff of each .got2 to be used in
3612 // PPC32PltCallStub::writeTo(). The purpose of this empty synthetic section is
3613 // to collect input sections named ".got2".
3614 uint32_t offset = 0;
3615 for (BaseCommand *base : getParent()->sectionCommands)
3616 if (auto *isd = dyn_cast<InputSectionDescription>(base)) {
3617 for (InputSection *isec : isd->sections) {
3618 if (isec == this)
3619 continue;
3620 isec->file->ppc32Got2OutSecOff = offset;
3621 offset += (uint32_t)isec->getSize();
3622 }
3623 }
3624 }
3625
3626 // If linking position-dependent code then the table will store the addresses
3627 // directly in the binary so the section has type SHT_PROGBITS. If linking
3628 // position-independent code the section has type SHT_NOBITS since it will be
3629 // allocated and filled in by the dynamic linker.
PPC64LongBranchTargetSection()3630 PPC64LongBranchTargetSection::PPC64LongBranchTargetSection()
3631 : SyntheticSection(SHF_ALLOC | SHF_WRITE,
3632 config->isPic ? SHT_NOBITS : SHT_PROGBITS, 8,
3633 ".branch_lt") {}
3634
getEntryVA(const Symbol * sym,int64_t addend)3635 uint64_t PPC64LongBranchTargetSection::getEntryVA(const Symbol *sym,
3636 int64_t addend) {
3637 return getVA() + entry_index.find({sym, addend})->second * 8;
3638 }
3639
addEntry(const Symbol * sym,int64_t addend)3640 Optional<uint32_t> PPC64LongBranchTargetSection::addEntry(const Symbol *sym,
3641 int64_t addend) {
3642 auto res =
3643 entry_index.try_emplace(std::make_pair(sym, addend), entries.size());
3644 if (!res.second)
3645 return None;
3646 entries.emplace_back(sym, addend);
3647 return res.first->second;
3648 }
3649
getSize() const3650 size_t PPC64LongBranchTargetSection::getSize() const {
3651 return entries.size() * 8;
3652 }
3653
writeTo(uint8_t * buf)3654 void PPC64LongBranchTargetSection::writeTo(uint8_t *buf) {
3655 // If linking non-pic we have the final addresses of the targets and they get
3656 // written to the table directly. For pic the dynamic linker will allocate
3657 // the section and fill it it.
3658 if (config->isPic)
3659 return;
3660
3661 for (auto entry : entries) {
3662 const Symbol *sym = entry.first;
3663 int64_t addend = entry.second;
3664 assert(sym->getVA());
3665 // Need calls to branch to the local entry-point since a long-branch
3666 // must be a local-call.
3667 write64(buf, sym->getVA(addend) +
3668 getPPC64GlobalEntryToLocalEntryOffset(sym->stOther));
3669 buf += 8;
3670 }
3671 }
3672
isNeeded() const3673 bool PPC64LongBranchTargetSection::isNeeded() const {
3674 // `removeUnusedSyntheticSections()` is called before thunk allocation which
3675 // is too early to determine if this section will be empty or not. We need
3676 // Finalized to keep the section alive until after thunk creation. Finalized
3677 // only gets set to true once `finalizeSections()` is called after thunk
3678 // creation. Because of this, if we don't create any long-branch thunks we end
3679 // up with an empty .branch_lt section in the binary.
3680 return !finalized || !entries.empty();
3681 }
3682
getAbiVersion()3683 static uint8_t getAbiVersion() {
3684 // MIPS non-PIC executable gets ABI version 1.
3685 if (config->emachine == EM_MIPS) {
3686 if (!config->isPic && !config->relocatable &&
3687 (config->eflags & (EF_MIPS_PIC | EF_MIPS_CPIC)) == EF_MIPS_CPIC)
3688 return 1;
3689 return 0;
3690 }
3691
3692 if (config->emachine == EM_AMDGPU) {
3693 uint8_t ver = objectFiles[0]->abiVersion;
3694 for (InputFile *file : makeArrayRef(objectFiles).slice(1))
3695 if (file->abiVersion != ver)
3696 error("incompatible ABI version: " + toString(file));
3697 return ver;
3698 }
3699
3700 return 0;
3701 }
3702
writeEhdr(uint8_t * buf,Partition & part)3703 template <typename ELFT> void elf::writeEhdr(uint8_t *buf, Partition &part) {
3704 // For executable segments, the trap instructions are written before writing
3705 // the header. Setting Elf header bytes to zero ensures that any unused bytes
3706 // in header are zero-cleared, instead of having trap instructions.
3707 memset(buf, 0, sizeof(typename ELFT::Ehdr));
3708 memcpy(buf, "\177ELF", 4);
3709
3710 auto *eHdr = reinterpret_cast<typename ELFT::Ehdr *>(buf);
3711 eHdr->e_ident[EI_CLASS] = config->is64 ? ELFCLASS64 : ELFCLASS32;
3712 eHdr->e_ident[EI_DATA] = config->isLE ? ELFDATA2LSB : ELFDATA2MSB;
3713 eHdr->e_ident[EI_VERSION] = EV_CURRENT;
3714 eHdr->e_ident[EI_OSABI] = config->osabi;
3715 eHdr->e_ident[EI_ABIVERSION] = getAbiVersion();
3716 eHdr->e_machine = config->emachine;
3717 eHdr->e_version = EV_CURRENT;
3718 eHdr->e_flags = config->eflags;
3719 eHdr->e_ehsize = sizeof(typename ELFT::Ehdr);
3720 eHdr->e_phnum = part.phdrs.size();
3721 eHdr->e_shentsize = sizeof(typename ELFT::Shdr);
3722
3723 if (!config->relocatable) {
3724 eHdr->e_phoff = sizeof(typename ELFT::Ehdr);
3725 eHdr->e_phentsize = sizeof(typename ELFT::Phdr);
3726 }
3727 }
3728
writePhdrs(uint8_t * buf,Partition & part)3729 template <typename ELFT> void elf::writePhdrs(uint8_t *buf, Partition &part) {
3730 // Write the program header table.
3731 auto *hBuf = reinterpret_cast<typename ELFT::Phdr *>(buf);
3732 for (PhdrEntry *p : part.phdrs) {
3733 hBuf->p_type = p->p_type;
3734 hBuf->p_flags = p->p_flags;
3735 hBuf->p_offset = p->p_offset;
3736 hBuf->p_vaddr = p->p_vaddr;
3737 hBuf->p_paddr = p->p_paddr;
3738 hBuf->p_filesz = p->p_filesz;
3739 hBuf->p_memsz = p->p_memsz;
3740 hBuf->p_align = p->p_align;
3741 ++hBuf;
3742 }
3743 }
3744
3745 template <typename ELFT>
PartitionElfHeaderSection()3746 PartitionElfHeaderSection<ELFT>::PartitionElfHeaderSection()
3747 : SyntheticSection(SHF_ALLOC, SHT_LLVM_PART_EHDR, 1, "") {}
3748
3749 template <typename ELFT>
getSize() const3750 size_t PartitionElfHeaderSection<ELFT>::getSize() const {
3751 return sizeof(typename ELFT::Ehdr);
3752 }
3753
3754 template <typename ELFT>
writeTo(uint8_t * buf)3755 void PartitionElfHeaderSection<ELFT>::writeTo(uint8_t *buf) {
3756 writeEhdr<ELFT>(buf, getPartition());
3757
3758 // Loadable partitions are always ET_DYN.
3759 auto *eHdr = reinterpret_cast<typename ELFT::Ehdr *>(buf);
3760 eHdr->e_type = ET_DYN;
3761 }
3762
3763 template <typename ELFT>
PartitionProgramHeadersSection()3764 PartitionProgramHeadersSection<ELFT>::PartitionProgramHeadersSection()
3765 : SyntheticSection(SHF_ALLOC, SHT_LLVM_PART_PHDR, 1, ".phdrs") {}
3766
3767 template <typename ELFT>
getSize() const3768 size_t PartitionProgramHeadersSection<ELFT>::getSize() const {
3769 return sizeof(typename ELFT::Phdr) * getPartition().phdrs.size();
3770 }
3771
3772 template <typename ELFT>
writeTo(uint8_t * buf)3773 void PartitionProgramHeadersSection<ELFT>::writeTo(uint8_t *buf) {
3774 writePhdrs<ELFT>(buf, getPartition());
3775 }
3776
PartitionIndexSection()3777 PartitionIndexSection::PartitionIndexSection()
3778 : SyntheticSection(SHF_ALLOC, SHT_PROGBITS, 4, ".rodata") {}
3779
getSize() const3780 size_t PartitionIndexSection::getSize() const {
3781 return 12 * (partitions.size() - 1);
3782 }
3783
finalizeContents()3784 void PartitionIndexSection::finalizeContents() {
3785 for (size_t i = 1; i != partitions.size(); ++i)
3786 partitions[i].nameStrTab = mainPart->dynStrTab->addString(partitions[i].name);
3787 }
3788
writeTo(uint8_t * buf)3789 void PartitionIndexSection::writeTo(uint8_t *buf) {
3790 uint64_t va = getVA();
3791 for (size_t i = 1; i != partitions.size(); ++i) {
3792 write32(buf, mainPart->dynStrTab->getVA() + partitions[i].nameStrTab - va);
3793 write32(buf + 4, partitions[i].elfHeader->getVA() - (va + 4));
3794
3795 SyntheticSection *next =
3796 i == partitions.size() - 1 ? in.partEnd : partitions[i + 1].elfHeader;
3797 write32(buf + 8, next->getVA() - partitions[i].elfHeader->getVA());
3798
3799 va += 12;
3800 buf += 12;
3801 }
3802 }
3803
3804 InStruct elf::in;
3805
3806 std::vector<Partition> elf::partitions;
3807 Partition *elf::mainPart;
3808
3809 template GdbIndexSection *GdbIndexSection::create<ELF32LE>();
3810 template GdbIndexSection *GdbIndexSection::create<ELF32BE>();
3811 template GdbIndexSection *GdbIndexSection::create<ELF64LE>();
3812 template GdbIndexSection *GdbIndexSection::create<ELF64BE>();
3813
3814 template void elf::splitSections<ELF32LE>();
3815 template void elf::splitSections<ELF32BE>();
3816 template void elf::splitSections<ELF64LE>();
3817 template void elf::splitSections<ELF64BE>();
3818
3819 template class elf::MipsAbiFlagsSection<ELF32LE>;
3820 template class elf::MipsAbiFlagsSection<ELF32BE>;
3821 template class elf::MipsAbiFlagsSection<ELF64LE>;
3822 template class elf::MipsAbiFlagsSection<ELF64BE>;
3823
3824 template class elf::MipsOptionsSection<ELF32LE>;
3825 template class elf::MipsOptionsSection<ELF32BE>;
3826 template class elf::MipsOptionsSection<ELF64LE>;
3827 template class elf::MipsOptionsSection<ELF64BE>;
3828
3829 template void EhFrameSection::iterateFDEWithLSDA<ELF32LE>(
3830 function_ref<void(InputSection &)>);
3831 template void EhFrameSection::iterateFDEWithLSDA<ELF32BE>(
3832 function_ref<void(InputSection &)>);
3833 template void EhFrameSection::iterateFDEWithLSDA<ELF64LE>(
3834 function_ref<void(InputSection &)>);
3835 template void EhFrameSection::iterateFDEWithLSDA<ELF64BE>(
3836 function_ref<void(InputSection &)>);
3837
3838 template class elf::MipsReginfoSection<ELF32LE>;
3839 template class elf::MipsReginfoSection<ELF32BE>;
3840 template class elf::MipsReginfoSection<ELF64LE>;
3841 template class elf::MipsReginfoSection<ELF64BE>;
3842
3843 template class elf::DynamicSection<ELF32LE>;
3844 template class elf::DynamicSection<ELF32BE>;
3845 template class elf::DynamicSection<ELF64LE>;
3846 template class elf::DynamicSection<ELF64BE>;
3847
3848 template class elf::RelocationSection<ELF32LE>;
3849 template class elf::RelocationSection<ELF32BE>;
3850 template class elf::RelocationSection<ELF64LE>;
3851 template class elf::RelocationSection<ELF64BE>;
3852
3853 template class elf::AndroidPackedRelocationSection<ELF32LE>;
3854 template class elf::AndroidPackedRelocationSection<ELF32BE>;
3855 template class elf::AndroidPackedRelocationSection<ELF64LE>;
3856 template class elf::AndroidPackedRelocationSection<ELF64BE>;
3857
3858 template class elf::RelrSection<ELF32LE>;
3859 template class elf::RelrSection<ELF32BE>;
3860 template class elf::RelrSection<ELF64LE>;
3861 template class elf::RelrSection<ELF64BE>;
3862
3863 template class elf::SymbolTableSection<ELF32LE>;
3864 template class elf::SymbolTableSection<ELF32BE>;
3865 template class elf::SymbolTableSection<ELF64LE>;
3866 template class elf::SymbolTableSection<ELF64BE>;
3867
3868 template class elf::VersionNeedSection<ELF32LE>;
3869 template class elf::VersionNeedSection<ELF32BE>;
3870 template class elf::VersionNeedSection<ELF64LE>;
3871 template class elf::VersionNeedSection<ELF64BE>;
3872
3873 template void elf::writeEhdr<ELF32LE>(uint8_t *Buf, Partition &Part);
3874 template void elf::writeEhdr<ELF32BE>(uint8_t *Buf, Partition &Part);
3875 template void elf::writeEhdr<ELF64LE>(uint8_t *Buf, Partition &Part);
3876 template void elf::writeEhdr<ELF64BE>(uint8_t *Buf, Partition &Part);
3877
3878 template void elf::writePhdrs<ELF32LE>(uint8_t *Buf, Partition &Part);
3879 template void elf::writePhdrs<ELF32BE>(uint8_t *Buf, Partition &Part);
3880 template void elf::writePhdrs<ELF64LE>(uint8_t *Buf, Partition &Part);
3881 template void elf::writePhdrs<ELF64BE>(uint8_t *Buf, Partition &Part);
3882
3883 template class elf::PartitionElfHeaderSection<ELF32LE>;
3884 template class elf::PartitionElfHeaderSection<ELF32BE>;
3885 template class elf::PartitionElfHeaderSection<ELF64LE>;
3886 template class elf::PartitionElfHeaderSection<ELF64BE>;
3887
3888 template class elf::PartitionProgramHeadersSection<ELF32LE>;
3889 template class elf::PartitionProgramHeadersSection<ELF32BE>;
3890 template class elf::PartitionProgramHeadersSection<ELF64LE>;
3891 template class elf::PartitionProgramHeadersSection<ELF64BE>;
3892