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,
648 target->gotEntrySize, ".got") {
649 numEntries = target->gotHeaderEntriesNum;
650 }
651
addEntry(Symbol & sym)652 void GotSection::addEntry(Symbol &sym) {
653 sym.gotIndex = numEntries;
654 ++numEntries;
655 }
656
addDynTlsEntry(Symbol & sym)657 bool GotSection::addDynTlsEntry(Symbol &sym) {
658 if (sym.globalDynIndex != -1U)
659 return false;
660 sym.globalDynIndex = numEntries;
661 // Global Dynamic TLS entries take two GOT slots.
662 numEntries += 2;
663 return true;
664 }
665
666 // Reserves TLS entries for a TLS module ID and a TLS block offset.
667 // In total it takes two GOT slots.
addTlsIndex()668 bool GotSection::addTlsIndex() {
669 if (tlsIndexOff != uint32_t(-1))
670 return false;
671 tlsIndexOff = numEntries * config->wordsize;
672 numEntries += 2;
673 return true;
674 }
675
getGlobalDynAddr(const Symbol & b) const676 uint64_t GotSection::getGlobalDynAddr(const Symbol &b) const {
677 return this->getVA() + b.globalDynIndex * config->wordsize;
678 }
679
getGlobalDynOffset(const Symbol & b) const680 uint64_t GotSection::getGlobalDynOffset(const Symbol &b) const {
681 return b.globalDynIndex * config->wordsize;
682 }
683
finalizeContents()684 void GotSection::finalizeContents() {
685 if (config->emachine == EM_PPC64 &&
686 numEntries <= target->gotHeaderEntriesNum && !ElfSym::globalOffsetTable)
687 size = 0;
688 else
689 size = numEntries * config->wordsize;
690 }
691
isNeeded() const692 bool GotSection::isNeeded() const {
693 // Needed if the GOT symbol is used or the number of entries is more than just
694 // the header. A GOT with just the header may not be needed.
695 return hasGotOffRel || numEntries > target->gotHeaderEntriesNum;
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 // When building a shared library we still need a dynamic relocation
989 // for the TP-relative offset as we don't know how much other data will
990 // be allocated before us in the static TLS block.
991 if (s->isPreemptible || config->shared)
992 mainPart->relaDyn->addReloc({target->tlsGotRel, this, offset,
993 DynamicReloc::AgainstSymbolWithTargetVA,
994 *s, 0, R_ABS});
995 }
996 for (std::pair<Symbol *, size_t> &p : got.dynTlsSymbols) {
997 Symbol *s = p.first;
998 uint64_t offset = p.second * config->wordsize;
999 if (s == nullptr) {
1000 if (!config->shared)
1001 continue;
1002 mainPart->relaDyn->addReloc({target->tlsModuleIndexRel, this, offset});
1003 } else {
1004 // When building a shared library we still need a dynamic relocation
1005 // for the module index. Therefore only checking for
1006 // S->isPreemptible is not sufficient (this happens e.g. for
1007 // thread-locals that have been marked as local through a linker script)
1008 if (!s->isPreemptible && !config->shared)
1009 continue;
1010 mainPart->relaDyn->addSymbolReloc(target->tlsModuleIndexRel, this,
1011 offset, *s);
1012 // However, we can skip writing the TLS offset reloc for non-preemptible
1013 // symbols since it is known even in shared libraries
1014 if (!s->isPreemptible)
1015 continue;
1016 offset += config->wordsize;
1017 mainPart->relaDyn->addSymbolReloc(target->tlsOffsetRel, this, offset,
1018 *s);
1019 }
1020 }
1021
1022 // Do not create dynamic relocations for non-TLS
1023 // entries in the primary GOT.
1024 if (&got == primGot)
1025 continue;
1026
1027 // Dynamic relocations for "global" entries.
1028 for (const std::pair<Symbol *, size_t> &p : got.global) {
1029 uint64_t offset = p.second * config->wordsize;
1030 mainPart->relaDyn->addSymbolReloc(target->relativeRel, this, offset,
1031 *p.first);
1032 }
1033 if (!config->isPic)
1034 continue;
1035 // Dynamic relocations for "local" entries in case of PIC.
1036 for (const std::pair<const OutputSection *, FileGot::PageBlock> &l :
1037 got.pagesMap) {
1038 size_t pageCount = l.second.count;
1039 for (size_t pi = 0; pi < pageCount; ++pi) {
1040 uint64_t offset = (l.second.firstIndex + pi) * config->wordsize;
1041 mainPart->relaDyn->addReloc({target->relativeRel, this, offset, l.first,
1042 int64_t(pi * 0x10000)});
1043 }
1044 }
1045 for (const std::pair<GotEntry, size_t> &p : got.local16) {
1046 uint64_t offset = p.second * config->wordsize;
1047 mainPart->relaDyn->addReloc({target->relativeRel, this, offset,
1048 DynamicReloc::AddendOnlyWithTargetVA,
1049 *p.first.first, p.first.second, R_ABS});
1050 }
1051 }
1052 }
1053
isNeeded() const1054 bool MipsGotSection::isNeeded() const {
1055 // We add the .got section to the result for dynamic MIPS target because
1056 // its address and properties are mentioned in the .dynamic section.
1057 return !config->relocatable;
1058 }
1059
getGp(const InputFile * f) const1060 uint64_t MipsGotSection::getGp(const InputFile *f) const {
1061 // For files without related GOT or files refer a primary GOT
1062 // returns "common" _gp value. For secondary GOTs calculate
1063 // individual _gp values.
1064 if (!f || !f->mipsGotIndex.hasValue() || *f->mipsGotIndex == 0)
1065 return ElfSym::mipsGp->getVA(0);
1066 return getVA() + gots[*f->mipsGotIndex].startIndex * config->wordsize +
1067 0x7ff0;
1068 }
1069
writeTo(uint8_t * buf)1070 void MipsGotSection::writeTo(uint8_t *buf) {
1071 // Set the MSB of the second GOT slot. This is not required by any
1072 // MIPS ABI documentation, though.
1073 //
1074 // There is a comment in glibc saying that "The MSB of got[1] of a
1075 // gnu object is set to identify gnu objects," and in GNU gold it
1076 // says "the second entry will be used by some runtime loaders".
1077 // But how this field is being used is unclear.
1078 //
1079 // We are not really willing to mimic other linkers behaviors
1080 // without understanding why they do that, but because all files
1081 // generated by GNU tools have this special GOT value, and because
1082 // we've been doing this for years, it is probably a safe bet to
1083 // keep doing this for now. We really need to revisit this to see
1084 // if we had to do this.
1085 writeUint(buf + config->wordsize, (uint64_t)1 << (config->wordsize * 8 - 1));
1086 for (const FileGot &g : gots) {
1087 auto write = [&](size_t i, const Symbol *s, int64_t a) {
1088 uint64_t va = a;
1089 if (s)
1090 va = s->getVA(a);
1091 writeUint(buf + i * config->wordsize, va);
1092 };
1093 // Write 'page address' entries to the local part of the GOT.
1094 for (const std::pair<const OutputSection *, FileGot::PageBlock> &l :
1095 g.pagesMap) {
1096 size_t pageCount = l.second.count;
1097 uint64_t firstPageAddr = getMipsPageAddr(l.first->addr);
1098 for (size_t pi = 0; pi < pageCount; ++pi)
1099 write(l.second.firstIndex + pi, nullptr, firstPageAddr + pi * 0x10000);
1100 }
1101 // Local, global, TLS, reloc-only entries.
1102 // If TLS entry has a corresponding dynamic relocations, leave it
1103 // initialized by zero. Write down adjusted TLS symbol's values otherwise.
1104 // To calculate the adjustments use offsets for thread-local storage.
1105 // http://web.archive.org/web/20190324223224/https://www.linux-mips.org/wiki/NPTL
1106 for (const std::pair<GotEntry, size_t> &p : g.local16)
1107 write(p.second, p.first.first, p.first.second);
1108 // Write VA to the primary GOT only. For secondary GOTs that
1109 // will be done by REL32 dynamic relocations.
1110 if (&g == &gots.front())
1111 for (const std::pair<Symbol *, size_t> &p : g.global)
1112 write(p.second, p.first, 0);
1113 for (const std::pair<Symbol *, size_t> &p : g.relocs)
1114 write(p.second, p.first, 0);
1115 for (const std::pair<Symbol *, size_t> &p : g.tls)
1116 write(p.second, p.first,
1117 p.first->isPreemptible || config->shared ? 0 : -0x7000);
1118 for (const std::pair<Symbol *, size_t> &p : g.dynTlsSymbols) {
1119 if (p.first == nullptr && !config->shared)
1120 write(p.second, nullptr, 1);
1121 else if (p.first && !p.first->isPreemptible) {
1122 // If we are emitting a shared libary with relocations we mustn't write
1123 // anything to the GOT here. When using Elf_Rel relocations the value
1124 // one will be treated as an addend and will cause crashes at runtime
1125 if (!config->shared)
1126 write(p.second, nullptr, 1);
1127 write(p.second + 1, p.first, -0x8000);
1128 }
1129 }
1130 }
1131 }
1132
1133 // On PowerPC the .plt section is used to hold the table of function addresses
1134 // instead of the .got.plt, and the type is SHT_NOBITS similar to a .bss
1135 // section. I don't know why we have a BSS style type for the section but it is
1136 // consistent across both 64-bit PowerPC ABIs as well as the 32-bit PowerPC ABI.
GotPltSection()1137 GotPltSection::GotPltSection()
1138 : SyntheticSection(SHF_ALLOC | SHF_WRITE, SHT_PROGBITS, config->wordsize,
1139 ".got.plt") {
1140 if (config->emachine == EM_PPC) {
1141 name = ".plt";
1142 } else if (config->emachine == EM_PPC64) {
1143 type = SHT_NOBITS;
1144 name = ".plt";
1145 }
1146 }
1147
addEntry(Symbol & sym)1148 void GotPltSection::addEntry(Symbol &sym) {
1149 assert(sym.pltIndex == entries.size());
1150 entries.push_back(&sym);
1151 }
1152
getSize() const1153 size_t GotPltSection::getSize() const {
1154 return (target->gotPltHeaderEntriesNum + entries.size()) *
1155 target->gotEntrySize;
1156 }
1157
writeTo(uint8_t * buf)1158 void GotPltSection::writeTo(uint8_t *buf) {
1159 target->writeGotPltHeader(buf);
1160 buf += target->gotPltHeaderEntriesNum * target->gotEntrySize;
1161 for (const Symbol *b : entries) {
1162 target->writeGotPlt(buf, *b);
1163 buf += target->gotEntrySize;
1164 }
1165 }
1166
isNeeded() const1167 bool GotPltSection::isNeeded() const {
1168 // We need to emit GOTPLT even if it's empty if there's a relocation relative
1169 // to it.
1170 return !entries.empty() || hasGotPltOffRel;
1171 }
1172
getIgotPltName()1173 static StringRef getIgotPltName() {
1174 // On ARM the IgotPltSection is part of the GotSection.
1175 if (config->emachine == EM_ARM)
1176 return ".got";
1177
1178 // On PowerPC64 the GotPltSection is renamed to '.plt' so the IgotPltSection
1179 // needs to be named the same.
1180 if (config->emachine == EM_PPC64)
1181 return ".plt";
1182
1183 return ".got.plt";
1184 }
1185
1186 // On PowerPC64 the GotPltSection type is SHT_NOBITS so we have to follow suit
1187 // with the IgotPltSection.
IgotPltSection()1188 IgotPltSection::IgotPltSection()
1189 : SyntheticSection(SHF_ALLOC | SHF_WRITE,
1190 config->emachine == EM_PPC64 ? SHT_NOBITS : SHT_PROGBITS,
1191 target->gotEntrySize, getIgotPltName()) {}
1192
addEntry(Symbol & sym)1193 void IgotPltSection::addEntry(Symbol &sym) {
1194 assert(sym.pltIndex == entries.size());
1195 entries.push_back(&sym);
1196 }
1197
getSize() const1198 size_t IgotPltSection::getSize() const {
1199 return entries.size() * target->gotEntrySize;
1200 }
1201
writeTo(uint8_t * buf)1202 void IgotPltSection::writeTo(uint8_t *buf) {
1203 for (const Symbol *b : entries) {
1204 target->writeIgotPlt(buf, *b);
1205 buf += target->gotEntrySize;
1206 }
1207 }
1208
StringTableSection(StringRef name,bool dynamic)1209 StringTableSection::StringTableSection(StringRef name, bool dynamic)
1210 : SyntheticSection(dynamic ? (uint64_t)SHF_ALLOC : 0, SHT_STRTAB, 1, name),
1211 dynamic(dynamic) {
1212 // ELF string tables start with a NUL byte.
1213 addString("");
1214 }
1215
1216 // Adds a string to the string table. If `hashIt` is true we hash and check for
1217 // duplicates. It is optional because the name of global symbols are already
1218 // uniqued and hashing them again has a big cost for a small value: uniquing
1219 // them with some other string that happens to be the same.
addString(StringRef s,bool hashIt)1220 unsigned StringTableSection::addString(StringRef s, bool hashIt) {
1221 if (hashIt) {
1222 auto r = stringMap.insert(std::make_pair(s, this->size));
1223 if (!r.second)
1224 return r.first->second;
1225 }
1226 unsigned ret = this->size;
1227 this->size = this->size + s.size() + 1;
1228 strings.push_back(s);
1229 return ret;
1230 }
1231
writeTo(uint8_t * buf)1232 void StringTableSection::writeTo(uint8_t *buf) {
1233 for (StringRef s : strings) {
1234 memcpy(buf, s.data(), s.size());
1235 buf[s.size()] = '\0';
1236 buf += s.size() + 1;
1237 }
1238 }
1239
1240 // Returns the number of entries in .gnu.version_d: the number of
1241 // non-VER_NDX_LOCAL-non-VER_NDX_GLOBAL definitions, plus 1.
1242 // Note that we don't support vd_cnt > 1 yet.
getVerDefNum()1243 static unsigned getVerDefNum() {
1244 return namedVersionDefs().size() + 1;
1245 }
1246
1247 template <class ELFT>
DynamicSection()1248 DynamicSection<ELFT>::DynamicSection()
1249 : SyntheticSection(SHF_ALLOC | SHF_WRITE, SHT_DYNAMIC, config->wordsize,
1250 ".dynamic") {
1251 this->entsize = ELFT::Is64Bits ? 16 : 8;
1252
1253 // .dynamic section is not writable on MIPS and on Fuchsia OS
1254 // which passes -z rodynamic.
1255 // See "Special Section" in Chapter 4 in the following document:
1256 // ftp://www.linux-mips.org/pub/linux/mips/doc/ABI/mipsabi.pdf
1257 if (config->emachine == EM_MIPS || config->zRodynamic)
1258 this->flags = SHF_ALLOC;
1259 }
1260
1261 template <class ELFT>
add(int32_t tag,std::function<uint64_t ()> fn)1262 void DynamicSection<ELFT>::add(int32_t tag, std::function<uint64_t()> fn) {
1263 entries.push_back({tag, fn});
1264 }
1265
1266 template <class ELFT>
addInt(int32_t tag,uint64_t val)1267 void DynamicSection<ELFT>::addInt(int32_t tag, uint64_t val) {
1268 entries.push_back({tag, [=] { return val; }});
1269 }
1270
1271 template <class ELFT>
addInSec(int32_t tag,InputSection * sec)1272 void DynamicSection<ELFT>::addInSec(int32_t tag, InputSection *sec) {
1273 entries.push_back({tag, [=] { return sec->getVA(0); }});
1274 }
1275
1276 template <class ELFT>
addInSecRelative(int32_t tag,InputSection * sec)1277 void DynamicSection<ELFT>::addInSecRelative(int32_t tag, InputSection *sec) {
1278 size_t tagOffset = entries.size() * entsize;
1279 entries.push_back(
1280 {tag, [=] { return sec->getVA(0) - (getVA() + tagOffset); }});
1281 }
1282
1283 template <class ELFT>
addOutSec(int32_t tag,OutputSection * sec)1284 void DynamicSection<ELFT>::addOutSec(int32_t tag, OutputSection *sec) {
1285 entries.push_back({tag, [=] { return sec->addr; }});
1286 }
1287
1288 template <class ELFT>
addSize(int32_t tag,OutputSection * sec)1289 void DynamicSection<ELFT>::addSize(int32_t tag, OutputSection *sec) {
1290 entries.push_back({tag, [=] { return sec->size; }});
1291 }
1292
1293 template <class ELFT>
addSym(int32_t tag,Symbol * sym)1294 void DynamicSection<ELFT>::addSym(int32_t tag, Symbol *sym) {
1295 entries.push_back({tag, [=] { return sym->getVA(); }});
1296 }
1297
1298 // The output section .rela.dyn may include these synthetic sections:
1299 //
1300 // - part.relaDyn
1301 // - in.relaIplt: this is included if in.relaIplt is named .rela.dyn
1302 // - in.relaPlt: this is included if a linker script places .rela.plt inside
1303 // .rela.dyn
1304 //
1305 // DT_RELASZ is the total size of the included sections.
addRelaSz(RelocationBaseSection * relaDyn)1306 static std::function<uint64_t()> addRelaSz(RelocationBaseSection *relaDyn) {
1307 return [=]() {
1308 size_t size = relaDyn->getSize();
1309 if (in.relaIplt->getParent() == relaDyn->getParent())
1310 size += in.relaIplt->getSize();
1311 if (in.relaPlt->getParent() == relaDyn->getParent())
1312 size += in.relaPlt->getSize();
1313 return size;
1314 };
1315 }
1316
1317 // A Linker script may assign the RELA relocation sections to the same
1318 // output section. When this occurs we cannot just use the OutputSection
1319 // Size. Moreover the [DT_JMPREL, DT_JMPREL + DT_PLTRELSZ) is permitted to
1320 // overlap with the [DT_RELA, DT_RELA + DT_RELASZ).
addPltRelSz()1321 static uint64_t addPltRelSz() {
1322 size_t size = in.relaPlt->getSize();
1323 if (in.relaIplt->getParent() == in.relaPlt->getParent() &&
1324 in.relaIplt->name == in.relaPlt->name)
1325 size += in.relaIplt->getSize();
1326 return size;
1327 }
1328
1329 // Add remaining entries to complete .dynamic contents.
finalizeContents()1330 template <class ELFT> void DynamicSection<ELFT>::finalizeContents() {
1331 elf::Partition &part = getPartition();
1332 bool isMain = part.name.empty();
1333
1334 for (StringRef s : config->filterList)
1335 addInt(DT_FILTER, part.dynStrTab->addString(s));
1336 for (StringRef s : config->auxiliaryList)
1337 addInt(DT_AUXILIARY, part.dynStrTab->addString(s));
1338
1339 if (!config->rpath.empty())
1340 addInt(config->enableNewDtags ? DT_RUNPATH : DT_RPATH,
1341 part.dynStrTab->addString(config->rpath));
1342
1343 for (SharedFile *file : sharedFiles)
1344 if (file->isNeeded)
1345 addInt(DT_NEEDED, part.dynStrTab->addString(file->soName));
1346
1347 if (isMain) {
1348 if (!config->soName.empty())
1349 addInt(DT_SONAME, part.dynStrTab->addString(config->soName));
1350 } else {
1351 if (!config->soName.empty())
1352 addInt(DT_NEEDED, part.dynStrTab->addString(config->soName));
1353 addInt(DT_SONAME, part.dynStrTab->addString(part.name));
1354 }
1355
1356 // Set DT_FLAGS and DT_FLAGS_1.
1357 uint32_t dtFlags = 0;
1358 uint32_t dtFlags1 = 0;
1359 if (config->bsymbolic == BsymbolicKind::All)
1360 dtFlags |= DF_SYMBOLIC;
1361 if (config->zGlobal)
1362 dtFlags1 |= DF_1_GLOBAL;
1363 if (config->zInitfirst)
1364 dtFlags1 |= DF_1_INITFIRST;
1365 if (config->zInterpose)
1366 dtFlags1 |= DF_1_INTERPOSE;
1367 if (config->zNodefaultlib)
1368 dtFlags1 |= DF_1_NODEFLIB;
1369 if (config->zNodelete)
1370 dtFlags1 |= DF_1_NODELETE;
1371 if (config->zNodlopen)
1372 dtFlags1 |= DF_1_NOOPEN;
1373 if (config->pie)
1374 dtFlags1 |= DF_1_PIE;
1375 if (config->zNow) {
1376 dtFlags |= DF_BIND_NOW;
1377 dtFlags1 |= DF_1_NOW;
1378 }
1379 if (config->zOrigin) {
1380 dtFlags |= DF_ORIGIN;
1381 dtFlags1 |= DF_1_ORIGIN;
1382 }
1383 if (!config->zText)
1384 dtFlags |= DF_TEXTREL;
1385 if (config->hasStaticTlsModel)
1386 dtFlags |= DF_STATIC_TLS;
1387
1388 if (dtFlags)
1389 addInt(DT_FLAGS, dtFlags);
1390 if (dtFlags1)
1391 addInt(DT_FLAGS_1, dtFlags1);
1392
1393 // DT_DEBUG is a pointer to debug information used by debuggers at runtime. We
1394 // need it for each process, so we don't write it for DSOs. The loader writes
1395 // the pointer into this entry.
1396 //
1397 // DT_DEBUG is the only .dynamic entry that needs to be written to. Some
1398 // systems (currently only Fuchsia OS) provide other means to give the
1399 // debugger this information. Such systems may choose make .dynamic read-only.
1400 // If the target is such a system (used -z rodynamic) don't write DT_DEBUG.
1401 if (!config->shared && !config->relocatable && !config->zRodynamic)
1402 addInt(DT_DEBUG, 0);
1403
1404 if (OutputSection *sec = part.dynStrTab->getParent())
1405 this->link = sec->sectionIndex;
1406
1407 if (part.relaDyn->isNeeded() ||
1408 (in.relaIplt->isNeeded() &&
1409 part.relaDyn->getParent() == in.relaIplt->getParent())) {
1410 addInSec(part.relaDyn->dynamicTag, part.relaDyn);
1411 entries.push_back({part.relaDyn->sizeDynamicTag, addRelaSz(part.relaDyn)});
1412
1413 bool isRela = config->isRela;
1414 addInt(isRela ? DT_RELAENT : DT_RELENT,
1415 isRela ? sizeof(Elf_Rela) : sizeof(Elf_Rel));
1416
1417 // MIPS dynamic loader does not support RELCOUNT tag.
1418 // The problem is in the tight relation between dynamic
1419 // relocations and GOT. So do not emit this tag on MIPS.
1420 if (config->emachine != EM_MIPS) {
1421 size_t numRelativeRels = part.relaDyn->getRelativeRelocCount();
1422 if (config->zCombreloc && numRelativeRels)
1423 addInt(isRela ? DT_RELACOUNT : DT_RELCOUNT, numRelativeRels);
1424 }
1425 }
1426 if (part.relrDyn && !part.relrDyn->relocs.empty()) {
1427 addInSec(config->useAndroidRelrTags ? DT_ANDROID_RELR : DT_RELR,
1428 part.relrDyn);
1429 addSize(config->useAndroidRelrTags ? DT_ANDROID_RELRSZ : DT_RELRSZ,
1430 part.relrDyn->getParent());
1431 addInt(config->useAndroidRelrTags ? DT_ANDROID_RELRENT : DT_RELRENT,
1432 sizeof(Elf_Relr));
1433 }
1434 // .rel[a].plt section usually consists of two parts, containing plt and
1435 // iplt relocations. It is possible to have only iplt relocations in the
1436 // output. In that case relaPlt is empty and have zero offset, the same offset
1437 // as relaIplt has. And we still want to emit proper dynamic tags for that
1438 // case, so here we always use relaPlt as marker for the beginning of
1439 // .rel[a].plt section.
1440 if (isMain && (in.relaPlt->isNeeded() || in.relaIplt->isNeeded())) {
1441 addInSec(DT_JMPREL, in.relaPlt);
1442 entries.push_back({DT_PLTRELSZ, addPltRelSz});
1443 switch (config->emachine) {
1444 case EM_MIPS:
1445 addInSec(DT_MIPS_PLTGOT, in.gotPlt);
1446 break;
1447 case EM_SPARCV9:
1448 addInSec(DT_PLTGOT, in.plt);
1449 break;
1450 case EM_AARCH64:
1451 if (llvm::find_if(in.relaPlt->relocs, [](const DynamicReloc &r) {
1452 return r.type == target->pltRel &&
1453 r.sym->stOther & STO_AARCH64_VARIANT_PCS;
1454 }) != in.relaPlt->relocs.end())
1455 addInt(DT_AARCH64_VARIANT_PCS, 0);
1456 LLVM_FALLTHROUGH;
1457 default:
1458 addInSec(DT_PLTGOT, in.gotPlt);
1459 break;
1460 }
1461 addInt(DT_PLTREL, config->isRela ? DT_RELA : DT_REL);
1462 }
1463
1464 if (config->emachine == EM_AARCH64) {
1465 if (config->andFeatures & GNU_PROPERTY_AARCH64_FEATURE_1_BTI)
1466 addInt(DT_AARCH64_BTI_PLT, 0);
1467 if (config->zPacPlt)
1468 addInt(DT_AARCH64_PAC_PLT, 0);
1469 }
1470
1471 addInSec(DT_SYMTAB, part.dynSymTab);
1472 addInt(DT_SYMENT, sizeof(Elf_Sym));
1473 addInSec(DT_STRTAB, part.dynStrTab);
1474 addInt(DT_STRSZ, part.dynStrTab->getSize());
1475 if (!config->zText)
1476 addInt(DT_TEXTREL, 0);
1477 if (part.gnuHashTab)
1478 addInSec(DT_GNU_HASH, part.gnuHashTab);
1479 if (part.hashTab)
1480 addInSec(DT_HASH, part.hashTab);
1481
1482 if (isMain) {
1483 if (Out::preinitArray) {
1484 addOutSec(DT_PREINIT_ARRAY, Out::preinitArray);
1485 addSize(DT_PREINIT_ARRAYSZ, Out::preinitArray);
1486 }
1487 if (Out::initArray) {
1488 addOutSec(DT_INIT_ARRAY, Out::initArray);
1489 addSize(DT_INIT_ARRAYSZ, Out::initArray);
1490 }
1491 if (Out::finiArray) {
1492 addOutSec(DT_FINI_ARRAY, Out::finiArray);
1493 addSize(DT_FINI_ARRAYSZ, Out::finiArray);
1494 }
1495
1496 if (Symbol *b = symtab->find(config->init))
1497 if (b->isDefined())
1498 addSym(DT_INIT, b);
1499 if (Symbol *b = symtab->find(config->fini))
1500 if (b->isDefined())
1501 addSym(DT_FINI, b);
1502 }
1503
1504 if (part.verSym && part.verSym->isNeeded())
1505 addInSec(DT_VERSYM, part.verSym);
1506 if (part.verDef && part.verDef->isLive()) {
1507 addInSec(DT_VERDEF, part.verDef);
1508 addInt(DT_VERDEFNUM, getVerDefNum());
1509 }
1510 if (part.verNeed && part.verNeed->isNeeded()) {
1511 addInSec(DT_VERNEED, part.verNeed);
1512 unsigned needNum = 0;
1513 for (SharedFile *f : sharedFiles)
1514 if (!f->vernauxs.empty())
1515 ++needNum;
1516 addInt(DT_VERNEEDNUM, needNum);
1517 }
1518
1519 if (config->emachine == EM_MIPS) {
1520 addInt(DT_MIPS_RLD_VERSION, 1);
1521 addInt(DT_MIPS_FLAGS, RHF_NOTPOT);
1522 addInt(DT_MIPS_BASE_ADDRESS, target->getImageBase());
1523 addInt(DT_MIPS_SYMTABNO, part.dynSymTab->getNumSymbols());
1524
1525 add(DT_MIPS_LOCAL_GOTNO, [] { return in.mipsGot->getLocalEntriesNum(); });
1526
1527 if (const Symbol *b = in.mipsGot->getFirstGlobalEntry())
1528 addInt(DT_MIPS_GOTSYM, b->dynsymIndex);
1529 else
1530 addInt(DT_MIPS_GOTSYM, part.dynSymTab->getNumSymbols());
1531 addInSec(DT_PLTGOT, in.mipsGot);
1532 if (in.mipsRldMap) {
1533 if (!config->pie)
1534 addInSec(DT_MIPS_RLD_MAP, in.mipsRldMap);
1535 // Store the offset to the .rld_map section
1536 // relative to the address of the tag.
1537 addInSecRelative(DT_MIPS_RLD_MAP_REL, in.mipsRldMap);
1538 }
1539 }
1540
1541 // DT_PPC_GOT indicates to glibc Secure PLT is used. If DT_PPC_GOT is absent,
1542 // glibc assumes the old-style BSS PLT layout which we don't support.
1543 if (config->emachine == EM_PPC)
1544 add(DT_PPC_GOT, [] { return in.got->getVA(); });
1545
1546 // Glink dynamic tag is required by the V2 abi if the plt section isn't empty.
1547 if (config->emachine == EM_PPC64 && in.plt->isNeeded()) {
1548 // The Glink tag points to 32 bytes before the first lazy symbol resolution
1549 // stub, which starts directly after the header.
1550 entries.push_back({DT_PPC64_GLINK, [=] {
1551 unsigned offset = target->pltHeaderSize - 32;
1552 return in.plt->getVA(0) + offset;
1553 }});
1554 }
1555
1556 addInt(DT_NULL, 0);
1557
1558 getParent()->link = this->link;
1559 this->size = entries.size() * this->entsize;
1560 }
1561
writeTo(uint8_t * buf)1562 template <class ELFT> void DynamicSection<ELFT>::writeTo(uint8_t *buf) {
1563 auto *p = reinterpret_cast<Elf_Dyn *>(buf);
1564
1565 for (std::pair<int32_t, std::function<uint64_t()>> &kv : entries) {
1566 p->d_tag = kv.first;
1567 p->d_un.d_val = kv.second();
1568 ++p;
1569 }
1570 }
1571
getOffset() const1572 uint64_t DynamicReloc::getOffset() const {
1573 return inputSec->getVA(offsetInSec);
1574 }
1575
computeAddend() const1576 int64_t DynamicReloc::computeAddend() const {
1577 switch (kind) {
1578 case AddendOnly:
1579 assert(sym == nullptr);
1580 return addend;
1581 case AgainstSymbol:
1582 assert(sym != nullptr);
1583 return addend;
1584 case AddendOnlyWithTargetVA:
1585 case AgainstSymbolWithTargetVA:
1586 return InputSection::getRelocTargetVA(inputSec->file, type, addend,
1587 getOffset(), *sym, expr);
1588 case MipsMultiGotPage:
1589 assert(sym == nullptr);
1590 return getMipsPageAddr(outputSec->addr) + addend;
1591 }
1592 llvm_unreachable("Unknown DynamicReloc::Kind enum");
1593 }
1594
getSymIndex(SymbolTableBaseSection * symTab) const1595 uint32_t DynamicReloc::getSymIndex(SymbolTableBaseSection *symTab) const {
1596 if (needsDynSymIndex())
1597 return symTab->getSymbolIndex(sym);
1598 return 0;
1599 }
1600
RelocationBaseSection(StringRef name,uint32_t type,int32_t dynamicTag,int32_t sizeDynamicTag)1601 RelocationBaseSection::RelocationBaseSection(StringRef name, uint32_t type,
1602 int32_t dynamicTag,
1603 int32_t sizeDynamicTag)
1604 : SyntheticSection(SHF_ALLOC, type, config->wordsize, name),
1605 dynamicTag(dynamicTag), sizeDynamicTag(sizeDynamicTag) {}
1606
addSymbolReloc(RelType dynType,InputSectionBase * isec,uint64_t offsetInSec,Symbol & sym,int64_t addend,Optional<RelType> addendRelType)1607 void RelocationBaseSection::addSymbolReloc(RelType dynType,
1608 InputSectionBase *isec,
1609 uint64_t offsetInSec, Symbol &sym,
1610 int64_t addend,
1611 Optional<RelType> addendRelType) {
1612 addReloc(DynamicReloc::AgainstSymbol, dynType, isec, offsetInSec, sym, addend,
1613 R_ADDEND, addendRelType ? *addendRelType : target->noneRel);
1614 }
1615
addRelativeReloc(RelType dynType,InputSectionBase * inputSec,uint64_t offsetInSec,Symbol & sym,int64_t addend,RelType addendRelType,RelExpr expr)1616 void RelocationBaseSection::addRelativeReloc(
1617 RelType dynType, InputSectionBase *inputSec, uint64_t offsetInSec,
1618 Symbol &sym, int64_t addend, RelType addendRelType, RelExpr expr) {
1619 // This function should only be called for non-preemptible symbols or
1620 // RelExpr values that refer to an address inside the output file (e.g. the
1621 // address of the GOT entry for a potentially preemptible symbol).
1622 assert((!sym.isPreemptible || expr == R_GOT) &&
1623 "cannot add relative relocation against preemptible symbol");
1624 assert(expr != R_ADDEND && "expected non-addend relocation expression");
1625 addReloc(DynamicReloc::AddendOnlyWithTargetVA, dynType, inputSec, offsetInSec,
1626 sym, addend, expr, addendRelType);
1627 }
1628
addAddendOnlyRelocIfNonPreemptible(RelType dynType,InputSectionBase * isec,uint64_t offsetInSec,Symbol & sym,RelType addendRelType)1629 void RelocationBaseSection::addAddendOnlyRelocIfNonPreemptible(
1630 RelType dynType, InputSectionBase *isec, uint64_t offsetInSec, Symbol &sym,
1631 RelType addendRelType) {
1632 // No need to write an addend to the section for preemptible symbols.
1633 if (sym.isPreemptible)
1634 addReloc({dynType, isec, offsetInSec, DynamicReloc::AgainstSymbol, sym, 0,
1635 R_ABS});
1636 else
1637 addReloc(DynamicReloc::AddendOnlyWithTargetVA, dynType, isec, offsetInSec,
1638 sym, 0, R_ABS, addendRelType);
1639 }
1640
addReloc(DynamicReloc::Kind kind,RelType dynType,InputSectionBase * inputSec,uint64_t offsetInSec,Symbol & sym,int64_t addend,RelExpr expr,RelType addendRelType)1641 void RelocationBaseSection::addReloc(DynamicReloc::Kind kind, RelType dynType,
1642 InputSectionBase *inputSec,
1643 uint64_t offsetInSec, Symbol &sym,
1644 int64_t addend, RelExpr expr,
1645 RelType addendRelType) {
1646 // Write the addends to the relocated address if required. We skip
1647 // it if the written value would be zero.
1648 if (config->writeAddends && (expr != R_ADDEND || addend != 0))
1649 inputSec->relocations.push_back(
1650 {expr, addendRelType, offsetInSec, addend, &sym});
1651 addReloc({dynType, inputSec, offsetInSec, kind, sym, addend, expr});
1652 }
1653
addReloc(const DynamicReloc & reloc)1654 void RelocationBaseSection::addReloc(const DynamicReloc &reloc) {
1655 if (reloc.type == target->relativeRel)
1656 ++numRelativeRelocs;
1657 relocs.push_back(reloc);
1658 }
1659
finalizeContents()1660 void RelocationBaseSection::finalizeContents() {
1661 SymbolTableBaseSection *symTab = getPartition().dynSymTab;
1662
1663 // When linking glibc statically, .rel{,a}.plt contains R_*_IRELATIVE
1664 // relocations due to IFUNC (e.g. strcpy). sh_link will be set to 0 in that
1665 // case.
1666 if (symTab && symTab->getParent())
1667 getParent()->link = symTab->getParent()->sectionIndex;
1668 else
1669 getParent()->link = 0;
1670
1671 if (in.relaPlt == this) {
1672 getParent()->flags |= ELF::SHF_INFO_LINK;
1673 getParent()->info = in.gotPlt->getParent()->sectionIndex;
1674 }
1675 if (in.relaIplt == this) {
1676 getParent()->flags |= ELF::SHF_INFO_LINK;
1677 getParent()->info = in.igotPlt->getParent()->sectionIndex;
1678 }
1679 }
1680
RelrBaseSection()1681 RelrBaseSection::RelrBaseSection()
1682 : SyntheticSection(SHF_ALLOC,
1683 config->useAndroidRelrTags ? SHT_ANDROID_RELR : SHT_RELR,
1684 config->wordsize, ".relr.dyn") {}
1685
1686 template <class ELFT>
encodeDynamicReloc(SymbolTableBaseSection * symTab,typename ELFT::Rela * p,const DynamicReloc & rel)1687 static void encodeDynamicReloc(SymbolTableBaseSection *symTab,
1688 typename ELFT::Rela *p,
1689 const DynamicReloc &rel) {
1690 if (config->isRela)
1691 p->r_addend = rel.computeAddend();
1692 p->r_offset = rel.getOffset();
1693 p->setSymbolAndType(rel.getSymIndex(symTab), rel.type, config->isMips64EL);
1694 }
1695
1696 template <class ELFT>
RelocationSection(StringRef name,bool sort)1697 RelocationSection<ELFT>::RelocationSection(StringRef name, bool sort)
1698 : RelocationBaseSection(name, config->isRela ? SHT_RELA : SHT_REL,
1699 config->isRela ? DT_RELA : DT_REL,
1700 config->isRela ? DT_RELASZ : DT_RELSZ),
1701 sort(sort) {
1702 this->entsize = config->isRela ? sizeof(Elf_Rela) : sizeof(Elf_Rel);
1703 }
1704
writeTo(uint8_t * buf)1705 template <class ELFT> void RelocationSection<ELFT>::writeTo(uint8_t *buf) {
1706 SymbolTableBaseSection *symTab = getPartition().dynSymTab;
1707
1708 // Sort by (!IsRelative,SymIndex,r_offset). DT_REL[A]COUNT requires us to
1709 // place R_*_RELATIVE first. SymIndex is to improve locality, while r_offset
1710 // is to make results easier to read.
1711 if (sort)
1712 llvm::stable_sort(
1713 relocs, [&](const DynamicReloc &a, const DynamicReloc &b) {
1714 return std::make_tuple(a.type != target->relativeRel,
1715 a.getSymIndex(symTab), a.getOffset()) <
1716 std::make_tuple(b.type != target->relativeRel,
1717 b.getSymIndex(symTab), b.getOffset());
1718 });
1719
1720 for (const DynamicReloc &rel : relocs) {
1721 encodeDynamicReloc<ELFT>(symTab, reinterpret_cast<Elf_Rela *>(buf), rel);
1722 buf += config->isRela ? sizeof(Elf_Rela) : sizeof(Elf_Rel);
1723 }
1724 }
1725
1726 template <class ELFT>
AndroidPackedRelocationSection(StringRef name)1727 AndroidPackedRelocationSection<ELFT>::AndroidPackedRelocationSection(
1728 StringRef name)
1729 : RelocationBaseSection(
1730 name, config->isRela ? SHT_ANDROID_RELA : SHT_ANDROID_REL,
1731 config->isRela ? DT_ANDROID_RELA : DT_ANDROID_REL,
1732 config->isRela ? DT_ANDROID_RELASZ : DT_ANDROID_RELSZ) {
1733 this->entsize = 1;
1734 }
1735
1736 template <class ELFT>
updateAllocSize()1737 bool AndroidPackedRelocationSection<ELFT>::updateAllocSize() {
1738 // This function computes the contents of an Android-format packed relocation
1739 // section.
1740 //
1741 // This format compresses relocations by using relocation groups to factor out
1742 // fields that are common between relocations and storing deltas from previous
1743 // relocations in SLEB128 format (which has a short representation for small
1744 // numbers). A good example of a relocation type with common fields is
1745 // R_*_RELATIVE, which is normally used to represent function pointers in
1746 // vtables. In the REL format, each relative relocation has the same r_info
1747 // field, and is only different from other relative relocations in terms of
1748 // the r_offset field. By sorting relocations by offset, grouping them by
1749 // r_info and representing each relocation with only the delta from the
1750 // previous offset, each 8-byte relocation can be compressed to as little as 1
1751 // byte (or less with run-length encoding). This relocation packer was able to
1752 // reduce the size of the relocation section in an Android Chromium DSO from
1753 // 2,911,184 bytes to 174,693 bytes, or 6% of the original size.
1754 //
1755 // A relocation section consists of a header containing the literal bytes
1756 // 'APS2' followed by a sequence of SLEB128-encoded integers. The first two
1757 // elements are the total number of relocations in the section and an initial
1758 // r_offset value. The remaining elements define a sequence of relocation
1759 // groups. Each relocation group starts with a header consisting of the
1760 // following elements:
1761 //
1762 // - the number of relocations in the relocation group
1763 // - flags for the relocation group
1764 // - (if RELOCATION_GROUPED_BY_OFFSET_DELTA_FLAG is set) the r_offset delta
1765 // for each relocation in the group.
1766 // - (if RELOCATION_GROUPED_BY_INFO_FLAG is set) the value of the r_info
1767 // field for each relocation in the group.
1768 // - (if RELOCATION_GROUP_HAS_ADDEND_FLAG and
1769 // RELOCATION_GROUPED_BY_ADDEND_FLAG are set) the r_addend delta for
1770 // each relocation in the group.
1771 //
1772 // Following the relocation group header are descriptions of each of the
1773 // relocations in the group. They consist of the following elements:
1774 //
1775 // - (if RELOCATION_GROUPED_BY_OFFSET_DELTA_FLAG is not set) the r_offset
1776 // delta for this relocation.
1777 // - (if RELOCATION_GROUPED_BY_INFO_FLAG is not set) the value of the r_info
1778 // field for this relocation.
1779 // - (if RELOCATION_GROUP_HAS_ADDEND_FLAG is set and
1780 // RELOCATION_GROUPED_BY_ADDEND_FLAG is not set) the r_addend delta for
1781 // this relocation.
1782
1783 size_t oldSize = relocData.size();
1784
1785 relocData = {'A', 'P', 'S', '2'};
1786 raw_svector_ostream os(relocData);
1787 auto add = [&](int64_t v) { encodeSLEB128(v, os); };
1788
1789 // The format header includes the number of relocations and the initial
1790 // offset (we set this to zero because the first relocation group will
1791 // perform the initial adjustment).
1792 add(relocs.size());
1793 add(0);
1794
1795 std::vector<Elf_Rela> relatives, nonRelatives;
1796
1797 for (const DynamicReloc &rel : relocs) {
1798 Elf_Rela r;
1799 encodeDynamicReloc<ELFT>(getPartition().dynSymTab, &r, rel);
1800
1801 if (r.getType(config->isMips64EL) == target->relativeRel)
1802 relatives.push_back(r);
1803 else
1804 nonRelatives.push_back(r);
1805 }
1806
1807 llvm::sort(relatives, [](const Elf_Rel &a, const Elf_Rel &b) {
1808 return a.r_offset < b.r_offset;
1809 });
1810
1811 // Try to find groups of relative relocations which are spaced one word
1812 // apart from one another. These generally correspond to vtable entries. The
1813 // format allows these groups to be encoded using a sort of run-length
1814 // encoding, but each group will cost 7 bytes in addition to the offset from
1815 // the previous group, so it is only profitable to do this for groups of
1816 // size 8 or larger.
1817 std::vector<Elf_Rela> ungroupedRelatives;
1818 std::vector<std::vector<Elf_Rela>> relativeGroups;
1819 for (auto i = relatives.begin(), e = relatives.end(); i != e;) {
1820 std::vector<Elf_Rela> group;
1821 do {
1822 group.push_back(*i++);
1823 } while (i != e && (i - 1)->r_offset + config->wordsize == i->r_offset);
1824
1825 if (group.size() < 8)
1826 ungroupedRelatives.insert(ungroupedRelatives.end(), group.begin(),
1827 group.end());
1828 else
1829 relativeGroups.emplace_back(std::move(group));
1830 }
1831
1832 // For non-relative relocations, we would like to:
1833 // 1. Have relocations with the same symbol offset to be consecutive, so
1834 // that the runtime linker can speed-up symbol lookup by implementing an
1835 // 1-entry cache.
1836 // 2. Group relocations by r_info to reduce the size of the relocation
1837 // section.
1838 // Since the symbol offset is the high bits in r_info, sorting by r_info
1839 // allows us to do both.
1840 //
1841 // For Rela, we also want to sort by r_addend when r_info is the same. This
1842 // enables us to group by r_addend as well.
1843 llvm::stable_sort(nonRelatives, [](const Elf_Rela &a, const Elf_Rela &b) {
1844 if (a.r_info != b.r_info)
1845 return a.r_info < b.r_info;
1846 if (config->isRela)
1847 return a.r_addend < b.r_addend;
1848 return false;
1849 });
1850
1851 // Group relocations with the same r_info. Note that each group emits a group
1852 // header and that may make the relocation section larger. It is hard to
1853 // estimate the size of a group header as the encoded size of that varies
1854 // based on r_info. However, we can approximate this trade-off by the number
1855 // of values encoded. Each group header contains 3 values, and each relocation
1856 // in a group encodes one less value, as compared to when it is not grouped.
1857 // Therefore, we only group relocations if there are 3 or more of them with
1858 // the same r_info.
1859 //
1860 // For Rela, the addend for most non-relative relocations is zero, and thus we
1861 // can usually get a smaller relocation section if we group relocations with 0
1862 // addend as well.
1863 std::vector<Elf_Rela> ungroupedNonRelatives;
1864 std::vector<std::vector<Elf_Rela>> nonRelativeGroups;
1865 for (auto i = nonRelatives.begin(), e = nonRelatives.end(); i != e;) {
1866 auto j = i + 1;
1867 while (j != e && i->r_info == j->r_info &&
1868 (!config->isRela || i->r_addend == j->r_addend))
1869 ++j;
1870 if (j - i < 3 || (config->isRela && i->r_addend != 0))
1871 ungroupedNonRelatives.insert(ungroupedNonRelatives.end(), i, j);
1872 else
1873 nonRelativeGroups.emplace_back(i, j);
1874 i = j;
1875 }
1876
1877 // Sort ungrouped relocations by offset to minimize the encoded length.
1878 llvm::sort(ungroupedNonRelatives, [](const Elf_Rela &a, const Elf_Rela &b) {
1879 return a.r_offset < b.r_offset;
1880 });
1881
1882 unsigned hasAddendIfRela =
1883 config->isRela ? RELOCATION_GROUP_HAS_ADDEND_FLAG : 0;
1884
1885 uint64_t offset = 0;
1886 uint64_t addend = 0;
1887
1888 // Emit the run-length encoding for the groups of adjacent relative
1889 // relocations. Each group is represented using two groups in the packed
1890 // format. The first is used to set the current offset to the start of the
1891 // group (and also encodes the first relocation), and the second encodes the
1892 // remaining relocations.
1893 for (std::vector<Elf_Rela> &g : relativeGroups) {
1894 // The first relocation in the group.
1895 add(1);
1896 add(RELOCATION_GROUPED_BY_OFFSET_DELTA_FLAG |
1897 RELOCATION_GROUPED_BY_INFO_FLAG | hasAddendIfRela);
1898 add(g[0].r_offset - offset);
1899 add(target->relativeRel);
1900 if (config->isRela) {
1901 add(g[0].r_addend - addend);
1902 addend = g[0].r_addend;
1903 }
1904
1905 // The remaining relocations.
1906 add(g.size() - 1);
1907 add(RELOCATION_GROUPED_BY_OFFSET_DELTA_FLAG |
1908 RELOCATION_GROUPED_BY_INFO_FLAG | hasAddendIfRela);
1909 add(config->wordsize);
1910 add(target->relativeRel);
1911 if (config->isRela) {
1912 for (auto i = g.begin() + 1, e = g.end(); i != e; ++i) {
1913 add(i->r_addend - addend);
1914 addend = i->r_addend;
1915 }
1916 }
1917
1918 offset = g.back().r_offset;
1919 }
1920
1921 // Now the ungrouped relatives.
1922 if (!ungroupedRelatives.empty()) {
1923 add(ungroupedRelatives.size());
1924 add(RELOCATION_GROUPED_BY_INFO_FLAG | hasAddendIfRela);
1925 add(target->relativeRel);
1926 for (Elf_Rela &r : ungroupedRelatives) {
1927 add(r.r_offset - offset);
1928 offset = r.r_offset;
1929 if (config->isRela) {
1930 add(r.r_addend - addend);
1931 addend = r.r_addend;
1932 }
1933 }
1934 }
1935
1936 // Grouped non-relatives.
1937 for (ArrayRef<Elf_Rela> g : nonRelativeGroups) {
1938 add(g.size());
1939 add(RELOCATION_GROUPED_BY_INFO_FLAG);
1940 add(g[0].r_info);
1941 for (const Elf_Rela &r : g) {
1942 add(r.r_offset - offset);
1943 offset = r.r_offset;
1944 }
1945 addend = 0;
1946 }
1947
1948 // Finally the ungrouped non-relative relocations.
1949 if (!ungroupedNonRelatives.empty()) {
1950 add(ungroupedNonRelatives.size());
1951 add(hasAddendIfRela);
1952 for (Elf_Rela &r : ungroupedNonRelatives) {
1953 add(r.r_offset - offset);
1954 offset = r.r_offset;
1955 add(r.r_info);
1956 if (config->isRela) {
1957 add(r.r_addend - addend);
1958 addend = r.r_addend;
1959 }
1960 }
1961 }
1962
1963 // Don't allow the section to shrink; otherwise the size of the section can
1964 // oscillate infinitely.
1965 if (relocData.size() < oldSize)
1966 relocData.append(oldSize - relocData.size(), 0);
1967
1968 // Returns whether the section size changed. We need to keep recomputing both
1969 // section layout and the contents of this section until the size converges
1970 // because changing this section's size can affect section layout, which in
1971 // turn can affect the sizes of the LEB-encoded integers stored in this
1972 // section.
1973 return relocData.size() != oldSize;
1974 }
1975
RelrSection()1976 template <class ELFT> RelrSection<ELFT>::RelrSection() {
1977 this->entsize = config->wordsize;
1978 }
1979
updateAllocSize()1980 template <class ELFT> bool RelrSection<ELFT>::updateAllocSize() {
1981 // This function computes the contents of an SHT_RELR packed relocation
1982 // section.
1983 //
1984 // Proposal for adding SHT_RELR sections to generic-abi is here:
1985 // https://groups.google.com/forum/#!topic/generic-abi/bX460iggiKg
1986 //
1987 // The encoded sequence of Elf64_Relr entries in a SHT_RELR section looks
1988 // like [ AAAAAAAA BBBBBBB1 BBBBBBB1 ... AAAAAAAA BBBBBB1 ... ]
1989 //
1990 // i.e. start with an address, followed by any number of bitmaps. The address
1991 // entry encodes 1 relocation. The subsequent bitmap entries encode up to 63
1992 // relocations each, at subsequent offsets following the last address entry.
1993 //
1994 // The bitmap entries must have 1 in the least significant bit. The assumption
1995 // here is that an address cannot have 1 in lsb. Odd addresses are not
1996 // supported.
1997 //
1998 // Excluding the least significant bit in the bitmap, each non-zero bit in
1999 // the bitmap represents a relocation to be applied to a corresponding machine
2000 // word that follows the base address word. The second least significant bit
2001 // represents the machine word immediately following the initial address, and
2002 // each bit that follows represents the next word, in linear order. As such,
2003 // a single bitmap can encode up to 31 relocations in a 32-bit object, and
2004 // 63 relocations in a 64-bit object.
2005 //
2006 // This encoding has a couple of interesting properties:
2007 // 1. Looking at any entry, it is clear whether it's an address or a bitmap:
2008 // even means address, odd means bitmap.
2009 // 2. Just a simple list of addresses is a valid encoding.
2010
2011 size_t oldSize = relrRelocs.size();
2012 relrRelocs.clear();
2013
2014 // Same as Config->Wordsize but faster because this is a compile-time
2015 // constant.
2016 const size_t wordsize = sizeof(typename ELFT::uint);
2017
2018 // Number of bits to use for the relocation offsets bitmap.
2019 // Must be either 63 or 31.
2020 const size_t nBits = wordsize * 8 - 1;
2021
2022 // Get offsets for all relative relocations and sort them.
2023 std::vector<uint64_t> offsets;
2024 for (const RelativeReloc &rel : relocs)
2025 offsets.push_back(rel.getOffset());
2026 llvm::sort(offsets);
2027
2028 // For each leading relocation, find following ones that can be folded
2029 // as a bitmap and fold them.
2030 for (size_t i = 0, e = offsets.size(); i < e;) {
2031 // Add a leading relocation.
2032 relrRelocs.push_back(Elf_Relr(offsets[i]));
2033 uint64_t base = offsets[i] + wordsize;
2034 ++i;
2035
2036 // Find foldable relocations to construct bitmaps.
2037 while (i < e) {
2038 uint64_t bitmap = 0;
2039
2040 while (i < e) {
2041 uint64_t delta = offsets[i] - base;
2042
2043 // If it is too far, it cannot be folded.
2044 if (delta >= nBits * wordsize)
2045 break;
2046
2047 // If it is not a multiple of wordsize away, it cannot be folded.
2048 if (delta % wordsize)
2049 break;
2050
2051 // Fold it.
2052 bitmap |= 1ULL << (delta / wordsize);
2053 ++i;
2054 }
2055
2056 if (!bitmap)
2057 break;
2058
2059 relrRelocs.push_back(Elf_Relr((bitmap << 1) | 1));
2060 base += nBits * wordsize;
2061 }
2062 }
2063
2064 // Don't allow the section to shrink; otherwise the size of the section can
2065 // oscillate infinitely. Trailing 1s do not decode to more relocations.
2066 if (relrRelocs.size() < oldSize) {
2067 log(".relr.dyn needs " + Twine(oldSize - relrRelocs.size()) +
2068 " padding word(s)");
2069 relrRelocs.resize(oldSize, Elf_Relr(1));
2070 }
2071
2072 return relrRelocs.size() != oldSize;
2073 }
2074
SymbolTableBaseSection(StringTableSection & strTabSec)2075 SymbolTableBaseSection::SymbolTableBaseSection(StringTableSection &strTabSec)
2076 : SyntheticSection(strTabSec.isDynamic() ? (uint64_t)SHF_ALLOC : 0,
2077 strTabSec.isDynamic() ? SHT_DYNSYM : SHT_SYMTAB,
2078 config->wordsize,
2079 strTabSec.isDynamic() ? ".dynsym" : ".symtab"),
2080 strTabSec(strTabSec) {}
2081
2082 // Orders symbols according to their positions in the GOT,
2083 // in compliance with MIPS ABI rules.
2084 // See "Global Offset Table" in Chapter 5 in the following document
2085 // for detailed description:
2086 // ftp://www.linux-mips.org/pub/linux/mips/doc/ABI/mipsabi.pdf
sortMipsSymbols(const SymbolTableEntry & l,const SymbolTableEntry & r)2087 static bool sortMipsSymbols(const SymbolTableEntry &l,
2088 const SymbolTableEntry &r) {
2089 // Sort entries related to non-local preemptible symbols by GOT indexes.
2090 // All other entries go to the beginning of a dynsym in arbitrary order.
2091 if (l.sym->isInGot() && r.sym->isInGot())
2092 return l.sym->gotIndex < r.sym->gotIndex;
2093 if (!l.sym->isInGot() && !r.sym->isInGot())
2094 return false;
2095 return !l.sym->isInGot();
2096 }
2097
finalizeContents()2098 void SymbolTableBaseSection::finalizeContents() {
2099 if (OutputSection *sec = strTabSec.getParent())
2100 getParent()->link = sec->sectionIndex;
2101
2102 if (this->type != SHT_DYNSYM) {
2103 sortSymTabSymbols();
2104 return;
2105 }
2106
2107 // If it is a .dynsym, there should be no local symbols, but we need
2108 // to do a few things for the dynamic linker.
2109
2110 // Section's Info field has the index of the first non-local symbol.
2111 // Because the first symbol entry is a null entry, 1 is the first.
2112 getParent()->info = 1;
2113
2114 if (getPartition().gnuHashTab) {
2115 // NB: It also sorts Symbols to meet the GNU hash table requirements.
2116 getPartition().gnuHashTab->addSymbols(symbols);
2117 } else if (config->emachine == EM_MIPS) {
2118 llvm::stable_sort(symbols, sortMipsSymbols);
2119 }
2120
2121 // Only the main partition's dynsym indexes are stored in the symbols
2122 // themselves. All other partitions use a lookup table.
2123 if (this == mainPart->dynSymTab) {
2124 size_t i = 0;
2125 for (const SymbolTableEntry &s : symbols)
2126 s.sym->dynsymIndex = ++i;
2127 }
2128 }
2129
2130 // The ELF spec requires that all local symbols precede global symbols, so we
2131 // sort symbol entries in this function. (For .dynsym, we don't do that because
2132 // symbols for dynamic linking are inherently all globals.)
2133 //
2134 // Aside from above, we put local symbols in groups starting with the STT_FILE
2135 // symbol. That is convenient for purpose of identifying where are local symbols
2136 // coming from.
sortSymTabSymbols()2137 void SymbolTableBaseSection::sortSymTabSymbols() {
2138 // Move all local symbols before global symbols.
2139 auto e = std::stable_partition(
2140 symbols.begin(), symbols.end(), [](const SymbolTableEntry &s) {
2141 return s.sym->isLocal() || s.sym->computeBinding() == STB_LOCAL;
2142 });
2143 size_t numLocals = e - symbols.begin();
2144 getParent()->info = numLocals + 1;
2145
2146 // We want to group the local symbols by file. For that we rebuild the local
2147 // part of the symbols vector. We do not need to care about the STT_FILE
2148 // symbols, they are already naturally placed first in each group. That
2149 // happens because STT_FILE is always the first symbol in the object and hence
2150 // precede all other local symbols we add for a file.
2151 MapVector<InputFile *, std::vector<SymbolTableEntry>> arr;
2152 for (const SymbolTableEntry &s : llvm::make_range(symbols.begin(), e))
2153 arr[s.sym->file].push_back(s);
2154
2155 auto i = symbols.begin();
2156 for (std::pair<InputFile *, std::vector<SymbolTableEntry>> &p : arr)
2157 for (SymbolTableEntry &entry : p.second)
2158 *i++ = entry;
2159 }
2160
addSymbol(Symbol * b)2161 void SymbolTableBaseSection::addSymbol(Symbol *b) {
2162 // Adding a local symbol to a .dynsym is a bug.
2163 assert(this->type != SHT_DYNSYM || !b->isLocal());
2164
2165 bool hashIt = b->isLocal();
2166 symbols.push_back({b, strTabSec.addString(b->getName(), hashIt)});
2167 }
2168
getSymbolIndex(Symbol * sym)2169 size_t SymbolTableBaseSection::getSymbolIndex(Symbol *sym) {
2170 if (this == mainPart->dynSymTab)
2171 return sym->dynsymIndex;
2172
2173 // Initializes symbol lookup tables lazily. This is used only for -r,
2174 // -emit-relocs and dynsyms in partitions other than the main one.
2175 llvm::call_once(onceFlag, [&] {
2176 symbolIndexMap.reserve(symbols.size());
2177 size_t i = 0;
2178 for (const SymbolTableEntry &e : symbols) {
2179 if (e.sym->type == STT_SECTION)
2180 sectionIndexMap[e.sym->getOutputSection()] = ++i;
2181 else
2182 symbolIndexMap[e.sym] = ++i;
2183 }
2184 });
2185
2186 // Section symbols are mapped based on their output sections
2187 // to maintain their semantics.
2188 if (sym->type == STT_SECTION)
2189 return sectionIndexMap.lookup(sym->getOutputSection());
2190 return symbolIndexMap.lookup(sym);
2191 }
2192
2193 template <class ELFT>
SymbolTableSection(StringTableSection & strTabSec)2194 SymbolTableSection<ELFT>::SymbolTableSection(StringTableSection &strTabSec)
2195 : SymbolTableBaseSection(strTabSec) {
2196 this->entsize = sizeof(Elf_Sym);
2197 }
2198
getCommonSec(Symbol * sym)2199 static BssSection *getCommonSec(Symbol *sym) {
2200 if (!config->defineCommon)
2201 if (auto *d = dyn_cast<Defined>(sym))
2202 return dyn_cast_or_null<BssSection>(d->section);
2203 return nullptr;
2204 }
2205
getSymSectionIndex(Symbol * sym)2206 static uint32_t getSymSectionIndex(Symbol *sym) {
2207 if (getCommonSec(sym))
2208 return SHN_COMMON;
2209 if (!isa<Defined>(sym) || sym->needsPltAddr)
2210 return SHN_UNDEF;
2211 if (const OutputSection *os = sym->getOutputSection())
2212 return os->sectionIndex >= SHN_LORESERVE ? (uint32_t)SHN_XINDEX
2213 : os->sectionIndex;
2214 return SHN_ABS;
2215 }
2216
2217 // Write the internal symbol table contents to the output symbol table.
writeTo(uint8_t * buf)2218 template <class ELFT> void SymbolTableSection<ELFT>::writeTo(uint8_t *buf) {
2219 // The first entry is a null entry as per the ELF spec.
2220 memset(buf, 0, sizeof(Elf_Sym));
2221 buf += sizeof(Elf_Sym);
2222
2223 auto *eSym = reinterpret_cast<Elf_Sym *>(buf);
2224
2225 for (SymbolTableEntry &ent : symbols) {
2226 Symbol *sym = ent.sym;
2227 bool isDefinedHere = type == SHT_SYMTAB || sym->partition == partition;
2228
2229 // Set st_info and st_other.
2230 eSym->st_other = 0;
2231 if (sym->isLocal()) {
2232 eSym->setBindingAndType(STB_LOCAL, sym->type);
2233 } else {
2234 eSym->setBindingAndType(sym->computeBinding(), sym->type);
2235 eSym->setVisibility(sym->visibility);
2236 }
2237
2238 // The 3 most significant bits of st_other are used by OpenPOWER ABI.
2239 // See getPPC64GlobalEntryToLocalEntryOffset() for more details.
2240 if (config->emachine == EM_PPC64)
2241 eSym->st_other |= sym->stOther & 0xe0;
2242 // The most significant bit of st_other is used by AArch64 ABI for the
2243 // variant PCS.
2244 else if (config->emachine == EM_AARCH64)
2245 eSym->st_other |= sym->stOther & STO_AARCH64_VARIANT_PCS;
2246
2247 eSym->st_name = ent.strTabOffset;
2248 if (isDefinedHere)
2249 eSym->st_shndx = getSymSectionIndex(ent.sym);
2250 else
2251 eSym->st_shndx = 0;
2252
2253 // Copy symbol size if it is a defined symbol. st_size is not significant
2254 // for undefined symbols, so whether copying it or not is up to us if that's
2255 // the case. We'll leave it as zero because by not setting a value, we can
2256 // get the exact same outputs for two sets of input files that differ only
2257 // in undefined symbol size in DSOs.
2258 if (eSym->st_shndx == SHN_UNDEF || !isDefinedHere)
2259 eSym->st_size = 0;
2260 else
2261 eSym->st_size = sym->getSize();
2262
2263 // st_value is usually an address of a symbol, but that has a special
2264 // meaning for uninstantiated common symbols (--no-define-common).
2265 if (BssSection *commonSec = getCommonSec(ent.sym))
2266 eSym->st_value = commonSec->alignment;
2267 else if (isDefinedHere)
2268 eSym->st_value = sym->getVA();
2269 else
2270 eSym->st_value = 0;
2271
2272 ++eSym;
2273 }
2274
2275 // On MIPS we need to mark symbol which has a PLT entry and requires
2276 // pointer equality by STO_MIPS_PLT flag. That is necessary to help
2277 // dynamic linker distinguish such symbols and MIPS lazy-binding stubs.
2278 // https://sourceware.org/ml/binutils/2008-07/txt00000.txt
2279 if (config->emachine == EM_MIPS) {
2280 auto *eSym = reinterpret_cast<Elf_Sym *>(buf);
2281
2282 for (SymbolTableEntry &ent : symbols) {
2283 Symbol *sym = ent.sym;
2284 if (sym->isInPlt() && sym->needsPltAddr)
2285 eSym->st_other |= STO_MIPS_PLT;
2286 if (isMicroMips()) {
2287 // We already set the less-significant bit for symbols
2288 // marked by the `STO_MIPS_MICROMIPS` flag and for microMIPS PLT
2289 // records. That allows us to distinguish such symbols in
2290 // the `MIPS<ELFT>::relocate()` routine. Now we should
2291 // clear that bit for non-dynamic symbol table, so tools
2292 // like `objdump` will be able to deal with a correct
2293 // symbol position.
2294 if (sym->isDefined() &&
2295 ((sym->stOther & STO_MIPS_MICROMIPS) || sym->needsPltAddr)) {
2296 if (!strTabSec.isDynamic())
2297 eSym->st_value &= ~1;
2298 eSym->st_other |= STO_MIPS_MICROMIPS;
2299 }
2300 }
2301 if (config->relocatable)
2302 if (auto *d = dyn_cast<Defined>(sym))
2303 if (isMipsPIC<ELFT>(d))
2304 eSym->st_other |= STO_MIPS_PIC;
2305 ++eSym;
2306 }
2307 }
2308 }
2309
SymtabShndxSection()2310 SymtabShndxSection::SymtabShndxSection()
2311 : SyntheticSection(0, SHT_SYMTAB_SHNDX, 4, ".symtab_shndx") {
2312 this->entsize = 4;
2313 }
2314
writeTo(uint8_t * buf)2315 void SymtabShndxSection::writeTo(uint8_t *buf) {
2316 // We write an array of 32 bit values, where each value has 1:1 association
2317 // with an entry in .symtab. If the corresponding entry contains SHN_XINDEX,
2318 // we need to write actual index, otherwise, we must write SHN_UNDEF(0).
2319 buf += 4; // Ignore .symtab[0] entry.
2320 for (const SymbolTableEntry &entry : in.symTab->getSymbols()) {
2321 if (getSymSectionIndex(entry.sym) == SHN_XINDEX)
2322 write32(buf, entry.sym->getOutputSection()->sectionIndex);
2323 buf += 4;
2324 }
2325 }
2326
isNeeded() const2327 bool SymtabShndxSection::isNeeded() const {
2328 // SHT_SYMTAB can hold symbols with section indices values up to
2329 // SHN_LORESERVE. If we need more, we want to use extension SHT_SYMTAB_SHNDX
2330 // section. Problem is that we reveal the final section indices a bit too
2331 // late, and we do not know them here. For simplicity, we just always create
2332 // a .symtab_shndx section when the amount of output sections is huge.
2333 size_t size = 0;
2334 for (BaseCommand *base : script->sectionCommands)
2335 if (isa<OutputSection>(base))
2336 ++size;
2337 return size >= SHN_LORESERVE;
2338 }
2339
finalizeContents()2340 void SymtabShndxSection::finalizeContents() {
2341 getParent()->link = in.symTab->getParent()->sectionIndex;
2342 }
2343
getSize() const2344 size_t SymtabShndxSection::getSize() const {
2345 return in.symTab->getNumSymbols() * 4;
2346 }
2347
2348 // .hash and .gnu.hash sections contain on-disk hash tables that map
2349 // symbol names to their dynamic symbol table indices. Their purpose
2350 // is to help the dynamic linker resolve symbols quickly. If ELF files
2351 // don't have them, the dynamic linker has to do linear search on all
2352 // dynamic symbols, which makes programs slower. Therefore, a .hash
2353 // section is added to a DSO by default. A .gnu.hash is added if you
2354 // give the -hash-style=gnu or -hash-style=both option.
2355 //
2356 // The Unix semantics of resolving dynamic symbols is somewhat expensive.
2357 // Each ELF file has a list of DSOs that the ELF file depends on and a
2358 // list of dynamic symbols that need to be resolved from any of the
2359 // DSOs. That means resolving all dynamic symbols takes O(m)*O(n)
2360 // where m is the number of DSOs and n is the number of dynamic
2361 // symbols. For modern large programs, both m and n are large. So
2362 // making each step faster by using hash tables substantially
2363 // improves time to load programs.
2364 //
2365 // (Note that this is not the only way to design the shared library.
2366 // For instance, the Windows DLL takes a different approach. On
2367 // Windows, each dynamic symbol has a name of DLL from which the symbol
2368 // has to be resolved. That makes the cost of symbol resolution O(n).
2369 // This disables some hacky techniques you can use on Unix such as
2370 // LD_PRELOAD, but this is arguably better semantics than the Unix ones.)
2371 //
2372 // Due to historical reasons, we have two different hash tables, .hash
2373 // and .gnu.hash. They are for the same purpose, and .gnu.hash is a new
2374 // and better version of .hash. .hash is just an on-disk hash table, but
2375 // .gnu.hash has a bloom filter in addition to a hash table to skip
2376 // DSOs very quickly. If you are sure that your dynamic linker knows
2377 // about .gnu.hash, you want to specify -hash-style=gnu. Otherwise, a
2378 // safe bet is to specify -hash-style=both for backward compatibility.
GnuHashTableSection()2379 GnuHashTableSection::GnuHashTableSection()
2380 : SyntheticSection(SHF_ALLOC, SHT_GNU_HASH, config->wordsize, ".gnu.hash") {
2381 }
2382
finalizeContents()2383 void GnuHashTableSection::finalizeContents() {
2384 if (OutputSection *sec = getPartition().dynSymTab->getParent())
2385 getParent()->link = sec->sectionIndex;
2386
2387 // Computes bloom filter size in word size. We want to allocate 12
2388 // bits for each symbol. It must be a power of two.
2389 if (symbols.empty()) {
2390 maskWords = 1;
2391 } else {
2392 uint64_t numBits = symbols.size() * 12;
2393 maskWords = NextPowerOf2(numBits / (config->wordsize * 8));
2394 }
2395
2396 size = 16; // Header
2397 size += config->wordsize * maskWords; // Bloom filter
2398 size += nBuckets * 4; // Hash buckets
2399 size += symbols.size() * 4; // Hash values
2400 }
2401
writeTo(uint8_t * buf)2402 void GnuHashTableSection::writeTo(uint8_t *buf) {
2403 // The output buffer is not guaranteed to be zero-cleared because we pre-
2404 // fill executable sections with trap instructions. This is a precaution
2405 // for that case, which happens only when -no-rosegment is given.
2406 memset(buf, 0, size);
2407
2408 // Write a header.
2409 write32(buf, nBuckets);
2410 write32(buf + 4, getPartition().dynSymTab->getNumSymbols() - symbols.size());
2411 write32(buf + 8, maskWords);
2412 write32(buf + 12, Shift2);
2413 buf += 16;
2414
2415 // Write a bloom filter and a hash table.
2416 writeBloomFilter(buf);
2417 buf += config->wordsize * maskWords;
2418 writeHashTable(buf);
2419 }
2420
2421 // This function writes a 2-bit bloom filter. This bloom filter alone
2422 // usually filters out 80% or more of all symbol lookups [1].
2423 // The dynamic linker uses the hash table only when a symbol is not
2424 // filtered out by a bloom filter.
2425 //
2426 // [1] Ulrich Drepper (2011), "How To Write Shared Libraries" (Ver. 4.1.2),
2427 // p.9, https://www.akkadia.org/drepper/dsohowto.pdf
writeBloomFilter(uint8_t * buf)2428 void GnuHashTableSection::writeBloomFilter(uint8_t *buf) {
2429 unsigned c = config->is64 ? 64 : 32;
2430 for (const Entry &sym : symbols) {
2431 // When C = 64, we choose a word with bits [6:...] and set 1 to two bits in
2432 // the word using bits [0:5] and [26:31].
2433 size_t i = (sym.hash / c) & (maskWords - 1);
2434 uint64_t val = readUint(buf + i * config->wordsize);
2435 val |= uint64_t(1) << (sym.hash % c);
2436 val |= uint64_t(1) << ((sym.hash >> Shift2) % c);
2437 writeUint(buf + i * config->wordsize, val);
2438 }
2439 }
2440
writeHashTable(uint8_t * buf)2441 void GnuHashTableSection::writeHashTable(uint8_t *buf) {
2442 uint32_t *buckets = reinterpret_cast<uint32_t *>(buf);
2443 uint32_t oldBucket = -1;
2444 uint32_t *values = buckets + nBuckets;
2445 for (auto i = symbols.begin(), e = symbols.end(); i != e; ++i) {
2446 // Write a hash value. It represents a sequence of chains that share the
2447 // same hash modulo value. The last element of each chain is terminated by
2448 // LSB 1.
2449 uint32_t hash = i->hash;
2450 bool isLastInChain = (i + 1) == e || i->bucketIdx != (i + 1)->bucketIdx;
2451 hash = isLastInChain ? hash | 1 : hash & ~1;
2452 write32(values++, hash);
2453
2454 if (i->bucketIdx == oldBucket)
2455 continue;
2456 // Write a hash bucket. Hash buckets contain indices in the following hash
2457 // value table.
2458 write32(buckets + i->bucketIdx,
2459 getPartition().dynSymTab->getSymbolIndex(i->sym));
2460 oldBucket = i->bucketIdx;
2461 }
2462 }
2463
hashGnu(StringRef name)2464 static uint32_t hashGnu(StringRef name) {
2465 uint32_t h = 5381;
2466 for (uint8_t c : name)
2467 h = (h << 5) + h + c;
2468 return h;
2469 }
2470
2471 // Add symbols to this symbol hash table. Note that this function
2472 // destructively sort a given vector -- which is needed because
2473 // GNU-style hash table places some sorting requirements.
addSymbols(std::vector<SymbolTableEntry> & v)2474 void GnuHashTableSection::addSymbols(std::vector<SymbolTableEntry> &v) {
2475 // We cannot use 'auto' for Mid because GCC 6.1 cannot deduce
2476 // its type correctly.
2477 std::vector<SymbolTableEntry>::iterator mid =
2478 std::stable_partition(v.begin(), v.end(), [&](const SymbolTableEntry &s) {
2479 return !s.sym->isDefined() || s.sym->partition != partition;
2480 });
2481
2482 // We chose load factor 4 for the on-disk hash table. For each hash
2483 // collision, the dynamic linker will compare a uint32_t hash value.
2484 // Since the integer comparison is quite fast, we believe we can
2485 // make the load factor even larger. 4 is just a conservative choice.
2486 //
2487 // Note that we don't want to create a zero-sized hash table because
2488 // Android loader as of 2018 doesn't like a .gnu.hash containing such
2489 // table. If that's the case, we create a hash table with one unused
2490 // dummy slot.
2491 nBuckets = std::max<size_t>((v.end() - mid) / 4, 1);
2492
2493 if (mid == v.end())
2494 return;
2495
2496 for (SymbolTableEntry &ent : llvm::make_range(mid, v.end())) {
2497 Symbol *b = ent.sym;
2498 uint32_t hash = hashGnu(b->getName());
2499 uint32_t bucketIdx = hash % nBuckets;
2500 symbols.push_back({b, ent.strTabOffset, hash, bucketIdx});
2501 }
2502
2503 llvm::stable_sort(symbols, [](const Entry &l, const Entry &r) {
2504 return l.bucketIdx < r.bucketIdx;
2505 });
2506
2507 v.erase(mid, v.end());
2508 for (const Entry &ent : symbols)
2509 v.push_back({ent.sym, ent.strTabOffset});
2510 }
2511
HashTableSection()2512 HashTableSection::HashTableSection()
2513 : SyntheticSection(SHF_ALLOC, SHT_HASH, 4, ".hash") {
2514 this->entsize = 4;
2515 }
2516
finalizeContents()2517 void HashTableSection::finalizeContents() {
2518 SymbolTableBaseSection *symTab = getPartition().dynSymTab;
2519
2520 if (OutputSection *sec = symTab->getParent())
2521 getParent()->link = sec->sectionIndex;
2522
2523 unsigned numEntries = 2; // nbucket and nchain.
2524 numEntries += symTab->getNumSymbols(); // The chain entries.
2525
2526 // Create as many buckets as there are symbols.
2527 numEntries += symTab->getNumSymbols();
2528 this->size = numEntries * 4;
2529 }
2530
writeTo(uint8_t * buf)2531 void HashTableSection::writeTo(uint8_t *buf) {
2532 SymbolTableBaseSection *symTab = getPartition().dynSymTab;
2533
2534 // See comment in GnuHashTableSection::writeTo.
2535 memset(buf, 0, size);
2536
2537 unsigned numSymbols = symTab->getNumSymbols();
2538
2539 uint32_t *p = reinterpret_cast<uint32_t *>(buf);
2540 write32(p++, numSymbols); // nbucket
2541 write32(p++, numSymbols); // nchain
2542
2543 uint32_t *buckets = p;
2544 uint32_t *chains = p + numSymbols;
2545
2546 for (const SymbolTableEntry &s : symTab->getSymbols()) {
2547 Symbol *sym = s.sym;
2548 StringRef name = sym->getName();
2549 unsigned i = sym->dynsymIndex;
2550 uint32_t hash = hashSysV(name) % numSymbols;
2551 chains[i] = buckets[hash];
2552 write32(buckets + hash, i);
2553 }
2554 }
2555
PltSection()2556 PltSection::PltSection()
2557 : SyntheticSection(SHF_ALLOC | SHF_EXECINSTR, SHT_PROGBITS, 16, ".plt"),
2558 headerSize(target->pltHeaderSize) {
2559 // On PowerPC, this section contains lazy symbol resolvers.
2560 if (config->emachine == EM_PPC64) {
2561 name = ".glink";
2562 alignment = 4;
2563 }
2564
2565 // On x86 when IBT is enabled, this section contains the second PLT (lazy
2566 // symbol resolvers).
2567 if ((config->emachine == EM_386 || config->emachine == EM_X86_64) &&
2568 (config->andFeatures & GNU_PROPERTY_X86_FEATURE_1_IBT))
2569 name = ".plt.sec";
2570
2571 // The PLT needs to be writable on SPARC as the dynamic linker will
2572 // modify the instructions in the PLT entries.
2573 if (config->emachine == EM_SPARCV9)
2574 this->flags |= SHF_WRITE;
2575 }
2576
writeTo(uint8_t * buf)2577 void PltSection::writeTo(uint8_t *buf) {
2578 // At beginning of PLT, we have code to call the dynamic
2579 // linker to resolve dynsyms at runtime. Write such code.
2580 target->writePltHeader(buf);
2581 size_t off = headerSize;
2582
2583 for (const Symbol *sym : entries) {
2584 target->writePlt(buf + off, *sym, getVA() + off);
2585 off += target->pltEntrySize;
2586 }
2587 }
2588
addEntry(Symbol & sym)2589 void PltSection::addEntry(Symbol &sym) {
2590 sym.pltIndex = entries.size();
2591 entries.push_back(&sym);
2592 }
2593
getSize() const2594 size_t PltSection::getSize() const {
2595 return headerSize + entries.size() * target->pltEntrySize;
2596 }
2597
isNeeded() const2598 bool PltSection::isNeeded() const {
2599 // For -z retpolineplt, .iplt needs the .plt header.
2600 return !entries.empty() || (config->zRetpolineplt && in.iplt->isNeeded());
2601 }
2602
2603 // Used by ARM to add mapping symbols in the PLT section, which aid
2604 // disassembly.
addSymbols()2605 void PltSection::addSymbols() {
2606 target->addPltHeaderSymbols(*this);
2607
2608 size_t off = headerSize;
2609 for (size_t i = 0; i < entries.size(); ++i) {
2610 target->addPltSymbols(*this, off);
2611 off += target->pltEntrySize;
2612 }
2613 }
2614
IpltSection()2615 IpltSection::IpltSection()
2616 : SyntheticSection(SHF_ALLOC | SHF_EXECINSTR, SHT_PROGBITS, 16, ".iplt") {
2617 if (config->emachine == EM_PPC || config->emachine == EM_PPC64) {
2618 name = ".glink";
2619 alignment = 4;
2620 }
2621 }
2622
writeTo(uint8_t * buf)2623 void IpltSection::writeTo(uint8_t *buf) {
2624 uint32_t off = 0;
2625 for (const Symbol *sym : entries) {
2626 target->writeIplt(buf + off, *sym, getVA() + off);
2627 off += target->ipltEntrySize;
2628 }
2629 }
2630
getSize() const2631 size_t IpltSection::getSize() const {
2632 return entries.size() * target->ipltEntrySize;
2633 }
2634
addEntry(Symbol & sym)2635 void IpltSection::addEntry(Symbol &sym) {
2636 sym.pltIndex = entries.size();
2637 entries.push_back(&sym);
2638 }
2639
2640 // ARM uses mapping symbols to aid disassembly.
addSymbols()2641 void IpltSection::addSymbols() {
2642 size_t off = 0;
2643 for (size_t i = 0, e = entries.size(); i != e; ++i) {
2644 target->addPltSymbols(*this, off);
2645 off += target->pltEntrySize;
2646 }
2647 }
2648
PPC32GlinkSection()2649 PPC32GlinkSection::PPC32GlinkSection() {
2650 name = ".glink";
2651 alignment = 4;
2652 }
2653
writeTo(uint8_t * buf)2654 void PPC32GlinkSection::writeTo(uint8_t *buf) {
2655 writePPC32GlinkSection(buf, entries.size());
2656 }
2657
getSize() const2658 size_t PPC32GlinkSection::getSize() const {
2659 return headerSize + entries.size() * target->pltEntrySize + footerSize;
2660 }
2661
2662 // This is an x86-only extra PLT section and used only when a security
2663 // enhancement feature called CET is enabled. In this comment, I'll explain what
2664 // the feature is and why we have two PLT sections if CET is enabled.
2665 //
2666 // So, what does CET do? CET introduces a new restriction to indirect jump
2667 // instructions. CET works this way. Assume that CET is enabled. Then, if you
2668 // execute an indirect jump instruction, the processor verifies that a special
2669 // "landing pad" instruction (which is actually a repurposed NOP instruction and
2670 // now called "endbr32" or "endbr64") is at the jump target. If the jump target
2671 // does not start with that instruction, the processor raises an exception
2672 // instead of continuing executing code.
2673 //
2674 // If CET is enabled, the compiler emits endbr to all locations where indirect
2675 // jumps may jump to.
2676 //
2677 // This mechanism makes it extremely hard to transfer the control to a middle of
2678 // a function that is not supporsed to be a indirect jump target, preventing
2679 // certain types of attacks such as ROP or JOP.
2680 //
2681 // Note that the processors in the market as of 2019 don't actually support the
2682 // feature. Only the spec is available at the moment.
2683 //
2684 // Now, I'll explain why we have this extra PLT section for CET.
2685 //
2686 // Since you can indirectly jump to a PLT entry, we have to make PLT entries
2687 // start with endbr. The problem is there's no extra space for endbr (which is 4
2688 // bytes long), as the PLT entry is only 16 bytes long and all bytes are already
2689 // used.
2690 //
2691 // In order to deal with the issue, we split a PLT entry into two PLT entries.
2692 // Remember that each PLT entry contains code to jump to an address read from
2693 // .got.plt AND code to resolve a dynamic symbol lazily. With the 2-PLT scheme,
2694 // the former code is written to .plt.sec, and the latter code is written to
2695 // .plt.
2696 //
2697 // Lazy symbol resolution in the 2-PLT scheme works in the usual way, except
2698 // that the regular .plt is now called .plt.sec and .plt is repurposed to
2699 // contain only code for lazy symbol resolution.
2700 //
2701 // In other words, this is how the 2-PLT scheme works. Application code is
2702 // supposed to jump to .plt.sec to call an external function. Each .plt.sec
2703 // entry contains code to read an address from a corresponding .got.plt entry
2704 // and jump to that address. Addresses in .got.plt initially point to .plt, so
2705 // when an application calls an external function for the first time, the
2706 // control is transferred to a function that resolves a symbol name from
2707 // external shared object files. That function then rewrites a .got.plt entry
2708 // with a resolved address, so that the subsequent function calls directly jump
2709 // to a desired location from .plt.sec.
2710 //
2711 // There is an open question as to whether the 2-PLT scheme was desirable or
2712 // not. We could have simply extended the PLT entry size to 32-bytes to
2713 // accommodate endbr, and that scheme would have been much simpler than the
2714 // 2-PLT scheme. One reason to split PLT was, by doing that, we could keep hot
2715 // code (.plt.sec) from cold code (.plt). But as far as I know no one proved
2716 // that the optimization actually makes a difference.
2717 //
2718 // That said, the 2-PLT scheme is a part of the ABI, debuggers and other tools
2719 // depend on it, so we implement the ABI.
IBTPltSection()2720 IBTPltSection::IBTPltSection()
2721 : SyntheticSection(SHF_ALLOC | SHF_EXECINSTR, SHT_PROGBITS, 16, ".plt") {}
2722
writeTo(uint8_t * buf)2723 void IBTPltSection::writeTo(uint8_t *buf) {
2724 target->writeIBTPlt(buf, in.plt->getNumEntries());
2725 }
2726
getSize() const2727 size_t IBTPltSection::getSize() const {
2728 // 16 is the header size of .plt.
2729 return 16 + in.plt->getNumEntries() * target->pltEntrySize;
2730 }
2731
2732 // The string hash function for .gdb_index.
computeGdbHash(StringRef s)2733 static uint32_t computeGdbHash(StringRef s) {
2734 uint32_t h = 0;
2735 for (uint8_t c : s)
2736 h = h * 67 + toLower(c) - 113;
2737 return h;
2738 }
2739
GdbIndexSection()2740 GdbIndexSection::GdbIndexSection()
2741 : SyntheticSection(0, SHT_PROGBITS, 1, ".gdb_index") {}
2742
2743 // Returns the desired size of an on-disk hash table for a .gdb_index section.
2744 // There's a tradeoff between size and collision rate. We aim 75% utilization.
computeSymtabSize() const2745 size_t GdbIndexSection::computeSymtabSize() const {
2746 return std::max<size_t>(NextPowerOf2(symbols.size() * 4 / 3), 1024);
2747 }
2748
2749 // Compute the output section size.
initOutputSize()2750 void GdbIndexSection::initOutputSize() {
2751 size = sizeof(GdbIndexHeader) + computeSymtabSize() * 8;
2752
2753 for (GdbChunk &chunk : chunks)
2754 size += chunk.compilationUnits.size() * 16 + chunk.addressAreas.size() * 20;
2755
2756 // Add the constant pool size if exists.
2757 if (!symbols.empty()) {
2758 GdbSymbol &sym = symbols.back();
2759 size += sym.nameOff + sym.name.size() + 1;
2760 }
2761 }
2762
readCuList(DWARFContext & dwarf)2763 static std::vector<GdbIndexSection::CuEntry> readCuList(DWARFContext &dwarf) {
2764 std::vector<GdbIndexSection::CuEntry> ret;
2765 for (std::unique_ptr<DWARFUnit> &cu : dwarf.compile_units())
2766 ret.push_back({cu->getOffset(), cu->getLength() + 4});
2767 return ret;
2768 }
2769
2770 static std::vector<GdbIndexSection::AddressEntry>
readAddressAreas(DWARFContext & dwarf,InputSection * sec)2771 readAddressAreas(DWARFContext &dwarf, InputSection *sec) {
2772 std::vector<GdbIndexSection::AddressEntry> ret;
2773
2774 uint32_t cuIdx = 0;
2775 for (std::unique_ptr<DWARFUnit> &cu : dwarf.compile_units()) {
2776 if (Error e = cu->tryExtractDIEsIfNeeded(false)) {
2777 warn(toString(sec) + ": " + toString(std::move(e)));
2778 return {};
2779 }
2780 Expected<DWARFAddressRangesVector> ranges = cu->collectAddressRanges();
2781 if (!ranges) {
2782 warn(toString(sec) + ": " + toString(ranges.takeError()));
2783 return {};
2784 }
2785
2786 ArrayRef<InputSectionBase *> sections = sec->file->getSections();
2787 for (DWARFAddressRange &r : *ranges) {
2788 if (r.SectionIndex == -1ULL)
2789 continue;
2790 // Range list with zero size has no effect.
2791 InputSectionBase *s = sections[r.SectionIndex];
2792 if (s && s != &InputSection::discarded && s->isLive())
2793 if (r.LowPC != r.HighPC)
2794 ret.push_back({cast<InputSection>(s), r.LowPC, r.HighPC, cuIdx});
2795 }
2796 ++cuIdx;
2797 }
2798
2799 return ret;
2800 }
2801
2802 template <class ELFT>
2803 static std::vector<GdbIndexSection::NameAttrEntry>
readPubNamesAndTypes(const LLDDwarfObj<ELFT> & obj,const std::vector<GdbIndexSection::CuEntry> & cus)2804 readPubNamesAndTypes(const LLDDwarfObj<ELFT> &obj,
2805 const std::vector<GdbIndexSection::CuEntry> &cus) {
2806 const LLDDWARFSection &pubNames = obj.getGnuPubnamesSection();
2807 const LLDDWARFSection &pubTypes = obj.getGnuPubtypesSection();
2808
2809 std::vector<GdbIndexSection::NameAttrEntry> ret;
2810 for (const LLDDWARFSection *pub : {&pubNames, &pubTypes}) {
2811 DWARFDataExtractor data(obj, *pub, config->isLE, config->wordsize);
2812 DWARFDebugPubTable table;
2813 table.extract(data, /*GnuStyle=*/true, [&](Error e) {
2814 warn(toString(pub->sec) + ": " + toString(std::move(e)));
2815 });
2816 for (const DWARFDebugPubTable::Set &set : table.getData()) {
2817 // The value written into the constant pool is kind << 24 | cuIndex. As we
2818 // don't know how many compilation units precede this object to compute
2819 // cuIndex, we compute (kind << 24 | cuIndexInThisObject) instead, and add
2820 // the number of preceding compilation units later.
2821 uint32_t i = llvm::partition_point(cus,
2822 [&](GdbIndexSection::CuEntry cu) {
2823 return cu.cuOffset < set.Offset;
2824 }) -
2825 cus.begin();
2826 for (const DWARFDebugPubTable::Entry &ent : set.Entries)
2827 ret.push_back({{ent.Name, computeGdbHash(ent.Name)},
2828 (ent.Descriptor.toBits() << 24) | i});
2829 }
2830 }
2831 return ret;
2832 }
2833
2834 // Create a list of symbols from a given list of symbol names and types
2835 // by uniquifying them by name.
2836 static std::vector<GdbIndexSection::GdbSymbol>
createSymbols(ArrayRef<std::vector<GdbIndexSection::NameAttrEntry>> nameAttrs,const std::vector<GdbIndexSection::GdbChunk> & chunks)2837 createSymbols(ArrayRef<std::vector<GdbIndexSection::NameAttrEntry>> nameAttrs,
2838 const std::vector<GdbIndexSection::GdbChunk> &chunks) {
2839 using GdbSymbol = GdbIndexSection::GdbSymbol;
2840 using NameAttrEntry = GdbIndexSection::NameAttrEntry;
2841
2842 // For each chunk, compute the number of compilation units preceding it.
2843 uint32_t cuIdx = 0;
2844 std::vector<uint32_t> cuIdxs(chunks.size());
2845 for (uint32_t i = 0, e = chunks.size(); i != e; ++i) {
2846 cuIdxs[i] = cuIdx;
2847 cuIdx += chunks[i].compilationUnits.size();
2848 }
2849
2850 // The number of symbols we will handle in this function is of the order
2851 // of millions for very large executables, so we use multi-threading to
2852 // speed it up.
2853 constexpr size_t numShards = 32;
2854 size_t concurrency = PowerOf2Floor(
2855 std::min<size_t>(hardware_concurrency(parallel::strategy.ThreadsRequested)
2856 .compute_thread_count(),
2857 numShards));
2858
2859 // A sharded map to uniquify symbols by name.
2860 std::vector<DenseMap<CachedHashStringRef, size_t>> map(numShards);
2861 size_t shift = 32 - countTrailingZeros(numShards);
2862
2863 // Instantiate GdbSymbols while uniqufying them by name.
2864 std::vector<std::vector<GdbSymbol>> symbols(numShards);
2865 parallelForEachN(0, concurrency, [&](size_t threadId) {
2866 uint32_t i = 0;
2867 for (ArrayRef<NameAttrEntry> entries : nameAttrs) {
2868 for (const NameAttrEntry &ent : entries) {
2869 size_t shardId = ent.name.hash() >> shift;
2870 if ((shardId & (concurrency - 1)) != threadId)
2871 continue;
2872
2873 uint32_t v = ent.cuIndexAndAttrs + cuIdxs[i];
2874 size_t &idx = map[shardId][ent.name];
2875 if (idx) {
2876 symbols[shardId][idx - 1].cuVector.push_back(v);
2877 continue;
2878 }
2879
2880 idx = symbols[shardId].size() + 1;
2881 symbols[shardId].push_back({ent.name, {v}, 0, 0});
2882 }
2883 ++i;
2884 }
2885 });
2886
2887 size_t numSymbols = 0;
2888 for (ArrayRef<GdbSymbol> v : symbols)
2889 numSymbols += v.size();
2890
2891 // The return type is a flattened vector, so we'll copy each vector
2892 // contents to Ret.
2893 std::vector<GdbSymbol> ret;
2894 ret.reserve(numSymbols);
2895 for (std::vector<GdbSymbol> &vec : symbols)
2896 for (GdbSymbol &sym : vec)
2897 ret.push_back(std::move(sym));
2898
2899 // CU vectors and symbol names are adjacent in the output file.
2900 // We can compute their offsets in the output file now.
2901 size_t off = 0;
2902 for (GdbSymbol &sym : ret) {
2903 sym.cuVectorOff = off;
2904 off += (sym.cuVector.size() + 1) * 4;
2905 }
2906 for (GdbSymbol &sym : ret) {
2907 sym.nameOff = off;
2908 off += sym.name.size() + 1;
2909 }
2910
2911 return ret;
2912 }
2913
2914 // Returns a newly-created .gdb_index section.
create()2915 template <class ELFT> GdbIndexSection *GdbIndexSection::create() {
2916 // Collect InputFiles with .debug_info. See the comment in
2917 // LLDDwarfObj<ELFT>::LLDDwarfObj. If we do lightweight parsing in the future,
2918 // note that isec->data() may uncompress the full content, which should be
2919 // parallelized.
2920 SetVector<InputFile *> files;
2921 for (InputSectionBase *s : inputSections) {
2922 InputSection *isec = dyn_cast<InputSection>(s);
2923 if (!isec)
2924 continue;
2925 // .debug_gnu_pub{names,types} are useless in executables.
2926 // They are present in input object files solely for creating
2927 // a .gdb_index. So we can remove them from the output.
2928 if (s->name == ".debug_gnu_pubnames" || s->name == ".debug_gnu_pubtypes")
2929 s->markDead();
2930 else if (isec->name == ".debug_info")
2931 files.insert(isec->file);
2932 }
2933 // Drop .rel[a].debug_gnu_pub{names,types} for --emit-relocs.
2934 llvm::erase_if(inputSections, [](InputSectionBase *s) {
2935 if (auto *isec = dyn_cast<InputSection>(s))
2936 if (InputSectionBase *rel = isec->getRelocatedSection())
2937 return !rel->isLive();
2938 return !s->isLive();
2939 });
2940
2941 std::vector<GdbChunk> chunks(files.size());
2942 std::vector<std::vector<NameAttrEntry>> nameAttrs(files.size());
2943
2944 parallelForEachN(0, files.size(), [&](size_t i) {
2945 // To keep memory usage low, we don't want to keep cached DWARFContext, so
2946 // avoid getDwarf() here.
2947 ObjFile<ELFT> *file = cast<ObjFile<ELFT>>(files[i]);
2948 DWARFContext dwarf(std::make_unique<LLDDwarfObj<ELFT>>(file));
2949 auto &dobj = static_cast<const LLDDwarfObj<ELFT> &>(dwarf.getDWARFObj());
2950
2951 // If the are multiple compile units .debug_info (very rare ld -r --unique),
2952 // this only picks the last one. Other address ranges are lost.
2953 chunks[i].sec = dobj.getInfoSection();
2954 chunks[i].compilationUnits = readCuList(dwarf);
2955 chunks[i].addressAreas = readAddressAreas(dwarf, chunks[i].sec);
2956 nameAttrs[i] = readPubNamesAndTypes<ELFT>(dobj, chunks[i].compilationUnits);
2957 });
2958
2959 auto *ret = make<GdbIndexSection>();
2960 ret->chunks = std::move(chunks);
2961 ret->symbols = createSymbols(nameAttrs, ret->chunks);
2962 ret->initOutputSize();
2963 return ret;
2964 }
2965
writeTo(uint8_t * buf)2966 void GdbIndexSection::writeTo(uint8_t *buf) {
2967 // Write the header.
2968 auto *hdr = reinterpret_cast<GdbIndexHeader *>(buf);
2969 uint8_t *start = buf;
2970 hdr->version = 7;
2971 buf += sizeof(*hdr);
2972
2973 // Write the CU list.
2974 hdr->cuListOff = buf - start;
2975 for (GdbChunk &chunk : chunks) {
2976 for (CuEntry &cu : chunk.compilationUnits) {
2977 write64le(buf, chunk.sec->outSecOff + cu.cuOffset);
2978 write64le(buf + 8, cu.cuLength);
2979 buf += 16;
2980 }
2981 }
2982
2983 // Write the address area.
2984 hdr->cuTypesOff = buf - start;
2985 hdr->addressAreaOff = buf - start;
2986 uint32_t cuOff = 0;
2987 for (GdbChunk &chunk : chunks) {
2988 for (AddressEntry &e : chunk.addressAreas) {
2989 // In the case of ICF there may be duplicate address range entries.
2990 const uint64_t baseAddr = e.section->repl->getVA(0);
2991 write64le(buf, baseAddr + e.lowAddress);
2992 write64le(buf + 8, baseAddr + e.highAddress);
2993 write32le(buf + 16, e.cuIndex + cuOff);
2994 buf += 20;
2995 }
2996 cuOff += chunk.compilationUnits.size();
2997 }
2998
2999 // Write the on-disk open-addressing hash table containing symbols.
3000 hdr->symtabOff = buf - start;
3001 size_t symtabSize = computeSymtabSize();
3002 uint32_t mask = symtabSize - 1;
3003
3004 for (GdbSymbol &sym : symbols) {
3005 uint32_t h = sym.name.hash();
3006 uint32_t i = h & mask;
3007 uint32_t step = ((h * 17) & mask) | 1;
3008
3009 while (read32le(buf + i * 8))
3010 i = (i + step) & mask;
3011
3012 write32le(buf + i * 8, sym.nameOff);
3013 write32le(buf + i * 8 + 4, sym.cuVectorOff);
3014 }
3015
3016 buf += symtabSize * 8;
3017
3018 // Write the string pool.
3019 hdr->constantPoolOff = buf - start;
3020 parallelForEach(symbols, [&](GdbSymbol &sym) {
3021 memcpy(buf + sym.nameOff, sym.name.data(), sym.name.size());
3022 });
3023
3024 // Write the CU vectors.
3025 for (GdbSymbol &sym : symbols) {
3026 write32le(buf, sym.cuVector.size());
3027 buf += 4;
3028 for (uint32_t val : sym.cuVector) {
3029 write32le(buf, val);
3030 buf += 4;
3031 }
3032 }
3033 }
3034
isNeeded() const3035 bool GdbIndexSection::isNeeded() const { return !chunks.empty(); }
3036
EhFrameHeader()3037 EhFrameHeader::EhFrameHeader()
3038 : SyntheticSection(SHF_ALLOC, SHT_PROGBITS, 4, ".eh_frame_hdr") {}
3039
writeTo(uint8_t * buf)3040 void EhFrameHeader::writeTo(uint8_t *buf) {
3041 // Unlike most sections, the EhFrameHeader section is written while writing
3042 // another section, namely EhFrameSection, which calls the write() function
3043 // below from its writeTo() function. This is necessary because the contents
3044 // of EhFrameHeader depend on the relocated contents of EhFrameSection and we
3045 // don't know which order the sections will be written in.
3046 }
3047
3048 // .eh_frame_hdr contains a binary search table of pointers to FDEs.
3049 // Each entry of the search table consists of two values,
3050 // the starting PC from where FDEs covers, and the FDE's address.
3051 // It is sorted by PC.
write()3052 void EhFrameHeader::write() {
3053 uint8_t *buf = Out::bufferStart + getParent()->offset + outSecOff;
3054 using FdeData = EhFrameSection::FdeData;
3055
3056 std::vector<FdeData> fdes = getPartition().ehFrame->getFdeData();
3057
3058 buf[0] = 1;
3059 buf[1] = DW_EH_PE_pcrel | DW_EH_PE_sdata4;
3060 buf[2] = DW_EH_PE_udata4;
3061 buf[3] = DW_EH_PE_datarel | DW_EH_PE_sdata4;
3062 write32(buf + 4,
3063 getPartition().ehFrame->getParent()->addr - this->getVA() - 4);
3064 write32(buf + 8, fdes.size());
3065 buf += 12;
3066
3067 for (FdeData &fde : fdes) {
3068 write32(buf, fde.pcRel);
3069 write32(buf + 4, fde.fdeVARel);
3070 buf += 8;
3071 }
3072 }
3073
getSize() const3074 size_t EhFrameHeader::getSize() const {
3075 // .eh_frame_hdr has a 12 bytes header followed by an array of FDEs.
3076 return 12 + getPartition().ehFrame->numFdes * 8;
3077 }
3078
isNeeded() const3079 bool EhFrameHeader::isNeeded() const {
3080 return isLive() && getPartition().ehFrame->isNeeded();
3081 }
3082
VersionDefinitionSection()3083 VersionDefinitionSection::VersionDefinitionSection()
3084 : SyntheticSection(SHF_ALLOC, SHT_GNU_verdef, sizeof(uint32_t),
3085 ".gnu.version_d") {}
3086
getFileDefName()3087 StringRef VersionDefinitionSection::getFileDefName() {
3088 if (!getPartition().name.empty())
3089 return getPartition().name;
3090 if (!config->soName.empty())
3091 return config->soName;
3092 return config->outputFile;
3093 }
3094
finalizeContents()3095 void VersionDefinitionSection::finalizeContents() {
3096 fileDefNameOff = getPartition().dynStrTab->addString(getFileDefName());
3097 for (const VersionDefinition &v : namedVersionDefs())
3098 verDefNameOffs.push_back(getPartition().dynStrTab->addString(v.name));
3099
3100 if (OutputSection *sec = getPartition().dynStrTab->getParent())
3101 getParent()->link = sec->sectionIndex;
3102
3103 // sh_info should be set to the number of definitions. This fact is missed in
3104 // documentation, but confirmed by binutils community:
3105 // https://sourceware.org/ml/binutils/2014-11/msg00355.html
3106 getParent()->info = getVerDefNum();
3107 }
3108
writeOne(uint8_t * buf,uint32_t index,StringRef name,size_t nameOff)3109 void VersionDefinitionSection::writeOne(uint8_t *buf, uint32_t index,
3110 StringRef name, size_t nameOff) {
3111 uint16_t flags = index == 1 ? VER_FLG_BASE : 0;
3112
3113 // Write a verdef.
3114 write16(buf, 1); // vd_version
3115 write16(buf + 2, flags); // vd_flags
3116 write16(buf + 4, index); // vd_ndx
3117 write16(buf + 6, 1); // vd_cnt
3118 write32(buf + 8, hashSysV(name)); // vd_hash
3119 write32(buf + 12, 20); // vd_aux
3120 write32(buf + 16, 28); // vd_next
3121
3122 // Write a veraux.
3123 write32(buf + 20, nameOff); // vda_name
3124 write32(buf + 24, 0); // vda_next
3125 }
3126
writeTo(uint8_t * buf)3127 void VersionDefinitionSection::writeTo(uint8_t *buf) {
3128 writeOne(buf, 1, getFileDefName(), fileDefNameOff);
3129
3130 auto nameOffIt = verDefNameOffs.begin();
3131 for (const VersionDefinition &v : namedVersionDefs()) {
3132 buf += EntrySize;
3133 writeOne(buf, v.id, v.name, *nameOffIt++);
3134 }
3135
3136 // Need to terminate the last version definition.
3137 write32(buf + 16, 0); // vd_next
3138 }
3139
getSize() const3140 size_t VersionDefinitionSection::getSize() const {
3141 return EntrySize * getVerDefNum();
3142 }
3143
3144 // .gnu.version is a table where each entry is 2 byte long.
VersionTableSection()3145 VersionTableSection::VersionTableSection()
3146 : SyntheticSection(SHF_ALLOC, SHT_GNU_versym, sizeof(uint16_t),
3147 ".gnu.version") {
3148 this->entsize = 2;
3149 }
3150
finalizeContents()3151 void VersionTableSection::finalizeContents() {
3152 // At the moment of june 2016 GNU docs does not mention that sh_link field
3153 // should be set, but Sun docs do. Also readelf relies on this field.
3154 getParent()->link = getPartition().dynSymTab->getParent()->sectionIndex;
3155 }
3156
getSize() const3157 size_t VersionTableSection::getSize() const {
3158 return (getPartition().dynSymTab->getSymbols().size() + 1) * 2;
3159 }
3160
writeTo(uint8_t * buf)3161 void VersionTableSection::writeTo(uint8_t *buf) {
3162 buf += 2;
3163 for (const SymbolTableEntry &s : getPartition().dynSymTab->getSymbols()) {
3164 // Use the original versionId for an unfetched lazy symbol (undefined weak),
3165 // which must be VER_NDX_GLOBAL (an undefined versioned symbol is an error).
3166 write16(buf, s.sym->isLazy() ? static_cast<uint16_t>(VER_NDX_GLOBAL)
3167 : s.sym->versionId);
3168 buf += 2;
3169 }
3170 }
3171
isNeeded() const3172 bool VersionTableSection::isNeeded() const {
3173 return isLive() &&
3174 (getPartition().verDef || getPartition().verNeed->isNeeded());
3175 }
3176
addVerneed(Symbol * ss)3177 void elf::addVerneed(Symbol *ss) {
3178 auto &file = cast<SharedFile>(*ss->file);
3179 if (ss->verdefIndex == VER_NDX_GLOBAL) {
3180 ss->versionId = VER_NDX_GLOBAL;
3181 return;
3182 }
3183
3184 if (file.vernauxs.empty())
3185 file.vernauxs.resize(file.verdefs.size());
3186
3187 // Select a version identifier for the vernaux data structure, if we haven't
3188 // already allocated one. The verdef identifiers cover the range
3189 // [1..getVerDefNum()]; this causes the vernaux identifiers to start from
3190 // getVerDefNum()+1.
3191 if (file.vernauxs[ss->verdefIndex] == 0)
3192 file.vernauxs[ss->verdefIndex] = ++SharedFile::vernauxNum + getVerDefNum();
3193
3194 ss->versionId = file.vernauxs[ss->verdefIndex];
3195 }
3196
3197 template <class ELFT>
VersionNeedSection()3198 VersionNeedSection<ELFT>::VersionNeedSection()
3199 : SyntheticSection(SHF_ALLOC, SHT_GNU_verneed, sizeof(uint32_t),
3200 ".gnu.version_r") {}
3201
finalizeContents()3202 template <class ELFT> void VersionNeedSection<ELFT>::finalizeContents() {
3203 for (SharedFile *f : sharedFiles) {
3204 if (f->vernauxs.empty())
3205 continue;
3206 verneeds.emplace_back();
3207 Verneed &vn = verneeds.back();
3208 vn.nameStrTab = getPartition().dynStrTab->addString(f->soName);
3209 for (unsigned i = 0; i != f->vernauxs.size(); ++i) {
3210 if (f->vernauxs[i] == 0)
3211 continue;
3212 auto *verdef =
3213 reinterpret_cast<const typename ELFT::Verdef *>(f->verdefs[i]);
3214 vn.vernauxs.push_back(
3215 {verdef->vd_hash, f->vernauxs[i],
3216 getPartition().dynStrTab->addString(f->getStringTable().data() +
3217 verdef->getAux()->vda_name)});
3218 }
3219 }
3220
3221 if (OutputSection *sec = getPartition().dynStrTab->getParent())
3222 getParent()->link = sec->sectionIndex;
3223 getParent()->info = verneeds.size();
3224 }
3225
writeTo(uint8_t * buf)3226 template <class ELFT> void VersionNeedSection<ELFT>::writeTo(uint8_t *buf) {
3227 // The Elf_Verneeds need to appear first, followed by the Elf_Vernauxs.
3228 auto *verneed = reinterpret_cast<Elf_Verneed *>(buf);
3229 auto *vernaux = reinterpret_cast<Elf_Vernaux *>(verneed + verneeds.size());
3230
3231 for (auto &vn : verneeds) {
3232 // Create an Elf_Verneed for this DSO.
3233 verneed->vn_version = 1;
3234 verneed->vn_cnt = vn.vernauxs.size();
3235 verneed->vn_file = vn.nameStrTab;
3236 verneed->vn_aux =
3237 reinterpret_cast<char *>(vernaux) - reinterpret_cast<char *>(verneed);
3238 verneed->vn_next = sizeof(Elf_Verneed);
3239 ++verneed;
3240
3241 // Create the Elf_Vernauxs for this Elf_Verneed.
3242 for (auto &vna : vn.vernauxs) {
3243 vernaux->vna_hash = vna.hash;
3244 vernaux->vna_flags = 0;
3245 vernaux->vna_other = vna.verneedIndex;
3246 vernaux->vna_name = vna.nameStrTab;
3247 vernaux->vna_next = sizeof(Elf_Vernaux);
3248 ++vernaux;
3249 }
3250
3251 vernaux[-1].vna_next = 0;
3252 }
3253 verneed[-1].vn_next = 0;
3254 }
3255
getSize() const3256 template <class ELFT> size_t VersionNeedSection<ELFT>::getSize() const {
3257 return verneeds.size() * sizeof(Elf_Verneed) +
3258 SharedFile::vernauxNum * sizeof(Elf_Vernaux);
3259 }
3260
isNeeded() const3261 template <class ELFT> bool VersionNeedSection<ELFT>::isNeeded() const {
3262 return isLive() && SharedFile::vernauxNum != 0;
3263 }
3264
addSection(MergeInputSection * ms)3265 void MergeSyntheticSection::addSection(MergeInputSection *ms) {
3266 ms->parent = this;
3267 sections.push_back(ms);
3268 assert(alignment == ms->alignment || !(ms->flags & SHF_STRINGS));
3269 alignment = std::max(alignment, ms->alignment);
3270 }
3271
MergeTailSection(StringRef name,uint32_t type,uint64_t flags,uint32_t alignment)3272 MergeTailSection::MergeTailSection(StringRef name, uint32_t type,
3273 uint64_t flags, uint32_t alignment)
3274 : MergeSyntheticSection(name, type, flags, alignment),
3275 builder(StringTableBuilder::RAW, alignment) {}
3276
getSize() const3277 size_t MergeTailSection::getSize() const { return builder.getSize(); }
3278
writeTo(uint8_t * buf)3279 void MergeTailSection::writeTo(uint8_t *buf) { builder.write(buf); }
3280
finalizeContents()3281 void MergeTailSection::finalizeContents() {
3282 // Add all string pieces to the string table builder to create section
3283 // contents.
3284 for (MergeInputSection *sec : sections)
3285 for (size_t i = 0, e = sec->pieces.size(); i != e; ++i)
3286 if (sec->pieces[i].live)
3287 builder.add(sec->getData(i));
3288
3289 // Fix the string table content. After this, the contents will never change.
3290 builder.finalize();
3291
3292 // finalize() fixed tail-optimized strings, so we can now get
3293 // offsets of strings. Get an offset for each string and save it
3294 // to a corresponding SectionPiece for easy access.
3295 for (MergeInputSection *sec : sections)
3296 for (size_t i = 0, e = sec->pieces.size(); i != e; ++i)
3297 if (sec->pieces[i].live)
3298 sec->pieces[i].outputOff = builder.getOffset(sec->getData(i));
3299 }
3300
writeTo(uint8_t * buf)3301 void MergeNoTailSection::writeTo(uint8_t *buf) {
3302 for (size_t i = 0; i < numShards; ++i)
3303 shards[i].write(buf + shardOffsets[i]);
3304 }
3305
3306 // This function is very hot (i.e. it can take several seconds to finish)
3307 // because sometimes the number of inputs is in an order of magnitude of
3308 // millions. So, we use multi-threading.
3309 //
3310 // For any strings S and T, we know S is not mergeable with T if S's hash
3311 // value is different from T's. If that's the case, we can safely put S and
3312 // T into different string builders without worrying about merge misses.
3313 // We do it in parallel.
finalizeContents()3314 void MergeNoTailSection::finalizeContents() {
3315 // Initializes string table builders.
3316 for (size_t i = 0; i < numShards; ++i)
3317 shards.emplace_back(StringTableBuilder::RAW, alignment);
3318
3319 // Concurrency level. Must be a power of 2 to avoid expensive modulo
3320 // operations in the following tight loop.
3321 size_t concurrency = PowerOf2Floor(
3322 std::min<size_t>(hardware_concurrency(parallel::strategy.ThreadsRequested)
3323 .compute_thread_count(),
3324 numShards));
3325
3326 // Add section pieces to the builders.
3327 parallelForEachN(0, concurrency, [&](size_t threadId) {
3328 for (MergeInputSection *sec : sections) {
3329 for (size_t i = 0, e = sec->pieces.size(); i != e; ++i) {
3330 if (!sec->pieces[i].live)
3331 continue;
3332 size_t shardId = getShardId(sec->pieces[i].hash);
3333 if ((shardId & (concurrency - 1)) == threadId)
3334 sec->pieces[i].outputOff = shards[shardId].add(sec->getData(i));
3335 }
3336 }
3337 });
3338
3339 // Compute an in-section offset for each shard.
3340 size_t off = 0;
3341 for (size_t i = 0; i < numShards; ++i) {
3342 shards[i].finalizeInOrder();
3343 if (shards[i].getSize() > 0)
3344 off = alignTo(off, alignment);
3345 shardOffsets[i] = off;
3346 off += shards[i].getSize();
3347 }
3348 size = off;
3349
3350 // So far, section pieces have offsets from beginning of shards, but
3351 // we want offsets from beginning of the whole section. Fix them.
3352 parallelForEach(sections, [&](MergeInputSection *sec) {
3353 for (size_t i = 0, e = sec->pieces.size(); i != e; ++i)
3354 if (sec->pieces[i].live)
3355 sec->pieces[i].outputOff +=
3356 shardOffsets[getShardId(sec->pieces[i].hash)];
3357 });
3358 }
3359
createMergeSynthetic(StringRef name,uint32_t type,uint64_t flags,uint32_t alignment)3360 MergeSyntheticSection *elf::createMergeSynthetic(StringRef name, uint32_t type,
3361 uint64_t flags,
3362 uint32_t alignment) {
3363 bool shouldTailMerge = (flags & SHF_STRINGS) && config->optimize >= 2;
3364 if (shouldTailMerge)
3365 return make<MergeTailSection>(name, type, flags, alignment);
3366 return make<MergeNoTailSection>(name, type, flags, alignment);
3367 }
3368
splitSections()3369 template <class ELFT> void elf::splitSections() {
3370 llvm::TimeTraceScope timeScope("Split sections");
3371 // splitIntoPieces needs to be called on each MergeInputSection
3372 // before calling finalizeContents().
3373 parallelForEach(inputSections, [](InputSectionBase *sec) {
3374 if (auto *s = dyn_cast<MergeInputSection>(sec))
3375 s->splitIntoPieces();
3376 else if (auto *eh = dyn_cast<EhInputSection>(sec))
3377 eh->split<ELFT>();
3378 });
3379 }
3380
MipsRldMapSection()3381 MipsRldMapSection::MipsRldMapSection()
3382 : SyntheticSection(SHF_ALLOC | SHF_WRITE, SHT_PROGBITS, config->wordsize,
3383 ".rld_map") {}
3384
ARMExidxSyntheticSection()3385 ARMExidxSyntheticSection::ARMExidxSyntheticSection()
3386 : SyntheticSection(SHF_ALLOC | SHF_LINK_ORDER, SHT_ARM_EXIDX,
3387 config->wordsize, ".ARM.exidx") {}
3388
findExidxSection(InputSection * isec)3389 static InputSection *findExidxSection(InputSection *isec) {
3390 for (InputSection *d : isec->dependentSections)
3391 if (d->type == SHT_ARM_EXIDX && d->isLive())
3392 return d;
3393 return nullptr;
3394 }
3395
isValidExidxSectionDep(InputSection * isec)3396 static bool isValidExidxSectionDep(InputSection *isec) {
3397 return (isec->flags & SHF_ALLOC) && (isec->flags & SHF_EXECINSTR) &&
3398 isec->getSize() > 0;
3399 }
3400
addSection(InputSection * isec)3401 bool ARMExidxSyntheticSection::addSection(InputSection *isec) {
3402 if (isec->type == SHT_ARM_EXIDX) {
3403 if (InputSection *dep = isec->getLinkOrderDep())
3404 if (isValidExidxSectionDep(dep)) {
3405 exidxSections.push_back(isec);
3406 // Every exidxSection is 8 bytes, we need an estimate of
3407 // size before assignAddresses can be called. Final size
3408 // will only be known after finalize is called.
3409 size += 8;
3410 }
3411 return true;
3412 }
3413
3414 if (isValidExidxSectionDep(isec)) {
3415 executableSections.push_back(isec);
3416 return false;
3417 }
3418
3419 // FIXME: we do not output a relocation section when --emit-relocs is used
3420 // as we do not have relocation sections for linker generated table entries
3421 // and we would have to erase at a late stage relocations from merged entries.
3422 // Given that exception tables are already position independent and a binary
3423 // analyzer could derive the relocations we choose to erase the relocations.
3424 if (config->emitRelocs && isec->type == SHT_REL)
3425 if (InputSectionBase *ex = isec->getRelocatedSection())
3426 if (isa<InputSection>(ex) && ex->type == SHT_ARM_EXIDX)
3427 return true;
3428
3429 return false;
3430 }
3431
3432 // References to .ARM.Extab Sections have bit 31 clear and are not the
3433 // special EXIDX_CANTUNWIND bit-pattern.
isExtabRef(uint32_t unwind)3434 static bool isExtabRef(uint32_t unwind) {
3435 return (unwind & 0x80000000) == 0 && unwind != 0x1;
3436 }
3437
3438 // Return true if the .ARM.exidx section Cur can be merged into the .ARM.exidx
3439 // section Prev, where Cur follows Prev in the table. This can be done if the
3440 // unwinding instructions in Cur are identical to Prev. Linker generated
3441 // EXIDX_CANTUNWIND entries are represented by nullptr as they do not have an
3442 // InputSection.
isDuplicateArmExidxSec(InputSection * prev,InputSection * cur)3443 static bool isDuplicateArmExidxSec(InputSection *prev, InputSection *cur) {
3444
3445 struct ExidxEntry {
3446 ulittle32_t fn;
3447 ulittle32_t unwind;
3448 };
3449 // Get the last table Entry from the previous .ARM.exidx section. If Prev is
3450 // nullptr then it will be a synthesized EXIDX_CANTUNWIND entry.
3451 ExidxEntry prevEntry = {ulittle32_t(0), ulittle32_t(1)};
3452 if (prev)
3453 prevEntry = prev->getDataAs<ExidxEntry>().back();
3454 if (isExtabRef(prevEntry.unwind))
3455 return false;
3456
3457 // We consider the unwind instructions of an .ARM.exidx table entry
3458 // a duplicate if the previous unwind instructions if:
3459 // - Both are the special EXIDX_CANTUNWIND.
3460 // - Both are the same inline unwind instructions.
3461 // We do not attempt to follow and check links into .ARM.extab tables as
3462 // consecutive identical entries are rare and the effort to check that they
3463 // are identical is high.
3464
3465 // If Cur is nullptr then this is synthesized EXIDX_CANTUNWIND entry.
3466 if (cur == nullptr)
3467 return prevEntry.unwind == 1;
3468
3469 for (const ExidxEntry entry : cur->getDataAs<ExidxEntry>())
3470 if (isExtabRef(entry.unwind) || entry.unwind != prevEntry.unwind)
3471 return false;
3472
3473 // All table entries in this .ARM.exidx Section can be merged into the
3474 // previous Section.
3475 return true;
3476 }
3477
3478 // The .ARM.exidx table must be sorted in ascending order of the address of the
3479 // functions the table describes. Optionally duplicate adjacent table entries
3480 // can be removed. At the end of the function the executableSections must be
3481 // sorted in ascending order of address, Sentinel is set to the InputSection
3482 // with the highest address and any InputSections that have mergeable
3483 // .ARM.exidx table entries are removed from it.
finalizeContents()3484 void ARMExidxSyntheticSection::finalizeContents() {
3485 // The executableSections and exidxSections that we use to derive the final
3486 // contents of this SyntheticSection are populated before
3487 // processSectionCommands() and ICF. A /DISCARD/ entry in SECTIONS command or
3488 // ICF may remove executable InputSections and their dependent .ARM.exidx
3489 // section that we recorded earlier.
3490 auto isDiscarded = [](const InputSection *isec) { return !isec->isLive(); };
3491 llvm::erase_if(exidxSections, isDiscarded);
3492 // We need to remove discarded InputSections and InputSections without
3493 // .ARM.exidx sections that if we generated the .ARM.exidx it would be out
3494 // of range.
3495 auto isDiscardedOrOutOfRange = [this](InputSection *isec) {
3496 if (!isec->isLive())
3497 return true;
3498 if (findExidxSection(isec))
3499 return false;
3500 int64_t off = static_cast<int64_t>(isec->getVA() - getVA());
3501 return off != llvm::SignExtend64(off, 31);
3502 };
3503 llvm::erase_if(executableSections, isDiscardedOrOutOfRange);
3504
3505 // Sort the executable sections that may or may not have associated
3506 // .ARM.exidx sections by order of ascending address. This requires the
3507 // relative positions of InputSections and OutputSections to be known.
3508 auto compareByFilePosition = [](const InputSection *a,
3509 const InputSection *b) {
3510 OutputSection *aOut = a->getParent();
3511 OutputSection *bOut = b->getParent();
3512
3513 if (aOut != bOut)
3514 return aOut->addr < bOut->addr;
3515 return a->outSecOff < b->outSecOff;
3516 };
3517 llvm::stable_sort(executableSections, compareByFilePosition);
3518 sentinel = executableSections.back();
3519 // Optionally merge adjacent duplicate entries.
3520 if (config->mergeArmExidx) {
3521 std::vector<InputSection *> selectedSections;
3522 selectedSections.reserve(executableSections.size());
3523 selectedSections.push_back(executableSections[0]);
3524 size_t prev = 0;
3525 for (size_t i = 1; i < executableSections.size(); ++i) {
3526 InputSection *ex1 = findExidxSection(executableSections[prev]);
3527 InputSection *ex2 = findExidxSection(executableSections[i]);
3528 if (!isDuplicateArmExidxSec(ex1, ex2)) {
3529 selectedSections.push_back(executableSections[i]);
3530 prev = i;
3531 }
3532 }
3533 executableSections = std::move(selectedSections);
3534 }
3535
3536 size_t offset = 0;
3537 size = 0;
3538 for (InputSection *isec : executableSections) {
3539 if (InputSection *d = findExidxSection(isec)) {
3540 d->outSecOff = offset;
3541 d->parent = getParent();
3542 offset += d->getSize();
3543 } else {
3544 offset += 8;
3545 }
3546 }
3547 // Size includes Sentinel.
3548 size = offset + 8;
3549 }
3550
getLinkOrderDep() const3551 InputSection *ARMExidxSyntheticSection::getLinkOrderDep() const {
3552 return executableSections.front();
3553 }
3554
3555 // To write the .ARM.exidx table from the ExecutableSections we have three cases
3556 // 1.) The InputSection has a .ARM.exidx InputSection in its dependent sections.
3557 // We write the .ARM.exidx section contents and apply its relocations.
3558 // 2.) The InputSection does not have a dependent .ARM.exidx InputSection. We
3559 // must write the contents of an EXIDX_CANTUNWIND directly. We use the
3560 // start of the InputSection as the purpose of the linker generated
3561 // section is to terminate the address range of the previous entry.
3562 // 3.) A trailing EXIDX_CANTUNWIND sentinel section is required at the end of
3563 // the table to terminate the address range of the final entry.
writeTo(uint8_t * buf)3564 void ARMExidxSyntheticSection::writeTo(uint8_t *buf) {
3565
3566 const uint8_t cantUnwindData[8] = {0, 0, 0, 0, // PREL31 to target
3567 1, 0, 0, 0}; // EXIDX_CANTUNWIND
3568
3569 uint64_t offset = 0;
3570 for (InputSection *isec : executableSections) {
3571 assert(isec->getParent() != nullptr);
3572 if (InputSection *d = findExidxSection(isec)) {
3573 memcpy(buf + offset, d->data().data(), d->data().size());
3574 d->relocateAlloc(buf + d->outSecOff, buf + d->outSecOff + d->getSize());
3575 offset += d->getSize();
3576 } else {
3577 // A Linker generated CANTUNWIND section.
3578 memcpy(buf + offset, cantUnwindData, sizeof(cantUnwindData));
3579 uint64_t s = isec->getVA();
3580 uint64_t p = getVA() + offset;
3581 target->relocateNoSym(buf + offset, R_ARM_PREL31, s - p);
3582 offset += 8;
3583 }
3584 }
3585 // Write Sentinel.
3586 memcpy(buf + offset, cantUnwindData, sizeof(cantUnwindData));
3587 uint64_t s = sentinel->getVA(sentinel->getSize());
3588 uint64_t p = getVA() + offset;
3589 target->relocateNoSym(buf + offset, R_ARM_PREL31, s - p);
3590 assert(size == offset + 8);
3591 }
3592
isNeeded() const3593 bool ARMExidxSyntheticSection::isNeeded() const {
3594 return llvm::find_if(exidxSections, [](InputSection *isec) {
3595 return isec->isLive();
3596 }) != exidxSections.end();
3597 }
3598
classof(const SectionBase * d)3599 bool ARMExidxSyntheticSection::classof(const SectionBase *d) {
3600 return d->kind() == InputSectionBase::Synthetic && d->type == SHT_ARM_EXIDX;
3601 }
3602
ThunkSection(OutputSection * os,uint64_t off)3603 ThunkSection::ThunkSection(OutputSection *os, uint64_t off)
3604 : SyntheticSection(SHF_ALLOC | SHF_EXECINSTR, SHT_PROGBITS,
3605 config->emachine == EM_PPC64 ? 16 : 4, ".text.thunk") {
3606 this->parent = os;
3607 this->outSecOff = off;
3608 }
3609
getSize() const3610 size_t ThunkSection::getSize() const {
3611 if (roundUpSizeForErrata)
3612 return alignTo(size, 4096);
3613 return size;
3614 }
3615
addThunk(Thunk * t)3616 void ThunkSection::addThunk(Thunk *t) {
3617 thunks.push_back(t);
3618 t->addSymbols(*this);
3619 }
3620
writeTo(uint8_t * buf)3621 void ThunkSection::writeTo(uint8_t *buf) {
3622 for (Thunk *t : thunks)
3623 t->writeTo(buf + t->offset);
3624 }
3625
getTargetInputSection() const3626 InputSection *ThunkSection::getTargetInputSection() const {
3627 if (thunks.empty())
3628 return nullptr;
3629 const Thunk *t = thunks.front();
3630 return t->getTargetInputSection();
3631 }
3632
assignOffsets()3633 bool ThunkSection::assignOffsets() {
3634 uint64_t off = 0;
3635 for (Thunk *t : thunks) {
3636 off = alignTo(off, t->alignment);
3637 t->setOffset(off);
3638 uint32_t size = t->size();
3639 t->getThunkTargetSym()->size = size;
3640 off += size;
3641 }
3642 bool changed = off != size;
3643 size = off;
3644 return changed;
3645 }
3646
PPC32Got2Section()3647 PPC32Got2Section::PPC32Got2Section()
3648 : SyntheticSection(SHF_ALLOC | SHF_WRITE, SHT_PROGBITS, 4, ".got2") {}
3649
isNeeded() const3650 bool PPC32Got2Section::isNeeded() const {
3651 // See the comment below. This is not needed if there is no other
3652 // InputSection.
3653 for (BaseCommand *base : getParent()->sectionCommands)
3654 if (auto *isd = dyn_cast<InputSectionDescription>(base))
3655 for (InputSection *isec : isd->sections)
3656 if (isec != this)
3657 return true;
3658 return false;
3659 }
3660
finalizeContents()3661 void PPC32Got2Section::finalizeContents() {
3662 // PPC32 may create multiple GOT sections for -fPIC/-fPIE, one per file in
3663 // .got2 . This function computes outSecOff of each .got2 to be used in
3664 // PPC32PltCallStub::writeTo(). The purpose of this empty synthetic section is
3665 // to collect input sections named ".got2".
3666 uint32_t offset = 0;
3667 for (BaseCommand *base : getParent()->sectionCommands)
3668 if (auto *isd = dyn_cast<InputSectionDescription>(base)) {
3669 for (InputSection *isec : isd->sections) {
3670 if (isec == this)
3671 continue;
3672 isec->file->ppc32Got2OutSecOff = offset;
3673 offset += (uint32_t)isec->getSize();
3674 }
3675 }
3676 }
3677
3678 // If linking position-dependent code then the table will store the addresses
3679 // directly in the binary so the section has type SHT_PROGBITS. If linking
3680 // position-independent code the section has type SHT_NOBITS since it will be
3681 // allocated and filled in by the dynamic linker.
PPC64LongBranchTargetSection()3682 PPC64LongBranchTargetSection::PPC64LongBranchTargetSection()
3683 : SyntheticSection(SHF_ALLOC | SHF_WRITE,
3684 config->isPic ? SHT_NOBITS : SHT_PROGBITS, 8,
3685 ".branch_lt") {}
3686
getEntryVA(const Symbol * sym,int64_t addend)3687 uint64_t PPC64LongBranchTargetSection::getEntryVA(const Symbol *sym,
3688 int64_t addend) {
3689 return getVA() + entry_index.find({sym, addend})->second * 8;
3690 }
3691
addEntry(const Symbol * sym,int64_t addend)3692 Optional<uint32_t> PPC64LongBranchTargetSection::addEntry(const Symbol *sym,
3693 int64_t addend) {
3694 auto res =
3695 entry_index.try_emplace(std::make_pair(sym, addend), entries.size());
3696 if (!res.second)
3697 return None;
3698 entries.emplace_back(sym, addend);
3699 return res.first->second;
3700 }
3701
getSize() const3702 size_t PPC64LongBranchTargetSection::getSize() const {
3703 return entries.size() * 8;
3704 }
3705
writeTo(uint8_t * buf)3706 void PPC64LongBranchTargetSection::writeTo(uint8_t *buf) {
3707 // If linking non-pic we have the final addresses of the targets and they get
3708 // written to the table directly. For pic the dynamic linker will allocate
3709 // the section and fill it it.
3710 if (config->isPic)
3711 return;
3712
3713 for (auto entry : entries) {
3714 const Symbol *sym = entry.first;
3715 int64_t addend = entry.second;
3716 assert(sym->getVA());
3717 // Need calls to branch to the local entry-point since a long-branch
3718 // must be a local-call.
3719 write64(buf, sym->getVA(addend) +
3720 getPPC64GlobalEntryToLocalEntryOffset(sym->stOther));
3721 buf += 8;
3722 }
3723 }
3724
isNeeded() const3725 bool PPC64LongBranchTargetSection::isNeeded() const {
3726 // `removeUnusedSyntheticSections()` is called before thunk allocation which
3727 // is too early to determine if this section will be empty or not. We need
3728 // Finalized to keep the section alive until after thunk creation. Finalized
3729 // only gets set to true once `finalizeSections()` is called after thunk
3730 // creation. Because of this, if we don't create any long-branch thunks we end
3731 // up with an empty .branch_lt section in the binary.
3732 return !finalized || !entries.empty();
3733 }
3734
getAbiVersion()3735 static uint8_t getAbiVersion() {
3736 // MIPS non-PIC executable gets ABI version 1.
3737 if (config->emachine == EM_MIPS) {
3738 if (!config->isPic && !config->relocatable &&
3739 (config->eflags & (EF_MIPS_PIC | EF_MIPS_CPIC)) == EF_MIPS_CPIC)
3740 return 1;
3741 return 0;
3742 }
3743
3744 if (config->emachine == EM_AMDGPU) {
3745 uint8_t ver = objectFiles[0]->abiVersion;
3746 for (InputFile *file : makeArrayRef(objectFiles).slice(1))
3747 if (file->abiVersion != ver)
3748 error("incompatible ABI version: " + toString(file));
3749 return ver;
3750 }
3751
3752 return 0;
3753 }
3754
writeEhdr(uint8_t * buf,Partition & part)3755 template <typename ELFT> void elf::writeEhdr(uint8_t *buf, Partition &part) {
3756 // For executable segments, the trap instructions are written before writing
3757 // the header. Setting Elf header bytes to zero ensures that any unused bytes
3758 // in header are zero-cleared, instead of having trap instructions.
3759 memset(buf, 0, sizeof(typename ELFT::Ehdr));
3760 memcpy(buf, "\177ELF", 4);
3761
3762 auto *eHdr = reinterpret_cast<typename ELFT::Ehdr *>(buf);
3763 eHdr->e_ident[EI_CLASS] = config->is64 ? ELFCLASS64 : ELFCLASS32;
3764 eHdr->e_ident[EI_DATA] = config->isLE ? ELFDATA2LSB : ELFDATA2MSB;
3765 eHdr->e_ident[EI_VERSION] = EV_CURRENT;
3766 eHdr->e_ident[EI_OSABI] = config->osabi;
3767 eHdr->e_ident[EI_ABIVERSION] = getAbiVersion();
3768 eHdr->e_machine = config->emachine;
3769 eHdr->e_version = EV_CURRENT;
3770 eHdr->e_flags = config->eflags;
3771 eHdr->e_ehsize = sizeof(typename ELFT::Ehdr);
3772 eHdr->e_phnum = part.phdrs.size();
3773 eHdr->e_shentsize = sizeof(typename ELFT::Shdr);
3774
3775 if (!config->relocatable) {
3776 eHdr->e_phoff = sizeof(typename ELFT::Ehdr);
3777 eHdr->e_phentsize = sizeof(typename ELFT::Phdr);
3778 }
3779 }
3780
writePhdrs(uint8_t * buf,Partition & part)3781 template <typename ELFT> void elf::writePhdrs(uint8_t *buf, Partition &part) {
3782 // Write the program header table.
3783 auto *hBuf = reinterpret_cast<typename ELFT::Phdr *>(buf);
3784 for (PhdrEntry *p : part.phdrs) {
3785 hBuf->p_type = p->p_type;
3786 hBuf->p_flags = p->p_flags;
3787 hBuf->p_offset = p->p_offset;
3788 hBuf->p_vaddr = p->p_vaddr;
3789 hBuf->p_paddr = p->p_paddr;
3790 hBuf->p_filesz = p->p_filesz;
3791 hBuf->p_memsz = p->p_memsz;
3792 hBuf->p_align = p->p_align;
3793 ++hBuf;
3794 }
3795 }
3796
3797 template <typename ELFT>
PartitionElfHeaderSection()3798 PartitionElfHeaderSection<ELFT>::PartitionElfHeaderSection()
3799 : SyntheticSection(SHF_ALLOC, SHT_LLVM_PART_EHDR, 1, "") {}
3800
3801 template <typename ELFT>
getSize() const3802 size_t PartitionElfHeaderSection<ELFT>::getSize() const {
3803 return sizeof(typename ELFT::Ehdr);
3804 }
3805
3806 template <typename ELFT>
writeTo(uint8_t * buf)3807 void PartitionElfHeaderSection<ELFT>::writeTo(uint8_t *buf) {
3808 writeEhdr<ELFT>(buf, getPartition());
3809
3810 // Loadable partitions are always ET_DYN.
3811 auto *eHdr = reinterpret_cast<typename ELFT::Ehdr *>(buf);
3812 eHdr->e_type = ET_DYN;
3813 }
3814
3815 template <typename ELFT>
PartitionProgramHeadersSection()3816 PartitionProgramHeadersSection<ELFT>::PartitionProgramHeadersSection()
3817 : SyntheticSection(SHF_ALLOC, SHT_LLVM_PART_PHDR, 1, ".phdrs") {}
3818
3819 template <typename ELFT>
getSize() const3820 size_t PartitionProgramHeadersSection<ELFT>::getSize() const {
3821 return sizeof(typename ELFT::Phdr) * getPartition().phdrs.size();
3822 }
3823
3824 template <typename ELFT>
writeTo(uint8_t * buf)3825 void PartitionProgramHeadersSection<ELFT>::writeTo(uint8_t *buf) {
3826 writePhdrs<ELFT>(buf, getPartition());
3827 }
3828
PartitionIndexSection()3829 PartitionIndexSection::PartitionIndexSection()
3830 : SyntheticSection(SHF_ALLOC, SHT_PROGBITS, 4, ".rodata") {}
3831
getSize() const3832 size_t PartitionIndexSection::getSize() const {
3833 return 12 * (partitions.size() - 1);
3834 }
3835
finalizeContents()3836 void PartitionIndexSection::finalizeContents() {
3837 for (size_t i = 1; i != partitions.size(); ++i)
3838 partitions[i].nameStrTab = mainPart->dynStrTab->addString(partitions[i].name);
3839 }
3840
writeTo(uint8_t * buf)3841 void PartitionIndexSection::writeTo(uint8_t *buf) {
3842 uint64_t va = getVA();
3843 for (size_t i = 1; i != partitions.size(); ++i) {
3844 write32(buf, mainPart->dynStrTab->getVA() + partitions[i].nameStrTab - va);
3845 write32(buf + 4, partitions[i].elfHeader->getVA() - (va + 4));
3846
3847 SyntheticSection *next =
3848 i == partitions.size() - 1 ? in.partEnd : partitions[i + 1].elfHeader;
3849 write32(buf + 8, next->getVA() - partitions[i].elfHeader->getVA());
3850
3851 va += 12;
3852 buf += 12;
3853 }
3854 }
3855
3856 InStruct elf::in;
3857
3858 std::vector<Partition> elf::partitions;
3859 Partition *elf::mainPart;
3860
3861 template GdbIndexSection *GdbIndexSection::create<ELF32LE>();
3862 template GdbIndexSection *GdbIndexSection::create<ELF32BE>();
3863 template GdbIndexSection *GdbIndexSection::create<ELF64LE>();
3864 template GdbIndexSection *GdbIndexSection::create<ELF64BE>();
3865
3866 template void elf::splitSections<ELF32LE>();
3867 template void elf::splitSections<ELF32BE>();
3868 template void elf::splitSections<ELF64LE>();
3869 template void elf::splitSections<ELF64BE>();
3870
3871 template class elf::MipsAbiFlagsSection<ELF32LE>;
3872 template class elf::MipsAbiFlagsSection<ELF32BE>;
3873 template class elf::MipsAbiFlagsSection<ELF64LE>;
3874 template class elf::MipsAbiFlagsSection<ELF64BE>;
3875
3876 template class elf::MipsOptionsSection<ELF32LE>;
3877 template class elf::MipsOptionsSection<ELF32BE>;
3878 template class elf::MipsOptionsSection<ELF64LE>;
3879 template class elf::MipsOptionsSection<ELF64BE>;
3880
3881 template void EhFrameSection::iterateFDEWithLSDA<ELF32LE>(
3882 function_ref<void(InputSection &)>);
3883 template void EhFrameSection::iterateFDEWithLSDA<ELF32BE>(
3884 function_ref<void(InputSection &)>);
3885 template void EhFrameSection::iterateFDEWithLSDA<ELF64LE>(
3886 function_ref<void(InputSection &)>);
3887 template void EhFrameSection::iterateFDEWithLSDA<ELF64BE>(
3888 function_ref<void(InputSection &)>);
3889
3890 template class elf::MipsReginfoSection<ELF32LE>;
3891 template class elf::MipsReginfoSection<ELF32BE>;
3892 template class elf::MipsReginfoSection<ELF64LE>;
3893 template class elf::MipsReginfoSection<ELF64BE>;
3894
3895 template class elf::DynamicSection<ELF32LE>;
3896 template class elf::DynamicSection<ELF32BE>;
3897 template class elf::DynamicSection<ELF64LE>;
3898 template class elf::DynamicSection<ELF64BE>;
3899
3900 template class elf::RelocationSection<ELF32LE>;
3901 template class elf::RelocationSection<ELF32BE>;
3902 template class elf::RelocationSection<ELF64LE>;
3903 template class elf::RelocationSection<ELF64BE>;
3904
3905 template class elf::AndroidPackedRelocationSection<ELF32LE>;
3906 template class elf::AndroidPackedRelocationSection<ELF32BE>;
3907 template class elf::AndroidPackedRelocationSection<ELF64LE>;
3908 template class elf::AndroidPackedRelocationSection<ELF64BE>;
3909
3910 template class elf::RelrSection<ELF32LE>;
3911 template class elf::RelrSection<ELF32BE>;
3912 template class elf::RelrSection<ELF64LE>;
3913 template class elf::RelrSection<ELF64BE>;
3914
3915 template class elf::SymbolTableSection<ELF32LE>;
3916 template class elf::SymbolTableSection<ELF32BE>;
3917 template class elf::SymbolTableSection<ELF64LE>;
3918 template class elf::SymbolTableSection<ELF64BE>;
3919
3920 template class elf::VersionNeedSection<ELF32LE>;
3921 template class elf::VersionNeedSection<ELF32BE>;
3922 template class elf::VersionNeedSection<ELF64LE>;
3923 template class elf::VersionNeedSection<ELF64BE>;
3924
3925 template void elf::writeEhdr<ELF32LE>(uint8_t *Buf, Partition &Part);
3926 template void elf::writeEhdr<ELF32BE>(uint8_t *Buf, Partition &Part);
3927 template void elf::writeEhdr<ELF64LE>(uint8_t *Buf, Partition &Part);
3928 template void elf::writeEhdr<ELF64BE>(uint8_t *Buf, Partition &Part);
3929
3930 template void elf::writePhdrs<ELF32LE>(uint8_t *Buf, Partition &Part);
3931 template void elf::writePhdrs<ELF32BE>(uint8_t *Buf, Partition &Part);
3932 template void elf::writePhdrs<ELF64LE>(uint8_t *Buf, Partition &Part);
3933 template void elf::writePhdrs<ELF64BE>(uint8_t *Buf, Partition &Part);
3934
3935 template class elf::PartitionElfHeaderSection<ELF32LE>;
3936 template class elf::PartitionElfHeaderSection<ELF32BE>;
3937 template class elf::PartitionElfHeaderSection<ELF64LE>;
3938 template class elf::PartitionElfHeaderSection<ELF64BE>;
3939
3940 template class elf::PartitionProgramHeadersSection<ELF32LE>;
3941 template class elf::PartitionProgramHeadersSection<ELF32BE>;
3942 template class elf::PartitionProgramHeadersSection<ELF64LE>;
3943 template class elf::PartitionProgramHeadersSection<ELF64BE>;
3944